Re: [PATCH] ceph: revalidate ki_pos for O_APPEND writes after cap acquisition

On Fri, 2026-07-17 at 17:33 +0800, Xiubo Li via B4 Relay wrote:
> From: Xiubo Li <[email protected]>
> 
> For O_APPEND writes, ki_pos is set to the current EOF via
> generic_write_checks() after fetching i_size from the MDS.  However,
> ceph_get_caps() may need to wait for Fwx exclusive caps if the write
> extends the file (endoff > i_max_size).  While waiting for Fwx, the
> previous Fwx holder (another client) may have already extended the
> file.  When the MDS grants us Fwx, the cap grant message updates the
> local i_size, but ki_pos remains at the old EOF, causing the append
> write to land at a stale offset and overwrite data from the other
> client.
> 
> Fix by re-reading i_size_read(inode) after ceph_get_caps() returns.
> At this point we hold Fwx exclusive caps, no other client can modify
> the file, and i_size reflects the true EOF from the MDS cap grant.
> No extra MDS round-trip is needed.  Only adjust ki_pos when the EOF
> has actually changed.
> 
> Link: https://tracker.ceph.com/issues/7333
> Signed-off-by: Xiubo Li <[email protected]>
> ---
> For O_APPEND writes, ki_pos is set to EOF via generic_write_checks()
> after fetching i_size from the MDS via ceph_do_getattr().  However,
> ceph_get_caps() may wait for Fwx exclusive caps (because endoff >
> i_max_size).  During that wait, the previous Fwx holder may have
> already extended the file.  When the MDS grants us Fwx, the local
> i_size is updated via the cap grant message, but ki_pos is still
> the old EOF, resulting in the append write overwriting data.
> 
> Fix by revalidating ki_pos against i_size_read(inode) after
> ceph_get_caps() returns.  We now hold Fwx — no other client can
> change the file, so i_size is stable and correct.
> 
> Fixes: 8e4473bb50a1 ("ceph: do not execute direct write in parallel
> if O_APPEND is specified")
> Link: https://tracker.ceph.com/issues/7333

The patch has two --- separators. The Fixes and Link will be finally
dropped.

> ---
>  fs/ceph/file.c | 26 ++++++++++++++++++++++++++
>  1 file changed, 26 insertions(+)
> 
> diff --git a/fs/ceph/file.c b/fs/ceph/file.c
> index 9d89d7fc1095..b35c9aee48dd 100644
> --- a/fs/ceph/file.c
> +++ b/fs/ceph/file.c
> @@ -2491,6 +2491,32 @@ static ssize_t ceph_write_iter(struct kiocb
> *iocb, struct iov_iter *from)
>  	if (err < 0)
>  		goto out;

ceph_get_caps() validated caps against the old endoff = pos + count.
After adjusting pos forward, the real write range is [cur_eof, cur_eof
+ count), which isn't re-validated against i_max_size. Could it be a
potential issue?

Thanks,
Slava.

>  
> +	/*
> +	 * For O_APPEND writes we may have waited for Fwx exclusive
> caps
> +	 * while the previous Fwx holder (another client) extended
> the
> +	 * file.  i_size has been updated via the cap grant message
> from
> +	 * the MDS, but ki_pos is still the old EOF.  Re-read i_size
> here
> +	 * (no extra MDS round-trip needed) and adjust ki_pos to the
> true
> +	 * EOF.  Since we hold Fwx, no other client can change the
> file.
> +	 */
> +	if (iocb->ki_flags & IOCB_APPEND) {
> +		loff_t cur_eof = i_size_read(inode);
> +
> +		if (cur_eof != pos) {
> +			doutc(cl,
> +			      "%p %llx.%llx O_APPEND: pos adjusted
> %lld -> %lld\n",
> +			      inode, ceph_vinop(inode), pos,
> cur_eof);
> +			iocb->ki_pos = cur_eof;
> +			pos = cur_eof;
> +			if (pos >= limit) {
> +				err = -EFBIG;
> +				goto out_caps;
> +			}
> +			iov_iter_truncate(from, limit - pos);
> +			count = iov_iter_count(from);
> +		}
> +	}
> +
>  	err = file_update_time(file);
>  	if (err)
>  		goto out_caps;
> 
> ---
> base-commit: df47791ab9f12cd9144aca477ddae195f68072a5
> change-id: 20260717-ceph-append-fix-9820e3b1a78c
> 
> Best regards,
> --  
> Xiubo Li <[email protected]>
>