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