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                 uint64_t req_size = *size;
 660
 661                 vva = vhost_user_iotlb_cache_find(vq, ra,
 662                                         size, VHOST_ACCESS_RW);
 663                 if (req_size != *size)
 664                         vhost_user_iotlb_miss(dev, (ra + *size),
 665                                               VHOST_ACCESS_RW);
 666
 667                 return vva;
 668         }
 669
 670         return qva_to_vva(dev, ra, size);
 671 }
 672
 673 /*
 674  * Converts vring log address to GPA
 675  * If IOMMU is enabled, the log address is IOVA
 676  * If IOMMU not enabled, the log address is already GPA
 677  */
 678 static uint64_t
 679 translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
 680                 uint64_t log_addr)
 681 {
 682         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 683                 const uint64_t exp_size = sizeof(struct vring_used) +
 684                         sizeof(struct vring_used_elem) * vq->size;
 685                 uint64_t hva, gpa;
 686                 uint64_t size = exp_size;
 687
 688                 hva = vhost_iova_to_vva(dev, vq, log_addr,
 689                                         &size, VHOST_ACCESS_RW);
 690                 if (size != exp_size)
 691                         return 0;
 692
 693                 gpa = hva_to_gpa(dev, hva, exp_size);
 694                 if (!gpa) {
 695                         VHOST_LOG_CONFIG(ERR,
 696                                 "VQ: Failed to find GPA for log_addr: 0x%" PRIx64 " hva: 0x%" PRIx64 "\n",
 697                                 log_addr, hva);
 698                         return 0;
 699                 }
 700                 return gpa;
 701
 702         } else
 703                 return log_addr;
 704 }
 705
 706 static struct virtio_net *
 707 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 708 {
 709         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 710         struct vhost_vring_addr *addr = &vq->ring_addrs;
 711         uint64_t len, expected_len;
 712
 713         if (addr->flags & (1 << VHOST_VRING_F_LOG)) {
 714                 vq->log_guest_addr =
 715                         translate_log_addr(dev, vq, addr->log_guest_addr);
 716                 if (vq->log_guest_addr == 0) {
 717                         VHOST_LOG_CONFIG(DEBUG,
 718                                 "(%d) failed to map log_guest_addr.\n",
 719                                 dev->vid);
 720                         return dev;
 721                 }
 722         }
 723
 724         if (vq_is_packed(dev)) {
 725                 len = sizeof(struct vring_packed_desc) * vq->size;
 726                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 727                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 728                 if (vq->desc_packed == NULL ||
 729                                 len != sizeof(struct vring_packed_desc) *
 730                                 vq->size) {
 731                         VHOST_LOG_CONFIG(DEBUG,
 732                                 "(%d) failed to map desc_packed ring.\n",
 733                                 dev->vid);
 734                         return dev;
 735                 }
 736
 737                 dev = numa_realloc(dev, vq_index);
 738                 vq = dev->virtqueue[vq_index];
 739                 addr = &vq->ring_addrs;
 740
 741                 len = sizeof(struct vring_packed_desc_event);
 742                 vq->driver_event = (struct vring_packed_desc_event *)
 743                                         (uintptr_t)ring_addr_to_vva(dev,
 744                                         vq, addr->avail_user_addr, &len);
 745                 if (vq->driver_event == NULL ||
 746                                 len != sizeof(struct vring_packed_desc_event)) {
 747                         VHOST_LOG_CONFIG(DEBUG,
 748                                 "(%d) failed to find driver area address.\n",
 749                                 dev->vid);
 750                         return dev;
 751                 }
 752
 753                 len = sizeof(struct vring_packed_desc_event);
 754                 vq->device_event = (struct vring_packed_desc_event *)
 755                                         (uintptr_t)ring_addr_to_vva(dev,
 756                                         vq, addr->used_user_addr, &len);
 757                 if (vq->device_event == NULL ||
 758                                 len != sizeof(struct vring_packed_desc_event)) {
 759                         VHOST_LOG_CONFIG(DEBUG,
 760                                 "(%d) failed to find device area address.\n",
 761                                 dev->vid);
 762                         return dev;
 763                 }
 764
 765                 vq->access_ok = 1;
 766                 return dev;
 767         }
 768
 769         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 770         if (vq->desc && vq->avail && vq->used)
 771                 return dev;
 772
 773         len = sizeof(struct vring_desc) * vq->size;
 774         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 775                         vq, addr->desc_user_addr, &len);
 776         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 777                 VHOST_LOG_CONFIG(DEBUG,
 778                         "(%d) failed to map desc ring.\n",
 779                         dev->vid);
 780                 return dev;
 781         }
 782
 783         dev = numa_realloc(dev, vq_index);
 784         vq = dev->virtqueue[vq_index];
 785         addr = &vq->ring_addrs;
 786
 787         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 788         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 789                 len += sizeof(uint16_t);
 790         expected_len = len;
 791         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 792                         vq, addr->avail_user_addr, &len);
 793         if (vq->avail == 0 || len != expected_len) {
 794                 VHOST_LOG_CONFIG(DEBUG,
 795                         "(%d) failed to map avail ring.\n",
 796                         dev->vid);
 797                 return dev;
 798         }
 799
 800         len = sizeof(struct vring_used) +
 801                 sizeof(struct vring_used_elem) * vq->size;
 802         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 803                 len += sizeof(uint16_t);
 804         expected_len = len;
 805         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 806                         vq, addr->used_user_addr, &len);
 807         if (vq->used == 0 || len != expected_len) {
 808                 VHOST_LOG_CONFIG(DEBUG,
 809                         "(%d) failed to map used ring.\n",
 810                         dev->vid);
 811                 return dev;
 812         }
 813
 814         if (vq->last_used_idx != vq->used->idx) {
 815                 VHOST_LOG_CONFIG(WARNING,
 816                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 817                         "some packets maybe resent for Tx and dropped for Rx\n",
 818                         vq->last_used_idx, vq->used->idx);
 819                 vq->last_used_idx  = vq->used->idx;
 820                 vq->last_avail_idx = vq->used->idx;
 821         }
 822
 823         vq->access_ok = 1;
 824
 825         VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address desc: %p\n",
 826                         dev->vid, vq->desc);
 827         VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address avail: %p\n",
 828                         dev->vid, vq->avail);
 829         VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address used: %p\n",
 830                         dev->vid, vq->used);
 831         VHOST_LOG_CONFIG(DEBUG, "(%d) log_guest_addr: %" PRIx64 "\n",
 832                         dev->vid, vq->log_guest_addr);
 833
 834         return dev;
 835 }
 836
 837 /*
 838  * The virtio device sends us the desc, used and avail ring addresses.
 839  * This function then converts these to our address space.
 840  */
 841 static int
 842 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 843                         int main_fd __rte_unused)
 844 {
 845         struct virtio_net *dev = *pdev;
 846         struct vhost_virtqueue *vq;
 847         struct vhost_vring_addr *addr = &msg->payload.addr;
 848         bool access_ok;
 849
 850         if (validate_msg_fds(msg, 0) != 0)
 851                 return RTE_VHOST_MSG_RESULT_ERR;
 852
 853         if (dev->mem == NULL)
 854                 return RTE_VHOST_MSG_RESULT_ERR;
 855
 856         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 857         vq = dev->virtqueue[msg->payload.addr.index];
 858
 859         access_ok = vq->access_ok;
 860
 861         /*
 862          * Rings addresses should not be interpreted as long as the ring is not
 863          * started and enabled
 864          */
 865         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 866
 867         vring_invalidate(dev, vq);
 868
 869         if ((vq->enabled && (dev->features &
 870                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) ||
 871                         access_ok) {
 872                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 873                 if (!dev)
 874                         return RTE_VHOST_MSG_RESULT_ERR;
 875
 876                 *pdev = dev;
 877         }
 878
 879         return RTE_VHOST_MSG_RESULT_OK;
 880 }
 881
 882 /*
 883  * The virtio device sends us the available ring last used index.
 884  */
 885 static int
 886 vhost_user_set_vring_base(struct virtio_net **pdev,
 887                         struct VhostUserMsg *msg,
 888                         int main_fd __rte_unused)
 889 {
 890         struct virtio_net *dev = *pdev;
 891         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 892         uint64_t val = msg->payload.state.num;
 893
 894         if (validate_msg_fds(msg, 0) != 0)
 895                 return RTE_VHOST_MSG_RESULT_ERR;
 896
 897         if (vq_is_packed(dev)) {
 898                 /*
 899                  * Bit[0:14]: avail index
 900                  * Bit[15]: avail wrap counter
 901                  */
 902                 vq->last_avail_idx = val & 0x7fff;
 903                 vq->avail_wrap_counter = !!(val & (0x1 << 15));
 904                 /*
 905                  * Set used index to same value as available one, as
 906                  * their values should be the same since ring processing
 907                  * was stopped at get time.
