This patch fixes a race in nodeid2con in cases that we parallel running
a lookup and both will create a connection structure for the same nodeid.
It's a rare case to create a new connection structure to keep reader
lockless we just do a lookup inside the protection area again and drop
previous work if this race happens.
Fixes: a47666eb76 ("fs: dlm: make connection hash lockless")
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
This patch reworks the current receive handling of dlm. As I tried to
change the send handling to fix reorder issues I took a look into the
receive handling and simplified it, it works as the following:
Each connection has a preallocated receive buffer with a minimum length of
4096. On receive, the upper layer protocol will process all dlm message
until there is not enough data anymore. If there exists "leftover" data at
the end of the receive buffer because the dlm message wasn't fully received
it will be copied to the begin of the preallocated receive buffer. Next
receive more data will be appended to the previous "leftover" data and
processing will begin again.
This will remove a lot of code of the current mechanism. Inside the
processing functionality we will ensure with a memmove() that the dlm
message should be memory aligned. To have a dlm message always started
at the beginning of the buffer will reduce some amount of memmove()
calls because src and dest pointers are the same.
The cluster attribute "buffer_size" becomes a new meaning, it's now the
size of application layer receive buffer size. If this is changed during
runtime the receive buffer will be reallocated. It's important that the
receive buffer size has at minimum the size of the maximum possible dlm
message size otherwise the received message cannot be placed inside
the receive buffer size.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
This patch fixes to set per nodeid mark configuration for accepted
sockets as well. Before this patch only the listen socket mark value was
used for all accepted connections. This patch will ensure that the
cluster mark attribute value will be always used for all sockets, if a
per nodeid mark value is specified dlm will use this value for the
specific node.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
During my experiments to make dlm robust against tcpkill application I
was able to run sometimes in a circular lock dependency warning between
clusters_root.subsys.su_mutex and con->sock_mutex. We don't need to
held the sock_mutex when getting the mark value which held the
clusters_root.subsys.su_mutex. This patch moves the specific handling
just before the sock_mutex will be held.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
This patch use free_con() functionality to free the listen connection if
listen fails. It also fixes an issue that a freed resource is still part
of the connection_hash as hlist_del() is not called in this case. The
only difference is that free_con() handles othercon as well, but this is
never been set for the listen connection.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
This patch adds free of possible other writequeue entries in othercon
member of struct connection.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
This patch just move the free of struct connection member writequeue
into the functionality when struct connection will be freed instead of
doing two iterations.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
There are some problems with the connections_lock. During my
experiements I saw sometimes circular dependencies with sock_lock.
The reason here might be code parts which runs nodeid2con() before
or after sock_lock is acquired.
Another issue are missing locks in for_conn() iteration. Maybe this
works fine because for_conn() is running in a context where
connection_hash cannot be manipulated by others anymore.
However this patch changes the connection_hash to be protected by
sleepable rcu. The hotpath function __find_con() is implemented
lockless as it is only a reader of connection_hash and this hopefully
fixes the circular locking dependencies. The iteration for_conn() will
still call some sleepable functionality, that's why we use sleepable rcu
in this case.
This patch removes the kmemcache functionality as I think I need to
make some free() functionality via call_rcu(). However allocation time
isn't here an issue. The dlm_allow_con will not be protected by a lock
anymore as I think it's enough to just set and flush workqueues
afterwards.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
This patch moves the dlm workqueue dlm synchronization before shutdown
handling. The patch just flushes all pending work before starting to
shutdown the connection. At least for the send_workqeue we should flush
the workqueue to make sure there is no new connection handling going on
as dlm_allow_conn switch is turned to false before.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
During my code inspection I saw there is no implementation of a graceful
shutdown for tcp. This patch will introduce a graceful shutdown for tcp
connections. The shutdown is implemented synchronized as
dlm_lowcomms_stop() is called to end all dlm communication. After shutdown
is done, a lot of flush and closing functionality will be called. However
I don't see a problem with that.
