[dm-devel] [PATCH] direct-io: use GFP_NOIO to avoid deadlock

[Matthew Wilcox]

On Thu, Aug 08, 2019 at 05:50:10AM -0400, Mikulas Patocka wrote:
> A deadlock with this stacktrace was observed.
> 
> The obvious problem here is that in the call chain 
> xfs_vm_direct_IO->__blockdev_direct_IO->do_blockdev_direct_IO->kmem_cache_alloc 
> we do a GFP_KERNEL allocation while we are in a filesystem driver and in a 
> block device driver.

But that's not the problem.  The problem is the loop driver calls into the
filesystem without calling memalloc_noio_save() / memalloc_noio_restore().
There are dozens of places in XFS which use GFP_KERNEL allocations and
all can trigger this same problem if called from the loop driver.

>   #14 [ffff88272f5af880] kmem_cache_alloc at ffffffff811f484b
>   #15 [ffff88272f5af8d0] do_blockdev_direct_IO at ffffffff812535b3
>   #16 [ffff88272f5afb00] __blockdev_direct_IO at ffffffff81255dc3
>   #17 [ffff88272f5afb30] xfs_vm_direct_IO at ffffffffa01fe3fc [xfs]
>   #18 [ffff88272f5afb90] generic_file_read_iter at ffffffff81198994
>   #19 [ffff88272f5afc50] __dta_xfs_file_read_iter_2398 at ffffffffa020c970 [xfs]
>   #20 [ffff88272f5afcc0] lo_rw_aio at ffffffffa0377042 [loop]
>   #21 [ffff88272f5afd70] loop_queue_work at ffffffffa0377c3b [loop]
>   #22 [ffff88272f5afe60] kthread_worker_fn at ffffffff810a8a0c
>   #23 [ffff88272f5afec0] kthread at ffffffff810a8428
>   #24 [ffff88272f5aff50] ret_from_fork at ffffffff81745242