common/mlx5: support DevX virtq stats operations
[dpdk.git] / drivers / vdpa / mlx5 / mlx5_vdpa_virtq.c
index 781bccf..bd48460 100644 (file)
@@ -2,9 +2,12 @@
  * Copyright 2019 Mellanox Technologies, Ltd
  */
 #include <string.h>
+#include <unistd.h>
+#include <sys/mman.h>
 
 #include <rte_malloc.h>
 #include <rte_errno.h>
+#include <rte_io.h>
 
 #include <mlx5_common.h>
 
 #include "mlx5_vdpa.h"
 
 
+static void
+mlx5_vdpa_virtq_handler(void *cb_arg)
+{
+       struct mlx5_vdpa_virtq *virtq = cb_arg;
+       struct mlx5_vdpa_priv *priv = virtq->priv;
+       uint64_t buf;
+       int nbytes;
+
+       do {
+               nbytes = read(virtq->intr_handle.fd, &buf, 8);
+               if (nbytes < 0) {
+                       if (errno == EINTR ||
+                           errno == EWOULDBLOCK ||
+                           errno == EAGAIN)
+                               continue;
+                       DRV_LOG(ERR,  "Failed to read kickfd of virtq %d: %s",
+                               virtq->index, strerror(errno));
+               }
+               break;
+       } while (1);
+       rte_write32(virtq->index, priv->virtq_db_addr);
+       DRV_LOG(DEBUG, "Ring virtq %u doorbell.", virtq->index);
+}
+
 static int
 mlx5_vdpa_virtq_unset(struct mlx5_vdpa_virtq *virtq)
 {
-       int i;
+       unsigned int i;
+       int retries = MLX5_VDPA_INTR_RETRIES;
+       int ret = -EAGAIN;
 
-       if (virtq->virtq) {
-               claim_zero(mlx5_devx_cmd_destroy(virtq->virtq));
-               virtq->virtq = NULL;
+       if (virtq->intr_handle.fd != -1) {
+               while (retries-- && ret == -EAGAIN) {
+                       ret = rte_intr_callback_unregister(&virtq->intr_handle,
+                                                       mlx5_vdpa_virtq_handler,
+                                                       virtq);
+                       if (ret == -EAGAIN) {
+                               DRV_LOG(DEBUG, "Try again to unregister fd %d "
+                                       "of virtq %d interrupt, retries = %d.",
+                                       virtq->intr_handle.fd,
+                                       (int)virtq->index, retries);
+                               usleep(MLX5_VDPA_INTR_RETRIES_USEC);
+                       }
+               }
+               virtq->intr_handle.fd = -1;
        }
-       for (i = 0; i < 3; ++i) {
+       if (virtq->virtq)
+               claim_zero(mlx5_devx_cmd_destroy(virtq->virtq));
+       virtq->virtq = NULL;
+       for (i = 0; i < RTE_DIM(virtq->umems); ++i) {
                if (virtq->umems[i].obj)
                        claim_zero(mlx5_glue->devx_umem_dereg
                                                         (virtq->umems[i].obj));
@@ -37,18 +80,12 @@ mlx5_vdpa_virtq_unset(struct mlx5_vdpa_virtq *virtq)
 void
 mlx5_vdpa_virtqs_release(struct mlx5_vdpa_priv *priv)
 {
-       struct mlx5_vdpa_virtq *entry;
-       struct mlx5_vdpa_virtq *next;
-
-       entry = SLIST_FIRST(&priv->virtq_list);
-       while (entry) {
-               next = SLIST_NEXT(entry, next);
-               mlx5_vdpa_virtq_unset(entry);
-               SLIST_REMOVE(&priv->virtq_list, entry, mlx5_vdpa_virtq, next);
-               rte_free(entry);
-               entry = next;
-       }
-       SLIST_INIT(&priv->virtq_list);
+       int i;
+
+       for (i = 0; i < priv->nr_virtqs; i++) {
+               mlx5_vdpa_virtq_unset(&priv->virtqs[i]);
+               priv->virtqs[i].enable = 0;
+       }
        if (priv->tis) {
                claim_zero(mlx5_devx_cmd_destroy(priv->tis));
                priv->tis = NULL;
@@ -57,6 +94,51 @@ mlx5_vdpa_virtqs_release(struct mlx5_vdpa_priv *priv)
                claim_zero(mlx5_devx_cmd_destroy(priv->td));
                priv->td = NULL;
        }
+       if (priv->virtq_db_addr) {
+               claim_zero(munmap(priv->virtq_db_addr, priv->var->length));
+               priv->virtq_db_addr = NULL;
+       }
+       priv->features = 0;
+       priv->nr_virtqs = 0;
+}
+
+int
+mlx5_vdpa_virtq_modify(struct mlx5_vdpa_virtq *virtq, int state)
+{
+       struct mlx5_devx_virtq_attr attr = {
+                       .type = MLX5_VIRTQ_MODIFY_TYPE_STATE,
+                       .state = state ? MLX5_VIRTQ_STATE_RDY :
+                                        MLX5_VIRTQ_STATE_SUSPEND,
+                       .queue_index = virtq->index,
+       };
+
+       return mlx5_devx_cmd_modify_virtq(virtq->virtq, &attr);
+}
+
+int
+mlx5_vdpa_virtq_stop(struct mlx5_vdpa_priv *priv, int index)
+{
+       struct mlx5_devx_virtq_attr attr = {0};
+       struct mlx5_vdpa_virtq *virtq = &priv->virtqs[index];
+       int ret = mlx5_vdpa_virtq_modify(virtq, 0);
+
+       if (ret)
+               return -1;
+       if (mlx5_devx_cmd_query_virtq(virtq->virtq, &attr)) {
+               DRV_LOG(ERR, "Failed to query virtq %d.", index);
+               return -1;
+       }
+       DRV_LOG(INFO, "Query vid %d vring %d: hw_available_idx=%d, "
+               "hw_used_index=%d", priv->vid, index,
+               attr.hw_available_index, attr.hw_used_index);
+       ret = rte_vhost_set_vring_base(priv->vid, index,
+                                      attr.hw_available_index,
+                                      attr.hw_used_index);
+       if (ret) {
+               DRV_LOG(ERR, "Failed to set virtq %d base.", index);
+               return -1;
+       }
+       return 0;
 }
 
