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 #ifdef F_ADD_SEALS /* if file sealing is supported, so is memfd */
  41 #include <linux/memfd.h>
  42 #define MEMFD_SUPPORTED
  43 #endif
  44
  45 #include <rte_common.h>
  46 #include <rte_malloc.h>
  47 #include <rte_log.h>
  48
  49 #include "iotlb.h"
  50 #include "vhost.h"
  51 #include "vhost_user.h"
  52
  53 #define VIRTIO_MIN_MTU 68
  54 #define VIRTIO_MAX_MTU 65535
  55
  56 #define INFLIGHT_ALIGNMENT      64
  57 #define INFLIGHT_VERSION        0x1
  58
  59 static const char *vhost_message_str[VHOST_USER_MAX] = {
  60         [VHOST_USER_NONE] = "VHOST_USER_NONE",
  61         [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
  62         [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
  63         [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
  64         [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
  65         [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
  66         [VHOST_USER_SET_LOG_BASE] = "VHOST_USER_SET_LOG_BASE",
  67         [VHOST_USER_SET_LOG_FD] = "VHOST_USER_SET_LOG_FD",
  68         [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
  69         [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
  70         [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
  71         [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
  72         [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
  73         [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
  74         [VHOST_USER_SET_VRING_ERR]  = "VHOST_USER_SET_VRING_ERR",
  75         [VHOST_USER_GET_PROTOCOL_FEATURES]  = "VHOST_USER_GET_PROTOCOL_FEATURES",
  76         [VHOST_USER_SET_PROTOCOL_FEATURES]  = "VHOST_USER_SET_PROTOCOL_FEATURES",
  77         [VHOST_USER_GET_QUEUE_NUM]  = "VHOST_USER_GET_QUEUE_NUM",
  78         [VHOST_USER_SET_VRING_ENABLE]  = "VHOST_USER_SET_VRING_ENABLE",
  79         [VHOST_USER_SEND_RARP]  = "VHOST_USER_SEND_RARP",
  80         [VHOST_USER_NET_SET_MTU]  = "VHOST_USER_NET_SET_MTU",
  81         [VHOST_USER_SET_SLAVE_REQ_FD]  = "VHOST_USER_SET_SLAVE_REQ_FD",
  82         [VHOST_USER_IOTLB_MSG]  = "VHOST_USER_IOTLB_MSG",
  83         [VHOST_USER_CRYPTO_CREATE_SESS] = "VHOST_USER_CRYPTO_CREATE_SESS",
  84         [VHOST_USER_CRYPTO_CLOSE_SESS] = "VHOST_USER_CRYPTO_CLOSE_SESS",
  85         [VHOST_USER_POSTCOPY_ADVISE]  = "VHOST_USER_POSTCOPY_ADVISE",
  86         [VHOST_USER_POSTCOPY_LISTEN]  = "VHOST_USER_POSTCOPY_LISTEN",
  87         [VHOST_USER_POSTCOPY_END]  = "VHOST_USER_POSTCOPY_END",
  88         [VHOST_USER_GET_INFLIGHT_FD] = "VHOST_USER_GET_INFLIGHT_FD",
  89         [VHOST_USER_SET_INFLIGHT_FD] = "VHOST_USER_SET_INFLIGHT_FD",
  90 };
  91
  92 static int send_vhost_reply(int sockfd, struct VhostUserMsg *msg);
  93 static int read_vhost_message(int sockfd, struct VhostUserMsg *msg);
  94
  95 static void
  96 close_msg_fds(struct VhostUserMsg *msg)
  97 {
  98         int i;
  99
 100         for (i = 0; i < msg->fd_num; i++)
 101                 close(msg->fds[i]);
 102 }
 103
 104 /*
 105  * Ensure the expected number of FDs is received,
 106  * close all FDs and return an error if this is not the case.
 107  */
 108 static int
 109 validate_msg_fds(struct VhostUserMsg *msg, int expected_fds)
 110 {
 111         if (msg->fd_num == expected_fds)
 112                 return 0;
 113
 114         VHOST_LOG_CONFIG(ERR,
 115                 " Expect %d FDs for request %s, received %d\n",
 116                 expected_fds,
 117                 vhost_message_str[msg->request.master],
 118                 msg->fd_num);
 119
 120         close_msg_fds(msg);
 121
 122         return -1;
 123 }
 124
 125 static uint64_t
 126 get_blk_size(int fd)
 127 {
 128         struct stat stat;
 129         int ret;
 130
 131         ret = fstat(fd, &stat);
 132         return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
 133 }
 134
 135 /*
 136  * Reclaim all the outstanding zmbufs for a virtqueue.
 137  */
 138 static void
 139 drain_zmbuf_list(struct vhost_virtqueue *vq)
 140 {
 141         struct zcopy_mbuf *zmbuf, *next;
 142
 143         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
 144              zmbuf != NULL; zmbuf = next) {
 145                 next = TAILQ_NEXT(zmbuf, next);
 146
 147                 while (!mbuf_is_consumed(zmbuf->mbuf))
 148                         usleep(1000);
 149
 150                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
 151                 restore_mbuf(zmbuf->mbuf);
 152                 rte_pktmbuf_free(zmbuf->mbuf);
 153                 put_zmbuf(zmbuf);
 154                 vq->nr_zmbuf -= 1;
 155         }
 156 }
 157
 158 static void
 159 free_mem_region(struct virtio_net *dev)
 160 {
 161         uint32_t i;
 162         struct rte_vhost_mem_region *reg;
 163         struct vhost_virtqueue *vq;
 164
 165         if (!dev || !dev->mem)
 166                 return;
 167
 168         if (dev->dequeue_zero_copy) {
 169                 for (i = 0; i < dev->nr_vring; i++) {
 170                         vq = dev->virtqueue[i];
 171                         if (vq)
 172                                 drain_zmbuf_list(vq);
 173                 }
 174         }
 175
 176         for (i = 0; i < dev->mem->nregions; i++) {
 177                 reg = &dev->mem->regions[i];
 178                 if (reg->host_user_addr) {
 179                         munmap(reg->mmap_addr, reg->mmap_size);
 180                         close(reg->fd);
 181                 }
 182         }
 183 }
 184
 185 void
 186 vhost_backend_cleanup(struct virtio_net *dev)
 187 {
 188         if (dev->mem) {
 189                 free_mem_region(dev);
 190                 rte_free(dev->mem);
 191                 dev->mem = NULL;
 192         }
 193
 194         free(dev->guest_pages);
 195         dev->guest_pages = NULL;
 196
 197         if (dev->log_addr) {
 198                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 199                 dev->log_addr = 0;
 200         }
 201
 202         if (dev->inflight_info) {
 203                 if (dev->inflight_info->addr) {
 204                         munmap(dev->inflight_info->addr,
 205                                dev->inflight_info->size);
 206                         dev->inflight_info->addr = NULL;
 207                 }
 208
 209                 if (dev->inflight_info->fd > 0) {
 210                         close(dev->inflight_info->fd);
 211                         dev->inflight_info->fd = -1;
 212                 }
 213
 214                 free(dev->inflight_info);
 215                 dev->inflight_info = NULL;
 216         }
 217
 218         if (dev->slave_req_fd >= 0) {
 219                 close(dev->slave_req_fd);
 220                 dev->slave_req_fd = -1;
 221         }
 222
 223         if (dev->postcopy_ufd >= 0) {
 224                 close(dev->postcopy_ufd);
 225                 dev->postcopy_ufd = -1;
 226         }
 227
 228         dev->postcopy_listening = 0;
 229 }
 230
 231 /*
 232  * This function just returns success at the moment unless
 233  * the device hasn't been initialised.
 234  */
 235 static int
 236 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
 237                         struct VhostUserMsg *msg,
 238                         int main_fd __rte_unused)
 239 {
 240         if (validate_msg_fds(msg, 0) != 0)
 241                 return RTE_VHOST_MSG_RESULT_ERR;
 242
 243         return RTE_VHOST_MSG_RESULT_OK;
 244 }
 245
 246 static int
 247 vhost_user_reset_owner(struct virtio_net **pdev,
 248                         struct VhostUserMsg *msg,
 249                         int main_fd __rte_unused)
 250 {
 251         struct virtio_net *dev = *pdev;
 252
 253         if (validate_msg_fds(msg, 0) != 0)
 254                 return RTE_VHOST_MSG_RESULT_ERR;
 255
 256         vhost_destroy_device_notify(dev);
 257
 258         cleanup_device(dev, 0);
 259         reset_device(dev);
 260         return RTE_VHOST_MSG_RESULT_OK;
 261 }
 262
 263 /*
 264  * The features that we support are requested.
 265  */
 266 static int
 267 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 268                         int main_fd __rte_unused)
 269 {
 270         struct virtio_net *dev = *pdev;
 271         uint64_t features = 0;
 272
 273         if (validate_msg_fds(msg, 0) != 0)
 274                 return RTE_VHOST_MSG_RESULT_ERR;
 275
 276         rte_vhost_driver_get_features(dev->ifname, &features);
 277
 278         msg->payload.u64 = features;
 279         msg->size = sizeof(msg->payload.u64);
 280         msg->fd_num = 0;
 281
 282         return RTE_VHOST_MSG_RESULT_REPLY;
 283 }
 284
 285 /*
 286  * The queue number that we support are requested.
 287  */
 288 static int
 289 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
 290                         int main_fd __rte_unused)
 291 {
 292         struct virtio_net *dev = *pdev;
 293         uint32_t queue_num = 0;
 294
 295         if (validate_msg_fds(msg, 0) != 0)
 296                 return RTE_VHOST_MSG_RESULT_ERR;
 297
 298         rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
 299
 300         msg->payload.u64 = (uint64_t)queue_num;
 301         msg->size = sizeof(msg->payload.u64);
 302         msg->fd_num = 0;
 303
 304         return RTE_VHOST_MSG_RESULT_REPLY;
 305 }
 306
 307 /*
 308  * We receive the negotiated features supported by us and the virtio device.
 309  */
 310 static int
 311 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 312                         int main_fd __rte_unused)
 313 {
 314         struct virtio_net *dev = *pdev;
 315         uint64_t features = msg->payload.u64;
 316         uint64_t vhost_features = 0;
 317         struct rte_vdpa_device *vdpa_dev;
 318         int did = -1;
 319
 320         if (validate_msg_fds(msg, 0) != 0)
 321                 return RTE_VHOST_MSG_RESULT_ERR;
 322
 323         rte_vhost_driver_get_features(dev->ifname, &vhost_features);
 324         if (features & ~vhost_features) {
 325                 VHOST_LOG_CONFIG(ERR,
 326                         "(%d) received invalid negotiated features.\n",
 327                         dev->vid);
 328                 return RTE_VHOST_MSG_RESULT_ERR;
 329         }
 330
 331         if (dev->flags & VIRTIO_DEV_RUNNING) {
 332                 if (dev->features == features)
 333                         return RTE_VHOST_MSG_RESULT_OK;
 334
 335                 /*
 336                  * Error out if master tries to change features while device is
 337                  * in running state. The exception being VHOST_F_LOG_ALL, which
 338                  * is enabled when the live-migration starts.
 339                  */
 340                 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
 341                         VHOST_LOG_CONFIG(ERR,
 342                                 "(%d) features changed while device is running.\n",
 343                                 dev->vid);
 344                         return RTE_VHOST_MSG_RESULT_ERR;
 345                 }
 346
 347                 if (dev->notify_ops->features_changed)
 348                         dev->notify_ops->features_changed(dev->vid, features);
 349         }
 350
 351         dev->features = features;
 352         if (dev->features &
 353                 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
 354                 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 355         } else {
 356                 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
 357         }
 358         VHOST_LOG_CONFIG(INFO,
 359                 "negotiated Virtio features: 0x%" PRIx64 "\n", dev->features);
 360         VHOST_LOG_CONFIG(DEBUG,
 361                 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
 362                 dev->vid,
 363                 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
 364                 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 365
 366         if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
 367             !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
 368                 /*
 369                  * Remove all but first queue pair if MQ hasn't been
 370                  * negotiated. This is safe because the device is not
 371                  * running at this stage.
 372                  */
 373                 while (dev->nr_vring > 2) {
 374                         struct vhost_virtqueue *vq;
 375
 376                         vq = dev->virtqueue[--dev->nr_vring];
 377                         if (!vq)
 378                                 continue;
 379
 380                         dev->virtqueue[dev->nr_vring] = NULL;
 381                         cleanup_vq(vq, 1);
 382                         cleanup_vq_inflight(dev, vq);
 383                         free_vq(dev, vq);
 384                 }
 385         }
 386
 387         did = dev->vdpa_dev_id;
 388         vdpa_dev = rte_vdpa_get_device(did);
 389         if (vdpa_dev && vdpa_dev->ops->set_features)
 390                 vdpa_dev->ops->set_features(dev->vid);
 391
 392         return RTE_VHOST_MSG_RESULT_OK;
 393 }
 394
 395 /*
 396  * The virtio device sends us the size of the descriptor ring.
