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         }
  63         if (virtq->virtq)
  64                 claim_zero(mlx5_devx_cmd_destroy(virtq->virtq));
  65         for (i = 0; i < RTE_DIM(virtq->umems); ++i) {
  66                 if (virtq->umems[i].obj)
  67                         claim_zero(mlx5_glue->devx_umem_dereg
  68                                                          (virtq->umems[i].obj));
  69                 if (virtq->umems[i].buf)
  70                         rte_free(virtq->umems[i].buf);
  71         }
  72         if (virtq->eqp.fw_qp)
  73                 mlx5_vdpa_event_qp_destroy(&virtq->eqp);
  74         memset(virtq, 0, sizeof(*virtq));
  75         virtq->intr_handle.fd = -1;
  76         return 0;
  77 }
  78
  79 void
  80 mlx5_vdpa_virtqs_release(struct mlx5_vdpa_priv *priv)
  81 {
  82         int i;
  83
  84         for (i = 0; i < priv->nr_virtqs; i++)
  85                 mlx5_vdpa_virtq_unset(&priv->virtqs[i]);
  86         if (priv->tis) {
  87                 claim_zero(mlx5_devx_cmd_destroy(priv->tis));
  88                 priv->tis = NULL;
  89         }
  90         if (priv->td) {
  91                 claim_zero(mlx5_devx_cmd_destroy(priv->td));
  92                 priv->td = NULL;
  93         }
  94         if (priv->virtq_db_addr) {
  95                 claim_zero(munmap(priv->virtq_db_addr, priv->var->length));
  96                 priv->virtq_db_addr = NULL;
  97         }
  98         priv->features = 0;
  99         priv->nr_virtqs = 0;
 100 }
 101
 102 int
 103 mlx5_vdpa_virtq_modify(struct mlx5_vdpa_virtq *virtq, int state)
 104 {
 105         struct mlx5_devx_virtq_attr attr = {
 106                         .type = MLX5_VIRTQ_MODIFY_TYPE_STATE,
 107                         .state = state ? MLX5_VIRTQ_STATE_RDY :
 108                                          MLX5_VIRTQ_STATE_SUSPEND,
 109                         .queue_index = virtq->index,
 110         };
 111
 112         return mlx5_devx_cmd_modify_virtq(virtq->virtq, &attr);
 113 }
 114
 115 int
 116 mlx5_vdpa_virtq_stop(struct mlx5_vdpa_priv *priv, int index)
 117 {
 118         struct mlx5_devx_virtq_attr attr = {0};
 119         struct mlx5_vdpa_virtq *virtq = &priv->virtqs[index];
 120         int ret = mlx5_vdpa_virtq_modify(virtq, 0);
 121
 122         if (ret)
 123                 return -1;
 124         if (mlx5_devx_cmd_query_virtq(virtq->virtq, &attr)) {
 125                 DRV_LOG(ERR, "Failed to query virtq %d.", index);
 126                 return -1;
 127         }
 128         DRV_LOG(INFO, "Query vid %d vring %d: hw_available_idx=%d, "
 129                 "hw_used_index=%d", priv->vid, index,
 130                 attr.hw_available_index, attr.hw_used_index);
 131         ret = rte_vhost_set_vring_base(priv->vid, index,
 132                                        attr.hw_available_index,
 133                                        attr.hw_used_index);
 134         if (ret) {
 135                 DRV_LOG(ERR, "Failed to set virtq %d base.", index);
 136                 return -1;
 137         }
 138         return 0;
 139 }
 140
 141 static uint64_t
 142 mlx5_vdpa_hva_to_gpa(struct rte_vhost_memory *mem, uint64_t hva)
 143 {
 144         struct rte_vhost_mem_region *reg;
 145         uint32_t i;
 146         uint64_t gpa = 0;
 147
 148         for (i = 0; i < mem->nregions; i++) {
 149                 reg = &mem->regions[i];
 150                 if (hva >= reg->host_user_addr &&
 151                     hva < reg->host_user_addr + reg->size) {
 152                         gpa = hva - reg->host_user_addr + reg->guest_phys_addr;
 153                         break;
 154                 }
 155         }
 156         return gpa;
 157 }
 158
 159 static int
 160 mlx5_vdpa_virtq_setup(struct mlx5_vdpa_priv *priv, int index)
 161 {
 162         struct mlx5_vdpa_virtq *virtq = &priv->virtqs[index];
 163         struct rte_vhost_vring vq;
 164         struct mlx5_devx_virtq_attr attr = {0};
 165         uint64_t gpa;
 166         int ret;
 167         unsigned int i;
 168         uint16_t last_avail_idx;
