drivers/net/virtio/virtio_user/vhost_vdpa.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2020 Red Hat Inc.
   3  */
   4
   5 #include <sys/ioctl.h>
   6 #include <sys/types.h>
   7 #include <sys/stat.h>
   8 #include <fcntl.h>
   9 #include <unistd.h>
  10
  11 #include <rte_memory.h>
  12
  13 #include "vhost.h"
  14 #include "virtio_user_dev.h"
  15
  16 /* vhost kernel & vdpa ioctls */
  17 #define VHOST_VIRTIO 0xAF
  18 #define VHOST_GET_FEATURES _IOR(VHOST_VIRTIO, 0x00, __u64)
  19 #define VHOST_SET_FEATURES _IOW(VHOST_VIRTIO, 0x00, __u64)
  20 #define VHOST_SET_OWNER _IO(VHOST_VIRTIO, 0x01)
  21 #define VHOST_RESET_OWNER _IO(VHOST_VIRTIO, 0x02)
  22 #define VHOST_SET_LOG_BASE _IOW(VHOST_VIRTIO, 0x04, __u64)
  23 #define VHOST_SET_LOG_FD _IOW(VHOST_VIRTIO, 0x07, int)
  24 #define VHOST_SET_VRING_NUM _IOW(VHOST_VIRTIO, 0x10, struct vhost_vring_state)
  25 #define VHOST_SET_VRING_ADDR _IOW(VHOST_VIRTIO, 0x11, struct vhost_vring_addr)
  26 #define VHOST_SET_VRING_BASE _IOW(VHOST_VIRTIO, 0x12, struct vhost_vring_state)
  27 #define VHOST_GET_VRING_BASE _IOWR(VHOST_VIRTIO, 0x12, struct vhost_vring_state)
  28 #define VHOST_SET_VRING_KICK _IOW(VHOST_VIRTIO, 0x20, struct vhost_vring_file)
  29 #define VHOST_SET_VRING_CALL _IOW(VHOST_VIRTIO, 0x21, struct vhost_vring_file)
  30 #define VHOST_SET_VRING_ERR _IOW(VHOST_VIRTIO, 0x22, struct vhost_vring_file)
  31 #define VHOST_NET_SET_BACKEND _IOW(VHOST_VIRTIO, 0x30, struct vhost_vring_file)
  32 #define VHOST_VDPA_GET_DEVICE_ID _IOR(VHOST_VIRTIO, 0x70, __u32)
  33 #define VHOST_VDPA_GET_STATUS _IOR(VHOST_VIRTIO, 0x71, __u8)
  34 #define VHOST_VDPA_SET_STATUS _IOW(VHOST_VIRTIO, 0x72, __u8)
  35 #define VHOST_VDPA_SET_VRING_ENABLE _IOW(VHOST_VIRTIO, 0x75, struct vhost_vring_state)
  36 #define VHOST_SET_BACKEND_FEATURES _IOW(VHOST_VIRTIO, 0x25, __u64)
  37 #define VHOST_GET_BACKEND_FEATURES _IOR(VHOST_VIRTIO, 0x26, __u64)
  38
  39 static uint64_t vhost_req_user_to_vdpa[] = {
  40         [VHOST_USER_RESET_OWNER] = VHOST_RESET_OWNER,
  41         [VHOST_USER_SET_VRING_ADDR] = VHOST_SET_VRING_ADDR,
  42         [VHOST_USER_SET_STATUS] = VHOST_VDPA_SET_STATUS,
  43         [VHOST_USER_GET_STATUS] = VHOST_VDPA_GET_STATUS,
  44 };
  45
  46 /* no alignment requirement */
  47 struct vhost_iotlb_msg {
  48         uint64_t iova;
  49         uint64_t size;
  50         uint64_t uaddr;
  51 #define VHOST_ACCESS_RO      0x1
  52 #define VHOST_ACCESS_WO      0x2
  53 #define VHOST_ACCESS_RW      0x3
  54         uint8_t perm;
  55 #define VHOST_IOTLB_MISS           1
  56 #define VHOST_IOTLB_UPDATE         2
  57 #define VHOST_IOTLB_INVALIDATE     3
  58 #define VHOST_IOTLB_ACCESS_FAIL    4
  59 #define VHOST_IOTLB_BATCH_BEGIN    5
  60 #define VHOST_IOTLB_BATCH_END      6
  61         uint8_t type;
  62 };
  63
  64 #define VHOST_IOTLB_MSG_V2 0x2
  65
  66 struct vhost_msg {
  67         uint32_t type;
  68         uint32_t reserved;
  69         union {
  70                 struct vhost_iotlb_msg iotlb;
  71                 uint8_t padding[64];
  72         };
  73 };
  74
  75
  76 static int
  77 vhost_vdpa_ioctl(int fd, uint64_t request, void *arg)
  78 {
  79         int ret;
  80
  81         ret = ioctl(fd, request, arg);
  82         if (ret) {
  83                 PMD_DRV_LOG(ERR, "Vhost-vDPA ioctl %"PRIu64" failed (%s)",
