lib/librte_vhost/vhost_user/virtio-net-user.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stdint.h>
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <unistd.h>
  38 #include <sys/mman.h>
  39 #include <sys/types.h>
  40 #include <sys/stat.h>
  41 #include <unistd.h>
  42
  43 #include <rte_common.h>
  44 #include <rte_log.h>
  45
  46 #include "virtio-net.h"
  47 #include "virtio-net-user.h"
  48 #include "vhost-net-user.h"
  49 #include "vhost-net.h"
  50
  51 struct orig_region_map {
  52         int fd;
  53         uint64_t mapped_address;
  54         uint64_t mapped_size;
  55         uint64_t blksz;
  56 };
  57
  58 #define orig_region(ptr, nregions) \
  59         ((struct orig_region_map *)RTE_PTR_ADD((ptr), \
  60                 sizeof(struct virtio_memory) + \
  61                 sizeof(struct virtio_memory_regions) * (nregions)))
  62
  63 static uint64_t
  64 get_blk_size(int fd)
  65 {
  66         struct stat stat;
  67
  68         fstat(fd, &stat);
  69         return (uint64_t)stat.st_blksize;
  70 }
  71
  72 static void
  73 free_mem_region(struct virtio_net *dev)
  74 {
  75         struct orig_region_map *region;
  76         unsigned int idx;
  77
  78         if (!dev || !dev->mem)
  79                 return;
  80
  81         region = orig_region(dev->mem, dev->mem->nregions);
  82         for (idx = 0; idx < dev->mem->nregions; idx++) {
  83                 if (region[idx].mapped_address) {
  84                         munmap((void *)(uintptr_t)region[idx].mapped_address,
  85                                         region[idx].mapped_size);
  86                         close(region[idx].fd);
  87                 }
  88         }
  89 }
  90
  91 int
  92 user_set_mem_table(struct vhost_device_ctx ctx, struct VhostUserMsg *pmsg)
  93 {
  94         struct VhostUserMemory memory = pmsg->payload.memory;
  95         struct virtio_memory_regions *pregion;
  96         uint64_t mapped_address, mapped_size;
  97         struct virtio_net *dev;
  98         unsigned int idx = 0;
  99         struct orig_region_map *pregion_orig;
 100         uint64_t alignment;
 101
 102         /* unmap old memory regions one by one*/
 103         dev = get_device(ctx);
 104         if (dev == NULL)
 105                 return -1;
 106
 107         /* Remove from the data plane. */
 108         if (dev->flags & VIRTIO_DEV_RUNNING)
 109                 notify_ops->destroy_device(dev);
 110
 111         if (dev->mem) {
 112                 free_mem_region(dev);
 113                 free(dev->mem);
 114                 dev->mem = NULL;
 115         }
 116
 117         dev->mem = calloc(1,
 118                 sizeof(struct virtio_memory) +
 119                 sizeof(struct virtio_memory_regions) * memory.nregions +
 120                 sizeof(struct orig_region_map) * memory.nregions);
 121         if (dev->mem == NULL) {
 122                 RTE_LOG(ERR, VHOST_CONFIG,
 123                         "(%"PRIu64") Failed to allocate memory for dev->mem\n",
 124                         dev->device_fh);
 125                 return -1;
 126         }
 127         dev->mem->nregions = memory.nregions;
 128
 129         pregion_orig = orig_region(dev->mem, memory.nregions);
 130         for (idx = 0; idx < memory.nregions; idx++) {
 131                 pregion = &dev->mem->regions[idx];
 132                 pregion->guest_phys_address =
 133                         memory.regions[idx].guest_phys_addr;
 134                 pregion->guest_phys_address_end =
 135                         memory.regions[idx].guest_phys_addr +
 136                         memory.regions[idx].memory_size;
 137                 pregion->memory_size =
 138                         memory.regions[idx].memory_size;
 139                 pregion->userspace_address =
 140                         memory.regions[idx].userspace_addr;
 141
 142                 /* This is ugly */
 143                 mapped_size = memory.regions[idx].memory_size +
 144                         memory.regions[idx].mmap_offset;
 145
 146                 /* mmap() without flag of MAP_ANONYMOUS, should be called
 147                  * with length argument aligned with hugepagesz at older
 148                  * longterm version Linux, like 2.6.32 and 3.2.72, or
 149                  * mmap() will fail with EINVAL.
 150                  *
 151                  * to avoid failure, make sure in caller to keep length
 152                  * aligned.