 169         uint16_t last_used_idx;
 170
 171         ret = rte_vhost_get_vhost_vring(priv->vid, index, &vq);
 172         if (ret)
 173                 return -1;
 174         virtq->index = index;
 175         virtq->vq_size = vq.size;
 176         attr.tso_ipv4 = !!(priv->features & (1ULL << VIRTIO_NET_F_HOST_TSO4));
 177         attr.tso_ipv6 = !!(priv->features & (1ULL << VIRTIO_NET_F_HOST_TSO6));
 178         attr.tx_csum = !!(priv->features & (1ULL << VIRTIO_NET_F_CSUM));
 179         attr.rx_csum = !!(priv->features & (1ULL << VIRTIO_NET_F_GUEST_CSUM));
 180         attr.virtio_version_1_0 = !!(priv->features & (1ULL <<
 181                                                         VIRTIO_F_VERSION_1));
 182         attr.type = (priv->features & (1ULL << VIRTIO_F_RING_PACKED)) ?
 183                         MLX5_VIRTQ_TYPE_PACKED : MLX5_VIRTQ_TYPE_SPLIT;
 184         /*
 185          * No need event QPs creation when the guest in poll mode or when the
 186          * capability allows it.
 187          */
 188         attr.event_mode = vq.callfd != -1 || !(priv->caps.event_mode & (1 <<
 189                                                MLX5_VIRTQ_EVENT_MODE_NO_MSIX)) ?
 190                                                       MLX5_VIRTQ_EVENT_MODE_QP :
 191                                                   MLX5_VIRTQ_EVENT_MODE_NO_MSIX;
 192         if (attr.event_mode == MLX5_VIRTQ_EVENT_MODE_QP) {
 193                 ret = mlx5_vdpa_event_qp_create(priv, vq.size, vq.callfd,
 194                                                 &virtq->eqp);
 195                 if (ret) {
 196                         DRV_LOG(ERR, "Failed to create event QPs for virtq %d.",
 197                                 index);
 198                         return -1;
 199                 }
 200                 attr.qp_id = virtq->eqp.fw_qp->id;
 201         } else {
 202                 DRV_LOG(INFO, "Virtq %d is, for sure, working by poll mode, no"
 203                         " need event QPs and event mechanism.", index);
 204         }
 205         /* Setup 3 UMEMs for each virtq. */
 206         for (i = 0; i < RTE_DIM(virtq->umems); ++i) {
 207                 virtq->umems[i].size = priv->caps.umems[i].a * vq.size +
 208                                                           priv->caps.umems[i].b;
 209                 virtq->umems[i].buf = rte_zmalloc(__func__,
 210                                                   virtq->umems[i].size, 4096);
 211                 if (!virtq->umems[i].buf) {
 212                         DRV_LOG(ERR, "Cannot allocate umem %d memory for virtq"
 213                                 " %u.", i, index);
 214                         goto error;
 215                 }
 216                 virtq->umems[i].obj = mlx5_glue->devx_umem_reg(priv->ctx,
 217                                                         virtq->umems[i].buf,
 218                                                         virtq->umems[i].size,
 219                                                         IBV_ACCESS_LOCAL_WRITE);
 220                 if (!virtq->umems[i].obj) {
 221                         DRV_LOG(ERR, "Failed to register umem %d for virtq %u.",
 222                                 i, index);
 223                         goto error;
 224                 }
 225                 attr.umems[i].id = virtq->umems[i].obj->umem_id;
 226                 attr.umems[i].offset = 0;
 227                 attr.umems[i].size = virtq->umems[i].size;
 228         }
 229         if (attr.type == MLX5_VIRTQ_TYPE_SPLIT) {
 230                 gpa = mlx5_vdpa_hva_to_gpa(priv->vmem,
 231                                            (uint64_t)(uintptr_t)vq.desc);
 232                 if (!gpa) {
 233                         DRV_LOG(ERR, "Failed to get descriptor ring GPA.");
 234                         goto error;
 235                 }
 236                 attr.desc_addr = gpa;
 237                 gpa = mlx5_vdpa_hva_to_gpa(priv->vmem,
 238                                            (uint64_t)(uintptr_t)vq.used);
 239                 if (!gpa) {
