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                 if (vq->zmbufs)
 390                         rte_free(vq->zmbufs);
 391                 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
 392                                          sizeof(struct zcopy_mbuf), 0);
 393                 if (vq->zmbufs == NULL) {
 394                         RTE_LOG(WARNING, VHOST_CONFIG,
 395                                 "failed to allocate mem for zero copy; "
 396                                 "zero copy is force disabled\n");
 397                         dev->dequeue_zero_copy = 0;
 398                 }
 399                 TAILQ_INIT(&vq->zmbuf_list);
 400         }
 401
 402         if (vq_is_packed(dev)) {
 403                 if (vq->shadow_used_packed)
 404                         rte_free(vq->shadow_used_packed);
 405                 vq->shadow_used_packed = rte_malloc(NULL,
 406                                 vq->size *
 407                                 sizeof(struct vring_used_elem_packed),
 408                                 RTE_CACHE_LINE_SIZE);
 409                 if (!vq->shadow_used_packed) {
 410                         RTE_LOG(ERR, VHOST_CONFIG,
 411                                         "failed to allocate memory for shadow used ring.\n");
 412                         return RTE_VHOST_MSG_RESULT_ERR;
 413                 }
 414
 415         } else {
 416                 if (vq->shadow_used_split)
 417                         rte_free(vq->shadow_used_split);
 418                 vq->shadow_used_split = rte_malloc(NULL,
 419                                 vq->size * sizeof(struct vring_used_elem),
 420                                 RTE_CACHE_LINE_SIZE);
 421                 if (!vq->shadow_used_split) {
 422                         RTE_LOG(ERR, VHOST_CONFIG,
 423                                         "failed to allocate memory for shadow used ring.\n");
 424                         return RTE_VHOST_MSG_RESULT_ERR;
 425                 }
 426         }
 427
 428         if (vq->batch_copy_elems)
 429                 rte_free(vq->batch_copy_elems);
 430         vq->batch_copy_elems = rte_malloc(NULL,
 431                                 vq->size * sizeof(struct batch_copy_elem),
 432                                 RTE_CACHE_LINE_SIZE);
 433         if (!vq->batch_copy_elems) {
 434                 RTE_LOG(ERR, VHOST_CONFIG,
 435                         "failed to allocate memory for batching copy.\n");
 436                 return RTE_VHOST_MSG_RESULT_ERR;
 437         }
 438
 439         return RTE_VHOST_MSG_RESULT_OK;
 440 }
 441
 442 /*
 443  * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
 444  * same numa node as the memory of vring descriptor.
 445  */
 446 #ifdef RTE_LIBRTE_VHOST_NUMA
 447 static struct virtio_net*
 448 numa_realloc(struct virtio_net *dev, int index)
 449 {
 450         int oldnode, newnode;
 451         struct virtio_net *old_dev;
 452         struct vhost_virtqueue *old_vq, *vq;
 453         struct zcopy_mbuf *new_zmbuf;
 454         struct vring_used_elem *new_shadow_used_split;
 455         struct vring_used_elem_packed *new_shadow_used_packed;
 456         struct batch_copy_elem *new_batch_copy_elems;
 457         int ret;
 458
 459         if (dev->flags & VIRTIO_DEV_RUNNING)
 460                 return dev;
 461
 462         old_dev = dev;
 463         vq = old_vq = dev->virtqueue[index];
 464
 465         ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
 466                             MPOL_F_NODE | MPOL_F_ADDR);
 467
 468         /* check if we need to reallocate vq */
 469         ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
 470                              MPOL_F_NODE | MPOL_F_ADDR);
 471         if (ret) {
 472                 RTE_LOG(ERR, VHOST_CONFIG,
 473                         "Unable to get vq numa information.\n");
 474                 return dev;
 475         }
 476         if (oldnode != newnode) {
 477                 RTE_LOG(INFO, VHOST_CONFIG,
 478                         "reallocate vq from %d to %d node\n", oldnode, newnode);
 479                 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
 480                 if (!vq)
 481                         return dev;
 482
 483                 memcpy(vq, old_vq, sizeof(*vq));
 484                 TAILQ_INIT(&vq->zmbuf_list);
 485
 486                 if (dev->dequeue_zero_copy) {
 487                         new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
