Pull networking updates from David Miller:
1) Seccomp BPF filters can now be JIT'd, from Alexei Starovoitov.
2) Multiqueue support in xen-netback and xen-netfront, from Andrew J
Benniston.
3) Allow tweaking of aggregation settings in cdc_ncm driver, from Bjørn
Mork.
4) BPF now has a "random" opcode, from Chema Gonzalez.
5) Add more BPF documentation and improve test framework, from Daniel
Borkmann.
6) Support TCP fastopen over ipv6, from Daniel Lee.
7) Add software TSO helper functions and use them to support software
TSO in mvneta and mv643xx_eth drivers. From Ezequiel Garcia.
8) Support software TSO in fec driver too, from Nimrod Andy.
9) Add Broadcom SYSTEMPORT driver, from Florian Fainelli.
10) Handle broadcasts more gracefully over macvlan when there are large
numbers of interfaces configured, from Herbert Xu.
11) Allow more control over fwmark used for non-socket based responses,
from Lorenzo Colitti.
12) Do TCP congestion window limiting based upon measurements, from Neal
Cardwell.
13) Support busy polling in SCTP, from Neal Horman.
14) Allow RSS key to be configured via ethtool, from Venkata Duvvuru.
15) Bridge promisc mode handling improvements from Vlad Yasevich.
16) Don't use inetpeer entries to implement ID generation any more, it
performs poorly, from Eric Dumazet.
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next: (1522 commits)
rtnetlink: fix userspace API breakage for iproute2 < v3.9.0
tcp: fixing TLP's FIN recovery
net: fec: Add software TSO support
net: fec: Add Scatter/gather support
net: fec: Increase buffer descriptor entry number
net: fec: Factorize feature setting
net: fec: Enable IP header hardware checksum
net: fec: Factorize the .xmit transmit function
bridge: fix compile error when compiling without IPv6 support
bridge: fix smatch warning / potential null pointer dereference
via-rhine: fix full-duplex with autoneg disable
bnx2x: Enlarge the dorq threshold for VFs
bnx2x: Check for UNDI in uncommon branch
bnx2x: Fix 1G-baseT link
bnx2x: Fix link for KR with swapped polarity lane
sctp: Fix sk_ack_backlog wrap-around problem
net/core: Add VF link state control policy
net/fsl: xgmac_mdio is dependent on OF_MDIO
net/fsl: Make xgmac_mdio read error message useful
net_sched: drr: warn when qdisc is not work conserving
...
Fixed a bug that shows up with recv window sizes that exceed the size of
the RCV_BUFSIZ field in opt0 (>= 1024K). If the recv window exceeds
this, then we specify the max possible in opt0, add add the rest in via
a RX_DATA_ACK credits.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Select the appropriate hw mtu index and initial sequence number to optimize
hw memory performance.
Add new cxgb4_best_aligned_mtu() which allows callers to provide enough
information to be used to [possibly] select an MTU which will result in the
TCP Data Segment Size (AKA Maximum Segment Size) to be an aligned value.
If an RTR message exhange is required, then align the ISS to 8B - 1 + 4, so
that after the SYN the send seqno will align on a 4B boundary. The RTR
message exchange will leave the send seqno aligned on an 8B boundary.
If an RTR is not required, then align the ISS to 8B - 1. The goal is
to have the send seqno be 8B aligned when we send the first FPDU.
Based on original work by Casey Leedom <leeedom@chelsio.com> and
Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Casey Leedom <leedom@chelsio.com>
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently indirect interrupts for RDMA CQs funnel through the LLD's RDMA
RXQs, which also handle direct interrupts for offload CPLs during RDMA
connection setup/teardown. The intended T4 usage model, however, is to
have indirect interrupts flow through dedicated IQs. IE not to mix
indirect interrupts with CPL messages in an IQ. This patch adds the
concept of RDMA concentrator IQs, or CIQs, setup and maintained by the
LLD and exported to iw_cxgb4 for use when creating CQs. RDMA CPLs will
flow through the LLD's RDMA RXQs, and CQ interrupts flow through the
CIQs.
Design:
cxgb4 creates and exports an array of CIQs for the RDMA ULD. These IQs
are sized according to the max available CQs available at adapter init.
In addition, these IQs don't need FL buffers since they only service
indirect interrupts. One CIQ is setup per RX channel similar to the
RDMA RXQs.
iw_cxgb4 will utilize these CIQs based on the vector value passed into
create_cq(). The num_comp_vectors advertised by iw_cxgb4 will be the
number of CIQs configured, and thus the vector value will be the index
into the array of CIQs.
Based on original work by Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Based on original work by Vipul Pandya.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
[ Fix htons -> ntohs to make sparse happy. - Roland ]
Signed-off-by: Roland Dreier <roland@purestorage.com>
The i386 ABI disagrees with most other ABIs regarding alignment of
data types larger than 4 bytes: on most ABIs a padding must be added
at end of the structures, while it is not required on i386.
So for most ABI struct c4iw_alloc_ucontext_resp gets implicitly padded
to be aligned on a 8 bytes multiple, while for i386, such padding is
not added.
The tool pahole can be used to find such implicit padding:
$ pahole --anon_include \
--nested_anon_include \
--recursive \
--class_name c4iw_alloc_ucontext_resp \
drivers/infiniband/hw/cxgb4/iw_cxgb4.o
Then, structure layout can be compared between i386 and x86_64:
+++ obj-i386/drivers/infiniband/hw/cxgb4/iw_cxgb4.o.pahole.txt 2014-03-28 11:43:05.547432195 +0100
--- obj-x86_64/drivers/infiniband/hw/cxgb4/iw_cxgb4.o.pahole.txt 2014-03-28 10:55:10.990133017 +0100
@@ -2,9 +2,8 @@ struct c4iw_alloc_ucontext_resp {
__u64 status_page_key; /* 0 8 */
__u32 status_page_size; /* 8 4 */
- /* size: 12, cachelines: 1, members: 2 */
- /* last cacheline: 12 bytes */
+ /* size: 16, cachelines: 1, members: 2 */
+ /* padding: 4 */
+ /* last cacheline: 16 bytes */
};
This ABI disagreement will make an x86_64 kernel try to write past the
buffer provided by an i386 binary.
When boundary check will be implemented, the x86_64 kernel will refuse
to write past the i386 userspace provided buffer and the uverbs will
fail.
If the structure is on a page boundary and the next page is not
mapped, ib_copy_to_udata() will fail and the uverb will fail.
Additionally, as reported by Dan Carpenter, without the implicit
padding being properly cleared, an information leak would take place
in most architectures.
This patch adds an explicit padding to struct c4iw_alloc_ucontext_resp,
and, like 92b0ca7cb1 ("IB/mlx5: Fix stack info leak in
mlx5_ib_alloc_ucontext()"), makes function c4iw_alloc_ucontext()
not writting this padding field to userspace. This way, x86_64 kernel
will be able to write struct c4iw_alloc_ucontext_resp as expected by
unpatched and patched i386 libcxgb4.
Link: http://marc.info/?i=cover.1399309513.git.ydroneaud@opteya.com
Link: http://marc.info/?i=1395848977.3297.15.camel@localhost.localdomain
Link: http://marc.info/?i=20140328082428.GH25192@mwanda
Cc: <stable@vger.kernel.org>
Fixes: 05eb23893c ("cxgb4/iw_cxgb4: Doorbell Drop Avoidance Bug Fixes")
Reported-by: Yann Droneaud <ydroneaud@opteya.com>
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Yann Droneaud <ydroneaud@opteya.com>
Acked-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The i386 ABI disagrees with most other ABIs regarding alignment of
data types larger than 4 bytes: on most ABIs a padding must be added
at end of the structures, while it is not required on i386.
So for most ABI struct c4iw_create_cq_resp gets implicitly padded
to be aligned on a 8 bytes multiple, while for i386, such padding
is not added.
