2006-04-07 06:13:41 +04:00
|
|
|
/*
|
|
|
|
* iSCSI lib functions
|
|
|
|
*
|
|
|
|
* Copyright (C) 2006 Red Hat, Inc. All rights reserved.
|
|
|
|
* Copyright (C) 2004 - 2006 Mike Christie
|
|
|
|
* Copyright (C) 2004 - 2005 Dmitry Yusupov
|
|
|
|
* Copyright (C) 2004 - 2005 Alex Aizman
|
|
|
|
* maintained by open-iscsi@googlegroups.com
|
|
|
|
*
|
|
|
|
* This program is free software; you can redistribute it and/or modify
|
|
|
|
* it under the terms of the GNU General Public License as published by
|
|
|
|
* the Free Software Foundation; either version 2 of the License, or
|
|
|
|
* (at your option) any later version.
|
|
|
|
*
|
|
|
|
* This program is distributed in the hope that it will be useful,
|
|
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
|
|
* GNU General Public License for more details.
|
|
|
|
*
|
|
|
|
* You should have received a copy of the GNU General Public License
|
|
|
|
* along with this program; if not, write to the Free Software
|
|
|
|
* Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
|
|
|
|
*/
|
|
|
|
#include <linux/types.h>
|
|
|
|
#include <linux/kfifo.h>
|
|
|
|
#include <linux/delay.h>
|
2007-12-13 21:43:41 +03:00
|
|
|
#include <linux/log2.h>
|
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 11:04:11 +03:00
|
|
|
#include <linux/slab.h>
|
2011-05-27 17:47:43 +04:00
|
|
|
#include <linux/module.h>
|
2007-03-01 02:32:19 +03:00
|
|
|
#include <asm/unaligned.h>
|
2006-04-07 06:13:41 +04:00
|
|
|
#include <net/tcp.h>
|
|
|
|
#include <scsi/scsi_cmnd.h>
|
|
|
|
#include <scsi/scsi_device.h>
|
|
|
|
#include <scsi/scsi_eh.h>
|
|
|
|
#include <scsi/scsi_tcq.h>
|
|
|
|
#include <scsi/scsi_host.h>
|
|
|
|
#include <scsi/scsi.h>
|
|
|
|
#include <scsi/iscsi_proto.h>
|
|
|
|
#include <scsi/scsi_transport.h>
|
|
|
|
#include <scsi/scsi_transport_iscsi.h>
|
|
|
|
#include <scsi/libiscsi.h>
|
|
|
|
|
2009-06-16 07:11:10 +04:00
|
|
|
static int iscsi_dbg_lib_conn;
|
|
|
|
module_param_named(debug_libiscsi_conn, iscsi_dbg_lib_conn, int,
|
|
|
|
S_IRUGO | S_IWUSR);
|
|
|
|
MODULE_PARM_DESC(debug_libiscsi_conn,
|
|
|
|
"Turn on debugging for connections in libiscsi module. "
|
|
|
|
"Set to 1 to turn on, and zero to turn off. Default is off.");
|
|
|
|
|
|
|
|
static int iscsi_dbg_lib_session;
|
|
|
|
module_param_named(debug_libiscsi_session, iscsi_dbg_lib_session, int,
|
|
|
|
S_IRUGO | S_IWUSR);
|
|
|
|
MODULE_PARM_DESC(debug_libiscsi_session,
|
|
|
|
"Turn on debugging for sessions in libiscsi module. "
|
|
|
|
"Set to 1 to turn on, and zero to turn off. Default is off.");
|
|
|
|
|
|
|
|
static int iscsi_dbg_lib_eh;
|
|
|
|
module_param_named(debug_libiscsi_eh, iscsi_dbg_lib_eh, int,
|
|
|
|
S_IRUGO | S_IWUSR);
|
|
|
|
MODULE_PARM_DESC(debug_libiscsi_eh,
|
|
|
|
"Turn on debugging for error handling in libiscsi module. "
|
|
|
|
"Set to 1 to turn on, and zero to turn off. Default is off.");
|
2009-03-05 23:45:58 +03:00
|
|
|
|
|
|
|
#define ISCSI_DBG_CONN(_conn, dbg_fmt, arg...) \
|
|
|
|
do { \
|
2009-06-16 07:11:10 +04:00
|
|
|
if (iscsi_dbg_lib_conn) \
|
2009-03-05 23:45:58 +03:00
|
|
|
iscsi_conn_printk(KERN_INFO, _conn, \
|
|
|
|
"%s " dbg_fmt, \
|
|
|
|
__func__, ##arg); \
|
|
|
|
} while (0);
|
|
|
|
|
|
|
|
#define ISCSI_DBG_SESSION(_session, dbg_fmt, arg...) \
|
|
|
|
do { \
|
2009-06-16 07:11:10 +04:00
|
|
|
if (iscsi_dbg_lib_session) \
|
|
|
|
iscsi_session_printk(KERN_INFO, _session, \
|
|
|
|
"%s " dbg_fmt, \
|
|
|
|
__func__, ##arg); \
|
|
|
|
} while (0);
|
|
|
|
|
|
|
|
#define ISCSI_DBG_EH(_session, dbg_fmt, arg...) \
|
|
|
|
do { \
|
|
|
|
if (iscsi_dbg_lib_eh) \
|
2009-03-05 23:45:58 +03:00
|
|
|
iscsi_session_printk(KERN_INFO, _session, \
|
|
|
|
"%s " dbg_fmt, \
|
|
|
|
__func__, ##arg); \
|
|
|
|
} while (0);
|
|
|
|
|
2009-03-05 23:46:03 +03:00
|
|
|
inline void iscsi_conn_queue_work(struct iscsi_conn *conn)
|
|
|
|
{
|
|
|
|
struct Scsi_Host *shost = conn->session->host;
|
|
|
|
struct iscsi_host *ihost = shost_priv(shost);
|
|
|
|
|
2009-05-14 02:57:48 +04:00
|
|
|
if (ihost->workq)
|
|
|
|
queue_work(ihost->workq, &conn->xmitwork);
|
2009-03-05 23:46:03 +03:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_conn_queue_work);
|
|
|
|
|
2009-09-05 06:04:23 +04:00
|
|
|
static void __iscsi_update_cmdsn(struct iscsi_session *session,
|
|
|
|
uint32_t exp_cmdsn, uint32_t max_cmdsn)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
2007-05-30 21:57:18 +04:00
|
|
|
/*
|
|
|
|
* standard specifies this check for when to update expected and
|
|
|
|
* max sequence numbers
|
|
|
|
*/
|
|
|
|
if (iscsi_sna_lt(max_cmdsn, exp_cmdsn - 1))
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (exp_cmdsn != session->exp_cmdsn &&
|
|
|
|
!iscsi_sna_lt(exp_cmdsn, session->exp_cmdsn))
|
2006-04-07 06:13:41 +04:00
|
|
|
session->exp_cmdsn = exp_cmdsn;
|
|
|
|
|
2007-05-30 21:57:18 +04:00
|
|
|
if (max_cmdsn != session->max_cmdsn &&
|
2014-02-07 10:41:39 +04:00
|
|
|
!iscsi_sna_lt(max_cmdsn, session->max_cmdsn))
|
2007-05-30 21:57:18 +04:00
|
|
|
session->max_cmdsn = max_cmdsn;
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
2009-09-05 06:04:23 +04:00
|
|
|
|
|
|
|
void iscsi_update_cmdsn(struct iscsi_session *session, struct iscsi_nopin *hdr)
|
|
|
|
{
|
|
|
|
__iscsi_update_cmdsn(session, be32_to_cpu(hdr->exp_cmdsn),
|
|
|
|
be32_to_cpu(hdr->max_cmdsn));
|
|
|
|
}
|
2007-05-30 21:57:18 +04:00
|
|
|
EXPORT_SYMBOL_GPL(iscsi_update_cmdsn);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-12-02 09:32:05 +03:00
|
|
|
/**
|
|
|
|
* iscsi_prep_data_out_pdu - initialize Data-Out
|
|
|
|
* @task: scsi command task
|
|
|
|
* @r2t: R2T info
|
|
|
|
* @hdr: iscsi data in pdu
|
|
|
|
*
|
|
|
|
* Notes:
|
|
|
|
* Initialize Data-Out within this R2T sequence and finds
|
|
|
|
* proper data_offset within this SCSI command.
|
|
|
|
*
|
|
|
|
* This function is called with connection lock taken.
|
|
|
|
**/
|
|
|
|
void iscsi_prep_data_out_pdu(struct iscsi_task *task, struct iscsi_r2t_info *r2t,
|
|
|
|
struct iscsi_data *hdr)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_conn *conn = task->conn;
|
2008-12-02 09:32:05 +03:00
|
|
|
unsigned int left = r2t->data_length - r2t->sent;
|
|
|
|
|
|
|
|
task->hdr_len = sizeof(struct iscsi_data);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
|
|
|
memset(hdr, 0, sizeof(struct iscsi_data));
|
2008-12-02 09:32:05 +03:00
|
|
|
hdr->ttt = r2t->ttt;
|
|
|
|
hdr->datasn = cpu_to_be32(r2t->datasn);
|
|
|
|
r2t->datasn++;
|
2006-04-07 06:13:41 +04:00
|
|
|
hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
|
2011-06-17 02:57:09 +04:00
|
|
|
hdr->lun = task->lun;
|
2008-12-02 09:32:05 +03:00
|
|
|
hdr->itt = task->hdr_itt;
|
|
|
|
hdr->exp_statsn = r2t->exp_statsn;
|
|
|
|
hdr->offset = cpu_to_be32(r2t->data_offset + r2t->sent);
|
|
|
|
if (left > conn->max_xmit_dlength) {
|
2006-04-07 06:13:41 +04:00
|
|
|
hton24(hdr->dlength, conn->max_xmit_dlength);
|
2008-12-02 09:32:05 +03:00
|
|
|
r2t->data_count = conn->max_xmit_dlength;
|
2006-04-07 06:13:41 +04:00
|
|
|
hdr->flags = 0;
|
|
|
|
} else {
|
2008-12-02 09:32:05 +03:00
|
|
|
hton24(hdr->dlength, left);
|
|
|
|
r2t->data_count = left;
|
2006-04-07 06:13:41 +04:00
|
|
|
hdr->flags = ISCSI_FLAG_CMD_FINAL;
|
|
|
|
}
|
2008-12-02 09:32:05 +03:00
|
|
|
conn->dataout_pdus_cnt++;
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
2008-12-02 09:32:05 +03:00
|
|
|
EXPORT_SYMBOL_GPL(iscsi_prep_data_out_pdu);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
static int iscsi_add_hdr(struct iscsi_task *task, unsigned len)
|
2007-12-13 21:43:23 +03:00
|
|
|
{
|
2008-05-22 00:54:09 +04:00
|
|
|
unsigned exp_len = task->hdr_len + len;
|
2007-12-13 21:43:23 +03:00
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
if (exp_len > task->hdr_max) {
|
2007-12-13 21:43:23 +03:00
|
|
|
WARN_ON(1);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
WARN_ON(len & (ISCSI_PAD_LEN - 1)); /* caller must pad the AHS */
|
2008-05-22 00:54:09 +04:00
|
|
|
task->hdr_len = exp_len;
|
2007-12-13 21:43:23 +03:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2008-04-18 19:11:51 +04:00
|
|
|
/*
|
|
|
|
* make an extended cdb AHS
|
|
|
|
*/
|
2008-05-22 00:54:09 +04:00
|
|
|
static int iscsi_prep_ecdb_ahs(struct iscsi_task *task)
|
2008-04-18 19:11:51 +04:00
|
|
|
{
|
2008-05-22 00:54:09 +04:00
|
|
|
struct scsi_cmnd *cmd = task->sc;
|
2008-04-18 19:11:51 +04:00
|
|
|
unsigned rlen, pad_len;
|
|
|
|
unsigned short ahslength;
|
|
|
|
struct iscsi_ecdb_ahdr *ecdb_ahdr;
|
|
|
|
int rc;
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
ecdb_ahdr = iscsi_next_hdr(task);
|
2008-04-18 19:11:51 +04:00
|
|
|
rlen = cmd->cmd_len - ISCSI_CDB_SIZE;
|
|
|
|
|
|
|
|
BUG_ON(rlen > sizeof(ecdb_ahdr->ecdb));
|
|
|
|
ahslength = rlen + sizeof(ecdb_ahdr->reserved);
|
|
|
|
|
|
|
|
pad_len = iscsi_padding(rlen);
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
rc = iscsi_add_hdr(task, sizeof(ecdb_ahdr->ahslength) +
|
2008-04-18 19:11:51 +04:00
|
|
|
sizeof(ecdb_ahdr->ahstype) + ahslength + pad_len);
|
|
|
|
if (rc)
|
|
|
|
return rc;
|
|
|
|
|
|
|
|
if (pad_len)
|
|
|
|
memset(&ecdb_ahdr->ecdb[rlen], 0, pad_len);
|
|
|
|
|
|
|
|
ecdb_ahdr->ahslength = cpu_to_be16(ahslength);
|
|
|
|
ecdb_ahdr->ahstype = ISCSI_AHSTYPE_CDB;
|
|
|
|
ecdb_ahdr->reserved = 0;
|
|
|
|
memcpy(ecdb_ahdr->ecdb, cmd->cmnd + ISCSI_CDB_SIZE, rlen);
|
|
|
|
|
2009-03-05 23:45:58 +03:00
|
|
|
ISCSI_DBG_SESSION(task->conn->session,
|
|
|
|
"iscsi_prep_ecdb_ahs: varlen_cdb_len %d "
|
|
|
|
"rlen %d pad_len %d ahs_length %d iscsi_headers_size "
|
|
|
|
"%u\n", cmd->cmd_len, rlen, pad_len, ahslength,
|
|
|
|
task->hdr_len);
|
2008-04-18 19:11:51 +04:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
static int iscsi_prep_bidi_ahs(struct iscsi_task *task)
|
2008-04-18 19:11:52 +04:00
|
|
|
{
|
2008-05-22 00:54:09 +04:00
|
|
|
struct scsi_cmnd *sc = task->sc;
|
2008-04-18 19:11:52 +04:00
|
|
|
struct iscsi_rlength_ahdr *rlen_ahdr;
|
|
|
|
int rc;
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
rlen_ahdr = iscsi_next_hdr(task);
|
|
|
|
rc = iscsi_add_hdr(task, sizeof(*rlen_ahdr));
|
2008-04-18 19:11:52 +04:00
|
|
|
if (rc)
|
|
|
|
return rc;
|
|
|
|
|
|
|
|
rlen_ahdr->ahslength =
|
|
|
|
cpu_to_be16(sizeof(rlen_ahdr->read_length) +
|
|
|
|
sizeof(rlen_ahdr->reserved));
|
|
|
|
rlen_ahdr->ahstype = ISCSI_AHSTYPE_RLENGTH;
|
|
|
|
rlen_ahdr->reserved = 0;
|
|
|
|
rlen_ahdr->read_length = cpu_to_be32(scsi_in(sc)->length);
|
|
|
|
|
2009-03-05 23:45:58 +03:00
|
|
|
ISCSI_DBG_SESSION(task->conn->session,
|
|
|
|
"bidi-in rlen_ahdr->read_length(%d) "
|
|
|
|
"rlen_ahdr->ahslength(%d)\n",
|
|
|
|
be32_to_cpu(rlen_ahdr->read_length),
|
|
|
|
be16_to_cpu(rlen_ahdr->ahslength));
|
2008-04-18 19:11:52 +04:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2009-11-12 01:34:32 +03:00
|
|
|
/**
|
|
|
|
* iscsi_check_tmf_restrictions - check if a task is affected by TMF
|
|
|
|
* @task: iscsi task
|
|
|
|
* @opcode: opcode to check for
|
|
|
|
*
|
|
|
|
* During TMF a task has to be checked if it's affected.
|
|
|
|
* All unrelated I/O can be passed through, but I/O to the
|
|
|
|
* affected LUN should be restricted.
|
|
|
|
* If 'fast_abort' is set we won't be sending any I/O to the
|
|
|
|
* affected LUN.
|
|
|
|
* Otherwise the target is waiting for all TTTs to be completed,
|
|
|
|
* so we have to send all outstanding Data-Out PDUs to the target.
|
|
|
|
*/
|
|
|
|
static int iscsi_check_tmf_restrictions(struct iscsi_task *task, int opcode)
|
|
|
|
{
|
|
|
|
struct iscsi_conn *conn = task->conn;
|
|
|
|
struct iscsi_tm *tmf = &conn->tmhdr;
|
2014-06-25 17:27:36 +04:00
|
|
|
u64 hdr_lun;
|
2009-11-12 01:34:32 +03:00
|
|
|
|
|
|
|
if (conn->tmf_state == TMF_INITIAL)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
if ((tmf->opcode & ISCSI_OPCODE_MASK) != ISCSI_OP_SCSI_TMFUNC)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
switch (ISCSI_TM_FUNC_VALUE(tmf)) {
|
|
|
|
case ISCSI_TM_FUNC_LOGICAL_UNIT_RESET:
|
|
|
|
/*
|
|
|
|
* Allow PDUs for unrelated LUNs
|
|
|
|
*/
|
2011-06-17 02:57:09 +04:00
|
|
|
hdr_lun = scsilun_to_int(&tmf->lun);
|
2009-11-12 01:34:32 +03:00
|
|
|
if (hdr_lun != task->sc->device->lun)
|
|
|
|
return 0;
|
2009-11-12 01:34:33 +03:00
|
|
|
/* fall through */
|
|
|
|
case ISCSI_TM_FUNC_TARGET_WARM_RESET:
|
2009-11-12 01:34:32 +03:00
|
|
|
/*
|
|
|
|
* Fail all SCSI cmd PDUs
|
|
|
|
*/
|
|
|
|
if (opcode != ISCSI_OP_SCSI_DATA_OUT) {
|
|
|
|
iscsi_conn_printk(KERN_INFO, conn,
|
|
|
|
"task [op %x/%x itt "
|
2009-11-12 01:34:33 +03:00
|
|
|
"0x%x/0x%x] "
|
2009-11-12 01:34:32 +03:00
|
|
|
"rejected.\n",
|
|
|
|
task->hdr->opcode, opcode,
|
2009-11-12 01:34:33 +03:00
|
|
|
task->itt, task->hdr_itt);
|
2009-11-12 01:34:32 +03:00
|
|
|
return -EACCES;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* And also all data-out PDUs in response to R2T
|
|
|
|
* if fast_abort is set.
|
|
|
|
*/
|
|
|
|
if (conn->session->fast_abort) {
|
|
|
|
iscsi_conn_printk(KERN_INFO, conn,
|
|
|
|
"task [op %x/%x itt "
|
2009-11-12 01:34:33 +03:00
|
|
|
"0x%x/0x%x] fast abort.\n",
|
2009-11-12 01:34:32 +03:00
|
|
|
task->hdr->opcode, opcode,
|
2009-11-12 01:34:33 +03:00
|
|
|
task->itt, task->hdr_itt);
|
2009-11-12 01:34:32 +03:00
|
|
|
return -EACCES;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case ISCSI_TM_FUNC_ABORT_TASK:
|
|
|
|
/*
|
|
|
|
* the caller has already checked if the task
|
|
|
|
* they want to abort was in the pending queue so if
|
|
|
|
* we are here the cmd pdu has gone out already, and
|
|
|
|
* we will only hit this for data-outs
|
|
|
|
*/
|
|
|
|
if (opcode == ISCSI_OP_SCSI_DATA_OUT &&
|
|
|
|
task->hdr_itt == tmf->rtt) {
|
|
|
|
ISCSI_DBG_SESSION(conn->session,
|
|
|
|
"Preventing task %x/%x from sending "
|
|
|
|
"data-out due to abort task in "
|
|
|
|
"progress\n", task->itt,
|
|
|
|
task->hdr_itt);
|
|
|
|
return -EACCES;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
/**
|
|
|
|
* iscsi_prep_scsi_cmd_pdu - prep iscsi scsi cmd pdu
|
2008-05-22 00:54:09 +04:00
|
|
|
* @task: iscsi task
|
2006-04-07 06:13:41 +04:00
|
|
|
*
|
|
|
|
* Prep basic iSCSI PDU fields for a scsi cmd pdu. The LLD should set
|
|
|
|
* fields like dlength or final based on how much data it sends
|
|
|
|
*/
|
2008-05-22 00:54:09 +04:00
|
|
|
static int iscsi_prep_scsi_cmd_pdu(struct iscsi_task *task)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_conn *conn = task->conn;
|
2006-04-07 06:13:41 +04:00
|
|
|
struct iscsi_session *session = conn->session;
|
2008-05-22 00:54:09 +04:00
|
|
|
struct scsi_cmnd *sc = task->sc;
|
2011-05-27 15:16:33 +04:00
|
|
|
struct iscsi_scsi_req *hdr;
|
2014-06-11 13:09:58 +04:00
|
|
|
unsigned hdrlength, cmd_len, transfer_length;
|
2008-12-02 09:32:13 +03:00
|
|
|
itt_t itt;
|
2007-12-13 21:43:23 +03:00
|
|
|
int rc;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2009-11-12 01:34:32 +03:00
|
|
|
rc = iscsi_check_tmf_restrictions(task, ISCSI_OP_SCSI_CMD);
|
|
|
|
if (rc)
|
|
|
|
return rc;
|
|
|
|
|
2009-05-14 02:57:39 +04:00
|
|
|
if (conn->session->tt->alloc_pdu) {
|
|
|
|
rc = conn->session->tt->alloc_pdu(task, ISCSI_OP_SCSI_CMD);
|
|
|
|
if (rc)
|
|
|
|
return rc;
|
|
|
|
}
|
2011-05-27 15:16:33 +04:00
|
|
|
hdr = (struct iscsi_scsi_req *)task->hdr;
|
2008-12-02 09:32:13 +03:00
|
|
|
itt = hdr->itt;
|
2008-12-02 09:32:05 +03:00
|
|
|
memset(hdr, 0, sizeof(*hdr));
|
|
|
|
|
2008-12-02 09:32:14 +03:00
|
|
|
if (session->tt->parse_pdu_itt)
|
|
|
|
hdr->itt = task->hdr_itt = itt;
|
|
|
|
else
|
|
|
|
hdr->itt = task->hdr_itt = build_itt(task->itt,
|
|
|
|
task->conn->session->age);
|
2008-05-22 00:54:09 +04:00
|
|
|
task->hdr_len = 0;
|
|
|
|
rc = iscsi_add_hdr(task, sizeof(*hdr));
|
2007-12-13 21:43:23 +03:00
|
|
|
if (rc)
|
|
|
|
return rc;
|
2007-12-13 21:43:35 +03:00
|
|
|
hdr->opcode = ISCSI_OP_SCSI_CMD;
|
|
|
|
hdr->flags = ISCSI_ATTR_SIMPLE;
|
2011-06-17 02:57:09 +04:00
|
|
|
int_to_scsilun(sc->device->lun, &hdr->lun);
|
|
|
|
task->lun = hdr->lun;
|
2007-12-13 21:43:35 +03:00
|
|
|
hdr->exp_statsn = cpu_to_be32(conn->exp_statsn);
|
2008-04-18 19:11:51 +04:00
|
|
|
cmd_len = sc->cmd_len;
|
|
|
|
if (cmd_len < ISCSI_CDB_SIZE)
|
|
|
|
memset(&hdr->cdb[cmd_len], 0, ISCSI_CDB_SIZE - cmd_len);
|
|
|
|
else if (cmd_len > ISCSI_CDB_SIZE) {
|
2008-05-22 00:54:09 +04:00
|
|
|
rc = iscsi_prep_ecdb_ahs(task);
|
2008-04-18 19:11:51 +04:00
|
|
|
if (rc)
|
|
|
|
return rc;
|
|
|
|
cmd_len = ISCSI_CDB_SIZE;
|
|
|
|
}
|
|
|
|
memcpy(hdr->cdb, sc->cmnd, cmd_len);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
task->imm_count = 0;
|
2008-04-18 19:11:52 +04:00
|
|
|
if (scsi_bidi_cmnd(sc)) {
|
|
|
|
hdr->flags |= ISCSI_FLAG_CMD_READ;
|
2008-05-22 00:54:09 +04:00
|
|
|
rc = iscsi_prep_bidi_ahs(task);
|
2008-04-18 19:11:52 +04:00
|
|
|
if (rc)
|
|
|
|
return rc;
|
|
|
|
}
|
2014-03-05 21:43:49 +04:00
|
|
|
|
|
|
|
if (scsi_get_prot_op(sc) != SCSI_PROT_NORMAL)
|
|
|
|
task->protected = true;
|
|
|
|
|
2014-06-11 13:09:58 +04:00
|
|
|
transfer_length = scsi_transfer_length(sc);
|
|
|
|
hdr->data_length = cpu_to_be32(transfer_length);
|
2006-04-07 06:13:41 +04:00
|
|
|
if (sc->sc_data_direction == DMA_TO_DEVICE) {
|
2008-12-02 09:32:05 +03:00
|
|
|
struct iscsi_r2t_info *r2t = &task->unsol_r2t;
|
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
hdr->flags |= ISCSI_FLAG_CMD_WRITE;
|
|
|
|
/*
|
|
|
|
* Write counters:
|
|
|
|
*
|
|
|
|
* imm_count bytes to be sent right after
|
|
|
|
* SCSI PDU Header
|
|
|
|
*
|
|
|
|
* unsol_count bytes(as Data-Out) to be sent
|
|
|
|
* without R2T ack right after
|
|
|
|
* immediate data
|
|
|
|
*
|
2008-12-02 09:32:05 +03:00
|
|
|
* r2t data_length bytes to be sent via R2T ack's
|
2006-04-07 06:13:41 +04:00
|
|
|
*
|
|
|
|
* pad_count bytes to be sent as zero-padding
|
|
|
|
*/
|
2008-12-02 09:32:05 +03:00
|
|
|
memset(r2t, 0, sizeof(*r2t));
|
2006-04-07 06:13:41 +04:00
|
|
|
|
|
|
|
if (session->imm_data_en) {
|
2014-06-11 13:09:58 +04:00
|
|
|
if (transfer_length >= session->first_burst)
|
2008-05-22 00:54:09 +04:00
|
|
|
task->imm_count = min(session->first_burst,
|
2006-04-07 06:13:41 +04:00
|
|
|
conn->max_xmit_dlength);
|
|
|
|
else
|
2014-06-11 13:09:58 +04:00
|
|
|
task->imm_count = min(transfer_length,
|
|
|
|
conn->max_xmit_dlength);
|
2008-05-22 00:54:09 +04:00
|
|
|
hton24(hdr->dlength, task->imm_count);
|
2006-04-07 06:13:41 +04:00
|
|
|
} else
|
2007-12-13 21:43:35 +03:00
|
|
|
zero_data(hdr->dlength);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2006-09-01 02:09:24 +04:00
|
|
|
if (!session->initial_r2t_en) {
|
2014-06-11 13:09:58 +04:00
|
|
|
r2t->data_length = min(session->first_burst,
|
|
|
|
transfer_length) -
|
2008-12-02 09:32:05 +03:00
|
|
|
task->imm_count;
|
|
|
|
r2t->data_offset = task->imm_count;
|
|
|
|
r2t->ttt = cpu_to_be32(ISCSI_RESERVED_TAG);
|
|
|
|
r2t->exp_statsn = cpu_to_be32(conn->exp_statsn);
|
2006-09-01 02:09:24 +04:00
|
|
|
}
|
|
|
|
|
2008-12-02 09:32:05 +03:00
|
|
|
if (!task->unsol_r2t.data_length)
|
2006-04-07 06:13:41 +04:00
|
|
|
/* No unsolicit Data-Out's */
|
2007-12-13 21:43:35 +03:00
|
|
|
hdr->flags |= ISCSI_FLAG_CMD_FINAL;
|
2006-04-07 06:13:41 +04:00
|
|
|
} else {
|
|
|
|
hdr->flags |= ISCSI_FLAG_CMD_FINAL;
|
|
|
|
zero_data(hdr->dlength);
|
|
|
|
|
|
|
|
if (sc->sc_data_direction == DMA_FROM_DEVICE)
|
|
|
|
hdr->flags |= ISCSI_FLAG_CMD_READ;
|
|
|
|
}
|
|
|
|
|
2007-12-13 21:43:23 +03:00
|
|
|
/* calculate size of additional header segments (AHSs) */
|
2008-05-22 00:54:09 +04:00
|
|
|
hdrlength = task->hdr_len - sizeof(*hdr);
|
2007-12-13 21:43:23 +03:00
|
|
|
|
|
|
|
WARN_ON(hdrlength & (ISCSI_PAD_LEN-1));
|
|
|
|
hdrlength /= ISCSI_PAD_LEN;
|
|
|
|
|
|
|
|
WARN_ON(hdrlength >= 256);
|
|
|
|
hdr->hlength = hdrlength & 0xFF;
|
2010-04-25 01:21:19 +04:00
|
|
|
hdr->cmdsn = task->cmdsn = cpu_to_be32(session->cmdsn);
|
2007-12-13 21:43:23 +03:00
|
|
|
|
2008-12-02 09:32:05 +03:00
|
|
|
if (session->tt->init_task && session->tt->init_task(task))
|
2008-05-22 00:54:05 +04:00
|
|
|
return -EIO;
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
task->state = ISCSI_TASK_RUNNING;
|
2009-08-21 00:10:58 +04:00
|
|
|
session->cmdsn++;
|
2007-05-30 21:57:18 +04:00
|
|
|
|
2007-12-13 21:43:35 +03:00
|
|
|
conn->scsicmd_pdus_cnt++;
|
2009-03-05 23:45:58 +03:00
|
|
|
ISCSI_DBG_SESSION(session, "iscsi prep [%s cid %d sc %p cdb 0x%x "
|
|
|
|
"itt 0x%x len %d bidi_len %d cmdsn %d win %d]\n",
|
|
|
|
scsi_bidi_cmnd(sc) ? "bidirectional" :
|
|
|
|
sc->sc_data_direction == DMA_TO_DEVICE ?
