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