raw/ioat: support limiting queues for idxd PCI device
[dpdk.git] / drivers / vdpa / mlx5 / mlx5_vdpa_event.c
index a055d20..404e135 100644 (file)
@@ -3,10 +3,12 @@
  */
 #include <unistd.h>
 #include <stdint.h>
+#include <sched.h>
 #include <fcntl.h>
 #include <sys/eventfd.h>
 
 #include <rte_malloc.h>
+#include <rte_memory.h>
 #include <rte_errno.h>
 #include <rte_lcore.h>
 #include <rte_atomic.h>
@@ -16,6 +18,7 @@
 
 #include <mlx5_common.h>
 #include <mlx5_common_os.h>
+#include <mlx5_common_devx.h>
 #include <mlx5_glue.h>
 
 #include "mlx5_vdpa_utils.h"
@@ -48,7 +51,6 @@ mlx5_vdpa_event_qp_global_release(struct mlx5_vdpa_priv *priv)
                priv->eventc = NULL;
        }
 #endif
-       priv->eqn = 0;
 }
 
 /* Prepare all the global resources for all the event objects.*/
@@ -59,11 +61,6 @@ mlx5_vdpa_event_qp_global_prepare(struct mlx5_vdpa_priv *priv)
 
        if (priv->eventc)
                return 0;
-       if (mlx5_glue->devx_query_eqn(priv->ctx, 0, &priv->eqn)) {
-               rte_errno = errno;
-               DRV_LOG(ERR, "Failed to query EQ number %d.", rte_errno);
-               return -1;
-       }
        priv->eventc = mlx5_os_devx_create_event_channel(priv->ctx,
                           MLX5DV_DEVX_CREATE_EVENT_CHANNEL_FLAGS_OMIT_EV_DATA);
        if (!priv->eventc) {
@@ -98,12 +95,7 @@ error:
 static void
 mlx5_vdpa_cq_destroy(struct mlx5_vdpa_cq *cq)
 {
-       if (cq->cq)
-               claim_zero(mlx5_devx_cmd_destroy(cq->cq));
-       if (cq->umem_obj)
-               claim_zero(mlx5_glue->devx_umem_dereg(cq->umem_obj));
-       if (cq->umem_buf)
-               rte_free((void *)(uintptr_t)cq->umem_buf);
+       mlx5_devx_cq_destroy(&cq->cq_obj);
        memset(cq, 0, sizeof(*cq));
 }
 
@@ -113,12 +105,12 @@ mlx5_vdpa_cq_arm(struct mlx5_vdpa_priv *priv, struct mlx5_vdpa_cq *cq)
        uint32_t arm_sn = cq->arm_sn << MLX5_CQ_SQN_OFFSET;
        uint32_t cq_ci = cq->cq_ci & MLX5_CI_MASK;
        uint32_t doorbell_hi = arm_sn | MLX5_CQ_DBR_CMD_ALL | cq_ci;
-       uint64_t doorbell = ((uint64_t)doorbell_hi << 32) | cq->cq->id;
+       uint64_t doorbell = ((uint64_t)doorbell_hi << 32) | cq->cq_obj.cq->id;
        uint64_t db_be = rte_cpu_to_be_64(doorbell);
        uint32_t *addr = RTE_PTR_ADD(priv->uar->base_addr, MLX5_CQ_DOORBELL);
 
