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 /*
  97  * Reclaim all the outstanding zmbufs for a virtqueue.
  98  */
  99 static void
 100 drain_zmbuf_list(struct vhost_virtqueue *vq)
 101 {
 102         struct zcopy_mbuf *zmbuf, *next;
 103
 104         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
 105              zmbuf != NULL; zmbuf = next) {
 106                 next = TAILQ_NEXT(zmbuf, next);
 107
 108                 while (!mbuf_is_consumed(zmbuf->mbuf))
 109                         usleep(1000);
 110
 111                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
 112                 restore_mbuf(zmbuf->mbuf);
 113                 rte_pktmbuf_free(zmbuf->mbuf);
 114                 put_zmbuf(zmbuf);
 115                 vq->nr_zmbuf -= 1;
 116         }
 117 }
 118
 119 static void
 120 free_mem_region(struct virtio_net *dev)
 121 {
 122         uint32_t i;
 123         struct rte_vhost_mem_region *reg;
 124         struct vhost_virtqueue *vq;
 125
 126         if (!dev || !dev->mem)
 127                 return;
 128
 129         if (dev->dequeue_zero_copy) {
 130                 for (i = 0; i < dev->nr_vring; i++) {
 131                         vq = dev->virtqueue[i];
 132                         if (vq)
 133                                 drain_zmbuf_list(vq);
 134                 }
 135         }
 136
 137         for (i = 0; i < dev->mem->nregions; i++) {
 138                 reg = &dev->mem->regions[i];
 139                 if (reg->host_user_addr) {
 140                         munmap(reg->mmap_addr, reg->mmap_size);
 141                         close(reg->fd);
 142                 }
 143         }
 144 }
 145
 146 void
 147 vhost_backend_cleanup(struct virtio_net *dev)
 148 {
 149         if (dev->mem) {
 150                 free_mem_region(dev);
 151                 rte_free(dev->mem);
 152                 dev->mem = NULL;
 153         }
 154
 155         free(dev->guest_pages);
 156         dev->guest_pages = NULL;
 157
 158         if (dev->log_addr) {
 159                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 160                 dev->log_addr = 0;
 161         }
 162
 163         if (dev->slave_req_fd >= 0) {
 164                 close(dev->slave_req_fd);
 165                 dev->slave_req_fd = -1;
 166         }
 167
 168         if (dev->postcopy_ufd >= 0) {
 169                 close(dev->postcopy_ufd);
 170                 dev->postcopy_ufd = -1;
 171         }
 172
 173         dev->postcopy_listening = 0;
 174 }
 175
 176 /*
 177  * This function just returns success at the moment unless
 178  * the device hasn't been initialised.
 179  */
 180 static int
 181 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
 182                         struct VhostUserMsg *msg __rte_unused,
 183                         int main_fd __rte_unused)
 184 {
 185         return RTE_VHOST_MSG_RESULT_OK;
 186 }
 187
 188 static int
 189 vhost_user_reset_owner(struct virtio_net **pdev,
 190                         struct VhostUserMsg *msg __rte_unused,
 191                         int main_fd __rte_unused)
 192 {
 193         struct virtio_net *dev = *pdev;
 194         vhost_destroy_device_notify(dev);
 195
 196         cleanup_device(dev, 0);
 197         reset_device(dev);
 198         return RTE_VHOST_MSG_RESULT_OK;
 199 }
 200
 201 /*
 202  * The features that we support are requested.
 203  */
 204 static int
 205 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 206                         int main_fd __rte_unused)
 207 {
 208         struct virtio_net *dev = *pdev;
 209         uint64_t features = 0;
 210
 211         rte_vhost_driver_get_features(dev->ifname, &features);
 212
 213         msg->payload.u64 = features;
 214         msg->size = sizeof(msg->payload.u64);
 215         msg->fd_num = 0;
 216
 217         return RTE_VHOST_MSG_RESULT_REPLY;
 218 }
 219
 220 /*
 221  * The queue number that we support are requested.
 222  */
 223 static int
 224 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
 225                         int main_fd __rte_unused)
 226 {
 227         struct virtio_net *dev = *pdev;
 228         uint32_t queue_num = 0;
 229
 230         rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
 231
 232         msg->payload.u64 = (uint64_t)queue_num;
 233         msg->size = sizeof(msg->payload.u64);
 234         msg->fd_num = 0;
 235
 236         return RTE_VHOST_MSG_RESULT_REPLY;
 237 }
 238
 239 /*
 240  * We receive the negotiated features supported by us and the virtio device.
 241  */
 242 static int
 243 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 244                         int main_fd __rte_unused)
 245 {
 246         struct virtio_net *dev = *pdev;
 247         uint64_t features = msg->payload.u64;
 248         uint64_t vhost_features = 0;
 249         struct rte_vdpa_device *vdpa_dev;
 250         int did = -1;
 251
 252         rte_vhost_driver_get_features(dev->ifname, &vhost_features);
 253         if (features & ~vhost_features) {
 254                 RTE_LOG(ERR, VHOST_CONFIG,
 255                         "(%d) received invalid negotiated features.\n",
 256                         dev->vid);
 257                 return RTE_VHOST_MSG_RESULT_ERR;
 258         }
 259
 260         if (dev->flags & VIRTIO_DEV_RUNNING) {
 261                 if (dev->features == features)
 262                         return RTE_VHOST_MSG_RESULT_OK;
 263
 264                 /*
 265                  * Error out if master tries to change features while device is
 266                  * in running state. The exception being VHOST_F_LOG_ALL, which
 267                  * is enabled when the live-migration starts.
 268                  */
 269                 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
 270                         RTE_LOG(ERR, VHOST_CONFIG,
 271                                 "(%d) features changed while device is running.\n",
 272                                 dev->vid);
 273                         return RTE_VHOST_MSG_RESULT_ERR;
 274                 }
 275
 276                 if (dev->notify_ops->features_changed)
 277                         dev->notify_ops->features_changed(dev->vid, features);
 278         }
 279
 280         dev->features = features;
 281         if (dev->features &
 282                 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
 283                 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 284         } else {
 285                 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
 286         }
 287         RTE_LOG(INFO, VHOST_CONFIG,
 288                 "negotiated Virtio features: 0x%" PRIx64 "\n", dev->features);
 289         VHOST_LOG_DEBUG(VHOST_CONFIG,
 290                 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
 291                 dev->vid,
 292                 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
 293                 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 294
 295         if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
 296             !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
 297                 /*
 298                  * Remove all but first queue pair if MQ hasn't been
 299                  * negotiated. This is safe because the device is not
 300                  * running at this stage.
 301                  */
 302                 while (dev->nr_vring > 2) {
 303                         struct vhost_virtqueue *vq;
 304
 305                         vq = dev->virtqueue[--dev->nr_vring];
 306                         if (!vq)
 307                                 continue;
 308
 309                         dev->virtqueue[dev->nr_vring] = NULL;
 310                         cleanup_vq(vq, 1);
 311                         free_vq(dev, vq);
 312                 }
 313         }
 314
 315         did = dev->vdpa_dev_id;
 316         vdpa_dev = rte_vdpa_get_device(did);
 317         if (vdpa_dev && vdpa_dev->ops->set_features)
 318                 vdpa_dev->ops->set_features(dev->vid);
 319
 320         return RTE_VHOST_MSG_RESULT_OK;
 321 }
 322
 323 /*
 324  * The virtio device sends us the size of the descriptor ring.
 325  */
 326 static int
 327 vhost_user_set_vring_num(struct virtio_net **pdev,
 328                         struct VhostUserMsg *msg,
 329                         int main_fd __rte_unused)
 330 {
 331         struct virtio_net *dev = *pdev;
 332         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 333
 334         vq->size = msg->payload.state.num;
 335
 336         /* VIRTIO 1.0, 2.4 Virtqueues says:
 337          *
 338          *   Queue Size value is always a power of 2. The maximum Queue Size
 339          *   value is 32768.
