lib/librte_vhost/vhost_user.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2018 Intel Corporation
   3  */
   4
   5 /* Security model
   6  * --------------
   7  * The vhost-user protocol connection is an external interface, so it must be
   8  * robust against invalid inputs.
   9  *
  10  * This is important because the vhost-user master is only one step removed
  11  * from the guest.  Malicious guests that have escaped will then launch further
  12  * attacks from the vhost-user master.
  13  *
  14  * Even in deployments where guests are trusted, a bug in the vhost-user master
  15  * can still cause invalid messages to be sent.  Such messages must not
  16  * compromise the stability of the DPDK application by causing crashes, memory
  17  * corruption, or other problematic behavior.
  18  *
  19  * Do not assume received VhostUserMsg fields contain sensible values!
  20  */
  21
  22 #include <stdint.h>
  23 #include <stdio.h>
  24 #include <stdlib.h>
  25 #include <string.h>
  26 #include <unistd.h>
  27 #include <fcntl.h>
  28 #include <sys/ioctl.h>
  29 #include <sys/mman.h>
  30 #include <sys/types.h>
  31 #include <sys/stat.h>
  32 #include <sys/syscall.h>
  33 #include <assert.h>
  34 #ifdef RTE_LIBRTE_VHOST_NUMA
  35 #include <numaif.h>
  36 #endif
  37 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
  38 #include <linux/userfaultfd.h>
  39 #endif
  40
  41 #include <rte_common.h>
  42 #include <rte_malloc.h>
  43 #include <rte_log.h>
  44
  45 #include "iotlb.h"
  46 #include "vhost.h"
  47 #include "vhost_user.h"
  48
  49 #define VIRTIO_MIN_MTU 68
  50 #define VIRTIO_MAX_MTU 65535
  51
  52 static const char *vhost_message_str[VHOST_USER_MAX] = {
  53         [VHOST_USER_NONE] = "VHOST_USER_NONE",
  54         [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
  55         [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
  56         [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
  57         [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
  58         [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
  59         [VHOST_USER_SET_LOG_BASE] = "VHOST_USER_SET_LOG_BASE",
  60         [VHOST_USER_SET_LOG_FD] = "VHOST_USER_SET_LOG_FD",
  61         [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
  62         [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
  63         [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
  64         [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
  65         [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
  66         [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
  67         [VHOST_USER_SET_VRING_ERR]  = "VHOST_USER_SET_VRING_ERR",
  68         [VHOST_USER_GET_PROTOCOL_FEATURES]  = "VHOST_USER_GET_PROTOCOL_FEATURES",
  69         [VHOST_USER_SET_PROTOCOL_FEATURES]  = "VHOST_USER_SET_PROTOCOL_FEATURES",
  70         [VHOST_USER_GET_QUEUE_NUM]  = "VHOST_USER_GET_QUEUE_NUM",
  71         [VHOST_USER_SET_VRING_ENABLE]  = "VHOST_USER_SET_VRING_ENABLE",
  72         [VHOST_USER_SEND_RARP]  = "VHOST_USER_SEND_RARP",
  73         [VHOST_USER_NET_SET_MTU]  = "VHOST_USER_NET_SET_MTU",
  74         [VHOST_USER_SET_SLAVE_REQ_FD]  = "VHOST_USER_SET_SLAVE_REQ_FD",
  75         [VHOST_USER_IOTLB_MSG]  = "VHOST_USER_IOTLB_MSG",
  76         [VHOST_USER_CRYPTO_CREATE_SESS] = "VHOST_USER_CRYPTO_CREATE_SESS",
  77         [VHOST_USER_CRYPTO_CLOSE_SESS] = "VHOST_USER_CRYPTO_CLOSE_SESS",
  78         [VHOST_USER_POSTCOPY_ADVISE]  = "VHOST_USER_POSTCOPY_ADVISE",
  79         [VHOST_USER_POSTCOPY_LISTEN]  = "VHOST_USER_POSTCOPY_LISTEN",
  80         [VHOST_USER_POSTCOPY_END]  = "VHOST_USER_POSTCOPY_END",
  81 };
  82
  83 static int send_vhost_reply(int sockfd, struct VhostUserMsg *msg);
  84 static int read_vhost_message(int sockfd, struct VhostUserMsg *msg);
  85
  86 static uint64_t
  87 get_blk_size(int fd)
  88 {
  89         struct stat stat;
  90         int ret;
  91
  92         ret = fstat(fd, &stat);
  93         return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
  94 }
  95
  96 static void
  97 free_mem_region(struct virtio_net *dev)
  98 {
  99         uint32_t i;
 100         struct rte_vhost_mem_region *reg;
 101
 102         if (!dev || !dev->mem)
 103                 return;
 104
 105         for (i = 0; i < dev->mem->nregions; i++) {
 106                 reg = &dev->mem->regions[i];
 107                 if (reg->host_user_addr) {
 108                         munmap(reg->mmap_addr, reg->mmap_size);
 109                         close(reg->fd);
 110                 }
 111         }
 112 }
 113
 114 void
 115 vhost_backend_cleanup(struct virtio_net *dev)
 116 {
 117         if (dev->mem) {
 118                 free_mem_region(dev);
 119                 rte_free(dev->mem);
 120                 dev->mem = NULL;
 121         }
 122
 123         free(dev->guest_pages);
 124         dev->guest_pages = NULL;
 125
 126         if (dev->log_addr) {
 127                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 128                 dev->log_addr = 0;
 129         }
 130
 131         if (dev->slave_req_fd >= 0) {
 132                 close(dev->slave_req_fd);
 133                 dev->slave_req_fd = -1;
 134         }
 135
 136         if (dev->postcopy_ufd >= 0) {
 137                 close(dev->postcopy_ufd);
 138                 dev->postcopy_ufd = -1;
 139         }
 140
 141         dev->postcopy_listening = 0;
 142 }
 143
 144 /*
 145  * This function just returns success at the moment unless
 146  * the device hasn't been initialised.
 147  */
 148 static int
 149 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
 150                         struct VhostUserMsg *msg __rte_unused,
 151                         int main_fd __rte_unused)
 152 {
 153         return VH_RESULT_OK;
 154 }
 155
 156 static int
 157 vhost_user_reset_owner(struct virtio_net **pdev,
 158                         struct VhostUserMsg *msg __rte_unused,
 159                         int main_fd __rte_unused)
 160 {
 161         struct virtio_net *dev = *pdev;
 162         vhost_destroy_device_notify(dev);
 163
 164         cleanup_device(dev, 0);
 165         reset_device(dev);
 166         return VH_RESULT_OK;
 167 }
 168
 169 /*
 170  * The features that we support are requested.
 171  */
 172 static int
 173 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 174                         int main_fd __rte_unused)
 175 {
 176         struct virtio_net *dev = *pdev;
 177         uint64_t features = 0;
 178
 179         rte_vhost_driver_get_features(dev->ifname, &features);
 180
 181         msg->payload.u64 = features;
 182         msg->size = sizeof(msg->payload.u64);
 183         msg->fd_num = 0;
 184
 185         return VH_RESULT_REPLY;
 186 }
 187
 188 /*
 189  * The queue number that we support are requested.
 190  */
 191 static int
 192 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
 193                         int main_fd __rte_unused)
 194 {
 195         struct virtio_net *dev = *pdev;
 196         uint32_t queue_num = 0;
 197
 198         rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
 199
 200         msg->payload.u64 = (uint64_t)queue_num;
 201         msg->size = sizeof(msg->payload.u64);
 202         msg->fd_num = 0;
 203
 204         return VH_RESULT_REPLY;
 205 }
 206
 207 /*
 208  * We receive the negotiated features supported by us and the virtio device.
 209  */
 210 static int
 211 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 212                         int main_fd __rte_unused)
 213 {
 214         struct virtio_net *dev = *pdev;
 215         uint64_t features = msg->payload.u64;
 216         uint64_t vhost_features = 0;
 217         struct rte_vdpa_device *vdpa_dev;
 218         int did = -1;
 219
 220         rte_vhost_driver_get_features(dev->ifname, &vhost_features);
 221         if (features & ~vhost_features) {
 222                 RTE_LOG(ERR, VHOST_CONFIG,
 223                         "(%d) received invalid negotiated features.\n",
 224                         dev->vid);
 225                 return VH_RESULT_ERR;
 226         }
 227
 228         if (dev->flags & VIRTIO_DEV_RUNNING) {
 229                 if (dev->features == features)
 230                         return VH_RESULT_OK;
 231
 232                 /*
 233                  * Error out if master tries to change features while device is
 234                  * in running state. The exception being VHOST_F_LOG_ALL, which
 235                  * is enabled when the live-migration starts.