 488                                         sizeof(struct zcopy_mbuf), 0, newnode);
 489                         if (new_zmbuf) {
 490                                 rte_free(vq->zmbufs);
 491                                 vq->zmbufs = new_zmbuf;
 492                         }
 493                 }
 494
 495                 if (vq_is_packed(dev)) {
 496                         new_shadow_used_packed = rte_malloc_socket(NULL,
 497                                         vq->size *
 498                                         sizeof(struct vring_used_elem_packed),
 499                                         RTE_CACHE_LINE_SIZE,
 500                                         newnode);
 501                         if (new_shadow_used_packed) {
 502                                 rte_free(vq->shadow_used_packed);
 503                                 vq->shadow_used_packed = new_shadow_used_packed;
 504                         }
 505                 } else {
 506                         new_shadow_used_split = rte_malloc_socket(NULL,
 507                                         vq->size *
 508                                         sizeof(struct vring_used_elem),
 509                                         RTE_CACHE_LINE_SIZE,
 510                                         newnode);
 511                         if (new_shadow_used_split) {
 512                                 rte_free(vq->shadow_used_split);
 513                                 vq->shadow_used_split = new_shadow_used_split;
 514                         }
 515                 }
 516
 517                 new_batch_copy_elems = rte_malloc_socket(NULL,
 518                         vq->size * sizeof(struct batch_copy_elem),
 519                         RTE_CACHE_LINE_SIZE,
 520                         newnode);
 521                 if (new_batch_copy_elems) {
 522                         rte_free(vq->batch_copy_elems);
 523                         vq->batch_copy_elems = new_batch_copy_elems;
 524                 }
 525
 526                 rte_free(old_vq);
 527         }
 528
 529         /* check if we need to reallocate dev */
 530         ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
 531                             MPOL_F_NODE | MPOL_F_ADDR);
 532         if (ret) {
 533                 RTE_LOG(ERR, VHOST_CONFIG,
 534                         "Unable to get dev numa information.\n");
 535                 goto out;
 536         }
 537         if (oldnode != newnode) {
 538                 RTE_LOG(INFO, VHOST_CONFIG,
 539                         "reallocate dev from %d to %d node\n",
 540                         oldnode, newnode);
 541                 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
 542                 if (!dev) {
 543                         dev = old_dev;
 544                         goto out;
 545                 }
 546
 547                 memcpy(dev, old_dev, sizeof(*dev));
 548                 rte_free(old_dev);
 549         }
 550
 551 out:
 552         dev->virtqueue[index] = vq;
 553         vhost_devices[dev->vid] = dev;
 554
 555         if (old_vq != vq)
 556                 vhost_user_iotlb_init(dev, index);
 557
 558         return dev;
 559 }
 560 #else
 561 static struct virtio_net*
 562 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 563 {
 564         return dev;
 565 }
 566 #endif
 567
 568 /* Converts QEMU virtual address to Vhost virtual address. */
 569 static uint64_t
 570 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
 571 {
 572         struct rte_vhost_mem_region *r;
 573         uint32_t i;
 574
 575         if (unlikely(!dev || !dev->mem))
 576                 goto out_error;
 577
 578         /* Find the region where the address lives. */
 579         for (i = 0; i < dev->mem->nregions; i++) {
 580                 r = &dev->mem->regions[i];
 581
 582                 if (qva >= r->guest_user_addr &&
 583                     qva <  r->guest_user_addr + r->size) {
 584
 585                         if (unlikely(*len > r->guest_user_addr + r->size - qva))
 586                                 *len = r->guest_user_addr + r->size - qva;
 587
 588                         return qva - r->guest_user_addr +
 589                                r->host_user_addr;
 590                 }
 591         }
 592 out_error:
 593         *len = 0;
 594
 595         return 0;
 596 }
 597
 598
 599 /*
 600  * Converts ring address to Vhost virtual address.
 601  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 602  * else it is a QEMU virtual address.