The tool pahole can be used to find such implicit padding:
$ pahole --anon_include \
--nested_anon_include \
--recursive \
--class_name c4iw_create_cq_resp \
drivers/infiniband/hw/cxgb4/iw_cxgb4.o
Then, structure layout can be compared between i386 and x86_64:
+++ obj-i386/drivers/infiniband/hw/cxgb4/iw_cxgb4.o.pahole.txt 2014-03-28 11:43:05.547432195 +0100
--- obj-x86_64/drivers/infiniband/hw/cxgb4/iw_cxgb4.o.pahole.txt 2014-03-28 10:55:10.990133017 +0100
@@ -14,9 +13,8 @@ struct c4iw_create_cq_resp {
__u32 size; /* 28 4 */
__u32 qid_mask; /* 32 4 */
- /* size: 36, cachelines: 1, members: 6 */
- /* last cacheline: 36 bytes */
+ /* size: 40, cachelines: 1, members: 6 */
+ /* padding: 4 */
+ /* last cacheline: 40 bytes */
};
This ABI disagreement will make an x86_64 kernel try to write past the
buffer provided by an i386 binary.
When boundary check will be implemented, the x86_64 kernel will refuse
to write past the i386 userspace provided buffer and the uverbs will
fail.
If the structure is on a page boundary and the next page is not
mapped, ib_copy_to_udata() will fail and the uverb will fail.
This patch adds an explicit padding at end of structure
c4iw_create_cq_resp, and, like 92b0ca7cb1 ("IB/mlx5: Fix stack info
leak in mlx5_ib_alloc_ucontext()"), makes function c4iw_create_cq()
not writting this padding field to userspace. This way, x86_64 kernel
will be able to write struct c4iw_create_cq_resp as expected by
unpatched and patched i386 libcxgb4.
Link: http://marc.info/?i=cover.1399309513.git.ydroneaud@opteya.com
Cc: <stable@vger.kernel.org>
Fixes: cfdda9d764 ("RDMA/cxgb4: Add driver for Chelsio T4 RNIC")
Fixes: e24a72a330 ("RDMA/cxgb4: Fix four byte info leak in c4iw_create_cq()")
Cc: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Yann Droneaud <ydroneaud@opteya.com>
Acked-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
RDMA connections over a vlan interface don't work due to
import_ep() not using the correct egress device.
- use the real device in import_ep()
- use rdma_vlan_dev_real_dev() in get_real_dev().
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
c4iw_alloc() bails out without freeing the storage that 'devp' points to.
Picked up by Coverity - CID 1204241.
Fixes: fa658a98a2 ("RDMA/cxgb4: Use the BAR2/WC path for kernel QPs and T5 devices")
Signed-off-by: Christoph Jaeger <christophjaeger@linux.com>
Acked-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The whole db drop avoidance stuff is for T4 only. So we cannot allow
that to be enabled for T5 devices.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
This is required to work around a T5 HW issue.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
In cases where the cm calls c4iw_modify_rc_qp() with the endpoint
mutex held, they must be called with internal == 1. rx_data() and
process_mpa_reply() are not doing this. This causes a deadlock
because c4iw_modify_rc_qp() might call c4iw_ep_disconnect() in some
!internal cases, and c4iw_ep_disconnect() acquires the endpoint mutex.
The design was intended to only do the disconnect for !internal calls.
Change rx_data(), FPDU_MODE case, to call c4iw_modify_rc_qp() with
internal == 1, and then disconnect only after releasing the mutex.
Change process_mpa_reply() to call c4iw_modify_rc_qp(TERMINATE) with
internal == 1 and set a new attr flag telling it to send a TERMINATE
message. Previously this was implied by !internal.
Change process_mpa_reply() to return whether the caller should
disconnect after releasing the endpoint mutex. Now rx_data() will do
the disconnect in the cases where process_mpa_reply() wants to
disconnect after the TERMINATE is sent.
Change c4iw_modify_rc_qp() RTS->TERM to only disconnect if !internal,
and to send a TERMINATE message if attrs->send_term is 1.