 397  */
 398 static int
 399 vhost_user_set_vring_num(struct virtio_net **pdev,
 400                         struct VhostUserMsg *msg,
 401                         int main_fd __rte_unused)
 402 {
 403         struct virtio_net *dev = *pdev;
 404         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 405
 406         if (validate_msg_fds(msg, 0) != 0)
 407                 return RTE_VHOST_MSG_RESULT_ERR;
 408
 409         vq->size = msg->payload.state.num;
 410
 411         /* VIRTIO 1.0, 2.4 Virtqueues says:
 412          *
 413          *   Queue Size value is always a power of 2. The maximum Queue Size
 414          *   value is 32768.
 415          *
 416          * VIRTIO 1.1 2.7 Virtqueues says:
 417          *
 418          *   Packed virtqueues support up to 2^15 entries each.
 419          */
 420         if (!vq_is_packed(dev)) {
 421                 if (vq->size & (vq->size - 1)) {
 422                         VHOST_LOG_CONFIG(ERR,
 423                                 "invalid virtqueue size %u\n", vq->size);
 424                         return RTE_VHOST_MSG_RESULT_ERR;
 425                 }
 426         }
 427
 428         if (vq->size > 32768) {
 429                 VHOST_LOG_CONFIG(ERR,
 430                         "invalid virtqueue size %u\n", vq->size);
 431                 return RTE_VHOST_MSG_RESULT_ERR;
 432         }
 433
 434         if (dev->dequeue_zero_copy) {
 435                 vq->nr_zmbuf = 0;
 436                 vq->last_zmbuf_idx = 0;
 437                 vq->zmbuf_size = vq->size;
 438                 if (vq->zmbufs)
 439                         rte_free(vq->zmbufs);
 440                 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
 441                                          sizeof(struct zcopy_mbuf), 0);
 442                 if (vq->zmbufs == NULL) {
 443                         VHOST_LOG_CONFIG(WARNING,
 444                                 "failed to allocate mem for zero copy; "
 445                                 "zero copy is force disabled\n");
 446                         dev->dequeue_zero_copy = 0;
 447                 }
 448                 TAILQ_INIT(&vq->zmbuf_list);
 449         }
 450
 451         if (vq_is_packed(dev)) {
 452                 if (vq->shadow_used_packed)
 453                         rte_free(vq->shadow_used_packed);
 454                 vq->shadow_used_packed = rte_malloc(NULL,
 455                                 vq->size *
 456                                 sizeof(struct vring_used_elem_packed),
 457                                 RTE_CACHE_LINE_SIZE);
 458                 if (!vq->shadow_used_packed) {
 459                         VHOST_LOG_CONFIG(ERR,
 460                                         "failed to allocate memory for shadow used ring.\n");
 461                         return RTE_VHOST_MSG_RESULT_ERR;
 462                 }
 463
 464         } else {
 465                 if (vq->shadow_used_split)
 466                         rte_free(vq->shadow_used_split);
 467                 vq->shadow_used_split = rte_malloc(NULL,
 468                                 vq->size * sizeof(struct vring_used_elem),
 469                                 RTE_CACHE_LINE_SIZE);
 470                 if (!vq->shadow_used_split) {
 471                         VHOST_LOG_CONFIG(ERR,
 472                                         "failed to allocate memory for shadow used ring.\n");
 473                         return RTE_VHOST_MSG_RESULT_ERR;
 474                 }
 475         }
 476
 477         if (vq->batch_copy_elems)
 478                 rte_free(vq->batch_copy_elems);
 479         vq->batch_copy_elems = rte_malloc(NULL,
 480                                 vq->size * sizeof(struct batch_copy_elem),
 481                                 RTE_CACHE_LINE_SIZE);
 482         if (!vq->batch_copy_elems) {
 483                 VHOST_LOG_CONFIG(ERR,
 484                         "failed to allocate memory for batching copy.\n");
 485                 return RTE_VHOST_MSG_RESULT_ERR;
 486         }
 487
 488         return RTE_VHOST_MSG_RESULT_OK;
 489 }
 490
 491 /*
 492  * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
 493  * same numa node as the memory of vring descriptor.
 494  */
 495 #ifdef RTE_LIBRTE_VHOST_NUMA
 496 static struct virtio_net*
 497 numa_realloc(struct virtio_net *dev, int index)
 498 {
 499         int oldnode, newnode;
 500         struct virtio_net *old_dev;
 501         struct vhost_virtqueue *old_vq, *vq;
 502         struct zcopy_mbuf *new_zmbuf;
 503         struct vring_used_elem *new_shadow_used_split;
 504         struct vring_used_elem_packed *new_shadow_used_packed;
 505         struct batch_copy_elem *new_batch_copy_elems;
 506         int ret;
 507
 508         if (dev->flags & VIRTIO_DEV_RUNNING)
 509                 return dev;
 510
 511         old_dev = dev;
 512         vq = old_vq = dev->virtqueue[index];
 513
 514         ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
 515                             MPOL_F_NODE | MPOL_F_ADDR);
 516
 517         /* check if we need to reallocate vq */
 518         ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
 519                              MPOL_F_NODE | MPOL_F_ADDR);
 520         if (ret) {
 521                 VHOST_LOG_CONFIG(ERR,
 522                         "Unable to get vq numa information.\n");
 523                 return dev;
 524         }
 525         if (oldnode != newnode) {
 526                 VHOST_LOG_CONFIG(INFO,
 527                         "reallocate vq from %d to %d node\n", oldnode, newnode);
 528                 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
 529                 if (!vq)
 530                         return dev;
 531
 532                 memcpy(vq, old_vq, sizeof(*vq));
 533                 TAILQ_INIT(&vq->zmbuf_list);
 534
 535                 if (dev->dequeue_zero_copy) {
 536                         new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
 537                                         sizeof(struct zcopy_mbuf), 0, newnode);
 538                         if (new_zmbuf) {
 539                                 rte_free(vq->zmbufs);
 540                                 vq->zmbufs = new_zmbuf;
 541                         }
 542                 }
 543
 544                 if (vq_is_packed(dev)) {
 545                         new_shadow_used_packed = rte_malloc_socket(NULL,
 546                                         vq->size *
 547                                         sizeof(struct vring_used_elem_packed),
 548                                         RTE_CACHE_LINE_SIZE,
 549                                         newnode);
 550                         if (new_shadow_used_packed) {
 551                                 rte_free(vq->shadow_used_packed);
 552                                 vq->shadow_used_packed = new_shadow_used_packed;
 553                         }
 554                 } else {
 555                         new_shadow_used_split = rte_malloc_socket(NULL,
 556                                         vq->size *
 557                                         sizeof(struct vring_used_elem),
 558                                         RTE_CACHE_LINE_SIZE,
 559                                         newnode);
 560                         if (new_shadow_used_split) {
 561                                 rte_free(vq->shadow_used_split);
 562                                 vq->shadow_used_split = new_shadow_used_split;
 563                         }
 564                 }
 565
 566                 new_batch_copy_elems = rte_malloc_socket(NULL,
 567                         vq->size * sizeof(struct batch_copy_elem),
 568                         RTE_CACHE_LINE_SIZE,
 569                         newnode);
 570                 if (new_batch_copy_elems) {
 571                         rte_free(vq->batch_copy_elems);
 572                         vq->batch_copy_elems = new_batch_copy_elems;
 573                 }
 574
 575                 rte_free(old_vq);
 576         }
 577
 578         /* check if we need to reallocate dev */
 579         ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
 580                             MPOL_F_NODE | MPOL_F_ADDR);
 581         if (ret) {
 582                 VHOST_LOG_CONFIG(ERR,
 583                         "Unable to get dev numa information.\n");
 584                 goto out;
 585         }
 586         if (oldnode != newnode) {
 587                 VHOST_LOG_CONFIG(INFO,
 588                         "reallocate dev from %d to %d node\n",
 589                         oldnode, newnode);
 590                 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
 591                 if (!dev) {
 592                         dev = old_dev;
 593                         goto out;
 594                 }
 595
 596                 memcpy(dev, old_dev, sizeof(*dev));
 597                 rte_free(old_dev);
 598         }
 599
 600 out:
 601         dev->virtqueue[index] = vq;
 602         vhost_devices[dev->vid] = dev;
 603
 604         if (old_vq != vq)
 605                 vhost_user_iotlb_init(dev, index);
 606
 607         return dev;
 608 }
 609 #else
 610 static struct virtio_net*
 611 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 612 {
 613         return dev;
 614 }
 615 #endif
 616
 617 /* Converts QEMU virtual address to Vhost virtual address. */
 618 static uint64_t
 619 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
 620 {
 621         struct rte_vhost_mem_region *r;
 622         uint32_t i;
 623
 624         if (unlikely(!dev || !dev->mem))
 625                 goto out_error;
 626
 627         /* Find the region where the address lives. */
 628         for (i = 0; i < dev->mem->nregions; i++) {
 629                 r = &dev->mem->regions[i];
 630
 631                 if (qva >= r->guest_user_addr &&
 632                     qva <  r->guest_user_addr + r->size) {
 633
 634                         if (unlikely(*len > r->guest_user_addr + r->size - qva))
 635                                 *len = r->guest_user_addr + r->size - qva;
 636
 637                         return qva - r->guest_user_addr +
 638                                r->host_user_addr;
 639                 }
 640         }
 641 out_error:
 642         *len = 0;
 643
 644         return 0;
 645 }
 646
 647
 648 /*
 649  * Converts ring address to Vhost virtual address.
 650  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 651  * else it is a QEMU virtual address.
 652  */
 653 static uint64_t
 654 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 655                 uint64_t ra, uint64_t *size)
 656 {
 657         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 658                 uint64_t vva;
 659
 660                 vhost_user_iotlb_rd_lock(vq);
 661                 vva = vhost_iova_to_vva(dev, vq, ra,
 662                                         size, VHOST_ACCESS_RW);
 663                 vhost_user_iotlb_rd_unlock(vq);
 664
 665                 return vva;
 666         }
 667
 668         return qva_to_vva(dev, ra, size);
 669 }
 670
 671 static uint64_t
 672 log_addr_to_gpa(struct virtio_net *dev, struct vhost_virtqueue *vq)
 673 {
 674         uint64_t log_gpa;
 675
 676         vhost_user_iotlb_rd_lock(vq);
 677         log_gpa = translate_log_addr(dev, vq, vq->ring_addrs.log_guest_addr);
 678         vhost_user_iotlb_rd_unlock(vq);
 679
 680         return log_gpa;
 681 }
 682
 683 static struct virtio_net *
 684 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 685 {
 686         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 687         struct vhost_vring_addr *addr = &vq->ring_addrs;
 688         uint64_t len, expected_len;
 689
 690         if (addr->flags & (1 << VHOST_VRING_F_LOG)) {
 691                 vq->log_guest_addr =
 692                         log_addr_to_gpa(dev, vq);
 693                 if (vq->log_guest_addr == 0) {
 694                         VHOST_LOG_CONFIG(DEBUG,
 695                                 "(%d) failed to map log_guest_addr.\n",
 696                                 dev->vid);
 697                         return dev;
 698                 }
 699         }
 700
 701         if (vq_is_packed(dev)) {
 702                 len = sizeof(struct vring_packed_desc) * vq->size;
 703                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 704                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 705                 if (vq->desc_packed == NULL ||
 706                                 len != sizeof(struct vring_packed_desc) *
 707                                 vq->size) {
 708                         VHOST_LOG_CONFIG(DEBUG,
 709                                 "(%d) failed to map desc_packed ring.\n",
 710                                 dev->vid);
 711                         return dev;
 712                 }
 713
 714                 dev = numa_realloc(dev, vq_index);
 715                 vq = dev->virtqueue[vq_index];
 716                 addr = &vq->ring_addrs;
 717
 718                 len = sizeof(struct vring_packed_desc_event);
 719                 vq->driver_event = (struct vring_packed_desc_event *)
 720                                         (uintptr_t)ring_addr_to_vva(dev,
 721                                         vq, addr->avail_user_addr, &len);
 722                 if (vq->driver_event == NULL ||
 723                                 len != sizeof(struct vring_packed_desc_event)) {
 724                         VHOST_LOG_CONFIG(DEBUG,
 725                                 "(%d) failed to find driver area address.\n",
 726                                 dev->vid);
 727                         return dev;
 728                 }
 729
 730                 len = sizeof(struct vring_packed_desc_event);
 731                 vq->device_event = (struct vring_packed_desc_event *)
 732                                         (uintptr_t)ring_addr_to_vva(dev,
 733                                         vq, addr->used_user_addr, &len);
 734                 if (vq->device_event == NULL ||
 735                                 len != sizeof(struct vring_packed_desc_event)) {
 736                         VHOST_LOG_CONFIG(DEBUG,
 737                                 "(%d) failed to find device area address.\n",
 738                                 dev->vid);
 739                         return dev;
 740                 }
 741
 742                 vq->access_ok = 1;
 743                 return dev;
 744         }
 745
 746         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 747         if (vq->desc && vq->avail && vq->used)
 748                 return dev;
 749
 750         len = sizeof(struct vring_desc) * vq->size;
 751         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 752                         vq, addr->desc_user_addr, &len);
 753         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 754                 VHOST_LOG_CONFIG(DEBUG,
 755                         "(%d) failed to map desc ring.\n",
 756                         dev->vid);
 757                 return dev;
 758         }
 759
 760         dev = numa_realloc(dev, vq_index);
 761         vq = dev->virtqueue[vq_index];
 762         addr = &vq->ring_addrs;
 763
 764         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 765         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 766                 len += sizeof(uint16_t);
 767         expected_len = len;
 768         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 769                         vq, addr->avail_user_addr, &len);
 770         if (vq->avail == 0 || len != expected_len) {
 771                 VHOST_LOG_CONFIG(DEBUG,
 772                         "(%d) failed to map avail ring.\n",
 773                         dev->vid);
 774                 return dev;
 775         }
 776
 777         len = sizeof(struct vring_used) +
 778                 sizeof(struct vring_used_elem) * vq->size;
 779         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 780                 len += sizeof(uint16_t);
 781         expected_len = len;
 782         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 783                         vq, addr->used_user_addr, &len);
 784         if (vq->used == 0 || len != expected_len) {
 785                 VHOST_LOG_CONFIG(DEBUG,
 786                         "(%d) failed to map used ring.\n",
 787                         dev->vid);
 788                 return dev;
 789         }
 790
 791         if (vq->last_used_idx != vq->used->idx) {
 792                 VHOST_LOG_CONFIG(WARNING,
 793                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 794                         "some packets maybe resent for Tx and dropped for Rx\n",
 795                         vq->last_used_idx, vq->used->idx);
 796                 vq->last_used_idx  = vq->used->idx;
 797                 vq->last_avail_idx = vq->used->idx;
 798         }
 799
 800         vq->access_ok = 1;
 801
 802         VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address desc: %p\n",
 803                         dev->vid, vq->desc);
 804         VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address avail: %p\n",
 805                         dev->vid, vq->avail);
 806         VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address used: %p\n",
 807                         dev->vid, vq->used);
 808         VHOST_LOG_CONFIG(DEBUG, "(%d) log_guest_addr: %" PRIx64 "\n",
 809                         dev->vid, vq->log_guest_addr);
 810
 811         return dev;
 812 }
 813
 814 /*
 815  * The virtio device sends us the desc, used and avail ring addresses.
