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         vq->signalled_used_valid = false;
1302
1303         if (dev->dequeue_zero_copy)
1304                 free_zmbufs(vq);
1305         if (vq_is_packed(dev)) {
1306                 rte_free(vq->shadow_used_packed);
1307                 vq->shadow_used_packed = NULL;
1308         } else {
1309                 rte_free(vq->shadow_used_split);
1310                 vq->shadow_used_split = NULL;
1311         }
1312
1313         rte_free(vq->batch_copy_elems);
1314         vq->batch_copy_elems = NULL;
1315
1316         msg->size = sizeof(msg->payload.state);
1317         msg->fd_num = 0;
1318
1319         return RTE_VHOST_MSG_RESULT_REPLY;
1320 }
1321
1322 /*
1323  * when virtio queues are ready to work, qemu will send us to
1324  * enable the virtio queue pair.
1325  */
1326 static int
1327 vhost_user_set_vring_enable(struct virtio_net **pdev,
1328                         struct VhostUserMsg *msg,
1329                         int main_fd __rte_unused)
1330 {
1331         struct virtio_net *dev = *pdev;
1332         int enable = (int)msg->payload.state.num;
1333         int index = (int)msg->payload.state.index;
1334         struct rte_vdpa_device *vdpa_dev;
1335         int did = -1;
1336
1337         RTE_LOG(INFO, VHOST_CONFIG,
1338                 "set queue enable: %d to qp idx: %d\n",
1339                 enable, index);
1340
1341         did = dev->vdpa_dev_id;
1342         vdpa_dev = rte_vdpa_get_device(did);
1343         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1344                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1345
1346         if (dev->notify_ops->vring_state_changed)
1347                 dev->notify_ops->vring_state_changed(dev->vid,
1348                                 index, enable);
1349
1350         /* On disable, rings have to be stopped being processed. */
1351         if (!enable && dev->dequeue_zero_copy)
1352                 drain_zmbuf_list(dev->virtqueue[index]);
1353
1354         dev->virtqueue[index]->enabled = enable;
1355
1356         return RTE_VHOST_MSG_RESULT_OK;
1357 }
1358
1359 static int
1360 vhost_user_get_protocol_features(struct virtio_net **pdev,
1361                         struct VhostUserMsg *msg,
1362                         int main_fd __rte_unused)
1363 {
1364         struct virtio_net *dev = *pdev;
1365         uint64_t features, protocol_features;
1366
1367         rte_vhost_driver_get_features(dev->ifname, &features);
1368         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1369
1370         /*
1371          * REPLY_ACK protocol feature is only mandatory for now
1372          * for IOMMU feature. If IOMMU is explicitly disabled by the
1373          * application, disable also REPLY_ACK feature for older buggy
1374          * Qemu versions (from v2.7.0 to v2.9.0).
1375          */
1376         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1377                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1378
1379         msg->payload.u64 = protocol_features;
1380         msg->size = sizeof(msg->payload.u64);
1381         msg->fd_num = 0;
1382
1383         return RTE_VHOST_MSG_RESULT_REPLY;
1384 }
1385
1386 static int
1387 vhost_user_set_protocol_features(struct virtio_net **pdev,
1388                         struct VhostUserMsg *msg,
1389                         int main_fd __rte_unused)
1390 {
1391         struct virtio_net *dev = *pdev;
1392         uint64_t protocol_features = msg->payload.u64;
1393         uint64_t slave_protocol_features = 0;
1394
1395         rte_vhost_driver_get_protocol_features(dev->ifname,
1396                         &slave_protocol_features);
1397         if (protocol_features & ~slave_protocol_features) {
1398                 RTE_LOG(ERR, VHOST_CONFIG,
1399                         "(%d) received invalid protocol features.\n",
1400                         dev->vid);
1401                 return RTE_VHOST_MSG_RESULT_ERR;
1402         }
1403
1404         dev->protocol_features = protocol_features;
1405
1406         return RTE_VHOST_MSG_RESULT_OK;
1407 }
1408
1409 static int
1410 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1411                         int main_fd __rte_unused)
1412 {
1413         struct virtio_net *dev = *pdev;
1414         int fd = msg->fds[0];
1415         uint64_t size, off;
1416         void *addr;
1417
1418         if (fd < 0) {
1419                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1420                 return RTE_VHOST_MSG_RESULT_ERR;
1421         }
1422
1423         if (msg->size != sizeof(VhostUserLog)) {
1424                 RTE_LOG(ERR, VHOST_CONFIG,
1425                         "invalid log base msg size: %"PRId32" != %d\n",
1426                         msg->size, (int)sizeof(VhostUserLog));
1427                 return RTE_VHOST_MSG_RESULT_ERR;
1428         }
1429
1430         size = msg->payload.log.mmap_size;
1431         off  = msg->payload.log.mmap_offset;
1432
1433         /* Don't allow mmap_offset to point outside the mmap region */
1434         if (off > size) {
1435                 RTE_LOG(ERR, VHOST_CONFIG,
1436                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1437                         off, size);
1438                 return RTE_VHOST_MSG_RESULT_ERR;
1439         }
1440
1441         RTE_LOG(INFO, VHOST_CONFIG,
1442                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1443                 size, off);
1444
1445         /*
1446          * mmap from 0 to workaround a hugepage mmap bug: mmap will
1447          * fail when offset is not page size aligned.
