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 uint64_t vhost_req_user_to_kernel[] = {
 152         [VHOST_USER_RESET_OWNER] = VHOST_RESET_OWNER,
 153         [VHOST_USER_SET_VRING_CALL] = VHOST_SET_VRING_CALL,
 154         [VHOST_USER_SET_VRING_NUM] = VHOST_SET_VRING_NUM,
 155         [VHOST_USER_SET_VRING_BASE] = VHOST_SET_VRING_BASE,
 156         [VHOST_USER_GET_VRING_BASE] = VHOST_GET_VRING_BASE,
 157         [VHOST_USER_SET_VRING_ADDR] = VHOST_SET_VRING_ADDR,
 158         [VHOST_USER_SET_VRING_KICK] = VHOST_SET_VRING_KICK,
 159         [VHOST_USER_SET_MEM_TABLE] = VHOST_SET_MEM_TABLE,
 160 };
 161
 162 static int
 163 add_memseg_list(const struct rte_memseg_list *msl, void *arg)
 164 {
 165         struct vhost_memory_kernel *vm = arg;
 166         struct vhost_memory_region *mr;
 167         void *start_addr;
 168         uint64_t len;
 169
 170         if (msl->external)
 171                 return 0;
 172
 173         if (vm->nregions >= max_regions)
 174                 return -1;
 175
 176         start_addr = msl->base_va;
 177         len = msl->page_sz * msl->memseg_arr.len;
 178
 179         mr = &vm->regions[vm->nregions++];
 180
 181         mr->guest_phys_addr = (uint64_t)(uintptr_t)start_addr;
 182         mr->userspace_addr = (uint64_t)(uintptr_t)start_addr;
 183         mr->memory_size = len;
 184         mr->mmap_offset = 0; /* flags_padding */
 185
 186         PMD_DRV_LOG(DEBUG, "index=%u addr=%p len=%" PRIu64,
 187                         vm->nregions - 1, start_addr, len);
 188
 189         return 0;
 190 }
 191
 192 /* By default, vhost kernel module allows 64 regions, but DPDK may
 193  * have much more memory regions. Below function will treat each
 194  * contiguous memory space reserved by DPDK as one region.
 195  */
 196 static struct vhost_memory_kernel *
 197 prepare_vhost_memory_kernel(void)
 198 {
 199         struct vhost_memory_kernel *vm;
 200
 201         vm = malloc(sizeof(struct vhost_memory_kernel) +
 202                         max_regions *
 203                         sizeof(struct vhost_memory_region));
 204         if (!vm)
 205                 return NULL;
 206
 207         vm->nregions = 0;
 208         vm->padding = 0;
 209
 210         /*
 211          * The memory lock has already been taken by memory subsystem
 212          * or virtio_user_start_device().
 213          */
 214         if (rte_memseg_list_walk_thread_unsafe(add_memseg_list, vm) < 0) {
 215                 free(vm);
 216                 return NULL;
 217         }
 218
 219         return vm;
 220 }
 221
 222 static int
 223 vhost_kernel_send_request(struct virtio_user_dev *dev,
 224                    enum vhost_user_request req,
 225                    void *arg)
 226 {
 227         int ret = -1;
 228         unsigned int i;
 229         uint64_t req_kernel;
 230         struct vhost_memory_kernel *vm = NULL;
 231         int vhostfd;
 232         unsigned int queue_sel;
 233
 234         PMD_DRV_LOG(INFO, "%s", vhost_msg_strings[req]);
 235
 236         req_kernel = vhost_req_user_to_kernel[req];
 237
 238         if (req_kernel == VHOST_SET_MEM_TABLE) {
 239                 vm = prepare_vhost_memory_kernel();
 240                 if (!vm)
 241                         return -1;
 242                 arg = (void *)vm;
 243         }
 244
 245         switch (req_kernel) {
 246         case VHOST_SET_VRING_NUM:
 247         case VHOST_SET_VRING_ADDR:
 248         case VHOST_SET_VRING_BASE:
 249         case VHOST_GET_VRING_BASE:
 250         case VHOST_SET_VRING_KICK:
 251         case VHOST_SET_VRING_CALL:
 252                 queue_sel = *(unsigned int *)arg;
 253                 vhostfd = dev->vhostfds[queue_sel / 2];
 254                 *(unsigned int *)arg = queue_sel % 2;
 255                 PMD_DRV_LOG(DEBUG, "vhostfd=%d, index=%u",
 256                             vhostfd, *(unsigned int *)arg);
 257                 break;
 258         default:
 259                 vhostfd = -1;
 260         }
 261         if (vhostfd == -1) {
 262                 for (i = 0; i < dev->max_queue_pairs; ++i) {
 263                         if (dev->vhostfds[i] < 0)
 264                                 continue;
 265
 266                         ret = ioctl(dev->vhostfds[i], req_kernel, arg);
 267                         if (ret < 0)
 268                                 break;
 269                 }
 270         } else {
 271                 ret = ioctl(vhostfd, req_kernel, arg);
 272         }
 273
 274         if (vm)
 275                 free(vm);
 276
 277         if (ret < 0)
 278                 PMD_DRV_LOG(ERR, "%s failed: %s",
 279                             vhost_msg_strings[req], strerror(errno));
 280
 281         return ret;
 282 }
 283
 284 /**
 285  * Set up environment to talk with a vhost kernel backend.
