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 #ifdef F_ADD_SEALS /* if file sealing is supported, so is memfd */
  41 #include <linux/memfd.h>
  42 #define MEMFD_SUPPORTED
  43 #endif
  44
  45 #include <rte_common.h>
  46 #include <rte_malloc.h>
  47 #include <rte_log.h>
  48
  49 #include "iotlb.h"
  50 #include "vhost.h"
  51 #include "vhost_user.h"
  52
  53 #define VIRTIO_MIN_MTU 68
  54 #define VIRTIO_MAX_MTU 65535
  55
  56 #define INFLIGHT_ALIGNMENT      64
  57 #define INFLIGHT_VERSION        0x1
  58
  59 static const char *vhost_message_str[VHOST_USER_MAX] = {
  60         [VHOST_USER_NONE] = "VHOST_USER_NONE",
  61         [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
  62         [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
  63         [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
  64         [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
  65         [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
  66         [VHOST_USER_SET_LOG_BASE] = "VHOST_USER_SET_LOG_BASE",
  67         [VHOST_USER_SET_LOG_FD] = "VHOST_USER_SET_LOG_FD",
  68         [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
  69         [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
  70         [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
  71         [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
  72         [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
  73         [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
  74         [VHOST_USER_SET_VRING_ERR]  = "VHOST_USER_SET_VRING_ERR",
  75         [VHOST_USER_GET_PROTOCOL_FEATURES]  = "VHOST_USER_GET_PROTOCOL_FEATURES",
  76         [VHOST_USER_SET_PROTOCOL_FEATURES]  = "VHOST_USER_SET_PROTOCOL_FEATURES",
  77         [VHOST_USER_GET_QUEUE_NUM]  = "VHOST_USER_GET_QUEUE_NUM",
  78         [VHOST_USER_SET_VRING_ENABLE]  = "VHOST_USER_SET_VRING_ENABLE",
  79         [VHOST_USER_SEND_RARP]  = "VHOST_USER_SEND_RARP",
  80         [VHOST_USER_NET_SET_MTU]  = "VHOST_USER_NET_SET_MTU",
  81         [VHOST_USER_SET_SLAVE_REQ_FD]  = "VHOST_USER_SET_SLAVE_REQ_FD",
  82         [VHOST_USER_IOTLB_MSG]  = "VHOST_USER_IOTLB_MSG",
  83         [VHOST_USER_CRYPTO_CREATE_SESS] = "VHOST_USER_CRYPTO_CREATE_SESS",
  84         [VHOST_USER_CRYPTO_CLOSE_SESS] = "VHOST_USER_CRYPTO_CLOSE_SESS",
  85         [VHOST_USER_POSTCOPY_ADVISE]  = "VHOST_USER_POSTCOPY_ADVISE",
  86         [VHOST_USER_POSTCOPY_LISTEN]  = "VHOST_USER_POSTCOPY_LISTEN",
  87         [VHOST_USER_POSTCOPY_END]  = "VHOST_USER_POSTCOPY_END",
  88         [VHOST_USER_GET_INFLIGHT_FD] = "VHOST_USER_GET_INFLIGHT_FD",
  89         [VHOST_USER_SET_INFLIGHT_FD] = "VHOST_USER_SET_INFLIGHT_FD",
  90 };
  91
  92 static int send_vhost_reply(int sockfd, struct VhostUserMsg *msg);
  93 static int read_vhost_message(int sockfd, struct VhostUserMsg *msg);
  94
  95 static uint64_t
  96 get_blk_size(int fd)
  97 {
  98         struct stat stat;
  99         int ret;
 100
 101         ret = fstat(fd, &stat);
 102         return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
 103 }
 104
 105 /*
 106  * Reclaim all the outstanding zmbufs for a virtqueue.
 107  */
 108 static void
 109 drain_zmbuf_list(struct vhost_virtqueue *vq)
 110 {
 111         struct zcopy_mbuf *zmbuf, *next;
 112
 113         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
 114              zmbuf != NULL; zmbuf = next) {
 115                 next = TAILQ_NEXT(zmbuf, next);
 116
 117                 while (!mbuf_is_consumed(zmbuf->mbuf))
 118                         usleep(1000);
 119
 120                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
 121                 restore_mbuf(zmbuf->mbuf);
 122                 rte_pktmbuf_free(zmbuf->mbuf);
 123                 put_zmbuf(zmbuf);
 124                 vq->nr_zmbuf -= 1;
 125         }
 126 }
 127
 128 static void
 129 free_mem_region(struct virtio_net *dev)
 130 {
 131         uint32_t i;
 132         struct rte_vhost_mem_region *reg;
 133         struct vhost_virtqueue *vq;
 134
 135         if (!dev || !dev->mem)
 136                 return;
 137
 138         if (dev->dequeue_zero_copy) {
 139                 for (i = 0; i < dev->nr_vring; i++) {
 140                         vq = dev->virtqueue[i];
 141                         if (vq)
 142                                 drain_zmbuf_list(vq);
 143                 }
 144         }
 145
 146         for (i = 0; i < dev->mem->nregions; i++) {
 147                 reg = &dev->mem->regions[i];
 148                 if (reg->host_user_addr) {
 149                         munmap(reg->mmap_addr, reg->mmap_size);
 150                         close(reg->fd);
 151                 }
 152         }
 153 }
 154
 155 void
 156 vhost_backend_cleanup(struct virtio_net *dev)
 157 {
 158         if (dev->mem) {
 159                 free_mem_region(dev);
 160                 rte_free(dev->mem);
 161                 dev->mem = NULL;
 162         }
 163
 164         free(dev->guest_pages);
 165         dev->guest_pages = NULL;
 166
 167         if (dev->log_addr) {
 168                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 169                 dev->log_addr = 0;
 170         }
 171
 172         if (dev->inflight_info) {
 173                 if (dev->inflight_info->addr) {
 174                         munmap(dev->inflight_info->addr,
 175                                dev->inflight_info->size);
 176                         dev->inflight_info->addr = NULL;
 177                 }
 178
 179                 if (dev->inflight_info->fd > 0) {
 180                         close(dev->inflight_info->fd);
 181                         dev->inflight_info->fd = -1;
 182                 }
 183
 184                 free(dev->inflight_info);
 185                 dev->inflight_info = NULL;
 186         }
 187
 188         if (dev->slave_req_fd >= 0) {
 189                 close(dev->slave_req_fd);
 190                 dev->slave_req_fd = -1;
 191         }
 192
 193         if (dev->postcopy_ufd >= 0) {
 194                 close(dev->postcopy_ufd);
 195                 dev->postcopy_ufd = -1;
 196         }
 197
 198         dev->postcopy_listening = 0;
 199 }
 200
 201 /*
 202  * This function just returns success at the moment unless
 203  * the device hasn't been initialised.
 204  */
 205 static int
 206 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
 207                         struct VhostUserMsg *msg __rte_unused,
 208                         int main_fd __rte_unused)
 209 {
 210         return RTE_VHOST_MSG_RESULT_OK;
 211 }
 212
 213 static int
 214 vhost_user_reset_owner(struct virtio_net **pdev,
 215                         struct VhostUserMsg *msg __rte_unused,
 216                         int main_fd __rte_unused)
 217 {
 218         struct virtio_net *dev = *pdev;
 219         vhost_destroy_device_notify(dev);
 220
 221         cleanup_device(dev, 0);
 222         reset_device(dev);
 223         return RTE_VHOST_MSG_RESULT_OK;
 224 }
 225
 226 /*
 227  * The features that we support are requested.
 228  */
 229 static int
 230 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 231                         int main_fd __rte_unused)
 232 {
 233         struct virtio_net *dev = *pdev;
 234         uint64_t features = 0;
 235
 236         rte_vhost_driver_get_features(dev->ifname, &features);
 237
 238         msg->payload.u64 = features;
 239         msg->size = sizeof(msg->payload.u64);
 240         msg->fd_num = 0;
 241
 242         return RTE_VHOST_MSG_RESULT_REPLY;
 243 }
 244
 245 /*
 246  * The queue number that we support are requested.
 247  */
 248 static int
 249 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
 250                         int main_fd __rte_unused)
 251 {
 252         struct virtio_net *dev = *pdev;
 253         uint32_t queue_num = 0;
 254
 255         rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
 256
 257         msg->payload.u64 = (uint64_t)queue_num;
 258         msg->size = sizeof(msg->payload.u64);
 259         msg->fd_num = 0;
 260
 261         return RTE_VHOST_MSG_RESULT_REPLY;
 262 }
 263
 264 /*
 265  * We receive the negotiated features supported by us and the virtio device.
 266  */
 267 static int
 268 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 269                         int main_fd __rte_unused)
 270 {
 271         struct virtio_net *dev = *pdev;
 272         uint64_t features = msg->payload.u64;
 273         uint64_t vhost_features = 0;
 274         struct rte_vdpa_device *vdpa_dev;
 275         int did = -1;
 276
 277         rte_vhost_driver_get_features(dev->ifname, &vhost_features);
 278         if (features & ~vhost_features) {
 279                 RTE_LOG(ERR, VHOST_CONFIG,
 280                         "(%d) received invalid negotiated features.\n",
 281                         dev->vid);
 282                 return RTE_VHOST_MSG_RESULT_ERR;
 283         }
 284
 285         if (dev->flags & VIRTIO_DEV_RUNNING) {
 286                 if (dev->features == features)
 287                         return RTE_VHOST_MSG_RESULT_OK;
 288
 289                 /*
 290                  * Error out if master tries to change features while device is
 291                  * in running state. The exception being VHOST_F_LOG_ALL, which
 292                  * is enabled when the live-migration starts.
 293                  */
 294                 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
 295                         RTE_LOG(ERR, VHOST_CONFIG,
 296                                 "(%d) features changed while device is running.\n",
 297                                 dev->vid);
 298                         return RTE_VHOST_MSG_RESULT_ERR;
 299                 }
 300
 301                 if (dev->notify_ops->features_changed)
 302                         dev->notify_ops->features_changed(dev->vid, features);
 303         }
 304
 305         dev->features = features;
 306         if (dev->features &
 307                 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
 308                 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 309         } else {
 310                 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
 311         }
 312         RTE_LOG(INFO, VHOST_CONFIG,
 313                 "negotiated Virtio features: 0x%" PRIx64 "\n", dev->features);
 314         VHOST_LOG_DEBUG(VHOST_CONFIG,
 315                 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
 316                 dev->vid,
 317                 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
 318                 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 319
 320         if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
 321             !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
 322                 /*
 323                  * Remove all but first queue pair if MQ hasn't been
 324                  * negotiated. This is safe because the device is not
 325                  * running at this stage.
 326                  */
 327                 while (dev->nr_vring > 2) {
 328                         struct vhost_virtqueue *vq;
 329
 330                         vq = dev->virtqueue[--dev->nr_vring];
 331                         if (!vq)
 332                                 continue;
 333
 334                         dev->virtqueue[dev->nr_vring] = NULL;
 335                         cleanup_vq(vq, 1);
 336                         cleanup_vq_inflight(dev, vq);
 337                         free_vq(dev, vq);
 338                 }
 339         }
 340
 341         did = dev->vdpa_dev_id;
 342         vdpa_dev = rte_vdpa_get_device(did);
 343         if (vdpa_dev && vdpa_dev->ops->set_features)
 344                 vdpa_dev->ops->set_features(dev->vid);
 345
 346         return RTE_VHOST_MSG_RESULT_OK;
 347 }
 348
 349 /*
 350  * The virtio device sends us the size of the descriptor ring.
 351  */
 352 static int
 353 vhost_user_set_vring_num(struct virtio_net **pdev,
 354                         struct VhostUserMsg *msg,
 355                         int main_fd __rte_unused)
 356 {
 357         struct virtio_net *dev = *pdev;
 358         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 359
 360         vq->size = msg->payload.state.num;
 361
 362         /* VIRTIO 1.0, 2.4 Virtqueues says:
 363          *
 364          *   Queue Size value is always a power of 2. The maximum Queue Size
 365          *   value is 32768.