Change abort_connection() to not aquire the ep mutex for setting the
state, and make all calls to abort_connection() do so with the mutex
held.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Need to get the endpoint reference before calling rdma_fini(), which
might fail causing us to not get the reference.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The max depth of a fastreg mr depends on whether the device supports
DSGL or not. So compute it dynamically based on the device support
and the module use_dsgl option.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
There is a race when moving a QP from RTS->CLOSING where a SQ work
request could be posted after the FW receives the RDMA_RI/FINI WR.
The SQ work request will never get processed, and should be completed
with FLUSHED status. Function c4iw_flush_sq(), however was dropping
the oldest SQ work request when in CLOSING or IDLE states, instead of
completing the pending work request. If that oldest pending work
request was actually complete and has a CQE in the CQ, then when that
CQE is proceessed in poll_cq, we'll BUG_ON() due to the inconsistent
SQ/CQ state.
This is a very small timing hole and has only been hit once so far.
The fix is two-fold:
1) c4iw_flush_sq() MUST always flush all non-completed WRs with FLUSHED
status regardless of the QP state.
2) In c4iw_modify_rc_qp(), always set the "in error" bit on the queue
before moving the state out of RTS. This ensures that the state
transition will not happen while another thread is in
post_rc_send(), because set_state() and post_rc_send() both aquire
the qp spinlock. Also, once we transition the state out of RTS,
subsequent calls to post_rc_send() will fail because the "in error"
bit is set. I don't think this fully closes the race where the FW
can get a FINI followed a SQ work request being posted (because
they are posted to differente EQs), but the #1 fix will handle the
issue by flushing the SQ work request.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Some HW platforms can reorder read operations, so we must rmb() after
we see a valid gen bit in a CQE but before we read any other fields
from the CQE.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
1) timedout endpoint processing can be starved. If there are continual
CPL messages flowing into the driver, the endpoint timeout
processing can be starved. This condition exposed the other bugs
below.
Solution: In process_work(), call process_timedout_eps() after each CPL
is processed.
2) Connection events can be processed even though the endpoint is on
the timeout list. If the endpoint is scheduled for timeout
processing, then we must ignore MPA Start Requests and Replies.
Solution: Change stop_ep_timer() to return 1 if the ep has already been
queued for timeout processing. All the callers of stop_ep_timer() need
to check this and act accordingly. There are just a few cases where
the caller needs to do something different if stop_ep_timer() returns 1:
1) in process_mpa_reply(), ignore the reply and process_timeout()
will abort the connection.
2) in process_mpa_request, ignore the request and process_timeout()
will abort the connection.
It is ok for callers of stop_ep_timer() to abort the connection since
that will leave the state in ABORTING or DEAD, and process_timeout()
now ignores timeouts when the ep is in these states.
3) Double insertion on the timeout list. Since the endpoint timers
are used for connection setup and teardown, we need to guard
against the possibility that an endpoint is already on the timeout
list. This is a rare condition and only seen under heavy load and
in the presense of the above 2 bugs.
Solution: In ep_timeout(), don't queue the endpoint if it is already on
the queue.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
[ Fix cast from u64* to integer. - Roland ]
Signed-off-by: Roland Dreier <roland@purestorage.com>
- The biggest change is core API extensions and mlx5 low-level driver
support for handling DIF/DIX-style protection information, and the
addition of PI support to the iSER initiator. Target support will be
arriving shortly through the SCSI target tree.
- A nice simplification to the "umem" memory pinning library now that
we have chained sg lists. Kudos to Yishai Hadas for realizing our
code didn't have to be so crazy.
- Another nice simplification to the sg wrappers used by qib, ipath and
ehca to handle their mapping of memory to adapter.
- The usual batch of fixes to bugs found by static checkers etc. from
intrepid people like Dan Carpenter and Yann Droneaud.
