drivers/vdpa/mlx5/mlx5_vdpa_virtq.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2019 Mellanox Technologies, Ltd
   3  */
   4 #include <string.h>
   5 #include <unistd.h>
   6 #include <sys/mman.h>
   7
   8 #include <rte_malloc.h>
   9 #include <rte_errno.h>
  10 #include <rte_io.h>
  11
  12 #include <mlx5_common.h>
  13
  14 #include "mlx5_vdpa_utils.h"
  15 #include "mlx5_vdpa.h"
  16
  17
  18 static void
  19 mlx5_vdpa_virtq_handler(void *cb_arg)
  20 {
  21         struct mlx5_vdpa_virtq *virtq = cb_arg;
  22         struct mlx5_vdpa_priv *priv = virtq->priv;
  23         uint64_t buf;
  24         int nbytes;
  25
  26         do {
  27                 nbytes = read(virtq->intr_handle.fd, &buf, 8);
  28                 if (nbytes < 0) {
  29                         if (errno == EINTR ||
  30                             errno == EWOULDBLOCK ||
  31                             errno == EAGAIN)
  32                                 continue;
  33                         DRV_LOG(ERR,  "Failed to read kickfd of virtq %d: %s",
  34                                 virtq->index, strerror(errno));
  35                 }
  36                 break;
  37         } while (1);
  38         rte_write32(virtq->index, priv->virtq_db_addr);
  39         DRV_LOG(DEBUG, "Ring virtq %u doorbell.", virtq->index);
  40 }
  41
  42 static int
  43 mlx5_vdpa_virtq_unset(struct mlx5_vdpa_virtq *virtq)
  44 {
  45         unsigned int i;
  46         int retries = MLX5_VDPA_INTR_RETRIES;
  47         int ret = -EAGAIN;
  48
  49         if (virtq->intr_handle.fd != -1) {
  50                 while (retries-- && ret == -EAGAIN) {
  51                         ret = rte_intr_callback_unregister(&virtq->intr_handle,
  52                                                         mlx5_vdpa_virtq_handler,
  53                                                         virtq);
  54                         if (ret == -EAGAIN) {
  55                                 DRV_LOG(DEBUG, "Try again to unregister fd %d "
  56                                         "of virtq %d interrupt, retries = %d.",
  57                                         virtq->intr_handle.fd,
  58                                         (int)virtq->index, retries);
  59                                 usleep(MLX5_VDPA_INTR_RETRIES_USEC);
  60                         }
  61                 }
  62                 virtq->intr_handle.fd = -1;
  63         }
  64         if (virtq->virtq) {
  65                 claim_zero(mlx5_devx_cmd_destroy(virtq->virtq));
  66                 virtq->virtq = NULL;
  67         }
  68         for (i = 0; i < RTE_DIM(virtq->umems); ++i) {
  69                 if (virtq->umems[i].obj)
  70                         claim_zero(mlx5_glue->devx_umem_dereg
  71                                                          (virtq->umems[i].obj));
  72                 if (virtq->umems[i].buf)
  73                         rte_free(virtq->umems[i].buf);
  74         }
  75         memset(&virtq->umems, 0, sizeof(virtq->umems));
  76         if (virtq->eqp.fw_qp)
  77                 mlx5_vdpa_event_qp_destroy(&virtq->eqp);
  78         return 0;
  79 }
  80
  81 void
  82 mlx5_vdpa_virtqs_release(struct mlx5_vdpa_priv *priv)
  83 {
  84         struct mlx5_vdpa_virtq *entry;
  85         struct mlx5_vdpa_virtq *next;
  86
  87         entry = SLIST_FIRST(&priv->virtq_list);
  88         while (entry) {
  89                 next = SLIST_NEXT(entry, next);
  90                 mlx5_vdpa_virtq_unset(entry);
  91                 SLIST_REMOVE(&priv->virtq_list, entry, mlx5_vdpa_virtq, next);
  92                 rte_free(entry);
  93                 entry = next;
  94         }
  95         SLIST_INIT(&priv->virtq_list);
  96         if (priv->tis) {
  97                 claim_zero(mlx5_devx_cmd_destroy(priv->tis));
  98                 priv->tis = NULL;
  99         }
 100         if (priv->td) {
 101                 claim_zero(mlx5_devx_cmd_destroy(priv->td));
 102                 priv->td = NULL;
 103         }
 104         if (priv->virtq_db_addr) {
 105                 claim_zero(munmap(priv->virtq_db_addr, priv->var->length));
 106                 priv->virtq_db_addr = NULL;
 107         }
 108         priv->features = 0;
 109 }
 110
 111 int
 112 mlx5_vdpa_virtq_modify(struct mlx5_vdpa_virtq *virtq, int state)
 113 {
 114         struct mlx5_devx_virtq_attr attr = {
 115                         .type = MLX5_VIRTQ_MODIFY_TYPE_STATE,
 116                         .state = state ? MLX5_VIRTQ_STATE_RDY :
