Re: NFSD: Refuse a direct READ whose aligned range exceeds the reply pages

[Mike Snitzer <snitzer@kernel.org>]

On Wed, Nov 11, 2026 at 10:20:50AM -0500, Chuck Lever wrote:
> nfsd_direct_read() expands the READ's byte range to the file's direct
> I/O alignment, then maps reply pages until the range is covered or
> rq_page_end is reached. A block device whose logical block size
> exceeds PAGE_SIZE reports an alignment larger than PAGE_SIZE. With
> 16KB blocks and 4KB pages, a maximum-size READ at an offset that is
> not a multiple of 16KB needs one more page than the svc_rqst has.
> The loop stops at rq_page_end and passes a truncated iterator to the
> filesystem. iomap rejects the unaligned length with -EINVAL, and NFSD
> returns NFSERR_SERVERFAULT for a READ that a buffered read would have
> satisfied.
> 
> Check the aligned range against the reply pages in
> nfsd_direct_read_ok(), and fall back to a buffered read when the
> range does not fit.
> 
> Reported-by: sashiko-bot <sashiko-bot@kernel.org>
> Closes: https://sashiko.dev/#/patchset/20261109-xdr-page-base-v1-0-a4f4411d760d@kernel.org?part=11
> Fixes: d686e64e931c ("NFSD: Implement NFSD_IO_DIRECT for NFS READ")
> Signed-off-by: Chuck Lever <cel@kernel.org>

Reviewed-by: Mike Snitzer <snitzer@kernel.org>