- A large batch of cxgb4, ocrdma, qib driver updates.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1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=Ww6H
-----END PGP SIGNATURE-----
Merge tag 'rdma-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland/infiniband
Pull infiniband updates from Roland Dreier:
"Main batch of InfiniBand/RDMA changes for 3.15:
- The biggest change is core API extensions and mlx5 low-level driver
support for handling DIF/DIX-style protection information, and the
addition of PI support to the iSER initiator. Target support will
be arriving shortly through the SCSI target tree.
- A nice simplification to the "umem" memory pinning library now that
we have chained sg lists. Kudos to Yishai Hadas for realizing our
code didn't have to be so crazy.
- Another nice simplification to the sg wrappers used by qib, ipath
and ehca to handle their mapping of memory to adapter.
- The usual batch of fixes to bugs found by static checkers etc.
from intrepid people like Dan Carpenter and Yann Droneaud.
- A large batch of cxgb4, ocrdma, qib driver updates"
* tag 'rdma-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland/infiniband: (102 commits)
RDMA/ocrdma: Unregister inet notifier when unloading ocrdma
RDMA/ocrdma: Fix warnings about pointer <-> integer casts
RDMA/ocrdma: Code clean-up
RDMA/ocrdma: Display FW version
RDMA/ocrdma: Query controller information
RDMA/ocrdma: Support non-embedded mailbox commands
RDMA/ocrdma: Handle CQ overrun error
RDMA/ocrdma: Display proper value for max_mw
RDMA/ocrdma: Use non-zero tag in SRQ posting
RDMA/ocrdma: Memory leak fix in ocrdma_dereg_mr()
RDMA/ocrdma: Increment abi version count
RDMA/ocrdma: Update version string
be2net: Add abi version between be2net and ocrdma
RDMA/ocrdma: ABI versioning between ocrdma and be2net
RDMA/ocrdma: Allow DPP QP creation
RDMA/ocrdma: Read ASIC_ID register to select asic_gen
RDMA/ocrdma: SQ and RQ doorbell offset clean up
RDMA/ocrdma: EQ full catastrophe avoidance
RDMA/cxgb4: Disable DSGL use by default
RDMA/cxgb4: rx_data() needs to hold the ep mutex
...
Current hardware doesn't correctly support DSGL.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
To avoid racing with other threads doing close/flush/whatever, rx_data()
should hold the endpoint mutex.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
There is a race between ULP threads doing an accept/reject, and the
ingress processing thread handling close/abort for the same connection.
The accept/reject path needs to hold the lock to serialize these paths.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
[ Fold in locking fix found by Dan Carpenter <dan.carpenter@oracle.com>.
- Roland ]
Signed-off-by: Roland Dreier <roland@purestorage.com>
If kmalloc() fails in c4iw_alloc_ucontext(), the function
leaves but does not set an error code in ret variable:
it will return 0 to the caller.
This patch set ret to -ENOMEM in such case.
Cc: Steve Wise <swise@opengridcomputing.com>
Cc: Steve Wise <swise@chelsio.com>
Signed-off-by: Yann Droneaud <ydroneaud@opteya.com>
Acked-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When processing an MPA Start Request, if the listening endpoint is
DEAD, then abort the connection.
If the IWCM returns an error, then we must abort the connection and
release resources. Also abort_connection() should not post a CLOSE
event, so clean that up too.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
These are generated by HW in some error cases and need to be
silently discarded.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
If cxgb4_ofld_send() returns < 0, then send_fw_pass_open_req() must
free the request skb and the saved skb with the tcp header.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
We cannot save the mapped length using the rdma max_page_list_len field
of the ib_fast_reg_page_list struct because the core code uses it. This
results in an incorrect unmap of the page list in c4iw_free_fastreg_pbl().
I found this with dma mapping debugging enabled in the kernel. The
fix is to save the length in the c4iw_fr_page_list struct.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Based on original work from Jay Hernandez <jay@chelsio.com>
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Always release the neigh entry in rx_pkt().