 117                                          MLX5_VIRTQ_STATE_SUSPEND,
 118                         .queue_index = virtq->index,
 119         };
 120
 121         return mlx5_devx_cmd_modify_virtq(virtq->virtq, &attr);
 122 }
 123
 124 static uint64_t
 125 mlx5_vdpa_hva_to_gpa(struct rte_vhost_memory *mem, uint64_t hva)
 126 {
 127         struct rte_vhost_mem_region *reg;
 128         uint32_t i;
 129         uint64_t gpa = 0;
 130
 131         for (i = 0; i < mem->nregions; i++) {
 132                 reg = &mem->regions[i];
 133                 if (hva >= reg->host_user_addr &&
 134                     hva < reg->host_user_addr + reg->size) {
 135                         gpa = hva - reg->host_user_addr + reg->guest_phys_addr;
 136                         break;
 137                 }
 138         }
 139         return gpa;
 140 }
 141
 142 static int
 143 mlx5_vdpa_virtq_setup(struct mlx5_vdpa_priv *priv,
 144                       struct mlx5_vdpa_virtq *virtq, int index)
 145 {
 146         struct rte_vhost_vring vq;
 147         struct mlx5_devx_virtq_attr attr = {0};
 148         uint64_t gpa;
 149         int ret;
 150         unsigned int i;
 151         uint16_t last_avail_idx;
 152         uint16_t last_used_idx;
 153
 154         ret = rte_vhost_get_vhost_vring(priv->vid, index, &vq);
 155         if (ret)
 156                 return -1;
 157         virtq->index = index;
 158         virtq->vq_size = vq.size;
 159         attr.tso_ipv4 = !!(priv->features & (1ULL << VIRTIO_NET_F_HOST_TSO4));
 160         attr.tso_ipv6 = !!(priv->features & (1ULL << VIRTIO_NET_F_HOST_TSO6));
 161         attr.tx_csum = !!(priv->features & (1ULL << VIRTIO_NET_F_CSUM));
 162         attr.rx_csum = !!(priv->features & (1ULL << VIRTIO_NET_F_GUEST_CSUM));
 163         attr.virtio_version_1_0 = !!(priv->features & (1ULL <<
 164                                                         VIRTIO_F_VERSION_1));
 165         attr.type = (priv->features & (1ULL << VIRTIO_F_RING_PACKED)) ?
 166                         MLX5_VIRTQ_TYPE_PACKED : MLX5_VIRTQ_TYPE_SPLIT;
 167         /*
 168          * No need event QPs creation when the guest in poll mode or when the
 169          * capability allows it.
 170          */
 171         attr.event_mode = vq.callfd != -1 || !(priv->caps.event_mode & (1 <<
 172                                                MLX5_VIRTQ_EVENT_MODE_NO_MSIX)) ?
 173                                                       MLX5_VIRTQ_EVENT_MODE_QP :
 174                                                   MLX5_VIRTQ_EVENT_MODE_NO_MSIX;
 175         if (attr.event_mode == MLX5_VIRTQ_EVENT_MODE_QP) {
 176                 ret = mlx5_vdpa_event_qp_create(priv, vq.size, vq.callfd,
 177                                                 &virtq->eqp);
 178                 if (ret) {
 179                         DRV_LOG(ERR, "Failed to create event QPs for virtq %d.",
 180                                 index);
 181                         return -1;
 182                 }
 183                 attr.qp_id = virtq->eqp.fw_qp->id;
 184         } else {
 185                 DRV_LOG(INFO, "Virtq %d is, for sure, working by poll mode, no"
 186                         " need event QPs and event mechanism.", index);
 187         }
 188         /* Setup 3 UMEMs for each virtq. */
 189         for (i = 0; i < RTE_DIM(virtq->umems); ++i) {
 190                 virtq->umems[i].size = priv->caps.umems[i].a * vq.size +
 191                                                           priv->caps.umems[i].b;
 192                 virtq->umems[i].buf = rte_zmalloc(__func__,
 193                                                   virtq->umems[i].size, 4096);
 194                 if (!virtq->umems[i].buf) {
 195                         DRV_LOG(ERR, "Cannot allocate umem %d memory for virtq"
 196                                 " %u.", i, index);
 197                         goto error;
 198                 }
 199                 virtq->umems[i].obj = mlx5_glue->devx_umem_reg(priv->ctx,
 200                                                         virtq->umems[i].buf,
 201                                                         virtq->umems[i].size,
 202                                                         IBV_ACCESS_LOCAL_WRITE);
 203                 if (!virtq->umems[i].obj) {
 204                         DRV_LOG(ERR, "Failed to register umem %d for virtq %u.",
 205                                 i, index);
 206                         goto error;
 207                 }