1448          */
1449         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1450         close(fd);
1451         if (addr == MAP_FAILED) {
1452                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1453                 return RTE_VHOST_MSG_RESULT_ERR;
1454         }
1455
1456         /*
1457          * Free previously mapped log memory on occasionally
1458          * multiple VHOST_USER_SET_LOG_BASE.
1459          */
1460         if (dev->log_addr) {
1461                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1462         }
1463         dev->log_addr = (uint64_t)(uintptr_t)addr;
1464         dev->log_base = dev->log_addr + off;
1465         dev->log_size = size;
1466
1467         /*
1468          * The spec is not clear about it (yet), but QEMU doesn't expect
1469          * any payload in the reply.
1470          */
1471         msg->size = 0;
1472         msg->fd_num = 0;
1473
1474         return RTE_VHOST_MSG_RESULT_REPLY;
1475 }
1476
1477 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1478                         struct VhostUserMsg *msg,
1479                         int main_fd __rte_unused)
1480 {
1481         close(msg->fds[0]);
1482         RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1483
1484         return RTE_VHOST_MSG_RESULT_OK;
1485 }
1486
1487 /*
1488  * An rarp packet is constructed and broadcasted to notify switches about
1489  * the new location of the migrated VM, so that packets from outside will
1490  * not be lost after migration.
1491  *
1492  * However, we don't actually "send" a rarp packet here, instead, we set
1493  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1494  */
1495 static int
1496 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
1497                         int main_fd __rte_unused)
1498 {
1499         struct virtio_net *dev = *pdev;
1500         uint8_t *mac = (uint8_t *)&msg->payload.u64;
1501         struct rte_vdpa_device *vdpa_dev;
1502         int did = -1;
1503
1504         RTE_LOG(DEBUG, VHOST_CONFIG,
1505                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1506                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1507         memcpy(dev->mac.addr_bytes, mac, 6);
1508
1509         /*
1510          * Set the flag to inject a RARP broadcast packet at
1511          * rte_vhost_dequeue_burst().
1512          *
1513          * rte_smp_wmb() is for making sure the mac is copied
1514          * before the flag is set.
