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
  10 #include <rte_memory.h>
  11 #include <rte_eal_memconfig.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 static uint64_t max_regions = 64;
  42
  43 static void
  44 get_vhost_kernel_max_regions(void)
  45 {
  46         int fd;
  47         char buf[20] = {'\0'};
  48
  49         fd = open("/sys/module/vhost/parameters/max_mem_regions", O_RDONLY);
  50         if (fd < 0)
  51                 return;
  52
  53         if (read(fd, buf, sizeof(buf) - 1) > 0)
  54                 max_regions = strtoull(buf, NULL, 10);
  55
  56         close(fd);
  57 }
  58
  59 static uint64_t vhost_req_user_to_kernel[] = {
  60         [VHOST_USER_SET_OWNER] = VHOST_SET_OWNER,
  61         [VHOST_USER_RESET_OWNER] = VHOST_RESET_OWNER,
  62         [VHOST_USER_SET_FEATURES] = VHOST_SET_FEATURES,
  63         [VHOST_USER_GET_FEATURES] = VHOST_GET_FEATURES,
  64         [VHOST_USER_SET_VRING_CALL] = VHOST_SET_VRING_CALL,
  65         [VHOST_USER_SET_VRING_NUM] = VHOST_SET_VRING_NUM,
  66         [VHOST_USER_SET_VRING_BASE] = VHOST_SET_VRING_BASE,
  67         [VHOST_USER_GET_VRING_BASE] = VHOST_GET_VRING_BASE,
  68         [VHOST_USER_SET_VRING_ADDR] = VHOST_SET_VRING_ADDR,
  69         [VHOST_USER_SET_VRING_KICK] = VHOST_SET_VRING_KICK,
  70         [VHOST_USER_SET_MEM_TABLE] = VHOST_SET_MEM_TABLE,
  71 };
  72
  73 static int
  74 add_memseg_list(const struct rte_memseg_list *msl, void *arg)
  75 {
  76         struct vhost_memory_kernel *vm = arg;
  77         struct vhost_memory_region *mr;
  78         void *start_addr;
  79         uint64_t len;
  80
  81         if (vm->nregions >= max_regions)
  82                 return -1;
  83
  84         start_addr = msl->base_va;
  85         len = msl->page_sz * msl->memseg_arr.len;
  86
  87         mr = &vm->regions[vm->nregions++];
  88
  89         mr->guest_phys_addr = (uint64_t)(uintptr_t)start_addr;
  90         mr->userspace_addr = (uint64_t)(uintptr_t)start_addr;
  91         mr->memory_size = len;
  92         mr->mmap_offset = 0; /* flags_padding */
  93
  94         PMD_DRV_LOG(DEBUG, "index=%u addr=%p len=%" PRIu64,
  95                         vm->nregions - 1, start_addr, len);
  96
  97         return 0;
  98 }
  99
 100 /* By default, vhost kernel module allows 64 regions, but DPDK may
 101  * have much more memory regions. Below function will treat each
 102  * contiguous memory space reserved by DPDK as one region.
 103  */
 104 static struct vhost_memory_kernel *
 105 prepare_vhost_memory_kernel(void)
 106 {
 107         struct vhost_memory_kernel *vm;
 108
 109         vm = malloc(sizeof(struct vhost_memory_kernel) +
 110                         max_regions *
 111                         sizeof(struct vhost_memory_region));
 112         if (!vm)
 113                 return NULL;
 114
 115         vm->nregions = 0;
 116         vm->padding = 0;
 117
 118         /*
 119          * The memory lock has already been taken by memory subsystem
 120          * or virtio_user_start_device().
 121          */
 122         if (rte_memseg_list_walk_thread_unsafe(add_memseg_list, vm) < 0) {
 123                 free(vm);
 124                 return NULL;
 125         }
 126
 127         return vm;
 128 }
 129
 130 /* with below features, vhost kernel does not need to do the checksum and TSO,
 131  * these info will be passed to virtio_user through virtio net header.
 132  */
 133 #define VHOST_KERNEL_GUEST_OFFLOADS_MASK        \
 134         ((1ULL << VIRTIO_NET_F_GUEST_CSUM) |    \
 135          (1ULL << VIRTIO_NET_F_GUEST_TSO4) |    \
 136          (1ULL << VIRTIO_NET_F_GUEST_TSO6) |    \
 137          (1ULL << VIRTIO_NET_F_GUEST_ECN)  |    \
 138          (1ULL << VIRTIO_NET_F_GUEST_UFO))
 139
 140 /* with below features, when flows from virtio_user to vhost kernel
 141  * (1) if flows goes up through the kernel networking stack, it does not need
 142  * to verify checksum, which can save CPU cycles;
 143  * (2) if flows goes through a Linux bridge and outside from an interface
 144  * (kernel driver), checksum and TSO will be done by GSO in kernel or even
 145  * offloaded into real physical device.
