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