 340          */
 341         if ((vq->size & (vq->size - 1)) || vq->size > 32768) {
 342                 RTE_LOG(ERR, VHOST_CONFIG,
 343                         "invalid virtqueue size %u\n", vq->size);
 344                 return RTE_VHOST_MSG_RESULT_ERR;
 345         }
 346
 347         if (dev->dequeue_zero_copy) {
 348                 vq->nr_zmbuf = 0;
 349                 vq->last_zmbuf_idx = 0;
 350                 vq->zmbuf_size = vq->size;
 351                 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
 352                                          sizeof(struct zcopy_mbuf), 0);
 353                 if (vq->zmbufs == NULL) {
 354                         RTE_LOG(WARNING, VHOST_CONFIG,
 355                                 "failed to allocate mem for zero copy; "
 356                                 "zero copy is force disabled\n");
 357                         dev->dequeue_zero_copy = 0;
 358                 }
 359                 TAILQ_INIT(&vq->zmbuf_list);
 360         }
 361
 362         if (vq_is_packed(dev)) {
 363                 vq->shadow_used_packed = rte_malloc(NULL,
 364                                 vq->size *
 365                                 sizeof(struct vring_used_elem_packed),
 366                                 RTE_CACHE_LINE_SIZE);
 367                 if (!vq->shadow_used_packed) {
 368                         RTE_LOG(ERR, VHOST_CONFIG,
 369                                         "failed to allocate memory for shadow used ring.\n");
 370                         return RTE_VHOST_MSG_RESULT_ERR;
 371                 }
 372
 373         } else {
 374                 vq->shadow_used_split = rte_malloc(NULL,
 375                                 vq->size * sizeof(struct vring_used_elem),
 376                                 RTE_CACHE_LINE_SIZE);
 377                 if (!vq->shadow_used_split) {
 378                         RTE_LOG(ERR, VHOST_CONFIG,
 379                                         "failed to allocate memory for shadow used ring.\n");
 380                         return RTE_VHOST_MSG_RESULT_ERR;
 381                 }
 382         }
 383
 384         vq->batch_copy_elems = rte_malloc(NULL,
 385                                 vq->size * sizeof(struct batch_copy_elem),
 386                                 RTE_CACHE_LINE_SIZE);
 387         if (!vq->batch_copy_elems) {
 388                 RTE_LOG(ERR, VHOST_CONFIG,
 389                         "failed to allocate memory for batching copy.\n");
 390                 return RTE_VHOST_MSG_RESULT_ERR;
 391         }
 392
 393         return RTE_VHOST_MSG_RESULT_OK;
 394 }
 395
 396 /*
 397  * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
 398  * same numa node as the memory of vring descriptor.
 399  */
 400 #ifdef RTE_LIBRTE_VHOST_NUMA
 401 static struct virtio_net*
 402 numa_realloc(struct virtio_net *dev, int index)
 403 {
 404         int oldnode, newnode;
 405         struct virtio_net *old_dev;
 406         struct vhost_virtqueue *old_vq, *vq;
 407         struct zcopy_mbuf *new_zmbuf;
 408         struct vring_used_elem *new_shadow_used_split;
 409         struct vring_used_elem_packed *new_shadow_used_packed;
 410         struct batch_copy_elem *new_batch_copy_elems;
 411         int ret;
 412
 413         if (dev->flags & VIRTIO_DEV_RUNNING)
 414                 return dev;
 415
 416         old_dev = dev;
 417         vq = old_vq = dev->virtqueue[index];
 418
 419         ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
 420                             MPOL_F_NODE | MPOL_F_ADDR);
 421
 422         /* check if we need to reallocate vq */
 423         ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
 424                              MPOL_F_NODE | MPOL_F_ADDR);
 425         if (ret) {
 426                 RTE_LOG(ERR, VHOST_CONFIG,
 427                         "Unable to get vq numa information.\n");
 428                 return dev;
 429         }
 430         if (oldnode != newnode) {
 431                 RTE_LOG(INFO, VHOST_CONFIG,
 432                         "reallocate vq from %d to %d node\n", oldnode, newnode);
 433                 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
 434                 if (!vq)
 435                         return dev;
 436
 437                 memcpy(vq, old_vq, sizeof(*vq));
 438                 TAILQ_INIT(&vq->zmbuf_list);
 439
 440                 if (dev->dequeue_zero_copy) {
 441                         new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
 442                                         sizeof(struct zcopy_mbuf), 0, newnode);
 443                         if (new_zmbuf) {
 444                                 rte_free(vq->zmbufs);
 445                                 vq->zmbufs = new_zmbuf;
 446                         }
 447                 }
 448
 449                 if (vq_is_packed(dev)) {
 450                         new_shadow_used_packed = rte_malloc_socket(NULL,
 451                                         vq->size *
 452                                         sizeof(struct vring_used_elem_packed),
 453                                         RTE_CACHE_LINE_SIZE,
 454                                         newnode);
 455                         if (new_shadow_used_packed) {
 456                                 rte_free(vq->shadow_used_packed);
 457                                 vq->shadow_used_packed = new_shadow_used_packed;
 458                         }
 459                 } else {
 460                         new_shadow_used_split = rte_malloc_socket(NULL,
 461                                         vq->size *
 462                                         sizeof(struct vring_used_elem),
 463                                         RTE_CACHE_LINE_SIZE,
 464                                         newnode);
 465                         if (new_shadow_used_split) {
 466                                 rte_free(vq->shadow_used_split);
 467                                 vq->shadow_used_split = new_shadow_used_split;
 468                         }
 469                 }
 470
 471                 new_batch_copy_elems = rte_malloc_socket(NULL,
 472                         vq->size * sizeof(struct batch_copy_elem),
 473                         RTE_CACHE_LINE_SIZE,
 474                         newnode);
 475                 if (new_batch_copy_elems) {
 476                         rte_free(vq->batch_copy_elems);
 477                         vq->batch_copy_elems = new_batch_copy_elems;
 478                 }
 479
 480                 rte_free(old_vq);
 481         }
 482
 483         /* check if we need to reallocate dev */
 484         ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
 485                             MPOL_F_NODE | MPOL_F_ADDR);
 486         if (ret) {
 487                 RTE_LOG(ERR, VHOST_CONFIG,
 488                         "Unable to get dev numa information.\n");
 489                 goto out;
 490         }
 491         if (oldnode != newnode) {
 492                 RTE_LOG(INFO, VHOST_CONFIG,
 493                         "reallocate dev from %d to %d node\n",
 494                         oldnode, newnode);
 495                 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
 496                 if (!dev) {
 497                         dev = old_dev;
 498                         goto out;
 499                 }
 500
 501                 memcpy(dev, old_dev, sizeof(*dev));
 502                 rte_free(old_dev);
 503         }
 504
 505 out:
 506         dev->virtqueue[index] = vq;
 507         vhost_devices[dev->vid] = dev;
 508
 509         if (old_vq != vq)
 510                 vhost_user_iotlb_init(dev, index);
 511
 512         return dev;
 513 }
 514 #else
 515 static struct virtio_net*
 516 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 517 {
 518         return dev;
 519 }
 520 #endif
 521
 522 /* Converts QEMU virtual address to Vhost virtual address. */
 523 static uint64_t
 524 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
 525 {
 526         struct rte_vhost_mem_region *r;
 527         uint32_t i;
 528
 529         if (unlikely(!dev || !dev->mem))
 530                 goto out_error;
 531
 532         /* Find the region where the address lives. */
 533         for (i = 0; i < dev->mem->nregions; i++) {
 534                 r = &dev->mem->regions[i];
 535
 536                 if (qva >= r->guest_user_addr &&
 537                     qva <  r->guest_user_addr + r->size) {
 538
 539                         if (unlikely(*len > r->guest_user_addr + r->size - qva))
 540                                 *len = r->guest_user_addr + r->size - qva;
 541
 542                         return qva - r->guest_user_addr +
 543                                r->host_user_addr;
 544                 }
 545         }
 546 out_error:
 547         *len = 0;
 548
 549         return 0;
 550 }
 551
 552
 553 /*
 554  * Converts ring address to Vhost virtual address.
 555  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 556  * else it is a QEMU virtual address.