  84                                 request, strerror(errno));
  85                 return -1;
  86         }
  87
  88         return 0;
  89 }
  90
  91 static int
  92 vhost_vdpa_set_owner(struct virtio_user_dev *dev)
  93 {
  94         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_OWNER, NULL);
  95 }
  96
  97 static int
  98 vhost_vdpa_get_backend_features(struct virtio_user_dev *dev, uint64_t *features)
  99 {
 100         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_GET_BACKEND_FEATURES, features);
 101 }
 102
 103 static int
 104 vhost_vdpa_set_backend_features(struct virtio_user_dev *dev, uint64_t features)
 105 {
 106         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_BACKEND_FEATURES, &features);
 107 }
 108
 109 static int
 110 vhost_vdpa_get_features(struct virtio_user_dev *dev, uint64_t *features)
 111 {
 112         int ret;
 113
 114         ret = vhost_vdpa_ioctl(dev->vhostfd, VHOST_GET_FEATURES, features);
 115         if (ret) {
 116                 PMD_DRV_LOG(ERR, "Failed to get features");
 117                 return -1;
 118         }
 119
 120         /* Multiqueue not supported for now */
 121         *features &= ~(1ULL << VIRTIO_NET_F_MQ);
 122
 123         return 0;
 124 }
 125
 126 static int
 127 vhost_vdpa_set_features(struct virtio_user_dev *dev, uint64_t features)
 128 {
 129         /* WORKAROUND */
 130         features |= 1ULL << VIRTIO_F_IOMMU_PLATFORM;
 131
 132         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_FEATURES, &features);
 133 }
 134
 135 static int
 136 vhost_vdpa_iotlb_batch_begin(struct virtio_user_dev *dev)
 137 {
 138         struct vhost_msg msg = {};
 139
 140         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_BATCH)))
 141                 return 0;
 142
 143         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2))) {
 144                 PMD_DRV_LOG(ERR, "IOTLB_MSG_V2 not supported by the backend.");
 145                 return -1;
 146         }
 147
 148         msg.type = VHOST_IOTLB_MSG_V2;
 149         msg.iotlb.type = VHOST_IOTLB_BATCH_BEGIN;
 150
 151         if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
 152                 PMD_DRV_LOG(ERR, "Failed to send IOTLB batch begin (%s)",
 153                                 strerror(errno));
 154                 return -1;
 155         }
 156
 157         return 0;
 158 }
 159
 160 static int
 161 vhost_vdpa_iotlb_batch_end(struct virtio_user_dev *dev)
 162 {
 163         struct vhost_msg msg = {};
 164
 165         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_BATCH)))
 166                 return 0;
 167
 168         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2))) {
 169                 PMD_DRV_LOG(ERR, "IOTLB_MSG_V2 not supported by the backend.");
 170                 return -1;
 171         }
 172
 173         msg.type = VHOST_IOTLB_MSG_V2;
 174         msg.iotlb.type = VHOST_IOTLB_BATCH_END;
 175
 176         if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
 177                 PMD_DRV_LOG(ERR, "Failed to send IOTLB batch end (%s)",
 178                                 strerror(errno));
 179                 return -1;
 180         }
 181
 182         return 0;
 183 }
 184
 185 static int
 186 vhost_vdpa_dma_map(struct virtio_user_dev *dev, void *addr,
