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_MEM_TABLE _IOW(VHOST_VIRTIO, 0x03, void *)
  23 #define VHOST_SET_LOG_BASE _IOW(VHOST_VIRTIO, 0x04, __u64)
  24 #define VHOST_SET_LOG_FD _IOW(VHOST_VIRTIO, 0x07, int)
  25 #define VHOST_SET_VRING_NUM _IOW(VHOST_VIRTIO, 0x10, struct vhost_vring_state)
  26 #define VHOST_SET_VRING_ADDR _IOW(VHOST_VIRTIO, 0x11, struct vhost_vring_addr)
  27 #define VHOST_SET_VRING_BASE _IOW(VHOST_VIRTIO, 0x12, struct vhost_vring_state)
  28 #define VHOST_GET_VRING_BASE _IOWR(VHOST_VIRTIO, 0x12, struct vhost_vring_state)
  29 #define VHOST_SET_VRING_KICK _IOW(VHOST_VIRTIO, 0x20, struct vhost_vring_file)
  30 #define VHOST_SET_VRING_CALL _IOW(VHOST_VIRTIO, 0x21, struct vhost_vring_file)
  31 #define VHOST_SET_VRING_ERR _IOW(VHOST_VIRTIO, 0x22, struct vhost_vring_file)
  32 #define VHOST_NET_SET_BACKEND _IOW(VHOST_VIRTIO, 0x30, struct vhost_vring_file)
  33 #define VHOST_VDPA_GET_DEVICE_ID _IOR(VHOST_VIRTIO, 0x70, __u32)
  34 #define VHOST_VDPA_GET_STATUS _IOR(VHOST_VIRTIO, 0x71, __u8)
  35 #define VHOST_VDPA_SET_STATUS _IOW(VHOST_VIRTIO, 0x72, __u8)
  36 #define VHOST_VDPA_SET_VRING_ENABLE     _IOW(VHOST_VIRTIO, 0x75, \
  37                                              struct vhost_vring_state)
  38 #define VHOST_SET_BACKEND_FEATURES _IOW(VHOST_VIRTIO, 0x25, __u64)
  39 #define VHOST_GET_BACKEND_FEATURES _IOR(VHOST_VIRTIO, 0x26, __u64)
  40
  41 static uint64_t vhost_req_user_to_vdpa[] = {
  42         [VHOST_USER_RESET_OWNER] = VHOST_RESET_OWNER,
  43         [VHOST_USER_SET_VRING_CALL] = VHOST_SET_VRING_CALL,
  44         [VHOST_USER_SET_VRING_NUM] = VHOST_SET_VRING_NUM,
  45         [VHOST_USER_SET_VRING_BASE] = VHOST_SET_VRING_BASE,
  46         [VHOST_USER_GET_VRING_BASE] = VHOST_GET_VRING_BASE,
  47         [VHOST_USER_SET_VRING_ADDR] = VHOST_SET_VRING_ADDR,
  48         [VHOST_USER_SET_VRING_KICK] = VHOST_SET_VRING_KICK,
  49         [VHOST_USER_SET_MEM_TABLE] = VHOST_SET_MEM_TABLE,
  50         [VHOST_USER_SET_STATUS] = VHOST_VDPA_SET_STATUS,
  51         [VHOST_USER_GET_STATUS] = VHOST_VDPA_GET_STATUS,
  52         [VHOST_USER_SET_VRING_ENABLE] = VHOST_VDPA_SET_VRING_ENABLE,
  53         [VHOST_USER_GET_PROTOCOL_FEATURES] = VHOST_GET_BACKEND_FEATURES,
  54         [VHOST_USER_SET_PROTOCOL_FEATURES] = VHOST_SET_BACKEND_FEATURES,
  55 };
  56
  57 /* no alignment requirement */
  58 struct vhost_iotlb_msg {
  59         uint64_t iova;
  60         uint64_t size;
  61         uint64_t uaddr;
  62 #define VHOST_ACCESS_RO      0x1
  63 #define VHOST_ACCESS_WO      0x2
  64 #define VHOST_ACCESS_RW      0x3
  65         uint8_t perm;
  66 #define VHOST_IOTLB_MISS           1
  67 #define VHOST_IOTLB_UPDATE         2
  68 #define VHOST_IOTLB_INVALIDATE     3
  69 #define VHOST_IOTLB_ACCESS_FAIL    4
  70 #define VHOST_IOTLB_BATCH_BEGIN    5
  71 #define VHOST_IOTLB_BATCH_END      6
  72         uint8_t type;
  73 };
  74
  75 #define VHOST_IOTLB_MSG_V2 0x2
  76
  77 struct vhost_msg {
  78         uint32_t type;
  79         uint32_t reserved;
  80         union {
  81                 struct vhost_iotlb_msg iotlb;
  82                 uint8_t padding[64];
  83         };
  84 };
  85
  86
  87 static int
  88 vhost_vdpa_ioctl(int fd, uint64_t request, void *arg)
  89 {
  90         int ret;
  91
  92         ret = ioctl(fd, request, arg);
  93         if (ret) {
  94                 PMD_DRV_LOG(ERR, "Vhost-vDPA ioctl %"PRIu64" failed (%s)",
  95                                 request, strerror(errno));
  96                 return -1;
  97         }
  98
  99         return 0;
 100 }
 101
 102 static int
 103 vhost_vdpa_set_owner(struct virtio_user_dev *dev)
 104 {
 105         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_OWNER, NULL);
 106 }
 107
 108 static int
 109 vhost_vdpa_get_backend_features(struct virtio_user_dev *dev, uint64_t *features)
 110 {
 111         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_GET_BACKEND_FEATURES, features);
 112 }
 113
 114 static int
 115 vhost_vdpa_set_backend_features(struct virtio_user_dev *dev, uint64_t features)