Based on original work by Santosh Rastapur <santosh@chelsio.com>.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
find_route() must treat loopback as a valid egress interface.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
There is a four byte hole at the end of the "uresp" struct after the
->qid_mask member.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
These sizes should be unsigned so we don't allow negative values and
have underflow bugs. These can come from the user so there may be
security implications, but I have not tested this.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The current logic suffers from a slow response time to disable user DB
usage, and also fails to avoid DB FIFO drops under heavy load. This commit
fixes these deficiencies and makes the avoidance logic more optimal.
This is done by more efficiently notifying the ULDs of potential DB
problems, and implements a smoother flow control algorithm in iw_cxgb4,
which is the ULD that puts the most load on the DB fifo.
Design:
cxgb4:
Direct ULD callback from the DB FULL/DROP interrupt handler. This allows
the ULD to stop doing user DB writes as quickly as possible.
While user DB usage is disabled, the LLD will accumulate DB write events
for its queues. Then once DB usage is reenabled, a single DB write is
done for each queue with its accumulated write count. This reduces the
load put on the DB fifo when reenabling.
iw_cxgb4:
Instead of marking each qp to indicate DB writes are disabled, we create
a device-global status page that each user process maps. This allows
iw_cxgb4 to only set this single bit to disable all DB writes for all
user QPs vs traversing the idr of all the active QPs. If the libcxgb4
doesn't support this, then we fall back to the old approach of marking
each QP. Thus we allow the new driver to work with an older libcxgb4.
When the LLD upcalls iw_cxgb4 indicating DB FULL, we disable all DB writes
via the status page and transition the DB state to STOPPED. As user
processes see that DB writes are disabled, they call into iw_cxgb4
to submit their DB write events. Since the DB state is in STOPPED,
the QP trying to write gets enqueued on a new DB "flow control" list.
As subsequent DB writes are submitted for this flow controlled QP, the
amount of writes are accumulated for each QP on the flow control list.
So all the user QPs that are actively ringing the DB get put on this
list and the number of writes they request are accumulated.
When the LLD upcalls iw_cxgb4 indicating DB EMPTY, which is in a workq
context, we change the DB state to FLOW_CONTROL, and begin resuming all
the QPs that are on the flow control list. This logic runs on until
the flow control list is empty or we exit FLOW_CONTROL mode (due to
a DB DROP upcall, for example). QPs are removed from this list, and
their accumulated DB write counts written to the DB FIFO. Sets of QPs,
called chunks in the code, are removed at one time. The chunk size is 64.
So 64 QPs are resumed at a time, and before the next chunk is resumed, the
logic waits (blocks) for the DB FIFO to drain. This prevents resuming to
quickly and overflowing the FIFO. Once the flow control list is empty,
the db state transitions back to NORMAL and user QPs are again allowed
to write directly to the user DB register.
The algorithm is designed such that if the DB write load is high enough,
then all the DB writes get submitted by the kernel using this flow
controlled approach to avoid DB drops. As the load lightens though, we
resume to normal DB writes directly by user applications.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Based on original work by Anand Priyadarshee <anandp@chelsio.com>.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch refactors the IB core umem code and vendor drivers to use a
linear (chained) SG table instead of chunk list. With this change the
relevant code becomes clearer—no need for nested loops to build and
use umem.
Signed-off-by: Shachar Raindel <raindel@mellanox.com>
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Building mem.o for 32 bits x86 triggers a GCC warning:
drivers/infiniband/hw/cxgb4/mem.c: In function '_c4iw_write_mem_dma_aligned':
drivers/infiniband/hw/cxgb4/mem.c:79:25: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast]
Silence that warning by casting "&wr_wait" to unsigned long before
casting it to __be64. That's what _c4iw_write_mem_inline() already does.
Signed-off-by: Paul Bolle <pebolle@tiscali.nl>
Acked-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Linus Torvalds
Hariprasad Shenai
Roland Dreier
Steve Wise
Yann Droneaud
Christoph Jaeger
Yann Droneaud
Dan Carpenter
Yishai Hadas
Kumar Sanghvi
Paul Bolle