The waitqueue for synchronize the shutdown has a timeout of 10 seconds, if
timeout a force close will be exectued.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
This patch changes the handling of reconnects. At first we only close
the connection related to the communication failure. If we get a new
connection for an already existing connection we close the existing
connection and take the new one.
This patch improves significantly the stability of tcp connections while
running "tcpkill -9 -i $IFACE port 21064" while generating a lot of dlm
messages e.g. on a gfs2 mount with many files. My test setup shows that a
deadlock is "more" unlikely. Before this patch I wasn't able to get
not a deadlock after 5 seconds. After this patch my observation is
that it's more likely to survive after 5 seconds and more, but still a
deadlock occurs after certain time. My guess is that there are still
"segments" inside the tcp writequeue or retransmit queue which get dropped
when receiving a tcp reset [1]. Hard to reproduce because the right message
need to be inside these queues, which might even be in the 5 first seconds
with this patch.
[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/net/ipv4/tcp_input.c?h=v5.8-rc6#n4122
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
This patch doesn't close sockets when there is an invalid dlm message
received. The connection will probably reconnect anyway so. To not
close the connection will reduce the number of possible failtures.
As we don't have a different strategy to react on such scenario
just keep going the connection and ignore the message.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
This patch adds support to set the skb mark value for the DLM tcp and
sctp socket per peer. The mark value will be offered as per comm value
of configfs. At creation time of the peer socket it will be set as
socket option.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
This patch adds support to set the skb mark value for the DLM listen
tcp and sctp sockets. The mark value will be offered as cluster
configuration. At creation time of the listen socket it will be set as
socket option.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
The SCTP protocol allows to bind multiple address to a socket. That
feature is currently only exposed as a socket option. Add a bind_add
method struct proto that allows to bind additional addresses, and
switch the dlm code to use the method instead of going through the
socket option from kernel space.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a helper to directly set the SCTP_NODELAY sockopt from kernel space
without going through a fake uaccess.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a helper to directly set the TCP_NODELAY sockopt from kernel space
without going through a fake uaccess. Cleanup the callers to avoid
pointless wrappers now that this is a simple function call.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Sagi Grimberg <sagi@grimberg.me>
Acked-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a helper to directly set the SO_RCVBUFFORCE sockopt from kernel space
without going through a fake uaccess.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a helper to directly set the SO_KEEPALIVE sockopt from kernel space
without going through a fake uaccess.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a helper to directly set the SO_SNDTIMEO_NEW sockopt from kernel
space without going through a fake uaccess. The interface is
simplified to only pass the seconds value, as that is the only
thing needed at the moment.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a helper to directly set the SO_REUSEADDR sockopt from kernel space
without going through a fake uaccess.
For this the iscsi target now has to formally depend on inet to avoid
a mostly theoretical compile failure. For actual operation it already
did depend on having ipv4 or ipv6 support.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: David S. Miller <davem@davemloft.net>
The only difference between a few missing fixes applied to the SCTP
one is that TCP uses ->getpeername to get the remote address, while
SCTP uses kernel_getsockopt(.. SCTP_PRIMARY_ADDR). But given that
getpeername is defined to return the primary address for sctp, there
doesn't seem to be any reason for the different way of quering the
peername, or all the code duplication.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Eliminate one more use of 'struct timeval' from the kernel so
we can eventually remove the definition as well.
The kernel supports the new format with a 64-bit time_t version
of timeval here, so use that instead of the old timeval.
Acked-by: Deepa Dinamani <deepa.kernel@gmail.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
If the DLM lowcomms stack is shut down before any DLM
traffic can be generated, flush_workqueue() and
destroy_workqueue() can be called on empty send and/or recv
workqueues.
Insert guard conditionals to only call flush_workqueue()
and destroy_workqueue() on workqueues that are not NULL.
Signed-off-by: David Windsor <dwindsor@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
Based on 1 normalized pattern(s):
this copyrighted material is made available to anyone wishing to use
modify copy or redistribute it subject to the terms and conditions
of the gnu general public license v 2
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-only
has been chosen to replace the boilerplate/reference in 45 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Richard Fontana <rfontana@redhat.com>
Reviewed-by: Allison Randal <allison@lohutok.net>
Reviewed-by: Steve Winslow <swinslow@gmail.com>
Reviewed-by: Alexios Zavras <alexios.zavras@intel.com>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190528170027.342746075@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
SO_RCVTIMEO and SO_SNDTIMEO socket options use struct timeval
as the time format. struct timeval is not y2038 safe.
