drivers/net/virtio/virtio_user/vhost_kernel.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2016 Intel Corporation
   3  */
   4
   5 #include <sys/types.h>
   6 #include <sys/stat.h>
   7 #include <fcntl.h>
   8 #include <unistd.h>
   9 #include <errno.h>
  10
  11 #include <rte_memory.h>
  12
  13 #include "vhost.h"
  14 #include "virtio_user_dev.h"
  15 #include "vhost_kernel_tap.h"
  16
  17 struct vhost_memory_kernel {
  18         uint32_t nregions;
  19         uint32_t padding;
  20         struct vhost_memory_region regions[0];
  21 };
  22
  23 /* vhost kernel ioctls */
  24 #define VHOST_VIRTIO 0xAF
  25 #define VHOST_GET_FEATURES _IOR(VHOST_VIRTIO, 0x00, __u64)
  26 #define VHOST_SET_FEATURES _IOW(VHOST_VIRTIO, 0x00, __u64)
  27 #define VHOST_SET_OWNER _IO(VHOST_VIRTIO, 0x01)
  28 #define VHOST_RESET_OWNER _IO(VHOST_VIRTIO, 0x02)
  29 #define VHOST_SET_MEM_TABLE _IOW(VHOST_VIRTIO, 0x03, struct vhost_memory_kernel)
  30 #define VHOST_SET_LOG_BASE _IOW(VHOST_VIRTIO, 0x04, __u64)
  31 #define VHOST_SET_LOG_FD _IOW(VHOST_VIRTIO, 0x07, int)
  32 #define VHOST_SET_VRING_NUM _IOW(VHOST_VIRTIO, 0x10, struct vhost_vring_state)
  33 #define VHOST_SET_VRING_ADDR _IOW(VHOST_VIRTIO, 0x11, struct vhost_vring_addr)
  34 #define VHOST_SET_VRING_BASE _IOW(VHOST_VIRTIO, 0x12, struct vhost_vring_state)
  35 #define VHOST_GET_VRING_BASE _IOWR(VHOST_VIRTIO, 0x12, struct vhost_vring_state)
  36 #define VHOST_SET_VRING_KICK _IOW(VHOST_VIRTIO, 0x20, struct vhost_vring_file)
  37 #define VHOST_SET_VRING_CALL _IOW(VHOST_VIRTIO, 0x21, struct vhost_vring_file)
  38 #define VHOST_SET_VRING_ERR _IOW(VHOST_VIRTIO, 0x22, struct vhost_vring_file)
  39 #define VHOST_NET_SET_BACKEND _IOW(VHOST_VIRTIO, 0x30, struct vhost_vring_file)
  40
  41 /* with below features, vhost kernel does not need to do the checksum and TSO,
  42  * these info will be passed to virtio_user through virtio net header.
  43  */
  44 #define VHOST_KERNEL_GUEST_OFFLOADS_MASK        \
  45         ((1ULL << VIRTIO_NET_F_GUEST_CSUM) |    \
  46          (1ULL << VIRTIO_NET_F_GUEST_TSO4) |    \
  47          (1ULL << VIRTIO_NET_F_GUEST_TSO6) |    \
  48          (1ULL << VIRTIO_NET_F_GUEST_ECN)  |    \
  49          (1ULL << VIRTIO_NET_F_GUEST_UFO))
  50
  51 /* with below features, when flows from virtio_user to vhost kernel
  52  * (1) if flows goes up through the kernel networking stack, it does not need
  53  * to verify checksum, which can save CPU cycles;
  54  * (2) if flows goes through a Linux bridge and outside from an interface
  55  * (kernel driver), checksum and TSO will be done by GSO in kernel or even
  56  * offloaded into real physical device.
