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
 565                 vva = vhost_user_iotlb_cache_find(vq, ra,
 566                                         size, VHOST_ACCESS_RW);
 567                 if (!vva)
 568                         vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
 569
 570                 return vva;
 571         }
 572
 573         return qva_to_vva(dev, ra, size);
 574 }
 575
 576 static struct virtio_net *
 577 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 578 {
 579         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 580         struct vhost_vring_addr *addr = &vq->ring_addrs;
 581         uint64_t len, expected_len;
 582
 583         if (vq_is_packed(dev)) {
 584                 len = sizeof(struct vring_packed_desc) * vq->size;
 585                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 586                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 587                 vq->log_guest_addr = 0;
 588                 if (vq->desc_packed == NULL ||
 589                                 len != sizeof(struct vring_packed_desc) *
 590                                 vq->size) {
 591                         RTE_LOG(DEBUG, VHOST_CONFIG,
 592                                 "(%d) failed to map desc_packed ring.\n",
 593                                 dev->vid);
 594                         return dev;
 595                 }
 596
 597                 dev = numa_realloc(dev, vq_index);
 598                 vq = dev->virtqueue[vq_index];
 599                 addr = &vq->ring_addrs;
 600
 601                 len = sizeof(struct vring_packed_desc_event);
 602                 vq->driver_event = (struct vring_packed_desc_event *)
 603                                         (uintptr_t)ring_addr_to_vva(dev,
 604                                         vq, addr->avail_user_addr, &len);
 605                 if (vq->driver_event == NULL ||
 606                                 len != sizeof(struct vring_packed_desc_event)) {
 607                         RTE_LOG(DEBUG, VHOST_CONFIG,
 608                                 "(%d) failed to find driver area address.\n",
 609                                 dev->vid);
 610                         return dev;
 611                 }
 612
 613                 len = sizeof(struct vring_packed_desc_event);
 614                 vq->device_event = (struct vring_packed_desc_event *)
 615                                         (uintptr_t)ring_addr_to_vva(dev,
 616                                         vq, addr->used_user_addr, &len);
 617                 if (vq->device_event == NULL ||
 618                                 len != sizeof(struct vring_packed_desc_event)) {
 619                         RTE_LOG(DEBUG, VHOST_CONFIG,
 620                                 "(%d) failed to find device area address.\n",
 621                                 dev->vid);
 622                         return dev;
 623                 }
 624
 625                 return dev;
 626         }
 627
 628         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 629         if (vq->desc && vq->avail && vq->used)
 630                 return dev;
 631
 632         len = sizeof(struct vring_desc) * vq->size;
 633         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 634                         vq, addr->desc_user_addr, &len);
 635         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 636                 RTE_LOG(DEBUG, VHOST_CONFIG,
 637                         "(%d) failed to map desc ring.\n",
 638                         dev->vid);
 639                 return dev;
 640         }
 641
 642         dev = numa_realloc(dev, vq_index);
 643         vq = dev->virtqueue[vq_index];
 644         addr = &vq->ring_addrs;
 645
 646         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 647         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 648                 len += sizeof(uint16_t);
 649         expected_len = len;
 650         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 651                         vq, addr->avail_user_addr, &len);
 652         if (vq->avail == 0 || len != expected_len) {
 653                 RTE_LOG(DEBUG, VHOST_CONFIG,
 654                         "(%d) failed to map avail ring.\n",
 655                         dev->vid);
 656                 return dev;
 657         }
 658
 659         len = sizeof(struct vring_used) +
 660                 sizeof(struct vring_used_elem) * vq->size;
 661         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 662                 len += sizeof(uint16_t);
 663         expected_len = len;
 664         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 665                         vq, addr->used_user_addr, &len);
 666         if (vq->used == 0 || len != expected_len) {
 667                 RTE_LOG(DEBUG, VHOST_CONFIG,
 668                         "(%d) failed to map used ring.\n",
 669                         dev->vid);
 670                 return dev;
 671         }
 672
 673         if (vq->last_used_idx != vq->used->idx) {
 674                 RTE_LOG(WARNING, VHOST_CONFIG,
 675                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 676                         "some packets maybe resent for Tx and dropped for Rx\n",
 677                         vq->last_used_idx, vq->used->idx);
 678                 vq->last_used_idx  = vq->used->idx;
 679                 vq->last_avail_idx = vq->used->idx;
 680         }
 681
 682         vq->log_guest_addr = addr->log_guest_addr;
 683
 684         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
 685                         dev->vid, vq->desc);
 686         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
 687                         dev->vid, vq->avail);
 688         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
 689                         dev->vid, vq->used);
 690         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
 691                         dev->vid, vq->log_guest_addr);
 692
 693         return dev;
 694 }
 695
 696 /*
 697  * The virtio device sends us the desc, used and avail ring addresses.
 698  * This function then converts these to our address space.
 699  */
 700 static int
 701 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 702                         int main_fd __rte_unused)
 703 {
 704         struct virtio_net *dev = *pdev;
 705         struct vhost_virtqueue *vq;
 706         struct vhost_vring_addr *addr = &msg->payload.addr;
 707
 708         if (dev->mem == NULL)
 709                 return RTE_VHOST_MSG_RESULT_ERR;
 710
 711         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 712         vq = dev->virtqueue[msg->payload.addr.index];
 713
 714         /*
 715          * Rings addresses should not be interpreted as long as the ring is not
 716          * started and enabled
 717          */
 718         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 719
 720         vring_invalidate(dev, vq);
 721
 722         if (vq->enabled && (dev->features &
 723                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
 724                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 725                 if (!dev)
 726                         return RTE_VHOST_MSG_RESULT_ERR;
 727
 728                 *pdev = dev;
 729         }
 730
 731         return RTE_VHOST_MSG_RESULT_OK;
 732 }
 733
 734 /*
 735  * The virtio device sends us the available ring last used index.
 736  */
 737 static int
 738 vhost_user_set_vring_base(struct virtio_net **pdev,
 739                         struct VhostUserMsg *msg,
 740                         int main_fd __rte_unused)
 741 {
 742         struct virtio_net *dev = *pdev;
 743         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 744         uint64_t val = msg->payload.state.num;
 745
 746         if (vq_is_packed(dev)) {
 747                 /*
 748                  * Bit[0:14]: avail index
 749                  * Bit[15]: avail wrap counter
 750                  */
 751                 vq->last_avail_idx = val & 0x7fff;
 752                 vq->avail_wrap_counter = !!(val & (0x1 << 15));
 753                 /*
 754                  * Set used index to same value as available one, as
 755                  * their values should be the same since ring processing
 756                  * was stopped at get time.