|
|
|
|
"write" : "read", conn->id, sc, sc->cmnd[0],
|
2014-06-11 13:09:58 +04:00
|
|
|
task->itt, transfer_length,
|
2009-03-05 23:45:58 +03:00
|
|
|
scsi_bidi_cmnd(sc) ? scsi_in(sc)->length : 0,
|
|
|
|
session->cmdsn,
|
|
|
|
session->max_cmdsn - session->exp_cmdsn + 1);
|
2007-12-13 21:43:23 +03:00
|
|
|
return 0;
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
2009-05-14 02:57:46 +04:00
|
|
|
* iscsi_free_task - free a task
|
2008-05-22 00:54:09 +04:00
|
|
|
* @task: iscsi cmd task
|
2006-04-07 06:13:41 +04:00
|
|
|
*
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
* Must be called with session back_lock.
|
2008-05-22 00:54:06 +04:00
|
|
|
* This function returns the scsi command to scsi-ml or cleans
|
|
|
|
* up mgmt tasks then returns the task to the pool.
|
2006-04-07 06:13:41 +04:00
|
|
|
*/
|
2009-05-14 02:57:46 +04:00
|
|
|
static void iscsi_free_task(struct iscsi_task *task)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_conn *conn = task->conn;
|
2007-12-13 21:43:33 +03:00
|
|
|
struct iscsi_session *session = conn->session;
|
2008-05-22 00:54:09 +04:00
|
|
|
struct scsi_cmnd *sc = task->sc;
|
2010-12-31 11:22:21 +03:00
|
|
|
int oldstate = task->state;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2009-05-14 02:57:50 +04:00
|
|
|
ISCSI_DBG_SESSION(session, "freeing task itt 0x%x state %d sc %p\n",
|
|
|
|
task->itt, task->state, task->sc);
|
|
|
|
|
2008-12-02 09:32:05 +03:00
|
|
|
session->tt->cleanup_task(task);
|
2009-05-14 02:57:46 +04:00
|
|
|
task->state = ISCSI_TASK_FREE;
|
2008-05-22 00:54:09 +04:00
|
|
|
task->sc = NULL;
|
2008-05-22 00:54:06 +04:00
|
|
|
/*
|
2008-05-22 00:54:09 +04:00
|
|
|
* login task is preallocated so do not free
|
2008-05-22 00:54:06 +04:00
|
|
|
*/
|
2008-05-22 00:54:09 +04:00
|
|
|
if (conn->login_task == task)
|
2008-05-22 00:54:06 +04:00
|
|
|
return;
|
|
|
|
|
2009-12-22 01:37:28 +03:00
|
|
|
kfifo_in(&session->cmdpool.queue, (void*)&task, sizeof(void*));
|
2008-05-22 00:54:05 +04:00
|
|
|
|
2008-05-22 00:54:06 +04:00
|
|
|
if (sc) {
|
|
|
|
/* SCSI eh reuses commands to verify us */
|
|
|
|
sc->SCp.ptr = NULL;
|
|
|
|
/*
|
2010-12-31 11:22:21 +03:00
|
|
|
* queue command may call this to free the task, so
|
|
|
|
* it will decide how to return sc to scsi-ml.
|
2008-05-22 00:54:06 +04:00
|
|
|
*/
|
2010-12-31 11:22:21 +03:00
|
|
|
if (oldstate != ISCSI_TASK_REQUEUE_SCSIQ)
|
2008-05-22 00:54:06 +04:00
|
|
|
sc->scsi_done(sc);
|
|
|
|
}
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
|
2008-05-22 00:54:18 +04:00
|
|
|
void __iscsi_get_task(struct iscsi_task *task)
|
2006-09-01 02:09:25 +04:00
|
|
|
{
|
2008-05-22 00:54:09 +04:00
|
|
|
atomic_inc(&task->refcount);
|
2006-09-01 02:09:25 +04:00
|
|
|
}
|
2008-05-22 00:54:18 +04:00
|
|
|
EXPORT_SYMBOL_GPL(__iscsi_get_task);
|
2006-09-01 02:09:25 +04:00
|
|
|
|
2010-11-24 02:29:21 +03:00
|
|
|
void __iscsi_put_task(struct iscsi_task *task)
|
2006-09-01 02:09:25 +04:00
|
|
|
{
|
2008-05-22 00:54:09 +04:00
|
|
|
if (atomic_dec_and_test(&task->refcount))
|
2009-05-14 02:57:46 +04:00
|
|
|
iscsi_free_task(task);
|
2006-09-01 02:09:25 +04:00
|
|
|
}
|
2010-11-24 02:29:21 +03:00
|
|
|
EXPORT_SYMBOL_GPL(__iscsi_put_task);
|
2006-09-01 02:09:25 +04:00
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
void iscsi_put_task(struct iscsi_task *task)
|
2008-05-22 00:54:06 +04:00
|
|
|
{
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_session *session = task->conn->session;
|
2008-05-22 00:54:06 +04:00
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
/* regular RX path uses back_lock */
|
|
|
|
spin_lock_bh(&session->back_lock);
|
2008-05-22 00:54:09 +04:00
|
|
|
__iscsi_put_task(task);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->back_lock);
|
2008-05-22 00:54:06 +04:00
|
|
|
}
|
2008-05-22 00:54:09 +04:00
|
|
|
EXPORT_SYMBOL_GPL(iscsi_put_task);
|
2008-05-22 00:54:06 +04:00
|
|
|
|
2009-05-14 02:57:46 +04:00
|
|
|
/**
|
|
|
|
* iscsi_complete_task - finish a task
|
|
|
|
* @task: iscsi cmd task
|
2009-05-14 02:57:49 +04:00
|
|
|
* @state: state to complete task with
|
2009-05-14 02:57:46 +04:00
|
|
|
*
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
* Must be called with session back_lock.
|
2009-05-14 02:57:46 +04:00
|
|
|
*/
|
2009-05-14 02:57:49 +04:00
|
|
|
static void iscsi_complete_task(struct iscsi_task *task, int state)
|
2009-05-14 02:57:46 +04:00
|
|
|
{
|
|
|
|
struct iscsi_conn *conn = task->conn;
|
|
|
|
|
2009-05-14 02:57:50 +04:00
|
|
|
ISCSI_DBG_SESSION(conn->session,
|
|
|
|
"complete task itt 0x%x state %d sc %p\n",
|
|
|
|
task->itt, task->state, task->sc);
|
2009-05-14 02:57:49 +04:00
|
|
|
if (task->state == ISCSI_TASK_COMPLETED ||
|
|
|
|
task->state == ISCSI_TASK_ABRT_TMF ||
|
2010-12-31 11:22:21 +03:00
|
|
|
task->state == ISCSI_TASK_ABRT_SESS_RECOV ||
|
|
|
|
task->state == ISCSI_TASK_REQUEUE_SCSIQ)
|
2009-05-14 02:57:46 +04:00
|
|
|
return;
|
|
|
|
WARN_ON_ONCE(task->state == ISCSI_TASK_FREE);
|
2009-05-14 02:57:49 +04:00
|
|
|
task->state = state;
|
2009-05-14 02:57:46 +04:00
|
|
|
|
|
|
|
if (!list_empty(&task->running))
|
|
|
|
list_del_init(&task->running);
|
|
|
|
|
|
|
|
if (conn->task == task)
|
|
|
|
conn->task = NULL;
|
|
|
|
|
|
|
|
if (conn->ping_task == task)
|
|
|
|
conn->ping_task = NULL;
|
|
|
|
|
|
|
|
/* release get from queueing */
|
|
|
|
__iscsi_put_task(task);
|
|
|
|
}
|
|
|
|
|
2009-09-05 06:04:23 +04:00
|
|
|
/**
|
|
|
|
* iscsi_complete_scsi_task - finish scsi task normally
|
|
|
|
* @task: iscsi task for scsi cmd
|
|
|
|
* @exp_cmdsn: expected cmd sn in cpu format
|
|
|
|
* @max_cmdsn: max cmd sn in cpu format
|
|
|
|
*
|
|
|
|
* This is used when drivers do not need or cannot perform
|
|
|
|
* lower level pdu processing.
|
|
|
|
*
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
* Called with session back_lock
|
2009-09-05 06:04:23 +04:00
|
|
|
*/
|
|
|
|
void iscsi_complete_scsi_task(struct iscsi_task *task,
|
|
|
|
uint32_t exp_cmdsn, uint32_t max_cmdsn)
|
|
|
|
{
|
|
|
|
struct iscsi_conn *conn = task->conn;
|
|
|
|
|
|
|
|
ISCSI_DBG_SESSION(conn->session, "[itt 0x%x]\n", task->itt);
|
|
|
|
|
|
|
|
conn->last_recv = jiffies;
|
|
|
|
__iscsi_update_cmdsn(conn->session, exp_cmdsn, max_cmdsn);
|
|
|
|
iscsi_complete_task(task, ISCSI_TASK_COMPLETED);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_complete_scsi_task);
|
|
|
|
|
|
|
|
|
2007-12-13 21:43:26 +03:00
|
|
|
/*
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
* session back_lock must be held and if not called for a task that is
|
2009-05-14 02:57:46 +04:00
|
|
|
* still pending or from the xmit thread, then xmit thread must
|
|
|
|
* be suspended.
|
2007-12-13 21:43:26 +03:00
|
|
|
*/
|
2009-05-14 02:57:46 +04:00
|
|
|
static void fail_scsi_task(struct iscsi_task *task, int err)
|
2007-12-13 21:43:26 +03:00
|
|
|
{
|
2009-05-14 02:57:46 +04:00
|
|
|
struct iscsi_conn *conn = task->conn;
|
2007-12-13 21:43:26 +03:00
|
|
|
struct scsi_cmnd *sc;
|
2009-05-14 02:57:49 +04:00
|
|
|
int state;
|
2007-12-13 21:43:26 +03:00
|
|
|
|
2009-05-14 02:57:46 +04:00
|
|
|
/*
|
|
|
|
* if a command completes and we get a successful tmf response
|
|
|
|
* we will hit this because the scsi eh abort code does not take
|
|
|
|
* a ref to the task.
|
|
|
|
*/
|
2008-05-22 00:54:09 +04:00
|
|
|
sc = task->sc;
|
2007-12-13 21:43:26 +03:00
|
|
|
if (!sc)
|
|
|
|
return;
|
|
|
|
|
2009-05-14 02:57:49 +04:00
|
|
|
if (task->state == ISCSI_TASK_PENDING) {
|
2007-12-13 21:43:26 +03:00
|
|
|
/*
|
|
|
|
* cmd never made it to the xmit thread, so we should not count
|
|
|
|
* the cmd in the sequencing
|
|
|
|
*/
|
|
|
|
conn->session->queued_cmdsn--;
|
2009-05-14 02:57:49 +04:00
|
|
|
/* it was never sent so just complete like normal */
|
|
|
|
state = ISCSI_TASK_COMPLETED;
|
|
|
|
} else if (err == DID_TRANSPORT_DISRUPTED)
|
|
|
|
state = ISCSI_TASK_ABRT_SESS_RECOV;
|
|
|
|
else
|
|
|
|
state = ISCSI_TASK_ABRT_TMF;
|
2007-12-13 21:43:26 +03:00
|
|
|
|
2009-05-14 02:57:49 +04:00
|
|
|
sc->result = err << 16;
|
2008-04-18 19:11:52 +04:00
|
|
|
if (!scsi_bidi_cmnd(sc))
|
|
|
|
scsi_set_resid(sc, scsi_bufflen(sc));
|
|
|
|
else {
|
|
|
|
scsi_out(sc)->resid = scsi_out(sc)->length;
|
|
|
|
scsi_in(sc)->resid = scsi_in(sc)->length;
|
|
|
|
}
|
2008-05-22 00:54:06 +04:00
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
/* regular RX path uses back_lock */
|
|
|
|
spin_lock_bh(&conn->session->back_lock);
|
2009-05-14 02:57:49 +04:00
|
|
|
iscsi_complete_task(task, state);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&conn->session->back_lock);
|
2007-12-13 21:43:26 +03:00
|
|
|
}
|
|
|
|
|
2008-05-22 00:54:06 +04:00
|
|
|
static int iscsi_prep_mgmt_task(struct iscsi_conn *conn,
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_task *task)
|
2008-05-22 00:54:05 +04:00
|
|
|
{
|
|
|
|
struct iscsi_session *session = conn->session;
|
2008-12-02 09:32:05 +03:00
|
|
|
struct iscsi_hdr *hdr = task->hdr;
|
2008-05-22 00:54:05 +04:00
|
|
|
struct iscsi_nopout *nop = (struct iscsi_nopout *)hdr;
|
2009-11-12 01:34:31 +03:00
|
|
|
uint8_t opcode = hdr->opcode & ISCSI_OPCODE_MASK;
|
2008-05-22 00:54:05 +04:00
|
|
|
|
|
|
|
if (conn->session->state == ISCSI_STATE_LOGGING_OUT)
|
|
|
|
return -ENOTCONN;
|
|
|
|
|
2009-11-12 01:34:31 +03:00
|
|
|
if (opcode != ISCSI_OP_LOGIN && opcode != ISCSI_OP_TEXT)
|
2008-05-22 00:54:05 +04:00
|
|
|
nop->exp_statsn = cpu_to_be32(conn->exp_statsn);
|
|
|
|
/*
|
|
|
|
* pre-format CmdSN for outgoing PDU.
|
|
|
|
*/
|
|
|
|
nop->cmdsn = cpu_to_be32(session->cmdsn);
|
|
|
|
if (hdr->itt != RESERVED_ITT) {
|
|
|
|
/*
|
2009-11-12 01:34:31 +03:00
|
|
|
* TODO: We always use immediate for normal session pdus.
|
2008-05-22 00:54:05 +04:00
|
|
|
* If we start to send tmfs or nops as non-immediate then
|
|
|
|
* we should start checking the cmdsn numbers for mgmt tasks.
|
2009-11-12 01:34:31 +03:00
|
|
|
*
|
|
|
|
* During discovery sessions iscsid sends TEXT as non immediate,
|
|
|
|
* but we always only send one PDU at a time.
|
2008-05-22 00:54:05 +04:00
|
|
|
*/
|
|
|
|
if (conn->c_stage == ISCSI_CONN_STARTED &&
|
|
|
|
!(hdr->opcode & ISCSI_OP_IMMEDIATE)) {
|
|
|
|
session->queued_cmdsn++;
|
|
|
|
session->cmdsn++;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-12-02 09:32:15 +03:00
|
|
|
if (session->tt->init_task && session->tt->init_task(task))
|
|
|
|
return -EIO;
|
2008-05-22 00:54:05 +04:00
|
|
|
|
|
|
|
if ((hdr->opcode & ISCSI_OPCODE_MASK) == ISCSI_OP_LOGOUT)
|
|
|
|
session->state = ISCSI_STATE_LOGGING_OUT;
|
|
|
|
|
2008-12-02 09:32:05 +03:00
|
|
|
task->state = ISCSI_TASK_RUNNING;
|
2009-03-05 23:45:58 +03:00
|
|
|
ISCSI_DBG_SESSION(session, "mgmtpdu [op 0x%x hdr->itt 0x%x "
|
|
|
|
"datalen %d]\n", hdr->opcode & ISCSI_OPCODE_MASK,
|
|
|
|
hdr->itt, task->data_count);
|
2008-05-22 00:54:05 +04:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
static struct iscsi_task *
|
2007-12-13 21:43:30 +03:00
|
|
|
__iscsi_conn_send_pdu(struct iscsi_conn *conn, struct iscsi_hdr *hdr,
|
|
|
|
char *data, uint32_t data_size)
|
|
|
|
{
|
|
|
|
struct iscsi_session *session = conn->session;
|
2009-05-14 02:57:48 +04:00
|
|
|
struct iscsi_host *ihost = shost_priv(session->host);
|
2009-11-12 01:34:31 +03:00
|
|
|
uint8_t opcode = hdr->opcode & ISCSI_OPCODE_MASK;
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_task *task;
|
2008-12-02 09:32:13 +03:00
|
|
|
itt_t itt;
|
2007-12-13 21:43:30 +03:00
|
|
|
|
|
|
|
if (session->state == ISCSI_STATE_TERMINATE)
|
|
|
|
return NULL;
|
|
|
|
|
2009-11-12 01:34:31 +03:00
|
|
|
if (opcode == ISCSI_OP_LOGIN || opcode == ISCSI_OP_TEXT) {
|
2007-12-13 21:43:30 +03:00
|
|
|
/*
|
|
|
|
* Login and Text are sent serially, in
|
|
|
|
* request-followed-by-response sequence.
|
2008-05-22 00:54:06 +04:00
|
|
|
* Same task can be used. Same ITT must be used.
|
|
|
|
* Note that login_task is preallocated at conn_create().
|
2007-12-13 21:43:30 +03:00
|
|
|
*/
|
2009-11-12 01:34:31 +03:00
|
|
|
if (conn->login_task->state != ISCSI_TASK_FREE) {
|
|
|
|
iscsi_conn_printk(KERN_ERR, conn, "Login/Text in "
|
|
|
|
"progress. Cannot start new task.\n");
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2014-09-03 09:00:39 +04:00
|
|
|
if (data_size > ISCSI_DEF_MAX_RECV_SEG_LEN) {
|
|
|
|
iscsi_conn_printk(KERN_ERR, conn, "Invalid buffer len of %u for login task. Max len is %u\n", data_size, ISCSI_DEF_MAX_RECV_SEG_LEN);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
task = conn->login_task;
|
2009-11-12 01:34:31 +03:00
|
|
|
} else {
|
2009-05-14 02:57:43 +04:00
|
|
|
if (session->state != ISCSI_STATE_LOGGED_IN)
|
|
|
|
return NULL;
|
|
|
|
|
2014-09-03 09:00:39 +04:00
|
|
|
if (data_size != 0) {
|
|
|
|
iscsi_conn_printk(KERN_ERR, conn, "Can not send data buffer of len %u for op 0x%x\n", data_size, opcode);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2007-12-13 21:43:30 +03:00
|
|
|
BUG_ON(conn->c_stage == ISCSI_CONN_INITIAL_STAGE);
|
|
|
|
BUG_ON(conn->c_stage == ISCSI_CONN_STOPPED);
|
|
|
|
|
2009-12-22 01:37:28 +03:00
|
|
|
if (!kfifo_out(&session->cmdpool.queue,
|
2008-05-22 00:54:09 +04:00
|
|
|
(void*)&task, sizeof(void*)))
|
2007-12-13 21:43:30 +03:00
|
|
|
return NULL;
|
|
|
|
}
|
2008-05-22 00:54:06 +04:00
|
|
|
/*
|
|
|
|
* released in complete pdu for task we expect a response for, and
|
|
|
|
* released by the lld when it has transmitted the task for
|
|
|
|
* pdus we do not expect a response for.
|
|
|
|
*/
|
2008-05-22 00:54:09 +04:00
|
|
|
atomic_set(&task->refcount, 1);
|
|
|
|
task->conn = conn;
|
|
|
|
task->sc = NULL;
|
2009-05-14 02:57:46 +04:00
|
|
|
INIT_LIST_HEAD(&task->running);
|
|
|
|
task->state = ISCSI_TASK_PENDING;
|
2007-12-13 21:43:30 +03:00
|
|
|
|
|
|
|
if (data_size) {
|
2008-05-22 00:54:09 +04:00
|
|
|
memcpy(task->data, data, data_size);
|
|
|
|
task->data_count = data_size;
|
2007-12-13 21:43:30 +03:00
|
|
|
} else
|
2008-05-22 00:54:09 +04:00
|
|
|
task->data_count = 0;
|
2007-12-13 21:43:30 +03:00
|
|
|
|
2009-05-14 02:57:39 +04:00
|
|
|
if (conn->session->tt->alloc_pdu) {
|
|
|
|
if (conn->session->tt->alloc_pdu(task, hdr->opcode)) {
|
|
|
|
iscsi_conn_printk(KERN_ERR, conn, "Could not allocate "
|
|
|
|
"pdu for mgmt task.\n");
|
2009-05-14 02:57:46 +04:00
|
|
|
goto free_task;
|
2009-05-14 02:57:39 +04:00
|
|
|
}
|
2008-12-02 09:32:05 +03:00
|
|
|
}
|
2009-05-14 02:57:39 +04:00
|
|
|
|
2008-12-02 09:32:13 +03:00
|
|
|
itt = task->hdr->itt;
|
2008-12-02 09:32:05 +03:00
|
|
|
task->hdr_len = sizeof(struct iscsi_hdr);
|
2008-05-22 00:54:09 +04:00
|
|
|
memcpy(task->hdr, hdr, sizeof(struct iscsi_hdr));
|
2008-12-02 09:32:13 +03:00
|
|
|
|
|
|
|
if (hdr->itt != RESERVED_ITT) {
|
|
|
|
if (session->tt->parse_pdu_itt)
|
|
|
|
task->hdr->itt = itt;
|
|
|
|
else
|
|
|
|
task->hdr->itt = build_itt(task->itt,
|
|
|
|
task->conn->session->age);
|
|
|
|
}
|
|
|
|
|
2009-05-14 02:57:48 +04:00
|
|
|
if (!ihost->workq) {
|
2008-12-02 09:32:05 +03:00
|
|
|
if (iscsi_prep_mgmt_task(conn, task))
|
|
|
|
goto free_task;
|
2008-05-22 00:54:05 +04:00
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
if (session->tt->xmit_task(task))
|
2008-12-02 09:32:05 +03:00
|
|
|
goto free_task;
|
2009-05-14 02:57:46 +04:00
|
|
|
} else {
|
|
|
|
list_add_tail(&task->running, &conn->mgmtqueue);
|
2009-03-05 23:46:03 +03:00
|
|
|
iscsi_conn_queue_work(conn);
|
2009-05-14 02:57:46 +04:00
|
|
|
}
|
2008-05-22 00:54:05 +04:00
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
return task;
|
2008-12-02 09:32:05 +03:00
|
|
|
|
|
|
|
free_task:
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
/* regular RX path uses back_lock */
|
|
|
|
spin_lock_bh(&session->back_lock);
|
2008-12-02 09:32:05 +03:00
|
|
|
__iscsi_put_task(task);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->back_lock);
|
2008-12-02 09:32:05 +03:00
|
|
|
return NULL;
|
2007-12-13 21:43:30 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
int iscsi_conn_send_pdu(struct iscsi_cls_conn *cls_conn, struct iscsi_hdr *hdr,
|
|
|
|
char *data, uint32_t data_size)
|
|
|
|
{
|
|
|
|
struct iscsi_conn *conn = cls_conn->dd_data;
|
|
|
|
struct iscsi_session *session = conn->session;
|
|
|
|
int err = 0;
|
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2007-12-13 21:43:30 +03:00
|
|
|
if (!__iscsi_conn_send_pdu(conn, hdr, data, data_size))
|
|
|
|
err = -EPERM;
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2007-12-13 21:43:30 +03:00
|
|
|
return err;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_conn_send_pdu);
|
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
/**
|
|
|
|
* iscsi_cmd_rsp - SCSI Command Response processing
|
|
|
|
* @conn: iscsi connection
|
|
|
|
* @hdr: iscsi header
|
2008-05-22 00:54:09 +04:00
|
|
|
* @task: scsi command task
|
2006-04-07 06:13:41 +04:00
|
|
|
* @data: cmd data buffer
|
|
|
|
* @datalen: len of buffer
|
|
|
|
*
|
|
|
|
* iscsi_cmd_rsp sets up the scsi_cmnd fields based on the PDU and
|
2008-05-22 00:54:09 +04:00
|
|
|
* then completes the command and task.
|
2006-04-07 06:13:41 +04:00
|
|
|
**/
|
2007-05-30 21:57:18 +04:00
|
|
|
static void iscsi_scsi_cmd_rsp(struct iscsi_conn *conn, struct iscsi_hdr *hdr,
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_task *task, char *data,
|
2007-05-30 21:57:18 +04:00
|
|
|
int datalen)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
2011-05-27 15:16:33 +04:00
|
|
|
struct iscsi_scsi_rsp *rhdr = (struct iscsi_scsi_rsp *)hdr;
|
2006-04-07 06:13:41 +04:00
|
|
|
struct iscsi_session *session = conn->session;
|
2008-05-22 00:54:09 +04:00
|
|
|
struct scsi_cmnd *sc = task->sc;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2007-05-30 21:57:18 +04:00
|
|
|
iscsi_update_cmdsn(session, (struct iscsi_nopin*)rhdr);
|
2006-04-07 06:13:41 +04:00
|
|
|
conn->exp_statsn = be32_to_cpu(rhdr->statsn) + 1;
|
|
|
|
|
|
|
|
sc->result = (DID_OK << 16) | rhdr->cmd_status;
|
|
|
|
|
2014-03-05 21:43:49 +04:00
|
|
|
if (task->protected) {
|
|
|
|
sector_t sector;
|
|
|
|
u8 ascq;
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Transports that didn't implement check_protection
|
|
|
|
* callback but still published T10-PI support to scsi-mid
|
|
|
|
* deserve this BUG_ON.