 208                 attr.umems[i].id = virtq->umems[i].obj->umem_id;
 209                 attr.umems[i].offset = 0;
 210                 attr.umems[i].size = virtq->umems[i].size;
 211         }
 212         if (attr.type == MLX5_VIRTQ_TYPE_SPLIT) {
 213                 gpa = mlx5_vdpa_hva_to_gpa(priv->vmem,
 214                                            (uint64_t)(uintptr_t)vq.desc);
 215                 if (!gpa) {
 216                         DRV_LOG(ERR, "Failed to get descriptor ring GPA.");
 217                         goto error;
 218                 }
 219                 attr.desc_addr = gpa;
 220                 gpa = mlx5_vdpa_hva_to_gpa(priv->vmem,
 221                                            (uint64_t)(uintptr_t)vq.used);
 222                 if (!gpa) {
 223                         DRV_LOG(ERR, "Failed to get GPA for used ring.");
 224                         goto error;
 225                 }
 226                 attr.used_addr = gpa;
 227                 gpa = mlx5_vdpa_hva_to_gpa(priv->vmem,
 228                                            (uint64_t)(uintptr_t)vq.avail);
 229                 if (!gpa) {
 230                         DRV_LOG(ERR, "Failed to get GPA for available ring.");
 231                         goto error;
 232                 }
 233                 attr.available_addr = gpa;
 234         }
 235         ret = rte_vhost_get_vring_base(priv->vid, index, &last_avail_idx,
 236                                  &last_used_idx);
 237         if (ret) {
 238                 last_avail_idx = 0;
 239                 last_used_idx = 0;
 240                 DRV_LOG(WARNING, "Couldn't get vring base, idx are set to 0");
 241         } else {
 242                 DRV_LOG(INFO, "vid %d: Init last_avail_idx=%d, last_used_idx=%d for "
 243                                 "virtq %d.", priv->vid, last_avail_idx,
 244                                 last_used_idx, index);
 245         }
 246         attr.hw_available_index = last_avail_idx;
 247         attr.hw_used_index = last_used_idx;
 248         attr.q_size = vq.size;
 249         attr.mkey = priv->gpa_mkey_index;
 250         attr.tis_id = priv->tis->id;
 251         attr.queue_index = index;
 252         virtq->virtq = mlx5_devx_cmd_create_virtq(priv->ctx, &attr);
 253         virtq->priv = priv;
 254         if (!virtq->virtq)
 255                 goto error;
 256         if (mlx5_vdpa_virtq_modify(virtq, 1))
 257                 goto error;
 258         virtq->enable = 1;
 259         virtq->priv = priv;
 260         /* Be sure notifications are not missed during configuration. */
 261         claim_zero(rte_vhost_enable_guest_notification(priv->vid, index, 1));
 262         rte_write32(virtq->index, priv->virtq_db_addr);
 263         /* Setup doorbell mapping. */
 264         virtq->intr_handle.fd = vq.kickfd;
 265         if (virtq->intr_handle.fd == -1) {
 266                 DRV_LOG(WARNING, "Virtq %d kickfd is invalid.", index);
 267                 if (!priv->direct_notifier) {
 268                         DRV_LOG(ERR, "Virtq %d cannot be notified.", index);
 269                         goto error;
 270                 }
 271         } else {
 272                 virtq->intr_handle.type = RTE_INTR_HANDLE_EXT;
 273                 if (rte_intr_callback_register(&virtq->intr_handle,
 274                                                mlx5_vdpa_virtq_handler,
 275                                                virtq)) {
 276                         virtq->intr_handle.fd = -1;
 277                         DRV_LOG(ERR, "Failed to register virtq %d interrupt.",
 278                                 index);
 279                         goto error;
 280                 } else {
 281                         DRV_LOG(DEBUG, "Register fd %d interrupt for virtq %d.",
 282                                 virtq->intr_handle.fd, index);
 283                 }
 284         }
 285         return 0;
 286 error:
 287         mlx5_vdpa_virtq_unset(virtq);
 288         return -1;
 289 }
 290
 291 static int
 292 mlx5_vdpa_features_validate(struct mlx5_vdpa_priv *priv)
 293 {
 294         if (priv->features & (1ULL << VIRTIO_F_RING_PACKED)) {
 295                 if (!(priv->caps.virtio_queue_type & (1 <<
 296                                                      MLX5_VIRTQ_TYPE_PACKED))) {
 297                         DRV_LOG(ERR, "Failed to configur PACKED mode for vdev "
 298                                 "%d - it was not reported by HW/driver"
 299                                 " capability.", priv->vid);
