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 RTE_VHOST_MSG_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 RTE_VHOST_MSG_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 RTE_VHOST_MSG_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 RTE_VHOST_MSG_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 RTE_VHOST_MSG_RESULT_ERR;
 226         }
 227
 228         if (dev->flags & VIRTIO_DEV_RUNNING) {
 229                 if (dev->features == features)
 230                         return RTE_VHOST_MSG_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 RTE_VHOST_MSG_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 RTE_VHOST_MSG_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 RTE_VHOST_MSG_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 RTE_VHOST_MSG_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 RTE_VHOST_MSG_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 RTE_VHOST_MSG_RESULT_ERR;
 357         }
 358
 359         return RTE_VHOST_MSG_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         if (unlikely(!dev || !dev->mem))
 493                 goto out_error;
 494
 495         /* Find the region where the address lives. */
 496         for (i = 0; i < dev->mem->nregions; i++) {
 497                 r = &dev->mem->regions[i];
 498
 499                 if (qva >= r->guest_user_addr &&
 500                     qva <  r->guest_user_addr + r->size) {
 501
 502                         if (unlikely(*len > r->guest_user_addr + r->size - qva))
 503                                 *len = r->guest_user_addr + r->size - qva;
 504
 505                         return qva - r->guest_user_addr +
 506                                r->host_user_addr;
 507                 }
 508         }
 509 out_error:
 510         *len = 0;
 511
 512         return 0;
 513 }
 514
 515
 516 /*
 517  * Converts ring address to Vhost virtual address.
 518  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 519  * else it is a QEMU virtual address.
 520  */
 521 static uint64_t
 522 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 523                 uint64_t ra, uint64_t *size)
 524 {
 525         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 526                 uint64_t vva;
 527
 528                 vva = vhost_user_iotlb_cache_find(vq, ra,
 529                                         size, VHOST_ACCESS_RW);
 530                 if (!vva)
 531                         vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
 532
 533                 return vva;
 534         }
 535
 536         return qva_to_vva(dev, ra, size);
 537 }
 538
 539 static struct virtio_net *
 540 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 541 {
 542         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 543         struct vhost_vring_addr *addr = &vq->ring_addrs;
 544         uint64_t len, expected_len;
 545
 546         if (vq_is_packed(dev)) {
 547                 len = sizeof(struct vring_packed_desc) * vq->size;
 548                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 549                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 550                 vq->log_guest_addr = 0;
 551                 if (vq->desc_packed == NULL ||
 552                                 len != sizeof(struct vring_packed_desc) *
 553                                 vq->size) {
 554                         RTE_LOG(DEBUG, VHOST_CONFIG,
 555                                 "(%d) failed to map desc_packed ring.\n",
 556                                 dev->vid);
 557                         return dev;
 558                 }
 559
 560                 dev = numa_realloc(dev, vq_index);
 561                 vq = dev->virtqueue[vq_index];
 562                 addr = &vq->ring_addrs;
 563
 564                 len = sizeof(struct vring_packed_desc_event);
 565                 vq->driver_event = (struct vring_packed_desc_event *)
 566                                         (uintptr_t)ring_addr_to_vva(dev,
 567                                         vq, addr->avail_user_addr, &len);
 568                 if (vq->driver_event == NULL ||
 569                                 len != sizeof(struct vring_packed_desc_event)) {
 570                         RTE_LOG(DEBUG, VHOST_CONFIG,
 571                                 "(%d) failed to find driver area address.\n",
 572                                 dev->vid);
 573                         return dev;
 574                 }
 575
 576                 len = sizeof(struct vring_packed_desc_event);
 577                 vq->device_event = (struct vring_packed_desc_event *)
 578                                         (uintptr_t)ring_addr_to_vva(dev,
 579                                         vq, addr->used_user_addr, &len);
 580                 if (vq->device_event == NULL ||
 581                                 len != sizeof(struct vring_packed_desc_event)) {
 582                         RTE_LOG(DEBUG, VHOST_CONFIG,
 583                                 "(%d) failed to find device area address.\n",
 584                                 dev->vid);
 585                         return dev;
 586                 }
 587
 588                 return dev;
 589         }
 590
 591         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 592         if (vq->desc && vq->avail && vq->used)
 593                 return dev;
 594
 595         len = sizeof(struct vring_desc) * vq->size;
 596         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 597                         vq, addr->desc_user_addr, &len);
 598         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 599                 RTE_LOG(DEBUG, VHOST_CONFIG,
 600                         "(%d) failed to map desc ring.\n",
 601                         dev->vid);
 602                 return dev;
 603         }
 604
 605         dev = numa_realloc(dev, vq_index);
 606         vq = dev->virtqueue[vq_index];
 607         addr = &vq->ring_addrs;
 608
 609         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 610         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 611                 len += sizeof(uint16_t);
 612         expected_len = len;
 613         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 614                         vq, addr->avail_user_addr, &len);
 615         if (vq->avail == 0 || len != expected_len) {
 616                 RTE_LOG(DEBUG, VHOST_CONFIG,
 617                         "(%d) failed to map avail ring.\n",
 618                         dev->vid);
 619                 return dev;
 620         }
 621
 622         len = sizeof(struct vring_used) +
 623                 sizeof(struct vring_used_elem) * vq->size;
 624         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 625                 len += sizeof(uint16_t);
 626         expected_len = len;
 627         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 628                         vq, addr->used_user_addr, &len);
 629         if (vq->used == 0 || len != expected_len) {
 630                 RTE_LOG(DEBUG, VHOST_CONFIG,
 631                         "(%d) failed to map used ring.\n",
 632                         dev->vid);
 633                 return dev;
 634         }
 635
 636         if (vq->last_used_idx != vq->used->idx) {
 637                 RTE_LOG(WARNING, VHOST_CONFIG,
 638                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 639                         "some packets maybe resent for Tx and dropped for Rx\n",
 640                         vq->last_used_idx, vq->used->idx);
 641                 vq->last_used_idx  = vq->used->idx;
 642                 vq->last_avail_idx = vq->used->idx;
 643         }
 644
 645         vq->log_guest_addr = addr->log_guest_addr;
 646
 647         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
 648                         dev->vid, vq->desc);
 649         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
 650                         dev->vid, vq->avail);
 651         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
 652                         dev->vid, vq->used);
 653         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
 654                         dev->vid, vq->log_guest_addr);
 655
 656         return dev;
 657 }
 658
 659 /*
 660  * The virtio device sends us the desc, used and avail ring addresses.
 661  * This function then converts these to our address space.
 662  */
 663 static int
 664 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 665                         int main_fd __rte_unused)
 666 {
 667         struct virtio_net *dev = *pdev;
 668         struct vhost_virtqueue *vq;
 669         struct vhost_vring_addr *addr = &msg->payload.addr;
 670
 671         if (dev->mem == NULL)
 672                 return RTE_VHOST_MSG_RESULT_ERR;
 673
 674         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 675         vq = dev->virtqueue[msg->payload.addr.index];
 676
 677         /*
 678          * Rings addresses should not be interpreted as long as the ring is not
 679          * started and enabled
 680          */
 681         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 682
 683         vring_invalidate(dev, vq);
 684
 685         if (vq->enabled && (dev->features &
 686                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
 687                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 688                 if (!dev)
 689                         return RTE_VHOST_MSG_RESULT_ERR;
 690
 691                 *pdev = dev;
 692         }
 693
 694         return RTE_VHOST_MSG_RESULT_OK;
 695 }
 696
 697 /*
 698  * The virtio device sends us the available ring last used index.
