On 13.12.18 г. 23:17 ч., fdmanana@xxxxxxxxxx wrote:
> From: Filipe Manana <fdmanana@xxxxxxxx>
>
> Several places allocate a device while holding the device list mutex. This
> can result in a deadlock if reclaim happens because the device, and its
> flush bio, are allocated using GFP_KERNEL mode (by __alloc_device() which
> is used by btrfs_alloc_device()). A transaction commit, which reclaim can
> trigger, needs to lock the device list mutex in its critical section, done
> at btrfs_update_commit_device_size().
>
> Some of these places are device_list_add(), which ends up being called
> through the device scan ioctl, and btrfs_close_one_device(), which ends up
> being called through the device remove ioctl.
>
> Since all the places that add elements to the list of resized devices (the
> device grow and shrink functions) only lock the chunk mutex before adding
> a device to the list, drop the need to acquire the device list mutex from
> btrfs_update_commit_device_size(), which is the only other place that uses
> this list and it already locks the chunk mutex.
>
> Fixes: 78f2c9e6dbb14 ("btrfs: device add and remove: use GFP_KERNEL")
> Fixes: e0ae999414238 ("btrfs: preallocate device flush bio")
> Signed-off-by: Filipe Manana <fdmanana@xxxxxxxx>
Reviewed-by: Nikolay Borisov <nborisov@xxxxxxxx>