|
|
|
|
**/
|
|
|
|
BUG_ON(!session->tt->check_protection);
|
|
|
|
|
|
|
|
ascq = session->tt->check_protection(task, §or);
|
|
|
|
if (ascq) {
|
|
|
|
sc->result = DRIVER_SENSE << 24 |
|
|
|
|
SAM_STAT_CHECK_CONDITION;
|
|
|
|
scsi_build_sense_buffer(1, sc->sense_buffer,
|
|
|
|
ILLEGAL_REQUEST, 0x10, ascq);
|
2015-07-15 10:55:39 +03:00
|
|
|
scsi_set_sense_information(sc->sense_buffer,
|
|
|
|
SCSI_SENSE_BUFFERSIZE,
|
|
|
|
sector);
|
2014-03-05 21:43:49 +04:00
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
if (rhdr->response != ISCSI_STATUS_CMD_COMPLETED) {
|
|
|
|
sc->result = DID_ERROR << 16;
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (rhdr->cmd_status == SAM_STAT_CHECK_CONDITION) {
|
2006-12-17 21:10:28 +03:00
|
|
|
uint16_t senselen;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
|
|
|
if (datalen < 2) {
|
|
|
|
invalid_datalen:
|
2008-01-31 22:36:52 +03:00
|
|
|
iscsi_conn_printk(KERN_ERR, conn,
|
|
|
|
"Got CHECK_CONDITION but invalid data "
|
|
|
|
"buffer size of %d\n", datalen);
|
2006-04-07 06:13:41 +04:00
|
|
|
sc->result = DID_BAD_TARGET << 16;
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
2008-05-22 00:54:20 +04:00
|
|
|
senselen = get_unaligned_be16(data);
|
2006-04-07 06:13:41 +04:00
|
|
|
if (datalen < senselen)
|
|
|
|
goto invalid_datalen;
|
|
|
|
|
|
|
|
memcpy(sc->sense_buffer, data + 2,
|
2006-12-17 21:10:28 +03:00
|
|
|
min_t(uint16_t, senselen, SCSI_SENSE_BUFFERSIZE));
|
2009-03-05 23:45:58 +03:00
|
|
|
ISCSI_DBG_SESSION(session, "copied %d bytes of sense\n",
|
|
|
|
min_t(uint16_t, senselen,
|
|
|
|
SCSI_SENSE_BUFFERSIZE));
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
|
2008-04-18 19:11:52 +04:00
|
|
|
if (rhdr->flags & (ISCSI_FLAG_CMD_BIDI_UNDERFLOW |
|
|
|
|
ISCSI_FLAG_CMD_BIDI_OVERFLOW)) {
|
|
|
|
int res_count = be32_to_cpu(rhdr->bi_residual_count);
|
|
|
|
|
|
|
|
if (scsi_bidi_cmnd(sc) && res_count > 0 &&
|
|
|
|
(rhdr->flags & ISCSI_FLAG_CMD_BIDI_OVERFLOW ||
|
|
|
|
res_count <= scsi_in(sc)->length))
|
|
|
|
scsi_in(sc)->resid = res_count;
|
|
|
|
else
|
|
|
|
sc->result = (DID_BAD_TARGET << 16) | rhdr->cmd_status;
|
|
|
|
}
|
|
|
|
|
2007-12-13 21:43:22 +03:00
|
|
|
if (rhdr->flags & (ISCSI_FLAG_CMD_UNDERFLOW |
|
|
|
|
ISCSI_FLAG_CMD_OVERFLOW)) {
|
2006-04-07 06:13:41 +04:00
|
|
|
int res_count = be32_to_cpu(rhdr->residual_count);
|
|
|
|
|
2007-12-13 21:43:22 +03:00
|
|
|
if (res_count > 0 &&
|
|
|
|
(rhdr->flags & ISCSI_FLAG_CMD_OVERFLOW ||
|
|
|
|
res_count <= scsi_bufflen(sc)))
|
2008-04-18 19:11:52 +04:00
|
|
|
/* write side for bidi or uni-io set_resid */
|
2007-06-14 17:13:17 +04:00
|
|
|
scsi_set_resid(sc, res_count);
|
2006-04-07 06:13:41 +04:00
|
|
|
else
|
|
|
|
sc->result = (DID_BAD_TARGET << 16) | rhdr->cmd_status;
|
2008-04-18 19:11:52 +04:00
|
|
|
}
|
2006-04-07 06:13:41 +04:00
|
|
|
out:
|
2009-05-14 02:57:46 +04:00
|
|
|
ISCSI_DBG_SESSION(session, "cmd rsp done [sc %p res %d itt 0x%x]\n",
|
2009-03-05 23:45:58 +03:00
|
|
|
sc, sc->result, task->itt);
|
2006-04-07 06:13:41 +04:00
|
|
|
conn->scsirsp_pdus_cnt++;
|
2009-05-14 02:57:49 +04:00
|
|
|
iscsi_complete_task(task, ISCSI_TASK_COMPLETED);
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
|
2008-09-24 20:46:09 +04:00
|
|
|
/**
|
|
|
|
* iscsi_data_in_rsp - SCSI Data-In Response processing
|
|
|
|
* @conn: iscsi connection
|
|
|
|
* @hdr: iscsi pdu
|
|
|
|
* @task: scsi command task
|
|
|
|
**/
|
|
|
|
static void
|
|
|
|
iscsi_data_in_rsp(struct iscsi_conn *conn, struct iscsi_hdr *hdr,
|
|
|
|
struct iscsi_task *task)
|
|
|
|
{
|
|
|
|
struct iscsi_data_rsp *rhdr = (struct iscsi_data_rsp *)hdr;
|
|
|
|
struct scsi_cmnd *sc = task->sc;
|
|
|
|
|
|
|
|
if (!(rhdr->flags & ISCSI_FLAG_DATA_STATUS))
|
|
|
|
return;
|
|
|
|
|
2009-05-14 02:57:42 +04:00
|
|
|
iscsi_update_cmdsn(conn->session, (struct iscsi_nopin *)hdr);
|
2008-09-24 20:46:09 +04:00
|
|
|
sc->result = (DID_OK << 16) | rhdr->cmd_status;
|
|
|
|
conn->exp_statsn = be32_to_cpu(rhdr->statsn) + 1;
|
|
|
|
if (rhdr->flags & (ISCSI_FLAG_DATA_UNDERFLOW |
|
|
|
|
ISCSI_FLAG_DATA_OVERFLOW)) {
|
|
|
|
int res_count = be32_to_cpu(rhdr->residual_count);
|
|
|
|
|
|
|
|
if (res_count > 0 &&
|
|
|
|
(rhdr->flags & ISCSI_FLAG_CMD_OVERFLOW ||
|
|
|
|
res_count <= scsi_in(sc)->length))
|
|
|
|
scsi_in(sc)->resid = res_count;
|
|
|
|
else
|
|
|
|
sc->result = (DID_BAD_TARGET << 16) | rhdr->cmd_status;
|
|
|
|
}
|
|
|
|
|
2009-05-14 02:57:46 +04:00
|
|
|
ISCSI_DBG_SESSION(conn->session, "data in with status done "
|
|
|
|
"[sc %p res %d itt 0x%x]\n",
|
|
|
|
sc, sc->result, task->itt);
|
2008-09-24 20:46:09 +04:00
|
|
|
conn->scsirsp_pdus_cnt++;
|
2009-05-14 02:57:49 +04:00
|
|
|
iscsi_complete_task(task, ISCSI_TASK_COMPLETED);
|
2008-09-24 20:46:09 +04:00
|
|
|
}
|
|
|
|
|
2006-07-25 00:47:22 +04:00
|
|
|
static void iscsi_tmf_rsp(struct iscsi_conn *conn, struct iscsi_hdr *hdr)
|
|
|
|
{
|
|
|
|
struct iscsi_tm_rsp *tmf = (struct iscsi_tm_rsp *)hdr;
|
|
|
|
|
|
|
|
conn->exp_statsn = be32_to_cpu(hdr->statsn) + 1;
|
|
|
|
conn->tmfrsp_pdus_cnt++;
|
|
|
|
|
2007-12-13 21:43:20 +03:00
|
|
|
if (conn->tmf_state != TMF_QUEUED)
|
2006-07-25 00:47:22 +04:00
|
|
|
return;
|
|
|
|
|
|
|
|
if (tmf->response == ISCSI_TMF_RSP_COMPLETE)
|
2007-12-13 21:43:20 +03:00
|
|
|
conn->tmf_state = TMF_SUCCESS;
|
2006-07-25 00:47:22 +04:00
|
|
|
else if (tmf->response == ISCSI_TMF_RSP_NO_TASK)
|
2007-12-13 21:43:20 +03:00
|
|
|
conn->tmf_state = TMF_NOT_FOUND;
|
2006-07-25 00:47:22 +04:00
|
|
|
else
|
2007-12-13 21:43:20 +03:00
|
|
|
conn->tmf_state = TMF_FAILED;
|
2006-07-25 00:47:22 +04:00
|
|
|
wake_up(&conn->ehwait);
|
|
|
|
}
|
|
|
|
|
2015-09-03 19:49:55 +03:00
|
|
|
static int iscsi_send_nopout(struct iscsi_conn *conn, struct iscsi_nopin *rhdr)
|
2007-12-13 21:43:30 +03:00
|
|
|
{
|
|
|
|
struct iscsi_nopout hdr;
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_task *task;
|
2007-12-13 21:43:30 +03:00
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
if (!rhdr && conn->ping_task)
|
2015-09-03 19:49:55 +03:00
|
|
|
return -EINVAL;
|
2007-12-13 21:43:30 +03:00
|
|
|
|
|
|
|
memset(&hdr, 0, sizeof(struct iscsi_nopout));
|
|
|
|
hdr.opcode = ISCSI_OP_NOOP_OUT | ISCSI_OP_IMMEDIATE;
|
|
|
|
hdr.flags = ISCSI_FLAG_CMD_FINAL;
|
|
|
|
|
|
|
|
if (rhdr) {
|
2011-06-17 02:57:09 +04:00
|
|
|
hdr.lun = rhdr->lun;
|
2007-12-13 21:43:30 +03:00
|
|
|
hdr.ttt = rhdr->ttt;
|
|
|
|
hdr.itt = RESERVED_ITT;
|
|
|
|
} else
|
|
|
|
hdr.ttt = RESERVED_ITT;
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
task = __iscsi_conn_send_pdu(conn, (struct iscsi_hdr *)&hdr, NULL, 0);
|
2015-09-03 19:49:55 +03:00
|
|
|
if (!task) {
|
2008-01-31 22:36:52 +03:00
|
|
|
iscsi_conn_printk(KERN_ERR, conn, "Could not send nopout\n");
|
2015-09-03 19:49:55 +03:00
|
|
|
return -EIO;
|
|
|
|
} else if (!rhdr) {
|
2008-12-01 21:13:00 +03:00
|
|
|
/* only track our nops */
|
|
|
|
conn->ping_task = task;
|
|
|
|
conn->last_ping = jiffies;
|
|
|
|
}
|
2015-09-03 19:49:55 +03:00
|
|
|
|
|
|
|
return 0;
|
2007-12-13 21:43:30 +03:00
|
|
|
}
|
|
|
|
|
2009-08-21 00:10:59 +04:00
|
|
|
static int iscsi_nop_out_rsp(struct iscsi_task *task,
|
|
|
|
struct iscsi_nopin *nop, char *data, int datalen)
|
|
|
|
{
|
|
|
|
struct iscsi_conn *conn = task->conn;
|
|
|
|
int rc = 0;
|
|
|
|
|
|
|
|
if (conn->ping_task != task) {
|
|
|
|
/*
|
|
|
|
* If this is not in response to one of our
|
|
|
|
* nops then it must be from userspace.
|
|
|
|
*/
|
|
|
|
if (iscsi_recv_pdu(conn->cls_conn, (struct iscsi_hdr *)nop,
|
|
|
|
data, datalen))
|
|
|
|
rc = ISCSI_ERR_CONN_FAILED;
|
|
|
|
} else
|
|
|
|
mod_timer(&conn->transport_timer, jiffies + conn->recv_timeout);
|
|
|
|
iscsi_complete_task(task, ISCSI_TASK_COMPLETED);
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
[SCSI] iscsi_tcp: fix padding, data digests, and IO at weird offsets
iscsi_tcp calculates padding by using the expected transfer length. This
has the problem where if we have immediate data = no and initial R2T =
yes, and the transfer length ended up needing padding then we send:
1. header
2. padding which should have gone after data
3. data
Besides this bug, we also assume the target will always ask for nice
transfer lengths and the first burst length will always be a nice value.
As far as I can tell form the RFC this is not a requirement. It would be
silly to do this, but if someone did it we will end doing bad things.
Finally the last bug in that bit of code is in our handling of the
recalculation of data digests when we do not send a whole iscsi_buf in
one try. The bug here is that we call crypto_digest_final on a
iscsi_sendpage error, then when we send the rest of the iscsi_buf, we
doiscsi_data_digest_init and this causes the previous data digest to be
lost.
And to make matters worse, some of these bugs are replicated over and
over and over again for immediate data, solicited data and unsolicited
data. So the attached patch made over the iscsi git tree (see
kernel.org/git for details) which I updated today to include the patches
I said I merged, consolidates the sending of data, padding and digests
and calculation of data digests and fixes the above bugs.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2006-09-01 02:09:27 +04:00
|
|
|
static int iscsi_handle_reject(struct iscsi_conn *conn, struct iscsi_hdr *hdr,
|
|
|
|
char *data, int datalen)
|
|
|
|
{
|
|
|
|
struct iscsi_reject *reject = (struct iscsi_reject *)hdr;
|
|
|
|
struct iscsi_hdr rejected_pdu;
|
2009-08-21 00:10:59 +04:00
|
|
|
int opcode, rc = 0;
|
[SCSI] iscsi_tcp: fix padding, data digests, and IO at weird offsets
iscsi_tcp calculates padding by using the expected transfer length. This
has the problem where if we have immediate data = no and initial R2T =
yes, and the transfer length ended up needing padding then we send:
1. header
2. padding which should have gone after data
3. data
Besides this bug, we also assume the target will always ask for nice
transfer lengths and the first burst length will always be a nice value.
As far as I can tell form the RFC this is not a requirement. It would be
silly to do this, but if someone did it we will end doing bad things.
Finally the last bug in that bit of code is in our handling of the
recalculation of data digests when we do not send a whole iscsi_buf in
one try. The bug here is that we call crypto_digest_final on a
iscsi_sendpage error, then when we send the rest of the iscsi_buf, we
doiscsi_data_digest_init and this causes the previous data digest to be
lost.
And to make matters worse, some of these bugs are replicated over and
over and over again for immediate data, solicited data and unsolicited
data. So the attached patch made over the iscsi git tree (see
kernel.org/git for details) which I updated today to include the patches
I said I merged, consolidates the sending of data, padding and digests
and calculation of data digests and fixes the above bugs.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2006-09-01 02:09:27 +04:00
|
|
|
|
|
|
|
conn->exp_statsn = be32_to_cpu(reject->statsn) + 1;
|
|
|
|
|
2009-08-21 00:10:59 +04:00
|
|
|
if (ntoh24(reject->dlength) > datalen ||
|
|
|
|
ntoh24(reject->dlength) < sizeof(struct iscsi_hdr)) {
|
|
|
|
iscsi_conn_printk(KERN_ERR, conn, "Cannot handle rejected "
|
|
|
|
"pdu. Invalid data length (pdu dlength "
|
|
|
|
"%u, datalen %d\n", ntoh24(reject->dlength),
|
|
|
|
datalen);
|
|
|
|
return ISCSI_ERR_PROTO;
|
|
|
|
}
|
|
|
|
memcpy(&rejected_pdu, data, sizeof(struct iscsi_hdr));
|
|
|
|
opcode = rejected_pdu.opcode & ISCSI_OPCODE_MASK;
|
|
|
|
|
|
|
|
switch (reject->reason) {
|
|
|
|
case ISCSI_REASON_DATA_DIGEST_ERROR:
|
|
|
|
iscsi_conn_printk(KERN_ERR, conn,
|
|
|
|
"pdu (op 0x%x itt 0x%x) rejected "
|
|
|
|
"due to DataDigest error.\n",
|
2014-01-09 06:21:34 +04:00
|
|
|
opcode, rejected_pdu.itt);
|
2009-08-21 00:10:59 +04:00
|
|
|
break;
|
|
|
|
case ISCSI_REASON_IMM_CMD_REJECT:
|
|
|
|
iscsi_conn_printk(KERN_ERR, conn,
|
|
|
|
"pdu (op 0x%x itt 0x%x) rejected. Too many "
|
|
|
|
"immediate commands.\n",
|
2014-01-09 06:21:34 +04:00
|
|
|
opcode, rejected_pdu.itt);
|
2009-08-21 00:10:59 +04:00
|
|
|
/*
|
|
|
|
* We only send one TMF at a time so if the target could not
|
|
|
|
* handle it, then it should get fixed (RFC mandates that
|
|
|
|
* a target can handle one immediate TMF per conn).
|
|
|
|
*
|
|
|
|
* For nops-outs, we could have sent more than one if
|
|
|
|
* the target is sending us lots of nop-ins
|
|
|
|
*/
|
|
|
|
if (opcode != ISCSI_OP_NOOP_OUT)
|
|
|
|
return 0;
|
[SCSI] iscsi_tcp: fix padding, data digests, and IO at weird offsets
iscsi_tcp calculates padding by using the expected transfer length. This
has the problem where if we have immediate data = no and initial R2T =
yes, and the transfer length ended up needing padding then we send:
1. header
2. padding which should have gone after data
3. data
Besides this bug, we also assume the target will always ask for nice
transfer lengths and the first burst length will always be a nice value.
As far as I can tell form the RFC this is not a requirement. It would be
silly to do this, but if someone did it we will end doing bad things.
Finally the last bug in that bit of code is in our handling of the
recalculation of data digests when we do not send a whole iscsi_buf in
one try. The bug here is that we call crypto_digest_final on a
iscsi_sendpage error, then when we send the rest of the iscsi_buf, we
doiscsi_data_digest_init and this causes the previous data digest to be
lost.
And to make matters worse, some of these bugs are replicated over and
over and over again for immediate data, solicited data and unsolicited
data. So the attached patch made over the iscsi git tree (see
kernel.org/git for details) which I updated today to include the patches
I said I merged, consolidates the sending of data, padding and digests
and calculation of data digests and fixes the above bugs.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2006-09-01 02:09:27 +04:00
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
if (rejected_pdu.itt == cpu_to_be32(ISCSI_RESERVED_TAG)) {
|
2009-08-21 00:10:59 +04:00
|
|
|
/*
|
|
|
|
* nop-out in response to target's nop-out rejected.
|
|
|
|
* Just resend.
|
|
|
|
*/
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
/* In RX path we are under back lock */
|
|
|
|
spin_unlock(&conn->session->back_lock);
|
|
|
|
spin_lock(&conn->session->frwd_lock);
|
2009-08-21 00:10:59 +04:00
|
|
|
iscsi_send_nopout(conn,
|
|
|
|
(struct iscsi_nopin*)&rejected_pdu);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock(&conn->session->frwd_lock);
|
|
|
|
spin_lock(&conn->session->back_lock);
|
|
|
|
} else {
|
2009-08-21 00:10:59 +04:00
|
|
|
struct iscsi_task *task;
|
|
|
|
/*
|
|
|
|
* Our nop as ping got dropped. We know the target
|
|
|
|
* and transport are ok so just clean up
|
|
|
|
*/
|
|
|
|
task = iscsi_itt_to_task(conn, rejected_pdu.itt);
|
|
|
|
if (!task) {
|
|
|
|
iscsi_conn_printk(KERN_ERR, conn,
|
|
|
|
"Invalid pdu reject. Could "
|
|
|
|
"not lookup rejected task.\n");
|
|
|
|
rc = ISCSI_ERR_BAD_ITT;
|
|
|
|
} else
|
|
|
|
rc = iscsi_nop_out_rsp(task,
|
|
|
|
(struct iscsi_nopin*)&rejected_pdu,
|
|
|
|
NULL, 0);
|
[SCSI] iscsi_tcp: fix padding, data digests, and IO at weird offsets
iscsi_tcp calculates padding by using the expected transfer length. This
has the problem where if we have immediate data = no and initial R2T =
yes, and the transfer length ended up needing padding then we send:
1. header
2. padding which should have gone after data
3. data
Besides this bug, we also assume the target will always ask for nice
transfer lengths and the first burst length will always be a nice value.
As far as I can tell form the RFC this is not a requirement. It would be
silly to do this, but if someone did it we will end doing bad things.
Finally the last bug in that bit of code is in our handling of the
recalculation of data digests when we do not send a whole iscsi_buf in
one try. The bug here is that we call crypto_digest_final on a
iscsi_sendpage error, then when we send the rest of the iscsi_buf, we
doiscsi_data_digest_init and this causes the previous data digest to be
lost.
And to make matters worse, some of these bugs are replicated over and
over and over again for immediate data, solicited data and unsolicited
data. So the attached patch made over the iscsi git tree (see
kernel.org/git for details) which I updated today to include the patches
I said I merged, consolidates the sending of data, padding and digests
and calculation of data digests and fixes the above bugs.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2006-09-01 02:09:27 +04:00
|
|
|
}
|
2009-08-21 00:10:59 +04:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
iscsi_conn_printk(KERN_ERR, conn,
|
|
|
|
"pdu (op 0x%x itt 0x%x) rejected. Reason "
|
2014-01-09 06:21:34 +04:00
|
|
|
"code 0x%x\n", rejected_pdu.opcode,
|
|
|
|
rejected_pdu.itt, reject->reason);
|
2009-08-21 00:10:59 +04:00
|
|
|
break;
|
[SCSI] iscsi_tcp: fix padding, data digests, and IO at weird offsets
iscsi_tcp calculates padding by using the expected transfer length. This
has the problem where if we have immediate data = no and initial R2T =
yes, and the transfer length ended up needing padding then we send:
1. header
2. padding which should have gone after data
3. data
Besides this bug, we also assume the target will always ask for nice
transfer lengths and the first burst length will always be a nice value.
As far as I can tell form the RFC this is not a requirement. It would be
silly to do this, but if someone did it we will end doing bad things.
Finally the last bug in that bit of code is in our handling of the
recalculation of data digests when we do not send a whole iscsi_buf in
one try. The bug here is that we call crypto_digest_final on a
iscsi_sendpage error, then when we send the rest of the iscsi_buf, we
doiscsi_data_digest_init and this causes the previous data digest to be
lost.
And to make matters worse, some of these bugs are replicated over and
over and over again for immediate data, solicited data and unsolicited
data. So the attached patch made over the iscsi git tree (see
kernel.org/git for details) which I updated today to include the patches
I said I merged, consolidates the sending of data, padding and digests
and calculation of data digests and fixes the above bugs.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2006-09-01 02:09:27 +04:00
|
|
|
}
|
2009-08-21 00:10:59 +04:00
|
|
|
return rc;
|
[SCSI] iscsi_tcp: fix padding, data digests, and IO at weird offsets
iscsi_tcp calculates padding by using the expected transfer length. This
has the problem where if we have immediate data = no and initial R2T =
yes, and the transfer length ended up needing padding then we send:
1. header
2. padding which should have gone after data
3. data
Besides this bug, we also assume the target will always ask for nice
transfer lengths and the first burst length will always be a nice value.
As far as I can tell form the RFC this is not a requirement. It would be
silly to do this, but if someone did it we will end doing bad things.
Finally the last bug in that bit of code is in our handling of the
recalculation of data digests when we do not send a whole iscsi_buf in
one try. The bug here is that we call crypto_digest_final on a
iscsi_sendpage error, then when we send the rest of the iscsi_buf, we
doiscsi_data_digest_init and this causes the previous data digest to be
lost.
And to make matters worse, some of these bugs are replicated over and
over and over again for immediate data, solicited data and unsolicited
data. So the attached patch made over the iscsi git tree (see
kernel.org/git for details) which I updated today to include the patches
I said I merged, consolidates the sending of data, padding and digests
and calculation of data digests and fixes the above bugs.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2006-09-01 02:09:27 +04:00
|
|
|
}
|
|
|
|
|
2008-05-22 00:54:18 +04:00
|
|
|
/**
|
|
|
|
* iscsi_itt_to_task - look up task by itt
|
|
|
|
* @conn: iscsi connection
|
|
|
|
* @itt: itt
|
|
|
|
*
|
|
|
|
* This should be used for mgmt tasks like login and nops, or if
|
|
|
|
* the LDD's itt space does not include the session age.
|
|
|
|
*
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
* The session back_lock must be held.
|
2008-05-22 00:54:18 +04:00
|
|
|
*/
|
2009-05-14 02:57:41 +04:00
|
|
|
struct iscsi_task *iscsi_itt_to_task(struct iscsi_conn *conn, itt_t itt)
|
2008-05-22 00:54:18 +04:00
|
|
|
{
|
|
|
|
struct iscsi_session *session = conn->session;
|
2008-12-02 09:32:13 +03:00
|
|
|
int i;
|
2008-05-22 00:54:18 +04:00
|
|
|
|
|
|
|
if (itt == RESERVED_ITT)
|
|
|
|
return NULL;
|
|
|
|
|
2008-12-02 09:32:13 +03:00
|
|
|
if (session->tt->parse_pdu_itt)
|
|
|
|
session->tt->parse_pdu_itt(conn, itt, &i, NULL);
|
|
|
|
else
|
|
|
|
i = get_itt(itt);
|
2008-05-22 00:54:18 +04:00
|
|
|
if (i >= session->cmds_max)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
return session->cmds[i];
|
|
|
|
}
|
2009-05-14 02:57:41 +04:00
|
|
|
EXPORT_SYMBOL_GPL(iscsi_itt_to_task);
|
2008-05-22 00:54:18 +04:00
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
/**
|
|
|
|
* __iscsi_complete_pdu - complete pdu
|
|
|
|
* @conn: iscsi conn
|
|
|
|
* @hdr: iscsi header
|
|
|
|
* @data: data buffer
|
|
|
|
* @datalen: len of data buffer
|
|
|
|
*
|
|
|
|
* Completes pdu processing by freeing any resources allocated at
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
* queuecommand or send generic. session back_lock must be held and verify
|
2006-04-07 06:13:41 +04:00
|
|
|
* itt must have been called.
|
|
|
|
*/
|
2008-05-22 00:54:18 +04:00
|
|
|
int __iscsi_complete_pdu(struct iscsi_conn *conn, struct iscsi_hdr *hdr,
|
|
|
|
char *data, int datalen)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
|
|
|
struct iscsi_session *session = conn->session;
|
|
|
|
int opcode = hdr->opcode & ISCSI_OPCODE_MASK, rc = 0;
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_task *task;
|
2006-04-07 06:13:41 +04:00
|
|
|
uint32_t itt;
|
|
|
|
|
2007-12-13 21:43:30 +03:00
|
|
|
conn->last_recv = jiffies;
|
2008-05-22 00:54:04 +04:00
|
|
|
rc = iscsi_verify_itt(conn, hdr->itt);
|
|
|
|
if (rc)
|
|
|
|
return rc;
|
|
|
|
|
2007-02-09 19:39:40 +03:00
|
|
|
if (hdr->itt != RESERVED_ITT)
|
|
|
|
itt = get_itt(hdr->itt);
|
2006-04-07 06:13:41 +04:00
|
|
|
else
|
2007-02-09 19:39:40 +03:00
|
|
|
itt = ~0U;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2009-03-05 23:45:58 +03:00
|
|
|
ISCSI_DBG_SESSION(session, "[op 0x%x cid %d itt 0x%x len %d]\n",
|
|
|
|
opcode, conn->id, itt, datalen);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:06 +04:00
|
|
|
if (itt == ~0U) {
|
2007-05-30 21:57:18 +04:00
|
|
|
iscsi_update_cmdsn(session, (struct iscsi_nopin*)hdr);
|
[SCSI] iscsi_tcp: fix padding, data digests, and IO at weird offsets
iscsi_tcp calculates padding by using the expected transfer length. This
has the problem where if we have immediate data = no and initial R2T =
yes, and the transfer length ended up needing padding then we send:
1. header
2. padding which should have gone after data
3. data
Besides this bug, we also assume the target will always ask for nice
transfer lengths and the first burst length will always be a nice value.
As far as I can tell form the RFC this is not a requirement. It would be
silly to do this, but if someone did it we will end doing bad things.
Finally the last bug in that bit of code is in our handling of the
recalculation of data digests when we do not send a whole iscsi_buf in
one try. The bug here is that we call crypto_digest_final on a
iscsi_sendpage error, then when we send the rest of the iscsi_buf, we
doiscsi_data_digest_init and this causes the previous data digest to be
lost.
And to make matters worse, some of these bugs are replicated over and
over and over again for immediate data, solicited data and unsolicited
data. So the attached patch made over the iscsi git tree (see
kernel.org/git for details) which I updated today to include the patches
I said I merged, consolidates the sending of data, padding and digests
and calculation of data digests and fixes the above bugs.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2006-09-01 02:09:27 +04:00
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
switch(opcode) {
|
|
|
|
case ISCSI_OP_NOOP_IN:
|
2006-07-25 00:47:45 +04:00
|
|
|
if (datalen) {
|
2006-04-07 06:13:41 +04:00
|
|
|
rc = ISCSI_ERR_PROTO;
|
2006-07-25 00:47:45 +04:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2007-02-09 19:39:40 +03:00
|
|
|
if (hdr->ttt == cpu_to_be32(ISCSI_RESERVED_TAG))
|
2006-07-25 00:47:45 +04:00
|
|
|
break;
|
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
/* In RX path we are under back lock */
|
|
|
|
spin_unlock(&session->back_lock);
|
|
|
|
spin_lock(&session->frwd_lock);
|
2007-12-13 21:43:30 +03:00
|
|
|
iscsi_send_nopout(conn, (struct iscsi_nopin*)hdr);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock(&session->frwd_lock);
|
|
|
|
spin_lock(&session->back_lock);
|
2006-04-07 06:13:41 +04:00
|
|
|
break;
|
|
|
|
case ISCSI_OP_REJECT:
|
[SCSI] iscsi_tcp: fix padding, data digests, and IO at weird offsets
iscsi_tcp calculates padding by using the expected transfer length. This
has the problem where if we have immediate data = no and initial R2T =
yes, and the transfer length ended up needing padding then we send:
1. header
2. padding which should have gone after data
3. data
Besides this bug, we also assume the target will always ask for nice
transfer lengths and the first burst length will always be a nice value.
As far as I can tell form the RFC this is not a requirement. It would be
silly to do this, but if someone did it we will end doing bad things.
Finally the last bug in that bit of code is in our handling of the
recalculation of data digests when we do not send a whole iscsi_buf in
one try. The bug here is that we call crypto_digest_final on a
iscsi_sendpage error, then when we send the rest of the iscsi_buf, we
doiscsi_data_digest_init and this causes the previous data digest to be
lost.
