lib/librte_vhost/vhost_user.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2018 Intel Corporation
   3  */
   4
   5 /* Security model
   6  * --------------
   7  * The vhost-user protocol connection is an external interface, so it must be
   8  * robust against invalid inputs.
   9  *
  10  * This is important because the vhost-user master is only one step removed
  11  * from the guest.  Malicious guests that have escaped will then launch further
  12  * attacks from the vhost-user master.
  13  *
  14  * Even in deployments where guests are trusted, a bug in the vhost-user master
  15  * can still cause invalid messages to be sent.  Such messages must not
  16  * compromise the stability of the DPDK application by causing crashes, memory
  17  * corruption, or other problematic behavior.
  18  *
  19  * Do not assume received VhostUserMsg fields contain sensible values!
  20  */
  21
  22 #include <stdint.h>
  23 #include <stdio.h>
  24 #include <stdlib.h>
  25 #include <string.h>
  26 #include <unistd.h>
  27 #include <fcntl.h>
  28 #include <sys/ioctl.h>
  29 #include <sys/mman.h>
  30 #include <sys/types.h>
  31 #include <sys/stat.h>
  32 #include <sys/syscall.h>
  33 #include <assert.h>
  34 #ifdef RTE_LIBRTE_VHOST_NUMA
  35 #include <numaif.h>
  36 #endif
  37 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
  38 #include <linux/userfaultfd.h>
  39 #endif
  40
  41 #include <rte_common.h>
  42 #include <rte_malloc.h>
  43 #include <rte_log.h>
  44
  45 #include "iotlb.h"
  46 #include "vhost.h"
  47 #include "vhost_user.h"
  48
  49 #define VIRTIO_MIN_MTU 68
  50 #define VIRTIO_MAX_MTU 65535
  51
  52 static const char *vhost_message_str[VHOST_USER_MAX] = {
  53         [VHOST_USER_NONE] = "VHOST_USER_NONE",
  54         [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
  55         [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
  56         [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
  57         [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
  58         [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
  59         [VHOST_USER_SET_LOG_BASE] = "VHOST_USER_SET_LOG_BASE",
  60         [VHOST_USER_SET_LOG_FD] = "VHOST_USER_SET_LOG_FD",
  61         [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
  62         [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
  63         [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
  64         [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
  65         [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
  66         [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
  67         [VHOST_USER_SET_VRING_ERR]  = "VHOST_USER_SET_VRING_ERR",
  68         [VHOST_USER_GET_PROTOCOL_FEATURES]  = "VHOST_USER_GET_PROTOCOL_FEATURES",
  69         [VHOST_USER_SET_PROTOCOL_FEATURES]  = "VHOST_USER_SET_PROTOCOL_FEATURES",
  70         [VHOST_USER_GET_QUEUE_NUM]  = "VHOST_USER_GET_QUEUE_NUM",
  71         [VHOST_USER_SET_VRING_ENABLE]  = "VHOST_USER_SET_VRING_ENABLE",
  72         [VHOST_USER_SEND_RARP]  = "VHOST_USER_SEND_RARP",
  73         [VHOST_USER_NET_SET_MTU]  = "VHOST_USER_NET_SET_MTU",
  74         [VHOST_USER_SET_SLAVE_REQ_FD]  = "VHOST_USER_SET_SLAVE_REQ_FD",
  75         [VHOST_USER_IOTLB_MSG]  = "VHOST_USER_IOTLB_MSG",
  76         [VHOST_USER_CRYPTO_CREATE_SESS] = "VHOST_USER_CRYPTO_CREATE_SESS",
  77         [VHOST_USER_CRYPTO_CLOSE_SESS] = "VHOST_USER_CRYPTO_CLOSE_SESS",
  78         [VHOST_USER_POSTCOPY_ADVISE]  = "VHOST_USER_POSTCOPY_ADVISE",
  79 };
  80
  81 static uint64_t
  82 get_blk_size(int fd)
  83 {
  84         struct stat stat;
  85         int ret;
  86
  87         ret = fstat(fd, &stat);
  88         return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
  89 }
  90
  91 static void
  92 free_mem_region(struct virtio_net *dev)
  93 {
  94         uint32_t i;
  95         struct rte_vhost_mem_region *reg;
  96
  97         if (!dev || !dev->mem)
  98                 return;
  99
 100         for (i = 0; i < dev->mem->nregions; i++) {
 101                 reg = &dev->mem->regions[i];
 102                 if (reg->host_user_addr) {
 103                         munmap(reg->mmap_addr, reg->mmap_size);
 104                         close(reg->fd);
 105                 }
 106         }
 107 }
 108
 109 void
 110 vhost_backend_cleanup(struct virtio_net *dev)
 111 {
 112         if (dev->mem) {
 113                 free_mem_region(dev);
 114                 rte_free(dev->mem);
 115                 dev->mem = NULL;
 116         }
 117
 118         free(dev->guest_pages);
 119         dev->guest_pages = NULL;
 120
 121         if (dev->log_addr) {
 122                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 123                 dev->log_addr = 0;
 124         }
 125
 126         if (dev->slave_req_fd >= 0) {
 127                 close(dev->slave_req_fd);
 128                 dev->slave_req_fd = -1;
 129         }
 130
 131         if (dev->postcopy_ufd >= 0) {
 132                 close(dev->postcopy_ufd);
 133                 dev->postcopy_ufd = -1;
 134         }
 135 }
 136
 137 /*
 138  * This function just returns success at the moment unless
 139  * the device hasn't been initialised.
 140  */
 141 static int
 142 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
 143                         struct VhostUserMsg *msg __rte_unused,
 144                         int main_fd __rte_unused)
 145 {
 146         return VH_RESULT_OK;
 147 }
 148
 149 static int
 150 vhost_user_reset_owner(struct virtio_net **pdev,
 151                         struct VhostUserMsg *msg __rte_unused,
 152                         int main_fd __rte_unused)
 153 {
 154         struct virtio_net *dev = *pdev;
 155         vhost_destroy_device_notify(dev);
 156
 157         cleanup_device(dev, 0);
 158         reset_device(dev);
 159         return VH_RESULT_OK;
 160 }
 161
 162 /*
 163  * The features that we support are requested.
 164  */
 165 static int
 166 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 167                         int main_fd __rte_unused)
 168 {
 169         struct virtio_net *dev = *pdev;
 170         uint64_t features = 0;
 171
 172         rte_vhost_driver_get_features(dev->ifname, &features);
 173
 174         msg->payload.u64 = features;
 175         msg->size = sizeof(msg->payload.u64);
 176         msg->fd_num = 0;
 177
 178         return VH_RESULT_REPLY;
 179 }
 180
 181 /*
 182  * The queue number that we support are requested.
 183  */
 184 static int
 185 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
 186                         int main_fd __rte_unused)
 187 {
 188         struct virtio_net *dev = *pdev;
 189         uint32_t queue_num = 0;
 190
 191         rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
 192
 193         msg->payload.u64 = (uint64_t)queue_num;
 194         msg->size = sizeof(msg->payload.u64);
 195         msg->fd_num = 0;
 196
 197         return VH_RESULT_REPLY;
 198 }
 199
 200 /*
 201  * We receive the negotiated features supported by us and the virtio device.
 202  */
 203 static int
 204 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 205                         int main_fd __rte_unused)
 206 {
 207         struct virtio_net *dev = *pdev;
 208         uint64_t features = msg->payload.u64;
 209         uint64_t vhost_features = 0;
 210         struct rte_vdpa_device *vdpa_dev;
 211         int did = -1;
 212
 213         rte_vhost_driver_get_features(dev->ifname, &vhost_features);
 214         if (features & ~vhost_features) {
 215                 RTE_LOG(ERR, VHOST_CONFIG,
 216                         "(%d) received invalid negotiated features.\n",
 217                         dev->vid);
 218                 return VH_RESULT_ERR;
 219         }
 220
 221         if (dev->flags & VIRTIO_DEV_RUNNING) {
 222                 if (dev->features == features)
 223                         return VH_RESULT_OK;
 224
 225                 /*
 226                  * Error out if master tries to change features while device is
 227                  * in running state. The exception being VHOST_F_LOG_ALL, which
 228                  * is enabled when the live-migration starts.
 229                  */
 230                 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
 231                         RTE_LOG(ERR, VHOST_CONFIG,
 232                                 "(%d) features changed while device is running.\n",
 233                                 dev->vid);
 234                         return VH_RESULT_ERR;
 235                 }
 236
 237                 if (dev->notify_ops->features_changed)
 238                         dev->notify_ops->features_changed(dev->vid, features);
 239         }
 240
 241         dev->features = features;
 242         if (dev->features &
 243                 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
 244                 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 245         } else {
 246                 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
 247         }
 248         VHOST_LOG_DEBUG(VHOST_CONFIG,
 249                 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
 250                 dev->vid,
 251                 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
 252                 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 253
 254         if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
 255             !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
 256                 /*
 257                  * Remove all but first queue pair if MQ hasn't been
 258                  * negotiated. This is safe because the device is not
 259                  * running at this stage.
 260                  */
 261                 while (dev->nr_vring > 2) {
 262                         struct vhost_virtqueue *vq;
 263
 264                         vq = dev->virtqueue[--dev->nr_vring];
 265                         if (!vq)
 266                                 continue;
 267
 268                         dev->virtqueue[dev->nr_vring] = NULL;
 269                         cleanup_vq(vq, 1);
 270                         free_vq(dev, vq);
 271                 }
 272         }
 273
 274         did = dev->vdpa_dev_id;
 275         vdpa_dev = rte_vdpa_get_device(did);
 276         if (vdpa_dev && vdpa_dev->ops->set_features)
 277                 vdpa_dev->ops->set_features(dev->vid);
 278
 279         return VH_RESULT_OK;
 280 }
 281
 282 /*
 283  * The virtio device sends us the size of the descriptor ring.
 284  */
 285 static int
 286 vhost_user_set_vring_num(struct virtio_net **pdev,
 287                         struct VhostUserMsg *msg,
 288                         int main_fd __rte_unused)
 289 {
 290         struct virtio_net *dev = *pdev;
 291         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 292
 293         vq->size = msg->payload.state.num;
 294
 295         /* VIRTIO 1.0, 2.4 Virtqueues says:
 296          *
 297          *   Queue Size value is always a power of 2. The maximum Queue Size
 298          *   value is 32768.
 299          */
 300         if ((vq->size & (vq->size - 1)) || vq->size > 32768) {
 301                 RTE_LOG(ERR, VHOST_CONFIG,
 302                         "invalid virtqueue size %u\n", vq->size);
 303                 return VH_RESULT_ERR;
 304         }
 305
 306         if (dev->dequeue_zero_copy) {
 307                 vq->nr_zmbuf = 0;
 308                 vq->last_zmbuf_idx = 0;
 309                 vq->zmbuf_size = vq->size;
 310                 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
 311                                          sizeof(struct zcopy_mbuf), 0);
 312                 if (vq->zmbufs == NULL) {
 313                         RTE_LOG(WARNING, VHOST_CONFIG,
 314                                 "failed to allocate mem for zero copy; "
 315                                 "zero copy is force disabled\n");
 316                         dev->dequeue_zero_copy = 0;
 317                 }
 318                 TAILQ_INIT(&vq->zmbuf_list);
 319         }
 320
 321         if (vq_is_packed(dev)) {
 322                 vq->shadow_used_packed = rte_malloc(NULL,
 323                                 vq->size *
 324                                 sizeof(struct vring_used_elem_packed),
 325                                 RTE_CACHE_LINE_SIZE);
 326                 if (!vq->shadow_used_packed) {
 327                         RTE_LOG(ERR, VHOST_CONFIG,
 328                                         "failed to allocate memory for shadow used ring.\n");
 329                         return VH_RESULT_ERR;
 330                 }
 331
 332         } else {
 333                 vq->shadow_used_split = rte_malloc(NULL,
 334                                 vq->size * sizeof(struct vring_used_elem),
 335                                 RTE_CACHE_LINE_SIZE);
 336                 if (!vq->shadow_used_split) {
 337                         RTE_LOG(ERR, VHOST_CONFIG,
 338                                         "failed to allocate memory for shadow used ring.\n");
 339                         return VH_RESULT_ERR;
 340                 }
 341         }
 342
 343         vq->batch_copy_elems = rte_malloc(NULL,
 344                                 vq->size * sizeof(struct batch_copy_elem),
 345                                 RTE_CACHE_LINE_SIZE);
 346         if (!vq->batch_copy_elems) {
 347                 RTE_LOG(ERR, VHOST_CONFIG,
 348                         "failed to allocate memory for batching copy.\n");
 349                 return VH_RESULT_ERR;
 350         }
 351
 352         return VH_RESULT_OK;
 353 }
 354
 355 /*
 356  * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
 357  * same numa node as the memory of vring descriptor.
 358  */
 359 #ifdef RTE_LIBRTE_VHOST_NUMA
 360 static struct virtio_net*
 361 numa_realloc(struct virtio_net *dev, int index)
 362 {
 363         int oldnode, newnode;
 364         struct virtio_net *old_dev;
 365         struct vhost_virtqueue *old_vq, *vq;
 366         struct zcopy_mbuf *new_zmbuf;
 367         struct vring_used_elem *new_shadow_used_split;
 368         struct vring_used_elem_packed *new_shadow_used_packed;
 369         struct batch_copy_elem *new_batch_copy_elems;
 370         int ret;
 371
 372         old_dev = dev;
 373         vq = old_vq = dev->virtqueue[index];
 374
 375         ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
 376                             MPOL_F_NODE | MPOL_F_ADDR);
 377
 378         /* check if we need to reallocate vq */
 379         ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
 380                              MPOL_F_NODE | MPOL_F_ADDR);
 381         if (ret) {
 382                 RTE_LOG(ERR, VHOST_CONFIG,
 383                         "Unable to get vq numa information.\n");
 384                 return dev;
 385         }
 386         if (oldnode != newnode) {
 387                 RTE_LOG(INFO, VHOST_CONFIG,
 388                         "reallocate vq from %d to %d node\n", oldnode, newnode);
 389                 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
 390                 if (!vq)
 391                         return dev;
 392
 393                 memcpy(vq, old_vq, sizeof(*vq));
 394                 TAILQ_INIT(&vq->zmbuf_list);
 395
 396                 if (dev->dequeue_zero_copy) {
 397                         new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
 398                                         sizeof(struct zcopy_mbuf), 0, newnode);
 399                         if (new_zmbuf) {
 400                                 rte_free(vq->zmbufs);
 401                                 vq->zmbufs = new_zmbuf;
 402                         }
 403                 }
 404
 405                 if (vq_is_packed(dev)) {
 406                         new_shadow_used_packed = rte_malloc_socket(NULL,
 407                                         vq->size *
 408                                         sizeof(struct vring_used_elem_packed),
 409                                         RTE_CACHE_LINE_SIZE,
 410                                         newnode);
 411                         if (new_shadow_used_packed) {
 412                                 rte_free(vq->shadow_used_packed);
 413                                 vq->shadow_used_packed = new_shadow_used_packed;
 414                         }
 415                 } else {
 416                         new_shadow_used_split = rte_malloc_socket(NULL,
 417                                         vq->size *
 418                                         sizeof(struct vring_used_elem),
 419                                         RTE_CACHE_LINE_SIZE,
 420                                         newnode);
 421                         if (new_shadow_used_split) {
 422                                 rte_free(vq->shadow_used_split);
 423                                 vq->shadow_used_split = new_shadow_used_split;
 424                         }
 425                 }
 426
 427                 new_batch_copy_elems = rte_malloc_socket(NULL,
 428                         vq->size * sizeof(struct batch_copy_elem),
 429                         RTE_CACHE_LINE_SIZE,
 430                         newnode);
 431                 if (new_batch_copy_elems) {
 432                         rte_free(vq->batch_copy_elems);
 433                         vq->batch_copy_elems = new_batch_copy_elems;
 434                 }
 435
 436                 rte_free(old_vq);
 437         }
 438
 439         /* check if we need to reallocate dev */
 440         ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
 441                             MPOL_F_NODE | MPOL_F_ADDR);
 442         if (ret) {
 443                 RTE_LOG(ERR, VHOST_CONFIG,
 444                         "Unable to get dev numa information.\n");
 445                 goto out;
 446         }
 447         if (oldnode != newnode) {
 448                 RTE_LOG(INFO, VHOST_CONFIG,
 449                         "reallocate dev from %d to %d node\n",
 450                         oldnode, newnode);
 451                 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
 452                 if (!dev) {
 453                         dev = old_dev;
 454                         goto out;
 455                 }
 456
 457                 memcpy(dev, old_dev, sizeof(*dev));
 458                 rte_free(old_dev);
 459         }
 460
 461 out:
 462         dev->virtqueue[index] = vq;
 463         vhost_devices[dev->vid] = dev;
 464
 465         if (old_vq != vq)
 466                 vhost_user_iotlb_init(dev, index);
 467
 468         return dev;
 469 }
 470 #else
 471 static struct virtio_net*
 472 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 473 {
 474         return dev;
 475 }
 476 #endif
 477
 478 /* Converts QEMU virtual address to Vhost virtual address. */
 479 static uint64_t
 480 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
 481 {
 482         struct rte_vhost_mem_region *r;
 483         uint32_t i;
 484
 485         /* Find the region where the address lives. */
 486         for (i = 0; i < dev->mem->nregions; i++) {
 487                 r = &dev->mem->regions[i];
 488
 489                 if (qva >= r->guest_user_addr &&
 490                     qva <  r->guest_user_addr + r->size) {
 491
 492                         if (unlikely(*len > r->guest_user_addr + r->size - qva))
 493                                 *len = r->guest_user_addr + r->size - qva;
 494
 495                         return qva - r->guest_user_addr +
 496                                r->host_user_addr;
 497                 }
 498         }
 499         *len = 0;
 500
 501         return 0;
 502 }
 503
 504
 505 /*
 506  * Converts ring address to Vhost virtual address.
 507  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 508  * else it is a QEMU virtual address.
 509  */
 510 static uint64_t
 511 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 512                 uint64_t ra, uint64_t *size)
 513 {
 514         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 515                 uint64_t vva;
 516
 517                 vva = vhost_user_iotlb_cache_find(vq, ra,
 518                                         size, VHOST_ACCESS_RW);
 519                 if (!vva)
 520                         vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
 521
 522                 return vva;
 523         }
 524
 525         return qva_to_vva(dev, ra, size);
 526 }
 527
 528 static struct virtio_net *
 529 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 530 {
 531         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 532         struct vhost_vring_addr *addr = &vq->ring_addrs;
 533         uint64_t len;
 534
 535         if (vq_is_packed(dev)) {
 536                 len = sizeof(struct vring_packed_desc) * vq->size;
 537                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 538                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 539                 vq->log_guest_addr = 0;
 540                 if (vq->desc_packed == NULL ||
 541                                 len != sizeof(struct vring_packed_desc) *
 542                                 vq->size) {
 543                         RTE_LOG(DEBUG, VHOST_CONFIG,
 544                                 "(%d) failed to map desc_packed ring.\n",
 545                                 dev->vid);
 546                         return dev;
 547                 }
 548
 549                 dev = numa_realloc(dev, vq_index);
 550                 vq = dev->virtqueue[vq_index];
 551                 addr = &vq->ring_addrs;
 552
 553                 len = sizeof(struct vring_packed_desc_event);
 554                 vq->driver_event = (struct vring_packed_desc_event *)
 555                                         (uintptr_t)ring_addr_to_vva(dev,
 556                                         vq, addr->avail_user_addr, &len);
 557                 if (vq->driver_event == NULL ||
 558                                 len != sizeof(struct vring_packed_desc_event)) {
 559                         RTE_LOG(DEBUG, VHOST_CONFIG,
 560                                 "(%d) failed to find driver area address.\n",
 561                                 dev->vid);
 562                         return dev;
 563                 }
 564
 565                 len = sizeof(struct vring_packed_desc_event);
 566                 vq->device_event = (struct vring_packed_desc_event *)
 567                                         (uintptr_t)ring_addr_to_vva(dev,
 568                                         vq, addr->used_user_addr, &len);
 569                 if (vq->device_event == NULL ||
 570                                 len != sizeof(struct vring_packed_desc_event)) {
 571                         RTE_LOG(DEBUG, VHOST_CONFIG,
 572                                 "(%d) failed to find device area address.\n",
 573                                 dev->vid);
 574                         return dev;
 575                 }
 576
 577                 return dev;
 578         }
 579
 580         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 581         if (vq->desc && vq->avail && vq->used)
 582                 return dev;
 583
 584         len = sizeof(struct vring_desc) * vq->size;
 585         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 586                         vq, addr->desc_user_addr, &len);
 587         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 588                 RTE_LOG(DEBUG, VHOST_CONFIG,
 589                         "(%d) failed to map desc ring.\n",
 590                         dev->vid);
 591                 return dev;
 592         }
 593
 594         dev = numa_realloc(dev, vq_index);
 595         vq = dev->virtqueue[vq_index];
 596         addr = &vq->ring_addrs;
 597
 598         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 599         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 600                         vq, addr->avail_user_addr, &len);
 601         if (vq->avail == 0 ||
 602                         len != sizeof(struct vring_avail) +
 603                         sizeof(uint16_t) * vq->size) {
 604                 RTE_LOG(DEBUG, VHOST_CONFIG,
 605                         "(%d) failed to map avail ring.\n",
 606                         dev->vid);
 607                 return dev;
 608         }
 609
 610         len = sizeof(struct vring_used) +
 611                 sizeof(struct vring_used_elem) * vq->size;
 612         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 613                         vq, addr->used_user_addr, &len);
 614         if (vq->used == 0 || len != sizeof(struct vring_used) +
 615                         sizeof(struct vring_used_elem) * vq->size) {
 616                 RTE_LOG(DEBUG, VHOST_CONFIG,
 617                         "(%d) failed to map used ring.\n",
 618                         dev->vid);
 619                 return dev;
 620         }
 621
 622         if (vq->last_used_idx != vq->used->idx) {
 623                 RTE_LOG(WARNING, VHOST_CONFIG,
 624                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 625                         "some packets maybe resent for Tx and dropped for Rx\n",
 626                         vq->last_used_idx, vq->used->idx);
 627                 vq->last_used_idx  = vq->used->idx;
 628                 vq->last_avail_idx = vq->used->idx;
 629         }
 630
 631         vq->log_guest_addr = addr->log_guest_addr;
 632
 633         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
 634                         dev->vid, vq->desc);
 635         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
 636                         dev->vid, vq->avail);
 637         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
 638                         dev->vid, vq->used);
 639         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
 640                         dev->vid, vq->log_guest_addr);
 641
 642         return dev;
 643 }
 644
 645 /*
 646  * The virtio device sends us the desc, used and avail ring addresses.
 647  * This function then converts these to our address space.
 648  */
 649 static int
 650 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 651                         int main_fd __rte_unused)
 652 {
 653         struct virtio_net *dev = *pdev;
 654         struct vhost_virtqueue *vq;
 655         struct vhost_vring_addr *addr = &msg->payload.addr;
 656
 657         if (dev->mem == NULL)
 658                 return VH_RESULT_ERR;
 659
 660         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 661         vq = dev->virtqueue[msg->payload.addr.index];
 662
 663         /*
 664          * Rings addresses should not be interpreted as long as the ring is not
 665          * started and enabled
 666          */
 667         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 668
 669         vring_invalidate(dev, vq);
 670
 671         if (vq->enabled && (dev->features &
 672                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
 673                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 674                 if (!dev)
 675                         return VH_RESULT_ERR;
 676
 677                 *pdev = dev;
 678         }
 679
 680         return VH_RESULT_OK;
 681 }
 682
 683 /*
 684  * The virtio device sends us the available ring last used index.
 685  */
 686 static int
 687 vhost_user_set_vring_base(struct virtio_net **pdev,
 688                         struct VhostUserMsg *msg,
 689                         int main_fd __rte_unused)
 690 {
 691         struct virtio_net *dev = *pdev;
 692         dev->virtqueue[msg->payload.state.index]->last_used_idx  =
 693                         msg->payload.state.num;
 694         dev->virtqueue[msg->payload.state.index]->last_avail_idx =
 695                         msg->payload.state.num;
 696
 697         return VH_RESULT_OK;
 698 }
 699
 700 static int
 701 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 702                    uint64_t host_phys_addr, uint64_t size)
 703 {
 704         struct guest_page *page, *last_page;
 705
 706         if (dev->nr_guest_pages == dev->max_guest_pages) {
 707                 dev->max_guest_pages *= 2;
 708                 dev->guest_pages = realloc(dev->guest_pages,
 709                                         dev->max_guest_pages * sizeof(*page));
 710                 if (!dev->guest_pages) {
 711                         RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
 712                         return -1;
 713                 }
 714         }
 715
 716         if (dev->nr_guest_pages > 0) {
 717                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 718                 /* merge if the two pages are continuous */
 719                 if (host_phys_addr == last_page->host_phys_addr +
 720                                       last_page->size) {
 721                         last_page->size += size;
 722                         return 0;
 723                 }
 724         }
 725
 726         page = &dev->guest_pages[dev->nr_guest_pages++];
 727         page->guest_phys_addr = guest_phys_addr;
 728         page->host_phys_addr  = host_phys_addr;
 729         page->size = size;
 730
 731         return 0;
 732 }
 733
 734 static int
 735 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 736                 uint64_t page_size)
 737 {
 738         uint64_t reg_size = reg->size;
 739         uint64_t host_user_addr  = reg->host_user_addr;
 740         uint64_t guest_phys_addr = reg->guest_phys_addr;
 741         uint64_t host_phys_addr;
 742         uint64_t size;
 743
 744         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 745         size = page_size - (guest_phys_addr & (page_size - 1));
 746         size = RTE_MIN(size, reg_size);
 747
 748         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 749                 return -1;
 750
 751         host_user_addr  += size;
 752         guest_phys_addr += size;
 753         reg_size -= size;
 754
 755         while (reg_size > 0) {
 756                 size = RTE_MIN(reg_size, page_size);
 757                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 758                                                   host_user_addr);
 759                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 760                                 size) < 0)
 761                         return -1;
 762
 763                 host_user_addr  += size;
 764                 guest_phys_addr += size;
 765                 reg_size -= size;
 766         }
 767
 768         return 0;
 769 }
 770
 771 #ifdef RTE_LIBRTE_VHOST_DEBUG
 772 /* TODO: enable it only in debug mode? */
 773 static void
 774 dump_guest_pages(struct virtio_net *dev)
 775 {
 776         uint32_t i;
 777         struct guest_page *page;
 778
 779         for (i = 0; i < dev->nr_guest_pages; i++) {
 780                 page = &dev->guest_pages[i];
 781
 782                 RTE_LOG(INFO, VHOST_CONFIG,
 783                         "guest physical page region %u\n"
 784                         "\t guest_phys_addr: %" PRIx64 "\n"
 785                         "\t host_phys_addr : %" PRIx64 "\n"
 786                         "\t size           : %" PRIx64 "\n",
 787                         i,
 788                         page->guest_phys_addr,
 789                         page->host_phys_addr,
 790                         page->size);
 791         }
 792 }
 793 #else
 794 #define dump_guest_pages(dev)
 795 #endif
 796
 797 static bool
 798 vhost_memory_changed(struct VhostUserMemory *new,
 799                      struct rte_vhost_memory *old)
 800 {
 801         uint32_t i;
 802
 803         if (new->nregions != old->nregions)
 804                 return true;
 805
 806         for (i = 0; i < new->nregions; ++i) {
 807                 VhostUserMemoryRegion *new_r = &new->regions[i];
 808                 struct rte_vhost_mem_region *old_r = &old->regions[i];
 809
 810                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
 811                         return true;
 812                 if (new_r->memory_size != old_r->size)
 813                         return true;
 814                 if (new_r->userspace_addr != old_r->guest_user_addr)
 815                         return true;
 816         }
 817
 818         return false;
 819 }
 820
 821 static int
 822 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
 823                         int main_fd __rte_unused)
 824 {
 825         struct virtio_net *dev = *pdev;
 826         struct VhostUserMemory memory = msg->payload.memory;
 827         struct rte_vhost_mem_region *reg;
 828         void *mmap_addr;
 829         uint64_t mmap_size;
 830         uint64_t mmap_offset;
 831         uint64_t alignment;
 832         uint32_t i;
 833         int populate;
 834         int fd;
 835
 836         if (memory.nregions > VHOST_MEMORY_MAX_NREGIONS) {
 837                 RTE_LOG(ERR, VHOST_CONFIG,
 838                         "too many memory regions (%u)\n", memory.nregions);
 839                 return VH_RESULT_ERR;
 840         }
 841
 842         if (dev->mem && !vhost_memory_changed(&memory, dev->mem)) {
 843                 RTE_LOG(INFO, VHOST_CONFIG,
 844                         "(%d) memory regions not changed\n", dev->vid);
 845
 846                 for (i = 0; i < memory.nregions; i++)
 847                         close(msg->fds[i]);
 848
 849                 return VH_RESULT_OK;
 850         }
 851
 852         if (dev->mem) {
 853                 free_mem_region(dev);
 854                 rte_free(dev->mem);
 855                 dev->mem = NULL;
 856         }
 857
 858         /* Flush IOTLB cache as previous HVAs are now invalid */
 859         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 860                 for (i = 0; i < dev->nr_vring; i++)
 861                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
 862
 863         dev->nr_guest_pages = 0;
 864         if (!dev->guest_pages) {
 865                 dev->max_guest_pages = 8;
 866                 dev->guest_pages = malloc(dev->max_guest_pages *
 867                                                 sizeof(struct guest_page));
 868                 if (dev->guest_pages == NULL) {
 869                         RTE_LOG(ERR, VHOST_CONFIG,
 870                                 "(%d) failed to allocate memory "
 871                                 "for dev->guest_pages\n",
 872                                 dev->vid);
 873                         return VH_RESULT_ERR;
 874                 }
 875         }
 876
 877         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
 878                 sizeof(struct rte_vhost_mem_region) * memory.nregions, 0);
 879         if (dev->mem == NULL) {
 880                 RTE_LOG(ERR, VHOST_CONFIG,
 881                         "(%d) failed to allocate memory for dev->mem\n",
 882                         dev->vid);
 883                 return VH_RESULT_ERR;
 884         }
 885         dev->mem->nregions = memory.nregions;
 886
 887         for (i = 0; i < memory.nregions; i++) {
 888                 fd  = msg->fds[i];
 889                 reg = &dev->mem->regions[i];
 890
 891                 reg->guest_phys_addr = memory.regions[i].guest_phys_addr;
 892                 reg->guest_user_addr = memory.regions[i].userspace_addr;
 893                 reg->size            = memory.regions[i].memory_size;
 894                 reg->fd              = fd;
 895
 896                 mmap_offset = memory.regions[i].mmap_offset;
 897
 898                 /* Check for memory_size + mmap_offset overflow */
 899                 if (mmap_offset >= -reg->size) {
 900                         RTE_LOG(ERR, VHOST_CONFIG,
 901                                 "mmap_offset (%#"PRIx64") and memory_size "
 902                                 "(%#"PRIx64") overflow\n",
 903                                 mmap_offset, reg->size);
 904                         goto err_mmap;
 905                 }
 906
 907                 mmap_size = reg->size + mmap_offset;
 908
 909                 /* mmap() without flag of MAP_ANONYMOUS, should be called
 910                  * with length argument aligned with hugepagesz at older
 911                  * longterm version Linux, like 2.6.32 and 3.2.72, or
 912                  * mmap() will fail with EINVAL.
 913                  *
 914                  * to avoid failure, make sure in caller to keep length
 915                  * aligned.
 916                  */
 917                 alignment = get_blk_size(fd);
 918                 if (alignment == (uint64_t)-1) {
 919                         RTE_LOG(ERR, VHOST_CONFIG,
 920                                 "couldn't get hugepage size through fstat\n");
 921                         goto err_mmap;
 922                 }
 923                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
 924
 925                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
 926                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
 927                                  MAP_SHARED | populate, fd, 0);
 928
 929                 if (mmap_addr == MAP_FAILED) {
 930                         RTE_LOG(ERR, VHOST_CONFIG,
 931                                 "mmap region %u failed.\n", i);
 932                         goto err_mmap;
 933                 }
 934
 935                 reg->mmap_addr = mmap_addr;
 936                 reg->mmap_size = mmap_size;
 937                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
 938                                       mmap_offset;
 939
 940                 if (dev->dequeue_zero_copy)
 941                         if (add_guest_pages(dev, reg, alignment) < 0) {
 942                                 RTE_LOG(ERR, VHOST_CONFIG,
 943                                         "adding guest pages to region %u failed.\n",
 944                                         i);
 945                                 goto err_mmap;
 946                         }
 947
 948                 RTE_LOG(INFO, VHOST_CONFIG,
 949                         "guest memory region %u, size: 0x%" PRIx64 "\n"
 950                         "\t guest physical addr: 0x%" PRIx64 "\n"
 951                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
 952                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
 953                         "\t mmap addr : 0x%" PRIx64 "\n"
 954                         "\t mmap size : 0x%" PRIx64 "\n"
 955                         "\t mmap align: 0x%" PRIx64 "\n"
 956                         "\t mmap off  : 0x%" PRIx64 "\n",
 957                         i, reg->size,
 958                         reg->guest_phys_addr,
 959                         reg->guest_user_addr,
 960                         reg->host_user_addr,
 961                         (uint64_t)(uintptr_t)mmap_addr,
 962                         mmap_size,
 963                         alignment,
 964                         mmap_offset);
 965         }
 966
 967         for (i = 0; i < dev->nr_vring; i++) {
 968                 struct vhost_virtqueue *vq = dev->virtqueue[i];
 969
 970                 if (vq->desc || vq->avail || vq->used) {
 971                         /*
 972                          * If the memory table got updated, the ring addresses
 973                          * need to be translated again as virtual addresses have
 974                          * changed.
 975                          */
 976                         vring_invalidate(dev, vq);
 977
 978                         dev = translate_ring_addresses(dev, i);
 979                         if (!dev) {
 980                                 dev = *pdev;
 981                                 goto err_mmap;
 982                         }
 983
 984                         *pdev = dev;
 985                 }
 986         }
 987
 988         dump_guest_pages(dev);
 989
 990         return VH_RESULT_OK;
 991
 992 err_mmap:
 993         free_mem_region(dev);
 994         rte_free(dev->mem);
 995         dev->mem = NULL;
 996         return VH_RESULT_ERR;
 997 }
 998
 999 static bool
1000 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1001 {
1002         bool rings_ok;
1003
1004         if (!vq)
1005                 return false;
1006
1007         if (vq_is_packed(dev))
1008                 rings_ok = !!vq->desc_packed;
1009         else
1010                 rings_ok = vq->desc && vq->avail && vq->used;
1011
1012         return rings_ok &&
1013                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1014                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1015 }
1016
1017 static int
1018 virtio_is_ready(struct virtio_net *dev)
1019 {
1020         struct vhost_virtqueue *vq;
1021         uint32_t i;
1022
1023         if (dev->nr_vring == 0)
1024                 return 0;
1025
1026         for (i = 0; i < dev->nr_vring; i++) {
1027                 vq = dev->virtqueue[i];
1028
1029                 if (!vq_is_ready(dev, vq))
1030                         return 0;
1031         }
1032
1033         RTE_LOG(INFO, VHOST_CONFIG,
1034                 "virtio is now ready for processing.\n");
1035         return 1;
1036 }
1037
1038 static int
1039 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1040                         int main_fd __rte_unused)
1041 {
1042         struct virtio_net *dev = *pdev;
1043         struct vhost_vring_file file;
1044         struct vhost_virtqueue *vq;
1045
1046         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1047         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1048                 file.fd = VIRTIO_INVALID_EVENTFD;
1049         else
1050                 file.fd = msg->fds[0];
1051         RTE_LOG(INFO, VHOST_CONFIG,
1052                 "vring call idx:%d file:%d\n", file.index, file.fd);
1053
1054         vq = dev->virtqueue[file.index];
1055         if (vq->callfd >= 0)
1056                 close(vq->callfd);
1057
1058         vq->callfd = file.fd;
1059
1060         return VH_RESULT_OK;
1061 }
1062
1063 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1064                         struct VhostUserMsg *msg,
1065                         int main_fd __rte_unused)
1066 {
1067         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1068                 close(msg->fds[0]);
1069         RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1070
1071         return VH_RESULT_OK;
1072 }
1073
1074 static int
1075 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1076                         int main_fd __rte_unused)
1077 {
1078         struct virtio_net *dev = *pdev;
1079         struct vhost_vring_file file;
1080         struct vhost_virtqueue *vq;
1081
1082         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1083         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1084                 file.fd = VIRTIO_INVALID_EVENTFD;
1085         else
1086                 file.fd = msg->fds[0];
1087         RTE_LOG(INFO, VHOST_CONFIG,
1088                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1089
1090         /* Interpret ring addresses only when ring is started. */
1091         dev = translate_ring_addresses(dev, file.index);
1092         if (!dev)
1093                 return VH_RESULT_ERR;
1094
1095         *pdev = dev;
1096
1097         vq = dev->virtqueue[file.index];
1098
1099         /*
1100          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1101          * the ring starts already enabled. Otherwise, it is enabled via
1102          * the SET_VRING_ENABLE message.
1103          */
1104         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)))
1105                 vq->enabled = 1;
1106
1107         if (vq->kickfd >= 0)
1108                 close(vq->kickfd);
1109         vq->kickfd = file.fd;
1110
1111         return VH_RESULT_OK;
1112 }
1113
1114 static void
1115 free_zmbufs(struct vhost_virtqueue *vq)
1116 {
1117         struct zcopy_mbuf *zmbuf, *next;
1118
1119         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1120              zmbuf != NULL; zmbuf = next) {
1121                 next = TAILQ_NEXT(zmbuf, next);
1122
1123                 rte_pktmbuf_free(zmbuf->mbuf);
1124                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1125         }
1126
1127         rte_free(vq->zmbufs);
1128 }
1129
1130 /*
1131  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1132  */
1133 static int
1134 vhost_user_get_vring_base(struct virtio_net **pdev,
1135                         struct VhostUserMsg *msg,
1136                         int main_fd __rte_unused)
1137 {
1138         struct virtio_net *dev = *pdev;
1139         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1140
1141         /* We have to stop the queue (virtio) if it is running. */
1142         vhost_destroy_device_notify(dev);
1143
1144         dev->flags &= ~VIRTIO_DEV_READY;
1145         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1146
1147         /* Here we are safe to get the last avail index */
1148         msg->payload.state.num = vq->last_avail_idx;
1149
1150         RTE_LOG(INFO, VHOST_CONFIG,
1151                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1152                 msg->payload.state.num);
1153         /*
1154          * Based on current qemu vhost-user implementation, this message is
1155          * sent and only sent in vhost_vring_stop.
1156          * TODO: cleanup the vring, it isn't usable since here.
1157          */
1158         if (vq->kickfd >= 0)
1159                 close(vq->kickfd);
1160
1161         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1162
1163         if (vq->callfd >= 0)
1164                 close(vq->callfd);
1165
1166         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1167
1168         if (dev->dequeue_zero_copy)
1169                 free_zmbufs(vq);
1170         if (vq_is_packed(dev)) {
1171                 rte_free(vq->shadow_used_packed);
1172                 vq->shadow_used_packed = NULL;
1173         } else {
1174                 rte_free(vq->shadow_used_split);
1175                 vq->shadow_used_split = NULL;
1176         }
1177
1178         rte_free(vq->batch_copy_elems);
1179         vq->batch_copy_elems = NULL;
1180
1181         msg->size = sizeof(msg->payload.state);
1182         msg->fd_num = 0;
1183
1184         return VH_RESULT_REPLY;
1185 }
1186
1187 /*
1188  * when virtio queues are ready to work, qemu will send us to
1189  * enable the virtio queue pair.
1190  */
1191 static int
1192 vhost_user_set_vring_enable(struct virtio_net **pdev,
1193                         struct VhostUserMsg *msg,
1194                         int main_fd __rte_unused)
1195 {
1196         struct virtio_net *dev = *pdev;
1197         int enable = (int)msg->payload.state.num;
1198         int index = (int)msg->payload.state.index;
1199         struct rte_vdpa_device *vdpa_dev;
1200         int did = -1;
1201
1202         RTE_LOG(INFO, VHOST_CONFIG,
1203                 "set queue enable: %d to qp idx: %d\n",
1204                 enable, index);
1205
1206         did = dev->vdpa_dev_id;
1207         vdpa_dev = rte_vdpa_get_device(did);
1208         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1209                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1210
1211         if (dev->notify_ops->vring_state_changed)
1212                 dev->notify_ops->vring_state_changed(dev->vid,
1213                                 index, enable);
1214
1215         dev->virtqueue[index]->enabled = enable;
1216
1217         return VH_RESULT_OK;
1218 }
1219
1220 static int
1221 vhost_user_get_protocol_features(struct virtio_net **pdev,
1222                         struct VhostUserMsg *msg,
1223                         int main_fd __rte_unused)
1224 {
1225         struct virtio_net *dev = *pdev;
1226         uint64_t features, protocol_features;
1227
1228         rte_vhost_driver_get_features(dev->ifname, &features);
1229         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1230
1231         /*
1232          * REPLY_ACK protocol feature is only mandatory for now
1233          * for IOMMU feature. If IOMMU is explicitly disabled by the
1234          * application, disable also REPLY_ACK feature for older buggy
1235          * Qemu versions (from v2.7.0 to v2.9.0).
1236          */
1237         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1238                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1239
1240         msg->payload.u64 = protocol_features;
1241         msg->size = sizeof(msg->payload.u64);
1242         msg->fd_num = 0;
1243
1244         return VH_RESULT_REPLY;
1245 }
1246
1247 static int
1248 vhost_user_set_protocol_features(struct virtio_net **pdev,
1249                         struct VhostUserMsg *msg,
1250                         int main_fd __rte_unused)
1251 {
1252         struct virtio_net *dev = *pdev;
1253         uint64_t protocol_features = msg->payload.u64;
1254         if (protocol_features & ~VHOST_USER_PROTOCOL_FEATURES) {
1255                 RTE_LOG(ERR, VHOST_CONFIG,
1256                         "(%d) received invalid protocol features.\n",
1257                         dev->vid);
1258                 return VH_RESULT_ERR;
1259         }
1260
1261         dev->protocol_features = protocol_features;
1262
1263         return VH_RESULT_OK;
1264 }
1265
1266 static int
1267 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1268                         int main_fd __rte_unused)
1269 {
1270         struct virtio_net *dev = *pdev;
1271         int fd = msg->fds[0];
1272         uint64_t size, off;
1273         void *addr;
1274
1275         if (fd < 0) {
1276                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1277                 return VH_RESULT_ERR;
1278         }
1279
1280         if (msg->size != sizeof(VhostUserLog)) {
1281                 RTE_LOG(ERR, VHOST_CONFIG,
1282                         "invalid log base msg size: %"PRId32" != %d\n",
1283                         msg->size, (int)sizeof(VhostUserLog));
1284                 return VH_RESULT_ERR;
1285         }
1286
1287         size = msg->payload.log.mmap_size;
1288         off  = msg->payload.log.mmap_offset;
1289
1290         /* Don't allow mmap_offset to point outside the mmap region */
1291         if (off > size) {
1292                 RTE_LOG(ERR, VHOST_CONFIG,
1293                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1294                         off, size);
1295                 return VH_RESULT_ERR;
1296         }
1297
1298         RTE_LOG(INFO, VHOST_CONFIG,
1299                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1300                 size, off);
1301
1302         /*
1303          * mmap from 0 to workaround a hugepage mmap bug: mmap will
1304          * fail when offset is not page size aligned.
1305          */
1306         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1307         close(fd);
1308         if (addr == MAP_FAILED) {
1309                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1310                 return VH_RESULT_ERR;
1311         }
1312
1313         /*
1314          * Free previously mapped log memory on occasionally
1315          * multiple VHOST_USER_SET_LOG_BASE.
1316          */
1317         if (dev->log_addr) {
1318                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1319         }
1320         dev->log_addr = (uint64_t)(uintptr_t)addr;
1321         dev->log_base = dev->log_addr + off;
1322         dev->log_size = size;
1323
1324         /*
1325          * The spec is not clear about it (yet), but QEMU doesn't expect
1326          * any payload in the reply.
1327          */
1328         msg->size = 0;
1329         msg->fd_num = 0;
1330
1331         return VH_RESULT_REPLY;
1332 }
1333
1334 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1335                         struct VhostUserMsg *msg,
1336                         int main_fd __rte_unused)
1337 {
1338         close(msg->fds[0]);
1339         RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1340
1341         return VH_RESULT_OK;
1342 }
1343
1344 /*
1345  * An rarp packet is constructed and broadcasted to notify switches about
1346  * the new location of the migrated VM, so that packets from outside will
1347  * not be lost after migration.
1348  *
1349  * However, we don't actually "send" a rarp packet here, instead, we set
1350  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1351  */
1352 static int
1353 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
1354                         int main_fd __rte_unused)
1355 {
1356         struct virtio_net *dev = *pdev;
1357         uint8_t *mac = (uint8_t *)&msg->payload.u64;
1358         struct rte_vdpa_device *vdpa_dev;
1359         int did = -1;
1360
1361         RTE_LOG(DEBUG, VHOST_CONFIG,
1362                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1363                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1364         memcpy(dev->mac.addr_bytes, mac, 6);
1365
1366         /*
1367          * Set the flag to inject a RARP broadcast packet at
1368          * rte_vhost_dequeue_burst().
1369          *
1370          * rte_smp_wmb() is for making sure the mac is copied
1371          * before the flag is set.
1372          */
1373         rte_smp_wmb();
1374         rte_atomic16_set(&dev->broadcast_rarp, 1);
1375         did = dev->vdpa_dev_id;
1376         vdpa_dev = rte_vdpa_get_device(did);
1377         if (vdpa_dev && vdpa_dev->ops->migration_done)
1378                 vdpa_dev->ops->migration_done(dev->vid);
1379
1380         return VH_RESULT_OK;
1381 }
1382
1383 static int
1384 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
1385                         int main_fd __rte_unused)
1386 {
1387         struct virtio_net *dev = *pdev;
1388         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1389                         msg->payload.u64 > VIRTIO_MAX_MTU) {
1390                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1391                                 msg->payload.u64);
1392
1393                 return VH_RESULT_ERR;
1394         }
1395
1396         dev->mtu = msg->payload.u64;
1397
1398         return VH_RESULT_OK;
1399 }
1400
1401 static int
1402 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
1403                         int main_fd __rte_unused)
1404 {
1405         struct virtio_net *dev = *pdev;
1406         int fd = msg->fds[0];
1407
1408         if (fd < 0) {
1409                 RTE_LOG(ERR, VHOST_CONFIG,
1410                                 "Invalid file descriptor for slave channel (%d)\n",
1411                                 fd);
1412                 return VH_RESULT_ERR;
1413         }
1414
1415         dev->slave_req_fd = fd;
1416
1417         return VH_RESULT_OK;
1418 }
1419
1420 static int
1421 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1422 {
1423         struct vhost_vring_addr *ra;
1424         uint64_t start, end;
1425
1426         start = imsg->iova;
1427         end = start + imsg->size;
1428
1429         ra = &vq->ring_addrs;
1430         if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1431                 return 1;
1432         if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1433                 return 1;
1434         if (ra->used_user_addr >= start && ra->used_user_addr < end)
1435                 return 1;
1436
1437         return 0;
1438 }
1439
1440 static int
1441 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1442                                 struct vhost_iotlb_msg *imsg)
1443 {
1444         uint64_t istart, iend, vstart, vend;
1445
1446         istart = imsg->iova;
1447         iend = istart + imsg->size - 1;
1448
1449         vstart = (uintptr_t)vq->desc;
1450         vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1451         if (vstart <= iend && istart <= vend)
1452                 return 1;
1453
1454         vstart = (uintptr_t)vq->avail;
1455         vend = vstart + sizeof(struct vring_avail);
1456         vend += sizeof(uint16_t) * vq->size - 1;
1457         if (vstart <= iend && istart <= vend)
1458                 return 1;
1459
1460         vstart = (uintptr_t)vq->used;
1461         vend = vstart + sizeof(struct vring_used);
1462         vend += sizeof(struct vring_used_elem) * vq->size - 1;
1463         if (vstart <= iend && istart <= vend)
1464                 return 1;
1465
1466         return 0;
1467 }
1468
1469 static int
1470 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
1471                         int main_fd __rte_unused)
1472 {
1473         struct virtio_net *dev = *pdev;
1474         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1475         uint16_t i;
1476         uint64_t vva, len;
1477
1478         switch (imsg->type) {
1479         case VHOST_IOTLB_UPDATE:
1480                 len = imsg->size;
1481                 vva = qva_to_vva(dev, imsg->uaddr, &len);
1482                 if (!vva)
1483                         return VH_RESULT_ERR;
1484
1485                 for (i = 0; i < dev->nr_vring; i++) {
1486                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1487
1488                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1489                                         len, imsg->perm);
1490
1491                         if (is_vring_iotlb_update(vq, imsg))
1492                                 *pdev = dev = translate_ring_addresses(dev, i);
1493                 }
1494                 break;
1495         case VHOST_IOTLB_INVALIDATE:
1496                 for (i = 0; i < dev->nr_vring; i++) {
1497                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1498
1499                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
1500                                         imsg->size);
1501
1502                         if (is_vring_iotlb_invalidate(vq, imsg))
1503                                 vring_invalidate(dev, vq);
1504                 }
1505                 break;
1506         default:
1507                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1508                                 imsg->type);
1509                 return VH_RESULT_ERR;
1510         }
1511
1512         return VH_RESULT_OK;
1513 }
1514
1515 static int
1516 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
1517                         struct VhostUserMsg *msg,
1518                         int main_fd __rte_unused)
1519 {
1520         struct virtio_net *dev = *pdev;
1521 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1522         struct uffdio_api api_struct;
1523
1524         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1525
1526         if (dev->postcopy_ufd == -1) {
1527                 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
1528                         strerror(errno));
1529                 return VH_RESULT_ERR;
1530         }
1531         api_struct.api = UFFD_API;
1532         api_struct.features = 0;
1533         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
1534                 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
1535                         strerror(errno));
1536                 close(dev->postcopy_ufd);
1537                 dev->postcopy_ufd = -1;
1538                 return VH_RESULT_ERR;
1539         }
1540         msg->fds[0] = dev->postcopy_ufd;
1541         msg->fd_num = 1;
1542
1543         return VH_RESULT_REPLY;
1544 #else
1545         dev->postcopy_ufd = -1;
1546         msg->fd_num = 0;
1547
1548         return VH_RESULT_ERR;
1549 #endif
1550 }
1551
1552 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
1553                                         struct VhostUserMsg *msg,
1554                                         int main_fd);
1555 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
1556         [VHOST_USER_NONE] = NULL,
1557         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
1558         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
1559         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
1560         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
1561         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
1562         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
1563         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
1564         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
1565         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
1566         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
1567         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
1568         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
1569         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
1570         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
1571         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
1572         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
1573         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
1574         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
1575         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
1576         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
1577         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
1578         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
1579         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
1580 };
1581
1582
1583 /* return bytes# of read on success or negative val on failure. */
1584 static int
1585 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1586 {
1587         int ret;
1588
1589         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1590                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
1591         if (ret <= 0)
1592                 return ret;
1593
1594         if (msg && msg->size) {
1595                 if (msg->size > sizeof(msg->payload)) {
1596                         RTE_LOG(ERR, VHOST_CONFIG,
1597                                 "invalid msg size: %d\n", msg->size);
1598                         return -1;
1599                 }
1600                 ret = read(sockfd, &msg->payload, msg->size);
1601                 if (ret <= 0)
1602                         return ret;
1603                 if (ret != (int)msg->size) {
1604                         RTE_LOG(ERR, VHOST_CONFIG,
1605                                 "read control message failed\n");
1606                         return -1;
1607                 }
1608         }
1609
1610         return ret;
1611 }
1612
1613 static int
1614 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1615 {
1616         if (!msg)
1617                 return 0;
1618
1619         return send_fd_message(sockfd, (char *)msg,
1620                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
1621 }
1622
1623 static int
1624 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1625 {
1626         if (!msg)
1627                 return 0;
1628
1629         msg->flags &= ~VHOST_USER_VERSION_MASK;
1630         msg->flags &= ~VHOST_USER_NEED_REPLY;
1631         msg->flags |= VHOST_USER_VERSION;
1632         msg->flags |= VHOST_USER_REPLY_MASK;
1633
1634         return send_vhost_message(sockfd, msg);
1635 }
1636
1637 static int
1638 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
1639 {
1640         int ret;
1641
1642         if (msg->flags & VHOST_USER_NEED_REPLY)
1643                 rte_spinlock_lock(&dev->slave_req_lock);
1644
1645         ret = send_vhost_message(dev->slave_req_fd, msg);
1646         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1647                 rte_spinlock_unlock(&dev->slave_req_lock);
1648
1649         return ret;
1650 }
1651
1652 /*
1653  * Allocate a queue pair if it hasn't been allocated yet
1654  */
1655 static int
1656 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1657                         struct VhostUserMsg *msg)
1658 {
1659         uint16_t vring_idx;
1660
1661         switch (msg->request.master) {
1662         case VHOST_USER_SET_VRING_KICK:
1663         case VHOST_USER_SET_VRING_CALL:
1664         case VHOST_USER_SET_VRING_ERR:
1665                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1666                 break;
1667         case VHOST_USER_SET_VRING_NUM:
1668         case VHOST_USER_SET_VRING_BASE:
1669         case VHOST_USER_SET_VRING_ENABLE:
1670                 vring_idx = msg->payload.state.index;
1671                 break;
1672         case VHOST_USER_SET_VRING_ADDR:
1673                 vring_idx = msg->payload.addr.index;
1674                 break;
1675         default:
1676                 return 0;
1677         }
1678
1679         if (vring_idx >= VHOST_MAX_VRING) {
1680                 RTE_LOG(ERR, VHOST_CONFIG,
1681                         "invalid vring index: %u\n", vring_idx);
1682                 return -1;
1683         }
1684
1685         if (dev->virtqueue[vring_idx])
1686                 return 0;
1687
1688         return alloc_vring_queue(dev, vring_idx);
1689 }
1690
1691 static void
1692 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1693 {
1694         unsigned int i = 0;
1695         unsigned int vq_num = 0;
1696
1697         while (vq_num < dev->nr_vring) {
1698                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1699
1700                 if (vq) {
1701                         rte_spinlock_lock(&vq->access_lock);
1702                         vq_num++;
1703                 }
1704                 i++;
1705         }
1706 }
1707
1708 static void
1709 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1710 {
1711         unsigned int i = 0;
1712         unsigned int vq_num = 0;
1713
1714         while (vq_num < dev->nr_vring) {
1715                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1716
1717                 if (vq) {
1718                         rte_spinlock_unlock(&vq->access_lock);
1719                         vq_num++;
1720                 }
1721                 i++;
1722         }
1723 }
1724
1725 int
1726 vhost_user_msg_handler(int vid, int fd)
1727 {
1728         struct virtio_net *dev;
1729         struct VhostUserMsg msg;
1730         struct rte_vdpa_device *vdpa_dev;
1731         int did = -1;
1732         int ret;
1733         int unlock_required = 0;
1734         uint32_t skip_master = 0;
1735         int request;
1736
1737         dev = get_device(vid);
1738         if (dev == NULL)
1739                 return -1;
1740
1741         if (!dev->notify_ops) {
1742                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1743                 if (!dev->notify_ops) {
1744                         RTE_LOG(ERR, VHOST_CONFIG,
1745                                 "failed to get callback ops for driver %s\n",
1746                                 dev->ifname);
1747                         return -1;
1748                 }
1749         }
1750
1751         ret = read_vhost_message(fd, &msg);
1752         if (ret <= 0 || msg.request.master >= VHOST_USER_MAX) {
1753                 if (ret < 0)
1754                         RTE_LOG(ERR, VHOST_CONFIG,
1755                                 "vhost read message failed\n");
1756                 else if (ret == 0)
1757                         RTE_LOG(INFO, VHOST_CONFIG,
1758                                 "vhost peer closed\n");
1759                 else
1760                         RTE_LOG(ERR, VHOST_CONFIG,
1761                                 "vhost read incorrect message\n");
1762
1763                 return -1;
1764         }
1765
1766         ret = 0;
1767         if (msg.request.master != VHOST_USER_IOTLB_MSG)
1768                 RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1769                         vhost_message_str[msg.request.master]);
1770         else
1771                 RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1772                         vhost_message_str[msg.request.master]);
1773
1774         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1775         if (ret < 0) {
1776                 RTE_LOG(ERR, VHOST_CONFIG,
1777                         "failed to alloc queue\n");
1778                 return -1;
1779         }
1780
1781         /*
1782          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1783          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1784          * and device is destroyed. destroy_device waits for queues to be
1785          * inactive, so it is safe. Otherwise taking the access_lock
1786          * would cause a dead lock.
1787          */
1788         switch (msg.request.master) {
1789         case VHOST_USER_SET_FEATURES:
1790         case VHOST_USER_SET_PROTOCOL_FEATURES:
1791         case VHOST_USER_SET_OWNER:
1792         case VHOST_USER_SET_MEM_TABLE:
1793         case VHOST_USER_SET_LOG_BASE:
1794         case VHOST_USER_SET_LOG_FD:
1795         case VHOST_USER_SET_VRING_NUM:
1796         case VHOST_USER_SET_VRING_ADDR:
1797         case VHOST_USER_SET_VRING_BASE:
1798         case VHOST_USER_SET_VRING_KICK:
1799         case VHOST_USER_SET_VRING_CALL:
1800         case VHOST_USER_SET_VRING_ERR:
1801         case VHOST_USER_SET_VRING_ENABLE:
1802         case VHOST_USER_SEND_RARP:
1803         case VHOST_USER_NET_SET_MTU:
1804         case VHOST_USER_SET_SLAVE_REQ_FD:
1805                 vhost_user_lock_all_queue_pairs(dev);
1806                 unlock_required = 1;
1807                 break;
1808         default:
1809                 break;
1810
1811         }
1812
1813         if (dev->extern_ops.pre_msg_handle) {
1814                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
1815                                 (void *)&msg, &skip_master);
1816                 if (ret == VH_RESULT_ERR)
1817                         goto skip_to_reply;
1818                 else if (ret == VH_RESULT_REPLY)
1819                         send_vhost_reply(fd, &msg);
1820
1821                 if (skip_master)
1822                         goto skip_to_post_handle;
1823         }
1824
1825         request = msg.request.master;
1826         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
1827                 if (!vhost_message_handlers[request])
1828                         goto skip_to_post_handle;
1829                 ret = vhost_message_handlers[request](&dev, &msg, fd);
1830
1831                 switch (ret) {
1832                 case VH_RESULT_ERR:
1833                         RTE_LOG(ERR, VHOST_CONFIG,
1834                                 "Processing %s failed.\n",
1835                                 vhost_message_str[request]);
1836                         break;
1837                 case VH_RESULT_OK:
1838                         RTE_LOG(DEBUG, VHOST_CONFIG,
1839                                 "Processing %s succeeded.\n",
1840                                 vhost_message_str[request]);
1841                         break;
1842                 case VH_RESULT_REPLY:
1843                         RTE_LOG(DEBUG, VHOST_CONFIG,
1844                                 "Processing %s succeeded and needs reply.\n",
1845                                 vhost_message_str[request]);
1846                         send_vhost_reply(fd, &msg);
1847                         break;
1848                 }
1849         } else {
1850                 RTE_LOG(ERR, VHOST_CONFIG,
1851                         "Requested invalid message type %d.\n", request);
1852                 ret = VH_RESULT_ERR;
1853         }
1854
1855 skip_to_post_handle:
1856         if (ret != VH_RESULT_ERR && dev->extern_ops.post_msg_handle) {
1857                 ret = (*dev->extern_ops.post_msg_handle)(
1858                                 dev->vid, (void *)&msg);
1859                 if (ret == VH_RESULT_ERR)
1860                         goto skip_to_reply;
1861                 else if (ret == VH_RESULT_REPLY)
1862                         send_vhost_reply(fd, &msg);
1863         }
1864
1865 skip_to_reply:
1866         if (unlock_required)
1867                 vhost_user_unlock_all_queue_pairs(dev);
1868
1869         /*
1870          * If the request required a reply that was already sent,
1871          * this optional reply-ack won't be sent as the
1872          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
1873          */
1874         if (msg.flags & VHOST_USER_NEED_REPLY) {
1875                 msg.payload.u64 = ret == VH_RESULT_ERR;
1876                 msg.size = sizeof(msg.payload.u64);
1877                 msg.fd_num = 0;
1878                 send_vhost_reply(fd, &msg);
1879         } else if (ret == VH_RESULT_ERR) {
1880                 RTE_LOG(ERR, VHOST_CONFIG,
1881                         "vhost message handling failed.\n");
1882                 return -1;
1883         }
1884
1885         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
1886                 dev->flags |= VIRTIO_DEV_READY;
1887
1888                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
1889                         if (dev->dequeue_zero_copy) {
1890                                 RTE_LOG(INFO, VHOST_CONFIG,
1891                                                 "dequeue zero copy is enabled\n");
1892                         }
1893
1894                         if (dev->notify_ops->new_device(dev->vid) == 0)
1895                                 dev->flags |= VIRTIO_DEV_RUNNING;
1896                 }
1897         }
1898
1899         did = dev->vdpa_dev_id;
1900         vdpa_dev = rte_vdpa_get_device(did);
1901         if (vdpa_dev && virtio_is_ready(dev) &&
1902                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
1903                         msg.request.master == VHOST_USER_SET_VRING_ENABLE) {
1904                 if (vdpa_dev->ops->dev_conf)
1905                         vdpa_dev->ops->dev_conf(dev->vid);
1906                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
1907                 if (vhost_user_host_notifier_ctrl(dev->vid, true) != 0) {
1908                         RTE_LOG(INFO, VHOST_CONFIG,
1909                                 "(%d) software relay is used for vDPA, performance may be low.\n",
1910                                 dev->vid);
1911                 }
1912         }
1913
1914         return 0;
1915 }
1916
1917 static int process_slave_message_reply(struct virtio_net *dev,
1918                                        const struct VhostUserMsg *msg)
1919 {
1920         struct VhostUserMsg msg_reply;
1921         int ret;
1922
1923         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
1924                 return 0;
1925
1926         if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
1927                 ret = -1;
1928                 goto out;
1929         }
1930
1931         if (msg_reply.request.slave != msg->request.slave) {
1932                 RTE_LOG(ERR, VHOST_CONFIG,
1933                         "Received unexpected msg type (%u), expected %u\n",
1934                         msg_reply.request.slave, msg->request.slave);
1935                 ret = -1;
1936                 goto out;
1937         }
1938
1939         ret = msg_reply.payload.u64 ? -1 : 0;
1940
1941 out:
1942         rte_spinlock_unlock(&dev->slave_req_lock);
1943         return ret;
1944 }
1945
1946 int
1947 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
1948 {
1949         int ret;
1950         struct VhostUserMsg msg = {
1951                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
1952                 .flags = VHOST_USER_VERSION,
1953                 .size = sizeof(msg.payload.iotlb),
1954                 .payload.iotlb = {
1955                         .iova = iova,
1956                         .perm = perm,
1957                         .type = VHOST_IOTLB_MISS,
1958                 },
1959         };
1960
1961         ret = send_vhost_message(dev->slave_req_fd, &msg);
1962         if (ret < 0) {
1963                 RTE_LOG(ERR, VHOST_CONFIG,
1964                                 "Failed to send IOTLB miss message (%d)\n",
1965                                 ret);
1966                 return ret;
1967         }
1968
1969         return 0;
1970 }
1971
1972 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
1973                                                     int index, int fd,
1974                                                     uint64_t offset,
1975                                                     uint64_t size)
1976 {
1977         int ret;
1978         struct VhostUserMsg msg = {
1979                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
1980                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
1981                 .size = sizeof(msg.payload.area),
1982                 .payload.area = {
1983                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
1984                         .size = size,
1985                         .offset = offset,
1986                 },
1987         };
1988
1989         if (fd < 0)
1990                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
1991         else {
1992                 msg.fds[0] = fd;
1993                 msg.fd_num = 1;
1994         }
1995
1996         ret = send_vhost_slave_message(dev, &msg);
1997         if (ret < 0) {
1998                 RTE_LOG(ERR, VHOST_CONFIG,
1999                         "Failed to set host notifier (%d)\n", ret);
2000                 return ret;
2001         }
2002
2003         return process_slave_message_reply(dev, &msg);
2004 }
2005
2006 int vhost_user_host_notifier_ctrl(int vid, bool enable)
2007 {
2008         struct virtio_net *dev;
2009         struct rte_vdpa_device *vdpa_dev;
2010         int vfio_device_fd, did, ret = 0;
2011         uint64_t offset, size;
2012         unsigned int i;
2013
2014         dev = get_device(vid);
2015         if (!dev)
2016                 return -ENODEV;
2017
2018         did = dev->vdpa_dev_id;
2019         if (did < 0)
2020                 return -EINVAL;
2021
2022         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2023             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2024             !(dev->protocol_features &
2025                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2026             !(dev->protocol_features &
2027                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2028             !(dev->protocol_features &
2029                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2030                 return -ENOTSUP;
2031
2032         vdpa_dev = rte_vdpa_get_device(did);
2033         if (!vdpa_dev)
2034                 return -ENODEV;
2035
2036         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2037         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2038
2039         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2040         if (vfio_device_fd < 0)
2041                 return -ENOTSUP;
2042
2043         if (enable) {
2044                 for (i = 0; i < dev->nr_vring; i++) {
2045                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2046                                         &size) < 0) {
2047                                 ret = -ENOTSUP;
2048                                 goto disable;
2049                         }
2050
2051                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2052                                         vfio_device_fd, offset, size) < 0) {
2053                                 ret = -EFAULT;
2054                                 goto disable;
2055                         }
2056                 }
2057         } else {
2058 disable:
2059                 for (i = 0; i < dev->nr_vring; i++) {
2060                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
2061                                         0, 0);
2062                 }
2063         }
2064
2065         return ret;
2066 }