 116 {
 117         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_BACKEND_FEATURES, &features);
 118 }
 119
 120 static int
 121 vhost_vdpa_get_features(struct virtio_user_dev *dev, uint64_t *features)
 122 {
 123         int ret;
 124
 125         ret = vhost_vdpa_ioctl(dev->vhostfd, VHOST_GET_FEATURES, features);
 126         if (ret) {
 127                 PMD_DRV_LOG(ERR, "Failed to get features");
 128                 return -1;
 129         }
 130
 131         /* Multiqueue not supported for now */
 132         *features &= ~(1ULL << VIRTIO_NET_F_MQ);
 133
 134         return 0;
 135 }
 136
 137 static int
 138 vhost_vdpa_set_features(struct virtio_user_dev *dev, uint64_t features)
 139 {
 140         /* WORKAROUND */
 141         features |= 1ULL << VIRTIO_F_IOMMU_PLATFORM;
 142
 143         return vhost_vdpa_ioctl(dev->vhostfd, VHOST_SET_FEATURES, &features);
 144 }
 145
 146 static int
 147 vhost_vdpa_iotlb_batch_begin(struct virtio_user_dev *dev)
 148 {
 149         struct vhost_msg msg = {};
 150
 151         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_BATCH)))
 152                 return 0;
 153
 154         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2))) {
 155                 PMD_DRV_LOG(ERR, "IOTLB_MSG_V2 not supported by the backend.");
 156                 return -1;
 157         }
 158
 159         msg.type = VHOST_IOTLB_MSG_V2;
 160         msg.iotlb.type = VHOST_IOTLB_BATCH_BEGIN;
 161
 162         if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
 163                 PMD_DRV_LOG(ERR, "Failed to send IOTLB batch begin (%s)",
 164                                 strerror(errno));
 165                 return -1;
 166         }
 167
 168         return 0;
 169 }
 170
 171 static int
 172 vhost_vdpa_iotlb_batch_end(struct virtio_user_dev *dev)
 173 {
 174         struct vhost_msg msg = {};
 175
 176         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_BATCH)))
 177                 return 0;
 178
 179         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2))) {
 180                 PMD_DRV_LOG(ERR, "IOTLB_MSG_V2 not supported by the backend.");
 181                 return -1;
 182         }
 183
 184         msg.type = VHOST_IOTLB_MSG_V2;
 185         msg.iotlb.type = VHOST_IOTLB_BATCH_END;
 186
 187         if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
 188                 PMD_DRV_LOG(ERR, "Failed to send IOTLB batch end (%s)",
 189                                 strerror(errno));
 190                 return -1;
 191         }
 192
 193         return 0;
 194 }
 195
 196 static int
 197 vhost_vdpa_dma_map(struct virtio_user_dev *dev, void *addr,
 198                                   uint64_t iova, size_t len)
 199 {
 200         struct vhost_msg msg = {};
 201
 202         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2))) {
 203                 PMD_DRV_LOG(ERR, "IOTLB_MSG_V2 not supported by the backend.");
 204                 return -1;
 205         }
 206
 207         msg.type = VHOST_IOTLB_MSG_V2;
 208         msg.iotlb.type = VHOST_IOTLB_UPDATE;
 209         msg.iotlb.iova = iova;
 210         msg.iotlb.uaddr = (uint64_t)(uintptr_t)addr;
 211         msg.iotlb.size = len;
 212         msg.iotlb.perm = VHOST_ACCESS_RW;
 213
 214         PMD_DRV_LOG(DEBUG, "%s: iova: 0x%" PRIx64 ", addr: %p, len: 0x%zx",
 215                         __func__, iova, addr, len);
 216
 217         if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
 218                 PMD_DRV_LOG(ERR, "Failed to send IOTLB update (%s)",
 219                                 strerror(errno));
 220                 return -1;
 221         }
 222
 223         return 0;
 224 }
 225
 226 static int
 227 vhost_vdpa_dma_unmap(struct virtio_user_dev *dev, __rte_unused void *addr,
 228                                   uint64_t iova, size_t len)
 229 {
 230         struct vhost_msg msg = {};
 231
 232         if (!(dev->protocol_features & (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2))) {