 557  */
 558 static uint64_t
 559 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 560                 uint64_t ra, uint64_t *size)
 561 {
 562         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 563                 uint64_t vva;
 564                 uint64_t req_size = *size;
 565
 566                 vva = vhost_user_iotlb_cache_find(vq, ra,
 567                                         size, VHOST_ACCESS_RW);
 568                 if (req_size != *size)
 569                         vhost_user_iotlb_miss(dev, (ra + *size),
 570                                               VHOST_ACCESS_RW);
 571
 572                 return vva;
 573         }
 574
 575         return qva_to_vva(dev, ra, size);
 576 }
 577
 578 /*
 579  * Converts vring log address to GPA
 580  * If IOMMU is enabled, the log address is IOVA
 581  * If IOMMU not enabled, the log address is already GPA
 582  */
 583 static uint64_t
 584 translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
 585                 uint64_t log_addr)
 586 {
 587         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 588                 const uint64_t exp_size = sizeof(struct vring_used) +
 589                         sizeof(struct vring_used_elem) * vq->size;
 590                 uint64_t hva, gpa;
 591                 uint64_t size = exp_size;
 592
 593                 hva = vhost_iova_to_vva(dev, vq, log_addr,
 594                                         &size, VHOST_ACCESS_RW);
 595                 if (size != exp_size)
 596                         return 0;
 597
 598                 gpa = hva_to_gpa(dev, hva, exp_size);
 599                 if (!gpa) {
 600                         RTE_LOG(ERR, VHOST_CONFIG,
 601                                 "VQ: Failed to find GPA for log_addr: 0x%" PRIx64 " hva: 0x%" PRIx64 "\n",
 602                                 log_addr, hva);
 603                         return 0;
 604                 }
 605                 return gpa;
 606
 607         } else
 608                 return log_addr;
 609 }
 610
 611 static struct virtio_net *
 612 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 613 {
 614         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 615         struct vhost_vring_addr *addr = &vq->ring_addrs;
 616         uint64_t len, expected_len;
 617
 618         if (vq_is_packed(dev)) {
 619                 len = sizeof(struct vring_packed_desc) * vq->size;
 620                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 621                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 622                 vq->log_guest_addr = 0;
 623                 if (vq->desc_packed == NULL ||
 624                                 len != sizeof(struct vring_packed_desc) *
 625                                 vq->size) {
 626                         RTE_LOG(DEBUG, VHOST_CONFIG,
 627                                 "(%d) failed to map desc_packed ring.\n",
 628                                 dev->vid);
 629                         return dev;
 630                 }
 631
 632                 dev = numa_realloc(dev, vq_index);
 633                 vq = dev->virtqueue[vq_index];
 634                 addr = &vq->ring_addrs;
 635
 636                 len = sizeof(struct vring_packed_desc_event);
 637                 vq->driver_event = (struct vring_packed_desc_event *)
 638                                         (uintptr_t)ring_addr_to_vva(dev,
 639                                         vq, addr->avail_user_addr, &len);
 640                 if (vq->driver_event == NULL ||
 641                                 len != sizeof(struct vring_packed_desc_event)) {
 642                         RTE_LOG(DEBUG, VHOST_CONFIG,
 643                                 "(%d) failed to find driver area address.\n",
 644                                 dev->vid);
 645                         return dev;
 646                 }
 647
 648                 len = sizeof(struct vring_packed_desc_event);
 649                 vq->device_event = (struct vring_packed_desc_event *)
 650                                         (uintptr_t)ring_addr_to_vva(dev,
 651                                         vq, addr->used_user_addr, &len);
 652                 if (vq->device_event == NULL ||
 653                                 len != sizeof(struct vring_packed_desc_event)) {
 654                         RTE_LOG(DEBUG, VHOST_CONFIG,
 655                                 "(%d) failed to find device area address.\n",
 656                                 dev->vid);
 657                         return dev;
 658                 }
 659
 660                 vq->access_ok = 1;
 661                 return dev;
 662         }
 663
 664         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 665         if (vq->desc && vq->avail && vq->used)
 666                 return dev;
 667
 668         len = sizeof(struct vring_desc) * vq->size;
 669         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 670                         vq, addr->desc_user_addr, &len);
 671         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 672                 RTE_LOG(DEBUG, VHOST_CONFIG,
 673                         "(%d) failed to map desc ring.\n",
 674                         dev->vid);
 675                 return dev;
 676         }
 677
 678         dev = numa_realloc(dev, vq_index);
 679         vq = dev->virtqueue[vq_index];
 680         addr = &vq->ring_addrs;
 681
 682         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 683         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 684                 len += sizeof(uint16_t);
 685         expected_len = len;
 686         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 687                         vq, addr->avail_user_addr, &len);
 688         if (vq->avail == 0 || len != expected_len) {
 689                 RTE_LOG(DEBUG, VHOST_CONFIG,
 690                         "(%d) failed to map avail ring.\n",
 691                         dev->vid);
 692                 return dev;
 693         }
 694
 695         len = sizeof(struct vring_used) +
 696                 sizeof(struct vring_used_elem) * vq->size;
 697         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 698                 len += sizeof(uint16_t);
 699         expected_len = len;
 700         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 701                         vq, addr->used_user_addr, &len);
 702         if (vq->used == 0 || len != expected_len) {
 703                 RTE_LOG(DEBUG, VHOST_CONFIG,
 704                         "(%d) failed to map used ring.\n",
 705                         dev->vid);
 706                 return dev;
 707         }
 708
 709         if (vq->last_used_idx != vq->used->idx) {
 710                 RTE_LOG(WARNING, VHOST_CONFIG,
 711                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 712                         "some packets maybe resent for Tx and dropped for Rx\n",
 713                         vq->last_used_idx, vq->used->idx);
 714                 vq->last_used_idx  = vq->used->idx;
 715                 vq->last_avail_idx = vq->used->idx;
 716         }
 717
 718         vq->log_guest_addr =
 719                 translate_log_addr(dev, vq, addr->log_guest_addr);
 720         if (vq->log_guest_addr == 0) {
 721                 RTE_LOG(DEBUG, VHOST_CONFIG,
 722                         "(%d) failed to map log_guest_addr .\n",
 723                         dev->vid);
 724                 return dev;
 725         }
 726         vq->access_ok = 1;
 727
 728         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
 729                         dev->vid, vq->desc);
 730         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
 731                         dev->vid, vq->avail);
 732         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
 733                         dev->vid, vq->used);
 734         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
 735                         dev->vid, vq->log_guest_addr);
 736
 737         return dev;
 738 }
 739
 740 /*
 741  * The virtio device sends us the desc, used and avail ring addresses.
 742  * This function then converts these to our address space.
 743  */
 744 static int
 745 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 746                         int main_fd __rte_unused)
 747 {
 748         struct virtio_net *dev = *pdev;
 749         struct vhost_virtqueue *vq;
 750         struct vhost_vring_addr *addr = &msg->payload.addr;
 751         bool access_ok;
 752
 753         if (dev->mem == NULL)
 754                 return RTE_VHOST_MSG_RESULT_ERR;
 755
 756         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 757         vq = dev->virtqueue[msg->payload.addr.index];
 758
 759         access_ok = vq->access_ok;
 760
 761         /*
 762          * Rings addresses should not be interpreted as long as the ring is not
 763          * started and enabled
 764          */
 765         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 766
 767         vring_invalidate(dev, vq);
 768
 769         if ((vq->enabled && (dev->features &
 770                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) ||
 771                         access_ok) {
 772                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 773                 if (!dev)
 774                         return RTE_VHOST_MSG_RESULT_ERR;
 775
 776                 *pdev = dev;
 777         }
 778
 779         return RTE_VHOST_MSG_RESULT_OK;
 780 }
 781
 782 /*
 783  * The virtio device sends us the available ring last used index.
 784  */
 785 static int
 786 vhost_user_set_vring_base(struct virtio_net **pdev,
 787                         struct VhostUserMsg *msg,
 788                         int main_fd __rte_unused)
 789 {
 790         struct virtio_net *dev = *pdev;
 791         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 792         uint64_t val = msg->payload.state.num;
 793
 794         if (vq_is_packed(dev)) {
 795                 /*
 796                  * Bit[0:14]: avail index
 797                  * Bit[15]: avail wrap counter
 798                  */
 799                 vq->last_avail_idx = val & 0x7fff;
 800                 vq->avail_wrap_counter = !!(val & (0x1 << 15));
 801                 /*
 802                  * Set used index to same value as available one, as
 803                  * their values should be the same since ring processing
 804                  * was stopped at get time.
 805                  */
 806                 vq->last_used_idx = vq->last_avail_idx;
 807                 vq->used_wrap_counter = vq->avail_wrap_counter;
 808         } else {
 809                 vq->last_used_idx = msg->payload.state.num;
 810                 vq->last_avail_idx = msg->payload.state.num;
 811         }
 812
 813         return RTE_VHOST_MSG_RESULT_OK;
 814 }
 815
 816 static int
 817 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 818                    uint64_t host_phys_addr, uint64_t size)
 819 {
 820         struct guest_page *page, *last_page;
 821         struct guest_page *old_pages;
 822
 823         if (dev->nr_guest_pages == dev->max_guest_pages) {
 824                 dev->max_guest_pages *= 2;
 825                 old_pages = dev->guest_pages;
 826                 dev->guest_pages = realloc(dev->guest_pages,
 827                                         dev->max_guest_pages * sizeof(*page));
 828                 if (!dev->guest_pages) {
 829                         RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
 830                         free(old_pages);
 831                         return -1;
 832                 }
 833         }
 834
 835         if (dev->nr_guest_pages > 0) {
 836                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 837                 /* merge if the two pages are continuous */
 838                 if (host_phys_addr == last_page->host_phys_addr +
 839                                       last_page->size) {
 840                         last_page->size += size;
 841                         return 0;
 842                 }
 843         }
 844
 845         page = &dev->guest_pages[dev->nr_guest_pages++];
 846         page->guest_phys_addr = guest_phys_addr;
 847         page->host_phys_addr  = host_phys_addr;
 848         page->size = size;
 849
 850         return 0;
 851 }
 852
 853 static int
 854 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 855                 uint64_t page_size)
 856 {
 857         uint64_t reg_size = reg->size;
 858         uint64_t host_user_addr  = reg->host_user_addr;
 859         uint64_t guest_phys_addr = reg->guest_phys_addr;
 860         uint64_t host_phys_addr;
 861         uint64_t size;
 862
 863         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 864         size = page_size - (guest_phys_addr & (page_size - 1));
 865         size = RTE_MIN(size, reg_size);
 866
 867         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 868                 return -1;