And to make matters worse, some of these bugs are replicated over and
over and over again for immediate data, solicited data and unsolicited
data. So the attached patch made over the iscsi git tree (see
kernel.org/git for details) which I updated today to include the patches
I said I merged, consolidates the sending of data, padding and digests
and calculation of data digests and fixes the above bugs.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2006-09-01 02:09:27 +04:00
|
|
|
rc = iscsi_handle_reject(conn, hdr, data, datalen);
|
|
|
|
break;
|
2006-04-07 06:13:41 +04:00
|
|
|
case ISCSI_OP_ASYNC_EVENT:
|
2006-05-03 04:46:47 +04:00
|
|
|
conn->exp_statsn = be32_to_cpu(hdr->statsn) + 1;
|
2006-10-17 02:09:41 +04:00
|
|
|
if (iscsi_recv_pdu(conn->cls_conn, hdr, data, datalen))
|
|
|
|
rc = ISCSI_ERR_CONN_FAILED;
|
2006-04-07 06:13:41 +04:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
rc = ISCSI_ERR_BAD_OPCODE;
|
|
|
|
break;
|
|
|
|
}
|
2008-05-22 00:54:06 +04:00
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
|
|
|
switch(opcode) {
|
|
|
|
case ISCSI_OP_SCSI_CMD_RSP:
|
2008-05-22 00:54:18 +04:00
|
|
|
case ISCSI_OP_SCSI_DATA_IN:
|
|
|
|
task = iscsi_itt_to_ctask(conn, hdr->itt);
|
|
|
|
if (!task)
|
|
|
|
return ISCSI_ERR_BAD_ITT;
|
2009-06-16 07:11:08 +04:00
|
|
|
task->last_xfer = jiffies;
|
2008-05-22 00:54:18 +04:00
|
|
|
break;
|
|
|
|
case ISCSI_OP_R2T:
|
|
|
|
/*
|
|
|
|
* LLD handles R2Ts if they need to.
|
|
|
|
*/
|
|
|
|
return 0;
|
|
|
|
case ISCSI_OP_LOGOUT_RSP:
|
|
|
|
case ISCSI_OP_LOGIN_RSP:
|
|
|
|
case ISCSI_OP_TEXT_RSP:
|
|
|
|
case ISCSI_OP_SCSI_TMFUNC_RSP:
|
|
|
|
case ISCSI_OP_NOOP_IN:
|
|
|
|
task = iscsi_itt_to_task(conn, hdr->itt);
|
|
|
|
if (!task)
|
|
|
|
return ISCSI_ERR_BAD_ITT;
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return ISCSI_ERR_BAD_OPCODE;
|
|
|
|
}
|
|
|
|
|
|
|
|
switch(opcode) {
|
|
|
|
case ISCSI_OP_SCSI_CMD_RSP:
|
2008-05-22 00:54:09 +04:00
|
|
|
iscsi_scsi_cmd_rsp(conn, hdr, task, data, datalen);
|
2008-05-22 00:54:06 +04:00
|
|
|
break;
|
|
|
|
case ISCSI_OP_SCSI_DATA_IN:
|
2008-09-24 20:46:09 +04:00
|
|
|
iscsi_data_in_rsp(conn, hdr, task);
|
2008-05-22 00:54:06 +04:00
|
|
|
break;
|
|
|
|
case ISCSI_OP_LOGOUT_RSP:
|
|
|
|
iscsi_update_cmdsn(session, (struct iscsi_nopin*)hdr);
|
|
|
|
if (datalen) {
|
|
|
|
rc = ISCSI_ERR_PROTO;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
conn->exp_statsn = be32_to_cpu(hdr->statsn) + 1;
|
|
|
|
goto recv_pdu;
|
|
|
|
case ISCSI_OP_LOGIN_RSP:
|
|
|
|
case ISCSI_OP_TEXT_RSP:
|
|
|
|
iscsi_update_cmdsn(session, (struct iscsi_nopin*)hdr);
|
|
|
|
/*
|
|
|
|
* login related PDU's exp_statsn is handled in
|
|
|
|
* userspace
|
|
|
|
*/
|
|
|
|
goto recv_pdu;
|
|
|
|
case ISCSI_OP_SCSI_TMFUNC_RSP:
|
|
|
|
iscsi_update_cmdsn(session, (struct iscsi_nopin*)hdr);
|
|
|
|
if (datalen) {
|
|
|
|
rc = ISCSI_ERR_PROTO;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
iscsi_tmf_rsp(conn, hdr);
|
2009-05-14 02:57:49 +04:00
|
|
|
iscsi_complete_task(task, ISCSI_TASK_COMPLETED);
|
2008-05-22 00:54:06 +04:00
|
|
|
break;
|
|
|
|
case ISCSI_OP_NOOP_IN:
|
|
|
|
iscsi_update_cmdsn(session, (struct iscsi_nopin*)hdr);
|
|
|
|
if (hdr->ttt != cpu_to_be32(ISCSI_RESERVED_TAG) || datalen) {
|
|
|
|
rc = ISCSI_ERR_PROTO;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
conn->exp_statsn = be32_to_cpu(hdr->statsn) + 1;
|
|
|
|
|
2009-08-21 00:10:59 +04:00
|
|
|
rc = iscsi_nop_out_rsp(task, (struct iscsi_nopin*)hdr,
|
|
|
|
data, datalen);
|
2008-05-22 00:54:06 +04:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
rc = ISCSI_ERR_BAD_OPCODE;
|
|
|
|
break;
|
|
|
|
}
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:06 +04:00
|
|
|
out:
|
|
|
|
return rc;
|
|
|
|
recv_pdu:
|
|
|
|
if (iscsi_recv_pdu(conn->cls_conn, hdr, data, datalen))
|
|
|
|
rc = ISCSI_ERR_CONN_FAILED;
|
2009-05-14 02:57:49 +04:00
|
|
|
iscsi_complete_task(task, ISCSI_TASK_COMPLETED);
|
2006-04-07 06:13:41 +04:00
|
|
|
return rc;
|
|
|
|
}
|
2008-05-22 00:54:18 +04:00
|
|
|
EXPORT_SYMBOL_GPL(__iscsi_complete_pdu);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
|
|
|
int iscsi_complete_pdu(struct iscsi_conn *conn, struct iscsi_hdr *hdr,
|
|
|
|
char *data, int datalen)
|
|
|
|
{
|
|
|
|
int rc;
|
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock(&conn->session->back_lock);
|
2006-04-07 06:13:41 +04:00
|
|
|
rc = __iscsi_complete_pdu(conn, hdr, data, datalen);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock(&conn->session->back_lock);
|
2006-04-07 06:13:41 +04:00
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_complete_pdu);
|
|
|
|
|
2008-05-22 00:54:04 +04:00
|
|
|
int iscsi_verify_itt(struct iscsi_conn *conn, itt_t itt)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
|
|
|
struct iscsi_session *session = conn->session;
|
2008-12-02 09:32:13 +03:00
|
|
|
int age = 0, i = 0;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:04 +04:00
|
|
|
if (itt == RESERVED_ITT)
|
|
|
|
return 0;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-12-02 09:32:13 +03:00
|
|
|
if (session->tt->parse_pdu_itt)
|
|
|
|
session->tt->parse_pdu_itt(conn, itt, &i, &age);
|
|
|
|
else {
|
|
|
|
i = get_itt(itt);
|
|
|
|
age = ((__force u32)itt >> ISCSI_AGE_SHIFT) & ISCSI_AGE_MASK;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (age != session->age) {
|
2008-05-22 00:54:04 +04:00
|
|
|
iscsi_conn_printk(KERN_ERR, conn,
|
|
|
|
"received itt %x expected session age (%x)\n",
|
2008-05-22 00:54:18 +04:00
|
|
|
(__force u32)itt, session->age);
|
2008-05-22 00:54:04 +04:00
|
|
|
return ISCSI_ERR_BAD_ITT;
|
|
|
|
}
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:06 +04:00
|
|
|
if (i >= session->cmds_max) {
|
|
|
|
iscsi_conn_printk(KERN_ERR, conn,
|
|
|
|
"received invalid itt index %u (max cmds "
|
|
|
|
"%u.\n", i, session->cmds_max);
|
|
|
|
return ISCSI_ERR_BAD_ITT;
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_verify_itt);
|
|
|
|
|
2008-05-22 00:54:18 +04:00
|
|
|
/**
|
|
|
|
* iscsi_itt_to_ctask - look up ctask by itt
|
|
|
|
* @conn: iscsi connection
|
|
|
|
* @itt: itt
|
|
|
|
*
|
|
|
|
* This should be used for cmd tasks.
|
|
|
|
*
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
* The session back_lock must be held.
|
2008-05-22 00:54:18 +04:00
|
|
|
*/
|
|
|
|
struct iscsi_task *iscsi_itt_to_ctask(struct iscsi_conn *conn, itt_t itt)
|
2008-05-22 00:54:04 +04:00
|
|
|
{
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_task *task;
|
2008-05-22 00:54:04 +04:00
|
|
|
|
|
|
|
if (iscsi_verify_itt(conn, itt))
|
|
|
|
return NULL;
|
|
|
|
|
2008-05-22 00:54:18 +04:00
|
|
|
task = iscsi_itt_to_task(conn, itt);
|
|
|
|
if (!task || !task->sc)
|
2008-05-22 00:54:04 +04:00
|
|
|
return NULL;
|
|
|
|
|
2008-05-22 00:54:18 +04:00
|
|
|
if (task->sc->SCp.phase != conn->session->age) {
|
|
|
|
iscsi_session_printk(KERN_ERR, conn->session,
|
|
|
|
"task's session age %d, expected %d\n",
|
|
|
|
task->sc->SCp.phase, conn->session->age);
|
2008-05-22 00:54:04 +04:00
|
|
|
return NULL;
|
2008-05-22 00:54:18 +04:00
|
|
|
}
|
2008-05-22 00:54:04 +04:00
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
return task;
|
2008-05-22 00:54:04 +04:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_itt_to_ctask);
|
|
|
|
|
2009-03-05 23:46:05 +03:00
|
|
|
void iscsi_session_failure(struct iscsi_session *session,
|
2008-09-24 20:46:10 +04:00
|
|
|
enum iscsi_err err)
|
|
|
|
{
|
|
|
|
struct iscsi_conn *conn;
|
|
|
|
struct device *dev;
|
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2008-09-24 20:46:10 +04:00
|
|
|
conn = session->leadconn;
|
|
|
|
if (session->state == ISCSI_STATE_TERMINATE || !conn) {
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2008-09-24 20:46:10 +04:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
dev = get_device(&conn->cls_conn->dev);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2008-09-24 20:46:10 +04:00
|
|
|
if (!dev)
|
|
|
|
return;
|
|
|
|
/*
|
|
|
|
* if the host is being removed bypass the connection
|
|
|
|
* recovery initialization because we are going to kill
|
|
|
|
* the session.
|
|
|
|
*/
|
|
|
|
if (err == ISCSI_ERR_INVALID_HOST)
|
|
|
|
iscsi_conn_error_event(conn->cls_conn, err);
|
|
|
|
else
|
|
|
|
iscsi_conn_failure(conn, err);
|
|
|
|
put_device(dev);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_session_failure);
|
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
void iscsi_conn_failure(struct iscsi_conn *conn, enum iscsi_err err)
|
|
|
|
{
|
|
|
|
struct iscsi_session *session = conn->session;
|
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2006-05-19 05:31:42 +04:00
|
|
|
if (session->state == ISCSI_STATE_FAILED) {
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2006-05-19 05:31:42 +04:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2006-05-30 09:37:20 +04:00
|
|
|
if (conn->stop_stage == 0)
|
2006-04-07 06:13:41 +04:00
|
|
|
session->state = ISCSI_STATE_FAILED;
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2008-09-24 20:46:10 +04:00
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
set_bit(ISCSI_SUSPEND_BIT, &conn->suspend_tx);
|
|
|
|
set_bit(ISCSI_SUSPEND_BIT, &conn->suspend_rx);
|
2008-09-24 20:46:10 +04:00
|
|
|
iscsi_conn_error_event(conn->cls_conn, err);
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_conn_failure);
|
|
|
|
|
2007-05-30 21:57:18 +04:00
|
|
|
static int iscsi_check_cmdsn_window_closed(struct iscsi_conn *conn)
|
|
|
|
{
|
|
|
|
struct iscsi_session *session = conn->session;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Check for iSCSI window and take care of CmdSN wrap-around
|
|
|
|
*/
|
2007-07-26 21:46:48 +04:00
|
|
|
if (!iscsi_sna_lte(session->queued_cmdsn, session->max_cmdsn)) {
|
2009-03-05 23:45:58 +03:00
|
|
|
ISCSI_DBG_SESSION(session, "iSCSI CmdSN closed. ExpCmdSn "
|
|
|
|
"%u MaxCmdSN %u CmdSN %u/%u\n",
|
|
|
|
session->exp_cmdsn, session->max_cmdsn,
|
|
|
|
session->cmdsn, session->queued_cmdsn);
|
2007-05-30 21:57:18 +04:00
|
|
|
return -ENOSPC;
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
static int iscsi_xmit_task(struct iscsi_conn *conn)
|
2007-05-30 21:57:18 +04:00
|
|
|
{
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_task *task = conn->task;
|
2007-12-13 21:43:20 +03:00
|
|
|
int rc;
|
2007-05-30 21:57:18 +04:00
|
|
|
|
2009-08-21 00:11:03 +04:00
|
|
|
if (test_bit(ISCSI_SUSPEND_BIT, &conn->suspend_tx))
|
|
|
|
return -ENODATA;
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
__iscsi_get_task(task);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&conn->session->frwd_lock);
|
2008-05-22 00:54:09 +04:00
|
|
|
rc = conn->session->tt->xmit_task(task);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&conn->session->frwd_lock);
|
2009-06-16 07:11:08 +04:00
|
|
|
if (!rc) {
|
2008-05-22 00:54:09 +04:00
|
|
|
/* done with this task */
|
2009-06-16 07:11:08 +04:00
|
|
|
task->last_xfer = jiffies;
|
2008-05-22 00:54:09 +04:00
|
|
|
conn->task = NULL;
|
2009-06-16 07:11:08 +04:00
|
|
|
}
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
/* regular RX path uses back_lock */
|
2014-03-30 16:26:29 +04:00
|
|
|
spin_lock(&conn->session->back_lock);
|
2009-06-16 07:11:08 +04:00
|
|
|
__iscsi_put_task(task);
|
2014-03-30 16:26:29 +04:00
|
|
|
spin_unlock(&conn->session->back_lock);
|
2007-05-30 21:57:18 +04:00
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
2007-12-13 21:43:20 +03:00
|
|
|
/**
|
2008-05-22 00:54:09 +04:00
|
|
|
* iscsi_requeue_task - requeue task to run from session workqueue
|
|
|
|
* @task: task to requeue
|
2007-12-13 21:43:20 +03:00
|
|
|
*
|
2008-05-22 00:54:09 +04:00
|
|
|
* LLDs that need to run a task from the session workqueue should call
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
* this. The session frwd_lock must be held. This should only be called
|
2008-05-22 00:54:05 +04:00
|
|
|
* by software drivers.
|
2007-12-13 21:43:20 +03:00
|
|
|
*/
|
2008-05-22 00:54:09 +04:00
|
|
|
void iscsi_requeue_task(struct iscsi_task *task)
|
2007-12-13 21:43:20 +03:00
|
|
|
{
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_conn *conn = task->conn;
|
2007-12-13 21:43:20 +03:00
|
|
|
|
2009-05-14 02:57:46 +04:00
|
|
|
/*
|
|
|
|
* this may be on the requeue list already if the xmit_task callout
|
|
|
|
* is handling the r2ts while we are adding new ones
|
|
|
|
*/
|
|
|
|
if (list_empty(&task->running))
|
|
|
|
list_add_tail(&task->running, &conn->requeue);
|
2009-03-05 23:46:03 +03:00
|
|
|
iscsi_conn_queue_work(conn);
|
2007-12-13 21:43:20 +03:00
|
|
|
}
|
2008-05-22 00:54:09 +04:00
|
|
|
EXPORT_SYMBOL_GPL(iscsi_requeue_task);
|
2007-12-13 21:43:20 +03:00
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
/**
|
|
|
|
* iscsi_data_xmit - xmit any command into the scheduled connection
|
|
|
|
* @conn: iscsi connection
|
|
|
|
*
|
|
|
|
* Notes:
|
|
|
|
* The function can return -EAGAIN in which case the caller must
|
|
|
|
* re-schedule it again later or recover. '0' return code means
|
|
|
|
* successful xmit.
|
|
|
|
**/
|
|
|
|
static int iscsi_data_xmit(struct iscsi_conn *conn)
|
|
|
|
{
|
2009-11-12 01:34:32 +03:00
|
|
|
struct iscsi_task *task;
|
2006-05-30 09:37:28 +04:00
|
|
|
int rc = 0;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&conn->session->frwd_lock);
|
2009-08-21 00:11:03 +04:00
|
|
|
if (test_bit(ISCSI_SUSPEND_BIT, &conn->suspend_tx)) {
|
2009-03-05 23:45:58 +03:00
|
|
|
ISCSI_DBG_SESSION(conn->session, "Tx suspended!\n");
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&conn->session->frwd_lock);
|
2006-05-30 09:37:28 +04:00
|
|
|
return -ENODATA;
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
if (conn->task) {
|
|
|
|
rc = iscsi_xmit_task(conn);
|
2006-05-30 09:37:28 +04:00
|
|
|
if (rc)
|
2009-08-21 00:11:03 +04:00
|
|
|
goto done;
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
|
2007-05-30 21:57:18 +04:00
|
|
|
/*
|
|
|
|
* process mgmt pdus like nops before commands since we should
|
|
|
|
* only have one nop-out as a ping from us and targets should not
|
|
|
|
* overflow us with nop-ins
|
|
|
|
*/
|
|
|
|
check_mgmt:
|
2007-12-13 21:43:20 +03:00
|
|
|
while (!list_empty(&conn->mgmtqueue)) {
|
2008-05-22 00:54:09 +04:00
|
|
|
conn->task = list_entry(conn->mgmtqueue.next,
|
|
|
|
struct iscsi_task, running);
|
2009-05-14 02:57:46 +04:00
|
|
|
list_del_init(&conn->task->running);
|
2008-05-22 00:54:09 +04:00
|
|
|
if (iscsi_prep_mgmt_task(conn, conn->task)) {
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
/* regular RX path uses back_lock */
|
|
|
|
spin_lock_bh(&conn->session->back_lock);
|
2008-05-22 00:54:09 +04:00
|
|
|
__iscsi_put_task(conn->task);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&conn->session->back_lock);
|
2008-05-22 00:54:09 +04:00
|
|
|
conn->task = NULL;
|
2007-12-13 21:43:26 +03:00
|
|
|
continue;
|
|
|
|
}
|
2008-05-22 00:54:09 +04:00
|
|
|
rc = iscsi_xmit_task(conn);
|
2007-05-30 21:57:18 +04:00
|
|
|
if (rc)
|
2009-08-21 00:11:03 +04:00
|
|
|
goto done;
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
|
2007-12-13 21:43:20 +03:00
|
|
|
/* process pending command queue */
|
2009-05-14 02:57:46 +04:00
|
|
|
while (!list_empty(&conn->cmdqueue)) {
|
2009-11-12 01:34:32 +03:00
|
|
|
conn->task = list_entry(conn->cmdqueue.next, struct iscsi_task,
|
|
|
|
running);
|
2009-05-14 02:57:46 +04:00
|
|
|
list_del_init(&conn->task->running);
|
2007-12-13 21:43:26 +03:00
|
|
|
if (conn->session->state == ISCSI_STATE_LOGGING_OUT) {
|
2009-05-14 02:57:49 +04:00
|
|
|
fail_scsi_task(conn->task, DID_IMM_RETRY);
|
2007-12-13 21:43:26 +03:00
|
|
|
continue;
|
|
|
|
}
|
2008-12-02 09:32:05 +03:00
|
|
|
rc = iscsi_prep_scsi_cmd_pdu(conn->task);
|
|
|
|
if (rc) {
|
2009-11-12 01:34:32 +03:00
|
|
|
if (rc == -ENOMEM || rc == -EACCES) {
|
2009-05-14 02:57:46 +04:00
|
|
|
list_add_tail(&conn->task->running,
|
|
|
|
&conn->cmdqueue);
|
2008-12-02 09:32:05 +03:00
|
|
|
conn->task = NULL;
|
2009-08-21 00:11:03 +04:00
|
|
|
goto done;
|
2008-12-02 09:32:05 +03:00
|
|
|
} else
|
2009-05-14 02:57:49 +04:00
|
|
|
fail_scsi_task(conn->task, DID_ABORT);
|
2007-12-13 21:43:23 +03:00
|
|
|
continue;
|
|
|
|
}
|
2008-05-22 00:54:09 +04:00
|
|
|
rc = iscsi_xmit_task(conn);
|
2007-05-30 21:57:18 +04:00
|
|
|
if (rc)
|
2009-08-21 00:11:03 +04:00
|
|
|
goto done;
|
2007-05-30 21:57:18 +04:00
|
|
|
/*
|
2008-05-22 00:54:09 +04:00
|
|
|
* we could continuously get new task requests so
|
2007-05-30 21:57:18 +04:00
|
|
|
* we need to check the mgmt queue for nops that need to
|
|
|
|
* be sent to aviod starvation
|
|
|
|
*/
|
2007-12-13 21:43:20 +03:00
|
|
|
if (!list_empty(&conn->mgmtqueue))
|
|
|
|
goto check_mgmt;
|
|
|
|
}
|
|
|
|
|
|
|
|
while (!list_empty(&conn->requeue)) {
|
2007-12-13 21:43:26 +03:00
|
|
|
/*
|
|
|
|
* we always do fastlogout - conn stop code will clean up.
|
|
|
|
*/
|
|
|
|
if (conn->session->state == ISCSI_STATE_LOGGING_OUT)
|
|
|
|
break;
|
|
|
|
|
2009-11-12 01:34:32 +03:00
|
|
|
task = list_entry(conn->requeue.next, struct iscsi_task,
|
|
|
|
running);
|
|
|
|
if (iscsi_check_tmf_restrictions(task, ISCSI_OP_SCSI_DATA_OUT))
|
|
|
|
break;
|
|
|
|
|
|
|
|
conn->task = task;
|
2009-05-14 02:57:46 +04:00
|
|
|
list_del_init(&conn->task->running);
|
2008-05-22 00:54:09 +04:00
|
|
|
conn->task->state = ISCSI_TASK_RUNNING;
|
|
|
|
rc = iscsi_xmit_task(conn);
|
2007-12-13 21:43:20 +03:00
|
|
|
if (rc)
|
2009-08-21 00:11:03 +04:00
|
|
|
goto done;
|
2007-12-13 21:43:20 +03:00
|
|
|
if (!list_empty(&conn->mgmtqueue))
|
2007-05-30 21:57:18 +04:00
|
|
|
goto check_mgmt;
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&conn->session->frwd_lock);
|
2006-05-30 09:37:28 +04:00
|
|
|
return -ENODATA;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2009-08-21 00:11:03 +04:00
|
|
|
done:
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&conn->session->frwd_lock);
|
2006-05-30 09:37:28 +04:00
|
|
|
return rc;
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
|
2006-11-22 17:57:56 +03:00
|
|
|
static void iscsi_xmitworker(struct work_struct *work)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
2006-11-22 17:57:56 +03:00
|
|
|
struct iscsi_conn *conn =
|
|
|
|
container_of(work, struct iscsi_conn, xmitwork);
|
2006-05-30 09:37:28 +04:00
|
|
|
int rc;
|
2006-04-07 06:13:41 +04:00
|
|
|
/*
|
|
|
|
* serialize Xmit worker on a per-connection basis.
|
|
|
|
*/
|
2006-05-30 09:37:28 +04:00
|
|
|
do {
|
|
|
|
rc = iscsi_data_xmit(conn);
|
|
|
|
} while (rc >= 0 || rc == -EAGAIN);
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
|
2008-12-02 09:32:05 +03:00
|
|
|
static inline struct iscsi_task *iscsi_alloc_task(struct iscsi_conn *conn,
|
|
|
|
struct scsi_cmnd *sc)
|
|
|
|
{
|
|
|
|
struct iscsi_task *task;
|
|
|
|
|
2009-12-22 01:37:28 +03:00
|
|
|
if (!kfifo_out(&conn->session->cmdpool.queue,
|
2008-12-02 09:32:05 +03:00
|
|
|
(void *) &task, sizeof(void *)))
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
sc->SCp.phase = conn->session->age;
|
|
|
|
sc->SCp.ptr = (char *) task;
|
|
|
|
|
|
|
|
atomic_set(&task->refcount, 1);
|
|
|
|
task->state = ISCSI_TASK_PENDING;
|
|
|
|
task->conn = conn;
|
|
|
|
task->sc = sc;
|
2009-06-16 07:11:08 +04:00
|
|
|
task->have_checked_conn = false;
|
|
|
|
task->last_timeout = jiffies;
|
|
|
|
task->last_xfer = jiffies;
|
2014-03-05 21:43:49 +04:00
|
|
|
task->protected = false;
|
2008-12-02 09:32:05 +03:00
|
|
|
INIT_LIST_HEAD(&task->running);
|
|
|
|
return task;
|
|
|
|
}
|
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
enum {
|
|
|
|
FAILURE_BAD_HOST = 1,
|
|
|
|
FAILURE_SESSION_FAILED,
|
|
|
|
FAILURE_SESSION_FREED,
|
|
|
|
FAILURE_WINDOW_CLOSED,
|
2006-09-01 02:09:25 +04:00
|
|
|
FAILURE_OOM,
|
2006-04-07 06:13:41 +04:00
|
|
|
FAILURE_SESSION_TERMINATE,
|
2006-05-19 05:31:42 +04:00
|
|
|
FAILURE_SESSION_IN_RECOVERY,
|
2006-04-07 06:13:41 +04:00
|
|
|
FAILURE_SESSION_RECOVERY_TIMEOUT,
|
2007-12-13 21:43:26 +03:00
|
|
|
FAILURE_SESSION_LOGGING_OUT,
|
2008-01-31 22:36:43 +03:00
|
|
|
FAILURE_SESSION_NOT_READY,
|
2006-04-07 06:13:41 +04:00
|
|
|
};
|
|
|
|
|
2010-12-31 11:22:21 +03:00
|
|
|
int iscsi_queuecommand(struct Scsi_Host *host, struct scsi_cmnd *sc)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
2008-05-22 00:53:59 +04:00
|
|
|
struct iscsi_cls_session *cls_session;
|
2009-05-14 02:57:48 +04:00
|
|
|
struct iscsi_host *ihost;
|
2006-04-07 06:13:41 +04:00
|
|
|
int reason = 0;
|
|
|
|
struct iscsi_session *session;
|
|
|
|
struct iscsi_conn *conn;
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_task *task = NULL;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
|
|
|
sc->result = 0;
|
2006-09-01 02:09:33 +04:00
|
|
|
sc->SCp.ptr = NULL;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2009-05-14 02:57:48 +04:00
|
|
|
ihost = shost_priv(host);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:53:59 +04:00
|
|
|
cls_session = starget_to_session(scsi_target(sc->device));
|
|
|
|
session = cls_session->dd_data;
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:53:59 +04:00
|
|
|
reason = iscsi_session_chkready(cls_session);
|
2008-01-31 22:36:43 +03:00
|
|
|
if (reason) {
|
|
|
|
sc->result = reason;
|
|
|
|
goto fault;
|
|
|
|
}
|
|
|
|
|
2009-05-14 02:57:47 +04:00
|
|
|
if (session->state != ISCSI_STATE_LOGGED_IN) {
|
2006-05-19 05:31:42 +04:00
|
|
|
/*
|
|
|
|
* to handle the race between when we set the recovery state
|
|
|
|
* and block the session we requeue here (commands could
|
|
|
|
* be entering our queuecommand while a block is starting
|
|
|
|
* up because the block code is not locked)
|
|
|
|
*/
|
2007-12-13 21:43:32 +03:00
|
|
|
switch (session->state) {
|
2009-05-14 02:57:47 +04:00
|
|
|
case ISCSI_STATE_FAILED:
|
2007-12-13 21:43:32 +03:00
|
|
|
case ISCSI_STATE_IN_RECOVERY:
|
2006-05-19 05:31:42 +04:00
|
|
|
reason = FAILURE_SESSION_IN_RECOVERY;
|
2009-05-14 02:57:47 +04:00
|
|
|
sc->result = DID_IMM_RETRY << 16;
|
|
|
|
break;
|
2007-12-13 21:43:32 +03:00
|
|
|
case ISCSI_STATE_LOGGING_OUT:
|
|
|
|
reason = FAILURE_SESSION_LOGGING_OUT;
|
2009-05-14 02:57:47 +04:00
|
|
|
sc->result = DID_IMM_RETRY << 16;
|
|
|
|
break;
|
2007-12-13 21:43:26 +03:00
|
|
|
case ISCSI_STATE_RECOVERY_FAILED:
|
2006-05-19 05:31:42 +04:00
|
|
|
reason = FAILURE_SESSION_RECOVERY_TIMEOUT;
|
2008-08-20 03:45:26 +04:00
|
|
|
sc->result = DID_TRANSPORT_FAILFAST << 16;
|
2007-12-13 21:43:26 +03:00
|
|
|
break;
|
|
|
|
case ISCSI_STATE_TERMINATE:
|
2006-05-19 05:31:42 +04:00
|
|
|
reason = FAILURE_SESSION_TERMINATE;
|
2008-01-31 22:36:43 +03:00
|
|
|
sc->result = DID_NO_CONNECT << 16;
|
2007-12-13 21:43:26 +03:00
|
|
|
break;
|
|
|
|
default:
|
2006-05-19 05:31:42 +04:00
|
|
|
reason = FAILURE_SESSION_FREED;
|
2008-01-31 22:36:43 +03:00
|
|
|
sc->result = DID_NO_CONNECT << 16;
|
2007-12-13 21:43:26 +03:00
|
|
|
}
|
2006-04-07 06:13:41 +04:00
|
|
|
goto fault;
|
|
|
|
}
|
|
|
|
|
|
|
|
conn = session->leadconn;
|
2006-10-17 02:09:39 +04:00
|
|
|
if (!conn) {
|
|
|
|
reason = FAILURE_SESSION_FREED;
|
2008-01-31 22:36:43 +03:00
|
|
|
sc->result = DID_NO_CONNECT << 16;
|
2006-10-17 02:09:39 +04:00
|
|
|
goto fault;
|
|
|
|
}
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2009-09-05 06:05:33 +04:00
|
|
|
if (test_bit(ISCSI_SUSPEND_BIT, &conn->suspend_tx)) {
|
|
|
|
reason = FAILURE_SESSION_IN_RECOVERY;
|
|
|
|
sc->result = DID_REQUEUE;
|
|
|
|
goto fault;
|
|
|
|
}
|
|
|
|
|
2007-05-30 21:57:18 +04:00
|
|
|
if (iscsi_check_cmdsn_window_closed(conn)) {
|
|
|
|
reason = FAILURE_WINDOW_CLOSED;
|
|
|
|
goto reject;
|
|
|
|
}
|
|
|
|
|
2008-12-02 09:32:05 +03:00
|
|
|
task = iscsi_alloc_task(conn, sc);
|
|
|
|
if (!task) {
|
2006-09-01 02:09:25 +04:00
|
|
|
reason = FAILURE_OOM;
|
|
|
|
goto reject;
|
|
|
|
}
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2009-05-14 02:57:48 +04:00
|
|
|
if (!ihost->workq) {
|
2008-12-02 09:32:05 +03:00
|
|
|
reason = iscsi_prep_scsi_cmd_pdu(task);
|
|
|
|
if (reason) {
|
2009-11-12 01:34:32 +03:00
|
|
|
if (reason == -ENOMEM || reason == -EACCES) {
|
2008-12-02 09:32:05 +03:00
|
|
|
reason = FAILURE_OOM;
|
|
|
|
goto prepd_reject;
|
|
|
|
} else {
|
|
|
|
sc->result = DID_ABORT << 16;
|
|
|
|
goto prepd_fault;
|
|
|
|
}
|
2008-05-22 00:54:05 +04:00
|
|
|
}
|
2008-05-22 00:54:09 +04:00
|
|
|
if (session->tt->xmit_task(task)) {
|
2009-08-21 00:10:58 +04:00
|
|
|
session->cmdsn--;
|
2008-05-22 00:54:05 +04:00
|
|
|
reason = FAILURE_SESSION_NOT_READY;
|
2008-12-02 09:32:05 +03:00
|
|
|
goto prepd_reject;
|
2008-05-22 00:54:05 +04:00
|
|
|
}
|
2009-05-14 02:57:46 +04:00
|
|
|
} else {
|
|
|
|
list_add_tail(&task->running, &conn->cmdqueue);
|
2009-03-05 23:46:03 +03:00
|
|
|
iscsi_conn_queue_work(conn);
|
2009-05-14 02:57:46 +04:00
|
|
|
}
|
2008-05-22 00:54:05 +04:00
|
|
|
|
|
|
|
session->queued_cmdsn++;
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2006-04-07 06:13:41 +04:00
|
|
|
return 0;
|
|
|
|
|
2008-12-02 09:32:05 +03:00
|
|
|
prepd_reject:
|
2010-12-31 11:22:21 +03:00
|
|
|
iscsi_complete_task(task, ISCSI_TASK_REQUEUE_SCSIQ);
|
2006-04-07 06:13:41 +04:00
|
|
|
reject:
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2009-03-05 23:45:58 +03:00
|
|
|
ISCSI_DBG_SESSION(session, "cmd 0x%x rejected (%d)\n",
|
|
|
|
sc->cmnd[0], reason);
|
2008-08-18 00:24:43 +04:00
|
|
|
return SCSI_MLQUEUE_TARGET_BUSY;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-12-02 09:32:05 +03:00
|
|
|
prepd_fault:
|
2010-12-31 11:22:21 +03:00
|
|
|
iscsi_complete_task(task, ISCSI_TASK_REQUEUE_SCSIQ);
|
2006-04-07 06:13:41 +04:00
|
|
|
fault:
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2009-03-05 23:45:58 +03:00
|
|
|
ISCSI_DBG_SESSION(session, "iscsi: cmd 0x%x is not queued (%d)\n",
|
|
|
|
sc->cmnd[0], reason);
|
2008-04-18 19:11:52 +04:00
|
|
|
if (!scsi_bidi_cmnd(sc))
|
|
|
|
scsi_set_resid(sc, scsi_bufflen(sc));
|
|
|
|
else {
|
|
|
|
scsi_out(sc)->resid = scsi_out(sc)->length;
|
|
|
|
scsi_in(sc)->resid = scsi_in(sc)->length;
|
|
|
|
}
|
2010-12-31 11:22:21 +03:00
|
|
|
sc->scsi_done(sc);
|
2006-04-07 06:13:41 +04:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_queuecommand);
|
|
|
|
|
2009-04-22 00:32:32 +04:00
|
|
|
int iscsi_target_alloc(struct scsi_target *starget)
|
|
|
|
{
|
|
|
|
struct iscsi_cls_session *cls_session = starget_to_session(starget);
|
|
|
|
struct iscsi_session *session = cls_session->dd_data;
|
|
|
|
|
|
|
|
starget->can_queue = session->scsi_cmds_max;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_target_alloc);
|
|
|
|
|
2007-12-13 21:43:20 +03:00
|
|
|
static void iscsi_tmf_timedout(unsigned long data)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
2007-12-13 21:43:20 +03:00
|
|
|
struct iscsi_conn *conn = (struct iscsi_conn *)data;
|
2006-04-07 06:13:41 +04:00
|
|
|
struct iscsi_session *session = conn->session;
|
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock(&session->frwd_lock);
|
2007-12-13 21:43:20 +03:00
|
|
|
if (conn->tmf_state == TMF_QUEUED) {
|
|
|
|
conn->tmf_state = TMF_TIMEDOUT;
|
2009-06-16 07:11:10 +04:00
|
|
|
ISCSI_DBG_EH(session, "tmf timedout\n");
|
2006-04-07 06:13:41 +04:00
|
|
|
/* unblock eh_abort() */
|
|
|
|
wake_up(&conn->ehwait);
|
|
|
|
}
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock(&session->frwd_lock);
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
static int iscsi_exec_task_mgmt_fn(struct iscsi_conn *conn,
|
2007-12-13 21:43:30 +03:00
|
|
|
struct iscsi_tm *hdr, int age,
|
|
|
|
int timeout)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
|
|
|
struct iscsi_session *session = conn->session;
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_task *task;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
task = __iscsi_conn_send_pdu(conn, (struct iscsi_hdr *)hdr,
|
2007-12-13 21:43:20 +03:00
|
|
|
NULL, 0);
|
2008-05-22 00:54:09 +04:00
|
|
|
if (!task) {
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2010-12-31 11:22:18 +03:00
|
|
|
iscsi_conn_printk(KERN_ERR, conn, "Could not send TMF.\n");
|
2006-04-07 06:13:41 +04:00
|
|
|
iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2007-05-30 21:57:18 +04:00
|
|
|
return -EPERM;
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
2007-12-13 21:43:20 +03:00
|
|
|
conn->tmfcmd_pdus_cnt++;
|
2007-12-13 21:43:30 +03:00
|
|
|
conn->tmf_timer.expires = timeout * HZ + jiffies;
|
2007-12-13 21:43:20 +03:00
|
|
|
conn->tmf_timer.function = iscsi_tmf_timedout;
|
|
|
|
conn->tmf_timer.data = (unsigned long)conn;
|
|
|
|
add_timer(&conn->tmf_timer);
|
2009-06-16 07:11:10 +04:00
|
|
|
ISCSI_DBG_EH(session, "tmf set timeout\n");
|
2006-04-07 06:13:41 +04:00
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2007-08-15 10:38:30 +04:00
|
|
|
mutex_unlock(&session->eh_mutex);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
|
|
|
/*
|
|
|
|
* block eh thread until:
|
|
|
|
*
|
2007-12-13 21:43:20 +03:00
|
|
|
* 1) tmf response
|
|
|
|
* 2) tmf timeout
|
2006-04-07 06:13:41 +04:00
|
|
|
* 3) session is terminated or restarted or userspace has
|
|
|
|
* given up on recovery
|
|
|
|
*/
|
2007-12-13 21:43:20 +03:00
|
|
|
wait_event_interruptible(conn->ehwait, age != session->age ||
|
2006-04-07 06:13:41 +04:00
|
|
|
session->state != ISCSI_STATE_LOGGED_IN ||
|
2007-12-13 21:43:20 +03:00
|
|
|
conn->tmf_state != TMF_QUEUED);
|
2006-04-07 06:13:41 +04:00
|
|
|
if (signal_pending(current))
|
|
|
|
flush_signals(current);
|
2007-12-13 21:43:20 +03:00
|
|
|
del_timer_sync(&conn->tmf_timer);
|
|
|
|
|
2007-08-15 10:38:30 +04:00
|
|
|
mutex_lock(&session->eh_mutex);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2008-05-22 00:54:09 +04:00
|
|
|
/* if the session drops it will clean up the task */
|
2007-12-13 21:43:20 +03:00
|
|
|
if (age != session->age ||
|
|
|
|
session->state != ISCSI_STATE_LOGGED_IN)
|
|
|
|
return -ENOTCONN;
|
2006-04-07 06:13:41 +04:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2007-12-13 21:43:20 +03:00
|
|
|
/*
|
|
|
|
* Fail commands. session lock held and recv side suspended and xmit
|
|
|
|
* thread flushed
|
|
|
|
*/
|
2014-06-25 17:27:36 +04:00
|
|
|
static void fail_scsi_tasks(struct iscsi_conn *conn, u64 lun, int error)
|
2007-12-13 21:43:20 +03:00
|
|
|
{
|
2009-05-14 02:57:46 +04:00
|
|
|
struct iscsi_task *task;
|
|
|
|
int i;
|
2007-12-13 21:43:20 +03:00
|
|
|
|
2009-05-14 02:57:46 +04:00
|
|
|
for (i = 0; i < conn->session->cmds_max; i++) {
|
|
|
|
task = conn->session->cmds[i];
|
|
|
|
if (!task->sc || task->state == ISCSI_TASK_FREE)
|
|
|
|
continue;
|
2007-12-13 21:43:20 +03:00
|
|
|
|
2009-05-14 02:57:46 +04:00
|
|
|
if (lun != -1 && lun != task->sc->device->lun)
|
|
|
|
continue;
|
2007-12-13 21:43:20 +03:00
|
|
|
|
2009-05-14 02:57:46 +04:00
|
|
|
ISCSI_DBG_SESSION(conn->session,
|
|
|
|
"failing sc %p itt 0x%x state %d\n",
|
|
|
|
task->sc, task->itt, task->state);
|
2009-05-14 02:57:49 +04:00
|
|
|
fail_scsi_task(task, error);
|
2007-12-13 21:43:20 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2009-09-05 06:05:33 +04:00
|
|
|
/**
|
|
|
|
* iscsi_suspend_queue - suspend iscsi_queuecommand
|
|
|
|
* @conn: iscsi conn to stop queueing IO on
|
|
|
|
*
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
* This grabs the session frwd_lock to make sure no one is in
|
2009-09-05 06:05:33 +04:00
|
|
|
* xmit_task/queuecommand, and then sets suspend to prevent
|
|
|
|
* new commands from being queued. This only needs to be called
|
|
|
|
* by offload drivers that need to sync a path like ep disconnect
|
|
|
|
* with the iscsi_queuecommand/xmit_task. To start IO again libiscsi
|
|
|
|
* will call iscsi_start_tx and iscsi_unblock_session when in FFP.