 699  */
 700 static int
 701 vhost_user_set_vring_base(struct virtio_net **pdev,
 702                         struct VhostUserMsg *msg,
 703                         int main_fd __rte_unused)
 704 {
 705         struct virtio_net *dev = *pdev;
 706         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 707         uint64_t val = msg->payload.state.num;
 708
 709         if (vq_is_packed(dev)) {
 710                 /*
 711                  * Bit[0:14]: avail index
 712                  * Bit[15]: avail wrap counter
 713                  */
 714                 vq->last_avail_idx = val & 0x7fff;
 715                 vq->avail_wrap_counter = !!(val & (0x1 << 15));
 716                 /*
 717                  * Set used index to same value as available one, as
 718                  * their values should be the same since ring processing
 719                  * was stopped at get time.
 720                  */
 721                 vq->last_used_idx = vq->last_avail_idx;
 722                 vq->used_wrap_counter = vq->avail_wrap_counter;
 723         } else {
 724                 vq->last_used_idx = msg->payload.state.num;
 725                 vq->last_avail_idx = msg->payload.state.num;
 726         }
 727
 728         return RTE_VHOST_MSG_RESULT_OK;
 729 }
 730
 731 static int
 732 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 733                    uint64_t host_phys_addr, uint64_t size)
 734 {
 735         struct guest_page *page, *last_page;
 736         struct guest_page *old_pages;
 737
 738         if (dev->nr_guest_pages == dev->max_guest_pages) {
 739                 dev->max_guest_pages *= 2;
 740                 old_pages = dev->guest_pages;
 741                 dev->guest_pages = realloc(dev->guest_pages,
 742                                         dev->max_guest_pages * sizeof(*page));
 743                 if (!dev->guest_pages) {
 744                         RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
 745                         free(old_pages);
 746                         return -1;
 747                 }
 748         }
 749
 750         if (dev->nr_guest_pages > 0) {
 751                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 752                 /* merge if the two pages are continuous */
 753                 if (host_phys_addr == last_page->host_phys_addr +
 754                                       last_page->size) {
 755                         last_page->size += size;
 756                         return 0;
 757                 }
 758         }
 759
 760         page = &dev->guest_pages[dev->nr_guest_pages++];
 761         page->guest_phys_addr = guest_phys_addr;
 762         page->host_phys_addr  = host_phys_addr;
 763         page->size = size;
 764
 765         return 0;
 766 }
 767
 768 static int
 769 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 770                 uint64_t page_size)
 771 {
 772         uint64_t reg_size = reg->size;
 773         uint64_t host_user_addr  = reg->host_user_addr;
 774         uint64_t guest_phys_addr = reg->guest_phys_addr;
 775         uint64_t host_phys_addr;
 776         uint64_t size;
 777
 778         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 779         size = page_size - (guest_phys_addr & (page_size - 1));
 780         size = RTE_MIN(size, reg_size);
 781
 782         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 783                 return -1;
 784
 785         host_user_addr  += size;
 786         guest_phys_addr += size;
 787         reg_size -= size;
 788
 789         while (reg_size > 0) {
 790                 size = RTE_MIN(reg_size, page_size);
 791                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 792                                                   host_user_addr);
 793                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 794                                 size) < 0)
 795                         return -1;
 796
 797                 host_user_addr  += size;
 798                 guest_phys_addr += size;
 799                 reg_size -= size;
 800         }
 801
 802         return 0;
 803 }
 804
 805 #ifdef RTE_LIBRTE_VHOST_DEBUG
 806 /* TODO: enable it only in debug mode? */
 807 static void
 808 dump_guest_pages(struct virtio_net *dev)
 809 {
 810         uint32_t i;
 811         struct guest_page *page;
 812
 813         for (i = 0; i < dev->nr_guest_pages; i++) {
 814                 page = &dev->guest_pages[i];
 815
 816                 RTE_LOG(INFO, VHOST_CONFIG,
 817                         "guest physical page region %u\n"
 818                         "\t guest_phys_addr: %" PRIx64 "\n"
 819                         "\t host_phys_addr : %" PRIx64 "\n"
 820                         "\t size           : %" PRIx64 "\n",
 821                         i,
 822                         page->guest_phys_addr,
 823                         page->host_phys_addr,
 824                         page->size);
 825         }
 826 }
 827 #else
 828 #define dump_guest_pages(dev)
 829 #endif
 830
 831 static bool
 832 vhost_memory_changed(struct VhostUserMemory *new,
 833                      struct rte_vhost_memory *old)
 834 {
 835         uint32_t i;
 836
 837         if (new->nregions != old->nregions)
 838                 return true;
 839
 840         for (i = 0; i < new->nregions; ++i) {
 841                 VhostUserMemoryRegion *new_r = &new->regions[i];
 842                 struct rte_vhost_mem_region *old_r = &old->regions[i];
 843
 844                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
 845                         return true;
 846                 if (new_r->memory_size != old_r->size)
 847                         return true;
 848                 if (new_r->userspace_addr != old_r->guest_user_addr)
 849                         return true;
 850         }
 851
 852         return false;
 853 }
 854
 855 static int
 856 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
 857                         int main_fd)
 858 {
 859         struct virtio_net *dev = *pdev;
 860         struct VhostUserMemory *memory = &msg->payload.memory;
 861         struct rte_vhost_mem_region *reg;
 862         void *mmap_addr;
 863         uint64_t mmap_size;
 864         uint64_t mmap_offset;
 865         uint64_t alignment;
 866         uint32_t i;
 867         int populate;
 868         int fd;
 869
 870         if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
 871                 RTE_LOG(ERR, VHOST_CONFIG,
 872                         "too many memory regions (%u)\n", memory->nregions);
 873                 return RTE_VHOST_MSG_RESULT_ERR;
 874         }
 875
 876         if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
 877                 RTE_LOG(INFO, VHOST_CONFIG,
 878                         "(%d) memory regions not changed\n", dev->vid);
 879
 880                 for (i = 0; i < memory->nregions; i++)
 881                         close(msg->fds[i]);
 882
 883                 return RTE_VHOST_MSG_RESULT_OK;
 884         }
 885
 886         if (dev->mem) {
 887                 free_mem_region(dev);
 888                 rte_free(dev->mem);
 889                 dev->mem = NULL;
 890         }
 891
 892         /* Flush IOTLB cache as previous HVAs are now invalid */
 893         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 894                 for (i = 0; i < dev->nr_vring; i++)
 895                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
 896
 897         dev->nr_guest_pages = 0;
 898         if (!dev->guest_pages) {
 899                 dev->max_guest_pages = 8;
 900                 dev->guest_pages = malloc(dev->max_guest_pages *
 901                                                 sizeof(struct guest_page));
 902                 if (dev->guest_pages == NULL) {
 903                         RTE_LOG(ERR, VHOST_CONFIG,
 904                                 "(%d) failed to allocate memory "
 905                                 "for dev->guest_pages\n",
 906                                 dev->vid);
 907                         return RTE_VHOST_MSG_RESULT_ERR;
 908                 }
 909         }
 910
 911         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
 912                 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
 913         if (dev->mem == NULL) {
 914                 RTE_LOG(ERR, VHOST_CONFIG,
 915                         "(%d) failed to allocate memory for dev->mem\n",
 916                         dev->vid);
 917                 return RTE_VHOST_MSG_RESULT_ERR;
 918         }
 919         dev->mem->nregions = memory->nregions;
 920
 921         for (i = 0; i < memory->nregions; i++) {
 922                 fd  = msg->fds[i];
 923                 reg = &dev->mem->regions[i];
 924
 925                 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
 926                 reg->guest_user_addr = memory->regions[i].userspace_addr;
 927                 reg->size            = memory->regions[i].memory_size;
 928                 reg->fd              = fd;
 929
 930                 mmap_offset = memory->regions[i].mmap_offset;
 931
 932                 /* Check for memory_size + mmap_offset overflow */
 933                 if (mmap_offset >= -reg->size) {
 934                         RTE_LOG(ERR, VHOST_CONFIG,
 935                                 "mmap_offset (%#"PRIx64") and memory_size "
 936                                 "(%#"PRIx64") overflow\n",
 937                                 mmap_offset, reg->size);
 938                         goto err_mmap;
 939                 }
 940
 941                 mmap_size = reg->size + mmap_offset;
 942
 943                 /* mmap() without flag of MAP_ANONYMOUS, should be called
 944                  * with length argument aligned with hugepagesz at older
 945                  * longterm version Linux, like 2.6.32 and 3.2.72, or
 946                  * mmap() will fail with EINVAL.