 816  * This function then converts these to our address space.
 817  */
 818 static int
 819 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 820                         int main_fd __rte_unused)
 821 {
 822         struct virtio_net *dev = *pdev;
 823         struct vhost_virtqueue *vq;
 824         struct vhost_vring_addr *addr = &msg->payload.addr;
 825         bool access_ok;
 826
 827         if (validate_msg_fds(msg, 0) != 0)
 828                 return RTE_VHOST_MSG_RESULT_ERR;
 829
 830         if (dev->mem == NULL)
 831                 return RTE_VHOST_MSG_RESULT_ERR;
 832
 833         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 834         vq = dev->virtqueue[msg->payload.addr.index];
 835
 836         access_ok = vq->access_ok;
 837
 838         /*
 839          * Rings addresses should not be interpreted as long as the ring is not
 840          * started and enabled
 841          */
 842         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 843
 844         vring_invalidate(dev, vq);
 845
 846         if ((vq->enabled && (dev->features &
 847                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) ||
 848                         access_ok) {
 849                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 850                 if (!dev)
 851                         return RTE_VHOST_MSG_RESULT_ERR;
 852
 853                 *pdev = dev;
 854         }
 855
 856         return RTE_VHOST_MSG_RESULT_OK;
 857 }
 858
 859 /*
 860  * The virtio device sends us the available ring last used index.
 861  */
 862 static int
 863 vhost_user_set_vring_base(struct virtio_net **pdev,
 864                         struct VhostUserMsg *msg,
 865                         int main_fd __rte_unused)
 866 {
 867         struct virtio_net *dev = *pdev;
 868         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 869         uint64_t val = msg->payload.state.num;
 870
 871         if (validate_msg_fds(msg, 0) != 0)
 872                 return RTE_VHOST_MSG_RESULT_ERR;
 873
 874         if (vq_is_packed(dev)) {
 875                 /*
 876                  * Bit[0:14]: avail index
 877                  * Bit[15]: avail wrap counter
 878                  */
 879                 vq->last_avail_idx = val & 0x7fff;
 880                 vq->avail_wrap_counter = !!(val & (0x1 << 15));
 881                 /*
 882                  * Set used index to same value as available one, as
 883                  * their values should be the same since ring processing
 884                  * was stopped at get time.
 885                  */
 886                 vq->last_used_idx = vq->last_avail_idx;
 887                 vq->used_wrap_counter = vq->avail_wrap_counter;
 888         } else {
 889                 vq->last_used_idx = msg->payload.state.num;
 890                 vq->last_avail_idx = msg->payload.state.num;
 891         }
 892
 893         return RTE_VHOST_MSG_RESULT_OK;
 894 }
 895
 896 static int
 897 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 898                    uint64_t host_phys_addr, uint64_t size)
 899 {
 900         struct guest_page *page, *last_page;
 901         struct guest_page *old_pages;
 902
 903         if (dev->nr_guest_pages == dev->max_guest_pages) {
 904                 dev->max_guest_pages *= 2;
 905                 old_pages = dev->guest_pages;
 906                 dev->guest_pages = realloc(dev->guest_pages,
 907                                         dev->max_guest_pages * sizeof(*page));
 908                 if (!dev->guest_pages) {
 909                         VHOST_LOG_CONFIG(ERR, "cannot realloc guest_pages\n");
 910                         free(old_pages);
 911                         return -1;
 912                 }
 913         }
 914
 915         if (dev->nr_guest_pages > 0) {
 916                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 917                 /* merge if the two pages are continuous */
 918                 if (host_phys_addr == last_page->host_phys_addr +
 919                                       last_page->size) {
 920                         last_page->size += size;
 921                         return 0;
 922                 }
 923         }
 924
 925         page = &dev->guest_pages[dev->nr_guest_pages++];
 926         page->guest_phys_addr = guest_phys_addr;
 927         page->host_phys_addr  = host_phys_addr;
 928         page->size = size;
 929
 930         return 0;
 931 }
 932
 933 static int
 934 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 935                 uint64_t page_size)
 936 {
 937         uint64_t reg_size = reg->size;
 938         uint64_t host_user_addr  = reg->host_user_addr;
 939         uint64_t guest_phys_addr = reg->guest_phys_addr;
 940         uint64_t host_phys_addr;
 941         uint64_t size;
 942
 943         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 944         size = page_size - (guest_phys_addr & (page_size - 1));
 945         size = RTE_MIN(size, reg_size);
 946
 947         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 948                 return -1;
 949
 950         host_user_addr  += size;
 951         guest_phys_addr += size;
 952         reg_size -= size;
 953
 954         while (reg_size > 0) {
 955                 size = RTE_MIN(reg_size, page_size);
 956                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 957                                                   host_user_addr);
 958                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 959                                 size) < 0)
 960                         return -1;
 961
 962                 host_user_addr  += size;
 963                 guest_phys_addr += size;
 964                 reg_size -= size;
 965         }
 966
 967         return 0;
 968 }
 969
 970 #ifdef RTE_LIBRTE_VHOST_DEBUG
 971 /* TODO: enable it only in debug mode? */
 972 static void
 973 dump_guest_pages(struct virtio_net *dev)
 974 {
 975         uint32_t i;
 976         struct guest_page *page;
 977
 978         for (i = 0; i < dev->nr_guest_pages; i++) {
 979                 page = &dev->guest_pages[i];
 980
 981                 VHOST_LOG_CONFIG(INFO,
 982                         "guest physical page region %u\n"
 983                         "\t guest_phys_addr: %" PRIx64 "\n"
 984                         "\t host_phys_addr : %" PRIx64 "\n"
 985                         "\t size           : %" PRIx64 "\n",
 986                         i,
 987                         page->guest_phys_addr,
 988                         page->host_phys_addr,
 989                         page->size);
 990         }
 991 }
 992 #else
 993 #define dump_guest_pages(dev)
 994 #endif
 995
 996 static bool
 997 vhost_memory_changed(struct VhostUserMemory *new,
 998                      struct rte_vhost_memory *old)
 999 {
1000         uint32_t i;
1001
1002         if (new->nregions != old->nregions)
1003                 return true;
1004
1005         for (i = 0; i < new->nregions; ++i) {
1006                 VhostUserMemoryRegion *new_r = &new->regions[i];
1007                 struct rte_vhost_mem_region *old_r = &old->regions[i];
1008
1009                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
1010                         return true;
1011                 if (new_r->memory_size != old_r->size)
1012                         return true;
1013                 if (new_r->userspace_addr != old_r->guest_user_addr)
1014                         return true;
1015         }
1016
1017         return false;
1018 }
1019
1020 static int
1021 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
1022                         int main_fd)
1023 {
1024         struct virtio_net *dev = *pdev;
1025         struct VhostUserMemory *memory = &msg->payload.memory;
1026         struct rte_vhost_mem_region *reg;
1027         void *mmap_addr;
1028         uint64_t mmap_size;
1029         uint64_t mmap_offset;
1030         uint64_t alignment;
1031         uint32_t i;
1032         int populate;
1033         int fd;
1034
1035         if (validate_msg_fds(msg, memory->nregions) != 0)
1036                 return RTE_VHOST_MSG_RESULT_ERR;
1037
1038         if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
1039                 VHOST_LOG_CONFIG(ERR,
1040                         "too many memory regions (%u)\n", memory->nregions);
1041                 return RTE_VHOST_MSG_RESULT_ERR;
1042         }
1043
1044         if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
1045                 VHOST_LOG_CONFIG(INFO,
1046                         "(%d) memory regions not changed\n", dev->vid);
1047
1048                 close_msg_fds(msg);
1049
1050                 return RTE_VHOST_MSG_RESULT_OK;
1051         }
1052
1053         if (dev->mem) {
1054                 free_mem_region(dev);
1055                 rte_free(dev->mem);
1056                 dev->mem = NULL;
1057         }
1058
1059         /* Flush IOTLB cache as previous HVAs are now invalid */
1060         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1061                 for (i = 0; i < dev->nr_vring; i++)
1062                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
1063
1064         dev->nr_guest_pages = 0;
1065         if (!dev->guest_pages) {
1066                 dev->max_guest_pages = 8;
1067                 dev->guest_pages = malloc(dev->max_guest_pages *
1068                                                 sizeof(struct guest_page));
1069                 if (dev->guest_pages == NULL) {
1070                         VHOST_LOG_CONFIG(ERR,
1071                                 "(%d) failed to allocate memory "
1072                                 "for dev->guest_pages\n",
1073                                 dev->vid);
1074                         return RTE_VHOST_MSG_RESULT_ERR;
1075                 }
1076         }
1077
1078         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
1079                 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
1080         if (dev->mem == NULL) {
1081                 VHOST_LOG_CONFIG(ERR,
1082                         "(%d) failed to allocate memory for dev->mem\n",
1083                         dev->vid);
1084                 return RTE_VHOST_MSG_RESULT_ERR;
1085         }
1086         dev->mem->nregions = memory->nregions;
1087
1088         for (i = 0; i < memory->nregions; i++) {
1089                 fd  = msg->fds[i];
1090                 reg = &dev->mem->regions[i];
1091
1092                 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
1093                 reg->guest_user_addr = memory->regions[i].userspace_addr;
1094                 reg->size            = memory->regions[i].memory_size;
1095                 reg->fd              = fd;
1096
1097                 mmap_offset = memory->regions[i].mmap_offset;
1098
1099                 /* Check for memory_size + mmap_offset overflow */
1100                 if (mmap_offset >= -reg->size) {
1101                         VHOST_LOG_CONFIG(ERR,
1102                                 "mmap_offset (%#"PRIx64") and memory_size "
1103                                 "(%#"PRIx64") overflow\n",
1104                                 mmap_offset, reg->size);
1105                         goto err_mmap;
1106                 }
1107
1108                 mmap_size = reg->size + mmap_offset;
1109
1110                 /* mmap() without flag of MAP_ANONYMOUS, should be called
1111                  * with length argument aligned with hugepagesz at older
1112                  * longterm version Linux, like 2.6.32 and 3.2.72, or
1113                  * mmap() will fail with EINVAL.
1114                  *
1115                  * to avoid failure, make sure in caller to keep length
1116                  * aligned.