|
|
|
|
*/
|
|
|
|
void iscsi_suspend_queue(struct iscsi_conn *conn)
|
|
|
|
{
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&conn->session->frwd_lock);
|
2009-09-05 06:05:33 +04:00
|
|
|
set_bit(ISCSI_SUSPEND_BIT, &conn->suspend_tx);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&conn->session->frwd_lock);
|
2009-09-05 06:05:33 +04:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_suspend_queue);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* iscsi_suspend_tx - suspend iscsi_data_xmit
|
|
|
|
* @conn: iscsi conn tp stop processing IO on.
|
|
|
|
*
|
|
|
|
* This function sets the suspend bit to prevent iscsi_data_xmit
|
|
|
|
* from sending new IO, and if work is queued on the xmit thread
|
|
|
|
* it will wait for it to be completed.
|
|
|
|
*/
|
2008-05-22 00:54:03 +04:00
|
|
|
void iscsi_suspend_tx(struct iscsi_conn *conn)
|
2007-08-15 10:38:30 +04:00
|
|
|
{
|
2009-03-05 23:46:03 +03:00
|
|
|
struct Scsi_Host *shost = conn->session->host;
|
|
|
|
struct iscsi_host *ihost = shost_priv(shost);
|
|
|
|
|
2007-08-15 10:38:30 +04:00
|
|
|
set_bit(ISCSI_SUSPEND_BIT, &conn->suspend_tx);
|
2009-05-14 02:57:48 +04:00
|
|
|
if (ihost->workq)
|
2009-03-05 23:46:03 +03:00
|
|
|
flush_workqueue(ihost->workq);
|
2007-08-15 10:38:30 +04:00
|
|
|
}
|
2008-05-22 00:54:03 +04:00
|
|
|
EXPORT_SYMBOL_GPL(iscsi_suspend_tx);
|
2007-08-15 10:38:30 +04:00
|
|
|
|
|
|
|
static void iscsi_start_tx(struct iscsi_conn *conn)
|
|
|
|
{
|
|
|
|
clear_bit(ISCSI_SUSPEND_BIT, &conn->suspend_tx);
|
2009-05-14 02:57:48 +04:00
|
|
|
iscsi_conn_queue_work(conn);
|
2007-08-15 10:38:30 +04:00
|
|
|
}
|
|
|
|
|
2009-05-14 02:57:45 +04:00
|
|
|
/*
|
|
|
|
* We want to make sure a ping is in flight. It has timed out.
|
|
|
|
* And we are not busy processing a pdu that is making
|
|
|
|
* progress but got started before the ping and is taking a while
|
|
|
|
* to complete so the ping is just stuck behind it in a queue.
|
|
|
|
*/
|
|
|
|
static int iscsi_has_ping_timed_out(struct iscsi_conn *conn)
|
|
|
|
{
|
|
|
|
if (conn->ping_task &&
|
|
|
|
time_before_eq(conn->last_recv + (conn->recv_timeout * HZ) +
|
|
|
|
(conn->ping_timeout * HZ), jiffies))
|
|
|
|
return 1;
|
|
|
|
else
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2009-06-16 07:11:08 +04:00
|
|
|
static enum blk_eh_timer_return iscsi_eh_cmd_timed_out(struct scsi_cmnd *sc)
|
2007-12-13 21:43:30 +03:00
|
|
|
{
|
2009-06-16 07:11:08 +04:00
|
|
|
enum blk_eh_timer_return rc = BLK_EH_NOT_HANDLED;
|
2010-02-11 01:51:45 +03:00
|
|
|
struct iscsi_task *task = NULL, *running_task;
|
2007-12-13 21:43:30 +03:00
|
|
|
struct iscsi_cls_session *cls_session;
|
|
|
|
struct iscsi_session *session;
|
|
|
|
struct iscsi_conn *conn;
|
2010-02-11 01:51:45 +03:00
|
|
|
int i;
|
2007-12-13 21:43:30 +03:00
|
|
|
|
2009-06-16 07:11:08 +04:00
|
|
|
cls_session = starget_to_session(scsi_target(sc->device));
|
2008-05-22 00:53:59 +04:00
|
|
|
session = cls_session->dd_data;
|
2007-12-13 21:43:30 +03:00
|
|
|
|
2009-06-16 07:11:10 +04:00
|
|
|
ISCSI_DBG_EH(session, "scsi cmd %p timedout\n", sc);
|
2007-12-13 21:43:30 +03:00
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock(&session->frwd_lock);
|
2012-01-27 07:13:11 +04:00
|
|
|
task = (struct iscsi_task *)sc->SCp.ptr;
|
|
|
|
if (!task) {
|
|
|
|
/*
|
|
|
|
* Raced with completion. Blk layer has taken ownership
|
|
|
|
* so let timeout code complete it now.
|
|
|
|
*/
|
|
|
|
rc = BLK_EH_HANDLED;
|
|
|
|
goto done;
|
|
|
|
}
|
|
|
|
|
2007-12-13 21:43:30 +03:00
|
|
|
if (session->state != ISCSI_STATE_LOGGED_IN) {
|
|
|
|
/*
|
|
|
|
* We are probably in the middle of iscsi recovery so let
|
|
|
|
* that complete and handle the error.
|
|
|
|
*/
|
2008-09-14 16:55:09 +04:00
|
|
|
rc = BLK_EH_RESET_TIMER;
|
2007-12-13 21:43:30 +03:00
|
|
|
goto done;
|
|
|
|
}
|
|
|
|
|
|
|
|
conn = session->leadconn;
|
|
|
|
if (!conn) {
|
|
|
|
/* In the middle of shuting down */
|
2008-09-14 16:55:09 +04:00
|
|
|
rc = BLK_EH_RESET_TIMER;
|
2007-12-13 21:43:30 +03:00
|
|
|
goto done;
|
|
|
|
}
|
|
|
|
|
2009-06-16 07:11:08 +04:00
|
|
|
/*
|
|
|
|
* If we have sent (at least queued to the network layer) a pdu or
|
|
|
|
* recvd one for the task since the last timeout ask for
|
|
|
|
* more time. If on the next timeout we have not made progress
|
|
|
|
* we can check if it is the task or connection when we send the
|
|
|
|
* nop as a ping.
|
|
|
|
*/
|
2010-02-11 01:51:45 +03:00
|
|
|
if (time_after(task->last_xfer, task->last_timeout)) {
|
2009-06-16 07:11:10 +04:00
|
|
|
ISCSI_DBG_EH(session, "Command making progress. Asking "
|
|
|
|
"scsi-ml for more time to complete. "
|
2010-02-11 01:51:45 +03:00
|
|
|
"Last data xfer at %lu. Last timeout was at "
|
2009-06-16 07:11:10 +04:00
|
|
|
"%lu\n.", task->last_xfer, task->last_timeout);
|
2009-06-16 07:11:08 +04:00
|
|
|
task->have_checked_conn = false;
|
|
|
|
rc = BLK_EH_RESET_TIMER;
|
|
|
|
goto done;
|
|
|
|
}
|
|
|
|
|
2007-12-13 21:43:30 +03:00
|
|
|
if (!conn->recv_timeout && !conn->ping_timeout)
|
|
|
|
goto done;
|
|
|
|
/*
|
|
|
|
* if the ping timedout then we are in the middle of cleaning up
|
|
|
|
* and can let the iscsi eh handle it
|
|
|
|
*/
|
2009-05-14 02:57:45 +04:00
|
|
|
if (iscsi_has_ping_timed_out(conn)) {
|
2008-09-14 16:55:09 +04:00
|
|
|
rc = BLK_EH_RESET_TIMER;
|
2009-05-14 02:57:45 +04:00
|
|
|
goto done;
|
|
|
|
}
|
2009-06-16 07:11:08 +04:00
|
|
|
|
2010-02-11 01:51:45 +03:00
|
|
|
for (i = 0; i < conn->session->cmds_max; i++) {
|
|
|
|
running_task = conn->session->cmds[i];
|
|
|
|
if (!running_task->sc || running_task == task ||
|
|
|
|
running_task->state != ISCSI_TASK_RUNNING)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Only check if cmds started before this one have made
|
|
|
|
* progress, or this could never fail
|
|
|
|
*/
|
|
|
|
if (time_after(running_task->sc->jiffies_at_alloc,
|
|
|
|
task->sc->jiffies_at_alloc))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (time_after(running_task->last_xfer, task->last_timeout)) {
|
|
|
|
/*
|
|
|
|
* This task has not made progress, but a task
|
|
|
|
* started before us has transferred data since
|
|
|
|
* we started/last-checked. We could be queueing
|
|
|
|
* too many tasks or the LU is bad.
|
|
|
|
*
|
|
|
|
* If the device is bad the cmds ahead of us on
|
|
|
|
* other devs will complete, and this loop will
|
|
|
|
* eventually fail starting the scsi eh.
|
|
|
|
*/
|
|
|
|
ISCSI_DBG_EH(session, "Command has not made progress "
|
|
|
|
"but commands ahead of it have. "
|
|
|
|
"Asking scsi-ml for more time to "
|
|
|
|
"complete. Our last xfer vs running task "
|
|
|
|
"last xfer %lu/%lu. Last check %lu.\n",
|
|
|
|
task->last_xfer, running_task->last_xfer,
|
|
|
|
task->last_timeout);
|
|
|
|
rc = BLK_EH_RESET_TIMER;
|
|
|
|
goto done;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2009-06-16 07:11:08 +04:00
|
|
|
/* Assumes nop timeout is shorter than scsi cmd timeout */
|
|
|
|
if (task->have_checked_conn)
|
|
|
|
goto done;
|
|
|
|
|
2007-12-13 21:43:30 +03:00
|
|
|
/*
|
2009-06-16 07:11:08 +04:00
|
|
|
* Checking the transport already or nop from a cmd timeout still
|
|
|
|
* running
|
2007-12-13 21:43:30 +03:00
|
|
|
*/
|
2009-06-16 07:11:08 +04:00
|
|
|
if (conn->ping_task) {
|
|
|
|
task->have_checked_conn = true;
|
2008-09-14 16:55:09 +04:00
|
|
|
rc = BLK_EH_RESET_TIMER;
|
2009-05-14 02:57:45 +04:00
|
|
|
goto done;
|
|
|
|
}
|
|
|
|
|
2009-06-16 07:11:08 +04:00
|
|
|
/* Make sure there is a transport check done */
|
|
|
|
iscsi_send_nopout(conn, NULL);
|
|
|
|
task->have_checked_conn = true;
|
|
|
|
rc = BLK_EH_RESET_TIMER;
|
|
|
|
|
2007-12-13 21:43:30 +03:00
|
|
|
done:
|
2009-06-16 07:11:08 +04:00
|
|
|
if (task)
|
|
|
|
task->last_timeout = jiffies;
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock(&session->frwd_lock);
|
2009-06-16 07:11:10 +04:00
|
|
|
ISCSI_DBG_EH(session, "return %s\n", rc == BLK_EH_RESET_TIMER ?
|
|
|
|
"timer reset" : "nh");
|
2007-12-13 21:43:30 +03:00
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void iscsi_check_transport_timeouts(unsigned long data)
|
|
|
|
{
|
|
|
|
struct iscsi_conn *conn = (struct iscsi_conn *)data;
|
|
|
|
struct iscsi_session *session = conn->session;
|
2008-05-08 05:43:52 +04:00
|
|
|
unsigned long recv_timeout, next_timeout = 0, last_recv;
|
2007-12-13 21:43:30 +03:00
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock(&session->frwd_lock);
|
2007-12-13 21:43:30 +03:00
|
|
|
if (session->state != ISCSI_STATE_LOGGED_IN)
|
|
|
|
goto done;
|
|
|
|
|
2008-05-08 05:43:52 +04:00
|
|
|
recv_timeout = conn->recv_timeout;
|
|
|
|
if (!recv_timeout)
|
2007-12-13 21:43:30 +03:00
|
|
|
goto done;
|
|
|
|
|
2008-05-08 05:43:52 +04:00
|
|
|
recv_timeout *= HZ;
|
2007-12-13 21:43:30 +03:00
|
|
|
last_recv = conn->last_recv;
|
2009-05-14 02:57:45 +04:00
|
|
|
|
|
|
|
if (iscsi_has_ping_timed_out(conn)) {
|
2008-01-31 22:36:52 +03:00
|
|
|
iscsi_conn_printk(KERN_ERR, conn, "ping timeout of %d secs "
|
2009-05-14 02:57:45 +04:00
|
|
|
"expired, recv timeout %d, last rx %lu, "
|
|
|
|
"last ping %lu, now %lu\n",
|
|
|
|
conn->ping_timeout, conn->recv_timeout,
|
|
|
|
last_recv, conn->last_ping, jiffies);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock(&session->frwd_lock);
|
2014-07-13 00:51:51 +04:00
|
|
|
iscsi_conn_failure(conn, ISCSI_ERR_NOP_TIMEDOUT);
|
2007-12-13 21:43:30 +03:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2008-05-08 05:43:52 +04:00
|
|
|
if (time_before_eq(last_recv + recv_timeout, jiffies)) {
|
2008-05-09 05:15:34 +04:00
|
|
|
/* send a ping to try to provoke some traffic */
|
2009-03-05 23:45:58 +03:00
|
|
|
ISCSI_DBG_CONN(conn, "Sending nopout as ping\n");
|
2015-09-03 19:49:55 +03:00
|
|
|
if (iscsi_send_nopout(conn, NULL))
|
|
|
|
next_timeout = jiffies + (1 * HZ);
|
|
|
|
else
|
|
|
|
next_timeout = conn->last_ping + (conn->ping_timeout * HZ);
|
2008-01-31 22:36:50 +03:00
|
|
|
} else
|
2008-05-08 05:43:52 +04:00
|
|
|
next_timeout = last_recv + recv_timeout;
|
2007-12-13 21:43:30 +03:00
|
|
|
|
2009-03-05 23:45:58 +03:00
|
|
|
ISCSI_DBG_CONN(conn, "Setting next tmo %lu\n", next_timeout);
|
2008-01-31 22:36:50 +03:00
|
|
|
mod_timer(&conn->transport_timer, next_timeout);
|
2007-12-13 21:43:30 +03:00
|
|
|
done:
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock(&session->frwd_lock);
|
2007-12-13 21:43:30 +03:00
|
|
|
}
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
static void iscsi_prep_abort_task_pdu(struct iscsi_task *task,
|
2007-12-13 21:43:20 +03:00
|
|
|
struct iscsi_tm *hdr)
|
|
|
|
{
|
|
|
|
memset(hdr, 0, sizeof(*hdr));
|
|
|
|
hdr->opcode = ISCSI_OP_SCSI_TMFUNC | ISCSI_OP_IMMEDIATE;
|
|
|
|
hdr->flags = ISCSI_TM_FUNC_ABORT_TASK & ISCSI_FLAG_TM_FUNC_MASK;
|
|
|
|
hdr->flags |= ISCSI_FLAG_CMD_FINAL;
|
2011-06-17 02:57:09 +04:00
|
|
|
hdr->lun = task->lun;
|
2008-12-02 09:32:05 +03:00
|
|
|
hdr->rtt = task->hdr_itt;
|
|
|
|
hdr->refcmdsn = task->cmdsn;
|
2007-12-13 21:43:20 +03:00
|
|
|
}
|
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
int iscsi_eh_abort(struct scsi_cmnd *sc)
|
|
|
|
{
|
2008-05-22 00:53:59 +04:00
|
|
|
struct iscsi_cls_session *cls_session;
|
|
|
|
struct iscsi_session *session;
|
2006-09-01 02:09:33 +04:00
|
|
|
struct iscsi_conn *conn;
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_task *task;
|
2007-12-13 21:43:20 +03:00
|
|
|
struct iscsi_tm *hdr;
|
2016-03-30 21:27:08 +03:00
|
|
|
int age;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:53:59 +04:00
|
|
|
cls_session = starget_to_session(scsi_target(sc->device));
|
|
|
|
session = cls_session->dd_data;
|
|
|
|
|
2009-06-16 07:11:10 +04:00
|
|
|
ISCSI_DBG_EH(session, "aborting sc %p\n", sc);
|
2009-05-14 02:57:50 +04:00
|
|
|
|
2007-08-15 10:38:30 +04:00
|
|
|
mutex_lock(&session->eh_mutex);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2006-09-01 02:09:33 +04:00
|
|
|
/*
|
|
|
|
* if session was ISCSI_STATE_IN_RECOVERY then we may not have
|
|
|
|
* got the command.
|
|
|
|
*/
|
|
|
|
if (!sc->SCp.ptr) {
|
2009-06-16 07:11:10 +04:00
|
|
|
ISCSI_DBG_EH(session, "sc never reached iscsi layer or "
|
|
|
|
"it completed.\n");
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2007-08-15 10:38:30 +04:00
|
|
|
mutex_unlock(&session->eh_mutex);
|
2006-09-01 02:09:33 +04:00
|
|
|
return SUCCESS;
|
|
|
|
}
|
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
/*
|
|
|
|
* If we are not logged in or we have started a new session
|
|
|
|
* then let the host reset code handle this
|
|
|
|
*/
|
2007-12-13 21:43:20 +03:00
|
|
|
if (!session->leadconn || session->state != ISCSI_STATE_LOGGED_IN ||
|
|
|
|
sc->SCp.phase != session->age) {
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2007-12-13 21:43:20 +03:00
|
|
|
mutex_unlock(&session->eh_mutex);
|
2009-06-16 07:11:10 +04:00
|
|
|
ISCSI_DBG_EH(session, "failing abort due to dropped "
|
2009-05-14 02:57:50 +04:00
|
|
|
"session.\n");
|
2007-12-13 21:43:20 +03:00
|
|
|
return FAILED;
|
|
|
|
}
|
|
|
|
|
|
|
|
conn = session->leadconn;
|
|
|
|
conn->eh_abort_cnt++;
|
|
|
|
age = session->age;
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
task = (struct iscsi_task *)sc->SCp.ptr;
|
2009-06-16 07:11:10 +04:00
|
|
|
ISCSI_DBG_EH(session, "aborting [sc %p itt 0x%x]\n",
|
|
|
|
sc, task->itt);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
/* task completed before time out */
|
|
|
|
if (!task->sc) {
|
2009-06-16 07:11:10 +04:00
|
|
|
ISCSI_DBG_EH(session, "sc completed while abort in progress\n");
|
2007-05-30 21:57:18 +04:00
|
|
|
goto success;
|
2006-07-25 00:47:22 +04:00
|
|
|
}
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
if (task->state == ISCSI_TASK_PENDING) {
|
2009-05-14 02:57:49 +04:00
|
|
|
fail_scsi_task(task, DID_ABORT);
|
2007-05-30 21:57:18 +04:00
|
|
|
goto success;
|
|
|
|
}
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2007-12-13 21:43:20 +03:00
|
|
|
/* only have one tmf outstanding at a time */
|
|
|
|
if (conn->tmf_state != TMF_INITIAL)
|
2006-04-07 06:13:41 +04:00
|
|
|
goto failed;
|
2007-12-13 21:43:20 +03:00
|
|
|
conn->tmf_state = TMF_QUEUED;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2007-12-13 21:43:20 +03:00
|
|
|
hdr = &conn->tmhdr;
|
2008-05-22 00:54:09 +04:00
|
|
|
iscsi_prep_abort_task_pdu(task, hdr);
|
2007-12-13 21:43:20 +03:00
|
|
|
|
2016-03-30 21:27:08 +03:00
|
|
|
if (iscsi_exec_task_mgmt_fn(conn, hdr, age, session->abort_timeout))
|
2007-12-13 21:43:20 +03:00
|
|
|
goto failed;
|
|
|
|
|
|
|
|
switch (conn->tmf_state) {
|
|
|
|
case TMF_SUCCESS:
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2008-05-22 00:54:18 +04:00
|
|
|
/*
|
|
|
|
* stop tx side incase the target had sent a abort rsp but
|
|
|
|
* the initiator was still writing out data.
|
|
|
|
*/
|
2007-08-15 10:38:30 +04:00
|
|
|
iscsi_suspend_tx(conn);
|
2007-05-30 21:57:18 +04:00
|
|
|
/*
|
2008-05-22 00:54:18 +04:00
|
|
|
* we do not stop the recv side because targets have been
|
|
|
|
* good and have never sent us a successful tmf response
|
|
|
|
* then sent more data for the cmd.