 908                  */
 909                 vq->last_used_idx = vq->last_avail_idx;
 910                 vq->used_wrap_counter = vq->avail_wrap_counter;
 911         } else {
 912                 vq->last_used_idx = msg->payload.state.num;
 913                 vq->last_avail_idx = msg->payload.state.num;
 914         }
 915
 916         return RTE_VHOST_MSG_RESULT_OK;
 917 }
 918
 919 static int
 920 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 921                    uint64_t host_phys_addr, uint64_t size)
 922 {
 923         struct guest_page *page, *last_page;
 924         struct guest_page *old_pages;
 925
 926         if (dev->nr_guest_pages == dev->max_guest_pages) {
 927                 dev->max_guest_pages *= 2;
 928                 old_pages = dev->guest_pages;
 929                 dev->guest_pages = realloc(dev->guest_pages,
 930                                         dev->max_guest_pages * sizeof(*page));
 931                 if (!dev->guest_pages) {
 932                         VHOST_LOG_CONFIG(ERR, "cannot realloc guest_pages\n");
 933                         free(old_pages);
 934                         return -1;
 935                 }
 936         }
 937
 938         if (dev->nr_guest_pages > 0) {
 939                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 940                 /* merge if the two pages are continuous */
 941                 if (host_phys_addr == last_page->host_phys_addr +
 942                                       last_page->size) {
 943                         last_page->size += size;
 944                         return 0;
 945                 }
 946         }
 947
 948         page = &dev->guest_pages[dev->nr_guest_pages++];
 949         page->guest_phys_addr = guest_phys_addr;
 950         page->host_phys_addr  = host_phys_addr;
 951         page->size = size;
 952
 953         return 0;
 954 }
 955
 956 static int
 957 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 958                 uint64_t page_size)
 959 {
 960         uint64_t reg_size = reg->size;
 961         uint64_t host_user_addr  = reg->host_user_addr;
 962         uint64_t guest_phys_addr = reg->guest_phys_addr;
 963         uint64_t host_phys_addr;
 964         uint64_t size;
 965
 966         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 967         size = page_size - (guest_phys_addr & (page_size - 1));
 968         size = RTE_MIN(size, reg_size);
 969
 970         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 971                 return -1;
 972
 973         host_user_addr  += size;
 974         guest_phys_addr += size;
 975         reg_size -= size;
 976
 977         while (reg_size > 0) {
 978                 size = RTE_MIN(reg_size, page_size);
 979                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 980                                                   host_user_addr);
 981                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 982                                 size) < 0)
 983                         return -1;
 984
 985                 host_user_addr  += size;
 986                 guest_phys_addr += size;
 987                 reg_size -= size;
 988         }
 989
 990         return 0;
 991 }
 992
 993 #ifdef RTE_LIBRTE_VHOST_DEBUG
 994 /* TODO: enable it only in debug mode? */
 995 static void
 996 dump_guest_pages(struct virtio_net *dev)
 997 {
 998         uint32_t i;
 999         struct guest_page *page;
1000
1001         for (i = 0; i < dev->nr_guest_pages; i++) {
1002                 page = &dev->guest_pages[i];
1003
1004                 VHOST_LOG_CONFIG(INFO,
1005                         "guest physical page region %u\n"
1006                         "\t guest_phys_addr: %" PRIx64 "\n"
1007                         "\t host_phys_addr : %" PRIx64 "\n"
1008                         "\t size           : %" PRIx64 "\n",
1009                         i,
1010                         page->guest_phys_addr,
1011                         page->host_phys_addr,
1012                         page->size);
1013         }
1014 }
1015 #else
1016 #define dump_guest_pages(dev)
1017 #endif
1018
1019 static bool
1020 vhost_memory_changed(struct VhostUserMemory *new,
1021                      struct rte_vhost_memory *old)
1022 {
1023         uint32_t i;
1024
1025         if (new->nregions != old->nregions)
1026                 return true;
1027
1028         for (i = 0; i < new->nregions; ++i) {
1029                 VhostUserMemoryRegion *new_r = &new->regions[i];
1030                 struct rte_vhost_mem_region *old_r = &old->regions[i];
1031
1032                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
1033                         return true;
1034                 if (new_r->memory_size != old_r->size)
1035                         return true;
1036                 if (new_r->userspace_addr != old_r->guest_user_addr)
1037                         return true;
1038         }
1039
1040         return false;
1041 }
1042
1043 static int
1044 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
1045                         int main_fd)
1046 {
1047         struct virtio_net *dev = *pdev;
1048         struct VhostUserMemory *memory = &msg->payload.memory;
1049         struct rte_vhost_mem_region *reg;
1050         void *mmap_addr;
1051         uint64_t mmap_size;
1052         uint64_t mmap_offset;
1053         uint64_t alignment;
1054         uint32_t i;
1055         int populate;
1056         int fd;
1057
1058         if (validate_msg_fds(msg, memory->nregions) != 0)
1059                 return RTE_VHOST_MSG_RESULT_ERR;
1060
1061         if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
1062                 VHOST_LOG_CONFIG(ERR,
1063                         "too many memory regions (%u)\n", memory->nregions);
1064                 return RTE_VHOST_MSG_RESULT_ERR;
1065         }
1066
1067         if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
1068                 VHOST_LOG_CONFIG(INFO,
1069                         "(%d) memory regions not changed\n", dev->vid);
1070
1071                 close_msg_fds(msg);
1072
1073                 return RTE_VHOST_MSG_RESULT_OK;
1074         }
1075
1076         if (dev->mem) {
1077                 free_mem_region(dev);
1078                 rte_free(dev->mem);
1079                 dev->mem = NULL;
1080         }
1081
1082         /* Flush IOTLB cache as previous HVAs are now invalid */
1083         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1084                 for (i = 0; i < dev->nr_vring; i++)
1085                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
1086
1087         dev->nr_guest_pages = 0;
1088         if (!dev->guest_pages) {
1089                 dev->max_guest_pages = 8;
1090                 dev->guest_pages = malloc(dev->max_guest_pages *
1091                                                 sizeof(struct guest_page));
1092                 if (dev->guest_pages == NULL) {
1093                         VHOST_LOG_CONFIG(ERR,
1094                                 "(%d) failed to allocate memory "
1095                                 "for dev->guest_pages\n",
1096                                 dev->vid);
1097                         return RTE_VHOST_MSG_RESULT_ERR;
1098                 }
1099         }
1100
1101         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
1102                 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
1103         if (dev->mem == NULL) {
1104                 VHOST_LOG_CONFIG(ERR,
1105                         "(%d) failed to allocate memory for dev->mem\n",
1106                         dev->vid);
1107                 return RTE_VHOST_MSG_RESULT_ERR;
1108         }
1109         dev->mem->nregions = memory->nregions;
1110
1111         for (i = 0; i < memory->nregions; i++) {
1112                 fd  = msg->fds[i];
1113                 reg = &dev->mem->regions[i];
1114
1115                 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
1116                 reg->guest_user_addr = memory->regions[i].userspace_addr;
1117                 reg->size            = memory->regions[i].memory_size;
1118                 reg->fd              = fd;
1119
1120                 mmap_offset = memory->regions[i].mmap_offset;
1121
1122                 /* Check for memory_size + mmap_offset overflow */
1123                 if (mmap_offset >= -reg->size) {
1124                         VHOST_LOG_CONFIG(ERR,
1125                                 "mmap_offset (%#"PRIx64") and memory_size "
1126                                 "(%#"PRIx64") overflow\n",
1127                                 mmap_offset, reg->size);
1128                         goto err_mmap;
1129                 }
1130
1131                 mmap_size = reg->size + mmap_offset;
1132
1133                 /* mmap() without flag of MAP_ANONYMOUS, should be called
1134                  * with length argument aligned with hugepagesz at older
1135                  * longterm version Linux, like 2.6.32 and 3.2.72, or
1136                  * mmap() will fail with EINVAL.
1137                  *
1138                  * to avoid failure, make sure in caller to keep length
1139                  * aligned.
1140                  */
1141                 alignment = get_blk_size(fd);
1142                 if (alignment == (uint64_t)-1) {
1143                         VHOST_LOG_CONFIG(ERR,
1144                                 "couldn't get hugepage size through fstat\n");
1145                         goto err_mmap;
1146                 }
1147                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
1148                 if (mmap_size == 0) {
1149                         /*
1150                          * It could happen if initial mmap_size + alignment
1151                          * overflows the sizeof uint64, which could happen if
1152                          * either mmap_size or alignment value is wrong.
1153                          *
1154                          * mmap() kernel implementation would return an error,
1155                          * but better catch it before and provide useful info
1156                          * in the logs.
1157                          */
1158                         VHOST_LOG_CONFIG(ERR, "mmap size (0x%" PRIx64 ") "
1159                                         "or alignment (0x%" PRIx64 ") is invalid\n",
1160                                         reg->size + mmap_offset, alignment);
1161                         goto err_mmap;
1162                 }
1163
1164                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
1165                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
1166                                  MAP_SHARED | populate, fd, 0);
1167
1168                 if (mmap_addr == MAP_FAILED) {
1169                         VHOST_LOG_CONFIG(ERR,
1170                                 "mmap region %u failed.\n", i);
1171                         goto err_mmap;
1172                 }
1173
1174                 reg->mmap_addr = mmap_addr;
1175                 reg->mmap_size = mmap_size;
1176                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
1177                                       mmap_offset;
1178
1179                 if (dev->dequeue_zero_copy)
1180                         if (add_guest_pages(dev, reg, alignment) < 0) {
1181                                 VHOST_LOG_CONFIG(ERR,
1182                                         "adding guest pages to region %u failed.\n",
1183                                         i);
1184                                 goto err_mmap;
1185                         }
1186
1187                 VHOST_LOG_CONFIG(INFO,
1188                         "guest memory region %u, size: 0x%" PRIx64 "\n"
1189                         "\t guest physical addr: 0x%" PRIx64 "\n"
1190                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
1191                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
1192                         "\t mmap addr : 0x%" PRIx64 "\n"
1193                         "\t mmap size : 0x%" PRIx64 "\n"
1194                         "\t mmap align: 0x%" PRIx64 "\n"
1195                         "\t mmap off  : 0x%" PRIx64 "\n",
1196                         i, reg->size,
1197                         reg->guest_phys_addr,
1198                         reg->guest_user_addr,
1199                         reg->host_user_addr,
1200                         (uint64_t)(uintptr_t)mmap_addr,
1201                         mmap_size,
1202                         alignment,
1203                         mmap_offset);
1204
1205                 if (dev->postcopy_listening) {
1206                         /*
1207                          * We haven't a better way right now than sharing
1208                          * DPDK's virtual address with Qemu, so that Qemu can
1209                          * retrieve the region offset when handling userfaults.
1210                          */
1211                         memory->regions[i].userspace_addr =
1212                                 reg->host_user_addr;
1213                 }
1214         }
1215         if (dev->postcopy_listening) {
1216                 /* Send the addresses back to qemu */
1217                 msg->fd_num = 0;
1218                 send_vhost_reply(main_fd, msg);
1219
1220                 /* Wait for qemu to acknolwedge it's got the addresses
1221                  * we've got to wait before we're allowed to generate faults.
