On Tue, Jul 09 2019 at 8:17pm -0400,
Junxiao Bi <junxiao.bi@xxxxxxxxxx> wrote:
> When thin-volume was built on loop device, if available memory is low,
> the following deadlock can be triggered.
>
> One process P1 was allocating memory with GFP_FS flag, direct alloc fail,
> memory reclaim invoked memory shrinker in dm_bufio, dm_bufio_shrink_scan()
> run, mutex dm_bufio_client->lock was acquired, then P1 wait for dm_buffer
> io done in __try_evict_buffer->()__try_evict_buffer().
>
> But this io may never done as it was issued to the underlying loop device
> who forward it using fs direct-io, there some memory allocation were using
> GFP_FS(like do_blockdev_direct_IO()), if direct alloc fail, memory reclaim
> will invoke memory shrinker in dm_bufio, where dm_bufio_shrink_scan()
> will be invoked, since the mutex was already hold by P1, the loop thread
> will hung, io will never done. ABBA deadlock.
>
> Signed-off-by: Junxiao Bi <junxiao.bi@xxxxxxxxxx>
> ---
> Changes in v2:
> - refine the commit log
>
> drivers/md/dm-bufio.c | 4 +---
> 1 file changed, 1 insertion(+), 3 deletions(-)
>
> diff --git a/drivers/md/dm-bufio.c b/drivers/md/dm-bufio.c
> index 2a48ea3f1b30..b6b5acc92ca2 100644
> --- a/drivers/md/dm-bufio.c
> +++ b/drivers/md/dm-bufio.c
> @@ -1599,9 +1599,7 @@ dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
> unsigned long freed;
>
> c = container_of(shrink, struct dm_bufio_client, shrinker);
> - if (sc->gfp_mask & __GFP_FS)
> - dm_bufio_lock(c);
> - else if (!dm_bufio_trylock(c))
> + if (!dm_bufio_trylock(c))
> return SHRINK_STOP;
>
> freed = __scan(c, sc->nr_to_scan, sc->gfp_mask);
> --
> 2.17.1
>
This needs more careful review and understanding.
I'll commit to getting that done (hopefully with Mikulas' assistance)
during the 5.3-rcX cycle. But I'm not ready to stage this change yet.
Revisiting dm-bufio on loop is needed. Commit 9d28eb12447ee ("dm bufio:
change __GFP_IO to __GFP_FS in shrinker callbacks") was meant to address
deadlocks reported whern running on loop. And __try_evict_buffer() has
a check for GFP_NOFS ("!(gfp & __GFP_FS"); but that is only relevant to
__scan() callers and dm_bufio_shrink_scan() is looking to take the lock
before __scan() is called. So it does seem like we have issues in the
bufio code in general. Needs a proper audit though.
Mike
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