1117                  */
1118                 alignment = get_blk_size(fd);
1119                 if (alignment == (uint64_t)-1) {
1120                         VHOST_LOG_CONFIG(ERR,
1121                                 "couldn't get hugepage size through fstat\n");
1122                         goto err_mmap;
1123                 }
1124                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
1125                 if (mmap_size == 0) {
1126                         /*
1127                          * It could happen if initial mmap_size + alignment
1128                          * overflows the sizeof uint64, which could happen if
1129                          * either mmap_size or alignment value is wrong.
1130                          *
1131                          * mmap() kernel implementation would return an error,
1132                          * but better catch it before and provide useful info
1133                          * in the logs.
1134                          */
1135                         VHOST_LOG_CONFIG(ERR, "mmap size (0x%" PRIx64 ") "
1136                                         "or alignment (0x%" PRIx64 ") is invalid\n",
1137                                         reg->size + mmap_offset, alignment);
1138                         goto err_mmap;
1139                 }
1140
1141                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
1142                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
1143                                  MAP_SHARED | populate, fd, 0);
1144
1145                 if (mmap_addr == MAP_FAILED) {
1146                         VHOST_LOG_CONFIG(ERR,
1147                                 "mmap region %u failed.\n", i);
1148                         goto err_mmap;
1149                 }
1150
1151                 reg->mmap_addr = mmap_addr;
1152                 reg->mmap_size = mmap_size;
1153                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
1154                                       mmap_offset;
1155
1156                 if (dev->dequeue_zero_copy)
1157                         if (add_guest_pages(dev, reg, alignment) < 0) {
1158                                 VHOST_LOG_CONFIG(ERR,
1159                                         "adding guest pages to region %u failed.\n",
1160                                         i);
1161                                 goto err_mmap;
1162                         }
1163
1164                 VHOST_LOG_CONFIG(INFO,
1165                         "guest memory region %u, size: 0x%" PRIx64 "\n"
1166                         "\t guest physical addr: 0x%" PRIx64 "\n"
1167                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
1168                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
1169                         "\t mmap addr : 0x%" PRIx64 "\n"
1170                         "\t mmap size : 0x%" PRIx64 "\n"
1171                         "\t mmap align: 0x%" PRIx64 "\n"
1172                         "\t mmap off  : 0x%" PRIx64 "\n",
1173                         i, reg->size,
1174                         reg->guest_phys_addr,
1175                         reg->guest_user_addr,
1176                         reg->host_user_addr,
1177                         (uint64_t)(uintptr_t)mmap_addr,
1178                         mmap_size,
1179                         alignment,
1180                         mmap_offset);
1181
1182                 if (dev->postcopy_listening) {
1183                         /*
1184                          * We haven't a better way right now than sharing
1185                          * DPDK's virtual address with Qemu, so that Qemu can
1186                          * retrieve the region offset when handling userfaults.
1187                          */
1188                         memory->regions[i].userspace_addr =
1189                                 reg->host_user_addr;
1190                 }
1191         }
1192         if (dev->postcopy_listening) {
1193                 /* Send the addresses back to qemu */
1194                 msg->fd_num = 0;
1195                 send_vhost_reply(main_fd, msg);
1196
1197                 /* Wait for qemu to acknolwedge it's got the addresses
1198                  * we've got to wait before we're allowed to generate faults.
1199                  */
1200                 VhostUserMsg ack_msg;
1201                 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1202                         VHOST_LOG_CONFIG(ERR,
1203                                 "Failed to read qemu ack on postcopy set-mem-table\n");
1204                         goto err_mmap;
1205                 }
1206
1207                 if (validate_msg_fds(&ack_msg, 0) != 0)
1208                         goto err_mmap;
1209
1210                 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1211                         VHOST_LOG_CONFIG(ERR,
1212                                 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1213                                 ack_msg.request.master);
1214                         goto err_mmap;
1215                 }
1216
1217                 /* Now userfault register and we can use the memory */
1218                 for (i = 0; i < memory->nregions; i++) {
1219 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1220                         reg = &dev->mem->regions[i];
1221                         struct uffdio_register reg_struct;
1222
1223                         /*
1224                          * Let's register all the mmap'ed area to ensure
1225                          * alignment on page boundary.
1226                          */
1227                         reg_struct.range.start =
1228                                 (uint64_t)(uintptr_t)reg->mmap_addr;
1229                         reg_struct.range.len = reg->mmap_size;
1230                         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1231
1232                         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1233                                                 &reg_struct)) {
1234                                 VHOST_LOG_CONFIG(ERR,
1235                                         "Failed to register ufd for region %d: (ufd = %d) %s\n",
1236                                         i, dev->postcopy_ufd,
1237                                         strerror(errno));
1238                                 goto err_mmap;
1239                         }
1240                         VHOST_LOG_CONFIG(INFO,
1241                                 "\t userfaultfd registered for range : "
1242                                 "%" PRIx64 " - %" PRIx64 "\n",
1243                                 (uint64_t)reg_struct.range.start,
1244                                 (uint64_t)reg_struct.range.start +
1245                                 (uint64_t)reg_struct.range.len - 1);
1246 #else
1247                         goto err_mmap;
1248 #endif
1249                 }
1250         }
1251
1252         for (i = 0; i < dev->nr_vring; i++) {
1253                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1254
1255                 if (vq->desc || vq->avail || vq->used) {
1256                         /*
1257                          * If the memory table got updated, the ring addresses
1258                          * need to be translated again as virtual addresses have
1259                          * changed.
1260                          */
1261                         vring_invalidate(dev, vq);
1262
1263                         dev = translate_ring_addresses(dev, i);
1264                         if (!dev) {
1265                                 dev = *pdev;
1266                                 goto err_mmap;
1267                         }
1268
1269                         *pdev = dev;
1270                 }
1271         }
1272
1273         dump_guest_pages(dev);
1274
1275         return RTE_VHOST_MSG_RESULT_OK;
1276
1277 err_mmap:
1278         free_mem_region(dev);
1279         rte_free(dev->mem);
1280         dev->mem = NULL;
1281         return RTE_VHOST_MSG_RESULT_ERR;
1282 }
1283
1284 static bool
1285 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1286 {
1287         bool rings_ok;
1288
1289         if (!vq)
1290                 return false;
1291
1292         if (vq_is_packed(dev))
1293                 rings_ok = vq->desc_packed && vq->driver_event &&
1294                         vq->device_event;
1295         else
1296                 rings_ok = vq->desc && vq->avail && vq->used;
1297
1298         return rings_ok &&
1299                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1300                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1301 }
1302
1303 static int
1304 virtio_is_ready(struct virtio_net *dev)
1305 {
1306         struct vhost_virtqueue *vq;
1307         uint32_t i;
1308
1309         if (dev->nr_vring == 0)
1310                 return 0;
1311
1312         for (i = 0; i < dev->nr_vring; i++) {
1313                 vq = dev->virtqueue[i];
1314
1315                 if (!vq_is_ready(dev, vq))
1316                         return 0;
1317         }
1318
1319         VHOST_LOG_CONFIG(INFO,
1320                 "virtio is now ready for processing.\n");
1321         return 1;
1322 }
1323
1324 static void *
1325 inflight_mem_alloc(const char *name, size_t size, int *fd)
1326 {
1327         void *ptr;
1328         int mfd = -1;
1329         char fname[20] = "/tmp/memfd-XXXXXX";
1330
1331         *fd = -1;
1332 #ifdef MEMFD_SUPPORTED
1333         mfd = memfd_create(name, MFD_CLOEXEC);
1334 #else
1335         RTE_SET_USED(name);
1336 #endif
1337         if (mfd == -1) {
1338                 mfd = mkstemp(fname);
1339                 if (mfd == -1) {
1340                         VHOST_LOG_CONFIG(ERR,
1341                                 "failed to get inflight buffer fd\n");
1342                         return NULL;
1343                 }
1344
1345                 unlink(fname);
1346         }
1347
1348         if (ftruncate(mfd, size) == -1) {
1349                 VHOST_LOG_CONFIG(ERR,
1350                         "failed to alloc inflight buffer\n");
1351                 close(mfd);
1352                 return NULL;
1353         }
1354
1355         ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd, 0);
1356         if (ptr == MAP_FAILED) {
1357                 VHOST_LOG_CONFIG(ERR,
1358                         "failed to mmap inflight buffer\n");
1359                 close(mfd);
1360                 return NULL;
1361         }
1362
1363         *fd = mfd;
1364         return ptr;
1365 }
1366
1367 static uint32_t
1368 get_pervq_shm_size_split(uint16_t queue_size)
1369 {
1370         return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_split) *
1371                                   queue_size + sizeof(uint64_t) +
1372                                   sizeof(uint16_t) * 4, INFLIGHT_ALIGNMENT);
1373 }
1374
1375 static uint32_t
1376 get_pervq_shm_size_packed(uint16_t queue_size)
1377 {
1378         return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_packed)
1379                                   * queue_size + sizeof(uint64_t) +
1380                                   sizeof(uint16_t) * 6 + sizeof(uint8_t) * 9,
1381                                   INFLIGHT_ALIGNMENT);
1382 }
1383
1384 static int
1385 vhost_user_get_inflight_fd(struct virtio_net **pdev,
1386                            VhostUserMsg *msg,
1387                            int main_fd __rte_unused)
1388 {
1389         struct rte_vhost_inflight_info_packed *inflight_packed;
1390         uint64_t pervq_inflight_size, mmap_size;
1391         uint16_t num_queues, queue_size;
1392         struct virtio_net *dev = *pdev;
1393         int fd, i, j;
1394         void *addr;
1395
1396         if (msg->size != sizeof(msg->payload.inflight)) {
1397                 VHOST_LOG_CONFIG(ERR,
1398                         "invalid get_inflight_fd message size is %d\n",
1399                         msg->size);
1400                 return RTE_VHOST_MSG_RESULT_ERR;
1401         }
1402
1403         if (dev->inflight_info == NULL) {
1404                 dev->inflight_info = calloc(1,
1405                                             sizeof(struct inflight_mem_info));
1406                 if (!dev->inflight_info) {
1407                         VHOST_LOG_CONFIG(ERR,
1408                                 "failed to alloc dev inflight area\n");
1409                         return RTE_VHOST_MSG_RESULT_ERR;
1410                 }
1411         }
1412
1413         num_queues = msg->payload.inflight.num_queues;
1414         queue_size = msg->payload.inflight.queue_size;
1415
1416         VHOST_LOG_CONFIG(INFO, "get_inflight_fd num_queues: %u\n",
1417                 msg->payload.inflight.num_queues);
1418         VHOST_LOG_CONFIG(INFO, "get_inflight_fd queue_size: %u\n",
1419                 msg->payload.inflight.queue_size);
1420
1421         if (vq_is_packed(dev))
1422                 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1423         else
1424                 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1425
1426         mmap_size = num_queues * pervq_inflight_size;
1427         addr = inflight_mem_alloc("vhost-inflight", mmap_size, &fd);
1428         if (!addr) {
1429                 VHOST_LOG_CONFIG(ERR,
1430                         "failed to alloc vhost inflight area\n");
1431                         msg->payload.inflight.mmap_size = 0;
1432                 return RTE_VHOST_MSG_RESULT_ERR;
1433         }
1434         memset(addr, 0, mmap_size);
1435
1436         dev->inflight_info->addr = addr;
1437         dev->inflight_info->size = msg->payload.inflight.mmap_size = mmap_size;
1438         dev->inflight_info->fd = msg->fds[0] = fd;
1439         msg->payload.inflight.mmap_offset = 0;
1440         msg->fd_num = 1;
1441
1442         if (vq_is_packed(dev)) {
1443                 for (i = 0; i < num_queues; i++) {
1444                         inflight_packed =
1445                                 (struct rte_vhost_inflight_info_packed *)addr;
1446                         inflight_packed->used_wrap_counter = 1;
1447                         inflight_packed->old_used_wrap_counter = 1;
1448                         for (j = 0; j < queue_size; j++)
1449                                 inflight_packed->desc[j].next = j + 1;
1450                         addr = (void *)((char *)addr + pervq_inflight_size);
1451                 }
1452         }
1453
1454         VHOST_LOG_CONFIG(INFO,
1455                 "send inflight mmap_size: %"PRIu64"\n",
1456                 msg->payload.inflight.mmap_size);
1457         VHOST_LOG_CONFIG(INFO,
1458                 "send inflight mmap_offset: %"PRIu64"\n",
1459                 msg->payload.inflight.mmap_offset);
1460         VHOST_LOG_CONFIG(INFO,
1461                 "send inflight fd: %d\n", msg->fds[0]);
1462
1463         return RTE_VHOST_MSG_RESULT_REPLY;
1464 }
1465
1466 static int