 869
 870         host_user_addr  += size;
 871         guest_phys_addr += size;
 872         reg_size -= size;
 873
 874         while (reg_size > 0) {
 875                 size = RTE_MIN(reg_size, page_size);
 876                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 877                                                   host_user_addr);
 878                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 879                                 size) < 0)
 880                         return -1;
 881
 882                 host_user_addr  += size;
 883                 guest_phys_addr += size;
 884                 reg_size -= size;
 885         }
 886
 887         return 0;
 888 }
 889
 890 #ifdef RTE_LIBRTE_VHOST_DEBUG
 891 /* TODO: enable it only in debug mode? */
 892 static void
 893 dump_guest_pages(struct virtio_net *dev)
 894 {
 895         uint32_t i;
 896         struct guest_page *page;
 897
 898         for (i = 0; i < dev->nr_guest_pages; i++) {
 899                 page = &dev->guest_pages[i];
 900
 901                 RTE_LOG(INFO, VHOST_CONFIG,
 902                         "guest physical page region %u\n"
 903                         "\t guest_phys_addr: %" PRIx64 "\n"
 904                         "\t host_phys_addr : %" PRIx64 "\n"
 905                         "\t size           : %" PRIx64 "\n",
 906                         i,
 907                         page->guest_phys_addr,
 908                         page->host_phys_addr,
 909                         page->size);
 910         }
 911 }
 912 #else
 913 #define dump_guest_pages(dev)
 914 #endif
 915
 916 static bool
 917 vhost_memory_changed(struct VhostUserMemory *new,
 918                      struct rte_vhost_memory *old)
 919 {
 920         uint32_t i;
 921
 922         if (new->nregions != old->nregions)
 923                 return true;
 924
 925         for (i = 0; i < new->nregions; ++i) {
 926                 VhostUserMemoryRegion *new_r = &new->regions[i];
 927                 struct rte_vhost_mem_region *old_r = &old->regions[i];
 928
 929                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
 930                         return true;
 931                 if (new_r->memory_size != old_r->size)
 932                         return true;
 933                 if (new_r->userspace_addr != old_r->guest_user_addr)
 934                         return true;
 935         }
 936
 937         return false;
 938 }
 939
 940 static int
 941 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
 942                         int main_fd)
 943 {
 944         struct virtio_net *dev = *pdev;
 945         struct VhostUserMemory *memory = &msg->payload.memory;
 946         struct rte_vhost_mem_region *reg;
 947         void *mmap_addr;
 948         uint64_t mmap_size;
 949         uint64_t mmap_offset;
 950         uint64_t alignment;
 951         uint32_t i;
 952         int populate;
 953         int fd;
 954
 955         if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
 956                 RTE_LOG(ERR, VHOST_CONFIG,
 957                         "too many memory regions (%u)\n", memory->nregions);
 958                 return RTE_VHOST_MSG_RESULT_ERR;
 959         }
 960
 961         if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
 962                 RTE_LOG(INFO, VHOST_CONFIG,
 963                         "(%d) memory regions not changed\n", dev->vid);
 964
 965                 for (i = 0; i < memory->nregions; i++)
 966                         close(msg->fds[i]);
 967
 968                 return RTE_VHOST_MSG_RESULT_OK;
 969         }
 970
 971         if (dev->mem) {
 972                 free_mem_region(dev);
 973                 rte_free(dev->mem);
 974                 dev->mem = NULL;
 975         }
 976
 977         /* Flush IOTLB cache as previous HVAs are now invalid */
 978         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 979                 for (i = 0; i < dev->nr_vring; i++)
 980                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
 981
 982         dev->nr_guest_pages = 0;
 983         if (!dev->guest_pages) {
 984                 dev->max_guest_pages = 8;
 985                 dev->guest_pages = malloc(dev->max_guest_pages *
 986                                                 sizeof(struct guest_page));
 987                 if (dev->guest_pages == NULL) {
 988                         RTE_LOG(ERR, VHOST_CONFIG,
 989                                 "(%d) failed to allocate memory "
 990                                 "for dev->guest_pages\n",
 991                                 dev->vid);
 992                         return RTE_VHOST_MSG_RESULT_ERR;
 993                 }
 994         }
 995
 996         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
 997                 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
 998         if (dev->mem == NULL) {
 999                 RTE_LOG(ERR, VHOST_CONFIG,
1000                         "(%d) failed to allocate memory for dev->mem\n",
1001                         dev->vid);
1002                 return RTE_VHOST_MSG_RESULT_ERR;
1003         }
1004         dev->mem->nregions = memory->nregions;
1005
1006         for (i = 0; i < memory->nregions; i++) {
1007                 fd  = msg->fds[i];
1008                 reg = &dev->mem->regions[i];
1009
1010                 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
1011                 reg->guest_user_addr = memory->regions[i].userspace_addr;
1012                 reg->size            = memory->regions[i].memory_size;
1013                 reg->fd              = fd;
1014
1015                 mmap_offset = memory->regions[i].mmap_offset;
1016
1017                 /* Check for memory_size + mmap_offset overflow */
1018                 if (mmap_offset >= -reg->size) {
1019                         RTE_LOG(ERR, VHOST_CONFIG,
1020                                 "mmap_offset (%#"PRIx64") and memory_size "
1021                                 "(%#"PRIx64") overflow\n",
1022                                 mmap_offset, reg->size);
1023                         goto err_mmap;
1024                 }
1025
1026                 mmap_size = reg->size + mmap_offset;
1027
1028                 /* mmap() without flag of MAP_ANONYMOUS, should be called
1029                  * with length argument aligned with hugepagesz at older
1030                  * longterm version Linux, like 2.6.32 and 3.2.72, or
1031                  * mmap() will fail with EINVAL.
1032                  *
1033                  * to avoid failure, make sure in caller to keep length
1034                  * aligned.
1035                  */
1036                 alignment = get_blk_size(fd);
1037                 if (alignment == (uint64_t)-1) {
1038                         RTE_LOG(ERR, VHOST_CONFIG,
1039                                 "couldn't get hugepage size through fstat\n");
1040                         goto err_mmap;
1041                 }
1042                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
1043
1044                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
1045                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
1046                                  MAP_SHARED | populate, fd, 0);
1047
1048                 if (mmap_addr == MAP_FAILED) {
1049                         RTE_LOG(ERR, VHOST_CONFIG,
1050                                 "mmap region %u failed.\n", i);
1051                         goto err_mmap;
1052                 }
1053
1054                 reg->mmap_addr = mmap_addr;
1055                 reg->mmap_size = mmap_size;
1056                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
1057                                       mmap_offset;
1058
1059                 if (dev->dequeue_zero_copy)
1060                         if (add_guest_pages(dev, reg, alignment) < 0) {
1061                                 RTE_LOG(ERR, VHOST_CONFIG,
1062                                         "adding guest pages to region %u failed.\n",
1063                                         i);
1064                                 goto err_mmap;
1065                         }
1066
1067                 RTE_LOG(INFO, VHOST_CONFIG,
1068                         "guest memory region %u, size: 0x%" PRIx64 "\n"
1069                         "\t guest physical addr: 0x%" PRIx64 "\n"
1070                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
1071                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
1072                         "\t mmap addr : 0x%" PRIx64 "\n"
1073                         "\t mmap size : 0x%" PRIx64 "\n"
1074                         "\t mmap align: 0x%" PRIx64 "\n"
1075                         "\t mmap off  : 0x%" PRIx64 "\n",
1076                         i, reg->size,
1077                         reg->guest_phys_addr,
1078                         reg->guest_user_addr,
1079                         reg->host_user_addr,
1080                         (uint64_t)(uintptr_t)mmap_addr,
1081                         mmap_size,
1082                         alignment,
1083                         mmap_offset);
1084
1085                 if (dev->postcopy_listening) {
1086                         /*
1087                          * We haven't a better way right now than sharing
1088                          * DPDK's virtual address with Qemu, so that Qemu can
1089                          * retrieve the region offset when handling userfaults.
1090                          */
1091                         memory->regions[i].userspace_addr =
1092                                 reg->host_user_addr;
1093                 }
1094         }
1095         if (dev->postcopy_listening) {
1096                 /* Send the addresses back to qemu */
1097                 msg->fd_num = 0;
1098                 send_vhost_reply(main_fd, msg);
1099
1100                 /* Wait for qemu to acknolwedge it's got the addresses
1101                  * we've got to wait before we're allowed to generate faults.
1102                  */
1103                 VhostUserMsg ack_msg;
1104                 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1105                         RTE_LOG(ERR, VHOST_CONFIG,
1106                                 "Failed to read qemu ack on postcopy set-mem-table\n");
1107                         goto err_mmap;
1108                 }
1109                 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1110                         RTE_LOG(ERR, VHOST_CONFIG,
1111                                 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1112                                 ack_msg.request.master);
1113                         goto err_mmap;
1114                 }
1115
1116                 /* Now userfault register and we can use the memory */
1117                 for (i = 0; i < memory->nregions; i++) {
1118 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1119                         reg = &dev->mem->regions[i];
1120                         struct uffdio_register reg_struct;
1121
1122                         /*
1123                          * Let's register all the mmap'ed area to ensure
1124                          * alignment on page boundary.
1125                          */
1126                         reg_struct.range.start =
1127                                 (uint64_t)(uintptr_t)reg->mmap_addr;
1128                         reg_struct.range.len = reg->mmap_size;
1129                         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1130
1131                         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1132                                                 &reg_struct)) {
1133                                 RTE_LOG(ERR, VHOST_CONFIG,
1134                                         "Failed to register ufd for region %d: (ufd = %d) %s\n",
1135                                         i, dev->postcopy_ufd,
1136                                         strerror(errno));
1137                                 goto err_mmap;
1138                         }
1139                         RTE_LOG(INFO, VHOST_CONFIG,
1140                                 "\t userfaultfd registered for range : %llx - %llx\n",
1141                                 reg_struct.range.start,
1142                                 reg_struct.range.start +
1143                                 reg_struct.range.len - 1);
1144 #else
1145                         goto err_mmap;
1146 #endif
1147                 }
1148         }
1149
1150         for (i = 0; i < dev->nr_vring; i++) {
1151                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1152
1153                 if (vq->desc || vq->avail || vq->used) {
1154                         /*
1155                          * If the memory table got updated, the ring addresses
1156                          * need to be translated again as virtual addresses have
1157                          * changed.