 366          *
 367          * VIRTIO 1.1 2.7 Virtqueues says:
 368          *
 369          *   Packed virtqueues support up to 2^15 entries each.
 370          */
 371         if (!vq_is_packed(dev)) {
 372                 if (vq->size & (vq->size - 1)) {
 373                         RTE_LOG(ERR, VHOST_CONFIG,
 374                                 "invalid virtqueue size %u\n", vq->size);
 375                         return RTE_VHOST_MSG_RESULT_ERR;
 376                 }
 377         }
 378
 379         if (vq->size > 32768) {
 380                 RTE_LOG(ERR, VHOST_CONFIG,
 381                         "invalid virtqueue size %u\n", vq->size);
 382                 return RTE_VHOST_MSG_RESULT_ERR;
 383         }
 384
 385         if (dev->dequeue_zero_copy) {
 386                 vq->nr_zmbuf = 0;
 387                 vq->last_zmbuf_idx = 0;
 388                 vq->zmbuf_size = vq->size;
 389                 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
 390                                          sizeof(struct zcopy_mbuf), 0);
 391                 if (vq->zmbufs == NULL) {
 392                         RTE_LOG(WARNING, VHOST_CONFIG,
 393                                 "failed to allocate mem for zero copy; "
 394                                 "zero copy is force disabled\n");
 395                         dev->dequeue_zero_copy = 0;
 396                 }
 397                 TAILQ_INIT(&vq->zmbuf_list);
 398         }
 399
 400         if (vq_is_packed(dev)) {
 401                 vq->shadow_used_packed = rte_malloc(NULL,
 402                                 vq->size *
 403                                 sizeof(struct vring_used_elem_packed),
 404                                 RTE_CACHE_LINE_SIZE);
 405                 if (!vq->shadow_used_packed) {
 406                         RTE_LOG(ERR, VHOST_CONFIG,
 407                                         "failed to allocate memory for shadow used ring.\n");
 408                         return RTE_VHOST_MSG_RESULT_ERR;
 409                 }
 410
 411         } else {
 412                 vq->shadow_used_split = rte_malloc(NULL,
 413                                 vq->size * sizeof(struct vring_used_elem),
 414                                 RTE_CACHE_LINE_SIZE);
 415                 if (!vq->shadow_used_split) {
 416                         RTE_LOG(ERR, VHOST_CONFIG,
 417                                         "failed to allocate memory for shadow used ring.\n");
 418                         return RTE_VHOST_MSG_RESULT_ERR;
 419                 }
 420         }
 421
 422         vq->batch_copy_elems = rte_malloc(NULL,
 423                                 vq->size * sizeof(struct batch_copy_elem),
 424                                 RTE_CACHE_LINE_SIZE);
 425         if (!vq->batch_copy_elems) {
 426                 RTE_LOG(ERR, VHOST_CONFIG,
 427                         "failed to allocate memory for batching copy.\n");
 428                 return RTE_VHOST_MSG_RESULT_ERR;
 429         }
 430
 431         return RTE_VHOST_MSG_RESULT_OK;
 432 }
 433
 434 /*
 435  * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
 436  * same numa node as the memory of vring descriptor.
 437  */
 438 #ifdef RTE_LIBRTE_VHOST_NUMA
 439 static struct virtio_net*
 440 numa_realloc(struct virtio_net *dev, int index)
 441 {
 442         int oldnode, newnode;
 443         struct virtio_net *old_dev;
 444         struct vhost_virtqueue *old_vq, *vq;
 445         struct zcopy_mbuf *new_zmbuf;
 446         struct vring_used_elem *new_shadow_used_split;
 447         struct vring_used_elem_packed *new_shadow_used_packed;
 448         struct batch_copy_elem *new_batch_copy_elems;
 449         int ret;
 450
 451         if (dev->flags & VIRTIO_DEV_RUNNING)
 452                 return dev;
 453
 454         old_dev = dev;
 455         vq = old_vq = dev->virtqueue[index];
 456
 457         ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
 458                             MPOL_F_NODE | MPOL_F_ADDR);
 459
 460         /* check if we need to reallocate vq */
 461         ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
 462                              MPOL_F_NODE | MPOL_F_ADDR);
 463         if (ret) {
 464                 RTE_LOG(ERR, VHOST_CONFIG,
 465                         "Unable to get vq numa information.\n");
 466                 return dev;
 467         }
 468         if (oldnode != newnode) {
 469                 RTE_LOG(INFO, VHOST_CONFIG,
 470                         "reallocate vq from %d to %d node\n", oldnode, newnode);
 471                 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
 472                 if (!vq)
 473                         return dev;
 474
 475                 memcpy(vq, old_vq, sizeof(*vq));
 476                 TAILQ_INIT(&vq->zmbuf_list);
 477
 478                 if (dev->dequeue_zero_copy) {
 479                         new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
 480                                         sizeof(struct zcopy_mbuf), 0, newnode);
 481                         if (new_zmbuf) {
 482                                 rte_free(vq->zmbufs);
 483                                 vq->zmbufs = new_zmbuf;
 484                         }
 485                 }
 486
 487                 if (vq_is_packed(dev)) {
 488                         new_shadow_used_packed = rte_malloc_socket(NULL,
 489                                         vq->size *
 490                                         sizeof(struct vring_used_elem_packed),
 491                                         RTE_CACHE_LINE_SIZE,
 492                                         newnode);
 493                         if (new_shadow_used_packed) {
 494                                 rte_free(vq->shadow_used_packed);
 495                                 vq->shadow_used_packed = new_shadow_used_packed;
 496                         }
 497                 } else {
 498                         new_shadow_used_split = rte_malloc_socket(NULL,
 499                                         vq->size *
 500                                         sizeof(struct vring_used_elem),
 501                                         RTE_CACHE_LINE_SIZE,
 502                                         newnode);
 503                         if (new_shadow_used_split) {
 504                                 rte_free(vq->shadow_used_split);
 505                                 vq->shadow_used_split = new_shadow_used_split;
 506                         }
 507                 }
 508
 509                 new_batch_copy_elems = rte_malloc_socket(NULL,
 510                         vq->size * sizeof(struct batch_copy_elem),
 511                         RTE_CACHE_LINE_SIZE,
 512                         newnode);
 513                 if (new_batch_copy_elems) {
 514                         rte_free(vq->batch_copy_elems);
 515                         vq->batch_copy_elems = new_batch_copy_elems;
 516                 }
 517
 518                 rte_free(old_vq);
 519         }
 520
 521         /* check if we need to reallocate dev */
 522         ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
 523                             MPOL_F_NODE | MPOL_F_ADDR);
 524         if (ret) {
 525                 RTE_LOG(ERR, VHOST_CONFIG,
 526                         "Unable to get dev numa information.\n");
 527                 goto out;
 528         }
 529         if (oldnode != newnode) {
 530                 RTE_LOG(INFO, VHOST_CONFIG,
 531                         "reallocate dev from %d to %d node\n",
 532                         oldnode, newnode);
 533                 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
 534                 if (!dev) {
 535                         dev = old_dev;
 536                         goto out;
 537                 }
 538
 539                 memcpy(dev, old_dev, sizeof(*dev));
 540                 rte_free(old_dev);
 541         }
 542
 543 out:
 544         dev->virtqueue[index] = vq;
 545         vhost_devices[dev->vid] = dev;
 546
 547         if (old_vq != vq)
 548                 vhost_user_iotlb_init(dev, index);
 549
 550         return dev;
 551 }
 552 #else
 553 static struct virtio_net*
 554 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 555 {
 556         return dev;
 557 }
 558 #endif
 559
 560 /* Converts QEMU virtual address to Vhost virtual address. */
 561 static uint64_t
 562 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
 563 {
 564         struct rte_vhost_mem_region *r;
 565         uint32_t i;
 566
 567         if (unlikely(!dev || !dev->mem))
 568                 goto out_error;
 569
 570         /* Find the region where the address lives. */
 571         for (i = 0; i < dev->mem->nregions; i++) {
 572                 r = &dev->mem->regions[i];
 573
 574                 if (qva >= r->guest_user_addr &&
 575                     qva <  r->guest_user_addr + r->size) {
 576
 577                         if (unlikely(*len > r->guest_user_addr + r->size - qva))
 578                                 *len = r->guest_user_addr + r->size - qva;
 579
 580                         return qva - r->guest_user_addr +
 581                                r->host_user_addr;
 582                 }
 583         }
 584 out_error:
 585         *len = 0;
 586
 587         return 0;
 588 }
 589
 590
 591 /*
 592  * Converts ring address to Vhost virtual address.
 593  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 594  * else it is a QEMU virtual address.
