Re: [PATCH v3 0/5] btrfs: fix io_uring encoded IO cleanup

[David Sterba <dsterba@suse.cz>]

On Wed, Aug 19, 2026 at 10:54:31AM +0800, Yang Xiuwei wrote:
> Hi,
> 
> Fixes for the io_uring ENCODED_READ / ENCODED_WRITE cleanup paths.
> 
> 1. Always return -EIOCBQUEUED after endio so the caller and
>    btrfs_uring_read_finished() do not both clean up.
> 
>    Link: https://lore.kernel.org/linux-btrfs/20260630091609.3414-1-samsun1006219@gmail.com/
> 
> 2. btrfs_uring_read_extent() failure skipped kfree(data->iov).
> 
> 3. Unlock inode/extent in the caller on read-extent failure.
> 
> 4. Do not stash encoded data in the cmd PDU across -EAGAIN.
> 
> 5. Drop the unused REISSUE restore.
> 
> v2 also exported __import_iovec() and passed IO_URING_F_COMPAT.
> That is not a bug with the current io_uring thread setup:
> create_io_thread() copies TS_COMPAT, so import_iovec() already
> sees compat iovecs. Passing the flag explicitly would still be a
> cleanup, but this series does not take it.
> 
> Testing:
> - xfstests btrfs/333 passed.
> - kmemleak after dropping the ring on NONBLOCK encoded write/read:
> 
>   unreferenced object (size 512):
>     comm "btrfs_uring_eag"
>     __kmalloc_cache_noprof+0x38f/0x4a0
>     btrfs_uring_encoded_write+0x78d/0x1790 [btrfs]
>     btrfs_uring_cmd+0x142/0x1d0 [btrfs]
>     io_uring_cmd+0x296/0x5e0
> 
>   unreferenced object (size 512):
>     comm "btrfs_uring_eag"
>     __kmalloc_cache_noprof+0x38f/0x4a0
>     btrfs_uring_encoded_read+0x89a/0x12b0 [btrfs]
>     btrfs_uring_cmd+0x173/0x1d0 [btrfs]
>     io_uring_cmd+0x296/0x5e0
> 
>   After 4/5 the same repro is clean.
> 
> Changes in v3:
> - Drop the __import_iovec / IO_URING_F_COMPAT patches. 
> - 3/5: drop a leftover goto out_fail from v2.

Added to for-next, thanks.

IORING_URING_CMD_REISSUE is now only used in block layer code.