1222                  */
1223                 VhostUserMsg ack_msg;
1224                 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1225                         VHOST_LOG_CONFIG(ERR,
1226                                 "Failed to read qemu ack on postcopy set-mem-table\n");
1227                         goto err_mmap;
1228                 }
1229
1230                 if (validate_msg_fds(&ack_msg, 0) != 0)
1231                         goto err_mmap;
1232
1233                 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1234                         VHOST_LOG_CONFIG(ERR,
1235                                 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1236                                 ack_msg.request.master);
1237                         goto err_mmap;
1238                 }
1239
1240                 /* Now userfault register and we can use the memory */
1241                 for (i = 0; i < memory->nregions; i++) {
1242 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1243                         reg = &dev->mem->regions[i];
1244                         struct uffdio_register reg_struct;
1245
1246                         /*
1247                          * Let's register all the mmap'ed area to ensure
1248                          * alignment on page boundary.
1249                          */
1250                         reg_struct.range.start =
1251                                 (uint64_t)(uintptr_t)reg->mmap_addr;
1252                         reg_struct.range.len = reg->mmap_size;
1253                         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1254
1255                         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1256                                                 &reg_struct)) {
1257                                 VHOST_LOG_CONFIG(ERR,
1258                                         "Failed to register ufd for region %d: (ufd = %d) %s\n",
1259                                         i, dev->postcopy_ufd,
1260                                         strerror(errno));
1261                                 goto err_mmap;
1262                         }
1263                         VHOST_LOG_CONFIG(INFO,
1264                                 "\t userfaultfd registered for range : "
1265                                 "%" PRIx64 " - %" PRIx64 "\n",
1266                                 (uint64_t)reg_struct.range.start,
1267                                 (uint64_t)reg_struct.range.start +
1268                                 (uint64_t)reg_struct.range.len - 1);
1269 #else
1270                         goto err_mmap;
1271 #endif
1272                 }
1273         }
1274
1275         for (i = 0; i < dev->nr_vring; i++) {
1276                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1277
1278                 if (vq->desc || vq->avail || vq->used) {
1279                         /*
1280                          * If the memory table got updated, the ring addresses
1281                          * need to be translated again as virtual addresses have
1282                          * changed.
1283                          */
1284                         vring_invalidate(dev, vq);
1285
1286                         dev = translate_ring_addresses(dev, i);
1287                         if (!dev) {
1288                                 dev = *pdev;
1289                                 goto err_mmap;
1290                         }
1291
1292                         *pdev = dev;
1293                 }
1294         }
1295
1296         dump_guest_pages(dev);
1297
1298         return RTE_VHOST_MSG_RESULT_OK;
1299
1300 err_mmap:
1301         free_mem_region(dev);
1302         rte_free(dev->mem);
1303         dev->mem = NULL;
1304         return RTE_VHOST_MSG_RESULT_ERR;
1305 }
1306
1307 static bool
1308 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1309 {
1310         bool rings_ok;
1311
1312         if (!vq)
1313                 return false;
1314
1315         if (vq_is_packed(dev))
1316                 rings_ok = vq->desc_packed && vq->driver_event &&
1317                         vq->device_event;
1318         else
1319                 rings_ok = vq->desc && vq->avail && vq->used;
1320
1321         return rings_ok &&
1322                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1323                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1324 }
1325
1326 static int
1327 virtio_is_ready(struct virtio_net *dev)
1328 {
1329         struct vhost_virtqueue *vq;
1330         uint32_t i;
1331
1332         if (dev->nr_vring == 0)
1333                 return 0;
1334
1335         for (i = 0; i < dev->nr_vring; i++) {
1336                 vq = dev->virtqueue[i];
1337
1338                 if (!vq_is_ready(dev, vq))
1339                         return 0;
1340         }
1341
1342         VHOST_LOG_CONFIG(INFO,
1343                 "virtio is now ready for processing.\n");
1344         return 1;
1345 }
1346
1347 static void *
1348 inflight_mem_alloc(const char *name, size_t size, int *fd)
1349 {
1350         void *ptr;
1351         int mfd = -1;
1352         char fname[20] = "/tmp/memfd-XXXXXX";
1353
1354         *fd = -1;
1355 #ifdef MEMFD_SUPPORTED
1356         mfd = memfd_create(name, MFD_CLOEXEC);
1357 #else
1358         RTE_SET_USED(name);
1359 #endif
1360         if (mfd == -1) {
1361                 mfd = mkstemp(fname);
1362                 if (mfd == -1) {
1363                         VHOST_LOG_CONFIG(ERR,
1364                                 "failed to get inflight buffer fd\n");
1365                         return NULL;
1366                 }
1367
1368                 unlink(fname);
1369         }
1370
1371         if (ftruncate(mfd, size) == -1) {
1372                 VHOST_LOG_CONFIG(ERR,
1373                         "failed to alloc inflight buffer\n");
1374                 close(mfd);
1375                 return NULL;
1376         }
1377
1378         ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd, 0);
1379         if (ptr == MAP_FAILED) {
1380                 VHOST_LOG_CONFIG(ERR,
1381                         "failed to mmap inflight buffer\n");
1382                 close(mfd);
1383                 return NULL;
1384         }
1385
1386         *fd = mfd;
1387         return ptr;
1388 }
1389
1390 static uint32_t
1391 get_pervq_shm_size_split(uint16_t queue_size)
1392 {
1393         return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_split) *
1394                                   queue_size + sizeof(uint64_t) +
1395                                   sizeof(uint16_t) * 4, INFLIGHT_ALIGNMENT);
1396 }
1397
1398 static uint32_t
1399 get_pervq_shm_size_packed(uint16_t queue_size)
1400 {
1401         return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_packed)
1402                                   * queue_size + sizeof(uint64_t) +
1403                                   sizeof(uint16_t) * 6 + sizeof(uint8_t) * 9,
1404                                   INFLIGHT_ALIGNMENT);
1405 }
1406
1407 static int
1408 vhost_user_get_inflight_fd(struct virtio_net **pdev,
1409                            VhostUserMsg *msg,
1410                            int main_fd __rte_unused)
1411 {
1412         struct rte_vhost_inflight_info_packed *inflight_packed;
1413         uint64_t pervq_inflight_size, mmap_size;
1414         uint16_t num_queues, queue_size;
1415         struct virtio_net *dev = *pdev;
1416         int fd, i, j;
1417         void *addr;
1418
1419         if (msg->size != sizeof(msg->payload.inflight)) {
1420                 VHOST_LOG_CONFIG(ERR,
1421                         "invalid get_inflight_fd message size is %d\n",
1422                         msg->size);
1423                 return RTE_VHOST_MSG_RESULT_ERR;
1424         }
1425
1426         if (dev->inflight_info == NULL) {
1427                 dev->inflight_info = calloc(1,
1428                                             sizeof(struct inflight_mem_info));
1429                 if (!dev->inflight_info) {
1430                         VHOST_LOG_CONFIG(ERR,
1431                                 "failed to alloc dev inflight area\n");
1432                         return RTE_VHOST_MSG_RESULT_ERR;
1433                 }
1434         }
1435
1436         num_queues = msg->payload.inflight.num_queues;
1437         queue_size = msg->payload.inflight.queue_size;
1438
1439         VHOST_LOG_CONFIG(INFO, "get_inflight_fd num_queues: %u\n",
1440                 msg->payload.inflight.num_queues);
1441         VHOST_LOG_CONFIG(INFO, "get_inflight_fd queue_size: %u\n",
1442                 msg->payload.inflight.queue_size);
1443
1444         if (vq_is_packed(dev))
1445                 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1446         else
1447                 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1448
1449         mmap_size = num_queues * pervq_inflight_size;
1450         addr = inflight_mem_alloc("vhost-inflight", mmap_size, &fd);
1451         if (!addr) {
1452                 VHOST_LOG_CONFIG(ERR,
1453                         "failed to alloc vhost inflight area\n");
1454                         msg->payload.inflight.mmap_size = 0;
1455                 return RTE_VHOST_MSG_RESULT_ERR;
1456         }
1457         memset(addr, 0, mmap_size);
1458
1459         dev->inflight_info->addr = addr;
1460         dev->inflight_info->size = msg->payload.inflight.mmap_size = mmap_size;
1461         dev->inflight_info->fd = msg->fds[0] = fd;
1462         msg->payload.inflight.mmap_offset = 0;
1463         msg->fd_num = 1;
1464
1465         if (vq_is_packed(dev)) {
1466                 for (i = 0; i < num_queues; i++) {
1467                         inflight_packed =
1468                                 (struct rte_vhost_inflight_info_packed *)addr;
1469                         inflight_packed->used_wrap_counter = 1;
1470                         inflight_packed->old_used_wrap_counter = 1;
1471                         for (j = 0; j < queue_size; j++)
1472                                 inflight_packed->desc[j].next = j + 1;
1473                         addr = (void *)((char *)addr + pervq_inflight_size);
1474                 }
1475         }
1476
1477         VHOST_LOG_CONFIG(INFO,
1478                 "send inflight mmap_size: %"PRIu64"\n",
1479                 msg->payload.inflight.mmap_size);