 153                  */
 154                 alignment = get_blk_size(pmsg->fds[idx]);
 155                 mapped_size = RTE_ALIGN_CEIL(mapped_size, alignment);
 156
 157                 mapped_address = (uint64_t)(uintptr_t)mmap(NULL,
 158                         mapped_size,
 159                         PROT_READ | PROT_WRITE, MAP_SHARED,
 160                         pmsg->fds[idx],
 161                         0);
 162
 163                 RTE_LOG(INFO, VHOST_CONFIG,
 164                         "mapped region %d fd:%d to:%p sz:0x%"PRIx64" "
 165                         "off:0x%"PRIx64" align:0x%"PRIx64"\n",
 166                         idx, pmsg->fds[idx], (void *)(uintptr_t)mapped_address,
 167                         mapped_size, memory.regions[idx].mmap_offset,
 168                         alignment);
 169
 170                 if (mapped_address == (uint64_t)(uintptr_t)MAP_FAILED) {
 171                         RTE_LOG(ERR, VHOST_CONFIG,
 172                                 "mmap qemu guest failed.\n");
 173                         goto err_mmap;
 174                 }
 175
 176                 pregion_orig[idx].mapped_address = mapped_address;
 177                 pregion_orig[idx].mapped_size = mapped_size;
 178                 pregion_orig[idx].blksz = alignment;
 179                 pregion_orig[idx].fd = pmsg->fds[idx];
 180
 181                 mapped_address +=  memory.regions[idx].mmap_offset;
 182
 183                 pregion->address_offset = mapped_address -
 184                         pregion->guest_phys_address;
 185
 186                 if (memory.regions[idx].guest_phys_addr == 0) {
 187                         dev->mem->base_address =
 188                                 memory.regions[idx].userspace_addr;
 189                         dev->mem->mapped_address =
 190                                 pregion->address_offset;
 191                 }
 192
 193                 LOG_DEBUG(VHOST_CONFIG,
 194                         "REGION: %u GPA: %p QEMU VA: %p SIZE (%"PRIu64")\n",
 195                         idx,
 196                         (void *)(uintptr_t)pregion->guest_phys_address,
 197                         (void *)(uintptr_t)pregion->userspace_address,
 198                          pregion->memory_size);
 199         }
 200
 201         return 0;
 202
 203 err_mmap:
 204         while (idx--) {
 205                 munmap((void *)(uintptr_t)pregion_orig[idx].mapped_address,
 206                                 pregion_orig[idx].mapped_size);
 207                 close(pregion_orig[idx].fd);
 208         }
 209         free(dev->mem);
 210         dev->mem = NULL;
 211         return -1;
 212 }
 213
 214 static int
 215 vq_is_ready(struct vhost_virtqueue *vq)
 216 {
 217         return vq && vq->desc   &&
 218                vq->kickfd != -1 &&
 219                vq->callfd != -1;
 220 }
 221
 222 static int
 223 virtio_is_ready(struct virtio_net *dev)
 224 {
 225         struct vhost_virtqueue *rvq, *tvq;
 226         uint32_t i;
 227
 228         for (i = 0; i < dev->virt_qp_nb; i++) {
 229                 rvq = dev->virtqueue[i * VIRTIO_QNUM + VIRTIO_RXQ];
 230                 tvq = dev->virtqueue[i * VIRTIO_QNUM + VIRTIO_TXQ];
 231
 232                 if (!vq_is_ready(rvq) || !vq_is_ready(tvq)) {
 233                         RTE_LOG(INFO, VHOST_CONFIG,
 234                                 "virtio is not ready for processing.\n");
 235                         return 0;
 236                 }
 237         }
 238
 239         RTE_LOG(INFO, VHOST_CONFIG,
 240                 "virtio is now ready for processing.\n");
 241         return 1;
 242 }
 243
 244 void
 245 user_set_vring_call(struct vhost_device_ctx ctx, struct VhostUserMsg *pmsg)
 246 {
 247         struct vhost_vring_file file;
 248
 249         file.index = pmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
 250         if (pmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
 251                 file.fd = -1;
 252         else
 253                 file.fd = pmsg->fds[0];
 254         RTE_LOG(INFO, VHOST_CONFIG,
 255                 "vring call idx:%d file:%d\n", file.index, file.fd);
 256         ops->set_vring_call(ctx, &file);
 257 }
 258
 259
 260 /*
 261  *  In vhost-user, when we receive kick message, will test whether virtio
 262  *  device is ready for packet processing.
