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                 vq->access_ok = 1;
 626                 return dev;
 627         }
 628
 629         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 630         if (vq->desc && vq->avail && vq->used)
 631                 return dev;
 632
 633         len = sizeof(struct vring_desc) * vq->size;
 634         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 635                         vq, addr->desc_user_addr, &len);
 636         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 637                 RTE_LOG(DEBUG, VHOST_CONFIG,
 638                         "(%d) failed to map desc ring.\n",
 639                         dev->vid);
 640                 return dev;
 641         }
 642
 643         dev = numa_realloc(dev, vq_index);
 644         vq = dev->virtqueue[vq_index];
 645         addr = &vq->ring_addrs;
 646
 647         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 648         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 649                 len += sizeof(uint16_t);
 650         expected_len = len;
 651         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 652                         vq, addr->avail_user_addr, &len);
 653         if (vq->avail == 0 || len != expected_len) {
 654                 RTE_LOG(DEBUG, VHOST_CONFIG,
 655                         "(%d) failed to map avail ring.\n",
 656                         dev->vid);
 657                 return dev;
 658         }
 659
 660         len = sizeof(struct vring_used) +
 661                 sizeof(struct vring_used_elem) * vq->size;
 662         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 663                 len += sizeof(uint16_t);
 664         expected_len = len;
 665         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 666                         vq, addr->used_user_addr, &len);
 667         if (vq->used == 0 || len != expected_len) {
 668                 RTE_LOG(DEBUG, VHOST_CONFIG,
 669                         "(%d) failed to map used ring.\n",
 670                         dev->vid);
 671                 return dev;
 672         }
 673
 674         if (vq->last_used_idx != vq->used->idx) {
 675                 RTE_LOG(WARNING, VHOST_CONFIG,
 676                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 677                         "some packets maybe resent for Tx and dropped for Rx\n",
 678                         vq->last_used_idx, vq->used->idx);
 679                 vq->last_used_idx  = vq->used->idx;
 680                 vq->last_avail_idx = vq->used->idx;
 681         }
 682
 683         vq->log_guest_addr = addr->log_guest_addr;
 684         vq->access_ok = 1;
 685
 686         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
 687                         dev->vid, vq->desc);
 688         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
 689                         dev->vid, vq->avail);
 690         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
 691                         dev->vid, vq->used);
 692         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
 693                         dev->vid, vq->log_guest_addr);
 694
 695         return dev;
 696 }
 697
 698 /*
 699  * The virtio device sends us the desc, used and avail ring addresses.
 700  * This function then converts these to our address space.
 701  */
 702 static int
 703 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 704                         int main_fd __rte_unused)
 705 {
 706         struct virtio_net *dev = *pdev;
 707         struct vhost_virtqueue *vq;
 708         struct vhost_vring_addr *addr = &msg->payload.addr;
 709         bool access_ok;
 710
 711         if (dev->mem == NULL)
 712                 return RTE_VHOST_MSG_RESULT_ERR;
 713
 714         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 715         vq = dev->virtqueue[msg->payload.addr.index];
 716
 717         access_ok = vq->access_ok;
 718
 719         /*
 720          * Rings addresses should not be interpreted as long as the ring is not
 721          * started and enabled
 722          */
 723         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 724
 725         vring_invalidate(dev, vq);
 726
 727         if ((vq->enabled && (dev->features &
 728                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) ||
 729                         access_ok) {
 730                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 731                 if (!dev)
 732                         return RTE_VHOST_MSG_RESULT_ERR;
 733
 734                 *pdev = dev;
 735         }
 736
 737         return RTE_VHOST_MSG_RESULT_OK;
 738 }
 739
 740 /*
 741  * The virtio device sends us the available ring last used index.
 742  */
 743 static int
 744 vhost_user_set_vring_base(struct virtio_net **pdev,
 745                         struct VhostUserMsg *msg,
 746                         int main_fd __rte_unused)
 747 {
 748         struct virtio_net *dev = *pdev;
 749         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 750         uint64_t val = msg->payload.state.num;
 751
 752         if (vq_is_packed(dev)) {
 753                 /*
 754                  * Bit[0:14]: avail index
 755                  * Bit[15]: avail wrap counter
 756                  */
 757                 vq->last_avail_idx = val & 0x7fff;
 758                 vq->avail_wrap_counter = !!(val & (0x1 << 15));
 759                 /*
 760                  * Set used index to same value as available one, as
 761                  * their values should be the same since ring processing
 762                  * was stopped at get time.
 763                  */
 764                 vq->last_used_idx = vq->last_avail_idx;
 765                 vq->used_wrap_counter = vq->avail_wrap_counter;
 766         } else {
 767                 vq->last_used_idx = msg->payload.state.num;
 768                 vq->last_avail_idx = msg->payload.state.num;
 769         }
 770
 771         return RTE_VHOST_MSG_RESULT_OK;
 772 }
 773
 774 static int
 775 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 776                    uint64_t host_phys_addr, uint64_t size)
 777 {
 778         struct guest_page *page, *last_page;
 779         struct guest_page *old_pages;
 780
 781         if (dev->nr_guest_pages == dev->max_guest_pages) {
 782                 dev->max_guest_pages *= 2;
 783                 old_pages = dev->guest_pages;
 784                 dev->guest_pages = realloc(dev->guest_pages,
 785                                         dev->max_guest_pages * sizeof(*page));
 786                 if (!dev->guest_pages) {
 787                         RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
 788                         free(old_pages);
 789                         return -1;
 790                 }
 791         }
 792
 793         if (dev->nr_guest_pages > 0) {
 794                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 795                 /* merge if the two pages are continuous */
 796                 if (host_phys_addr == last_page->host_phys_addr +
 797                                       last_page->size) {
 798                         last_page->size += size;
 799                         return 0;
 800                 }
 801         }
 802
 803         page = &dev->guest_pages[dev->nr_guest_pages++];
 804         page->guest_phys_addr = guest_phys_addr;
 805         page->host_phys_addr  = host_phys_addr;
 806         page->size = size;
 807
 808         return 0;
 809 }
 810
 811 static int
 812 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 813                 uint64_t page_size)
 814 {
 815         uint64_t reg_size = reg->size;
 816         uint64_t host_user_addr  = reg->host_user_addr;
 817         uint64_t guest_phys_addr = reg->guest_phys_addr;
 818         uint64_t host_phys_addr;
 819         uint64_t size;
 820
 821         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 822         size = page_size - (guest_phys_addr & (page_size - 1));
 823         size = RTE_MIN(size, reg_size);
 824
 825         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 826                 return -1;
 827
 828         host_user_addr  += size;
 829         guest_phys_addr += size;
 830         reg_size -= size;
 831
 832         while (reg_size > 0) {
 833                 size = RTE_MIN(reg_size, page_size);
 834                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 835                                                   host_user_addr);
 836                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 837                                 size) < 0)
 838                         return -1;
 839
 840                 host_user_addr  += size;
 841                 guest_phys_addr += size;
 842                 reg_size -= size;
 843         }
 844
 845         return 0;
 846 }
 847
 848 #ifdef RTE_LIBRTE_VHOST_DEBUG
 849 /* TODO: enable it only in debug mode? */
 850 static void
 851 dump_guest_pages(struct virtio_net *dev)
 852 {
 853         uint32_t i;
 854         struct guest_page *page;
 855
 856         for (i = 0; i < dev->nr_guest_pages; i++) {
 857                 page = &dev->guest_pages[i];
 858
 859                 RTE_LOG(INFO, VHOST_CONFIG,
 860                         "guest physical page region %u\n"
 861                         "\t guest_phys_addr: %" PRIx64 "\n"
 862                         "\t host_phys_addr : %" PRIx64 "\n"
 863                         "\t size           : %" PRIx64 "\n",
 864                         i,
 865                         page->guest_phys_addr,
 866                         page->host_phys_addr,
 867                         page->size);
 868         }
 869 }
 870 #else
 871 #define dump_guest_pages(dev)
 872 #endif
 873
 874 static bool
 875 vhost_memory_changed(struct VhostUserMemory *new,
 876                      struct rte_vhost_memory *old)
 877 {
 878         uint32_t i;
 879
 880         if (new->nregions != old->nregions)
 881                 return true;
 882
 883         for (i = 0; i < new->nregions; ++i) {
 884                 VhostUserMemoryRegion *new_r = &new->regions[i];
 885                 struct rte_vhost_mem_region *old_r = &old->regions[i];
 886
 887                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
 888                         return true;
 889                 if (new_r->memory_size != old_r->size)
 890                         return true;
 891                 if (new_r->userspace_addr != old_r->guest_user_addr)
 892                         return true;
 893         }
 894
 895         return false;
 896 }
 897
 898 static int
 899 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
 900                         int main_fd)
 901 {
 902         struct virtio_net *dev = *pdev;
 903         struct VhostUserMemory *memory = &msg->payload.memory;
 904         struct rte_vhost_mem_region *reg;
 905         void *mmap_addr;
 906         uint64_t mmap_size;
 907         uint64_t mmap_offset;
 908         uint64_t alignment;
 909         uint32_t i;
 910         int populate;
 911         int fd;
 912
 913         if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
 914                 RTE_LOG(ERR, VHOST_CONFIG,
 915                         "too many memory regions (%u)\n", memory->nregions);
 916                 return RTE_VHOST_MSG_RESULT_ERR;
 917         }
 918
 919         if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
 920                 RTE_LOG(INFO, VHOST_CONFIG,
 921                         "(%d) memory regions not changed\n", dev->vid);
 922
 923                 for (i = 0; i < memory->nregions; i++)
 924                         close(msg->fds[i]);
 925
 926                 return RTE_VHOST_MSG_RESULT_OK;
 927         }
 928
 929         if (dev->mem) {
 930                 free_mem_region(dev);
 931                 rte_free(dev->mem);
 932                 dev->mem = NULL;
 933         }
 934
 935         /* Flush IOTLB cache as previous HVAs are now invalid */
 936         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 937                 for (i = 0; i < dev->nr_vring; i++)
 938                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
 939
 940         dev->nr_guest_pages = 0;
 941         if (!dev->guest_pages) {
 942                 dev->max_guest_pages = 8;
 943                 dev->guest_pages = malloc(dev->max_guest_pages *
 944                                                 sizeof(struct guest_page));
 945                 if (dev->guest_pages == NULL) {
 946                         RTE_LOG(ERR, VHOST_CONFIG,
 947                                 "(%d) failed to allocate memory "
 948                                 "for dev->guest_pages\n",
 949                                 dev->vid);
 950                         return RTE_VHOST_MSG_RESULT_ERR;
 951                 }
 952         }
 953
 954         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
 955                 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
 956         if (dev->mem == NULL) {
 957                 RTE_LOG(ERR, VHOST_CONFIG,
 958                         "(%d) failed to allocate memory for dev->mem\n",
 959                         dev->vid);
 960                 return RTE_VHOST_MSG_RESULT_ERR;
 961         }
 962         dev->mem->nregions = memory->nregions;
 963
 964         for (i = 0; i < memory->nregions; i++) {
 965                 fd  = msg->fds[i];
 966                 reg = &dev->mem->regions[i];
 967
 968                 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
 969                 reg->guest_user_addr = memory->regions[i].userspace_addr;
 970                 reg->size            = memory->regions[i].memory_size;
 971                 reg->fd              = fd;
 972
 973                 mmap_offset = memory->regions[i].mmap_offset;
 974
 975                 /* Check for memory_size + mmap_offset overflow */
 976                 if (mmap_offset >= -reg->size) {
 977                         RTE_LOG(ERR, VHOST_CONFIG,
 978                                 "mmap_offset (%#"PRIx64") and memory_size "
 979                                 "(%#"PRIx64") overflow\n",
 980                                 mmap_offset, reg->size);
 981                         goto err_mmap;
 982                 }
 983
 984                 mmap_size = reg->size + mmap_offset;
 985
 986                 /* mmap() without flag of MAP_ANONYMOUS, should be called
 987                  * with length argument aligned with hugepagesz at older
 988                  * longterm version Linux, like 2.6.32 and 3.2.72, or
 989                  * mmap() will fail with EINVAL.
 990                  *
 991                  * to avoid failure, make sure in caller to keep length
 992                  * aligned.
 993                  */
 994                 alignment = get_blk_size(fd);
 995                 if (alignment == (uint64_t)-1) {
 996                         RTE_LOG(ERR, VHOST_CONFIG,
 997                                 "couldn't get hugepage size through fstat\n");
 998                         goto err_mmap;
 999                 }
1000                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
1001
1002                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
1003                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
1004                                  MAP_SHARED | populate, fd, 0);
1005
1006                 if (mmap_addr == MAP_FAILED) {
1007                         RTE_LOG(ERR, VHOST_CONFIG,
1008                                 "mmap region %u failed.\n", i);
1009                         goto err_mmap;
1010                 }
1011
1012                 reg->mmap_addr = mmap_addr;
1013                 reg->mmap_size = mmap_size;
1014                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
1015                                       mmap_offset;
1016
1017                 if (dev->dequeue_zero_copy)
1018                         if (add_guest_pages(dev, reg, alignment) < 0) {
1019                                 RTE_LOG(ERR, VHOST_CONFIG,
1020                                         "adding guest pages to region %u failed.\n",
1021                                         i);
1022                                 goto err_mmap;
1023                         }
1024
1025                 RTE_LOG(INFO, VHOST_CONFIG,
1026                         "guest memory region %u, size: 0x%" PRIx64 "\n"
1027                         "\t guest physical addr: 0x%" PRIx64 "\n"
1028                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
1029                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
1030                         "\t mmap addr : 0x%" PRIx64 "\n"
1031                         "\t mmap size : 0x%" PRIx64 "\n"
1032                         "\t mmap align: 0x%" PRIx64 "\n"
1033                         "\t mmap off  : 0x%" PRIx64 "\n",
1034                         i, reg->size,
1035                         reg->guest_phys_addr,
1036                         reg->guest_user_addr,
1037                         reg->host_user_addr,
1038                         (uint64_t)(uintptr_t)mmap_addr,
1039                         mmap_size,
1040                         alignment,
1041                         mmap_offset);
1042
1043                 if (dev->postcopy_listening) {
1044                         /*
1045                          * We haven't a better way right now than sharing
1046                          * DPDK's virtual address with Qemu, so that Qemu can
1047                          * retrieve the region offset when handling userfaults.
1048                          */
1049                         memory->regions[i].userspace_addr =
1050                                 reg->host_user_addr;
1051                 }
1052         }
1053         if (dev->postcopy_listening) {
1054                 /* Send the addresses back to qemu */
1055                 msg->fd_num = 0;
1056                 send_vhost_reply(main_fd, msg);
1057
1058                 /* Wait for qemu to acknolwedge it's got the addresses
1059                  * we've got to wait before we're allowed to generate faults.
1060                  */
1061                 VhostUserMsg ack_msg;
1062                 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1063                         RTE_LOG(ERR, VHOST_CONFIG,
1064                                 "Failed to read qemu ack on postcopy set-mem-table\n");
1065                         goto err_mmap;
1066                 }
1067                 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1068                         RTE_LOG(ERR, VHOST_CONFIG,
1069                                 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1070                                 ack_msg.request.master);
1071                         goto err_mmap;
1072                 }
1073
1074                 /* Now userfault register and we can use the memory */
1075                 for (i = 0; i < memory->nregions; i++) {
1076 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1077                         reg = &dev->mem->regions[i];
1078                         struct uffdio_register reg_struct;
1079
1080                         /*
1081                          * Let's register all the mmap'ed area to ensure
1082                          * alignment on page boundary.
1083                          */
1084                         reg_struct.range.start =
1085                                 (uint64_t)(uintptr_t)reg->mmap_addr;
1086                         reg_struct.range.len = reg->mmap_size;
1087                         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1088
1089                         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1090                                                 &reg_struct)) {
1091                                 RTE_LOG(ERR, VHOST_CONFIG,
1092                                         "Failed to register ufd for region %d: (ufd = %d) %s\n",
1093                                         i, dev->postcopy_ufd,
1094                                         strerror(errno));
1095                                 goto err_mmap;
1096                         }
1097                         RTE_LOG(INFO, VHOST_CONFIG,
1098                                 "\t userfaultfd registered for range : %llx - %llx\n",
1099                                 reg_struct.range.start,
1100                                 reg_struct.range.start +
1101                                 reg_struct.range.len - 1);
1102 #else
1103                         goto err_mmap;
1104 #endif
1105                 }
1106         }
1107
1108         for (i = 0; i < dev->nr_vring; i++) {
1109                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1110
1111                 if (vq->desc || vq->avail || vq->used) {
1112                         /*
1113                          * If the memory table got updated, the ring addresses
1114                          * need to be translated again as virtual addresses have
1115                          * changed.
1116                          */
1117                         vring_invalidate(dev, vq);
1118
1119                         dev = translate_ring_addresses(dev, i);
1120                         if (!dev) {
1121                                 dev = *pdev;
1122                                 goto err_mmap;
1123                         }
1124
1125                         *pdev = dev;
1126                 }
1127         }
1128
1129         dump_guest_pages(dev);
1130
1131         return RTE_VHOST_MSG_RESULT_OK;
1132
1133 err_mmap:
1134         free_mem_region(dev);
1135         rte_free(dev->mem);
1136         dev->mem = NULL;
1137         return RTE_VHOST_MSG_RESULT_ERR;
1138 }
1139
1140 static bool
1141 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1142 {
1143         bool rings_ok;
1144
1145         if (!vq)
1146                 return false;
1147
1148         if (vq_is_packed(dev))
1149                 rings_ok = !!vq->desc_packed;
1150         else
1151                 rings_ok = vq->desc && vq->avail && vq->used;
1152
1153         return rings_ok &&
1154                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1155                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1156 }
1157
1158 static int
1159 virtio_is_ready(struct virtio_net *dev)
1160 {
1161         struct vhost_virtqueue *vq;
1162         uint32_t i;
1163
1164         if (dev->nr_vring == 0)
1165                 return 0;
1166
1167         for (i = 0; i < dev->nr_vring; i++) {
1168                 vq = dev->virtqueue[i];
1169
1170                 if (!vq_is_ready(dev, vq))
1171                         return 0;
1172         }
1173
1174         RTE_LOG(INFO, VHOST_CONFIG,
1175                 "virtio is now ready for processing.\n");
1176         return 1;
1177 }
1178
1179 static int
1180 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1181                         int main_fd __rte_unused)
1182 {
1183         struct virtio_net *dev = *pdev;
1184         struct vhost_vring_file file;
1185         struct vhost_virtqueue *vq;
1186
1187         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1188         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1189                 file.fd = VIRTIO_INVALID_EVENTFD;
1190         else
1191                 file.fd = msg->fds[0];
1192         RTE_LOG(INFO, VHOST_CONFIG,
1193                 "vring call idx:%d file:%d\n", file.index, file.fd);
1194
1195         vq = dev->virtqueue[file.index];
1196         if (vq->callfd >= 0)
1197                 close(vq->callfd);
1198
1199         vq->callfd = file.fd;
1200
1201         return RTE_VHOST_MSG_RESULT_OK;
1202 }
1203
1204 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1205                         struct VhostUserMsg *msg,
1206                         int main_fd __rte_unused)
1207 {
1208         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1209                 close(msg->fds[0]);
1210         RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1211
1212         return RTE_VHOST_MSG_RESULT_OK;
1213 }
1214
1215 static int
1216 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1217                         int main_fd __rte_unused)
1218 {
1219         struct virtio_net *dev = *pdev;
1220         struct vhost_vring_file file;
1221         struct vhost_virtqueue *vq;
1222
1223         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1224         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1225                 file.fd = VIRTIO_INVALID_EVENTFD;
1226         else
1227                 file.fd = msg->fds[0];
1228         RTE_LOG(INFO, VHOST_CONFIG,
1229                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1230
1231         /* Interpret ring addresses only when ring is started. */
1232         dev = translate_ring_addresses(dev, file.index);
1233         if (!dev)
1234                 return RTE_VHOST_MSG_RESULT_ERR;
1235
1236         *pdev = dev;
1237
1238         vq = dev->virtqueue[file.index];
1239
1240         /*
1241          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1242          * the ring starts already enabled. Otherwise, it is enabled via
1243          * the SET_VRING_ENABLE message.
1244          */
1245         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
1246                 vq->enabled = 1;
1247                 if (dev->notify_ops->vring_state_changed)
1248                         dev->notify_ops->vring_state_changed(
1249                                 dev->vid, file.index, 1);
1250         }
1251
1252         if (vq->kickfd >= 0)
1253                 close(vq->kickfd);
1254         vq->kickfd = file.fd;
1255
1256         return RTE_VHOST_MSG_RESULT_OK;
1257 }
1258
1259 static void
1260 free_zmbufs(struct vhost_virtqueue *vq)
1261 {
1262         drain_zmbuf_list(vq);
1263
1264         rte_free(vq->zmbufs);
1265 }
1266
1267 /*
1268  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1269  */
1270 static int
1271 vhost_user_get_vring_base(struct virtio_net **pdev,
1272                         struct VhostUserMsg *msg,
1273                         int main_fd __rte_unused)
1274 {
1275         struct virtio_net *dev = *pdev;
1276         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1277         uint64_t val;
1278
1279         /* We have to stop the queue (virtio) if it is running. */
1280         vhost_destroy_device_notify(dev);
1281
1282         dev->flags &= ~VIRTIO_DEV_READY;
1283         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1284
1285         /* Here we are safe to get the indexes */
1286         if (vq_is_packed(dev)) {
1287                 /*
1288                  * Bit[0:14]: avail index
1289                  * Bit[15]: avail wrap counter
1290                  */
1291                 val = vq->last_avail_idx & 0x7fff;
1292                 val |= vq->avail_wrap_counter << 15;
1293                 msg->payload.state.num = val;
1294         } else {
1295                 msg->payload.state.num = vq->last_avail_idx;
1296         }
1297
1298         RTE_LOG(INFO, VHOST_CONFIG,
1299                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1300                 msg->payload.state.num);
1301         /*
1302          * Based on current qemu vhost-user implementation, this message is
1303          * sent and only sent in vhost_vring_stop.
1304          * TODO: cleanup the vring, it isn't usable since here.
1305          */
1306         if (vq->kickfd >= 0)
1307                 close(vq->kickfd);
1308
1309         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1310
1311         if (vq->callfd >= 0)
1312                 close(vq->callfd);
1313
1314         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1315
1316         vq->signalled_used_valid = false;
1317
1318         if (dev->dequeue_zero_copy)
1319                 free_zmbufs(vq);
1320         if (vq_is_packed(dev)) {
1321                 rte_free(vq->shadow_used_packed);
1322                 vq->shadow_used_packed = NULL;
1323         } else {
1324                 rte_free(vq->shadow_used_split);
1325                 vq->shadow_used_split = NULL;
1326         }
1327
1328         rte_free(vq->batch_copy_elems);
1329         vq->batch_copy_elems = NULL;
1330
1331         msg->size = sizeof(msg->payload.state);
1332         msg->fd_num = 0;
1333
1334         vring_invalidate(dev, vq);
1335
1336         return RTE_VHOST_MSG_RESULT_REPLY;
1337 }
1338
1339 /*
1340  * when virtio queues are ready to work, qemu will send us to
1341  * enable the virtio queue pair.
1342  */
1343 static int
1344 vhost_user_set_vring_enable(struct virtio_net **pdev,
1345                         struct VhostUserMsg *msg,
1346                         int main_fd __rte_unused)
1347 {
1348         struct virtio_net *dev = *pdev;
1349         int enable = (int)msg->payload.state.num;
1350         int index = (int)msg->payload.state.index;
1351         struct rte_vdpa_device *vdpa_dev;
1352         int did = -1;
1353
1354         RTE_LOG(INFO, VHOST_CONFIG,
1355                 "set queue enable: %d to qp idx: %d\n",
1356                 enable, index);
1357
1358         did = dev->vdpa_dev_id;
1359         vdpa_dev = rte_vdpa_get_device(did);
1360         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1361                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1362
1363         if (dev->notify_ops->vring_state_changed)
1364                 dev->notify_ops->vring_state_changed(dev->vid,
1365                                 index, enable);
1366
1367         /* On disable, rings have to be stopped being processed. */
1368         if (!enable && dev->dequeue_zero_copy)
1369                 drain_zmbuf_list(dev->virtqueue[index]);
1370
1371         dev->virtqueue[index]->enabled = enable;
1372
1373         return RTE_VHOST_MSG_RESULT_OK;
1374 }
1375
1376 static int
1377 vhost_user_get_protocol_features(struct virtio_net **pdev,
1378                         struct VhostUserMsg *msg,
1379                         int main_fd __rte_unused)
1380 {
1381         struct virtio_net *dev = *pdev;
1382         uint64_t features, protocol_features;
1383
1384         rte_vhost_driver_get_features(dev->ifname, &features);
1385         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1386
1387         /*
1388          * REPLY_ACK protocol feature is only mandatory for now
1389          * for IOMMU feature. If IOMMU is explicitly disabled by the
1390          * application, disable also REPLY_ACK feature for older buggy
1391          * Qemu versions (from v2.7.0 to v2.9.0).
1392          */
1393         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1394                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1395
1396         msg->payload.u64 = protocol_features;
1397         msg->size = sizeof(msg->payload.u64);
1398         msg->fd_num = 0;
1399
1400         return RTE_VHOST_MSG_RESULT_REPLY;
1401 }
1402
1403 static int
1404 vhost_user_set_protocol_features(struct virtio_net **pdev,
1405                         struct VhostUserMsg *msg,
1406                         int main_fd __rte_unused)
1407 {
1408         struct virtio_net *dev = *pdev;
1409         uint64_t protocol_features = msg->payload.u64;
1410         uint64_t slave_protocol_features = 0;
1411
1412         rte_vhost_driver_get_protocol_features(dev->ifname,
1413                         &slave_protocol_features);
1414         if (protocol_features & ~slave_protocol_features) {
1415                 RTE_LOG(ERR, VHOST_CONFIG,
1416                         "(%d) received invalid protocol features.\n",
1417                         dev->vid);
1418                 return RTE_VHOST_MSG_RESULT_ERR;
1419         }
1420
1421         dev->protocol_features = protocol_features;
1422         RTE_LOG(INFO, VHOST_CONFIG,
1423                 "negotiated Vhost-user protocol features: 0x%" PRIx64 "\n",
1424                 dev->protocol_features);
1425
1426         return RTE_VHOST_MSG_RESULT_OK;
1427 }
1428
1429 static int
1430 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1431                         int main_fd __rte_unused)
1432 {
1433         struct virtio_net *dev = *pdev;
1434         int fd = msg->fds[0];
1435         uint64_t size, off;
1436         void *addr;
1437
1438         if (fd < 0) {
1439                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1440                 return RTE_VHOST_MSG_RESULT_ERR;
1441         }
1442
1443         if (msg->size != sizeof(VhostUserLog)) {
1444                 RTE_LOG(ERR, VHOST_CONFIG,
1445                         "invalid log base msg size: %"PRId32" != %d\n",
1446                         msg->size, (int)sizeof(VhostUserLog));
1447                 return RTE_VHOST_MSG_RESULT_ERR;
1448         }
1449
1450         size = msg->payload.log.mmap_size;
1451         off  = msg->payload.log.mmap_offset;
1452
1453         /* Don't allow mmap_offset to point outside the mmap region */
1454         if (off > size) {
1455                 RTE_LOG(ERR, VHOST_CONFIG,
1456                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1457                         off, size);
1458                 return RTE_VHOST_MSG_RESULT_ERR;
1459         }
1460
1461         RTE_LOG(INFO, VHOST_CONFIG,
1462                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1463                 size, off);
1464
1465         /*
1466          * mmap from 0 to workaround a hugepage mmap bug: mmap will
1467          * fail when offset is not page size aligned.
1468          */
1469         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1470         close(fd);
1471         if (addr == MAP_FAILED) {
1472                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1473                 return RTE_VHOST_MSG_RESULT_ERR;
1474         }
1475
1476         /*
1477          * Free previously mapped log memory on occasionally
1478          * multiple VHOST_USER_SET_LOG_BASE.
1479          */
1480         if (dev->log_addr) {
1481                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1482         }
1483         dev->log_addr = (uint64_t)(uintptr_t)addr;
1484         dev->log_base = dev->log_addr + off;
1485         dev->log_size = size;
1486
1487         /*
1488          * The spec is not clear about it (yet), but QEMU doesn't expect
1489          * any payload in the reply.
1490          */
1491         msg->size = 0;
1492         msg->fd_num = 0;
1493
1494         return RTE_VHOST_MSG_RESULT_REPLY;
1495 }
1496
1497 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1498                         struct VhostUserMsg *msg,
1499                         int main_fd __rte_unused)
1500 {
1501         close(msg->fds[0]);
1502         RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1503
1504         return RTE_VHOST_MSG_RESULT_OK;
1505 }
1506
1507 /*
1508  * An rarp packet is constructed and broadcasted to notify switches about
1509  * the new location of the migrated VM, so that packets from outside will
1510  * not be lost after migration.
1511  *
1512  * However, we don't actually "send" a rarp packet here, instead, we set
1513  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1514  */
1515 static int
1516 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
1517                         int main_fd __rte_unused)
1518 {
1519         struct virtio_net *dev = *pdev;
1520         uint8_t *mac = (uint8_t *)&msg->payload.u64;
1521         struct rte_vdpa_device *vdpa_dev;
1522         int did = -1;
1523
1524         RTE_LOG(DEBUG, VHOST_CONFIG,
1525                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1526                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1527         memcpy(dev->mac.addr_bytes, mac, 6);
1528
1529         /*
1530          * Set the flag to inject a RARP broadcast packet at
1531          * rte_vhost_dequeue_burst().
1532          *
1533          * rte_smp_wmb() is for making sure the mac is copied
1534          * before the flag is set.
1535          */
1536         rte_smp_wmb();
1537         rte_atomic16_set(&dev->broadcast_rarp, 1);
1538         did = dev->vdpa_dev_id;
1539         vdpa_dev = rte_vdpa_get_device(did);
1540         if (vdpa_dev && vdpa_dev->ops->migration_done)
1541                 vdpa_dev->ops->migration_done(dev->vid);
1542
1543         return RTE_VHOST_MSG_RESULT_OK;
1544 }
1545
1546 static int
1547 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
1548                         int main_fd __rte_unused)
1549 {
1550         struct virtio_net *dev = *pdev;
1551         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1552                         msg->payload.u64 > VIRTIO_MAX_MTU) {
1553                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1554                                 msg->payload.u64);
1555
1556                 return RTE_VHOST_MSG_RESULT_ERR;
1557         }
1558
1559         dev->mtu = msg->payload.u64;
1560
1561         return RTE_VHOST_MSG_RESULT_OK;
1562 }
1563
1564 static int
1565 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
1566                         int main_fd __rte_unused)
1567 {
1568         struct virtio_net *dev = *pdev;
1569         int fd = msg->fds[0];
1570
1571         if (fd < 0) {
1572                 RTE_LOG(ERR, VHOST_CONFIG,
1573                                 "Invalid file descriptor for slave channel (%d)\n",
1574                                 fd);
1575                 return RTE_VHOST_MSG_RESULT_ERR;
1576         }
1577
1578         if (dev->slave_req_fd >= 0)
1579                 close(dev->slave_req_fd);
1580
1581         dev->slave_req_fd = fd;
1582
1583         return RTE_VHOST_MSG_RESULT_OK;
1584 }
1585
1586 static int
1587 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1588 {
1589         struct vhost_vring_addr *ra;
1590         uint64_t start, end;
1591
1592         start = imsg->iova;
1593         end = start + imsg->size;
1594
1595         ra = &vq->ring_addrs;
1596         if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1597                 return 1;
1598         if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1599                 return 1;
1600         if (ra->used_user_addr >= start && ra->used_user_addr < end)
1601                 return 1;
1602
1603         return 0;
1604 }
1605
1606 static int
1607 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1608                                 struct vhost_iotlb_msg *imsg)
1609 {
1610         uint64_t istart, iend, vstart, vend;
1611
1612         istart = imsg->iova;
1613         iend = istart + imsg->size - 1;
1614
1615         vstart = (uintptr_t)vq->desc;
1616         vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1617         if (vstart <= iend && istart <= vend)
1618                 return 1;
1619
1620         vstart = (uintptr_t)vq->avail;
1621         vend = vstart + sizeof(struct vring_avail);
1622         vend += sizeof(uint16_t) * vq->size - 1;
1623         if (vstart <= iend && istart <= vend)
1624                 return 1;
1625
1626         vstart = (uintptr_t)vq->used;
1627         vend = vstart + sizeof(struct vring_used);
1628         vend += sizeof(struct vring_used_elem) * vq->size - 1;
1629         if (vstart <= iend && istart <= vend)
1630                 return 1;
1631
1632         return 0;
1633 }
1634
1635 static int
1636 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
1637                         int main_fd __rte_unused)
1638 {
1639         struct virtio_net *dev = *pdev;
1640         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1641         uint16_t i;
1642         uint64_t vva, len;
1643
1644         switch (imsg->type) {
1645         case VHOST_IOTLB_UPDATE:
1646                 len = imsg->size;
1647                 vva = qva_to_vva(dev, imsg->uaddr, &len);
1648                 if (!vva)
1649                         return RTE_VHOST_MSG_RESULT_ERR;
1650
1651                 for (i = 0; i < dev->nr_vring; i++) {
1652                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1653
1654                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1655                                         len, imsg->perm);
1656
1657                         if (is_vring_iotlb_update(vq, imsg))
1658                                 *pdev = dev = translate_ring_addresses(dev, i);
1659                 }
1660                 break;
1661         case VHOST_IOTLB_INVALIDATE:
1662                 for (i = 0; i < dev->nr_vring; i++) {
1663                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1664
1665                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
1666                                         imsg->size);
1667
1668                         if (is_vring_iotlb_invalidate(vq, imsg))
1669                                 vring_invalidate(dev, vq);
1670                 }
1671                 break;
1672         default:
1673                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1674                                 imsg->type);
1675                 return RTE_VHOST_MSG_RESULT_ERR;
1676         }
1677
1678         return RTE_VHOST_MSG_RESULT_OK;
1679 }
1680
1681 static int
1682 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
1683                         struct VhostUserMsg *msg,
1684                         int main_fd __rte_unused)
1685 {
1686         struct virtio_net *dev = *pdev;
1687 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1688         struct uffdio_api api_struct;
1689
1690         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1691
1692         if (dev->postcopy_ufd == -1) {
1693                 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
1694                         strerror(errno));
1695                 return RTE_VHOST_MSG_RESULT_ERR;
1696         }
1697         api_struct.api = UFFD_API;
1698         api_struct.features = 0;
1699         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
1700                 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
1701                         strerror(errno));
1702                 close(dev->postcopy_ufd);
1703                 dev->postcopy_ufd = -1;
1704                 return RTE_VHOST_MSG_RESULT_ERR;
1705         }
1706         msg->fds[0] = dev->postcopy_ufd;
1707         msg->fd_num = 1;
1708
1709         return RTE_VHOST_MSG_RESULT_REPLY;
1710 #else
1711         dev->postcopy_ufd = -1;
1712         msg->fd_num = 0;
1713
1714         return RTE_VHOST_MSG_RESULT_ERR;
1715 #endif
1716 }
1717
1718 static int
1719 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
1720                         struct VhostUserMsg *msg __rte_unused,
1721                         int main_fd __rte_unused)
1722 {
1723         struct virtio_net *dev = *pdev;
1724
1725         if (dev->mem && dev->mem->nregions) {
1726                 RTE_LOG(ERR, VHOST_CONFIG,
1727                         "Regions already registered at postcopy-listen\n");
1728                 return RTE_VHOST_MSG_RESULT_ERR;
1729         }
1730         dev->postcopy_listening = 1;
1731
1732         return RTE_VHOST_MSG_RESULT_OK;
1733 }
1734
1735 static int
1736 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
1737                         int main_fd __rte_unused)
1738 {
1739         struct virtio_net *dev = *pdev;
1740
1741         dev->postcopy_listening = 0;
1742         if (dev->postcopy_ufd >= 0) {
1743                 close(dev->postcopy_ufd);
1744                 dev->postcopy_ufd = -1;
1745         }
1746
1747         msg->payload.u64 = 0;
1748         msg->size = sizeof(msg->payload.u64);
1749         msg->fd_num = 0;
1750
1751         return RTE_VHOST_MSG_RESULT_REPLY;
1752 }
1753
1754 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
1755                                         struct VhostUserMsg *msg,
1756                                         int main_fd);
1757 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
1758         [VHOST_USER_NONE] = NULL,
1759         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
1760         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
1761         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
1762         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
1763         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
1764         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
1765         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
1766         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
1767         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
1768         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
1769         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
1770         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
1771         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
1772         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
1773         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
1774         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
1775         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
1776         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
1777         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
1778         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
1779         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
1780         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
1781         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
1782         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
1783         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
1784 };
1785
1786
1787 /* return bytes# of read on success or negative val on failure. */
1788 static int
1789 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1790 {
1791         int ret;
1792
1793         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1794                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
1795         if (ret <= 0)
1796                 return ret;
1797
1798         if (msg->size) {
1799                 if (msg->size > sizeof(msg->payload)) {
1800                         RTE_LOG(ERR, VHOST_CONFIG,
1801                                 "invalid msg size: %d\n", msg->size);
1802                         return -1;
1803                 }
1804                 ret = read(sockfd, &msg->payload, msg->size);
1805                 if (ret <= 0)
1806                         return ret;
1807                 if (ret != (int)msg->size) {
1808                         RTE_LOG(ERR, VHOST_CONFIG,
1809                                 "read control message failed\n");
1810                         return -1;
1811                 }
1812         }
1813
1814         return ret;
1815 }
1816
1817 static int
1818 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1819 {
1820         if (!msg)
1821                 return 0;
1822
1823         return send_fd_message(sockfd, (char *)msg,
1824                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
1825 }
1826
1827 static int
1828 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1829 {
1830         if (!msg)
1831                 return 0;
1832
1833         msg->flags &= ~VHOST_USER_VERSION_MASK;
1834         msg->flags &= ~VHOST_USER_NEED_REPLY;
1835         msg->flags |= VHOST_USER_VERSION;
1836         msg->flags |= VHOST_USER_REPLY_MASK;
1837
1838         return send_vhost_message(sockfd, msg);
1839 }
1840
1841 static int
1842 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
1843 {
1844         int ret;
1845
1846         if (msg->flags & VHOST_USER_NEED_REPLY)
1847                 rte_spinlock_lock(&dev->slave_req_lock);
1848
1849         ret = send_vhost_message(dev->slave_req_fd, msg);
1850         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1851                 rte_spinlock_unlock(&dev->slave_req_lock);
1852
1853         return ret;
1854 }
1855
1856 /*
1857  * Allocate a queue pair if it hasn't been allocated yet
1858  */
1859 static int
1860 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1861                         struct VhostUserMsg *msg)
1862 {
1863         uint16_t vring_idx;
1864
1865         switch (msg->request.master) {
1866         case VHOST_USER_SET_VRING_KICK:
1867         case VHOST_USER_SET_VRING_CALL:
1868         case VHOST_USER_SET_VRING_ERR:
1869                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1870                 break;
1871         case VHOST_USER_SET_VRING_NUM:
1872         case VHOST_USER_SET_VRING_BASE:
1873         case VHOST_USER_SET_VRING_ENABLE:
1874                 vring_idx = msg->payload.state.index;
1875                 break;
1876         case VHOST_USER_SET_VRING_ADDR:
1877                 vring_idx = msg->payload.addr.index;
1878                 break;
1879         default:
1880                 return 0;
1881         }
1882
1883         if (vring_idx >= VHOST_MAX_VRING) {
1884                 RTE_LOG(ERR, VHOST_CONFIG,
1885                         "invalid vring index: %u\n", vring_idx);
1886                 return -1;
1887         }
1888
1889         if (dev->virtqueue[vring_idx])
1890                 return 0;
1891
1892         return alloc_vring_queue(dev, vring_idx);
1893 }
1894
1895 static void
1896 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1897 {
1898         unsigned int i = 0;
1899         unsigned int vq_num = 0;
1900
1901         while (vq_num < dev->nr_vring) {
1902                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1903
1904                 if (vq) {
1905                         rte_spinlock_lock(&vq->access_lock);
1906                         vq_num++;
1907                 }
1908                 i++;
1909         }
1910 }
1911
1912 static void
1913 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1914 {
1915         unsigned int i = 0;
1916         unsigned int vq_num = 0;
1917
1918         while (vq_num < dev->nr_vring) {
1919                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1920
1921                 if (vq) {
1922                         rte_spinlock_unlock(&vq->access_lock);
1923                         vq_num++;
1924                 }
1925                 i++;
1926         }
1927 }
1928
1929 int
1930 vhost_user_msg_handler(int vid, int fd)
1931 {
1932         struct virtio_net *dev;
1933         struct VhostUserMsg msg;
1934         struct rte_vdpa_device *vdpa_dev;
1935         int did = -1;
1936         int ret;
1937         int unlock_required = 0;
1938         bool handled;
1939         int request;
1940
1941         dev = get_device(vid);
1942         if (dev == NULL)
1943                 return -1;
1944
1945         if (!dev->notify_ops) {
1946                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1947                 if (!dev->notify_ops) {
1948                         RTE_LOG(ERR, VHOST_CONFIG,
1949                                 "failed to get callback ops for driver %s\n",
1950                                 dev->ifname);
1951                         return -1;
1952                 }
1953         }
1954
1955         ret = read_vhost_message(fd, &msg);
1956         if (ret <= 0) {
1957                 if (ret < 0)
1958                         RTE_LOG(ERR, VHOST_CONFIG,
1959                                 "vhost read message failed\n");
1960                 else
1961                         RTE_LOG(INFO, VHOST_CONFIG,
1962                                 "vhost peer closed\n");
1963
1964                 return -1;
1965         }
1966
1967         ret = 0;
1968         request = msg.request.master;
1969         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
1970                         vhost_message_str[request]) {
1971                 if (request != VHOST_USER_IOTLB_MSG)
1972                         RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1973                                 vhost_message_str[request]);
1974                 else
1975                         RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1976                                 vhost_message_str[request]);
1977         } else {
1978                 RTE_LOG(DEBUG, VHOST_CONFIG, "External request %d\n", request);
1979         }
1980
1981         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1982         if (ret < 0) {
1983                 RTE_LOG(ERR, VHOST_CONFIG,
1984                         "failed to alloc queue\n");
1985                 return -1;
1986         }
1987
1988         /*
1989          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1990          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1991          * and device is destroyed. destroy_device waits for queues to be
1992          * inactive, so it is safe. Otherwise taking the access_lock
1993          * would cause a dead lock.
1994          */
1995         switch (request) {
1996         case VHOST_USER_SET_FEATURES:
1997         case VHOST_USER_SET_PROTOCOL_FEATURES:
1998         case VHOST_USER_SET_OWNER:
1999         case VHOST_USER_SET_MEM_TABLE:
2000         case VHOST_USER_SET_LOG_BASE:
2001         case VHOST_USER_SET_LOG_FD:
2002         case VHOST_USER_SET_VRING_NUM:
2003         case VHOST_USER_SET_VRING_ADDR:
2004         case VHOST_USER_SET_VRING_BASE:
2005         case VHOST_USER_SET_VRING_KICK:
2006         case VHOST_USER_SET_VRING_CALL:
2007         case VHOST_USER_SET_VRING_ERR:
2008         case VHOST_USER_SET_VRING_ENABLE:
2009         case VHOST_USER_SEND_RARP:
2010         case VHOST_USER_NET_SET_MTU:
2011         case VHOST_USER_SET_SLAVE_REQ_FD:
2012                 vhost_user_lock_all_queue_pairs(dev);
2013                 unlock_required = 1;
2014                 break;
2015         default:
2016                 break;
2017
2018         }
2019
2020         handled = false;
2021         if (dev->extern_ops.pre_msg_handle) {
2022                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
2023                                 (void *)&msg);
2024                 switch (ret) {
2025                 case RTE_VHOST_MSG_RESULT_REPLY:
2026                         send_vhost_reply(fd, &msg);
2027                         /* Fall-through */
2028                 case RTE_VHOST_MSG_RESULT_ERR:
2029                 case RTE_VHOST_MSG_RESULT_OK:
2030                         handled = true;
2031                         goto skip_to_post_handle;
2032                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2033                 default:
2034                         break;
2035                 }
2036         }
2037
2038         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
2039                 if (!vhost_message_handlers[request])
2040                         goto skip_to_post_handle;
2041                 ret = vhost_message_handlers[request](&dev, &msg, fd);
2042
2043                 switch (ret) {
2044                 case RTE_VHOST_MSG_RESULT_ERR:
2045                         RTE_LOG(ERR, VHOST_CONFIG,
2046                                 "Processing %s failed.\n",
2047                                 vhost_message_str[request]);
2048                         handled = true;
2049                         break;
2050                 case RTE_VHOST_MSG_RESULT_OK:
2051                         RTE_LOG(DEBUG, VHOST_CONFIG,
2052                                 "Processing %s succeeded.\n",
2053                                 vhost_message_str[request]);
2054                         handled = true;
2055                         break;
2056                 case RTE_VHOST_MSG_RESULT_REPLY:
2057                         RTE_LOG(DEBUG, VHOST_CONFIG,
2058                                 "Processing %s succeeded and needs reply.\n",
2059                                 vhost_message_str[request]);
2060                         send_vhost_reply(fd, &msg);
2061                         handled = true;
2062                         break;
2063                 default:
2064                         break;
2065                 }
2066         }
2067
2068 skip_to_post_handle:
2069         if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2070                         dev->extern_ops.post_msg_handle) {
2071                 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
2072                                 (void *)&msg);
2073                 switch (ret) {
2074                 case RTE_VHOST_MSG_RESULT_REPLY:
2075                         send_vhost_reply(fd, &msg);
2076                         /* Fall-through */
2077                 case RTE_VHOST_MSG_RESULT_ERR:
2078                 case RTE_VHOST_MSG_RESULT_OK:
2079                         handled = true;
2080                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2081                 default:
2082                         break;
2083                 }
2084         }
2085
2086         if (unlock_required)
2087                 vhost_user_unlock_all_queue_pairs(dev);
2088
2089         /* If message was not handled at this stage, treat it as an error */
2090         if (!handled) {
2091                 RTE_LOG(ERR, VHOST_CONFIG,
2092                         "vhost message (req: %d) was not handled.\n", request);
2093                 ret = RTE_VHOST_MSG_RESULT_ERR;
2094         }
2095
2096         /*
2097          * If the request required a reply that was already sent,
2098          * this optional reply-ack won't be sent as the
2099          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2100          */
2101         if (msg.flags & VHOST_USER_NEED_REPLY) {
2102                 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2103                 msg.size = sizeof(msg.payload.u64);
2104                 msg.fd_num = 0;
2105                 send_vhost_reply(fd, &msg);
2106         } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2107                 RTE_LOG(ERR, VHOST_CONFIG,
2108                         "vhost message handling failed.\n");
2109                 return -1;
2110         }
2111
2112         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2113                 dev->flags |= VIRTIO_DEV_READY;
2114
2115                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2116                         if (dev->dequeue_zero_copy) {
2117                                 RTE_LOG(INFO, VHOST_CONFIG,
2118                                                 "dequeue zero copy is enabled\n");
2119                         }
2120
2121                         if (dev->notify_ops->new_device(dev->vid) == 0)
2122                                 dev->flags |= VIRTIO_DEV_RUNNING;
2123                 }
2124         }
2125
2126         did = dev->vdpa_dev_id;
2127         vdpa_dev = rte_vdpa_get_device(did);
2128         if (vdpa_dev && virtio_is_ready(dev) &&
2129                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2130                         msg.request.master == VHOST_USER_SET_VRING_CALL) {
2131                 if (vdpa_dev->ops->dev_conf)
2132                         vdpa_dev->ops->dev_conf(dev->vid);
2133                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2134         }
2135
2136         return 0;
2137 }
2138
2139 static int process_slave_message_reply(struct virtio_net *dev,
2140                                        const struct VhostUserMsg *msg)
2141 {
2142         struct VhostUserMsg msg_reply;
2143         int ret;
2144
2145         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2146                 return 0;
2147
2148         if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
2149                 ret = -1;
2150                 goto out;
2151         }
2152
2153         if (msg_reply.request.slave != msg->request.slave) {
2154                 RTE_LOG(ERR, VHOST_CONFIG,
2155                         "Received unexpected msg type (%u), expected %u\n",
2156                         msg_reply.request.slave, msg->request.slave);
2157                 ret = -1;
2158                 goto out;
2159         }
2160
2161         ret = msg_reply.payload.u64 ? -1 : 0;
2162
2163 out:
2164         rte_spinlock_unlock(&dev->slave_req_lock);
2165         return ret;
2166 }
2167
2168 int
2169 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2170 {
2171         int ret;
2172         struct VhostUserMsg msg = {
2173                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2174                 .flags = VHOST_USER_VERSION,
2175                 .size = sizeof(msg.payload.iotlb),
2176                 .payload.iotlb = {
2177                         .iova = iova,
2178                         .perm = perm,
2179                         .type = VHOST_IOTLB_MISS,
2180                 },
2181         };
2182
2183         ret = send_vhost_message(dev->slave_req_fd, &msg);
2184         if (ret < 0) {
2185                 RTE_LOG(ERR, VHOST_CONFIG,
2186                                 "Failed to send IOTLB miss message (%d)\n",
2187                                 ret);
2188                 return ret;
2189         }
2190
2191         return 0;
2192 }
2193
2194 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2195                                                     int index, int fd,
2196                                                     uint64_t offset,
2197                                                     uint64_t size)
2198 {
2199         int ret;
2200         struct VhostUserMsg msg = {
2201                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2202                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2203                 .size = sizeof(msg.payload.area),
2204                 .payload.area = {
2205                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
2206                         .size = size,
2207                         .offset = offset,
2208                 },
2209         };
2210
2211         if (fd < 0)
2212                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2213         else {
2214                 msg.fds[0] = fd;
2215                 msg.fd_num = 1;
2216         }
2217
2218         ret = send_vhost_slave_message(dev, &msg);
2219         if (ret < 0) {
2220                 RTE_LOG(ERR, VHOST_CONFIG,
2221                         "Failed to set host notifier (%d)\n", ret);
2222                 return ret;
2223         }
2224
2225         return process_slave_message_reply(dev, &msg);
2226 }
2227
2228 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2229 {
2230         struct virtio_net *dev;
2231         struct rte_vdpa_device *vdpa_dev;
2232         int vfio_device_fd, did, ret = 0;
2233         uint64_t offset, size;
2234         unsigned int i;
2235
2236         dev = get_device(vid);
2237         if (!dev)
2238                 return -ENODEV;
2239
2240         did = dev->vdpa_dev_id;
2241         if (did < 0)
2242                 return -EINVAL;
2243
2244         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2245             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2246             !(dev->protocol_features &
2247                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2248             !(dev->protocol_features &
2249                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2250             !(dev->protocol_features &
2251                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2252                 return -ENOTSUP;
2253
2254         vdpa_dev = rte_vdpa_get_device(did);
2255         if (!vdpa_dev)
2256                 return -ENODEV;
2257
2258         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2259         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2260
2261         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2262         if (vfio_device_fd < 0)
2263                 return -ENOTSUP;
2264
2265         if (enable) {
2266                 for (i = 0; i < dev->nr_vring; i++) {
2267                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2268                                         &size) < 0) {
2269                                 ret = -ENOTSUP;
2270                                 goto disable;
2271                         }
2272
2273                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2274                                         vfio_device_fd, offset, size) < 0) {
2275                                 ret = -EFAULT;
2276                                 goto disable;
2277                         }
2278                 }
2279         } else {
2280 disable:
2281                 for (i = 0; i < dev->nr_vring; i++) {
2282                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
2283                                         0, 0);
2284                 }
2285         }
2286
2287         return ret;
2288 }