Re: [PATCH rdma-next] RDMA/rxe: Require a responder resource for FLUSH
From: Leon Romanovsky
Date: Sun Nov 22 2026 - 10:11:16 EST
On Thu, Nov 19, 2026 at 10:13:32AM -0500, Andrew Stellman wrote:
> process_flush() keeps its state in a responder resource from
> rxe_prepare_res(), as READ and ATOMIC requests do. check_resource()
> refuses those requests when max_dest_rd_atomic is 0, but it does not
> check FLUSH. With max_dest_rd_atomic 0, qp->resp.resources is
> ZERO_SIZE_PTR, and rxe_prepare_res() writes through it in
> free_rd_atomic_resource().
>
> A FLUSH sent to an RC QP with max_dest_rd_atomic 0 oopses the
> responder:
>
> KASAN: null-ptr-deref in range [0x0000000000000010-0x0000000000000017]
> Workqueue: rxe_wq do_work
> RIP: 0010:free_rd_atomic_resource+0x15/0x50
> Call Trace:
> rxe_receiver+0x5f6e/0x9a20
> do_work+0x144/0x470
>
> The responder task dies, and destroying the QP later warns on a timeout
> in __rxe_cleanup() and then hangs.
Please resend this patch with untrimmed dmesg stack trace.
Thanks
>
> Add RXE_FLUSH_MASK to the check, so a FLUSH gets the same Invalid
> Request NAK a READ would, and the requester sees IB_WC_REM_INV_REQ_ERR.
>
> 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 the same tool and reviewed
> by me.
>
> Fixes: ea1bb00ee9a5 ("RDMA/rxe: Implement flush execution in responder side")
> Signed-off-by: Andrew Stellman <astellman@xxxxxxxxxxxxxxxxxxx>
> ---
>
> Tested in QEMU with KASAN on v7.4-rc3, and on v7.4-rc3 with the RDMA
> code from rdma for-next (7404c82cbf12). The reproducer is a short
> libibverbs program; I can send it off-list.
>
> drivers/infiniband/sw/rxe/rxe_resp.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/infiniband/sw/rxe/rxe_resp.c b/drivers/infiniband/sw/rxe/rxe_resp.c
> index 1b7f31737225..38b5b0c2abe3 100644
> --- a/drivers/infiniband/sw/rxe/rxe_resp.c
> +++ b/drivers/infiniband/sw/rxe/rxe_resp.c
> @@ -340,7 +340,8 @@ static enum resp_states check_resource(struct rxe_qp *qp,
> {
> struct rxe_srq *srq = qp->srq;
>
> - if (pkt->mask & (RXE_READ_OR_ATOMIC_MASK | RXE_ATOMIC_WRITE_MASK)) {
> + if (pkt->mask & (RXE_READ_OR_ATOMIC_MASK | RXE_ATOMIC_WRITE_MASK |
> + RXE_FLUSH_MASK)) {
> /* it is the requesters job to not send
> * too many read/atomic ops, we just
> * recycle the responder resource queue
>
> base-commit: 7404c82cbf1246e0e84e2ec21834d23c26f9a8ed
> --
> 2.43.0
>
>