 595  */
 596 static uint64_t
 597 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 598                 uint64_t ra, uint64_t *size)
 599 {
 600         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 601                 uint64_t vva;
 602                 uint64_t req_size = *size;
 603
 604                 vva = vhost_user_iotlb_cache_find(vq, ra,
 605                                         size, VHOST_ACCESS_RW);
 606                 if (req_size != *size)
 607                         vhost_user_iotlb_miss(dev, (ra + *size),
 608                                               VHOST_ACCESS_RW);
 609
 610                 return vva;
 611         }
 612
 613         return qva_to_vva(dev, ra, size);
 614 }
 615
 616 /*
 617  * Converts vring log address to GPA
 618  * If IOMMU is enabled, the log address is IOVA
 619  * If IOMMU not enabled, the log address is already GPA
 620  */
 621 static uint64_t
 622 translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
 623                 uint64_t log_addr)
 624 {
 625         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 626                 const uint64_t exp_size = sizeof(struct vring_used) +
 627                         sizeof(struct vring_used_elem) * vq->size;
 628                 uint64_t hva, gpa;
 629                 uint64_t size = exp_size;
 630
 631                 hva = vhost_iova_to_vva(dev, vq, log_addr,
 632                                         &size, VHOST_ACCESS_RW);
 633                 if (size != exp_size)
 634                         return 0;
 635
 636                 gpa = hva_to_gpa(dev, hva, exp_size);
 637                 if (!gpa) {
 638                         RTE_LOG(ERR, VHOST_CONFIG,
 639                                 "VQ: Failed to find GPA for log_addr: 0x%" PRIx64 " hva: 0x%" PRIx64 "\n",
 640                                 log_addr, hva);
 641                         return 0;
 642                 }
 643                 return gpa;
 644
 645         } else
 646                 return log_addr;
 647 }
 648
 649 static struct virtio_net *
 650 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 651 {
 652         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 653         struct vhost_vring_addr *addr = &vq->ring_addrs;
 654         uint64_t len, expected_len;
 655
 656         if (addr->flags & (1 << VHOST_VRING_F_LOG)) {
 657                 vq->log_guest_addr =
 658                         translate_log_addr(dev, vq, addr->log_guest_addr);
 659                 if (vq->log_guest_addr == 0) {
 660                         RTE_LOG(DEBUG, VHOST_CONFIG,
 661                                 "(%d) failed to map log_guest_addr.\n",
 662                                 dev->vid);
 663                         return dev;
 664                 }
 665         }
 666
 667         if (vq_is_packed(dev)) {
 668                 len = sizeof(struct vring_packed_desc) * vq->size;
 669                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 670                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 671                 if (vq->desc_packed == NULL ||
 672                                 len != sizeof(struct vring_packed_desc) *
 673                                 vq->size) {
 674                         RTE_LOG(DEBUG, VHOST_CONFIG,
 675                                 "(%d) failed to map desc_packed ring.\n",
 676                                 dev->vid);
 677                         return dev;
 678                 }
 679
 680                 dev = numa_realloc(dev, vq_index);
 681                 vq = dev->virtqueue[vq_index];
 682                 addr = &vq->ring_addrs;
 683
 684                 len = sizeof(struct vring_packed_desc_event);
 685                 vq->driver_event = (struct vring_packed_desc_event *)
 686                                         (uintptr_t)ring_addr_to_vva(dev,
 687                                         vq, addr->avail_user_addr, &len);
 688                 if (vq->driver_event == NULL ||
 689                                 len != sizeof(struct vring_packed_desc_event)) {
 690                         RTE_LOG(DEBUG, VHOST_CONFIG,
 691                                 "(%d) failed to find driver area address.\n",
 692                                 dev->vid);
 693                         return dev;
 694                 }
 695
 696                 len = sizeof(struct vring_packed_desc_event);
 697                 vq->device_event = (struct vring_packed_desc_event *)
 698                                         (uintptr_t)ring_addr_to_vva(dev,
 699                                         vq, addr->used_user_addr, &len);
 700                 if (vq->device_event == NULL ||
 701                                 len != sizeof(struct vring_packed_desc_event)) {
 702                         RTE_LOG(DEBUG, VHOST_CONFIG,
 703                                 "(%d) failed to find device area address.\n",
 704                                 dev->vid);
 705                         return dev;
 706                 }
 707
 708                 vq->access_ok = 1;
 709                 return dev;
 710         }
 711
 712         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 713         if (vq->desc && vq->avail && vq->used)
 714                 return dev;
 715
 716         len = sizeof(struct vring_desc) * vq->size;
 717         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 718                         vq, addr->desc_user_addr, &len);
 719         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 720                 RTE_LOG(DEBUG, VHOST_CONFIG,
 721                         "(%d) failed to map desc ring.\n",
 722                         dev->vid);
 723                 return dev;
 724         }
 725
 726         dev = numa_realloc(dev, vq_index);
 727         vq = dev->virtqueue[vq_index];
 728         addr = &vq->ring_addrs;
 729
 730         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 731         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 732                 len += sizeof(uint16_t);
 733         expected_len = len;
 734         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 735                         vq, addr->avail_user_addr, &len);
 736         if (vq->avail == 0 || len != expected_len) {
 737                 RTE_LOG(DEBUG, VHOST_CONFIG,
 738                         "(%d) failed to map avail ring.\n",
 739                         dev->vid);
 740                 return dev;
 741         }
 742
 743         len = sizeof(struct vring_used) +
 744                 sizeof(struct vring_used_elem) * vq->size;
 745         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 746                 len += sizeof(uint16_t);
 747         expected_len = len;
 748         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 749                         vq, addr->used_user_addr, &len);
 750         if (vq->used == 0 || len != expected_len) {
 751                 RTE_LOG(DEBUG, VHOST_CONFIG,
 752                         "(%d) failed to map used ring.\n",
 753                         dev->vid);
 754                 return dev;
 755         }
 756
 757         if (vq->last_used_idx != vq->used->idx) {
 758                 RTE_LOG(WARNING, VHOST_CONFIG,
 759                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 760                         "some packets maybe resent for Tx and dropped for Rx\n",
 761                         vq->last_used_idx, vq->used->idx);
 762                 vq->last_used_idx  = vq->used->idx;
 763                 vq->last_avail_idx = vq->used->idx;
 764         }
 765
 766         vq->access_ok = 1;
 767
 768         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
 769                         dev->vid, vq->desc);
 770         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
 771                         dev->vid, vq->avail);
 772         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
 773                         dev->vid, vq->used);
 774         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
 775                         dev->vid, vq->log_guest_addr);
 776
 777         return dev;
 778 }
 779
 780 /*
 781  * The virtio device sends us the desc, used and avail ring addresses.
 782  * This function then converts these to our address space.
 783  */
 784 static int
 785 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 786                         int main_fd __rte_unused)
 787 {
 788         struct virtio_net *dev = *pdev;
 789         struct vhost_virtqueue *vq;
 790         struct vhost_vring_addr *addr = &msg->payload.addr;
 791         bool access_ok;
 792
 793         if (dev->mem == NULL)
 794                 return RTE_VHOST_MSG_RESULT_ERR;
 795
 796         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 797         vq = dev->virtqueue[msg->payload.addr.index];
 798
 799         access_ok = vq->access_ok;
 800
 801         /*
 802          * Rings addresses should not be interpreted as long as the ring is not
 803          * started and enabled
 804          */
 805         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 806
 807         vring_invalidate(dev, vq);
 808
 809         if ((vq->enabled && (dev->features &
 810                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) ||
 811                         access_ok) {
 812                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 813                 if (!dev)
 814                         return RTE_VHOST_MSG_RESULT_ERR;
 815
 816                 *pdev = dev;
 817         }
 818
 819         return RTE_VHOST_MSG_RESULT_OK;
 820 }
 821
 822 /*
 823  * The virtio device sends us the available ring last used index.
 824  */
 825 static int
 826 vhost_user_set_vring_base(struct virtio_net **pdev,
 827                         struct VhostUserMsg *msg,
 828                         int main_fd __rte_unused)
 829 {
 830         struct virtio_net *dev = *pdev;
 831         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 832         uint64_t val = msg->payload.state.num;
 833
 834         if (vq_is_packed(dev)) {
 835                 /*
 836                  * Bit[0:14]: avail index
 837                  * Bit[15]: avail wrap counter
 838                  */
 839                 vq->last_avail_idx = val & 0x7fff;
 840                 vq->avail_wrap_counter = !!(val & (0x1 << 15));
 841                 /*
 842                  * Set used index to same value as available one, as
 843                  * their values should be the same since ring processing
 844                  * was stopped at get time.
 845                  */
 846                 vq->last_used_idx = vq->last_avail_idx;
 847                 vq->used_wrap_counter = vq->avail_wrap_counter;
 848         } else {
 849                 vq->last_used_idx = msg->payload.state.num;
 850                 vq->last_avail_idx = msg->payload.state.num;
 851         }
 852
 853         return RTE_VHOST_MSG_RESULT_OK;
 854 }
 855
 856 static int
 857 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 858                    uint64_t host_phys_addr, uint64_t size)
 859 {
 860         struct guest_page *page, *last_page;
 861         struct guest_page *old_pages;
 862
 863         if (dev->nr_guest_pages == dev->max_guest_pages) {
 864                 dev->max_guest_pages *= 2;
 865                 old_pages = dev->guest_pages;
 866                 dev->guest_pages = realloc(dev->guest_pages,
 867                                         dev->max_guest_pages * sizeof(*page));
 868                 if (!dev->guest_pages) {
 869                         RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
 870                         free(old_pages);
 871                         return -1;
 872                 }
 873         }
 874
 875         if (dev->nr_guest_pages > 0) {
 876                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 877                 /* merge if the two pages are continuous */
 878                 if (host_phys_addr == last_page->host_phys_addr +
 879                                       last_page->size) {
 880                         last_page->size += size;
 881                         return 0;
 882                 }
 883         }
 884
 885         page = &dev->guest_pages[dev->nr_guest_pages++];
 886         page->guest_phys_addr = guest_phys_addr;
 887         page->host_phys_addr  = host_phys_addr;
 888         page->size = size;
 889
 890         return 0;
 891 }
 892
 893 static int
 894 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 895                 uint64_t page_size)
 896 {
 897         uint64_t reg_size = reg->size;
 898         uint64_t host_user_addr  = reg->host_user_addr;
 899         uint64_t guest_phys_addr = reg->guest_phys_addr;
 900         uint64_t host_phys_addr;
 901         uint64_t size;
 902
 903         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 904         size = page_size - (guest_phys_addr & (page_size - 1));
 905         size = RTE_MIN(size, reg_size);
 906
 907         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 908                 return -1;
 909
 910         host_user_addr  += size;
 911         guest_phys_addr += size;
 912         reg_size -= size;
 913
 914         while (reg_size > 0) {
 915                 size = RTE_MIN(reg_size, page_size);
 916                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 917                                                   host_user_addr);
 918                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 919                                 size) < 0)
 920                         return -1;
 921
 922                 host_user_addr  += size;
 923                 guest_phys_addr += size;
 924                 reg_size -= size;
 925         }
 926
 927         return 0;
 928 }
 929
 930 #ifdef RTE_LIBRTE_VHOST_DEBUG
 931 /* TODO: enable it only in debug mode? */
 932 static void
 933 dump_guest_pages(struct virtio_net *dev)
 934 {
 935         uint32_t i;
 936         struct guest_page *page;
 937
 938         for (i = 0; i < dev->nr_guest_pages; i++) {
 939                 page = &dev->guest_pages[i];
 940
 941                 RTE_LOG(INFO, VHOST_CONFIG,
 942                         "guest physical page region %u\n"
 943                         "\t guest_phys_addr: %" PRIx64 "\n"
 944                         "\t host_phys_addr : %" PRIx64 "\n"
 945                         "\t size           : %" PRIx64 "\n",
 946                         i,
 947                         page->guest_phys_addr,
 948                         page->host_phys_addr,
 949                         page->size);
 950         }
 951 }
 952 #else
 953 #define dump_guest_pages(dev)
 954 #endif
 955
 956 static bool
 957 vhost_memory_changed(struct VhostUserMemory *new,
 958                      struct rte_vhost_memory *old)
 959 {
 960         uint32_t i;
 961
 962         if (new->nregions != old->nregions)
 963                 return true;
 964
 965         for (i = 0; i < new->nregions; ++i) {
 966                 VhostUserMemoryRegion *new_r = &new->regions[i];
 967                 struct rte_vhost_mem_region *old_r = &old->regions[i];
 968
 969                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
 970                         return true;
 971                 if (new_r->memory_size != old_r->size)
 972                         return true;
 973                 if (new_r->userspace_addr != old_r->guest_user_addr)
 974                         return true;
 975         }
 976
 977         return false;
 978 }
 979
 980 static int
 981 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
 982                         int main_fd)
 983 {
 984         struct virtio_net *dev = *pdev;
 985         struct VhostUserMemory *memory = &msg->payload.memory;
 986         struct rte_vhost_mem_region *reg;
 987         void *mmap_addr;
 988         uint64_t mmap_size;
 989         uint64_t mmap_offset;
 990         uint64_t alignment;
 991         uint32_t i;
 992         int populate;
 993         int fd;
 994
 995         if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
 996                 RTE_LOG(ERR, VHOST_CONFIG,
 997                         "too many memory regions (%u)\n", memory->nregions);
 998                 return RTE_VHOST_MSG_RESULT_ERR;
 999         }
1000
1001         if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
1002                 RTE_LOG(INFO, VHOST_CONFIG,
1003                         "(%d) memory regions not changed\n", dev->vid);
1004
1005                 for (i = 0; i < memory->nregions; i++)
1006                         close(msg->fds[i]);
1007
1008                 return RTE_VHOST_MSG_RESULT_OK;
1009         }
1010
1011         if (dev->mem) {
1012                 free_mem_region(dev);
1013                 rte_free(dev->mem);
1014                 dev->mem = NULL;
1015         }
1016
1017         /* Flush IOTLB cache as previous HVAs are now invalid */
1018         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1019                 for (i = 0; i < dev->nr_vring; i++)
1020                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
1021
1022         dev->nr_guest_pages = 0;
1023         if (!dev->guest_pages) {
1024                 dev->max_guest_pages = 8;
1025                 dev->guest_pages = malloc(dev->max_guest_pages *
1026                                                 sizeof(struct guest_page));
1027                 if (dev->guest_pages == NULL) {
1028                         RTE_LOG(ERR, VHOST_CONFIG,
1029                                 "(%d) failed to allocate memory "
1030                                 "for dev->guest_pages\n",
1031                                 dev->vid);
1032                         return RTE_VHOST_MSG_RESULT_ERR;
1033                 }
1034         }
1035
1036         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
1037                 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
1038         if (dev->mem == NULL) {
1039                 RTE_LOG(ERR, VHOST_CONFIG,
1040                         "(%d) failed to allocate memory for dev->mem\n",
1041                         dev->vid);
1042                 return RTE_VHOST_MSG_RESULT_ERR;
1043         }
1044         dev->mem->nregions = memory->nregions;
1045
1046         for (i = 0; i < memory->nregions; i++) {
1047                 fd  = msg->fds[i];
1048                 reg = &dev->mem->regions[i];
1049
1050                 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
1051                 reg->guest_user_addr = memory->regions[i].userspace_addr;
1052                 reg->size            = memory->regions[i].memory_size;
1053                 reg->fd              = fd;
1054
1055                 mmap_offset = memory->regions[i].mmap_offset;
1056
1057                 /* Check for memory_size + mmap_offset overflow */
1058                 if (mmap_offset >= -reg->size) {
1059                         RTE_LOG(ERR, VHOST_CONFIG,
1060                                 "mmap_offset (%#"PRIx64") and memory_size "
1061                                 "(%#"PRIx64") overflow\n",
1062                                 mmap_offset, reg->size);
1063                         goto err_mmap;
1064                 }
1065
1066                 mmap_size = reg->size + mmap_offset;
1067
1068                 /* mmap() without flag of MAP_ANONYMOUS, should be called
1069                  * with length argument aligned with hugepagesz at older
1070                  * longterm version Linux, like 2.6.32 and 3.2.72, or
1071                  * mmap() will fail with EINVAL.