1158                          */
1159                         vring_invalidate(dev, vq);
1160
1161                         dev = translate_ring_addresses(dev, i);
1162                         if (!dev) {
1163                                 dev = *pdev;
1164                                 goto err_mmap;
1165                         }
1166
1167                         *pdev = dev;
1168                 }
1169         }
1170
1171         dump_guest_pages(dev);
1172
1173         return RTE_VHOST_MSG_RESULT_OK;
1174
1175 err_mmap:
1176         free_mem_region(dev);
1177         rte_free(dev->mem);
1178         dev->mem = NULL;
1179         return RTE_VHOST_MSG_RESULT_ERR;
1180 }
1181
1182 static bool
1183 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1184 {
1185         bool rings_ok;
1186
1187         if (!vq)
1188                 return false;
1189
1190         if (vq_is_packed(dev))
1191                 rings_ok = !!vq->desc_packed;
1192         else
1193                 rings_ok = vq->desc && vq->avail && vq->used;
1194
1195         return rings_ok &&
1196                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1197                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1198 }
1199
1200 static int
1201 virtio_is_ready(struct virtio_net *dev)
1202 {
1203         struct vhost_virtqueue *vq;
1204         uint32_t i;
1205
1206         if (dev->nr_vring == 0)
1207                 return 0;
1208
1209         for (i = 0; i < dev->nr_vring; i++) {
1210                 vq = dev->virtqueue[i];
1211
1212                 if (!vq_is_ready(dev, vq))
1213                         return 0;
1214         }
1215
1216         RTE_LOG(INFO, VHOST_CONFIG,
1217                 "virtio is now ready for processing.\n");
1218         return 1;
1219 }
1220
1221 static int
1222 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1223                         int main_fd __rte_unused)
1224 {
1225         struct virtio_net *dev = *pdev;
1226         struct vhost_vring_file file;
1227         struct vhost_virtqueue *vq;
1228
1229         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1230         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1231                 file.fd = VIRTIO_INVALID_EVENTFD;
1232         else
1233                 file.fd = msg->fds[0];
1234         RTE_LOG(INFO, VHOST_CONFIG,
1235                 "vring call idx:%d file:%d\n", file.index, file.fd);
1236
1237         vq = dev->virtqueue[file.index];
1238         if (vq->callfd >= 0)
1239                 close(vq->callfd);
1240
1241         vq->callfd = file.fd;
1242
1243         return RTE_VHOST_MSG_RESULT_OK;
1244 }
1245
1246 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1247                         struct VhostUserMsg *msg,
1248                         int main_fd __rte_unused)
1249 {
1250         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1251                 close(msg->fds[0]);
1252         RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1253
1254         return RTE_VHOST_MSG_RESULT_OK;
1255 }
1256
1257 static int
1258 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1259                         int main_fd __rte_unused)
1260 {
1261         struct virtio_net *dev = *pdev;
1262         struct vhost_vring_file file;
1263         struct vhost_virtqueue *vq;
1264
1265         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1266         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1267                 file.fd = VIRTIO_INVALID_EVENTFD;
1268         else
1269                 file.fd = msg->fds[0];
1270         RTE_LOG(INFO, VHOST_CONFIG,
1271                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1272
1273         /* Interpret ring addresses only when ring is started. */
1274         dev = translate_ring_addresses(dev, file.index);
1275         if (!dev)
1276                 return RTE_VHOST_MSG_RESULT_ERR;
1277
1278         *pdev = dev;
1279
1280         vq = dev->virtqueue[file.index];
1281
1282         /*
1283          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1284          * the ring starts already enabled. Otherwise, it is enabled via
1285          * the SET_VRING_ENABLE message.
1286          */
1287         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
1288                 vq->enabled = 1;
1289                 if (dev->notify_ops->vring_state_changed)
1290                         dev->notify_ops->vring_state_changed(
1291                                 dev->vid, file.index, 1);
1292         }
1293
1294         if (vq->kickfd >= 0)
1295                 close(vq->kickfd);
1296         vq->kickfd = file.fd;
1297
1298         return RTE_VHOST_MSG_RESULT_OK;
1299 }
1300
1301 static void
1302 free_zmbufs(struct vhost_virtqueue *vq)
1303 {
1304         drain_zmbuf_list(vq);
1305
1306         rte_free(vq->zmbufs);
1307 }
1308
1309 /*
1310  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1311  */
1312 static int
1313 vhost_user_get_vring_base(struct virtio_net **pdev,
1314                         struct VhostUserMsg *msg,
1315                         int main_fd __rte_unused)
1316 {
1317         struct virtio_net *dev = *pdev;
1318         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1319         uint64_t val;
1320
1321         /* We have to stop the queue (virtio) if it is running. */
1322         vhost_destroy_device_notify(dev);
1323
1324         dev->flags &= ~VIRTIO_DEV_READY;
1325         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1326
1327         /* Here we are safe to get the indexes */
1328         if (vq_is_packed(dev)) {
1329                 /*
1330                  * Bit[0:14]: avail index
1331                  * Bit[15]: avail wrap counter
1332                  */
1333                 val = vq->last_avail_idx & 0x7fff;
1334                 val |= vq->avail_wrap_counter << 15;
1335                 msg->payload.state.num = val;
1336         } else {
1337                 msg->payload.state.num = vq->last_avail_idx;
1338         }
1339
1340         RTE_LOG(INFO, VHOST_CONFIG,
1341                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1342                 msg->payload.state.num);
1343         /*
1344          * Based on current qemu vhost-user implementation, this message is
1345          * sent and only sent in vhost_vring_stop.
1346          * TODO: cleanup the vring, it isn't usable since here.
1347          */
1348         if (vq->kickfd >= 0)
1349                 close(vq->kickfd);
1350
1351         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1352
1353         if (vq->callfd >= 0)
1354                 close(vq->callfd);
1355
1356         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1357
1358         vq->signalled_used_valid = false;
1359
1360         if (dev->dequeue_zero_copy)
1361                 free_zmbufs(vq);
1362         if (vq_is_packed(dev)) {
1363                 rte_free(vq->shadow_used_packed);
1364                 vq->shadow_used_packed = NULL;
1365         } else {
1366                 rte_free(vq->shadow_used_split);
1367                 vq->shadow_used_split = NULL;
1368         }
1369
1370         rte_free(vq->batch_copy_elems);
1371         vq->batch_copy_elems = NULL;
1372
1373         msg->size = sizeof(msg->payload.state);
1374         msg->fd_num = 0;
1375
1376         vring_invalidate(dev, vq);
1377
1378         return RTE_VHOST_MSG_RESULT_REPLY;
1379 }
1380
1381 /*
1382  * when virtio queues are ready to work, qemu will send us to
1383  * enable the virtio queue pair.
1384  */
1385 static int
1386 vhost_user_set_vring_enable(struct virtio_net **pdev,
1387                         struct VhostUserMsg *msg,
1388                         int main_fd __rte_unused)
1389 {
1390         struct virtio_net *dev = *pdev;
1391         int enable = (int)msg->payload.state.num;
1392         int index = (int)msg->payload.state.index;
1393         struct rte_vdpa_device *vdpa_dev;
1394         int did = -1;
1395
1396         RTE_LOG(INFO, VHOST_CONFIG,
1397                 "set queue enable: %d to qp idx: %d\n",
1398                 enable, index);
1399
1400         did = dev->vdpa_dev_id;
1401         vdpa_dev = rte_vdpa_get_device(did);
1402         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1403                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1404
1405         if (dev->notify_ops->vring_state_changed)
1406                 dev->notify_ops->vring_state_changed(dev->vid,
1407                                 index, enable);
1408
1409         /* On disable, rings have to be stopped being processed. */
1410         if (!enable && dev->dequeue_zero_copy)
1411                 drain_zmbuf_list(dev->virtqueue[index]);
1412
1413         dev->virtqueue[index]->enabled = enable;
1414
1415         return RTE_VHOST_MSG_RESULT_OK;
1416 }
1417
1418 static int
1419 vhost_user_get_protocol_features(struct virtio_net **pdev,
1420                         struct VhostUserMsg *msg,
1421                         int main_fd __rte_unused)
1422 {
1423         struct virtio_net *dev = *pdev;
1424         uint64_t features, protocol_features;
1425
1426         rte_vhost_driver_get_features(dev->ifname, &features);
1427         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1428
1429         /*
1430          * REPLY_ACK protocol feature is only mandatory for now
1431          * for IOMMU feature. If IOMMU is explicitly disabled by the
1432          * application, disable also REPLY_ACK feature for older buggy
1433          * Qemu versions (from v2.7.0 to v2.9.0).
