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 };
  42
  43 /* no alignment requirement */
  44 struct vhost_iotlb_msg {
  45         uint64_t iova;
  46         uint64_t size;
  47         uint64_t uaddr;
  48 #define VHOST_ACCESS_RO      0x1
  49 #define VHOST_ACCESS_WO      0x2
  50 #define VHOST_ACCESS_RW      0x3
  51         uint8_t perm;
  52 #define VHOST_IOTLB_MISS           1
  53 #define VHOST_IOTLB_UPDATE         2
  54 #define VHOST_IOTLB_INVALIDATE     3
  55 #define VHOST_IOTLB_ACCESS_FAIL    4
  56 #define VHOST_IOTLB_BATCH_BEGIN    5
  57 #define VHOST_IOTLB_BATCH_END      6
  58         uint8_t type;
  59 };
  60
  61 #define VHOST_IOTLB_MSG_V2 0x2
  62
  63 struct vhost_msg {
  64         uint32_t type;
  65         uint32_t reserved;
  66         union {
  67                 struct vhost_iotlb_msg iotlb;
  68                 uint8_t padding[64];
  69         };
  70 };
  71
  72
  73 static int
  74 vhost_vdpa_ioctl(int fd, uint64_t request, void *arg)
  75 {
  76         int ret;
  77
  78         ret = ioctl(fd, request, arg);
  79         if (ret) {
  80                 PMD_DRV_LOG(ERR, "Vhost-vDPA ioctl %"PRIu64" failed (%s)",
  81                                 request, strerror(errno));
  82                 return -1;
  83         }
  84
  85         return 0;
  86 }
  87
  88 static int
  89 vhost_vdpa_set_owner(struct virtio_user_dev *dev)
  90 {
  91         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_OWNER, NULL);
  92 }
  93
  94 static int
  95 vhost_vdpa_get_backend_features(struct virtio_user_dev *dev, uint64_t *features)
  96 {
  97         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_GET_BACKEND_FEATURES, features);
  98 }
  99
 100 static int
 101 vhost_vdpa_set_backend_features(struct virtio_user_dev *dev, uint64_t features)
 102 {
 103         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_BACKEND_FEATURES, &features);
 104 }
 105
 106 static int
 107 vhost_vdpa_get_features(struct virtio_user_dev *dev, uint64_t *features)
 108 {
 109         int ret;
 110
 111         ret = vhost_vdpa_ioctl(dev->vhostfd, VHOST_GET_FEATURES, features);
 112         if (ret) {
 113                 PMD_DRV_LOG(ERR, "Failed to get features");
 114                 return -1;
 115         }
 116
 117         /* Multiqueue not supported for now */
 118         *features &= ~(1ULL << VIRTIO_NET_F_MQ);
 119
 120         return 0;
 121 }
 122
 123 static int
 124 vhost_vdpa_set_features(struct virtio_user_dev *dev, uint64_t features)
 125 {
 126         /* WORKAROUND */
 127         features |= 1ULL << VIRTIO_F_IOMMU_PLATFORM;
 128
 129         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_FEATURES, &features);
 130 }
 131
 132 static int
 133 vhost_vdpa_iotlb_batch_begin(struct virtio_user_dev *dev)
 134 {
 135         struct vhost_msg msg = {};
 136
 137         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_BATCH)))
 138                 return 0;
 139
 140         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2))) {
 141                 PMD_DRV_LOG(ERR, "IOTLB_MSG_V2 not supported by the backend.");
 142                 return -1;
 143         }
 144
 145         msg.type = VHOST_IOTLB_MSG_V2;
 146         msg.iotlb.type = VHOST_IOTLB_BATCH_BEGIN;
 147
 148         if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
 149                 PMD_DRV_LOG(ERR, "Failed to send IOTLB batch begin (%s)",
 150                                 strerror(errno));
 151                 return -1;
 152         }
 153
 154         return 0;
 155 }
 156
 157 static int
 158 vhost_vdpa_iotlb_batch_end(struct virtio_user_dev *dev)
 159 {
 160         struct vhost_msg msg = {};
 161
 162         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_BATCH)))
 163                 return 0;
 164
 165         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2))) {
 166                 PMD_DRV_LOG(ERR, "IOTLB_MSG_V2 not supported by the backend.");
 167                 return -1;
 168         }
 169
 170         msg.type = VHOST_IOTLB_MSG_V2;
 171         msg.iotlb.type = VHOST_IOTLB_BATCH_END;
 172
 173         if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
 174                 PMD_DRV_LOG(ERR, "Failed to send IOTLB batch end (%s)",
 175                                 strerror(errno));
 176                 return -1;
 177         }
 178
 179         return 0;
 180 }
 181
 182 static int
 183 vhost_vdpa_dma_map(struct virtio_user_dev *dev, void *addr,
 184                                   uint64_t iova, size_t len)
