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         rte_free(dev->guest_pages);
 195         dev->guest_pages = NULL;
 196
 197         if (dev->log_addr) {
 198                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 199                 dev->log_addr = 0;
 200         }
 201
 202         if (dev->inflight_info) {
 203                 if (dev->inflight_info->addr) {
 204                         munmap(dev->inflight_info->addr,
 205                                dev->inflight_info->size);
 206                         dev->inflight_info->addr = NULL;
 207                 }
 208
 209                 if (dev->inflight_info->fd > 0) {
 210                         close(dev->inflight_info->fd);
 211                         dev->inflight_info->fd = -1;
 212                 }
 213
 214                 free(dev->inflight_info);
 215                 dev->inflight_info = NULL;
 216         }
 217
 218         if (dev->slave_req_fd >= 0) {
 219                 close(dev->slave_req_fd);
 220                 dev->slave_req_fd = -1;
 221         }
 222
 223         if (dev->postcopy_ufd >= 0) {
 224                 close(dev->postcopy_ufd);
 225                 dev->postcopy_ufd = -1;
 226         }
 227
 228         dev->postcopy_listening = 0;
 229 }
 230
 231 /*
 232  * This function just returns success at the moment unless
 233  * the device hasn't been initialised.
 234  */
 235 static int
 236 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
 237                         struct VhostUserMsg *msg,
 238                         int main_fd __rte_unused)
 239 {
 240         if (validate_msg_fds(msg, 0) != 0)
 241                 return RTE_VHOST_MSG_RESULT_ERR;
 242
 243         return RTE_VHOST_MSG_RESULT_OK;
 244 }
 245
 246 static int
 247 vhost_user_reset_owner(struct virtio_net **pdev,
 248                         struct VhostUserMsg *msg,
 249                         int main_fd __rte_unused)
 250 {
 251         struct virtio_net *dev = *pdev;
 252
 253         if (validate_msg_fds(msg, 0) != 0)
 254                 return RTE_VHOST_MSG_RESULT_ERR;
 255
 256         vhost_destroy_device_notify(dev);
 257
 258         cleanup_device(dev, 0);
 259         reset_device(dev);
 260         return RTE_VHOST_MSG_RESULT_OK;
 261 }
 262
 263 /*
 264  * The features that we support are requested.
 265  */
 266 static int
 267 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 268                         int main_fd __rte_unused)
 269 {
 270         struct virtio_net *dev = *pdev;
 271         uint64_t features = 0;
 272
 273         if (validate_msg_fds(msg, 0) != 0)
 274                 return RTE_VHOST_MSG_RESULT_ERR;
 275
 276         rte_vhost_driver_get_features(dev->ifname, &features);
 277
 278         msg->payload.u64 = features;
 279         msg->size = sizeof(msg->payload.u64);
 280         msg->fd_num = 0;
 281
 282         return RTE_VHOST_MSG_RESULT_REPLY;
 283 }
 284
 285 /*
 286  * The queue number that we support are requested.
 287  */
 288 static int
 289 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
 290                         int main_fd __rte_unused)
 291 {
 292         struct virtio_net *dev = *pdev;
 293         uint32_t queue_num = 0;
 294
 295         if (validate_msg_fds(msg, 0) != 0)
 296                 return RTE_VHOST_MSG_RESULT_ERR;
 297
 298         rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
 299
 300         msg->payload.u64 = (uint64_t)queue_num;
 301         msg->size = sizeof(msg->payload.u64);
 302         msg->fd_num = 0;
 303
 304         return RTE_VHOST_MSG_RESULT_REPLY;
 305 }
 306
 307 /*
 308  * We receive the negotiated features supported by us and the virtio device.
 309  */
 310 static int
 311 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 312                         int main_fd __rte_unused)
 313 {
 314         struct virtio_net *dev = *pdev;
 315         uint64_t features = msg->payload.u64;
 316         uint64_t vhost_features = 0;
 317         struct rte_vdpa_device *vdpa_dev;
 318         int did = -1;
 319
 320         if (validate_msg_fds(msg, 0) != 0)
 321                 return RTE_VHOST_MSG_RESULT_ERR;
 322
 323         rte_vhost_driver_get_features(dev->ifname, &vhost_features);
 324         if (features & ~vhost_features) {
 325                 VHOST_LOG_CONFIG(ERR,
 326                         "(%d) received invalid negotiated features.\n",
 327                         dev->vid);
 328                 return RTE_VHOST_MSG_RESULT_ERR;
 329         }
 330
 331         if (dev->flags & VIRTIO_DEV_RUNNING) {
 332                 if (dev->features == features)
 333                         return RTE_VHOST_MSG_RESULT_OK;
 334
 335                 /*
 336                  * Error out if master tries to change features while device is
 337                  * in running state. The exception being VHOST_F_LOG_ALL, which
 338                  * is enabled when the live-migration starts.
 339                  */
 340                 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
 341                         VHOST_LOG_CONFIG(ERR,
 342                                 "(%d) features changed while device is running.\n",
 343                                 dev->vid);
 344                         return RTE_VHOST_MSG_RESULT_ERR;
 345                 }
 346
 347                 if (dev->notify_ops->features_changed)
 348                         dev->notify_ops->features_changed(dev->vid, features);
 349         }
 350
 351         dev->features = features;
 352         if (dev->features &
 353                 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
 354                 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 355         } else {
 356                 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
 357         }
 358         VHOST_LOG_CONFIG(INFO,
 359                 "negotiated Virtio features: 0x%" PRIx64 "\n", dev->features);
 360         VHOST_LOG_CONFIG(DEBUG,
 361                 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
 362                 dev->vid,
 363                 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
 364                 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 365
 366         if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
 367             !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
 368                 /*
 369                  * Remove all but first queue pair if MQ hasn't been
 370                  * negotiated. This is safe because the device is not
 371                  * running at this stage.
 372                  */
 373                 while (dev->nr_vring > 2) {
 374                         struct vhost_virtqueue *vq;
 375
 376                         vq = dev->virtqueue[--dev->nr_vring];
 377                         if (!vq)
 378                                 continue;
 379
 380                         dev->virtqueue[dev->nr_vring] = NULL;
 381                         cleanup_vq(vq, 1);
 382                         cleanup_vq_inflight(dev, vq);
 383                         free_vq(dev, vq);
 384                 }
 385         }
 386
 387         did = dev->vdpa_dev_id;
 388         vdpa_dev = rte_vdpa_get_device(did);
 389         if (vdpa_dev && vdpa_dev->ops->set_features)
 390                 vdpa_dev->ops->set_features(dev->vid);
 391
 392         return RTE_VHOST_MSG_RESULT_OK;
 393 }
 394
 395 /*
 396  * The virtio device sends us the size of the descriptor ring.
 397  */
 398 static int
 399 vhost_user_set_vring_num(struct virtio_net **pdev,
 400                         struct VhostUserMsg *msg,
 401                         int main_fd __rte_unused)
 402 {
 403         struct virtio_net *dev = *pdev;
 404         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 405
 406         if (validate_msg_fds(msg, 0) != 0)
 407                 return RTE_VHOST_MSG_RESULT_ERR;
 408
 409         vq->size = msg->payload.state.num;
 410
 411         /* VIRTIO 1.0, 2.4 Virtqueues says:
 412          *
 413          *   Queue Size value is always a power of 2. The maximum Queue Size
 414          *   value is 32768.
 415          *
 416          * VIRTIO 1.1 2.7 Virtqueues says:
 417          *
 418          *   Packed virtqueues support up to 2^15 entries each.
 419          */
 420         if (!vq_is_packed(dev)) {
 421                 if (vq->size & (vq->size - 1)) {
 422                         VHOST_LOG_CONFIG(ERR,
 423                                 "invalid virtqueue size %u\n", vq->size);
 424                         return RTE_VHOST_MSG_RESULT_ERR;
 425                 }
 426         }
 427
 428         if (vq->size > 32768) {
 429                 VHOST_LOG_CONFIG(ERR,
 430                         "invalid virtqueue size %u\n", vq->size);
 431                 return RTE_VHOST_MSG_RESULT_ERR;
 432         }
 433
 434         if (dev->dequeue_zero_copy) {
 435                 vq->nr_zmbuf = 0;
 436                 vq->last_zmbuf_idx = 0;
 437                 vq->zmbuf_size = vq->size;
 438                 if (vq->zmbufs)
 439                         rte_free(vq->zmbufs);
 440                 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
 441                                          sizeof(struct zcopy_mbuf), 0);
 442                 if (vq->zmbufs == NULL) {
 443                         VHOST_LOG_CONFIG(WARNING,
 444                                 "failed to allocate mem for zero copy; "
 445                                 "zero copy is force disabled\n");
 446                         dev->dequeue_zero_copy = 0;
 447                 }
 448                 TAILQ_INIT(&vq->zmbuf_list);
 449         }
 450
 451         if (vq_is_packed(dev)) {
 452                 if (vq->shadow_used_packed)
 453                         rte_free(vq->shadow_used_packed);
 454                 vq->shadow_used_packed = rte_malloc(NULL,
 455                                 vq->size *
 456                                 sizeof(struct vring_used_elem_packed),
 457                                 RTE_CACHE_LINE_SIZE);
 458                 if (!vq->shadow_used_packed) {
 459                         VHOST_LOG_CONFIG(ERR,
 460                                         "failed to allocate memory for shadow used ring.\n");
 461                         return RTE_VHOST_MSG_RESULT_ERR;
 462                 }
 463
 464         } else {
 465                 if (vq->shadow_used_split)
 466                         rte_free(vq->shadow_used_split);
 467                 vq->shadow_used_split = rte_malloc(NULL,
 468                                 vq->size * sizeof(struct vring_used_elem),
 469                                 RTE_CACHE_LINE_SIZE);
 470                 if (!vq->shadow_used_split) {
 471                         VHOST_LOG_CONFIG(ERR,
 472                                         "failed to allocate memory for shadow used ring.\n");
 473                         return RTE_VHOST_MSG_RESULT_ERR;
 474                 }
 475         }
 476
 477         if (vq->batch_copy_elems)
 478                 rte_free(vq->batch_copy_elems);
 479         vq->batch_copy_elems = rte_malloc(NULL,
 480                                 vq->size * sizeof(struct batch_copy_elem),
 481                                 RTE_CACHE_LINE_SIZE);
 482         if (!vq->batch_copy_elems) {
 483                 VHOST_LOG_CONFIG(ERR,
 484                         "failed to allocate memory for batching copy.\n");
 485                 return RTE_VHOST_MSG_RESULT_ERR;
 486         }
 487
 488         return RTE_VHOST_MSG_RESULT_OK;
 489 }
 490
 491 /*
 492  * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
 493  * same numa node as the memory of vring descriptor.
 494  */
 495 #ifdef RTE_LIBRTE_VHOST_NUMA
 496 static struct virtio_net*
 497 numa_realloc(struct virtio_net *dev, int index)
 498 {
 499         int oldnode, newnode;
 500         struct virtio_net *old_dev;
 501         struct vhost_virtqueue *old_vq, *vq;
 502         struct zcopy_mbuf *new_zmbuf;
 503         struct vring_used_elem *new_shadow_used_split;
 504         struct vring_used_elem_packed *new_shadow_used_packed;
 505         struct batch_copy_elem *new_batch_copy_elems;
 506         int ret;
 507
 508         if (dev->flags & VIRTIO_DEV_RUNNING)
 509                 return dev;
 510
 511         old_dev = dev;
 512         vq = old_vq = dev->virtqueue[index];
 513
 514         ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
 515                             MPOL_F_NODE | MPOL_F_ADDR);
 516
 517         /* check if we need to reallocate vq */
 518         ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
 519                              MPOL_F_NODE | MPOL_F_ADDR);
 520         if (ret) {
 521                 VHOST_LOG_CONFIG(ERR,
 522                         "Unable to get vq numa information.\n");
 523                 return dev;
 524         }
 525         if (oldnode != newnode) {
 526                 VHOST_LOG_CONFIG(INFO,
 527                         "reallocate vq from %d to %d node\n", oldnode, newnode);
 528                 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
 529                 if (!vq)
 530                         return dev;
 531
 532                 memcpy(vq, old_vq, sizeof(*vq));
 533                 TAILQ_INIT(&vq->zmbuf_list);
 534
 535                 if (dev->dequeue_zero_copy) {
 536                         new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
 537                                         sizeof(struct zcopy_mbuf), 0, newnode);
 538                         if (new_zmbuf) {
 539                                 rte_free(vq->zmbufs);
 540                                 vq->zmbufs = new_zmbuf;
 541                         }
 542                 }
 543
 544                 if (vq_is_packed(dev)) {
 545                         new_shadow_used_packed = rte_malloc_socket(NULL,
 546                                         vq->size *
 547                                         sizeof(struct vring_used_elem_packed),
 548                                         RTE_CACHE_LINE_SIZE,
 549                                         newnode);
 550                         if (new_shadow_used_packed) {
 551                                 rte_free(vq->shadow_used_packed);
 552                                 vq->shadow_used_packed = new_shadow_used_packed;
 553                         }
 554                 } else {
 555                         new_shadow_used_split = rte_malloc_socket(NULL,
 556                                         vq->size *
 557                                         sizeof(struct vring_used_elem),
 558                                         RTE_CACHE_LINE_SIZE,
 559                                         newnode);
 560                         if (new_shadow_used_split) {
 561                                 rte_free(vq->shadow_used_split);
 562                                 vq->shadow_used_split = new_shadow_used_split;
 563                         }
 564                 }
 565
 566                 new_batch_copy_elems = rte_malloc_socket(NULL,
 567                         vq->size * sizeof(struct batch_copy_elem),
 568                         RTE_CACHE_LINE_SIZE,
 569                         newnode);
 570                 if (new_batch_copy_elems) {
 571                         rte_free(vq->batch_copy_elems);
 572                         vq->batch_copy_elems = new_batch_copy_elems;
 573                 }
 574
 575                 rte_free(old_vq);
 576         }
 577
 578         /* check if we need to reallocate dev */
 579         ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
 580                             MPOL_F_NODE | MPOL_F_ADDR);
 581         if (ret) {
 582                 VHOST_LOG_CONFIG(ERR,
 583                         "Unable to get dev numa information.\n");
 584                 goto out;
 585         }
 586         if (oldnode != newnode) {
 587                 VHOST_LOG_CONFIG(INFO,
 588                         "reallocate dev from %d to %d node\n",
 589                         oldnode, newnode);
 590                 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
 591                 if (!dev) {
 592                         dev = old_dev;
 593                         goto out;
 594                 }
 595
 596                 memcpy(dev, old_dev, sizeof(*dev));
 597                 rte_free(old_dev);
 598         }
 599
 600 out:
 601         dev->virtqueue[index] = vq;
 602         vhost_devices[dev->vid] = dev;
 603
 604         if (old_vq != vq)
 605                 vhost_user_iotlb_init(dev, index);
 606
 607         return dev;
 608 }
 609 #else
 610 static struct virtio_net*
 611 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 612 {
 613         return dev;
 614 }
 615 #endif
 616
 617 /* Converts QEMU virtual address to Vhost virtual address. */
 618 static uint64_t
 619 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
 620 {
 621         struct rte_vhost_mem_region *r;
 622         uint32_t i;
 623
 624         if (unlikely(!dev || !dev->mem))
 625                 goto out_error;
 626
 627         /* Find the region where the address lives. */
 628         for (i = 0; i < dev->mem->nregions; i++) {
 629                 r = &dev->mem->regions[i];
 630
 631                 if (qva >= r->guest_user_addr &&
 632                     qva <  r->guest_user_addr + r->size) {
 633
 634                         if (unlikely(*len > r->guest_user_addr + r->size - qva))
 635                                 *len = r->guest_user_addr + r->size - qva;
 636
 637                         return qva - r->guest_user_addr +
 638                                r->host_user_addr;
 639                 }
 640         }
 641 out_error:
 642         *len = 0;
 643
 644         return 0;
 645 }
 646
 647
 648 /*
 649  * Converts ring address to Vhost virtual address.
 650  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 651  * else it is a QEMU virtual address.
 652  */
 653 static uint64_t
 654 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 655                 uint64_t ra, uint64_t *size)
 656 {
 657         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 658                 uint64_t vva;
 659
 660                 vhost_user_iotlb_rd_lock(vq);
 661                 vva = vhost_iova_to_vva(dev, vq, ra,
 662                                         size, VHOST_ACCESS_RW);
 663                 vhost_user_iotlb_rd_unlock(vq);
 664
 665                 return vva;
 666         }
 667
 668         return qva_to_vva(dev, ra, size);
 669 }
 670
 671 static uint64_t
 672 log_addr_to_gpa(struct virtio_net *dev, struct vhost_virtqueue *vq)
 673 {
 674         uint64_t log_gpa;
 675
 676         vhost_user_iotlb_rd_lock(vq);
 677         log_gpa = translate_log_addr(dev, vq, vq->ring_addrs.log_guest_addr);
 678         vhost_user_iotlb_rd_unlock(vq);
 679
 680         return log_gpa;
 681 }
 682
 683 static struct virtio_net *
 684 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 685 {
 686         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 687         struct vhost_vring_addr *addr = &vq->ring_addrs;
 688         uint64_t len, expected_len;
 689
 690         if (addr->flags & (1 << VHOST_VRING_F_LOG)) {
 691                 vq->log_guest_addr =
 692                         log_addr_to_gpa(dev, vq);
 693                 if (vq->log_guest_addr == 0) {
 694                         VHOST_LOG_CONFIG(DEBUG,
 695                                 "(%d) failed to map log_guest_addr.\n",
 696                                 dev->vid);
 697                         return dev;
 698                 }
 699         }
 700
 701         if (vq_is_packed(dev)) {
 702                 len = sizeof(struct vring_packed_desc) * vq->size;
 703                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 704                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 705                 if (vq->desc_packed == NULL ||
 706                                 len != sizeof(struct vring_packed_desc) *
 707                                 vq->size) {
 708                         VHOST_LOG_CONFIG(DEBUG,
 709                                 "(%d) failed to map desc_packed ring.\n",
 710                                 dev->vid);
 711                         return dev;
 712                 }
 713
 714                 dev = numa_realloc(dev, vq_index);
 715                 vq = dev->virtqueue[vq_index];
 716                 addr = &vq->ring_addrs;
 717
 718                 len = sizeof(struct vring_packed_desc_event);
 719                 vq->driver_event = (struct vring_packed_desc_event *)
 720                                         (uintptr_t)ring_addr_to_vva(dev,
 721                                         vq, addr->avail_user_addr, &len);
 722                 if (vq->driver_event == NULL ||
 723                                 len != sizeof(struct vring_packed_desc_event)) {
 724                         VHOST_LOG_CONFIG(DEBUG,
 725                                 "(%d) failed to find driver area address.\n",
 726                                 dev->vid);
 727                         return dev;
 728                 }
 729
 730                 len = sizeof(struct vring_packed_desc_event);
 731                 vq->device_event = (struct vring_packed_desc_event *)
 732                                         (uintptr_t)ring_addr_to_vva(dev,
 733                                         vq, addr->used_user_addr, &len);
 734                 if (vq->device_event == NULL ||
 735                                 len != sizeof(struct vring_packed_desc_event)) {
 736                         VHOST_LOG_CONFIG(DEBUG,
 737                                 "(%d) failed to find device area address.\n",
 738                                 dev->vid);
 739                         return dev;
 740                 }
 741
 742                 vq->access_ok = 1;
 743                 return dev;
 744         }
 745
 746         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 747         if (vq->desc && vq->avail && vq->used)
 748                 return dev;
 749
 750         len = sizeof(struct vring_desc) * vq->size;
 751         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 752                         vq, addr->desc_user_addr, &len);
 753         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 754                 VHOST_LOG_CONFIG(DEBUG,
 755                         "(%d) failed to map desc ring.\n",
 756                         dev->vid);
 757                 return dev;
 758         }
 759
 760         dev = numa_realloc(dev, vq_index);
 761         vq = dev->virtqueue[vq_index];
 762         addr = &vq->ring_addrs;
 763
 764         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 765         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 766                 len += sizeof(uint16_t);
 767         expected_len = len;
 768         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 769                         vq, addr->avail_user_addr, &len);
 770         if (vq->avail == 0 || len != expected_len) {
 771                 VHOST_LOG_CONFIG(DEBUG,
 772                         "(%d) failed to map avail ring.\n",
 773                         dev->vid);
 774                 return dev;
 775         }
 776
 777         len = sizeof(struct vring_used) +
 778                 sizeof(struct vring_used_elem) * vq->size;
 779         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 780                 len += sizeof(uint16_t);
 781         expected_len = len;
 782         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 783                         vq, addr->used_user_addr, &len);
 784         if (vq->used == 0 || len != expected_len) {
 785                 VHOST_LOG_CONFIG(DEBUG,
 786                         "(%d) failed to map used ring.\n",
 787                         dev->vid);
 788                 return dev;
 789         }
 790
 791         if (vq->last_used_idx != vq->used->idx) {
 792                 VHOST_LOG_CONFIG(WARNING,
 793                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 794                         "some packets maybe resent for Tx and dropped for Rx\n",
 795                         vq->last_used_idx, vq->used->idx);
 796                 vq->last_used_idx  = vq->used->idx;
 797                 vq->last_avail_idx = vq->used->idx;
 798         }
 799
 800         vq->access_ok = 1;
 801
 802         VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address desc: %p\n",
 803                         dev->vid, vq->desc);
 804         VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address avail: %p\n",
 805                         dev->vid, vq->avail);
 806         VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address used: %p\n",
 807                         dev->vid, vq->used);
 808         VHOST_LOG_CONFIG(DEBUG, "(%d) log_guest_addr: %" PRIx64 "\n",
 809                         dev->vid, vq->log_guest_addr);
 810
 811         return dev;
 812 }
 813
 814 /*
 815  * The virtio device sends us the desc, used and avail ring addresses.
 816  * This function then converts these to our address space.
 817  */
 818 static int
 819 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 820                         int main_fd __rte_unused)
 821 {
 822         struct virtio_net *dev = *pdev;
 823         struct vhost_virtqueue *vq;
 824         struct vhost_vring_addr *addr = &msg->payload.addr;
 825         bool access_ok;
 826
 827         if (validate_msg_fds(msg, 0) != 0)
 828                 return RTE_VHOST_MSG_RESULT_ERR;
 829
 830         if (dev->mem == NULL)
 831                 return RTE_VHOST_MSG_RESULT_ERR;
 832
 833         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 834         vq = dev->virtqueue[msg->payload.addr.index];
 835
 836         access_ok = vq->access_ok;
 837
 838         /*
 839          * Rings addresses should not be interpreted as long as the ring is not
 840          * started and enabled
 841          */
 842         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 843
 844         vring_invalidate(dev, vq);
 845
 846         if ((vq->enabled && (dev->features &
 847                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) ||
 848                         access_ok) {
 849                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 850                 if (!dev)
 851                         return RTE_VHOST_MSG_RESULT_ERR;
 852
 853                 *pdev = dev;
 854         }
 855
 856         return RTE_VHOST_MSG_RESULT_OK;
 857 }
 858
 859 /*
 860  * The virtio device sends us the available ring last used index.
 861  */
 862 static int
 863 vhost_user_set_vring_base(struct virtio_net **pdev,
 864                         struct VhostUserMsg *msg,
 865                         int main_fd __rte_unused)
 866 {
 867         struct virtio_net *dev = *pdev;
 868         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 869         uint64_t val = msg->payload.state.num;
 870
 871         if (validate_msg_fds(msg, 0) != 0)
 872                 return RTE_VHOST_MSG_RESULT_ERR;
 873
 874         if (vq_is_packed(dev)) {
 875                 /*
 876                  * Bit[0:14]: avail index
 877                  * Bit[15]: avail wrap counter
 878                  */
 879                 vq->last_avail_idx = val & 0x7fff;
 880                 vq->avail_wrap_counter = !!(val & (0x1 << 15));
 881                 /*
 882                  * Set used index to same value as available one, as
 883                  * their values should be the same since ring processing
 884                  * was stopped at get time.
