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