 185 {
 186         struct vhost_msg msg = {};
 187
 188         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2))) {
 189                 PMD_DRV_LOG(ERR, "IOTLB_MSG_V2 not supported by the backend.");
 190                 return -1;
 191         }
 192
 193         msg.type = VHOST_IOTLB_MSG_V2;
 194         msg.iotlb.type = VHOST_IOTLB_UPDATE;
 195         msg.iotlb.iova = iova;
 196         msg.iotlb.uaddr = (uint64_t)(uintptr_t)addr;
 197         msg.iotlb.size = len;
 198         msg.iotlb.perm = VHOST_ACCESS_RW;
 199
 200         PMD_DRV_LOG(DEBUG, "%s: iova: 0x%" PRIx64 ", addr: %p, len: 0x%zx",
 201                         __func__, iova, addr, len);
 202
 203         if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
 204                 PMD_DRV_LOG(ERR, "Failed to send IOTLB update (%s)",
 205                                 strerror(errno));
 206                 return -1;
 207         }
 208
 209         return 0;
 210 }
 211
 212 static int
 213 vhost_vdpa_dma_unmap(struct virtio_user_dev *dev, __rte_unused void *addr,
 214                                   uint64_t iova, size_t len)
 215 {
 216         struct vhost_msg msg = {};
 217
 218         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2))) {
 219                 PMD_DRV_LOG(ERR, "IOTLB_MSG_V2 not supported by the backend.");
 220                 return -1;
 221         }
 222
 223         msg.type = VHOST_IOTLB_MSG_V2;
 224         msg.iotlb.type = VHOST_IOTLB_INVALIDATE;
 225         msg.iotlb.iova = iova;
 226         msg.iotlb.size = len;
 227
 228         PMD_DRV_LOG(DEBUG, "%s: iova: 0x%" PRIx64 ", len: 0x%zx",
 229                         __func__, iova, len);
 230
 231         if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
 232                 PMD_DRV_LOG(ERR, "Failed to send IOTLB invalidate (%s)",
 233                                 strerror(errno));
 234                 return -1;
 235         }
 236
 237         return 0;
 238 }
 239
 240 static int
 241 vhost_vdpa_dma_map_batch(struct virtio_user_dev *dev, void *addr,
 242                                   uint64_t iova, size_t len)
 243 {
 244         int ret;
 245
 246         if (vhost_vdpa_iotlb_batch_begin(dev) < 0)
 247                 return -1;
 248
 249         ret = vhost_vdpa_dma_map(dev, addr, iova, len);
 250
 251         if (vhost_vdpa_iotlb_batch_end(dev) < 0)
 252                 return -1;
 253
 254         return ret;
 255 }
 256
 257 static int
 258 vhost_vdpa_dma_unmap_batch(struct virtio_user_dev *dev, void *addr,
 259                                   uint64_t iova, size_t len)
 260 {
 261         int ret;
 262
 263         if (vhost_vdpa_iotlb_batch_begin(dev) < 0)
 264                 return -1;
 265
 266         ret = vhost_vdpa_dma_unmap(dev, addr, iova, len);
 267
 268         if (vhost_vdpa_iotlb_batch_end(dev) < 0)
 269                 return -1;
 270
 271         return ret;
 272 }
 273
 274 static int
 275 vhost_vdpa_map_contig(const struct rte_memseg_list *msl,
 276                 const struct rte_memseg *ms, size_t len, void *arg)
 277 {
 278         struct virtio_user_dev *dev = arg;
 279
 280         if (msl->external)
 281                 return 0;
 282
 283         return vhost_vdpa_dma_map(dev, ms->addr, ms->iova, len);
 284 }
 285
 286 static int
 287 vhost_vdpa_map(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
 288                 void *arg)
 289 {
 290         struct virtio_user_dev *dev = arg;
 291
 292         /* skip external memory that isn't a heap */
 293         if (msl->external && !msl->heap)
 294                 return 0;
 295
 296         /* skip any segments with invalid IOVA addresses */
 297         if (ms->iova == RTE_BAD_IOVA)
 298                 return 0;
 299
 300         /* if IOVA mode is VA, we've already mapped the internal segments */
 301         if (!msl->external && rte_eal_iova_mode() == RTE_IOVA_VA)
 302                 return 0;
 303
 304         return vhost_vdpa_dma_map(dev, ms->addr, ms->iova, ms->len);
 305 }
 306
 307 static int
 308 vhost_vdpa_set_memory_table(struct virtio_user_dev *dev)
 309 {
 310         int ret;
 311
 312         if (vhost_vdpa_iotlb_batch_begin(dev) < 0)
 313                 return -1;
 314
 315         vhost_vdpa_dma_unmap(dev, NULL, 0, SIZE_MAX);
 316
 317         if (rte_eal_iova_mode() == RTE_IOVA_VA) {
 318                 /* with IOVA as VA mode, we can get away with mapping contiguous
 319                  * chunks rather than going page-by-page.