The subsequent patches in the series add support for new socket
timeout options with _NEW suffix that will use y2038 safe
data structures. Although the existing struct timeval layout
is sufficiently wide to represent timeouts, because of the way
libc will interpret time_t based on user defined flag, these
new flags provide a way of having a structure that is the same
for all architectures consistently.
Rename the existing options with _OLD suffix forms so that the
right option is enabled for userspace applications according
to the architecture and time_t definition of libc.
Signed-off-by: Deepa Dinamani <deepa.kernel@gmail.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Cc: ccaulfie@redhat.com
Cc: deller@gmx.de
Cc: paulus@samba.org
Cc: ralf@linux-mips.org
Cc: rth@twiddle.net
Cc: cluster-devel@redhat.com
Cc: linuxppc-dev@lists.ozlabs.org
Cc: linux-alpha@vger.kernel.org
Cc: linux-arch@vger.kernel.org
Cc: linux-mips@vger.kernel.org
Cc: linux-parisc@vger.kernel.org
Cc: sparclinux@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
In the iov_iter struct, separate the iterator type from the iterator
direction and use accessor functions to access them in most places.
Convert a bunch of places to use switch-statements to access them rather
then chains of bitwise-AND statements. This makes it easier to add further
iterator types. Also, this can be more efficient as to implement a switch
of small contiguous integers, the compiler can use ~50% fewer compare
instructions than it has to use bitwise-and instructions.
Further, cease passing the iterator type into the iterator setup function.
The iterator function can set that itself. Only the direction is required.
Signed-off-by: David Howells <dhowells@redhat.com>
We should remove O_NONBLOCK flag when calling sock->ops->connect()
in sctp_connect_to_sock() function.
Why?
1. up to now, sctp socket connect() function ignores the flag argument,
that means O_NONBLOCK flag does not take effect, then we should remove
it to avoid the confusion (but is not urgent).
2. for the future, there will be a patch to fix this problem, then the flag
argument will take effect, the patch has been queued at https://git.kernel.o
rg/pub/scm/linux/kernel/git/davem/net.git/commit/net/sctp?id=644fbdeacf1d3ed
d366e44b8ba214de9d1dd66a9.
But, the O_NONBLOCK flag will make sock->ops->connect() directly return
without any wait time, then the connection will not be established, DLM kernel
module will call sock->ops->connect() again and again, the bad results are,
CPU usage is almost 100%, even trigger soft_lockup problem if the related
configurations are enabled,
DLM kernel module also prints lots of messages like,
[Fri Apr 27 11:23:43 2018] dlm: connecting to 172167592
[Fri Apr 27 11:23:43 2018] dlm: connecting to 172167592
[Fri Apr 27 11:23:43 2018] dlm: connecting to 172167592
[Fri Apr 27 11:23:43 2018] dlm: connecting to 172167592
The upper application (e.g. ocfs2 mount command) is hanged at new_lockspace(),
the whole backtrace is as below,
tb0307-nd2:~ # cat /proc/2935/stack
[<0>] new_lockspace+0x957/0xac0 [dlm]
[<0>] dlm_new_lockspace+0xae/0x140 [dlm]
[<0>] user_cluster_connect+0xc3/0x3a0 [ocfs2_stack_user]
[<0>] ocfs2_cluster_connect+0x144/0x220 [ocfs2_stackglue]
[<0>] ocfs2_dlm_init+0x215/0x440 [ocfs2]
[<0>] ocfs2_fill_super+0xcb0/0x1290 [ocfs2]
[<0>] mount_bdev+0x173/0x1b0
[<0>] mount_fs+0x35/0x150
[<0>] vfs_kern_mount.part.23+0x54/0x100
[<0>] do_mount+0x59a/0xc40
[<0>] SyS_mount+0x80/0xd0
[<0>] do_syscall_64+0x76/0x140
[<0>] entry_SYSCALL_64_after_hwframe+0x42/0xb7
[<0>] 0xffffffffffffffff
So, I think we should remove O_NONBLOCK flag here, since DLM kernel module can
not handle non-block sockect in connect() properly.