1467 vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
1468                            int main_fd __rte_unused)
1469 {
1470         uint64_t mmap_size, mmap_offset;
1471         uint16_t num_queues, queue_size;
1472         struct virtio_net *dev = *pdev;
1473         uint32_t pervq_inflight_size;
1474         struct vhost_virtqueue *vq;
1475         void *addr;
1476         int fd, i;
1477
1478         fd = msg->fds[0];
1479         if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
1480                 VHOST_LOG_CONFIG(ERR,
1481                         "invalid set_inflight_fd message size is %d,fd is %d\n",
1482                         msg->size, fd);
1483                 return RTE_VHOST_MSG_RESULT_ERR;
1484         }
1485
1486         mmap_size = msg->payload.inflight.mmap_size;
1487         mmap_offset = msg->payload.inflight.mmap_offset;
1488         num_queues = msg->payload.inflight.num_queues;
1489         queue_size = msg->payload.inflight.queue_size;
1490
1491         if (vq_is_packed(dev))
1492                 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1493         else
1494                 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1495
1496         VHOST_LOG_CONFIG(INFO,
1497                 "set_inflight_fd mmap_size: %"PRIu64"\n", mmap_size);
1498         VHOST_LOG_CONFIG(INFO,
1499                 "set_inflight_fd mmap_offset: %"PRIu64"\n", mmap_offset);
1500         VHOST_LOG_CONFIG(INFO,
1501                 "set_inflight_fd num_queues: %u\n", num_queues);
1502         VHOST_LOG_CONFIG(INFO,
1503                 "set_inflight_fd queue_size: %u\n", queue_size);
1504         VHOST_LOG_CONFIG(INFO,
1505                 "set_inflight_fd fd: %d\n", fd);
1506         VHOST_LOG_CONFIG(INFO,
1507                 "set_inflight_fd pervq_inflight_size: %d\n",
1508                 pervq_inflight_size);
1509
1510         if (!dev->inflight_info) {
1511                 dev->inflight_info = calloc(1,
1512                                             sizeof(struct inflight_mem_info));
1513                 if (dev->inflight_info == NULL) {
1514                         VHOST_LOG_CONFIG(ERR,
1515                                 "failed to alloc dev inflight area\n");
1516                         return RTE_VHOST_MSG_RESULT_ERR;
1517                 }
1518         }
1519
1520         if (dev->inflight_info->addr)
1521                 munmap(dev->inflight_info->addr, dev->inflight_info->size);
1522
1523         addr = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
1524                     fd, mmap_offset);
1525         if (addr == MAP_FAILED) {
1526                 VHOST_LOG_CONFIG(ERR, "failed to mmap share memory.\n");
1527                 return RTE_VHOST_MSG_RESULT_ERR;
1528         }
1529
1530         if (dev->inflight_info->fd)
1531                 close(dev->inflight_info->fd);
1532
1533         dev->inflight_info->fd = fd;
1534         dev->inflight_info->addr = addr;
1535         dev->inflight_info->size = mmap_size;
1536
1537         for (i = 0; i < num_queues; i++) {
1538                 vq = dev->virtqueue[i];
1539                 if (vq_is_packed(dev)) {
1540                         vq->inflight_packed = addr;
1541                         vq->inflight_packed->desc_num = queue_size;
1542                 } else {
1543                         vq->inflight_split = addr;
1544                         vq->inflight_split->desc_num = queue_size;
1545                 }
1546                 addr = (void *)((char *)addr + pervq_inflight_size);
1547         }
1548
1549         return RTE_VHOST_MSG_RESULT_OK;
1550 }
1551
1552 static int
1553 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1554                         int main_fd __rte_unused)
1555 {
1556         struct virtio_net *dev = *pdev;
1557         struct vhost_vring_file file;
1558         struct vhost_virtqueue *vq;
1559         int expected_fds;
1560
1561         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1562         if (validate_msg_fds(msg, expected_fds) != 0)
1563                 return RTE_VHOST_MSG_RESULT_ERR;
1564
1565         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1566         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1567                 file.fd = VIRTIO_INVALID_EVENTFD;
1568         else
1569                 file.fd = msg->fds[0];
1570         VHOST_LOG_CONFIG(INFO,
1571                 "vring call idx:%d file:%d\n", file.index, file.fd);
1572
1573         vq = dev->virtqueue[file.index];
1574         if (vq->callfd >= 0)
1575                 close(vq->callfd);
1576
1577         vq->callfd = file.fd;
1578
1579         return RTE_VHOST_MSG_RESULT_OK;
1580 }
1581
1582 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1583                         struct VhostUserMsg *msg,
1584                         int main_fd __rte_unused)
1585 {
1586         int expected_fds;
1587
1588         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1589         if (validate_msg_fds(msg, expected_fds) != 0)
1590                 return RTE_VHOST_MSG_RESULT_ERR;
1591
1592         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1593                 close(msg->fds[0]);
1594         VHOST_LOG_CONFIG(INFO, "not implemented\n");
1595
1596         return RTE_VHOST_MSG_RESULT_OK;
1597 }
1598
1599 static int
1600 resubmit_desc_compare(const void *a, const void *b)
1601 {
1602         const struct rte_vhost_resubmit_desc *desc0 = a;
1603         const struct rte_vhost_resubmit_desc *desc1 = b;
1604
1605         if (desc1->counter > desc0->counter)
1606                 return 1;
1607
1608         return -1;
1609 }
1610
1611 static int
1612 vhost_check_queue_inflights_split(struct virtio_net *dev,
1613                                   struct vhost_virtqueue *vq)
1614 {
1615         uint16_t i;
1616         uint16_t resubmit_num = 0, last_io, num;
1617         struct vring_used *used = vq->used;
1618         struct rte_vhost_resubmit_info *resubmit;
1619         struct rte_vhost_inflight_info_split *inflight_split;
1620
1621         if (!(dev->protocol_features &
1622             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1623                 return RTE_VHOST_MSG_RESULT_OK;
1624
1625         /* The frontend may still not support the inflight feature
1626          * although we negotiate the protocol feature.
1627          */
1628         if ((!vq->inflight_split))
1629                 return RTE_VHOST_MSG_RESULT_OK;
1630
1631         if (!vq->inflight_split->version) {
1632                 vq->inflight_split->version = INFLIGHT_VERSION;
1633                 return RTE_VHOST_MSG_RESULT_OK;
1634         }
1635
1636         if (vq->resubmit_inflight)
1637                 return RTE_VHOST_MSG_RESULT_OK;
1638
1639         inflight_split = vq->inflight_split;
1640         vq->global_counter = 0;
1641         last_io = inflight_split->last_inflight_io;
1642
1643         if (inflight_split->used_idx != used->idx) {
1644                 inflight_split->desc[last_io].inflight = 0;
1645                 rte_smp_mb();
1646                 inflight_split->used_idx = used->idx;
1647         }
1648
1649         for (i = 0; i < inflight_split->desc_num; i++) {
1650                 if (inflight_split->desc[i].inflight == 1)
1651                         resubmit_num++;
1652         }
1653
1654         vq->last_avail_idx += resubmit_num;
1655
1656         if (resubmit_num) {
1657                 resubmit  = calloc(1, sizeof(struct rte_vhost_resubmit_info));
1658                 if (!resubmit) {
1659                         VHOST_LOG_CONFIG(ERR,
1660                                 "failed to allocate memory for resubmit info.\n");
1661                         return RTE_VHOST_MSG_RESULT_ERR;
1662                 }
1663
1664                 resubmit->resubmit_list = calloc(resubmit_num,
1665                         sizeof(struct rte_vhost_resubmit_desc));
1666                 if (!resubmit->resubmit_list) {
1667                         VHOST_LOG_CONFIG(ERR,
1668                                 "failed to allocate memory for inflight desc.\n");
1669                         free(resubmit);
1670                         return RTE_VHOST_MSG_RESULT_ERR;
1671                 }
1672
1673                 num = 0;
1674                 for (i = 0; i < vq->inflight_split->desc_num; i++) {
1675                         if (vq->inflight_split->desc[i].inflight == 1) {
1676                                 resubmit->resubmit_list[num].index = i;
1677                                 resubmit->resubmit_list[num].counter =
1678                                         inflight_split->desc[i].counter;
1679                                 num++;
1680                         }
1681                 }
1682                 resubmit->resubmit_num = num;
1683
1684                 if (resubmit->resubmit_num > 1)
1685                         qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1686                               sizeof(struct rte_vhost_resubmit_desc),
1687                               resubmit_desc_compare);
1688
1689                 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1690                 vq->resubmit_inflight = resubmit;
1691         }
1692
1693         return RTE_VHOST_MSG_RESULT_OK;
1694 }
1695
1696 static int
1697 vhost_check_queue_inflights_packed(struct virtio_net *dev,
1698                                    struct vhost_virtqueue *vq)
1699 {
1700         uint16_t i;
1701         uint16_t resubmit_num = 0, old_used_idx, num;
1702         struct rte_vhost_resubmit_info *resubmit;
1703         struct rte_vhost_inflight_info_packed *inflight_packed;
1704
1705         if (!(dev->protocol_features &
1706             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1707                 return RTE_VHOST_MSG_RESULT_OK;
1708
1709         /* The frontend may still not support the inflight feature
1710          * although we negotiate the protocol feature.
1711          */
1712         if ((!vq->inflight_packed))
1713                 return RTE_VHOST_MSG_RESULT_OK;
1714
1715         if (!vq->inflight_packed->version) {
1716                 vq->inflight_packed->version = INFLIGHT_VERSION;
1717                 return RTE_VHOST_MSG_RESULT_OK;
1718         }
1719
1720         if (vq->resubmit_inflight)
1721                 return RTE_VHOST_MSG_RESULT_OK;
1722
1723         inflight_packed = vq->inflight_packed;
1724         vq->global_counter = 0;
1725         old_used_idx = inflight_packed->old_used_idx;
1726
1727         if (inflight_packed->used_idx != old_used_idx) {
1728                 if (inflight_packed->desc[old_used_idx].inflight == 0) {
1729                         inflight_packed->old_used_idx =
1730                                 inflight_packed->used_idx;
1731                         inflight_packed->old_used_wrap_counter =
1732                                 inflight_packed->used_wrap_counter;
1733                         inflight_packed->old_free_head =
1734                                 inflight_packed->free_head;
1735                 } else {
1736                         inflight_packed->used_idx =
1737                                 inflight_packed->old_used_idx;
1738                         inflight_packed->used_wrap_counter =
1739                                 inflight_packed->old_used_wrap_counter;
1740                         inflight_packed->free_head =
1741                                 inflight_packed->old_free_head;
1742                 }
1743         }
1744
1745         for (i = 0; i < inflight_packed->desc_num; i++) {
1746                 if (inflight_packed->desc[i].inflight == 1)
1747                         resubmit_num++;
1748         }
1749
1750         if (resubmit_num) {
1751                 resubmit = calloc(1, sizeof(struct rte_vhost_resubmit_info));
1752                 if (resubmit == NULL) {
1753                         VHOST_LOG_CONFIG(ERR,
1754                                 "failed to allocate memory for resubmit info.\n");
1755                         return RTE_VHOST_MSG_RESULT_ERR;
1756                 }
1757
1758                 resubmit->resubmit_list = calloc(resubmit_num,
1759                         sizeof(struct rte_vhost_resubmit_desc));
1760                 if (resubmit->resubmit_list == NULL) {
1761                         VHOST_LOG_CONFIG(ERR,
1762                                 "failed to allocate memory for resubmit desc.\n");
1763                         free(resubmit);
1764                         return RTE_VHOST_MSG_RESULT_ERR;
1765                 }
1766
1767                 num = 0;
1768                 for (i = 0; i < inflight_packed->desc_num; i++) {
1769                         if (vq->inflight_packed->desc[i].inflight == 1) {
1770                                 resubmit->resubmit_list[num].index = i;
1771                                 resubmit->resubmit_list[num].counter =
1772                                         inflight_packed->desc[i].counter;
1773                                 num++;
1774                         }
1775                 }
1776                 resubmit->resubmit_num = num;
1777
1778                 if (resubmit->resubmit_num > 1)
1779                         qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1780                               sizeof(struct rte_vhost_resubmit_desc),
1781                               resubmit_desc_compare);
1782
1783                 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1784                 vq->resubmit_inflight = resubmit;
1785         }
1786
1787         return RTE_VHOST_MSG_RESULT_OK;
1788 }
1789
1790 static int
1791 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1792                         int main_fd __rte_unused)
1793 {
1794         struct virtio_net *dev = *pdev;
1795         struct vhost_vring_file file;
1796         struct vhost_virtqueue *vq;
1797         int expected_fds;
1798
1799         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1800         if (validate_msg_fds(msg, expected_fds) != 0)
1801                 return RTE_VHOST_MSG_RESULT_ERR;
1802
1803         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1804         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1805                 file.fd = VIRTIO_INVALID_EVENTFD;
1806         else
1807                 file.fd = msg->fds[0];
1808         VHOST_LOG_CONFIG(INFO,
1809                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1810
1811         /* Interpret ring addresses only when ring is started. */
1812         dev = translate_ring_addresses(dev, file.index);
1813         if (!dev)
1814                 return RTE_VHOST_MSG_RESULT_ERR;
1815
1816         *pdev = dev;
1817
1818         vq = dev->virtqueue[file.index];
1819
1820         /*
1821          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1822          * the ring starts already enabled. Otherwise, it is enabled via
1823          * the SET_VRING_ENABLE message.