1434          */
1435         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1436                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1437
1438         msg->payload.u64 = protocol_features;
1439         msg->size = sizeof(msg->payload.u64);
1440         msg->fd_num = 0;
1441
1442         return RTE_VHOST_MSG_RESULT_REPLY;
1443 }
1444
1445 static int
1446 vhost_user_set_protocol_features(struct virtio_net **pdev,
1447                         struct VhostUserMsg *msg,
1448                         int main_fd __rte_unused)
1449 {
1450         struct virtio_net *dev = *pdev;
1451         uint64_t protocol_features = msg->payload.u64;
1452         uint64_t slave_protocol_features = 0;
1453
1454         rte_vhost_driver_get_protocol_features(dev->ifname,
1455                         &slave_protocol_features);
1456         if (protocol_features & ~slave_protocol_features) {
1457                 RTE_LOG(ERR, VHOST_CONFIG,
1458                         "(%d) received invalid protocol features.\n",
1459                         dev->vid);
1460                 return RTE_VHOST_MSG_RESULT_ERR;
1461         }
1462
1463         dev->protocol_features = protocol_features;
1464         RTE_LOG(INFO, VHOST_CONFIG,
1465                 "negotiated Vhost-user protocol features: 0x%" PRIx64 "\n",
1466                 dev->protocol_features);
1467
1468         return RTE_VHOST_MSG_RESULT_OK;
1469 }
1470
1471 static int
1472 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1473                         int main_fd __rte_unused)
1474 {
1475         struct virtio_net *dev = *pdev;
1476         int fd = msg->fds[0];
1477         uint64_t size, off;
1478         void *addr;
1479
1480         if (fd < 0) {
1481                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1482                 return RTE_VHOST_MSG_RESULT_ERR;
1483         }
1484
1485         if (msg->size != sizeof(VhostUserLog)) {
1486                 RTE_LOG(ERR, VHOST_CONFIG,
1487                         "invalid log base msg size: %"PRId32" != %d\n",
1488                         msg->size, (int)sizeof(VhostUserLog));
1489                 return RTE_VHOST_MSG_RESULT_ERR;
1490         }
1491
1492         size = msg->payload.log.mmap_size;
1493         off  = msg->payload.log.mmap_offset;
1494
1495         /* Don't allow mmap_offset to point outside the mmap region */
1496         if (off > size) {
1497                 RTE_LOG(ERR, VHOST_CONFIG,
1498                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1499                         off, size);
1500                 return RTE_VHOST_MSG_RESULT_ERR;
1501         }
1502
1503         RTE_LOG(INFO, VHOST_CONFIG,
1504                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1505                 size, off);
1506
1507         /*
1508          * mmap from 0 to workaround a hugepage mmap bug: mmap will
1509          * fail when offset is not page size aligned.
1510          */
1511         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1512         close(fd);
1513         if (addr == MAP_FAILED) {
1514                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1515                 return RTE_VHOST_MSG_RESULT_ERR;
1516         }
1517
1518         /*
1519          * Free previously mapped log memory on occasionally
1520          * multiple VHOST_USER_SET_LOG_BASE.
1521          */
1522         if (dev->log_addr) {
1523                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1524         }
1525         dev->log_addr = (uint64_t)(uintptr_t)addr;
1526         dev->log_base = dev->log_addr + off;
1527         dev->log_size = size;
1528
1529         /*
1530          * The spec is not clear about it (yet), but QEMU doesn't expect
1531          * any payload in the reply.
1532          */
1533         msg->size = 0;
1534         msg->fd_num = 0;
1535
1536         return RTE_VHOST_MSG_RESULT_REPLY;
1537 }
1538
1539 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1540                         struct VhostUserMsg *msg,
1541                         int main_fd __rte_unused)
1542 {
1543         close(msg->fds[0]);
1544         RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1545
1546         return RTE_VHOST_MSG_RESULT_OK;
1547 }
1548
1549 /*
1550  * An rarp packet is constructed and broadcasted to notify switches about
1551  * the new location of the migrated VM, so that packets from outside will
1552  * not be lost after migration.
1553  *
1554  * However, we don't actually "send" a rarp packet here, instead, we set
1555  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1556  */
1557 static int
1558 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
1559                         int main_fd __rte_unused)
1560 {
1561         struct virtio_net *dev = *pdev;
1562         uint8_t *mac = (uint8_t *)&msg->payload.u64;
1563         struct rte_vdpa_device *vdpa_dev;
1564         int did = -1;
1565
1566         RTE_LOG(DEBUG, VHOST_CONFIG,
1567                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1568                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1569         memcpy(dev->mac.addr_bytes, mac, 6);
1570
1571         /*
1572          * Set the flag to inject a RARP broadcast packet at
1573          * rte_vhost_dequeue_burst().
1574          *
1575          * rte_smp_wmb() is for making sure the mac is copied
1576          * before the flag is set.
1577          */
1578         rte_smp_wmb();
1579         rte_atomic16_set(&dev->broadcast_rarp, 1);
1580         did = dev->vdpa_dev_id;
1581         vdpa_dev = rte_vdpa_get_device(did);
1582         if (vdpa_dev && vdpa_dev->ops->migration_done)
1583                 vdpa_dev->ops->migration_done(dev->vid);
1584
1585         return RTE_VHOST_MSG_RESULT_OK;
1586 }
1587
1588 static int
1589 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
1590                         int main_fd __rte_unused)
1591 {
1592         struct virtio_net *dev = *pdev;
1593         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1594                         msg->payload.u64 > VIRTIO_MAX_MTU) {
1595                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1596                                 msg->payload.u64);
1597
1598                 return RTE_VHOST_MSG_RESULT_ERR;
1599         }
1600
1601         dev->mtu = msg->payload.u64;
1602
1603         return RTE_VHOST_MSG_RESULT_OK;
1604 }
1605
1606 static int
1607 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
1608                         int main_fd __rte_unused)
1609 {
1610         struct virtio_net *dev = *pdev;
1611         int fd = msg->fds[0];
1612
1613         if (fd < 0) {
1614                 RTE_LOG(ERR, VHOST_CONFIG,
1615                                 "Invalid file descriptor for slave channel (%d)\n",
1616                                 fd);
1617                 return RTE_VHOST_MSG_RESULT_ERR;
1618         }
1619
1620         if (dev->slave_req_fd >= 0)
1621                 close(dev->slave_req_fd);
1622
1623         dev->slave_req_fd = fd;
1624
1625         return RTE_VHOST_MSG_RESULT_OK;
1626 }
1627
1628 static int
1629 is_vring_iotlb_split(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1630 {
1631         struct vhost_vring_addr *ra;
1632         uint64_t start, end, len;
1633
1634         start = imsg->iova;
1635         end = start + imsg->size;
1636
1637         ra = &vq->ring_addrs;
1638         len = sizeof(struct vring_desc) * vq->size;
1639         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
1640                 return 1;
1641
1642         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
1643         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
1644                 return 1;
1645
1646         len = sizeof(struct vring_used) +
1647                sizeof(struct vring_used_elem) * vq->size;
1648         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
1649                 return 1;
1650
1651         return 0;
1652 }
1653
1654 static int
1655 is_vring_iotlb_packed(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1656 {
1657         struct vhost_vring_addr *ra;
1658         uint64_t start, end, len;
1659
1660         start = imsg->iova;
1661         end = start + imsg->size;
1662
1663         ra = &vq->ring_addrs;
1664         len = sizeof(struct vring_packed_desc) * vq->size;
1665         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
1666                 return 1;
1667
1668         len = sizeof(struct vring_packed_desc_event);
1669         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
1670                 return 1;
1671
1672         len = sizeof(struct vring_packed_desc_event);
1673         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
1674                 return 1;
1675
1676         return 0;
1677 }
1678
1679 static int is_vring_iotlb(struct virtio_net *dev,
1680                           struct vhost_virtqueue *vq,
1681                           struct vhost_iotlb_msg *imsg)
1682 {
1683         if (vq_is_packed(dev))
1684                 return is_vring_iotlb_packed(vq, imsg);
1685         else
1686                 return is_vring_iotlb_split(vq, imsg);
1687 }
1688
1689 static int
1690 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
1691                         int main_fd __rte_unused)
1692 {
1693         struct virtio_net *dev = *pdev;
1694         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1695         uint16_t i;
1696         uint64_t vva, len;
1697
1698         switch (imsg->type) {
1699         case VHOST_IOTLB_UPDATE:
1700                 len = imsg->size;
1701                 vva = qva_to_vva(dev, imsg->uaddr, &len);
1702                 if (!vva)
1703                         return RTE_VHOST_MSG_RESULT_ERR;
1704
1705                 for (i = 0; i < dev->nr_vring; i++) {
1706                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1707
1708                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1709                                         len, imsg->perm);
1710
1711                         if (is_vring_iotlb(dev, vq, imsg))
1712                                 *pdev = dev = translate_ring_addresses(dev, i);
1713                 }
1714                 break;
1715         case VHOST_IOTLB_INVALIDATE:
1716                 for (i = 0; i < dev->nr_vring; i++) {
1717                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1718
1719                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
1720                                         imsg->size);
1721
1722                         if (is_vring_iotlb(dev, vq, imsg))
1723                                 vring_invalidate(dev, vq);
1724                 }
1725                 break;
1726         default:
1727                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1728                                 imsg->type);
1729                 return RTE_VHOST_MSG_RESULT_ERR;
1730         }
1731
1732         return RTE_VHOST_MSG_RESULT_OK;
1733 }
1734
1735 static int
1736 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
1737                         struct VhostUserMsg *msg,
1738                         int main_fd __rte_unused)
1739 {
1740         struct virtio_net *dev = *pdev;
1741 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1742         struct uffdio_api api_struct;
