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         [VHOST_USER_POSTCOPY_LISTEN]  = "VHOST_USER_POSTCOPY_LISTEN",
  80         [VHOST_USER_POSTCOPY_END]  = "VHOST_USER_POSTCOPY_END",
  81 };
  82
  83 static int send_vhost_reply(int sockfd, struct VhostUserMsg *msg);
  84 static int read_vhost_message(int sockfd, struct VhostUserMsg *msg);
  85
  86 static uint64_t
  87 get_blk_size(int fd)
  88 {
  89         struct stat stat;
  90         int ret;
  91
  92         ret = fstat(fd, &stat);
  93         return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
  94 }
  95
  96 /*
  97  * Reclaim all the outstanding zmbufs for a virtqueue.
  98  */
  99 static void
 100 drain_zmbuf_list(struct vhost_virtqueue *vq)
 101 {
 102         struct zcopy_mbuf *zmbuf, *next;
 103
 104         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
 105              zmbuf != NULL; zmbuf = next) {
 106                 next = TAILQ_NEXT(zmbuf, next);
 107
 108                 while (!mbuf_is_consumed(zmbuf->mbuf))
 109                         usleep(1000);
 110
 111                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
 112                 restore_mbuf(zmbuf->mbuf);
 113                 rte_pktmbuf_free(zmbuf->mbuf);
 114                 put_zmbuf(zmbuf);
 115                 vq->nr_zmbuf -= 1;
 116         }
 117 }
 118
 119 static void
 120 free_mem_region(struct virtio_net *dev)
 121 {
 122         uint32_t i;
 123         struct rte_vhost_mem_region *reg;
 124         struct vhost_virtqueue *vq;
 125
 126         if (!dev || !dev->mem)
 127                 return;
 128
 129         if (dev->dequeue_zero_copy) {
 130                 for (i = 0; i < dev->nr_vring; i++) {
 131                         vq = dev->virtqueue[i];
 132                         if (vq)
 133                                 drain_zmbuf_list(vq);
 134                 }
 135         }
 136
 137         for (i = 0; i < dev->mem->nregions; i++) {
 138                 reg = &dev->mem->regions[i];
 139                 if (reg->host_user_addr) {
 140                         munmap(reg->mmap_addr, reg->mmap_size);
 141                         close(reg->fd);
 142                 }
 143         }
 144 }
 145
 146 void
 147 vhost_backend_cleanup(struct virtio_net *dev)
 148 {
 149         if (dev->mem) {
 150                 free_mem_region(dev);
 151                 rte_free(dev->mem);
 152                 dev->mem = NULL;
 153         }
 154
 155         free(dev->guest_pages);
 156         dev->guest_pages = NULL;
 157
 158         if (dev->log_addr) {
 159                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 160                 dev->log_addr = 0;
 161         }
 162
 163         if (dev->slave_req_fd >= 0) {
 164                 close(dev->slave_req_fd);
 165                 dev->slave_req_fd = -1;
 166         }
 167
 168         if (dev->postcopy_ufd >= 0) {
 169                 close(dev->postcopy_ufd);
 170                 dev->postcopy_ufd = -1;
 171         }
 172
 173         dev->postcopy_listening = 0;
 174 }
 175
 176 /*
 177  * This function just returns success at the moment unless
 178  * the device hasn't been initialised.
 179  */
 180 static int
 181 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
 182                         struct VhostUserMsg *msg __rte_unused,
 183                         int main_fd __rte_unused)
 184 {
 185         return RTE_VHOST_MSG_RESULT_OK;
 186 }
 187
 188 static int
 189 vhost_user_reset_owner(struct virtio_net **pdev,
 190                         struct VhostUserMsg *msg __rte_unused,
 191                         int main_fd __rte_unused)
 192 {
 193         struct virtio_net *dev = *pdev;
 194         vhost_destroy_device_notify(dev);
 195
 196         cleanup_device(dev, 0);
 197         reset_device(dev);
 198         return RTE_VHOST_MSG_RESULT_OK;
 199 }
 200
 201 /*
 202  * The features that we support are requested.
 203  */
 204 static int
 205 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 206                         int main_fd __rte_unused)
 207 {
 208         struct virtio_net *dev = *pdev;
 209         uint64_t features = 0;
 210
 211         rte_vhost_driver_get_features(dev->ifname, &features);
 212
 213         msg->payload.u64 = features;
 214         msg->size = sizeof(msg->payload.u64);
 215         msg->fd_num = 0;
 216
 217         return RTE_VHOST_MSG_RESULT_REPLY;
 218 }
 219
 220 /*
 221  * The queue number that we support are requested.
 222  */
 223 static int
 224 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
 225                         int main_fd __rte_unused)
 226 {
 227         struct virtio_net *dev = *pdev;
 228         uint32_t queue_num = 0;
 229
 230         rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
 231
 232         msg->payload.u64 = (uint64_t)queue_num;
 233         msg->size = sizeof(msg->payload.u64);
 234         msg->fd_num = 0;
 235
 236         return RTE_VHOST_MSG_RESULT_REPLY;
 237 }
 238
 239 /*
 240  * We receive the negotiated features supported by us and the virtio device.
 241  */
 242 static int
 243 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 244                         int main_fd __rte_unused)
 245 {
 246         struct virtio_net *dev = *pdev;
 247         uint64_t features = msg->payload.u64;
 248         uint64_t vhost_features = 0;
 249         struct rte_vdpa_device *vdpa_dev;
 250         int did = -1;
 251
 252         rte_vhost_driver_get_features(dev->ifname, &vhost_features);
 253         if (features & ~vhost_features) {
 254                 RTE_LOG(ERR, VHOST_CONFIG,
 255                         "(%d) received invalid negotiated features.\n",
 256                         dev->vid);
 257                 return RTE_VHOST_MSG_RESULT_ERR;
 258         }
 259
 260         if (dev->flags & VIRTIO_DEV_RUNNING) {
 261                 if (dev->features == features)
 262                         return RTE_VHOST_MSG_RESULT_OK;
 263
 264                 /*
 265                  * Error out if master tries to change features while device is
 266                  * in running state. The exception being VHOST_F_LOG_ALL, which
 267                  * is enabled when the live-migration starts.
 268                  */
 269                 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
 270                         RTE_LOG(ERR, VHOST_CONFIG,
 271                                 "(%d) features changed while device is running.\n",
 272                                 dev->vid);
 273                         return RTE_VHOST_MSG_RESULT_ERR;
 274                 }
 275
 276                 if (dev->notify_ops->features_changed)
 277                         dev->notify_ops->features_changed(dev->vid, features);
 278         }
 279
 280         dev->features = features;
 281         if (dev->features &
 282                 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
 283                 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 284         } else {
 285                 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
 286         }
 287         RTE_LOG(INFO, VHOST_CONFIG,
 288                 "negotiated Virtio features: 0x%" PRIx64 "\n", dev->features);
 289         VHOST_LOG_DEBUG(VHOST_CONFIG,
 290                 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
 291                 dev->vid,
 292                 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
 293                 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 294
 295         if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
 296             !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
 297                 /*
 298                  * Remove all but first queue pair if MQ hasn't been
 299                  * negotiated. This is safe because the device is not
 300                  * running at this stage.
 301                  */
 302                 while (dev->nr_vring > 2) {
 303                         struct vhost_virtqueue *vq;
 304
 305                         vq = dev->virtqueue[--dev->nr_vring];
 306                         if (!vq)
 307                                 continue;
 308
 309                         dev->virtqueue[dev->nr_vring] = NULL;
 310                         cleanup_vq(vq, 1);
 311                         free_vq(dev, vq);
 312                 }
 313         }
 314
 315         did = dev->vdpa_dev_id;
 316         vdpa_dev = rte_vdpa_get_device(did);
 317         if (vdpa_dev && vdpa_dev->ops->set_features)
 318                 vdpa_dev->ops->set_features(dev->vid);
 319
 320         return RTE_VHOST_MSG_RESULT_OK;
 321 }
 322
 323 /*
 324  * The virtio device sends us the size of the descriptor ring.
 325  */
 326 static int
 327 vhost_user_set_vring_num(struct virtio_net **pdev,
 328                         struct VhostUserMsg *msg,
 329                         int main_fd __rte_unused)
 330 {
 331         struct virtio_net *dev = *pdev;
 332         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 333
 334         vq->size = msg->payload.state.num;
 335
 336         /* VIRTIO 1.0, 2.4 Virtqueues says:
 337          *
 338          *   Queue Size value is always a power of 2. The maximum Queue Size
 339          *   value is 32768.
 340          */
 341         if ((vq->size & (vq->size - 1)) || vq->size > 32768) {
 342                 RTE_LOG(ERR, VHOST_CONFIG,
 343                         "invalid virtqueue size %u\n", vq->size);
 344                 return RTE_VHOST_MSG_RESULT_ERR;
 345         }
 346
 347         if (dev->dequeue_zero_copy) {
 348                 vq->nr_zmbuf = 0;
 349                 vq->last_zmbuf_idx = 0;
 350                 vq->zmbuf_size = vq->size;
 351                 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
 352                                          sizeof(struct zcopy_mbuf), 0);
 353                 if (vq->zmbufs == NULL) {
 354                         RTE_LOG(WARNING, VHOST_CONFIG,
 355                                 "failed to allocate mem for zero copy; "
 356                                 "zero copy is force disabled\n");
 357                         dev->dequeue_zero_copy = 0;
 358                 }
 359                 TAILQ_INIT(&vq->zmbuf_list);
 360         }
 361
 362         if (vq_is_packed(dev)) {
 363                 vq->shadow_used_packed = rte_malloc(NULL,
 364                                 vq->size *
 365                                 sizeof(struct vring_used_elem_packed),
 366                                 RTE_CACHE_LINE_SIZE);
 367                 if (!vq->shadow_used_packed) {
 368                         RTE_LOG(ERR, VHOST_CONFIG,
 369                                         "failed to allocate memory for shadow used ring.\n");
 370                         return RTE_VHOST_MSG_RESULT_ERR;
 371                 }
 372
 373         } else {
 374                 vq->shadow_used_split = rte_malloc(NULL,
 375                                 vq->size * sizeof(struct vring_used_elem),
 376                                 RTE_CACHE_LINE_SIZE);
 377                 if (!vq->shadow_used_split) {
 378                         RTE_LOG(ERR, VHOST_CONFIG,
 379                                         "failed to allocate memory for shadow used ring.\n");
 380                         return RTE_VHOST_MSG_RESULT_ERR;
 381                 }
 382         }
 383
 384         vq->batch_copy_elems = rte_malloc(NULL,
 385                                 vq->size * sizeof(struct batch_copy_elem),
 386                                 RTE_CACHE_LINE_SIZE);
 387         if (!vq->batch_copy_elems) {
 388                 RTE_LOG(ERR, VHOST_CONFIG,
 389                         "failed to allocate memory for batching copy.\n");
 390                 return RTE_VHOST_MSG_RESULT_ERR;
 391         }
 392
 393         return RTE_VHOST_MSG_RESULT_OK;
 394 }
 395
 396 /*
 397  * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
 398  * same numa node as the memory of vring descriptor.
 399  */
 400 #ifdef RTE_LIBRTE_VHOST_NUMA
 401 static struct virtio_net*
 402 numa_realloc(struct virtio_net *dev, int index)
 403 {
 404         int oldnode, newnode;
 405         struct virtio_net *old_dev;
 406         struct vhost_virtqueue *old_vq, *vq;
 407         struct zcopy_mbuf *new_zmbuf;
 408         struct vring_used_elem *new_shadow_used_split;
 409         struct vring_used_elem_packed *new_shadow_used_packed;
 410         struct batch_copy_elem *new_batch_copy_elems;
 411         int ret;
 412
 413         if (dev->flags & VIRTIO_DEV_RUNNING)
 414                 return dev;
 415
 416         old_dev = dev;
 417         vq = old_vq = dev->virtqueue[index];
 418
 419         ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
 420                             MPOL_F_NODE | MPOL_F_ADDR);
 421
 422         /* check if we need to reallocate vq */
 423         ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
 424                              MPOL_F_NODE | MPOL_F_ADDR);
 425         if (ret) {
 426                 RTE_LOG(ERR, VHOST_CONFIG,
 427                         "Unable to get vq numa information.\n");
 428                 return dev;
 429         }
 430         if (oldnode != newnode) {
 431                 RTE_LOG(INFO, VHOST_CONFIG,
 432                         "reallocate vq from %d to %d node\n", oldnode, newnode);
 433                 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
 434                 if (!vq)
 435                         return dev;
 436
 437                 memcpy(vq, old_vq, sizeof(*vq));
 438                 TAILQ_INIT(&vq->zmbuf_list);
 439
 440                 if (dev->dequeue_zero_copy) {
 441                         new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
 442                                         sizeof(struct zcopy_mbuf), 0, newnode);
 443                         if (new_zmbuf) {
 444                                 rte_free(vq->zmbufs);
 445                                 vq->zmbufs = new_zmbuf;
 446                         }
 447                 }
 448
 449                 if (vq_is_packed(dev)) {
 450                         new_shadow_used_packed = rte_malloc_socket(NULL,
 451                                         vq->size *
 452                                         sizeof(struct vring_used_elem_packed),
 453                                         RTE_CACHE_LINE_SIZE,
 454                                         newnode);
 455                         if (new_shadow_used_packed) {
 456                                 rte_free(vq->shadow_used_packed);
 457                                 vq->shadow_used_packed = new_shadow_used_packed;
 458                         }
 459                 } else {
 460                         new_shadow_used_split = rte_malloc_socket(NULL,
 461                                         vq->size *
 462                                         sizeof(struct vring_used_elem),
 463                                         RTE_CACHE_LINE_SIZE,
 464                                         newnode);
 465                         if (new_shadow_used_split) {
 466                                 rte_free(vq->shadow_used_split);
 467                                 vq->shadow_used_split = new_shadow_used_split;
 468                         }
 469                 }
 470
 471                 new_batch_copy_elems = rte_malloc_socket(NULL,
 472                         vq->size * sizeof(struct batch_copy_elem),
 473                         RTE_CACHE_LINE_SIZE,
 474                         newnode);
 475                 if (new_batch_copy_elems) {
 476                         rte_free(vq->batch_copy_elems);
 477                         vq->batch_copy_elems = new_batch_copy_elems;
 478                 }
 479
 480                 rte_free(old_vq);
 481         }
 482
 483         /* check if we need to reallocate dev */
 484         ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
 485                             MPOL_F_NODE | MPOL_F_ADDR);
 486         if (ret) {
 487                 RTE_LOG(ERR, VHOST_CONFIG,
 488                         "Unable to get dev numa information.\n");
 489                 goto out;
 490         }
 491         if (oldnode != newnode) {
 492                 RTE_LOG(INFO, VHOST_CONFIG,
 493                         "reallocate dev from %d to %d node\n",
 494                         oldnode, newnode);
 495                 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
 496                 if (!dev) {
 497                         dev = old_dev;
 498                         goto out;
 499                 }
 500
 501                 memcpy(dev, old_dev, sizeof(*dev));
 502                 rte_free(old_dev);
 503         }
 504
 505 out:
 506         dev->virtqueue[index] = vq;
 507         vhost_devices[dev->vid] = dev;
 508
 509         if (old_vq != vq)
 510                 vhost_user_iotlb_init(dev, index);
 511
 512         return dev;
 513 }
 514 #else
 515 static struct virtio_net*
 516 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 517 {
 518         return dev;
 519 }
 520 #endif
 521
 522 /* Converts QEMU virtual address to Vhost virtual address. */
 523 static uint64_t
 524 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
 525 {
 526         struct rte_vhost_mem_region *r;
 527         uint32_t i;
 528
 529         if (unlikely(!dev || !dev->mem))
 530                 goto out_error;
 531
 532         /* Find the region where the address lives. */
 533         for (i = 0; i < dev->mem->nregions; i++) {
 534                 r = &dev->mem->regions[i];
 535
 536                 if (qva >= r->guest_user_addr &&
 537                     qva <  r->guest_user_addr + r->size) {
 538
 539                         if (unlikely(*len > r->guest_user_addr + r->size - qva))
 540                                 *len = r->guest_user_addr + r->size - qva;
 541
 542                         return qva - r->guest_user_addr +
 543                                r->host_user_addr;
 544                 }
 545         }
 546 out_error:
 547         *len = 0;
 548
 549         return 0;
 550 }
 551
 552
 553 /*
 554  * Converts ring address to Vhost virtual address.
 555  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 556  * else it is a QEMU virtual address.
 557  */
 558 static uint64_t
 559 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 560                 uint64_t ra, uint64_t *size)
 561 {
 562         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 563                 uint64_t vva;
 564                 uint64_t req_size = *size;
 565
 566                 vva = vhost_user_iotlb_cache_find(vq, ra,
 567                                         size, VHOST_ACCESS_RW);
 568                 if (req_size != *size)
 569                         vhost_user_iotlb_miss(dev, (ra + *size),
 570                                               VHOST_ACCESS_RW);
 571
 572                 return vva;
 573         }
 574
 575         return qva_to_vva(dev, ra, size);
 576 }
 577
 578 static struct virtio_net *
 579 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 580 {
 581         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 582         struct vhost_vring_addr *addr = &vq->ring_addrs;
 583         uint64_t len, expected_len;
 584
 585         if (vq_is_packed(dev)) {
 586                 len = sizeof(struct vring_packed_desc) * vq->size;
 587                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 588                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 589                 vq->log_guest_addr = 0;
 590                 if (vq->desc_packed == NULL ||
 591                                 len != sizeof(struct vring_packed_desc) *
 592                                 vq->size) {
 593                         RTE_LOG(DEBUG, VHOST_CONFIG,
 594                                 "(%d) failed to map desc_packed ring.\n",
 595                                 dev->vid);
 596                         return dev;
 597                 }
 598
 599                 dev = numa_realloc(dev, vq_index);
 600                 vq = dev->virtqueue[vq_index];
 601                 addr = &vq->ring_addrs;
 602
 603                 len = sizeof(struct vring_packed_desc_event);
 604                 vq->driver_event = (struct vring_packed_desc_event *)
 605                                         (uintptr_t)ring_addr_to_vva(dev,
 606                                         vq, addr->avail_user_addr, &len);
 607                 if (vq->driver_event == NULL ||
 608                                 len != sizeof(struct vring_packed_desc_event)) {
 609                         RTE_LOG(DEBUG, VHOST_CONFIG,
 610                                 "(%d) failed to find driver area address.\n",
 611                                 dev->vid);
 612                         return dev;
 613                 }
 614
 615                 len = sizeof(struct vring_packed_desc_event);
 616                 vq->device_event = (struct vring_packed_desc_event *)
 617                                         (uintptr_t)ring_addr_to_vva(dev,
 618                                         vq, addr->used_user_addr, &len);
 619                 if (vq->device_event == NULL ||
 620                                 len != sizeof(struct vring_packed_desc_event)) {
 621                         RTE_LOG(DEBUG, VHOST_CONFIG,
 622                                 "(%d) failed to find device area address.\n",
 623                                 dev->vid);
 624                         return dev;
 625                 }
 626
 627                 vq->access_ok = 1;
 628                 return dev;
 629         }
 630
 631         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 632         if (vq->desc && vq->avail && vq->used)
 633                 return dev;
 634
 635         len = sizeof(struct vring_desc) * vq->size;
 636         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 637                         vq, addr->desc_user_addr, &len);
 638         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 639                 RTE_LOG(DEBUG, VHOST_CONFIG,
 640                         "(%d) failed to map desc ring.\n",
 641                         dev->vid);
 642                 return dev;
 643         }
 644
 645         dev = numa_realloc(dev, vq_index);
 646         vq = dev->virtqueue[vq_index];
 647         addr = &vq->ring_addrs;
 648
 649         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 650         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 651                 len += sizeof(uint16_t);
 652         expected_len = len;
 653         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 654                         vq, addr->avail_user_addr, &len);
 655         if (vq->avail == 0 || len != expected_len) {
 656                 RTE_LOG(DEBUG, VHOST_CONFIG,
 657                         "(%d) failed to map avail ring.\n",
 658                         dev->vid);
 659                 return dev;
 660         }
 661
 662         len = sizeof(struct vring_used) +
 663                 sizeof(struct vring_used_elem) * vq->size;
 664         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 665                 len += sizeof(uint16_t);
 666         expected_len = len;
 667         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 668                         vq, addr->used_user_addr, &len);
 669         if (vq->used == 0 || len != expected_len) {
 670                 RTE_LOG(DEBUG, VHOST_CONFIG,
 671                         "(%d) failed to map used ring.\n",
 672                         dev->vid);
 673                 return dev;
 674         }
 675
 676         if (vq->last_used_idx != vq->used->idx) {
 677                 RTE_LOG(WARNING, VHOST_CONFIG,
 678                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 679                         "some packets maybe resent for Tx and dropped for Rx\n",
 680                         vq->last_used_idx, vq->used->idx);
 681                 vq->last_used_idx  = vq->used->idx;
 682                 vq->last_avail_idx = vq->used->idx;
 683         }
 684
 685         vq->log_guest_addr = addr->log_guest_addr;
 686         vq->access_ok = 1;
 687
 688         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
 689                         dev->vid, vq->desc);
 690         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
 691                         dev->vid, vq->avail);
 692         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
 693                         dev->vid, vq->used);
 694         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
 695                         dev->vid, vq->log_guest_addr);
 696
 697         return dev;
 698 }
 699
 700 /*
 701  * The virtio device sends us the desc, used and avail ring addresses.
 702  * This function then converts these to our address space.
 703  */
 704 static int
 705 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 706                         int main_fd __rte_unused)
 707 {
 708         struct virtio_net *dev = *pdev;
 709         struct vhost_virtqueue *vq;
 710         struct vhost_vring_addr *addr = &msg->payload.addr;
 711         bool access_ok;
 712
 713         if (dev->mem == NULL)
 714                 return RTE_VHOST_MSG_RESULT_ERR;
 715
 716         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 717         vq = dev->virtqueue[msg->payload.addr.index];
 718
 719         access_ok = vq->access_ok;
 720
 721         /*
 722          * Rings addresses should not be interpreted as long as the ring is not
 723          * started and enabled
 724          */
 725         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 726
 727         vring_invalidate(dev, vq);
 728
 729         if ((vq->enabled && (dev->features &
 730                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) ||
 731                         access_ok) {
 732                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 733                 if (!dev)
 734                         return RTE_VHOST_MSG_RESULT_ERR;
 735
 736                 *pdev = dev;
 737         }
 738
 739         return RTE_VHOST_MSG_RESULT_OK;
 740 }
 741
 742 /*
 743  * The virtio device sends us the available ring last used index.
 744  */
 745 static int
 746 vhost_user_set_vring_base(struct virtio_net **pdev,
 747                         struct VhostUserMsg *msg,
 748                         int main_fd __rte_unused)
 749 {
 750         struct virtio_net *dev = *pdev;
 751         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 752         uint64_t val = msg->payload.state.num;
 753
 754         if (vq_is_packed(dev)) {
 755                 /*
 756                  * Bit[0:14]: avail index
 757                  * Bit[15]: avail wrap counter
 758                  */
 759                 vq->last_avail_idx = val & 0x7fff;
 760                 vq->avail_wrap_counter = !!(val & (0x1 << 15));
 761                 /*
 762                  * Set used index to same value as available one, as
 763                  * their values should be the same since ring processing
 764                  * was stopped at get time.
 765                  */
 766                 vq->last_used_idx = vq->last_avail_idx;
 767                 vq->used_wrap_counter = vq->avail_wrap_counter;
 768         } else {
 769                 vq->last_used_idx = msg->payload.state.num;
 770                 vq->last_avail_idx = msg->payload.state.num;
 771         }
 772
 773         return RTE_VHOST_MSG_RESULT_OK;
 774 }
 775
 776 static int
 777 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 778                    uint64_t host_phys_addr, uint64_t size)
 779 {
 780         struct guest_page *page, *last_page;
 781         struct guest_page *old_pages;
 782
 783         if (dev->nr_guest_pages == dev->max_guest_pages) {
 784                 dev->max_guest_pages *= 2;
 785                 old_pages = dev->guest_pages;
 786                 dev->guest_pages = realloc(dev->guest_pages,
 787                                         dev->max_guest_pages * sizeof(*page));
 788                 if (!dev->guest_pages) {
 789                         RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
 790                         free(old_pages);
 791                         return -1;
 792                 }
 793         }
 794
 795         if (dev->nr_guest_pages > 0) {
 796                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 797                 /* merge if the two pages are continuous */
 798                 if (host_phys_addr == last_page->host_phys_addr +
 799                                       last_page->size) {
 800                         last_page->size += size;
 801                         return 0;
 802                 }
 803         }
 804
 805         page = &dev->guest_pages[dev->nr_guest_pages++];
 806         page->guest_phys_addr = guest_phys_addr;
 807         page->host_phys_addr  = host_phys_addr;
 808         page->size = size;
 809
 810         return 0;
 811 }
 812
 813 static int
 814 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 815                 uint64_t page_size)
 816 {
 817         uint64_t reg_size = reg->size;
 818         uint64_t host_user_addr  = reg->host_user_addr;
 819         uint64_t guest_phys_addr = reg->guest_phys_addr;
 820         uint64_t host_phys_addr;
 821         uint64_t size;
 822
 823         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 824         size = page_size - (guest_phys_addr & (page_size - 1));
 825         size = RTE_MIN(size, reg_size);
 826
 827         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 828                 return -1;
 829
 830         host_user_addr  += size;
 831         guest_phys_addr += size;
 832         reg_size -= size;
 833
 834         while (reg_size > 0) {
 835                 size = RTE_MIN(reg_size, page_size);
 836                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 837                                                   host_user_addr);
 838                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 839                                 size) < 0)
 840                         return -1;
 841
 842                 host_user_addr  += size;
 843                 guest_phys_addr += size;
 844                 reg_size -= size;
 845         }
 846
 847         return 0;
 848 }
 849
 850 #ifdef RTE_LIBRTE_VHOST_DEBUG
 851 /* TODO: enable it only in debug mode? */
 852 static void
 853 dump_guest_pages(struct virtio_net *dev)
 854 {
 855         uint32_t i;
 856         struct guest_page *page;
 857
 858         for (i = 0; i < dev->nr_guest_pages; i++) {
 859                 page = &dev->guest_pages[i];
 860
 861                 RTE_LOG(INFO, VHOST_CONFIG,
 862                         "guest physical page region %u\n"
 863                         "\t guest_phys_addr: %" PRIx64 "\n"
 864                         "\t host_phys_addr : %" PRIx64 "\n"
 865                         "\t size           : %" PRIx64 "\n",
 866                         i,
 867                         page->guest_phys_addr,
 868                         page->host_phys_addr,
 869                         page->size);
 870         }
 871 }
 872 #else
 873 #define dump_guest_pages(dev)
 874 #endif
 875
 876 static bool
 877 vhost_memory_changed(struct VhostUserMemory *new,
 878                      struct rte_vhost_memory *old)
 879 {
 880         uint32_t i;
 881
 882         if (new->nregions != old->nregions)
 883                 return true;
 884
 885         for (i = 0; i < new->nregions; ++i) {
 886                 VhostUserMemoryRegion *new_r = &new->regions[i];
 887                 struct rte_vhost_mem_region *old_r = &old->regions[i];
 888
 889                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
 890                         return true;
 891                 if (new_r->memory_size != old_r->size)
 892                         return true;
 893                 if (new_r->userspace_addr != old_r->guest_user_addr)
 894                         return true;
 895         }
 896
 897         return false;
 898 }
 899
 900 static int
 901 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
 902                         int main_fd)
 903 {
 904         struct virtio_net *dev = *pdev;
 905         struct VhostUserMemory *memory = &msg->payload.memory;
 906         struct rte_vhost_mem_region *reg;
 907         void *mmap_addr;
 908         uint64_t mmap_size;
 909         uint64_t mmap_offset;
 910         uint64_t alignment;
 911         uint32_t i;
 912         int populate;
 913         int fd;
 914
 915         if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
 916                 RTE_LOG(ERR, VHOST_CONFIG,
 917                         "too many memory regions (%u)\n", memory->nregions);
 918                 return RTE_VHOST_MSG_RESULT_ERR;
 919         }
 920
 921         if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
 922                 RTE_LOG(INFO, VHOST_CONFIG,
 923                         "(%d) memory regions not changed\n", dev->vid);
 924
 925                 for (i = 0; i < memory->nregions; i++)
 926                         close(msg->fds[i]);
 927
 928                 return RTE_VHOST_MSG_RESULT_OK;
 929         }
 930
 931         if (dev->mem) {
 932                 free_mem_region(dev);
 933                 rte_free(dev->mem);
 934                 dev->mem = NULL;
 935         }
 936
 937         /* Flush IOTLB cache as previous HVAs are now invalid */
 938         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 939                 for (i = 0; i < dev->nr_vring; i++)
 940                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
 941
 942         dev->nr_guest_pages = 0;
 943         if (!dev->guest_pages) {
 944                 dev->max_guest_pages = 8;
 945                 dev->guest_pages = malloc(dev->max_guest_pages *
 946                                                 sizeof(struct guest_page));
 947                 if (dev->guest_pages == NULL) {
 948                         RTE_LOG(ERR, VHOST_CONFIG,
 949                                 "(%d) failed to allocate memory "
 950                                 "for dev->guest_pages\n",
 951                                 dev->vid);
 952                         return RTE_VHOST_MSG_RESULT_ERR;
 953                 }
 954         }
 955
 956         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
 957                 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
 958         if (dev->mem == NULL) {
 959                 RTE_LOG(ERR, VHOST_CONFIG,
 960                         "(%d) failed to allocate memory for dev->mem\n",
 961                         dev->vid);
 962                 return RTE_VHOST_MSG_RESULT_ERR;
 963         }
 964         dev->mem->nregions = memory->nregions;
 965
 966         for (i = 0; i < memory->nregions; i++) {
 967                 fd  = msg->fds[i];
 968                 reg = &dev->mem->regions[i];
 969
 970                 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
 971                 reg->guest_user_addr = memory->regions[i].userspace_addr;
 972                 reg->size            = memory->regions[i].memory_size;
 973                 reg->fd              = fd;
 974
 975                 mmap_offset = memory->regions[i].mmap_offset;
 976
 977                 /* Check for memory_size + mmap_offset overflow */
 978                 if (mmap_offset >= -reg->size) {
 979                         RTE_LOG(ERR, VHOST_CONFIG,
 980                                 "mmap_offset (%#"PRIx64") and memory_size "
 981                                 "(%#"PRIx64") overflow\n",
 982                                 mmap_offset, reg->size);
 983                         goto err_mmap;
 984                 }
 985
 986                 mmap_size = reg->size + mmap_offset;
 987
 988                 /* mmap() without flag of MAP_ANONYMOUS, should be called
 989                  * with length argument aligned with hugepagesz at older
 990                  * longterm version Linux, like 2.6.32 and 3.2.72, or
 991                  * mmap() will fail with EINVAL.
 992                  *
 993                  * to avoid failure, make sure in caller to keep length
 994                  * aligned.
 995                  */
 996                 alignment = get_blk_size(fd);
 997                 if (alignment == (uint64_t)-1) {
 998                         RTE_LOG(ERR, VHOST_CONFIG,
 999                                 "couldn't get hugepage size through fstat\n");
1000                         goto err_mmap;
1001                 }
1002                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
1003
1004                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
1005                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
1006                                  MAP_SHARED | populate, fd, 0);
1007
1008                 if (mmap_addr == MAP_FAILED) {
1009                         RTE_LOG(ERR, VHOST_CONFIG,
1010                                 "mmap region %u failed.\n", i);
1011                         goto err_mmap;
1012                 }
1013
1014                 reg->mmap_addr = mmap_addr;
1015                 reg->mmap_size = mmap_size;
1016                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
1017                                       mmap_offset;
1018
1019                 if (dev->dequeue_zero_copy)
1020                         if (add_guest_pages(dev, reg, alignment) < 0) {
1021                                 RTE_LOG(ERR, VHOST_CONFIG,
1022                                         "adding guest pages to region %u failed.\n",
1023                                         i);
1024                                 goto err_mmap;
1025                         }
1026
1027                 RTE_LOG(INFO, VHOST_CONFIG,
1028                         "guest memory region %u, size: 0x%" PRIx64 "\n"
1029                         "\t guest physical addr: 0x%" PRIx64 "\n"
1030                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
1031                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
1032                         "\t mmap addr : 0x%" PRIx64 "\n"
1033                         "\t mmap size : 0x%" PRIx64 "\n"
1034                         "\t mmap align: 0x%" PRIx64 "\n"
1035                         "\t mmap off  : 0x%" PRIx64 "\n",
1036                         i, reg->size,
1037                         reg->guest_phys_addr,
1038                         reg->guest_user_addr,
1039                         reg->host_user_addr,
1040                         (uint64_t)(uintptr_t)mmap_addr,
1041                         mmap_size,
1042                         alignment,
1043                         mmap_offset);
1044
1045                 if (dev->postcopy_listening) {
1046                         /*
1047                          * We haven't a better way right now than sharing
1048                          * DPDK's virtual address with Qemu, so that Qemu can
1049                          * retrieve the region offset when handling userfaults.
1050                          */
1051                         memory->regions[i].userspace_addr =
1052                                 reg->host_user_addr;
1053                 }
1054         }
1055         if (dev->postcopy_listening) {
1056                 /* Send the addresses back to qemu */
1057                 msg->fd_num = 0;
1058                 send_vhost_reply(main_fd, msg);
1059
1060                 /* Wait for qemu to acknolwedge it's got the addresses
1061                  * we've got to wait before we're allowed to generate faults.
1062                  */
1063                 VhostUserMsg ack_msg;
1064                 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1065                         RTE_LOG(ERR, VHOST_CONFIG,
1066                                 "Failed to read qemu ack on postcopy set-mem-table\n");
1067                         goto err_mmap;
1068                 }
1069                 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1070                         RTE_LOG(ERR, VHOST_CONFIG,
1071                                 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1072                                 ack_msg.request.master);
1073                         goto err_mmap;
1074                 }
1075
1076                 /* Now userfault register and we can use the memory */
1077                 for (i = 0; i < memory->nregions; i++) {
1078 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1079                         reg = &dev->mem->regions[i];
1080                         struct uffdio_register reg_struct;
1081
1082                         /*
1083                          * Let's register all the mmap'ed area to ensure
1084                          * alignment on page boundary.
1085                          */
1086                         reg_struct.range.start =
1087                                 (uint64_t)(uintptr_t)reg->mmap_addr;
1088                         reg_struct.range.len = reg->mmap_size;
1089                         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1090
1091                         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1092                                                 &reg_struct)) {
1093                                 RTE_LOG(ERR, VHOST_CONFIG,
1094                                         "Failed to register ufd for region %d: (ufd = %d) %s\n",
1095                                         i, dev->postcopy_ufd,
1096                                         strerror(errno));
1097                                 goto err_mmap;
1098                         }
1099                         RTE_LOG(INFO, VHOST_CONFIG,
1100                                 "\t userfaultfd registered for range : %llx - %llx\n",
1101                                 reg_struct.range.start,
1102                                 reg_struct.range.start +
1103                                 reg_struct.range.len - 1);
1104 #else
1105                         goto err_mmap;
1106 #endif
1107                 }
1108         }
1109
1110         for (i = 0; i < dev->nr_vring; i++) {
1111                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1112
1113                 if (vq->desc || vq->avail || vq->used) {
1114                         /*
1115                          * If the memory table got updated, the ring addresses
1116                          * need to be translated again as virtual addresses have
1117                          * changed.
1118                          */
1119                         vring_invalidate(dev, vq);
1120
1121                         dev = translate_ring_addresses(dev, i);
1122                         if (!dev) {
1123                                 dev = *pdev;
1124                                 goto err_mmap;
1125                         }
1126
1127                         *pdev = dev;
1128                 }
1129         }
1130
1131         dump_guest_pages(dev);
1132
1133         return RTE_VHOST_MSG_RESULT_OK;
1134
1135 err_mmap:
1136         free_mem_region(dev);
1137         rte_free(dev->mem);
1138         dev->mem = NULL;
1139         return RTE_VHOST_MSG_RESULT_ERR;
1140 }
1141
1142 static bool
1143 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1144 {
1145         bool rings_ok;
1146
1147         if (!vq)
1148                 return false;
1149
1150         if (vq_is_packed(dev))
1151                 rings_ok = !!vq->desc_packed;
1152         else
1153                 rings_ok = vq->desc && vq->avail && vq->used;
1154
1155         return rings_ok &&
1156                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1157                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1158 }
1159
1160 static int
1161 virtio_is_ready(struct virtio_net *dev)
1162 {
1163         struct vhost_virtqueue *vq;
1164         uint32_t i;
1165
1166         if (dev->nr_vring == 0)
1167                 return 0;
1168
1169         for (i = 0; i < dev->nr_vring; i++) {
1170                 vq = dev->virtqueue[i];
1171
1172                 if (!vq_is_ready(dev, vq))
1173                         return 0;
1174         }
1175
1176         RTE_LOG(INFO, VHOST_CONFIG,
1177                 "virtio is now ready for processing.\n");
1178         return 1;
1179 }
1180
1181 static int
1182 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1183                         int main_fd __rte_unused)
1184 {
1185         struct virtio_net *dev = *pdev;
1186         struct vhost_vring_file file;
1187         struct vhost_virtqueue *vq;
1188
1189         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1190         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1191                 file.fd = VIRTIO_INVALID_EVENTFD;
1192         else
1193                 file.fd = msg->fds[0];
1194         RTE_LOG(INFO, VHOST_CONFIG,
1195                 "vring call idx:%d file:%d\n", file.index, file.fd);
1196
1197         vq = dev->virtqueue[file.index];
1198         if (vq->callfd >= 0)
1199                 close(vq->callfd);
1200
1201         vq->callfd = file.fd;
1202
1203         return RTE_VHOST_MSG_RESULT_OK;
1204 }
1205
1206 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1207                         struct VhostUserMsg *msg,
1208                         int main_fd __rte_unused)
1209 {
1210         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1211                 close(msg->fds[0]);
1212         RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1213
1214         return RTE_VHOST_MSG_RESULT_OK;
1215 }
1216
1217 static int
1218 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1219                         int main_fd __rte_unused)
1220 {
1221         struct virtio_net *dev = *pdev;
1222         struct vhost_vring_file file;
1223         struct vhost_virtqueue *vq;
1224
1225         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1226         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1227                 file.fd = VIRTIO_INVALID_EVENTFD;
1228         else
1229                 file.fd = msg->fds[0];
1230         RTE_LOG(INFO, VHOST_CONFIG,
1231                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1232
1233         /* Interpret ring addresses only when ring is started. */
1234         dev = translate_ring_addresses(dev, file.index);
1235         if (!dev)
1236                 return RTE_VHOST_MSG_RESULT_ERR;
1237
1238         *pdev = dev;
1239
1240         vq = dev->virtqueue[file.index];
1241
1242         /*
1243          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1244          * the ring starts already enabled. Otherwise, it is enabled via
1245          * the SET_VRING_ENABLE message.
1246          */
1247         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
1248                 vq->enabled = 1;
1249                 if (dev->notify_ops->vring_state_changed)
1250                         dev->notify_ops->vring_state_changed(
1251                                 dev->vid, file.index, 1);
1252         }
1253
1254         if (vq->kickfd >= 0)
1255                 close(vq->kickfd);
1256         vq->kickfd = file.fd;
1257
1258         return RTE_VHOST_MSG_RESULT_OK;
1259 }
1260
1261 static void
1262 free_zmbufs(struct vhost_virtqueue *vq)
1263 {
1264         drain_zmbuf_list(vq);
1265
1266         rte_free(vq->zmbufs);
1267 }
1268
1269 /*
1270  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1271  */
1272 static int
1273 vhost_user_get_vring_base(struct virtio_net **pdev,
1274                         struct VhostUserMsg *msg,
1275                         int main_fd __rte_unused)
1276 {
1277         struct virtio_net *dev = *pdev;
1278         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1279         uint64_t val;
1280
1281         /* We have to stop the queue (virtio) if it is running. */
1282         vhost_destroy_device_notify(dev);
1283
1284         dev->flags &= ~VIRTIO_DEV_READY;
1285         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1286
1287         /* Here we are safe to get the indexes */
1288         if (vq_is_packed(dev)) {
1289                 /*
1290                  * Bit[0:14]: avail index
1291                  * Bit[15]: avail wrap counter
1292                  */
1293                 val = vq->last_avail_idx & 0x7fff;
1294                 val |= vq->avail_wrap_counter << 15;
1295                 msg->payload.state.num = val;
1296         } else {
1297                 msg->payload.state.num = vq->last_avail_idx;
1298         }
1299
1300         RTE_LOG(INFO, VHOST_CONFIG,
1301                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1302                 msg->payload.state.num);
1303         /*
1304          * Based on current qemu vhost-user implementation, this message is
1305          * sent and only sent in vhost_vring_stop.
1306          * TODO: cleanup the vring, it isn't usable since here.
1307          */
1308         if (vq->kickfd >= 0)
1309                 close(vq->kickfd);
1310
1311         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1312
1313         if (vq->callfd >= 0)
1314                 close(vq->callfd);
1315
1316         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1317
1318         vq->signalled_used_valid = false;
1319
1320         if (dev->dequeue_zero_copy)
1321                 free_zmbufs(vq);
1322         if (vq_is_packed(dev)) {
1323                 rte_free(vq->shadow_used_packed);
1324                 vq->shadow_used_packed = NULL;
1325         } else {
1326                 rte_free(vq->shadow_used_split);
1327                 vq->shadow_used_split = NULL;
1328         }
1329
1330         rte_free(vq->batch_copy_elems);
1331         vq->batch_copy_elems = NULL;
1332
1333         msg->size = sizeof(msg->payload.state);
1334         msg->fd_num = 0;
1335
1336         vring_invalidate(dev, vq);
1337
1338         return RTE_VHOST_MSG_RESULT_REPLY;
1339 }
1340
1341 /*
1342  * when virtio queues are ready to work, qemu will send us to
1343  * enable the virtio queue pair.
1344  */
1345 static int
1346 vhost_user_set_vring_enable(struct virtio_net **pdev,
1347                         struct VhostUserMsg *msg,
1348                         int main_fd __rte_unused)
1349 {
1350         struct virtio_net *dev = *pdev;
1351         int enable = (int)msg->payload.state.num;
1352         int index = (int)msg->payload.state.index;
1353         struct rte_vdpa_device *vdpa_dev;
1354         int did = -1;
1355
1356         RTE_LOG(INFO, VHOST_CONFIG,
1357                 "set queue enable: %d to qp idx: %d\n",
1358                 enable, index);
1359
1360         did = dev->vdpa_dev_id;
1361         vdpa_dev = rte_vdpa_get_device(did);
1362         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1363                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1364
1365         if (dev->notify_ops->vring_state_changed)
1366                 dev->notify_ops->vring_state_changed(dev->vid,
1367                                 index, enable);
1368
1369         /* On disable, rings have to be stopped being processed. */
1370         if (!enable && dev->dequeue_zero_copy)
1371                 drain_zmbuf_list(dev->virtqueue[index]);
1372
1373         dev->virtqueue[index]->enabled = enable;
1374
1375         return RTE_VHOST_MSG_RESULT_OK;
1376 }
1377
1378 static int
1379 vhost_user_get_protocol_features(struct virtio_net **pdev,
1380                         struct VhostUserMsg *msg,
1381                         int main_fd __rte_unused)
1382 {
1383         struct virtio_net *dev = *pdev;
1384         uint64_t features, protocol_features;
1385
1386         rte_vhost_driver_get_features(dev->ifname, &features);
1387         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1388
1389         /*
1390          * REPLY_ACK protocol feature is only mandatory for now
1391          * for IOMMU feature. If IOMMU is explicitly disabled by the
1392          * application, disable also REPLY_ACK feature for older buggy
1393          * Qemu versions (from v2.7.0 to v2.9.0).
1394          */
1395         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1396                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1397
1398         msg->payload.u64 = protocol_features;
1399         msg->size = sizeof(msg->payload.u64);
1400         msg->fd_num = 0;
1401
1402         return RTE_VHOST_MSG_RESULT_REPLY;
1403 }
1404
1405 static int
1406 vhost_user_set_protocol_features(struct virtio_net **pdev,
1407                         struct VhostUserMsg *msg,
1408                         int main_fd __rte_unused)
1409 {
1410         struct virtio_net *dev = *pdev;
1411         uint64_t protocol_features = msg->payload.u64;
1412         uint64_t slave_protocol_features = 0;
1413
1414         rte_vhost_driver_get_protocol_features(dev->ifname,
1415                         &slave_protocol_features);
1416         if (protocol_features & ~slave_protocol_features) {
1417                 RTE_LOG(ERR, VHOST_CONFIG,
1418                         "(%d) received invalid protocol features.\n",
1419                         dev->vid);
1420                 return RTE_VHOST_MSG_RESULT_ERR;
1421         }
1422
1423         dev->protocol_features = protocol_features;
1424         RTE_LOG(INFO, VHOST_CONFIG,
1425                 "negotiated Vhost-user protocol features: 0x%" PRIx64 "\n",
1426                 dev->protocol_features);
1427
1428         return RTE_VHOST_MSG_RESULT_OK;
1429 }
1430
1431 static int
1432 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1433                         int main_fd __rte_unused)
1434 {
1435         struct virtio_net *dev = *pdev;
1436         int fd = msg->fds[0];
1437         uint64_t size, off;
1438         void *addr;
1439
1440         if (fd < 0) {
1441                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1442                 return RTE_VHOST_MSG_RESULT_ERR;
1443         }
1444
1445         if (msg->size != sizeof(VhostUserLog)) {
1446                 RTE_LOG(ERR, VHOST_CONFIG,
1447                         "invalid log base msg size: %"PRId32" != %d\n",
1448                         msg->size, (int)sizeof(VhostUserLog));
1449                 return RTE_VHOST_MSG_RESULT_ERR;
1450         }
1451
1452         size = msg->payload.log.mmap_size;
1453         off  = msg->payload.log.mmap_offset;
1454
1455         /* Don't allow mmap_offset to point outside the mmap region */
1456         if (off > size) {
1457                 RTE_LOG(ERR, VHOST_CONFIG,
1458                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1459                         off, size);
1460                 return RTE_VHOST_MSG_RESULT_ERR;
1461         }
1462
1463         RTE_LOG(INFO, VHOST_CONFIG,
1464                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1465                 size, off);
1466
1467         /*
1468          * mmap from 0 to workaround a hugepage mmap bug: mmap will
1469          * fail when offset is not page size aligned.
1470          */
1471         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1472         close(fd);
1473         if (addr == MAP_FAILED) {
1474                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1475                 return RTE_VHOST_MSG_RESULT_ERR;
1476         }
1477
1478         /*
1479          * Free previously mapped log memory on occasionally
1480          * multiple VHOST_USER_SET_LOG_BASE.
1481          */
1482         if (dev->log_addr) {
1483                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1484         }
1485         dev->log_addr = (uint64_t)(uintptr_t)addr;
1486         dev->log_base = dev->log_addr + off;
1487         dev->log_size = size;
1488
1489         /*
1490          * The spec is not clear about it (yet), but QEMU doesn't expect
1491          * any payload in the reply.
1492          */
1493         msg->size = 0;
1494         msg->fd_num = 0;
1495
1496         return RTE_VHOST_MSG_RESULT_REPLY;
1497 }
1498
1499 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1500                         struct VhostUserMsg *msg,
1501                         int main_fd __rte_unused)
1502 {
1503         close(msg->fds[0]);
1504         RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1505
1506         return RTE_VHOST_MSG_RESULT_OK;
1507 }
1508
1509 /*
1510  * An rarp packet is constructed and broadcasted to notify switches about
1511  * the new location of the migrated VM, so that packets from outside will
1512  * not be lost after migration.
1513  *
1514  * However, we don't actually "send" a rarp packet here, instead, we set
1515  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1516  */
1517 static int
1518 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
1519                         int main_fd __rte_unused)
1520 {
1521         struct virtio_net *dev = *pdev;
1522         uint8_t *mac = (uint8_t *)&msg->payload.u64;
1523         struct rte_vdpa_device *vdpa_dev;
1524         int did = -1;
1525
1526         RTE_LOG(DEBUG, VHOST_CONFIG,
1527                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1528                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1529         memcpy(dev->mac.addr_bytes, mac, 6);
1530
1531         /*
1532          * Set the flag to inject a RARP broadcast packet at
1533          * rte_vhost_dequeue_burst().
1534          *
1535          * rte_smp_wmb() is for making sure the mac is copied
1536          * before the flag is set.
1537          */
1538         rte_smp_wmb();
1539         rte_atomic16_set(&dev->broadcast_rarp, 1);
1540         did = dev->vdpa_dev_id;
1541         vdpa_dev = rte_vdpa_get_device(did);
1542         if (vdpa_dev && vdpa_dev->ops->migration_done)
1543                 vdpa_dev->ops->migration_done(dev->vid);
1544
1545         return RTE_VHOST_MSG_RESULT_OK;
1546 }
1547
1548 static int
1549 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
1550                         int main_fd __rte_unused)
1551 {
1552         struct virtio_net *dev = *pdev;
1553         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1554                         msg->payload.u64 > VIRTIO_MAX_MTU) {
1555                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1556                                 msg->payload.u64);
1557
1558                 return RTE_VHOST_MSG_RESULT_ERR;
1559         }
1560
1561         dev->mtu = msg->payload.u64;
1562
1563         return RTE_VHOST_MSG_RESULT_OK;
1564 }
1565
1566 static int
1567 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
1568                         int main_fd __rte_unused)
1569 {
1570         struct virtio_net *dev = *pdev;
1571         int fd = msg->fds[0];
1572
1573         if (fd < 0) {
1574                 RTE_LOG(ERR, VHOST_CONFIG,
1575                                 "Invalid file descriptor for slave channel (%d)\n",
1576                                 fd);
1577                 return RTE_VHOST_MSG_RESULT_ERR;
1578         }
1579
1580         if (dev->slave_req_fd >= 0)
1581                 close(dev->slave_req_fd);
1582
1583         dev->slave_req_fd = fd;
1584
1585         return RTE_VHOST_MSG_RESULT_OK;
1586 }
1587
1588 static int
1589 is_vring_iotlb_split(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1590 {
1591         struct vhost_vring_addr *ra;
1592         uint64_t start, end, len;
1593
1594         start = imsg->iova;
1595         end = start + imsg->size;
1596
1597         ra = &vq->ring_addrs;
1598         len = sizeof(struct vring_desc) * vq->size;
1599         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
1600                 return 1;
1601
1602         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
1603         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
1604                 return 1;
1605
1606         len = sizeof(struct vring_used) +
1607                sizeof(struct vring_used_elem) * vq->size;
1608         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
1609                 return 1;
1610
1611         return 0;
1612 }
1613
1614 static int
1615 is_vring_iotlb_packed(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1616 {
1617         struct vhost_vring_addr *ra;
1618         uint64_t start, end, len;
1619
1620         start = imsg->iova;
1621         end = start + imsg->size;
1622
1623         ra = &vq->ring_addrs;
1624         len = sizeof(struct vring_packed_desc) * vq->size;
1625         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
1626                 return 1;
1627
1628         len = sizeof(struct vring_packed_desc_event);
1629         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
1630                 return 1;
1631
1632         len = sizeof(struct vring_packed_desc_event);
1633         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
1634                 return 1;
1635
1636         return 0;
1637 }
1638
1639 static int is_vring_iotlb(struct virtio_net *dev,
1640                           struct vhost_virtqueue *vq,
1641                           struct vhost_iotlb_msg *imsg)
1642 {
1643         if (vq_is_packed(dev))
1644                 return is_vring_iotlb_packed(vq, imsg);
1645         else
1646                 return is_vring_iotlb_split(vq, imsg);
1647 }
1648
1649 static int
1650 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
1651                         int main_fd __rte_unused)
1652 {
1653         struct virtio_net *dev = *pdev;
1654         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1655         uint16_t i;
1656         uint64_t vva, len;
1657
1658         switch (imsg->type) {
1659         case VHOST_IOTLB_UPDATE:
1660                 len = imsg->size;
1661                 vva = qva_to_vva(dev, imsg->uaddr, &len);
1662                 if (!vva)
1663                         return RTE_VHOST_MSG_RESULT_ERR;
1664
1665                 for (i = 0; i < dev->nr_vring; i++) {
1666                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1667
1668                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1669                                         len, imsg->perm);
1670
1671                         if (is_vring_iotlb(dev, vq, imsg))
1672                                 *pdev = dev = translate_ring_addresses(dev, i);
1673                 }
1674                 break;
1675         case VHOST_IOTLB_INVALIDATE:
1676                 for (i = 0; i < dev->nr_vring; i++) {
1677                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1678
1679                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
1680                                         imsg->size);
1681
1682                         if (is_vring_iotlb(dev, vq, imsg))
1683                                 vring_invalidate(dev, vq);
1684                 }
1685                 break;
1686         default:
1687                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1688                                 imsg->type);
1689                 return RTE_VHOST_MSG_RESULT_ERR;
1690         }
1691
1692         return RTE_VHOST_MSG_RESULT_OK;
1693 }
1694
1695 static int
1696 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
1697                         struct VhostUserMsg *msg,
1698                         int main_fd __rte_unused)
1699 {
1700         struct virtio_net *dev = *pdev;
1701 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1702         struct uffdio_api api_struct;
1703
1704         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1705
1706         if (dev->postcopy_ufd == -1) {
1707                 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
1708                         strerror(errno));
1709                 return RTE_VHOST_MSG_RESULT_ERR;
1710         }
1711         api_struct.api = UFFD_API;
1712         api_struct.features = 0;
1713         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
1714                 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
1715                         strerror(errno));
1716                 close(dev->postcopy_ufd);
1717                 dev->postcopy_ufd = -1;
1718                 return RTE_VHOST_MSG_RESULT_ERR;
1719         }
1720         msg->fds[0] = dev->postcopy_ufd;
1721         msg->fd_num = 1;
1722
1723         return RTE_VHOST_MSG_RESULT_REPLY;
1724 #else
1725         dev->postcopy_ufd = -1;
1726         msg->fd_num = 0;
1727
1728         return RTE_VHOST_MSG_RESULT_ERR;
1729 #endif
1730 }
1731
1732 static int
1733 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
1734                         struct VhostUserMsg *msg __rte_unused,
1735                         int main_fd __rte_unused)
1736 {
1737         struct virtio_net *dev = *pdev;
1738
1739         if (dev->mem && dev->mem->nregions) {
1740                 RTE_LOG(ERR, VHOST_CONFIG,
1741                         "Regions already registered at postcopy-listen\n");
1742                 return RTE_VHOST_MSG_RESULT_ERR;
1743         }
1744         dev->postcopy_listening = 1;
1745
1746         return RTE_VHOST_MSG_RESULT_OK;
1747 }
1748
1749 static int
1750 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
1751                         int main_fd __rte_unused)
1752 {
1753         struct virtio_net *dev = *pdev;
1754
1755         dev->postcopy_listening = 0;
1756         if (dev->postcopy_ufd >= 0) {
1757                 close(dev->postcopy_ufd);
1758                 dev->postcopy_ufd = -1;
1759         }
1760
1761         msg->payload.u64 = 0;
1762         msg->size = sizeof(msg->payload.u64);
1763         msg->fd_num = 0;
1764
1765         return RTE_VHOST_MSG_RESULT_REPLY;
1766 }
1767
1768 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
1769                                         struct VhostUserMsg *msg,
1770                                         int main_fd);
1771 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
1772         [VHOST_USER_NONE] = NULL,
1773         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
1774         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
1775         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
1776         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
1777         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
1778         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
1779         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
1780         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
1781         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
1782         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
1783         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
1784         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
1785         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
1786         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
1787         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
1788         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
1789         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
1790         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
1791         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
1792         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
1793         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
1794         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
1795         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
1796         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
1797         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
1798 };
1799
1800
1801 /* return bytes# of read on success or negative val on failure. */
1802 static int
1803 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1804 {
1805         int ret;
1806
1807         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1808                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
1809         if (ret <= 0)
1810                 return ret;
1811
1812         if (msg->size) {
1813                 if (msg->size > sizeof(msg->payload)) {
1814                         RTE_LOG(ERR, VHOST_CONFIG,
1815                                 "invalid msg size: %d\n", msg->size);
1816                         return -1;
1817                 }
1818                 ret = read(sockfd, &msg->payload, msg->size);
1819                 if (ret <= 0)
1820                         return ret;
1821                 if (ret != (int)msg->size) {
1822                         RTE_LOG(ERR, VHOST_CONFIG,
1823                                 "read control message failed\n");
1824                         return -1;
1825                 }
1826         }
1827
1828         return ret;
1829 }
1830
1831 static int
1832 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1833 {
1834         if (!msg)
1835                 return 0;
1836
1837         return send_fd_message(sockfd, (char *)msg,
1838                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
1839 }
1840
1841 static int
1842 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1843 {
1844         if (!msg)
1845                 return 0;
1846
1847         msg->flags &= ~VHOST_USER_VERSION_MASK;
1848         msg->flags &= ~VHOST_USER_NEED_REPLY;
1849         msg->flags |= VHOST_USER_VERSION;
1850         msg->flags |= VHOST_USER_REPLY_MASK;
1851
1852         return send_vhost_message(sockfd, msg);
1853 }
1854
1855 static int
1856 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
1857 {
1858         int ret;
1859
1860         if (msg->flags & VHOST_USER_NEED_REPLY)
1861                 rte_spinlock_lock(&dev->slave_req_lock);
1862
1863         ret = send_vhost_message(dev->slave_req_fd, msg);
1864         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1865                 rte_spinlock_unlock(&dev->slave_req_lock);
1866
1867         return ret;
1868 }
1869
1870 /*
1871  * Allocate a queue pair if it hasn't been allocated yet
1872  */
1873 static int
1874 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1875                         struct VhostUserMsg *msg)
1876 {
1877         uint16_t vring_idx;
1878
1879         switch (msg->request.master) {
1880         case VHOST_USER_SET_VRING_KICK:
1881         case VHOST_USER_SET_VRING_CALL:
1882         case VHOST_USER_SET_VRING_ERR:
1883                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1884                 break;
1885         case VHOST_USER_SET_VRING_NUM:
1886         case VHOST_USER_SET_VRING_BASE:
1887         case VHOST_USER_SET_VRING_ENABLE:
1888                 vring_idx = msg->payload.state.index;
1889                 break;
1890         case VHOST_USER_SET_VRING_ADDR:
1891                 vring_idx = msg->payload.addr.index;
1892                 break;
1893         default:
1894                 return 0;
1895         }
1896
1897         if (vring_idx >= VHOST_MAX_VRING) {
1898                 RTE_LOG(ERR, VHOST_CONFIG,
1899                         "invalid vring index: %u\n", vring_idx);
1900                 return -1;
1901         }
1902
1903         if (dev->virtqueue[vring_idx])
1904                 return 0;
1905
1906         return alloc_vring_queue(dev, vring_idx);
1907 }
1908
1909 static void
1910 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1911 {
1912         unsigned int i = 0;
1913         unsigned int vq_num = 0;
1914
1915         while (vq_num < dev->nr_vring) {
1916                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1917
1918                 if (vq) {
1919                         rte_spinlock_lock(&vq->access_lock);
1920                         vq_num++;
1921                 }
1922                 i++;
1923         }
1924 }
1925
1926 static void
1927 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1928 {
1929         unsigned int i = 0;
1930         unsigned int vq_num = 0;
1931
1932         while (vq_num < dev->nr_vring) {
1933                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1934
1935                 if (vq) {
1936                         rte_spinlock_unlock(&vq->access_lock);
1937                         vq_num++;
1938                 }
1939                 i++;
1940         }
1941 }
1942
1943 int
1944 vhost_user_msg_handler(int vid, int fd)
1945 {
1946         struct virtio_net *dev;
1947         struct VhostUserMsg msg;
1948         struct rte_vdpa_device *vdpa_dev;
1949         int did = -1;
1950         int ret;
1951         int unlock_required = 0;
1952         bool handled;
1953         int request;
1954
1955         dev = get_device(vid);
1956         if (dev == NULL)
1957                 return -1;
1958
1959         if (!dev->notify_ops) {
1960                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1961                 if (!dev->notify_ops) {
1962                         RTE_LOG(ERR, VHOST_CONFIG,
1963                                 "failed to get callback ops for driver %s\n",
1964                                 dev->ifname);
1965                         return -1;
1966                 }
1967         }
1968
1969         ret = read_vhost_message(fd, &msg);
1970         if (ret <= 0) {
1971                 if (ret < 0)
1972                         RTE_LOG(ERR, VHOST_CONFIG,
1973                                 "vhost read message failed\n");
1974                 else
1975                         RTE_LOG(INFO, VHOST_CONFIG,
1976                                 "vhost peer closed\n");
1977
1978                 return -1;
1979         }
1980
1981         ret = 0;
1982         request = msg.request.master;
1983         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
1984                         vhost_message_str[request]) {
1985                 if (request != VHOST_USER_IOTLB_MSG)
1986                         RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1987                                 vhost_message_str[request]);
1988                 else
1989                         RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1990                                 vhost_message_str[request]);
1991         } else {
1992                 RTE_LOG(DEBUG, VHOST_CONFIG, "External request %d\n", request);
1993         }
1994
1995         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1996         if (ret < 0) {
1997                 RTE_LOG(ERR, VHOST_CONFIG,
1998                         "failed to alloc queue\n");
1999                 return -1;
2000         }
2001
2002         /*
2003          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
2004          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
2005          * and device is destroyed. destroy_device waits for queues to be
2006          * inactive, so it is safe. Otherwise taking the access_lock
2007          * would cause a dead lock.
2008          */
2009         switch (request) {
2010         case VHOST_USER_SET_FEATURES:
2011         case VHOST_USER_SET_PROTOCOL_FEATURES:
2012         case VHOST_USER_SET_OWNER:
2013         case VHOST_USER_SET_MEM_TABLE:
2014         case VHOST_USER_SET_LOG_BASE:
2015         case VHOST_USER_SET_LOG_FD:
2016         case VHOST_USER_SET_VRING_NUM:
2017         case VHOST_USER_SET_VRING_ADDR:
2018         case VHOST_USER_SET_VRING_BASE:
2019         case VHOST_USER_SET_VRING_KICK:
2020         case VHOST_USER_SET_VRING_CALL:
2021         case VHOST_USER_SET_VRING_ERR:
2022         case VHOST_USER_SET_VRING_ENABLE:
2023         case VHOST_USER_SEND_RARP:
2024         case VHOST_USER_NET_SET_MTU:
2025         case VHOST_USER_SET_SLAVE_REQ_FD:
2026                 vhost_user_lock_all_queue_pairs(dev);
2027                 unlock_required = 1;
2028                 break;
2029         default:
2030                 break;
2031
2032         }
2033
2034         handled = false;
2035         if (dev->extern_ops.pre_msg_handle) {
2036                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
2037                                 (void *)&msg);
2038                 switch (ret) {
2039                 case RTE_VHOST_MSG_RESULT_REPLY:
2040                         send_vhost_reply(fd, &msg);
2041                         /* Fall-through */
2042                 case RTE_VHOST_MSG_RESULT_ERR:
2043                 case RTE_VHOST_MSG_RESULT_OK:
2044                         handled = true;
2045                         goto skip_to_post_handle;
2046                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2047                 default:
2048                         break;
2049                 }
2050         }
2051
2052         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
2053                 if (!vhost_message_handlers[request])
2054                         goto skip_to_post_handle;
2055                 ret = vhost_message_handlers[request](&dev, &msg, fd);
2056
2057                 switch (ret) {
2058                 case RTE_VHOST_MSG_RESULT_ERR:
2059                         RTE_LOG(ERR, VHOST_CONFIG,
2060                                 "Processing %s failed.\n",
2061                                 vhost_message_str[request]);
2062                         handled = true;
2063                         break;
2064                 case RTE_VHOST_MSG_RESULT_OK:
2065                         RTE_LOG(DEBUG, VHOST_CONFIG,
2066                                 "Processing %s succeeded.\n",
2067                                 vhost_message_str[request]);
2068                         handled = true;
2069                         break;
2070                 case RTE_VHOST_MSG_RESULT_REPLY:
2071                         RTE_LOG(DEBUG, VHOST_CONFIG,
2072                                 "Processing %s succeeded and needs reply.\n",
2073                                 vhost_message_str[request]);
2074                         send_vhost_reply(fd, &msg);
2075                         handled = true;
2076                         break;
2077                 default:
2078                         break;
2079                 }
2080         }
2081
2082 skip_to_post_handle:
2083         if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2084                         dev->extern_ops.post_msg_handle) {
2085                 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
2086                                 (void *)&msg);
2087                 switch (ret) {
2088                 case RTE_VHOST_MSG_RESULT_REPLY:
2089                         send_vhost_reply(fd, &msg);
2090                         /* Fall-through */
2091                 case RTE_VHOST_MSG_RESULT_ERR:
2092                 case RTE_VHOST_MSG_RESULT_OK:
2093                         handled = true;
2094                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2095                 default:
2096                         break;
2097                 }
2098         }
2099
2100         if (unlock_required)
2101                 vhost_user_unlock_all_queue_pairs(dev);
2102
2103         /* If message was not handled at this stage, treat it as an error */
2104         if (!handled) {
2105                 RTE_LOG(ERR, VHOST_CONFIG,
2106                         "vhost message (req: %d) was not handled.\n", request);
2107                 ret = RTE_VHOST_MSG_RESULT_ERR;
2108         }
2109
2110         /*
2111          * If the request required a reply that was already sent,
2112          * this optional reply-ack won't be sent as the
2113          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2114          */
2115         if (msg.flags & VHOST_USER_NEED_REPLY) {
2116                 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2117                 msg.size = sizeof(msg.payload.u64);
2118                 msg.fd_num = 0;
2119                 send_vhost_reply(fd, &msg);
2120         } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2121                 RTE_LOG(ERR, VHOST_CONFIG,
2122                         "vhost message handling failed.\n");
2123                 return -1;
2124         }
2125
2126         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2127                 dev->flags |= VIRTIO_DEV_READY;
2128
2129                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2130                         if (dev->dequeue_zero_copy) {
2131                                 RTE_LOG(INFO, VHOST_CONFIG,
2132                                                 "dequeue zero copy is enabled\n");
2133                         }
2134
2135                         if (dev->notify_ops->new_device(dev->vid) == 0)
2136                                 dev->flags |= VIRTIO_DEV_RUNNING;
2137                 }
2138         }
2139
2140         did = dev->vdpa_dev_id;
2141         vdpa_dev = rte_vdpa_get_device(did);
2142         if (vdpa_dev && virtio_is_ready(dev) &&
2143                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2144                         msg.request.master == VHOST_USER_SET_VRING_CALL) {
2145                 if (vdpa_dev->ops->dev_conf)
2146                         vdpa_dev->ops->dev_conf(dev->vid);
2147                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2148         }
2149
2150         return 0;
2151 }
2152
2153 static int process_slave_message_reply(struct virtio_net *dev,
2154                                        const struct VhostUserMsg *msg)
2155 {
2156         struct VhostUserMsg msg_reply;
2157         int ret;
2158
2159         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2160                 return 0;
2161
2162         if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
2163                 ret = -1;
2164                 goto out;
2165         }
2166
2167         if (msg_reply.request.slave != msg->request.slave) {
2168                 RTE_LOG(ERR, VHOST_CONFIG,
2169                         "Received unexpected msg type (%u), expected %u\n",
2170                         msg_reply.request.slave, msg->request.slave);
2171                 ret = -1;
2172                 goto out;
2173         }
2174
2175         ret = msg_reply.payload.u64 ? -1 : 0;
2176
2177 out:
2178         rte_spinlock_unlock(&dev->slave_req_lock);
2179         return ret;
2180 }
2181
2182 int
2183 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2184 {
2185         int ret;
2186         struct VhostUserMsg msg = {
2187                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2188                 .flags = VHOST_USER_VERSION,
2189                 .size = sizeof(msg.payload.iotlb),
2190                 .payload.iotlb = {
2191                         .iova = iova,
2192                         .perm = perm,
2193                         .type = VHOST_IOTLB_MISS,
2194                 },
2195         };
2196
2197         ret = send_vhost_message(dev->slave_req_fd, &msg);
2198         if (ret < 0) {
2199                 RTE_LOG(ERR, VHOST_CONFIG,
2200                                 "Failed to send IOTLB miss message (%d)\n",
2201                                 ret);
2202                 return ret;
2203         }
2204
2205         return 0;
2206 }
2207
2208 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2209                                                     int index, int fd,
2210                                                     uint64_t offset,
2211                                                     uint64_t size)
2212 {
2213         int ret;
2214         struct VhostUserMsg msg = {
2215                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2216                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2217                 .size = sizeof(msg.payload.area),
2218                 .payload.area = {
2219                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
2220                         .size = size,
2221                         .offset = offset,
2222                 },
2223         };
2224
2225         if (fd < 0)
2226                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2227         else {
2228                 msg.fds[0] = fd;
2229                 msg.fd_num = 1;
2230         }
2231
2232         ret = send_vhost_slave_message(dev, &msg);
2233         if (ret < 0) {
2234                 RTE_LOG(ERR, VHOST_CONFIG,
2235                         "Failed to set host notifier (%d)\n", ret);
2236                 return ret;
2237         }
2238
2239         return process_slave_message_reply(dev, &msg);
2240 }
2241
2242 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2243 {
2244         struct virtio_net *dev;
2245         struct rte_vdpa_device *vdpa_dev;
2246         int vfio_device_fd, did, ret = 0;
2247         uint64_t offset, size;
2248         unsigned int i;
2249
2250         dev = get_device(vid);
2251         if (!dev)
2252                 return -ENODEV;
2253
2254         did = dev->vdpa_dev_id;
2255         if (did < 0)
2256                 return -EINVAL;
2257
2258         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2259             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2260             !(dev->protocol_features &
2261                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2262             !(dev->protocol_features &
2263                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2264             !(dev->protocol_features &
2265                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2266                 return -ENOTSUP;
2267
2268         vdpa_dev = rte_vdpa_get_device(did);
2269         if (!vdpa_dev)
2270                 return -ENODEV;
2271
2272         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2273         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2274
2275         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2276         if (vfio_device_fd < 0)
2277                 return -ENOTSUP;
2278
2279         if (enable) {
2280                 for (i = 0; i < dev->nr_vring; i++) {
2281                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2282                                         &size) < 0) {
2283                                 ret = -ENOTSUP;
2284                                 goto disable;
2285                         }
2286
2287                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2288                                         vfio_device_fd, offset, size) < 0) {
2289                                 ret = -EFAULT;
2290                                 goto disable;
2291                         }
2292                 }
2293         } else {
2294 disable:
2295                 for (i = 0; i < dev->nr_vring; i++) {
2296                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
2297                                         0, 0);
2298                 }
2299         }
2300
2301         return ret;
2302 }