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