 603  */
 604 static uint64_t
 605 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 606                 uint64_t ra, uint64_t *size)
 607 {
 608         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 609                 uint64_t vva;
 610                 uint64_t req_size = *size;
 611
 612                 vva = vhost_user_iotlb_cache_find(vq, ra,
 613                                         size, VHOST_ACCESS_RW);
 614                 if (req_size != *size)
 615                         vhost_user_iotlb_miss(dev, (ra + *size),
 616                                               VHOST_ACCESS_RW);
 617
 618                 return vva;
 619         }
 620
 621         return qva_to_vva(dev, ra, size);
 622 }
 623
 624 /*
 625  * Converts vring log address to GPA
 626  * If IOMMU is enabled, the log address is IOVA
 627  * If IOMMU not enabled, the log address is already GPA
 628  */
 629 static uint64_t
 630 translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
 631                 uint64_t log_addr)
 632 {
 633         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 634                 const uint64_t exp_size = sizeof(struct vring_used) +
 635                         sizeof(struct vring_used_elem) * vq->size;
 636                 uint64_t hva, gpa;
 637                 uint64_t size = exp_size;
 638
 639                 hva = vhost_iova_to_vva(dev, vq, log_addr,
 640                                         &size, VHOST_ACCESS_RW);
 641                 if (size != exp_size)
 642                         return 0;
 643
 644                 gpa = hva_to_gpa(dev, hva, exp_size);
 645                 if (!gpa) {
 646                         RTE_LOG(ERR, VHOST_CONFIG,
 647                                 "VQ: Failed to find GPA for log_addr: 0x%" PRIx64 " hva: 0x%" PRIx64 "\n",
 648                                 log_addr, hva);
 649                         return 0;
 650                 }
 651                 return gpa;
 652
 653         } else
 654                 return log_addr;
 655 }
 656
 657 static struct virtio_net *
 658 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 659 {
 660         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 661         struct vhost_vring_addr *addr = &vq->ring_addrs;
 662         uint64_t len, expected_len;
 663
 664         if (addr->flags & (1 << VHOST_VRING_F_LOG)) {
 665                 vq->log_guest_addr =
 666                         translate_log_addr(dev, vq, addr->log_guest_addr);
 667                 if (vq->log_guest_addr == 0) {
 668                         RTE_LOG(DEBUG, VHOST_CONFIG,
 669                                 "(%d) failed to map log_guest_addr.\n",
 670                                 dev->vid);
 671                         return dev;
 672                 }
 673         }
 674
 675         if (vq_is_packed(dev)) {
 676                 len = sizeof(struct vring_packed_desc) * vq->size;
 677                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 678                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 679                 if (vq->desc_packed == NULL ||
 680                                 len != sizeof(struct vring_packed_desc) *
 681                                 vq->size) {
 682                         RTE_LOG(DEBUG, VHOST_CONFIG,
 683                                 "(%d) failed to map desc_packed ring.\n",
 684                                 dev->vid);
 685                         return dev;
 686                 }
 687
 688                 dev = numa_realloc(dev, vq_index);
 689                 vq = dev->virtqueue[vq_index];
 690                 addr = &vq->ring_addrs;
 691
 692                 len = sizeof(struct vring_packed_desc_event);
 693                 vq->driver_event = (struct vring_packed_desc_event *)
 694                                         (uintptr_t)ring_addr_to_vva(dev,
 695                                         vq, addr->avail_user_addr, &len);
 696                 if (vq->driver_event == NULL ||
 697                                 len != sizeof(struct vring_packed_desc_event)) {
 698                         RTE_LOG(DEBUG, VHOST_CONFIG,
 699                                 "(%d) failed to find driver area address.\n",
 700                                 dev->vid);
 701                         return dev;
 702                 }
 703
 704                 len = sizeof(struct vring_packed_desc_event);
 705                 vq->device_event = (struct vring_packed_desc_event *)
 706                                         (uintptr_t)ring_addr_to_vva(dev,
 707                                         vq, addr->used_user_addr, &len);
 708                 if (vq->device_event == NULL ||
 709                                 len != sizeof(struct vring_packed_desc_event)) {
 710                         RTE_LOG(DEBUG, VHOST_CONFIG,
 711                                 "(%d) failed to find device area address.\n",
 712                                 dev->vid);
 713                         return dev;
 714                 }
 715
 716                 vq->access_ok = 1;
 717                 return dev;
 718         }
 719
 720         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 721         if (vq->desc && vq->avail && vq->used)
 722                 return dev;
 723