 236                  */
 237                 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
 238                         RTE_LOG(ERR, VHOST_CONFIG,
 239                                 "(%d) features changed while device is running.\n",
 240                                 dev->vid);
 241                         return VH_RESULT_ERR;
 242                 }
 243
 244                 if (dev->notify_ops->features_changed)
 245                         dev->notify_ops->features_changed(dev->vid, features);
 246         }
 247
 248         dev->features = features;
 249         if (dev->features &
 250                 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
 251                 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 252         } else {
 253                 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
 254         }
 255         VHOST_LOG_DEBUG(VHOST_CONFIG,
 256                 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
 257                 dev->vid,
 258                 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
 259                 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 260
 261         if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
 262             !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
 263                 /*
 264                  * Remove all but first queue pair if MQ hasn't been
 265                  * negotiated. This is safe because the device is not
 266                  * running at this stage.
 267                  */
 268                 while (dev->nr_vring > 2) {
 269                         struct vhost_virtqueue *vq;
 270
 271                         vq = dev->virtqueue[--dev->nr_vring];
 272                         if (!vq)
 273                                 continue;
 274
 275                         dev->virtqueue[dev->nr_vring] = NULL;
 276                         cleanup_vq(vq, 1);
 277                         free_vq(dev, vq);
 278                 }
 279         }
 280
 281         did = dev->vdpa_dev_id;
 282         vdpa_dev = rte_vdpa_get_device(did);
 283         if (vdpa_dev && vdpa_dev->ops->set_features)
 284                 vdpa_dev->ops->set_features(dev->vid);
 285
 286         return VH_RESULT_OK;
 287 }
 288
 289 /*
 290  * The virtio device sends us the size of the descriptor ring.
 291  */
 292 static int
 293 vhost_user_set_vring_num(struct virtio_net **pdev,
 294                         struct VhostUserMsg *msg,
 295                         int main_fd __rte_unused)
 296 {
 297         struct virtio_net *dev = *pdev;
 298         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 299
 300         vq->size = msg->payload.state.num;
 301
 302         /* VIRTIO 1.0, 2.4 Virtqueues says:
 303          *
 304          *   Queue Size value is always a power of 2. The maximum Queue Size
 305          *   value is 32768.
 306          */
 307         if ((vq->size & (vq->size - 1)) || vq->size > 32768) {
 308                 RTE_LOG(ERR, VHOST_CONFIG,
 309                         "invalid virtqueue size %u\n", vq->size);
 310                 return VH_RESULT_ERR;
 311         }
 312
 313         if (dev->dequeue_zero_copy) {
 314                 vq->nr_zmbuf = 0;
 315                 vq->last_zmbuf_idx = 0;
 316                 vq->zmbuf_size = vq->size;
 317                 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
 318                                          sizeof(struct zcopy_mbuf), 0);
 319                 if (vq->zmbufs == NULL) {
 320                         RTE_LOG(WARNING, VHOST_CONFIG,
 321                                 "failed to allocate mem for zero copy; "
 322                                 "zero copy is force disabled\n");
 323                         dev->dequeue_zero_copy = 0;
 324                 }
 325                 TAILQ_INIT(&vq->zmbuf_list);
 326         }
 327
 328         if (vq_is_packed(dev)) {
 329                 vq->shadow_used_packed = rte_malloc(NULL,
 330                                 vq->size *
 331                                 sizeof(struct vring_used_elem_packed),
 332                                 RTE_CACHE_LINE_SIZE);
 333                 if (!vq->shadow_used_packed) {
 334                         RTE_LOG(ERR, VHOST_CONFIG,
 335                                         "failed to allocate memory for shadow used ring.\n");
 336                         return VH_RESULT_ERR;
 337                 }
 338
 339         } else {
 340                 vq->shadow_used_split = rte_malloc(NULL,
 341                                 vq->size * sizeof(struct vring_used_elem),
 342                                 RTE_CACHE_LINE_SIZE);
 343                 if (!vq->shadow_used_split) {
 344                         RTE_LOG(ERR, VHOST_CONFIG,
 345                                         "failed to allocate memory for shadow used ring.\n");
 346                         return VH_RESULT_ERR;
 347                 }
 348         }
 349
 350         vq->batch_copy_elems = rte_malloc(NULL,
 351                                 vq->size * sizeof(struct batch_copy_elem),
 352                                 RTE_CACHE_LINE_SIZE);
 353         if (!vq->batch_copy_elems) {
 354                 RTE_LOG(ERR, VHOST_CONFIG,
 355                         "failed to allocate memory for batching copy.\n");
 356                 return VH_RESULT_ERR;
 357         }
 358
 359         return VH_RESULT_OK;
 360 }
 361
 362 /*
 363  * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
 364  * same numa node as the memory of vring descriptor.
 365  */
 366 #ifdef RTE_LIBRTE_VHOST_NUMA
 367 static struct virtio_net*
 368 numa_realloc(struct virtio_net *dev, int index)
 369 {
 370         int oldnode, newnode;
 371         struct virtio_net *old_dev;
 372         struct vhost_virtqueue *old_vq, *vq;
 373         struct zcopy_mbuf *new_zmbuf;
 374         struct vring_used_elem *new_shadow_used_split;
 375         struct vring_used_elem_packed *new_shadow_used_packed;
 376         struct batch_copy_elem *new_batch_copy_elems;
 377         int ret;
 378
 379         old_dev = dev;
 380         vq = old_vq = dev->virtqueue[index];
 381
 382         ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
 383                             MPOL_F_NODE | MPOL_F_ADDR);
 384
 385         /* check if we need to reallocate vq */
 386         ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
 387                              MPOL_F_NODE | MPOL_F_ADDR);
 388         if (ret) {
 389                 RTE_LOG(ERR, VHOST_CONFIG,
 390                         "Unable to get vq numa information.\n");
 391                 return dev;
 392         }
 393         if (oldnode != newnode) {
 394                 RTE_LOG(INFO, VHOST_CONFIG,
 395                         "reallocate vq from %d to %d node\n", oldnode, newnode);
 396                 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
 397                 if (!vq)
 398                         return dev;
 399
 400                 memcpy(vq, old_vq, sizeof(*vq));
 401                 TAILQ_INIT(&vq->zmbuf_list);
 402
 403                 if (dev->dequeue_zero_copy) {
 404                         new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
 405                                         sizeof(struct zcopy_mbuf), 0, newnode);
 406                         if (new_zmbuf) {
 407                                 rte_free(vq->zmbufs);
 408                                 vq->zmbufs = new_zmbuf;
 409                         }
 410                 }
 411
 412                 if (vq_is_packed(dev)) {
 413                         new_shadow_used_packed = rte_malloc_socket(NULL,
 414                                         vq->size *
 415                                         sizeof(struct vring_used_elem_packed),
 416                                         RTE_CACHE_LINE_SIZE,
 417                                         newnode);
 418                         if (new_shadow_used_packed) {
 419                                 rte_free(vq->shadow_used_packed);
 420                                 vq->shadow_used_packed = new_shadow_used_packed;
 421                         }
 422                 } else {
 423                         new_shadow_used_split = rte_malloc_socket(NULL,
 424                                         vq->size *
 425                                         sizeof(struct vring_used_elem),
 426                                         RTE_CACHE_LINE_SIZE,
 427                                         newnode);
 428                         if (new_shadow_used_split) {
 429                                 rte_free(vq->shadow_used_split);
 430                                 vq->shadow_used_split = new_shadow_used_split;
 431                         }
 432                 }
 433
 434                 new_batch_copy_elems = rte_malloc_socket(NULL,
 435                         vq->size * sizeof(struct batch_copy_elem),
 436                         RTE_CACHE_LINE_SIZE,
 437                         newnode);
 438                 if (new_batch_copy_elems) {
 439                         rte_free(vq->batch_copy_elems);
 440                         vq->batch_copy_elems = new_batch_copy_elems;
 441                 }
 442
 443                 rte_free(old_vq);
 444         }
 445
 446         /* check if we need to reallocate dev */
 447         ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
 448                             MPOL_F_NODE | MPOL_F_ADDR);
 449         if (ret) {
 450                 RTE_LOG(ERR, VHOST_CONFIG,
 451                         "Unable to get dev numa information.\n");
 452                 goto out;
 453         }
 454         if (oldnode != newnode) {
 455                 RTE_LOG(INFO, VHOST_CONFIG,
 456                         "reallocate dev from %d to %d node\n",
 457                         oldnode, newnode);
 458                 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
 459                 if (!dev) {
 460                         dev = old_dev;
 461                         goto out;
 462                 }
 463
 464                 memcpy(dev, old_dev, sizeof(*dev));
 465                 rte_free(old_dev);
 466         }
 467
 468 out:
 469         dev->virtqueue[index] = vq;
 470         vhost_devices[dev->vid] = dev;
 471
 472         if (old_vq != vq)
 473                 vhost_user_iotlb_init(dev, index);
 474
 475         return dev;
 476 }
 477 #else
 478 static struct virtio_net*
 479 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 480 {
 481         return dev;
 482 }
 483 #endif
 484
 485 /* Converts QEMU virtual address to Vhost virtual address. */
 486 static uint64_t
 487 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
 488 {
 489         struct rte_vhost_mem_region *r;
 490         uint32_t i;
 491
 492         /* Find the region where the address lives. */
 493         for (i = 0; i < dev->mem->nregions; i++) {
 494                 r = &dev->mem->regions[i];
 495
 496                 if (qva >= r->guest_user_addr &&
 497                     qva <  r->guest_user_addr + r->size) {
 498
 499                         if (unlikely(*len > r->guest_user_addr + r->size - qva))
 500                                 *len = r->guest_user_addr + r->size - qva;
 501
 502                         return qva - r->guest_user_addr +
 503                                r->host_user_addr;
 504                 }
 505         }
 506         *len = 0;
 507
 508         return 0;
 509 }
 510
 511
 512 /*
 513  * Converts ring address to Vhost virtual address.