1480         VHOST_LOG_CONFIG(INFO,
1481                 "send inflight mmap_offset: %"PRIu64"\n",
1482                 msg->payload.inflight.mmap_offset);
1483         VHOST_LOG_CONFIG(INFO,
1484                 "send inflight fd: %d\n", msg->fds[0]);
1485
1486         return RTE_VHOST_MSG_RESULT_REPLY;
1487 }
1488
1489 static int
1490 vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
1491                            int main_fd __rte_unused)
1492 {
1493         uint64_t mmap_size, mmap_offset;
1494         uint16_t num_queues, queue_size;
1495         struct virtio_net *dev = *pdev;
1496         uint32_t pervq_inflight_size;
1497         struct vhost_virtqueue *vq;
1498         void *addr;
1499         int fd, i;
1500
1501         fd = msg->fds[0];
1502         if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
1503                 VHOST_LOG_CONFIG(ERR,
1504                         "invalid set_inflight_fd message size is %d,fd is %d\n",
1505                         msg->size, fd);
1506                 return RTE_VHOST_MSG_RESULT_ERR;
1507         }
1508
1509         mmap_size = msg->payload.inflight.mmap_size;
1510         mmap_offset = msg->payload.inflight.mmap_offset;
1511         num_queues = msg->payload.inflight.num_queues;
1512         queue_size = msg->payload.inflight.queue_size;
1513
1514         if (vq_is_packed(dev))
1515                 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1516         else
1517                 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1518
1519         VHOST_LOG_CONFIG(INFO,
1520                 "set_inflight_fd mmap_size: %"PRIu64"\n", mmap_size);
1521         VHOST_LOG_CONFIG(INFO,
1522                 "set_inflight_fd mmap_offset: %"PRIu64"\n", mmap_offset);
1523         VHOST_LOG_CONFIG(INFO,
1524                 "set_inflight_fd num_queues: %u\n", num_queues);
1525         VHOST_LOG_CONFIG(INFO,
1526                 "set_inflight_fd queue_size: %u\n", queue_size);
1527         VHOST_LOG_CONFIG(INFO,
1528                 "set_inflight_fd fd: %d\n", fd);
1529         VHOST_LOG_CONFIG(INFO,
1530                 "set_inflight_fd pervq_inflight_size: %d\n",
1531                 pervq_inflight_size);
1532
1533         if (!dev->inflight_info) {
1534                 dev->inflight_info = calloc(1,
1535                                             sizeof(struct inflight_mem_info));
1536                 if (dev->inflight_info == NULL) {
1537                         VHOST_LOG_CONFIG(ERR,
1538                                 "failed to alloc dev inflight area\n");
1539                         return RTE_VHOST_MSG_RESULT_ERR;
1540                 }
1541         }
1542
1543         if (dev->inflight_info->addr)
1544                 munmap(dev->inflight_info->addr, dev->inflight_info->size);
1545
1546         addr = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
1547                     fd, mmap_offset);
1548         if (addr == MAP_FAILED) {
1549                 VHOST_LOG_CONFIG(ERR, "failed to mmap share memory.\n");
1550                 return RTE_VHOST_MSG_RESULT_ERR;
1551         }
1552
1553         if (dev->inflight_info->fd)
1554                 close(dev->inflight_info->fd);
1555
1556         dev->inflight_info->fd = fd;
1557         dev->inflight_info->addr = addr;
1558         dev->inflight_info->size = mmap_size;
1559
1560         for (i = 0; i < num_queues; i++) {
1561                 vq = dev->virtqueue[i];
1562                 if (vq_is_packed(dev)) {
1563                         vq->inflight_packed = addr;
1564                         vq->inflight_packed->desc_num = queue_size;
1565                 } else {
1566                         vq->inflight_split = addr;
1567                         vq->inflight_split->desc_num = queue_size;
1568                 }
1569                 addr = (void *)((char *)addr + pervq_inflight_size);
1570         }
1571
1572         return RTE_VHOST_MSG_RESULT_OK;
1573 }
1574
1575 static int
1576 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1577                         int main_fd __rte_unused)
1578 {
1579         struct virtio_net *dev = *pdev;
1580         struct vhost_vring_file file;
1581         struct vhost_virtqueue *vq;
1582         int expected_fds;
1583
1584         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1585         if (validate_msg_fds(msg, expected_fds) != 0)
1586                 return RTE_VHOST_MSG_RESULT_ERR;
1587
1588         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1589         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1590                 file.fd = VIRTIO_INVALID_EVENTFD;
1591         else
1592                 file.fd = msg->fds[0];
1593         VHOST_LOG_CONFIG(INFO,
1594                 "vring call idx:%d file:%d\n", file.index, file.fd);
1595
1596         vq = dev->virtqueue[file.index];
1597         if (vq->callfd >= 0)
1598                 close(vq->callfd);
1599
1600         vq->callfd = file.fd;
1601
1602         return RTE_VHOST_MSG_RESULT_OK;
1603 }
1604
1605 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1606                         struct VhostUserMsg *msg,
1607                         int main_fd __rte_unused)
1608 {
1609         int expected_fds;
1610
1611         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1612         if (validate_msg_fds(msg, expected_fds) != 0)
1613                 return RTE_VHOST_MSG_RESULT_ERR;
1614
1615         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1616                 close(msg->fds[0]);
1617         VHOST_LOG_CONFIG(INFO, "not implemented\n");
1618
1619         return RTE_VHOST_MSG_RESULT_OK;
1620 }
1621
1622 static int
1623 resubmit_desc_compare(const void *a, const void *b)
1624 {
1625         const struct rte_vhost_resubmit_desc *desc0 = a;
1626         const struct rte_vhost_resubmit_desc *desc1 = b;
1627
1628         if (desc1->counter > desc0->counter)
1629                 return 1;
1630
1631         return -1;
1632 }
1633
1634 static int
1635 vhost_check_queue_inflights_split(struct virtio_net *dev,
1636                                   struct vhost_virtqueue *vq)
1637 {
1638         uint16_t i;
1639         uint16_t resubmit_num = 0, last_io, num;
1640         struct vring_used *used = vq->used;
1641         struct rte_vhost_resubmit_info *resubmit;
1642         struct rte_vhost_inflight_info_split *inflight_split;
1643
1644         if (!(dev->protocol_features &
1645             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1646                 return RTE_VHOST_MSG_RESULT_OK;
1647
1648         /* The frontend may still not support the inflight feature
1649          * although we negotiate the protocol feature.
1650          */
1651         if ((!vq->inflight_split))
1652                 return RTE_VHOST_MSG_RESULT_OK;
1653
1654         if (!vq->inflight_split->version) {
1655                 vq->inflight_split->version = INFLIGHT_VERSION;
1656                 return RTE_VHOST_MSG_RESULT_OK;
1657         }
1658
1659         if (vq->resubmit_inflight)
1660                 return RTE_VHOST_MSG_RESULT_OK;
1661
1662         inflight_split = vq->inflight_split;
1663         vq->global_counter = 0;
1664         last_io = inflight_split->last_inflight_io;
1665
1666         if (inflight_split->used_idx != used->idx) {
1667                 inflight_split->desc[last_io].inflight = 0;
1668                 rte_smp_mb();
1669                 inflight_split->used_idx = used->idx;
1670         }
1671
1672         for (i = 0; i < inflight_split->desc_num; i++) {
1673                 if (inflight_split->desc[i].inflight == 1)
1674                         resubmit_num++;
1675         }
1676
1677         vq->last_avail_idx += resubmit_num;
1678
1679         if (resubmit_num) {
1680                 resubmit  = calloc(1, sizeof(struct rte_vhost_resubmit_info));
1681                 if (!resubmit) {
1682                         VHOST_LOG_CONFIG(ERR,
1683                                 "failed to allocate memory for resubmit info.\n");
1684                         return RTE_VHOST_MSG_RESULT_ERR;
1685                 }
1686
1687                 resubmit->resubmit_list = calloc(resubmit_num,
1688                         sizeof(struct rte_vhost_resubmit_desc));
1689                 if (!resubmit->resubmit_list) {
1690                         VHOST_LOG_CONFIG(ERR,
1691                                 "failed to allocate memory for inflight desc.\n");
1692                         free(resubmit);
1693                         return RTE_VHOST_MSG_RESULT_ERR;
1694                 }
1695
1696                 num = 0;
1697                 for (i = 0; i < vq->inflight_split->desc_num; i++) {
1698                         if (vq->inflight_split->desc[i].inflight == 1) {
1699                                 resubmit->resubmit_list[num].index = i;
1700                                 resubmit->resubmit_list[num].counter =
1701                                         inflight_split->desc[i].counter;
1702                                 num++;
1703                         }
1704                 }
1705                 resubmit->resubmit_num = num;
1706
1707                 if (resubmit->resubmit_num > 1)
1708                         qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1709                               sizeof(struct rte_vhost_resubmit_desc),
1710                               resubmit_desc_compare);
1711
1712                 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1713                 vq->resubmit_inflight = resubmit;
1714         }
1715
1716         return RTE_VHOST_MSG_RESULT_OK;
1717 }
1718
1719 static int
1720 vhost_check_queue_inflights_packed(struct virtio_net *dev,
1721                                    struct vhost_virtqueue *vq)
1722 {
1723         uint16_t i;
1724         uint16_t resubmit_num = 0, old_used_idx, num;
1725         struct rte_vhost_resubmit_info *resubmit;
1726         struct rte_vhost_inflight_info_packed *inflight_packed;
1727
1728         if (!(dev->protocol_features &
1729             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1730                 return RTE_VHOST_MSG_RESULT_OK;
1731
1732         /* The frontend may still not support the inflight feature
1733          * although we negotiate the protocol feature.