 724         len = sizeof(struct vring_desc) * vq->size;
 725         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 726                         vq, addr->desc_user_addr, &len);
 727         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 728                 RTE_LOG(DEBUG, VHOST_CONFIG,
 729                         "(%d) failed to map desc ring.\n",
 730                         dev->vid);
 731                 return dev;
 732         }
 733
 734         dev = numa_realloc(dev, vq_index);
 735         vq = dev->virtqueue[vq_index];
 736         addr = &vq->ring_addrs;
 737
 738         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 739         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 740                 len += sizeof(uint16_t);
 741         expected_len = len;
 742         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 743                         vq, addr->avail_user_addr, &len);
 744         if (vq->avail == 0 || len != expected_len) {
 745                 RTE_LOG(DEBUG, VHOST_CONFIG,
 746                         "(%d) failed to map avail ring.\n",
 747                         dev->vid);
 748                 return dev;
 749         }
 750
 751         len = sizeof(struct vring_used) +
 752                 sizeof(struct vring_used_elem) * vq->size;
 753         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 754                 len += sizeof(uint16_t);
 755         expected_len = len;
 756         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 757                         vq, addr->used_user_addr, &len);
 758         if (vq->used == 0 || len != expected_len) {
 759                 RTE_LOG(DEBUG, VHOST_CONFIG,
 760                         "(%d) failed to map used ring.\n",
 761                         dev->vid);
 762                 return dev;
 763         }
 764
 765         if (vq->last_used_idx != vq->used->idx) {
 766                 RTE_LOG(WARNING, VHOST_CONFIG,
 767                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 768                         "some packets maybe resent for Tx and dropped for Rx\n",
 769                         vq->last_used_idx, vq->used->idx);
 770                 vq->last_used_idx  = vq->used->idx;
 771                 vq->last_avail_idx = vq->used->idx;
 772         }
 773
 774         vq->access_ok = 1;
 775
 776         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
 777                         dev->vid, vq->desc);
 778         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
 779                         dev->vid, vq->avail);
 780         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
 781                         dev->vid, vq->used);
 782         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
 783                         dev->vid, vq->log_guest_addr);
 784
 785         return dev;
 786 }
 787
 788 /*
 789  * The virtio device sends us the desc, used and avail ring addresses.
 790  * This function then converts these to our address space.
 791  */
 792 static int
 793 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 794                         int main_fd __rte_unused)
 795 {
 796         struct virtio_net *dev = *pdev;
 797         struct vhost_virtqueue *vq;
 798         struct vhost_vring_addr *addr = &msg->payload.addr;
 799         bool access_ok;
 800
 801         if (dev->mem == NULL)
 802                 return RTE_VHOST_MSG_RESULT_ERR;
 803
 804         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 805         vq = dev->virtqueue[msg->payload.addr.index];
 806
 807         access_ok = vq->access_ok;
 808
 809         /*
 810          * Rings addresses should not be interpreted as long as the ring is not
 811          * started and enabled
 812          */
 813         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 814
 815         vring_invalidate(dev, vq);
 816
 817         if ((vq->enabled && (dev->features &
 818                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) ||
 819                         access_ok) {
 820                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 821                 if (!dev)
 822                         return RTE_VHOST_MSG_RESULT_ERR;
 823
 824                 *pdev = dev;
 825         }
 826
 827         return RTE_VHOST_MSG_RESULT_OK;
 828 }
 829
 830 /*
 831  * The virtio device sends us the available ring last used index.
 832  */
 833 static int
 834 vhost_user_set_vring_base(struct virtio_net **pdev,
 835                         struct VhostUserMsg *msg,
 836                         int main_fd __rte_unused)
 837 {
 838         struct virtio_net *dev = *pdev;
 839         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 840         uint64_t val = msg->payload.state.num;
 841
 842         if (vq_is_packed(dev)) {
 843                 /*
 844                  * Bit[0:14]: avail index
 845                  * Bit[15]: avail wrap counter
 846                  */
 847                 vq->last_avail_idx = val & 0x7fff;
 848                 vq->avail_wrap_counter = !!(val & (0x1 << 15));
 849                 /*
 850                  * Set used index to same value as available one, as
 851                  * their values should be the same since ring processing
 852                  * was stopped at get time.