1072                  *
1073                  * to avoid failure, make sure in caller to keep length
1074                  * aligned.
1075                  */
1076                 alignment = get_blk_size(fd);
1077                 if (alignment == (uint64_t)-1) {
1078                         RTE_LOG(ERR, VHOST_CONFIG,
1079                                 "couldn't get hugepage size through fstat\n");
1080                         goto err_mmap;
1081                 }
1082                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
1083
1084                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
1085                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
1086                                  MAP_SHARED | populate, fd, 0);
1087
1088                 if (mmap_addr == MAP_FAILED) {
1089                         RTE_LOG(ERR, VHOST_CONFIG,
1090                                 "mmap region %u failed.\n", i);
1091                         goto err_mmap;
1092                 }
1093
1094                 reg->mmap_addr = mmap_addr;
1095                 reg->mmap_size = mmap_size;
1096                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
1097                                       mmap_offset;
1098
1099                 if (dev->dequeue_zero_copy)
1100                         if (add_guest_pages(dev, reg, alignment) < 0) {
1101                                 RTE_LOG(ERR, VHOST_CONFIG,
1102                                         "adding guest pages to region %u failed.\n",
1103                                         i);
1104                                 goto err_mmap;
1105                         }
1106
1107                 RTE_LOG(INFO, VHOST_CONFIG,
1108                         "guest memory region %u, size: 0x%" PRIx64 "\n"
1109                         "\t guest physical addr: 0x%" PRIx64 "\n"
1110                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
1111                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
1112                         "\t mmap addr : 0x%" PRIx64 "\n"
1113                         "\t mmap size : 0x%" PRIx64 "\n"
1114                         "\t mmap align: 0x%" PRIx64 "\n"
1115                         "\t mmap off  : 0x%" PRIx64 "\n",
1116                         i, reg->size,
1117                         reg->guest_phys_addr,
1118                         reg->guest_user_addr,
1119                         reg->host_user_addr,
1120                         (uint64_t)(uintptr_t)mmap_addr,
1121                         mmap_size,
1122                         alignment,
1123                         mmap_offset);
1124
1125                 if (dev->postcopy_listening) {
1126                         /*
1127                          * We haven't a better way right now than sharing
1128                          * DPDK's virtual address with Qemu, so that Qemu can
1129                          * retrieve the region offset when handling userfaults.
1130                          */
1131                         memory->regions[i].userspace_addr =
1132                                 reg->host_user_addr;
1133                 }
1134         }
1135         if (dev->postcopy_listening) {
1136                 /* Send the addresses back to qemu */
1137                 msg->fd_num = 0;
1138                 send_vhost_reply(main_fd, msg);
1139
1140                 /* Wait for qemu to acknolwedge it's got the addresses
1141                  * we've got to wait before we're allowed to generate faults.
1142                  */
1143                 VhostUserMsg ack_msg;
1144                 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1145                         RTE_LOG(ERR, VHOST_CONFIG,
1146                                 "Failed to read qemu ack on postcopy set-mem-table\n");
1147                         goto err_mmap;
1148                 }
1149                 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1150                         RTE_LOG(ERR, VHOST_CONFIG,
1151                                 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1152                                 ack_msg.request.master);
1153                         goto err_mmap;
1154                 }
1155
1156                 /* Now userfault register and we can use the memory */
1157                 for (i = 0; i < memory->nregions; i++) {
1158 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1159                         reg = &dev->mem->regions[i];
1160                         struct uffdio_register reg_struct;
1161
1162                         /*
1163                          * Let's register all the mmap'ed area to ensure
1164                          * alignment on page boundary.
1165                          */
1166                         reg_struct.range.start =
1167                                 (uint64_t)(uintptr_t)reg->mmap_addr;
1168                         reg_struct.range.len = reg->mmap_size;
1169                         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1170
1171                         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1172                                                 &reg_struct)) {
1173                                 RTE_LOG(ERR, VHOST_CONFIG,
1174                                         "Failed to register ufd for region %d: (ufd = %d) %s\n",
1175                                         i, dev->postcopy_ufd,
1176                                         strerror(errno));
1177                                 goto err_mmap;
1178                         }
1179                         RTE_LOG(INFO, VHOST_CONFIG,
1180                                 "\t userfaultfd registered for range : "
1181                                 "%" PRIx64 " - %" PRIx64 "\n",
1182                                 (uint64_t)reg_struct.range.start,
1183                                 (uint64_t)reg_struct.range.start +
1184                                 (uint64_t)reg_struct.range.len - 1);
1185 #else
1186                         goto err_mmap;
1187 #endif
1188                 }
1189         }
1190
1191         for (i = 0; i < dev->nr_vring; i++) {
1192                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1193
1194                 if (vq->desc || vq->avail || vq->used) {
1195                         /*
1196                          * If the memory table got updated, the ring addresses
1197                          * need to be translated again as virtual addresses have
1198                          * changed.
1199                          */
1200                         vring_invalidate(dev, vq);
1201
1202                         dev = translate_ring_addresses(dev, i);
1203                         if (!dev) {
1204                                 dev = *pdev;
1205                                 goto err_mmap;
1206                         }
1207
1208                         *pdev = dev;
1209                 }
1210         }
1211
1212         dump_guest_pages(dev);
1213
1214         return RTE_VHOST_MSG_RESULT_OK;
1215
1216 err_mmap:
1217         free_mem_region(dev);
1218         rte_free(dev->mem);
1219         dev->mem = NULL;
1220         return RTE_VHOST_MSG_RESULT_ERR;
1221 }
1222
1223 static bool
1224 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1225 {
1226         bool rings_ok;
1227
1228         if (!vq)
1229                 return false;
1230
1231         if (vq_is_packed(dev))
1232                 rings_ok = !!vq->desc_packed;
1233         else
1234                 rings_ok = vq->desc && vq->avail && vq->used;
1235
1236         return rings_ok &&
1237                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1238                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1239 }
1240
1241 static int
1242 virtio_is_ready(struct virtio_net *dev)
1243 {
1244         struct vhost_virtqueue *vq;
1245         uint32_t i;
1246
1247         if (dev->nr_vring == 0)
1248                 return 0;
1249
1250         for (i = 0; i < dev->nr_vring; i++) {
1251                 vq = dev->virtqueue[i];
1252
1253                 if (!vq_is_ready(dev, vq))
1254                         return 0;
1255         }
1256
1257         RTE_LOG(INFO, VHOST_CONFIG,
1258                 "virtio is now ready for processing.\n");
1259         return 1;
1260 }
1261
1262 static void *
1263 inflight_mem_alloc(const char *name, size_t size, int *fd)
1264 {
1265         void *ptr;
1266         int mfd = -1;
1267         char fname[20] = "/tmp/memfd-XXXXXX";
1268
1269         *fd = -1;
1270 #ifdef MEMFD_SUPPORTED
1271         mfd = memfd_create(name, MFD_CLOEXEC);
1272 #else
1273         RTE_SET_USED(name);
1274 #endif
1275         if (mfd == -1) {
1276                 mfd = mkstemp(fname);
1277                 if (mfd == -1) {
1278                         RTE_LOG(ERR, VHOST_CONFIG,
1279                                 "failed to get inflight buffer fd\n");
1280                         return NULL;
1281                 }
1282
1283                 unlink(fname);
1284         }
1285
1286         if (ftruncate(mfd, size) == -1) {
1287                 RTE_LOG(ERR, VHOST_CONFIG,
1288                         "failed to alloc inflight buffer\n");
1289                 close(mfd);
1290                 return NULL;
1291         }
1292
1293         ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd, 0);
1294         if (ptr == MAP_FAILED) {
1295                 RTE_LOG(ERR, VHOST_CONFIG,
1296                         "failed to mmap inflight buffer\n");
1297                 close(mfd);
1298                 return NULL;
1299         }
1300
1301         *fd = mfd;
1302         return ptr;
1303 }
1304
1305 static uint32_t
1306 get_pervq_shm_size_split(uint16_t queue_size)
1307 {
1308         return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_split) *
1309                                   queue_size + sizeof(uint64_t) +
1310                                   sizeof(uint16_t) * 4, INFLIGHT_ALIGNMENT);
1311 }
1312
1313 static uint32_t
1314 get_pervq_shm_size_packed(uint16_t queue_size)
1315 {
1316         return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_packed)
1317                                   * queue_size + sizeof(uint64_t) +
1318                                   sizeof(uint16_t) * 6 + sizeof(uint8_t) * 9,
1319                                   INFLIGHT_ALIGNMENT);
1320 }
1321
1322 static int
1323 vhost_user_get_inflight_fd(struct virtio_net **pdev,
1324                            VhostUserMsg *msg,
1325                            int main_fd __rte_unused)
1326 {
1327         struct rte_vhost_inflight_info_packed *inflight_packed;
1328         uint64_t pervq_inflight_size, mmap_size;
1329         uint16_t num_queues, queue_size;
1330         struct virtio_net *dev = *pdev;
1331         int fd, i, j;
1332         void *addr;
1333
1334         if (msg->size != sizeof(msg->payload.inflight)) {
1335                 RTE_LOG(ERR, VHOST_CONFIG,
1336                         "invalid get_inflight_fd message size is %d\n",
1337                         msg->size);
1338                 return RTE_VHOST_MSG_RESULT_ERR;
1339         }
1340
1341         if (dev->inflight_info == NULL) {
1342                 dev->inflight_info = calloc(1,
1343                                             sizeof(struct inflight_mem_info));
1344                 if (!dev->inflight_info) {
1345                         RTE_LOG(ERR, VHOST_CONFIG,
1346                                 "failed to alloc dev inflight area\n");
1347                         return RTE_VHOST_MSG_RESULT_ERR;
1348                 }
1349         }
1350
1351         num_queues = msg->payload.inflight.num_queues;
1352         queue_size = msg->payload.inflight.queue_size;
1353
1354         RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd num_queues: %u\n",
1355                 msg->payload.inflight.num_queues);
1356         RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd queue_size: %u\n",
1357                 msg->payload.inflight.queue_size);
1358
1359         if (vq_is_packed(dev))
1360                 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1361         else
1362                 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1363
1364         mmap_size = num_queues * pervq_inflight_size;
1365         addr = inflight_mem_alloc("vhost-inflight", mmap_size, &fd);
1366         if (!addr) {
1367                 RTE_LOG(ERR, VHOST_CONFIG,
1368                         "failed to alloc vhost inflight area\n");
1369                         msg->payload.inflight.mmap_size = 0;
1370                 return RTE_VHOST_MSG_RESULT_ERR;
1371         }
1372         memset(addr, 0, mmap_size);
1373
1374         dev->inflight_info->addr = addr;
1375         dev->inflight_info->size = msg->payload.inflight.mmap_size = mmap_size;
1376         dev->inflight_info->fd = msg->fds[0] = fd;
1377         msg->payload.inflight.mmap_offset = 0;
1378         msg->fd_num = 1;
1379
1380         if (vq_is_packed(dev)) {
1381                 for (i = 0; i < num_queues; i++) {
1382                         inflight_packed =
1383                                 (struct rte_vhost_inflight_info_packed *)addr;
1384                         inflight_packed->used_wrap_counter = 1;
1385                         inflight_packed->old_used_wrap_counter = 1;
1386                         for (j = 0; j < queue_size; j++)
1387                                 inflight_packed->desc[j].next = j + 1;
1388                         addr = (void *)((char *)addr + pervq_inflight_size);
1389                 }
1390         }
1391
1392         RTE_LOG(INFO, VHOST_CONFIG,
1393                 "send inflight mmap_size: %"PRIu64"\n",
1394                 msg->payload.inflight.mmap_size);
1395         RTE_LOG(INFO, VHOST_CONFIG,
1396                 "send inflight mmap_offset: %"PRIu64"\n",
1397                 msg->payload.inflight.mmap_offset);
1398         RTE_LOG(INFO, VHOST_CONFIG,
1399                 "send inflight fd: %d\n", msg->fds[0]);
1400
1401         return RTE_VHOST_MSG_RESULT_REPLY;
1402 }
1403
1404 static int
