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