 853                  */
 854                 vq->last_used_idx = vq->last_avail_idx;
 855                 vq->used_wrap_counter = vq->avail_wrap_counter;
 856         } else {
 857                 vq->last_used_idx = msg->payload.state.num;
 858                 vq->last_avail_idx = msg->payload.state.num;
 859         }
 860
 861         return RTE_VHOST_MSG_RESULT_OK;
 862 }
 863
 864 static int
 865 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 866                    uint64_t host_phys_addr, uint64_t size)
 867 {
 868         struct guest_page *page, *last_page;
 869         struct guest_page *old_pages;
 870
 871         if (dev->nr_guest_pages == dev->max_guest_pages) {
 872                 dev->max_guest_pages *= 2;
 873                 old_pages = dev->guest_pages;
 874                 dev->guest_pages = realloc(dev->guest_pages,
 875                                         dev->max_guest_pages * sizeof(*page));
 876                 if (!dev->guest_pages) {
 877                         RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
 878                         free(old_pages);
 879                         return -1;
 880                 }
 881         }
 882
 883         if (dev->nr_guest_pages > 0) {
 884                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 885                 /* merge if the two pages are continuous */
 886                 if (host_phys_addr == last_page->host_phys_addr +
 887                                       last_page->size) {
 888                         last_page->size += size;
 889                         return 0;
 890                 }
 891         }
 892
 893         page = &dev->guest_pages[dev->nr_guest_pages++];
 894         page->guest_phys_addr = guest_phys_addr;
 895         page->host_phys_addr  = host_phys_addr;
 896         page->size = size;
 897
 898         return 0;
 899 }
 900
 901 static int
 902 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 903                 uint64_t page_size)
 904 {
 905         uint64_t reg_size = reg->size;
 906         uint64_t host_user_addr  = reg->host_user_addr;
 907         uint64_t guest_phys_addr = reg->guest_phys_addr;
 908         uint64_t host_phys_addr;
 909         uint64_t size;
 910
 911         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 912         size = page_size - (guest_phys_addr & (page_size - 1));
 913         size = RTE_MIN(size, reg_size);
 914
 915         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 916                 return -1;
 917
 918         host_user_addr  += size;
 919         guest_phys_addr += size;
 920         reg_size -= size;
 921
 922         while (reg_size > 0) {
 923                 size = RTE_MIN(reg_size, page_size);
 924                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 925                                                   host_user_addr);
 926                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 927                                 size) < 0)
 928                         return -1;
 929
 930                 host_user_addr  += size;
 931                 guest_phys_addr += size;
 932                 reg_size -= size;
 933         }
 934
 935         return 0;
 936 }
 937
 938 #ifdef RTE_LIBRTE_VHOST_DEBUG
 939 /* TODO: enable it only in debug mode? */
 940 static void
 941 dump_guest_pages(struct virtio_net *dev)
 942 {
 943         uint32_t i;
 944         struct guest_page *page;
 945
 946         for (i = 0; i < dev->nr_guest_pages; i++) {
 947                 page = &dev->guest_pages[i];
 948
 949                 RTE_LOG(INFO, VHOST_CONFIG,
 950                         "guest physical page region %u\n"
 951                         "\t guest_phys_addr: %" PRIx64 "\n"
 952                         "\t host_phys_addr : %" PRIx64 "\n"
 953                         "\t size           : %" PRIx64 "\n",
 954                         i,
 955                         page->guest_phys_addr,
 956                         page->host_phys_addr,
 957                         page->size);
 958         }
 959 }
 960 #else
 961 #define dump_guest_pages(dev)
 962 #endif
 963
 964 static bool
 965 vhost_memory_changed(struct VhostUserMemory *new,
 966                      struct rte_vhost_memory *old)
 967 {
 968         uint32_t i;
 969
 970         if (new->nregions != old->nregions)
 971                 return true;
 972
 973         for (i = 0; i < new->nregions; ++i) {
 974                 VhostUserMemoryRegion *new_r = &new->regions[i];
 975                 struct rte_vhost_mem_region *old_r = &old->regions[i];
