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