 263  */
 264 void
 265 user_set_vring_kick(struct vhost_device_ctx ctx, struct VhostUserMsg *pmsg)
 266 {
 267         struct vhost_vring_file file;
 268         struct virtio_net *dev = get_device(ctx);
 269
 270         file.index = pmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
 271         if (pmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
 272                 file.fd = -1;
 273         else
 274                 file.fd = pmsg->fds[0];
 275         RTE_LOG(INFO, VHOST_CONFIG,
 276                 "vring kick idx:%d file:%d\n", file.index, file.fd);
 277         ops->set_vring_kick(ctx, &file);
 278
 279         if (virtio_is_ready(dev) &&
 280                 !(dev->flags & VIRTIO_DEV_RUNNING))
 281                         notify_ops->new_device(dev);
 282 }
 283
 284 /*
 285  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
 286  */
 287 int
 288 user_get_vring_base(struct vhost_device_ctx ctx,
 289         struct vhost_vring_state *state)
 290 {
 291         struct virtio_net *dev = get_device(ctx);
 292
 293         if (dev == NULL)
 294                 return -1;
 295         /* We have to stop the queue (virtio) if it is running. */
 296         if (dev->flags & VIRTIO_DEV_RUNNING)
 297                 notify_ops->destroy_device(dev);
 298
 299         /* Here we are safe to get the last used index */
 300         ops->get_vring_base(ctx, state->index, state);
 301
 302         RTE_LOG(INFO, VHOST_CONFIG,
 303                 "vring base idx:%d file:%d\n", state->index, state->num);
 304         /*
 305          * Based on current qemu vhost-user implementation, this message is
 306          * sent and only sent in vhost_vring_stop.
 307          * TODO: cleanup the vring, it isn't usable since here.
 308          */
 309         if (dev->virtqueue[state->index]->kickfd >= 0) {
 310                 close(dev->virtqueue[state->index]->kickfd);
 311                 dev->virtqueue[state->index]->kickfd = -1;
 312         }
 313
 314         return 0;
 315 }
 316
 317 /*
 318  * when virtio queues are ready to work, qemu will send us to
 319  * enable the virtio queue pair.
 320  */
 321 int
 322 user_set_vring_enable(struct vhost_device_ctx ctx,
 323                       struct vhost_vring_state *state)
 324 {
 325         struct virtio_net *dev = get_device(ctx);
 326         int enable = (int)state->num;
 327
 328         RTE_LOG(INFO, VHOST_CONFIG,
 329                 "set queue enable: %d to qp idx: %d\n",
 330                 enable, state->index);
 331
 332         if (notify_ops->vring_state_changed) {
 333                 notify_ops->vring_state_changed(dev, state->index, enable);
 334         }
 335
 336         dev->virtqueue[state->index]->enabled = enable;
 337
 338         return 0;
 339 }
 340
 341 void
 342 user_destroy_device(struct vhost_device_ctx ctx)
 343 {
 344         struct virtio_net *dev = get_device(ctx);
 345
 346         if (dev && (dev->flags & VIRTIO_DEV_RUNNING))
 347                 notify_ops->destroy_device(dev);
 348
 349         if (dev && dev->mem) {
 350                 free_mem_region(dev);
 351                 free(dev->mem);
 352                 dev->mem = NULL;
 353         }
 354 }
 355
 356 void
 357 user_set_protocol_features(struct vhost_device_ctx ctx,
 358                            uint64_t protocol_features)
 359 {
 360         struct virtio_net *dev;
 361
 362         dev = get_device(ctx);
 363         if (dev == NULL || protocol_features & ~VHOST_USER_PROTOCOL_FEATURES)
 364                 return;
 365
 366         dev->protocol_features = protocol_features;
 367 }
 368
 369 int
 370 user_set_log_base(struct vhost_device_ctx ctx,
 371                  struct VhostUserMsg *msg)
 372 {
 373         struct virtio_net *dev;
 374         int fd = msg->fds[0];
 375         uint64_t size, off;
 376         void *addr;
 377
 378         dev = get_device(ctx);
 379         if (!dev)
 380                 return -1;
 381
 382         if (fd < 0) {
 383                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
 384                 return -1;
 385         }
 386
 387         if (msg->size != sizeof(VhostUserLog)) {
 388                 RTE_LOG(ERR, VHOST_CONFIG,
 389                         "invalid log base msg size: %"PRId32" != %d\n",
 390                         msg->size, (int)sizeof(VhostUserLog));
 391                 return -1;
 392         }
 393
 394         size = msg->payload.log.mmap_size;
 395         off  = msg->payload.log.mmap_offset;
 396         RTE_LOG(INFO, VHOST_CONFIG,
 397                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
 398                 size, off);
 399
 400         /*
 401          * mmap from 0 to workaround a hugepage mmap bug: mmap will
 402          * fail when offset is not page size aligned.
 403          */
 404         addr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 405         if (addr == MAP_FAILED) {
 406                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
 407                 return -1;
 408         }
 409
 410         /* TODO: unmap on stop */
 411         dev->log_base = (uint64_t)(uintptr_t)addr + off;
 412         dev->log_size = size;
 413
 414         return 0;
 415 }