1515          */
1516         rte_smp_wmb();
1517         rte_atomic16_set(&dev->broadcast_rarp, 1);
1518         did = dev->vdpa_dev_id;
1519         vdpa_dev = rte_vdpa_get_device(did);
1520         if (vdpa_dev && vdpa_dev->ops->migration_done)
1521                 vdpa_dev->ops->migration_done(dev->vid);
1522
1523         return RTE_VHOST_MSG_RESULT_OK;
1524 }
1525
1526 static int
1527 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
1528                         int main_fd __rte_unused)
1529 {
1530         struct virtio_net *dev = *pdev;
1531         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1532                         msg->payload.u64 > VIRTIO_MAX_MTU) {
1533                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1534                                 msg->payload.u64);
1535
1536                 return RTE_VHOST_MSG_RESULT_ERR;
1537         }
1538
1539         dev->mtu = msg->payload.u64;
1540
1541         return RTE_VHOST_MSG_RESULT_OK;
1542 }
1543
1544 static int
1545 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
1546                         int main_fd __rte_unused)
1547 {
1548         struct virtio_net *dev = *pdev;
1549         int fd = msg->fds[0];
1550
1551         if (fd < 0) {
1552                 RTE_LOG(ERR, VHOST_CONFIG,
1553                                 "Invalid file descriptor for slave channel (%d)\n",
1554                                 fd);
1555                 return RTE_VHOST_MSG_RESULT_ERR;
1556         }
1557
1558         dev->slave_req_fd = fd;
1559
1560         return RTE_VHOST_MSG_RESULT_OK;
1561 }
1562
1563 static int
1564 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1565 {
1566         struct vhost_vring_addr *ra;
1567         uint64_t start, end;
1568
1569         start = imsg->iova;
1570         end = start + imsg->size;
1571
1572         ra = &vq->ring_addrs;
1573         if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1574                 return 1;
1575         if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1576                 return 1;
1577         if (ra->used_user_addr >= start && ra->used_user_addr < end)
1578                 return 1;
1579
1580         return 0;
1581 }
1582
1583 static int
1584 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1585                                 struct vhost_iotlb_msg *imsg)
1586 {
1587         uint64_t istart, iend, vstart, vend;
1588
1589         istart = imsg->iova;
1590         iend = istart + imsg->size - 1;
1591
1592         vstart = (uintptr_t)vq->desc;
1593         vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1594         if (vstart <= iend && istart <= vend)
1595                 return 1;
1596
1597         vstart = (uintptr_t)vq->avail;
1598         vend = vstart + sizeof(struct vring_avail);
1599         vend += sizeof(uint16_t) * vq->size - 1;
1600         if (vstart <= iend && istart <= vend)
1601                 return 1;
1602
1603         vstart = (uintptr_t)vq->used;
1604         vend = vstart + sizeof(struct vring_used);
1605         vend += sizeof(struct vring_used_elem) * vq->size - 1;
1606         if (vstart <= iend && istart <= vend)
1607                 return 1;
1608
1609         return 0;
1610 }
1611
1612 static int
1613 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
1614                         int main_fd __rte_unused)
1615 {
1616         struct virtio_net *dev = *pdev;
1617         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1618         uint16_t i;
1619         uint64_t vva, len;
1620
1621         switch (imsg->type) {
1622         case VHOST_IOTLB_UPDATE:
1623                 len = imsg->size;
1624                 vva = qva_to_vva(dev, imsg->uaddr, &len);
1625                 if (!vva)
1626                         return RTE_VHOST_MSG_RESULT_ERR;
1627
1628                 for (i = 0; i < dev->nr_vring; i++) {
1629                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1630
1631                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1632                                         len, imsg->perm);
1633
1634                         if (is_vring_iotlb_update(vq, imsg))
1635                                 *pdev = dev = translate_ring_addresses(dev, i);
1636                 }
1637                 break;
1638         case VHOST_IOTLB_INVALIDATE:
1639                 for (i = 0; i < dev->nr_vring; i++) {
1640                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1641
1642                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
1643                                         imsg->size);
1644
1645                         if (is_vring_iotlb_invalidate(vq, imsg))
1646                                 vring_invalidate(dev, vq);
1647                 }
1648                 break;
1649         default:
1650                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1651                                 imsg->type);
1652                 return RTE_VHOST_MSG_RESULT_ERR;
1653         }
1654
1655         return RTE_VHOST_MSG_RESULT_OK;
1656 }
1657
1658 static int
1659 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
1660                         struct VhostUserMsg *msg,
1661                         int main_fd __rte_unused)
1662 {
1663         struct virtio_net *dev = *pdev;
1664 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1665         struct uffdio_api api_struct;
1666
1667         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1668
1669         if (dev->postcopy_ufd == -1) {
1670                 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