        rte_io_wmb();
-       cq->db_rec[MLX5_CQ_ARM_DB] = rte_cpu_to_be_32(doorbell_hi);
+       cq->cq_obj.db_rec[MLX5_CQ_ARM_DB] = rte_cpu_to_be_32(doorbell_hi);
        rte_wmb();
 #ifdef RTE_ARCH_64
        *(uint64_t *)addr = db_be;
@@ -135,52 +127,25 @@ static int
 mlx5_vdpa_cq_create(struct mlx5_vdpa_priv *priv, uint16_t log_desc_n,
                    int callfd, struct mlx5_vdpa_cq *cq)
 {
-       struct mlx5_devx_cq_attr attr = {0};
-       size_t pgsize = sysconf(_SC_PAGESIZE);
-       uint32_t umem_size;
+       struct mlx5_devx_cq_attr attr = {
+               .use_first_only = 1,
+               .uar_page_id = priv->uar->page_id,
+       };
        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) * 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.");
-               rte_errno = ENOMEM;
-               return -ENOMEM;
-       }
-       cq->umem_obj = mlx5_glue->devx_umem_reg(priv->ctx,
-                                               (void *)(uintptr_t)cq->umem_buf,
-                                               umem_size,
-                                               IBV_ACCESS_LOCAL_WRITE);
-       if (!cq->umem_obj) {
-               DRV_LOG(ERR, "Failed to register umem for CQ.");
-               goto error;
-       }
-       attr.q_umem_valid = 1;
-       attr.db_umem_valid = 1;
-       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) * cq_size;
-       attr.eqn = priv->eqn;
-       attr.log_cq_size = log_desc_n;
-       attr.log_page_size = rte_log2_u32(pgsize);
-       cq->cq = mlx5_devx_cmd_create_cq(priv->ctx, &attr);
-       if (!cq->cq)
+       ret = mlx5_devx_cq_create(priv->ctx, &cq->cq_obj, log_desc_n, &attr,
+                                 SOCKET_ID_ANY);
+       if (ret)
                goto error;
-       cq->db_rec = RTE_PTR_ADD(cq->umem_buf, (uintptr_t)attr.db_umem_offset);
        cq->cq_ci = 0;
+       cq->log_desc_n = log_desc_n;
        rte_spinlock_init(&cq->sl);
        /* Subscribe CQ event to the event channel controlled by the driver. */
-       ret = mlx5_os_devx_subscribe_devx_event(priv->eventc, cq->cq->obj,
-                                                  sizeof(event_nums),
-                                                  event_nums,
-                                                  (uint64_t)(uintptr_t)cq);
+       ret = mlx5_os_devx_subscribe_devx_event(priv->eventc,
+                                               cq->cq_obj.cq->obj,
+                                               sizeof(event_nums), event_nums,
+                                               (uint64_t)(uintptr_t)cq);
        if (ret) {
                DRV_LOG(ERR, "Failed to subscribe CQE event.");
                rte_errno = errno;
@@ -188,8 +153,8 @@ 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(UINT16_MAX);
+       cq->cq_obj.cqes[0].op_own = MLX5_CQE_OWNER_MASK;
+       cq->cq_obj.cqes[0].wqe_counter = rte_cpu_to_be_16(UINT16_MAX);
        /* First arming. */
        mlx5_vdpa_cq_arm(priv, cq);
        return 0;
@@ -216,7 +181,7 @@ mlx5_vdpa_cq_poll(struct mlx5_vdpa_cq *cq)
        uint16_t cur_wqe_counter;
        uint16_t comp;
 
-       last_word.word = rte_read32(&cq->cqes[0].wqe_counter);
+       last_word.word = rte_read32(&cq->cq_obj.cqes[0].wqe_counter);
        cur_wqe_counter = rte_be_to_cpu_16(last_word.wqe_counter);
        comp = cur_wqe_counter + (uint16_t)1 - next_wqe_counter;
        if (comp) {
@@ -230,7 +195,7 @@ mlx5_vdpa_cq_poll(struct mlx5_vdpa_cq *cq)
                        cq->errors++;
                rte_io_wmb();
                /* Ring CQ doorbell record. */
-               cq->db_rec[0] = rte_cpu_to_be_32(cq->cq_ci);
+               cq->cq_obj.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);
@@ -246,7 +211,7 @@ mlx5_vdpa_arm_all_cqs(struct mlx5_vdpa_priv *priv)
 
        for (i = 0; i < priv->nr_virtqs; i++) {
                cq = &priv->virtqs[i].eqp.cq;
-               if (cq->cq && !cq->armed)
+               if (cq->cq_obj.cq && !cq->armed)
                        mlx5_vdpa_cq_arm(priv, cq);
        }
 }
@@ -268,6 +233,9 @@ mlx5_vdpa_timer_sleep(struct mlx5_vdpa_priv *priv, uint32_t max)
        }
        if (priv->timer_delay_us)
                usleep(priv->timer_delay_us);
+       else
+               /* Give-up CPU to improve polling threads scheduling. */
+               sched_yield();
 }
 
 static void *
@@ -292,7 +260,7 @@ mlx5_vdpa_poll_handle(void *arg)
                pthread_mutex_lock(&priv->vq_config_lock);
                for (i = 0; i < priv->nr_virtqs; i++) {
                        cq = &priv->virtqs[i].eqp.cq;
-                       if (cq->cq && !cq->armed) {
+                       if (cq->cq_obj.cq && !cq->armed) {
                                uint32_t comp = mlx5_vdpa_cq_poll(cq);
 