 947                  *
 948                  * to avoid failure, make sure in caller to keep length
 949                  * aligned.
 950                  */
 951                 alignment = get_blk_size(fd);
 952                 if (alignment == (uint64_t)-1) {
 953                         RTE_LOG(ERR, VHOST_CONFIG,
 954                                 "couldn't get hugepage size through fstat\n");
 955                         goto err_mmap;
 956                 }
 957                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
 958
 959                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
 960                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
 961                                  MAP_SHARED | populate, fd, 0);
 962
 963                 if (mmap_addr == MAP_FAILED) {
 964                         RTE_LOG(ERR, VHOST_CONFIG,
 965                                 "mmap region %u failed.\n", i);
 966                         goto err_mmap;
 967                 }
 968
 969                 reg->mmap_addr = mmap_addr;
 970                 reg->mmap_size = mmap_size;
 971                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
 972                                       mmap_offset;
 973
 974                 if (dev->dequeue_zero_copy)
 975                         if (add_guest_pages(dev, reg, alignment) < 0) {
 976                                 RTE_LOG(ERR, VHOST_CONFIG,
 977                                         "adding guest pages to region %u failed.\n",
 978                                         i);
 979                                 goto err_mmap;
 980                         }
 981
 982                 RTE_LOG(INFO, VHOST_CONFIG,
 983                         "guest memory region %u, size: 0x%" PRIx64 "\n"
 984                         "\t guest physical addr: 0x%" PRIx64 "\n"
 985                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
 986                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
 987                         "\t mmap addr : 0x%" PRIx64 "\n"
 988                         "\t mmap size : 0x%" PRIx64 "\n"
 989                         "\t mmap align: 0x%" PRIx64 "\n"
 990                         "\t mmap off  : 0x%" PRIx64 "\n",
 991                         i, reg->size,
 992                         reg->guest_phys_addr,
 993                         reg->guest_user_addr,
 994                         reg->host_user_addr,
 995                         (uint64_t)(uintptr_t)mmap_addr,
 996                         mmap_size,
 997                         alignment,
 998                         mmap_offset);
 999
1000                 if (dev->postcopy_listening) {
1001                         /*
1002                          * We haven't a better way right now than sharing
1003                          * DPDK's virtual address with Qemu, so that Qemu can
1004                          * retrieve the region offset when handling userfaults.
1005                          */
1006                         memory->regions[i].userspace_addr =
1007                                 reg->host_user_addr;
1008                 }
1009         }
1010         if (dev->postcopy_listening) {
1011                 /* Send the addresses back to qemu */
1012                 msg->fd_num = 0;
1013                 send_vhost_reply(main_fd, msg);
1014
1015                 /* Wait for qemu to acknolwedge it's got the addresses
1016                  * we've got to wait before we're allowed to generate faults.
1017                  */
1018                 VhostUserMsg ack_msg;
1019                 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1020                         RTE_LOG(ERR, VHOST_CONFIG,
1021                                 "Failed to read qemu ack on postcopy set-mem-table\n");
1022                         goto err_mmap;
1023                 }
1024                 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1025                         RTE_LOG(ERR, VHOST_CONFIG,
1026                                 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1027                                 ack_msg.request.master);
1028                         goto err_mmap;
1029                 }
1030
1031                 /* Now userfault register and we can use the memory */
1032                 for (i = 0; i < memory->nregions; i++) {
1033 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1034                         reg = &dev->mem->regions[i];
1035                         struct uffdio_register reg_struct;
1036
1037                         /*
1038                          * Let's register all the mmap'ed area to ensure
1039                          * alignment on page boundary.
1040                          */
1041                         reg_struct.range.start =
1042                                 (uint64_t)(uintptr_t)reg->mmap_addr;
1043                         reg_struct.range.len = reg->mmap_size;
1044                         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1045
1046                         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1047                                                 &reg_struct)) {
1048                                 RTE_LOG(ERR, VHOST_CONFIG,
1049                                         "Failed to register ufd for region %d: (ufd = %d) %s\n",
1050                                         i, dev->postcopy_ufd,
1051                                         strerror(errno));
1052                                 goto err_mmap;
1053                         }
1054                         RTE_LOG(INFO, VHOST_CONFIG,
1055                                 "\t userfaultfd registered for range : %llx - %llx\n",
1056                                 reg_struct.range.start,
1057                                 reg_struct.range.start +
1058                                 reg_struct.range.len - 1);
1059 #else
1060                         goto err_mmap;
1061 #endif
1062                 }
1063         }
1064
1065         for (i = 0; i < dev->nr_vring; i++) {
1066                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1067
1068                 if (vq->desc || vq->avail || vq->used) {
1069                         /*
1070                          * If the memory table got updated, the ring addresses
1071                          * need to be translated again as virtual addresses have
1072                          * changed.
