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