  57  */
  58 #define VHOST_KERNEL_HOST_OFFLOADS_MASK         \
  59         ((1ULL << VIRTIO_NET_F_HOST_TSO4) |     \
  60          (1ULL << VIRTIO_NET_F_HOST_TSO6) |     \
  61          (1ULL << VIRTIO_NET_F_CSUM))
  62
  63 static uint64_t max_regions = 64;
  64
  65 static void
  66 get_vhost_kernel_max_regions(void)
  67 {
  68         int fd;
  69         char buf[20] = {'\0'};
  70
  71         fd = open("/sys/module/vhost/parameters/max_mem_regions", O_RDONLY);
  72         if (fd < 0)
  73                 return;
  74
  75         if (read(fd, buf, sizeof(buf) - 1) > 0)
  76                 max_regions = strtoull(buf, NULL, 10);
  77
  78         close(fd);
  79 }
  80
  81 static int
  82 vhost_kernel_ioctl(int fd, uint64_t request, void *arg)
  83 {
  84         int ret;
  85
  86         ret = ioctl(fd, request, arg);
  87         if (ret) {
  88                 PMD_DRV_LOG(ERR, "Vhost-kernel ioctl %"PRIu64" failed (%s)",
  89                                 request, strerror(errno));
  90                 return -1;
  91         }
  92
  93         return 0;
  94 }
  95
  96 static int
  97 vhost_kernel_set_owner(struct virtio_user_dev *dev)
  98 {
  99         return vhost_kernel_ioctl(dev->vhostfds[0], VHOST_SET_OWNER, NULL);
 100 }
 101
 102 static int
 103 vhost_kernel_get_features(struct virtio_user_dev *dev, uint64_t *features)
 104 {
 105         int ret;
 106         unsigned int tap_features;
 107
 108         ret = vhost_kernel_ioctl(dev->vhostfds[0], VHOST_GET_FEATURES, features);
 109         if (ret < 0) {
 110                 PMD_DRV_LOG(ERR, "Failed to get features");
 111                 return -1;
 112         }
 113
 114         ret = tap_support_features(&tap_features);
 115         if (ret < 0) {
 116                 PMD_DRV_LOG(ERR, "Failed to get TAP features");
 117                 return -1;
 118         }
 119
 120         /* with tap as the backend, all these features are supported
 121          * but not claimed by vhost-net, so we add them back when
 122          * reporting to upper layer.
 123          */
 124         if (tap_features & IFF_VNET_HDR) {
 125                 *features |= VHOST_KERNEL_GUEST_OFFLOADS_MASK;
 126                 *features |= VHOST_KERNEL_HOST_OFFLOADS_MASK;
 127         }
 128
 129         /* vhost_kernel will not declare this feature, but it does
 130          * support multi-queue.
 131          */
 132         if (tap_features & IFF_MULTI_QUEUE)
 133                 *features |= (1ull << VIRTIO_NET_F_MQ);
 134
 135         return 0;
 136 }
 137
 138 static int
 139 vhost_kernel_set_features(struct virtio_user_dev *dev, uint64_t features)
 140 {
 141         /* We don't need memory protection here */
 142         features &= ~(1ULL << VIRTIO_F_IOMMU_PLATFORM);
 143         /* VHOST kernel does not know about below flags */
 144         features &= ~VHOST_KERNEL_GUEST_OFFLOADS_MASK;
 145         features &= ~VHOST_KERNEL_HOST_OFFLOADS_MASK;
 146         features &= ~(1ULL << VIRTIO_NET_F_MQ);
 147
 148         return vhost_kernel_ioctl(dev->vhostfds[0], VHOST_SET_FEATURES, &features);
 149 }
 150
 151 static int
 152 add_memseg_list(const struct rte_memseg_list *msl, void *arg)
 153 {
 154         struct vhost_memory_kernel *vm = arg;
 155         struct vhost_memory_region *mr;
 156         void *start_addr;
 157         uint64_t len;
 158
 159         if (msl->external)
 160                 return 0;
 161
 162         if (vm->nregions >= max_regions)
 163                 return -1;
 164
 165         start_addr = msl->base_va;
 166         len = msl->page_sz * msl->memseg_arr.len;
 167
 168         mr = &vm->regions[vm->nregions++];
 169
 170         mr->guest_phys_addr = (uint64_t)(uintptr_t)start_addr;
 171         mr->userspace_addr = (uint64_t)(uintptr_t)start_addr;
 172         mr->memory_size = len;
 173         mr->mmap_offset = 0; /* flags_padding */
 174
 175         PMD_DRV_LOG(DEBUG, "index=%u addr=%p len=%" PRIu64,
 176                         vm->nregions - 1, start_addr, len);
 177
 178         return 0;
 179 }
 180
 181 /* By default, vhost kernel module allows 64 regions, but DPDK may
 182  * have much more memory regions. Below function will treat each
 183  * contiguous memory space reserved by DPDK as one region.
 184  */
 185 static int
 186 vhost_kernel_set_memory_table(struct virtio_user_dev *dev)
 187 {
 188         struct vhost_memory_kernel *vm;
 189         int ret;
 190
 191         vm = malloc(sizeof(struct vhost_memory_kernel) +
 192                         max_regions *
 193                         sizeof(struct vhost_memory_region));
 194         if (!vm)
 195                 goto err;
 196
 197         vm->nregions = 0;
 198         vm->padding = 0;
 199
 200         /*
 201          * The memory lock has already been taken by memory subsystem
 202          * or virtio_user_start_device().