 514  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 515  * else it is a QEMU virtual address.
 516  */
 517 static uint64_t
 518 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 519                 uint64_t ra, uint64_t *size)
 520 {
 521         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 522                 uint64_t vva;
 523
 524                 vva = vhost_user_iotlb_cache_find(vq, ra,
 525                                         size, VHOST_ACCESS_RW);
 526                 if (!vva)
 527                         vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
 528
 529                 return vva;
 530         }
 531
 532         return qva_to_vva(dev, ra, size);
 533 }
 534
 535 static struct virtio_net *
 536 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 537 {
 538         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 539         struct vhost_vring_addr *addr = &vq->ring_addrs;
 540         uint64_t len;
 541
 542         if (vq_is_packed(dev)) {
 543                 len = sizeof(struct vring_packed_desc) * vq->size;
 544                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 545                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 546                 vq->log_guest_addr = 0;
 547                 if (vq->desc_packed == NULL ||
 548                                 len != sizeof(struct vring_packed_desc) *
 549                                 vq->size) {
 550                         RTE_LOG(DEBUG, VHOST_CONFIG,
 551                                 "(%d) failed to map desc_packed ring.\n",
 552                                 dev->vid);
 553                         return dev;
 554                 }
 555
 556                 dev = numa_realloc(dev, vq_index);
 557                 vq = dev->virtqueue[vq_index];
 558                 addr = &vq->ring_addrs;
 559
 560                 len = sizeof(struct vring_packed_desc_event);
 561                 vq->driver_event = (struct vring_packed_desc_event *)
 562                                         (uintptr_t)ring_addr_to_vva(dev,
 563                                         vq, addr->avail_user_addr, &len);
 564                 if (vq->driver_event == NULL ||
 565                                 len != sizeof(struct vring_packed_desc_event)) {
 566                         RTE_LOG(DEBUG, VHOST_CONFIG,
 567                                 "(%d) failed to find driver area address.\n",
 568                                 dev->vid);
 569                         return dev;
 570                 }
 571
 572                 len = sizeof(struct vring_packed_desc_event);
 573                 vq->device_event = (struct vring_packed_desc_event *)
 574                                         (uintptr_t)ring_addr_to_vva(dev,
 575                                         vq, addr->used_user_addr, &len);
 576                 if (vq->device_event == NULL ||
 577                                 len != sizeof(struct vring_packed_desc_event)) {
 578                         RTE_LOG(DEBUG, VHOST_CONFIG,
 579                                 "(%d) failed to find device area address.\n",
 580                                 dev->vid);
 581                         return dev;
 582                 }
 583
 584                 return dev;
 585         }
 586
 587         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 588         if (vq->desc && vq->avail && vq->used)
 589                 return dev;
 590
 591         len = sizeof(struct vring_desc) * vq->size;
 592         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 593                         vq, addr->desc_user_addr, &len);
 594         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 595                 RTE_LOG(DEBUG, VHOST_CONFIG,
 596                         "(%d) failed to map desc ring.\n",
 597                         dev->vid);
 598                 return dev;
 599         }
 600
 601         dev = numa_realloc(dev, vq_index);
 602         vq = dev->virtqueue[vq_index];
 603         addr = &vq->ring_addrs;
 604
 605         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 606         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 607                         vq, addr->avail_user_addr, &len);
 608         if (vq->avail == 0 ||
 609                         len != sizeof(struct vring_avail) +
 610                         sizeof(uint16_t) * vq->size) {
 611                 RTE_LOG(DEBUG, VHOST_CONFIG,
 612                         "(%d) failed to map avail ring.\n",
 613                         dev->vid);
 614                 return dev;
 615         }
 616
 617         len = sizeof(struct vring_used) +
 618                 sizeof(struct vring_used_elem) * vq->size;
 619         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 620                         vq, addr->used_user_addr, &len);
 621         if (vq->used == 0 || len != sizeof(struct vring_used) +
 622                         sizeof(struct vring_used_elem) * vq->size) {
 623                 RTE_LOG(DEBUG, VHOST_CONFIG,
 624                         "(%d) failed to map used ring.\n",
 625                         dev->vid);
 626                 return dev;
 627         }
 628
 629         if (vq->last_used_idx != vq->used->idx) {
 630                 RTE_LOG(WARNING, VHOST_CONFIG,
 631                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 632                         "some packets maybe resent for Tx and dropped for Rx\n",
 633                         vq->last_used_idx, vq->used->idx);
 634                 vq->last_used_idx  = vq->used->idx;
 635                 vq->last_avail_idx = vq->used->idx;
 636         }
 637
 638         vq->log_guest_addr = addr->log_guest_addr;
 639
 640         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
 641                         dev->vid, vq->desc);
 642         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
 643                         dev->vid, vq->avail);
 644         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
 645                         dev->vid, vq->used);
 646         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
 647                         dev->vid, vq->log_guest_addr);
 648
 649         return dev;
 650 }
 651
 652 /*
 653  * The virtio device sends us the desc, used and avail ring addresses.
 654  * This function then converts these to our address space.
 655  */
 656 static int
 657 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 658                         int main_fd __rte_unused)
 659 {
 660         struct virtio_net *dev = *pdev;
 661         struct vhost_virtqueue *vq;
 662         struct vhost_vring_addr *addr = &msg->payload.addr;
 663
 664         if (dev->mem == NULL)
 665                 return VH_RESULT_ERR;
 666
 667         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 668         vq = dev->virtqueue[msg->payload.addr.index];
 669
 670         /*
 671          * Rings addresses should not be interpreted as long as the ring is not
 672          * started and enabled
 673          */
 674         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 675
 676         vring_invalidate(dev, vq);
 677
 678         if (vq->enabled && (dev->features &
 679                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
 680                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 681                 if (!dev)
 682                         return VH_RESULT_ERR;
 683
 684                 *pdev = dev;
 685         }
 686
 687         return VH_RESULT_OK;
 688 }
 689
 690 /*
 691  * The virtio device sends us the available ring last used index.
