vdpa/mlx5: prepare virtio queues
authorMatan Azrad <matan@mellanox.com>
Sun, 2 Feb 2020 16:03:46 +0000 (16:03 +0000)
committerFerruh Yigit <ferruh.yigit@intel.com>
Wed, 5 Feb 2020 08:51:21 +0000 (09:51 +0100)
The HW virtq object represents an emulated context for a VIRTIO_NET
virtqueue which was created and managed by a VIRTIO_NET driver as
defined in VIRTIO Specification.

Add support to prepare and release all the basic HW resources needed
the user virtqs emulation according to the rte_vhost configurations.

This patch prepares the basic configurations needed by DevX commands to
create a virtq.

Add new file mlx5_vdpa_virtq.c to manage virtq operations.

Signed-off-by: Matan Azrad <matan@mellanox.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
drivers/vdpa/mlx5/Makefile
drivers/vdpa/mlx5/meson.build
drivers/vdpa/mlx5/mlx5_vdpa.c
drivers/vdpa/mlx5/mlx5_vdpa.h
drivers/vdpa/mlx5/mlx5_vdpa_virtq.c [new file with mode: 0644]

index 086af1b..1b24400 100644 (file)
@@ -10,6 +10,7 @@ LIB = librte_pmd_mlx5_vdpa.a
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_VDPA_PMD) += mlx5_vdpa.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_VDPA_PMD) += mlx5_vdpa_mem.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_VDPA_PMD) += mlx5_vdpa_event.c
+SRCS-$(CONFIG_RTE_LIBRTE_MLX5_VDPA_PMD) += mlx5_vdpa_virtq.c
 
 # Basic CFLAGS.
 CFLAGS += -O3
index 3da0d76..732ddce 100644 (file)
@@ -14,6 +14,7 @@ sources = files(
        'mlx5_vdpa.c',
        'mlx5_vdpa_mem.c',
        'mlx5_vdpa_event.c',
+       'mlx5_vdpa_virtq.c',
 )
 cflags_options = [
        '-std=c11',
index 16107cf..d76c3aa 100644 (file)
@@ -230,6 +230,7 @@ mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
                goto error;
        }
        SLIST_INIT(&priv->mr_list);
+       SLIST_INIT(&priv->virtq_list);
        pthread_mutex_lock(&priv_list_lock);
        TAILQ_INSERT_TAIL(&priv_list, priv, next);
        pthread_mutex_unlock(&priv_list_lock);
index 6282635..9284420 100644 (file)
@@ -59,6 +59,19 @@ struct mlx5_vdpa_query_mr {
        int is_indirect;
 };
 
+struct mlx5_vdpa_virtq {
+       SLIST_ENTRY(mlx5_vdpa_virtq) next;
+       uint16_t index;
+       uint16_t vq_size;
+       struct mlx5_devx_obj *virtq;
+       struct mlx5_vdpa_event_qp eqp;
+       struct {
+               struct mlx5dv_devx_umem *obj;
+               void *buf;
+               uint32_t size;
+       } umems[3];
+};
+
 struct mlx5_vdpa_priv {
        TAILQ_ENTRY(mlx5_vdpa_priv) next;
        int id; /* vDPA device id. */
@@ -75,6 +88,10 @@ struct mlx5_vdpa_priv {
        struct mlx5dv_devx_event_channel *eventc;
        struct mlx5dv_devx_uar *uar;
        struct rte_intr_handle intr_handle;
+       struct mlx5_devx_obj *td;
+       struct mlx5_devx_obj *tis;
+       uint16_t nr_virtqs;
+       SLIST_HEAD(virtq_list, mlx5_vdpa_virtq) virtq_list;
        SLIST_HEAD(mr_list, mlx5_vdpa_query_mr) mr_list;
 };
 
@@ -152,4 +169,23 @@ int mlx5_vdpa_cqe_event_setup(struct mlx5_vdpa_priv *priv);
  */
 void mlx5_vdpa_cqe_event_unset(struct mlx5_vdpa_priv *priv);
 