1073                          */
1074                         vring_invalidate(dev, vq);
1075
1076                         dev = translate_ring_addresses(dev, i);
1077                         if (!dev) {
1078                                 dev = *pdev;
1079                                 goto err_mmap;
1080                         }
1081
1082                         *pdev = dev;
1083                 }
1084         }
1085
1086         dump_guest_pages(dev);
1087
1088         return RTE_VHOST_MSG_RESULT_OK;
1089
1090 err_mmap:
1091         free_mem_region(dev);
1092         rte_free(dev->mem);
1093         dev->mem = NULL;
1094         return RTE_VHOST_MSG_RESULT_ERR;
1095 }
1096
1097 static bool
1098 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1099 {
1100         bool rings_ok;
1101
1102         if (!vq)
1103                 return false;
1104
1105         if (vq_is_packed(dev))
1106                 rings_ok = !!vq->desc_packed;
1107         else
1108                 rings_ok = vq->desc && vq->avail && vq->used;
1109
1110         return rings_ok &&
1111                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1112                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1113 }
1114
1115 static int
1116 virtio_is_ready(struct virtio_net *dev)
1117 {
1118         struct vhost_virtqueue *vq;
1119         uint32_t i;
1120
1121         if (dev->nr_vring == 0)
1122                 return 0;
1123
1124         for (i = 0; i < dev->nr_vring; i++) {
1125                 vq = dev->virtqueue[i];
1126
1127                 if (!vq_is_ready(dev, vq))
1128                         return 0;
1129         }
1130
1131         RTE_LOG(INFO, VHOST_CONFIG,
1132                 "virtio is now ready for processing.\n");
1133         return 1;
1134 }
1135
1136 static int
1137 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1138                         int main_fd __rte_unused)
1139 {
1140         struct virtio_net *dev = *pdev;
1141         struct vhost_vring_file file;
1142         struct vhost_virtqueue *vq;
1143
1144         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1145         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1146                 file.fd = VIRTIO_INVALID_EVENTFD;
1147         else
1148                 file.fd = msg->fds[0];
1149         RTE_LOG(INFO, VHOST_CONFIG,
1150                 "vring call idx:%d file:%d\n", file.index, file.fd);
1151
1152         vq = dev->virtqueue[file.index];
1153         if (vq->callfd >= 0)
1154                 close(vq->callfd);
1155
1156         vq->callfd = file.fd;
1157
1158         return RTE_VHOST_MSG_RESULT_OK;
1159 }
1160
1161 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1162                         struct VhostUserMsg *msg,
1163                         int main_fd __rte_unused)
1164 {
1165         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1166                 close(msg->fds[0]);
1167         RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1168
1169         return RTE_VHOST_MSG_RESULT_OK;
1170 }
1171
1172 static int
1173 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1174                         int main_fd __rte_unused)
1175 {
1176         struct virtio_net *dev = *pdev;
1177         struct vhost_vring_file file;
1178         struct vhost_virtqueue *vq;
1179
1180         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1181         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1182                 file.fd = VIRTIO_INVALID_EVENTFD;
1183         else
1184                 file.fd = msg->fds[0];
1185         RTE_LOG(INFO, VHOST_CONFIG,
1186                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1187
1188         /* Interpret ring addresses only when ring is started. */
1189         dev = translate_ring_addresses(dev, file.index);
1190         if (!dev)
1191                 return RTE_VHOST_MSG_RESULT_ERR;
1192
1193         *pdev = dev;
1194
1195         vq = dev->virtqueue[file.index];
1196
1197         /*
1198          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1199          * the ring starts already enabled. Otherwise, it is enabled via
1200          * the SET_VRING_ENABLE message.
1201          */
1202         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)))
1203                 vq->enabled = 1;
1204
1205         if (vq->kickfd >= 0)
1206                 close(vq->kickfd);
1207         vq->kickfd = file.fd;
1208
1209         return RTE_VHOST_MSG_RESULT_OK;
1210 }
1211
1212 static void
1213 free_zmbufs(struct vhost_virtqueue *vq)
1214 {
1215         struct zcopy_mbuf *zmbuf, *next;
1216
1217         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1218              zmbuf != NULL; zmbuf = next) {
1219                 next = TAILQ_NEXT(zmbuf, next);
1220
1221                 while (!mbuf_is_consumed(zmbuf->mbuf))
1222                         usleep(1000);
1223
1224                 restore_mbuf(zmbuf->mbuf);
1225                 rte_pktmbuf_free(zmbuf->mbuf);
1226                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1227         }
1228
1229         rte_free(vq->zmbufs);
1230 }
1231
1232 /*
1233  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1234  */
1235 static int
1236 vhost_user_get_vring_base(struct virtio_net **pdev,
1237                         struct VhostUserMsg *msg,
1238                         int main_fd __rte_unused)
1239 {
1240         struct virtio_net *dev = *pdev;
1241         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1242         uint64_t val;
1243
1244         /* We have to stop the queue (virtio) if it is running. */
1245         vhost_destroy_device_notify(dev);
1246
1247         dev->flags &= ~VIRTIO_DEV_READY;
1248         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1249
1250         /* Here we are safe to get the indexes */
1251         if (vq_is_packed(dev)) {
1252                 /*
1253                  * Bit[0:14]: avail index
1254                  * Bit[15]: avail wrap counter
1255                  */
1256                 val = vq->last_avail_idx & 0x7fff;
1257                 val |= vq->avail_wrap_counter << 15;
1258                 msg->payload.state.num = val;
1259         } else {
1260                 msg->payload.state.num = vq->last_avail_idx;
1261         }
1262
1263         RTE_LOG(INFO, VHOST_CONFIG,
1264                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1265                 msg->payload.state.num);
1266         /*
1267          * Based on current qemu vhost-user implementation, this message is
1268          * sent and only sent in vhost_vring_stop.
1269          * TODO: cleanup the vring, it isn't usable since here.
1270          */
1271         if (vq->kickfd >= 0)
1272                 close(vq->kickfd);
1273
1274         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1275
1276         if (vq->callfd >= 0)
1277                 close(vq->callfd);
1278
1279         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1280
1281         if (dev->dequeue_zero_copy)
1282                 free_zmbufs(vq);
1283         if (vq_is_packed(dev)) {
1284                 rte_free(vq->shadow_used_packed);
1285                 vq->shadow_used_packed = NULL;
1286         } else {
1287                 rte_free(vq->shadow_used_split);
1288                 vq->shadow_used_split = NULL;
1289         }
1290
1291         rte_free(vq->batch_copy_elems);
1292         vq->batch_copy_elems = NULL;
1293
1294         msg->size = sizeof(msg->payload.state);
1295         msg->fd_num = 0;
1296
1297         return RTE_VHOST_MSG_RESULT_REPLY;
1298 }
1299
1300 /*
1301  * when virtio queues are ready to work, qemu will send us to
1302  * enable the virtio queue pair.