1405 vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
1406                            int main_fd __rte_unused)
1407 {
1408         uint64_t mmap_size, mmap_offset;
1409         uint16_t num_queues, queue_size;
1410         struct virtio_net *dev = *pdev;
1411         uint32_t pervq_inflight_size;
1412         struct vhost_virtqueue *vq;
1413         void *addr;
1414         int fd, i;
1415
1416         fd = msg->fds[0];
1417         if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
1418                 RTE_LOG(ERR, VHOST_CONFIG,
1419                         "invalid set_inflight_fd message size is %d,fd is %d\n",
1420                         msg->size, fd);
1421                 return RTE_VHOST_MSG_RESULT_ERR;
1422         }
1423
1424         mmap_size = msg->payload.inflight.mmap_size;
1425         mmap_offset = msg->payload.inflight.mmap_offset;
1426         num_queues = msg->payload.inflight.num_queues;
1427         queue_size = msg->payload.inflight.queue_size;
1428
1429         if (vq_is_packed(dev))
1430                 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1431         else
1432                 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1433
1434         RTE_LOG(INFO, VHOST_CONFIG,
1435                 "set_inflight_fd mmap_size: %"PRIu64"\n", mmap_size);
1436         RTE_LOG(INFO, VHOST_CONFIG,
1437                 "set_inflight_fd mmap_offset: %"PRIu64"\n", mmap_offset);
1438         RTE_LOG(INFO, VHOST_CONFIG,
1439                 "set_inflight_fd num_queues: %u\n", num_queues);
1440         RTE_LOG(INFO, VHOST_CONFIG,
1441                 "set_inflight_fd queue_size: %u\n", queue_size);
1442         RTE_LOG(INFO, VHOST_CONFIG,
1443                 "set_inflight_fd fd: %d\n", fd);
1444         RTE_LOG(INFO, VHOST_CONFIG,
1445                 "set_inflight_fd pervq_inflight_size: %d\n",
1446                 pervq_inflight_size);
1447
1448         if (!dev->inflight_info) {
1449                 dev->inflight_info = calloc(1,
1450                                             sizeof(struct inflight_mem_info));
1451                 if (dev->inflight_info == NULL) {
1452                         RTE_LOG(ERR, VHOST_CONFIG,
1453                                 "failed to alloc dev inflight area\n");
1454                         return RTE_VHOST_MSG_RESULT_ERR;
1455                 }
1456         }
1457
1458         if (dev->inflight_info->addr)
1459                 munmap(dev->inflight_info->addr, dev->inflight_info->size);
1460
1461         addr = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
1462                     fd, mmap_offset);
1463         if (addr == MAP_FAILED) {
1464                 RTE_LOG(ERR, VHOST_CONFIG, "failed to mmap share memory.\n");
1465                 return RTE_VHOST_MSG_RESULT_ERR;
1466         }
1467
1468         if (dev->inflight_info->fd)
1469                 close(dev->inflight_info->fd);
1470
1471         dev->inflight_info->fd = fd;
1472         dev->inflight_info->addr = addr;
1473         dev->inflight_info->size = mmap_size;
1474
1475         for (i = 0; i < num_queues; i++) {
1476                 vq = dev->virtqueue[i];
1477                 if (vq_is_packed(dev)) {
1478                         vq->inflight_packed = addr;
1479                         vq->inflight_packed->desc_num = queue_size;
1480                 } else {
1481                         vq->inflight_split = addr;
1482                         vq->inflight_split->desc_num = queue_size;
1483                 }
1484                 addr = (void *)((char *)addr + pervq_inflight_size);
1485         }
1486
1487         return RTE_VHOST_MSG_RESULT_OK;
1488 }
1489
1490 static int
1491 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1492                         int main_fd __rte_unused)
1493 {
1494         struct virtio_net *dev = *pdev;
1495         struct vhost_vring_file file;
1496         struct vhost_virtqueue *vq;
1497
1498         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1499         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1500                 file.fd = VIRTIO_INVALID_EVENTFD;
1501         else
1502                 file.fd = msg->fds[0];
1503         RTE_LOG(INFO, VHOST_CONFIG,
1504                 "vring call idx:%d file:%d\n", file.index, file.fd);
1505
1506         vq = dev->virtqueue[file.index];
1507         if (vq->callfd >= 0)
1508                 close(vq->callfd);
1509
1510         vq->callfd = file.fd;
1511
1512         return RTE_VHOST_MSG_RESULT_OK;
1513 }
1514
1515 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1516                         struct VhostUserMsg *msg,
1517                         int main_fd __rte_unused)
1518 {
1519         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1520                 close(msg->fds[0]);
1521         RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1522
1523         return RTE_VHOST_MSG_RESULT_OK;
1524 }
1525
1526 static int
1527 resubmit_desc_compare(const void *a, const void *b)
1528 {
1529         const struct rte_vhost_resubmit_desc *desc0 = a;
1530         const struct rte_vhost_resubmit_desc *desc1 = b;
1531
1532         if (desc1->counter > desc0->counter)
1533                 return 1;
1534
1535         return -1;
1536 }
1537
1538 static int
1539 vhost_check_queue_inflights_split(struct virtio_net *dev,
1540                                   struct vhost_virtqueue *vq)
1541 {
1542         uint16_t i;
1543         uint16_t resubmit_num = 0, last_io, num;
1544         struct vring_used *used = vq->used;
1545         struct rte_vhost_resubmit_info *resubmit;
1546         struct rte_vhost_inflight_info_split *inflight_split;
1547
1548         if (!(dev->protocol_features &
1549             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1550                 return RTE_VHOST_MSG_RESULT_OK;
1551
1552         if ((!vq->inflight_split))
1553                 return RTE_VHOST_MSG_RESULT_ERR;
1554
1555         if (!vq->inflight_split->version) {
1556                 vq->inflight_split->version = INFLIGHT_VERSION;
1557                 return RTE_VHOST_MSG_RESULT_OK;
1558         }
1559
1560         if (vq->resubmit_inflight)
1561                 return RTE_VHOST_MSG_RESULT_OK;
1562
1563         inflight_split = vq->inflight_split;
1564         vq->global_counter = 0;
1565         last_io = inflight_split->last_inflight_io;
1566
1567         if (inflight_split->used_idx != used->idx) {
1568                 inflight_split->desc[last_io].inflight = 0;
1569                 rte_smp_mb();
1570                 inflight_split->used_idx = used->idx;
1571         }
1572
1573         for (i = 0; i < inflight_split->desc_num; i++) {
1574                 if (inflight_split->desc[i].inflight == 1)
1575                         resubmit_num++;
1576         }
1577
1578         vq->last_avail_idx += resubmit_num;
1579
1580         if (resubmit_num) {
1581                 resubmit  = calloc(1, sizeof(struct rte_vhost_resubmit_info));
1582                 if (!resubmit) {
1583                         RTE_LOG(ERR, VHOST_CONFIG,
1584                                 "failed to allocate memory for resubmit info.\n");
1585                         return RTE_VHOST_MSG_RESULT_ERR;
1586                 }
1587
1588                 resubmit->resubmit_list = calloc(resubmit_num,
1589                         sizeof(struct rte_vhost_resubmit_desc));
1590                 if (!resubmit->resubmit_list) {
1591                         RTE_LOG(ERR, VHOST_CONFIG,
1592                                 "failed to allocate memory for inflight desc.\n");
1593                         free(resubmit);
1594                         return RTE_VHOST_MSG_RESULT_ERR;
1595                 }
1596
1597                 num = 0;
1598                 for (i = 0; i < vq->inflight_split->desc_num; i++) {
1599                         if (vq->inflight_split->desc[i].inflight == 1) {
1600                                 resubmit->resubmit_list[num].index = i;
1601                                 resubmit->resubmit_list[num].counter =
1602                                         inflight_split->desc[i].counter;
1603                                 num++;
1604                         }
1605                 }
1606                 resubmit->resubmit_num = num;
1607
1608                 if (resubmit->resubmit_num > 1)
1609                         qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1610                               sizeof(struct rte_vhost_resubmit_desc),
1611                               resubmit_desc_compare);
1612
1613                 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1614                 vq->resubmit_inflight = resubmit;
1615         }
1616
1617         return RTE_VHOST_MSG_RESULT_OK;
1618 }
1619
1620 static int
1621 vhost_check_queue_inflights_packed(struct virtio_net *dev,
1622                                    struct vhost_virtqueue *vq)
1623 {
1624         uint16_t i;
1625         uint16_t resubmit_num = 0, old_used_idx, num;
1626         struct rte_vhost_resubmit_info *resubmit;
1627         struct rte_vhost_inflight_info_packed *inflight_packed;
1628
1629         if (!(dev->protocol_features &
1630             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1631                 return RTE_VHOST_MSG_RESULT_OK;
1632
1633         if ((!vq->inflight_packed))
1634                 return RTE_VHOST_MSG_RESULT_ERR;
1635
1636         if (!vq->inflight_packed->version) {
1637                 vq->inflight_packed->version = INFLIGHT_VERSION;
1638                 return RTE_VHOST_MSG_RESULT_OK;
1639         }
1640
1641         if (vq->resubmit_inflight)
1642                 return RTE_VHOST_MSG_RESULT_OK;
1643
1644         inflight_packed = vq->inflight_packed;
1645         vq->global_counter = 0;
1646         old_used_idx = inflight_packed->old_used_idx;
1647
1648         if (inflight_packed->used_idx != old_used_idx) {
1649                 if (inflight_packed->desc[old_used_idx].inflight == 0) {
1650                         inflight_packed->old_used_idx =
1651                                 inflight_packed->used_idx;
1652                         inflight_packed->old_used_wrap_counter =
1653                                 inflight_packed->used_wrap_counter;
1654                         inflight_packed->old_free_head =
1655                                 inflight_packed->free_head;
1656                 } else {
1657                         inflight_packed->used_idx =
1658                                 inflight_packed->old_used_idx;
1659                         inflight_packed->used_wrap_counter =
1660                                 inflight_packed->old_used_wrap_counter;
1661                         inflight_packed->free_head =
1662                                 inflight_packed->old_free_head;
1663                 }
1664         }
1665
1666         for (i = 0; i < inflight_packed->desc_num; i++) {
1667                 if (inflight_packed->desc[i].inflight == 1)
1668                         resubmit_num++;
1669         }
1670
1671         if (resubmit_num) {
1672                 resubmit = calloc(1, sizeof(struct rte_vhost_resubmit_info));
1673                 if (resubmit == NULL) {
1674                         RTE_LOG(ERR, VHOST_CONFIG,
1675                                 "failed to allocate memory for resubmit info.\n");
1676                         return RTE_VHOST_MSG_RESULT_ERR;
1677                 }
1678
1679                 resubmit->resubmit_list = calloc(resubmit_num,
1680                         sizeof(struct rte_vhost_resubmit_desc));
1681                 if (resubmit->resubmit_list == NULL) {
1682                         RTE_LOG(ERR, VHOST_CONFIG,
1683                                 "failed to allocate memory for resubmit desc.\n");
1684                         free(resubmit);
1685                         return RTE_VHOST_MSG_RESULT_ERR;
1686                 }
1687
1688                 num = 0;
1689                 for (i = 0; i < inflight_packed->desc_num; i++) {
1690                         if (vq->inflight_packed->desc[i].inflight == 1) {
1691                                 resubmit->resubmit_list[num].index = i;
1692                                 resubmit->resubmit_list[num].counter =
1693                                         inflight_packed->desc[i].counter;
1694                                 num++;
1695                         }
1696                 }
1697                 resubmit->resubmit_num = num;
1698
1699                 if (resubmit->resubmit_num > 1)
1700                         qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1701                               sizeof(struct rte_vhost_resubmit_desc),
1702                               resubmit_desc_compare);
1703
1704                 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1705                 vq->resubmit_inflight = resubmit;
1706         }
1707
1708         return RTE_VHOST_MSG_RESULT_OK;
1709 }
1710
1711 static int
1712 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1713                         int main_fd __rte_unused)
1714 {
1715         struct virtio_net *dev = *pdev;
1716         struct vhost_vring_file file;
1717         struct vhost_virtqueue *vq;
1718
1719         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1720         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1721                 file.fd = VIRTIO_INVALID_EVENTFD;
1722         else
1723                 file.fd = msg->fds[0];
1724         RTE_LOG(INFO, VHOST_CONFIG,
1725                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1726
1727         /* Interpret ring addresses only when ring is started. */
1728         dev = translate_ring_addresses(dev, file.index);
1729         if (!dev)
1730                 return RTE_VHOST_MSG_RESULT_ERR;
1731
1732         *pdev = dev;
1733
1734         vq = dev->virtqueue[file.index];
1735
1736         /*
1737          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1738          * the ring starts already enabled. Otherwise, it is enabled via
1739          * the SET_VRING_ENABLE message.