1743
1744         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1745
1746         if (dev->postcopy_ufd == -1) {
1747                 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
1748                         strerror(errno));
1749                 return RTE_VHOST_MSG_RESULT_ERR;
1750         }
1751         api_struct.api = UFFD_API;
1752         api_struct.features = 0;
1753         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
1754                 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
1755                         strerror(errno));
1756                 close(dev->postcopy_ufd);
1757                 dev->postcopy_ufd = -1;
1758                 return RTE_VHOST_MSG_RESULT_ERR;
1759         }
1760         msg->fds[0] = dev->postcopy_ufd;
1761         msg->fd_num = 1;
1762
1763         return RTE_VHOST_MSG_RESULT_REPLY;
1764 #else
1765         dev->postcopy_ufd = -1;
1766         msg->fd_num = 0;
1767
1768         return RTE_VHOST_MSG_RESULT_ERR;
1769 #endif
1770 }
1771
1772 static int
1773 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
1774                         struct VhostUserMsg *msg __rte_unused,
1775                         int main_fd __rte_unused)
1776 {
1777         struct virtio_net *dev = *pdev;
1778
1779         if (dev->mem && dev->mem->nregions) {
1780                 RTE_LOG(ERR, VHOST_CONFIG,
1781                         "Regions already registered at postcopy-listen\n");
1782                 return RTE_VHOST_MSG_RESULT_ERR;
1783         }
1784         dev->postcopy_listening = 1;
1785
1786         return RTE_VHOST_MSG_RESULT_OK;
1787 }
1788
1789 static int
1790 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
1791                         int main_fd __rte_unused)
1792 {
1793         struct virtio_net *dev = *pdev;
1794
1795         dev->postcopy_listening = 0;
1796         if (dev->postcopy_ufd >= 0) {
1797                 close(dev->postcopy_ufd);
1798                 dev->postcopy_ufd = -1;
1799         }
1800
1801         msg->payload.u64 = 0;
1802         msg->size = sizeof(msg->payload.u64);
1803         msg->fd_num = 0;
1804
1805         return RTE_VHOST_MSG_RESULT_REPLY;
1806 }
1807
1808 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
1809                                         struct VhostUserMsg *msg,
1810                                         int main_fd);
1811 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
1812         [VHOST_USER_NONE] = NULL,
1813         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
1814         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
1815         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
1816         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
1817         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
1818         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
1819         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
1820         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
1821         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
1822         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
1823         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
1824         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
1825         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
1826         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
1827         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
1828         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
1829         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
1830         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
1831         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
1832         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
1833         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
1834         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
1835         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
1836         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
1837         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
1838 };
1839
1840
1841 /* return bytes# of read on success or negative val on failure. */
1842 static int
1843 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1844 {
1845         int ret;
1846
1847         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1848                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
1849         if (ret <= 0)
1850                 return ret;
1851
1852         if (msg->size) {
1853                 if (msg->size > sizeof(msg->payload)) {
1854                         RTE_LOG(ERR, VHOST_CONFIG,
1855                                 "invalid msg size: %d\n", msg->size);
1856                         return -1;
1857                 }
1858                 ret = read(sockfd, &msg->payload, msg->size);
1859                 if (ret <= 0)
1860                         return ret;
1861                 if (ret != (int)msg->size) {
1862                         RTE_LOG(ERR, VHOST_CONFIG,
1863                                 "read control message failed\n");
1864                         return -1;
1865                 }
1866         }
1867
1868         return ret;
1869 }
1870
1871 static int
1872 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1873 {
1874         if (!msg)
1875                 return 0;
1876
1877         return send_fd_message(sockfd, (char *)msg,
1878                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
1879 }
1880
1881 static int
1882 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1883 {
1884         if (!msg)
1885                 return 0;
1886
1887         msg->flags &= ~VHOST_USER_VERSION_MASK;
1888         msg->flags &= ~VHOST_USER_NEED_REPLY;
1889         msg->flags |= VHOST_USER_VERSION;
1890         msg->flags |= VHOST_USER_REPLY_MASK;
1891
1892         return send_vhost_message(sockfd, msg);
1893 }
1894
1895 static int
1896 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
1897 {
1898         int ret;
1899
1900         if (msg->flags & VHOST_USER_NEED_REPLY)
1901                 rte_spinlock_lock(&dev->slave_req_lock);
1902
1903         ret = send_vhost_message(dev->slave_req_fd, msg);
1904         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1905                 rte_spinlock_unlock(&dev->slave_req_lock);
1906
1907         return ret;
1908 }
1909
1910 /*
1911  * Allocate a queue pair if it hasn't been allocated yet
1912  */
1913 static int
1914 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1915                         struct VhostUserMsg *msg)
1916 {
1917         uint16_t vring_idx;
1918
1919         switch (msg->request.master) {
1920         case VHOST_USER_SET_VRING_KICK:
1921         case VHOST_USER_SET_VRING_CALL:
1922         case VHOST_USER_SET_VRING_ERR:
1923                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1924                 break;
1925         case VHOST_USER_SET_VRING_NUM:
1926         case VHOST_USER_SET_VRING_BASE:
1927         case VHOST_USER_SET_VRING_ENABLE:
1928                 vring_idx = msg->payload.state.index;
1929                 break;
1930         case VHOST_USER_SET_VRING_ADDR:
1931                 vring_idx = msg->payload.addr.index;
1932                 break;
1933         default:
1934                 return 0;
1935         }
1936
1937         if (vring_idx >= VHOST_MAX_VRING) {
1938                 RTE_LOG(ERR, VHOST_CONFIG,
1939                         "invalid vring index: %u\n", vring_idx);
1940                 return -1;
1941         }
1942
1943         if (dev->virtqueue[vring_idx])
1944                 return 0;
1945
1946         return alloc_vring_queue(dev, vring_idx);
1947 }
1948
1949 static void
1950 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1951 {
1952         unsigned int i = 0;
1953         unsigned int vq_num = 0;
1954
1955         while (vq_num < dev->nr_vring) {
1956                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1957
1958                 if (vq) {
1959                         rte_spinlock_lock(&vq->access_lock);
1960                         vq_num++;
1961                 }
1962                 i++;
1963         }
1964 }
1965
1966 static void
1967 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1968 {
1969         unsigned int i = 0;
1970         unsigned int vq_num = 0;
1971
1972         while (vq_num < dev->nr_vring) {
1973                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1974
1975                 if (vq) {
1976                         rte_spinlock_unlock(&vq->access_lock);
1977                         vq_num++;
1978                 }
1979                 i++;
1980         }
1981 }
1982
1983 int
1984 vhost_user_msg_handler(int vid, int fd)
1985 {
1986         struct virtio_net *dev;
1987         struct VhostUserMsg msg;
1988         struct rte_vdpa_device *vdpa_dev;
1989         int did = -1;
1990         int ret;
1991         int unlock_required = 0;
1992         bool handled;
1993         int request;
1994
1995         dev = get_device(vid);
1996         if (dev == NULL)
1997                 return -1;
1998
1999         if (!dev->notify_ops) {
2000                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
2001                 if (!dev->notify_ops) {
2002                         RTE_LOG(ERR, VHOST_CONFIG,
2003                                 "failed to get callback ops for driver %s\n",
2004                                 dev->ifname);
2005                         return -1;
2006                 }
2007         }
2008
2009         ret = read_vhost_message(fd, &msg);
2010         if (ret <= 0) {
2011                 if (ret < 0)
2012                         RTE_LOG(ERR, VHOST_CONFIG,
2013                                 "vhost read message failed\n");
2014                 else
2015                         RTE_LOG(INFO, VHOST_CONFIG,
2016                                 "vhost peer closed\n");
2017
2018                 return -1;
2019         }
2020
2021         ret = 0;
2022         request = msg.request.master;
2023         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
2024                         vhost_message_str[request]) {
2025                 if (request != VHOST_USER_IOTLB_MSG)
2026                         RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
2027                                 vhost_message_str[request]);
2028                 else
2029                         RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
2030                                 vhost_message_str[request]);
2031         } else {
2032                 RTE_LOG(DEBUG, VHOST_CONFIG, "External request %d\n", request);
2033         }
2034
2035         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
2036         if (ret < 0) {
2037                 RTE_LOG(ERR, VHOST_CONFIG,
2038                         "failed to alloc queue\n");
2039                 return -1;
2040         }
2041
2042         /*
2043          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
2044          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
2045          * and device is destroyed. destroy_device waits for queues to be
2046          * inactive, so it is safe. Otherwise taking the access_lock
2047          * would cause a dead lock.