 203          */
 204         ret = rte_memseg_list_walk_thread_unsafe(add_memseg_list, vm);
 205         if (ret < 0)
 206                 goto err_free;
 207
 208         ret = vhost_kernel_ioctl(dev->vhostfds[0], VHOST_SET_MEM_TABLE, vm);
 209         if (ret < 0)
 210                 goto err_free;
 211
 212         free(vm);
 213
 214         return 0;
 215 err_free:
 216         free(vm);
 217 err:
 218         PMD_DRV_LOG(ERR, "Failed to set memory table");
 219         return -1;
 220 }
 221
 222 static int
 223 vhost_kernel_set_vring(struct virtio_user_dev *dev, uint64_t req, struct vhost_vring_state *state)
 224 {
 225         int ret, fd;
 226         unsigned int index = state->index;
 227
 228         /* Convert from queue index to queue-pair & offset */
 229         fd = dev->vhostfds[state->index / 2];
 230         state->index %= 2;
 231
 232         ret = vhost_kernel_ioctl(fd, req, state);
 233         if (ret < 0) {
 234                 PMD_DRV_LOG(ERR, "Failed to set vring (request %" PRIu64 ")", req);
 235                 return -1;
 236         }
 237
 238         /* restore index back to queue index */
 239         state->index = index;
 240
 241         return 0;
 242 }
 243
 244 static int
 245 vhost_kernel_set_vring_num(struct virtio_user_dev *dev, struct vhost_vring_state *state)
 246 {
 247         return vhost_kernel_set_vring(dev, VHOST_SET_VRING_NUM, state);
 248 }
 249
 250 static int
 251 vhost_kernel_set_vring_base(struct virtio_user_dev *dev, struct vhost_vring_state *state)
 252 {
 253         return vhost_kernel_set_vring(dev, VHOST_SET_VRING_BASE, state);
 254 }
 255
 256 static int
 257 vhost_kernel_get_vring_base(struct virtio_user_dev *dev, struct vhost_vring_state *state)
 258 {
 259         return vhost_kernel_set_vring(dev, VHOST_GET_VRING_BASE, state);
 260 }
 261
 262 static uint64_t vhost_req_user_to_kernel[] = {
 263         [VHOST_USER_RESET_OWNER] = VHOST_RESET_OWNER,
 264         [VHOST_USER_SET_VRING_CALL] = VHOST_SET_VRING_CALL,
 265         [VHOST_USER_SET_VRING_ADDR] = VHOST_SET_VRING_ADDR,
 266         [VHOST_USER_SET_VRING_KICK] = VHOST_SET_VRING_KICK,
 267 };
 268
 269 static int
 270 vhost_kernel_send_request(struct virtio_user_dev *dev,
 271                    enum vhost_user_request req,
 272                    void *arg)
 273 {
 274         int ret = -1;
 275         unsigned int i;
 276         uint64_t req_kernel;
 277         int vhostfd;
 278         unsigned int queue_sel;
 279
 280         PMD_DRV_LOG(INFO, "%s", vhost_msg_strings[req]);
 281
 282         req_kernel = vhost_req_user_to_kernel[req];
 283
 284         switch (req_kernel) {
 285         case VHOST_SET_VRING_ADDR:
 286         case VHOST_SET_VRING_KICK:
 287         case VHOST_SET_VRING_CALL:
 288                 queue_sel = *(unsigned int *)arg;
 289                 vhostfd = dev->vhostfds[queue_sel / 2];
 290                 *(unsigned int *)arg = queue_sel % 2;
 291                 PMD_DRV_LOG(DEBUG, "vhostfd=%d, index=%u",
 292                             vhostfd, *(unsigned int *)arg);
 293                 break;
 294         default:
 295                 vhostfd = -1;
 296         }
 297         if (vhostfd == -1) {
 298                 for (i = 0; i < dev->max_queue_pairs; ++i) {
 299                         if (dev->vhostfds[i] < 0)
 300                                 continue;
 301
 302                         ret = ioctl(dev->vhostfds[i], req_kernel, arg);
 303                         if (ret < 0)
 304                                 break;
 305                 }
 306         } else {
 307                 ret = ioctl(vhostfd, req_kernel, arg);
 308         }
 309
 310         if (ret < 0)
 311                 PMD_DRV_LOG(ERR, "%s failed: %s",
 312                             vhost_msg_strings[req], strerror(errno));
 313
 314         return ret;
 315 }
 316
 317 /**
 318  * Set up environment to talk with a vhost kernel backend.
 319  *
 320  * @return
 321  *   - (-1) if fail to set up;
 322  *   - (>=0) if successful.
