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