lib/librte_vhost/vhost_user.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2016 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 <string.h>
  38 #include <unistd.h>
  39 #include <sys/mman.h>
  40 #include <sys/types.h>
  41 #include <sys/stat.h>
  42 #include <assert.h>
  43 #ifdef RTE_LIBRTE_VHOST_NUMA
  44 #include <numaif.h>
  45 #endif
  46
  47 #include <rte_common.h>
  48 #include <rte_malloc.h>
  49 #include <rte_log.h>
  50
  51 #include "vhost.h"
  52 #include "vhost_user.h"
  53
  54 #define VIRTIO_MIN_MTU 68
  55 #define VIRTIO_MAX_MTU 65535
  56
  57 static const char *vhost_message_str[VHOST_USER_MAX] = {
  58         [VHOST_USER_NONE] = "VHOST_USER_NONE",
  59         [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
  60         [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
  61         [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
  62         [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
  63         [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
  64         [VHOST_USER_SET_LOG_BASE] = "VHOST_USER_SET_LOG_BASE",
  65         [VHOST_USER_SET_LOG_FD] = "VHOST_USER_SET_LOG_FD",
  66         [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
  67         [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
  68         [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
  69         [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
  70         [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
  71         [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
  72         [VHOST_USER_SET_VRING_ERR]  = "VHOST_USER_SET_VRING_ERR",
  73         [VHOST_USER_GET_PROTOCOL_FEATURES]  = "VHOST_USER_GET_PROTOCOL_FEATURES",
  74         [VHOST_USER_SET_PROTOCOL_FEATURES]  = "VHOST_USER_SET_PROTOCOL_FEATURES",
  75         [VHOST_USER_GET_QUEUE_NUM]  = "VHOST_USER_GET_QUEUE_NUM",
  76         [VHOST_USER_SET_VRING_ENABLE]  = "VHOST_USER_SET_VRING_ENABLE",
  77         [VHOST_USER_SEND_RARP]  = "VHOST_USER_SEND_RARP",
  78         [VHOST_USER_NET_SET_MTU]  = "VHOST_USER_NET_SET_MTU",
  79 };
  80
  81 static uint64_t
  82 get_blk_size(int fd)
  83 {
  84         struct stat stat;
  85         int ret;
  86
  87         ret = fstat(fd, &stat);
  88         return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
  89 }
  90
  91 static void
  92 free_mem_region(struct virtio_net *dev)
  93 {
  94         uint32_t i;
  95         struct rte_vhost_mem_region *reg;
  96
  97         if (!dev || !dev->mem)
  98                 return;
  99
 100         for (i = 0; i < dev->mem->nregions; i++) {
 101                 reg = &dev->mem->regions[i];
 102                 if (reg->host_user_addr) {
 103                         munmap(reg->mmap_addr, reg->mmap_size);
 104                         close(reg->fd);
 105                 }
 106         }
 107 }
 108
 109 void
 110 vhost_backend_cleanup(struct virtio_net *dev)
 111 {
 112         if (dev->mem) {
 113                 free_mem_region(dev);
 114                 rte_free(dev->mem);
 115                 dev->mem = NULL;
 116         }
 117
 118         free(dev->guest_pages);
 119         dev->guest_pages = NULL;
 120
 121         if (dev->log_addr) {
 122                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 123                 dev->log_addr = 0;
 124         }
 125 }
 126
 127 /*
 128  * This function just returns success at the moment unless
 129  * the device hasn't been initialised.
 130  */
 131 static int
 132 vhost_user_set_owner(void)
 133 {
 134         return 0;
 135 }
 136
 137 static int
 138 vhost_user_reset_owner(struct virtio_net *dev)
 139 {
 140         if (dev->flags & VIRTIO_DEV_RUNNING) {
 141                 dev->flags &= ~VIRTIO_DEV_RUNNING;
 142                 dev->notify_ops->destroy_device(dev->vid);
 143         }
 144
 145         cleanup_device(dev, 0);
 146         reset_device(dev);
 147         return 0;
 148 }
 149
 150 /*
 151  * The features that we support are requested.
 152  */
 153 static uint64_t
 154 vhost_user_get_features(struct virtio_net *dev)
 155 {
 156         uint64_t features = 0;
 157
 158         rte_vhost_driver_get_features(dev->ifname, &features);
 159         return features;
 160 }
 161
 162 /*
 163  * We receive the negotiated features supported by us and the virtio device.
 164  */
 165 static int
 166 vhost_user_set_features(struct virtio_net *dev, uint64_t features)
 167 {
 168         uint64_t vhost_features = 0;
 169
 170         rte_vhost_driver_get_features(dev->ifname, &vhost_features);
 171         if (features & ~vhost_features) {
 172                 RTE_LOG(ERR, VHOST_CONFIG,
 173                         "(%d) received invalid negotiated features.\n",
 174                         dev->vid);
 175                 return -1;
 176         }
 177
 178         if ((dev->flags & VIRTIO_DEV_RUNNING) && dev->features != features) {
 179                 if (dev->notify_ops->features_changed)
 180                         dev->notify_ops->features_changed(dev->vid, features);
 181         }
 182
 183         dev->features = features;
 184         if (dev->features &
 185                 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
 186                 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 187         } else {
 188                 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
 189         }
 190         LOG_DEBUG(VHOST_CONFIG,
 191                 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
 192                 dev->vid,
 193                 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
 194                 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 195
 196         return 0;
 197 }
 198
 199 /*
 200  * The virtio device sends us the size of the descriptor ring.
 201  */
 202 static int
 203 vhost_user_set_vring_num(struct virtio_net *dev,
 204                          VhostUserMsg *msg)
 205 {
 206         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 207
 208         vq->size = msg->payload.state.num;
 209
 210         if (dev->dequeue_zero_copy) {
 211                 vq->nr_zmbuf = 0;
 212                 vq->last_zmbuf_idx = 0;
 213                 vq->zmbuf_size = vq->size;
 214                 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
 215                                          sizeof(struct zcopy_mbuf), 0);
 216                 if (vq->zmbufs == NULL) {
 217                         RTE_LOG(WARNING, VHOST_CONFIG,
 218                                 "failed to allocate mem for zero copy; "
 219                                 "zero copy is force disabled\n");
 220                         dev->dequeue_zero_copy = 0;
 221                 }
 222         }
 223
 224         vq->shadow_used_ring = rte_malloc(NULL,
 225                                 vq->size * sizeof(struct vring_used_elem),
 226                                 RTE_CACHE_LINE_SIZE);
 227         if (!vq->shadow_used_ring) {
 228                 RTE_LOG(ERR, VHOST_CONFIG,
 229                         "failed to allocate memory for shadow used ring.\n");
 230                 return -1;
 231         }
 232
 233         return 0;
 234 }
 235
 236 /*
 237  * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
 238  * same numa node as the memory of vring descriptor.
 239  */
 240 #ifdef RTE_LIBRTE_VHOST_NUMA
 241 static struct virtio_net*
 242 numa_realloc(struct virtio_net *dev, int index)
 243 {
 244         int oldnode, newnode;
 245         struct virtio_net *old_dev;
 246         struct vhost_virtqueue *old_vq, *vq;
 247         int ret;
 248
 249         old_dev = dev;
 250         vq = old_vq = dev->virtqueue[index];
 251
 252         ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
 253                             MPOL_F_NODE | MPOL_F_ADDR);
 254
 255         /* check if we need to reallocate vq */
 256         ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
 257                              MPOL_F_NODE | MPOL_F_ADDR);
 258         if (ret) {
 259                 RTE_LOG(ERR, VHOST_CONFIG,
 260                         "Unable to get vq numa information.\n");
 261                 return dev;
 262         }
 263         if (oldnode != newnode) {
 264                 RTE_LOG(INFO, VHOST_CONFIG,
 265                         "reallocate vq from %d to %d node\n", oldnode, newnode);
 266                 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
 267                 if (!vq)
 268                         return dev;
 269
 270                 memcpy(vq, old_vq, sizeof(*vq));
 271                 rte_free(old_vq);
 272         }
 273
 274         /* check if we need to reallocate dev */
 275         ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
 276                             MPOL_F_NODE | MPOL_F_ADDR);
 277         if (ret) {
 278                 RTE_LOG(ERR, VHOST_CONFIG,
 279                         "Unable to get dev numa information.\n");
 280                 goto out;
 281         }
 282         if (oldnode != newnode) {
 283                 RTE_LOG(INFO, VHOST_CONFIG,
 284                         "reallocate dev from %d to %d node\n",
 285                         oldnode, newnode);
 286                 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
 287                 if (!dev) {
 288                         dev = old_dev;
 289                         goto out;
 290                 }
 291
 292                 memcpy(dev, old_dev, sizeof(*dev));
 293                 rte_free(old_dev);
 294         }
 295
 296 out:
 297         dev->virtqueue[index] = vq;
 298         vhost_devices[dev->vid] = dev;
 299
 300         return dev;
 301 }
 302 #else
 303 static struct virtio_net*
 304 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 305 {
 306         return dev;
 307 }
 308 #endif
 309
 310 /*
 311  * Converts QEMU virtual address to Vhost virtual address. This function is
 312  * used to convert the ring addresses to our address space.
 313  */
 314 static uint64_t
 315 qva_to_vva(struct virtio_net *dev, uint64_t qva)
 316 {
 317         struct rte_vhost_mem_region *reg;
 318         uint32_t i;
 319
 320         /* Find the region where the address lives. */
 321         for (i = 0; i < dev->mem->nregions; i++) {
 322                 reg = &dev->mem->regions[i];
 323
 324                 if (qva >= reg->guest_user_addr &&
 325                     qva <  reg->guest_user_addr + reg->size) {
 326                         return qva - reg->guest_user_addr +
 327                                reg->host_user_addr;
 328                 }
 329         }
 330
 331         return 0;
 332 }
 333
 334 /*
 335  * The virtio device sends us the desc, used and avail ring addresses.
 336  * This function then converts these to our address space.
 337  */
 338 static int
 339 vhost_user_set_vring_addr(struct virtio_net *dev, VhostUserMsg *msg)
 340 {
 341         struct vhost_virtqueue *vq;
 342
 343         if (dev->mem == NULL)
 344                 return -1;
 345
 346         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 347         vq = dev->virtqueue[msg->payload.addr.index];
 348
 349         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 350         vq->desc = (struct vring_desc *)(uintptr_t)qva_to_vva(dev,
 351                         msg->payload.addr.desc_user_addr);
 352         if (vq->desc == 0) {
 353                 RTE_LOG(ERR, VHOST_CONFIG,
 354                         "(%d) failed to find desc ring address.\n",
 355                         dev->vid);
 356                 return -1;
 357         }
 358
 359         dev = numa_realloc(dev, msg->payload.addr.index);
 360         vq = dev->virtqueue[msg->payload.addr.index];
 361
 362         vq->avail = (struct vring_avail *)(uintptr_t)qva_to_vva(dev,
 363                         msg->payload.addr.avail_user_addr);
 364         if (vq->avail == 0) {
 365                 RTE_LOG(ERR, VHOST_CONFIG,
 366                         "(%d) failed to find avail ring address.\n",
 367                         dev->vid);
 368                 return -1;
 369         }
 370
 371         vq->used = (struct vring_used *)(uintptr_t)qva_to_vva(dev,
 372                         msg->payload.addr.used_user_addr);
 373         if (vq->used == 0) {
 374                 RTE_LOG(ERR, VHOST_CONFIG,
 375                         "(%d) failed to find used ring address.\n",
 376                         dev->vid);
 377                 return -1;
 378         }
 379
 380         if (vq->last_used_idx != vq->used->idx) {
 381                 RTE_LOG(WARNING, VHOST_CONFIG,
 382                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 383                         "some packets maybe resent for Tx and dropped for Rx\n",
 384                         vq->last_used_idx, vq->used->idx);
 385                 vq->last_used_idx  = vq->used->idx;
 386                 vq->last_avail_idx = vq->used->idx;
 387         }
 388
 389         vq->log_guest_addr = msg->payload.addr.log_guest_addr;
 390
 391         LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
 392                         dev->vid, vq->desc);
 393         LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
 394                         dev->vid, vq->avail);
 395         LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
 396                         dev->vid, vq->used);
 397         LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
 398                         dev->vid, vq->log_guest_addr);
 399
 400         return 0;
 401 }
 402
 403 /*
 404  * The virtio device sends us the available ring last used index.
 405  */
 406 static int
 407 vhost_user_set_vring_base(struct virtio_net *dev,
 408                           VhostUserMsg *msg)
 409 {
 410         dev->virtqueue[msg->payload.state.index]->last_used_idx  =
 411                         msg->payload.state.num;
 412         dev->virtqueue[msg->payload.state.index]->last_avail_idx =
 413                         msg->payload.state.num;
 414
 415         return 0;
 416 }
 417
 418 static void
 419 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 420                    uint64_t host_phys_addr, uint64_t size)
 421 {
 422         struct guest_page *page, *last_page;
 423
 424         if (dev->nr_guest_pages == dev->max_guest_pages) {
 425                 dev->max_guest_pages *= 2;
 426                 dev->guest_pages = realloc(dev->guest_pages,
 427                                         dev->max_guest_pages * sizeof(*page));
 428         }
 429
 430         if (dev->nr_guest_pages > 0) {
 431                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 432                 /* merge if the two pages are continuous */
 433                 if (host_phys_addr == last_page->host_phys_addr +
 434                                       last_page->size) {
 435                         last_page->size += size;
 436                         return;
 437                 }
 438         }
 439
 440         page = &dev->guest_pages[dev->nr_guest_pages++];
 441         page->guest_phys_addr = guest_phys_addr;
 442         page->host_phys_addr  = host_phys_addr;
 443         page->size = size;
 444 }
 445
 446 static void
 447 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 448                 uint64_t page_size)
 449 {
 450         uint64_t reg_size = reg->size;
 451         uint64_t host_user_addr  = reg->host_user_addr;
 452         uint64_t guest_phys_addr = reg->guest_phys_addr;
 453         uint64_t host_phys_addr;
 454         uint64_t size;
 455
 456         host_phys_addr = rte_mem_virt2phy((void *)(uintptr_t)host_user_addr);
 457         size = page_size - (guest_phys_addr & (page_size - 1));
 458         size = RTE_MIN(size, reg_size);
 459
 460         add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size);
 461         host_user_addr  += size;
 462         guest_phys_addr += size;
 463         reg_size -= size;
 464
 465         while (reg_size > 0) {
 466                 size = RTE_MIN(reg_size, page_size);
 467                 host_phys_addr = rte_mem_virt2phy((void *)(uintptr_t)
 468                                                   host_user_addr);
 469                 add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size);
 470
 471                 host_user_addr  += size;
 472                 guest_phys_addr += size;
 473                 reg_size -= size;
 474         }
 475 }
 476
 477 #ifdef RTE_LIBRTE_VHOST_DEBUG
 478 /* TODO: enable it only in debug mode? */
 479 static void
 480 dump_guest_pages(struct virtio_net *dev)
 481 {
 482         uint32_t i;
 483         struct guest_page *page;
 484
 485         for (i = 0; i < dev->nr_guest_pages; i++) {
 486                 page = &dev->guest_pages[i];
 487
 488                 RTE_LOG(INFO, VHOST_CONFIG,
 489                         "guest physical page region %u\n"
 490                         "\t guest_phys_addr: %" PRIx64 "\n"
 491                         "\t host_phys_addr : %" PRIx64 "\n"
 492                         "\t size           : %" PRIx64 "\n",
 493                         i,
 494                         page->guest_phys_addr,
 495                         page->host_phys_addr,
 496                         page->size);
 497         }
 498 }
 499 #else
 500 #define dump_guest_pages(dev)
 501 #endif
 502
 503 static int
 504 vhost_user_set_mem_table(struct virtio_net *dev, struct VhostUserMsg *pmsg)
 505 {
 506         struct VhostUserMemory memory = pmsg->payload.memory;
 507         struct rte_vhost_mem_region *reg;
 508         void *mmap_addr;
 509         uint64_t mmap_size;
 510         uint64_t mmap_offset;
 511         uint64_t alignment;
 512         uint32_t i;
 513         int fd;
 514
 515         if (dev->mem) {
 516                 free_mem_region(dev);
 517                 rte_free(dev->mem);
 518                 dev->mem = NULL;
 519         }
 520
 521         dev->nr_guest_pages = 0;
 522         if (!dev->guest_pages) {
 523                 dev->max_guest_pages = 8;
 524                 dev->guest_pages = malloc(dev->max_guest_pages *
 525                                                 sizeof(struct guest_page));
 526                 if (dev->guest_pages == NULL) {
 527                         RTE_LOG(ERR, VHOST_CONFIG,
 528                                 "(%d) failed to allocate memory "
 529                                 "for dev->guest_pages\n",
 530                                 dev->vid);
 531                         return -1;
 532                 }
 533         }
 534
 535         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
 536                 sizeof(struct rte_vhost_mem_region) * memory.nregions, 0);
 537         if (dev->mem == NULL) {
 538                 RTE_LOG(ERR, VHOST_CONFIG,
 539                         "(%d) failed to allocate memory for dev->mem\n",
 540                         dev->vid);
 541                 return -1;
 542         }
 543         dev->mem->nregions = memory.nregions;
 544
 545         for (i = 0; i < memory.nregions; i++) {
 546                 fd  = pmsg->fds[i];
 547                 reg = &dev->mem->regions[i];
 548
 549                 reg->guest_phys_addr = memory.regions[i].guest_phys_addr;
 550                 reg->guest_user_addr = memory.regions[i].userspace_addr;
 551                 reg->size            = memory.regions[i].memory_size;
 552                 reg->fd              = fd;
 553
 554                 mmap_offset = memory.regions[i].mmap_offset;
 555                 mmap_size   = reg->size + mmap_offset;
 556
 557                 /* mmap() without flag of MAP_ANONYMOUS, should be called
 558                  * with length argument aligned with hugepagesz at older
 559                  * longterm version Linux, like 2.6.32 and 3.2.72, or
 560                  * mmap() will fail with EINVAL.
 561                  *
 562                  * to avoid failure, make sure in caller to keep length
 563                  * aligned.
 564                  */
 565                 alignment = get_blk_size(fd);
 566                 if (alignment == (uint64_t)-1) {
 567                         RTE_LOG(ERR, VHOST_CONFIG,
 568                                 "couldn't get hugepage size through fstat\n");
 569                         goto err_mmap;
 570                 }
 571                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
 572
 573                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
 574                                  MAP_SHARED | MAP_POPULATE, fd, 0);
 575
 576                 if (mmap_addr == MAP_FAILED) {
 577                         RTE_LOG(ERR, VHOST_CONFIG,
 578                                 "mmap region %u failed.\n", i);
 579                         goto err_mmap;
 580                 }
 581
 582                 reg->mmap_addr = mmap_addr;
 583                 reg->mmap_size = mmap_size;
 584                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
 585                                       mmap_offset;
 586
 587                 if (dev->dequeue_zero_copy)
 588                         add_guest_pages(dev, reg, alignment);
 589
 590                 RTE_LOG(INFO, VHOST_CONFIG,
 591                         "guest memory region %u, size: 0x%" PRIx64 "\n"
 592                         "\t guest physical addr: 0x%" PRIx64 "\n"
 593                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
 594                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
 595                         "\t mmap addr : 0x%" PRIx64 "\n"
 596                         "\t mmap size : 0x%" PRIx64 "\n"
 597                         "\t mmap align: 0x%" PRIx64 "\n"
 598                         "\t mmap off  : 0x%" PRIx64 "\n",
 599                         i, reg->size,
 600                         reg->guest_phys_addr,
 601                         reg->guest_user_addr,
 602                         reg->host_user_addr,
 603                         (uint64_t)(uintptr_t)mmap_addr,
 604                         mmap_size,
 605                         alignment,
 606                         mmap_offset);
 607         }
 608
 609         dump_guest_pages(dev);
 610
 611         return 0;
 612
 613 err_mmap:
 614         free_mem_region(dev);
 615         rte_free(dev->mem);
 616         dev->mem = NULL;
 617         return -1;
 618 }
 619
 620 static int
 621 vq_is_ready(struct vhost_virtqueue *vq)
 622 {
 623         return vq && vq->desc   &&
 624                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
 625                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
 626 }
 627
 628 static int
 629 virtio_is_ready(struct virtio_net *dev)
 630 {
 631         struct vhost_virtqueue *vq;
 632         uint32_t i;
 633
 634         if (dev->nr_vring == 0)
 635                 return 0;
 636
 637         for (i = 0; i < dev->nr_vring; i++) {
 638                 vq = dev->virtqueue[i];
 639
 640                 if (!vq_is_ready(vq))
 641                         return 0;
 642         }
 643
 644         RTE_LOG(INFO, VHOST_CONFIG,
 645                 "virtio is now ready for processing.\n");
 646         return 1;
 647 }
 648
 649 static void
 650 vhost_user_set_vring_call(struct virtio_net *dev, struct VhostUserMsg *pmsg)
 651 {
 652         struct vhost_vring_file file;
 653         struct vhost_virtqueue *vq;
 654
 655         file.index = pmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
 656         if (pmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
 657                 file.fd = VIRTIO_INVALID_EVENTFD;
 658         else
 659                 file.fd = pmsg->fds[0];
 660         RTE_LOG(INFO, VHOST_CONFIG,
 661                 "vring call idx:%d file:%d\n", file.index, file.fd);
 662
 663         vq = dev->virtqueue[file.index];
 664         if (vq->callfd >= 0)
 665                 close(vq->callfd);
 666
 667         vq->callfd = file.fd;
 668 }
 669
 670 static void
 671 vhost_user_set_vring_kick(struct virtio_net *dev, struct VhostUserMsg *pmsg)
 672 {
 673         struct vhost_vring_file file;
 674         struct vhost_virtqueue *vq;
 675
 676         file.index = pmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
 677         if (pmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
 678                 file.fd = VIRTIO_INVALID_EVENTFD;
 679         else
 680                 file.fd = pmsg->fds[0];
 681         RTE_LOG(INFO, VHOST_CONFIG,
 682                 "vring kick idx:%d file:%d\n", file.index, file.fd);
 683
 684         vq = dev->virtqueue[file.index];
 685         if (vq->kickfd >= 0)
 686                 close(vq->kickfd);
 687         vq->kickfd = file.fd;
 688 }
 689
 690 static void
 691 free_zmbufs(struct vhost_virtqueue *vq)
 692 {
 693         struct zcopy_mbuf *zmbuf, *next;
 694
 695         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
 696              zmbuf != NULL; zmbuf = next) {
 697                 next = TAILQ_NEXT(zmbuf, next);
 698
 699                 rte_pktmbuf_free(zmbuf->mbuf);
 700                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
 701         }
 702
 703         rte_free(vq->zmbufs);
 704 }
 705
 706 /*
 707  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
 708  */
 709 static int
 710 vhost_user_get_vring_base(struct virtio_net *dev,
 711                           VhostUserMsg *msg)
 712 {
 713         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 714
 715         /* We have to stop the queue (virtio) if it is running. */
 716         if (dev->flags & VIRTIO_DEV_RUNNING) {
 717                 dev->flags &= ~VIRTIO_DEV_RUNNING;
 718                 dev->notify_ops->destroy_device(dev->vid);
 719         }
 720
 721         dev->flags &= ~VIRTIO_DEV_READY;
 722
 723         /* Here we are safe to get the last used index */
 724         msg->payload.state.num = vq->last_used_idx;
 725
 726         RTE_LOG(INFO, VHOST_CONFIG,
 727                 "vring base idx:%d file:%d\n", msg->payload.state.index,
 728                 msg->payload.state.num);
 729         /*
 730          * Based on current qemu vhost-user implementation, this message is
 731          * sent and only sent in vhost_vring_stop.
 732          * TODO: cleanup the vring, it isn't usable since here.
 733          */
 734         if (vq->kickfd >= 0)
 735                 close(vq->kickfd);
 736
 737         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
 738
 739         if (dev->dequeue_zero_copy)
 740                 free_zmbufs(vq);
 741         rte_free(vq->shadow_used_ring);
 742         vq->shadow_used_ring = NULL;
 743
 744         return 0;
 745 }
 746
 747 /*
 748  * when virtio queues are ready to work, qemu will send us to
 749  * enable the virtio queue pair.
 750  */
 751 static int
 752 vhost_user_set_vring_enable(struct virtio_net *dev,
 753                             VhostUserMsg *msg)
 754 {
 755         int enable = (int)msg->payload.state.num;
 756
 757         RTE_LOG(INFO, VHOST_CONFIG,
 758                 "set queue enable: %d to qp idx: %d\n",
 759                 enable, msg->payload.state.index);
 760
 761         if (dev->notify_ops->vring_state_changed)
 762                 dev->notify_ops->vring_state_changed(dev->vid,
 763                                 msg->payload.state.index, enable);
 764
 765         dev->virtqueue[msg->payload.state.index]->enabled = enable;
 766
 767         return 0;
 768 }
 769
 770 static void
 771 vhost_user_set_protocol_features(struct virtio_net *dev,
 772                                  uint64_t protocol_features)
 773 {
 774         if (protocol_features & ~VHOST_USER_PROTOCOL_FEATURES)
 775                 return;
 776
 777         dev->protocol_features = protocol_features;
 778 }
 779
 780 static int
 781 vhost_user_set_log_base(struct virtio_net *dev, struct VhostUserMsg *msg)
 782 {
 783         int fd = msg->fds[0];
 784         uint64_t size, off;
 785         void *addr;
 786
 787         if (fd < 0) {
 788                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
 789                 return -1;
 790         }
 791
 792         if (msg->size != sizeof(VhostUserLog)) {
 793                 RTE_LOG(ERR, VHOST_CONFIG,
 794                         "invalid log base msg size: %"PRId32" != %d\n",
 795                         msg->size, (int)sizeof(VhostUserLog));
 796                 return -1;
 797         }
 798
 799         size = msg->payload.log.mmap_size;
 800         off  = msg->payload.log.mmap_offset;
 801         RTE_LOG(INFO, VHOST_CONFIG,
 802                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
 803                 size, off);
 804
 805         /*
 806          * mmap from 0 to workaround a hugepage mmap bug: mmap will
 807          * fail when offset is not page size aligned.
 808          */
 809         addr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 810         close(fd);
 811         if (addr == MAP_FAILED) {
 812                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
 813                 return -1;
 814         }
 815
 816         /*
 817          * Free previously mapped log memory on occasionally
 818          * multiple VHOST_USER_SET_LOG_BASE.
 819          */
 820         if (dev->log_addr) {
 821                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 822         }
 823         dev->log_addr = (uint64_t)(uintptr_t)addr;
 824         dev->log_base = dev->log_addr + off;
 825         dev->log_size = size;
 826
 827         return 0;
 828 }
 829
 830 /*
 831  * An rarp packet is constructed and broadcasted to notify switches about
 832  * the new location of the migrated VM, so that packets from outside will
 833  * not be lost after migration.
 834  *
 835  * However, we don't actually "send" a rarp packet here, instead, we set
 836  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
 837  */
 838 static int
 839 vhost_user_send_rarp(struct virtio_net *dev, struct VhostUserMsg *msg)
 840 {
 841         uint8_t *mac = (uint8_t *)&msg->payload.u64;
 842
 843         RTE_LOG(DEBUG, VHOST_CONFIG,
 844                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
 845                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
 846         memcpy(dev->mac.addr_bytes, mac, 6);
 847
 848         /*
 849          * Set the flag to inject a RARP broadcast packet at
 850          * rte_vhost_dequeue_burst().
 851          *
 852          * rte_smp_wmb() is for making sure the mac is copied
 853          * before the flag is set.
 854          */
 855         rte_smp_wmb();
 856         rte_atomic16_set(&dev->broadcast_rarp, 1);
 857
 858         return 0;
 859 }
 860
 861 static int
 862 vhost_user_net_set_mtu(struct virtio_net *dev, struct VhostUserMsg *msg)
 863 {
 864         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
 865                         msg->payload.u64 > VIRTIO_MAX_MTU) {
 866                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
 867                                 msg->payload.u64);
 868
 869                 return -1;
 870         }
 871
 872         dev->mtu = msg->payload.u64;
 873
 874         return 0;
 875 }
 876
 877 /* return bytes# of read on success or negative val on failure. */
 878 static int
 879 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
 880 {
 881         int ret;
 882
 883         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
 884                 msg->fds, VHOST_MEMORY_MAX_NREGIONS);
 885         if (ret <= 0)
 886                 return ret;
 887
 888         if (msg && msg->size) {
 889                 if (msg->size > sizeof(msg->payload)) {
 890                         RTE_LOG(ERR, VHOST_CONFIG,
 891                                 "invalid msg size: %d\n", msg->size);
 892                         return -1;
 893                 }
 894                 ret = read(sockfd, &msg->payload, msg->size);
 895                 if (ret <= 0)
 896                         return ret;
 897                 if (ret != (int)msg->size) {
 898                         RTE_LOG(ERR, VHOST_CONFIG,
 899                                 "read control message failed\n");
 900                         return -1;
 901                 }
 902         }
 903
 904         return ret;
 905 }
 906
 907 static int
 908 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
 909 {
 910         int ret;
 911
 912         if (!msg)
 913                 return 0;
 914
 915         msg->flags &= ~VHOST_USER_VERSION_MASK;
 916         msg->flags &= ~VHOST_USER_NEED_REPLY;
 917         msg->flags |= VHOST_USER_VERSION;
 918         msg->flags |= VHOST_USER_REPLY_MASK;
 919
 920         ret = send_fd_message(sockfd, (char *)msg,
 921                 VHOST_USER_HDR_SIZE + msg->size, NULL, 0);
 922
 923         return ret;
 924 }
 925
 926 /*
 927  * Allocate a queue pair if it hasn't been allocated yet
 928  */
 929 static int
 930 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev, VhostUserMsg *msg)
 931 {
 932         uint16_t vring_idx;
 933
 934         switch (msg->request) {
 935         case VHOST_USER_SET_VRING_KICK:
 936         case VHOST_USER_SET_VRING_CALL:
 937         case VHOST_USER_SET_VRING_ERR:
 938                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
 939                 break;
 940         case VHOST_USER_SET_VRING_NUM:
 941         case VHOST_USER_SET_VRING_BASE:
 942         case VHOST_USER_SET_VRING_ENABLE:
 943                 vring_idx = msg->payload.state.index;
 944                 break;
 945         case VHOST_USER_SET_VRING_ADDR:
 946                 vring_idx = msg->payload.addr.index;
 947                 break;
 948         default:
 949                 return 0;
 950         }
 951
 952         if (vring_idx >= VHOST_MAX_VRING) {
 953                 RTE_LOG(ERR, VHOST_CONFIG,
 954                         "invalid vring index: %u\n", vring_idx);
 955                 return -1;
 956         }
 957
 958         if (dev->virtqueue[vring_idx])
 959                 return 0;
 960
 961         return alloc_vring_queue(dev, vring_idx);
 962 }
 963
 964 int
 965 vhost_user_msg_handler(int vid, int fd)
 966 {
 967         struct virtio_net *dev;
 968         struct VhostUserMsg msg;
 969         int ret;
 970
 971         dev = get_device(vid);
 972         if (dev == NULL)
 973                 return -1;
 974
 975         if (!dev->notify_ops) {
 976                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
 977                 if (!dev->notify_ops) {
 978                         RTE_LOG(ERR, VHOST_CONFIG,
 979                                 "failed to get callback ops for driver %s\n",
 980                                 dev->ifname);
 981                         return -1;
 982                 }
 983         }
 984
 985         ret = read_vhost_message(fd, &msg);
 986         if (ret <= 0 || msg.request >= VHOST_USER_MAX) {
 987                 if (ret < 0)
 988                         RTE_LOG(ERR, VHOST_CONFIG,
 989                                 "vhost read message failed\n");
 990                 else if (ret == 0)
 991                         RTE_LOG(INFO, VHOST_CONFIG,
 992                                 "vhost peer closed\n");
 993                 else
 994                         RTE_LOG(ERR, VHOST_CONFIG,
 995                                 "vhost read incorrect message\n");
 996
 997                 return -1;
 998         }
 999
1000         ret = 0;
1001         RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1002                 vhost_message_str[msg.request]);
1003
1004         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1005         if (ret < 0) {
1006                 RTE_LOG(ERR, VHOST_CONFIG,
1007                         "failed to alloc queue\n");
1008                 return -1;
1009         }
1010
1011         switch (msg.request) {
1012         case VHOST_USER_GET_FEATURES:
1013                 msg.payload.u64 = vhost_user_get_features(dev);
1014                 msg.size = sizeof(msg.payload.u64);
1015                 send_vhost_message(fd, &msg);
1016                 break;
1017         case VHOST_USER_SET_FEATURES:
1018                 vhost_user_set_features(dev, msg.payload.u64);
1019                 break;
1020
1021         case VHOST_USER_GET_PROTOCOL_FEATURES:
1022                 msg.payload.u64 = VHOST_USER_PROTOCOL_FEATURES;
1023                 msg.size = sizeof(msg.payload.u64);
1024                 send_vhost_message(fd, &msg);
1025                 break;
1026         case VHOST_USER_SET_PROTOCOL_FEATURES:
1027                 vhost_user_set_protocol_features(dev, msg.payload.u64);
1028                 break;
1029
1030         case VHOST_USER_SET_OWNER:
1031                 vhost_user_set_owner();
1032                 break;
1033         case VHOST_USER_RESET_OWNER:
1034                 vhost_user_reset_owner(dev);
1035                 break;
1036
1037         case VHOST_USER_SET_MEM_TABLE:
1038                 ret = vhost_user_set_mem_table(dev, &msg);
1039                 break;
1040
1041         case VHOST_USER_SET_LOG_BASE:
1042                 vhost_user_set_log_base(dev, &msg);
1043
1044                 /* it needs a reply */
1045                 msg.size = sizeof(msg.payload.u64);
1046                 send_vhost_message(fd, &msg);
1047                 break;
1048         case VHOST_USER_SET_LOG_FD:
1049                 close(msg.fds[0]);
1050                 RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1051                 break;
1052
1053         case VHOST_USER_SET_VRING_NUM:
1054                 vhost_user_set_vring_num(dev, &msg);
1055                 break;
1056         case VHOST_USER_SET_VRING_ADDR:
1057                 vhost_user_set_vring_addr(dev, &msg);
1058                 break;
1059         case VHOST_USER_SET_VRING_BASE:
1060                 vhost_user_set_vring_base(dev, &msg);
1061                 break;
1062
1063         case VHOST_USER_GET_VRING_BASE:
1064                 vhost_user_get_vring_base(dev, &msg);
1065                 msg.size = sizeof(msg.payload.state);
1066                 send_vhost_message(fd, &msg);
1067                 break;
1068
1069         case VHOST_USER_SET_VRING_KICK:
1070                 vhost_user_set_vring_kick(dev, &msg);
1071                 break;
1072         case VHOST_USER_SET_VRING_CALL:
1073                 vhost_user_set_vring_call(dev, &msg);
1074                 break;
1075
1076         case VHOST_USER_SET_VRING_ERR:
1077                 if (!(msg.payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1078                         close(msg.fds[0]);
1079                 RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1080                 break;
1081
1082         case VHOST_USER_GET_QUEUE_NUM:
1083                 msg.payload.u64 = VHOST_MAX_QUEUE_PAIRS;
1084                 msg.size = sizeof(msg.payload.u64);
1085                 send_vhost_message(fd, &msg);
1086                 break;
1087
1088         case VHOST_USER_SET_VRING_ENABLE:
1089                 vhost_user_set_vring_enable(dev, &msg);
1090                 break;
1091         case VHOST_USER_SEND_RARP:
1092                 vhost_user_send_rarp(dev, &msg);
1093                 break;
1094
1095         case VHOST_USER_NET_SET_MTU:
1096                 ret = vhost_user_net_set_mtu(dev, &msg);
1097                 break;
1098
1099         default:
1100                 ret = -1;
1101                 break;
1102
1103         }
1104
1105         if (msg.flags & VHOST_USER_NEED_REPLY) {
1106                 msg.payload.u64 = !!ret;
1107                 msg.size = sizeof(msg.payload.u64);
1108                 send_vhost_message(fd, &msg);
1109         }
1110
1111         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
1112                 dev->flags |= VIRTIO_DEV_READY;
1113
1114                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
1115                         if (dev->dequeue_zero_copy) {
1116                                 RTE_LOG(INFO, VHOST_CONFIG,
1117                                                 "dequeue zero copy is enabled\n");
1118                         }
1119
1120                         if (dev->notify_ops->new_device(dev->vid) == 0)
1121                                 dev->flags |= VIRTIO_DEV_RUNNING;
1122                 }
1123         }
1124
1125         return 0;
1126 }