 692  */
 693 static int
 694 vhost_user_set_vring_base(struct virtio_net **pdev,
 695                         struct VhostUserMsg *msg,
 696                         int main_fd __rte_unused)
 697 {
 698         struct virtio_net *dev = *pdev;
 699         dev->virtqueue[msg->payload.state.index]->last_used_idx  =
 700                         msg->payload.state.num;
 701         dev->virtqueue[msg->payload.state.index]->last_avail_idx =
 702                         msg->payload.state.num;
 703
 704         return VH_RESULT_OK;
 705 }
 706
 707 static int
 708 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 709                    uint64_t host_phys_addr, uint64_t size)
 710 {
 711         struct guest_page *page, *last_page;
 712
 713         if (dev->nr_guest_pages == dev->max_guest_pages) {
 714                 dev->max_guest_pages *= 2;
 715                 dev->guest_pages = realloc(dev->guest_pages,
 716                                         dev->max_guest_pages * sizeof(*page));
 717                 if (!dev->guest_pages) {
 718                         RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
 719                         return -1;
 720                 }
 721         }
 722
 723         if (dev->nr_guest_pages > 0) {
 724                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 725                 /* merge if the two pages are continuous */
 726                 if (host_phys_addr == last_page->host_phys_addr +
 727                                       last_page->size) {
 728                         last_page->size += size;
 729                         return 0;
 730                 }
 731         }
 732
 733         page = &dev->guest_pages[dev->nr_guest_pages++];
 734         page->guest_phys_addr = guest_phys_addr;
 735         page->host_phys_addr  = host_phys_addr;
 736         page->size = size;
 737
 738         return 0;
 739 }
 740
 741 static int
 742 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 743                 uint64_t page_size)
 744 {
 745         uint64_t reg_size = reg->size;
 746         uint64_t host_user_addr  = reg->host_user_addr;
 747         uint64_t guest_phys_addr = reg->guest_phys_addr;
 748         uint64_t host_phys_addr;
 749         uint64_t size;
 750
 751         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 752         size = page_size - (guest_phys_addr & (page_size - 1));
 753         size = RTE_MIN(size, reg_size);
 754
 755         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 756                 return -1;
 757
 758         host_user_addr  += size;
 759         guest_phys_addr += size;
 760         reg_size -= size;
 761
 762         while (reg_size > 0) {
 763                 size = RTE_MIN(reg_size, page_size);
 764                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 765                                                   host_user_addr);
 766                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 767                                 size) < 0)
 768                         return -1;
 769
 770                 host_user_addr  += size;
 771                 guest_phys_addr += size;
 772                 reg_size -= size;
 773         }
 774
 775         return 0;
 776 }
 777
 778 #ifdef RTE_LIBRTE_VHOST_DEBUG
 779 /* TODO: enable it only in debug mode? */
 780 static void
 781 dump_guest_pages(struct virtio_net *dev)
 782 {
 783         uint32_t i;
 784         struct guest_page *page;
 785
 786         for (i = 0; i < dev->nr_guest_pages; i++) {
 787                 page = &dev->guest_pages[i];
 788
 789                 RTE_LOG(INFO, VHOST_CONFIG,
 790                         "guest physical page region %u\n"
 791                         "\t guest_phys_addr: %" PRIx64 "\n"
 792                         "\t host_phys_addr : %" PRIx64 "\n"
 793                         "\t size           : %" PRIx64 "\n",
 794                         i,
 795                         page->guest_phys_addr,
 796                         page->host_phys_addr,
 797                         page->size);
 798         }
 799 }
 800 #else
 801 #define dump_guest_pages(dev)
 802 #endif
 803
 804 static bool
 805 vhost_memory_changed(struct VhostUserMemory *new,
 806                      struct rte_vhost_memory *old)
 807 {
 808         uint32_t i;
 809
 810         if (new->nregions != old->nregions)
 811                 return true;
 812
 813         for (i = 0; i < new->nregions; ++i) {
 814                 VhostUserMemoryRegion *new_r = &new->regions[i];
 815                 struct rte_vhost_mem_region *old_r = &old->regions[i];
 816
 817                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
 818                         return true;
 819                 if (new_r->memory_size != old_r->size)
 820                         return true;
 821                 if (new_r->userspace_addr != old_r->guest_user_addr)
 822                         return true;
 823         }
 824
 825         return false;
 826 }
 827
 828 static int
 829 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
 830                         int main_fd)
 831 {
 832         struct virtio_net *dev = *pdev;
 833         struct VhostUserMemory *memory = &msg->payload.memory;
 834         struct rte_vhost_mem_region *reg;
 835         void *mmap_addr;
 836         uint64_t mmap_size;
 837         uint64_t mmap_offset;
 838         uint64_t alignment;
 839         uint32_t i;
 840         int populate;
 841         int fd;
 842
 843         if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
 844                 RTE_LOG(ERR, VHOST_CONFIG,
 845                         "too many memory regions (%u)\n", memory->nregions);
 846                 return VH_RESULT_ERR;
 847         }
 848
 849         if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
 850                 RTE_LOG(INFO, VHOST_CONFIG,
 851                         "(%d) memory regions not changed\n", dev->vid);
 852
 853                 for (i = 0; i < memory->nregions; i++)
 854                         close(msg->fds[i]);
 855
 856                 return VH_RESULT_OK;
 857         }
 858
 859         if (dev->mem) {
 860                 free_mem_region(dev);
 861                 rte_free(dev->mem);
 862                 dev->mem = NULL;
 863         }
 864
 865         /* Flush IOTLB cache as previous HVAs are now invalid */
 866         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 867                 for (i = 0; i < dev->nr_vring; i++)
 868                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
 869
 870         dev->nr_guest_pages = 0;
 871         if (!dev->guest_pages) {
 872                 dev->max_guest_pages = 8;
 873                 dev->guest_pages = malloc(dev->max_guest_pages *
 874                                                 sizeof(struct guest_page));
 875                 if (dev->guest_pages == NULL) {
 876                         RTE_LOG(ERR, VHOST_CONFIG,
 877                                 "(%d) failed to allocate memory "
 878                                 "for dev->guest_pages\n",
 879                                 dev->vid);
 880                         return VH_RESULT_ERR;
 881                 }
 882         }
 883
 884         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
 885                 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
 886         if (dev->mem == NULL) {
 887                 RTE_LOG(ERR, VHOST_CONFIG,
 888                         "(%d) failed to allocate memory for dev->mem\n",
 889                         dev->vid);
 890                 return VH_RESULT_ERR;
 891         }
 892         dev->mem->nregions = memory->nregions;
 893
 894         for (i = 0; i < memory->nregions; i++) {
 895                 fd  = msg->fds[i];
 896                 reg = &dev->mem->regions[i];
 897
 898                 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
 899                 reg->guest_user_addr = memory->regions[i].userspace_addr;
 900                 reg->size            = memory->regions[i].memory_size;
 901                 reg->fd              = fd;
 902
 903                 mmap_offset = memory->regions[i].mmap_offset;
 904
 905                 /* Check for memory_size + mmap_offset overflow */
 906                 if (mmap_offset >= -reg->size) {
 907                         RTE_LOG(ERR, VHOST_CONFIG,
 908                                 "mmap_offset (%#"PRIx64") and memory_size "
 909                                 "(%#"PRIx64") overflow\n",
 910                                 mmap_offset, reg->size);
 911                         goto err_mmap;
 912                 }
 913
 914                 mmap_size = reg->size + mmap_offset;
 915
 916                 /* mmap() without flag of MAP_ANONYMOUS, should be called
 917                  * with length argument aligned with hugepagesz at older
 918                  * longterm version Linux, like 2.6.32 and 3.2.72, or
 919                  * mmap() will fail with EINVAL.
 920                  *
 921                  * to avoid failure, make sure in caller to keep length
 922                  * aligned.
 923                  */
 924                 alignment = get_blk_size(fd);
 925                 if (alignment == (uint64_t)-1) {
 926                         RTE_LOG(ERR, VHOST_CONFIG,
 927                                 "couldn't get hugepage size through fstat\n");
 928                         goto err_mmap;
 929                 }
 930                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
 931
 932                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
 933                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
 934                                  MAP_SHARED | populate, fd, 0);
 935
 936                 if (mmap_addr == MAP_FAILED) {
 937                         RTE_LOG(ERR, VHOST_CONFIG,
 938                                 "mmap region %u failed.\n", i);
 939                         goto err_mmap;
 940                 }
 941
 942                 reg->mmap_addr = mmap_addr;
 943                 reg->mmap_size = mmap_size;
 944                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
 945                                       mmap_offset;
 946
 947                 if (dev->dequeue_zero_copy)
 948                         if (add_guest_pages(dev, reg, alignment) < 0) {
 949                                 RTE_LOG(ERR, VHOST_CONFIG,
 950                                         "adding guest pages to region %u failed.\n",
 951                                         i);
 952                                 goto err_mmap;
 953                         }
 954
 955                 RTE_LOG(INFO, VHOST_CONFIG,
 956                         "guest memory region %u, size: 0x%" PRIx64 "\n"
 957                         "\t guest physical addr: 0x%" PRIx64 "\n"
 958                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
 959                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
 960                         "\t mmap addr : 0x%" PRIx64 "\n"
 961                         "\t mmap size : 0x%" PRIx64 "\n"
 962                         "\t mmap align: 0x%" PRIx64 "\n"
 963                         "\t mmap off  : 0x%" PRIx64 "\n",
 964                         i, reg->size,
 965                         reg->guest_phys_addr,
 966                         reg->guest_user_addr,
 967                         reg->host_user_addr,
 968                         (uint64_t)(uintptr_t)mmap_addr,
 969                         mmap_size,
 970                         alignment,
 971                         mmap_offset);
 972
 973                 if (dev->postcopy_listening) {
 974                         /*
 975                          * We haven't a better way right now than sharing
 976                          * DPDK's virtual address with Qemu, so that Qemu can
 977                          * retrieve the region offset when handling userfaults.
 978                          */
 979                         memory->regions[i].userspace_addr =
 980                                 reg->host_user_addr;
 981                 }
 982         }
 983         if (dev->postcopy_listening) {
 984                 /* Send the addresses back to qemu */
 985                 msg->fd_num = 0;
 986                 send_vhost_reply(main_fd, msg);
 987
 988                 /* Wait for qemu to acknolwedge it's got the addresses
 989                  * we've got to wait before we're allowed to generate faults.
 990                  */
 991                 VhostUserMsg ack_msg;
 992                 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
 993                         RTE_LOG(ERR, VHOST_CONFIG,
 994                                 "Failed to read qemu ack on postcopy set-mem-table\n");
 995                         goto err_mmap;
 996                 }
 997                 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
 998                         RTE_LOG(ERR, VHOST_CONFIG,
 999                                 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1000                                 ack_msg.request.master);
1001                         goto err_mmap;
1002                 }
1003
1004                 /* Now userfault register and we can use the memory */
1005                 for (i = 0; i < memory->nregions; i++) {
1006 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1007                         reg = &dev->mem->regions[i];
1008                         struct uffdio_register reg_struct;
1009
1010                         /*
1011                          * Let's register all the mmap'ed area to ensure
1012                          * alignment on page boundary.