1734          */
1735         if ((!vq->inflight_packed))
1736                 return RTE_VHOST_MSG_RESULT_OK;
1737
1738         if (!vq->inflight_packed->version) {
1739                 vq->inflight_packed->version = INFLIGHT_VERSION;
1740                 return RTE_VHOST_MSG_RESULT_OK;
1741         }
1742
1743         if (vq->resubmit_inflight)
1744                 return RTE_VHOST_MSG_RESULT_OK;
1745
1746         inflight_packed = vq->inflight_packed;
1747         vq->global_counter = 0;
1748         old_used_idx = inflight_packed->old_used_idx;
1749
1750         if (inflight_packed->used_idx != old_used_idx) {
1751                 if (inflight_packed->desc[old_used_idx].inflight == 0) {
1752                         inflight_packed->old_used_idx =
1753                                 inflight_packed->used_idx;
1754                         inflight_packed->old_used_wrap_counter =
1755                                 inflight_packed->used_wrap_counter;
1756                         inflight_packed->old_free_head =
1757                                 inflight_packed->free_head;
1758                 } else {
1759                         inflight_packed->used_idx =
1760                                 inflight_packed->old_used_idx;
1761                         inflight_packed->used_wrap_counter =
1762                                 inflight_packed->old_used_wrap_counter;
1763                         inflight_packed->free_head =
1764                                 inflight_packed->old_free_head;
1765                 }
1766         }
1767
1768         for (i = 0; i < inflight_packed->desc_num; i++) {
1769                 if (inflight_packed->desc[i].inflight == 1)
1770                         resubmit_num++;
1771         }
1772
1773         if (resubmit_num) {
1774                 resubmit = calloc(1, sizeof(struct rte_vhost_resubmit_info));
1775                 if (resubmit == NULL) {
1776                         VHOST_LOG_CONFIG(ERR,
1777                                 "failed to allocate memory for resubmit info.\n");
1778                         return RTE_VHOST_MSG_RESULT_ERR;
1779                 }
1780
1781                 resubmit->resubmit_list = calloc(resubmit_num,
1782                         sizeof(struct rte_vhost_resubmit_desc));
1783                 if (resubmit->resubmit_list == NULL) {
1784                         VHOST_LOG_CONFIG(ERR,
1785                                 "failed to allocate memory for resubmit desc.\n");
1786                         free(resubmit);
1787                         return RTE_VHOST_MSG_RESULT_ERR;
1788                 }
1789
1790                 num = 0;
1791                 for (i = 0; i < inflight_packed->desc_num; i++) {
1792                         if (vq->inflight_packed->desc[i].inflight == 1) {
1793                                 resubmit->resubmit_list[num].index = i;
1794                                 resubmit->resubmit_list[num].counter =
1795                                         inflight_packed->desc[i].counter;
1796                                 num++;
1797                         }
1798                 }
1799                 resubmit->resubmit_num = num;
1800
1801                 if (resubmit->resubmit_num > 1)
1802                         qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1803                               sizeof(struct rte_vhost_resubmit_desc),
1804                               resubmit_desc_compare);
1805
1806                 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1807                 vq->resubmit_inflight = resubmit;
1808         }
1809
1810         return RTE_VHOST_MSG_RESULT_OK;
1811 }
1812
1813 static int
1814 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1815                         int main_fd __rte_unused)
1816 {
1817         struct virtio_net *dev = *pdev;
1818         struct vhost_vring_file file;
1819         struct vhost_virtqueue *vq;
1820         int expected_fds;
1821
1822         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1823         if (validate_msg_fds(msg, expected_fds) != 0)
1824                 return RTE_VHOST_MSG_RESULT_ERR;
1825
1826         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1827         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1828                 file.fd = VIRTIO_INVALID_EVENTFD;
1829         else
1830                 file.fd = msg->fds[0];
1831         VHOST_LOG_CONFIG(INFO,
1832                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1833
1834         /* Interpret ring addresses only when ring is started. */
1835         dev = translate_ring_addresses(dev, file.index);
1836         if (!dev)
1837                 return RTE_VHOST_MSG_RESULT_ERR;
1838
1839         *pdev = dev;
1840
1841         vq = dev->virtqueue[file.index];
1842
1843         /*
1844          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1845          * the ring starts already enabled. Otherwise, it is enabled via
1846          * the SET_VRING_ENABLE message.
1847          */
1848         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
1849                 vq->enabled = 1;
1850                 if (dev->notify_ops->vring_state_changed)
1851                         dev->notify_ops->vring_state_changed(
1852                                 dev->vid, file.index, 1);
1853         }
1854
1855         if (vq->kickfd >= 0)
1856                 close(vq->kickfd);
1857         vq->kickfd = file.fd;
1858
1859         if (vq_is_packed(dev)) {
1860                 if (vhost_check_queue_inflights_packed(dev, vq)) {
1861                         VHOST_LOG_CONFIG(ERR,
1862                                 "failed to inflights for vq: %d\n", file.index);
1863                         return RTE_VHOST_MSG_RESULT_ERR;
1864                 }
1865         } else {
1866                 if (vhost_check_queue_inflights_split(dev, vq)) {
1867                         VHOST_LOG_CONFIG(ERR,
1868                                 "failed to inflights for vq: %d\n", file.index);
1869                         return RTE_VHOST_MSG_RESULT_ERR;
1870                 }
1871         }
1872
1873         return RTE_VHOST_MSG_RESULT_OK;
1874 }
1875
1876 static void
1877 free_zmbufs(struct vhost_virtqueue *vq)
1878 {
1879         drain_zmbuf_list(vq);
1880
1881         rte_free(vq->zmbufs);
1882 }
1883
1884 /*
1885  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1886  */
1887 static int
1888 vhost_user_get_vring_base(struct virtio_net **pdev,
1889                         struct VhostUserMsg *msg,
1890                         int main_fd __rte_unused)
1891 {
1892         struct virtio_net *dev = *pdev;
1893         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1894         uint64_t val;
1895
1896         if (validate_msg_fds(msg, 0) != 0)
1897                 return RTE_VHOST_MSG_RESULT_ERR;
1898
1899         /* We have to stop the queue (virtio) if it is running. */
1900         vhost_destroy_device_notify(dev);
1901
1902         dev->flags &= ~VIRTIO_DEV_READY;
1903         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1904
1905         /* Here we are safe to get the indexes */
1906         if (vq_is_packed(dev)) {
1907                 /*
1908                  * Bit[0:14]: avail index
1909                  * Bit[15]: avail wrap counter
1910                  */
1911                 val = vq->last_avail_idx & 0x7fff;
1912                 val |= vq->avail_wrap_counter << 15;
1913                 msg->payload.state.num = val;
1914         } else {
1915                 msg->payload.state.num = vq->last_avail_idx;
1916         }
1917
1918         VHOST_LOG_CONFIG(INFO,
1919                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1920                 msg->payload.state.num);
1921         /*
1922          * Based on current qemu vhost-user implementation, this message is
1923          * sent and only sent in vhost_vring_stop.
1924          * TODO: cleanup the vring, it isn't usable since here.
1925          */
1926         if (vq->kickfd >= 0)
1927                 close(vq->kickfd);
1928
1929         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1930
1931         if (vq->callfd >= 0)
1932                 close(vq->callfd);
1933
1934         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1935
1936         vq->signalled_used_valid = false;
1937
1938         if (dev->dequeue_zero_copy)
1939                 free_zmbufs(vq);
1940         if (vq_is_packed(dev)) {
1941                 rte_free(vq->shadow_used_packed);
1942                 vq->shadow_used_packed = NULL;
1943         } else {
1944                 rte_free(vq->shadow_used_split);
1945                 vq->shadow_used_split = NULL;
1946         }
1947
1948         rte_free(vq->batch_copy_elems);
1949         vq->batch_copy_elems = NULL;
1950
1951         msg->size = sizeof(msg->payload.state);
1952         msg->fd_num = 0;
1953
1954         vring_invalidate(dev, vq);
1955
1956         return RTE_VHOST_MSG_RESULT_REPLY;
1957 }
1958
1959 /*
1960  * when virtio queues are ready to work, qemu will send us to
1961  * enable the virtio queue pair.
1962  */
1963 static int
1964 vhost_user_set_vring_enable(struct virtio_net **pdev,
1965                         struct VhostUserMsg *msg,
1966                         int main_fd __rte_unused)
1967 {
1968         struct virtio_net *dev = *pdev;
1969         int enable = (int)msg->payload.state.num;
1970         int index = (int)msg->payload.state.index;
1971         struct rte_vdpa_device *vdpa_dev;
1972         int did = -1;
1973
1974         if (validate_msg_fds(msg, 0) != 0)
1975                 return RTE_VHOST_MSG_RESULT_ERR;
1976
1977         VHOST_LOG_CONFIG(INFO,
1978                 "set queue enable: %d to qp idx: %d\n",
1979                 enable, index);
1980
1981         did = dev->vdpa_dev_id;
1982         vdpa_dev = rte_vdpa_get_device(did);
1983         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1984                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1985
1986         if (dev->notify_ops->vring_state_changed)
1987                 dev->notify_ops->vring_state_changed(dev->vid,
1988                                 index, enable);
1989
1990         /* On disable, rings have to be stopped being processed. */
1991         if (!enable && dev->dequeue_zero_copy)
1992                 drain_zmbuf_list(dev->virtqueue[index]);
1993
1994         dev->virtqueue[index]->enabled = enable;
1995
1996         return RTE_VHOST_MSG_RESULT_OK;
1997 }
1998
1999 static int
2000 vhost_user_get_protocol_features(struct virtio_net **pdev,
2001                         struct VhostUserMsg *msg,
2002                         int main_fd __rte_unused)
2003 {
2004         struct virtio_net *dev = *pdev;
2005         uint64_t features, protocol_features;
2006
2007         if (validate_msg_fds(msg, 0) != 0)
2008                 return RTE_VHOST_MSG_RESULT_ERR;
2009
2010         rte_vhost_driver_get_features(dev->ifname, &features);
2011         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
2012
2013         /*
2014          * REPLY_ACK protocol feature is only mandatory for now
2015          * for IOMMU feature. If IOMMU is explicitly disabled by the
2016          * application, disable also REPLY_ACK feature for older buggy
2017          * Qemu versions (from v2.7.0 to v2.9.0).