 static uint64_t
@@ -78,14 +160,14 @@ mlx5_vdpa_hva_to_gpa(struct rte_vhost_memory *mem, uint64_t hva)
 }
 
 static int
-mlx5_vdpa_virtq_setup(struct mlx5_vdpa_priv *priv,
-                     struct mlx5_vdpa_virtq *virtq, int index)
+mlx5_vdpa_virtq_setup(struct mlx5_vdpa_priv *priv, int index)
 {
+       struct mlx5_vdpa_virtq *virtq = &priv->virtqs[index];
        struct rte_vhost_vring vq;
        struct mlx5_devx_virtq_attr attr = {0};
        uint64_t gpa;
        int ret;
-       int i;
+       unsigned int i;
        uint16_t last_avail_idx;
        uint16_t last_used_idx;
 
@@ -94,6 +176,14 @@ mlx5_vdpa_virtq_setup(struct mlx5_vdpa_priv *priv,
                return -1;
        virtq->index = index;
        virtq->vq_size = vq.size;
+       attr.tso_ipv4 = !!(priv->features & (1ULL << VIRTIO_NET_F_HOST_TSO4));
+       attr.tso_ipv6 = !!(priv->features & (1ULL << VIRTIO_NET_F_HOST_TSO6));
+       attr.tx_csum = !!(priv->features & (1ULL << VIRTIO_NET_F_CSUM));
+       attr.rx_csum = !!(priv->features & (1ULL << VIRTIO_NET_F_GUEST_CSUM));
+       attr.virtio_version_1_0 = !!(priv->features & (1ULL <<
+                                                       VIRTIO_F_VERSION_1));
+       attr.type = (priv->features & (1ULL << VIRTIO_F_RING_PACKED)) ?
+                       MLX5_VIRTQ_TYPE_PACKED : MLX5_VIRTQ_TYPE_SPLIT;
        /*
         * No need event QPs creation when the guest in poll mode or when the
         * capability allows it.
@@ -116,7 +206,7 @@ mlx5_vdpa_virtq_setup(struct mlx5_vdpa_priv *priv,
                        " need event QPs and event mechanism.", index);
        }
        /* Setup 3 UMEMs for each virtq. */
-       for (i = 0; i < 3; ++i) {
+       for (i = 0; i < RTE_DIM(virtq->umems); ++i) {
                virtq->umems[i].size = priv->caps.umems[i].a * vq.size +
                                                          priv->caps.umems[i].b;
                virtq->umems[i].buf = rte_zmalloc(__func__,
@@ -139,28 +229,40 @@ mlx5_vdpa_virtq_setup(struct mlx5_vdpa_priv *priv,
                attr.umems[i].offset = 0;
                attr.umems[i].size = virtq->umems[i].size;
        }
-       gpa = mlx5_vdpa_hva_to_gpa(priv->vmem, (uint64_t)(uintptr_t)vq.desc);
-       if (!gpa) {
-               DRV_LOG(ERR, "Fail to get GPA for descriptor ring.");
-               goto error;
-       }
-       attr.desc_addr = gpa;
-       gpa = mlx5_vdpa_hva_to_gpa(priv->vmem, (uint64_t)(uintptr_t)vq.used);
-       if (!gpa) {
-               DRV_LOG(ERR, "Fail to get GPA for used ring.");
-               goto error;
-       }
-       attr.used_addr = gpa;
-       gpa = mlx5_vdpa_hva_to_gpa(priv->vmem, (uint64_t)(uintptr_t)vq.avail);
-       if (!gpa) {
-               DRV_LOG(ERR, "Fail to get GPA for available ring.");
-               goto error;
+       if (attr.type == MLX5_VIRTQ_TYPE_SPLIT) {
+               gpa = mlx5_vdpa_hva_to_gpa(priv->vmem,
+                                          (uint64_t)(uintptr_t)vq.desc);
+               if (!gpa) {
+                       DRV_LOG(ERR, "Failed to get descriptor ring GPA.");
+                       goto error;
+               }
+               attr.desc_addr = gpa;