 233                 PMD_DRV_LOG(ERR, "IOTLB_MSG_V2 not supported by the backend.");
 234                 return -1;
 235         }
 236
 237         msg.type = VHOST_IOTLB_MSG_V2;
 238         msg.iotlb.type = VHOST_IOTLB_INVALIDATE;
 239         msg.iotlb.iova = iova;
 240         msg.iotlb.size = len;
 241
 242         PMD_DRV_LOG(DEBUG, "%s: iova: 0x%" PRIx64 ", len: 0x%zx",
 243                         __func__, iova, len);
 244
 245         if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
 246                 PMD_DRV_LOG(ERR, "Failed to send IOTLB invalidate (%s)",
 247                                 strerror(errno));
 248                 return -1;
 249         }
 250
 251         return 0;
 252 }
 253
 254 static int
 255 vhost_vdpa_dma_map_batch(struct virtio_user_dev *dev, void *addr,
 256                                   uint64_t iova, size_t len)
 257 {
 258         int ret;
 259
 260         if (vhost_vdpa_iotlb_batch_begin(dev) < 0)
 261                 return -1;
 262
 263         ret = vhost_vdpa_dma_map(dev, addr, iova, len);
 264
 265         if (vhost_vdpa_iotlb_batch_end(dev) < 0)
 266                 return -1;
 267
 268         return ret;
 269 }
 270
 271 static int
 272 vhost_vdpa_dma_unmap_batch(struct virtio_user_dev *dev, void *addr,
 273                                   uint64_t iova, size_t len)
 274 {
 275         int ret;
 276
 277         if (vhost_vdpa_iotlb_batch_begin(dev) < 0)
 278                 return -1;
 279
 280         ret = vhost_vdpa_dma_unmap(dev, addr, iova, len);
 281
 282         if (vhost_vdpa_iotlb_batch_end(dev) < 0)
 283                 return -1;
 284
 285         return ret;
 286 }
 287
 288 static int
 289 vhost_vdpa_map_contig(const struct rte_memseg_list *msl,
 290                 const struct rte_memseg *ms, size_t len, void *arg)
 291 {
 292         struct virtio_user_dev *dev = arg;
 293
 294         if (msl->external)
 295                 return 0;
 296
 297         return vhost_vdpa_dma_map(dev, ms->addr, ms->iova, len);
 298 }
 299
 300 static int
 301 vhost_vdpa_map(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
 302                 void *arg)
 303 {
 304         struct virtio_user_dev *dev = arg;
 305
 306         /* skip external memory that isn't a heap */
 307         if (msl->external && !msl->heap)
 308                 return 0;
 309
 310         /* skip any segments with invalid IOVA addresses */
 311         if (ms->iova == RTE_BAD_IOVA)
 312                 return 0;
 313
 314         /* if IOVA mode is VA, we've already mapped the internal segments */
 315         if (!msl->external && rte_eal_iova_mode() == RTE_IOVA_VA)
 316                 return 0;
 317
 318         return vhost_vdpa_dma_map(dev, ms->addr, ms->iova, ms->len);
 319 }
 320
 321 static int
 322 vhost_vdpa_dma_map_all(struct virtio_user_dev *dev)
 323 {
 324         int ret;
 325
 326         if (vhost_vdpa_iotlb_batch_begin(dev) < 0)
 327                 return -1;
 328
 329         vhost_vdpa_dma_unmap(dev, NULL, 0, SIZE_MAX);
 330
 331         if (rte_eal_iova_mode() == RTE_IOVA_VA) {
 332                 /* with IOVA as VA mode, we can get away with mapping contiguous
 333                  * chunks rather than going page-by-page.
 334                  */
 335                 ret = rte_memseg_contig_walk_thread_unsafe(
 336                                 vhost_vdpa_map_contig, dev);
 337                 if (ret)
 338                         goto batch_end;
 339                 /* we have to continue the walk because we've skipped the
 340                  * external segments during the config walk.
 341                  */
 342         }
 343         ret = rte_memseg_walk_thread_unsafe(vhost_vdpa_map, dev);
 344
 345 batch_end:
 346         if (vhost_vdpa_iotlb_batch_end(dev) < 0)
 347                 return -1;
 348
 349         return ret;
 350 }
 351
 352 /* with below features, vhost vdpa does not need to do the checksum and TSO,
 353  * these info will be passed to virtio_user through virtio net header.