1740          */
1741         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
1742                 vq->enabled = 1;
1743                 if (dev->notify_ops->vring_state_changed)
1744                         dev->notify_ops->vring_state_changed(
1745                                 dev->vid, file.index, 1);
1746         }
1747
1748         if (vq->kickfd >= 0)
1749                 close(vq->kickfd);
1750         vq->kickfd = file.fd;
1751
1752         if (vq_is_packed(dev)) {
1753                 if (vhost_check_queue_inflights_packed(dev, vq)) {
1754                         RTE_LOG(ERR, VHOST_CONFIG,
1755                                 "failed to inflights for vq: %d\n", file.index);
1756                         return RTE_VHOST_MSG_RESULT_ERR;
1757                 }
1758         } else {
1759                 if (vhost_check_queue_inflights_split(dev, vq)) {
1760                         RTE_LOG(ERR, VHOST_CONFIG,
1761                                 "failed to inflights for vq: %d\n", file.index);
1762                         return RTE_VHOST_MSG_RESULT_ERR;
1763                 }
1764         }
1765
1766         return RTE_VHOST_MSG_RESULT_OK;
1767 }
1768
1769 static void
1770 free_zmbufs(struct vhost_virtqueue *vq)
1771 {
1772         drain_zmbuf_list(vq);
1773
1774         rte_free(vq->zmbufs);
1775 }
1776
1777 /*
1778  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1779  */
1780 static int
1781 vhost_user_get_vring_base(struct virtio_net **pdev,
1782                         struct VhostUserMsg *msg,
1783                         int main_fd __rte_unused)
1784 {
1785         struct virtio_net *dev = *pdev;
1786         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1787         uint64_t val;
1788
1789         /* We have to stop the queue (virtio) if it is running. */
1790         vhost_destroy_device_notify(dev);
1791
1792         dev->flags &= ~VIRTIO_DEV_READY;
1793         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1794
1795         /* Here we are safe to get the indexes */
1796         if (vq_is_packed(dev)) {
1797                 /*
1798                  * Bit[0:14]: avail index
1799                  * Bit[15]: avail wrap counter
1800                  */
1801                 val = vq->last_avail_idx & 0x7fff;
1802                 val |= vq->avail_wrap_counter << 15;
1803                 msg->payload.state.num = val;
1804         } else {
1805                 msg->payload.state.num = vq->last_avail_idx;
1806         }
1807
1808         RTE_LOG(INFO, VHOST_CONFIG,
1809                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1810                 msg->payload.state.num);
1811         /*
1812          * Based on current qemu vhost-user implementation, this message is
1813          * sent and only sent in vhost_vring_stop.
1814          * TODO: cleanup the vring, it isn't usable since here.
1815          */
1816         if (vq->kickfd >= 0)
1817                 close(vq->kickfd);
1818
1819         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1820
1821         if (vq->callfd >= 0)
1822                 close(vq->callfd);
1823
1824         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1825
1826         vq->signalled_used_valid = false;
1827
1828         if (dev->dequeue_zero_copy)
1829                 free_zmbufs(vq);
1830         if (vq_is_packed(dev)) {
1831                 rte_free(vq->shadow_used_packed);
1832                 vq->shadow_used_packed = NULL;
1833         } else {
1834                 rte_free(vq->shadow_used_split);
1835                 vq->shadow_used_split = NULL;
1836         }
1837
1838         rte_free(vq->batch_copy_elems);
1839         vq->batch_copy_elems = NULL;
1840
1841         msg->size = sizeof(msg->payload.state);
1842         msg->fd_num = 0;
1843
1844         vring_invalidate(dev, vq);
1845
1846         return RTE_VHOST_MSG_RESULT_REPLY;
1847 }
1848
1849 /*
1850  * when virtio queues are ready to work, qemu will send us to
1851  * enable the virtio queue pair.
1852  */
1853 static int
1854 vhost_user_set_vring_enable(struct virtio_net **pdev,
1855                         struct VhostUserMsg *msg,
1856                         int main_fd __rte_unused)
1857 {
1858         struct virtio_net *dev = *pdev;
1859         int enable = (int)msg->payload.state.num;
1860         int index = (int)msg->payload.state.index;
1861         struct rte_vdpa_device *vdpa_dev;
1862         int did = -1;
1863
1864         RTE_LOG(INFO, VHOST_CONFIG,
1865                 "set queue enable: %d to qp idx: %d\n",
1866                 enable, index);
1867
1868         did = dev->vdpa_dev_id;
1869         vdpa_dev = rte_vdpa_get_device(did);
1870         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1871                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1872
1873         if (dev->notify_ops->vring_state_changed)
1874                 dev->notify_ops->vring_state_changed(dev->vid,
1875                                 index, enable);
1876
1877         /* On disable, rings have to be stopped being processed. */
1878         if (!enable && dev->dequeue_zero_copy)
1879                 drain_zmbuf_list(dev->virtqueue[index]);
1880
1881         dev->virtqueue[index]->enabled = enable;
1882
1883         return RTE_VHOST_MSG_RESULT_OK;
1884 }
1885
1886 static int
1887 vhost_user_get_protocol_features(struct virtio_net **pdev,
1888                         struct VhostUserMsg *msg,
1889                         int main_fd __rte_unused)
1890 {
1891         struct virtio_net *dev = *pdev;
1892         uint64_t features, protocol_features;
1893
1894         rte_vhost_driver_get_features(dev->ifname, &features);
1895         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1896
1897         /*
1898          * REPLY_ACK protocol feature is only mandatory for now
1899          * for IOMMU feature. If IOMMU is explicitly disabled by the
1900          * application, disable also REPLY_ACK feature for older buggy
1901          * Qemu versions (from v2.7.0 to v2.9.0).
1902          */
1903         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1904                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1905
1906         msg->payload.u64 = protocol_features;
1907         msg->size = sizeof(msg->payload.u64);
1908         msg->fd_num = 0;
1909
1910         return RTE_VHOST_MSG_RESULT_REPLY;
1911 }
1912
1913 static int
1914 vhost_user_set_protocol_features(struct virtio_net **pdev,
1915                         struct VhostUserMsg *msg,
1916                         int main_fd __rte_unused)
1917 {
1918         struct virtio_net *dev = *pdev;
1919         uint64_t protocol_features = msg->payload.u64;
1920         uint64_t slave_protocol_features = 0;
1921
1922         rte_vhost_driver_get_protocol_features(dev->ifname,
1923                         &slave_protocol_features);
1924         if (protocol_features & ~slave_protocol_features) {
1925                 RTE_LOG(ERR, VHOST_CONFIG,
1926                         "(%d) received invalid protocol features.\n",
1927                         dev->vid);
1928                 return RTE_VHOST_MSG_RESULT_ERR;
1929         }
1930
1931         dev->protocol_features = protocol_features;
1932         RTE_LOG(INFO, VHOST_CONFIG,
1933                 "negotiated Vhost-user protocol features: 0x%" PRIx64 "\n",
1934                 dev->protocol_features);
1935
1936         return RTE_VHOST_MSG_RESULT_OK;
1937 }
1938
1939 static int
1940 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1941                         int main_fd __rte_unused)
1942 {
1943         struct virtio_net *dev = *pdev;
1944         int fd = msg->fds[0];
1945         uint64_t size, off;
1946         void *addr;
1947
1948         if (fd < 0) {
1949                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1950                 return RTE_VHOST_MSG_RESULT_ERR;
1951         }
1952
1953         if (msg->size != sizeof(VhostUserLog)) {
1954                 RTE_LOG(ERR, VHOST_CONFIG,
1955                         "invalid log base msg size: %"PRId32" != %d\n",
1956                         msg->size, (int)sizeof(VhostUserLog));
1957                 return RTE_VHOST_MSG_RESULT_ERR;
1958         }
1959
1960         size = msg->payload.log.mmap_size;
1961         off  = msg->payload.log.mmap_offset;
1962
1963         /* Don't allow mmap_offset to point outside the mmap region */
1964         if (off > size) {
1965                 RTE_LOG(ERR, VHOST_CONFIG,
1966                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1967                         off, size);
1968                 return RTE_VHOST_MSG_RESULT_ERR;
1969         }
1970
1971         RTE_LOG(INFO, VHOST_CONFIG,
1972                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1973                 size, off);
1974
1975         /*
1976          * mmap from 0 to workaround a hugepage mmap bug: mmap will
1977          * fail when offset is not page size aligned.
1978          */
1979         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1980         close(fd);
1981         if (addr == MAP_FAILED) {
1982                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1983                 return RTE_VHOST_MSG_RESULT_ERR;
1984         }
1985
1986         /*
1987          * Free previously mapped log memory on occasionally
1988          * multiple VHOST_USER_SET_LOG_BASE.
1989          */
1990         if (dev->log_addr) {
1991                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1992         }
1993         dev->log_addr = (uint64_t)(uintptr_t)addr;
1994         dev->log_base = dev->log_addr + off;
1995         dev->log_size = size;
1996
1997         /*
1998          * The spec is not clear about it (yet), but QEMU doesn't expect
1999          * any payload in the reply.
2000          */
2001         msg->size = 0;
2002         msg->fd_num = 0;
2003
2004         return RTE_VHOST_MSG_RESULT_REPLY;
2005 }
2006
2007 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
2008                         struct VhostUserMsg *msg,
2009                         int main_fd __rte_unused)
2010 {
2011         close(msg->fds[0]);
2012         RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
2013
2014         return RTE_VHOST_MSG_RESULT_OK;
2015 }
2016
2017 /*
2018  * An rarp packet is constructed and broadcasted to notify switches about
2019  * the new location of the migrated VM, so that packets from outside will
2020  * not be lost after migration.