1013                          */
1014                         reg_struct.range.start =
1015                                 (uint64_t)(uintptr_t)reg->mmap_addr;
1016                         reg_struct.range.len = reg->mmap_size;
1017                         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1018
1019                         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1020                                                 &reg_struct)) {
1021                                 RTE_LOG(ERR, VHOST_CONFIG,
1022                                         "Failed to register ufd for region %d: (ufd = %d) %s\n",
1023                                         i, dev->postcopy_ufd,
1024                                         strerror(errno));
1025                                 goto err_mmap;
1026                         }
1027                         RTE_LOG(INFO, VHOST_CONFIG,
1028                                 "\t userfaultfd registered for range : %llx - %llx\n",
1029                                 reg_struct.range.start,
1030                                 reg_struct.range.start +
1031                                 reg_struct.range.len - 1);
1032 #else
1033                         goto err_mmap;
1034 #endif
1035                 }
1036         }
1037
1038         for (i = 0; i < dev->nr_vring; i++) {
1039                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1040
1041                 if (vq->desc || vq->avail || vq->used) {
1042                         /*
1043                          * If the memory table got updated, the ring addresses
1044                          * need to be translated again as virtual addresses have
1045                          * changed.
1046                          */
1047                         vring_invalidate(dev, vq);
1048
1049                         dev = translate_ring_addresses(dev, i);
1050                         if (!dev) {
1051                                 dev = *pdev;
1052                                 goto err_mmap;
1053                         }
1054
1055                         *pdev = dev;
1056                 }
1057         }
1058
1059         dump_guest_pages(dev);
1060
1061         return VH_RESULT_OK;
1062
1063 err_mmap:
1064         free_mem_region(dev);
1065         rte_free(dev->mem);
1066         dev->mem = NULL;
1067         return VH_RESULT_ERR;
1068 }
1069
1070 static bool
1071 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1072 {
1073         bool rings_ok;
1074
1075         if (!vq)
1076                 return false;
1077
1078         if (vq_is_packed(dev))
1079                 rings_ok = !!vq->desc_packed;
1080         else
1081                 rings_ok = vq->desc && vq->avail && vq->used;
1082
1083         return rings_ok &&
1084                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1085                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1086 }
1087
1088 static int
1089 virtio_is_ready(struct virtio_net *dev)
1090 {
1091         struct vhost_virtqueue *vq;
1092         uint32_t i;
1093
1094         if (dev->nr_vring == 0)
1095                 return 0;
1096
1097         for (i = 0; i < dev->nr_vring; i++) {
1098                 vq = dev->virtqueue[i];
1099
1100                 if (!vq_is_ready(dev, vq))
1101                         return 0;
1102         }
1103
1104         RTE_LOG(INFO, VHOST_CONFIG,
1105                 "virtio is now ready for processing.\n");
1106         return 1;
1107 }
1108
1109 static int
1110 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1111                         int main_fd __rte_unused)
1112 {
1113         struct virtio_net *dev = *pdev;
1114         struct vhost_vring_file file;
1115         struct vhost_virtqueue *vq;
1116
1117         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1118         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1119                 file.fd = VIRTIO_INVALID_EVENTFD;
1120         else
1121                 file.fd = msg->fds[0];
1122         RTE_LOG(INFO, VHOST_CONFIG,
1123                 "vring call idx:%d file:%d\n", file.index, file.fd);
1124
1125         vq = dev->virtqueue[file.index];
1126         if (vq->callfd >= 0)
1127                 close(vq->callfd);
1128
1129         vq->callfd = file.fd;
1130
1131         return VH_RESULT_OK;
1132 }
1133
1134 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1135                         struct VhostUserMsg *msg,
1136                         int main_fd __rte_unused)
1137 {
1138         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1139                 close(msg->fds[0]);
1140         RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1141
1142         return VH_RESULT_OK;
1143 }
1144
1145 static int
1146 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1147                         int main_fd __rte_unused)
1148 {
1149         struct virtio_net *dev = *pdev;
1150         struct vhost_vring_file file;
1151         struct vhost_virtqueue *vq;
1152
1153         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1154         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1155                 file.fd = VIRTIO_INVALID_EVENTFD;
1156         else
1157                 file.fd = msg->fds[0];
1158         RTE_LOG(INFO, VHOST_CONFIG,
1159                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1160
1161         /* Interpret ring addresses only when ring is started. */
1162         dev = translate_ring_addresses(dev, file.index);
1163         if (!dev)
1164                 return VH_RESULT_ERR;
1165
1166         *pdev = dev;
1167
1168         vq = dev->virtqueue[file.index];
1169
1170         /*
1171          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1172          * the ring starts already enabled. Otherwise, it is enabled via
1173          * the SET_VRING_ENABLE message.
1174          */
1175         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)))
1176                 vq->enabled = 1;
1177
1178         if (vq->kickfd >= 0)
1179                 close(vq->kickfd);
1180         vq->kickfd = file.fd;
1181
1182         return VH_RESULT_OK;
1183 }
1184
1185 static void
1186 free_zmbufs(struct vhost_virtqueue *vq)
1187 {
1188         struct zcopy_mbuf *zmbuf, *next;
1189
1190         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1191              zmbuf != NULL; zmbuf = next) {
1192                 next = TAILQ_NEXT(zmbuf, next);
1193
1194                 rte_pktmbuf_free(zmbuf->mbuf);
1195                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1196         }
1197
1198         rte_free(vq->zmbufs);
1199 }
1200
1201 /*
1202  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1203  */
1204 static int
1205 vhost_user_get_vring_base(struct virtio_net **pdev,
1206                         struct VhostUserMsg *msg,
1207                         int main_fd __rte_unused)
1208 {
1209         struct virtio_net *dev = *pdev;
1210         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1211
1212         /* We have to stop the queue (virtio) if it is running. */
1213         vhost_destroy_device_notify(dev);
1214
1215         dev->flags &= ~VIRTIO_DEV_READY;
1216         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1217
1218         /* Here we are safe to get the last avail index */
1219         msg->payload.state.num = vq->last_avail_idx;
1220
1221         RTE_LOG(INFO, VHOST_CONFIG,
1222                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1223                 msg->payload.state.num);
1224         /*
1225          * Based on current qemu vhost-user implementation, this message is
1226          * sent and only sent in vhost_vring_stop.
1227          * TODO: cleanup the vring, it isn't usable since here.
1228          */
1229         if (vq->kickfd >= 0)
1230                 close(vq->kickfd);
1231
1232         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1233
1234         if (vq->callfd >= 0)
1235                 close(vq->callfd);
1236
1237         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1238
1239         if (dev->dequeue_zero_copy)
1240                 free_zmbufs(vq);
1241         if (vq_is_packed(dev)) {
1242                 rte_free(vq->shadow_used_packed);
1243                 vq->shadow_used_packed = NULL;
1244         } else {
1245                 rte_free(vq->shadow_used_split);
1246                 vq->shadow_used_split = NULL;
1247         }
1248
1249         rte_free(vq->batch_copy_elems);
1250         vq->batch_copy_elems = NULL;
1251
1252         msg->size = sizeof(msg->payload.state);
1253         msg->fd_num = 0;
1254
1255         return VH_RESULT_REPLY;
1256 }
1257
1258 /*
1259  * when virtio queues are ready to work, qemu will send us to
1260  * enable the virtio queue pair.
1261  */
1262 static int
1263 vhost_user_set_vring_enable(struct virtio_net **pdev,
1264                         struct VhostUserMsg *msg,
1265                         int main_fd __rte_unused)
1266 {
1267         struct virtio_net *dev = *pdev;
1268         int enable = (int)msg->payload.state.num;
1269         int index = (int)msg->payload.state.index;
1270         struct rte_vdpa_device *vdpa_dev;
1271         int did = -1;
1272
1273         RTE_LOG(INFO, VHOST_CONFIG,
1274                 "set queue enable: %d to qp idx: %d\n",
1275                 enable, index);
1276
1277         did = dev->vdpa_dev_id;
1278         vdpa_dev = rte_vdpa_get_device(did);
1279         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1280                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1281
1282         if (dev->notify_ops->vring_state_changed)
1283                 dev->notify_ops->vring_state_changed(dev->vid,
1284                                 index, enable);
1285
1286         dev->virtqueue[index]->enabled = enable;
1287
1288         return VH_RESULT_OK;
1289 }
1290
1291 static int
1292 vhost_user_get_protocol_features(struct virtio_net **pdev,
1293                         struct VhostUserMsg *msg,
1294                         int main_fd __rte_unused)
1295 {
1296         struct virtio_net *dev = *pdev;
1297         uint64_t features, protocol_features;
1298
1299         rte_vhost_driver_get_features(dev->ifname, &features);
1300         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1301
1302         /*
1303          * REPLY_ACK protocol feature is only mandatory for now
1304          * for IOMMU feature. If IOMMU is explicitly disabled by the
1305          * application, disable also REPLY_ACK feature for older buggy
1306          * Qemu versions (from v2.7.0 to v2.9.0).