+/**
+ * Release a virtq and all its related resources.
+ *
+ * @param[in] priv
+ *   The vdpa driver private structure.
+ */
+void mlx5_vdpa_virtqs_release(struct mlx5_vdpa_priv *priv);
+
+/**
+ * Create all the HW virtqs resources and all their related resources.
+ *
+ * @param[in] priv
+ *   The vdpa driver private structure.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+int mlx5_vdpa_virtqs_prepare(struct mlx5_vdpa_priv *priv);
+
 #endif /* RTE_PMD_MLX5_VDPA_H_ */
diff --git a/drivers/vdpa/mlx5/mlx5_vdpa_virtq.c b/drivers/vdpa/mlx5/mlx5_vdpa_virtq.c
new file mode 100644 (file)
index 0000000..781bccf
--- /dev/null
@@ -0,0 +1,212 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright 2019 Mellanox Technologies, Ltd
+ */
+#include <string.h>
+
+#include <rte_malloc.h>
+#include <rte_errno.h>
+
+#include <mlx5_common.h>
+
+#include "mlx5_vdpa_utils.h"
+#include "mlx5_vdpa.h"
+
+
+static int
+mlx5_vdpa_virtq_unset(struct mlx5_vdpa_virtq *virtq)
+{
+       int i;
+
+       if (virtq->virtq) {
+               claim_zero(mlx5_devx_cmd_destroy(virtq->virtq));
+               virtq->virtq = NULL;
+       }
+       for (i = 0; i < 3; ++i) {
+               if (virtq->umems[i].obj)
+                       claim_zero(mlx5_glue->devx_umem_dereg
+                                                        (virtq->umems[i].obj));
+               if (virtq->umems[i].buf)
+                       rte_free(virtq->umems[i].buf);
+       }
+       memset(&virtq->umems, 0, sizeof(virtq->umems));
+       if (virtq->eqp.fw_qp)
+               mlx5_vdpa_event_qp_destroy(&virtq->eqp);
+       return 0;
+}
+
+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);
+       if (priv->tis) {
+               claim_zero(mlx5_devx_cmd_destroy(priv->tis));
+               priv->tis = NULL;
+       }
+       if (priv->td) {
+               claim_zero(mlx5_devx_cmd_destroy(priv->td));
+               priv->td = NULL;
+       }
+}
+
+static uint64_t
+mlx5_vdpa_hva_to_gpa(struct rte_vhost_memory *mem, uint64_t hva)
+{
+       struct rte_vhost_mem_region *reg;
+       uint32_t i;
+       uint64_t gpa = 0;
+
+       for (i = 0; i < mem->nregions; i++) {
+               reg = &mem->regions[i];
+               if (hva >= reg->host_user_addr &&
+                   hva < reg->host_user_addr + reg->size) {
+                       gpa = hva - reg->host_user_addr + reg->guest_phys_addr;
+                       break;
+               }
+       }
+       return gpa;
+}
+
+static int
+mlx5_vdpa_virtq_setup(struct mlx5_vdpa_priv *priv,
+                     struct mlx5_vdpa_virtq *virtq, int index)
+{
+       struct rte_vhost_vring vq;
+       struct mlx5_devx_virtq_attr attr = {0};
+       uint64_t gpa;
+       int ret;
+       int i;
+       uint16_t last_avail_idx;
+       uint16_t last_used_idx;
+
+       ret = rte_vhost_get_vhost_vring(priv->vid, index, &vq);
+       if (ret)
+               return -1;
+       virtq->index = index;
+       virtq->vq_size = vq.size;
+       /*
+        * No need event QPs creation when the guest in poll mode or when the
+        * capability allows it.
+        */
+       attr.event_mode = vq.callfd != -1 || !(priv->caps.event_mode & (1 <<
+                                              MLX5_VIRTQ_EVENT_MODE_NO_MSIX)) ?
+                                                     MLX5_VIRTQ_EVENT_MODE_QP :
+                                                 MLX5_VIRTQ_EVENT_MODE_NO_MSIX;
+       if (attr.event_mode == MLX5_VIRTQ_EVENT_MODE_QP) {
+               ret = mlx5_vdpa_event_qp_create(priv, vq.size, vq.callfd,
+                                               &virtq->eqp);
+               if (ret) {
+                       DRV_LOG(ERR, "Failed to create event QPs for virtq %d.",
+                               index);
+                       return -1;
+               }
+               attr.qp_id = virtq->eqp.fw_qp->id;
+       } else {
+               DRV_LOG(INFO, "Virtq %d is, for sure, working by poll mode, no"
+                       " need event QPs and event mechanism.", index);
+       }
+       /* Setup 3 UMEMs for each virtq. */
+       for (i = 0; i < 3; ++i) {
+               virtq->umems[i].size = priv->caps.umems[i].a * vq.size +
+                                                         priv->caps.umems[i].b;
+               virtq->umems[i].buf = rte_zmalloc(__func__,
+                                                 virtq->umems[i].size, 4096);
+               if (!virtq->umems[i].buf) {
+                       DRV_LOG(ERR, "Cannot allocate umem %d memory for virtq"
+                               " %u.", i, index);
+                       goto error;
+               }
+               virtq->umems[i].obj = mlx5_glue->devx_umem_reg(priv->ctx,
+                                                       virtq->umems[i].buf,
+                                                       virtq->umems[i].size,
+                                                       IBV_ACCESS_LOCAL_WRITE);
+               if (!virtq->umems[i].obj) {
+                       DRV_LOG(ERR, "Failed to register umem %d for virtq %u.",
+                               i, index);
+                       goto error;
+               }
+               attr.umems[i].id = virtq->umems[i].obj->umem_id;
+               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;
+       }
+       attr.available_addr = gpa;
+       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);
+       attr.hw_available_index = last_avail_idx;
+       attr.hw_used_index = last_used_idx;
+       attr.q_size = vq.size;
+       attr.mkey = priv->gpa_mkey_index;
+       attr.tis_id = priv->tis->id;
+       attr.queue_index = index;
+       virtq->virtq = mlx5_devx_cmd_create_virtq(priv->ctx, &attr);
+       if (!virtq->virtq)
+               goto error;
+       return 0;
+error:
+       mlx5_vdpa_virtq_unset(virtq);
+       return -1;
+}
+
+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);
+
+       priv->td = mlx5_devx_cmd_create_td(priv->ctx);
+       if (!priv->td) {
+               DRV_LOG(ERR, "Failed to create transport domain.");
+               return -rte_errno;
+       }
+       tis_attr.transport_domain = priv->td->id;
+       priv->tis = mlx5_devx_cmd_create_tis(priv->ctx, &tis_attr);
+       if (!priv->tis) {
+               DRV_LOG(ERR, "Failed to create TIS.");
+               goto error;
+       }
+       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);
+                       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;
+}