1671                         strerror(errno));
1672                 return RTE_VHOST_MSG_RESULT_ERR;
1673         }
1674         api_struct.api = UFFD_API;
1675         api_struct.features = 0;
1676         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
1677                 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
1678                         strerror(errno));
1679                 close(dev->postcopy_ufd);
1680                 dev->postcopy_ufd = -1;
1681                 return RTE_VHOST_MSG_RESULT_ERR;
1682         }
1683         msg->fds[0] = dev->postcopy_ufd;
1684         msg->fd_num = 1;
1685
1686         return RTE_VHOST_MSG_RESULT_REPLY;
1687 #else
1688         dev->postcopy_ufd = -1;
1689         msg->fd_num = 0;
1690
1691         return RTE_VHOST_MSG_RESULT_ERR;
1692 #endif
1693 }
1694
1695 static int
1696 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
1697                         struct VhostUserMsg *msg __rte_unused,
1698                         int main_fd __rte_unused)
1699 {
1700         struct virtio_net *dev = *pdev;
1701
1702         if (dev->mem && dev->mem->nregions) {
1703                 RTE_LOG(ERR, VHOST_CONFIG,
1704                         "Regions already registered at postcopy-listen\n");
1705                 return RTE_VHOST_MSG_RESULT_ERR;
1706         }
1707         dev->postcopy_listening = 1;
1708
1709         return RTE_VHOST_MSG_RESULT_OK;
1710 }
1711
1712 static int
1713 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
1714                         int main_fd __rte_unused)
1715 {
1716         struct virtio_net *dev = *pdev;
1717
1718         dev->postcopy_listening = 0;
1719         if (dev->postcopy_ufd >= 0) {
1720                 close(dev->postcopy_ufd);
1721                 dev->postcopy_ufd = -1;
1722         }
1723
1724         msg->payload.u64 = 0;
1725         msg->size = sizeof(msg->payload.u64);
1726         msg->fd_num = 0;
1727
1728         return RTE_VHOST_MSG_RESULT_REPLY;
1729 }
1730
1731 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
1732                                         struct VhostUserMsg *msg,
1733                                         int main_fd);
1734 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
1735         [VHOST_USER_NONE] = NULL,
1736         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
1737         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
1738         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
1739         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
1740         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
1741         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
1742         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
1743         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
1744         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
1745         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
1746         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
1747         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
1748         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
1749         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
1750         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
1751         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
1752         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
1753         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
1754         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
1755         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
1756         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
1757         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
1758         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
1759         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
1760         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
1761 };
1762
1763
1764 /* return bytes# of read on success or negative val on failure. */
1765 static int
1766 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1767 {
1768         int ret;
1769
1770         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1771                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
1772         if (ret <= 0)
1773                 return ret;
1774
1775         if (msg->size) {
1776                 if (msg->size > sizeof(msg->payload)) {
1777                         RTE_LOG(ERR, VHOST_CONFIG,
1778                                 "invalid msg size: %d\n", msg->size);
1779                         return -1;
1780                 }
1781                 ret = read(sockfd, &msg->payload, msg->size);
1782                 if (ret <= 0)
1783                         return ret;
1784                 if (ret != (int)msg->size) {
1785                         RTE_LOG(ERR, VHOST_CONFIG,
1786                                 "read control message failed\n");
1787                         return -1;
1788                 }
1789         }
1790
1791         return ret;
1792 }
1793
1794 static int
1795 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1796 {
1797         if (!msg)
1798                 return 0;
1799
1800         return send_fd_message(sockfd, (char *)msg,
1801                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
1802 }
1803
1804 static int
1805 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1806 {
1807         if (!msg)
1808                 return 0;
1809
1810         msg->flags &= ~VHOST_USER_VERSION_MASK;