1307          */
1308         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1309                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1310
1311         msg->payload.u64 = protocol_features;
1312         msg->size = sizeof(msg->payload.u64);
1313         msg->fd_num = 0;
1314
1315         return VH_RESULT_REPLY;
1316 }
1317
1318 static int
1319 vhost_user_set_protocol_features(struct virtio_net **pdev,
1320                         struct VhostUserMsg *msg,
1321                         int main_fd __rte_unused)
1322 {
1323         struct virtio_net *dev = *pdev;
1324         uint64_t protocol_features = msg->payload.u64;
1325         if (protocol_features & ~VHOST_USER_PROTOCOL_FEATURES) {
1326                 RTE_LOG(ERR, VHOST_CONFIG,
1327                         "(%d) received invalid protocol features.\n",
1328                         dev->vid);
1329                 return VH_RESULT_ERR;
1330         }
1331
1332         dev->protocol_features = protocol_features;
1333
1334         return VH_RESULT_OK;
1335 }
1336
1337 static int
1338 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1339                         int main_fd __rte_unused)
1340 {
1341         struct virtio_net *dev = *pdev;
1342         int fd = msg->fds[0];
1343         uint64_t size, off;
1344         void *addr;
1345
1346         if (fd < 0) {
1347                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1348                 return VH_RESULT_ERR;
1349         }
1350
1351         if (msg->size != sizeof(VhostUserLog)) {
1352                 RTE_LOG(ERR, VHOST_CONFIG,
1353                         "invalid log base msg size: %"PRId32" != %d\n",
1354                         msg->size, (int)sizeof(VhostUserLog));
1355                 return VH_RESULT_ERR;
1356         }
1357
1358         size = msg->payload.log.mmap_size;
1359         off  = msg->payload.log.mmap_offset;
1360
1361         /* Don't allow mmap_offset to point outside the mmap region */
1362         if (off > size) {
1363                 RTE_LOG(ERR, VHOST_CONFIG,
1364                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1365                         off, size);
1366                 return VH_RESULT_ERR;
1367         }
1368
1369         RTE_LOG(INFO, VHOST_CONFIG,
1370                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1371                 size, off);
1372
1373         /*
1374          * mmap from 0 to workaround a hugepage mmap bug: mmap will
1375          * fail when offset is not page size aligned.
1376          */
1377         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1378         close(fd);
1379         if (addr == MAP_FAILED) {
1380                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1381                 return VH_RESULT_ERR;
1382         }
1383
1384         /*
1385          * Free previously mapped log memory on occasionally
1386          * multiple VHOST_USER_SET_LOG_BASE.
1387          */
1388         if (dev->log_addr) {
1389                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1390         }
1391         dev->log_addr = (uint64_t)(uintptr_t)addr;
1392         dev->log_base = dev->log_addr + off;
1393         dev->log_size = size;
1394
1395         /*
1396          * The spec is not clear about it (yet), but QEMU doesn't expect
1397          * any payload in the reply.
1398          */
1399         msg->size = 0;
1400         msg->fd_num = 0;
1401
1402         return VH_RESULT_REPLY;
1403 }
1404
1405 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1406                         struct VhostUserMsg *msg,
1407                         int main_fd __rte_unused)
1408 {
1409         close(msg->fds[0]);
1410         RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1411
1412         return VH_RESULT_OK;
1413 }
1414
1415 /*
1416  * An rarp packet is constructed and broadcasted to notify switches about
1417  * the new location of the migrated VM, so that packets from outside will
1418  * not be lost after migration.
1419  *
1420  * However, we don't actually "send" a rarp packet here, instead, we set
1421  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1422  */
1423 static int
1424 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
1425                         int main_fd __rte_unused)
1426 {
1427         struct virtio_net *dev = *pdev;
1428         uint8_t *mac = (uint8_t *)&msg->payload.u64;
1429         struct rte_vdpa_device *vdpa_dev;
1430         int did = -1;
1431
1432         RTE_LOG(DEBUG, VHOST_CONFIG,
1433                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1434                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1435         memcpy(dev->mac.addr_bytes, mac, 6);
1436
1437         /*
1438          * Set the flag to inject a RARP broadcast packet at
1439          * rte_vhost_dequeue_burst().
1440          *
1441          * rte_smp_wmb() is for making sure the mac is copied
1442          * before the flag is set.
1443          */
1444         rte_smp_wmb();
1445         rte_atomic16_set(&dev->broadcast_rarp, 1);
1446         did = dev->vdpa_dev_id;
1447         vdpa_dev = rte_vdpa_get_device(did);
1448         if (vdpa_dev && vdpa_dev->ops->migration_done)
1449                 vdpa_dev->ops->migration_done(dev->vid);
1450
1451         return VH_RESULT_OK;
1452 }
1453
1454 static int
1455 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
1456                         int main_fd __rte_unused)
1457 {
1458         struct virtio_net *dev = *pdev;
1459         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1460                         msg->payload.u64 > VIRTIO_MAX_MTU) {
1461                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1462                                 msg->payload.u64);
1463
1464                 return VH_RESULT_ERR;
1465         }
1466
1467         dev->mtu = msg->payload.u64;
1468
1469         return VH_RESULT_OK;
1470 }
1471
1472 static int
1473 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
1474                         int main_fd __rte_unused)
1475 {
1476         struct virtio_net *dev = *pdev;
1477         int fd = msg->fds[0];
1478
1479         if (fd < 0) {
1480                 RTE_LOG(ERR, VHOST_CONFIG,
1481                                 "Invalid file descriptor for slave channel (%d)\n",
1482                                 fd);
1483                 return VH_RESULT_ERR;
1484         }
1485
1486         dev->slave_req_fd = fd;
1487
1488         return VH_RESULT_OK;
1489 }
1490
1491 static int
1492 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1493 {
1494         struct vhost_vring_addr *ra;
1495         uint64_t start, end;
1496
1497         start = imsg->iova;
1498         end = start + imsg->size;
1499
1500         ra = &vq->ring_addrs;
1501         if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1502                 return 1;
1503         if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1504                 return 1;
1505         if (ra->used_user_addr >= start && ra->used_user_addr < end)
1506                 return 1;
1507
1508         return 0;
1509 }
1510
1511 static int
1512 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1513                                 struct vhost_iotlb_msg *imsg)
1514 {
1515         uint64_t istart, iend, vstart, vend;
1516
1517         istart = imsg->iova;
1518         iend = istart + imsg->size - 1;
1519
1520         vstart = (uintptr_t)vq->desc;
1521         vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1522         if (vstart <= iend && istart <= vend)
1523                 return 1;
1524
1525         vstart = (uintptr_t)vq->avail;
1526         vend = vstart + sizeof(struct vring_avail);
1527         vend += sizeof(uint16_t) * vq->size - 1;
1528         if (vstart <= iend && istart <= vend)
1529                 return 1;
1530
1531         vstart = (uintptr_t)vq->used;
1532         vend = vstart + sizeof(struct vring_used);
1533         vend += sizeof(struct vring_used_elem) * vq->size - 1;
1534         if (vstart <= iend && istart <= vend)
1535                 return 1;
1536
1537         return 0;
1538 }
1539
1540 static int
1541 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
1542                         int main_fd __rte_unused)
1543 {
1544         struct virtio_net *dev = *pdev;
1545         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1546         uint16_t i;
1547         uint64_t vva, len;
1548
1549         switch (imsg->type) {
1550         case VHOST_IOTLB_UPDATE:
1551                 len = imsg->size;
1552                 vva = qva_to_vva(dev, imsg->uaddr, &len);
1553                 if (!vva)
1554                         return VH_RESULT_ERR;
1555
1556                 for (i = 0; i < dev->nr_vring; i++) {
1557                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1558
1559                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1560                                         len, imsg->perm);
1561
1562                         if (is_vring_iotlb_update(vq, imsg))
1563                                 *pdev = dev = translate_ring_addresses(dev, i);
1564                 }
1565                 break;
1566         case VHOST_IOTLB_INVALIDATE:
1567                 for (i = 0; i < dev->nr_vring; i++) {
1568                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1569
1570                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
1571                                         imsg->size);
1572
1573                         if (is_vring_iotlb_invalidate(vq, imsg))
1574                                 vring_invalidate(dev, vq);
1575                 }
1576                 break;
1577         default:
1578                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1579                                 imsg->type);
1580                 return VH_RESULT_ERR;
1581         }
1582
1583         return VH_RESULT_OK;
1584 }
1585
1586 static int
1587 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
1588                         struct VhostUserMsg *msg,
1589                         int main_fd __rte_unused)
1590 {
1591         struct virtio_net *dev = *pdev;
1592 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1593         struct uffdio_api api_struct;
1594
1595         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1596
1597         if (dev->postcopy_ufd == -1) {
1598                 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
1599                         strerror(errno));
1600                 return VH_RESULT_ERR;
1601         }
1602         api_struct.api = UFFD_API;
1603         api_struct.features = 0;
1604         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