 885                  */
 886                 vq->last_used_idx = vq->last_avail_idx;
 887                 vq->used_wrap_counter = vq->avail_wrap_counter;
 888         } else {
 889                 vq->last_used_idx = msg->payload.state.num;
 890                 vq->last_avail_idx = msg->payload.state.num;
 891         }
 892
 893         return RTE_VHOST_MSG_RESULT_OK;
 894 }
 895
 896 static int
 897 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 898                    uint64_t host_phys_addr, uint64_t size)
 899 {
 900         struct guest_page *page, *last_page;
 901         struct guest_page *old_pages;
 902
 903         if (dev->nr_guest_pages == dev->max_guest_pages) {
 904                 dev->max_guest_pages *= 2;
 905                 old_pages = dev->guest_pages;
 906                 dev->guest_pages = rte_realloc(dev->guest_pages,
 907                                         dev->max_guest_pages * sizeof(*page),
 908                                         RTE_CACHE_LINE_SIZE);
 909                 if (dev->guest_pages == NULL) {
 910                         VHOST_LOG_CONFIG(ERR, "cannot realloc guest_pages\n");
 911                         rte_free(old_pages);
 912                         return -1;
 913                 }
 914         }
 915
 916         if (dev->nr_guest_pages > 0) {
 917                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 918                 /* merge if the two pages are continuous */
 919                 if (host_phys_addr == last_page->host_phys_addr +
 920                                       last_page->size) {
 921                         last_page->size += size;
 922                         return 0;
 923                 }
 924         }
 925
 926         page = &dev->guest_pages[dev->nr_guest_pages++];
 927         page->guest_phys_addr = guest_phys_addr;
 928         page->host_phys_addr  = host_phys_addr;
 929         page->size = size;
 930
 931         return 0;
 932 }
 933
 934 static int
 935 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 936                 uint64_t page_size)
 937 {
 938         uint64_t reg_size = reg->size;
 939         uint64_t host_user_addr  = reg->host_user_addr;
 940         uint64_t guest_phys_addr = reg->guest_phys_addr;
 941         uint64_t host_phys_addr;
 942         uint64_t size;
 943
 944         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 945         size = page_size - (guest_phys_addr & (page_size - 1));
 946         size = RTE_MIN(size, reg_size);
 947
 948         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 949                 return -1;
 950
 951         host_user_addr  += size;
 952         guest_phys_addr += size;
 953         reg_size -= size;
 954
 955         while (reg_size > 0) {
 956                 size = RTE_MIN(reg_size, page_size);
 957                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 958                                                   host_user_addr);
 959                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 960                                 size) < 0)
 961                         return -1;
 962
 963                 host_user_addr  += size;
 964                 guest_phys_addr += size;
 965                 reg_size -= size;
 966         }
 967
 968         /* sort guest page array if over binary search threshold */
 969         if (dev->nr_guest_pages >= VHOST_BINARY_SEARCH_THRESH) {
 970                 qsort((void *)dev->guest_pages, dev->nr_guest_pages,
 971                         sizeof(struct guest_page), guest_page_addrcmp);
 972         }
 973
 974         return 0;
 975 }
 976
 977 #ifdef RTE_LIBRTE_VHOST_DEBUG
 978 /* TODO: enable it only in debug mode? */
 979 static void
 980 dump_guest_pages(struct virtio_net *dev)
 981 {
 982         uint32_t i;
 983         struct guest_page *page;
 984
 985         for (i = 0; i < dev->nr_guest_pages; i++) {
 986                 page = &dev->guest_pages[i];
 987
 988                 VHOST_LOG_CONFIG(INFO,
 989                         "guest physical page region %u\n"
 990                         "\t guest_phys_addr: %" PRIx64 "\n"
 991                         "\t host_phys_addr : %" PRIx64 "\n"
 992                         "\t size           : %" PRIx64 "\n",
 993                         i,
 994                         page->guest_phys_addr,
 995                         page->host_phys_addr,
 996                         page->size);
 997         }
 998 }
 999 #else
1000 #define dump_guest_pages(dev)
1001 #endif
1002
1003 static bool
1004 vhost_memory_changed(struct VhostUserMemory *new,
1005                      struct rte_vhost_memory *old)
1006 {
1007         uint32_t i;
1008
1009         if (new->nregions != old->nregions)
1010                 return true;
1011
1012         for (i = 0; i < new->nregions; ++i) {
1013                 VhostUserMemoryRegion *new_r = &new->regions[i];
1014                 struct rte_vhost_mem_region *old_r = &old->regions[i];
1015
1016                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
1017                         return true;
1018                 if (new_r->memory_size != old_r->size)
1019                         return true;
1020                 if (new_r->userspace_addr != old_r->guest_user_addr)
1021                         return true;
1022         }
1023
1024         return false;
1025 }
1026
1027 static int
1028 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
1029                         int main_fd)
1030 {
1031         struct virtio_net *dev = *pdev;
1032         struct VhostUserMemory *memory = &msg->payload.memory;
1033         struct rte_vhost_mem_region *reg;
1034         void *mmap_addr;
1035         uint64_t mmap_size;
1036         uint64_t mmap_offset;
1037         uint64_t alignment;
1038         uint32_t i;
1039         int populate;
1040         int fd;
1041
1042         if (validate_msg_fds(msg, memory->nregions) != 0)
1043                 return RTE_VHOST_MSG_RESULT_ERR;
1044
1045         if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
1046                 VHOST_LOG_CONFIG(ERR,
1047                         "too many memory regions (%u)\n", memory->nregions);
1048                 return RTE_VHOST_MSG_RESULT_ERR;
1049         }
1050
1051         if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
1052                 VHOST_LOG_CONFIG(INFO,
1053                         "(%d) memory regions not changed\n", dev->vid);
1054
1055                 close_msg_fds(msg);
1056
1057                 return RTE_VHOST_MSG_RESULT_OK;
1058         }
1059
1060         if (dev->mem) {
1061                 free_mem_region(dev);
1062                 rte_free(dev->mem);
1063                 dev->mem = NULL;
1064         }
1065
1066         /* Flush IOTLB cache as previous HVAs are now invalid */
1067         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1068                 for (i = 0; i < dev->nr_vring; i++)
1069                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
1070
1071         dev->nr_guest_pages = 0;
1072         if (dev->guest_pages == NULL) {
1073                 dev->max_guest_pages = 8;
1074                 dev->guest_pages = rte_zmalloc(NULL,
1075                                         dev->max_guest_pages *
1076                                         sizeof(struct guest_page),
1077                                         RTE_CACHE_LINE_SIZE);
1078                 if (dev->guest_pages == NULL) {
1079                         VHOST_LOG_CONFIG(ERR,
1080                                 "(%d) failed to allocate memory "
1081                                 "for dev->guest_pages\n",
1082                                 dev->vid);
1083                         return RTE_VHOST_MSG_RESULT_ERR;
1084                 }
1085         }
1086
1087         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
1088                 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
1089         if (dev->mem == NULL) {
1090                 VHOST_LOG_CONFIG(ERR,
1091                         "(%d) failed to allocate memory for dev->mem\n",
1092                         dev->vid);
1093                 return RTE_VHOST_MSG_RESULT_ERR;
1094         }
1095         dev->mem->nregions = memory->nregions;
1096
1097         for (i = 0; i < memory->nregions; i++) {
1098                 fd  = msg->fds[i];
1099                 reg = &dev->mem->regions[i];
1100
1101                 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
1102                 reg->guest_user_addr = memory->regions[i].userspace_addr;
1103                 reg->size            = memory->regions[i].memory_size;
1104                 reg->fd              = fd;
1105
1106                 mmap_offset = memory->regions[i].mmap_offset;
1107
1108                 /* Check for memory_size + mmap_offset overflow */
1109                 if (mmap_offset >= -reg->size) {
1110                         VHOST_LOG_CONFIG(ERR,
1111                                 "mmap_offset (%#"PRIx64") and memory_size "
1112                                 "(%#"PRIx64") overflow\n",
1113                                 mmap_offset, reg->size);
1114                         goto err_mmap;
1115                 }
1116
1117                 mmap_size = reg->size + mmap_offset;
1118
1119                 /* mmap() without flag of MAP_ANONYMOUS, should be called
1120                  * with length argument aligned with hugepagesz at older
1121                  * longterm version Linux, like 2.6.32 and 3.2.72, or
1122                  * mmap() will fail with EINVAL.
1123                  *
1124                  * to avoid failure, make sure in caller to keep length
1125                  * aligned.
1126                  */
1127                 alignment = get_blk_size(fd);
1128                 if (alignment == (uint64_t)-1) {
1129                         VHOST_LOG_CONFIG(ERR,
1130                                 "couldn't get hugepage size through fstat\n");
1131                         goto err_mmap;
1132                 }
1133                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
1134                 if (mmap_size == 0) {
1135                         /*
1136                          * It could happen if initial mmap_size + alignment
1137                          * overflows the sizeof uint64, which could happen if
1138                          * either mmap_size or alignment value is wrong.
1139                          *
1140                          * mmap() kernel implementation would return an error,
1141                          * but better catch it before and provide useful info
1142                          * in the logs.
1143                          */
1144                         VHOST_LOG_CONFIG(ERR, "mmap size (0x%" PRIx64 ") "
1145                                         "or alignment (0x%" PRIx64 ") is invalid\n",
1146                                         reg->size + mmap_offset, alignment);
1147                         goto err_mmap;
1148                 }
1149
1150                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
1151                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
1152                                  MAP_SHARED | populate, fd, 0);
1153
1154                 if (mmap_addr == MAP_FAILED) {
1155                         VHOST_LOG_CONFIG(ERR,
1156                                 "mmap region %u failed.\n", i);
1157                         goto err_mmap;
1158                 }
1159
1160                 reg->mmap_addr = mmap_addr;
1161                 reg->mmap_size = mmap_size;
1162                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
1163                                       mmap_offset;
1164
1165                 if (dev->dequeue_zero_copy)
1166                         if (add_guest_pages(dev, reg, alignment) < 0) {
1167                                 VHOST_LOG_CONFIG(ERR,
1168                                         "adding guest pages to region %u failed.\n",
1169                                         i);
1170                                 goto err_mmap;
1171                         }
1172
1173                 VHOST_LOG_CONFIG(INFO,
1174                         "guest memory region %u, size: 0x%" PRIx64 "\n"
1175                         "\t guest physical addr: 0x%" PRIx64 "\n"
1176                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
1177                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
1178                         "\t mmap addr : 0x%" PRIx64 "\n"
1179                         "\t mmap size : 0x%" PRIx64 "\n"
1180                         "\t mmap align: 0x%" PRIx64 "\n"
1181                         "\t mmap off  : 0x%" PRIx64 "\n",
1182                         i, reg->size,
1183                         reg->guest_phys_addr,
1184                         reg->guest_user_addr,
1185                         reg->host_user_addr,
1186                         (uint64_t)(uintptr_t)mmap_addr,
1187                         mmap_size,
1188                         alignment,
1189                         mmap_offset);
1190
1191                 if (dev->postcopy_listening) {
1192                         /*
1193                          * We haven't a better way right now than sharing
1194                          * DPDK's virtual address with Qemu, so that Qemu can
1195                          * retrieve the region offset when handling userfaults.
1196                          */
1197                         memory->regions[i].userspace_addr =
1198                                 reg->host_user_addr;
1199                 }
1200         }
1201         if (dev->postcopy_listening) {
1202                 /* Send the addresses back to qemu */
1203                 msg->fd_num = 0;
1204                 send_vhost_reply(main_fd, msg);
1205
1206                 /* Wait for qemu to acknolwedge it's got the addresses
1207                  * we've got to wait before we're allowed to generate faults.
1208                  */
1209                 VhostUserMsg ack_msg;
1210                 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1211                         VHOST_LOG_CONFIG(ERR,
1212                                 "Failed to read qemu ack on postcopy set-mem-table\n");
1213                         goto err_mmap;
1214                 }
1215
1216                 if (validate_msg_fds(&ack_msg, 0) != 0)
1217                         goto err_mmap;
1218
1219                 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1220                         VHOST_LOG_CONFIG(ERR,
1221                                 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1222                                 ack_msg.request.master);
1223                         goto err_mmap;
1224                 }
1225
1226                 /* Now userfault register and we can use the memory */
1227                 for (i = 0; i < memory->nregions; i++) {
1228 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1229                         reg = &dev->mem->regions[i];
1230                         struct uffdio_register reg_struct;
1231
1232                         /*
1233                          * Let's register all the mmap'ed area to ensure
1234                          * alignment on page boundary.
1235                          */
1236                         reg_struct.range.start =
1237                                 (uint64_t)(uintptr_t)reg->mmap_addr;
1238                         reg_struct.range.len = reg->mmap_size;
1239                         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1240
1241                         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1242                                                 &reg_struct)) {
1243                                 VHOST_LOG_CONFIG(ERR,
1244                                         "Failed to register ufd for region %d: (ufd = %d) %s\n",
1245                                         i, dev->postcopy_ufd,
1246                                         strerror(errno));
1247                                 goto err_mmap;
1248                         }
1249                         VHOST_LOG_CONFIG(INFO,
1250                                 "\t userfaultfd registered for range : "
1251                                 "%" PRIx64 " - %" PRIx64 "\n",
1252                                 (uint64_t)reg_struct.range.start,
1253                                 (uint64_t)reg_struct.range.start +
1254                                 (uint64_t)reg_struct.range.len - 1);
1255 #else
1256                         goto err_mmap;
1257 #endif
1258                 }
1259         }
1260
1261         for (i = 0; i < dev->nr_vring; i++) {
1262                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1263
1264                 if (vq->desc || vq->avail || vq->used) {
1265                         /*
1266                          * If the memory table got updated, the ring addresses
1267                          * need to be translated again as virtual addresses have
1268                          * changed.
1269                          */
1270                         vring_invalidate(dev, vq);
1271
1272                         dev = translate_ring_addresses(dev, i);
1273                         if (!dev) {
1274                                 dev = *pdev;
1275                                 goto err_mmap;
1276                         }
1277
1278                         *pdev = dev;
1279                 }
1280         }
1281
1282         dump_guest_pages(dev);
1283
1284         return RTE_VHOST_MSG_RESULT_OK;
1285
1286 err_mmap:
1287         free_mem_region(dev);
1288         rte_free(dev->mem);
1289         dev->mem = NULL;
1290         return RTE_VHOST_MSG_RESULT_ERR;
1291 }
1292
1293 static bool
1294 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1295 {
1296         bool rings_ok;
1297
1298         if (!vq)
1299                 return false;
1300
1301         if (vq_is_packed(dev))
1302                 rings_ok = vq->desc_packed && vq->driver_event &&
1303                         vq->device_event;
1304         else
1305                 rings_ok = vq->desc && vq->avail && vq->used;
1306
1307         return rings_ok &&
1308                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1309                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1310 }
1311
1312 static int
1313 virtio_is_ready(struct virtio_net *dev)
1314 {
1315         struct vhost_virtqueue *vq;
1316         uint32_t i;
1317
1318         if (dev->nr_vring == 0)
1319                 return 0;
1320
1321         for (i = 0; i < dev->nr_vring; i++) {
1322                 vq = dev->virtqueue[i];
1323
1324                 if (!vq_is_ready(dev, vq))
1325                         return 0;
1326         }
1327
1328         VHOST_LOG_CONFIG(INFO,
1329                 "virtio is now ready for processing.\n");
1330         return 1;
1331 }
1332
1333 static void *
1334 inflight_mem_alloc(const char *name, size_t size, int *fd)
1335 {
1336         void *ptr;
1337         int mfd = -1;
1338         char fname[20] = "/tmp/memfd-XXXXXX";
1339
1340         *fd = -1;
1341 #ifdef MEMFD_SUPPORTED
1342         mfd = memfd_create(name, MFD_CLOEXEC);
1343 #else
1344         RTE_SET_USED(name);
1345 #endif
1346         if (mfd == -1) {
1347                 mfd = mkstemp(fname);
1348                 if (mfd == -1) {
1349                         VHOST_LOG_CONFIG(ERR,
1350                                 "failed to get inflight buffer fd\n");
1351                         return NULL;
1352                 }
1353
1354                 unlink(fname);
1355         }
1356
1357         if (ftruncate(mfd, size) == -1) {
1358                 VHOST_LOG_CONFIG(ERR,
1359                         "failed to alloc inflight buffer\n");
1360                 close(mfd);
1361                 return NULL;
1362         }
1363
1364         ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd, 0);
1365         if (ptr == MAP_FAILED) {
1366                 VHOST_LOG_CONFIG(ERR,
1367                         "failed to mmap inflight buffer\n");
1368                 close(mfd);
1369                 return NULL;
1370         }
1371
1372         *fd = mfd;
1373         return ptr;
1374 }
1375
1376 static uint32_t
1377 get_pervq_shm_size_split(uint16_t queue_size)
1378 {
1379         return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_split) *
1380                                   queue_size + sizeof(uint64_t) +
1381                                   sizeof(uint16_t) * 4, INFLIGHT_ALIGNMENT);
1382 }
1383
1384 static uint32_t
1385 get_pervq_shm_size_packed(uint16_t queue_size)
1386 {
1387         return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_packed)
1388                                   * queue_size + sizeof(uint64_t) +
1389                                   sizeof(uint16_t) * 6 + sizeof(uint8_t) * 9,
1390                                   INFLIGHT_ALIGNMENT);
1391 }
1392
1393 static int
1394 vhost_user_get_inflight_fd(struct virtio_net **pdev,
1395                            VhostUserMsg *msg,
1396                            int main_fd __rte_unused)
1397 {
1398         struct rte_vhost_inflight_info_packed *inflight_packed;
1399         uint64_t pervq_inflight_size, mmap_size;
1400         uint16_t num_queues, queue_size;
1401         struct virtio_net *dev = *pdev;
1402         int fd, i, j;
1403         void *addr;
1404
1405         if (msg->size != sizeof(msg->payload.inflight)) {
1406                 VHOST_LOG_CONFIG(ERR,
1407                         "invalid get_inflight_fd message size is %d\n",
1408                         msg->size);
1409                 return RTE_VHOST_MSG_RESULT_ERR;
1410         }
1411
1412         if (dev->inflight_info == NULL) {
1413                 dev->inflight_info = calloc(1,
1414                                             sizeof(struct inflight_mem_info));
1415                 if (!dev->inflight_info) {
1416                         VHOST_LOG_CONFIG(ERR,
1417                                 "failed to alloc dev inflight area\n");
1418                         return RTE_VHOST_MSG_RESULT_ERR;
1419                 }
1420         }
1421
1422         num_queues = msg->payload.inflight.num_queues;
1423         queue_size = msg->payload.inflight.queue_size;
1424
1425         VHOST_LOG_CONFIG(INFO, "get_inflight_fd num_queues: %u\n",
1426                 msg->payload.inflight.num_queues);
1427         VHOST_LOG_CONFIG(INFO, "get_inflight_fd queue_size: %u\n",
1428                 msg->payload.inflight.queue_size);
1429
1430         if (vq_is_packed(dev))
1431                 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1432         else
1433                 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1434
1435         mmap_size = num_queues * pervq_inflight_size;
1436         addr = inflight_mem_alloc("vhost-inflight", mmap_size, &fd);
1437         if (!addr) {
1438                 VHOST_LOG_CONFIG(ERR,
1439                         "failed to alloc vhost inflight area\n");
1440                         msg->payload.inflight.mmap_size = 0;
1441                 return RTE_VHOST_MSG_RESULT_ERR;
1442         }
1443         memset(addr, 0, mmap_size);
1444
1445         dev->inflight_info->addr = addr;
1446         dev->inflight_info->size = msg->payload.inflight.mmap_size = mmap_size;
1447         dev->inflight_info->fd = msg->fds[0] = fd;
1448         msg->payload.inflight.mmap_offset = 0;
1449         msg->fd_num = 1;
1450
1451         if (vq_is_packed(dev)) {
1452                 for (i = 0; i < num_queues; i++) {
1453                         inflight_packed =
1454                                 (struct rte_vhost_inflight_info_packed *)addr;
1455                         inflight_packed->used_wrap_counter = 1;
1456                         inflight_packed->old_used_wrap_counter = 1;
1457                         for (j = 0; j < queue_size; j++)
1458                                 inflight_packed->desc[j].next = j + 1;
1459                         addr = (void *)((char *)addr + pervq_inflight_size);
1460                 }
1461         }
1462
1463         VHOST_LOG_CONFIG(INFO,
1464                 "send inflight mmap_size: %"PRIu64"\n",
1465                 msg->payload.inflight.mmap_size);
1466         VHOST_LOG_CONFIG(INFO,
1467                 "send inflight mmap_offset: %"PRIu64"\n",
1468                 msg->payload.inflight.mmap_offset);
1469         VHOST_LOG_CONFIG(INFO,
1470                 "send inflight fd: %d\n", msg->fds[0]);
1471
1472         return RTE_VHOST_MSG_RESULT_REPLY;
1473 }
1474
1475 static int
1476 vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
1477                            int main_fd __rte_unused)
1478 {
1479         uint64_t mmap_size, mmap_offset;
1480         uint16_t num_queues, queue_size;
1481         struct virtio_net *dev = *pdev;
1482         uint32_t pervq_inflight_size;
1483         struct vhost_virtqueue *vq;
1484         void *addr;
1485         int fd, i;
1486
1487         fd = msg->fds[0];
1488         if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
1489                 VHOST_LOG_CONFIG(ERR,
1490                         "invalid set_inflight_fd message size is %d,fd is %d\n",
1491                         msg->size, fd);
1492                 return RTE_VHOST_MSG_RESULT_ERR;
1493         }
1494
1495         mmap_size = msg->payload.inflight.mmap_size;
1496         mmap_offset = msg->payload.inflight.mmap_offset;
1497         num_queues = msg->payload.inflight.num_queues;
1498         queue_size = msg->payload.inflight.queue_size;
1499
1500         if (vq_is_packed(dev))
1501                 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1502         else
1503                 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1504
1505         VHOST_LOG_CONFIG(INFO,
1506                 "set_inflight_fd mmap_size: %"PRIu64"\n", mmap_size);
1507         VHOST_LOG_CONFIG(INFO,
1508                 "set_inflight_fd mmap_offset: %"PRIu64"\n", mmap_offset);
1509         VHOST_LOG_CONFIG(INFO,
1510                 "set_inflight_fd num_queues: %u\n", num_queues);
1511         VHOST_LOG_CONFIG(INFO,
1512                 "set_inflight_fd queue_size: %u\n", queue_size);
1513         VHOST_LOG_CONFIG(INFO,
1514                 "set_inflight_fd fd: %d\n", fd);
1515         VHOST_LOG_CONFIG(INFO,
1516                 "set_inflight_fd pervq_inflight_size: %d\n",
1517                 pervq_inflight_size);
1518
1519         if (!dev->inflight_info) {
1520                 dev->inflight_info = calloc(1,
1521                                             sizeof(struct inflight_mem_info));
1522                 if (dev->inflight_info == NULL) {
1523                         VHOST_LOG_CONFIG(ERR,
1524                                 "failed to alloc dev inflight area\n");
1525                         return RTE_VHOST_MSG_RESULT_ERR;
1526                 }
1527         }
1528
1529         if (dev->inflight_info->addr)
1530                 munmap(dev->inflight_info->addr, dev->inflight_info->size);
1531
1532         addr = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
1533                     fd, mmap_offset);
1534         if (addr == MAP_FAILED) {
1535                 VHOST_LOG_CONFIG(ERR, "failed to mmap share memory.\n");
1536                 return RTE_VHOST_MSG_RESULT_ERR;
1537         }
1538
1539         if (dev->inflight_info->fd)
1540                 close(dev->inflight_info->fd);
1541
1542         dev->inflight_info->fd = fd;
1543         dev->inflight_info->addr = addr;
1544         dev->inflight_info->size = mmap_size;
1545
1546         for (i = 0; i < num_queues; i++) {
1547                 vq = dev->virtqueue[i];
1548                 if (vq_is_packed(dev)) {
1549                         vq->inflight_packed = addr;
1550                         vq->inflight_packed->desc_num = queue_size;
1551                 } else {
1552                         vq->inflight_split = addr;
1553                         vq->inflight_split->desc_num = queue_size;
1554                 }
1555                 addr = (void *)((char *)addr + pervq_inflight_size);
1556         }
1557
1558         return RTE_VHOST_MSG_RESULT_OK;
1559 }
1560
1561 static int
1562 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1563                         int main_fd __rte_unused)
1564 {
1565         struct virtio_net *dev = *pdev;
1566         struct vhost_vring_file file;
1567         struct vhost_virtqueue *vq;
1568         int expected_fds;
1569
1570         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1571         if (validate_msg_fds(msg, expected_fds) != 0)
1572                 return RTE_VHOST_MSG_RESULT_ERR;
1573
1574         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1575         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1576                 file.fd = VIRTIO_INVALID_EVENTFD;
1577         else
1578                 file.fd = msg->fds[0];
1579         VHOST_LOG_CONFIG(INFO,
1580                 "vring call idx:%d file:%d\n", file.index, file.fd);
1581
1582         vq = dev->virtqueue[file.index];
1583         if (vq->callfd >= 0)
1584                 close(vq->callfd);
1585
1586         vq->callfd = file.fd;
1587
1588         return RTE_VHOST_MSG_RESULT_OK;
1589 }
1590
1591 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1592                         struct VhostUserMsg *msg,
1593                         int main_fd __rte_unused)
1594 {
1595         int expected_fds;
1596
1597         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1598         if (validate_msg_fds(msg, expected_fds) != 0)
1599                 return RTE_VHOST_MSG_RESULT_ERR;
1600
1601         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1602                 close(msg->fds[0]);
1603         VHOST_LOG_CONFIG(INFO, "not implemented\n");
1604
1605         return RTE_VHOST_MSG_RESULT_OK;
1606 }
1607
1608 static int
1609 resubmit_desc_compare(const void *a, const void *b)
1610 {
1611         const struct rte_vhost_resubmit_desc *desc0 = a;
1612         const struct rte_vhost_resubmit_desc *desc1 = b;
1613
1614         if (desc1->counter > desc0->counter)
1615                 return 1;
1616
1617         return -1;
1618 }
1619
1620 static int
1621 vhost_check_queue_inflights_split(struct virtio_net *dev,
1622                                   struct vhost_virtqueue *vq)
1623 {
1624         uint16_t i;
1625         uint16_t resubmit_num = 0, last_io, num;
1626         struct vring_used *used = vq->used;
1627         struct rte_vhost_resubmit_info *resubmit;
1628         struct rte_vhost_inflight_info_split *inflight_split;
1629
1630         if (!(dev->protocol_features &
1631             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1632                 return RTE_VHOST_MSG_RESULT_OK;
1633
1634         /* The frontend may still not support the inflight feature
1635          * although we negotiate the protocol feature.