1811         msg->flags &= ~VHOST_USER_NEED_REPLY;
1812         msg->flags |= VHOST_USER_VERSION;
1813         msg->flags |= VHOST_USER_REPLY_MASK;
1814
1815         return send_vhost_message(sockfd, msg);
1816 }
1817
1818 static int
1819 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
1820 {
1821         int ret;
1822
1823         if (msg->flags & VHOST_USER_NEED_REPLY)
1824                 rte_spinlock_lock(&dev->slave_req_lock);
1825
1826         ret = send_vhost_message(dev->slave_req_fd, msg);
1827         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1828                 rte_spinlock_unlock(&dev->slave_req_lock);
1829
1830         return ret;
1831 }
1832
1833 /*
1834  * Allocate a queue pair if it hasn't been allocated yet
1835  */
1836 static int
1837 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1838                         struct VhostUserMsg *msg)
1839 {
1840         uint16_t vring_idx;
1841
1842         switch (msg->request.master) {
1843         case VHOST_USER_SET_VRING_KICK:
1844         case VHOST_USER_SET_VRING_CALL:
1845         case VHOST_USER_SET_VRING_ERR:
1846                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1847                 break;
1848         case VHOST_USER_SET_VRING_NUM:
1849         case VHOST_USER_SET_VRING_BASE:
1850         case VHOST_USER_SET_VRING_ENABLE:
1851                 vring_idx = msg->payload.state.index;
1852                 break;
1853         case VHOST_USER_SET_VRING_ADDR:
1854                 vring_idx = msg->payload.addr.index;
1855                 break;
1856         default:
1857                 return 0;
1858         }
1859
1860         if (vring_idx >= VHOST_MAX_VRING) {
1861                 RTE_LOG(ERR, VHOST_CONFIG,
1862                         "invalid vring index: %u\n", vring_idx);
1863                 return -1;
1864         }
1865
1866         if (dev->virtqueue[vring_idx])
1867                 return 0;
1868
1869         return alloc_vring_queue(dev, vring_idx);
1870 }
1871
1872 static void
1873 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1874 {
1875         unsigned int i = 0;
1876         unsigned int vq_num = 0;
1877
1878         while (vq_num < dev->nr_vring) {
1879                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1880
1881                 if (vq) {
1882                         rte_spinlock_lock(&vq->access_lock);
1883                         vq_num++;
1884                 }
1885                 i++;
1886         }
1887 }
1888
1889 static void
1890 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1891 {
1892         unsigned int i = 0;
1893         unsigned int vq_num = 0;
1894
1895         while (vq_num < dev->nr_vring) {
1896                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1897
1898                 if (vq) {
1899                         rte_spinlock_unlock(&vq->access_lock);
1900                         vq_num++;
1901                 }
1902                 i++;
1903         }
1904 }
1905
1906 int
1907 vhost_user_msg_handler(int vid, int fd)
1908 {
1909         struct virtio_net *dev;
1910         struct VhostUserMsg msg;
1911         struct rte_vdpa_device *vdpa_dev;
1912         int did = -1;
1913         int ret;
1914         int unlock_required = 0;
1915         bool handled;
1916         int request;
1917
1918         dev = get_device(vid);
1919         if (dev == NULL)
1920                 return -1;
1921
1922         if (!dev->notify_ops) {
1923                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1924                 if (!dev->notify_ops) {
1925                         RTE_LOG(ERR, VHOST_CONFIG,
1926                                 "failed to get callback ops for driver %s\n",
1927                                 dev->ifname);
1928                         return -1;
1929                 }
1930         }
1931
1932         ret = read_vhost_message(fd, &msg);
1933         if (ret <= 0) {
1934                 if (ret < 0)
1935                         RTE_LOG(ERR, VHOST_CONFIG,
1936                                 "vhost read message failed\n");
1937                 else
1938                         RTE_LOG(INFO, VHOST_CONFIG,
1939                                 "vhost peer closed\n");
1940
1941                 return -1;
1942         }
1943
1944         ret = 0;
1945         request = msg.request.master;
1946         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
1947                         vhost_message_str[request]) {
1948                 if (request != VHOST_USER_IOTLB_MSG)
1949                         RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1950                                 vhost_message_str[request]);
1951                 else
1952                         RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1953                                 vhost_message_str[request]);
1954         } else {
1955                 RTE_LOG(DEBUG, VHOST_CONFIG, "External request %d\n", request);
1956         }
1957
1958         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1959         if (ret < 0) {
1960                 RTE_LOG(ERR, VHOST_CONFIG,
1961                         "failed to alloc queue\n");
1962                 return -1;
1963         }
1964
1965         /*
1966          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1967          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1968          * and device is destroyed. destroy_device waits for queues to be
1969          * inactive, so it is safe. Otherwise taking the access_lock
1970          * would cause a dead lock.