 240                         DRV_LOG(ERR, "Failed to get GPA for used ring.");
 241                         goto error;
 242                 }
 243                 attr.used_addr = gpa;
 244                 gpa = mlx5_vdpa_hva_to_gpa(priv->vmem,
 245                                            (uint64_t)(uintptr_t)vq.avail);
 246                 if (!gpa) {
 247                         DRV_LOG(ERR, "Failed to get GPA for available ring.");
 248                         goto error;
 249                 }
 250                 attr.available_addr = gpa;
 251         }
 252         ret = rte_vhost_get_vring_base(priv->vid, index, &last_avail_idx,
 253                                  &last_used_idx);
 254         if (ret) {
 255                 last_avail_idx = 0;
 256                 last_used_idx = 0;
 257                 DRV_LOG(WARNING, "Couldn't get vring base, idx are set to 0");
 258         } else {
 259                 DRV_LOG(INFO, "vid %d: Init last_avail_idx=%d, last_used_idx=%d for "
 260                                 "virtq %d.", priv->vid, last_avail_idx,
 261                                 last_used_idx, index);
 262         }
 263         attr.hw_available_index = last_avail_idx;
 264         attr.hw_used_index = last_used_idx;
 265         attr.q_size = vq.size;
 266         attr.mkey = priv->gpa_mkey_index;
 267         attr.tis_id = priv->tis->id;
 268         attr.queue_index = index;
 269         virtq->virtq = mlx5_devx_cmd_create_virtq(priv->ctx, &attr);
 270         virtq->priv = priv;
 271         if (!virtq->virtq)
 272                 goto error;
 273         if (mlx5_vdpa_virtq_modify(virtq, 1))
 274                 goto error;
 275         virtq->enable = 1;
 276         virtq->priv = priv;
 277         /* Be sure notifications are not missed during configuration. */
 278         claim_zero(rte_vhost_enable_guest_notification(priv->vid, index, 1));
 279         rte_write32(virtq->index, priv->virtq_db_addr);
 280         /* Setup doorbell mapping. */
 281         virtq->intr_handle.fd = vq.kickfd;
 282         if (virtq->intr_handle.fd == -1) {
 283                 DRV_LOG(WARNING, "Virtq %d kickfd is invalid.", index);
 284                 if (!priv->direct_notifier) {
 285                         DRV_LOG(ERR, "Virtq %d cannot be notified.", index);
 286                         goto error;
 287                 }
 288         } else {
 289                 virtq->intr_handle.type = RTE_INTR_HANDLE_EXT;
 290                 if (rte_intr_callback_register(&virtq->intr_handle,
 291                                                mlx5_vdpa_virtq_handler,
 292                                                virtq)) {
 293                         virtq->intr_handle.fd = -1;
 294                         DRV_LOG(ERR, "Failed to register virtq %d interrupt.",
 295                                 index);
 296                         goto error;
 297                 } else {
 298                         DRV_LOG(DEBUG, "Register fd %d interrupt for virtq %d.",
 299                                 virtq->intr_handle.fd, index);
 300                 }
 301         }
 302         return 0;
 303 error:
 304         mlx5_vdpa_virtq_unset(virtq);
 305         return -1;
 306 }
 307
 308 static int
 309 mlx5_vdpa_features_validate(struct mlx5_vdpa_priv *priv)
 310 {
 311         if (priv->features & (1ULL << VIRTIO_F_RING_PACKED)) {
 312                 if (!(priv->caps.virtio_queue_type & (1 <<
 313                                                      MLX5_VIRTQ_TYPE_PACKED))) {
 314                         DRV_LOG(ERR, "Failed to configur PACKED mode for vdev "
 315                                 "%d - it was not reported by HW/driver"
 316                                 " capability.", priv->vid);
 317                         return -ENOTSUP;
 318                 }
 319         }
 320         if (priv->features & (1ULL << VIRTIO_NET_F_HOST_TSO4)) {
 321                 if (!priv->caps.tso_ipv4) {
 322                         DRV_LOG(ERR, "Failed to enable TSO4 for vdev %d - TSO4"
 323                                 " was not reported by HW/driver capability.",
