vdpa/mlx5: fix notification timing
[dpdk.git] / drivers / vdpa / mlx5 / mlx5_vdpa_event.c
index 743e52a..0414c91 100644 (file)
@@ -20,9 +20,6 @@
 #include "mlx5_vdpa.h"
 
 
-#define MLX5_VDPA_DEFAULT_TIMER_DELAY_US 500u
-#define MLX5_VDPA_NO_TRAFFIC_TIME_S 2LLU
-
 void
 mlx5_vdpa_event_qp_global_release(struct mlx5_vdpa_priv *priv)
 {
@@ -127,12 +124,12 @@ mlx5_vdpa_cq_create(struct mlx5_vdpa_priv *priv, uint16_t log_desc_n,
        struct mlx5_devx_cq_attr attr;
        size_t pgsize = sysconf(_SC_PAGESIZE);
        uint32_t umem_size;
-       int ret;
        uint16_t event_nums[1] = {0};
+       uint16_t cq_size = 1 << log_desc_n;
+       int ret;
 
        cq->log_desc_n = log_desc_n;
-       umem_size = sizeof(struct mlx5_cqe) * (1 << log_desc_n) +
-                                                       sizeof(*cq->db_rec) * 2;
+       umem_size = sizeof(struct mlx5_cqe) * cq_size + sizeof(*cq->db_rec) * 2;
        cq->umem_buf = rte_zmalloc(__func__, umem_size, 4096);
        if (!cq->umem_buf) {
                DRV_LOG(ERR, "Failed to allocate memory for CQ.");
@@ -149,13 +146,13 @@ mlx5_vdpa_cq_create(struct mlx5_vdpa_priv *priv, uint16_t log_desc_n,
        }
        attr.q_umem_valid = 1;
        attr.db_umem_valid = 1;
-       attr.use_first_only = 0;
+       attr.use_first_only = 1;
        attr.overrun_ignore = 0;
        attr.uar_page_id = priv->uar->page_id;
        attr.q_umem_id = cq->umem_obj->umem_id;
        attr.q_umem_offset = 0;
        attr.db_umem_id = cq->umem_obj->umem_id;
-       attr.db_umem_offset = sizeof(struct mlx5_cqe) * (1 << log_desc_n);
+       attr.db_umem_offset = sizeof(struct mlx5_cqe) * cq_size;
        attr.eqn = priv->eqn;
        attr.log_cq_size = log_desc_n;
        attr.log_page_size = rte_log2_u32(pgsize);
@@ -175,19 +172,10 @@ mlx5_vdpa_cq_create(struct mlx5_vdpa_priv *priv, uint16_t log_desc_n,
                rte_errno = errno;
                goto error;
        }
-       if (callfd != -1) {
-               ret = mlx5_glue->devx_subscribe_devx_event_fd(priv->eventc,
-                                                             callfd,
-                                                             cq->cq->obj, 0);
-               if (ret) {
-                       DRV_LOG(ERR, "Failed to subscribe CQE event fd.");
-                       rte_errno = errno;
-                       goto error;
-               }
-       }
        cq->callfd = callfd;
        /* Init CQ to ones to be in HW owner in the start. */
-       memset((void *)(uintptr_t)cq->umem_buf, 0xFF, attr.db_umem_offset);
+       cq->cqes[0].op_own = MLX5_CQE_OWNER_MASK;
+       cq->cqes[0].wqe_counter = rte_cpu_to_be_16(cq_size - 1);
        /* First arming. */
        mlx5_vdpa_cq_arm(priv, cq);
        return 0;
@@ -203,34 +191,40 @@ mlx5_vdpa_cq_poll(struct mlx5_vdpa_cq *cq)
                                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;
-       uint32_t total = 0;
-       int ret;
-
-       do {
-               volatile struct mlx5_cqe *cqe = cq->cqes + ((cq->cq_ci + total)
-                                                           & cq_mask);
-
-               ret = check_cqe(cqe, cq_size, cq->cq_ci + total);
-               switch (ret) {
-               case MLX5_CQE_STATUS_ERR:
+       union {
+               struct {
+                       uint16_t wqe_counter;
+                       uint8_t rsvd5;
+                       uint8_t op_own;
+               };
+               uint32_t word;
+       } last_word;
+       uint16_t next_wqe_counter = cq->cq_ci & cq_mask;
+       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;
+       if (comp) {
+               cq->cq_ci += comp;
+               MLX5_ASSERT(!!(cq->cq_ci & cq_size) ==
+                           MLX5_CQE_OWNER(last_word.op_own));
+               MLX5_ASSERT(MLX5_CQE_OPCODE(last_word.op_own) !=
+                           MLX5_CQE_INVALID);
+               if (unlikely(!(MLX5_CQE_OPCODE(last_word.op_own) ==
+                              MLX5_CQE_RESP_ERR ||
+                              MLX5_CQE_OPCODE(last_word.op_own) ==
+                              MLX5_CQE_REQ_ERR)))
                        cq->errors++;
-                       /*fall-through*/
-               case MLX5_CQE_STATUS_SW_OWN:
-                       total++;
-                       break;
-               case MLX5_CQE_STATUS_HW_OWN:
-               default:
-                       break;
-               }
-       } while (ret != MLX5_CQE_STATUS_HW_OWN);
-       rte_io_wmb();
-       cq->cq_ci += total;
-       /* Ring CQ doorbell record. */
-       cq->db_rec[0] = rte_cpu_to_be_32(cq->cq_ci);
-       rte_io_wmb();
-       /* Ring SW QP doorbell record. */
-       eqp->db_rec[0] = rte_cpu_to_be_32(cq->cq_ci + cq_size);
-       return total;
+               rte_io_wmb();
+               /* Ring CQ doorbell record. */
+               cq->db_rec[0] = rte_cpu_to_be_32(cq->cq_ci);
+               rte_io_wmb();
+               /* Ring SW QP doorbell record. */
+               eqp->db_rec[0] = rte_cpu_to_be_32(cq->cq_ci + cq_size);
+       }
+       return comp;
 }
 
 static void
@@ -246,21 +240,43 @@ mlx5_vdpa_arm_all_cqs(struct mlx5_vdpa_priv *priv)
        }
 }
 