 187                                   uint64_t iova, size_t len)
 188 {
 189         struct vhost_msg msg = {};
 190
 191         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2))) {
 192                 PMD_DRV_LOG(ERR, "IOTLB_MSG_V2 not supported by the backend.");
 193                 return -1;
 194         }
 195
 196         msg.type = VHOST_IOTLB_MSG_V2;
 197         msg.iotlb.type = VHOST_IOTLB_UPDATE;
 198         msg.iotlb.iova = iova;
 199         msg.iotlb.uaddr = (uint64_t)(uintptr_t)addr;
 200         msg.iotlb.size = len;
 201         msg.iotlb.perm = VHOST_ACCESS_RW;
 202
 203         PMD_DRV_LOG(DEBUG, "%s: iova: 0x%" PRIx64 ", addr: %p, len: 0x%zx",
 204                         __func__, iova, addr, len);
 205
 206         if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
 207                 PMD_DRV_LOG(ERR, "Failed to send IOTLB update (%s)",
 208                                 strerror(errno));
 209                 return -1;
 210         }
 211
 212         return 0;
 213 }
 214
 215 static int
 216 vhost_vdpa_dma_unmap(struct virtio_user_dev *dev, __rte_unused void *addr,
 217                                   uint64_t iova, size_t len)
 218 {
 219         struct vhost_msg msg = {};
 220
 221         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2))) {
 222                 PMD_DRV_LOG(ERR, "IOTLB_MSG_V2 not supported by the backend.");
 223                 return -1;
 224         }
 225
 226         msg.type = VHOST_IOTLB_MSG_V2;
 227         msg.iotlb.type = VHOST_IOTLB_INVALIDATE;
 228         msg.iotlb.iova = iova;
 229         msg.iotlb.size = len;
 230
 231         PMD_DRV_LOG(DEBUG, "%s: iova: 0x%" PRIx64 ", len: 0x%zx",
 232                         __func__, iova, len);
 233
 234         if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
 235                 PMD_DRV_LOG(ERR, "Failed to send IOTLB invalidate (%s)",
 236                                 strerror(errno));
 237                 return -1;
 238         }
 239
 240         return 0;
 241 }
 242
 243 static int
 244 vhost_vdpa_dma_map_batch(struct virtio_user_dev *dev, void *addr,
 245                                   uint64_t iova, size_t len)
 246 {
 247         int ret;
 248
 249         if (vhost_vdpa_iotlb_batch_begin(dev) < 0)
 250                 return -1;
 251
 252         ret = vhost_vdpa_dma_map(dev, addr, iova, len);
 253
 254         if (vhost_vdpa_iotlb_batch_end(dev) < 0)
 255                 return -1;
 256
 257         return ret;
 258 }
 259
 260 static int
 261 vhost_vdpa_dma_unmap_batch(struct virtio_user_dev *dev, void *addr,
 262                                   uint64_t iova, size_t len)
 263 {
 264         int ret;
 265
 266         if (vhost_vdpa_iotlb_batch_begin(dev) < 0)
 267                 return -1;
 268
 269         ret = vhost_vdpa_dma_unmap(dev, addr, iova, len);
 270
 271         if (vhost_vdpa_iotlb_batch_end(dev) < 0)
 272                 return -1;
 273
 274         return ret;
 275 }
 276
 277 static int
 278 vhost_vdpa_map_contig(const struct rte_memseg_list *msl,
 279                 const struct rte_memseg *ms, size_t len, void *arg)
 280 {
 281         struct virtio_user_dev *dev = arg;
 282
 283         if (msl->external)
 284                 return 0;
 285
 286         return vhost_vdpa_dma_map(dev, ms->addr, ms->iova, len);