 320                  */
 321                 ret = rte_memseg_contig_walk_thread_unsafe(
 322                                 vhost_vdpa_map_contig, dev);
 323                 if (ret)
 324                         goto batch_end;
 325                 /* we have to continue the walk because we've skipped the
 326                  * external segments during the config walk.
 327                  */
 328         }
 329         ret = rte_memseg_walk_thread_unsafe(vhost_vdpa_map, dev);
 330
 331 batch_end:
 332         if (vhost_vdpa_iotlb_batch_end(dev) < 0)
 333                 return -1;
 334
 335         return ret;
 336 }
 337
 338 static int
 339 vhost_vdpa_set_vring_enable(struct virtio_user_dev *dev, struct vhost_vring_state *state)
 340 {
 341         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_VDPA_SET_VRING_ENABLE, state);
 342 }
 343
 344 static int
 345 vhost_vdpa_set_vring_num(struct virtio_user_dev *dev, struct vhost_vring_state *state)
 346 {
 347         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_VRING_NUM, state);
 348 }
 349
 350 static int
 351 vhost_vdpa_set_vring_base(struct virtio_user_dev *dev, struct vhost_vring_state *state)
 352 {
 353         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_VRING_BASE, state);
 354 }
 355
 356 static int
 357 vhost_vdpa_get_vring_base(struct virtio_user_dev *dev, struct vhost_vring_state *state)
 358 {
 359         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_GET_VRING_BASE, state);
 360 }
 361
 362 static int
 363 vhost_vdpa_set_vring_call(struct virtio_user_dev *dev, struct vhost_vring_file *file)
 364 {
 365         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_VRING_CALL, file);
 366 }
 367
 368 static int
 369 vhost_vdpa_set_vring_kick(struct virtio_user_dev *dev, struct vhost_vring_file *file)
 370 {
 371         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_VRING_KICK, file);
 372 }
 373
 374 static int
 375 vhost_vdpa_set_vring_addr(struct virtio_user_dev *dev, struct vhost_vring_addr *addr)
 376 {
 377         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_VRING_ADDR, addr);
 378 }
 379
 380 static int
 381 vhost_vdpa_get_status(struct virtio_user_dev *dev, uint8_t *status)
 382 {
 383         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_VDPA_GET_STATUS, status);
 384 }
 385
 386 static int
 387 vhost_vdpa_set_status(struct virtio_user_dev *dev, uint8_t status)
 388 {
 389         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_VDPA_SET_STATUS, &status);
 390 }
 391
 392 /* with below features, vhost vdpa does not need to do the checksum and TSO,
 393  * these info will be passed to virtio_user through virtio net header.