2048          */
2049         switch (request) {
2050         case VHOST_USER_SET_FEATURES:
2051         case VHOST_USER_SET_PROTOCOL_FEATURES:
2052         case VHOST_USER_SET_OWNER:
2053         case VHOST_USER_SET_MEM_TABLE:
2054         case VHOST_USER_SET_LOG_BASE:
2055         case VHOST_USER_SET_LOG_FD:
2056         case VHOST_USER_SET_VRING_NUM:
2057         case VHOST_USER_SET_VRING_ADDR:
2058         case VHOST_USER_SET_VRING_BASE:
2059         case VHOST_USER_SET_VRING_KICK:
2060         case VHOST_USER_SET_VRING_CALL:
2061         case VHOST_USER_SET_VRING_ERR:
2062         case VHOST_USER_SET_VRING_ENABLE:
2063         case VHOST_USER_SEND_RARP:
2064         case VHOST_USER_NET_SET_MTU:
2065         case VHOST_USER_SET_SLAVE_REQ_FD:
2066                 vhost_user_lock_all_queue_pairs(dev);
2067                 unlock_required = 1;
2068                 break;
2069         default:
2070                 break;
2071
2072         }
2073
2074         handled = false;
2075         if (dev->extern_ops.pre_msg_handle) {
2076                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
2077                                 (void *)&msg);
2078                 switch (ret) {
2079                 case RTE_VHOST_MSG_RESULT_REPLY:
2080                         send_vhost_reply(fd, &msg);
2081                         /* Fall-through */
2082                 case RTE_VHOST_MSG_RESULT_ERR:
2083                 case RTE_VHOST_MSG_RESULT_OK:
2084                         handled = true;
2085                         goto skip_to_post_handle;
2086                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2087                 default:
2088                         break;
2089                 }
2090         }
2091
2092         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
2093                 if (!vhost_message_handlers[request])
2094                         goto skip_to_post_handle;
2095                 ret = vhost_message_handlers[request](&dev, &msg, fd);
2096
2097                 switch (ret) {
2098                 case RTE_VHOST_MSG_RESULT_ERR:
2099                         RTE_LOG(ERR, VHOST_CONFIG,
2100                                 "Processing %s failed.\n",
2101                                 vhost_message_str[request]);
2102                         handled = true;
2103                         break;
2104                 case RTE_VHOST_MSG_RESULT_OK:
2105                         RTE_LOG(DEBUG, VHOST_CONFIG,
2106                                 "Processing %s succeeded.\n",
2107                                 vhost_message_str[request]);
2108                         handled = true;
2109                         break;
2110                 case RTE_VHOST_MSG_RESULT_REPLY:
2111                         RTE_LOG(DEBUG, VHOST_CONFIG,
2112                                 "Processing %s succeeded and needs reply.\n",
2113                                 vhost_message_str[request]);
2114                         send_vhost_reply(fd, &msg);
2115                         handled = true;
2116                         break;
2117                 default:
2118                         break;
2119                 }
2120         }
2121
2122 skip_to_post_handle:
2123         if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2124                         dev->extern_ops.post_msg_handle) {
2125                 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
2126                                 (void *)&msg);
2127                 switch (ret) {
2128                 case RTE_VHOST_MSG_RESULT_REPLY:
2129                         send_vhost_reply(fd, &msg);
2130                         /* Fall-through */
2131                 case RTE_VHOST_MSG_RESULT_ERR:
2132                 case RTE_VHOST_MSG_RESULT_OK:
2133                         handled = true;
2134                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2135                 default:
2136                         break;
2137                 }
2138         }
2139
2140         if (unlock_required)
2141                 vhost_user_unlock_all_queue_pairs(dev);
2142
2143         /* If message was not handled at this stage, treat it as an error */
2144         if (!handled) {
2145                 RTE_LOG(ERR, VHOST_CONFIG,
2146                         "vhost message (req: %d) was not handled.\n", request);
2147                 ret = RTE_VHOST_MSG_RESULT_ERR;
2148         }
2149
2150         /*
2151          * If the request required a reply that was already sent,
2152          * this optional reply-ack won't be sent as the
2153          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2154          */
2155         if (msg.flags & VHOST_USER_NEED_REPLY) {
2156                 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2157                 msg.size = sizeof(msg.payload.u64);
2158                 msg.fd_num = 0;
2159                 send_vhost_reply(fd, &msg);
2160         } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2161                 RTE_LOG(ERR, VHOST_CONFIG,
2162                         "vhost message handling failed.\n");
2163                 return -1;
2164         }
2165
2166         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2167                 dev->flags |= VIRTIO_DEV_READY;
2168
2169                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2170                         if (dev->dequeue_zero_copy) {
2171                                 RTE_LOG(INFO, VHOST_CONFIG,
2172                                                 "dequeue zero copy is enabled\n");
2173                         }
2174
2175                         if (dev->notify_ops->new_device(dev->vid) == 0)
2176                                 dev->flags |= VIRTIO_DEV_RUNNING;
2177                 }
2178         }
2179
2180         did = dev->vdpa_dev_id;
2181         vdpa_dev = rte_vdpa_get_device(did);
2182         if (vdpa_dev && virtio_is_ready(dev) &&
2183                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2184                         msg.request.master == VHOST_USER_SET_VRING_CALL) {
2185                 if (vdpa_dev->ops->dev_conf)
2186                         vdpa_dev->ops->dev_conf(dev->vid);
2187                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2188         }
2189
2190         return 0;
2191 }
2192
2193 static int process_slave_message_reply(struct virtio_net *dev,
2194                                        const struct VhostUserMsg *msg)
2195 {
2196         struct VhostUserMsg msg_reply;
2197         int ret;
2198
2199         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2200                 return 0;
2201
2202         if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
2203                 ret = -1;
2204                 goto out;
2205         }
2206
2207         if (msg_reply.request.slave != msg->request.slave) {
2208                 RTE_LOG(ERR, VHOST_CONFIG,
2209                         "Received unexpected msg type (%u), expected %u\n",
2210                         msg_reply.request.slave, msg->request.slave);
2211                 ret = -1;
2212                 goto out;
2213         }
2214
2215         ret = msg_reply.payload.u64 ? -1 : 0;
2216
2217 out:
2218         rte_spinlock_unlock(&dev->slave_req_lock);
2219         return ret;
2220 }
2221
2222 int
2223 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2224 {
2225         int ret;
2226         struct VhostUserMsg msg = {
2227                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2228                 .flags = VHOST_USER_VERSION,
2229                 .size = sizeof(msg.payload.iotlb),
2230                 .payload.iotlb = {
2231                         .iova = iova,
2232                         .perm = perm,
2233                         .type = VHOST_IOTLB_MISS,
2234                 },
2235         };
2236
2237         ret = send_vhost_message(dev->slave_req_fd, &msg);
2238         if (ret < 0) {
2239                 RTE_LOG(ERR, VHOST_CONFIG,
2240                                 "Failed to send IOTLB miss message (%d)\n",
2241                                 ret);
2242                 return ret;
2243         }
2244
2245         return 0;
2246 }
2247
2248 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2249                                                     int index, int fd,
2250                                                     uint64_t offset,
2251                                                     uint64_t size)
2252 {
2253         int ret;
2254         struct VhostUserMsg msg = {
2255                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2256                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2257                 .size = sizeof(msg.payload.area),
2258                 .payload.area = {
2259                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
2260                         .size = size,
2261                         .offset = offset,
2262                 },
2263         };
2264
2265         if (fd < 0)
2266                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2267         else {
2268                 msg.fds[0] = fd;
2269                 msg.fd_num = 1;
2270         }
2271
2272         ret = send_vhost_slave_message(dev, &msg);
2273         if (ret < 0) {
2274                 RTE_LOG(ERR, VHOST_CONFIG,
2275                         "Failed to set host notifier (%d)\n", ret);
2276                 return ret;
2277         }
2278
2279         return process_slave_message_reply(dev, &msg);
2280 }
2281
2282 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2283 {
2284         struct virtio_net *dev;
2285         struct rte_vdpa_device *vdpa_dev;
2286         int vfio_device_fd, did, ret = 0;
2287         uint64_t offset, size;
2288         unsigned int i;
2289
2290         dev = get_device(vid);
2291         if (!dev)
2292                 return -ENODEV;
2293
2294         did = dev->vdpa_dev_id;
2295         if (did < 0)
2296                 return -EINVAL;
2297
2298         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2299             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2300             !(dev->protocol_features &
2301                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2302             !(dev->protocol_features &
2303                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2304             !(dev->protocol_features &
2305                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2306                 return -ENOTSUP;
2307
2308         vdpa_dev = rte_vdpa_get_device(did);
2309         if (!vdpa_dev)
2310                 return -ENODEV;
2311
2312         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2313         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2314
2315         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2316         if (vfio_device_fd < 0)
2317                 return -ENOTSUP;
2318
2319         if (enable) {
2320                 for (i = 0; i < dev->nr_vring; i++) {
2321                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2322                                         &size) < 0) {
2323                                 ret = -ENOTSUP;
2324                                 goto disable;
2325                         }
2326
2327                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2328                                         vfio_device_fd, offset, size) < 0) {
2329                                 ret = -EFAULT;
2330                                 goto disable;
2331                         }
2332                 }
2333         } else {
2334 disable:
2335                 for (i = 0; i < dev->nr_vring; i++) {
2336                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
2337                                         0, 0);
2338                 }
2339         }
2340
2341         return ret;
2342 }