 354  */
 355 #define VHOST_VDPA_GUEST_OFFLOADS_MASK  \
 356         ((1ULL << VIRTIO_NET_F_GUEST_CSUM) |    \
 357          (1ULL << VIRTIO_NET_F_GUEST_TSO4) |    \
 358          (1ULL << VIRTIO_NET_F_GUEST_TSO6) |    \
 359          (1ULL << VIRTIO_NET_F_GUEST_ECN)  |    \
 360          (1ULL << VIRTIO_NET_F_GUEST_UFO))
 361
 362 #define VHOST_VDPA_HOST_OFFLOADS_MASK           \
 363         ((1ULL << VIRTIO_NET_F_HOST_TSO4) |     \
 364          (1ULL << VIRTIO_NET_F_HOST_TSO6) |     \
 365          (1ULL << VIRTIO_NET_F_CSUM))
 366
 367 static int
 368 vhost_vdpa_send_request(struct virtio_user_dev *dev,
 369                    enum vhost_user_request req,
 370                    void *arg)
 371 {
 372         int ret = -1;
 373         uint64_t req_vdpa;
 374
 375         PMD_DRV_LOG(INFO, "%s", vhost_msg_strings[req]);
 376
 377         req_vdpa = vhost_req_user_to_vdpa[req];
 378
 379         if (req_vdpa == VHOST_SET_MEM_TABLE)
 380                 return vhost_vdpa_dma_map_all(dev);
 381
 382         switch (req_vdpa) {
 383         case VHOST_SET_VRING_NUM:
 384         case VHOST_SET_VRING_ADDR:
 385         case VHOST_SET_VRING_BASE:
 386         case VHOST_GET_VRING_BASE:
 387         case VHOST_SET_VRING_KICK:
 388         case VHOST_SET_VRING_CALL:
 389                 PMD_DRV_LOG(DEBUG, "vhostfd=%d, index=%u",
 390                             dev->vhostfd, *(unsigned int *)arg);
 391                 break;
 392         default:
 393                 break;
 394         }
 395
 396         ret = ioctl(dev->vhostfd, req_vdpa, arg);
 397         if (ret < 0)
 398                 PMD_DRV_LOG(ERR, "%s failed: %s",
 399                             vhost_msg_strings[req], strerror(errno));
 400
 401         return ret;
 402 }
 403
 404 /**
 405  * Set up environment to talk with a vhost vdpa backend.
 406  *
 407  * @return
 408  *   - (-1) if fail to set up;
 409  *   - (>=0) if successful.
 410  */
 411 static int
 412 vhost_vdpa_setup(struct virtio_user_dev *dev)
 413 {
 414         uint32_t did = (uint32_t)-1;
 415
 416         dev->vhostfd = open(dev->path, O_RDWR);
 417         if (dev->vhostfd < 0) {
 418                 PMD_DRV_LOG(ERR, "Failed to open %s: %s\n",
 419                                 dev->path, strerror(errno));
 420                 return -1;
 421         }
 422
 423         if (ioctl(dev->vhostfd, VHOST_VDPA_GET_DEVICE_ID, &did) < 0 ||
 424                         did != VIRTIO_ID_NETWORK) {
 425                 PMD_DRV_LOG(ERR, "Invalid vdpa device ID: %u\n", did);
 426                 return -1;
 427         }
 428
 429         return 0;
 430 }
 431
 432 static int
 433 vhost_vdpa_enable_queue_pair(struct virtio_user_dev *dev,
 434                                uint16_t pair_idx,
 435                                int enable)
 436 {
 437         int i;
 438
 439         if (dev->qp_enabled[pair_idx] == enable)
 440                 return 0;
 441
 442         for (i = 0; i < 2; ++i) {
 443                 struct vhost_vring_state state = {
 444                         .index = pair_idx * 2 + i,
 445                         .num   = enable,
 446                 };
 447
 448                 if (vhost_vdpa_send_request(dev, VHOST_USER_SET_VRING_ENABLE, &state))
 449                         return -1;
 450         }
 451
 452         dev->qp_enabled[pair_idx] = enable;
 453
 454         return 0;
 455 }
 456
 457 struct virtio_user_backend_ops virtio_ops_vdpa = {
 458         .setup = vhost_vdpa_setup,
 459         .set_owner = vhost_vdpa_set_owner,
 460         .get_features = vhost_vdpa_get_features,
 461         .set_features = vhost_vdpa_set_features,
 462         .get_protocol_features = vhost_vdpa_get_backend_features,
 463         .set_protocol_features = vhost_vdpa_set_backend_features,
 464         .send_request = vhost_vdpa_send_request,
 465         .enable_qp = vhost_vdpa_enable_queue_pair,
 466         .dma_map = vhost_vdpa_dma_map_batch,
 467         .dma_unmap = vhost_vdpa_dma_unmap_batch,
 468 };