 976
 977                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
 978                         return true;
 979                 if (new_r->memory_size != old_r->size)
 980                         return true;
 981                 if (new_r->userspace_addr != old_r->guest_user_addr)
 982                         return true;
 983         }
 984
 985         return false;
 986 }
 987
 988 static int
 989 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
 990                         int main_fd)
 991 {
 992         struct virtio_net *dev = *pdev;
 993         struct VhostUserMemory *memory = &msg->payload.memory;
 994         struct rte_vhost_mem_region *reg;
 995         void *mmap_addr;
 996         uint64_t mmap_size;
 997         uint64_t mmap_offset;
 998         uint64_t alignment;
 999         uint32_t i;
1000         int populate;
1001         int fd;
1002
1003         if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
1004                 RTE_LOG(ERR, VHOST_CONFIG,
1005                         "too many memory regions (%u)\n", memory->nregions);
1006                 return RTE_VHOST_MSG_RESULT_ERR;
1007         }
1008
1009         if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
1010                 RTE_LOG(INFO, VHOST_CONFIG,
1011                         "(%d) memory regions not changed\n", dev->vid);
1012
1013                 for (i = 0; i < memory->nregions; i++)
1014                         close(msg->fds[i]);
1015
1016                 return RTE_VHOST_MSG_RESULT_OK;
1017         }
1018
1019         if (dev->mem) {
1020                 free_mem_region(dev);
1021                 rte_free(dev->mem);
1022                 dev->mem = NULL;
1023         }
1024
1025         /* Flush IOTLB cache as previous HVAs are now invalid */
1026         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1027                 for (i = 0; i < dev->nr_vring; i++)
1028                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
1029
1030         dev->nr_guest_pages = 0;
1031         if (!dev->guest_pages) {
1032                 dev->max_guest_pages = 8;
1033                 dev->guest_pages = malloc(dev->max_guest_pages *
1034                                                 sizeof(struct guest_page));
1035                 if (dev->guest_pages == NULL) {
1036                         RTE_LOG(ERR, VHOST_CONFIG,
1037                                 "(%d) failed to allocate memory "
1038                                 "for dev->guest_pages\n",
1039                                 dev->vid);
1040                         return RTE_VHOST_MSG_RESULT_ERR;
1041                 }
1042         }
1043
1044         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
1045                 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
1046         if (dev->mem == NULL) {
1047                 RTE_LOG(ERR, VHOST_CONFIG,
1048                         "(%d) failed to allocate memory for dev->mem\n",
1049                         dev->vid);
1050                 return RTE_VHOST_MSG_RESULT_ERR;
1051         }
1052         dev->mem->nregions = memory->nregions;
1053
1054         for (i = 0; i < memory->nregions; i++) {
1055                 fd  = msg->fds[i];
1056                 reg = &dev->mem->regions[i];
1057
1058                 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
1059                 reg->guest_user_addr = memory->regions[i].userspace_addr;
1060                 reg->size            = memory->regions[i].memory_size;
1061                 reg->fd              = fd;
1062
1063                 mmap_offset = memory->regions[i].mmap_offset;
1064
1065                 /* Check for memory_size + mmap_offset overflow */
1066                 if (mmap_offset >= -reg->size) {
1067                         RTE_LOG(ERR, VHOST_CONFIG,
1068                                 "mmap_offset (%#"PRIx64") and memory_size "
1069                                 "(%#"PRIx64") overflow\n",
1070                                 mmap_offset, reg->size);
1071                         goto err_mmap;
1072                 }
1073
1074                 mmap_size = reg->size + mmap_offset;
1075
1076                 /* mmap() without flag of MAP_ANONYMOUS, should be called
1077                  * with length argument aligned with hugepagesz at older
1078                  * longterm version Linux, like 2.6.32 and 3.2.72, or
1079                  * mmap() will fail with EINVAL.
1080                  *
1081                  * to avoid failure, make sure in caller to keep length
1082                  * aligned.