2018          */
2019         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
2020                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
2021
2022         msg->payload.u64 = protocol_features;
2023         msg->size = sizeof(msg->payload.u64);
2024         msg->fd_num = 0;
2025
2026         return RTE_VHOST_MSG_RESULT_REPLY;
2027 }
2028
2029 static int
2030 vhost_user_set_protocol_features(struct virtio_net **pdev,
2031                         struct VhostUserMsg *msg,
2032                         int main_fd __rte_unused)
2033 {
2034         struct virtio_net *dev = *pdev;
2035         uint64_t protocol_features = msg->payload.u64;
2036         uint64_t slave_protocol_features = 0;
2037
2038         if (validate_msg_fds(msg, 0) != 0)
2039                 return RTE_VHOST_MSG_RESULT_ERR;
2040
2041         rte_vhost_driver_get_protocol_features(dev->ifname,
2042                         &slave_protocol_features);
2043         if (protocol_features & ~slave_protocol_features) {
2044                 VHOST_LOG_CONFIG(ERR,
2045                         "(%d) received invalid protocol features.\n",
2046                         dev->vid);
2047                 return RTE_VHOST_MSG_RESULT_ERR;
2048         }
2049
2050         dev->protocol_features = protocol_features;
2051         VHOST_LOG_CONFIG(INFO,
2052                 "negotiated Vhost-user protocol features: 0x%" PRIx64 "\n",
2053                 dev->protocol_features);
2054
2055         return RTE_VHOST_MSG_RESULT_OK;
2056 }
2057
2058 static int
2059 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
2060                         int main_fd __rte_unused)
2061 {
2062         struct virtio_net *dev = *pdev;
2063         int fd = msg->fds[0];
2064         uint64_t size, off;
2065         void *addr;
2066
2067         if (validate_msg_fds(msg, 1) != 0)
2068                 return RTE_VHOST_MSG_RESULT_ERR;
2069
2070         if (fd < 0) {
2071                 VHOST_LOG_CONFIG(ERR, "invalid log fd: %d\n", fd);
2072                 return RTE_VHOST_MSG_RESULT_ERR;
2073         }
2074
2075         if (msg->size != sizeof(VhostUserLog)) {
2076                 VHOST_LOG_CONFIG(ERR,
2077                         "invalid log base msg size: %"PRId32" != %d\n",
2078                         msg->size, (int)sizeof(VhostUserLog));
2079                 return RTE_VHOST_MSG_RESULT_ERR;
2080         }
2081
2082         size = msg->payload.log.mmap_size;
2083         off  = msg->payload.log.mmap_offset;
2084
2085         /* Don't allow mmap_offset to point outside the mmap region */
2086         if (off > size) {
2087                 VHOST_LOG_CONFIG(ERR,
2088                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
2089                         off, size);
2090                 return RTE_VHOST_MSG_RESULT_ERR;
2091         }
2092
2093         VHOST_LOG_CONFIG(INFO,
2094                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
2095                 size, off);
2096
2097         /*
2098          * mmap from 0 to workaround a hugepage mmap bug: mmap will
2099          * fail when offset is not page size aligned.
2100          */
2101         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
2102         close(fd);
2103         if (addr == MAP_FAILED) {
2104                 VHOST_LOG_CONFIG(ERR, "mmap log base failed!\n");
2105                 return RTE_VHOST_MSG_RESULT_ERR;
2106         }
2107
2108         /*
2109          * Free previously mapped log memory on occasionally
2110          * multiple VHOST_USER_SET_LOG_BASE.
2111          */
2112         if (dev->log_addr) {
2113                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
2114         }
2115         dev->log_addr = (uint64_t)(uintptr_t)addr;
2116         dev->log_base = dev->log_addr + off;
2117         dev->log_size = size;
2118
2119         /*
2120          * The spec is not clear about it (yet), but QEMU doesn't expect
2121          * any payload in the reply.
2122          */
2123         msg->size = 0;
2124         msg->fd_num = 0;
2125
2126         return RTE_VHOST_MSG_RESULT_REPLY;
2127 }
2128
2129 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
2130                         struct VhostUserMsg *msg,
2131                         int main_fd __rte_unused)
2132 {
2133         if (validate_msg_fds(msg, 1) != 0)
2134                 return RTE_VHOST_MSG_RESULT_ERR;
2135
2136         close(msg->fds[0]);
2137         VHOST_LOG_CONFIG(INFO, "not implemented.\n");
2138
2139         return RTE_VHOST_MSG_RESULT_OK;
2140 }
2141
2142 /*
2143  * An rarp packet is constructed and broadcasted to notify switches about
2144  * the new location of the migrated VM, so that packets from outside will
2145  * not be lost after migration.
2146  *
2147  * However, we don't actually "send" a rarp packet here, instead, we set
2148  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
2149  */
2150 static int
2151 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
2152                         int main_fd __rte_unused)
2153 {
2154         struct virtio_net *dev = *pdev;
2155         uint8_t *mac = (uint8_t *)&msg->payload.u64;
2156         struct rte_vdpa_device *vdpa_dev;
2157         int did = -1;
2158
2159         if (validate_msg_fds(msg, 0) != 0)
2160                 return RTE_VHOST_MSG_RESULT_ERR;
2161
2162         VHOST_LOG_CONFIG(DEBUG,
2163                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
2164                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
2165         memcpy(dev->mac.addr_bytes, mac, 6);
2166
2167         /*
2168          * Set the flag to inject a RARP broadcast packet at
2169          * rte_vhost_dequeue_burst().
2170          *
2171          * rte_smp_wmb() is for making sure the mac is copied
2172          * before the flag is set.
2173          */
2174         rte_smp_wmb();
2175         rte_atomic16_set(&dev->broadcast_rarp, 1);
2176         did = dev->vdpa_dev_id;
2177         vdpa_dev = rte_vdpa_get_device(did);
2178         if (vdpa_dev && vdpa_dev->ops->migration_done)
2179                 vdpa_dev->ops->migration_done(dev->vid);
2180
2181         return RTE_VHOST_MSG_RESULT_OK;
2182 }
2183
2184 static int
2185 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
2186                         int main_fd __rte_unused)
2187 {
2188         struct virtio_net *dev = *pdev;
2189
2190         if (validate_msg_fds(msg, 0) != 0)
2191                 return RTE_VHOST_MSG_RESULT_ERR;
2192
2193         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
2194                         msg->payload.u64 > VIRTIO_MAX_MTU) {
2195                 VHOST_LOG_CONFIG(ERR, "Invalid MTU size (%"PRIu64")\n",
2196                                 msg->payload.u64);
2197
2198                 return RTE_VHOST_MSG_RESULT_ERR;
2199         }
2200
2201         dev->mtu = msg->payload.u64;
2202
2203         return RTE_VHOST_MSG_RESULT_OK;
2204 }
2205
2206 static int
2207 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
2208                         int main_fd __rte_unused)
2209 {
2210         struct virtio_net *dev = *pdev;
2211         int fd = msg->fds[0];
2212
2213         if (validate_msg_fds(msg, 1) != 0)
2214                 return RTE_VHOST_MSG_RESULT_ERR;
2215
2216         if (fd < 0) {
2217                 VHOST_LOG_CONFIG(ERR,
2218                                 "Invalid file descriptor for slave channel (%d)\n",
2219                                 fd);
2220                 return RTE_VHOST_MSG_RESULT_ERR;
2221         }
2222
2223         if (dev->slave_req_fd >= 0)
2224                 close(dev->slave_req_fd);
2225
2226         dev->slave_req_fd = fd;
2227
2228         return RTE_VHOST_MSG_RESULT_OK;
2229 }
2230
2231 static int
2232 is_vring_iotlb_split(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2233 {
2234         struct vhost_vring_addr *ra;
2235         uint64_t start, end, len;
2236
2237         start = imsg->iova;
2238         end = start + imsg->size;
2239
2240         ra = &vq->ring_addrs;
2241         len = sizeof(struct vring_desc) * vq->size;
2242         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2243                 return 1;
2244
2245         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
2246         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2247                 return 1;
2248
2249         len = sizeof(struct vring_used) +
2250                sizeof(struct vring_used_elem) * vq->size;
2251         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2252                 return 1;
2253
2254         return 0;
2255 }
2256
2257 static int
2258 is_vring_iotlb_packed(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2259 {
2260         struct vhost_vring_addr *ra;
2261         uint64_t start, end, len;
2262
2263         start = imsg->iova;
2264         end = start + imsg->size;
2265
2266         ra = &vq->ring_addrs;
2267         len = sizeof(struct vring_packed_desc) * vq->size;
2268         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2269                 return 1;
2270
2271         len = sizeof(struct vring_packed_desc_event);
2272         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2273                 return 1;
2274
2275         len = sizeof(struct vring_packed_desc_event);
2276         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2277                 return 1;
2278
2279         return 0;
2280 }
2281
2282 static int is_vring_iotlb(struct virtio_net *dev,
2283                           struct vhost_virtqueue *vq,
2284                           struct vhost_iotlb_msg *imsg)
2285 {
2286         if (vq_is_packed(dev))
2287                 return is_vring_iotlb_packed(vq, imsg);
2288         else
2289                 return is_vring_iotlb_split(vq, imsg);
2290 }
2291
2292 static int
2293 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
2294                         int main_fd __rte_unused)
2295 {
2296         struct virtio_net *dev = *pdev;
2297         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
2298         uint16_t i;
2299         uint64_t vva, len;
2300
2301         if (validate_msg_fds(msg, 0) != 0)
2302                 return RTE_VHOST_MSG_RESULT_ERR;
2303
2304         switch (imsg->type) {
2305         case VHOST_IOTLB_UPDATE:
2306                 len = imsg->size;
2307                 vva = qva_to_vva(dev, imsg->uaddr, &len);
2308                 if (!vva)
2309                         return RTE_VHOST_MSG_RESULT_ERR;
2310
2311                 for (i = 0; i < dev->nr_vring; i++) {
2312                         struct vhost_virtqueue *vq = dev->virtqueue[i];
2313
2314                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
2315                                         len, imsg->perm);
2316
2317                         if (is_vring_iotlb(dev, vq, imsg))
2318                                 *pdev = dev = translate_ring_addresses(dev, i);
2319                 }
2320                 break;
2321         case VHOST_IOTLB_INVALIDATE:
2322                 for (i = 0; i < dev->nr_vring; i++) {
2323                         struct vhost_virtqueue *vq = dev->virtqueue[i];
2324
2325                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
2326                                         imsg->size);
2327
2328                         if (is_vring_iotlb(dev, vq, imsg))
2329                                 vring_invalidate(dev, vq);
2330                 }
2331                 break;
2332         default:
2333                 VHOST_LOG_CONFIG(ERR, "Invalid IOTLB message type (%d)\n",
2334                                 imsg->type);
2335                 return RTE_VHOST_MSG_RESULT_ERR;
2336         }
2337
2338         return RTE_VHOST_MSG_RESULT_OK;
2339 }
2340
2341 static int
2342 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
2343                         struct VhostUserMsg *msg,
2344                         int main_fd __rte_unused)
2345 {
2346         struct virtio_net *dev = *pdev;
2347 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