1824          */
1825         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
1826                 vq->enabled = 1;
1827                 if (dev->notify_ops->vring_state_changed)
1828                         dev->notify_ops->vring_state_changed(
1829                                 dev->vid, file.index, 1);
1830         }
1831
1832         if (vq->kickfd >= 0)
1833                 close(vq->kickfd);
1834         vq->kickfd = file.fd;
1835
1836         if (vq_is_packed(dev)) {
1837                 if (vhost_check_queue_inflights_packed(dev, vq)) {
1838                         VHOST_LOG_CONFIG(ERR,
1839                                 "failed to inflights for vq: %d\n", file.index);
1840                         return RTE_VHOST_MSG_RESULT_ERR;
1841                 }
1842         } else {
1843                 if (vhost_check_queue_inflights_split(dev, vq)) {
1844                         VHOST_LOG_CONFIG(ERR,
1845                                 "failed to inflights for vq: %d\n", file.index);
1846                         return RTE_VHOST_MSG_RESULT_ERR;
1847                 }
1848         }
1849
1850         return RTE_VHOST_MSG_RESULT_OK;
1851 }
1852
1853 static void
1854 free_zmbufs(struct vhost_virtqueue *vq)
1855 {
1856         drain_zmbuf_list(vq);
1857
1858         rte_free(vq->zmbufs);
1859 }
1860
1861 /*
1862  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1863  */
1864 static int
1865 vhost_user_get_vring_base(struct virtio_net **pdev,
1866                         struct VhostUserMsg *msg,
1867                         int main_fd __rte_unused)
1868 {
1869         struct virtio_net *dev = *pdev;
1870         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1871         uint64_t val;
1872
1873         if (validate_msg_fds(msg, 0) != 0)
1874                 return RTE_VHOST_MSG_RESULT_ERR;
1875
1876         /* We have to stop the queue (virtio) if it is running. */
1877         vhost_destroy_device_notify(dev);
1878
1879         dev->flags &= ~VIRTIO_DEV_READY;
1880         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1881
1882         /* Here we are safe to get the indexes */
1883         if (vq_is_packed(dev)) {
1884                 /*
1885                  * Bit[0:14]: avail index
1886                  * Bit[15]: avail wrap counter
1887                  */
1888                 val = vq->last_avail_idx & 0x7fff;
1889                 val |= vq->avail_wrap_counter << 15;
1890                 msg->payload.state.num = val;
1891         } else {
1892                 msg->payload.state.num = vq->last_avail_idx;
1893         }
1894
1895         VHOST_LOG_CONFIG(INFO,
1896                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1897                 msg->payload.state.num);
1898         /*
1899          * Based on current qemu vhost-user implementation, this message is
1900          * sent and only sent in vhost_vring_stop.
1901          * TODO: cleanup the vring, it isn't usable since here.
1902          */
1903         if (vq->kickfd >= 0)
1904                 close(vq->kickfd);
1905
1906         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1907
1908         if (vq->callfd >= 0)
1909                 close(vq->callfd);
1910
1911         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1912
1913         vq->signalled_used_valid = false;
1914
1915         if (dev->dequeue_zero_copy)
1916                 free_zmbufs(vq);
1917         if (vq_is_packed(dev)) {
1918                 rte_free(vq->shadow_used_packed);
1919                 vq->shadow_used_packed = NULL;
1920         } else {
1921                 rte_free(vq->shadow_used_split);
1922                 vq->shadow_used_split = NULL;
1923         }
1924
1925         rte_free(vq->batch_copy_elems);
1926         vq->batch_copy_elems = NULL;
1927
1928         msg->size = sizeof(msg->payload.state);
1929         msg->fd_num = 0;
1930
1931         vring_invalidate(dev, vq);
1932
1933         return RTE_VHOST_MSG_RESULT_REPLY;
1934 }
1935
1936 /*
1937  * when virtio queues are ready to work, qemu will send us to
1938  * enable the virtio queue pair.
1939  */
1940 static int
1941 vhost_user_set_vring_enable(struct virtio_net **pdev,
1942                         struct VhostUserMsg *msg,
1943                         int main_fd __rte_unused)
1944 {
1945         struct virtio_net *dev = *pdev;
1946         int enable = (int)msg->payload.state.num;
1947         int index = (int)msg->payload.state.index;
1948         struct rte_vdpa_device *vdpa_dev;
1949         int did = -1;
1950
1951         if (validate_msg_fds(msg, 0) != 0)
1952                 return RTE_VHOST_MSG_RESULT_ERR;
1953
1954         VHOST_LOG_CONFIG(INFO,
1955                 "set queue enable: %d to qp idx: %d\n",
1956                 enable, index);
1957
1958         did = dev->vdpa_dev_id;
1959         vdpa_dev = rte_vdpa_get_device(did);
1960         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1961                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1962
1963         if (dev->notify_ops->vring_state_changed)
1964                 dev->notify_ops->vring_state_changed(dev->vid,
1965                                 index, enable);
1966
1967         /* On disable, rings have to be stopped being processed. */
1968         if (!enable && dev->dequeue_zero_copy)
1969                 drain_zmbuf_list(dev->virtqueue[index]);
1970
1971         dev->virtqueue[index]->enabled = enable;
1972
1973         return RTE_VHOST_MSG_RESULT_OK;
1974 }
1975
1976 static int
1977 vhost_user_get_protocol_features(struct virtio_net **pdev,
1978                         struct VhostUserMsg *msg,
1979                         int main_fd __rte_unused)
1980 {
1981         struct virtio_net *dev = *pdev;
1982         uint64_t features, protocol_features;
1983
1984         if (validate_msg_fds(msg, 0) != 0)
1985                 return RTE_VHOST_MSG_RESULT_ERR;
1986
1987         rte_vhost_driver_get_features(dev->ifname, &features);
1988         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1989
1990         /*
1991          * REPLY_ACK protocol feature is only mandatory for now
1992          * for IOMMU feature. If IOMMU is explicitly disabled by the
1993          * application, disable also REPLY_ACK feature for older buggy
1994          * Qemu versions (from v2.7.0 to v2.9.0).
1995          */
1996         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1997                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1998
1999         msg->payload.u64 = protocol_features;
2000         msg->size = sizeof(msg->payload.u64);
2001         msg->fd_num = 0;
2002
2003         return RTE_VHOST_MSG_RESULT_REPLY;
2004 }
2005
2006 static int
2007 vhost_user_set_protocol_features(struct virtio_net **pdev,
2008                         struct VhostUserMsg *msg,
2009                         int main_fd __rte_unused)
2010 {
2011         struct virtio_net *dev = *pdev;
2012         uint64_t protocol_features = msg->payload.u64;
2013         uint64_t slave_protocol_features = 0;
2014
2015         if (validate_msg_fds(msg, 0) != 0)
2016                 return RTE_VHOST_MSG_RESULT_ERR;
2017
2018         rte_vhost_driver_get_protocol_features(dev->ifname,
2019                         &slave_protocol_features);
2020         if (protocol_features & ~slave_protocol_features) {
2021                 VHOST_LOG_CONFIG(ERR,
2022                         "(%d) received invalid protocol features.\n",
2023                         dev->vid);
2024                 return RTE_VHOST_MSG_RESULT_ERR;
2025         }
2026
2027         dev->protocol_features = protocol_features;
2028         VHOST_LOG_CONFIG(INFO,
2029                 "negotiated Vhost-user protocol features: 0x%" PRIx64 "\n",
2030                 dev->protocol_features);
2031
2032         return RTE_VHOST_MSG_RESULT_OK;
2033 }
2034
2035 static int
2036 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
2037                         int main_fd __rte_unused)
2038 {
2039         struct virtio_net *dev = *pdev;
2040         int fd = msg->fds[0];
2041         uint64_t size, off;
2042         void *addr;
2043
2044         if (validate_msg_fds(msg, 1) != 0)
2045                 return RTE_VHOST_MSG_RESULT_ERR;
2046
2047         if (fd < 0) {
2048                 VHOST_LOG_CONFIG(ERR, "invalid log fd: %d\n", fd);
2049                 return RTE_VHOST_MSG_RESULT_ERR;
2050         }
2051
2052         if (msg->size != sizeof(VhostUserLog)) {
2053                 VHOST_LOG_CONFIG(ERR,
2054                         "invalid log base msg size: %"PRId32" != %d\n",
2055                         msg->size, (int)sizeof(VhostUserLog));
2056                 return RTE_VHOST_MSG_RESULT_ERR;
2057         }
2058
2059         size = msg->payload.log.mmap_size;
2060         off  = msg->payload.log.mmap_offset;
2061
2062         /* Don't allow mmap_offset to point outside the mmap region */
2063         if (off > size) {
2064                 VHOST_LOG_CONFIG(ERR,
2065                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
2066                         off, size);
2067                 return RTE_VHOST_MSG_RESULT_ERR;
2068         }
2069
2070         VHOST_LOG_CONFIG(INFO,
2071                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
2072                 size, off);
2073
2074         /*
2075          * mmap from 0 to workaround a hugepage mmap bug: mmap will
2076          * fail when offset is not page size aligned.
2077          */
2078         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
2079         close(fd);
2080         if (addr == MAP_FAILED) {
2081                 VHOST_LOG_CONFIG(ERR, "mmap log base failed!\n");
2082                 return RTE_VHOST_MSG_RESULT_ERR;
2083         }
2084
2085         /*
2086          * Free previously mapped log memory on occasionally
2087          * multiple VHOST_USER_SET_LOG_BASE.
2088          */
2089         if (dev->log_addr) {
2090                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
2091         }
2092         dev->log_addr = (uint64_t)(uintptr_t)addr;
2093         dev->log_base = dev->log_addr + off;
2094         dev->log_size = size;
2095
2096         /*
2097          * The spec is not clear about it (yet), but QEMU doesn't expect
2098          * any payload in the reply.
2099          */
2100         msg->size = 0;
2101         msg->fd_num = 0;
2102
2103         return RTE_VHOST_MSG_RESULT_REPLY;
2104 }
2105
2106 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
2107                         struct VhostUserMsg *msg,
2108                         int main_fd __rte_unused)
2109 {
2110         if (validate_msg_fds(msg, 1) != 0)
2111                 return RTE_VHOST_MSG_RESULT_ERR;
2112
2113         close(msg->fds[0]);
2114         VHOST_LOG_CONFIG(INFO, "not implemented.\n");
2115
2116         return RTE_VHOST_MSG_RESULT_OK;
2117 }
2118
2119 /*
2120  * An rarp packet is constructed and broadcasted to notify switches about
2121  * the new location of the migrated VM, so that packets from outside will
2122  * not be lost after migration.
2123  *
2124  * However, we don't actually "send" a rarp packet here, instead, we set
2125  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
2126  */
2127 static int
2128 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
2129                         int main_fd __rte_unused)
2130 {
2131         struct virtio_net *dev = *pdev;
2132         uint8_t *mac = (uint8_t *)&msg->payload.u64;
2133         struct rte_vdpa_device *vdpa_dev;
2134         int did = -1;
2135
2136         if (validate_msg_fds(msg, 0) != 0)
2137                 return RTE_VHOST_MSG_RESULT_ERR;
2138
2139         VHOST_LOG_CONFIG(DEBUG,
2140                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
2141                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
2142         memcpy(dev->mac.addr_bytes, mac, 6);
2143
2144         /*
2145          * Set the flag to inject a RARP broadcast packet at
2146          * rte_vhost_dequeue_burst().
2147          *
2148          * rte_smp_wmb() is for making sure the mac is copied
2149          * before the flag is set.