+               gpa = mlx5_vdpa_hva_to_gpa(priv->vmem,
+                                          (uint64_t)(uintptr_t)vq.used);
+               if (!gpa) {
+                       DRV_LOG(ERR, "Failed to get GPA for used ring.");
+                       goto error;
+               }
+               attr.used_addr = gpa;
+               gpa = mlx5_vdpa_hva_to_gpa(priv->vmem,
+                                          (uint64_t)(uintptr_t)vq.avail);
+               if (!gpa) {
+                       DRV_LOG(ERR, "Failed to get GPA for available ring.");
+                       goto error;
+               }
+               attr.available_addr = gpa;
        }
-       attr.available_addr = gpa;
-       rte_vhost_get_vring_base(priv->vid, index, &last_avail_idx,
+       ret = rte_vhost_get_vring_base(priv->vid, index, &last_avail_idx,
                                 &last_used_idx);
-       DRV_LOG(INFO, "vid %d: Init last_avail_idx=%d, last_used_idx=%d for "
-               "virtq %d.", priv->vid, last_avail_idx, last_used_idx, index);
+       if (ret) {
+               last_avail_idx = 0;
+               last_used_idx = 0;
+               DRV_LOG(WARNING, "Couldn't get vring base, idx are set to 0");
+       } else {
+               DRV_LOG(INFO, "vid %d: Init last_avail_idx=%d, last_used_idx=%d for "
+                               "virtq %d.", priv->vid, last_avail_idx,
+                               last_used_idx, index);
+       }
        attr.hw_available_index = last_avail_idx;
        attr.hw_used_index = last_used_idx;
        attr.q_size = vq.size;
@@ -168,22 +270,126 @@ mlx5_vdpa_virtq_setup(struct mlx5_vdpa_priv *priv,
        attr.tis_id = priv->tis->id;
        attr.queue_index = index;
        virtq->virtq = mlx5_devx_cmd_create_virtq(priv->ctx, &attr);
+       virtq->priv = priv;
        if (!virtq->virtq)
                goto error;
+       if (mlx5_vdpa_virtq_modify(virtq, 1))
+               goto error;
+       virtq->priv = priv;
+       rte_write32(virtq->index, priv->virtq_db_addr);
+       /* Setup doorbell mapping. */
+       virtq->intr_handle.fd = vq.kickfd;
+       if (virtq->intr_handle.fd == -1) {
+               DRV_LOG(WARNING, "Virtq %d kickfd is invalid.", index);
+               if (!priv->direct_notifier) {
+                       DRV_LOG(ERR, "Virtq %d cannot be notified.", index);
+                       goto error;
+               }
+       } else {
+               virtq->intr_handle.type = RTE_INTR_HANDLE_EXT;
+               if (rte_intr_callback_register(&virtq->intr_handle,
+                                              mlx5_vdpa_virtq_handler,
+                                              virtq)) {
+                       virtq->intr_handle.fd = -1;
+                       DRV_LOG(ERR, "Failed to register virtq %d interrupt.",
+                               index);
+                       goto error;
+               } else {
+                       DRV_LOG(DEBUG, "Register fd %d interrupt for virtq %d.",
+                               virtq->intr_handle.fd, index);
+               }
+       }
        return 0;
 error:
        mlx5_vdpa_virtq_unset(virtq);
        return -1;
 }
 