1303  */
1304 static int
1305 vhost_user_set_vring_enable(struct virtio_net **pdev,
1306                         struct VhostUserMsg *msg,
1307                         int main_fd __rte_unused)
1308 {
1309         struct virtio_net *dev = *pdev;
1310         int enable = (int)msg->payload.state.num;
1311         int index = (int)msg->payload.state.index;
1312         struct rte_vdpa_device *vdpa_dev;
1313         int did = -1;
1314
1315         RTE_LOG(INFO, VHOST_CONFIG,
1316                 "set queue enable: %d to qp idx: %d\n",
1317                 enable, index);
1318
1319         did = dev->vdpa_dev_id;
1320         vdpa_dev = rte_vdpa_get_device(did);
1321         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1322                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1323
1324         if (dev->notify_ops->vring_state_changed)
1325                 dev->notify_ops->vring_state_changed(dev->vid,
1326                                 index, enable);
1327
1328         dev->virtqueue[index]->enabled = enable;
1329
1330         return RTE_VHOST_MSG_RESULT_OK;
1331 }
1332
1333 static int
1334 vhost_user_get_protocol_features(struct virtio_net **pdev,
1335                         struct VhostUserMsg *msg,
1336                         int main_fd __rte_unused)
1337 {
1338         struct virtio_net *dev = *pdev;
1339         uint64_t features, protocol_features;
1340
1341         rte_vhost_driver_get_features(dev->ifname, &features);
1342         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1343
1344         /*
1345          * REPLY_ACK protocol feature is only mandatory for now
1346          * for IOMMU feature. If IOMMU is explicitly disabled by the
1347          * application, disable also REPLY_ACK feature for older buggy
1348          * Qemu versions (from v2.7.0 to v2.9.0).
1349          */
1350         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1351                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1352
1353         msg->payload.u64 = protocol_features;
1354         msg->size = sizeof(msg->payload.u64);
1355         msg->fd_num = 0;
1356
1357         return RTE_VHOST_MSG_RESULT_REPLY;
1358 }
1359
1360 static int
1361 vhost_user_set_protocol_features(struct virtio_net **pdev,
1362                         struct VhostUserMsg *msg,
1363                         int main_fd __rte_unused)
1364 {
1365         struct virtio_net *dev = *pdev;
1366         uint64_t protocol_features = msg->payload.u64;
1367         uint64_t slave_protocol_features = 0;
1368
1369         rte_vhost_driver_get_protocol_features(dev->ifname,
1370                         &slave_protocol_features);
1371         if (protocol_features & ~slave_protocol_features) {
1372                 RTE_LOG(ERR, VHOST_CONFIG,
1373                         "(%d) received invalid protocol features.\n",
1374                         dev->vid);
1375                 return RTE_VHOST_MSG_RESULT_ERR;
1376         }
1377
1378         dev->protocol_features = protocol_features;
1379
1380         return RTE_VHOST_MSG_RESULT_OK;
1381 }
1382
1383 static int
1384 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1385                         int main_fd __rte_unused)
1386 {
1387         struct virtio_net *dev = *pdev;
1388         int fd = msg->fds[0];
1389         uint64_t size, off;
1390         void *addr;
1391
1392         if (fd < 0) {
1393                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1394                 return RTE_VHOST_MSG_RESULT_ERR;
1395         }
1396
1397         if (msg->size != sizeof(VhostUserLog)) {
1398                 RTE_LOG(ERR, VHOST_CONFIG,
1399                         "invalid log base msg size: %"PRId32" != %d\n",
1400                         msg->size, (int)sizeof(VhostUserLog));
1401                 return RTE_VHOST_MSG_RESULT_ERR;
1402         }
1403
1404         size = msg->payload.log.mmap_size;
1405         off  = msg->payload.log.mmap_offset;
1406
1407         /* Don't allow mmap_offset to point outside the mmap region */
1408         if (off > size) {
1409                 RTE_LOG(ERR, VHOST_CONFIG,
1410                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1411                         off, size);
1412                 return RTE_VHOST_MSG_RESULT_ERR;
1413         }
1414
1415         RTE_LOG(INFO, VHOST_CONFIG,
1416                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1417                 size, off);
1418
1419         /*
1420          * mmap from 0 to workaround a hugepage mmap bug: mmap will
1421          * fail when offset is not page size aligned.
1422          */
1423         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1424         close(fd);
1425         if (addr == MAP_FAILED) {
1426                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1427                 return RTE_VHOST_MSG_RESULT_ERR;
1428         }
1429
1430         /*
1431          * Free previously mapped log memory on occasionally
1432          * multiple VHOST_USER_SET_LOG_BASE.
1433          */
1434         if (dev->log_addr) {
1435                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1436         }
1437         dev->log_addr = (uint64_t)(uintptr_t)addr;
1438         dev->log_base = dev->log_addr + off;
1439         dev->log_size = size;
1440
1441         /*
1442          * The spec is not clear about it (yet), but QEMU doesn't expect
1443          * any payload in the reply.
1444          */
1445         msg->size = 0;
1446         msg->fd_num = 0;
1447
1448         return RTE_VHOST_MSG_RESULT_REPLY;
1449 }
1450
1451 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1452                         struct VhostUserMsg *msg,
1453                         int main_fd __rte_unused)
1454 {
1455         close(msg->fds[0]);
1456         RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1457
1458         return RTE_VHOST_MSG_RESULT_OK;
1459 }
1460
1461 /*
1462  * An rarp packet is constructed and broadcasted to notify switches about
1463  * the new location of the migrated VM, so that packets from outside will
1464  * not be lost after migration.
1465  *
1466  * However, we don't actually "send" a rarp packet here, instead, we set
1467  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1468  */
1469 static int
1470 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
1471                         int main_fd __rte_unused)
1472 {
1473         struct virtio_net *dev = *pdev;
1474         uint8_t *mac = (uint8_t *)&msg->payload.u64;
1475         struct rte_vdpa_device *vdpa_dev;
1476         int did = -1;
1477
1478         RTE_LOG(DEBUG, VHOST_CONFIG,
1479                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1480                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1481         memcpy(dev->mac.addr_bytes, mac, 6);
1482
1483         /*
1484          * Set the flag to inject a RARP broadcast packet at
1485          * rte_vhost_dequeue_burst().
1486          *
1487          * rte_smp_wmb() is for making sure the mac is copied
1488          * before the flag is set.