1971          */
1972         switch (request) {
1973         case VHOST_USER_SET_FEATURES:
1974         case VHOST_USER_SET_PROTOCOL_FEATURES:
1975         case VHOST_USER_SET_OWNER:
1976         case VHOST_USER_SET_MEM_TABLE:
1977         case VHOST_USER_SET_LOG_BASE:
1978         case VHOST_USER_SET_LOG_FD:
1979         case VHOST_USER_SET_VRING_NUM:
1980         case VHOST_USER_SET_VRING_ADDR:
1981         case VHOST_USER_SET_VRING_BASE:
1982         case VHOST_USER_SET_VRING_KICK:
1983         case VHOST_USER_SET_VRING_CALL:
1984         case VHOST_USER_SET_VRING_ERR:
1985         case VHOST_USER_SET_VRING_ENABLE:
1986         case VHOST_USER_SEND_RARP:
1987         case VHOST_USER_NET_SET_MTU:
1988         case VHOST_USER_SET_SLAVE_REQ_FD:
1989                 vhost_user_lock_all_queue_pairs(dev);
1990                 unlock_required = 1;
1991                 break;
1992         default:
1993                 break;
1994
1995         }
1996
1997         handled = false;
1998         if (dev->extern_ops.pre_msg_handle) {
1999                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
2000                                 (void *)&msg);
2001                 switch (ret) {
2002                 case RTE_VHOST_MSG_RESULT_REPLY:
2003                         send_vhost_reply(fd, &msg);
2004                         /* Fall-through */
2005                 case RTE_VHOST_MSG_RESULT_ERR:
2006                 case RTE_VHOST_MSG_RESULT_OK:
2007                         handled = true;
2008                         goto skip_to_post_handle;
2009                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2010                 default:
2011                         break;
2012                 }
2013         }
2014
2015         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
2016                 if (!vhost_message_handlers[request])
2017                         goto skip_to_post_handle;
2018                 ret = vhost_message_handlers[request](&dev, &msg, fd);
2019
2020                 switch (ret) {
2021                 case RTE_VHOST_MSG_RESULT_ERR:
2022                         RTE_LOG(ERR, VHOST_CONFIG,
2023                                 "Processing %s failed.\n",
2024                                 vhost_message_str[request]);
2025                         handled = true;
2026                         break;
2027                 case RTE_VHOST_MSG_RESULT_OK:
2028                         RTE_LOG(DEBUG, VHOST_CONFIG,
2029                                 "Processing %s succeeded.\n",
2030                                 vhost_message_str[request]);
2031                         handled = true;
2032                         break;
2033                 case RTE_VHOST_MSG_RESULT_REPLY:
2034                         RTE_LOG(DEBUG, VHOST_CONFIG,
2035                                 "Processing %s succeeded and needs reply.\n",
2036                                 vhost_message_str[request]);
2037                         send_vhost_reply(fd, &msg);
2038                         handled = true;
2039                         break;
2040                 default:
2041                         break;
2042                 }
2043         }
2044
2045 skip_to_post_handle:
2046         if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2047                         dev->extern_ops.post_msg_handle) {
2048                 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
2049                                 (void *)&msg);
2050                 switch (ret) {
2051                 case RTE_VHOST_MSG_RESULT_REPLY:
2052                         send_vhost_reply(fd, &msg);
2053                         /* Fall-through */
2054                 case RTE_VHOST_MSG_RESULT_ERR:
2055                 case RTE_VHOST_MSG_RESULT_OK:
2056                         handled = true;
2057                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2058                 default:
2059                         break;
2060                 }
2061         }
2062
2063         if (unlock_required)
2064                 vhost_user_unlock_all_queue_pairs(dev);
2065
2066         /* If message was not handled at this stage, treat it as an error */
2067         if (!handled) {
2068                 RTE_LOG(ERR, VHOST_CONFIG,
2069                         "vhost message (req: %d) was not handled.\n", request);
2070                 ret = RTE_VHOST_MSG_RESULT_ERR;
2071         }
2072
2073         /*
2074          * If the request required a reply that was already sent,
2075          * this optional reply-ack won't be sent as the
2076          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2077          */
2078         if (msg.flags & VHOST_USER_NEED_REPLY) {
2079                 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2080                 msg.size = sizeof(msg.payload.u64);
2081                 msg.fd_num = 0;
2082                 send_vhost_reply(fd, &msg);
2083         } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2084                 RTE_LOG(ERR, VHOST_CONFIG,
2085                         "vhost message handling failed.\n");
2086                 return -1;
2087         }
2088
2089         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2090                 dev->flags |= VIRTIO_DEV_READY;
2091
2092                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2093                         if (dev->dequeue_zero_copy) {
2094                                 RTE_LOG(INFO, VHOST_CONFIG,
2095                                                 "dequeue zero copy is enabled\n");
2096                         }
2097
2098                         if (dev->notify_ops->new_device(dev->vid) == 0)
2099                                 dev->flags |= VIRTIO_DEV_RUNNING;
2100                 }
2101         }
2102
2103         did = dev->vdpa_dev_id;
2104         vdpa_dev = rte_vdpa_get_device(did);
2105         if (vdpa_dev && virtio_is_ready(dev) &&
