Re: [PATCH] NFS: invalidate i_blocks after COMMIT to fix stale block count on NFSv4

[Trond Myklebust <trondmy@kernel.org>]

On Mon, 2026-06-22 at 14:00 +0800, Jingbo Xu wrote:
> NFSv4 COMMIT compound does not include GETATTR, and
> nfs4_commit_done_cb
> does not refresh inode attributes. Meanwhile, every WRITE marks
> NFS_INO_INVALID_BLOCKS via nfs_post_op_update_inode_force_wcc_locked.
> 
> After COMMIT, i_blocks remains stale until the next stat() triggers a
> full revalidation. In writeback-heavy workloads where COMMITs happen
> without intervening stat() calls, the cached block count can stay
> indefinitely wrong.
> 
> Mark NFS_INO_INVALID_BLOCKS on successful COMMIT completion so that
> the
> next nfs_getattr requesting STATX_BLOCKS will issue a GETATTR with
> SPACE_USED, fetching the correct value from the server.
> 
> This matches NFSv3 behavior where nfs3_commit_done already calls
> nfs_refresh_inode with the wcc_data post-op attributes.
> 
> Reproduce with xfstests generic/694 on NFSv4.0 loopback:
> 
>   Server:
>     mount /dev/vdc /data/test
>     mount /dev/vdd /data/scratch
>     exportfs -o insecure,rw,sync,no_root_squash,fsid=1
> 127.0.0.1:/data/test
>     exportfs -o insecure,rw,sync,no_root_squash,fsid=2
> 127.0.0.1:/data/scratch
> 
>   Client:
>     mount -t nfs -o vers=4.0 localhost:/data/test /mnt/test
>     mount -t nfs -o vers=4.0 localhost:/data/scratch /mnt/scratch
> 
>   local.config:
>     export TEST_FS_MOUNT_OPTS="-o vers=4.0"
>     export MOUNT_OPTIONS="-o vers=4.0"
>     export FSTYP=nfs
>     export TEST_DEV=localhost:/data/test
>     export SCRATCH_DEV=localhost:/data/scratch
>     export TEST_DIR=/mnt/test
>     export SCRATCH_MNT=/mnt/scratch
> 
> This fixes xfstests generic/694.
> 
> Assisted-by: Qoder:Qwen3.7-Max
> Signed-off-by: Jingbo Xu <jefflexu@linux.alibaba.com>
> ---
>  fs/nfs/write.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/fs/nfs/write.c b/fs/nfs/write.c
> index d7c399763ad9..88c5c9f7434c 100644
> --- a/fs/nfs/write.c
> +++ b/fs/nfs/write.c
> @@ -1851,6 +1851,8 @@ static void nfs_commit_release_pages(struct
> nfs_commit_data *data)
>  		/* Latency breaker */
>  		cond_resched();
>  	}
> +	if (status >= 0)
> +		nfs_set_cache_invalid(data->inode,
> NFS_INO_INVALID_BLOCKS);
>  
>  	nfs_init_cinfo(&cinfo, data->inode, data->dreq);
>  	nfs_commit_end(cinfo.mds);

That sounds like an XFS bug, not an NFS bug. COMMIT isn't changing the
data contents of the file: it is just ensuring that the existing data
is persisted onto disk.

-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trondmy@kernel.org, trond.myklebust@hammerspace.com