 287 }
 288
 289 static int
 290 vhost_vdpa_map(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
 291                 void *arg)
 292 {
 293         struct virtio_user_dev *dev = arg;
 294
 295         /* skip external memory that isn't a heap */
 296         if (msl->external && !msl->heap)
 297                 return 0;
 298
 299         /* skip any segments with invalid IOVA addresses */
 300         if (ms->iova == RTE_BAD_IOVA)
 301                 return 0;
 302
 303         /* if IOVA mode is VA, we've already mapped the internal segments */
 304         if (!msl->external && rte_eal_iova_mode() == RTE_IOVA_VA)
 305                 return 0;
 306
 307         return vhost_vdpa_dma_map(dev, ms->addr, ms->iova, ms->len);
 308 }
 309
 310 static int
 311 vhost_vdpa_set_memory_table(struct virtio_user_dev *dev)
 312 {
 313         int ret;
 314
 315         if (vhost_vdpa_iotlb_batch_begin(dev) < 0)
 316                 return -1;
 317
 318         vhost_vdpa_dma_unmap(dev, NULL, 0, SIZE_MAX);
 319
 320         if (rte_eal_iova_mode() == RTE_IOVA_VA) {
 321                 /* with IOVA as VA mode, we can get away with mapping contiguous
 322                  * chunks rather than going page-by-page.
 323                  */
 324                 ret = rte_memseg_contig_walk_thread_unsafe(
 325                                 vhost_vdpa_map_contig, dev);
 326                 if (ret)
 327                         goto batch_end;
 328                 /* we have to continue the walk because we've skipped the
 329                  * external segments during the config walk.
 330                  */
 331         }
 332         ret = rte_memseg_walk_thread_unsafe(vhost_vdpa_map, dev);
 333
 334 batch_end:
 335         if (vhost_vdpa_iotlb_batch_end(dev) < 0)
 336                 return -1;
 337
 338         return ret;
 339 }
 340
 341 static int
 342 vhost_vdpa_set_vring_enable(struct virtio_user_dev *dev, struct vhost_vring_state *state)
 343 {
 344         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_VDPA_SET_VRING_ENABLE, state);
 345 }
 346
 347 static int
 348 vhost_vdpa_set_vring_num(struct virtio_user_dev *dev, struct vhost_vring_state *state)
 349 {
 350         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_VRING_NUM, state);
 351 }
 352
 353 static int
 354 vhost_vdpa_set_vring_base(struct virtio_user_dev *dev, struct vhost_vring_state *state)
 355 {
 356         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_VRING_BASE, state);
 357 }
 358
 359 static int
 360 vhost_vdpa_get_vring_base(struct virtio_user_dev *dev, struct vhost_vring_state *state)
 361 {
 362         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_GET_VRING_BASE, state);
 363 }
 364
 365 static int
 366 vhost_vdpa_set_vring_call(struct virtio_user_dev *dev, struct vhost_vring_file *file)
 367 {
 368         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_VRING_CALL, file);
 369 }
 370
 371 static int
 372 vhost_vdpa_set_vring_kick(struct virtio_user_dev *dev, struct vhost_vring_file *file)
 373 {
 374         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_VRING_KICK, file);
 375 }
 376
 377 /* with below features, vhost vdpa does not need to do the checksum and TSO,
 378  * these info will be passed to virtio_user through virtio net header.