 323  */
 324 static int
 325 vhost_kernel_setup(struct virtio_user_dev *dev)
 326 {
 327         int vhostfd;
 328         uint32_t i;
 329
 330         get_vhost_kernel_max_regions();
 331
 332         for (i = 0; i < dev->max_queue_pairs; ++i) {
 333                 vhostfd = open(dev->path, O_RDWR);
 334                 if (vhostfd < 0) {
 335                         PMD_DRV_LOG(ERR, "fail to open %s, %s",
 336                                     dev->path, strerror(errno));
 337                         return -1;
 338                 }
 339
 340                 dev->vhostfds[i] = vhostfd;
 341         }
 342
 343         return 0;
 344 }
 345
 346 static int
 347 vhost_kernel_set_backend(int vhostfd, int tapfd)
 348 {
 349         struct vhost_vring_file f;
 350
 351         f.fd = tapfd;
 352         f.index = 0;
 353         if (ioctl(vhostfd, VHOST_NET_SET_BACKEND, &f) < 0) {
 354                 PMD_DRV_LOG(ERR, "VHOST_NET_SET_BACKEND fails, %s",
 355                                 strerror(errno));
 356                 return -1;
 357         }
 358
 359         f.index = 1;
 360         if (ioctl(vhostfd, VHOST_NET_SET_BACKEND, &f) < 0) {
 361                 PMD_DRV_LOG(ERR, "VHOST_NET_SET_BACKEND fails, %s",
 362                                 strerror(errno));
 363                 return -1;
 364         }
 365
 366         return 0;
 367 }
 368
 369 static int
 370 vhost_kernel_enable_queue_pair(struct virtio_user_dev *dev,
 371                                uint16_t pair_idx,
 372                                int enable)
 373 {
 374         int hdr_size;
 375         int vhostfd;
 376         int tapfd;
 377         int req_mq = (dev->max_queue_pairs > 1);
 378
 379         vhostfd = dev->vhostfds[pair_idx];
 380
 381         if (dev->qp_enabled[pair_idx] == enable)
 382                 return 0;
 383
 384         if (!enable) {
 385                 tapfd = dev->tapfds[pair_idx];
 386                 if (vhost_kernel_set_backend(vhostfd, -1) < 0) {
 387                         PMD_DRV_LOG(ERR, "fail to set backend for vhost kernel");
 388                         return -1;
 389                 }
 390                 if (req_mq && vhost_kernel_tap_set_queue(tapfd, false) < 0) {
 391                         PMD_DRV_LOG(ERR, "fail to disable tap for vhost kernel");
 392                         return -1;
 393                 }
 394                 dev->qp_enabled[pair_idx] = false;
 395                 return 0;
 396         }
 397
 398         if (dev->tapfds[pair_idx] >= 0) {
 399                 tapfd = dev->tapfds[pair_idx];
 400                 if (vhost_kernel_tap_set_offload(tapfd, dev->features) == -1)
 401                         return -1;
 402                 if (req_mq && vhost_kernel_tap_set_queue(tapfd, true) < 0) {
 403                         PMD_DRV_LOG(ERR, "fail to enable tap for vhost kernel");
 404                         return -1;
 405                 }
 406                 goto set_backend;
 407         }
 408
 409         if ((dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF)) ||
 410             (dev->features & (1ULL << VIRTIO_F_VERSION_1)))
 411                 hdr_size = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 412         else
 413                 hdr_size = sizeof(struct virtio_net_hdr);
 414
 415         tapfd = vhost_kernel_open_tap(&dev->ifname, hdr_size, req_mq,
 416                          (char *)dev->mac_addr, dev->features);
 417         if (tapfd < 0) {
 418                 PMD_DRV_LOG(ERR, "fail to open tap for vhost kernel");
 419                 return -1;
 420         }
 421
 422         dev->tapfds[pair_idx] = tapfd;
 423
 424 set_backend:
 425         if (vhost_kernel_set_backend(vhostfd, tapfd) < 0) {
 426                 PMD_DRV_LOG(ERR, "fail to set backend for vhost kernel");
 427                 return -1;
 428         }
 429
 430         dev->qp_enabled[pair_idx] = true;
 431         return 0;
 432 }
 433
 434 struct virtio_user_backend_ops virtio_ops_kernel = {
 435         .setup = vhost_kernel_setup,
 436         .set_owner = vhost_kernel_set_owner,
 437         .get_features = vhost_kernel_get_features,
 438         .set_features = vhost_kernel_set_features,
 439         .set_memory_table = vhost_kernel_set_memory_table,
 440         .set_vring_num = vhost_kernel_set_vring_num,
 441         .set_vring_base = vhost_kernel_set_vring_base,
 442         .get_vring_base = vhost_kernel_get_vring_base,
 443         .send_request = vhost_kernel_send_request,
 444         .enable_qp = vhost_kernel_enable_queue_pair
 445 };