d/numam-dpdk
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

vdpa/mlx5: fix completion queue polling

Browse Source 
The CQ polling is done in order to notify the guest about new traffic
bursts and to release FW resources for the next bursts management.

When HW is faster than SW, it may be that all the FW resources are busy
in SW due to late polling.
In this case, due to wrong WQE counter masking, the fullness
calculation of the completions number is 0 while the queue is full.

Change the WQE counter masking to 16-bit wideness instead of the CQ
size mask as defined by the CQE format.

Fixes: c5f714e50b0e ("vdpa/mlx5: optimize completion queue poll")
Cc: stable@dpdk.org

Signed-off-by: Matan Azrad <matan@nvidia.com>
Acked-by: Xueming Li <xuemingl@nvidia.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
This commit is contained in:
Matan Azrad 2020-09-10 07:20:34 +00:00 committed by Ferruh Yigit
parent 9c0e15a117
commit 4fb86eb5e8
1 changed files with 3 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
drivers/vdpa/mlx5/mlx5_vdpa_event.c
View File

@ -180,7 +180,7 @@ mlx5_vdpa_cq_create(struct mlx5_vdpa_priv *priv, uint16_t log_desc_n,
cq->callfd = callfd;
/* Init CQ to ones to be in HW owner in the start. */
cq->cqes[0].op_own = MLX5_CQE_OWNER_MASK;
cq->cqes[0].wqe_counter = rte_cpu_to_be_16(cq_size - 1);
cq->cqes[0].wqe_counter = rte_cpu_to_be_16(UINT16_MAX);
/* First arming. */
mlx5_vdpa_cq_arm(priv, cq);
return 0;
@ -195,7 +195,6 @@ mlx5_vdpa_cq_poll(struct mlx5_vdpa_cq *cq)
struct mlx5_vdpa_event_qp *eqp =
container_of(cq, struct mlx5_vdpa_event_qp, cq);
const unsigned int cq_size = 1 << cq->log_desc_n;
const unsigned int cq_mask = cq_size - 1;
union {
struct {
uint16_t wqe_counter;
@ -204,13 +203,13 @@ mlx5_vdpa_cq_poll(struct mlx5_vdpa_cq *cq)
};
uint32_t word;
} last_word;
uint16_t next_wqe_counter = cq->cq_ci & cq_mask;
uint16_t next_wqe_counter = cq->cq_ci;
uint16_t cur_wqe_counter;
uint16_t comp;
last_word.word = rte_read32(&cq->cqes[0].wqe_counter);
cur_wqe_counter = rte_be_to_cpu_16(last_word.wqe_counter);
comp = (cur_wqe_counter + 1u - next_wqe_counter) & cq_mask;
comp = cur_wqe_counter + (uint16_t)1 - next_wqe_counter;
if (comp) {
cq->cq_ci += comp;
MLX5_ASSERT(MLX5_CQE_OPCODE(last_word.op_own) !=