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