1489          */
1490         rte_smp_wmb();
1491         rte_atomic16_set(&dev->broadcast_rarp, 1);
1492         did = dev->vdpa_dev_id;
1493         vdpa_dev = rte_vdpa_get_device(did);
1494         if (vdpa_dev && vdpa_dev->ops->migration_done)
1495                 vdpa_dev->ops->migration_done(dev->vid);
1496
1497         return RTE_VHOST_MSG_RESULT_OK;
1498 }
1499
1500 static int
1501 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
1502                         int main_fd __rte_unused)
1503 {
1504         struct virtio_net *dev = *pdev;
1505         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1506                         msg->payload.u64 > VIRTIO_MAX_MTU) {
1507                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1508                                 msg->payload.u64);
1509
1510                 return RTE_VHOST_MSG_RESULT_ERR;
1511         }
1512
1513         dev->mtu = msg->payload.u64;
1514
1515         return RTE_VHOST_MSG_RESULT_OK;
1516 }
1517
1518 static int
1519 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
1520                         int main_fd __rte_unused)
1521 {
1522         struct virtio_net *dev = *pdev;
1523         int fd = msg->fds[0];
1524
1525         if (fd < 0) {
1526                 RTE_LOG(ERR, VHOST_CONFIG,
1527                                 "Invalid file descriptor for slave channel (%d)\n",
1528                                 fd);
1529                 return RTE_VHOST_MSG_RESULT_ERR;
1530         }
1531
1532         dev->slave_req_fd = fd;
1533
1534         return RTE_VHOST_MSG_RESULT_OK;
1535 }
1536
1537 static int
1538 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1539 {
1540         struct vhost_vring_addr *ra;
1541         uint64_t start, end;
1542
1543         start = imsg->iova;
1544         end = start + imsg->size;
1545
1546         ra = &vq->ring_addrs;
1547         if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1548                 return 1;
1549         if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1550                 return 1;
1551         if (ra->used_user_addr >= start && ra->used_user_addr < end)
1552                 return 1;
1553
1554         return 0;
1555 }
1556
1557 static int
1558 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1559                                 struct vhost_iotlb_msg *imsg)
1560 {
1561         uint64_t istart, iend, vstart, vend;
1562
1563         istart = imsg->iova;
1564         iend = istart + imsg->size - 1;
1565
1566         vstart = (uintptr_t)vq->desc;
1567         vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1568         if (vstart <= iend && istart <= vend)
1569                 return 1;
1570
1571         vstart = (uintptr_t)vq->avail;
1572         vend = vstart + sizeof(struct vring_avail);
1573         vend += sizeof(uint16_t) * vq->size - 1;
1574         if (vstart <= iend && istart <= vend)
1575                 return 1;
1576
1577         vstart = (uintptr_t)vq->used;
1578         vend = vstart + sizeof(struct vring_used);
1579         vend += sizeof(struct vring_used_elem) * vq->size - 1;
1580         if (vstart <= iend && istart <= vend)
1581                 return 1;
1582
1583         return 0;
1584 }
1585
1586 static int
1587 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
1588                         int main_fd __rte_unused)
1589 {
1590         struct virtio_net *dev = *pdev;
1591         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1592         uint16_t i;
1593         uint64_t vva, len;
1594
1595         switch (imsg->type) {
1596         case VHOST_IOTLB_UPDATE:
1597                 len = imsg->size;
1598                 vva = qva_to_vva(dev, imsg->uaddr, &len);
1599                 if (!vva)
1600                         return RTE_VHOST_MSG_RESULT_ERR;
1601
1602                 for (i = 0; i < dev->nr_vring; i++) {
1603                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1604
1605                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1606                                         len, imsg->perm);
1607
1608                         if (is_vring_iotlb_update(vq, imsg))
1609                                 *pdev = dev = translate_ring_addresses(dev, i);
1610                 }
1611                 break;
1612         case VHOST_IOTLB_INVALIDATE:
1613                 for (i = 0; i < dev->nr_vring; i++) {
1614                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1615
1616                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
1617                                         imsg->size);
1618
1619                         if (is_vring_iotlb_invalidate(vq, imsg))
1620                                 vring_invalidate(dev, vq);
1621                 }
1622                 break;
1623         default:
1624                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1625                                 imsg->type);
1626                 return RTE_VHOST_MSG_RESULT_ERR;
1627         }
1628
1629         return RTE_VHOST_MSG_RESULT_OK;
1630 }
1631
1632 static int
1633 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
1634                         struct VhostUserMsg *msg,
1635                         int main_fd __rte_unused)
1636 {
1637         struct virtio_net *dev = *pdev;
1638 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1639         struct uffdio_api api_struct;
1640
1641         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1642
1643         if (dev->postcopy_ufd == -1) {
1644                 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
1645                         strerror(errno));
1646                 return RTE_VHOST_MSG_RESULT_ERR;
1647         }
1648         api_struct.api = UFFD_API;
1649         api_struct.features = 0;
1650         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
1651                 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
1652                         strerror(errno));
1653                 close(dev->postcopy_ufd);
1654                 dev->postcopy_ufd = -1;
1655                 return RTE_VHOST_MSG_RESULT_ERR;
1656         }
1657         msg->fds[0] = dev->postcopy_ufd;
1658         msg->fd_num = 1;
1659
1660         return RTE_VHOST_MSG_RESULT_REPLY;
1661 #else
1662         dev->postcopy_ufd = -1;
1663         msg->fd_num = 0;
1664
1665         return RTE_VHOST_MSG_RESULT_ERR;
1666 #endif
1667 }
1668
1669 static int
1670 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
1671                         struct VhostUserMsg *msg __rte_unused,
1672                         int main_fd __rte_unused)
1673 {
1674         struct virtio_net *dev = *pdev;
1675
1676         if (dev->mem && dev->mem->nregions) {
1677                 RTE_LOG(ERR, VHOST_CONFIG,
1678                         "Regions already registered at postcopy-listen\n");
1679                 return RTE_VHOST_MSG_RESULT_ERR;
1680         }
1681         dev->postcopy_listening = 1;
1682
1683         return RTE_VHOST_MSG_RESULT_OK;
1684 }
1685
1686 static int
1687 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
1688                         int main_fd __rte_unused)
1689 {
1690         struct virtio_net *dev = *pdev;
1691
1692         dev->postcopy_listening = 0;
1693         if (dev->postcopy_ufd >= 0) {
1694                 close(dev->postcopy_ufd);
1695                 dev->postcopy_ufd = -1;
1696         }
1697
1698         msg->payload.u64 = 0;
1699         msg->size = sizeof(msg->payload.u64);
1700         msg->fd_num = 0;
1701
1702         return RTE_VHOST_MSG_RESULT_REPLY;
1703 }
1704
1705 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
1706                                         struct VhostUserMsg *msg,
1707                                         int main_fd);
1708 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
1709         [VHOST_USER_NONE] = NULL,
1710         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
1711         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
1712         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
1713         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
1714         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
1715         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
1716         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
