[lkml]   [2014]   [Mar]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Patch in this message
Subject[RFC PATCH] cifs: Fix possible deadlock with cifs and work queues
We just had a customer report a deadlock in their 3.8-rt kernel.
Looking into this, it is very possible to have the same deadlock in
mainline in Linus's tree as it stands today.

It is much easier to deadlock in the -rt kernel because reader locks
are serialized, where a down_read() can block another down_read(). But
because rwsems are fair locks, if a writer is waiting, a new reader
will then block. This means that if it is possible for a reader to
deadlock another reader, this can happen if a write comes along and
blocks on a current reader. That will prevent another reader from
running, and if that new reader requires to wake up a reader that owns
the lock, you have your deadlock.

Here's the situation with CIFS and workqueues:

The cifs system has several workqueues used in file.c and other places.
One of them is used for completion of a read and to release the
page_lock which wakes up the reader. There are several other workqueues
that do various other tasks.

A holder of the reader lock can sleep on a page_lock() and expect the
reader workqueue to wake it up (page_unlock()). The reader workqueue
takes no locks so this does not seem to be a problem (but it is).

The other workqueues can take the rwsem for read or for write. But our
issue that we tripped over was that it grabs it for read (remember in
-rt readers are serialized). But this can also happen if a separate
writer is waiting on the lock as that would cause a reader to block on
another reader too.

All the workqueue callbacks are executed on the same workqueue:

queue_work(cifsiod_wq, &rdata->work);
queue_work(cifsiod_wq, &cfile->oplock_break);

Now if the reader workqueue callback is queued after one of these
workqueues that can take the rwsem, we can hit a deadlock. The
workqueue code looks to be able to prevent deadlocks of these kinds,
but I do not totally understand the workqueue scheduled work structure
and perhaps if the kworker thread structure blocks hard it wont move
works around.

Here's what we see:

rdata->work is scheduled after cfile->oplock_break

---- ----





[ note, cifs_oplock_break() can
also call cifs_push_locks which takes
the lock with down_write() ]

But we noticed that the rdata->work was queued to run under the same
workqueue task and this work is to wake up the owner of the semaphore.
But because the workqueue task is blocked waiting on that lock, it will
never wake it up.

My question to Tejun is, if we create another workqueue, to add the
rdata->work to, would that prevent the above problem? Or what other
fixes can we do?

This is only compiled tested, we have not given it to our customer yet.

Signed-off-by: Steven Rostedt <>

diff --git a/fs/cifs/cifsfs.c b/fs/cifs/cifsfs.c
index 849f613..6656058 100644
--- a/fs/cifs/cifsfs.c
+++ b/fs/cifs/cifsfs.c
@@ -86,6 +86,7 @@ extern mempool_t *cifs_req_poolp;
extern mempool_t *cifs_mid_poolp;

struct workqueue_struct *cifsiod_wq;
+struct workqueue_struct *cifsiord_wq;

__u8 cifs_client_guid[SMB2_CLIENT_GUID_SIZE];
@@ -1199,9 +1200,15 @@ init_cifs(void)
goto out_clean_proc;

+ cifsiord_wq = alloc_workqueue("cifsiord", WQ_FREEZABLE|WQ_MEM_RECLAIM, 0);
+ if (!cifsiord_wq) {
+ rc = -ENOMEM;
+ goto out_destroy_wq;
+ }
rc = cifs_fscache_register();
if (rc)
- goto out_destroy_wq;
+ goto out_destroy_rwq;

rc = cifs_init_inodecache();
if (rc)
@@ -1249,6 +1256,8 @@ out_destroy_inodecache:
+ destroy_workqueue(cifsiord_wq);
@@ -1273,6 +1282,7 @@ exit_cifs(void)
+ destroy_workqueue(cifsiord_wq);

diff --git a/fs/cifs/cifsglob.h b/fs/cifs/cifsglob.h
index c0f3718..75d1941 100644
--- a/fs/cifs/cifsglob.h
+++ b/fs/cifs/cifsglob.h
@@ -1561,6 +1561,7 @@ void cifs_oplock_break(struct work_struct *work);

extern const struct slow_work_ops cifs_oplock_break_ops;
extern struct workqueue_struct *cifsiod_wq;
+extern struct workqueue_struct *cifsiord_wq;

extern mempool_t *cifs_mid_poolp;

diff --git a/fs/cifs/cifssmb.c b/fs/cifs/cifssmb.c
index f3264bd..ca04a2e 100644
--- a/fs/cifs/cifssmb.c
+++ b/fs/cifs/cifssmb.c
@@ -1571,7 +1571,7 @@ cifs_readv_callback(struct mid_q_entry *mid)
rdata->result = -EIO;

- queue_work(cifsiod_wq, &rdata->work);
+ queue_work(cifsiord_wq, &rdata->work);
add_credits(server, 1, 0);
diff --git a/fs/cifs/smb2pdu.c b/fs/cifs/smb2pdu.c
index 8603447..b74bf61 100644
--- a/fs/cifs/smb2pdu.c
+++ b/fs/cifs/smb2pdu.c
@@ -1742,7 +1742,7 @@ smb2_readv_callback(struct mid_q_entry *mid)
if (rdata->result)
cifs_stats_fail_inc(tcon, SMB2_READ_HE);

- queue_work(cifsiod_wq, &rdata->work);
+ queue_work(cifsiord_wq, &rdata->work);
add_credits(server, credits_received, 0);

 \ /
  Last update: 2014-03-19 20:41    [W:0.126 / U:1.316 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site