 286  *
 287  * @return
 288  *   - (-1) if fail to set up;
 289  *   - (>=0) if successful.
 290  */
 291 static int
 292 vhost_kernel_setup(struct virtio_user_dev *dev)
 293 {
 294         int vhostfd;
 295         uint32_t i;
 296
 297         get_vhost_kernel_max_regions();
 298
 299         for (i = 0; i < dev->max_queue_pairs; ++i) {
 300                 vhostfd = open(dev->path, O_RDWR);
 301                 if (vhostfd < 0) {
 302                         PMD_DRV_LOG(ERR, "fail to open %s, %s",
 303                                     dev->path, strerror(errno));
 304                         return -1;
 305                 }
 306
 307                 dev->vhostfds[i] = vhostfd;
 308         }
 309
 310         return 0;
 311 }
 312
 313 static int
 314 vhost_kernel_set_backend(int vhostfd, int tapfd)
 315 {
 316         struct vhost_vring_file f;
 317
 318         f.fd = tapfd;
 319         f.index = 0;
 320         if (ioctl(vhostfd, VHOST_NET_SET_BACKEND, &f) < 0) {
 321                 PMD_DRV_LOG(ERR, "VHOST_NET_SET_BACKEND fails, %s",
 322                                 strerror(errno));
 323                 return -1;
 324         }
 325
 326         f.index = 1;
 327         if (ioctl(vhostfd, VHOST_NET_SET_BACKEND, &f) < 0) {
 328                 PMD_DRV_LOG(ERR, "VHOST_NET_SET_BACKEND fails, %s",
 329                                 strerror(errno));
 330                 return -1;
 331         }
 332
 333         return 0;
 334 }
 335
 336 static int
 337 vhost_kernel_enable_queue_pair(struct virtio_user_dev *dev,
 338                                uint16_t pair_idx,
 339                                int enable)
 340 {
 341         int hdr_size;
 342         int vhostfd;
 343         int tapfd;
 344         int req_mq = (dev->max_queue_pairs > 1);
 345
 346         vhostfd = dev->vhostfds[pair_idx];
 347
 348         if (dev->qp_enabled[pair_idx] == enable)
 349                 return 0;
 350
 351         if (!enable) {
 352                 tapfd = dev->tapfds[pair_idx];
 353                 if (vhost_kernel_set_backend(vhostfd, -1) < 0) {
 354                         PMD_DRV_LOG(ERR, "fail to set backend for vhost kernel");
 355                         return -1;
 356                 }
 357                 if (req_mq && vhost_kernel_tap_set_queue(tapfd, false) < 0) {
 358                         PMD_DRV_LOG(ERR, "fail to disable tap for vhost kernel");
 359                         return -1;
 360                 }
 361                 dev->qp_enabled[pair_idx] = false;
 362                 return 0;
 363         }
 364
 365         if (dev->tapfds[pair_idx] >= 0) {
 366                 tapfd = dev->tapfds[pair_idx];
 367                 if (vhost_kernel_tap_set_offload(tapfd, dev->features) == -1)
 368                         return -1;
 369                 if (req_mq && vhost_kernel_tap_set_queue(tapfd, true) < 0) {
 370                         PMD_DRV_LOG(ERR, "fail to enable tap for vhost kernel");
 371                         return -1;
 372                 }
 373                 goto set_backend;
 374         }
 375
 376         if ((dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF)) ||
 377             (dev->features & (1ULL << VIRTIO_F_VERSION_1)))
 378                 hdr_size = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 379         else
 380                 hdr_size = sizeof(struct virtio_net_hdr);
 381
 382         tapfd = vhost_kernel_open_tap(&dev->ifname, hdr_size, req_mq,
 383                          (char *)dev->mac_addr, dev->features);
 384         if (tapfd < 0) {
 385                 PMD_DRV_LOG(ERR, "fail to open tap for vhost kernel");
 386                 return -1;
 387         }
 388
 389         dev->tapfds[pair_idx] = tapfd;
 390
 391 set_backend:
 392         if (vhost_kernel_set_backend(vhostfd, tapfd) < 0) {
 393                 PMD_DRV_LOG(ERR, "fail to set backend for vhost kernel");
 394                 return -1;
 395         }
 396
 397         dev->qp_enabled[pair_idx] = true;
 398         return 0;
 399 }
 400
 401 struct virtio_user_backend_ops virtio_ops_kernel = {
 402         .setup = vhost_kernel_setup,
 403         .set_owner = vhost_kernel_set_owner,
 404         .get_features = vhost_kernel_get_features,
 405         .set_features = vhost_kernel_set_features,
 406         .send_request = vhost_kernel_send_request,
 407         .enable_qp = vhost_kernel_enable_queue_pair
 408 };