1717         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
1718         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
1719         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
1720         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
1721         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
1722         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
1723         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
1724         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
1725         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
1726         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
1727         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
1728         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
1729         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
1730         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
1731         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
1732         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
1733         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
1734         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
1735 };
1736
1737
1738 /* return bytes# of read on success or negative val on failure. */
1739 static int
1740 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1741 {
1742         int ret;
1743
1744         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1745                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
1746         if (ret <= 0)
1747                 return ret;
1748
1749         if (msg->size) {
1750                 if (msg->size > sizeof(msg->payload)) {
1751                         RTE_LOG(ERR, VHOST_CONFIG,
1752                                 "invalid msg size: %d\n", msg->size);
1753                         return -1;
1754                 }
1755                 ret = read(sockfd, &msg->payload, msg->size);
1756                 if (ret <= 0)
1757                         return ret;
1758                 if (ret != (int)msg->size) {
1759                         RTE_LOG(ERR, VHOST_CONFIG,
1760                                 "read control message failed\n");
1761                         return -1;
1762                 }
1763         }
1764
1765         return ret;
1766 }
1767
1768 static int
1769 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1770 {
1771         if (!msg)
1772                 return 0;
1773
1774         return send_fd_message(sockfd, (char *)msg,
1775                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
1776 }
1777
1778 static int
1779 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1780 {
1781         if (!msg)
1782                 return 0;
1783
1784         msg->flags &= ~VHOST_USER_VERSION_MASK;
1785         msg->flags &= ~VHOST_USER_NEED_REPLY;
1786         msg->flags |= VHOST_USER_VERSION;
1787         msg->flags |= VHOST_USER_REPLY_MASK;
1788
1789         return send_vhost_message(sockfd, msg);
1790 }
1791
1792 static int
1793 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
1794 {
1795         int ret;
1796
1797         if (msg->flags & VHOST_USER_NEED_REPLY)
1798                 rte_spinlock_lock(&dev->slave_req_lock);
1799
1800         ret = send_vhost_message(dev->slave_req_fd, msg);
1801         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1802                 rte_spinlock_unlock(&dev->slave_req_lock);
1803
1804         return ret;
1805 }
1806
1807 /*
1808  * Allocate a queue pair if it hasn't been allocated yet
1809  */
1810 static int
1811 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1812                         struct VhostUserMsg *msg)
1813 {
1814         uint16_t vring_idx;
1815
1816         switch (msg->request.master) {
1817         case VHOST_USER_SET_VRING_KICK:
1818         case VHOST_USER_SET_VRING_CALL:
1819         case VHOST_USER_SET_VRING_ERR:
1820                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1821                 break;
1822         case VHOST_USER_SET_VRING_NUM:
1823         case VHOST_USER_SET_VRING_BASE:
1824         case VHOST_USER_SET_VRING_ENABLE:
1825                 vring_idx = msg->payload.state.index;
1826                 break;
1827         case VHOST_USER_SET_VRING_ADDR:
1828                 vring_idx = msg->payload.addr.index;
1829                 break;
1830         default:
1831                 return 0;
1832         }
1833
1834         if (vring_idx >= VHOST_MAX_VRING) {
1835                 RTE_LOG(ERR, VHOST_CONFIG,
1836                         "invalid vring index: %u\n", vring_idx);
1837                 return -1;
1838         }
1839
1840         if (dev->virtqueue[vring_idx])
1841                 return 0;
1842
1843         return alloc_vring_queue(dev, vring_idx);
1844 }
1845
1846 static void
1847 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1848 {
1849         unsigned int i = 0;
1850         unsigned int vq_num = 0;
1851
1852         while (vq_num < dev->nr_vring) {
1853                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1854
1855                 if (vq) {
1856                         rte_spinlock_lock(&vq->access_lock);
1857                         vq_num++;
1858                 }
1859                 i++;
1860         }
1861 }
1862
1863 static void
1864 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1865 {
1866         unsigned int i = 0;
1867         unsigned int vq_num = 0;
1868
1869         while (vq_num < dev->nr_vring) {
1870                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1871
1872                 if (vq) {
1873                         rte_spinlock_unlock(&vq->access_lock);
1874                         vq_num++;
1875                 }
1876                 i++;
1877         }
1878 }
1879
1880 int
1881 vhost_user_msg_handler(int vid, int fd)
1882 {
1883         struct virtio_net *dev;
1884         struct VhostUserMsg msg;
1885         struct rte_vdpa_device *vdpa_dev;
1886         int did = -1;
1887         int ret;
1888         int unlock_required = 0;
1889         uint32_t skip_master = 0;
1890         int request;
1891
1892         dev = get_device(vid);
1893         if (dev == NULL)
1894                 return -1;
1895
1896         if (!dev->notify_ops) {
1897                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1898                 if (!dev->notify_ops) {
1899                         RTE_LOG(ERR, VHOST_CONFIG,
1900                                 "failed to get callback ops for driver %s\n",
1901                                 dev->ifname);
1902                         return -1;
1903                 }
1904         }
1905
1906         ret = read_vhost_message(fd, &msg);
1907         if (ret <= 0 || msg.request.master >= VHOST_USER_MAX) {
1908                 if (ret < 0)
1909                         RTE_LOG(ERR, VHOST_CONFIG,
1910                                 "vhost read message failed\n");
1911                 else if (ret == 0)
1912                         RTE_LOG(INFO, VHOST_CONFIG,
1913                                 "vhost peer closed\n");
1914                 else
1915                         RTE_LOG(ERR, VHOST_CONFIG,
1916                                 "vhost read incorrect message\n");
1917
1918                 return -1;
1919         }
1920
1921         ret = 0;
1922         if (msg.request.master != VHOST_USER_IOTLB_MSG)
1923                 RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1924                         vhost_message_str[msg.request.master]);
1925         else
1926                 RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1927                         vhost_message_str[msg.request.master]);
1928
1929         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1930         if (ret < 0) {
1931                 RTE_LOG(ERR, VHOST_CONFIG,
1932                         "failed to alloc queue\n");
1933                 return -1;
1934         }
1935
1936         /*
1937          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1938          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1939          * and device is destroyed. destroy_device waits for queues to be
1940          * inactive, so it is safe. Otherwise taking the access_lock
1941          * would cause a dead lock.