|
2007-05-30 21:57:18 +04:00
|
|
|
*/
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2009-05-14 02:57:49 +04:00
|
|
|
fail_scsi_task(task, DID_ABORT);
|
2007-12-13 21:43:20 +03:00
|
|
|
conn->tmf_state = TMF_INITIAL;
|
2009-11-12 01:34:32 +03:00
|
|
|
memset(hdr, 0, sizeof(*hdr));
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2007-08-15 10:38:30 +04:00
|
|
|
iscsi_start_tx(conn);
|
2007-05-30 21:57:18 +04:00
|
|
|
goto success_unlocked;
|
2007-12-13 21:43:20 +03:00
|
|
|
case TMF_TIMEDOUT:
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2010-12-31 11:22:18 +03:00
|
|
|
iscsi_conn_failure(conn, ISCSI_ERR_SCSI_EH_SESSION_RST);
|
2007-12-13 21:43:20 +03:00
|
|
|
goto failed_unlocked;
|
|
|
|
case TMF_NOT_FOUND:
|
|
|
|
if (!sc->SCp.ptr) {
|
|
|
|
conn->tmf_state = TMF_INITIAL;
|
2009-11-12 01:34:32 +03:00
|
|
|
memset(hdr, 0, sizeof(*hdr));
|
2008-05-22 00:54:09 +04:00
|
|
|
/* task completed before tmf abort response */
|
2009-06-16 07:11:10 +04:00
|
|
|
ISCSI_DBG_EH(session, "sc completed while abort in "
|
|
|
|
"progress\n");
|
2007-05-30 21:57:18 +04:00
|
|
|
goto success;
|
2006-07-25 00:47:22 +04:00
|
|
|
}
|
|
|
|
/* fall through */
|
|
|
|
default:
|
2007-12-13 21:43:20 +03:00
|
|
|
conn->tmf_state = TMF_INITIAL;
|
|
|
|
goto failed;
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
|
2007-05-30 21:57:18 +04:00
|
|
|
success:
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2007-05-30 21:57:18 +04:00
|
|
|
success_unlocked:
|
2009-06-16 07:11:10 +04:00
|
|
|
ISCSI_DBG_EH(session, "abort success [sc %p itt 0x%x]\n",
|
|
|
|
sc, task->itt);
|
2007-08-15 10:38:30 +04:00
|
|
|
mutex_unlock(&session->eh_mutex);
|
2006-04-07 06:13:41 +04:00
|
|
|
return SUCCESS;
|
|
|
|
|
|
|
|
failed:
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2007-05-30 21:57:18 +04:00
|
|
|
failed_unlocked:
|
2009-06-16 07:11:10 +04:00
|
|
|
ISCSI_DBG_EH(session, "abort failed [sc %p itt 0x%x]\n", sc,
|
|
|
|
task ? task->itt : 0);
|
2007-08-15 10:38:30 +04:00
|
|
|
mutex_unlock(&session->eh_mutex);
|
2006-04-07 06:13:41 +04:00
|
|
|
return FAILED;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_eh_abort);
|
|
|
|
|
2007-12-13 21:43:20 +03:00
|
|
|
static void iscsi_prep_lun_reset_pdu(struct scsi_cmnd *sc, struct iscsi_tm *hdr)
|
|
|
|
{
|
|
|
|
memset(hdr, 0, sizeof(*hdr));
|
|
|
|
hdr->opcode = ISCSI_OP_SCSI_TMFUNC | ISCSI_OP_IMMEDIATE;
|
|
|
|
hdr->flags = ISCSI_TM_FUNC_LOGICAL_UNIT_RESET & ISCSI_FLAG_TM_FUNC_MASK;
|
|
|
|
hdr->flags |= ISCSI_FLAG_CMD_FINAL;
|
2011-06-17 02:57:09 +04:00
|
|
|
int_to_scsilun(sc->device->lun, &hdr->lun);
|
2007-12-13 21:43:30 +03:00
|
|
|
hdr->rtt = RESERVED_ITT;
|
2007-12-13 21:43:20 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
int iscsi_eh_device_reset(struct scsi_cmnd *sc)
|
|
|
|
{
|
2008-05-22 00:53:59 +04:00
|
|
|
struct iscsi_cls_session *cls_session;
|
|
|
|
struct iscsi_session *session;
|
2007-12-13 21:43:20 +03:00
|
|
|
struct iscsi_conn *conn;
|
|
|
|
struct iscsi_tm *hdr;
|
|
|
|
int rc = FAILED;
|
|
|
|
|
2008-05-22 00:53:59 +04:00
|
|
|
cls_session = starget_to_session(scsi_target(sc->device));
|
|
|
|
session = cls_session->dd_data;
|
|
|
|
|
2014-06-25 17:27:36 +04:00
|
|
|
ISCSI_DBG_EH(session, "LU Reset [sc %p lun %llu]\n", sc,
|
|
|
|
sc->device->lun);
|
2007-12-13 21:43:20 +03:00
|
|
|
|
|
|
|
mutex_lock(&session->eh_mutex);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2007-12-13 21:43:20 +03:00
|
|
|
/*
|
|
|
|
* Just check if we are not logged in. We cannot check for
|
|
|
|
* the phase because the reset could come from a ioctl.
|
|
|
|
*/
|
|
|
|
if (!session->leadconn || session->state != ISCSI_STATE_LOGGED_IN)
|
|
|
|
goto unlock;
|
|
|
|
conn = session->leadconn;
|
|
|
|
|
|
|
|
/* only have one tmf outstanding at a time */
|
|
|
|
if (conn->tmf_state != TMF_INITIAL)
|
|
|
|
goto unlock;
|
|
|
|
conn->tmf_state = TMF_QUEUED;
|
|
|
|
|
|
|
|
hdr = &conn->tmhdr;
|
|
|
|
iscsi_prep_lun_reset_pdu(sc, hdr);
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
if (iscsi_exec_task_mgmt_fn(conn, hdr, session->age,
|
2007-12-13 21:43:30 +03:00
|
|
|
session->lu_reset_timeout)) {
|
2007-12-13 21:43:20 +03:00
|
|
|
rc = FAILED;
|
|
|
|
goto unlock;
|
|
|
|
}
|
|
|
|
|
|
|
|
switch (conn->tmf_state) {
|
|
|
|
case TMF_SUCCESS:
|
|
|
|
break;
|
|
|
|
case TMF_TIMEDOUT:
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2010-12-31 11:22:18 +03:00
|
|
|
iscsi_conn_failure(conn, ISCSI_ERR_SCSI_EH_SESSION_RST);
|
2007-12-13 21:43:20 +03:00
|
|
|
goto done;
|
|
|
|
default:
|
|
|
|
conn->tmf_state = TMF_INITIAL;
|
|
|
|
goto unlock;
|
|
|
|
}
|
|
|
|
|
|
|
|
rc = SUCCESS;
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2007-12-13 21:43:20 +03:00
|
|
|
|
|
|
|
iscsi_suspend_tx(conn);
|
2008-05-22 00:54:18 +04:00
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2009-11-12 01:34:32 +03:00
|
|
|
memset(hdr, 0, sizeof(*hdr));
|
2009-05-14 02:57:46 +04:00
|
|
|
fail_scsi_tasks(conn, sc->device->lun, DID_ERROR);
|
2007-12-13 21:43:20 +03:00
|
|
|
conn->tmf_state = TMF_INITIAL;
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2007-12-13 21:43:20 +03:00
|
|
|
|
|
|
|
iscsi_start_tx(conn);
|
|
|
|
goto done;
|
|
|
|
|
|
|
|
unlock:
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2007-12-13 21:43:20 +03:00
|
|
|
done:
|
2009-06-16 07:11:10 +04:00
|
|
|
ISCSI_DBG_EH(session, "dev reset result = %s\n",
|
|
|
|
rc == SUCCESS ? "SUCCESS" : "FAILED");
|
2007-12-13 21:43:20 +03:00
|
|
|
mutex_unlock(&session->eh_mutex);
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_eh_device_reset);
|
|
|
|
|
2009-11-12 01:34:33 +03:00
|
|
|
void iscsi_session_recovery_timedout(struct iscsi_cls_session *cls_session)
|
|
|
|
{
|
|
|
|
struct iscsi_session *session = cls_session->dd_data;
|
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2009-11-12 01:34:33 +03:00
|
|
|
if (session->state != ISCSI_STATE_LOGGED_IN) {
|
|
|
|
session->state = ISCSI_STATE_RECOVERY_FAILED;
|
|
|
|
if (session->leadconn)
|
|
|
|
wake_up(&session->leadconn->ehwait);
|
|
|
|
}
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2009-11-12 01:34:33 +03:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_session_recovery_timedout);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* iscsi_eh_session_reset - drop session and attempt relogin
|
|
|
|
* @sc: scsi command
|
|
|
|
*
|
|
|
|
* This function will wait for a relogin, session termination from
|
|
|
|
* userspace, or a recovery/replacement timeout.
|
|
|
|
*/
|
2010-02-20 05:32:10 +03:00
|
|
|
int iscsi_eh_session_reset(struct scsi_cmnd *sc)
|
2009-11-12 01:34:33 +03:00
|
|
|
{
|
|
|
|
struct iscsi_cls_session *cls_session;
|
|
|
|
struct iscsi_session *session;
|
|
|
|
struct iscsi_conn *conn;
|
|
|
|
|
|
|
|
cls_session = starget_to_session(scsi_target(sc->device));
|
|
|
|
session = cls_session->dd_data;
|
|
|
|
conn = session->leadconn;
|
|
|
|
|
|
|
|
mutex_lock(&session->eh_mutex);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2009-11-12 01:34:33 +03:00
|
|
|
if (session->state == ISCSI_STATE_TERMINATE) {
|
|
|
|
failed:
|
|
|
|
ISCSI_DBG_EH(session,
|
|
|
|
"failing session reset: Could not log back into "
|
|
|
|
"%s, %s [age %d]\n", session->targetname,
|
|
|
|
conn->persistent_address, session->age);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2009-11-12 01:34:33 +03:00
|
|
|
mutex_unlock(&session->eh_mutex);
|
|
|
|
return FAILED;
|
|
|
|
}
|
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2009-11-12 01:34:33 +03:00
|
|
|
mutex_unlock(&session->eh_mutex);
|
|
|
|
/*
|
|
|
|
* we drop the lock here but the leadconn cannot be destoyed while
|
|
|
|
* we are in the scsi eh
|
|
|
|
*/
|
2010-12-31 11:22:18 +03:00
|
|
|
iscsi_conn_failure(conn, ISCSI_ERR_SCSI_EH_SESSION_RST);
|
2009-11-12 01:34:33 +03:00
|
|
|
|
|
|
|
ISCSI_DBG_EH(session, "wait for relogin\n");
|
|
|
|
wait_event_interruptible(conn->ehwait,
|
|
|
|
session->state == ISCSI_STATE_TERMINATE ||
|
|
|
|
session->state == ISCSI_STATE_LOGGED_IN ||
|
|
|
|
session->state == ISCSI_STATE_RECOVERY_FAILED);
|
|
|
|
if (signal_pending(current))
|
|
|
|
flush_signals(current);
|
|
|
|
|
|
|
|
mutex_lock(&session->eh_mutex);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2009-11-12 01:34:33 +03:00
|
|
|
if (session->state == ISCSI_STATE_LOGGED_IN) {
|
|
|
|
ISCSI_DBG_EH(session,
|
|
|
|
"session reset succeeded for %s,%s\n",
|
|
|
|
session->targetname, conn->persistent_address);
|
|
|
|
} else
|
|
|
|
goto failed;
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2009-11-12 01:34:33 +03:00
|
|
|
mutex_unlock(&session->eh_mutex);
|
|
|
|
return SUCCESS;
|
|
|
|
}
|
2010-02-20 05:32:10 +03:00
|
|
|
EXPORT_SYMBOL_GPL(iscsi_eh_session_reset);
|
2009-11-12 01:34:33 +03:00
|
|
|
|
|
|
|
static void iscsi_prep_tgt_reset_pdu(struct scsi_cmnd *sc, struct iscsi_tm *hdr)
|
|
|
|
{
|
|
|
|
memset(hdr, 0, sizeof(*hdr));
|
|
|
|
hdr->opcode = ISCSI_OP_SCSI_TMFUNC | ISCSI_OP_IMMEDIATE;
|
|
|
|
hdr->flags = ISCSI_TM_FUNC_TARGET_WARM_RESET & ISCSI_FLAG_TM_FUNC_MASK;
|
|
|
|
hdr->flags |= ISCSI_FLAG_CMD_FINAL;
|
|
|
|
hdr->rtt = RESERVED_ITT;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* iscsi_eh_target_reset - reset target
|
|
|
|
* @sc: scsi command
|
|
|
|
*
|
2010-02-20 05:32:10 +03:00
|
|
|
* This will attempt to send a warm target reset.
|
2009-11-12 01:34:33 +03:00
|
|
|
*/
|
2016-03-30 21:26:46 +03:00
|
|
|
static int iscsi_eh_target_reset(struct scsi_cmnd *sc)
|
2009-11-12 01:34:33 +03:00
|
|
|
{
|
|
|
|
struct iscsi_cls_session *cls_session;
|
|
|
|
struct iscsi_session *session;
|
|
|
|
struct iscsi_conn *conn;
|
|
|
|
struct iscsi_tm *hdr;
|
|
|
|
int rc = FAILED;
|
|
|
|
|
|
|
|
cls_session = starget_to_session(scsi_target(sc->device));
|
|
|
|
session = cls_session->dd_data;
|
|
|
|
|
|
|
|
ISCSI_DBG_EH(session, "tgt Reset [sc %p tgt %s]\n", sc,
|
|
|
|
session->targetname);
|
|
|
|
|
|
|
|
mutex_lock(&session->eh_mutex);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2009-11-12 01:34:33 +03:00
|
|
|
/*
|
|
|
|
* Just check if we are not logged in. We cannot check for
|
|
|
|
* the phase because the reset could come from a ioctl.
|
|
|
|
*/
|
|
|
|
if (!session->leadconn || session->state != ISCSI_STATE_LOGGED_IN)
|
|
|
|
goto unlock;
|
|
|
|
conn = session->leadconn;
|
|
|
|
|
|
|
|
/* only have one tmf outstanding at a time */
|
|
|
|
if (conn->tmf_state != TMF_INITIAL)
|
|
|
|
goto unlock;
|
|
|
|
conn->tmf_state = TMF_QUEUED;
|
|
|
|
|
|
|
|
hdr = &conn->tmhdr;
|
|
|
|
iscsi_prep_tgt_reset_pdu(sc, hdr);
|
|
|
|
|
|
|
|
if (iscsi_exec_task_mgmt_fn(conn, hdr, session->age,
|
|
|
|
session->tgt_reset_timeout)) {
|
|
|
|
rc = FAILED;
|
|
|
|
goto unlock;
|
|
|
|
}
|
|
|
|
|
|
|
|
switch (conn->tmf_state) {
|
|
|
|
case TMF_SUCCESS:
|
|
|
|
break;
|
|
|
|
case TMF_TIMEDOUT:
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2010-12-31 11:22:18 +03:00
|
|
|
iscsi_conn_failure(conn, ISCSI_ERR_SCSI_EH_SESSION_RST);
|
2009-11-12 01:34:33 +03:00
|
|
|
goto done;
|
|
|
|
default:
|
|
|
|
conn->tmf_state = TMF_INITIAL;
|
|
|
|
goto unlock;
|
|
|
|
}
|
|
|
|
|
|
|
|
rc = SUCCESS;
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2009-11-12 01:34:33 +03:00
|
|
|
|
|
|
|
iscsi_suspend_tx(conn);
|
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2009-11-12 01:34:33 +03:00
|
|
|
memset(hdr, 0, sizeof(*hdr));
|
|
|
|
fail_scsi_tasks(conn, -1, DID_ERROR);
|
|
|
|
conn->tmf_state = TMF_INITIAL;
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2009-11-12 01:34:33 +03:00
|
|
|
|
|
|
|
iscsi_start_tx(conn);
|
|
|
|
goto done;
|
|
|
|
|
|
|
|
unlock:
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2009-11-12 01:34:33 +03:00
|
|
|
done:
|
|
|
|
ISCSI_DBG_EH(session, "tgt %s reset result = %s\n", session->targetname,
|
|
|
|
rc == SUCCESS ? "SUCCESS" : "FAILED");
|
|
|
|
mutex_unlock(&session->eh_mutex);
|
2010-02-20 05:32:10 +03:00
|
|
|
return rc;
|
|
|
|
}
|
2009-11-12 01:34:33 +03:00
|
|
|
|
2010-02-20 05:32:10 +03:00
|
|
|
/**
|
|
|
|
* iscsi_eh_recover_target - reset target and possibly the session
|
|
|
|
* @sc: scsi command
|
|
|
|
*
|
|
|
|
* This will attempt to send a warm target reset. If that fails,
|
|
|
|
* we will escalate to ERL0 session recovery.
|
|
|
|
*/
|
|
|
|
int iscsi_eh_recover_target(struct scsi_cmnd *sc)
|
|
|
|
{
|
|
|
|
int rc;
|
|
|
|
|
|
|
|
rc = iscsi_eh_target_reset(sc);
|
2009-11-12 01:34:33 +03:00
|
|
|
if (rc == FAILED)
|
|
|
|
rc = iscsi_eh_session_reset(sc);
|
|
|
|
return rc;
|
|
|
|
}
|
2010-02-20 05:32:10 +03:00
|
|
|
EXPORT_SYMBOL_GPL(iscsi_eh_recover_target);
|
2009-11-12 01:34:33 +03:00
|
|
|
|
2007-12-13 21:43:25 +03:00
|
|
|
/*
|
|
|
|
* Pre-allocate a pool of @max items of @item_size. By default, the pool
|
|
|
|
* should be accessed via kfifo_{get,put} on q->queue.
|
|
|
|
* Optionally, the caller can obtain the array of object pointers
|
|
|
|
* by passing in a non-NULL @items pointer
|
|
|
|
*/
|
2006-04-07 06:13:41 +04:00
|
|
|
int
|
2007-12-13 21:43:25 +03:00
|
|
|
iscsi_pool_init(struct iscsi_pool *q, int max, void ***items, int item_size)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
2007-12-13 21:43:25 +03:00
|
|
|
int i, num_arrays = 1;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2007-12-13 21:43:25 +03:00
|
|
|
memset(q, 0, sizeof(*q));
|
2006-04-07 06:13:41 +04:00
|
|
|
|
|
|
|
q->max = max;
|
2007-12-13 21:43:25 +03:00
|
|
|
|
|
|
|
/* If the user passed an items pointer, he wants a copy of
|
|
|
|
* the array. */
|
|
|
|
if (items)
|
|
|
|
num_arrays++;
|
|
|
|
q->pool = kzalloc(num_arrays * max * sizeof(void*), GFP_KERNEL);
|
|
|
|
if (q->pool == NULL)
|
2009-03-05 23:45:55 +03:00
|
|
|
return -ENOMEM;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2009-12-22 01:37:27 +03:00
|
|
|
kfifo_init(&q->queue, (void*)q->pool, max * sizeof(void*));
|
2006-04-07 06:13:41 +04:00
|
|
|
|
|
|
|
for (i = 0; i < max; i++) {
|
2007-12-13 21:43:25 +03:00
|
|
|
q->pool[i] = kzalloc(item_size, GFP_KERNEL);
|
2006-04-07 06:13:41 +04:00
|
|
|
if (q->pool[i] == NULL) {
|
2007-12-13 21:43:25 +03:00
|
|
|
q->max = i;
|
|
|
|
goto enomem;
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
2009-12-22 01:37:28 +03:00
|
|
|
kfifo_in(&q->queue, (void*)&q->pool[i], sizeof(void*));
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
2007-12-13 21:43:25 +03:00
|
|
|
|
|
|
|
if (items) {
|
|
|
|
*items = q->pool + max;
|
|
|
|
memcpy(*items, q->pool, max * sizeof(void *));
|
|
|
|
}
|
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
return 0;
|
2007-12-13 21:43:25 +03:00
|
|
|
|
|
|
|
enomem:
|
|
|
|
iscsi_pool_free(q);
|
|
|
|
return -ENOMEM;
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_pool_init);
|
|
|
|
|
2007-12-13 21:43:25 +03:00
|
|
|
void iscsi_pool_free(struct iscsi_pool *q)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = 0; i < q->max; i++)
|
2007-12-13 21:43:25 +03:00
|
|
|
kfree(q->pool[i]);
|
2009-03-05 23:45:55 +03:00
|
|
|
kfree(q->pool);
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_pool_free);
|
|
|
|
|
2008-05-22 00:54:00 +04:00
|
|
|
/**
|
|
|
|
* iscsi_host_add - add host to system
|
|
|
|
* @shost: scsi host
|
|
|
|
* @pdev: parent device
|
|
|
|
*
|
|
|
|
* This should be called by partial offload and software iscsi drivers
|
|
|
|
* to add a host to the system.
|
|
|
|
*/
|
|
|
|
int iscsi_host_add(struct Scsi_Host *shost, struct device *pdev)
|
|
|
|
{
|
2008-06-16 19:11:33 +04:00
|
|
|
if (!shost->can_queue)
|
|
|
|
shost->can_queue = ISCSI_DEF_XMIT_CMDS_MAX;
|
|
|
|
|
2009-03-05 23:46:04 +03:00
|
|
|
if (!shost->cmd_per_lun)
|
|
|
|
shost->cmd_per_lun = ISCSI_DEF_CMD_PER_LUN;
|
|
|
|
|
[SCSI] libiscsi: Fix scsi command timeout oops in iscsi_eh_timed_out
Yanling Qi from LSI found the root cause of the panic, below is his
analysis:
Problem description: the open iscsi driver installs eh_timed_out handler
to the
blank_transport_template of the scsi middle level that causes panic of
timed
out command of other host
Here are the details
Iscsi Session creation
During iscsi session creation time, the iscsi_tcp_session_create() of
iscsi_tpc.c will create a scsi-host for the session. See the statement
marked
with the label A. The statement B replaces the shost->transportt point
with a
local struct variable.
static struct iscsi_cls_session *
iscsi_tcp_session_create(struct iscsi_endpoint *ep, uint16_t cmds_max,
uint16_t qdepth, uint32_t initial_cmdsn,
uint32_t *hostno)
{
struct iscsi_cls_session *cls_session;
struct iscsi_session *session;
struct Scsi_Host *shost;
int cmd_i;
if (ep) {
printk(KERN_ERR "iscsi_tcp: invalid ep %p.\n", ep);
return NULL;
}
A shost = iscsi_host_alloc(&iscsi_sht, 0, qdepth);
if (!shost)
return NULL;
B shost->transportt = iscsi_tcp_scsi_transport;
shost->max_lun = iscsi_max_lun;
Please note the scsi host is allocated by invoking isccsi_host_alloc()
in
libiscsi.c
Polluting the middle level blank_transport_template in
iscsi_host_alloc() of
libiscsi.c
The iscsi_host_alloc() invokes the middle level function
scsi_host_alloc() in
hosts.c for allocating a scsi_host. Then the statement marked with C
assigns
the iscsi_eh_cmd_timed_out handler to the eh_timed_out callback
function.
struct Scsi_Host *iscsi_host_alloc(struct scsi_host_template *sht,
int dd_data_size, uint16_t qdepth)
{
struct Scsi_Host *shost;
struct iscsi_host *ihost;
shost = scsi_host_alloc(sht, sizeof(struct iscsi_host) +
dd_data_size);
if (!shost)
return NULL;
C shost->transportt->eh_timed_out = iscsi_eh_cmd_timed_out;
Please note the shost->transport is the middle level
blank_transport_template
as shown in the code segment below. We see two problems here. 1.
iscsi_eh_cmd_timed_out is installed to the blank_transport_template that
will
cause some body else problem. 2. iscsi_eh_cmd_timed_out will never be
invoked
when iscsi command gets timeout because the statement B resets the
pointer.
Middle level blank_transport_template
In the middle level function scsi_host_alloc() of hosts.c, the middle
level
assigns a blank_transport_template for those hosts not implementing its
transport layer. All HBAs without supporting a specific scsi_transport
will
share the middle level blank_transport_template. Please see the
statement D
struct Scsi_Host *scsi_host_alloc(struct scsi_host_template *sht, int
privsize)
{
struct Scsi_Host *shost;
gfp_t gfp_mask = GFP_KERNEL;
int rval;
if (sht->unchecked_isa_dma && privsize)
gfp_mask |= __GFP_DMA;
shost = kzalloc(sizeof(struct Scsi_Host) + privsize, gfp_mask);
if (!shost)
return NULL;
shost->host_lock = &shost->default_lock;
spin_lock_init(shost->host_lock);
shost->shost_state = SHOST_CREATED;
INIT_LIST_HEAD(&shost->__devices);
INIT_LIST_HEAD(&shost->__targets);
INIT_LIST_HEAD(&shost->eh_cmd_q);
INIT_LIST_HEAD(&shost->starved_list);
init_waitqueue_head(&shost->host_wait);
mutex_init(&shost->scan_mutex);
shost->host_no = scsi_host_next_hn++; /* XXX(hch): still racy */
shost->dma_channel = 0xff;
/* These three are default values which can be overridden */
shost->max_channel = 0;
shost->max_id = 8;
shost->max_lun = 8;
/* Give each shost a default transportt */
D shost->transportt = &blank_transport_template;
Why we see panic at iscsi_eh_cmd_timed_out()
The mpp virtual HBA doesn’t have a specific scsi_transport. Therefore,
the
blank_transport_template will be assigned to the virtual host of the MPP
virtual HBA by SCSI middle level. Please note that the statement C has
assigned
iscsi-transport eh_timedout handler to the blank_transport_template.
When a mpp
virtual command gets timedout, the iscsi_eh_cmd_timed_out() will be
invoked to
handle mpp virtual command timeout from the middle level
scsi_times_out()
function of the scsi_error.c.
enum blk_eh_timer_return scsi_times_out(struct request *req)
{
struct scsi_cmnd *scmd = req->special;
enum blk_eh_timer_return (*eh_timed_out)(struct scsi_cmnd *);
enum blk_eh_timer_return rtn = BLK_EH_NOT_HANDLED;
scsi_log_completion(scmd, TIMEOUT_ERROR);
if (scmd->device->host->transportt->eh_timed_out)
E eh_timed_out =
scmd->device->host->transportt->eh_timed_out;
else if (scmd->device->host->hostt->eh_timed_out)
eh_timed_out = scmd->device->host->hostt->eh_timed_out;
else
eh_timed_out = NULL;
if (eh_timed_out) {
rtn = eh_timed_out(scmd);
It is very easy to understand why we get panic in the
iscsi_eh_cmd_timed_out().
A scsi_cmnd from a no-iscsi device definitely can not resolve out a
session and
session->lock. The panic can be happed anywhere during the differencing.
static enum blk_eh_timer_return iscsi_eh_cmd_timed_out(struct scsi_cmnd
*scmd)
{
struct iscsi_cls_session *cls_session;
struct iscsi_session *session;
struct iscsi_conn *conn;
enum blk_eh_timer_return rc = BLK_EH_NOT_HANDLED;
cls_session = starget_to_session(scsi_target(scmd->device));
session = cls_session->dd_data;
debug_scsi("scsi cmd %p timedout\n", scmd);
spin_lock(&session->lock);
This patch fixes the problem by moving the setting of the
iscsi_eh_cmd_timed_out to iscsi_add_host, which is after the LLDs
have set their transport template to shost->transportt.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
2009-02-06 21:06:20 +03:00
|
|
|
if (!shost->transportt->eh_timed_out)
|
|
|
|
shost->transportt->eh_timed_out = iscsi_eh_cmd_timed_out;
|
2008-05-22 00:54:00 +04:00
|
|
|
return scsi_add_host(shost, pdev);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_host_add);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* iscsi_host_alloc - allocate a host and driver data
|
|
|
|
* @sht: scsi host template
|
|
|
|
* @dd_data_size: driver host data size
|
2009-03-05 23:46:03 +03:00
|
|
|
* @xmit_can_sleep: bool indicating if LLD will queue IO from a work queue
|
2008-05-22 00:54:00 +04:00
|
|
|
*
|
|
|
|
* This should be called by partial offload and software iscsi drivers.
|
|
|
|
* To access the driver specific memory use the iscsi_host_priv() macro.
|
|
|
|
*/
|
|
|
|
struct Scsi_Host *iscsi_host_alloc(struct scsi_host_template *sht,
|
2009-03-05 23:46:04 +03:00
|
|
|
int dd_data_size, bool xmit_can_sleep)
|
2008-05-22 00:53:59 +04:00
|
|
|
{
|
2008-05-22 00:54:00 +04:00
|
|
|
struct Scsi_Host *shost;
|
2008-09-24 20:46:10 +04:00
|
|
|
struct iscsi_host *ihost;
|
2008-05-22 00:54:00 +04:00
|
|
|
|
|
|
|
shost = scsi_host_alloc(sht, sizeof(struct iscsi_host) + dd_data_size);
|
|
|
|
if (!shost)
|
|
|
|
return NULL;
|
2008-09-24 20:46:10 +04:00
|
|
|
ihost = shost_priv(shost);
|
2009-03-05 23:46:03 +03:00
|
|
|
|
|
|
|
if (xmit_can_sleep) {
|
|
|
|
snprintf(ihost->workq_name, sizeof(ihost->workq_name),
|
|
|
|
"iscsi_q_%d", shost->host_no);
|
|
|
|
ihost->workq = create_singlethread_workqueue(ihost->workq_name);
|
|
|
|
if (!ihost->workq)
|
|
|
|
goto free_host;
|
|
|
|
}
|
|
|
|
|
2008-09-24 20:46:10 +04:00
|
|
|
spin_lock_init(&ihost->lock);
|
|
|
|
ihost->state = ISCSI_HOST_SETUP;
|
|
|
|
ihost->num_sessions = 0;
|
|
|
|
init_waitqueue_head(&ihost->session_removal_wq);
|
2008-05-22 00:54:00 +04:00
|
|
|
return shost;
|
2009-03-05 23:46:03 +03:00
|
|
|
|
|
|
|
free_host:
|
|
|
|
scsi_host_put(shost);
|
|
|
|
return NULL;
|
2008-05-22 00:54:00 +04:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_host_alloc);
|
|
|
|
|
2008-09-24 20:46:10 +04:00
|
|
|
static void iscsi_notify_host_removed(struct iscsi_cls_session *cls_session)
|
|
|
|
{
|
2009-03-05 23:46:05 +03:00
|
|
|
iscsi_session_failure(cls_session->dd_data, ISCSI_ERR_INVALID_HOST);
|
2008-09-24 20:46:10 +04:00
|
|
|
}
|
|
|
|
|
2008-05-22 00:54:00 +04:00
|
|
|
/**
|
|
|
|
* iscsi_host_remove - remove host and sessions
|
|
|
|
* @shost: scsi host
|
|
|
|
*
|
2008-09-24 20:46:10 +04:00
|
|
|
* If there are any sessions left, this will initiate the removal and wait
|
|
|
|
* for the completion.