2021  *
2022  * However, we don't actually "send" a rarp packet here, instead, we set
2023  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
2024  */
2025 static int
2026 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
2027                         int main_fd __rte_unused)
2028 {
2029         struct virtio_net *dev = *pdev;
2030         uint8_t *mac = (uint8_t *)&msg->payload.u64;
2031         struct rte_vdpa_device *vdpa_dev;
2032         int did = -1;
2033
2034         RTE_LOG(DEBUG, VHOST_CONFIG,
2035                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
2036                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
2037         memcpy(dev->mac.addr_bytes, mac, 6);
2038
2039         /*
2040          * Set the flag to inject a RARP broadcast packet at
2041          * rte_vhost_dequeue_burst().
2042          *
2043          * rte_smp_wmb() is for making sure the mac is copied
2044          * before the flag is set.
2045          */
2046         rte_smp_wmb();
2047         rte_atomic16_set(&dev->broadcast_rarp, 1);
2048         did = dev->vdpa_dev_id;
2049         vdpa_dev = rte_vdpa_get_device(did);
2050         if (vdpa_dev && vdpa_dev->ops->migration_done)
2051                 vdpa_dev->ops->migration_done(dev->vid);
2052
2053         return RTE_VHOST_MSG_RESULT_OK;
2054 }
2055
2056 static int
2057 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
2058                         int main_fd __rte_unused)
2059 {
2060         struct virtio_net *dev = *pdev;
2061         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
2062                         msg->payload.u64 > VIRTIO_MAX_MTU) {
2063                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
2064                                 msg->payload.u64);
2065
2066                 return RTE_VHOST_MSG_RESULT_ERR;
2067         }
2068
2069         dev->mtu = msg->payload.u64;
2070
2071         return RTE_VHOST_MSG_RESULT_OK;
2072 }
2073
2074 static int
2075 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
2076                         int main_fd __rte_unused)
2077 {
2078         struct virtio_net *dev = *pdev;
2079         int fd = msg->fds[0];
2080
2081         if (fd < 0) {
2082                 RTE_LOG(ERR, VHOST_CONFIG,
2083                                 "Invalid file descriptor for slave channel (%d)\n",
2084                                 fd);
2085                 return RTE_VHOST_MSG_RESULT_ERR;
2086         }
2087
2088         if (dev->slave_req_fd >= 0)
2089                 close(dev->slave_req_fd);
2090
2091         dev->slave_req_fd = fd;
2092
2093         return RTE_VHOST_MSG_RESULT_OK;
2094 }
2095
2096 static int
2097 is_vring_iotlb_split(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2098 {
2099         struct vhost_vring_addr *ra;
2100         uint64_t start, end, len;
2101
2102         start = imsg->iova;
2103         end = start + imsg->size;
2104
2105         ra = &vq->ring_addrs;
2106         len = sizeof(struct vring_desc) * vq->size;
2107         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2108                 return 1;
2109
2110         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
2111         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2112                 return 1;
2113
2114         len = sizeof(struct vring_used) +
2115                sizeof(struct vring_used_elem) * vq->size;
2116         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2117                 return 1;
2118
2119         return 0;
2120 }
2121
2122 static int
2123 is_vring_iotlb_packed(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2124 {
2125         struct vhost_vring_addr *ra;
2126         uint64_t start, end, len;
2127
2128         start = imsg->iova;
2129         end = start + imsg->size;
2130
2131         ra = &vq->ring_addrs;
2132         len = sizeof(struct vring_packed_desc) * vq->size;
2133         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2134                 return 1;
2135
2136         len = sizeof(struct vring_packed_desc_event);
2137         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2138                 return 1;
2139
2140         len = sizeof(struct vring_packed_desc_event);
2141         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2142                 return 1;
2143
2144         return 0;
2145 }
2146
2147 static int is_vring_iotlb(struct virtio_net *dev,
2148                           struct vhost_virtqueue *vq,
2149                           struct vhost_iotlb_msg *imsg)
2150 {
2151         if (vq_is_packed(dev))
2152                 return is_vring_iotlb_packed(vq, imsg);
2153         else
2154                 return is_vring_iotlb_split(vq, imsg);
2155 }
2156
2157 static int
2158 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
2159                         int main_fd __rte_unused)
2160 {
2161         struct virtio_net *dev = *pdev;
2162         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
2163         uint16_t i;
2164         uint64_t vva, len;
2165
2166         switch (imsg->type) {
2167         case VHOST_IOTLB_UPDATE:
2168                 len = imsg->size;
2169                 vva = qva_to_vva(dev, imsg->uaddr, &len);
2170                 if (!vva)
2171                         return RTE_VHOST_MSG_RESULT_ERR;
2172
2173                 for (i = 0; i < dev->nr_vring; i++) {
2174                         struct vhost_virtqueue *vq = dev->virtqueue[i];
2175
2176                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
2177                                         len, imsg->perm);
2178
2179                         if (is_vring_iotlb(dev, vq, imsg))
2180                                 *pdev = dev = translate_ring_addresses(dev, i);
2181                 }
2182                 break;
2183         case VHOST_IOTLB_INVALIDATE:
2184                 for (i = 0; i < dev->nr_vring; i++) {
2185                         struct vhost_virtqueue *vq = dev->virtqueue[i];
2186
2187                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
2188                                         imsg->size);
2189
2190                         if (is_vring_iotlb(dev, vq, imsg))
2191                                 vring_invalidate(dev, vq);
2192                 }
2193                 break;
2194         default:
2195                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
2196                                 imsg->type);
2197                 return RTE_VHOST_MSG_RESULT_ERR;
2198         }
2199
2200         return RTE_VHOST_MSG_RESULT_OK;
2201 }
2202
2203 static int
2204 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
2205                         struct VhostUserMsg *msg,
2206                         int main_fd __rte_unused)
2207 {
2208         struct virtio_net *dev = *pdev;
2209 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
2210         struct uffdio_api api_struct;
2211
2212         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
2213
2214         if (dev->postcopy_ufd == -1) {
2215                 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
2216                         strerror(errno));
2217                 return RTE_VHOST_MSG_RESULT_ERR;
2218         }
2219         api_struct.api = UFFD_API;
2220         api_struct.features = 0;
2221         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
2222                 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
2223                         strerror(errno));
2224                 close(dev->postcopy_ufd);
2225                 dev->postcopy_ufd = -1;
2226                 return RTE_VHOST_MSG_RESULT_ERR;
2227         }
2228         msg->fds[0] = dev->postcopy_ufd;
2229         msg->fd_num = 1;
2230
2231         return RTE_VHOST_MSG_RESULT_REPLY;
2232 #else
2233         dev->postcopy_ufd = -1;
2234         msg->fd_num = 0;
2235
2236         return RTE_VHOST_MSG_RESULT_ERR;
2237 #endif
2238 }
2239
2240 static int
2241 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
2242                         struct VhostUserMsg *msg __rte_unused,
2243                         int main_fd __rte_unused)
2244 {
2245         struct virtio_net *dev = *pdev;
2246
2247         if (dev->mem && dev->mem->nregions) {
2248                 RTE_LOG(ERR, VHOST_CONFIG,
2249                         "Regions already registered at postcopy-listen\n");
2250                 return RTE_VHOST_MSG_RESULT_ERR;
2251         }
2252         dev->postcopy_listening = 1;
2253
2254         return RTE_VHOST_MSG_RESULT_OK;
2255 }
2256
2257 static int
2258 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
2259                         int main_fd __rte_unused)
2260 {
2261         struct virtio_net *dev = *pdev;
2262
2263         dev->postcopy_listening = 0;
2264         if (dev->postcopy_ufd >= 0) {
2265                 close(dev->postcopy_ufd);
2266                 dev->postcopy_ufd = -1;
2267         }
2268
2269         msg->payload.u64 = 0;
2270         msg->size = sizeof(msg->payload.u64);
2271         msg->fd_num = 0;
2272
2273         return RTE_VHOST_MSG_RESULT_REPLY;
2274 }
2275
2276 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
2277                                         struct VhostUserMsg *msg,
2278                                         int main_fd);
2279 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
2280         [VHOST_USER_NONE] = NULL,
2281         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
2282         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
2283         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
2284         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
2285         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
2286         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
2287         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
2288         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
2289         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
2290         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
2291         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
2292         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
2293         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
2294         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
2295         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
2296         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
2297         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
2298         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
2299         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
2300         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
2301         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
2302         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
2303         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
2304         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
2305         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
2306         [VHOST_USER_GET_INFLIGHT_FD] = vhost_user_get_inflight_fd,
2307         [VHOST_USER_SET_INFLIGHT_FD] = vhost_user_set_inflight_fd,
2308 };
2309
2310 /* return bytes# of read on success or negative val on failure. */
2311 static int
2312 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
2313 {
2314         int ret;
2315
2316         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
2317                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
2318         if (ret <= 0)
2319                 return ret;
2320
2321         if (msg->size) {
2322                 if (msg->size > sizeof(msg->payload)) {
2323                         RTE_LOG(ERR, VHOST_CONFIG,
2324                                 "invalid msg size: %d\n", msg->size);
2325                         return -1;
2326                 }
2327                 ret = read(sockfd, &msg->payload, msg->size);
2328                 if (ret <= 0)
2329                         return ret;
2330                 if (ret != (int)msg->size) {
2331                         RTE_LOG(ERR, VHOST_CONFIG,
2332                                 "read control message failed\n");
2333                         return -1;
2334                 }
2335         }
2336
2337         return ret;
2338 }
2339
2340 static int
2341 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
2342 {
2343         if (!msg)
2344                 return 0;
2345
2346         return send_fd_message(sockfd, (char *)msg,
2347                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
2348 }
2349
2350 static int
2351 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
2352 {
2353         if (!msg)
2354                 return 0;
2355
2356         msg->flags &= ~VHOST_USER_VERSION_MASK;
2357         msg->flags &= ~VHOST_USER_NEED_REPLY;
2358         msg->flags |= VHOST_USER_VERSION;
2359         msg->flags |= VHOST_USER_REPLY_MASK;
2360
2361         return send_vhost_message(sockfd, msg);
2362 }
2363
2364 static int
2365 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
2366 {
2367         int ret;
2368
2369         if (msg->flags & VHOST_USER_NEED_REPLY)
2370                 rte_spinlock_lock(&dev->slave_req_lock);
2371
2372         ret = send_vhost_message(dev->slave_req_fd, msg);
2373         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
2374                 rte_spinlock_unlock(&dev->slave_req_lock);
2375
2376         return ret;
2377 }
2378
2379 /*
2380  * Allocate a queue pair if it hasn't been allocated yet
2381  */
2382 static int
2383 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
2384                         struct VhostUserMsg *msg)
2385 {
2386         uint16_t vring_idx;
2387
2388         switch (msg->request.master) {
2389         case VHOST_USER_SET_VRING_KICK:
2390         case VHOST_USER_SET_VRING_CALL:
2391         case VHOST_USER_SET_VRING_ERR:
2392                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
2393                 break;
2394         case VHOST_USER_SET_VRING_NUM:
2395         case VHOST_USER_SET_VRING_BASE:
2396         case VHOST_USER_SET_VRING_ENABLE:
2397                 vring_idx = msg->payload.state.index;
2398                 break;
2399         case VHOST_USER_SET_VRING_ADDR:
2400                 vring_idx = msg->payload.addr.index;
2401                 break;
2402         default:
2403                 return 0;
2404         }
2405
2406         if (vring_idx >= VHOST_MAX_VRING) {
2407                 RTE_LOG(ERR, VHOST_CONFIG,
2408                         "invalid vring index: %u\n", vring_idx);
2409                 return -1;
2410         }
2411
2412         if (dev->virtqueue[vring_idx])
2413                 return 0;
2414
2415         return alloc_vring_queue(dev, vring_idx);
2416 }
2417
2418 static void
2419 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
2420 {
2421         unsigned int i = 0;
2422         unsigned int vq_num = 0;
2423
2424         while (vq_num < dev->nr_vring) {
2425                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2426
2427                 if (vq) {
2428                         rte_spinlock_lock(&vq->access_lock);
2429                         vq_num++;
2430                 }
2431                 i++;
2432         }
2433 }
2434
2435 static void
2436 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
2437 {
2438         unsigned int i = 0;
2439         unsigned int vq_num = 0;
2440
2441         while (vq_num < dev->nr_vring) {
2442                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2443
2444                 if (vq) {
2445                         rte_spinlock_unlock(&vq->access_lock);
2446                         vq_num++;
2447                 }
2448                 i++;
2449         }
2450 }
2451
2452 int
2453 vhost_user_msg_handler(int vid, int fd)
2454 {
2455         struct virtio_net *dev;
2456         struct VhostUserMsg msg;
2457         struct rte_vdpa_device *vdpa_dev;
2458         int did = -1;
2459         int ret;
2460         int unlock_required = 0;
2461         bool handled;
2462         int request;
2463
2464         dev = get_device(vid);
2465         if (dev == NULL)
2466                 return -1;
2467
2468         if (!dev->notify_ops) {
2469                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
2470                 if (!dev->notify_ops) {
2471                         RTE_LOG(ERR, VHOST_CONFIG,
2472                                 "failed to get callback ops for driver %s\n",
2473                                 dev->ifname);
2474                         return -1;
2475                 }
2476         }
2477
2478         ret = read_vhost_message(fd, &msg);
2479         if (ret <= 0) {
2480                 if (ret < 0)
2481                         RTE_LOG(ERR, VHOST_CONFIG,
2482                                 "vhost read message failed\n");
2483                 else
2484                         RTE_LOG(INFO, VHOST_CONFIG,
2485                                 "vhost peer closed\n");
2486
2487                 return -1;
2488         }
2489
2490         ret = 0;
2491         request = msg.request.master;
2492         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
2493                         vhost_message_str[request]) {
2494                 if (request != VHOST_USER_IOTLB_MSG)
2495                         RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
2496                                 vhost_message_str[request]);
2497                 else
2498                         RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
2499                                 vhost_message_str[request]);
2500         } else {
2501                 RTE_LOG(DEBUG, VHOST_CONFIG, "External request %d\n", request);
2502         }
2503
2504         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
2505         if (ret < 0) {
2506                 RTE_LOG(ERR, VHOST_CONFIG,
2507                         "failed to alloc queue\n");
2508                 return -1;
2509         }
2510
2511         /*
2512          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
2513          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
2514          * and device is destroyed. destroy_device waits for queues to be
2515          * inactive, so it is safe. Otherwise taking the access_lock
2516          * would cause a dead lock.