+static void
+mlx5_vdpa_timer_sleep(struct mlx5_vdpa_priv *priv, uint32_t max)
+{
+       if (priv->event_mode == MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER) {
+               switch (max) {
+               case 0:
+                       priv->timer_delay_us += priv->event_us;
+                       break;
+               case 1:
+                       break;
+               default:
+                       priv->timer_delay_us /= max;
+                       break;
+               }
+       }
+       usleep(priv->timer_delay_us);
+}
+
 static void *
 mlx5_vdpa_poll_handle(void *arg)
 {
        struct mlx5_vdpa_priv *priv = arg;
        int i;
        struct mlx5_vdpa_cq *cq;
-       uint32_t total;
+       uint32_t max;
        uint64_t current_tic;
 
        pthread_mutex_lock(&priv->timer_lock);
        while (!priv->timer_on)
                pthread_cond_wait(&priv->timer_cond, &priv->timer_lock);
        pthread_mutex_unlock(&priv->timer_lock);
+       priv->timer_delay_us = priv->event_mode ==
+                                           MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER ?
+                                             MLX5_VDPA_DEFAULT_TIMER_DELAY_US :
+                                                                priv->event_us;
        while (1) {
-               total = 0;
+               max = 0;
                for (i = 0; i < priv->nr_virtqs; i++) {
                        cq = &priv->virtqs[i].eqp.cq;
                        if (cq->cq && !cq->armed) {
@@ -271,15 +287,16 @@ mlx5_vdpa_poll_handle(void *arg)
                                        if (cq->callfd != -1)
                                                eventfd_write(cq->callfd,
                                                              (eventfd_t)1);
-                                       total += comp;
+                                       if (comp > max)
+                                               max = comp;
                                }
                        }
                }
                current_tic = rte_rdtsc();
-               if (!total) {
+               if (!max) {
                        /* No traffic ? stop timer and load interrupts. */
                        if (current_tic - priv->last_traffic_tic >=
-                           rte_get_timer_hz() * MLX5_VDPA_NO_TRAFFIC_TIME_S) {
+                           rte_get_timer_hz() * priv->no_traffic_time_s) {
                                DRV_LOG(DEBUG, "Device %s traffic was stopped.",
                                        priv->vdev->device->name);
                                mlx5_vdpa_arm_all_cqs(priv);
@@ -289,12 +306,16 @@ mlx5_vdpa_poll_handle(void *arg)
                                        pthread_cond_wait(&priv->timer_cond,
                                                          &priv->timer_lock);
                                pthread_mutex_unlock(&priv->timer_lock);
+                               priv->timer_delay_us = priv->event_mode ==
+                                           MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER ?
+                                             MLX5_VDPA_DEFAULT_TIMER_DELAY_US :
+                                                                priv->event_us;
                                continue;
                        }
                } else {
                        priv->last_traffic_tic = current_tic;
                }
-               usleep(priv->timer_delay_us);
+               mlx5_vdpa_timer_sleep(priv, max);
        }
        return NULL;
 }
@@ -320,6 +341,13 @@ mlx5_vdpa_interrupt_handler(void *cb_arg)
                                                   struct mlx5_vdpa_virtq, eqp);
 
                mlx5_vdpa_cq_poll(cq);
+               /* Notify guest for descs consuming. */
+               if (cq->callfd != -1)
+                       eventfd_write(cq->callfd, (eventfd_t)1);
+               if (priv->event_mode == MLX5_VDPA_EVENT_MODE_ONLY_INTERRUPT) {
+                       mlx5_vdpa_cq_arm(priv, cq);
+                       return;
+               }
                /* Don't arm again - timer will take control. */
                DRV_LOG(DEBUG, "Device %s virtq %d cq %d event was captured."
                        " Timer is %s, cq ci is %u.\n",
@@ -349,15 +377,16 @@ mlx5_vdpa_cqe_event_setup(struct mlx5_vdpa_priv *priv)
        if (!priv->eventc)
                /* All virtqs are in poll mode. */
                return 0;
-       pthread_mutex_init(&priv->timer_lock, NULL);
-       pthread_cond_init(&priv->timer_cond, NULL);
-       priv->timer_on = 0;
-       priv->timer_delay_us = MLX5_VDPA_DEFAULT_TIMER_DELAY_US;
-       ret = pthread_create(&priv->timer_tid, NULL, mlx5_vdpa_poll_handle,
-                            (void *)priv);
-       if (ret) {
-               DRV_LOG(ERR, "Failed to create timer thread.");
-               return -1;
+       if (priv->event_mode != MLX5_VDPA_EVENT_MODE_ONLY_INTERRUPT) {
+               pthread_mutex_init(&priv->timer_lock, NULL);
+               pthread_cond_init(&priv->timer_cond, NULL);
+               priv->timer_on = 0;
+               ret = pthread_create(&priv->timer_tid, NULL,
+                                    mlx5_vdpa_poll_handle, (void *)priv);
+               if (ret) {
+                       DRV_LOG(ERR, "Failed to create timer thread.");
+                       return -1;
+               }
        }
        flags = fcntl(priv->eventc->fd, F_GETFL);
        ret = fcntl(priv->eventc->fd, F_SETFL, flags | O_NONBLOCK);