[PATCH] net/mana: fix Tx stall from send queue free-space unit mismatch

[Stephen Hemminger]

On Mon, 14 Sep 2026 13:58:08 +0300
Rita Ruvinsky <rita.ruvinsky at weka.io> wrote:

> gdma_post_work_request() subtracted a unit count from an entry count:
> 
>   queue_free_units = queue->count - (queue->head - queue->tail);
> 
> queue->count is in entries, while head and tail are in WQE alignment
> units. On a 512-entry, 128KB send queue the check saw 512 units of
> capacity instead of queue->size / GDMA_WQE_ALIGNMENT_UNIT_SIZE = 4096,
> and returned -EBUSY with the queue one eighth full. A workload that
> fills that window faster than it drains makes rte_eth_tx_burst() return
> 0 for long enough to look like a dead port.
> 
> Derive the capacity from queue->size, which is also what the ring wrap
> in gdma_get_wqe_pointer() uses. Rx is unaffected: its WQEs occupy
> exactly one unit, so entries and units coincide.
> 
> Fixes: 56dd45c0ce7b ("net/mana: implement hardware layer operations")
> Cc: stable at dpdk.org
> 
> Signed-off-by: Rita Ruvinsky <rita.ruvinsky at weka.io>
> ---

Applied to next-net

The long form AI review had some observations worth including:

On Mon, 14 Sep 2026 13:58:08 +0300
Rita Ruvinsky <rita.ruvinsky at weka.io> wrote:

> gdma_post_work_request() subtracted a unit count from an entry count:

The unit analysis is right.  head/tail are advanced in alignment
units (queue->head += wqe_size / GDMA_WQE_ALIGNMENT_UNIT_SIZE, and
gdma_get_wqe_pointer() multiplies head by the same constant), while
sq_count comes from rdma-core as attr->cap.max_send_wr and sq_size as
align_hw_size(max_send_wr * get_wqe_size(max_send_sge)).  Deriving the
capacity from size is the only self-consistent choice, and it is what
mana_gd_wq_avail_space() in the kernel driver does.

Info:

1. The debug line in the -EBUSY path still reports queue->count:

	DP_LOG(DEBUG, "WQE size %u queue count %u head %u tail %u",
	       wqe_size, queue->count, queue->head, queue->tail);

After this patch count no longer takes part in the decision for the
send or receive queue; only gdma_poll_completion_queue() still uses it,
for the CQ.  The one line printed when a post is rejected no longer
shows what it was rejected against.  Suggest:

	DP_LOG(DEBUG, "WQE size %u queue size %u free %u head %u tail %u",
	       wqe_size, queue->size, queue_free_units,
	       queue->head, queue->tail);

2. The comment describes the old bug rather than the invariant:

	/* head/tail count WQE alignment units, so the capacity they are
	 * compared against must too: queue->count is in entries and
	 * undercounts the queue, stalling Tx well below capacity.
	 */

The stall belongs in the commit message, where it already is.  In the
source the invariant is enough:

	/* head and tail are in WQE alignment units, so the capacity must
	 * come from the queue size in bytes, not the entry count.
	 */

3. Worth a sentence in the commit message that the kernel mana driver
computes the same limit in mana_gd_wq_avail_space(), in bytes:

	u32 used_space = (wq->head - wq->tail) * GDMA_WQE_BU_SIZE;
	return wq->queue_size - used_space;

It is independent confirmation of the unit convention and tells anyone
backporting this that the two drivers now agree.