1942          */
1943         switch (msg.request.master) {
1944         case VHOST_USER_SET_FEATURES:
1945         case VHOST_USER_SET_PROTOCOL_FEATURES:
1946         case VHOST_USER_SET_OWNER:
1947         case VHOST_USER_SET_MEM_TABLE:
1948         case VHOST_USER_SET_LOG_BASE:
1949         case VHOST_USER_SET_LOG_FD:
1950         case VHOST_USER_SET_VRING_NUM:
1951         case VHOST_USER_SET_VRING_ADDR:
1952         case VHOST_USER_SET_VRING_BASE:
1953         case VHOST_USER_SET_VRING_KICK:
1954         case VHOST_USER_SET_VRING_CALL:
1955         case VHOST_USER_SET_VRING_ERR:
1956         case VHOST_USER_SET_VRING_ENABLE:
1957         case VHOST_USER_SEND_RARP:
1958         case VHOST_USER_NET_SET_MTU:
1959         case VHOST_USER_SET_SLAVE_REQ_FD:
1960                 vhost_user_lock_all_queue_pairs(dev);
1961                 unlock_required = 1;
1962                 break;
1963         default:
1964                 break;
1965
1966         }
1967
1968         if (dev->extern_ops.pre_msg_handle) {
1969                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
1970                                 (void *)&msg, &skip_master);
1971                 if (ret == RTE_VHOST_MSG_RESULT_ERR)
1972                         goto skip_to_reply;
1973                 else if (ret == RTE_VHOST_MSG_RESULT_REPLY)
1974                         send_vhost_reply(fd, &msg);
1975
1976                 if (skip_master)
1977                         goto skip_to_post_handle;
1978         }
1979
1980         request = msg.request.master;
1981         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
1982                 if (!vhost_message_handlers[request])
1983                         goto skip_to_post_handle;
1984                 ret = vhost_message_handlers[request](&dev, &msg, fd);
1985
1986                 switch (ret) {
1987                 case RTE_VHOST_MSG_RESULT_ERR:
1988                         RTE_LOG(ERR, VHOST_CONFIG,
1989                                 "Processing %s failed.\n",
1990                                 vhost_message_str[request]);
1991                         break;
1992                 case RTE_VHOST_MSG_RESULT_OK:
1993                         RTE_LOG(DEBUG, VHOST_CONFIG,
1994                                 "Processing %s succeeded.\n",
1995                                 vhost_message_str[request]);
1996                         break;
1997                 case RTE_VHOST_MSG_RESULT_REPLY:
1998                         RTE_LOG(DEBUG, VHOST_CONFIG,
1999                                 "Processing %s succeeded and needs reply.\n",
2000                                 vhost_message_str[request]);
2001                         send_vhost_reply(fd, &msg);
2002                         break;
2003                 }
2004         } else {
2005                 RTE_LOG(ERR, VHOST_CONFIG,
2006                         "Requested invalid message type %d.\n", request);
2007                 ret = RTE_VHOST_MSG_RESULT_ERR;
2008         }
2009
2010 skip_to_post_handle:
2011         if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2012                         dev->extern_ops.post_msg_handle) {
2013                 ret = (*dev->extern_ops.post_msg_handle)(
2014                                 dev->vid, (void *)&msg);
2015                 if (ret == RTE_VHOST_MSG_RESULT_ERR)
2016                         goto skip_to_reply;
2017                 else if (ret == RTE_VHOST_MSG_RESULT_REPLY)
2018                         send_vhost_reply(fd, &msg);
2019         }
2020
2021 skip_to_reply:
2022         if (unlock_required)
2023                 vhost_user_unlock_all_queue_pairs(dev);
2024
2025         /*
2026          * If the request required a reply that was already sent,
2027          * this optional reply-ack won't be sent as the
2028          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2029          */
2030         if (msg.flags & VHOST_USER_NEED_REPLY) {
2031                 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2032                 msg.size = sizeof(msg.payload.u64);
2033                 msg.fd_num = 0;
2034                 send_vhost_reply(fd, &msg);
2035         } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2036                 RTE_LOG(ERR, VHOST_CONFIG,
2037                         "vhost message handling failed.\n");
2038                 return -1;
2039         }
2040
2041         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2042                 dev->flags |= VIRTIO_DEV_READY;
2043
2044                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2045                         if (dev->dequeue_zero_copy) {
2046                                 RTE_LOG(INFO, VHOST_CONFIG,
2047                                                 "dequeue zero copy is enabled\n");
2048                         }
2049
2050                         if (dev->notify_ops->new_device(dev->vid) == 0)
2051                                 dev->flags |= VIRTIO_DEV_RUNNING;
2052                 }
2053         }
2054
2055         did = dev->vdpa_dev_id;
2056         vdpa_dev = rte_vdpa_get_device(did);
2057         if (vdpa_dev && virtio_is_ready(dev) &&
2058                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2059                         msg.request.master == VHOST_USER_SET_VRING_CALL) {
2060                 if (vdpa_dev->ops->dev_conf)
2061                         vdpa_dev->ops->dev_conf(dev->vid);
2062                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2063         }
2064
2065         return 0;
2066 }
2067
2068 static int process_slave_message_reply(struct virtio_net *dev,
2069                                        const struct VhostUserMsg *msg)
2070 {
2071         struct VhostUserMsg msg_reply;
2072         int ret;
2073
2074         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2075                 return 0;
2076
2077         if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
2078                 ret = -1;
2079                 goto out;
2080         }
2081
2082         if (msg_reply.request.slave != msg->request.slave) {
2083                 RTE_LOG(ERR, VHOST_CONFIG,
2084                         "Received unexpected msg type (%u), expected %u\n",
2085                         msg_reply.request.slave, msg->request.slave);
2086                 ret = -1;
2087                 goto out;
2088         }
2089
2090         ret = msg_reply.payload.u64 ? -1 : 0;
2091
2092 out:
2093         rte_spinlock_unlock(&dev->slave_req_lock);
2094         return ret;
2095 }
2096
2097 int
2098 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2099 {
2100         int ret;
2101         struct VhostUserMsg msg = {
2102                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2103                 .flags = VHOST_USER_VERSION,
2104                 .size = sizeof(msg.payload.iotlb),
2105                 .payload.iotlb = {
2106                         .iova = iova,
2107                         .perm = perm,
2108                         .type = VHOST_IOTLB_MISS,
2109                 },
2110         };
2111
2112         ret = send_vhost_message(dev->slave_req_fd, &msg);
2113         if (ret < 0) {
2114                 RTE_LOG(ERR, VHOST_CONFIG,
2115                                 "Failed to send IOTLB miss message (%d)\n",
2116                                 ret);
2117                 return ret;
2118         }
2119
2120         return 0;
2121 }
2122
2123 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2124                                                     int index, int fd,
2125                                                     uint64_t offset,
2126                                                     uint64_t size)
2127 {
2128         int ret;
2129         struct VhostUserMsg msg = {
2130                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2131                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2132                 .size = sizeof(msg.payload.area),
2133                 .payload.area = {
2134                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
2135                         .size = size,
2136                         .offset = offset,
2137                 },
2138         };
2139
2140         if (fd < 0)
2141                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2142         else {
2143                 msg.fds[0] = fd;
2144                 msg.fd_num = 1;
2145         }
2146
2147         ret = send_vhost_slave_message(dev, &msg);
2148         if (ret < 0) {
2149                 RTE_LOG(ERR, VHOST_CONFIG,
2150                         "Failed to set host notifier (%d)\n", ret);
2151                 return ret;
2152         }
2153
2154         return process_slave_message_reply(dev, &msg);
2155 }
2156
2157 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2158 {
2159         struct virtio_net *dev;
2160         struct rte_vdpa_device *vdpa_dev;
2161         int vfio_device_fd, did, ret = 0;
2162         uint64_t offset, size;
2163         unsigned int i;
2164
2165         dev = get_device(vid);
2166         if (!dev)
2167                 return -ENODEV;
2168
2169         did = dev->vdpa_dev_id;
2170         if (did < 0)
2171                 return -EINVAL;
2172
2173         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2174             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2175             !(dev->protocol_features &
2176                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2177             !(dev->protocol_features &
2178                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2179             !(dev->protocol_features &
2180                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2181                 return -ENOTSUP;
2182
2183         vdpa_dev = rte_vdpa_get_device(did);
2184         if (!vdpa_dev)
2185                 return -ENODEV;
2186
2187         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2188         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2189
2190         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2191         if (vfio_device_fd < 0)
2192                 return -ENOTSUP;
2193
2194         if (enable) {
2195                 for (i = 0; i < dev->nr_vring; i++) {
2196                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2197                                         &size) < 0) {
2198                                 ret = -ENOTSUP;
2199                                 goto disable;
2200                         }
2201
2202                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2203                                         vfio_device_fd, offset, size) < 0) {
2204                                 ret = -EFAULT;
2205                                 goto disable;
2206                         }
2207                 }
2208         } else {
2209 disable:
2210                 for (i = 0; i < dev->nr_vring; i++) {
2211                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
2212                                         0, 0);
2213                 }
2214         }
2215
2216         return ret;
2217 }