1605                 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
1606                         strerror(errno));
1607                 close(dev->postcopy_ufd);
1608                 dev->postcopy_ufd = -1;
1609                 return VH_RESULT_ERR;
1610         }
1611         msg->fds[0] = dev->postcopy_ufd;
1612         msg->fd_num = 1;
1613
1614         return VH_RESULT_REPLY;
1615 #else
1616         dev->postcopy_ufd = -1;
1617         msg->fd_num = 0;
1618
1619         return VH_RESULT_ERR;
1620 #endif
1621 }
1622
1623 static int
1624 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
1625                         struct VhostUserMsg *msg __rte_unused,
1626                         int main_fd __rte_unused)
1627 {
1628         struct virtio_net *dev = *pdev;
1629
1630         if (dev->mem && dev->mem->nregions) {
1631                 RTE_LOG(ERR, VHOST_CONFIG,
1632                         "Regions already registered at postcopy-listen\n");
1633                 return VH_RESULT_ERR;
1634         }
1635         dev->postcopy_listening = 1;
1636
1637         return VH_RESULT_OK;
1638 }
1639
1640 static int
1641 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
1642                         int main_fd __rte_unused)
1643 {
1644         struct virtio_net *dev = *pdev;
1645
1646         dev->postcopy_listening = 0;
1647         if (dev->postcopy_ufd >= 0) {
1648                 close(dev->postcopy_ufd);
1649                 dev->postcopy_ufd = -1;
1650         }
1651
1652         msg->payload.u64 = 0;
1653         msg->size = sizeof(msg->payload.u64);
1654         msg->fd_num = 0;
1655
1656         return VH_RESULT_REPLY;
1657 }
1658
1659 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
1660                                         struct VhostUserMsg *msg,
1661                                         int main_fd);
1662 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
1663         [VHOST_USER_NONE] = NULL,
1664         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
1665         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
1666         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
1667         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
1668         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
1669         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
1670         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
1671         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
1672         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
1673         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
1674         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
1675         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
1676         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
1677         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
1678         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
1679         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
1680         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
1681         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
1682         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
1683         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
1684         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
1685         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
1686         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
1687         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
1688         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
1689 };
1690
1691
1692 /* return bytes# of read on success or negative val on failure. */
1693 static int
1694 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1695 {
1696         int ret;
1697
1698         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1699                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
1700         if (ret <= 0)
1701                 return ret;
1702
1703         if (msg && msg->size) {
1704                 if (msg->size > sizeof(msg->payload)) {
1705                         RTE_LOG(ERR, VHOST_CONFIG,
1706                                 "invalid msg size: %d\n", msg->size);
1707                         return -1;
1708                 }
1709                 ret = read(sockfd, &msg->payload, msg->size);
1710                 if (ret <= 0)
1711                         return ret;
1712                 if (ret != (int)msg->size) {
1713                         RTE_LOG(ERR, VHOST_CONFIG,
1714                                 "read control message failed\n");
1715                         return -1;
1716                 }
1717         }
1718
1719         return ret;
1720 }
1721
1722 static int
1723 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1724 {
1725         if (!msg)
1726                 return 0;
1727
1728         return send_fd_message(sockfd, (char *)msg,
1729                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
1730 }
1731
1732 static int
1733 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1734 {
1735         if (!msg)
1736                 return 0;
1737
1738         msg->flags &= ~VHOST_USER_VERSION_MASK;
1739         msg->flags &= ~VHOST_USER_NEED_REPLY;
1740         msg->flags |= VHOST_USER_VERSION;
1741         msg->flags |= VHOST_USER_REPLY_MASK;
1742
1743         return send_vhost_message(sockfd, msg);
1744 }
1745
1746 static int
1747 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
1748 {
1749         int ret;
1750
1751         if (msg->flags & VHOST_USER_NEED_REPLY)
1752                 rte_spinlock_lock(&dev->slave_req_lock);
1753
1754         ret = send_vhost_message(dev->slave_req_fd, msg);
1755         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1756                 rte_spinlock_unlock(&dev->slave_req_lock);
1757
1758         return ret;
1759 }
1760
1761 /*
1762  * Allocate a queue pair if it hasn't been allocated yet
1763  */
1764 static int
1765 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1766                         struct VhostUserMsg *msg)
1767 {
1768         uint16_t vring_idx;
1769
1770         switch (msg->request.master) {
1771         case VHOST_USER_SET_VRING_KICK:
1772         case VHOST_USER_SET_VRING_CALL:
1773         case VHOST_USER_SET_VRING_ERR:
1774                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1775                 break;
1776         case VHOST_USER_SET_VRING_NUM:
1777         case VHOST_USER_SET_VRING_BASE:
1778         case VHOST_USER_SET_VRING_ENABLE:
1779                 vring_idx = msg->payload.state.index;
1780                 break;
1781         case VHOST_USER_SET_VRING_ADDR:
1782                 vring_idx = msg->payload.addr.index;
1783                 break;
1784         default:
1785                 return 0;
1786         }
1787
1788         if (vring_idx >= VHOST_MAX_VRING) {
1789                 RTE_LOG(ERR, VHOST_CONFIG,
1790                         "invalid vring index: %u\n", vring_idx);
1791                 return -1;
1792         }
1793
1794         if (dev->virtqueue[vring_idx])
1795                 return 0;
1796
1797         return alloc_vring_queue(dev, vring_idx);
1798 }
1799
1800 static void
1801 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1802 {
1803         unsigned int i = 0;
1804         unsigned int vq_num = 0;
1805
1806         while (vq_num < dev->nr_vring) {
1807                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1808
1809                 if (vq) {
1810                         rte_spinlock_lock(&vq->access_lock);
1811                         vq_num++;
1812                 }
1813                 i++;
1814         }
1815 }
1816
1817 static void
1818 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1819 {
1820         unsigned int i = 0;
1821         unsigned int vq_num = 0;
1822
1823         while (vq_num < dev->nr_vring) {
1824                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1825
1826                 if (vq) {
1827                         rte_spinlock_unlock(&vq->access_lock);
1828                         vq_num++;
1829                 }
1830                 i++;
1831         }
1832 }
1833
1834 int
1835 vhost_user_msg_handler(int vid, int fd)
1836 {
1837         struct virtio_net *dev;
1838         struct VhostUserMsg msg;
1839         struct rte_vdpa_device *vdpa_dev;
1840         int did = -1;
1841         int ret;
1842         int unlock_required = 0;
1843         uint32_t skip_master = 0;
1844         int request;
1845
1846         dev = get_device(vid);
1847         if (dev == NULL)
1848                 return -1;
1849
1850         if (!dev->notify_ops) {
1851                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1852                 if (!dev->notify_ops) {
1853                         RTE_LOG(ERR, VHOST_CONFIG,
1854                                 "failed to get callback ops for driver %s\n",
1855                                 dev->ifname);
1856                         return -1;
1857                 }
1858         }
1859
1860         ret = read_vhost_message(fd, &msg);
1861         if (ret <= 0 || msg.request.master >= VHOST_USER_MAX) {
1862                 if (ret < 0)
1863                         RTE_LOG(ERR, VHOST_CONFIG,
1864                                 "vhost read message failed\n");
1865                 else if (ret == 0)
1866                         RTE_LOG(INFO, VHOST_CONFIG,
1867                                 "vhost peer closed\n");
1868                 else
1869                         RTE_LOG(ERR, VHOST_CONFIG,
1870                                 "vhost read incorrect message\n");
1871
1872                 return -1;
1873         }
1874
1875         ret = 0;
1876         if (msg.request.master != VHOST_USER_IOTLB_MSG)
1877                 RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1878                         vhost_message_str[msg.request.master]);
1879         else
1880                 RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1881                         vhost_message_str[msg.request.master]);
1882
1883         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1884         if (ret < 0) {
1885                 RTE_LOG(ERR, VHOST_CONFIG,
1886                         "failed to alloc queue\n");
1887                 return -1;
1888         }
1889
1890         /*
1891          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1892          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1893          * and device is destroyed. destroy_device waits for queues to be
1894          * inactive, so it is safe. Otherwise taking the access_lock
1895          * would cause a dead lock.