2150          */
2151         rte_smp_wmb();
2152         rte_atomic16_set(&dev->broadcast_rarp, 1);
2153         did = dev->vdpa_dev_id;
2154         vdpa_dev = rte_vdpa_get_device(did);
2155         if (vdpa_dev && vdpa_dev->ops->migration_done)
2156                 vdpa_dev->ops->migration_done(dev->vid);
2157
2158         return RTE_VHOST_MSG_RESULT_OK;
2159 }
2160
2161 static int
2162 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
2163                         int main_fd __rte_unused)
2164 {
2165         struct virtio_net *dev = *pdev;
2166
2167         if (validate_msg_fds(msg, 0) != 0)
2168                 return RTE_VHOST_MSG_RESULT_ERR;
2169
2170         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
2171                         msg->payload.u64 > VIRTIO_MAX_MTU) {
2172                 VHOST_LOG_CONFIG(ERR, "Invalid MTU size (%"PRIu64")\n",
2173                                 msg->payload.u64);
2174
2175                 return RTE_VHOST_MSG_RESULT_ERR;
2176         }
2177
2178         dev->mtu = msg->payload.u64;
2179
2180         return RTE_VHOST_MSG_RESULT_OK;
2181 }
2182
2183 static int
2184 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
2185                         int main_fd __rte_unused)
2186 {
2187         struct virtio_net *dev = *pdev;
2188         int fd = msg->fds[0];
2189
2190         if (validate_msg_fds(msg, 1) != 0)
2191                 return RTE_VHOST_MSG_RESULT_ERR;
2192
2193         if (fd < 0) {
2194                 VHOST_LOG_CONFIG(ERR,
2195                                 "Invalid file descriptor for slave channel (%d)\n",
2196                                 fd);
2197                 return RTE_VHOST_MSG_RESULT_ERR;
2198         }
2199
2200         if (dev->slave_req_fd >= 0)
2201                 close(dev->slave_req_fd);
2202
2203         dev->slave_req_fd = fd;
2204
2205         return RTE_VHOST_MSG_RESULT_OK;
2206 }
2207
2208 static int
2209 is_vring_iotlb_split(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2210 {
2211         struct vhost_vring_addr *ra;
2212         uint64_t start, end, len;
2213
2214         start = imsg->iova;
2215         end = start + imsg->size;
2216
2217         ra = &vq->ring_addrs;
2218         len = sizeof(struct vring_desc) * vq->size;
2219         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2220                 return 1;
2221
2222         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
2223         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2224                 return 1;
2225
2226         len = sizeof(struct vring_used) +
2227                sizeof(struct vring_used_elem) * vq->size;
2228         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2229                 return 1;
2230
2231         if (ra->flags & (1 << VHOST_VRING_F_LOG)) {
2232                 len = sizeof(uint64_t);
2233                 if (ra->log_guest_addr < end &&
2234                     (ra->log_guest_addr + len) > start)
2235                         return 1;
2236         }
2237
2238         return 0;
2239 }
2240
2241 static int
2242 is_vring_iotlb_packed(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2243 {
2244         struct vhost_vring_addr *ra;
2245         uint64_t start, end, len;
2246
2247         start = imsg->iova;
2248         end = start + imsg->size;
2249
2250         ra = &vq->ring_addrs;
2251         len = sizeof(struct vring_packed_desc) * vq->size;
2252         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2253                 return 1;
2254
2255         len = sizeof(struct vring_packed_desc_event);
2256         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2257                 return 1;
2258
2259         len = sizeof(struct vring_packed_desc_event);
2260         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2261                 return 1;
2262
2263         if (ra->flags & (1 << VHOST_VRING_F_LOG)) {
2264                 len = sizeof(uint64_t);
2265                 if (ra->log_guest_addr < end &&
2266                     (ra->log_guest_addr + len) > start)
2267                         return 1;
2268         }
2269
2270         return 0;
2271 }
2272
2273 static int is_vring_iotlb(struct virtio_net *dev,
2274                           struct vhost_virtqueue *vq,
2275                           struct vhost_iotlb_msg *imsg)
2276 {
2277         if (vq_is_packed(dev))
2278                 return is_vring_iotlb_packed(vq, imsg);
2279         else
2280                 return is_vring_iotlb_split(vq, imsg);
2281 }
2282
2283 static int
2284 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
2285                         int main_fd __rte_unused)
2286 {
2287         struct virtio_net *dev = *pdev;
2288         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
2289         uint16_t i;
2290         uint64_t vva, len;
2291
2292         if (validate_msg_fds(msg, 0) != 0)
2293                 return RTE_VHOST_MSG_RESULT_ERR;
2294
2295         switch (imsg->type) {
2296         case VHOST_IOTLB_UPDATE:
2297                 len = imsg->size;
2298                 vva = qva_to_vva(dev, imsg->uaddr, &len);
2299                 if (!vva)
2300                         return RTE_VHOST_MSG_RESULT_ERR;
2301
2302                 for (i = 0; i < dev->nr_vring; i++) {
2303                         struct vhost_virtqueue *vq = dev->virtqueue[i];
2304
2305                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
2306                                         len, imsg->perm);
2307
2308                         if (is_vring_iotlb(dev, vq, imsg))
2309                                 *pdev = dev = translate_ring_addresses(dev, i);
2310                 }
2311                 break;
2312         case VHOST_IOTLB_INVALIDATE:
2313                 for (i = 0; i < dev->nr_vring; i++) {
2314                         struct vhost_virtqueue *vq = dev->virtqueue[i];
2315
2316                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
2317                                         imsg->size);
2318
2319                         if (is_vring_iotlb(dev, vq, imsg))
2320                                 vring_invalidate(dev, vq);
2321                 }
2322                 break;
2323         default:
2324                 VHOST_LOG_CONFIG(ERR, "Invalid IOTLB message type (%d)\n",
2325                                 imsg->type);
2326                 return RTE_VHOST_MSG_RESULT_ERR;
2327         }
2328
2329         return RTE_VHOST_MSG_RESULT_OK;
2330 }
2331
2332 static int
2333 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
2334                         struct VhostUserMsg *msg,
2335                         int main_fd __rte_unused)
2336 {
2337         struct virtio_net *dev = *pdev;
2338 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
2339         struct uffdio_api api_struct;
2340
2341         if (validate_msg_fds(msg, 0) != 0)
2342                 return RTE_VHOST_MSG_RESULT_ERR;
2343
2344         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
2345
2346         if (dev->postcopy_ufd == -1) {
2347                 VHOST_LOG_CONFIG(ERR, "Userfaultfd not available: %s\n",
2348                         strerror(errno));
2349                 return RTE_VHOST_MSG_RESULT_ERR;
2350         }
2351         api_struct.api = UFFD_API;
2352         api_struct.features = 0;
2353         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
2354                 VHOST_LOG_CONFIG(ERR, "UFFDIO_API ioctl failure: %s\n",
2355                         strerror(errno));
2356                 close(dev->postcopy_ufd);
2357                 dev->postcopy_ufd = -1;
2358                 return RTE_VHOST_MSG_RESULT_ERR;
2359         }
2360         msg->fds[0] = dev->postcopy_ufd;
2361         msg->fd_num = 1;
2362
2363         return RTE_VHOST_MSG_RESULT_REPLY;
2364 #else
2365         dev->postcopy_ufd = -1;
2366         msg->fd_num = 0;
2367
2368         return RTE_VHOST_MSG_RESULT_ERR;
2369 #endif
2370 }
2371
2372 static int
2373 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
2374                         struct VhostUserMsg *msg __rte_unused,
2375                         int main_fd __rte_unused)
2376 {
2377         struct virtio_net *dev = *pdev;
2378
2379         if (validate_msg_fds(msg, 0) != 0)
2380                 return RTE_VHOST_MSG_RESULT_ERR;
2381
2382         if (dev->mem && dev->mem->nregions) {
2383                 VHOST_LOG_CONFIG(ERR,
2384                         "Regions already registered at postcopy-listen\n");
2385                 return RTE_VHOST_MSG_RESULT_ERR;
2386         }
2387         dev->postcopy_listening = 1;
2388
2389         return RTE_VHOST_MSG_RESULT_OK;
2390 }
2391
2392 static int
2393 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
2394                         int main_fd __rte_unused)
2395 {
2396         struct virtio_net *dev = *pdev;
2397
2398         if (validate_msg_fds(msg, 0) != 0)
2399                 return RTE_VHOST_MSG_RESULT_ERR;
2400
2401         dev->postcopy_listening = 0;
2402         if (dev->postcopy_ufd >= 0) {
2403                 close(dev->postcopy_ufd);
2404                 dev->postcopy_ufd = -1;
2405         }
2406
2407         msg->payload.u64 = 0;
2408         msg->size = sizeof(msg->payload.u64);
2409         msg->fd_num = 0;
2410
2411         return RTE_VHOST_MSG_RESULT_REPLY;
2412 }
2413
2414 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
2415                                         struct VhostUserMsg *msg,
2416                                         int main_fd);
2417 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
2418         [VHOST_USER_NONE] = NULL,
2419         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
2420         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
2421         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
2422         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
2423         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
2424         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
2425         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
2426         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
2427         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
2428         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
2429         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
2430         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
2431         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
2432         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
2433         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
2434         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
2435         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
2436         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
2437         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
2438         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
2439         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
2440         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
2441         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
2442         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
2443         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
2444         [VHOST_USER_GET_INFLIGHT_FD] = vhost_user_get_inflight_fd,
2445         [VHOST_USER_SET_INFLIGHT_FD] = vhost_user_set_inflight_fd,
2446 };
2447
2448 /* return bytes# of read on success or negative val on failure. */
2449 static int
2450 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
2451 {
2452         int ret;
2453
2454         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
2455                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
2456         if (ret <= 0) {
2457                 return ret;
2458         } else if (ret != VHOST_USER_HDR_SIZE) {
2459                 VHOST_LOG_CONFIG(ERR, "Unexpected header size read\n");
2460                 close_msg_fds(msg);
2461                 return -1;
2462         }
2463
2464         if (msg->size) {
2465                 if (msg->size > sizeof(msg->payload)) {
2466                         VHOST_LOG_CONFIG(ERR,
2467                                 "invalid msg size: %d\n", msg->size);
2468                         return -1;
2469                 }
2470                 ret = read(sockfd, &msg->payload, msg->size);
2471                 if (ret <= 0)
2472                         return ret;
2473                 if (ret != (int)msg->size) {
2474                         VHOST_LOG_CONFIG(ERR,
2475                                 "read control message failed\n");
2476                         return -1;
2477                 }
2478         }
2479
2480         return ret;
2481 }
2482
2483 static int
2484 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
2485 {
2486         if (!msg)
2487                 return 0;
2488
2489         return send_fd_message(sockfd, (char *)msg,
2490                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
2491 }
2492
2493 static int
2494 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
2495 {
2496         if (!msg)
2497                 return 0;
2498
2499         msg->flags &= ~VHOST_USER_VERSION_MASK;
2500         msg->flags &= ~VHOST_USER_NEED_REPLY;
2501         msg->flags |= VHOST_USER_VERSION;
2502         msg->flags |= VHOST_USER_REPLY_MASK;
2503
2504         return send_vhost_message(sockfd, msg);
2505 }
2506
2507 static int
2508 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
2509 {
2510         int ret;
2511
2512         if (msg->flags & VHOST_USER_NEED_REPLY)
2513                 rte_spinlock_lock(&dev->slave_req_lock);
2514
2515         ret = send_vhost_message(dev->slave_req_fd, msg);
2516         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
2517                 rte_spinlock_unlock(&dev->slave_req_lock);
2518
2519         return ret;
2520 }
2521
2522 /*
2523  * Allocate a queue pair if it hasn't been allocated yet
2524  */
2525 static int
2526 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
2527                         struct VhostUserMsg *msg)
2528 {
2529         uint16_t vring_idx;
2530
2531         switch (msg->request.master) {
2532         case VHOST_USER_SET_VRING_KICK:
2533         case VHOST_USER_SET_VRING_CALL:
2534         case VHOST_USER_SET_VRING_ERR:
2535                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
2536                 break;
2537         case VHOST_USER_SET_VRING_NUM:
2538         case VHOST_USER_SET_VRING_BASE:
2539         case VHOST_USER_SET_VRING_ENABLE:
2540                 vring_idx = msg->payload.state.index;
2541                 break;
2542         case VHOST_USER_SET_VRING_ADDR:
2543                 vring_idx = msg->payload.addr.index;
2544                 break;
2545         default:
2546                 return 0;
2547         }
2548
2549         if (vring_idx >= VHOST_MAX_VRING) {
2550                 VHOST_LOG_CONFIG(ERR,
2551                         "invalid vring index: %u\n", vring_idx);
2552                 return -1;
2553         }
2554
2555         if (dev->virtqueue[vring_idx])
2556                 return 0;
2557
2558         return alloc_vring_queue(dev, vring_idx);
2559 }
2560
2561 static void
2562 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
2563 {
2564         unsigned int i = 0;
2565         unsigned int vq_num = 0;
2566
2567         while (vq_num < dev->nr_vring) {
2568                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2569
2570                 if (vq) {
2571                         rte_spinlock_lock(&vq->access_lock);
2572                         vq_num++;
2573                 }
2574                 i++;
2575         }
2576 }
2577
2578 static void
2579 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
2580 {
2581         unsigned int i = 0;
2582         unsigned int vq_num = 0;
2583
2584         while (vq_num < dev->nr_vring) {
2585                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2586
2587                 if (vq) {
2588                         rte_spinlock_unlock(&vq->access_lock);
2589                         vq_num++;
2590                 }
2591                 i++;
2592         }
2593 }
2594
2595 int
2596 vhost_user_msg_handler(int vid, int fd)
2597 {
2598         struct virtio_net *dev;
2599         struct VhostUserMsg msg;
2600         struct rte_vdpa_device *vdpa_dev;
2601         int did = -1;
2602         int ret;
2603         int unlock_required = 0;
2604         bool handled;
2605         int request;
2606
2607         dev = get_device(vid);
2608         if (dev == NULL)
2609                 return -1;
2610
2611         if (!dev->notify_ops) {
2612                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
2613                 if (!dev->notify_ops) {
2614                         VHOST_LOG_CONFIG(ERR,
2615                                 "failed to get callback ops for driver %s\n",
2616                                 dev->ifname);
2617                         return -1;
2618                 }
2619         }
2620
2621         ret = read_vhost_message(fd, &msg);
2622         if (ret <= 0) {
2623                 if (ret < 0)
2624                         VHOST_LOG_CONFIG(ERR,
2625                                 "vhost read message failed\n");
2626                 else
2627                         VHOST_LOG_CONFIG(INFO,
2628                                 "vhost peer closed\n");
2629
2630                 return -1;
2631         }
2632
2633         ret = 0;
2634         request = msg.request.master;
2635         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
2636                         vhost_message_str[request]) {
2637                 if (request != VHOST_USER_IOTLB_MSG)
2638                         VHOST_LOG_CONFIG(INFO, "read message %s\n",
2639                                 vhost_message_str[request]);
2640                 else
2641                         VHOST_LOG_CONFIG(DEBUG, "read message %s\n",
2642                                 vhost_message_str[request]);
2643         } else {
2644                 VHOST_LOG_CONFIG(DEBUG, "External request %d\n", request);
2645         }
2646
2647         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
2648         if (ret < 0) {
2649                 VHOST_LOG_CONFIG(ERR,
2650                         "failed to alloc queue\n");
2651                 return -1;
2652         }
2653
2654         /*
2655          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
2656          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
2657          * and device is destroyed. destroy_device waits for queues to be
2658          * inactive, so it is safe. Otherwise taking the access_lock
2659          * would cause a dead lock.
