Linux RDMA and InfiniBand development
 help / color / mirror / Atom feed
* [PATCH rdma-rc] RDMA/siw: Don't send stale data for a queued zero length Read Request
@ 2026-11-22 19:11 Andrew Stellman
  2026-11-22 19:15 ` sashiko-bot
  0 siblings, 1 reply; 2+ messages in thread
From: Andrew Stellman @ 2026-11-22 19:11 UTC (permalink / raw)
  To: Bernard Metzler
  Cc: Jason Gunthorpe, Leon Romanovsky, linux-rdma, linux-kernel

siw_init_rresp() queues an inbound RDMA Read Request in the inbound read
queue (IRQ) when the TX work request is busy. For a zero length Read
Request the IRQ entry has num_sge 0, and siw_activate_tx() then leaves
wqe->sqe.sge[0] as the previous work request left it.
siw_qp_sq_proc_tx() sets num_sge to 1 for every Read Response, so the
Read Response is built from the previous work request's sge[0]. RFC
5040, section 5.2.1, requires a zero length Read Response.

A plain TCP peer that speaks MPA rev 1 sent two Read Requests in one
write(): 64 bytes from an MR with remote read access, then a zero
length Read Request. The second Read Response carried the same 64
bytes of the MR:

  Read Response, sink STag 0x11110000: 64 bytes "SIW016-REMOTE-RD..."
  Read Response, sink STag 0x22220000: 64 bytes "SIW016-REMOTE-RD..."

When the zero length Read Request was sent 500 ms after the first one,
its Read Response had no payload.

Copy sge[0] from the IRQ entry whatever its length, as the immediate
path in siw_init_rresp() does. With this change the second Read
Response has no payload in both cases.

The issue was found by Claude Opus 5.5 running Quality Playbook, an
LLM-driven code review tool:
https://github.com/andrewstellman/quality-playbook
The fix and this changelog were drafted with Claude Opus 5.5 and
reviewed by me.

Fixes: f29dd55b0236 ("rdma/siw: queue pair methods")
Signed-off-by: Andrew Stellman <astellman@stellman-greene.com>
---
Tested in QEMU on v7.4-rc3 with KASAN, with a plain TCP peer that speaks
MPA rev 1. A reproducer is available on request.

 drivers/infiniband/sw/siw/siw_qp.c | 17 +++++++++--------
 1 file changed, 9 insertions(+), 8 deletions(-)

diff --git a/drivers/infiniband/sw/siw/siw_qp.c b/drivers/infiniband/sw/siw/siw_qp.c
index bb780e3904a..f3eb1c85f3e 100644
--- a/drivers/infiniband/sw/siw/siw_qp.c
+++ b/drivers/infiniband/sw/siw/siw_qp.c
@@ -972,14 +972,15 @@ int siw_activate_tx(struct siw_qp *qp)
 	/* start READ RESPONSE */
 	wqe->sqe.opcode = SIW_OP_READ_RESPONSE;
 	wqe->sqe.flags = 0;
-	if (irqe->num_sge) {
-		wqe->sqe.num_sge = 1;
-		wqe->sqe.sge[0].length = irqe->sge[0].length;
-		wqe->sqe.sge[0].laddr = irqe->sge[0].laddr;
-		wqe->sqe.sge[0].lkey = irqe->sge[0].lkey;
-	} else {
-		wqe->sqe.num_sge = 0;
-	}
+	/*
+	 * Copy sge[0] also for a zero length Read Request. The TX path
+	 * sends sge[0] of every Read Response, and the previous work
+	 * request's sge[0] must not be sent in its place.
+	 */
+	wqe->sqe.num_sge = irqe->num_sge;
+	wqe->sqe.sge[0].length = irqe->sge[0].length;
+	wqe->sqe.sge[0].laddr = irqe->sge[0].laddr;
+	wqe->sqe.sge[0].lkey = irqe->sge[0].lkey;
 
 	/* Retain original RREQ's message sequence number for
 	 * potential error reporting cases.
-- 
2.43.0


^ permalink raw reply related	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2026-11-22 19:15 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-22 19:11 [PATCH rdma-rc] RDMA/siw: Don't send stale data for a queued zero length Read Request Andrew Stellman
2026-11-22 19:15 ` sashiko-bot

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox