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