2660          */
2661         switch (request) {
2662         case VHOST_USER_SET_FEATURES:
2663         case VHOST_USER_SET_PROTOCOL_FEATURES:
2664         case VHOST_USER_SET_OWNER:
2665         case VHOST_USER_SET_MEM_TABLE:
2666         case VHOST_USER_SET_LOG_BASE:
2667         case VHOST_USER_SET_LOG_FD:
2668         case VHOST_USER_SET_VRING_NUM:
2669         case VHOST_USER_SET_VRING_ADDR:
2670         case VHOST_USER_SET_VRING_BASE:
2671         case VHOST_USER_SET_VRING_KICK:
2672         case VHOST_USER_SET_VRING_CALL:
2673         case VHOST_USER_SET_VRING_ERR:
2674         case VHOST_USER_SET_VRING_ENABLE:
2675         case VHOST_USER_SEND_RARP:
2676         case VHOST_USER_NET_SET_MTU:
2677         case VHOST_USER_SET_SLAVE_REQ_FD:
2678                 vhost_user_lock_all_queue_pairs(dev);
2679                 unlock_required = 1;
2680                 break;
2681         default:
2682                 break;
2683
2684         }
2685
2686         handled = false;
2687         if (dev->extern_ops.pre_msg_handle) {
2688                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
2689                                 (void *)&msg);
2690                 switch (ret) {
2691                 case RTE_VHOST_MSG_RESULT_REPLY:
2692                         send_vhost_reply(fd, &msg);
2693                         /* Fall-through */
2694                 case RTE_VHOST_MSG_RESULT_ERR:
2695                 case RTE_VHOST_MSG_RESULT_OK:
2696                         handled = true;
2697                         goto skip_to_post_handle;
2698                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2699                 default:
2700                         break;
2701                 }
2702         }
2703
2704         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
2705                 if (!vhost_message_handlers[request])
2706                         goto skip_to_post_handle;
2707                 ret = vhost_message_handlers[request](&dev, &msg, fd);
2708
2709                 switch (ret) {
2710                 case RTE_VHOST_MSG_RESULT_ERR:
2711                         VHOST_LOG_CONFIG(ERR,
2712                                 "Processing %s failed.\n",
2713                                 vhost_message_str[request]);
2714                         handled = true;
2715                         break;
2716                 case RTE_VHOST_MSG_RESULT_OK:
2717                         VHOST_LOG_CONFIG(DEBUG,
2718                                 "Processing %s succeeded.\n",
2719                                 vhost_message_str[request]);
2720                         handled = true;
2721                         break;
2722                 case RTE_VHOST_MSG_RESULT_REPLY:
2723                         VHOST_LOG_CONFIG(DEBUG,
2724                                 "Processing %s succeeded and needs reply.\n",
2725                                 vhost_message_str[request]);
2726                         send_vhost_reply(fd, &msg);
2727                         handled = true;
2728                         break;
2729                 default:
2730                         break;
2731                 }
2732         }
2733
2734 skip_to_post_handle:
2735         if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2736                         dev->extern_ops.post_msg_handle) {
2737                 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
2738                                 (void *)&msg);
2739                 switch (ret) {
2740                 case RTE_VHOST_MSG_RESULT_REPLY:
2741                         send_vhost_reply(fd, &msg);
2742                         /* Fall-through */
2743                 case RTE_VHOST_MSG_RESULT_ERR:
2744                 case RTE_VHOST_MSG_RESULT_OK:
2745                         handled = true;
2746                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2747                 default:
2748                         break;
2749                 }
2750         }
2751
2752         if (unlock_required)
2753                 vhost_user_unlock_all_queue_pairs(dev);
2754
2755         /* If message was not handled at this stage, treat it as an error */
2756         if (!handled) {
2757                 VHOST_LOG_CONFIG(ERR,
2758                         "vhost message (req: %d) was not handled.\n", request);
2759                 close_msg_fds(&msg);
2760                 ret = RTE_VHOST_MSG_RESULT_ERR;
2761         }
2762
2763         /*
2764          * If the request required a reply that was already sent,
2765          * this optional reply-ack won't be sent as the
2766          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2767          */
2768         if (msg.flags & VHOST_USER_NEED_REPLY) {
2769                 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2770                 msg.size = sizeof(msg.payload.u64);
2771                 msg.fd_num = 0;
2772                 send_vhost_reply(fd, &msg);
2773         } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2774                 VHOST_LOG_CONFIG(ERR,
2775                         "vhost message handling failed.\n");
2776                 return -1;
2777         }
2778
2779         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2780                 dev->flags |= VIRTIO_DEV_READY;
2781
2782                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2783                         if (dev->dequeue_zero_copy) {
2784                                 VHOST_LOG_CONFIG(INFO,
2785                                                 "dequeue zero copy is enabled\n");
2786                         }
2787
2788                         if (dev->notify_ops->new_device(dev->vid) == 0)
2789                                 dev->flags |= VIRTIO_DEV_RUNNING;
2790                 }
2791         }
2792
2793         did = dev->vdpa_dev_id;
2794         vdpa_dev = rte_vdpa_get_device(did);
2795         if (vdpa_dev && virtio_is_ready(dev) &&
2796                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2797                         msg.request.master == VHOST_USER_SET_VRING_CALL) {
2798                 if (vdpa_dev->ops->dev_conf)
2799                         vdpa_dev->ops->dev_conf(dev->vid);
2800                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2801         }
2802
2803         return 0;
2804 }
2805
2806 static int process_slave_message_reply(struct virtio_net *dev,
2807                                        const struct VhostUserMsg *msg)
2808 {
2809         struct VhostUserMsg msg_reply;
2810         int ret;
2811
2812         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2813                 return 0;
2814
2815         ret = read_vhost_message(dev->slave_req_fd, &msg_reply);
2816         if (ret <= 0) {
2817                 if (ret < 0)
2818                         VHOST_LOG_CONFIG(ERR,
2819                                 "vhost read slave message reply failed\n");
2820                 else
2821                         VHOST_LOG_CONFIG(INFO,
2822                                 "vhost peer closed\n");
2823                 ret = -1;
2824                 goto out;
2825         }
2826
2827         ret = 0;
2828         if (msg_reply.request.slave != msg->request.slave) {
2829                 VHOST_LOG_CONFIG(ERR,
2830                         "Received unexpected msg type (%u), expected %u\n",
2831                         msg_reply.request.slave, msg->request.slave);
2832                 ret = -1;
2833                 goto out;
2834         }
2835
2836         ret = msg_reply.payload.u64 ? -1 : 0;
2837
2838 out:
2839         rte_spinlock_unlock(&dev->slave_req_lock);
2840         return ret;
2841 }
2842
2843 int
2844 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2845 {
2846         int ret;
2847         struct VhostUserMsg msg = {
2848                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2849                 .flags = VHOST_USER_VERSION,
2850                 .size = sizeof(msg.payload.iotlb),
2851                 .payload.iotlb = {
2852                         .iova = iova,
2853                         .perm = perm,
2854                         .type = VHOST_IOTLB_MISS,
2855                 },
2856         };
2857
2858         ret = send_vhost_message(dev->slave_req_fd, &msg);
2859         if (ret < 0) {
2860                 VHOST_LOG_CONFIG(ERR,
2861                                 "Failed to send IOTLB miss message (%d)\n",
2862                                 ret);
2863                 return ret;
2864         }
2865
2866         return 0;
2867 }
2868
2869 static int
2870 vhost_user_slave_config_change(struct virtio_net *dev, bool need_reply)
2871 {
2872         int ret;
2873         struct VhostUserMsg msg = {
2874                 .request.slave = VHOST_USER_SLAVE_CONFIG_CHANGE_MSG,
2875                 .flags = VHOST_USER_VERSION,
2876                 .size = 0,
2877         };
2878
2879         if (need_reply)
2880                 msg.flags |= VHOST_USER_NEED_REPLY;
2881
2882         ret = send_vhost_slave_message(dev, &msg);
2883         if (ret < 0) {
2884                 VHOST_LOG_CONFIG(ERR,
2885                                 "Failed to send config change (%d)\n",
2886                                 ret);
2887                 return ret;
2888         }
2889
2890         return process_slave_message_reply(dev, &msg);
2891 }
2892
2893 int
2894 rte_vhost_slave_config_change(int vid, bool need_reply)
2895 {
2896         struct virtio_net *dev;
2897
2898         dev = get_device(vid);
2899         if (!dev)
2900                 return -ENODEV;
2901
2902         return vhost_user_slave_config_change(dev, need_reply);
2903 }
2904
2905 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2906                                                     int index, int fd,
2907                                                     uint64_t offset,
2908                                                     uint64_t size)
2909 {
2910         int ret;
2911         struct VhostUserMsg msg = {
2912                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2913                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2914                 .size = sizeof(msg.payload.area),
2915                 .payload.area = {
2916                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
2917                         .size = size,
2918                         .offset = offset,
2919                 },
2920         };
2921
2922         if (fd < 0)
2923                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2924         else {
2925                 msg.fds[0] = fd;
2926                 msg.fd_num = 1;
2927         }
2928
2929         ret = send_vhost_slave_message(dev, &msg);
2930         if (ret < 0) {
2931                 VHOST_LOG_CONFIG(ERR,
2932                         "Failed to set host notifier (%d)\n", ret);
2933                 return ret;
2934         }
2935
2936         return process_slave_message_reply(dev, &msg);
2937 }
2938
2939 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2940 {
2941         struct virtio_net *dev;
2942         struct rte_vdpa_device *vdpa_dev;
2943         int vfio_device_fd, did, ret = 0;
2944         uint64_t offset, size;
2945         unsigned int i;
2946
2947         dev = get_device(vid);
2948         if (!dev)
2949                 return -ENODEV;
2950
2951         did = dev->vdpa_dev_id;
2952         if (did < 0)
2953                 return -EINVAL;
2954
2955         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2956             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2957             !(dev->protocol_features &
2958                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2959             !(dev->protocol_features &
2960                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2961             !(dev->protocol_features &
2962                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2963                 return -ENOTSUP;
2964
2965         vdpa_dev = rte_vdpa_get_device(did);
2966         if (!vdpa_dev)
2967                 return -ENODEV;
2968
2969         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2970         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2971
2972         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2973         if (vfio_device_fd < 0)
2974                 return -ENOTSUP;
2975
2976         if (enable) {
2977                 for (i = 0; i < dev->nr_vring; i++) {
2978                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2979                                         &size) < 0) {
2980                                 ret = -ENOTSUP;
2981                                 goto disable;
2982                         }
2983
2984                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2985                                         vfio_device_fd, offset, size) < 0) {
2986                                 ret = -EFAULT;
2987                                 goto disable;
2988                         }
2989                 }
2990         } else {
2991 disable:
2992                 for (i = 0; i < dev->nr_vring; i++) {
2993                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
2994                                         0, 0);
2995                 }
2996         }
2997
2998         return ret;
2999 }