Signed-off-by: Gang He <ghe@suse.com>
Signed-off-by: David Teigland <teigland@redhat.com>
When the user setup a two-ring cluster, DLM kernel module
will automatically selects to use SCTP protocol to communicate
between each node. There will be about 5 minute hang in DLM
kernel module, in case one ring is broken before switching to
another ring, this will potentially affect the dependent upper
applications, e.g. ocfs2, gfs2, clvm and clustered-MD, etc.
Unfortunately, if the user setup a two-ring cluster, we can not
specify DLM communication protocol with TCP explicitly, since
DLM kernel module only supports SCTP protocol for multiple
ring cluster.
Base on my investigation, the time is spent in sock->ops->connect()
function before returns ETIMEDOUT(-110) error, since O_NONBLOCK
argument in connect() function does not work here, then we should
make sock->ops->connect() function return in specified time via
setting socket SO_SNDTIMEO atrribute.
Signed-off-by: Gang He <ghe@suse.com>
Signed-off-by: David Teigland <teigland@redhat.com>
There is a clerical error when turn off Nagle's algorithm in
sctp_connect_to_sock() function, this results in turn off
Nagle's algorithm failure.
After this correction, DLM performance will be improved obviously
when using SCTP procotol.
Signed-off-by: Gang He <ghe@suse.com>
Signed-off-by: Michal Kubecek <mkubecek@suse.cz>
Signed-off-by: David Teigland <teigland@redhat.com>
Changes since v1:
Added changes in these files:
drivers/infiniband/hw/usnic/usnic_transport.c
drivers/staging/lustre/lnet/lnet/lib-socket.c
drivers/target/iscsi/iscsi_target_login.c
drivers/vhost/net.c
fs/dlm/lowcomms.c
fs/ocfs2/cluster/tcp.c
security/tomoyo/network.c
Before:
All these functions either return a negative error indicator,
or store length of sockaddr into "int *socklen" parameter
and return zero on success.
"int *socklen" parameter is awkward. For example, if caller does not
care, it still needs to provide on-stack storage for the value
it does not need.
None of the many FOO_getname() functions of various protocols
ever used old value of *socklen. They always just overwrite it.
This change drops this parameter, and makes all these functions, on success,
return length of sockaddr. It's always >= 0 and can be differentiated
from an error.
Tests in callers are changed from "if (err)" to "if (err < 0)", where needed.
rpc_sockname() lost "int buflen" parameter, since its only use was
to be passed to kernel_getsockname() as &buflen and subsequently
not used in any way.
Userspace API is not changed.
text data bss dec hex filename
30108430 2633624 873672 33615726 200ef6e vmlinux.before.o
30108109 2633612 873672 33615393 200ee21 vmlinux.o
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
CC: David S. Miller <davem@davemloft.net>
CC: linux-kernel@vger.kernel.org
CC: netdev@vger.kernel.org
CC: linux-bluetooth@vger.kernel.org
CC: linux-decnet-user@lists.sourceforge.net
CC: linux-wireless@vger.kernel.org
CC: linux-rdma@vger.kernel.org
CC: linux-sctp@vger.kernel.org
CC: linux-nfs@vger.kernel.org
CC: linux-x25@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
The writequeue and writequeue_lock member of othercon was not initialized.
If lowcomms_state_change() is called from network layer, othercon->swork
may be scheduled. In this case, send_to_sock() will generate a NULL pointer
reference. We avoid this problem by correctly initializing writequeue and
writequeue_lock member of othercon.