 379  */
 380 #define VHOST_VDPA_GUEST_OFFLOADS_MASK  \
 381         ((1ULL << VIRTIO_NET_F_GUEST_CSUM) |    \
 382          (1ULL << VIRTIO_NET_F_GUEST_TSO4) |    \
 383          (1ULL << VIRTIO_NET_F_GUEST_TSO6) |    \
 384          (1ULL << VIRTIO_NET_F_GUEST_ECN)  |    \
 385          (1ULL << VIRTIO_NET_F_GUEST_UFO))
 386
 387 #define VHOST_VDPA_HOST_OFFLOADS_MASK           \
 388         ((1ULL << VIRTIO_NET_F_HOST_TSO4) |     \
 389          (1ULL << VIRTIO_NET_F_HOST_TSO6) |     \
 390          (1ULL << VIRTIO_NET_F_CSUM))
 391
 392 static int
 393 vhost_vdpa_send_request(struct virtio_user_dev *dev,
 394                    enum vhost_user_request req,
 395                    void *arg)
 396 {
 397         int ret = -1;
 398         uint64_t req_vdpa;
 399
 400         PMD_DRV_LOG(INFO, "%s", vhost_msg_strings[req]);
 401
 402         req_vdpa = vhost_req_user_to_vdpa[req];
 403
 404         switch (req_vdpa) {
 405         case VHOST_SET_VRING_ADDR:
 406                 PMD_DRV_LOG(DEBUG, "vhostfd=%d, index=%u",
 407                             dev->vhostfd, *(unsigned int *)arg);
 408                 break;
 409         default:
 410                 break;
 411         }
 412
 413         ret = ioctl(dev->vhostfd, req_vdpa, arg);
 414         if (ret < 0)
 415                 PMD_DRV_LOG(ERR, "%s failed: %s",
 416                             vhost_msg_strings[req], strerror(errno));
 417
 418         return ret;
 419 }
 420
 421 /**
 422  * Set up environment to talk with a vhost vdpa backend.
 423  *
 424  * @return
 425  *   - (-1) if fail to set up;
 426  *   - (>=0) if successful.
 427  */
 428 static int
 429 vhost_vdpa_setup(struct virtio_user_dev *dev)
 430 {
 431         uint32_t did = (uint32_t)-1;
 432
 433         dev->vhostfd = open(dev->path, O_RDWR);
 434         if (dev->vhostfd < 0) {
 435                 PMD_DRV_LOG(ERR, "Failed to open %s: %s\n",
 436                                 dev->path, strerror(errno));
 437                 return -1;
 438         }
 439
 440         if (ioctl(dev->vhostfd, VHOST_VDPA_GET_DEVICE_ID, &did) < 0 ||
 441                         did != VIRTIO_ID_NETWORK) {
 442                 PMD_DRV_LOG(ERR, "Invalid vdpa device ID: %u\n", did);
 443                 return -1;
 444         }
 445
 446         return 0;
 447 }
 448
 449 static int
 450 vhost_vdpa_enable_queue_pair(struct virtio_user_dev *dev,
 451                                uint16_t pair_idx,
 452                                int enable)
 453 {
 454         int i;
 455
 456         if (dev->qp_enabled[pair_idx] == enable)
 457                 return 0;
 458
 459         for (i = 0; i < 2; ++i) {
 460                 struct vhost_vring_state state = {
 461                         .index = pair_idx * 2 + i,
 462                         .num   = enable,
 463                 };
 464
 465                 if (vhost_vdpa_set_vring_enable(dev, &state))
 466                         return -1;
 467         }
 468
 469         dev->qp_enabled[pair_idx] = enable;
 470
 471         return 0;
 472 }
 473
 474 struct virtio_user_backend_ops virtio_ops_vdpa = {
 475         .setup = vhost_vdpa_setup,
 476         .set_owner = vhost_vdpa_set_owner,
 477         .get_features = vhost_vdpa_get_features,
 478         .set_features = vhost_vdpa_set_features,
 479         .get_protocol_features = vhost_vdpa_get_backend_features,
 480         .set_protocol_features = vhost_vdpa_set_backend_features,
 481         .set_memory_table = vhost_vdpa_set_memory_table,
 482         .set_vring_num = vhost_vdpa_set_vring_num,
 483         .set_vring_base = vhost_vdpa_set_vring_base,
 484         .get_vring_base = vhost_vdpa_get_vring_base,
 485         .set_vring_call = vhost_vdpa_set_vring_call,
 486         .set_vring_kick = vhost_vdpa_set_vring_kick,
 487         .send_request = vhost_vdpa_send_request,
 488         .enable_qp = vhost_vdpa_enable_queue_pair,
 489         .dma_map = vhost_vdpa_dma_map_batch,
 490         .dma_unmap = vhost_vdpa_dma_unmap_batch,
 491 };