 324                                 priv->vid);
 325                         return -ENOTSUP;
 326                 }
 327         }
 328         if (priv->features & (1ULL << VIRTIO_NET_F_HOST_TSO6)) {
 329                 if (!priv->caps.tso_ipv6) {
 330                         DRV_LOG(ERR, "Failed to enable TSO6 for vdev %d - TSO6"
 331                                 " was not reported by HW/driver capability.",
 332                                 priv->vid);
 333                         return -ENOTSUP;
 334                 }
 335         }
 336         if (priv->features & (1ULL << VIRTIO_NET_F_CSUM)) {
 337                 if (!priv->caps.tx_csum) {
 338                         DRV_LOG(ERR, "Failed to enable CSUM for vdev %d - CSUM"
 339                                 " was not reported by HW/driver capability.",
 340                                 priv->vid);
 341                         return -ENOTSUP;
 342                 }
 343         }
 344         if (priv->features & (1ULL << VIRTIO_NET_F_GUEST_CSUM)) {
 345                 if (!priv->caps.rx_csum) {
 346                         DRV_LOG(ERR, "Failed to enable GUEST CSUM for vdev %d"
 347                                 " GUEST CSUM was not reported by HW/driver "
 348                                 "capability.", priv->vid);
 349                         return -ENOTSUP;
 350                 }
 351         }
 352         if (priv->features & (1ULL << VIRTIO_F_VERSION_1)) {
 353                 if (!priv->caps.virtio_version_1_0) {
 354                         DRV_LOG(ERR, "Failed to enable version 1 for vdev %d "
 355                                 "version 1 was not reported by HW/driver"
 356                                 " capability.", priv->vid);
 357                         return -ENOTSUP;
 358                 }
 359         }
 360         return 0;
 361 }
 362
 363 int
 364 mlx5_vdpa_virtqs_prepare(struct mlx5_vdpa_priv *priv)
 365 {
 366         struct mlx5_devx_tis_attr tis_attr = {0};
 367         uint32_t i;
 368         uint16_t nr_vring = rte_vhost_get_vring_num(priv->vid);
 369         int ret = rte_vhost_get_negotiated_features(priv->vid, &priv->features);
 370
 371         if (ret || mlx5_vdpa_features_validate(priv)) {
 372                 DRV_LOG(ERR, "Failed to configure negotiated features.");
 373                 return -1;
 374         }
 375         if (nr_vring > priv->caps.max_num_virtio_queues * 2) {
 376                 DRV_LOG(ERR, "Do not support more than %d virtqs(%d).",
 377                         (int)priv->caps.max_num_virtio_queues * 2,
 378                         (int)nr_vring);
 379                 return -1;
 380         }
 381         /* Always map the entire page. */
 382         priv->virtq_db_addr = mmap(NULL, priv->var->length, PROT_READ |
 383                                    PROT_WRITE, MAP_SHARED, priv->ctx->cmd_fd,
 384                                    priv->var->mmap_off);
 385         if (priv->virtq_db_addr == MAP_FAILED) {
 386                 DRV_LOG(ERR, "Failed to map doorbell page %u.", errno);
 387                 priv->virtq_db_addr = NULL;
 388                 goto error;
 389         } else {
 390                 DRV_LOG(DEBUG, "VAR address of doorbell mapping is %p.",
 391                         priv->virtq_db_addr);
 392         }
 393         priv->td = mlx5_devx_cmd_create_td(priv->ctx);
 394         if (!priv->td) {
 395                 DRV_LOG(ERR, "Failed to create transport domain.");
 396                 return -rte_errno;
 397         }
 398         tis_attr.transport_domain = priv->td->id;
 399         priv->tis = mlx5_devx_cmd_create_tis(priv->ctx, &tis_attr);
 400         if (!priv->tis) {
 401                 DRV_LOG(ERR, "Failed to create TIS.");
 402                 goto error;
 403         }
 404         priv->nr_virtqs = nr_vring;
 405         for (i = 0; i < nr_vring; i++)
 406                 if (mlx5_vdpa_virtq_setup(priv, i))
 407                         goto error;
 408         return 0;
 409 error:
 410         mlx5_vdpa_virtqs_release(priv);
 411         return -1;
 412 }