Signed-off-by: Tadashi Miyauchi <miyauchi@toshiba-tops.co.jp>
Signed-off-by: Tsutomu Owa <tsutomu.owa@toshiba.co.jp>
Signed-off-by: David Teigland <teigland@redhat.com>
When an error occurs in kernel_recvmsg or kernel_sendpage and
close_connection is called and receive work is already scheduled,
receive work is canceled. In that case, the receive work will not
be scheduled forever after reconnection, because CF_READ_PENDING
flag is established.
Signed-off-by: Tadashi Miyauchi <miyauchi@toshiba-tops.co.jp>
Signed-off-by: Tsutomu Owa <tsutomu.owa@toshiba.co.jp>
Signed-off-by: David Teigland <teigland@redhat.com>
In the current implementation, we think that exclusion control between
processing to set the callback function to the connection structure and
processing to refer to the connection structure from the callback function
was not enough. We fix them.
Signed-off-by: Tadashi Miyauchi <miyauchi@toshiba-tops.co.jp>
Signed-off-by: Tsutomu Owa <tsutomu.owa@toshiba.co.jp>
Signed-off-by: David Teigland <teigland@redhat.com>
The sk member of the socket generated by sock_create_kern() is overwritten
by ops->accept(). So the previous sk will not be released.
We use kernel_accept() instead of sock_create_kern() and ops->accept().
Signed-off-by: Tadashi Miyauchi <miyauchi@toshiba-tops.co.jp>
Signed-off-by: Tsutomu Owa <tsutomu.owa@toshiba.co.jp>
Signed-off-by: David Teigland <teigland@redhat.com>
If reconnection fails while executing dlm_lowcomms_stop,
dlm_send will not stop.
Signed-off-by: Tadashi Miyauchi <miyauchi@toshiba-tops.co.jp>
Signed-off-by: Tsutomu Owa <tsutomu.owa@toshiba.co.jp>
Signed-off-by: David Teigland <teigland@redhat.com>
CF_WRITE_PENDING flag has been reanimated to make dlm_send stop properly
when running dlm_lowcomms_stop.
Signed-off-by: Tadashi Miyauchi <miyauchi@toshiba-tops.co.jp>
Signed-off-by: Tsutomu Owa <tsutomu.owa@toshiba.co.jp>
Signed-off-by: David Teigland <teigland@redhat.com>
If an error occurs in the sending / receiving process, if othercon
exists, sending / receiving processing using othercon may also result
in an error. We fix to pre-close othercon as well.
Signed-off-by: Tadashi Miyauchi <miyauchi@toshiba-tops.co.jp>
Signed-off-by: Tsutomu Owa <tsutomu.owa@toshiba.co.jp>
Signed-off-by: David Teigland <teigland@redhat.com>
In the current implementation, we think that exclusion control
for othercon in tcp_accept_from_sock() and sctp_accept_from_sock()
was not enough. We fix them.
Signed-off-by: Tadashi Miyauchi <miyauchi@toshiba-tops.co.jp>
Signed-off-by: Tsutomu Owa <tsutomu.owa@toshiba.co.jp>
Signed-off-by: David Teigland <teigland@redhat.com>
When kernel_sendpage(in send_to_sock) and kernel_recvmsg
(in receive_from_sock) return error, close_connection may works at the
same time. At that time, they may wait for each other by cancel_work_sync.
Signed-off-by: Tadashi Miyauchi <miayuchi@toshiba-tops.co.jp>
Signed-off-by: Tsutomu Owa <tsutomu.owa@toshiba.co.jp>
Signed-off-by: David Teigland <teigland@redhat.com>
save_cb argument is not used. We remove them.
Signed-off-by: Tadashi Miyauchi <miyauchi@toshiba-tops.co.jp>
Signed-off-by: Tsutomu Owa <tsutomu.owa@toshiba.co.jp>
Signed-off-by: David Teigland <teigland@redhat.com>
In a previous patch I noted that accept() often copies the struct
sock (sk) which overwrites the sock callbacks. However, in testing
we discovered that the dlm connection structures (con) are sometimes
deleted and recreated as connections come and go, and since they're
zeroed out by kmem_cache_zalloc, the saved callback pointers are
also initialized to zero. But with today's DLM code, the callbacks
are only saved when a socket is added.