|
2008-05-22 00:54:00 +04:00
|
|
|
*/
|
|
|
|
void iscsi_host_remove(struct Scsi_Host *shost)
|
|
|
|
{
|
2008-09-24 20:46:10 +04:00
|
|
|
struct iscsi_host *ihost = shost_priv(shost);
|
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
spin_lock_irqsave(&ihost->lock, flags);
|
|
|
|
ihost->state = ISCSI_HOST_REMOVED;
|
|
|
|
spin_unlock_irqrestore(&ihost->lock, flags);
|
|
|
|
|
|
|
|
iscsi_host_for_each_session(shost, iscsi_notify_host_removed);
|
|
|
|
wait_event_interruptible(ihost->session_removal_wq,
|
|
|
|
ihost->num_sessions == 0);
|
|
|
|
if (signal_pending(current))
|
|
|
|
flush_signals(current);
|
|
|
|
|
2008-05-22 00:54:00 +04:00
|
|
|
scsi_remove_host(shost);
|
2009-03-05 23:46:03 +03:00
|
|
|
if (ihost->workq)
|
|
|
|
destroy_workqueue(ihost->workq);
|
2008-05-22 00:53:59 +04:00
|
|
|
}
|
2008-05-22 00:54:00 +04:00
|
|
|
EXPORT_SYMBOL_GPL(iscsi_host_remove);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:00 +04:00
|
|
|
void iscsi_host_free(struct Scsi_Host *shost)
|
2008-05-22 00:53:59 +04:00
|
|
|
{
|
|
|
|
struct iscsi_host *ihost = shost_priv(shost);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:53:59 +04:00
|
|
|
kfree(ihost->netdev);
|
|
|
|
kfree(ihost->hwaddress);
|
|
|
|
kfree(ihost->initiatorname);
|
2008-05-22 00:54:00 +04:00
|
|
|
scsi_host_put(shost);
|
2008-05-22 00:53:59 +04:00
|
|
|
}
|
2008-05-22 00:54:00 +04:00
|
|
|
EXPORT_SYMBOL_GPL(iscsi_host_free);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-09-24 20:46:10 +04:00
|
|
|
static void iscsi_host_dec_session_cnt(struct Scsi_Host *shost)
|
|
|
|
{
|
|
|
|
struct iscsi_host *ihost = shost_priv(shost);
|
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
shost = scsi_host_get(shost);
|
|
|
|
if (!shost) {
|
|
|
|
printk(KERN_ERR "Invalid state. Cannot notify host removal "
|
|
|
|
"of session teardown event because host already "
|
|
|
|
"removed.\n");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
spin_lock_irqsave(&ihost->lock, flags);
|
|
|
|
ihost->num_sessions--;
|
|
|
|
if (ihost->num_sessions == 0)
|
|
|
|
wake_up(&ihost->session_removal_wq);
|
|
|
|
spin_unlock_irqrestore(&ihost->lock, flags);
|
|
|
|
scsi_host_put(shost);
|
|
|
|
}
|
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
/**
|
|
|
|
* iscsi_session_setup - create iscsi cls session and host and session
|
|
|
|
* @iscsit: iscsi transport template
|
2008-05-22 00:53:59 +04:00
|
|
|
* @shost: scsi host
|
|
|
|
* @cmds_max: session can queue
|
2008-05-22 00:54:09 +04:00
|
|
|
* @cmd_task_size: LLD task private data size
|
2006-04-07 06:13:41 +04:00
|
|
|
* @initial_cmdsn: initial CmdSN
|
|
|
|
*
|
|
|
|
* This can be used by software iscsi_transports that allocate
|
|
|
|
* a session per scsi host.
|
2008-05-22 00:54:17 +04:00
|
|
|
*
|
|
|
|
* Callers should set cmds_max to the largest total numer (mgmt + scsi) of
|
|
|
|
* tasks they support. The iscsi layer reserves ISCSI_MGMT_CMDS_MAX tasks
|
|
|
|
* for nop handling and login/logout requests.
|
2008-05-22 00:53:59 +04:00
|
|
|
*/
|
2006-04-07 06:13:41 +04:00
|
|
|
struct iscsi_cls_session *
|
2008-05-22 00:53:59 +04:00
|
|
|
iscsi_session_setup(struct iscsi_transport *iscsit, struct Scsi_Host *shost,
|
2009-09-22 06:51:22 +04:00
|
|
|
uint16_t cmds_max, int dd_size, int cmd_task_size,
|
2008-05-22 00:54:12 +04:00
|
|
|
uint32_t initial_cmdsn, unsigned int id)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
2008-09-24 20:46:10 +04:00
|
|
|
struct iscsi_host *ihost = shost_priv(shost);
|
2006-04-07 06:13:41 +04:00
|
|
|
struct iscsi_session *session;
|
|
|
|
struct iscsi_cls_session *cls_session;
|
2008-05-22 00:54:17 +04:00
|
|
|
int cmd_i, scsi_cmds, total_cmds = cmds_max;
|
2008-09-24 20:46:10 +04:00
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
spin_lock_irqsave(&ihost->lock, flags);
|
|
|
|
if (ihost->state == ISCSI_HOST_REMOVED) {
|
|
|
|
spin_unlock_irqrestore(&ihost->lock, flags);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
ihost->num_sessions++;
|
|
|
|
spin_unlock_irqrestore(&ihost->lock, flags);
|
2008-06-16 19:11:33 +04:00
|
|
|
|
|
|
|
if (!total_cmds)
|
|
|
|
total_cmds = ISCSI_DEF_XMIT_CMDS_MAX;
|
2008-05-22 00:54:06 +04:00
|
|
|
/*
|
2008-05-22 00:54:17 +04:00
|
|
|
* The iscsi layer needs some tasks for nop handling and tmfs,
|
|
|
|
* so the cmds_max must at least be greater than ISCSI_MGMT_CMDS_MAX
|
|
|
|
* + 1 command for scsi IO.
|
2008-05-22 00:54:06 +04:00
|
|
|
*/
|
2008-05-22 00:54:17 +04:00
|
|
|
if (total_cmds < ISCSI_TOTAL_CMDS_MIN) {
|
|
|
|
printk(KERN_ERR "iscsi: invalid can_queue of %d. can_queue "
|
|
|
|
"must be a power of two that is at least %d.\n",
|
|
|
|
total_cmds, ISCSI_TOTAL_CMDS_MIN);
|
2008-09-24 20:46:10 +04:00
|
|
|
goto dec_session_count;
|
2008-05-22 00:54:17 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
if (total_cmds > ISCSI_TOTAL_CMDS_MAX) {
|
|
|
|
printk(KERN_ERR "iscsi: invalid can_queue of %d. can_queue "
|
|
|
|
"must be a power of 2 less than or equal to %d.\n",
|
|
|
|
cmds_max, ISCSI_TOTAL_CMDS_MAX);
|
|
|
|
total_cmds = ISCSI_TOTAL_CMDS_MAX;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!is_power_of_2(total_cmds)) {
|
|
|
|
printk(KERN_ERR "iscsi: invalid can_queue of %d. can_queue "
|
|
|
|
"must be a power of 2.\n", total_cmds);
|
|
|
|
total_cmds = rounddown_pow_of_two(total_cmds);
|
|
|
|
if (total_cmds < ISCSI_TOTAL_CMDS_MIN)
|
|
|
|
return NULL;
|
|
|
|
printk(KERN_INFO "iscsi: Rounding can_queue to %d.\n",
|
|
|
|
total_cmds);
|
2007-05-30 21:57:19 +04:00
|
|
|
}
|
2008-05-22 00:54:17 +04:00
|
|
|
scsi_cmds = total_cmds - ISCSI_MGMT_CMDS_MAX;
|
2007-05-30 21:57:19 +04:00
|
|
|
|
2008-05-22 00:54:01 +04:00
|
|
|
cls_session = iscsi_alloc_session(shost, iscsit,
|
2009-09-22 06:51:22 +04:00
|
|
|
sizeof(struct iscsi_session) +
|
|
|
|
dd_size);
|
2008-05-22 00:53:59 +04:00
|
|
|
if (!cls_session)
|
2008-09-24 20:46:10 +04:00
|
|
|
goto dec_session_count;
|
2008-05-22 00:53:59 +04:00
|
|
|
session = cls_session->dd_data;
|
|
|
|
session->cls_session = cls_session;
|
2006-04-07 06:13:41 +04:00
|
|
|
session->host = shost;
|
|
|
|
session->state = ISCSI_STATE_FREE;
|
2007-12-13 21:43:30 +03:00
|
|
|
session->fast_abort = 1;
|
2009-11-12 01:34:33 +03:00
|
|
|
session->tgt_reset_timeout = 30;
|
2007-12-13 21:43:38 +03:00
|
|
|
session->lu_reset_timeout = 15;
|
|
|
|
session->abort_timeout = 10;
|
2008-05-22 00:54:17 +04:00
|
|
|
session->scsi_cmds_max = scsi_cmds;
|
|
|
|
session->cmds_max = total_cmds;
|
2007-07-26 21:46:48 +04:00
|
|
|
session->queued_cmdsn = session->cmdsn = initial_cmdsn;
|
2006-04-07 06:13:41 +04:00
|
|
|
session->exp_cmdsn = initial_cmdsn + 1;
|
|
|
|
session->max_cmdsn = initial_cmdsn + 1;
|
|
|
|
session->max_r2t = 1;
|
|
|
|
session->tt = iscsit;
|
2009-09-22 06:51:22 +04:00
|
|
|
session->dd_data = cls_session->dd_data + sizeof(*session);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
|
2007-08-15 10:38:30 +04:00
|
|
|
mutex_init(&session->eh_mutex);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_init(&session->frwd_lock);
|
|
|
|
spin_lock_init(&session->back_lock);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
|
|
|
/* initialize SCSI PDU commands pool */
|
|
|
|
if (iscsi_pool_init(&session->cmdpool, session->cmds_max,
|
|
|
|
(void***)&session->cmds,
|
2008-05-22 00:54:09 +04:00
|
|
|
cmd_task_size + sizeof(struct iscsi_task)))
|
2006-04-07 06:13:41 +04:00
|
|
|
goto cmdpool_alloc_fail;
|
|
|
|
|
|
|
|
/* pre-format cmds pool with ITT */
|
|
|
|
for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
|
2008-05-22 00:54:09 +04:00
|
|
|
struct iscsi_task *task = session->cmds[cmd_i];
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
if (cmd_task_size)
|
|
|
|
task->dd_data = &task[1];
|
|
|
|
task->itt = cmd_i;
|
2009-05-14 02:57:46 +04:00
|
|
|
task->state = ISCSI_TASK_FREE;
|
2008-05-22 00:54:09 +04:00
|
|
|
INIT_LIST_HEAD(&task->running);
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
|
2006-06-28 21:00:27 +04:00
|
|
|
if (!try_module_get(iscsit->owner))
|
2008-05-22 00:53:59 +04:00
|
|
|
goto module_get_fail;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:12 +04:00
|
|
|
if (iscsi_add_session(cls_session, id))
|
2008-05-22 00:53:59 +04:00
|
|
|
goto cls_session_fail;
|
2008-09-24 20:46:10 +04:00
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
return cls_session;
|
|
|
|
|
|
|
|
cls_session_fail:
|
2008-05-22 00:53:59 +04:00
|
|
|
module_put(iscsit->owner);
|
|
|
|
module_get_fail:
|
2007-12-13 21:43:25 +03:00
|
|
|
iscsi_pool_free(&session->cmdpool);
|
2006-04-07 06:13:41 +04:00
|
|
|
cmdpool_alloc_fail:
|
2008-05-22 00:53:59 +04:00
|
|
|
iscsi_free_session(cls_session);
|
2008-09-24 20:46:10 +04:00
|
|
|
dec_session_count:
|
|
|
|
iscsi_host_dec_session_cnt(shost);
|
2006-04-07 06:13:41 +04:00
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_session_setup);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* iscsi_session_teardown - destroy session, host, and cls_session
|
2008-05-22 00:53:59 +04:00
|
|
|
* @cls_session: iscsi session
|
2006-04-07 06:13:41 +04:00
|
|
|
*
|
2008-05-22 00:53:59 +04:00
|
|
|
* The driver must have called iscsi_remove_session before
|
|
|
|
* calling this.
|
|
|
|
*/
|
2006-04-07 06:13:41 +04:00
|
|
|
void iscsi_session_teardown(struct iscsi_cls_session *cls_session)
|
|
|
|
{
|
2008-05-22 00:53:59 +04:00
|
|
|
struct iscsi_session *session = cls_session->dd_data;
|
2006-07-25 00:47:29 +04:00
|
|
|
struct module *owner = cls_session->transport->owner;
|
2008-09-24 20:46:10 +04:00
|
|
|
struct Scsi_Host *shost = session->host;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2007-12-13 21:43:25 +03:00
|
|
|
iscsi_pool_free(&session->cmdpool);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2007-05-30 21:57:16 +04:00
|
|
|
kfree(session->password);
|
|
|
|
kfree(session->password_in);
|
|
|
|
kfree(session->username);
|
|
|
|
kfree(session->username_in);
|
2006-07-25 00:47:50 +04:00
|
|
|
kfree(session->targetname);
|
2012-01-19 15:06:53 +04:00
|
|
|
kfree(session->targetalias);
|
2008-05-22 00:54:16 +04:00
|
|
|
kfree(session->initiatorname);
|
2013-06-20 21:21:26 +04:00
|
|
|
kfree(session->boot_root);
|
|
|
|
kfree(session->boot_nic);
|
|
|
|
kfree(session->boot_target);
|
2008-05-22 00:54:16 +04:00
|
|
|
kfree(session->ifacename);
|
2013-07-01 13:54:12 +04:00
|
|
|
kfree(session->portal_type);
|
|
|
|
kfree(session->discovery_parent_type);
|
2006-07-25 00:47:50 +04:00
|
|
|
|
2008-05-22 00:53:59 +04:00
|
|
|
iscsi_destroy_session(cls_session);
|
2008-09-24 20:46:10 +04:00
|
|
|
iscsi_host_dec_session_cnt(shost);
|
2006-07-25 00:47:29 +04:00
|
|
|
module_put(owner);
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_session_teardown);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* iscsi_conn_setup - create iscsi_cls_conn and iscsi_conn
|
|
|
|
* @cls_session: iscsi_cls_session
|
2008-05-22 00:54:01 +04:00
|
|
|
* @dd_size: private driver data size
|
2006-04-07 06:13:41 +04:00
|
|
|
* @conn_idx: cid
|
2008-05-22 00:54:01 +04:00
|
|
|
*/
|
2006-04-07 06:13:41 +04:00
|
|
|
struct iscsi_cls_conn *
|
2008-05-22 00:54:01 +04:00
|
|
|
iscsi_conn_setup(struct iscsi_cls_session *cls_session, int dd_size,
|
|
|
|
uint32_t conn_idx)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
2008-05-22 00:53:59 +04:00
|
|
|
struct iscsi_session *session = cls_session->dd_data;
|
2006-04-07 06:13:41 +04:00
|
|
|
struct iscsi_conn *conn;
|
|
|
|
struct iscsi_cls_conn *cls_conn;
|
2006-05-19 05:31:34 +04:00
|
|
|
char *data;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:01 +04:00
|
|
|
cls_conn = iscsi_create_conn(cls_session, sizeof(*conn) + dd_size,
|
|
|
|
conn_idx);
|
2006-04-07 06:13:41 +04:00
|
|
|
if (!cls_conn)
|
|
|
|
return NULL;
|
|
|
|
conn = cls_conn->dd_data;
|
2008-05-22 00:54:01 +04:00
|
|
|
memset(conn, 0, sizeof(*conn) + dd_size);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:01 +04:00
|
|
|
conn->dd_data = cls_conn->dd_data + sizeof(*conn);
|
2006-04-07 06:13:41 +04:00
|
|
|
conn->session = session;
|
|
|
|
conn->cls_conn = cls_conn;
|
|
|
|
conn->c_stage = ISCSI_CONN_INITIAL_STAGE;
|
|
|
|
conn->id = conn_idx;
|
|
|
|
conn->exp_statsn = 0;
|
2007-12-13 21:43:20 +03:00
|
|
|
conn->tmf_state = TMF_INITIAL;
|
2007-12-13 21:43:30 +03:00
|
|
|
|
|
|
|
init_timer(&conn->transport_timer);
|
|
|
|
conn->transport_timer.data = (unsigned long)conn;
|
|
|
|
conn->transport_timer.function = iscsi_check_transport_timeouts;
|
|
|
|
|
2007-12-13 21:43:20 +03:00
|
|
|
INIT_LIST_HEAD(&conn->mgmtqueue);
|
2009-05-14 02:57:46 +04:00
|
|
|
INIT_LIST_HEAD(&conn->cmdqueue);
|
2007-12-13 21:43:20 +03:00
|
|
|
INIT_LIST_HEAD(&conn->requeue);
|
2006-11-22 17:57:56 +03:00
|
|
|
INIT_WORK(&conn->xmitwork, iscsi_xmitworker);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
/* allocate login_task used for the login/text sequences */
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2009-12-22 01:37:28 +03:00
|
|
|
if (!kfifo_out(&session->cmdpool.queue,
|
2008-05-22 00:54:09 +04:00
|
|
|
(void*)&conn->login_task,
|
2006-04-07 06:13:41 +04:00
|
|
|
sizeof(void*))) {
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2008-05-22 00:54:09 +04:00
|
|
|
goto login_task_alloc_fail;
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-12-02 09:32:09 +03:00
|
|
|
data = (char *) __get_free_pages(GFP_KERNEL,
|
|
|
|
get_order(ISCSI_DEF_MAX_RECV_SEG_LEN));
|
2006-05-19 05:31:34 +04:00
|
|
|
if (!data)
|
2008-05-22 00:54:09 +04:00
|
|
|
goto login_task_data_alloc_fail;
|
|
|
|
conn->login_task->data = conn->data = data;
|
2006-05-19 05:31:34 +04:00
|
|
|
|
2007-12-13 21:43:20 +03:00
|
|
|
init_timer(&conn->tmf_timer);
|
2006-04-07 06:13:41 +04:00
|
|
|
init_waitqueue_head(&conn->ehwait);
|
|
|
|
|
|
|
|
return cls_conn;
|
|
|
|
|
2008-05-22 00:54:09 +04:00
|
|
|
login_task_data_alloc_fail:
|
2009-12-22 01:37:28 +03:00
|
|
|
kfifo_in(&session->cmdpool.queue, (void*)&conn->login_task,
|
2006-05-19 05:31:34 +04:00
|
|
|
sizeof(void*));
|
2008-05-22 00:54:09 +04:00
|
|
|
login_task_alloc_fail:
|
2006-04-07 06:13:41 +04:00
|
|
|
iscsi_destroy_conn(cls_conn);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_conn_setup);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* iscsi_conn_teardown - teardown iscsi connection
|
|
|
|
* cls_conn: iscsi class connection
|
|
|
|
*
|
|
|
|
* TODO: we may need to make this into a two step process
|
|
|
|
* like scsi-mls remove + put host
|
|
|
|
*/
|
|
|
|
void iscsi_conn_teardown(struct iscsi_cls_conn *cls_conn)
|
|
|
|
{
|
|
|
|
struct iscsi_conn *conn = cls_conn->dd_data;
|
|
|
|
struct iscsi_session *session = conn->session;
|
|
|
|
|
2007-12-13 21:43:30 +03:00
|
|
|
del_timer_sync(&conn->transport_timer);
|
|
|
|
|
2015-06-24 04:11:58 +03:00
|
|
|
mutex_lock(&session->eh_mutex);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2006-04-07 06:13:41 +04:00
|
|
|
conn->c_stage = ISCSI_CONN_CLEANUP_WAIT;
|
|
|
|
if (session->leadconn == conn) {
|
|
|
|
/*
|
|
|
|
* leading connection? then give up on recovery.
|
|
|
|
*/
|
|
|
|
session->state = ISCSI_STATE_TERMINATE;
|
|
|
|
wake_up(&conn->ehwait);
|
|
|
|
}
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2007-03-01 02:32:15 +03:00
|
|
|
/* flush queued up work because we free the connection below */
|
2007-12-13 21:43:20 +03:00
|
|
|
iscsi_suspend_tx(conn);
|
2007-03-01 02:32:15 +03:00
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2008-12-02 09:32:09 +03:00
|
|
|
free_pages((unsigned long) conn->data,
|
|
|
|
get_order(ISCSI_DEF_MAX_RECV_SEG_LEN));
|
2006-07-25 00:47:50 +04:00
|
|
|
kfree(conn->persistent_address);
|
2013-11-22 14:28:21 +04:00
|
|
|
kfree(conn->local_ipaddr);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
/* regular RX path uses back_lock */
|
|
|
|
spin_lock_bh(&session->back_lock);
|
2009-12-22 01:37:28 +03:00
|
|
|
kfifo_in(&session->cmdpool.queue, (void*)&conn->login_task,
|
2006-04-07 06:13:41 +04:00
|
|
|
sizeof(void*));
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->back_lock);
|
2007-07-26 21:46:48 +04:00
|
|
|
if (session->leadconn == conn)
|
2006-04-07 06:13:41 +04:00
|
|
|
session->leadconn = NULL;
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2015-06-24 04:11:58 +03:00
|
|
|
mutex_unlock(&session->eh_mutex);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
|
|
|
iscsi_destroy_conn(cls_conn);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_conn_teardown);
|
|
|
|
|
|
|
|
int iscsi_conn_start(struct iscsi_cls_conn *cls_conn)
|
|
|
|
{
|
|
|
|
struct iscsi_conn *conn = cls_conn->dd_data;
|
|
|
|
struct iscsi_session *session = conn->session;
|
|
|
|
|
2006-09-01 02:09:24 +04:00
|
|
|
if (!session) {
|
2008-01-31 22:36:52 +03:00
|
|
|
iscsi_conn_printk(KERN_ERR, conn,
|
|
|
|
"can't start unbound connection\n");
|
2006-04-07 06:13:41 +04:00
|
|
|
return -EPERM;
|
|
|
|
}
|
|
|
|
|
2006-09-01 02:09:31 +04:00
|
|
|
if ((session->imm_data_en || !session->initial_r2t_en) &&
|
|
|
|
session->first_burst > session->max_burst) {
|
2008-01-31 22:36:52 +03:00
|
|
|
iscsi_conn_printk(KERN_INFO, conn, "invalid burst lengths: "
|
|
|
|
"first_burst %d max_burst %d\n",
|
|
|
|
session->first_burst, session->max_burst);
|
2006-09-01 02:09:24 +04:00
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
2007-12-13 21:43:30 +03:00
|
|
|
if (conn->ping_timeout && !conn->recv_timeout) {
|
2008-01-31 22:36:52 +03:00
|
|
|
iscsi_conn_printk(KERN_ERR, conn, "invalid recv timeout of "
|
|
|
|
"zero. Using 5 seconds\n.");
|
2007-12-13 21:43:30 +03:00
|
|
|
conn->recv_timeout = 5;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (conn->recv_timeout && !conn->ping_timeout) {
|
2008-01-31 22:36:52 +03:00
|
|
|
iscsi_conn_printk(KERN_ERR, conn, "invalid ping timeout of "
|
|
|
|
"zero. Using 5 seconds.\n");
|
2007-12-13 21:43:30 +03:00
|
|
|
conn->ping_timeout = 5;
|
|
|
|
}
|
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2006-04-07 06:13:41 +04:00
|
|
|
conn->c_stage = ISCSI_CONN_STARTED;
|
|
|
|
session->state = ISCSI_STATE_LOGGED_IN;
|
2007-07-26 21:46:48 +04:00
|
|
|
session->queued_cmdsn = session->cmdsn;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2007-12-13 21:43:30 +03:00
|
|
|
conn->last_recv = jiffies;
|
|
|
|
conn->last_ping = jiffies;
|
|
|
|
if (conn->recv_timeout && conn->ping_timeout)
|
|
|
|
mod_timer(&conn->transport_timer,
|
|
|
|
jiffies + (conn->recv_timeout * HZ));
|
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
switch(conn->stop_stage) {
|
|
|
|
case STOP_CONN_RECOVER:
|
|
|
|
/*
|
|
|
|
* unblock eh_abort() if it is blocked. re-try all
|
|
|
|
* commands after successful recovery
|
|
|
|
*/
|
|
|
|
conn->stop_stage = 0;
|
2007-12-13 21:43:20 +03:00
|
|
|
conn->tmf_state = TMF_INITIAL;
|
2006-04-07 06:13:41 +04:00
|
|
|
session->age++;
|
2008-01-31 22:36:53 +03:00
|
|
|
if (session->age == 16)
|
|
|
|
session->age = 0;
|
2008-01-31 22:36:43 +03:00
|
|
|
break;
|
2006-04-07 06:13:41 +04:00
|
|
|
case STOP_CONN_TERM:
|
|
|
|
conn->stop_stage = 0;
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2008-05-22 00:53:59 +04:00
|
|
|
iscsi_unblock_session(session->cls_session);
|
2008-01-31 22:36:43 +03:00
|
|
|
wake_up(&conn->ehwait);
|
2006-04-07 06:13:41 +04:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_conn_start);
|
|
|
|
|
|
|
|
static void
|
2009-05-14 02:57:46 +04:00
|
|
|
fail_mgmt_tasks(struct iscsi_session *session, struct iscsi_conn *conn)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
2009-05-14 02:57:46 +04:00
|
|
|
struct iscsi_task *task;
|
2009-05-14 02:57:49 +04:00
|
|
|
int i, state;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2009-05-14 02:57:46 +04:00
|
|
|
for (i = 0; i < conn->session->cmds_max; i++) {
|
|
|
|
task = conn->session->cmds[i];
|
|
|
|
if (task->sc)
|
|
|
|
continue;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2009-05-14 02:57:46 +04:00
|
|
|
if (task->state == ISCSI_TASK_FREE)
|
|
|
|
continue;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
2009-05-14 02:57:46 +04:00
|
|
|
ISCSI_DBG_SESSION(conn->session,
|
|
|
|
"failing mgmt itt 0x%x state %d\n",
|
|
|
|
task->itt, task->state);
|
2009-05-14 02:57:49 +04:00
|
|
|
state = ISCSI_TASK_ABRT_SESS_RECOV;
|
|
|
|
if (task->state == ISCSI_TASK_PENDING)
|
|
|
|
state = ISCSI_TASK_COMPLETED;
|
|
|
|
iscsi_complete_task(task, state);
|
|
|
|
|
2009-05-14 02:57:46 +04:00
|
|
|
}
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
|
2006-05-19 05:31:42 +04:00
|
|
|
static void iscsi_start_session_recovery(struct iscsi_session *session,
|
|
|
|
struct iscsi_conn *conn, int flag)
|
2006-04-07 06:13:41 +04:00
|
|
|
{
|
2006-05-03 04:46:40 +04:00
|
|
|
int old_stop_stage;
|
|
|
|
|
2007-08-15 10:38:30 +04:00
|
|
|
mutex_lock(&session->eh_mutex);
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2006-05-03 04:46:40 +04:00
|
|
|
if (conn->stop_stage == STOP_CONN_TERM) {
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2007-08-15 10:38:30 +04:00
|
|
|
mutex_unlock(&session->eh_mutex);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2006-05-03 04:46:40 +04:00
|
|
|
/*
|
|
|
|
* When this is called for the in_login state, we only want to clean
|
2008-05-22 00:54:09 +04:00
|
|
|
* up the login task and connection. We do not need to block and set
|
2006-05-30 09:37:20 +04:00
|
|
|
* the recovery state again
|
2006-05-03 04:46:40 +04:00
|
|
|
*/
|
2006-05-30 09:37:20 +04:00
|
|
|
if (flag == STOP_CONN_TERM)
|
|
|
|
session->state = ISCSI_STATE_TERMINATE;
|
|
|
|
else if (conn->stop_stage != STOP_CONN_RECOVER)
|
|
|
|
session->state = ISCSI_STATE_IN_RECOVERY;
|
2010-03-09 23:14:51 +03:00
|
|
|
|
|
|
|
old_stop_stage = conn->stop_stage;
|
|
|
|
conn->stop_stage = flag;
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2006-05-03 04:46:40 +04:00
|
|
|
|
2009-05-14 02:57:43 +04:00
|
|
|
del_timer_sync(&conn->transport_timer);
|
|
|
|
iscsi_suspend_tx(conn);
|
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2006-05-30 09:37:20 +04:00
|
|
|
conn->c_stage = ISCSI_CONN_STOPPED;
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2007-08-15 10:38:30 +04:00
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
/*
|
|
|
|
* for connection level recovery we should not calculate
|
|
|
|
* header digest. conn->hdr_size used for optimization
|
|
|
|
* in hdr_extract() and will be re-negotiated at
|
|
|
|
* set_param() time.
|
|
|
|
*/
|
|
|
|
if (flag == STOP_CONN_RECOVER) {
|
|
|
|
conn->hdrdgst_en = 0;
|
|
|
|
conn->datadgst_en = 0;
|
2006-05-19 05:31:42 +04:00
|
|
|
if (session->state == ISCSI_STATE_IN_RECOVERY &&
|
2006-05-30 09:37:20 +04:00
|
|
|
old_stop_stage != STOP_CONN_RECOVER) {
|
2009-03-05 23:45:58 +03:00
|
|
|
ISCSI_DBG_SESSION(session, "blocking session\n");
|
2008-05-22 00:53:59 +04:00
|
|
|
iscsi_block_session(session->cls_session);
|
2006-05-30 09:37:20 +04:00
|
|
|
}
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
2006-05-19 05:31:42 +04:00
|
|
|
|
|
|
|
/*
|
|
|
|
* flush queues.
|
|
|
|
*/
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2009-05-14 02:57:49 +04:00
|
|
|
fail_scsi_tasks(conn, -1, DID_TRANSPORT_DISRUPTED);
|
2009-05-14 02:57:46 +04:00
|
|
|
fail_mgmt_tasks(session, conn);
|
2009-11-12 01:34:32 +03:00
|
|
|
memset(&conn->tmhdr, 0, sizeof(conn->tmhdr));
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2007-08-15 10:38:30 +04:00
|
|
|
mutex_unlock(&session->eh_mutex);
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
void iscsi_conn_stop(struct iscsi_cls_conn *cls_conn, int flag)
|
|
|
|
{
|
|
|
|
struct iscsi_conn *conn = cls_conn->dd_data;
|
|
|
|
struct iscsi_session *session = conn->session;
|
|
|
|
|
|
|
|
switch (flag) {
|
|
|
|
case STOP_CONN_RECOVER:
|
|
|
|
case STOP_CONN_TERM:
|
|
|
|
iscsi_start_session_recovery(session, conn, flag);
|
2006-05-03 04:46:47 +04:00
|
|
|
break;
|
2006-04-07 06:13:41 +04:00
|
|
|
default:
|
2008-01-31 22:36:52 +03:00
|
|
|
iscsi_conn_printk(KERN_ERR, conn,
|
|
|
|
"invalid stop flag %d\n", flag);
|
2006-04-07 06:13:41 +04:00
|
|
|
}
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_conn_stop);
|
|
|
|
|
|
|
|
int iscsi_conn_bind(struct iscsi_cls_session *cls_session,
|
|
|
|
struct iscsi_cls_conn *cls_conn, int is_leading)
|
|
|
|
{
|
2008-05-22 00:53:59 +04:00
|
|
|
struct iscsi_session *session = cls_session->dd_data;
|
2006-10-17 02:09:39 +04:00
|
|
|
struct iscsi_conn *conn = cls_conn->dd_data;
|
2006-04-07 06:13:41 +04:00
|
|
|
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_lock_bh(&session->frwd_lock);
|
2006-04-07 06:13:41 +04:00
|
|
|
if (is_leading)
|
|
|
|
session->leadconn = conn;
|
[SCSI] libiscsi: Reduce locking contention in fast path
Replace the session lock with two locks, a forward lock and
a backwards lock named frwd_lock and back_lock respectively.