1083                  */
1084                 alignment = get_blk_size(fd);
1085                 if (alignment == (uint64_t)-1) {
1086                         RTE_LOG(ERR, VHOST_CONFIG,
1087                                 "couldn't get hugepage size through fstat\n");
1088                         goto err_mmap;
1089                 }
1090                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
1091
1092                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
1093                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
1094                                  MAP_SHARED | populate, fd, 0);
1095
1096                 if (mmap_addr == MAP_FAILED) {
1097                         RTE_LOG(ERR, VHOST_CONFIG,
1098                                 "mmap region %u failed.\n", i);
1099                         goto err_mmap;
1100                 }
1101
1102                 reg->mmap_addr = mmap_addr;
1103                 reg->mmap_size = mmap_size;
1104                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
1105                                       mmap_offset;
1106
1107                 if (dev->dequeue_zero_copy)
1108                         if (add_guest_pages(dev, reg, alignment) < 0) {
1109                                 RTE_LOG(ERR, VHOST_CONFIG,
1110                                         "adding guest pages to region %u failed.\n",
1111                                         i);
1112                                 goto err_mmap;
1113                         }
1114
1115                 RTE_LOG(INFO, VHOST_CONFIG,
1116                         "guest memory region %u, size: 0x%" PRIx64 "\n"
1117                         "\t guest physical addr: 0x%" PRIx64 "\n"
1118                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
1119                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
1120                         "\t mmap addr : 0x%" PRIx64 "\n"
1121                         "\t mmap size : 0x%" PRIx64 "\n"
1122                         "\t mmap align: 0x%" PRIx64 "\n"
1123                         "\t mmap off  : 0x%" PRIx64 "\n",
1124                         i, reg->size,
1125                         reg->guest_phys_addr,
1126                         reg->guest_user_addr,
1127                         reg->host_user_addr,
1128                         (uint64_t)(uintptr_t)mmap_addr,
1129                         mmap_size,
1130                         alignment,
1131                         mmap_offset);
1132
1133                 if (dev->postcopy_listening) {
1134                         /*
1135                          * We haven't a better way right now than sharing
1136                          * DPDK's virtual address with Qemu, so that Qemu can
1137                          * retrieve the region offset when handling userfaults.
1138                          */
1139                         memory->regions[i].userspace_addr =
1140                                 reg->host_user_addr;
1141                 }
1142         }
1143         if (dev->postcopy_listening) {
1144                 /* Send the addresses back to qemu */
1145                 msg->fd_num = 0;
1146                 send_vhost_reply(main_fd, msg);
1147
1148                 /* Wait for qemu to acknolwedge it's got the addresses
1149                  * we've got to wait before we're allowed to generate faults.
1150                  */
1151                 VhostUserMsg ack_msg;
1152                 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1153                         RTE_LOG(ERR, VHOST_CONFIG,
1154                                 "Failed to read qemu ack on postcopy set-mem-table\n");
1155                         goto err_mmap;
1156                 }
1157                 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1158                         RTE_LOG(ERR, VHOST_CONFIG,
1159                                 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1160                                 ack_msg.request.master);
1161                         goto err_mmap;
1162                 }
1163
1164                 /* Now userfault register and we can use the memory */
1165                 for (i = 0; i < memory->nregions; i++) {
1166 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1167                         reg = &dev->mem->regions[i];
1168                         struct uffdio_register reg_struct;
1169
1170                         /*
1171                          * Let's register all the mmap'ed area to ensure
1172                          * alignment on page boundary.
1173                          */
1174                         reg_struct.range.start =
1175                                 (uint64_t)(uintptr_t)reg->mmap_addr;
1176                         reg_struct.range.len = reg->mmap_size;
1177                         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1178
1179                         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1180                                                 &reg_struct)) {
1181                                 RTE_LOG(ERR, VHOST_CONFIG,
1182                                         "Failed to register ufd for region %d: (ufd = %d) %s\n",
1183                                         i, dev->postcopy_ufd,
1184                                         strerror(errno));
1185                                 goto err_mmap;
1186                         }
1187                         RTE_LOG(INFO, VHOST_CONFIG,
1188                                 "\t userfaultfd registered for range : "
1189                                 "%" PRIx64 " - %" PRIx64 "\n",
1190                                 (uint64_t)reg_struct.range.start,
1191                                 (uint64_t)reg_struct.range.start +
1192                                 (uint64_t)reg_struct.range.len - 1);
1193 #else
1194                         goto err_mmap;
1195 #endif
1196                 }
1197         }
1198
1199         for (i = 0; i < dev->nr_vring; i++) {
1200                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1201
1202                 if (vq->desc || vq->avail || vq->used) {
1203                         /*
1204                          * If the memory table got updated, the ring addresses
1205                          * need to be translated again as virtual addresses have
1206                          * changed.
