[dm-devel] [PATCH 8/9] dm: Fix two race conditions related to stopping and starting queues

[Bart Van Assche]

On 09/01/2016 01:33 PM, Mike Snitzer wrote:
> I'm able to easily reproduce this 100% cpu usage using mptest's
> test_02_sdev_delete.
>
> 'dmsetup suspend --nolockfs --noflush mp' hangs, seems rooted in your
> use of blk_mq_freeze_queue():
>
> [  298.136930] dmsetup         D ffff880142cb3b70     0  9478   9414 0x00000080
> [  298.144831]  ffff880142cb3b70 ffff880142cb3b28 ffff880330d6cb00 ffff88032d0022f8
> [  298.153132]  ffff880142cb4000 ffff88032d0022f8 ffff88032b161800 0000000000000001
> [  298.161438]  0000000000000001 ffff880142cb3b88 ffffffff816c06e5 ffff88032d001aa0
> [  298.169740] Call Trace:
> [  298.172473]  [<ffffffff816c06e5>] schedule+0x35/0x80
> [  298.178019]  [<ffffffff8131b937>] blk_mq_freeze_queue_wait+0x57/0xc0
> [  298.185116]  [<ffffffff810c58c0>] ? prepare_to_wait_event+0xf0/0xf0
> [  298.192117]  [<ffffffff8131d92a>] blk_mq_freeze_queue+0x1a/0x20
> [  298.198734]  [<ffffffffa000e910>] dm_stop_queue+0x50/0xc0 [dm_mod]
> [  298.205644]  [<ffffffffa0001824>] __dm_suspend+0x134/0x1f0 [dm_mod]
> [  298.212649]  [<ffffffffa00035b8>] dm_suspend+0xb8/0xd0 [dm_mod]
> [  298.219270]  [<ffffffffa000882e>] dev_suspend+0x18e/0x240 [dm_mod]
> [  298.226175]  [<ffffffffa00086a0>] ? table_load+0x380/0x380 [dm_mod]
> [  298.233180]  [<ffffffffa0009027>] ctl_ioctl+0x1e7/0x4d0 [dm_mod]
> [  298.239890]  [<ffffffff81197f00>] ? lru_cache_add_active_or_unevictable+0x10/0xb0
> [  298.248253]  [<ffffffffa0009323>] dm_ctl_ioctl+0x13/0x20 [dm_mod]
> [  298.255049]  [<ffffffff81227937>] do_vfs_ioctl+0xa7/0x5d0
> [  298.261081]  [<ffffffff8112787f>] ? __audit_syscall_entry+0xaf/0x100
> [  298.268178]  [<ffffffff8100365d>] ? syscall_trace_enter+0x1dd/0x2c0
> [  298.275179]  [<ffffffff81227ed9>] SyS_ioctl+0x79/0x90
> [  298.280821]  [<ffffffff81003a47>] do_syscall_64+0x67/0x160
> [  298.286950]  [<ffffffff816c4921>] entry_SYSCALL64_slow_path+0x25/0x25
>

Hello Mike,

In your dm-4.9 branch I see that you call blk_mq_freeze_queue() while 
holding the block layer queue lock. Please don't do this. 
blk_mq_freeze_queue() can sleep and as you know calling a sleeping 
function while holding a spinlock is not allowed.

Bart.