1896          */
1897         switch (msg.request.master) {
1898         case VHOST_USER_SET_FEATURES:
1899         case VHOST_USER_SET_PROTOCOL_FEATURES:
1900         case VHOST_USER_SET_OWNER:
1901         case VHOST_USER_SET_MEM_TABLE:
1902         case VHOST_USER_SET_LOG_BASE:
1903         case VHOST_USER_SET_LOG_FD:
1904         case VHOST_USER_SET_VRING_NUM:
1905         case VHOST_USER_SET_VRING_ADDR:
1906         case VHOST_USER_SET_VRING_BASE:
1907         case VHOST_USER_SET_VRING_KICK:
1908         case VHOST_USER_SET_VRING_CALL:
1909         case VHOST_USER_SET_VRING_ERR:
1910         case VHOST_USER_SET_VRING_ENABLE:
1911         case VHOST_USER_SEND_RARP:
1912         case VHOST_USER_NET_SET_MTU:
1913         case VHOST_USER_SET_SLAVE_REQ_FD:
1914                 vhost_user_lock_all_queue_pairs(dev);
1915                 unlock_required = 1;
1916                 break;
1917         default:
1918                 break;
1919
1920         }
1921
1922         if (dev->extern_ops.pre_msg_handle) {
1923                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
1924                                 (void *)&msg, &skip_master);
1925                 if (ret == VH_RESULT_ERR)
1926                         goto skip_to_reply;
1927                 else if (ret == VH_RESULT_REPLY)
1928                         send_vhost_reply(fd, &msg);
1929
1930                 if (skip_master)
1931                         goto skip_to_post_handle;
1932         }
1933
1934         request = msg.request.master;
1935         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
1936                 if (!vhost_message_handlers[request])
1937                         goto skip_to_post_handle;
1938                 ret = vhost_message_handlers[request](&dev, &msg, fd);
1939
1940                 switch (ret) {
1941                 case VH_RESULT_ERR:
1942                         RTE_LOG(ERR, VHOST_CONFIG,
1943                                 "Processing %s failed.\n",
1944                                 vhost_message_str[request]);
1945                         break;
1946                 case VH_RESULT_OK:
1947                         RTE_LOG(DEBUG, VHOST_CONFIG,
1948                                 "Processing %s succeeded.\n",
1949                                 vhost_message_str[request]);
1950                         break;
1951                 case VH_RESULT_REPLY:
1952                         RTE_LOG(DEBUG, VHOST_CONFIG,
1953                                 "Processing %s succeeded and needs reply.\n",
1954                                 vhost_message_str[request]);
1955                         send_vhost_reply(fd, &msg);
1956                         break;
1957                 }
1958         } else {
1959                 RTE_LOG(ERR, VHOST_CONFIG,
1960                         "Requested invalid message type %d.\n", request);
1961                 ret = VH_RESULT_ERR;
1962         }
1963
1964 skip_to_post_handle:
1965         if (ret != VH_RESULT_ERR && dev->extern_ops.post_msg_handle) {
1966                 ret = (*dev->extern_ops.post_msg_handle)(
1967                                 dev->vid, (void *)&msg);
1968                 if (ret == VH_RESULT_ERR)
1969                         goto skip_to_reply;
1970                 else if (ret == VH_RESULT_REPLY)
1971                         send_vhost_reply(fd, &msg);
1972         }
1973
1974 skip_to_reply:
1975         if (unlock_required)
1976                 vhost_user_unlock_all_queue_pairs(dev);
1977
1978         /*
1979          * If the request required a reply that was already sent,
1980          * this optional reply-ack won't be sent as the
1981          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
1982          */
1983         if (msg.flags & VHOST_USER_NEED_REPLY) {
1984                 msg.payload.u64 = ret == VH_RESULT_ERR;
1985                 msg.size = sizeof(msg.payload.u64);
1986                 msg.fd_num = 0;
1987                 send_vhost_reply(fd, &msg);
1988         } else if (ret == VH_RESULT_ERR) {
1989                 RTE_LOG(ERR, VHOST_CONFIG,
1990                         "vhost message handling failed.\n");
1991                 return -1;
1992         }
1993
1994         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
1995                 dev->flags |= VIRTIO_DEV_READY;
1996
1997                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
1998                         if (dev->dequeue_zero_copy) {
1999                                 RTE_LOG(INFO, VHOST_CONFIG,
2000                                                 "dequeue zero copy is enabled\n");
2001                         }
2002
2003                         if (dev->notify_ops->new_device(dev->vid) == 0)
2004                                 dev->flags |= VIRTIO_DEV_RUNNING;
2005                 }
2006         }
2007
2008         did = dev->vdpa_dev_id;
2009         vdpa_dev = rte_vdpa_get_device(did);
2010         if (vdpa_dev && virtio_is_ready(dev) &&
2011                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2012                         msg.request.master == VHOST_USER_SET_VRING_ENABLE) {
2013                 if (vdpa_dev->ops->dev_conf)
2014                         vdpa_dev->ops->dev_conf(dev->vid);
2015                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2016                 if (vhost_user_host_notifier_ctrl(dev->vid, true) != 0) {
2017                         RTE_LOG(INFO, VHOST_CONFIG,
2018                                 "(%d) software relay is used for vDPA, performance may be low.\n",
2019                                 dev->vid);
2020                 }
2021         }
2022
2023         return 0;
2024 }
2025
2026 static int process_slave_message_reply(struct virtio_net *dev,
2027                                        const struct VhostUserMsg *msg)
2028 {
2029         struct VhostUserMsg msg_reply;
2030         int ret;
2031
2032         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2033                 return 0;
2034
2035         if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
2036                 ret = -1;
2037                 goto out;
2038         }
2039
2040         if (msg_reply.request.slave != msg->request.slave) {
2041                 RTE_LOG(ERR, VHOST_CONFIG,
2042                         "Received unexpected msg type (%u), expected %u\n",
2043                         msg_reply.request.slave, msg->request.slave);
2044                 ret = -1;
2045                 goto out;
2046         }
2047
2048         ret = msg_reply.payload.u64 ? -1 : 0;
2049
2050 out:
2051         rte_spinlock_unlock(&dev->slave_req_lock);
2052         return ret;
2053 }
2054
2055 int
2056 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2057 {
2058         int ret;
2059         struct VhostUserMsg msg = {
2060                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2061                 .flags = VHOST_USER_VERSION,
2062                 .size = sizeof(msg.payload.iotlb),
2063                 .payload.iotlb = {
2064                         .iova = iova,
2065                         .perm = perm,
2066                         .type = VHOST_IOTLB_MISS,
2067                 },
2068         };
2069
2070         ret = send_vhost_message(dev->slave_req_fd, &msg);
2071         if (ret < 0) {
2072                 RTE_LOG(ERR, VHOST_CONFIG,
2073                                 "Failed to send IOTLB miss message (%d)\n",
2074                                 ret);
2075                 return ret;
2076         }
2077
2078         return 0;
2079 }
2080
2081 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2082                                                     int index, int fd,
2083                                                     uint64_t offset,
2084                                                     uint64_t size)
2085 {
2086         int ret;
2087         struct VhostUserMsg msg = {
2088                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2089                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2090                 .size = sizeof(msg.payload.area),
2091                 .payload.area = {
2092                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
2093                         .size = size,
2094                         .offset = offset,
2095                 },
2096         };
2097
2098         if (fd < 0)
2099                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2100         else {
2101                 msg.fds[0] = fd;
2102                 msg.fd_num = 1;
2103         }
2104
2105         ret = send_vhost_slave_message(dev, &msg);
2106         if (ret < 0) {
2107                 RTE_LOG(ERR, VHOST_CONFIG,
2108                         "Failed to set host notifier (%d)\n", ret);
2109                 return ret;
2110         }
2111
2112         return process_slave_message_reply(dev, &msg);
2113 }
2114
2115 int vhost_user_host_notifier_ctrl(int vid, bool enable)
2116 {
2117         struct virtio_net *dev;
2118         struct rte_vdpa_device *vdpa_dev;
2119         int vfio_device_fd, did, ret = 0;
2120         uint64_t offset, size;
2121         unsigned int i;
2122
2123         dev = get_device(vid);
2124         if (!dev)
2125                 return -ENODEV;
2126
2127         did = dev->vdpa_dev_id;
2128         if (did < 0)
2129                 return -EINVAL;
2130
2131         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2132             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2133             !(dev->protocol_features &
2134                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2135             !(dev->protocol_features &
2136                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2137             !(dev->protocol_features &
2138                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2139                 return -ENOTSUP;
2140
2141         vdpa_dev = rte_vdpa_get_device(did);
2142         if (!vdpa_dev)
2143                 return -ENODEV;
2144
2145         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2146         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2147
2148         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2149         if (vfio_device_fd < 0)
2150                 return -ENOTSUP;
2151
2152         if (enable) {
2153                 for (i = 0; i < dev->nr_vring; i++) {
2154                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2155                                         &size) < 0) {
2156                                 ret = -ENOTSUP;
2157                                 goto disable;
2158                         }
2159
2160                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2161                                         vfio_device_fd, offset, size) < 0) {
2162                                 ret = -EFAULT;
2163                                 goto disable;
2164                         }
2165                 }
2166         } else {
2167 disable:
2168                 for (i = 0; i < dev->nr_vring; i++) {
2169                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
2170                                         0, 0);
2171                 }
2172         }
2173
2174         return ret;
2175 }