                                if (comp) {
@@ -371,7 +339,7 @@ mlx5_vdpa_interrupt_handler(void *cb_arg)
                DRV_LOG(DEBUG, "Device %s virtq %d cq %d event was captured."
                        " Timer is %s, cq ci is %u.\n",
                        priv->vdev->device->name,
-                       (int)virtq->index, cq->cq->id,
+                       (int)virtq->index, cq->cq_obj.cq->id,
                        priv->timer_on ? "on" : "off", cq->cq_ci);
                cq->armed = 0;
        }
@@ -536,6 +504,9 @@ mlx5_vdpa_cqe_event_setup(struct mlx5_vdpa_priv *priv)
        rte_cpuset_t cpuset;
        pthread_attr_t attr;
        char name[16];
+       const struct sched_param sp = {
+               .sched_priority = sched_get_priority_max(SCHED_RR),
+       };
 
        if (!priv->eventc)
                /* All virtqs are in poll mode. */
@@ -545,15 +516,14 @@ mlx5_vdpa_cqe_event_setup(struct mlx5_vdpa_priv *priv)
                pthread_cond_init(&priv->timer_cond, NULL);
                priv->timer_on = 0;
                pthread_attr_init(&attr);
-               CPU_ZERO(&cpuset);
-               if (priv->event_core != -1)
-                       CPU_SET(priv->event_core, &cpuset);
-               else
-                       cpuset = rte_lcore_cpuset(rte_get_main_lcore());
-               ret = pthread_attr_setaffinity_np(&attr, sizeof(cpuset),
-                                                 &cpuset);
+               ret = pthread_attr_setschedpolicy(&attr, SCHED_RR);
                if (ret) {
-                       DRV_LOG(ERR, "Failed to set thread affinity.");
+                       DRV_LOG(ERR, "Failed to set thread sched policy = RR.");
+                       return -1;
+               }
+               ret = pthread_attr_setschedparam(&attr, &sp);
+               if (ret) {
+                       DRV_LOG(ERR, "Failed to set thread priority.");
                        return -1;
                }
                ret = pthread_create(&priv->timer_tid, &attr,
@@ -562,6 +532,17 @@ mlx5_vdpa_cqe_event_setup(struct mlx5_vdpa_priv *priv)
                        DRV_LOG(ERR, "Failed to create timer thread.");
                        return -1;
                }
+               CPU_ZERO(&cpuset);
+               if (priv->event_core != -1)
+                       CPU_SET(priv->event_core, &cpuset);
+               else
+                       cpuset = rte_lcore_cpuset(rte_get_main_lcore());
+               ret = pthread_setaffinity_np(priv->timer_tid,
+                                            sizeof(cpuset), &cpuset);
+               if (ret) {
+                       DRV_LOG(ERR, "Failed to set thread affinity.");
+                       goto error;
+               }
                snprintf(name, sizeof(name), "vDPA-mlx5-%d", priv->vid);
                ret = pthread_setname_np(priv->timer_tid, name);
                if (ret) {
@@ -682,6 +663,7 @@ mlx5_vdpa_event_qp_create(struct mlx5_vdpa_priv *priv, uint16_t desc_n,
        if (mlx5_vdpa_cq_create(priv, log_desc_n, callfd, &eqp->cq))
                return -1;
        attr.pd = priv->pdn;
+       attr.ts_format = mlx5_ts_format_conv(priv->qp_ts_format);
        eqp->fw_qp = mlx5_devx_cmd_create_qp(priv->ctx, &attr);
        if (!eqp->fw_qp) {
                DRV_LOG(ERR, "Failed to create FW QP(%u).", rte_errno);
@@ -702,7 +684,7 @@ mlx5_vdpa_event_qp_create(struct mlx5_vdpa_priv *priv, uint16_t desc_n,
                goto error;
        }
        attr.uar_index = priv->uar->page_id;
-       attr.cqn = eqp->cq.cq->id;
+       attr.cqn = eqp->cq.cq_obj.cq->id;
        attr.log_page_size = rte_log2_u32(sysconf(_SC_PAGESIZE));
        attr.rq_size = 1 << log_desc_n;
        attr.log_rq_stride = rte_log2_u32(MLX5_WSEG_SIZE);
@@ -712,6 +694,7 @@ mlx5_vdpa_event_qp_create(struct mlx5_vdpa_priv *priv, uint16_t desc_n,
        attr.wq_umem_offset = 0;
        attr.dbr_umem_id = eqp->umem_obj->umem_id;
        attr.dbr_address = (1 << log_desc_n) * MLX5_WSEG_SIZE;
+       attr.ts_format = mlx5_ts_format_conv(priv->qp_ts_format);
        eqp->sw_qp = mlx5_devx_cmd_create_qp(priv->ctx, &attr);
        if (!eqp->sw_qp) {
                DRV_LOG(ERR, "Failed to create SW QP(%u).", rte_errno);