During recovery testing, we discovered a common situation in which
the new con is initialized to zero, then a socket is added after
accept(). In this case, the sock's saved values are all NULL, but
the saved values are wiped out, due to accept(). Therefore, we
don't have a known good copy of the callbacks from which we can
restore.
Since the struct sock callbacks are always good after listen(),
this patch saves the known good values after listen(). These good
values are then used for subsequent restores.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Reviewed-by: Tadashi Miyauchi <miyauchi@toshiba-tops.co.jp>
Signed-off-by: David Teigland <teigland@redhat.com>
Before this patch, there was a flag in the con structure that was
used to determine whether or not a connect was needed. The bit was
set here and there, and cleared here and there, so it left some
race conditions: the bit was set, work was queued, then the worker
cleared the bit, allowing someone else to set it while the worker
ran. For the most part, this worked okay, but we got into trouble
if connections were lost and it needed to reconnect.
This patch eliminates the flag in favor of simply checking if we
actually have a sock pointer while protected by the mutex.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Reviewed-by: Tadashi Miyauchi <miyauchi@toshiba-tops.co.jp>
Signed-off-by: David Teigland <teigland@redhat.com>
Lockdep issues a circular dependency warning when AFS issues an operation
through AF_RXRPC from a context in which the VFS/VM holds the mmap_sem.
The theory lockdep comes up with is as follows:
(1) If the pagefault handler decides it needs to read pages from AFS, it
calls AFS with mmap_sem held and AFS begins an AF_RXRPC call, but
creating a call requires the socket lock:
mmap_sem must be taken before sk_lock-AF_RXRPC
(2) afs_open_socket() opens an AF_RXRPC socket and binds it. rxrpc_bind()
binds the underlying UDP socket whilst holding its socket lock.
inet_bind() takes its own socket lock:
sk_lock-AF_RXRPC must be taken before sk_lock-AF_INET
(3) Reading from a TCP socket into a userspace buffer might cause a fault
and thus cause the kernel to take the mmap_sem, but the TCP socket is
locked whilst doing this:
sk_lock-AF_INET must be taken before mmap_sem
However, lockdep's theory is wrong in this instance because it deals only
with lock classes and not individual locks. The AF_INET lock in (2) isn't
really equivalent to the AF_INET lock in (3) as the former deals with a
socket entirely internal to the kernel that never sees userspace. This is
a limitation in the design of lockdep.
Fix the general case by:
(1) Double up all the locking keys used in sockets so that one set are
used if the socket is created by userspace and the other set is used
if the socket is created by the kernel.
(2) Store the kern parameter passed to sk_alloc() in a variable in the
sock struct (sk_kern_sock). This informs sock_lock_init(),
sock_init_data() and sk_clone_lock() as to the lock keys to be used.
Note that the child created by sk_clone_lock() inherits the parent's
kern setting.
(3) Add a 'kern' parameter to ->accept() that is analogous to the one
passed in to ->create() that distinguishes whether kernel_accept() or
sys_accept4() was the caller and can be passed to sk_alloc().
Note that a lot of accept functions merely dequeue an already
allocated socket. I haven't touched these as the new socket already
exists before we get the parameter.
Note also that there are a couple of places where I've made the accepted
socket unconditionally kernel-based:
irda_accept()
rds_rcp_accept_one()
tcp_accept_from_sock()
because they follow a sock_create_kern() and accept off of that.
Whilst creating this, I noticed that lustre and ocfs don't create sockets
through sock_create_kern() and thus they aren't marked as for-kernel,
though they appear to be internal. I wonder if these should do that so
that they use the new set of lock keys.
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix to return a negative error code from the error handling
case instead of 0, as done elsewhere in this function.
Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Signed-off-by: David Teigland <teigland@redhat.com>
Alexander Aring
Christoph Hellwig
Arnd Bergmann
David Windsor
Thomas Gleixner
Deepa Dinamani
David Howells
Gang He
Denys Vlasenko
Al Viro
tsutomu.owa@toshiba.co.jp
Bob Peterson
Guoqing Jiang
Wei Yongjun