 394  */
 395 #define VHOST_VDPA_GUEST_OFFLOADS_MASK  \
 396         ((1ULL << VIRTIO_NET_F_GUEST_CSUM) |    \
 397          (1ULL << VIRTIO_NET_F_GUEST_TSO4) |    \
 398          (1ULL << VIRTIO_NET_F_GUEST_TSO6) |    \
 399          (1ULL << VIRTIO_NET_F_GUEST_ECN)  |    \
 400          (1ULL << VIRTIO_NET_F_GUEST_UFO))
 401
 402 #define VHOST_VDPA_HOST_OFFLOADS_MASK           \
 403         ((1ULL << VIRTIO_NET_F_HOST_TSO4) |     \
 404          (1ULL << VIRTIO_NET_F_HOST_TSO6) |     \
 405          (1ULL << VIRTIO_NET_F_CSUM))
 406
 407 static int
 408 vhost_vdpa_send_request(struct virtio_user_dev *dev,
 409                    enum vhost_user_request req,
 410                    void *arg)
 411 {
 412         int ret = -1;
 413         uint64_t req_vdpa;
 414
 415         PMD_DRV_LOG(INFO, "%s", vhost_msg_strings[req]);
 416
 417         req_vdpa = vhost_req_user_to_vdpa[req];
 418
 419         switch (req_vdpa) {
 420         case VHOST_SET_VRING_ADDR:
 421                 PMD_DRV_LOG(DEBUG, "vhostfd=%d, index=%u",
 422                             dev->vhostfd, *(unsigned int *)arg);
 423                 break;
 424         default:
 425                 break;
 426         }
 427
 428         ret = ioctl(dev->vhostfd, req_vdpa, arg);
 429         if (ret < 0)
 430                 PMD_DRV_LOG(ERR, "%s failed: %s",
 431                             vhost_msg_strings[req], strerror(errno));
 432
 433         return ret;
 434 }
 435
 436 /**
 437  * Set up environment to talk with a vhost vdpa backend.
 438  *
 439  * @return
 440  *   - (-1) if fail to set up;
 441  *   - (>=0) if successful.
 442  */
 443 static int
 444 vhost_vdpa_setup(struct virtio_user_dev *dev)
 445 {
 446         uint32_t did = (uint32_t)-1;
 447
 448         dev->vhostfd = open(dev->path, O_RDWR);
 449         if (dev->vhostfd < 0) {
 450                 PMD_DRV_LOG(ERR, "Failed to open %s: %s\n",
 451                                 dev->path, strerror(errno));
 452                 return -1;
 453         }
 454
 455         if (ioctl(dev->vhostfd, VHOST_VDPA_GET_DEVICE_ID, &did) < 0 ||
 456                         did != VIRTIO_ID_NETWORK) {
 457                 PMD_DRV_LOG(ERR, "Invalid vdpa device ID: %u\n", did);
 458                 return -1;
 459         }
 460
 461         return 0;
 462 }
 463
 464 static int
 465 vhost_vdpa_enable_queue_pair(struct virtio_user_dev *dev,
 466                                uint16_t pair_idx,
 467                                int enable)
 468 {
 469         int i;
 470
 471         if (dev->qp_enabled[pair_idx] == enable)
 472                 return 0;
 473
 474         for (i = 0; i < 2; ++i) {
 475                 struct vhost_vring_state state = {
 476                         .index = pair_idx * 2 + i,
 477                         .num   = enable,
 478                 };
 479
 480                 if (vhost_vdpa_set_vring_enable(dev, &state))
 481                         return -1;
 482         }
 483
 484         dev->qp_enabled[pair_idx] = enable;
 485
 486         return 0;
 487 }
 488
 489 struct virtio_user_backend_ops virtio_ops_vdpa = {
 490         .setup = vhost_vdpa_setup,
 491         .set_owner = vhost_vdpa_set_owner,
 492         .get_features = vhost_vdpa_get_features,
 493         .set_features = vhost_vdpa_set_features,
 494         .get_protocol_features = vhost_vdpa_get_backend_features,
 495         .set_protocol_features = vhost_vdpa_set_backend_features,
 496         .set_memory_table = vhost_vdpa_set_memory_table,
 497         .set_vring_num = vhost_vdpa_set_vring_num,
 498         .set_vring_base = vhost_vdpa_set_vring_base,
 499         .get_vring_base = vhost_vdpa_get_vring_base,
 500         .set_vring_call = vhost_vdpa_set_vring_call,
 501         .set_vring_kick = vhost_vdpa_set_vring_kick,
 502         .set_vring_addr = vhost_vdpa_set_vring_addr,
 503         .get_status = vhost_vdpa_get_status,
 504         .set_status = vhost_vdpa_set_status,
 505         .send_request = vhost_vdpa_send_request,
 506         .enable_qp = vhost_vdpa_enable_queue_pair,
 507         .dma_map = vhost_vdpa_dma_map_batch,
 508         .dma_unmap = vhost_vdpa_dma_unmap_batch,
 509 };