 757                  */
 758                 vq->last_used_idx = vq->last_avail_idx;
 759                 vq->used_wrap_counter = vq->avail_wrap_counter;
 760         } else {
 761                 vq->last_used_idx = msg->payload.state.num;
 762                 vq->last_avail_idx = msg->payload.state.num;
 763         }
 764
 765         return RTE_VHOST_MSG_RESULT_OK;
 766 }
 767
 768 static int
 769 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 770                    uint64_t host_phys_addr, uint64_t size)
 771 {
 772         struct guest_page *page, *last_page;
 773         struct guest_page *old_pages;
 774
 775         if (dev->nr_guest_pages == dev->max_guest_pages) {
 776                 dev->max_guest_pages *= 2;
 777                 old_pages = dev->guest_pages;
 778                 dev->guest_pages = realloc(dev->guest_pages,
 779                                         dev->max_guest_pages * sizeof(*page));
 780                 if (!dev->guest_pages) {
 781                         RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
 782                         free(old_pages);
 783                         return -1;
 784                 }
 785         }
 786
 787         if (dev->nr_guest_pages > 0) {
 788                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 789                 /* merge if the two pages are continuous */
 790                 if (host_phys_addr == last_page->host_phys_addr +
 791                                       last_page->size) {
 792                         last_page->size += size;
 793                         return 0;
 794                 }
 795         }
 796
 797         page = &dev->guest_pages[dev->nr_guest_pages++];
 798         page->guest_phys_addr = guest_phys_addr;
 799         page->host_phys_addr  = host_phys_addr;
 800         page->size = size;
 801
 802         return 0;
 803 }
 804
 805 static int
 806 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 807                 uint64_t page_size)
 808 {
 809         uint64_t reg_size = reg->size;
 810         uint64_t host_user_addr  = reg->host_user_addr;
 811         uint64_t guest_phys_addr = reg->guest_phys_addr;
 812         uint64_t host_phys_addr;
 813         uint64_t size;
 814
 815         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 816         size = page_size - (guest_phys_addr & (page_size - 1));
 817         size = RTE_MIN(size, reg_size);
 818
 819         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 820                 return -1;
 821
 822         host_user_addr  += size;
 823         guest_phys_addr += size;
 824         reg_size -= size;
 825
 826         while (reg_size > 0) {
 827                 size = RTE_MIN(reg_size, page_size);
 828                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 829                                                   host_user_addr);
 830                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 831                                 size) < 0)
 832                         return -1;
 833
 834                 host_user_addr  += size;
 835                 guest_phys_addr += size;
 836                 reg_size -= size;
 837         }
 838
 839         return 0;
 840 }
 841
 842 #ifdef RTE_LIBRTE_VHOST_DEBUG
 843 /* TODO: enable it only in debug mode? */
 844 static void
 845 dump_guest_pages(struct virtio_net *dev)
 846 {
 847         uint32_t i;
 848         struct guest_page *page;
 849
 850         for (i = 0; i < dev->nr_guest_pages; i++) {
 851                 page = &dev->guest_pages[i];
 852
 853                 RTE_LOG(INFO, VHOST_CONFIG,
 854                         "guest physical page region %u\n"
 855                         "\t guest_phys_addr: %" PRIx64 "\n"
 856                         "\t host_phys_addr : %" PRIx64 "\n"
 857                         "\t size           : %" PRIx64 "\n",
 858                         i,
 859                         page->guest_phys_addr,
 860                         page->host_phys_addr,
 861                         page->size);
 862         }
 863 }
 864 #else
 865 #define dump_guest_pages(dev)
 866 #endif
 867
 868 static bool
 869 vhost_memory_changed(struct VhostUserMemory *new,
 870                      struct rte_vhost_memory *old)
 871 {
 872         uint32_t i;
 873
 874         if (new->nregions != old->nregions)
 875                 return true;
 876
 877         for (i = 0; i < new->nregions; ++i) {
 878                 VhostUserMemoryRegion *new_r = &new->regions[i];
 879                 struct rte_vhost_mem_region *old_r = &old->regions[i];
 880
 881                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
 882                         return true;
 883                 if (new_r->memory_size != old_r->size)
 884                         return true;
 885                 if (new_r->userspace_addr != old_r->guest_user_addr)
 886                         return true;
 887         }
 888
 889         return false;
 890 }
 891
 892 static int
 893 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
 894                         int main_fd)
 895 {
 896         struct virtio_net *dev = *pdev;
 897         struct VhostUserMemory *memory = &msg->payload.memory;
 898         struct rte_vhost_mem_region *reg;
 899         void *mmap_addr;
 900         uint64_t mmap_size;
 901         uint64_t mmap_offset;
 902         uint64_t alignment;
 903         uint32_t i;
 904         int populate;
 905         int fd;
 906
 907         if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
 908                 RTE_LOG(ERR, VHOST_CONFIG,
 909                         "too many memory regions (%u)\n", memory->nregions);
 910                 return RTE_VHOST_MSG_RESULT_ERR;
 911         }
 912
 913         if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
 914                 RTE_LOG(INFO, VHOST_CONFIG,
 915                         "(%d) memory regions not changed\n", dev->vid);
 916
 917                 for (i = 0; i < memory->nregions; i++)
 918                         close(msg->fds[i]);
 919
 920                 return RTE_VHOST_MSG_RESULT_OK;
 921         }
 922
 923         if (dev->mem) {
 924                 free_mem_region(dev);
 925                 rte_free(dev->mem);
 926                 dev->mem = NULL;
 927         }
 928
 929         /* Flush IOTLB cache as previous HVAs are now invalid */
 930         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 931                 for (i = 0; i < dev->nr_vring; i++)
 932                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
 933
 934         dev->nr_guest_pages = 0;
 935         if (!dev->guest_pages) {
 936                 dev->max_guest_pages = 8;
 937                 dev->guest_pages = malloc(dev->max_guest_pages *
 938                                                 sizeof(struct guest_page));
 939                 if (dev->guest_pages == NULL) {
 940                         RTE_LOG(ERR, VHOST_CONFIG,
 941                                 "(%d) failed to allocate memory "
 942                                 "for dev->guest_pages\n",
 943                                 dev->vid);
 944                         return RTE_VHOST_MSG_RESULT_ERR;
 945                 }
 946         }
 947
 948         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
 949                 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
 950         if (dev->mem == NULL) {
 951                 RTE_LOG(ERR, VHOST_CONFIG,
 952                         "(%d) failed to allocate memory for dev->mem\n",
 953                         dev->vid);
 954                 return RTE_VHOST_MSG_RESULT_ERR;
 955         }
 956         dev->mem->nregions = memory->nregions;
 957
 958         for (i = 0; i < memory->nregions; i++) {
 959                 fd  = msg->fds[i];
 960                 reg = &dev->mem->regions[i];
 961
 962                 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
 963                 reg->guest_user_addr = memory->regions[i].userspace_addr;
 964                 reg->size            = memory->regions[i].memory_size;
 965                 reg->fd              = fd;
 966
 967                 mmap_offset = memory->regions[i].mmap_offset;
 968
 969                 /* Check for memory_size + mmap_offset overflow */
 970                 if (mmap_offset >= -reg->size) {
 971                         RTE_LOG(ERR, VHOST_CONFIG,
 972                                 "mmap_offset (%#"PRIx64") and memory_size "
 973                                 "(%#"PRIx64") overflow\n",
 974                                 mmap_offset, reg->size);
 975                         goto err_mmap;
 976                 }
 977
 978                 mmap_size = reg->size + mmap_offset;
 979
 980                 /* mmap() without flag of MAP_ANONYMOUS, should be called
 981                  * with length argument aligned with hugepagesz at older
 982                  * longterm version Linux, like 2.6.32 and 3.2.72, or
 983                  * mmap() will fail with EINVAL.