 300                         return -ENOTSUP;
 301                 }
 302         }
 303         if (priv->features & (1ULL << VIRTIO_NET_F_HOST_TSO4)) {
 304                 if (!priv->caps.tso_ipv4) {
 305                         DRV_LOG(ERR, "Failed to enable TSO4 for vdev %d - TSO4"
 306                                 " was not reported by HW/driver capability.",
 307                                 priv->vid);
 308                         return -ENOTSUP;
 309                 }
 310         }
 311         if (priv->features & (1ULL << VIRTIO_NET_F_HOST_TSO6)) {
 312                 if (!priv->caps.tso_ipv6) {
 313                         DRV_LOG(ERR, "Failed to enable TSO6 for vdev %d - TSO6"
 314                                 " was not reported by HW/driver capability.",
 315                                 priv->vid);
 316                         return -ENOTSUP;
 317                 }
 318         }
 319         if (priv->features & (1ULL << VIRTIO_NET_F_CSUM)) {
 320                 if (!priv->caps.tx_csum) {
 321                         DRV_LOG(ERR, "Failed to enable CSUM for vdev %d - CSUM"
 322                                 " was not reported by HW/driver capability.",
 323                                 priv->vid);
 324                         return -ENOTSUP;
 325                 }
 326         }
 327         if (priv->features & (1ULL << VIRTIO_NET_F_GUEST_CSUM)) {
 328                 if (!priv->caps.rx_csum) {
 329                         DRV_LOG(ERR, "Failed to enable GUEST CSUM for vdev %d"
 330                                 " GUEST CSUM was not reported by HW/driver "
 331                                 "capability.", priv->vid);
 332                         return -ENOTSUP;
 333                 }
 334         }
 335         if (priv->features & (1ULL << VIRTIO_F_VERSION_1)) {
 336                 if (!priv->caps.virtio_version_1_0) {
 337                         DRV_LOG(ERR, "Failed to enable version 1 for vdev %d "
 338                                 "version 1 was not reported by HW/driver"
 339                                 " capability.", priv->vid);
 340                         return -ENOTSUP;
 341                 }
 342         }
 343         return 0;
 344 }
 345
 346 int
 347 mlx5_vdpa_virtqs_prepare(struct mlx5_vdpa_priv *priv)
 348 {
 349         struct mlx5_devx_tis_attr tis_attr = {0};
 350         struct mlx5_vdpa_virtq *virtq;
 351         uint32_t i;
 352         uint16_t nr_vring = rte_vhost_get_vring_num(priv->vid);
 353         int ret = rte_vhost_get_negotiated_features(priv->vid, &priv->features);
 354
 355         if (ret || mlx5_vdpa_features_validate(priv)) {
 356                 DRV_LOG(ERR, "Failed to configure negotiated features.");
 357                 return -1;
 358         }
 359         /* Always map the entire page. */
 360         priv->virtq_db_addr = mmap(NULL, priv->var->length, PROT_READ |
 361                                    PROT_WRITE, MAP_SHARED, priv->ctx->cmd_fd,
 362                                    priv->var->mmap_off);
 363         if (priv->virtq_db_addr == MAP_FAILED) {
 364                 DRV_LOG(ERR, "Failed to map doorbell page %u.", errno);
 365                 priv->virtq_db_addr = NULL;
 366                 goto error;
 367         } else {
 368                 DRV_LOG(DEBUG, "VAR address of doorbell mapping is %p.",
 369                         priv->virtq_db_addr);
 370         }
 371         priv->td = mlx5_devx_cmd_create_td(priv->ctx);
 372         if (!priv->td) {
 373                 DRV_LOG(ERR, "Failed to create transport domain.");
 374                 return -rte_errno;
 375         }
 376         tis_attr.transport_domain = priv->td->id;
 377         priv->tis = mlx5_devx_cmd_create_tis(priv->ctx, &tis_attr);
 378         if (!priv->tis) {
 379                 DRV_LOG(ERR, "Failed to create TIS.");
 380                 goto error;
 381         }
 382         for (i = 0; i < nr_vring; i++) {
 383                 virtq = rte_zmalloc(__func__, sizeof(*virtq), 0);
 384                 if (!virtq || mlx5_vdpa_virtq_setup(priv, virtq, i)) {
 385                         if (virtq)
 386                                 rte_free(virtq);
 387                         goto error;
 388                 }
 389                 SLIST_INSERT_HEAD(&priv->virtq_list, virtq, next);
 390         }
 391         priv->nr_virtqs = nr_vring;
 392         return 0;
 393 error:
 394         mlx5_vdpa_virtqs_release(priv);
 395         return -1;
 396 }