Re: [PATCH] btrfs: set the block size before starting a transaction

[Qu Wenruo]

在 2026/10/28 22:44, Qing Zhang 写道:
> From: Qing Zhang <diasyzhang@xxxxxxxxxxx>
>
> When adding a device, btrfs_init_new_device() calls set_blocksize()
> with a transaction handle held. If the block size changes, this takes
> the bdev inode's i_rwsem and triggers a lockdep dependency cycle:
>
>    bdev i_rwsem -> mmap_lock -> trans_num_writers
>                -> trans_num_extwriters -> bdev i_rwsem
>
> The buffered read path nests mmap_lock under the bdev i_rwsem, while
> mmapping a btrfs file can join a transaction through atime updates.
> Since commit c0e473a0d226 ("block: fix race between set_blocksize and
> read paths"), set_blocksize() takes the bdev inode's i_rwsem when
> changing the block size.
>
> Move set_blocksize() before btrfs_start_transaction() to break the
> cycle. Also check its return value to reject devices whose logical
> block size exceeds BTRFS_BDEV_BLOCKSIZE.
>
> Fixes: c0e473a0d226 ("block: fix race between set_blocksize and read paths")
> Reported-by: syzbot+5c01398781fb51df783a@xxxxxxxxxxxxxxxxxxxxxxxxx
> Closes: https://syzkaller.appspot.com/bug?extid=5c01398781fb51df783a

There is already a better solution by not calling the stupid 
ste_blocksize() at all:

https://lore.kernel.org/linux-btrfs/3d98566e8aa2cd99a23ce89665f79281c614f05e.1792885061.git.wqu@xxxxxxxx/


> Signed-off-by: Qing Zhang <diasyzhang@xxxxxxxxxxx>
> ---
> #syz test: git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
>   fs/btrfs/volumes.c | 5 ++++-
>   1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
> index d3e6a9b429dd..a7f0607a5b86 100644
> --- a/fs/btrfs/volumes.c
> +++ b/fs/btrfs/volumes.c
> @@ -3019,6 +3019,10 @@ int btrfs_init_new_device(struct btrfs_fs_info *fs_info, const char *device_path
>   	if (ret)
>   		goto error_free_device;
>   
> +	ret = set_blocksize(device->bdev_file, BTRFS_BDEV_BLOCKSIZE);
> +	if (ret)
> +		goto error_free_zone;
> +
>   	trans = btrfs_start_transaction(root, 0);
>   	if (IS_ERR(trans)) {
>   		ret = PTR_ERR(trans);
> @@ -3043,7 +3047,6 @@ int btrfs_init_new_device(struct btrfs_fs_info *fs_info, const char *device_path
>   	 */
>   	atomic_inc(&device->dev_stats_ccnt);
>   	device->dev_stats_valid = 1;
> -	set_blocksize(device->bdev_file, BTRFS_BDEV_BLOCKSIZE);
>   
>   	if (seeding_dev) {
>   		/* GFP_KERNEL allocation must not be under device_list_mutex */