Re: [PATCH rdma-next v2 5/5] RDMA/mana_ib: Poll UD completions and flush software error QPs

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] RDMA/mana_ib: Send Queue deadlock when handling unsignaled Work Requests
--

commit d64ff9be78f7ee54ec40f144bfa6a3c96bf9df22
Author: Konstantin Taranov <kotaranov@microsoft.com>
RDMA/mana_ib: Poll UD completions and flush software error QPs

This commit updates the driver to build UD and GSI work completions directly
from each hardware CQE and shadow entry. It introduces a WC-budgeted poll
context and replaces the all-QP CQ membership list with persistent send and
receive error-QP lists.

> diff --git a/drivers/infiniband/hw/mana/wr.c b/drivers/infiniband/hw/mana/wr.c
> index eb671b45b2976..15faca66179d1 100644
> --- a/drivers/infiniband/hw/mana/wr.c
> +++ b/drivers/infiniband/hw/mana/wr.c
[ ... ]
> @@ -144,6 +145,7 @@ static int mana_ib_post_send_ud_one(struct mana_ib_qp *qp, const struct ib_ud_wr
>  	memset(shadow_wqe, 0, sizeof(*shadow_wqe));
>  	shadow_wqe->wr_id = wr->wr.wr_id;
>  	shadow_wqe->wqe_size_in_bu = wqe_info.wqe_size_in_bu;
> +	shadow_wqe->wqe_offset_or_psn = wqe_info.wqe_offset;
>  	shadow_queue_advance_producer(&qp->shadow_sq);
>  
>  	return 0;

[Severity: High]
Does this code lose track of which Work Requests were posted as unsignaled?

In mana_ib_post_send_ud_one(), shadow_wqe is zeroed out, but the
MANA_WQ_NO_SIGNAL_WC flag is never conditionally set in shadow_wqe->flags based
on the unsignaled status of the WR.

> diff --git a/drivers/infiniband/hw/mana/cq.c b/drivers/infiniband/hw/mana/cq.c
> index 6764e2de0c9f9..55d7a7d1cd69b 100644
> --- a/drivers/infiniband/hw/mana/cq.c
> +++ b/drivers/infiniband/hw/mana/cq.c
[ ... ]
> @@ -236,126 +288,156 @@ int mana_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
[ ... ]
> +static void handle_ud_sq_cqe(struct mana_ib_qp *qp, struct mana_rdma_cqe *rdma_cqe,
> +			     struct mana_cq_poll *poll)
>  {
> +	struct shadow_wqe_header *wqe = shadow_queue_get_next_to_consume(&qp->shadow_sq);
> +	u32 offset = rdma_cqe->ud_send.tx_wqe_offset & MANA_WQE_OFFSET_MASK;
>  
> +	if (!wqe || wqe->wqe_offset_or_psn != offset)
>  		return;
>  
> +	mana_complete_send(qp, poll, rdma_cqe->ud_send.vendor_error);
>  }

[Severity: High]
Can this strict offset check cause the Send Queue to become permanently
deadlocked if the user posts an unsignaled Work Request?

When an unsignaled WR is posted and succeeds, the hardware will skip
generating a CQE for it. The next CQE generated will belong to a subsequent
signaled WR.

Because there is no loop here to advance the shadow queue past unsignaled WRs,
wqe will continue to point to the older unsignaled WR, while offset belongs
to the newer signaled WR. This mismatch causes handle_ud_sq_cqe() to return
immediately. The CQE is discarded and the shadow queue is never advanced,
permanently stalling the queue.

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260918140407.2922819-1-kotaranov@linux.microsoft.com?part=5