* [PATCH] SUNRPC: Honor page_base beyond the first page in xdr_buf_to_bvec()
@ 2026-11-06 15:01 Chuck Lever
2026-11-06 19:13 ` Chuck Lever
0 siblings, 1 reply; 2+ messages in thread
From: Chuck Lever @ 2026-11-06 15:01 UTC (permalink / raw)
To: NeilBrown, Jeff Layton, Olga Kornievskaia, Dai Ngo, Tom Talpey
Cc: linux-nfs, stable
xdr_buf_to_bvec() takes offset_in_page(page_base) but starts its
walk at pages[0], so the page index that page_base carries above
PAGE_SIZE is dropped. nfsd_direct_read() sets rq_res.page_base to
the pad below the file's DIO offset alignment. When that alignment
exceeds PAGE_SIZE, as with 8K to 64K logical blocks under xfs or
ext4 on a 4K-page NFSD, a direct READ whose pad reaches a full page
sends the page before the one that holds the data. The defect was
found by inspection, not reproduced.
Advance the page pointer by page_base >> PAGE_SHIFT before the walk.
Commit 2eb2b9358181 ("SUNRPC: Convert svc_tcp_sendmsg to use
bio_vecs directly") introduced the dropped shift, but no caller
produced a page_base of a page or more until direct READ.
This fix depends at run time on nfsd4_encode_operation() deriving
rq_next_page from page_base as well as page_len. Otherwise the READ
encoder accounts for one page fewer than xdr_buf_to_bvec() now
transmits, and the last page sent sits outside the release range
while TCP still references the page.
Fixes: d686e64e931c ("NFSD: Implement NFSD_IO_DIRECT for NFS READ")
Cc: <stable@vger.kernel.org> # v6.19+
Signed-off-by: Chuck Lever <cel@kernel.org>
---
net/sunrpc/xdr.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/net/sunrpc/xdr.c b/net/sunrpc/xdr.c
index 7353109fdbae..7eef50634175 100644
--- a/net/sunrpc/xdr.c
+++ b/net/sunrpc/xdr.c
@@ -161,7 +161,7 @@ int xdr_buf_to_bvec(struct bio_vec *bvec, unsigned int bvec_size,
if (xdr->page_len) {
unsigned int offset, len, remaining;
- struct page **pages = xdr->pages;
+ struct page **pages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
offset = offset_in_page(xdr->page_base);
remaining = xdr->page_len;
--
2.56.0
^ permalink raw reply related [flat|nested] 2+ messages in thread* Re: [PATCH] SUNRPC: Honor page_base beyond the first page in xdr_buf_to_bvec()
2026-11-06 15:01 [PATCH] SUNRPC: Honor page_base beyond the first page in xdr_buf_to_bvec() Chuck Lever
@ 2026-11-06 19:13 ` Chuck Lever
0 siblings, 0 replies; 2+ messages in thread
From: Chuck Lever @ 2026-11-06 19:13 UTC (permalink / raw)
To: NeilBrown, Jeff Layton, Olga Kornievskaia, Dai Ngo, Tom Talpey
Cc: linux-nfs, stable
On Fri, Nov 6, 2026, at 10:01 AM, Chuck Lever wrote:
> xdr_buf_to_bvec() takes offset_in_page(page_base) but starts its
> walk at pages[0], so the page index that page_base carries above
> PAGE_SIZE is dropped. nfsd_direct_read() sets rq_res.page_base to
> the pad below the file's DIO offset alignment. When that alignment
> exceeds PAGE_SIZE, as with 8K to 64K logical blocks under xfs or
> ext4 on a 4K-page NFSD, a direct READ whose pad reaches a full page
> sends the page before the one that holds the data. The defect was
> found by inspection, not reproduced.
>
> Advance the page pointer by page_base >> PAGE_SHIFT before the walk.
>
> Commit 2eb2b9358181 ("SUNRPC: Convert svc_tcp_sendmsg to use
> bio_vecs directly") introduced the dropped shift, but no caller
> produced a page_base of a page or more until direct READ.
>
> This fix depends at run time on nfsd4_encode_operation() deriving
> rq_next_page from page_base as well as page_len. Otherwise the READ
> encoder accounts for one page fewer than xdr_buf_to_bvec() now
> transmits, and the last page sent sits outside the release range
> while TCP still references the page.
>
> Fixes: d686e64e931c ("NFSD: Implement NFSD_IO_DIRECT for NFS READ")
> Cc: <stable@vger.kernel.org> # v6.19+
> Signed-off-by: Chuck Lever <cel@kernel.org>
> ---
> net/sunrpc/xdr.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/net/sunrpc/xdr.c b/net/sunrpc/xdr.c
> index 7353109fdbae..7eef50634175 100644
> --- a/net/sunrpc/xdr.c
> +++ b/net/sunrpc/xdr.c
> @@ -161,7 +161,7 @@ int xdr_buf_to_bvec(struct bio_vec *bvec, unsigned
> int bvec_size,
>
> if (xdr->page_len) {
> unsigned int offset, len, remaining;
> - struct page **pages = xdr->pages;
> + struct page **pages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
>
> offset = offset_in_page(xdr->page_base);
> remaining = xdr->page_len;
> --
> 2.56.0
To address the sashiko.dev finding for this patch, I've folded this fix into
"nfsd: account for page_base when advancing rq_next_page" which is currently
in nfsd-next.
--
Chuck Lever (Come to NFS bake-a-thon! https://nfsv4bat.org)
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2026-11-06 19:14 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-06 15:01 [PATCH] SUNRPC: Honor page_base beyond the first page in xdr_buf_to_bvec() Chuck Lever
2026-11-06 19:13 ` Chuck Lever
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox