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