The forward lock protects resources that change while sending a
request to the target, such as cmdsn, queued_cmdsn, and allocating
task from the commands' pool with kfifo_out.
The backward lock protects resources that change while processing
a response or in error path, such as cmdsn_exp, cmdsn_max, and
returning tasks to the commands' pool with kfifo_in.
Under a steady state fast-path situation, that is when one
or more processes/threads submit IO to an iscsi device and
a single kernel upcall (e.g softirq) is dealing with processing
of responses without errors, this patch eliminates the contention
between the queuecommand()/request response/scsi_done() flows
associated with iscsi sessions.
Between the forward and the backward locks exists a strict locking
hierarchy. The mutual exclusion zone protected by the forward lock can
enclose the mutual exclusion zone protected by the backward lock but not
vice versa.
For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
a failure and __iscsi_put_task is called, the backward lock is taken while
the forward lock is still taken. On the other hand, if in the RX path a nop
is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
than the forward lock is released and the backward lock is taken for the
duration of iscsi_send_nopout, later the backward lock is released and the
forward lock is retaken.
libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
The insertion and deletion from these queues didn't corespond to the
assumption taken by the new forward/backwards session locking paradigm.
That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
path, r2t is taken out from r2t queue and inserted to the r2t pool.
In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
is also inserted to the r2t pool and another r2t is pulled from r2t
queue.
Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
to the TX path, r2t is taken from the r2t pool and inserted to the r2t
queue.
In order to cope with this situation, two spin locks were added,
pool2queue and queue2pool. The former protects extracting from the
r2t pool and inserting to the r2t queue, and the later protects the
extracing from the r2t queue and inserting to the r2t pool.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
[minor fix up to apply cleanly and compile fix]
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
2014-02-07 10:41:38 +04:00
|
|
|
spin_unlock_bh(&session->frwd_lock);
|
2006-04-07 06:13:41 +04:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Unblock xmitworker(), Login Phase will pass through.
|
|
|
|
*/
|
|
|
|
clear_bit(ISCSI_SUSPEND_BIT, &conn->suspend_rx);
|
|
|
|
clear_bit(ISCSI_SUSPEND_BIT, &conn->suspend_tx);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_conn_bind);
|
|
|
|
|
2013-03-22 15:41:30 +04:00
|
|
|
int iscsi_switch_str_param(char **param, char *new_val_buf)
|
2009-05-14 02:57:40 +04:00
|
|
|
{
|
|
|
|
char *new_val;
|
|
|
|
|
|
|
|
if (*param) {
|
|
|
|
if (!strcmp(*param, new_val_buf))
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
new_val = kstrdup(new_val_buf, GFP_NOIO);
|
|
|
|
if (!new_val)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
kfree(*param);
|
|
|
|
*param = new_val;
|
|
|
|
return 0;
|
|
|
|
}
|
2013-03-22 15:41:30 +04:00
|
|
|
EXPORT_SYMBOL_GPL(iscsi_switch_str_param);
|
2006-06-28 21:00:23 +04:00
|
|
|
|
|
|
|
int iscsi_set_param(struct iscsi_cls_conn *cls_conn,
|
|
|
|
enum iscsi_param param, char *buf, int buflen)
|
|
|
|
{
|
|
|
|
struct iscsi_conn *conn = cls_conn->dd_data;
|
|
|
|
struct iscsi_session *session = conn->session;
|
2013-08-08 14:44:29 +04:00
|
|
|
int val;
|
2006-06-28 21:00:23 +04:00
|
|
|
|
|
|
|
switch(param) {
|
2007-12-13 21:43:20 +03:00
|
|
|
case ISCSI_PARAM_FAST_ABORT:
|
|
|
|
sscanf(buf, "%d", &session->fast_abort);
|
|
|
|
break;
|
2007-12-13 21:43:30 +03:00
|
|
|
case ISCSI_PARAM_ABORT_TMO:
|
|
|
|
sscanf(buf, "%d", &session->abort_timeout);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_LU_RESET_TMO:
|
|
|
|
sscanf(buf, "%d", &session->lu_reset_timeout);
|
|
|
|
break;
|
2009-11-12 01:34:33 +03:00
|
|
|
case ISCSI_PARAM_TGT_RESET_TMO:
|
|
|
|
sscanf(buf, "%d", &session->tgt_reset_timeout);
|
|
|
|
break;
|
2007-12-13 21:43:30 +03:00
|
|
|
case ISCSI_PARAM_PING_TMO:
|
|
|
|
sscanf(buf, "%d", &conn->ping_timeout);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_RECV_TMO:
|
|
|
|
sscanf(buf, "%d", &conn->recv_timeout);
|
|
|
|
break;
|
2006-06-28 21:00:23 +04:00
|
|
|
case ISCSI_PARAM_MAX_RECV_DLENGTH:
|
|
|
|
sscanf(buf, "%d", &conn->max_recv_dlength);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_MAX_XMIT_DLENGTH:
|
|
|
|
sscanf(buf, "%d", &conn->max_xmit_dlength);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_HDRDGST_EN:
|
|
|
|
sscanf(buf, "%d", &conn->hdrdgst_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_DATADGST_EN:
|
|
|
|
sscanf(buf, "%d", &conn->datadgst_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_INITIAL_R2T_EN:
|
|
|
|
sscanf(buf, "%d", &session->initial_r2t_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_MAX_R2T:
|
2012-01-27 07:13:10 +04:00
|
|
|
sscanf(buf, "%hu", &session->max_r2t);
|
2006-06-28 21:00:23 +04:00
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_IMM_DATA_EN:
|
|
|
|
sscanf(buf, "%d", &session->imm_data_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_FIRST_BURST:
|
|
|
|
sscanf(buf, "%d", &session->first_burst);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_MAX_BURST:
|
|
|
|
sscanf(buf, "%d", &session->max_burst);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_PDU_INORDER_EN:
|
|
|
|
sscanf(buf, "%d", &session->pdu_inorder_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_DATASEQ_INORDER_EN:
|
|
|
|
sscanf(buf, "%d", &session->dataseq_inorder_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_ERL:
|
|
|
|
sscanf(buf, "%d", &session->erl);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_EXP_STATSN:
|
|
|
|
sscanf(buf, "%u", &conn->exp_statsn);
|
|
|
|
break;
|
2007-05-30 21:57:16 +04:00
|
|
|
case ISCSI_PARAM_USERNAME:
|
2009-05-14 02:57:40 +04:00
|
|
|
return iscsi_switch_str_param(&session->username, buf);
|
2007-05-30 21:57:16 +04:00
|
|
|
case ISCSI_PARAM_USERNAME_IN:
|
2009-05-14 02:57:40 +04:00
|
|
|
return iscsi_switch_str_param(&session->username_in, buf);
|
2007-05-30 21:57:16 +04:00
|
|
|
case ISCSI_PARAM_PASSWORD:
|
2009-05-14 02:57:40 +04:00
|
|
|
return iscsi_switch_str_param(&session->password, buf);
|
2007-05-30 21:57:16 +04:00
|
|
|
case ISCSI_PARAM_PASSWORD_IN:
|
2009-05-14 02:57:40 +04:00
|
|
|
return iscsi_switch_str_param(&session->password_in, buf);
|
2006-06-28 21:00:23 +04:00
|
|
|
case ISCSI_PARAM_TARGET_NAME:
|
2009-05-14 02:57:40 +04:00
|
|
|
return iscsi_switch_str_param(&session->targetname, buf);
|
2012-01-19 15:06:53 +04:00
|
|
|
case ISCSI_PARAM_TARGET_ALIAS:
|
|
|
|
return iscsi_switch_str_param(&session->targetalias, buf);
|
2006-06-28 21:00:23 +04:00
|
|
|
case ISCSI_PARAM_TPGT:
|
|
|
|
sscanf(buf, "%d", &session->tpgt);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_PERSISTENT_PORT:
|
|
|
|
sscanf(buf, "%d", &conn->persistent_port);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_PERSISTENT_ADDRESS:
|
2009-05-14 02:57:40 +04:00
|
|
|
return iscsi_switch_str_param(&conn->persistent_address, buf);
|
2008-05-22 00:54:16 +04:00
|
|
|
case ISCSI_PARAM_IFACE_NAME:
|
2009-05-14 02:57:40 +04:00
|
|
|
return iscsi_switch_str_param(&session->ifacename, buf);
|
2008-05-22 00:54:16 +04:00
|
|
|
case ISCSI_PARAM_INITIATOR_NAME:
|
2009-05-14 02:57:40 +04:00
|
|
|
return iscsi_switch_str_param(&session->initiatorname, buf);
|
2013-06-20 21:21:26 +04:00
|
|
|
case ISCSI_PARAM_BOOT_ROOT:
|
|
|
|
return iscsi_switch_str_param(&session->boot_root, buf);
|
|
|
|
case ISCSI_PARAM_BOOT_NIC:
|
|
|
|
return iscsi_switch_str_param(&session->boot_nic, buf);
|
|
|
|
case ISCSI_PARAM_BOOT_TARGET:
|
|
|
|
return iscsi_switch_str_param(&session->boot_target, buf);
|
2013-07-01 13:54:12 +04:00
|
|
|
case ISCSI_PARAM_PORTAL_TYPE:
|
|
|
|
return iscsi_switch_str_param(&session->portal_type, buf);
|
|
|
|
case ISCSI_PARAM_DISCOVERY_PARENT_TYPE:
|
|
|
|
return iscsi_switch_str_param(&session->discovery_parent_type,
|
|
|
|
buf);
|
2013-08-08 14:44:29 +04:00
|
|
|
case ISCSI_PARAM_DISCOVERY_SESS:
|
|
|
|
sscanf(buf, "%d", &val);
|
|
|
|
session->discovery_sess = !!val;
|
|
|
|
break;
|
2013-11-22 14:28:21 +04:00
|
|
|
case ISCSI_PARAM_LOCAL_IPADDR:
|
|
|
|
return iscsi_switch_str_param(&conn->local_ipaddr, buf);
|
2006-06-28 21:00:23 +04:00
|
|
|
default:
|
|
|
|
return -ENOSYS;
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_set_param);
|
|
|
|
|
|
|
|
int iscsi_session_get_param(struct iscsi_cls_session *cls_session,
|
|
|
|
enum iscsi_param param, char *buf)
|
|
|
|
{
|
2008-05-22 00:53:59 +04:00
|
|
|
struct iscsi_session *session = cls_session->dd_data;
|
2006-06-28 21:00:23 +04:00
|
|
|
int len;
|
|
|
|
|
|
|
|
switch(param) {
|
2007-12-13 21:43:20 +03:00
|
|
|
case ISCSI_PARAM_FAST_ABORT:
|
|
|
|
len = sprintf(buf, "%d\n", session->fast_abort);
|
|
|
|
break;
|
2007-12-13 21:43:30 +03:00
|
|
|
case ISCSI_PARAM_ABORT_TMO:
|
|
|
|
len = sprintf(buf, "%d\n", session->abort_timeout);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_LU_RESET_TMO:
|
|
|
|
len = sprintf(buf, "%d\n", session->lu_reset_timeout);
|
|
|
|
break;
|
2009-11-12 01:34:33 +03:00
|
|
|
case ISCSI_PARAM_TGT_RESET_TMO:
|
|
|
|
len = sprintf(buf, "%d\n", session->tgt_reset_timeout);
|
|
|
|
break;
|
2006-06-28 21:00:23 +04:00
|
|
|
case ISCSI_PARAM_INITIAL_R2T_EN:
|
|
|
|
len = sprintf(buf, "%d\n", session->initial_r2t_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_MAX_R2T:
|
|
|
|
len = sprintf(buf, "%hu\n", session->max_r2t);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_IMM_DATA_EN:
|
|
|
|
len = sprintf(buf, "%d\n", session->imm_data_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_FIRST_BURST:
|
|
|
|
len = sprintf(buf, "%u\n", session->first_burst);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_MAX_BURST:
|
|
|
|
len = sprintf(buf, "%u\n", session->max_burst);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_PDU_INORDER_EN:
|
|
|
|
len = sprintf(buf, "%d\n", session->pdu_inorder_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_DATASEQ_INORDER_EN:
|
|
|
|
len = sprintf(buf, "%d\n", session->dataseq_inorder_en);
|
|
|
|
break;
|
2013-07-22 15:46:10 +04:00
|
|
|
case ISCSI_PARAM_DEF_TASKMGMT_TMO:
|
|
|
|
len = sprintf(buf, "%d\n", session->def_taskmgmt_tmo);
|
|
|
|
break;
|
2006-06-28 21:00:23 +04:00
|
|
|
case ISCSI_PARAM_ERL:
|
|
|
|
len = sprintf(buf, "%d\n", session->erl);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_TARGET_NAME:
|
|
|
|
len = sprintf(buf, "%s\n", session->targetname);
|
|
|
|
break;
|
2012-01-19 15:06:53 +04:00
|
|
|
case ISCSI_PARAM_TARGET_ALIAS:
|
|
|
|
len = sprintf(buf, "%s\n", session->targetalias);
|
|
|
|
break;
|
2006-06-28 21:00:23 +04:00
|
|
|
case ISCSI_PARAM_TPGT:
|
|
|
|
len = sprintf(buf, "%d\n", session->tpgt);
|
|
|
|
break;
|
2007-05-30 21:57:16 +04:00
|
|
|
case ISCSI_PARAM_USERNAME:
|
|
|
|
len = sprintf(buf, "%s\n", session->username);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_USERNAME_IN:
|
|
|
|
len = sprintf(buf, "%s\n", session->username_in);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_PASSWORD:
|
|
|
|
len = sprintf(buf, "%s\n", session->password);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_PASSWORD_IN:
|
|
|
|
len = sprintf(buf, "%s\n", session->password_in);
|
|
|
|
break;
|
2008-05-22 00:54:16 +04:00
|
|
|
case ISCSI_PARAM_IFACE_NAME:
|
|
|
|
len = sprintf(buf, "%s\n", session->ifacename);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_INITIATOR_NAME:
|
2009-05-14 02:57:40 +04:00
|
|
|
len = sprintf(buf, "%s\n", session->initiatorname);
|
2008-05-22 00:54:16 +04:00
|
|
|
break;
|
2013-06-20 21:21:26 +04:00
|
|
|
case ISCSI_PARAM_BOOT_ROOT:
|
|
|
|
len = sprintf(buf, "%s\n", session->boot_root);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_BOOT_NIC:
|
|
|
|
len = sprintf(buf, "%s\n", session->boot_nic);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_BOOT_TARGET:
|
|
|
|
len = sprintf(buf, "%s\n", session->boot_target);
|
2013-07-22 15:46:09 +04:00
|
|
|
break;
|
2013-07-01 13:54:12 +04:00
|
|
|
case ISCSI_PARAM_AUTO_SND_TGT_DISABLE:
|
|
|
|
len = sprintf(buf, "%u\n", session->auto_snd_tgt_disable);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_DISCOVERY_SESS:
|
|
|
|
len = sprintf(buf, "%u\n", session->discovery_sess);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_PORTAL_TYPE:
|
|
|
|
len = sprintf(buf, "%s\n", session->portal_type);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_CHAP_AUTH_EN:
|
|
|
|
len = sprintf(buf, "%u\n", session->chap_auth_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_DISCOVERY_LOGOUT_EN:
|
|
|
|
len = sprintf(buf, "%u\n", session->discovery_logout_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_BIDI_CHAP_EN:
|
|
|
|
len = sprintf(buf, "%u\n", session->bidi_chap_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_DISCOVERY_AUTH_OPTIONAL:
|
|
|
|
len = sprintf(buf, "%u\n", session->discovery_auth_optional);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_DEF_TIME2WAIT:
|
|
|
|
len = sprintf(buf, "%d\n", session->time2wait);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_DEF_TIME2RETAIN:
|
|
|
|
len = sprintf(buf, "%d\n", session->time2retain);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_TSID:
|
|
|
|
len = sprintf(buf, "%u\n", session->tsid);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_ISID:
|
|
|
|
len = sprintf(buf, "%02x%02x%02x%02x%02x%02x\n",
|
|
|
|
session->isid[0], session->isid[1],
|
|
|
|
session->isid[2], session->isid[3],
|
|
|
|
session->isid[4], session->isid[5]);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_DISCOVERY_PARENT_IDX:
|
|
|
|
len = sprintf(buf, "%u\n", session->discovery_parent_idx);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_DISCOVERY_PARENT_TYPE:
|
|
|
|
if (session->discovery_parent_type)
|
|
|
|
len = sprintf(buf, "%s\n",
|
|
|
|
session->discovery_parent_type);
|
|
|
|
else
|
|
|
|
len = sprintf(buf, "\n");
|
2013-06-20 21:21:26 +04:00
|
|
|
break;
|
2006-06-28 21:00:23 +04:00
|
|
|
default:
|
|
|
|
return -ENOSYS;
|
|
|
|
}
|
|
|
|
|
|
|
|
return len;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_session_get_param);
|
|
|
|
|
2011-02-17 00:04:35 +03:00
|
|
|
int iscsi_conn_get_addr_param(struct sockaddr_storage *addr,
|
|
|
|
enum iscsi_param param, char *buf)
|
|
|
|
{
|
|
|
|
struct sockaddr_in6 *sin6 = NULL;
|
|
|
|
struct sockaddr_in *sin = NULL;
|
|
|
|
int len;
|
|
|
|
|
|
|
|
switch (addr->ss_family) {
|
|
|
|
case AF_INET:
|
|
|
|
sin = (struct sockaddr_in *)addr;
|
|
|
|
break;
|
|
|
|
case AF_INET6:
|
|
|
|
sin6 = (struct sockaddr_in6 *)addr;
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
switch (param) {
|
|
|
|
case ISCSI_PARAM_CONN_ADDRESS:
|
|
|
|
case ISCSI_HOST_PARAM_IPADDRESS:
|
|
|
|
if (sin)
|
|
|
|
len = sprintf(buf, "%pI4\n", &sin->sin_addr.s_addr);
|
|
|
|
else
|
|
|
|
len = sprintf(buf, "%pI6\n", &sin6->sin6_addr);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_CONN_PORT:
|
2014-09-29 22:55:42 +04:00
|
|
|
case ISCSI_PARAM_LOCAL_PORT:
|
2011-02-17 00:04:35 +03:00
|
|
|
if (sin)
|
|
|
|
len = sprintf(buf, "%hu\n", be16_to_cpu(sin->sin_port));
|
|
|
|
else
|
|
|
|
len = sprintf(buf, "%hu\n",
|
|
|
|
be16_to_cpu(sin6->sin6_port));
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
return len;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_conn_get_addr_param);
|
|
|
|
|
2006-06-28 21:00:23 +04:00
|
|
|
int iscsi_conn_get_param(struct iscsi_cls_conn *cls_conn,
|
|
|
|
enum iscsi_param param, char *buf)
|
|
|
|
{
|
|
|
|
struct iscsi_conn *conn = cls_conn->dd_data;
|
|
|
|
int len;
|
|
|
|
|
|
|
|
switch(param) {
|
2007-12-13 21:43:30 +03:00
|
|
|
case ISCSI_PARAM_PING_TMO:
|
|
|
|
len = sprintf(buf, "%u\n", conn->ping_timeout);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_RECV_TMO:
|
|
|
|
len = sprintf(buf, "%u\n", conn->recv_timeout);
|
|
|
|
break;
|
2006-06-28 21:00:23 +04:00
|
|
|
case ISCSI_PARAM_MAX_RECV_DLENGTH:
|
|
|
|
len = sprintf(buf, "%u\n", conn->max_recv_dlength);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_MAX_XMIT_DLENGTH:
|
|
|
|
len = sprintf(buf, "%u\n", conn->max_xmit_dlength);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_HDRDGST_EN:
|
|
|
|
len = sprintf(buf, "%d\n", conn->hdrdgst_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_DATADGST_EN:
|
|
|
|
len = sprintf(buf, "%d\n", conn->datadgst_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_IFMARKER_EN:
|
|
|
|
len = sprintf(buf, "%d\n", conn->ifmarker_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_OFMARKER_EN:
|
|
|
|
len = sprintf(buf, "%d\n", conn->ofmarker_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_EXP_STATSN:
|
|
|
|
len = sprintf(buf, "%u\n", conn->exp_statsn);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_PERSISTENT_PORT:
|
|
|
|
len = sprintf(buf, "%d\n", conn->persistent_port);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_PERSISTENT_ADDRESS:
|
|
|
|
len = sprintf(buf, "%s\n", conn->persistent_address);
|
|
|
|
break;
|
2013-07-01 13:54:12 +04:00
|
|
|
case ISCSI_PARAM_STATSN:
|
|
|
|
len = sprintf(buf, "%u\n", conn->statsn);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_MAX_SEGMENT_SIZE:
|
|
|
|
len = sprintf(buf, "%u\n", conn->max_segment_size);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_KEEPALIVE_TMO:
|
|
|
|
len = sprintf(buf, "%u\n", conn->keepalive_tmo);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_LOCAL_PORT:
|
|
|
|
len = sprintf(buf, "%u\n", conn->local_port);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_TCP_TIMESTAMP_STAT:
|
|
|
|
len = sprintf(buf, "%u\n", conn->tcp_timestamp_stat);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_TCP_NAGLE_DISABLE:
|
|
|
|
len = sprintf(buf, "%u\n", conn->tcp_nagle_disable);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_TCP_WSF_DISABLE:
|
|
|
|
len = sprintf(buf, "%u\n", conn->tcp_wsf_disable);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_TCP_TIMER_SCALE:
|
|
|
|
len = sprintf(buf, "%u\n", conn->tcp_timer_scale);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_TCP_TIMESTAMP_EN:
|
|
|
|
len = sprintf(buf, "%u\n", conn->tcp_timestamp_en);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_IP_FRAGMENT_DISABLE:
|
|
|
|
len = sprintf(buf, "%u\n", conn->fragment_disable);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_IPV4_TOS:
|
|
|
|
len = sprintf(buf, "%u\n", conn->ipv4_tos);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_IPV6_TC:
|
|
|
|
len = sprintf(buf, "%u\n", conn->ipv6_traffic_class);
|
|
|
|
break;
|
2013-07-22 15:46:10 +04:00
|
|
|
case ISCSI_PARAM_IPV6_FLOW_LABEL:
|
|
|
|
len = sprintf(buf, "%u\n", conn->ipv6_flow_label);
|
|
|
|
break;
|
2013-07-01 13:54:12 +04:00
|
|
|
case ISCSI_PARAM_IS_FW_ASSIGNED_IPV6:
|
|
|
|
len = sprintf(buf, "%u\n", conn->is_fw_assigned_ipv6);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_TCP_XMIT_WSF:
|
|
|
|
len = sprintf(buf, "%u\n", conn->tcp_xmit_wsf);
|
|
|
|
break;
|
|
|
|
case ISCSI_PARAM_TCP_RECV_WSF:
|
|
|
|
len = sprintf(buf, "%u\n", conn->tcp_recv_wsf);
|
|
|
|
break;
|
2013-11-22 14:28:21 +04:00
|
|
|
case ISCSI_PARAM_LOCAL_IPADDR:
|
|
|
|
len = sprintf(buf, "%s\n", conn->local_ipaddr);
|
|
|
|
break;
|
2006-06-28 21:00:23 +04:00
|
|
|
default:
|
|
|
|
return -ENOSYS;
|
|
|
|
}
|
|
|
|
|
|
|
|
return len;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_conn_get_param);
|
|
|
|
|
2007-05-30 21:57:12 +04:00
|
|
|
int iscsi_host_get_param(struct Scsi_Host *shost, enum iscsi_host_param param,
|
|
|
|
char *buf)
|
|
|
|
{
|
2008-05-22 00:53:59 +04:00
|
|
|
struct iscsi_host *ihost = shost_priv(shost);
|
2007-05-30 21:57:12 +04:00
|
|
|
int len;
|
|
|
|
|
|
|
|
switch (param) {
|
[SCSI] iscsi class, iscsi_tcp, iser, qla4xxx: add netdevname sysfs attr
iSCSI must support software iscsi (iscsi_tcp, iser), hardware iscsi (qla4xxx),
and partial offload (broadcom). To be able to allow each stack or driver
or port (virtual or physical) to be able to log into the same target portal
we use the initiator tuple [[HWADDRESS | NETDEVNAME], INITIATOR_NAME] and
the target tuple [TARGETNAME, CONN_ADDRESS, CONN_PORT] to id a session.
This patch adds the netdev name, which is used by software iscsi when
it binds a session to a netdevice using the SO_BINDTODEVICE sock opt.
It cannot use HWADDRESS because if someone did vlans then the same netdevice
will have the same mac and the initiator,target id will not be unique.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Cc: Roland Dreier <rdreier@cisco.com>
Cc: David C Somayajulu <david.somayajulu@qlogic.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2007-05-30 21:57:25 +04:00
|
|
|
case ISCSI_HOST_PARAM_NETDEV_NAME:
|
2009-05-14 02:57:40 +04:00
|
|
|
len = sprintf(buf, "%s\n", ihost->netdev);
|
[SCSI] iscsi class, iscsi_tcp, iser, qla4xxx: add netdevname sysfs attr
iSCSI must support software iscsi (iscsi_tcp, iser), hardware iscsi (qla4xxx),
and partial offload (broadcom). To be able to allow each stack or driver
or port (virtual or physical) to be able to log into the same target portal
we use the initiator tuple [[HWADDRESS | NETDEVNAME], INITIATOR_NAME] and
the target tuple [TARGETNAME, CONN_ADDRESS, CONN_PORT] to id a session.
This patch adds the netdev name, which is used by software iscsi when
it binds a session to a netdevice using the SO_BINDTODEVICE sock opt.
It cannot use HWADDRESS because if someone did vlans then the same netdevice
will have the same mac and the initiator,target id will not be unique.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Cc: Roland Dreier <rdreier@cisco.com>
Cc: David C Somayajulu <david.somayajulu@qlogic.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2007-05-30 21:57:25 +04:00
|
|
|
break;
|
2007-05-30 21:57:12 +04:00
|
|
|
case ISCSI_HOST_PARAM_HWADDRESS:
|
2009-05-14 02:57:40 +04:00
|
|
|
len = sprintf(buf, "%s\n", ihost->hwaddress);
|
2007-05-30 21:57:12 +04:00
|
|
|
break;
|
2007-05-30 21:57:13 +04:00
|
|
|
case ISCSI_HOST_PARAM_INITIATOR_NAME:
|
2009-05-14 02:57:40 +04:00
|
|
|
len = sprintf(buf, "%s\n", ihost->initiatorname);
|
2007-05-30 21:57:13 +04:00
|
|
|
break;
|
2007-05-30 21:57:12 +04:00
|
|
|
default:
|
|
|
|
return -ENOSYS;
|
|
|
|
}
|
|
|
|
|
|
|
|
return len;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_host_get_param);
|
|
|
|
|
|
|
|
int iscsi_host_set_param(struct Scsi_Host *shost, enum iscsi_host_param param,
|
|
|
|
char *buf, int buflen)
|
|
|
|
{
|
2008-05-22 00:53:59 +04:00
|
|
|
struct iscsi_host *ihost = shost_priv(shost);
|
2007-05-30 21:57:12 +04:00
|
|
|
|
|
|
|
switch (param) {
|
[SCSI] iscsi class, iscsi_tcp, iser, qla4xxx: add netdevname sysfs attr
iSCSI must support software iscsi (iscsi_tcp, iser), hardware iscsi (qla4xxx),
and partial offload (broadcom). To be able to allow each stack or driver
or port (virtual or physical) to be able to log into the same target portal
we use the initiator tuple [[HWADDRESS | NETDEVNAME], INITIATOR_NAME] and
the target tuple [TARGETNAME, CONN_ADDRESS, CONN_PORT] to id a session.
This patch adds the netdev name, which is used by software iscsi when
it binds a session to a netdevice using the SO_BINDTODEVICE sock opt.
It cannot use HWADDRESS because if someone did vlans then the same netdevice
will have the same mac and the initiator,target id will not be unique.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Cc: Roland Dreier <rdreier@cisco.com>
Cc: David C Somayajulu <david.somayajulu@qlogic.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
2007-05-30 21:57:25 +04:00
|
|
|
case ISCSI_HOST_PARAM_NETDEV_NAME:
|
2009-05-14 02:57:40 +04:00
|
|
|
return iscsi_switch_str_param(&ihost->netdev, buf);
|
2007-05-30 21:57:12 +04:00
|
|
|
case ISCSI_HOST_PARAM_HWADDRESS:
|
2009-05-14 02:57:40 +04:00
|
|
|
return iscsi_switch_str_param(&ihost->hwaddress, buf);
|
2007-05-30 21:57:13 +04:00
|
|
|
case ISCSI_HOST_PARAM_INITIATOR_NAME:
|
2009-05-14 02:57:40 +04:00
|
|
|
return iscsi_switch_str_param(&ihost->initiatorname, buf);
|
2007-05-30 21:57:12 +04:00
|
|
|
default:
|
|
|
|
return -ENOSYS;
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(iscsi_host_set_param);
|
|
|
|
|
2006-04-07 06:13:41 +04:00
|
|
|
MODULE_AUTHOR("Mike Christie");
|
|
|
|
MODULE_DESCRIPTION("iSCSI library functions");
|
|
|
|
MODULE_LICENSE("GPL");
|