2517          */
2518         switch (request) {
2519         case VHOST_USER_SET_FEATURES:
2520         case VHOST_USER_SET_PROTOCOL_FEATURES:
2521         case VHOST_USER_SET_OWNER:
2522         case VHOST_USER_SET_MEM_TABLE:
2523         case VHOST_USER_SET_LOG_BASE:
2524         case VHOST_USER_SET_LOG_FD:
2525         case VHOST_USER_SET_VRING_NUM:
2526         case VHOST_USER_SET_VRING_ADDR:
2527         case VHOST_USER_SET_VRING_BASE:
2528         case VHOST_USER_SET_VRING_KICK:
2529         case VHOST_USER_SET_VRING_CALL:
2530         case VHOST_USER_SET_VRING_ERR:
2531         case VHOST_USER_SET_VRING_ENABLE:
2532         case VHOST_USER_SEND_RARP:
2533         case VHOST_USER_NET_SET_MTU:
2534         case VHOST_USER_SET_SLAVE_REQ_FD:
2535                 vhost_user_lock_all_queue_pairs(dev);
2536                 unlock_required = 1;
2537                 break;
2538         default:
2539                 break;
2540
2541         }
2542
2543         handled = false;
2544         if (dev->extern_ops.pre_msg_handle) {
2545                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
2546                                 (void *)&msg);
2547                 switch (ret) {
2548                 case RTE_VHOST_MSG_RESULT_REPLY:
2549                         send_vhost_reply(fd, &msg);
2550                         /* Fall-through */
2551                 case RTE_VHOST_MSG_RESULT_ERR:
2552                 case RTE_VHOST_MSG_RESULT_OK:
2553                         handled = true;
2554                         goto skip_to_post_handle;
2555                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2556                 default:
2557                         break;
2558                 }
2559         }
2560
2561         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
2562                 if (!vhost_message_handlers[request])
2563                         goto skip_to_post_handle;
2564                 ret = vhost_message_handlers[request](&dev, &msg, fd);
2565
2566                 switch (ret) {
2567                 case RTE_VHOST_MSG_RESULT_ERR:
2568                         RTE_LOG(ERR, VHOST_CONFIG,
2569                                 "Processing %s failed.\n",
2570                                 vhost_message_str[request]);
2571                         handled = true;
2572                         break;
2573                 case RTE_VHOST_MSG_RESULT_OK:
2574                         RTE_LOG(DEBUG, VHOST_CONFIG,
2575                                 "Processing %s succeeded.\n",
2576                                 vhost_message_str[request]);
2577                         handled = true;
2578                         break;
2579                 case RTE_VHOST_MSG_RESULT_REPLY:
2580                         RTE_LOG(DEBUG, VHOST_CONFIG,
2581                                 "Processing %s succeeded and needs reply.\n",
2582                                 vhost_message_str[request]);
2583                         send_vhost_reply(fd, &msg);
2584                         handled = true;
2585                         break;
2586                 default:
2587                         break;
2588                 }
2589         }
2590
2591 skip_to_post_handle:
2592         if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2593                         dev->extern_ops.post_msg_handle) {
2594                 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
2595                                 (void *)&msg);
2596                 switch (ret) {
2597                 case RTE_VHOST_MSG_RESULT_REPLY:
2598                         send_vhost_reply(fd, &msg);
2599                         /* Fall-through */
2600                 case RTE_VHOST_MSG_RESULT_ERR:
2601                 case RTE_VHOST_MSG_RESULT_OK:
2602                         handled = true;
2603                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2604                 default:
2605                         break;
2606                 }
2607         }
2608
2609         if (unlock_required)
2610                 vhost_user_unlock_all_queue_pairs(dev);
2611
2612         /* If message was not handled at this stage, treat it as an error */
2613         if (!handled) {
2614                 RTE_LOG(ERR, VHOST_CONFIG,
2615                         "vhost message (req: %d) was not handled.\n", request);
2616                 ret = RTE_VHOST_MSG_RESULT_ERR;
2617         }
2618
2619         /*
2620          * If the request required a reply that was already sent,
2621          * this optional reply-ack won't be sent as the
2622          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2623          */
2624         if (msg.flags & VHOST_USER_NEED_REPLY) {
2625                 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2626                 msg.size = sizeof(msg.payload.u64);
2627                 msg.fd_num = 0;
2628                 send_vhost_reply(fd, &msg);
2629         } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2630                 RTE_LOG(ERR, VHOST_CONFIG,
2631                         "vhost message handling failed.\n");
2632                 return -1;
2633         }
2634
2635         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2636                 dev->flags |= VIRTIO_DEV_READY;
2637
2638                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2639                         if (dev->dequeue_zero_copy) {
2640                                 RTE_LOG(INFO, VHOST_CONFIG,
2641                                                 "dequeue zero copy is enabled\n");
2642                         }
2643
2644                         if (dev->notify_ops->new_device(dev->vid) == 0)
2645                                 dev->flags |= VIRTIO_DEV_RUNNING;
2646                 }
2647         }
2648
2649         did = dev->vdpa_dev_id;
2650         vdpa_dev = rte_vdpa_get_device(did);
2651         if (vdpa_dev && virtio_is_ready(dev) &&
2652                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2653                         msg.request.master == VHOST_USER_SET_VRING_CALL) {
2654                 if (vdpa_dev->ops->dev_conf)
2655                         vdpa_dev->ops->dev_conf(dev->vid);
2656                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2657         }
2658
2659         return 0;
2660 }
2661
2662 static int process_slave_message_reply(struct virtio_net *dev,
2663                                        const struct VhostUserMsg *msg)
2664 {
2665         struct VhostUserMsg msg_reply;
2666         int ret;
2667
2668         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2669                 return 0;
2670
2671         if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
2672                 ret = -1;
2673                 goto out;
2674         }
2675
2676         if (msg_reply.request.slave != msg->request.slave) {
2677                 RTE_LOG(ERR, VHOST_CONFIG,
2678                         "Received unexpected msg type (%u), expected %u\n",
2679                         msg_reply.request.slave, msg->request.slave);
2680                 ret = -1;
2681                 goto out;
2682         }
2683
2684         ret = msg_reply.payload.u64 ? -1 : 0;
2685
2686 out:
2687         rte_spinlock_unlock(&dev->slave_req_lock);
2688         return ret;
2689 }
2690
2691 int
2692 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2693 {
2694         int ret;
2695         struct VhostUserMsg msg = {
2696                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2697                 .flags = VHOST_USER_VERSION,
2698                 .size = sizeof(msg.payload.iotlb),
2699                 .payload.iotlb = {
2700                         .iova = iova,
2701                         .perm = perm,
2702                         .type = VHOST_IOTLB_MISS,
2703                 },
2704         };
2705
2706         ret = send_vhost_message(dev->slave_req_fd, &msg);
2707         if (ret < 0) {
2708                 RTE_LOG(ERR, VHOST_CONFIG,
2709                                 "Failed to send IOTLB miss message (%d)\n",
2710                                 ret);
2711                 return ret;
2712         }
2713
2714         return 0;
2715 }
2716
2717 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2718                                                     int index, int fd,
2719                                                     uint64_t offset,
2720                                                     uint64_t size)
2721 {
2722         int ret;
2723         struct VhostUserMsg msg = {
2724                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2725                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2726                 .size = sizeof(msg.payload.area),
2727                 .payload.area = {
2728                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
2729                         .size = size,
2730                         .offset = offset,
2731                 },
2732         };
2733
2734         if (fd < 0)
2735                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2736         else {
2737                 msg.fds[0] = fd;
2738                 msg.fd_num = 1;
2739         }
2740
2741         ret = send_vhost_slave_message(dev, &msg);
2742         if (ret < 0) {
2743                 RTE_LOG(ERR, VHOST_CONFIG,
2744                         "Failed to set host notifier (%d)\n", ret);
2745                 return ret;
2746         }
2747
2748         return process_slave_message_reply(dev, &msg);
2749 }
2750
2751 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2752 {
2753         struct virtio_net *dev;
2754         struct rte_vdpa_device *vdpa_dev;
2755         int vfio_device_fd, did, ret = 0;
2756         uint64_t offset, size;
2757         unsigned int i;
2758
2759         dev = get_device(vid);
2760         if (!dev)
2761                 return -ENODEV;
2762
2763         did = dev->vdpa_dev_id;
2764         if (did < 0)
2765                 return -EINVAL;
2766
2767         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2768             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2769             !(dev->protocol_features &
2770                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2771             !(dev->protocol_features &
2772                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2773             !(dev->protocol_features &
2774                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2775                 return -ENOTSUP;
2776
2777         vdpa_dev = rte_vdpa_get_device(did);
2778         if (!vdpa_dev)
2779                 return -ENODEV;
2780
2781         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2782         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2783
2784         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2785         if (vfio_device_fd < 0)
2786                 return -ENOTSUP;
2787
2788         if (enable) {
2789                 for (i = 0; i < dev->nr_vring; i++) {
2790                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2791                                         &size) < 0) {
2792                                 ret = -ENOTSUP;
2793                                 goto disable;
2794                         }
2795
2796                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2797                                         vfio_device_fd, offset, size) < 0) {
2798                                 ret = -EFAULT;
2799                                 goto disable;
2800                         }
2801                 }
2802         } else {
2803 disable:
2804                 for (i = 0; i < dev->nr_vring; i++) {
2805                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
2806                                         0, 0);
2807                 }
2808         }
2809
2810         return ret;
2811 }