 984                  *
 985                  * to avoid failure, make sure in caller to keep length
 986                  * aligned.
 987                  */
 988                 alignment = get_blk_size(fd);
 989                 if (alignment == (uint64_t)-1) {
 990                         RTE_LOG(ERR, VHOST_CONFIG,
 991                                 "couldn't get hugepage size through fstat\n");
 992                         goto err_mmap;
 993                 }
 994                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
 995
 996                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
 997                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
 998                                  MAP_SHARED | populate, fd, 0);
 999
1000                 if (mmap_addr == MAP_FAILED) {
1001                         RTE_LOG(ERR, VHOST_CONFIG,
1002                                 "mmap region %u failed.\n", i);
1003                         goto err_mmap;
1004                 }
1005
1006                 reg->mmap_addr = mmap_addr;
1007                 reg->mmap_size = mmap_size;
1008                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
1009                                       mmap_offset;
1010
1011                 if (dev->dequeue_zero_copy)
1012                         if (add_guest_pages(dev, reg, alignment) < 0) {
1013                                 RTE_LOG(ERR, VHOST_CONFIG,
1014                                         "adding guest pages to region %u failed.\n",
1015                                         i);
1016                                 goto err_mmap;
1017                         }
1018
1019                 RTE_LOG(INFO, VHOST_CONFIG,
1020                         "guest memory region %u, size: 0x%" PRIx64 "\n"
1021                         "\t guest physical addr: 0x%" PRIx64 "\n"
1022                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
1023                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
1024                         "\t mmap addr : 0x%" PRIx64 "\n"
1025                         "\t mmap size : 0x%" PRIx64 "\n"
1026                         "\t mmap align: 0x%" PRIx64 "\n"
1027                         "\t mmap off  : 0x%" PRIx64 "\n",
1028                         i, reg->size,
1029                         reg->guest_phys_addr,
1030                         reg->guest_user_addr,
1031                         reg->host_user_addr,
1032                         (uint64_t)(uintptr_t)mmap_addr,
1033                         mmap_size,
1034                         alignment,
1035                         mmap_offset);
1036
1037                 if (dev->postcopy_listening) {
1038                         /*
1039                          * We haven't a better way right now than sharing
1040                          * DPDK's virtual address with Qemu, so that Qemu can
1041                          * retrieve the region offset when handling userfaults.
1042                          */
1043                         memory->regions[i].userspace_addr =
1044                                 reg->host_user_addr;
1045                 }
1046         }
1047         if (dev->postcopy_listening) {
1048                 /* Send the addresses back to qemu */
1049                 msg->fd_num = 0;
1050                 send_vhost_reply(main_fd, msg);
1051
1052                 /* Wait for qemu to acknolwedge it's got the addresses
1053                  * we've got to wait before we're allowed to generate faults.
1054                  */
1055                 VhostUserMsg ack_msg;
1056                 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1057                         RTE_LOG(ERR, VHOST_CONFIG,
1058                                 "Failed to read qemu ack on postcopy set-mem-table\n");
1059                         goto err_mmap;
1060                 }
1061                 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1062                         RTE_LOG(ERR, VHOST_CONFIG,
1063                                 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1064                                 ack_msg.request.master);
1065                         goto err_mmap;
1066                 }
1067
1068                 /* Now userfault register and we can use the memory */
1069                 for (i = 0; i < memory->nregions; i++) {
1070 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1071                         reg = &dev->mem->regions[i];
1072                         struct uffdio_register reg_struct;
1073
1074                         /*
1075                          * Let's register all the mmap'ed area to ensure
1076                          * alignment on page boundary.
1077                          */
1078                         reg_struct.range.start =
1079                                 (uint64_t)(uintptr_t)reg->mmap_addr;
1080                         reg_struct.range.len = reg->mmap_size;
1081                         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1082
1083                         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1084                                                 &reg_struct)) {
1085                                 RTE_LOG(ERR, VHOST_CONFIG,
1086                                         "Failed to register ufd for region %d: (ufd = %d) %s\n",
1087                                         i, dev->postcopy_ufd,
1088                                         strerror(errno));
1089                                 goto err_mmap;
1090                         }
1091                         RTE_LOG(INFO, VHOST_CONFIG,
1092                                 "\t userfaultfd registered for range : %llx - %llx\n",
1093                                 reg_struct.range.start,
1094                                 reg_struct.range.start +
1095                                 reg_struct.range.len - 1);
1096 #else
1097                         goto err_mmap;
1098 #endif
1099                 }
1100         }
1101
1102         for (i = 0; i < dev->nr_vring; i++) {
1103                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1104
1105                 if (vq->desc || vq->avail || vq->used) {
1106                         /*
1107                          * If the memory table got updated, the ring addresses
1108                          * need to be translated again as virtual addresses have
1109                          * changed.
1110                          */
1111                         vring_invalidate(dev, vq);
1112
1113                         dev = translate_ring_addresses(dev, i);
1114                         if (!dev) {
1115                                 dev = *pdev;
1116                                 goto err_mmap;
1117                         }
1118
1119                         *pdev = dev;
1120                 }
1121         }
1122
1123         dump_guest_pages(dev);
1124
1125         return RTE_VHOST_MSG_RESULT_OK;
1126
1127 err_mmap:
1128         free_mem_region(dev);
1129         rte_free(dev->mem);
1130         dev->mem = NULL;
1131         return RTE_VHOST_MSG_RESULT_ERR;
1132 }
1133
1134 static bool
1135 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1136 {
1137         bool rings_ok;
1138
1139         if (!vq)
1140                 return false;
1141
1142         if (vq_is_packed(dev))
1143                 rings_ok = !!vq->desc_packed;
1144         else
1145                 rings_ok = vq->desc && vq->avail && vq->used;
1146
1147         return rings_ok &&
1148                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1149                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1150 }
1151
1152 static int
1153 virtio_is_ready(struct virtio_net *dev)
1154 {
1155         struct vhost_virtqueue *vq;
1156         uint32_t i;
1157
1158         if (dev->nr_vring == 0)
1159                 return 0;
1160
1161         for (i = 0; i < dev->nr_vring; i++) {
1162                 vq = dev->virtqueue[i];
1163
1164                 if (!vq_is_ready(dev, vq))
1165                         return 0;
1166         }
1167
1168         RTE_LOG(INFO, VHOST_CONFIG,
1169                 "virtio is now ready for processing.\n");
1170         return 1;
1171 }
1172
1173 static int
1174 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1175                         int main_fd __rte_unused)
1176 {
1177         struct virtio_net *dev = *pdev;
1178         struct vhost_vring_file file;
1179         struct vhost_virtqueue *vq;
1180
1181         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1182         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1183                 file.fd = VIRTIO_INVALID_EVENTFD;
1184         else
1185                 file.fd = msg->fds[0];
1186         RTE_LOG(INFO, VHOST_CONFIG,
1187                 "vring call idx:%d file:%d\n", file.index, file.fd);
1188
1189         vq = dev->virtqueue[file.index];
1190         if (vq->callfd >= 0)
1191                 close(vq->callfd);
1192
1193         vq->callfd = file.fd;
1194
1195         return RTE_VHOST_MSG_RESULT_OK;
1196 }
1197
1198 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1199                         struct VhostUserMsg *msg,
1200                         int main_fd __rte_unused)
1201 {
1202         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1203                 close(msg->fds[0]);
1204         RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1205
1206         return RTE_VHOST_MSG_RESULT_OK;
1207 }
1208
1209 static int
1210 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1211                         int main_fd __rte_unused)
1212 {
1213         struct virtio_net *dev = *pdev;
1214         struct vhost_vring_file file;
1215         struct vhost_virtqueue *vq;
1216
1217         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1218         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1219                 file.fd = VIRTIO_INVALID_EVENTFD;
1220         else
1221                 file.fd = msg->fds[0];
1222         RTE_LOG(INFO, VHOST_CONFIG,
1223                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1224
1225         /* Interpret ring addresses only when ring is started. */
1226         dev = translate_ring_addresses(dev, file.index);
1227         if (!dev)
1228                 return RTE_VHOST_MSG_RESULT_ERR;
1229
1230         *pdev = dev;
1231
1232         vq = dev->virtqueue[file.index];
1233
1234         /*
1235          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1236          * the ring starts already enabled. Otherwise, it is enabled via
1237          * the SET_VRING_ENABLE message.
1238          */
1239         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
1240                 vq->enabled = 1;
1241                 if (dev->notify_ops->vring_state_changed)
1242                         dev->notify_ops->vring_state_changed(
1243                                 dev->vid, file.index, 1);
1244         }
1245
1246         if (vq->kickfd >= 0)
1247                 close(vq->kickfd);
1248         vq->kickfd = file.fd;
1249
1250         return RTE_VHOST_MSG_RESULT_OK;
1251 }
1252
1253 static void
1254 free_zmbufs(struct vhost_virtqueue *vq)
1255 {
1256         drain_zmbuf_list(vq);
1257
1258         rte_free(vq->zmbufs);
1259 }
1260
1261 /*
1262  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1263  */
1264 static int
1265 vhost_user_get_vring_base(struct virtio_net **pdev,
1266                         struct VhostUserMsg *msg,
1267                         int main_fd __rte_unused)
1268 {
1269         struct virtio_net *dev = *pdev;
1270         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1271         uint64_t val;
1272
1273         /* We have to stop the queue (virtio) if it is running. */
1274         vhost_destroy_device_notify(dev);
1275
1276         dev->flags &= ~VIRTIO_DEV_READY;
1277         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1278
1279         /* Here we are safe to get the indexes */
1280         if (vq_is_packed(dev)) {
1281                 /*
1282                  * Bit[0:14]: avail index
1283                  * Bit[15]: avail wrap counter
1284                  */
1285                 val = vq->last_avail_idx & 0x7fff;
1286                 val |= vq->avail_wrap_counter << 15;
1287                 msg->payload.state.num = val;
1288         } else {
1289                 msg->payload.state.num = vq->last_avail_idx;
1290         }
1291
1292         RTE_LOG(INFO, VHOST_CONFIG,
1293                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1294                 msg->payload.state.num);
1295         /*
1296          * Based on current qemu vhost-user implementation, this message is
1297          * sent and only sent in vhost_vring_stop.
1298          * TODO: cleanup the vring, it isn't usable since here.
1299          */
1300         if (vq->kickfd >= 0)
1301                 close(vq->kickfd);
1302
1303         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1304
1305         if (vq->callfd >= 0)
1306                 close(vq->callfd);
1307
1308         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1309
1310         vq->signalled_used_valid = false;
1311
1312         if (dev->dequeue_zero_copy)
1313                 free_zmbufs(vq);
1314         if (vq_is_packed(dev)) {
1315                 rte_free(vq->shadow_used_packed);
1316                 vq->shadow_used_packed = NULL;
1317         } else {
1318                 rte_free(vq->shadow_used_split);
1319                 vq->shadow_used_split = NULL;
1320         }
1321
1322         rte_free(vq->batch_copy_elems);
1323         vq->batch_copy_elems = NULL;
1324
1325         msg->size = sizeof(msg->payload.state);
1326         msg->fd_num = 0;
1327
1328         return RTE_VHOST_MSG_RESULT_REPLY;
1329 }
1330
1331 /*
1332  * when virtio queues are ready to work, qemu will send us to
1333  * enable the virtio queue pair.
1334  */
1335 static int
1336 vhost_user_set_vring_enable(struct virtio_net **pdev,
1337                         struct VhostUserMsg *msg,
1338                         int main_fd __rte_unused)
1339 {
1340         struct virtio_net *dev = *pdev;
1341         int enable = (int)msg->payload.state.num;
1342         int index = (int)msg->payload.state.index;
1343         struct rte_vdpa_device *vdpa_dev;
1344         int did = -1;
1345
1346         RTE_LOG(INFO, VHOST_CONFIG,
1347                 "set queue enable: %d to qp idx: %d\n",
1348                 enable, index);
1349
1350         did = dev->vdpa_dev_id;
1351         vdpa_dev = rte_vdpa_get_device(did);
1352         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1353                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1354
1355         if (dev->notify_ops->vring_state_changed)
1356                 dev->notify_ops->vring_state_changed(dev->vid,
1357                                 index, enable);
1358
1359         /* On disable, rings have to be stopped being processed. */
1360         if (!enable && dev->dequeue_zero_copy)
1361                 drain_zmbuf_list(dev->virtqueue[index]);
1362
1363         dev->virtqueue[index]->enabled = enable;
1364
1365         return RTE_VHOST_MSG_RESULT_OK;
1366 }
1367
1368 static int
1369 vhost_user_get_protocol_features(struct virtio_net **pdev,
1370                         struct VhostUserMsg *msg,
1371                         int main_fd __rte_unused)
1372 {
1373         struct virtio_net *dev = *pdev;
1374         uint64_t features, protocol_features;
1375
1376         rte_vhost_driver_get_features(dev->ifname, &features);
1377         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1378
1379         /*
1380          * REPLY_ACK protocol feature is only mandatory for now
1381          * for IOMMU feature. If IOMMU is explicitly disabled by the
1382          * application, disable also REPLY_ACK feature for older buggy
1383          * Qemu versions (from v2.7.0 to v2.9.0).
1384          */
1385         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1386                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1387
1388         msg->payload.u64 = protocol_features;
1389         msg->size = sizeof(msg->payload.u64);
1390         msg->fd_num = 0;
1391
1392         return RTE_VHOST_MSG_RESULT_REPLY;
1393 }
1394
1395 static int
1396 vhost_user_set_protocol_features(struct virtio_net **pdev,
1397                         struct VhostUserMsg *msg,
1398                         int main_fd __rte_unused)
1399 {
1400         struct virtio_net *dev = *pdev;
1401         uint64_t protocol_features = msg->payload.u64;
1402         uint64_t slave_protocol_features = 0;
1403
1404         rte_vhost_driver_get_protocol_features(dev->ifname,
1405                         &slave_protocol_features);
1406         if (protocol_features & ~slave_protocol_features) {
1407                 RTE_LOG(ERR, VHOST_CONFIG,
1408                         "(%d) received invalid protocol features.\n",
1409                         dev->vid);
1410                 return RTE_VHOST_MSG_RESULT_ERR;
1411         }
1412
1413         dev->protocol_features = protocol_features;
1414         RTE_LOG(INFO, VHOST_CONFIG,
1415                 "negotiated Vhost-user protocol features: 0x%" PRIx64 "\n",
1416                 dev->protocol_features);
1417
1418         return RTE_VHOST_MSG_RESULT_OK;
1419 }
1420
1421 static int
1422 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1423                         int main_fd __rte_unused)
1424 {
1425         struct virtio_net *dev = *pdev;
1426         int fd = msg->fds[0];
1427         uint64_t size, off;
1428         void *addr;
1429
1430         if (fd < 0) {
1431                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1432                 return RTE_VHOST_MSG_RESULT_ERR;
1433         }
1434
1435         if (msg->size != sizeof(VhostUserLog)) {
1436                 RTE_LOG(ERR, VHOST_CONFIG,
1437                         "invalid log base msg size: %"PRId32" != %d\n",
1438                         msg->size, (int)sizeof(VhostUserLog));
1439                 return RTE_VHOST_MSG_RESULT_ERR;
1440         }
1441
1442         size = msg->payload.log.mmap_size;
1443         off  = msg->payload.log.mmap_offset;
1444
1445         /* Don't allow mmap_offset to point outside the mmap region */
1446         if (off > size) {
1447                 RTE_LOG(ERR, VHOST_CONFIG,
1448                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1449                         off, size);
1450                 return RTE_VHOST_MSG_RESULT_ERR;
1451         }
1452
1453         RTE_LOG(INFO, VHOST_CONFIG,
1454                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1455                 size, off);
1456
1457         /*
1458          * mmap from 0 to workaround a hugepage mmap bug: mmap will
1459          * fail when offset is not page size aligned.
1460          */
1461         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1462         close(fd);
1463         if (addr == MAP_FAILED) {
1464                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1465                 return RTE_VHOST_MSG_RESULT_ERR;
1466         }
1467
1468         /*
1469          * Free previously mapped log memory on occasionally
1470          * multiple VHOST_USER_SET_LOG_BASE.
1471          */
1472         if (dev->log_addr) {
1473                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1474         }
1475         dev->log_addr = (uint64_t)(uintptr_t)addr;
1476         dev->log_base = dev->log_addr + off;
1477         dev->log_size = size;
1478
1479         /*
1480          * The spec is not clear about it (yet), but QEMU doesn't expect
1481          * any payload in the reply.
1482          */
1483         msg->size = 0;
1484         msg->fd_num = 0;
1485
1486         return RTE_VHOST_MSG_RESULT_REPLY;
1487 }
1488
1489 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1490                         struct VhostUserMsg *msg,
1491                         int main_fd __rte_unused)
1492 {
1493         close(msg->fds[0]);
1494         RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1495
1496         return RTE_VHOST_MSG_RESULT_OK;
1497 }
1498
1499 /*
1500  * An rarp packet is constructed and broadcasted to notify switches about
1501  * the new location of the migrated VM, so that packets from outside will
1502  * not be lost after migration.
1503  *
1504  * However, we don't actually "send" a rarp packet here, instead, we set
1505  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1506  */
1507 static int
1508 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
1509                         int main_fd __rte_unused)
1510 {
1511         struct virtio_net *dev = *pdev;
1512         uint8_t *mac = (uint8_t *)&msg->payload.u64;
1513         struct rte_vdpa_device *vdpa_dev;
1514         int did = -1;
1515
1516         RTE_LOG(DEBUG, VHOST_CONFIG,
1517                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1518                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1519         memcpy(dev->mac.addr_bytes, mac, 6);
1520
1521         /*
1522          * Set the flag to inject a RARP broadcast packet at
1523          * rte_vhost_dequeue_burst().
1524          *
1525          * rte_smp_wmb() is for making sure the mac is copied
1526          * before the flag is set.
1527          */
1528         rte_smp_wmb();
1529         rte_atomic16_set(&dev->broadcast_rarp, 1);
1530         did = dev->vdpa_dev_id;
1531         vdpa_dev = rte_vdpa_get_device(did);
1532         if (vdpa_dev && vdpa_dev->ops->migration_done)
1533                 vdpa_dev->ops->migration_done(dev->vid);
1534
1535         return RTE_VHOST_MSG_RESULT_OK;
1536 }
1537
1538 static int
1539 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
1540                         int main_fd __rte_unused)
1541 {
1542         struct virtio_net *dev = *pdev;
1543         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1544                         msg->payload.u64 > VIRTIO_MAX_MTU) {
1545                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1546                                 msg->payload.u64);
1547
1548                 return RTE_VHOST_MSG_RESULT_ERR;
1549         }
1550
1551         dev->mtu = msg->payload.u64;
1552
1553         return RTE_VHOST_MSG_RESULT_OK;
1554 }
1555
1556 static int
1557 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
1558                         int main_fd __rte_unused)
1559 {
1560         struct virtio_net *dev = *pdev;
1561         int fd = msg->fds[0];
1562
1563         if (fd < 0) {
1564                 RTE_LOG(ERR, VHOST_CONFIG,
1565                                 "Invalid file descriptor for slave channel (%d)\n",
1566                                 fd);
1567                 return RTE_VHOST_MSG_RESULT_ERR;
1568         }
1569
1570         if (dev->slave_req_fd >= 0)
1571                 close(dev->slave_req_fd);
1572
1573         dev->slave_req_fd = fd;
1574
1575         return RTE_VHOST_MSG_RESULT_OK;
1576 }
1577
1578 static int
1579 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1580 {
1581         struct vhost_vring_addr *ra;
1582         uint64_t start, end;
1583
1584         start = imsg->iova;
1585         end = start + imsg->size;
1586
1587         ra = &vq->ring_addrs;
1588         if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1589                 return 1;
1590         if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1591                 return 1;
1592         if (ra->used_user_addr >= start && ra->used_user_addr < end)
1593                 return 1;
1594
1595         return 0;
1596 }
1597
1598 static int
1599 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1600                                 struct vhost_iotlb_msg *imsg)
1601 {
1602         uint64_t istart, iend, vstart, vend;
1603
1604         istart = imsg->iova;
1605         iend = istart + imsg->size - 1;
1606
1607         vstart = (uintptr_t)vq->desc;
1608         vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1609         if (vstart <= iend && istart <= vend)
1610                 return 1;
1611
1612         vstart = (uintptr_t)vq->avail;
1613         vend = vstart + sizeof(struct vring_avail);
1614         vend += sizeof(uint16_t) * vq->size - 1;
1615         if (vstart <= iend && istart <= vend)
1616                 return 1;
1617
1618         vstart = (uintptr_t)vq->used;
1619         vend = vstart + sizeof(struct vring_used);
1620         vend += sizeof(struct vring_used_elem) * vq->size - 1;
1621         if (vstart <= iend && istart <= vend)
1622                 return 1;
1623
1624         return 0;
1625 }
1626
1627 static int
1628 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
1629                         int main_fd __rte_unused)
1630 {
1631         struct virtio_net *dev = *pdev;
1632         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1633         uint16_t i;
1634         uint64_t vva, len;
1635
1636         switch (imsg->type) {
1637         case VHOST_IOTLB_UPDATE:
1638                 len = imsg->size;
1639                 vva = qva_to_vva(dev, imsg->uaddr, &len);
1640                 if (!vva)
1641                         return RTE_VHOST_MSG_RESULT_ERR;
1642
1643                 for (i = 0; i < dev->nr_vring; i++) {
1644                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1645
1646                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1647                                         len, imsg->perm);
1648
1649                         if (is_vring_iotlb_update(vq, imsg))
1650                                 *pdev = dev = translate_ring_addresses(dev, i);
1651                 }
1652                 break;
1653         case VHOST_IOTLB_INVALIDATE:
1654                 for (i = 0; i < dev->nr_vring; i++) {
1655                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1656
1657                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
1658                                         imsg->size);
1659
1660                         if (is_vring_iotlb_invalidate(vq, imsg))
1661                                 vring_invalidate(dev, vq);
1662                 }
1663                 break;
1664         default:
1665                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1666                                 imsg->type);
1667                 return RTE_VHOST_MSG_RESULT_ERR;
1668         }
1669
1670         return RTE_VHOST_MSG_RESULT_OK;
1671 }
1672
1673 static int
1674 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
1675                         struct VhostUserMsg *msg,
1676                         int main_fd __rte_unused)
1677 {
1678         struct virtio_net *dev = *pdev;
1679 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1680         struct uffdio_api api_struct;
1681
1682         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1683
1684         if (dev->postcopy_ufd == -1) {
1685                 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
1686                         strerror(errno));
1687                 return RTE_VHOST_MSG_RESULT_ERR;
1688         }
1689         api_struct.api = UFFD_API;
1690         api_struct.features = 0;
1691         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
1692                 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
1693                         strerror(errno));
1694                 close(dev->postcopy_ufd);
1695                 dev->postcopy_ufd = -1;
1696                 return RTE_VHOST_MSG_RESULT_ERR;
1697         }
1698         msg->fds[0] = dev->postcopy_ufd;
1699         msg->fd_num = 1;
1700
1701         return RTE_VHOST_MSG_RESULT_REPLY;
1702 #else
1703         dev->postcopy_ufd = -1;
1704         msg->fd_num = 0;
1705
1706         return RTE_VHOST_MSG_RESULT_ERR;
1707 #endif
1708 }
1709
1710 static int
1711 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
1712                         struct VhostUserMsg *msg __rte_unused,
1713                         int main_fd __rte_unused)
1714 {
1715         struct virtio_net *dev = *pdev;
1716
1717         if (dev->mem && dev->mem->nregions) {
1718                 RTE_LOG(ERR, VHOST_CONFIG,
1719                         "Regions already registered at postcopy-listen\n");
1720                 return RTE_VHOST_MSG_RESULT_ERR;
1721         }
1722         dev->postcopy_listening = 1;
1723
1724         return RTE_VHOST_MSG_RESULT_OK;
1725 }
1726
1727 static int
1728 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
1729                         int main_fd __rte_unused)
1730 {
1731         struct virtio_net *dev = *pdev;
1732
1733         dev->postcopy_listening = 0;
1734         if (dev->postcopy_ufd >= 0) {
1735                 close(dev->postcopy_ufd);
1736                 dev->postcopy_ufd = -1;
1737         }
1738
1739         msg->payload.u64 = 0;
1740         msg->size = sizeof(msg->payload.u64);
1741         msg->fd_num = 0;
1742
1743         return RTE_VHOST_MSG_RESULT_REPLY;
1744 }
1745
1746 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
1747                                         struct VhostUserMsg *msg,
1748                                         int main_fd);
1749 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
1750         [VHOST_USER_NONE] = NULL,
1751         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
1752         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
1753         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
1754         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
1755         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
1756         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
1757         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
1758         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
1759         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
1760         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
1761         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
1762         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
1763         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
1764         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
1765         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
1766         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
1767         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
1768         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
1769         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
1770         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
1771         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
1772         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
1773         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
1774         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
1775         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
1776 };
1777
1778
1779 /* return bytes# of read on success or negative val on failure. */
1780 static int
1781 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1782 {
1783         int ret;
1784
1785         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1786                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
1787         if (ret <= 0)
1788                 return ret;
1789
1790         if (msg->size) {
1791                 if (msg->size > sizeof(msg->payload)) {
1792                         RTE_LOG(ERR, VHOST_CONFIG,
1793                                 "invalid msg size: %d\n", msg->size);
1794                         return -1;
1795                 }
1796                 ret = read(sockfd, &msg->payload, msg->size);
1797                 if (ret <= 0)
1798                         return ret;
1799                 if (ret != (int)msg->size) {
1800                         RTE_LOG(ERR, VHOST_CONFIG,
1801                                 "read control message failed\n");
1802                         return -1;
1803                 }
1804         }
1805
1806         return ret;
1807 }
1808
1809 static int
1810 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1811 {
1812         if (!msg)
1813                 return 0;
1814
1815         return send_fd_message(sockfd, (char *)msg,
1816                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
1817 }
1818
1819 static int
1820 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1821 {
1822         if (!msg)
1823                 return 0;
1824
1825         msg->flags &= ~VHOST_USER_VERSION_MASK;
1826         msg->flags &= ~VHOST_USER_NEED_REPLY;
1827         msg->flags |= VHOST_USER_VERSION;
1828         msg->flags |= VHOST_USER_REPLY_MASK;
1829
1830         return send_vhost_message(sockfd, msg);
1831 }
1832
1833 static int
1834 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
1835 {
1836         int ret;
1837
1838         if (msg->flags & VHOST_USER_NEED_REPLY)
1839                 rte_spinlock_lock(&dev->slave_req_lock);
1840
1841         ret = send_vhost_message(dev->slave_req_fd, msg);
1842         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1843                 rte_spinlock_unlock(&dev->slave_req_lock);
1844
1845         return ret;
1846 }
1847
1848 /*
1849  * Allocate a queue pair if it hasn't been allocated yet
1850  */
1851 static int
1852 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1853                         struct VhostUserMsg *msg)
1854 {
1855         uint16_t vring_idx;
1856
1857         switch (msg->request.master) {
1858         case VHOST_USER_SET_VRING_KICK:
1859         case VHOST_USER_SET_VRING_CALL:
1860         case VHOST_USER_SET_VRING_ERR:
1861                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1862                 break;
1863         case VHOST_USER_SET_VRING_NUM:
1864         case VHOST_USER_SET_VRING_BASE:
1865         case VHOST_USER_SET_VRING_ENABLE:
1866                 vring_idx = msg->payload.state.index;
1867                 break;
1868         case VHOST_USER_SET_VRING_ADDR:
1869                 vring_idx = msg->payload.addr.index;
1870                 break;
1871         default:
1872                 return 0;
1873         }
1874
1875         if (vring_idx >= VHOST_MAX_VRING) {
1876                 RTE_LOG(ERR, VHOST_CONFIG,
1877                         "invalid vring index: %u\n", vring_idx);
1878                 return -1;
1879         }
1880
1881         if (dev->virtqueue[vring_idx])
1882                 return 0;
1883
1884         return alloc_vring_queue(dev, vring_idx);
1885 }
1886
1887 static void
1888 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1889 {
1890         unsigned int i = 0;
1891         unsigned int vq_num = 0;
1892
1893         while (vq_num < dev->nr_vring) {
1894                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1895
1896                 if (vq) {
1897                         rte_spinlock_lock(&vq->access_lock);
1898                         vq_num++;
1899                 }
1900                 i++;
1901         }
1902 }
1903
1904 static void
1905 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1906 {
1907         unsigned int i = 0;
1908         unsigned int vq_num = 0;
1909
1910         while (vq_num < dev->nr_vring) {
1911                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1912
1913                 if (vq) {
1914                         rte_spinlock_unlock(&vq->access_lock);
1915                         vq_num++;
1916                 }
1917                 i++;
1918         }
1919 }
1920
1921 int
1922 vhost_user_msg_handler(int vid, int fd)
1923 {
1924         struct virtio_net *dev;
1925         struct VhostUserMsg msg;
1926         struct rte_vdpa_device *vdpa_dev;
1927         int did = -1;
1928         int ret;
1929         int unlock_required = 0;
1930         bool handled;
1931         int request;
1932
1933         dev = get_device(vid);
1934         if (dev == NULL)
1935                 return -1;
1936
1937         if (!dev->notify_ops) {
1938                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1939                 if (!dev->notify_ops) {
1940                         RTE_LOG(ERR, VHOST_CONFIG,
1941                                 "failed to get callback ops for driver %s\n",
1942                                 dev->ifname);
1943                         return -1;
1944                 }
1945         }
1946
1947         ret = read_vhost_message(fd, &msg);
1948         if (ret <= 0) {
1949                 if (ret < 0)
1950                         RTE_LOG(ERR, VHOST_CONFIG,
1951                                 "vhost read message failed\n");
1952                 else
1953                         RTE_LOG(INFO, VHOST_CONFIG,
1954                                 "vhost peer closed\n");
1955
1956                 return -1;
1957         }
1958
1959         ret = 0;
1960         request = msg.request.master;
1961         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
1962                         vhost_message_str[request]) {
1963                 if (request != VHOST_USER_IOTLB_MSG)
1964                         RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1965                                 vhost_message_str[request]);
1966                 else
1967                         RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1968                                 vhost_message_str[request]);
1969         } else {
1970                 RTE_LOG(DEBUG, VHOST_CONFIG, "External request %d\n", request);
1971         }
1972
1973         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1974         if (ret < 0) {
1975                 RTE_LOG(ERR, VHOST_CONFIG,
1976                         "failed to alloc queue\n");
1977                 return -1;
1978         }
1979
1980         /*
1981          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1982          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1983          * and device is destroyed. destroy_device waits for queues to be
1984          * inactive, so it is safe. Otherwise taking the access_lock
1985          * would cause a dead lock.
1986          */
1987         switch (request) {
1988         case VHOST_USER_SET_FEATURES:
1989         case VHOST_USER_SET_PROTOCOL_FEATURES:
1990         case VHOST_USER_SET_OWNER:
1991         case VHOST_USER_SET_MEM_TABLE:
1992         case VHOST_USER_SET_LOG_BASE:
1993         case VHOST_USER_SET_LOG_FD:
1994         case VHOST_USER_SET_VRING_NUM:
1995         case VHOST_USER_SET_VRING_ADDR:
1996         case VHOST_USER_SET_VRING_BASE:
1997         case VHOST_USER_SET_VRING_KICK:
1998         case VHOST_USER_SET_VRING_CALL:
1999         case VHOST_USER_SET_VRING_ERR:
2000         case VHOST_USER_SET_VRING_ENABLE:
2001         case VHOST_USER_SEND_RARP:
2002         case VHOST_USER_NET_SET_MTU:
2003         case VHOST_USER_SET_SLAVE_REQ_FD:
2004                 vhost_user_lock_all_queue_pairs(dev);
2005                 unlock_required = 1;
2006                 break;
2007         default:
2008                 break;
2009
2010         }
2011
2012         handled = false;
2013         if (dev->extern_ops.pre_msg_handle) {
2014                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
2015                                 (void *)&msg);
2016                 switch (ret) {
2017                 case RTE_VHOST_MSG_RESULT_REPLY:
2018                         send_vhost_reply(fd, &msg);
2019                         /* Fall-through */
2020                 case RTE_VHOST_MSG_RESULT_ERR:
2021                 case RTE_VHOST_MSG_RESULT_OK:
2022                         handled = true;
2023                         goto skip_to_post_handle;
2024                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2025                 default:
2026                         break;
2027                 }
2028         }
2029
2030         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
2031                 if (!vhost_message_handlers[request])
2032                         goto skip_to_post_handle;
2033                 ret = vhost_message_handlers[request](&dev, &msg, fd);
2034
2035                 switch (ret) {
2036                 case RTE_VHOST_MSG_RESULT_ERR:
2037                         RTE_LOG(ERR, VHOST_CONFIG,
2038                                 "Processing %s failed.\n",
2039                                 vhost_message_str[request]);
2040                         handled = true;
2041                         break;
2042                 case RTE_VHOST_MSG_RESULT_OK:
2043                         RTE_LOG(DEBUG, VHOST_CONFIG,
2044                                 "Processing %s succeeded.\n",
2045                                 vhost_message_str[request]);
2046                         handled = true;
2047                         break;
2048                 case RTE_VHOST_MSG_RESULT_REPLY:
2049                         RTE_LOG(DEBUG, VHOST_CONFIG,
2050                                 "Processing %s succeeded and needs reply.\n",
2051                                 vhost_message_str[request]);
2052                         send_vhost_reply(fd, &msg);
2053                         handled = true;
2054                         break;
2055                 default:
2056                         break;
2057                 }
2058         }
2059
2060 skip_to_post_handle:
2061         if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2062                         dev->extern_ops.post_msg_handle) {
2063                 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
2064                                 (void *)&msg);
2065                 switch (ret) {
2066                 case RTE_VHOST_MSG_RESULT_REPLY:
2067                         send_vhost_reply(fd, &msg);
2068                         /* Fall-through */
2069                 case RTE_VHOST_MSG_RESULT_ERR:
2070                 case RTE_VHOST_MSG_RESULT_OK:
2071                         handled = true;
2072                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2073                 default:
2074                         break;
2075                 }
2076         }
2077
2078         if (unlock_required)
2079                 vhost_user_unlock_all_queue_pairs(dev);
2080
2081         /* If message was not handled at this stage, treat it as an error */
2082         if (!handled) {
2083                 RTE_LOG(ERR, VHOST_CONFIG,
2084                         "vhost message (req: %d) was not handled.\n", request);
2085                 ret = RTE_VHOST_MSG_RESULT_ERR;
2086         }
2087
2088         /*
2089          * If the request required a reply that was already sent,
2090          * this optional reply-ack won't be sent as the
2091          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2092          */
2093         if (msg.flags & VHOST_USER_NEED_REPLY) {
2094                 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2095                 msg.size = sizeof(msg.payload.u64);
2096                 msg.fd_num = 0;
2097                 send_vhost_reply(fd, &msg);
2098         } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2099                 RTE_LOG(ERR, VHOST_CONFIG,
2100                         "vhost message handling failed.\n");
2101                 return -1;
2102         }
2103
2104         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2105                 dev->flags |= VIRTIO_DEV_READY;
2106
2107                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2108                         if (dev->dequeue_zero_copy) {
2109                                 RTE_LOG(INFO, VHOST_CONFIG,
2110                                                 "dequeue zero copy is enabled\n");
2111                         }
2112
2113                         if (dev->notify_ops->new_device(dev->vid) == 0)
2114                                 dev->flags |= VIRTIO_DEV_RUNNING;
2115                 }
2116         }
2117
2118         did = dev->vdpa_dev_id;
2119         vdpa_dev = rte_vdpa_get_device(did);
2120         if (vdpa_dev && virtio_is_ready(dev) &&
2121                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2122                         msg.request.master == VHOST_USER_SET_VRING_CALL) {
2123                 if (vdpa_dev->ops->dev_conf)
2124                         vdpa_dev->ops->dev_conf(dev->vid);
2125                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2126         }
2127
2128         return 0;
2129 }
2130
2131 static int process_slave_message_reply(struct virtio_net *dev,
2132                                        const struct VhostUserMsg *msg)
2133 {
2134         struct VhostUserMsg msg_reply;
2135         int ret;
2136
2137         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2138                 return 0;
2139
2140         if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
2141                 ret = -1;
2142                 goto out;
2143         }
2144
2145         if (msg_reply.request.slave != msg->request.slave) {
2146                 RTE_LOG(ERR, VHOST_CONFIG,
2147                         "Received unexpected msg type (%u), expected %u\n",
2148                         msg_reply.request.slave, msg->request.slave);
2149                 ret = -1;
2150                 goto out;
2151         }
2152
2153         ret = msg_reply.payload.u64 ? -1 : 0;
2154
2155 out:
2156         rte_spinlock_unlock(&dev->slave_req_lock);
2157         return ret;
2158 }
2159
2160 int
2161 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2162 {
2163         int ret;
2164         struct VhostUserMsg msg = {
2165                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2166                 .flags = VHOST_USER_VERSION,
2167                 .size = sizeof(msg.payload.iotlb),
2168                 .payload.iotlb = {
2169                         .iova = iova,
2170                         .perm = perm,
2171                         .type = VHOST_IOTLB_MISS,
2172                 },
2173         };
2174
2175         ret = send_vhost_message(dev->slave_req_fd, &msg);
2176         if (ret < 0) {
2177                 RTE_LOG(ERR, VHOST_CONFIG,
2178                                 "Failed to send IOTLB miss message (%d)\n",
2179                                 ret);
2180                 return ret;
2181         }
2182
2183         return 0;
2184 }
2185
2186 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2187                                                     int index, int fd,
2188                                                     uint64_t offset,
2189                                                     uint64_t size)
2190 {
2191         int ret;
2192         struct VhostUserMsg msg = {
2193                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2194                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2195                 .size = sizeof(msg.payload.area),
2196                 .payload.area = {
2197                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
2198                         .size = size,
2199                         .offset = offset,
2200                 },
2201         };
2202
2203         if (fd < 0)
2204                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2205         else {
2206                 msg.fds[0] = fd;
2207                 msg.fd_num = 1;
2208         }
2209
2210         ret = send_vhost_slave_message(dev, &msg);
2211         if (ret < 0) {
2212                 RTE_LOG(ERR, VHOST_CONFIG,
2213                         "Failed to set host notifier (%d)\n", ret);
2214                 return ret;
2215         }
2216
2217         return process_slave_message_reply(dev, &msg);
2218 }
2219
2220 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2221 {
2222         struct virtio_net *dev;
2223         struct rte_vdpa_device *vdpa_dev;
2224         int vfio_device_fd, did, ret = 0;
2225         uint64_t offset, size;
2226         unsigned int i;
2227
2228         dev = get_device(vid);
2229         if (!dev)
2230                 return -ENODEV;
2231
2232         did = dev->vdpa_dev_id;
2233         if (did < 0)
2234                 return -EINVAL;
2235
2236         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2237             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2238             !(dev->protocol_features &
2239                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2240             !(dev->protocol_features &
2241                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2242             !(dev->protocol_features &
2243                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2244                 return -ENOTSUP;
2245
2246         vdpa_dev = rte_vdpa_get_device(did);
2247         if (!vdpa_dev)
2248                 return -ENODEV;
2249
2250         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2251         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2252
2253         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2254         if (vfio_device_fd < 0)
2255                 return -ENOTSUP;
2256
2257         if (enable) {
2258                 for (i = 0; i < dev->nr_vring; i++) {
2259                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2260                                         &size) < 0) {
2261                                 ret = -ENOTSUP;
2262                                 goto disable;
2263                         }
2264
2265                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2266                                         vfio_device_fd, offset, size) < 0) {
2267                                 ret = -EFAULT;
2268                                 goto disable;
2269                         }
2270                 }
2271         } else {
2272 disable:
2273                 for (i = 0; i < dev->nr_vring; i++) {
2274                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
2275                                         0, 0);
2276                 }
2277         }
2278
2279         return ret;
2280 }