2348         struct uffdio_api api_struct;
2349
2350         if (validate_msg_fds(msg, 0) != 0)
2351                 return RTE_VHOST_MSG_RESULT_ERR;
2352
2353         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
2354
2355         if (dev->postcopy_ufd == -1) {
2356                 VHOST_LOG_CONFIG(ERR, "Userfaultfd not available: %s\n",
2357                         strerror(errno));
2358                 return RTE_VHOST_MSG_RESULT_ERR;
2359         }
2360         api_struct.api = UFFD_API;
2361         api_struct.features = 0;
2362         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
2363                 VHOST_LOG_CONFIG(ERR, "UFFDIO_API ioctl failure: %s\n",
2364                         strerror(errno));
2365                 close(dev->postcopy_ufd);
2366                 dev->postcopy_ufd = -1;
2367                 return RTE_VHOST_MSG_RESULT_ERR;
2368         }
2369         msg->fds[0] = dev->postcopy_ufd;
2370         msg->fd_num = 1;
2371
2372         return RTE_VHOST_MSG_RESULT_REPLY;
2373 #else
2374         dev->postcopy_ufd = -1;
2375         msg->fd_num = 0;
2376
2377         return RTE_VHOST_MSG_RESULT_ERR;
2378 #endif
2379 }
2380
2381 static int
2382 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
2383                         struct VhostUserMsg *msg __rte_unused,
2384                         int main_fd __rte_unused)
2385 {
2386         struct virtio_net *dev = *pdev;
2387
2388         if (validate_msg_fds(msg, 0) != 0)
2389                 return RTE_VHOST_MSG_RESULT_ERR;
2390
2391         if (dev->mem && dev->mem->nregions) {
2392                 VHOST_LOG_CONFIG(ERR,
2393                         "Regions already registered at postcopy-listen\n");
2394                 return RTE_VHOST_MSG_RESULT_ERR;
2395         }
2396         dev->postcopy_listening = 1;
2397
2398         return RTE_VHOST_MSG_RESULT_OK;
2399 }
2400
2401 static int
2402 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
2403                         int main_fd __rte_unused)
2404 {
2405         struct virtio_net *dev = *pdev;
2406
2407         if (validate_msg_fds(msg, 0) != 0)
2408                 return RTE_VHOST_MSG_RESULT_ERR;
2409
2410         dev->postcopy_listening = 0;
2411         if (dev->postcopy_ufd >= 0) {
2412                 close(dev->postcopy_ufd);
2413                 dev->postcopy_ufd = -1;
2414         }
2415
2416         msg->payload.u64 = 0;
2417         msg->size = sizeof(msg->payload.u64);
2418         msg->fd_num = 0;
2419
2420         return RTE_VHOST_MSG_RESULT_REPLY;
2421 }
2422
2423 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
2424                                         struct VhostUserMsg *msg,
2425                                         int main_fd);
2426 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
2427         [VHOST_USER_NONE] = NULL,
2428         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
2429         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
2430         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
2431         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
2432         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
2433         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
2434         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
2435         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
2436         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
2437         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
2438         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
2439         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
2440         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
2441         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
2442         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
2443         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
2444         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
2445         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
2446         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
2447         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
2448         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
2449         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
2450         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
2451         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
2452         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
2453         [VHOST_USER_GET_INFLIGHT_FD] = vhost_user_get_inflight_fd,
2454         [VHOST_USER_SET_INFLIGHT_FD] = vhost_user_set_inflight_fd,
2455 };
2456
2457 /* return bytes# of read on success or negative val on failure. */
2458 static int
2459 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
2460 {
2461         int ret;
2462
2463         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
2464                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
2465         if (ret <= 0) {
2466                 return ret;
2467         } else if (ret != VHOST_USER_HDR_SIZE) {
2468                 VHOST_LOG_CONFIG(ERR, "Unexpected header size read\n");
2469                 close_msg_fds(msg);
2470                 return -1;
2471         }
2472
2473         if (msg->size) {
2474                 if (msg->size > sizeof(msg->payload)) {
2475                         VHOST_LOG_CONFIG(ERR,
2476                                 "invalid msg size: %d\n", msg->size);
2477                         return -1;
2478                 }
2479                 ret = read(sockfd, &msg->payload, msg->size);
2480                 if (ret <= 0)
2481                         return ret;
2482                 if (ret != (int)msg->size) {
2483                         VHOST_LOG_CONFIG(ERR,
2484                                 "read control message failed\n");
2485                         return -1;
2486                 }
2487         }
2488
2489         return ret;
2490 }
2491
2492 static int
2493 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
2494 {
2495         if (!msg)
2496                 return 0;
2497
2498         return send_fd_message(sockfd, (char *)msg,
2499                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
2500 }
2501
2502 static int
2503 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
2504 {
2505         if (!msg)
2506                 return 0;
2507
2508         msg->flags &= ~VHOST_USER_VERSION_MASK;
2509         msg->flags &= ~VHOST_USER_NEED_REPLY;
2510         msg->flags |= VHOST_USER_VERSION;
2511         msg->flags |= VHOST_USER_REPLY_MASK;
2512
2513         return send_vhost_message(sockfd, msg);
2514 }
2515
2516 static int
2517 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
2518 {
2519         int ret;
2520
2521         if (msg->flags & VHOST_USER_NEED_REPLY)
2522                 rte_spinlock_lock(&dev->slave_req_lock);
2523
2524         ret = send_vhost_message(dev->slave_req_fd, msg);
2525         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
2526                 rte_spinlock_unlock(&dev->slave_req_lock);
2527
2528         return ret;
2529 }
2530
2531 /*
2532  * Allocate a queue pair if it hasn't been allocated yet
2533  */
2534 static int
2535 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
2536                         struct VhostUserMsg *msg)
2537 {
2538         uint16_t vring_idx;
2539
2540         switch (msg->request.master) {
2541         case VHOST_USER_SET_VRING_KICK:
2542         case VHOST_USER_SET_VRING_CALL:
2543         case VHOST_USER_SET_VRING_ERR:
2544                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
2545                 break;
2546         case VHOST_USER_SET_VRING_NUM:
2547         case VHOST_USER_SET_VRING_BASE:
2548         case VHOST_USER_SET_VRING_ENABLE:
2549                 vring_idx = msg->payload.state.index;
2550                 break;
2551         case VHOST_USER_SET_VRING_ADDR:
2552                 vring_idx = msg->payload.addr.index;
2553                 break;
2554         default:
2555                 return 0;
2556         }
2557
2558         if (vring_idx >= VHOST_MAX_VRING) {
2559                 VHOST_LOG_CONFIG(ERR,
2560                         "invalid vring index: %u\n", vring_idx);
2561                 return -1;
2562         }
2563
2564         if (dev->virtqueue[vring_idx])
2565                 return 0;
2566
2567         return alloc_vring_queue(dev, vring_idx);
2568 }
2569
2570 static void
2571 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
2572 {
2573         unsigned int i = 0;
2574         unsigned int vq_num = 0;
2575
2576         while (vq_num < dev->nr_vring) {
2577                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2578
2579                 if (vq) {
2580                         rte_spinlock_lock(&vq->access_lock);
2581                         vq_num++;
2582                 }
2583                 i++;
2584         }
2585 }
2586
2587 static void
2588 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
2589 {
2590         unsigned int i = 0;
2591         unsigned int vq_num = 0;
2592
2593         while (vq_num < dev->nr_vring) {
2594                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2595
2596                 if (vq) {
2597                         rte_spinlock_unlock(&vq->access_lock);
2598                         vq_num++;
2599                 }
2600                 i++;
2601         }
2602 }
2603
2604 int
2605 vhost_user_msg_handler(int vid, int fd)
2606 {
2607         struct virtio_net *dev;
2608         struct VhostUserMsg msg;
2609         struct rte_vdpa_device *vdpa_dev;
2610         int did = -1;
2611         int ret;
2612         int unlock_required = 0;
2613         bool handled;
2614         int request;
2615
2616         dev = get_device(vid);
2617         if (dev == NULL)
2618                 return -1;
2619
2620         if (!dev->notify_ops) {
2621                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
2622                 if (!dev->notify_ops) {
2623                         VHOST_LOG_CONFIG(ERR,
2624                                 "failed to get callback ops for driver %s\n",
2625                                 dev->ifname);
2626                         return -1;
2627                 }
2628         }
2629
2630         ret = read_vhost_message(fd, &msg);
2631         if (ret <= 0) {
2632                 if (ret < 0)
2633                         VHOST_LOG_CONFIG(ERR,
2634                                 "vhost read message failed\n");
2635                 else
2636                         VHOST_LOG_CONFIG(INFO,
2637                                 "vhost peer closed\n");
2638
2639                 return -1;
2640         }
2641
2642         ret = 0;
2643         request = msg.request.master;
2644         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
2645                         vhost_message_str[request]) {
2646                 if (request != VHOST_USER_IOTLB_MSG)
2647                         VHOST_LOG_CONFIG(INFO, "read message %s\n",
2648                                 vhost_message_str[request]);
2649                 else
2650                         VHOST_LOG_CONFIG(DEBUG, "read message %s\n",
2651                                 vhost_message_str[request]);
2652         } else {
2653                 VHOST_LOG_CONFIG(DEBUG, "External request %d\n", request);
2654         }
2655
2656         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
2657         if (ret < 0) {
2658                 VHOST_LOG_CONFIG(ERR,
2659                         "failed to alloc queue\n");
2660                 return -1;
2661         }
2662
2663         /*
2664          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
2665          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
2666          * and device is destroyed. destroy_device waits for queues to be
2667          * inactive, so it is safe. Otherwise taking the access_lock
2668          * would cause a dead lock.