+static int
+mlx5_vdpa_features_validate(struct mlx5_vdpa_priv *priv)
+{
+       if (priv->features & (1ULL << VIRTIO_F_RING_PACKED)) {
+               if (!(priv->caps.virtio_queue_type & (1 <<
+                                                    MLX5_VIRTQ_TYPE_PACKED))) {
+                       DRV_LOG(ERR, "Failed to configur PACKED mode for vdev "
+                               "%d - it was not reported by HW/driver"
+                               " capability.", priv->vid);
+                       return -ENOTSUP;
+               }
+       }
+       if (priv->features & (1ULL << VIRTIO_NET_F_HOST_TSO4)) {
+               if (!priv->caps.tso_ipv4) {
+                       DRV_LOG(ERR, "Failed to enable TSO4 for vdev %d - TSO4"
+                               " was not reported by HW/driver capability.",
+                               priv->vid);
+                       return -ENOTSUP;
+               }
+       }
+       if (priv->features & (1ULL << VIRTIO_NET_F_HOST_TSO6)) {
+               if (!priv->caps.tso_ipv6) {
+                       DRV_LOG(ERR, "Failed to enable TSO6 for vdev %d - TSO6"
+                               " was not reported by HW/driver capability.",
+                               priv->vid);
+                       return -ENOTSUP;
+               }
+       }
+       if (priv->features & (1ULL << VIRTIO_NET_F_CSUM)) {
+               if (!priv->caps.tx_csum) {
+                       DRV_LOG(ERR, "Failed to enable CSUM for vdev %d - CSUM"
+                               " was not reported by HW/driver capability.",
+                               priv->vid);
+                       return -ENOTSUP;
+               }
+       }
+       if (priv->features & (1ULL << VIRTIO_NET_F_GUEST_CSUM)) {
+               if (!priv->caps.rx_csum) {
+                       DRV_LOG(ERR, "Failed to enable GUEST CSUM for vdev %d"
+                               " GUEST CSUM was not reported by HW/driver "
+                               "capability.", priv->vid);
+                       return -ENOTSUP;
+               }
+       }
+       if (priv->features & (1ULL << VIRTIO_F_VERSION_1)) {
+               if (!priv->caps.virtio_version_1_0) {
+                       DRV_LOG(ERR, "Failed to enable version 1 for vdev %d "
+                               "version 1 was not reported by HW/driver"
+                               " capability.", priv->vid);
+                       return -ENOTSUP;
+               }
+       }
+       return 0;
+}
+
 int
 mlx5_vdpa_virtqs_prepare(struct mlx5_vdpa_priv *priv)
 {
        struct mlx5_devx_tis_attr tis_attr = {0};
-       struct mlx5_vdpa_virtq *virtq;
        uint32_t i;
        uint16_t nr_vring = rte_vhost_get_vring_num(priv->vid);
+       int ret = rte_vhost_get_negotiated_features(priv->vid, &priv->features);
 
