pnfs: fix BUG in filelayout_recover_commit_reqs
authorWeston Andros Adamson <dros@primarydata.com>
Tue, 21 Jan 2014 20:21:33 +0000 (15:21 -0500)
committerTrond Myklebust <trond.myklebust@primarydata.com>
Tue, 21 Jan 2014 21:12:18 +0000 (14:12 -0700)
cond_resched_lock(cinfo->lock) is called everywhere else while holding
the cinfo->lock spinlock.  Not holding this lock while calling
transfer_commit_list in filelayout_recover_commit_reqs causes the BUG
below.

It's true that we can't hold this lock while calling pnfs_put_lseg,
because that might try to lock the inode lock - which might be the
same lock as cinfo->lock.

To reproduce, mount a 2 DS pynfs server and run an O_DIRECT command
that crosses a stripe boundary and is not page aligned, such as:

 dd if=/dev/zero of=/mnt/f bs=17000 count=1 oflag=direct

BUG: sleeping function called from invalid context at linux/fs/nfs/nfs4filelayout.c:1161
in_atomic(): 0, irqs_disabled(): 0, pid: 27, name: kworker/0:1
2 locks held by kworker/0:1/27:
 #0:  (events){.+.+.+}, at: [<ffffffff810501d7>] process_one_work+0x175/0x3a5
 #1:  ((&dreq->work)){+.+...}, at: [<ffffffff810501d7>] process_one_work+0x175/0x3a5
CPU: 0 PID: 27 Comm: kworker/0:1 Not tainted 3.13.0-rc3-branch-dros_testing+ #21
Hardware name: VMware, Inc. VMware Virtual Platform/440BX Desktop Reference Platform, BIOS 6.00 07/31/2013
Workqueue: events nfs_direct_write_schedule_work [nfs]
 0000000000000000 ffff88007a39bbb8 ffffffff81491256 ffff88007b87a130  ffff88007a39bbd8 ffffffff8105f103 ffff880079614000 ffff880079617d40  ffff88007a39bc20 ffffffffa011603e ffff880078988b98 0000000000000000
Call Trace:
 [<ffffffff81491256>] dump_stack+0x4d/0x66
 [<ffffffff8105f103>] __might_sleep+0x100/0x105
 [<ffffffffa011603e>] transfer_commit_list+0x94/0xf1 [nfs_layout_nfsv41_files]
 [<ffffffffa01160d6>] filelayout_recover_commit_reqs+0x3b/0x68 [nfs_layout_nfsv41_files]
 [<ffffffffa00ba53a>] nfs_direct_write_reschedule+0x9f/0x1d6 [nfs]
 [<ffffffff810705df>] ? mark_lock+0x1df/0x224
 [<ffffffff8106e617>] ? trace_hardirqs_off_caller+0x37/0xa4
 [<ffffffff8106e691>] ? trace_hardirqs_off+0xd/0xf
 [<ffffffffa00ba8f8>] nfs_direct_write_schedule_work+0x9d/0xb7 [nfs]
 [<ffffffff810501d7>] ? process_one_work+0x175/0x3a5
 [<ffffffff81050258>] process_one_work+0x1f6/0x3a5
 [<ffffffff810501d7>] ? process_one_work+0x175/0x3a5
 [<ffffffff8105187e>] worker_thread+0x149/0x1f5
 [<ffffffff81051735>] ? rescuer_thread+0x28d/0x28d
 [<ffffffff81056d74>] kthread+0xd2/0xda
 [<ffffffff81056ca2>] ? __kthread_parkme+0x61/0x61
 [<ffffffff8149e66c>] ret_from_fork+0x7c/0xb0
 [<ffffffff81056ca2>] ? __kthread_parkme+0x61/0x61

Signed-off-by: Weston Andros Adamson <dros@primarydata.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
fs/nfs/nfs4filelayout.c

index 0a93e798df3a854011bbdd258c9e910d3fd17de3..03fd8be8c0c50b3455ee94ae115fb77fa3d5e4fe 100644 (file)
@@ -1216,17 +1216,17 @@ static void filelayout_recover_commit_reqs(struct list_head *dst,
        struct pnfs_commit_bucket *b;
        int i;
 
-       /* NOTE cinfo->lock is NOT held, relying on fact that this is
-        * only called on single thread per dreq.
-        * Can't take the lock because need to do pnfs_put_lseg
-        */
+       spin_lock(cinfo->lock);
        for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
                if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
+                       spin_unlock(cinfo->lock);
                        pnfs_put_lseg(b->wlseg);
                        b->wlseg = NULL;
+                       spin_lock(cinfo->lock);
                }
        }
        cinfo->ds->nwritten = 0;
+       spin_unlock(cinfo->lock);
 }
 
 static unsigned int