1636          */
1637         if ((!vq->inflight_split))
1638                 return RTE_VHOST_MSG_RESULT_OK;
1639
1640         if (!vq->inflight_split->version) {
1641                 vq->inflight_split->version = INFLIGHT_VERSION;
1642                 return RTE_VHOST_MSG_RESULT_OK;
1643         }
1644
1645         if (vq->resubmit_inflight)
1646                 return RTE_VHOST_MSG_RESULT_OK;
1647
1648         inflight_split = vq->inflight_split;
1649         vq->global_counter = 0;
1650         last_io = inflight_split->last_inflight_io;
1651
1652         if (inflight_split->used_idx != used->idx) {
1653                 inflight_split->desc[last_io].inflight = 0;
1654                 rte_smp_mb();
1655                 inflight_split->used_idx = used->idx;
1656         }
1657
1658         for (i = 0; i < inflight_split->desc_num; i++) {
1659                 if (inflight_split->desc[i].inflight == 1)
1660                         resubmit_num++;
1661         }
1662
1663         vq->last_avail_idx += resubmit_num;
1664
1665         if (resubmit_num) {
1666                 resubmit  = calloc(1, sizeof(struct rte_vhost_resubmit_info));
1667                 if (!resubmit) {
1668                         VHOST_LOG_CONFIG(ERR,
1669                                 "failed to allocate memory for resubmit info.\n");
1670                         return RTE_VHOST_MSG_RESULT_ERR;
1671                 }
1672
1673                 resubmit->resubmit_list = calloc(resubmit_num,
1674                         sizeof(struct rte_vhost_resubmit_desc));
1675                 if (!resubmit->resubmit_list) {
1676                         VHOST_LOG_CONFIG(ERR,
1677                                 "failed to allocate memory for inflight desc.\n");
1678                         free(resubmit);
1679                         return RTE_VHOST_MSG_RESULT_ERR;
1680                 }
1681
1682                 num = 0;
1683                 for (i = 0; i < vq->inflight_split->desc_num; i++) {
1684                         if (vq->inflight_split->desc[i].inflight == 1) {
1685                                 resubmit->resubmit_list[num].index = i;
1686                                 resubmit->resubmit_list[num].counter =
1687                                         inflight_split->desc[i].counter;
1688                                 num++;
1689                         }
1690                 }
1691                 resubmit->resubmit_num = num;
1692
1693                 if (resubmit->resubmit_num > 1)
1694                         qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1695                               sizeof(struct rte_vhost_resubmit_desc),
1696                               resubmit_desc_compare);
1697
1698                 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1699                 vq->resubmit_inflight = resubmit;
1700         }
1701
1702         return RTE_VHOST_MSG_RESULT_OK;
1703 }
1704
1705 static int
1706 vhost_check_queue_inflights_packed(struct virtio_net *dev,
1707                                    struct vhost_virtqueue *vq)
1708 {
1709         uint16_t i;
1710         uint16_t resubmit_num = 0, old_used_idx, num;
1711         struct rte_vhost_resubmit_info *resubmit;
1712         struct rte_vhost_inflight_info_packed *inflight_packed;
1713
1714         if (!(dev->protocol_features &
1715             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1716                 return RTE_VHOST_MSG_RESULT_OK;
1717
1718         /* The frontend may still not support the inflight feature
1719          * although we negotiate the protocol feature.
1720          */
1721         if ((!vq->inflight_packed))
1722                 return RTE_VHOST_MSG_RESULT_OK;
1723
1724         if (!vq->inflight_packed->version) {
1725                 vq->inflight_packed->version = INFLIGHT_VERSION;
1726                 return RTE_VHOST_MSG_RESULT_OK;
1727         }
1728
1729         if (vq->resubmit_inflight)
1730                 return RTE_VHOST_MSG_RESULT_OK;
1731
1732         inflight_packed = vq->inflight_packed;
1733         vq->global_counter = 0;
1734         old_used_idx = inflight_packed->old_used_idx;
1735
1736         if (inflight_packed->used_idx != old_used_idx) {
1737                 if (inflight_packed->desc[old_used_idx].inflight == 0) {
1738                         inflight_packed->old_used_idx =
1739                                 inflight_packed->used_idx;
1740                         inflight_packed->old_used_wrap_counter =
1741                                 inflight_packed->used_wrap_counter;
1742                         inflight_packed->old_free_head =
1743                                 inflight_packed->free_head;
1744                 } else {
1745                         inflight_packed->used_idx =
1746                                 inflight_packed->old_used_idx;
1747                         inflight_packed->used_wrap_counter =
1748                                 inflight_packed->old_used_wrap_counter;
1749                         inflight_packed->free_head =
1750                                 inflight_packed->old_free_head;
1751                 }
1752         }
1753
1754         for (i = 0; i < inflight_packed->desc_num; i++) {
1755                 if (inflight_packed->desc[i].inflight == 1)
1756                         resubmit_num++;
1757         }
1758
1759         if (resubmit_num) {
1760                 resubmit = calloc(1, sizeof(struct rte_vhost_resubmit_info));
1761                 if (resubmit == NULL) {
1762                         VHOST_LOG_CONFIG(ERR,
1763                                 "failed to allocate memory for resubmit info.\n");
1764                         return RTE_VHOST_MSG_RESULT_ERR;
1765                 }
1766
1767                 resubmit->resubmit_list = calloc(resubmit_num,
1768                         sizeof(struct rte_vhost_resubmit_desc));
1769                 if (resubmit->resubmit_list == NULL) {
1770                         VHOST_LOG_CONFIG(ERR,
1771                                 "failed to allocate memory for resubmit desc.\n");
1772                         free(resubmit);
1773                         return RTE_VHOST_MSG_RESULT_ERR;
1774                 }
1775
1776                 num = 0;
1777                 for (i = 0; i < inflight_packed->desc_num; i++) {
1778                         if (vq->inflight_packed->desc[i].inflight == 1) {
1779                                 resubmit->resubmit_list[num].index = i;
1780                                 resubmit->resubmit_list[num].counter =
1781                                         inflight_packed->desc[i].counter;
1782                                 num++;
1783                         }
1784                 }
1785                 resubmit->resubmit_num = num;
1786
1787                 if (resubmit->resubmit_num > 1)
1788                         qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1789                               sizeof(struct rte_vhost_resubmit_desc),
1790                               resubmit_desc_compare);
1791
1792                 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1793                 vq->resubmit_inflight = resubmit;
1794         }
1795
1796         return RTE_VHOST_MSG_RESULT_OK;
1797 }
1798
1799 static int
1800 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1801                         int main_fd __rte_unused)
1802 {
1803         struct virtio_net *dev = *pdev;
1804         struct vhost_vring_file file;
1805         struct vhost_virtqueue *vq;
1806         int expected_fds;
1807
1808         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1809         if (validate_msg_fds(msg, expected_fds) != 0)
1810                 return RTE_VHOST_MSG_RESULT_ERR;
1811
1812         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1813         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1814                 file.fd = VIRTIO_INVALID_EVENTFD;
1815         else
1816                 file.fd = msg->fds[0];
1817         VHOST_LOG_CONFIG(INFO,
1818                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1819
1820         /* Interpret ring addresses only when ring is started. */
1821         dev = translate_ring_addresses(dev, file.index);
1822         if (!dev)
1823                 return RTE_VHOST_MSG_RESULT_ERR;
1824
1825         *pdev = dev;
1826
1827         vq = dev->virtqueue[file.index];
1828
1829         /*
1830          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1831          * the ring starts already enabled. Otherwise, it is enabled via
1832          * the SET_VRING_ENABLE message.
1833          */
1834         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
1835                 vq->enabled = 1;
1836                 if (dev->notify_ops->vring_state_changed)
1837                         dev->notify_ops->vring_state_changed(
1838                                 dev->vid, file.index, 1);
1839         }
1840
1841         if (vq->kickfd >= 0)
1842                 close(vq->kickfd);
1843         vq->kickfd = file.fd;
1844
1845         if (vq_is_packed(dev)) {
1846                 if (vhost_check_queue_inflights_packed(dev, vq)) {
1847                         VHOST_LOG_CONFIG(ERR,
1848                                 "failed to inflights for vq: %d\n", file.index);
1849                         return RTE_VHOST_MSG_RESULT_ERR;
1850                 }
1851         } else {
1852                 if (vhost_check_queue_inflights_split(dev, vq)) {
1853                         VHOST_LOG_CONFIG(ERR,
1854                                 "failed to inflights for vq: %d\n", file.index);
1855                         return RTE_VHOST_MSG_RESULT_ERR;
1856                 }
1857         }
1858
1859         return RTE_VHOST_MSG_RESULT_OK;
1860 }
1861
1862 static void
1863 free_zmbufs(struct vhost_virtqueue *vq)
1864 {
1865         drain_zmbuf_list(vq);
1866
1867         rte_free(vq->zmbufs);
1868 }
1869
1870 /*
1871  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1872  */
1873 static int
1874 vhost_user_get_vring_base(struct virtio_net **pdev,
1875                         struct VhostUserMsg *msg,
1876                         int main_fd __rte_unused)
1877 {
1878         struct virtio_net *dev = *pdev;
1879         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1880         uint64_t val;
1881
1882         if (validate_msg_fds(msg, 0) != 0)
1883                 return RTE_VHOST_MSG_RESULT_ERR;
1884
1885         /* We have to stop the queue (virtio) if it is running. */
1886         vhost_destroy_device_notify(dev);
1887
1888         dev->flags &= ~VIRTIO_DEV_READY;
1889         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1890
1891         /* Here we are safe to get the indexes */
1892         if (vq_is_packed(dev)) {
1893                 /*
1894                  * Bit[0:14]: avail index
1895                  * Bit[15]: avail wrap counter
1896                  */
1897                 val = vq->last_avail_idx & 0x7fff;
1898                 val |= vq->avail_wrap_counter << 15;
1899                 msg->payload.state.num = val;
1900         } else {
1901                 msg->payload.state.num = vq->last_avail_idx;
1902         }
1903
1904         VHOST_LOG_CONFIG(INFO,
1905                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1906                 msg->payload.state.num);
1907         /*
1908          * Based on current qemu vhost-user implementation, this message is
1909          * sent and only sent in vhost_vring_stop.
1910          * TODO: cleanup the vring, it isn't usable since here.
1911          */
1912         if (vq->kickfd >= 0)
1913                 close(vq->kickfd);
1914
1915         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1916
1917         if (vq->callfd >= 0)
1918                 close(vq->callfd);
1919
1920         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1921
1922         vq->signalled_used_valid = false;
1923
1924         if (dev->dequeue_zero_copy)
1925                 free_zmbufs(vq);
1926         if (vq_is_packed(dev)) {
1927                 rte_free(vq->shadow_used_packed);
1928                 vq->shadow_used_packed = NULL;
1929         } else {
1930                 rte_free(vq->shadow_used_split);
1931                 vq->shadow_used_split = NULL;
1932         }
1933
1934         rte_free(vq->batch_copy_elems);
1935         vq->batch_copy_elems = NULL;
1936
1937         msg->size = sizeof(msg->payload.state);
1938         msg->fd_num = 0;
1939
1940         vring_invalidate(dev, vq);
1941
1942         return RTE_VHOST_MSG_RESULT_REPLY;
1943 }
1944
1945 /*
1946  * when virtio queues are ready to work, qemu will send us to
1947  * enable the virtio queue pair.
1948  */
1949 static int
1950 vhost_user_set_vring_enable(struct virtio_net **pdev,
1951                         struct VhostUserMsg *msg,
1952                         int main_fd __rte_unused)
1953 {
1954         struct virtio_net *dev = *pdev;
1955         int enable = (int)msg->payload.state.num;
1956         int index = (int)msg->payload.state.index;
1957         struct rte_vdpa_device *vdpa_dev;
1958         int did = -1;
1959
1960         if (validate_msg_fds(msg, 0) != 0)
1961                 return RTE_VHOST_MSG_RESULT_ERR;
1962
1963         VHOST_LOG_CONFIG(INFO,
1964                 "set queue enable: %d to qp idx: %d\n",
1965                 enable, index);
1966
1967         did = dev->vdpa_dev_id;
1968         vdpa_dev = rte_vdpa_get_device(did);
1969         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1970                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1971
1972         if (dev->notify_ops->vring_state_changed)
1973                 dev->notify_ops->vring_state_changed(dev->vid,
1974                                 index, enable);
1975
1976         /* On disable, rings have to be stopped being processed. */
1977         if (!enable && dev->dequeue_zero_copy)
1978                 drain_zmbuf_list(dev->virtqueue[index]);
1979
1980         dev->virtqueue[index]->enabled = enable;
1981
1982         return RTE_VHOST_MSG_RESULT_OK;
1983 }
1984
1985 static int
1986 vhost_user_get_protocol_features(struct virtio_net **pdev,
1987                         struct VhostUserMsg *msg,
1988                         int main_fd __rte_unused)
1989 {
1990         struct virtio_net *dev = *pdev;
1991         uint64_t features, protocol_features;
1992
1993         if (validate_msg_fds(msg, 0) != 0)
1994                 return RTE_VHOST_MSG_RESULT_ERR;
1995
1996         rte_vhost_driver_get_features(dev->ifname, &features);
1997         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1998
1999         /*
2000          * REPLY_ACK protocol feature is only mandatory for now
2001          * for IOMMU feature. If IOMMU is explicitly disabled by the
2002          * application, disable also REPLY_ACK feature for older buggy
2003          * Qemu versions (from v2.7.0 to v2.9.0).
2004          */
2005         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
2006                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
2007
2008         msg->payload.u64 = protocol_features;
2009         msg->size = sizeof(msg->payload.u64);
2010         msg->fd_num = 0;
2011
2012         return RTE_VHOST_MSG_RESULT_REPLY;
2013 }
2014
2015 static int
2016 vhost_user_set_protocol_features(struct virtio_net **pdev,
2017                         struct VhostUserMsg *msg,
2018                         int main_fd __rte_unused)
2019 {
2020         struct virtio_net *dev = *pdev;
2021         uint64_t protocol_features = msg->payload.u64;
2022         uint64_t slave_protocol_features = 0;
2023
2024         if (validate_msg_fds(msg, 0) != 0)
2025                 return RTE_VHOST_MSG_RESULT_ERR;
2026
2027         rte_vhost_driver_get_protocol_features(dev->ifname,
2028                         &slave_protocol_features);
2029         if (protocol_features & ~slave_protocol_features) {
2030                 VHOST_LOG_CONFIG(ERR,
2031                         "(%d) received invalid protocol features.\n",
2032                         dev->vid);
2033                 return RTE_VHOST_MSG_RESULT_ERR;
2034         }
2035
2036         dev->protocol_features = protocol_features;
2037         VHOST_LOG_CONFIG(INFO,
2038                 "negotiated Vhost-user protocol features: 0x%" PRIx64 "\n",
2039                 dev->protocol_features);
2040
2041         return RTE_VHOST_MSG_RESULT_OK;
2042 }
2043
2044 static int
2045 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
2046                         int main_fd __rte_unused)
2047 {
2048         struct virtio_net *dev = *pdev;
2049         int fd = msg->fds[0];
2050         uint64_t size, off;
2051         void *addr;
2052
2053         if (validate_msg_fds(msg, 1) != 0)
2054                 return RTE_VHOST_MSG_RESULT_ERR;
2055
2056         if (fd < 0) {
2057                 VHOST_LOG_CONFIG(ERR, "invalid log fd: %d\n", fd);
2058                 return RTE_VHOST_MSG_RESULT_ERR;
2059         }
2060
2061         if (msg->size != sizeof(VhostUserLog)) {
2062                 VHOST_LOG_CONFIG(ERR,
2063                         "invalid log base msg size: %"PRId32" != %d\n",
2064                         msg->size, (int)sizeof(VhostUserLog));
2065                 return RTE_VHOST_MSG_RESULT_ERR;
2066         }
2067
2068         size = msg->payload.log.mmap_size;
2069         off  = msg->payload.log.mmap_offset;
2070
2071         /* Don't allow mmap_offset to point outside the mmap region */
2072         if (off > size) {
2073                 VHOST_LOG_CONFIG(ERR,
2074                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
2075                         off, size);
2076                 return RTE_VHOST_MSG_RESULT_ERR;
2077         }
2078
2079         VHOST_LOG_CONFIG(INFO,
2080                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
2081                 size, off);
2082
2083         /*
2084          * mmap from 0 to workaround a hugepage mmap bug: mmap will
2085          * fail when offset is not page size aligned.
2086          */
2087         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
2088         close(fd);
2089         if (addr == MAP_FAILED) {
2090                 VHOST_LOG_CONFIG(ERR, "mmap log base failed!\n");
2091                 return RTE_VHOST_MSG_RESULT_ERR;
2092         }
2093
2094         /*
2095          * Free previously mapped log memory on occasionally
2096          * multiple VHOST_USER_SET_LOG_BASE.
2097          */
2098         if (dev->log_addr) {
2099                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
2100         }
2101         dev->log_addr = (uint64_t)(uintptr_t)addr;
2102         dev->log_base = dev->log_addr + off;
2103         dev->log_size = size;
2104
2105         /*
2106          * The spec is not clear about it (yet), but QEMU doesn't expect
2107          * any payload in the reply.
2108          */
2109         msg->size = 0;
2110         msg->fd_num = 0;
2111
2112         return RTE_VHOST_MSG_RESULT_REPLY;
2113 }
2114
2115 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
2116                         struct VhostUserMsg *msg,
2117                         int main_fd __rte_unused)
2118 {
2119         if (validate_msg_fds(msg, 1) != 0)
2120                 return RTE_VHOST_MSG_RESULT_ERR;
2121
2122         close(msg->fds[0]);
2123         VHOST_LOG_CONFIG(INFO, "not implemented.\n");
2124
2125         return RTE_VHOST_MSG_RESULT_OK;
2126 }
2127
2128 /*
2129  * An rarp packet is constructed and broadcasted to notify switches about
2130  * the new location of the migrated VM, so that packets from outside will
2131  * not be lost after migration.
2132  *
2133  * However, we don't actually "send" a rarp packet here, instead, we set
2134  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
2135  */
2136 static int
2137 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
2138                         int main_fd __rte_unused)
2139 {
2140         struct virtio_net *dev = *pdev;
2141         uint8_t *mac = (uint8_t *)&msg->payload.u64;
2142         struct rte_vdpa_device *vdpa_dev;
2143         int did = -1;
2144
2145         if (validate_msg_fds(msg, 0) != 0)
2146                 return RTE_VHOST_MSG_RESULT_ERR;
2147
2148         VHOST_LOG_CONFIG(DEBUG,
2149                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
2150                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
2151         memcpy(dev->mac.addr_bytes, mac, 6);
2152
2153         /*
2154          * Set the flag to inject a RARP broadcast packet at
2155          * rte_vhost_dequeue_burst().
2156          *
2157          * __ATOMIC_RELEASE ordering is for making sure the mac is
2158          * copied before the flag is set.
2159          */
2160         __atomic_store_n(&dev->broadcast_rarp, 1, __ATOMIC_RELEASE);
2161         did = dev->vdpa_dev_id;
2162         vdpa_dev = rte_vdpa_get_device(did);
2163         if (vdpa_dev && vdpa_dev->ops->migration_done)
2164                 vdpa_dev->ops->migration_done(dev->vid);
2165
2166         return RTE_VHOST_MSG_RESULT_OK;
2167 }
2168
2169 static int
2170 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
2171                         int main_fd __rte_unused)
2172 {
2173         struct virtio_net *dev = *pdev;
2174
2175         if (validate_msg_fds(msg, 0) != 0)
2176                 return RTE_VHOST_MSG_RESULT_ERR;
2177
2178         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
2179                         msg->payload.u64 > VIRTIO_MAX_MTU) {
2180                 VHOST_LOG_CONFIG(ERR, "Invalid MTU size (%"PRIu64")\n",
2181                                 msg->payload.u64);
2182
2183                 return RTE_VHOST_MSG_RESULT_ERR;
2184         }
2185
2186         dev->mtu = msg->payload.u64;
2187
2188         return RTE_VHOST_MSG_RESULT_OK;
2189 }
2190
2191 static int
2192 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
2193                         int main_fd __rte_unused)
2194 {
2195         struct virtio_net *dev = *pdev;
2196         int fd = msg->fds[0];
2197
2198         if (validate_msg_fds(msg, 1) != 0)
2199                 return RTE_VHOST_MSG_RESULT_ERR;
2200
2201         if (fd < 0) {
2202                 VHOST_LOG_CONFIG(ERR,
2203                                 "Invalid file descriptor for slave channel (%d)\n",
2204                                 fd);
2205                 return RTE_VHOST_MSG_RESULT_ERR;
2206         }
2207
2208         if (dev->slave_req_fd >= 0)
2209                 close(dev->slave_req_fd);
2210
2211         dev->slave_req_fd = fd;
2212
2213         return RTE_VHOST_MSG_RESULT_OK;
2214 }
2215
2216 static int
2217 is_vring_iotlb_split(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2218 {
2219         struct vhost_vring_addr *ra;
2220         uint64_t start, end, len;
2221
2222         start = imsg->iova;
2223         end = start + imsg->size;
2224
2225         ra = &vq->ring_addrs;
2226         len = sizeof(struct vring_desc) * vq->size;
2227         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2228                 return 1;
2229
2230         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
2231         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2232                 return 1;
2233
2234         len = sizeof(struct vring_used) +
2235                sizeof(struct vring_used_elem) * vq->size;
2236         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2237                 return 1;
2238
2239         if (ra->flags & (1 << VHOST_VRING_F_LOG)) {
2240                 len = sizeof(uint64_t);
2241                 if (ra->log_guest_addr < end &&
2242                     (ra->log_guest_addr + len) > start)
2243                         return 1;
2244         }
2245
2246         return 0;
2247 }
2248
2249 static int
2250 is_vring_iotlb_packed(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2251 {
2252         struct vhost_vring_addr *ra;
2253         uint64_t start, end, len;
2254
2255         start = imsg->iova;
2256         end = start + imsg->size;
2257
2258         ra = &vq->ring_addrs;
2259         len = sizeof(struct vring_packed_desc) * vq->size;
2260         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2261                 return 1;
2262
2263         len = sizeof(struct vring_packed_desc_event);
2264         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2265                 return 1;
2266
2267         len = sizeof(struct vring_packed_desc_event);
2268         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2269                 return 1;
2270
2271         if (ra->flags & (1 << VHOST_VRING_F_LOG)) {
2272                 len = sizeof(uint64_t);
2273                 if (ra->log_guest_addr < end &&
2274                     (ra->log_guest_addr + len) > start)
2275                         return 1;
2276         }
2277
2278         return 0;
2279 }
2280
2281 static int is_vring_iotlb(struct virtio_net *dev,
2282                           struct vhost_virtqueue *vq,
2283                           struct vhost_iotlb_msg *imsg)
2284 {
2285         if (vq_is_packed(dev))
2286                 return is_vring_iotlb_packed(vq, imsg);
2287         else
2288                 return is_vring_iotlb_split(vq, imsg);
2289 }
2290
2291 static int
2292 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
2293                         int main_fd __rte_unused)
2294 {
2295         struct virtio_net *dev = *pdev;
2296         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
2297         uint16_t i;
2298         uint64_t vva, len;
2299
2300         if (validate_msg_fds(msg, 0) != 0)
2301                 return RTE_VHOST_MSG_RESULT_ERR;
2302
2303         switch (imsg->type) {
2304         case VHOST_IOTLB_UPDATE:
2305                 len = imsg->size;
2306                 vva = qva_to_vva(dev, imsg->uaddr, &len);
2307                 if (!vva)
2308                         return RTE_VHOST_MSG_RESULT_ERR;
2309
2310                 for (i = 0; i < dev->nr_vring; i++) {
2311                         struct vhost_virtqueue *vq = dev->virtqueue[i];
2312
2313                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
2314                                         len, imsg->perm);
2315
2316                         if (is_vring_iotlb(dev, vq, imsg))
2317                                 *pdev = dev = translate_ring_addresses(dev, i);
2318                 }
2319                 break;
2320         case VHOST_IOTLB_INVALIDATE:
2321                 for (i = 0; i < dev->nr_vring; i++) {
2322                         struct vhost_virtqueue *vq = dev->virtqueue[i];
2323
2324                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
2325                                         imsg->size);
2326
2327                         if (is_vring_iotlb(dev, vq, imsg))
2328                                 vring_invalidate(dev, vq);
2329                 }
2330                 break;
2331         default:
2332                 VHOST_LOG_CONFIG(ERR, "Invalid IOTLB message type (%d)\n",
2333                                 imsg->type);
2334                 return RTE_VHOST_MSG_RESULT_ERR;
2335         }
2336
2337         return RTE_VHOST_MSG_RESULT_OK;
2338 }
2339
2340 static int
2341 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
2342                         struct VhostUserMsg *msg,
2343                         int main_fd __rte_unused)
2344 {
2345         struct virtio_net *dev = *pdev;
2346 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
2347         struct uffdio_api api_struct;
2348
2349         if (validate_msg_fds(msg, 0) != 0)
2350                 return RTE_VHOST_MSG_RESULT_ERR;
2351
2352         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
2353
2354         if (dev->postcopy_ufd == -1) {
2355                 VHOST_LOG_CONFIG(ERR, "Userfaultfd not available: %s\n",
2356                         strerror(errno));
2357                 return RTE_VHOST_MSG_RESULT_ERR;
2358         }
2359         api_struct.api = UFFD_API;
2360         api_struct.features = 0;
2361         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
2362                 VHOST_LOG_CONFIG(ERR, "UFFDIO_API ioctl failure: %s\n",
2363                         strerror(errno));
2364                 close(dev->postcopy_ufd);
2365                 dev->postcopy_ufd = -1;
2366                 return RTE_VHOST_MSG_RESULT_ERR;
2367         }
2368         msg->fds[0] = dev->postcopy_ufd;
2369         msg->fd_num = 1;
2370
2371         return RTE_VHOST_MSG_RESULT_REPLY;
2372 #else
2373         dev->postcopy_ufd = -1;
2374         msg->fd_num = 0;
2375
2376         return RTE_VHOST_MSG_RESULT_ERR;
2377 #endif
2378 }
2379
2380 static int
2381 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
2382                         struct VhostUserMsg *msg __rte_unused,
2383                         int main_fd __rte_unused)
2384 {
2385         struct virtio_net *dev = *pdev;
2386
2387         if (validate_msg_fds(msg, 0) != 0)
2388                 return RTE_VHOST_MSG_RESULT_ERR;
2389
2390         if (dev->mem && dev->mem->nregions) {
2391                 VHOST_LOG_CONFIG(ERR,
2392                         "Regions already registered at postcopy-listen\n");
2393                 return RTE_VHOST_MSG_RESULT_ERR;
2394         }
2395         dev->postcopy_listening = 1;
2396
2397         return RTE_VHOST_MSG_RESULT_OK;
2398 }
2399
2400 static int
2401 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
2402                         int main_fd __rte_unused)
2403 {
2404         struct virtio_net *dev = *pdev;
2405
2406         if (validate_msg_fds(msg, 0) != 0)
2407                 return RTE_VHOST_MSG_RESULT_ERR;
2408
2409         dev->postcopy_listening = 0;
2410         if (dev->postcopy_ufd >= 0) {
2411                 close(dev->postcopy_ufd);
2412                 dev->postcopy_ufd = -1;
2413         }
2414
2415         msg->payload.u64 = 0;
2416         msg->size = sizeof(msg->payload.u64);
2417         msg->fd_num = 0;
2418
2419         return RTE_VHOST_MSG_RESULT_REPLY;
2420 }
2421
2422 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
2423                                         struct VhostUserMsg *msg,
2424                                         int main_fd);
2425 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
2426         [VHOST_USER_NONE] = NULL,
2427         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
2428         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
2429         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
2430         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
2431         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
2432         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
2433         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
2434         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
2435         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
2436         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
2437         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
2438         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
2439         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
2440         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
2441         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
2442         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
2443         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
2444         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
2445         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
2446         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
2447         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
2448         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
2449         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
2450         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
2451         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
2452         [VHOST_USER_GET_INFLIGHT_FD] = vhost_user_get_inflight_fd,
2453         [VHOST_USER_SET_INFLIGHT_FD] = vhost_user_set_inflight_fd,
2454 };
2455
2456 /* return bytes# of read on success or negative val on failure. */
2457 static int
2458 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
2459 {
2460         int ret;
2461
2462         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
2463                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
2464         if (ret <= 0) {
2465                 return ret;
2466         } else if (ret != VHOST_USER_HDR_SIZE) {
2467                 VHOST_LOG_CONFIG(ERR, "Unexpected header size read\n");
2468                 close_msg_fds(msg);
2469                 return -1;
2470         }
2471
2472         if (msg->size) {
2473                 if (msg->size > sizeof(msg->payload)) {
2474                         VHOST_LOG_CONFIG(ERR,
2475                                 "invalid msg size: %d\n", msg->size);
2476                         return -1;
2477                 }
2478                 ret = read(sockfd, &msg->payload, msg->size);
2479                 if (ret <= 0)
2480                         return ret;
2481                 if (ret != (int)msg->size) {
2482                         VHOST_LOG_CONFIG(ERR,
2483                                 "read control message failed\n");
2484                         return -1;
2485                 }
2486         }
2487
2488         return ret;
2489 }
2490
2491 static int
2492 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
2493 {
2494         if (!msg)
2495                 return 0;
2496
2497         return send_fd_message(sockfd, (char *)msg,
2498                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
2499 }
2500
2501 static int
2502 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
2503 {
2504         if (!msg)
2505                 return 0;
2506
2507         msg->flags &= ~VHOST_USER_VERSION_MASK;
2508         msg->flags &= ~VHOST_USER_NEED_REPLY;
2509         msg->flags |= VHOST_USER_VERSION;
2510         msg->flags |= VHOST_USER_REPLY_MASK;
2511
2512         return send_vhost_message(sockfd, msg);
2513 }
2514
2515 static int
2516 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
2517 {
2518         int ret;
2519
2520         if (msg->flags & VHOST_USER_NEED_REPLY)
2521                 rte_spinlock_lock(&dev->slave_req_lock);
2522
2523         ret = send_vhost_message(dev->slave_req_fd, msg);
2524         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
2525                 rte_spinlock_unlock(&dev->slave_req_lock);
2526
2527         return ret;
2528 }
2529
2530 /*
2531  * Allocate a queue pair if it hasn't been allocated yet
2532  */
2533 static int
2534 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
2535                         struct VhostUserMsg *msg)
2536 {
2537         uint16_t vring_idx;
2538
2539         switch (msg->request.master) {
2540         case VHOST_USER_SET_VRING_KICK:
2541         case VHOST_USER_SET_VRING_CALL:
2542         case VHOST_USER_SET_VRING_ERR:
2543                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
2544                 break;
2545         case VHOST_USER_SET_VRING_NUM:
2546         case VHOST_USER_SET_VRING_BASE:
2547         case VHOST_USER_SET_VRING_ENABLE:
2548                 vring_idx = msg->payload.state.index;
2549                 break;
2550         case VHOST_USER_SET_VRING_ADDR:
2551                 vring_idx = msg->payload.addr.index;
2552                 break;
2553         default:
2554                 return 0;
2555         }
2556
2557         if (vring_idx >= VHOST_MAX_VRING) {
2558                 VHOST_LOG_CONFIG(ERR,
2559                         "invalid vring index: %u\n", vring_idx);
2560                 return -1;
2561         }
2562
2563         if (dev->virtqueue[vring_idx])
2564                 return 0;
2565
2566         return alloc_vring_queue(dev, vring_idx);
2567 }
2568
2569 static void
2570 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
2571 {
2572         unsigned int i = 0;
2573         unsigned int vq_num = 0;
2574
2575         while (vq_num < dev->nr_vring) {
2576                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2577
2578                 if (vq) {
2579                         rte_spinlock_lock(&vq->access_lock);
2580                         vq_num++;
2581                 }
2582                 i++;
2583         }
2584 }
2585
2586 static void
2587 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
2588 {
2589         unsigned int i = 0;
2590         unsigned int vq_num = 0;
2591
2592         while (vq_num < dev->nr_vring) {
2593                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2594
2595                 if (vq) {
2596                         rte_spinlock_unlock(&vq->access_lock);
2597                         vq_num++;
2598                 }
2599                 i++;
2600         }
2601 }
2602
2603 int
2604 vhost_user_msg_handler(int vid, int fd)
2605 {
2606         struct virtio_net *dev;
2607         struct VhostUserMsg msg;
2608         struct rte_vdpa_device *vdpa_dev;
2609         int did = -1;
2610         int ret;
2611         int unlock_required = 0;
2612         bool handled;
2613         int request;
2614
2615         dev = get_device(vid);
2616         if (dev == NULL)
2617                 return -1;
2618
2619         if (!dev->notify_ops) {
2620                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
2621                 if (!dev->notify_ops) {
2622                         VHOST_LOG_CONFIG(ERR,
2623                                 "failed to get callback ops for driver %s\n",
2624                                 dev->ifname);
2625                         return -1;
2626                 }
2627         }
2628
2629         ret = read_vhost_message(fd, &msg);
2630         if (ret <= 0) {
2631                 if (ret < 0)
2632                         VHOST_LOG_CONFIG(ERR,
2633                                 "vhost read message failed\n");
2634                 else
2635                         VHOST_LOG_CONFIG(INFO,
2636                                 "vhost peer closed\n");
2637
2638                 return -1;
2639         }
2640
2641         ret = 0;
2642         request = msg.request.master;
2643         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
2644                         vhost_message_str[request]) {
2645                 if (request != VHOST_USER_IOTLB_MSG)
2646                         VHOST_LOG_CONFIG(INFO, "read message %s\n",
2647                                 vhost_message_str[request]);
2648                 else
2649                         VHOST_LOG_CONFIG(DEBUG, "read message %s\n",
2650                                 vhost_message_str[request]);
2651         } else {
2652                 VHOST_LOG_CONFIG(DEBUG, "External request %d\n", request);
2653         }
2654
2655         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
2656         if (ret < 0) {
2657                 VHOST_LOG_CONFIG(ERR,
2658                         "failed to alloc queue\n");
2659                 return -1;
2660         }
2661
2662         /*
2663          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
2664          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
2665          * and device is destroyed. destroy_device waits for queues to be
2666          * inactive, so it is safe. Otherwise taking the access_lock
2667          * would cause a dead lock.