+       if (ret || mlx5_vdpa_features_validate(priv)) {
+               DRV_LOG(ERR, "Failed to configure negotiated features.");
+               return -1;
+       }
+       if (nr_vring > priv->caps.max_num_virtio_queues * 2) {
+               DRV_LOG(ERR, "Do not support more than %d virtqs(%d).",
+                       (int)priv->caps.max_num_virtio_queues * 2,
+                       (int)nr_vring);
+               return -1;
+       }
+       /* Always map the entire page. */
+       priv->virtq_db_addr = mmap(NULL, priv->var->length, PROT_READ |
+                                  PROT_WRITE, MAP_SHARED, priv->ctx->cmd_fd,
+                                  priv->var->mmap_off);
+       if (priv->virtq_db_addr == MAP_FAILED) {
+               DRV_LOG(ERR, "Failed to map doorbell page %u.", errno);
+               priv->virtq_db_addr = NULL;
+               goto error;
+       } else {
+               DRV_LOG(DEBUG, "VAR address of doorbell mapping is %p.",
+                       priv->virtq_db_addr);
+       }
        priv->td = mlx5_devx_cmd_create_td(priv->ctx);
        if (!priv->td) {
                DRV_LOG(ERR, "Failed to create transport domain.");
@@ -195,18 +401,57 @@ mlx5_vdpa_virtqs_prepare(struct mlx5_vdpa_priv *priv)
                DRV_LOG(ERR, "Failed to create TIS.");
                goto error;
        }
+       priv->nr_virtqs = nr_vring;
        for (i = 0; i < nr_vring; i++) {
-               virtq = rte_zmalloc(__func__, sizeof(*virtq), 0);
-               if (!virtq || mlx5_vdpa_virtq_setup(priv, virtq, i)) {
-                       if (virtq)
-                               rte_free(virtq);
+               claim_zero(rte_vhost_enable_guest_notification(priv->vid, i,
+                                                              1));
+               if (mlx5_vdpa_virtq_setup(priv, i))
                        goto error;
-               }
-               SLIST_INSERT_HEAD(&priv->virtq_list, virtq, next);
        }
-       priv->nr_virtqs = nr_vring;
        return 0;
 error:
        mlx5_vdpa_virtqs_release(priv);
        return -1;
 }
+
+int
+mlx5_vdpa_virtq_enable(struct mlx5_vdpa_priv *priv, int index, int enable)
+{
+       struct mlx5_vdpa_virtq *virtq = &priv->virtqs[index];
+       int ret;
+
+       DRV_LOG(INFO, "Update virtq %d status %sable -> %sable.", index,
+               virtq->enable ? "en" : "dis", enable ? "en" : "dis");
+       if (virtq->enable == !!enable)
+               return 0;
+       if (!priv->configured) {
+               virtq->enable = !!enable;
+               return 0;
+       }
+       if (enable) {
+               /* Configuration might have been updated - reconfigure virtq. */
+               if (virtq->virtq) {
+                       ret = mlx5_vdpa_virtq_stop(priv, index);
+                       if (ret)
+                               DRV_LOG(WARNING, "Failed to stop virtq %d.",
+                                       index);
+                       mlx5_vdpa_virtq_unset(virtq);
+               }
+               ret = mlx5_vdpa_virtq_setup(priv, index);
+               if (ret) {
+                       DRV_LOG(ERR, "Failed to setup virtq %d.", index);
+                       return ret;
+                       /* The only case virtq can stay invalid. */
+               }
+       }
+       virtq->enable = !!enable;
+       if (is_virtq_recvq(virtq->index, priv->nr_virtqs)) {
+               /* Need to add received virtq to the RQT table of the TIRs. */
+               ret = mlx5_vdpa_steer_update(priv);
+               if (ret) {
+                       virtq->enable = !enable;
+                       return ret;
+               }
+       }
+       return 0;
+}