 146  */
 147 #define VHOST_KERNEL_HOST_OFFLOADS_MASK         \
 148         ((1ULL << VIRTIO_NET_F_HOST_TSO4) |     \
 149          (1ULL << VIRTIO_NET_F_HOST_TSO6) |     \
 150          (1ULL << VIRTIO_NET_F_CSUM))
 151
 152 static unsigned int
 153 tap_support_features(void)
 154 {
 155         int tapfd;
 156         unsigned int tap_features;
 157
 158         tapfd = open(PATH_NET_TUN, O_RDWR);
 159         if (tapfd < 0) {
 160                 PMD_DRV_LOG(ERR, "fail to open %s: %s",
 161                             PATH_NET_TUN, strerror(errno));
 162                 return -1;
 163         }
 164
 165         if (ioctl(tapfd, TUNGETFEATURES, &tap_features) == -1) {
 166                 PMD_DRV_LOG(ERR, "TUNGETFEATURES failed: %s", strerror(errno));
 167                 close(tapfd);
 168                 return -1;
 169         }
 170
 171         close(tapfd);
 172         return tap_features;
 173 }
 174
 175 static int
 176 vhost_kernel_ioctl(struct virtio_user_dev *dev,
 177                    enum vhost_user_request req,
 178                    void *arg)
 179 {
 180         int ret = -1;
 181         unsigned int i;
 182         uint64_t req_kernel;
 183         struct vhost_memory_kernel *vm = NULL;
 184         int vhostfd;
 185         unsigned int queue_sel;
 186         unsigned int features;
 187
 188         PMD_DRV_LOG(INFO, "%s", vhost_msg_strings[req]);
 189
 190         req_kernel = vhost_req_user_to_kernel[req];
 191
 192         if (req_kernel == VHOST_SET_MEM_TABLE) {
 193                 vm = prepare_vhost_memory_kernel();
 194                 if (!vm)
 195                         return -1;
 196                 arg = (void *)vm;
 197         }
 198
 199         if (req_kernel == VHOST_SET_FEATURES) {
 200                 /* We don't need memory protection here */
 201                 *(uint64_t *)arg &= ~(1ULL << VIRTIO_F_IOMMU_PLATFORM);
 202
 203                 /* VHOST kernel does not know about below flags */
 204                 *(uint64_t *)arg &= ~VHOST_KERNEL_GUEST_OFFLOADS_MASK;
 205                 *(uint64_t *)arg &= ~VHOST_KERNEL_HOST_OFFLOADS_MASK;
 206
 207                 *(uint64_t *)arg &= ~(1ULL << VIRTIO_NET_F_MQ);
 208         }
 209
 210         switch (req_kernel) {
 211         case VHOST_SET_VRING_NUM:
 212         case VHOST_SET_VRING_ADDR:
 213         case VHOST_SET_VRING_BASE:
 214         case VHOST_GET_VRING_BASE:
 215         case VHOST_SET_VRING_KICK:
 216         case VHOST_SET_VRING_CALL:
 217                 queue_sel = *(unsigned int *)arg;
 218                 vhostfd = dev->vhostfds[queue_sel / 2];
 219                 *(unsigned int *)arg = queue_sel % 2;
 220                 PMD_DRV_LOG(DEBUG, "vhostfd=%d, index=%u",
 221                             vhostfd, *(unsigned int *)arg);
 222                 break;
 223         default:
 224                 vhostfd = -1;
 225         }
 226         if (vhostfd == -1) {
 227                 for (i = 0; i < dev->max_queue_pairs; ++i) {
 228                         if (dev->vhostfds[i] < 0)
 229                                 continue;
 230
 231                         ret = ioctl(dev->vhostfds[i], req_kernel, arg);
 232                         if (ret < 0)
 233                                 break;
 234                 }
 235         } else {
 236                 ret = ioctl(vhostfd, req_kernel, arg);
 237         }
 238
 239         if (!ret && req_kernel == VHOST_GET_FEATURES) {
 240                 features = tap_support_features();
 241                 /* with tap as the backend, all these features are supported
 242                  * but not claimed by vhost-net, so we add them back when
 243                  * reporting to upper layer.