2668          */
2669         switch (request) {
2670         case VHOST_USER_SET_FEATURES:
2671         case VHOST_USER_SET_PROTOCOL_FEATURES:
2672         case VHOST_USER_SET_OWNER:
2673         case VHOST_USER_SET_MEM_TABLE:
2674         case VHOST_USER_SET_LOG_BASE:
2675         case VHOST_USER_SET_LOG_FD:
2676         case VHOST_USER_SET_VRING_NUM:
2677         case VHOST_USER_SET_VRING_ADDR:
2678         case VHOST_USER_SET_VRING_BASE:
2679         case VHOST_USER_SET_VRING_KICK:
2680         case VHOST_USER_SET_VRING_CALL:
2681         case VHOST_USER_SET_VRING_ERR:
2682         case VHOST_USER_SET_VRING_ENABLE:
2683         case VHOST_USER_SEND_RARP:
2684         case VHOST_USER_NET_SET_MTU:
2685         case VHOST_USER_SET_SLAVE_REQ_FD:
2686                 vhost_user_lock_all_queue_pairs(dev);
2687                 unlock_required = 1;
2688                 break;
2689         default:
2690                 break;
2691
2692         }
2693
2694         handled = false;
2695         if (dev->extern_ops.pre_msg_handle) {
2696                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
2697                                 (void *)&msg);
2698                 switch (ret) {
2699                 case RTE_VHOST_MSG_RESULT_REPLY:
2700                         send_vhost_reply(fd, &msg);
2701                         /* Fall-through */
2702                 case RTE_VHOST_MSG_RESULT_ERR:
2703                 case RTE_VHOST_MSG_RESULT_OK:
2704                         handled = true;
2705                         goto skip_to_post_handle;
2706                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2707                 default:
2708                         break;
2709                 }
2710         }
2711
2712         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
2713                 if (!vhost_message_handlers[request])
2714                         goto skip_to_post_handle;
2715                 ret = vhost_message_handlers[request](&dev, &msg, fd);
2716
2717                 switch (ret) {
2718                 case RTE_VHOST_MSG_RESULT_ERR:
2719                         VHOST_LOG_CONFIG(ERR,
2720                                 "Processing %s failed.\n",
2721                                 vhost_message_str[request]);
2722                         handled = true;
2723                         break;
2724                 case RTE_VHOST_MSG_RESULT_OK:
2725                         VHOST_LOG_CONFIG(DEBUG,
2726                                 "Processing %s succeeded.\n",
2727                                 vhost_message_str[request]);
2728                         handled = true;
2729                         break;
2730                 case RTE_VHOST_MSG_RESULT_REPLY:
2731                         VHOST_LOG_CONFIG(DEBUG,
2732                                 "Processing %s succeeded and needs reply.\n",
2733                                 vhost_message_str[request]);
2734                         send_vhost_reply(fd, &msg);
2735                         handled = true;
2736                         break;
2737                 default:
2738                         break;
2739                 }
2740         }
2741
2742 skip_to_post_handle:
2743         if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2744                         dev->extern_ops.post_msg_handle) {
2745                 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
2746                                 (void *)&msg);
2747                 switch (ret) {
2748                 case RTE_VHOST_MSG_RESULT_REPLY:
2749                         send_vhost_reply(fd, &msg);
2750                         /* Fall-through */
2751                 case RTE_VHOST_MSG_RESULT_ERR:
2752                 case RTE_VHOST_MSG_RESULT_OK:
2753                         handled = true;
2754                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2755                 default:
2756                         break;
2757                 }
2758         }
2759
2760         if (unlock_required)
2761                 vhost_user_unlock_all_queue_pairs(dev);
2762
2763         /* If message was not handled at this stage, treat it as an error */
2764         if (!handled) {
2765                 VHOST_LOG_CONFIG(ERR,
2766                         "vhost message (req: %d) was not handled.\n", request);
2767                 close_msg_fds(&msg);
2768                 ret = RTE_VHOST_MSG_RESULT_ERR;
2769         }
2770
2771         /*
2772          * If the request required a reply that was already sent,
2773          * this optional reply-ack won't be sent as the
2774          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2775          */
2776         if (msg.flags & VHOST_USER_NEED_REPLY) {
2777                 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2778                 msg.size = sizeof(msg.payload.u64);
2779                 msg.fd_num = 0;
2780                 send_vhost_reply(fd, &msg);
2781         } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2782                 VHOST_LOG_CONFIG(ERR,
2783                         "vhost message handling failed.\n");
2784                 return -1;
2785         }
2786
2787         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2788                 dev->flags |= VIRTIO_DEV_READY;
2789
2790                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2791                         if (dev->dequeue_zero_copy) {
2792                                 VHOST_LOG_CONFIG(INFO,
2793                                                 "dequeue zero copy is enabled\n");
2794                         }
2795
2796                         if (dev->notify_ops->new_device(dev->vid) == 0)
2797                                 dev->flags |= VIRTIO_DEV_RUNNING;
2798                 }
2799         }
2800
2801         did = dev->vdpa_dev_id;
2802         vdpa_dev = rte_vdpa_get_device(did);
2803         if (vdpa_dev && virtio_is_ready(dev) &&
2804                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2805                         msg.request.master == VHOST_USER_SET_VRING_CALL) {
2806                 if (vdpa_dev->ops->dev_conf)
2807                         vdpa_dev->ops->dev_conf(dev->vid);
2808                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2809         }
2810
2811         return 0;
2812 }
2813
2814 static int process_slave_message_reply(struct virtio_net *dev,
2815                                        const struct VhostUserMsg *msg)
2816 {
2817         struct VhostUserMsg msg_reply;
2818         int ret;
2819
2820         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2821                 return 0;
2822
2823         ret = read_vhost_message(dev->slave_req_fd, &msg_reply);
2824         if (ret <= 0) {
2825                 if (ret < 0)
2826                         VHOST_LOG_CONFIG(ERR,
2827                                 "vhost read slave message reply failed\n");
2828                 else
2829                         VHOST_LOG_CONFIG(INFO,
2830                                 "vhost peer closed\n");
2831                 ret = -1;
2832                 goto out;
2833         }
2834
2835         ret = 0;
2836         if (msg_reply.request.slave != msg->request.slave) {
2837                 VHOST_LOG_CONFIG(ERR,
2838                         "Received unexpected msg type (%u), expected %u\n",
2839                         msg_reply.request.slave, msg->request.slave);
2840                 ret = -1;
2841                 goto out;
2842         }
2843
2844         ret = msg_reply.payload.u64 ? -1 : 0;
2845
2846 out:
2847         rte_spinlock_unlock(&dev->slave_req_lock);
2848         return ret;
2849 }
2850
2851 int
2852 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2853 {
2854         int ret;
2855         struct VhostUserMsg msg = {
2856                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2857                 .flags = VHOST_USER_VERSION,
2858                 .size = sizeof(msg.payload.iotlb),
2859                 .payload.iotlb = {
2860                         .iova = iova,
2861                         .perm = perm,
2862                         .type = VHOST_IOTLB_MISS,
2863                 },
2864         };
2865
2866         ret = send_vhost_message(dev->slave_req_fd, &msg);
2867         if (ret < 0) {
2868                 VHOST_LOG_CONFIG(ERR,
2869                                 "Failed to send IOTLB miss message (%d)\n",
2870                                 ret);
2871                 return ret;
2872         }
2873
2874         return 0;
2875 }
2876
2877 static int
2878 vhost_user_slave_config_change(struct virtio_net *dev, bool need_reply)
2879 {
2880         int ret;
2881         struct VhostUserMsg msg = {
2882                 .request.slave = VHOST_USER_SLAVE_CONFIG_CHANGE_MSG,
2883                 .flags = VHOST_USER_VERSION,
2884                 .size = 0,
2885         };
2886
2887         if (need_reply)
2888                 msg.flags |= VHOST_USER_NEED_REPLY;
2889
2890         ret = send_vhost_slave_message(dev, &msg);
2891         if (ret < 0) {
2892                 VHOST_LOG_CONFIG(ERR,
2893                                 "Failed to send config change (%d)\n",
2894                                 ret);
2895                 return ret;
2896         }
2897
2898         return process_slave_message_reply(dev, &msg);
2899 }
2900
2901 int
2902 rte_vhost_slave_config_change(int vid, bool need_reply)
2903 {
2904         struct virtio_net *dev;
2905
2906         dev = get_device(vid);
2907         if (!dev)
2908                 return -ENODEV;
2909
2910         return vhost_user_slave_config_change(dev, need_reply);
2911 }
2912
2913 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2914                                                     int index, int fd,
2915                                                     uint64_t offset,
2916                                                     uint64_t size)
2917 {
2918         int ret;
2919         struct VhostUserMsg msg = {
2920                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2921                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2922                 .size = sizeof(msg.payload.area),
2923                 .payload.area = {
2924                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
2925                         .size = size,
2926                         .offset = offset,
2927                 },
2928         };
2929
2930         if (fd < 0)
2931                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2932         else {
2933                 msg.fds[0] = fd;
2934                 msg.fd_num = 1;
2935         }
2936
2937         ret = send_vhost_slave_message(dev, &msg);
2938         if (ret < 0) {
2939                 VHOST_LOG_CONFIG(ERR,
2940                         "Failed to set host notifier (%d)\n", ret);
2941                 return ret;
2942         }
2943
2944         return process_slave_message_reply(dev, &msg);
2945 }
2946
2947 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2948 {
2949         struct virtio_net *dev;
2950         struct rte_vdpa_device *vdpa_dev;
2951         int vfio_device_fd, did, ret = 0;
2952         uint64_t offset, size;
2953         unsigned int i;
2954
2955         dev = get_device(vid);
2956         if (!dev)
2957                 return -ENODEV;
2958
2959         did = dev->vdpa_dev_id;
2960         if (did < 0)
2961                 return -EINVAL;
2962
2963         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2964             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2965             !(dev->protocol_features &
2966                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2967             !(dev->protocol_features &
2968                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2969             !(dev->protocol_features &
2970                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2971                 return -ENOTSUP;
2972
2973         vdpa_dev = rte_vdpa_get_device(did);
2974         if (!vdpa_dev)
2975                 return -ENODEV;
2976
2977         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2978         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2979
2980         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2981         if (vfio_device_fd < 0)
2982                 return -ENOTSUP;
2983
2984         if (enable) {
2985                 for (i = 0; i < dev->nr_vring; i++) {
2986                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2987                                         &size) < 0) {
2988                                 ret = -ENOTSUP;
2989                                 goto disable;
2990                         }
2991
2992                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2993                                         vfio_device_fd, offset, size) < 0) {
2994                                 ret = -EFAULT;
2995                                 goto disable;
2996                         }
2997                 }
2998         } else {
2999 disable:
3000                 for (i = 0; i < dev->nr_vring; i++) {
3001                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
3002                                         0, 0);
3003                 }
3004         }
3005
3006         return ret;
3007 }