2669          */
2670         switch (request) {
2671         case VHOST_USER_SET_FEATURES:
2672         case VHOST_USER_SET_PROTOCOL_FEATURES:
2673         case VHOST_USER_SET_OWNER:
2674         case VHOST_USER_SET_MEM_TABLE:
2675         case VHOST_USER_SET_LOG_BASE:
2676         case VHOST_USER_SET_LOG_FD:
2677         case VHOST_USER_SET_VRING_NUM:
2678         case VHOST_USER_SET_VRING_ADDR:
2679         case VHOST_USER_SET_VRING_BASE:
2680         case VHOST_USER_SET_VRING_KICK:
2681         case VHOST_USER_SET_VRING_CALL:
2682         case VHOST_USER_SET_VRING_ERR:
2683         case VHOST_USER_SET_VRING_ENABLE:
2684         case VHOST_USER_SEND_RARP:
2685         case VHOST_USER_NET_SET_MTU:
2686         case VHOST_USER_SET_SLAVE_REQ_FD:
2687                 vhost_user_lock_all_queue_pairs(dev);
2688                 unlock_required = 1;
2689                 break;
2690         default:
2691                 break;
2692
2693         }
2694
2695         handled = false;
2696         if (dev->extern_ops.pre_msg_handle) {
2697                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
2698                                 (void *)&msg);
2699                 switch (ret) {
2700                 case RTE_VHOST_MSG_RESULT_REPLY:
2701                         send_vhost_reply(fd, &msg);
2702                         /* Fall-through */
2703                 case RTE_VHOST_MSG_RESULT_ERR:
2704                 case RTE_VHOST_MSG_RESULT_OK:
2705                         handled = true;
2706                         goto skip_to_post_handle;
2707                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2708                 default:
2709                         break;
2710                 }
2711         }
2712
2713         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
2714                 if (!vhost_message_handlers[request])
2715                         goto skip_to_post_handle;
2716                 ret = vhost_message_handlers[request](&dev, &msg, fd);
2717
2718                 switch (ret) {
2719                 case RTE_VHOST_MSG_RESULT_ERR:
2720                         VHOST_LOG_CONFIG(ERR,
2721                                 "Processing %s failed.\n",
2722                                 vhost_message_str[request]);
2723                         handled = true;
2724                         break;
2725                 case RTE_VHOST_MSG_RESULT_OK:
2726                         VHOST_LOG_CONFIG(DEBUG,
2727                                 "Processing %s succeeded.\n",
2728                                 vhost_message_str[request]);
2729                         handled = true;
2730                         break;
2731                 case RTE_VHOST_MSG_RESULT_REPLY:
2732                         VHOST_LOG_CONFIG(DEBUG,
2733                                 "Processing %s succeeded and needs reply.\n",
2734                                 vhost_message_str[request]);
2735                         send_vhost_reply(fd, &msg);
2736                         handled = true;
2737                         break;
2738                 default:
2739                         break;
2740                 }
2741         }
2742
2743 skip_to_post_handle:
2744         if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2745                         dev->extern_ops.post_msg_handle) {
2746                 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
2747                                 (void *)&msg);
2748                 switch (ret) {
2749                 case RTE_VHOST_MSG_RESULT_REPLY:
2750                         send_vhost_reply(fd, &msg);
2751                         /* Fall-through */
2752                 case RTE_VHOST_MSG_RESULT_ERR:
2753                 case RTE_VHOST_MSG_RESULT_OK:
2754                         handled = true;
2755                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2756                 default:
2757                         break;
2758                 }
2759         }
2760
2761         if (unlock_required)
2762                 vhost_user_unlock_all_queue_pairs(dev);
2763
2764         /* If message was not handled at this stage, treat it as an error */
2765         if (!handled) {
2766                 VHOST_LOG_CONFIG(ERR,
2767                         "vhost message (req: %d) was not handled.\n", request);
2768                 close_msg_fds(&msg);
2769                 ret = RTE_VHOST_MSG_RESULT_ERR;
2770         }
2771
2772         /*
2773          * If the request required a reply that was already sent,
2774          * this optional reply-ack won't be sent as the
2775          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2776          */
2777         if (msg.flags & VHOST_USER_NEED_REPLY) {
2778                 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2779                 msg.size = sizeof(msg.payload.u64);
2780                 msg.fd_num = 0;
2781                 send_vhost_reply(fd, &msg);
2782         } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2783                 VHOST_LOG_CONFIG(ERR,
2784                         "vhost message handling failed.\n");
2785                 return -1;
2786         }
2787
2788         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2789                 dev->flags |= VIRTIO_DEV_READY;
2790
2791                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2792                         if (dev->dequeue_zero_copy) {
2793                                 VHOST_LOG_CONFIG(INFO,
2794                                                 "dequeue zero copy is enabled\n");
2795                         }
2796
2797                         if (dev->notify_ops->new_device(dev->vid) == 0)
2798                                 dev->flags |= VIRTIO_DEV_RUNNING;
2799                 }
2800         }
2801
2802         did = dev->vdpa_dev_id;
2803         vdpa_dev = rte_vdpa_get_device(did);
2804         if (vdpa_dev && virtio_is_ready(dev) &&
2805                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2806                         msg.request.master == VHOST_USER_SET_VRING_CALL) {
2807                 if (vdpa_dev->ops->dev_conf)
2808                         vdpa_dev->ops->dev_conf(dev->vid);
2809                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2810         }
2811
2812         return 0;
2813 }
2814
2815 static int process_slave_message_reply(struct virtio_net *dev,
2816                                        const struct VhostUserMsg *msg)
2817 {
2818         struct VhostUserMsg msg_reply;
2819         int ret;
2820
2821         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2822                 return 0;
2823
2824         if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
2825                 ret = -1;
2826                 goto out;
2827         }
2828
2829         if (msg_reply.request.slave != msg->request.slave) {
2830                 VHOST_LOG_CONFIG(ERR,
2831                         "Received unexpected msg type (%u), expected %u\n",
2832                         msg_reply.request.slave, msg->request.slave);
2833                 ret = -1;
2834                 goto out;
2835         }
2836
2837         ret = msg_reply.payload.u64 ? -1 : 0;
2838
2839 out:
2840         rte_spinlock_unlock(&dev->slave_req_lock);
2841         return ret;
2842 }
2843
2844 int
2845 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2846 {
2847         int ret;
2848         struct VhostUserMsg msg = {
2849                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2850                 .flags = VHOST_USER_VERSION,
2851                 .size = sizeof(msg.payload.iotlb),
2852                 .payload.iotlb = {
2853                         .iova = iova,
2854                         .perm = perm,
2855                         .type = VHOST_IOTLB_MISS,
2856                 },
2857         };
2858
2859         ret = send_vhost_message(dev->slave_req_fd, &msg);
2860         if (ret < 0) {
2861                 VHOST_LOG_CONFIG(ERR,
2862                                 "Failed to send IOTLB miss message (%d)\n",
2863                                 ret);
2864                 return ret;
2865         }
2866
2867         return 0;
2868 }
2869
2870 static int
2871 vhost_user_slave_config_change(struct virtio_net *dev, bool need_reply)
2872 {
2873         int ret;
2874         struct VhostUserMsg msg = {
2875                 .request.slave = VHOST_USER_SLAVE_CONFIG_CHANGE_MSG,
2876                 .flags = VHOST_USER_VERSION,
2877                 .size = 0,
2878         };
2879
2880         if (need_reply)
2881                 msg.flags |= VHOST_USER_NEED_REPLY;
2882
2883         ret = send_vhost_slave_message(dev, &msg);
2884         if (ret < 0) {
2885                 VHOST_LOG_CONFIG(ERR,
2886                                 "Failed to send config change (%d)\n",
2887                                 ret);
2888                 return ret;
2889         }
2890
2891         return process_slave_message_reply(dev, &msg);
2892 }
2893
2894 int
2895 rte_vhost_slave_config_change(int vid, bool need_reply)
2896 {
2897         struct virtio_net *dev;
2898
2899         dev = get_device(vid);
2900         if (!dev)
2901                 return -ENODEV;
2902
2903         return vhost_user_slave_config_change(dev, need_reply);
2904 }
2905
2906 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2907                                                     int index, int fd,
2908                                                     uint64_t offset,
2909                                                     uint64_t size)
2910 {
2911         int ret;
2912         struct VhostUserMsg msg = {
2913                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2914                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2915                 .size = sizeof(msg.payload.area),
2916                 .payload.area = {
2917                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
2918                         .size = size,
2919                         .offset = offset,
2920                 },
2921         };
2922
2923         if (fd < 0)
2924                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2925         else {
2926                 msg.fds[0] = fd;
2927                 msg.fd_num = 1;
2928         }
2929
2930         ret = send_vhost_slave_message(dev, &msg);
2931         if (ret < 0) {
2932                 VHOST_LOG_CONFIG(ERR,
2933                         "Failed to set host notifier (%d)\n", ret);
2934                 return ret;
2935         }
2936
2937         return process_slave_message_reply(dev, &msg);
2938 }
2939
2940 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2941 {
2942         struct virtio_net *dev;
2943         struct rte_vdpa_device *vdpa_dev;
2944         int vfio_device_fd, did, ret = 0;
2945         uint64_t offset, size;
2946         unsigned int i;
2947
2948         dev = get_device(vid);
2949         if (!dev)
2950                 return -ENODEV;
2951
2952         did = dev->vdpa_dev_id;
2953         if (did < 0)
2954                 return -EINVAL;
2955
2956         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2957             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2958             !(dev->protocol_features &
2959                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2960             !(dev->protocol_features &
2961                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2962             !(dev->protocol_features &
2963                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2964                 return -ENOTSUP;
2965
2966         vdpa_dev = rte_vdpa_get_device(did);
2967         if (!vdpa_dev)
2968                 return -ENODEV;
2969
2970         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2971         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2972
2973         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2974         if (vfio_device_fd < 0)
2975                 return -ENOTSUP;
2976
2977         if (enable) {
2978                 for (i = 0; i < dev->nr_vring; i++) {
2979                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2980                                         &size) < 0) {
2981                                 ret = -ENOTSUP;
2982                                 goto disable;
2983                         }
2984
2985                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2986                                         vfio_device_fd, offset, size) < 0) {
2987                                 ret = -EFAULT;
2988                                 goto disable;
2989                         }
2990                 }
2991         } else {
2992 disable:
2993                 for (i = 0; i < dev->nr_vring; i++) {
2994                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
2995                                         0, 0);
2996                 }
2997         }
2998
2999         return ret;
3000 }