 244                  */
 245                 if (features & IFF_VNET_HDR) {
 246                         *((uint64_t *)arg) |= VHOST_KERNEL_GUEST_OFFLOADS_MASK;
 247                         *((uint64_t *)arg) |= VHOST_KERNEL_HOST_OFFLOADS_MASK;
 248                 }
 249
 250                 /* vhost_kernel will not declare this feature, but it does
 251                  * support multi-queue.
 252                  */
 253                 if (features & IFF_MULTI_QUEUE)
 254                         *(uint64_t *)arg |= (1ull << VIRTIO_NET_F_MQ);
 255         }
 256
 257         if (vm)
 258                 free(vm);
 259
 260         if (ret < 0)
 261                 PMD_DRV_LOG(ERR, "%s failed: %s",
 262                             vhost_msg_strings[req], strerror(errno));
 263
 264         return ret;
 265 }
 266
 267 /**
 268  * Set up environment to talk with a vhost kernel backend.
 269  *
 270  * @return
 271  *   - (-1) if fail to set up;
 272  *   - (>=0) if successful.
 273  */
 274 static int
 275 vhost_kernel_setup(struct virtio_user_dev *dev)
 276 {
 277         int vhostfd;
 278         uint32_t i;
 279
 280         get_vhost_kernel_max_regions();
 281
 282         for (i = 0; i < dev->max_queue_pairs; ++i) {
 283                 vhostfd = open(dev->path, O_RDWR);
 284                 if (vhostfd < 0) {
 285                         PMD_DRV_LOG(ERR, "fail to open %s, %s",
 286                                     dev->path, strerror(errno));
 287                         return -1;
 288                 }
 289
 290                 dev->vhostfds[i] = vhostfd;
 291         }
 292
 293         return 0;
 294 }
 295
 296 static int
 297 vhost_kernel_set_backend(int vhostfd, int tapfd)
 298 {
 299         struct vhost_vring_file f;
 300
 301         f.fd = tapfd;
 302         f.index = 0;
 303         if (ioctl(vhostfd, VHOST_NET_SET_BACKEND, &f) < 0) {
 304                 PMD_DRV_LOG(ERR, "VHOST_NET_SET_BACKEND fails, %s",
 305                                 strerror(errno));
 306                 return -1;
 307         }
 308
 309         f.index = 1;
 310         if (ioctl(vhostfd, VHOST_NET_SET_BACKEND, &f) < 0) {
 311                 PMD_DRV_LOG(ERR, "VHOST_NET_SET_BACKEND fails, %s",
 312                                 strerror(errno));
 313                 return -1;
 314         }
 315
 316         return 0;
 317 }
 318
 319 static int
 320 vhost_kernel_enable_queue_pair(struct virtio_user_dev *dev,
 321                                uint16_t pair_idx,
 322                                int enable)
 323 {
 324         int hdr_size;
 325         int vhostfd;
 326         int tapfd;
 327         int req_mq = (dev->max_queue_pairs > 1);
 328
 329         vhostfd = dev->vhostfds[pair_idx];
 330
 331         if (!enable) {
 332                 if (dev->tapfds[pair_idx] >= 0) {
 333                         close(dev->tapfds[pair_idx]);
 334                         dev->tapfds[pair_idx] = -1;
 335                 }
 336                 return vhost_kernel_set_backend(vhostfd, -1);
 337         } else if (dev->tapfds[pair_idx] >= 0) {
 338                 return 0;
 339         }
 340
 341         if ((dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF)) ||
 342             (dev->features & (1ULL << VIRTIO_F_VERSION_1)))
 343                 hdr_size = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 344         else
 345                 hdr_size = sizeof(struct virtio_net_hdr);
 346
 347         tapfd = vhost_kernel_open_tap(&dev->ifname, hdr_size, req_mq,
 348                          (char *)dev->mac_addr, dev->features);
 349         if (tapfd < 0) {
 350                 PMD_DRV_LOG(ERR, "fail to open tap for vhost kernel");
 351                 return -1;
 352         }
 353
 354         if (vhost_kernel_set_backend(vhostfd, tapfd) < 0) {
 355                 PMD_DRV_LOG(ERR, "fail to set backend for vhost kernel");
 356                 close(tapfd);
 357                 return -1;
 358         }
 359
 360         dev->tapfds[pair_idx] = tapfd;
 361         return 0;
 362 }
 363
 364 struct virtio_user_backend_ops ops_kernel = {
 365         .setup = vhost_kernel_setup,
 366         .send_request = vhost_kernel_ioctl,
 367         .enable_qp = vhost_kernel_enable_queue_pair
 368 };