2106                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2107                         msg.request.master == VHOST_USER_SET_VRING_CALL) {
2108                 if (vdpa_dev->ops->dev_conf)
2109                         vdpa_dev->ops->dev_conf(dev->vid);
2110                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2111         }
2112
2113         return 0;
2114 }
2115
2116 static int process_slave_message_reply(struct virtio_net *dev,
2117                                        const struct VhostUserMsg *msg)
2118 {
2119         struct VhostUserMsg msg_reply;
2120         int ret;
2121
2122         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2123                 return 0;
2124
2125         if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
2126                 ret = -1;
2127                 goto out;
2128         }
2129
2130         if (msg_reply.request.slave != msg->request.slave) {
2131                 RTE_LOG(ERR, VHOST_CONFIG,
2132                         "Received unexpected msg type (%u), expected %u\n",
2133                         msg_reply.request.slave, msg->request.slave);
2134                 ret = -1;
2135                 goto out;
2136         }
2137
2138         ret = msg_reply.payload.u64 ? -1 : 0;
2139
2140 out:
2141         rte_spinlock_unlock(&dev->slave_req_lock);
2142         return ret;
2143 }
2144
2145 int
2146 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2147 {
2148         int ret;
2149         struct VhostUserMsg msg = {
2150                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2151                 .flags = VHOST_USER_VERSION,
2152                 .size = sizeof(msg.payload.iotlb),
2153                 .payload.iotlb = {
2154                         .iova = iova,
2155                         .perm = perm,
2156                         .type = VHOST_IOTLB_MISS,
2157                 },
2158         };
2159
2160         ret = send_vhost_message(dev->slave_req_fd, &msg);
2161         if (ret < 0) {
2162                 RTE_LOG(ERR, VHOST_CONFIG,
2163                                 "Failed to send IOTLB miss message (%d)\n",
2164                                 ret);
2165                 return ret;
2166         }
2167
2168         return 0;
2169 }
2170
2171 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2172                                                     int index, int fd,
2173                                                     uint64_t offset,
2174                                                     uint64_t size)
2175 {
2176         int ret;
2177         struct VhostUserMsg msg = {
2178                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2179                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2180                 .size = sizeof(msg.payload.area),
2181                 .payload.area = {
2182                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
2183                         .size = size,
2184                         .offset = offset,
2185                 },
2186         };
2187
2188         if (fd < 0)
2189                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2190         else {
2191                 msg.fds[0] = fd;
2192                 msg.fd_num = 1;
2193         }
2194
2195         ret = send_vhost_slave_message(dev, &msg);
2196         if (ret < 0) {
2197                 RTE_LOG(ERR, VHOST_CONFIG,
2198                         "Failed to set host notifier (%d)\n", ret);
2199                 return ret;
2200         }
2201
2202         return process_slave_message_reply(dev, &msg);
2203 }
2204
2205 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2206 {
2207         struct virtio_net *dev;
2208         struct rte_vdpa_device *vdpa_dev;
2209         int vfio_device_fd, did, ret = 0;
2210         uint64_t offset, size;
2211         unsigned int i;
2212
2213         dev = get_device(vid);
2214         if (!dev)
2215                 return -ENODEV;
2216
2217         did = dev->vdpa_dev_id;
2218         if (did < 0)
2219                 return -EINVAL;
2220
2221         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2222             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2223             !(dev->protocol_features &
2224                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2225             !(dev->protocol_features &
2226                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2227             !(dev->protocol_features &
2228                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2229                 return -ENOTSUP;
2230
2231         vdpa_dev = rte_vdpa_get_device(did);
2232         if (!vdpa_dev)
2233                 return -ENODEV;
2234
2235         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2236         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2237
2238         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2239         if (vfio_device_fd < 0)
2240                 return -ENOTSUP;
2241
2242         if (enable) {
2243                 for (i = 0; i < dev->nr_vring; i++) {
2244                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2245                                         &size) < 0) {
2246                                 ret = -ENOTSUP;
2247                                 goto disable;
2248                         }
2249
2250                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2251                                         vfio_device_fd, offset, size) < 0) {
2252                                 ret = -EFAULT;
2253                                 goto disable;
2254                         }
2255                 }
2256         } else {
2257 disable:
2258                 for (i = 0; i < dev->nr_vring; i++) {
2259                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
2260                                         0, 0);
2261                 }
2262         }
2263
2264         return ret;
2265 }