1207                          */
1208                         vring_invalidate(dev, vq);
1209
1210                         dev = translate_ring_addresses(dev, i);
1211                         if (!dev) {
1212                                 dev = *pdev;
1213                                 goto err_mmap;
1214                         }
1215
1216                         *pdev = dev;
1217                 }
1218         }
1219
1220         dump_guest_pages(dev);
1221
1222         return RTE_VHOST_MSG_RESULT_OK;
1223
1224 err_mmap:
1225         free_mem_region(dev);
1226         rte_free(dev->mem);
1227         dev->mem = NULL;
1228         return RTE_VHOST_MSG_RESULT_ERR;
1229 }
1230
1231 static bool
1232 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1233 {
1234         bool rings_ok;
1235
1236         if (!vq)
1237                 return false;
1238
1239         if (vq_is_packed(dev))
1240                 rings_ok = !!vq->desc_packed;
1241         else
1242                 rings_ok = vq->desc && vq->avail && vq->used;
1243
1244         return rings_ok &&
1245                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1246                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1247 }
1248
1249 static int
1250 virtio_is_ready(struct virtio_net *dev)
1251 {
1252         struct vhost_virtqueue *vq;
1253         uint32_t i;
1254
1255         if (dev->nr_vring == 0)
1256                 return 0;
1257
1258         for (i = 0; i < dev->nr_vring; i++) {
1259                 vq = dev->virtqueue[i];
1260
1261                 if (!vq_is_ready(dev, vq))
1262                         return 0;
1263         }
1264
1265         RTE_LOG(INFO, VHOST_CONFIG,
1266                 "virtio is now ready for processing.\n");
1267         return 1;
1268 }
1269
1270 static void *
1271 inflight_mem_alloc(const char *name, size_t size, int *fd)
1272 {
1273         void *ptr;
1274         int mfd = -1;
1275         char fname[20] = "/tmp/memfd-XXXXXX";
1276
1277         *fd = -1;
1278 #ifdef MEMFD_SUPPORTED
1279         mfd = memfd_create(name, MFD_CLOEXEC);
1280 #else
1281         RTE_SET_USED(name);
1282 #endif
1283         if (mfd == -1) {
1284                 mfd = mkstemp(fname);
1285                 if (mfd == -1) {
1286                         RTE_LOG(ERR, VHOST_CONFIG,
1287                                 "failed to get inflight buffer fd\n");
1288                         return NULL;
1289                 }
1290
1291                 unlink(fname);
1292         }
1293
1294         if (ftruncate(mfd, size) == -1) {
1295                 RTE_LOG(ERR, VHOST_CONFIG,
1296                         "failed to alloc inflight buffer\n");
1297                 close(mfd);
1298                 return NULL;
1299         }
1300
1301         ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd, 0);
1302         if (ptr == MAP_FAILED) {
1303                 RTE_LOG(ERR, VHOST_CONFIG,
1304                         "failed to mmap inflight buffer\n");
1305                 close(mfd);
1306                 return NULL;
1307         }
1308
1309         *fd = mfd;
1310         return ptr;
1311 }
1312
1313 static uint32_t
1314 get_pervq_shm_size_split(uint16_t queue_size)
1315 {
1316         return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_split) *
1317                                   queue_size + sizeof(uint64_t) +
1318                                   sizeof(uint16_t) * 4, INFLIGHT_ALIGNMENT);
1319 }
1320
1321 static uint32_t
1322 get_pervq_shm_size_packed(uint16_t queue_size)
1323 {
1324         return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_packed)
1325                                   * queue_size + sizeof(uint64_t) +
1326                                   sizeof(uint16_t) * 6 + sizeof(uint8_t) * 9,
1327                                   INFLIGHT_ALIGNMENT);
1328 }
1329
1330 static int
1331 vhost_user_get_inflight_fd(struct virtio_net **pdev,
1332                            VhostUserMsg *msg,
1333                            int main_fd __rte_unused)
1334 {
1335         struct rte_vhost_inflight_info_packed *inflight_packed;
1336         uint64_t pervq_inflight_size, mmap_size;
1337         uint16_t num_queues, queue_size;
1338         struct virtio_net *dev = *pdev;
1339         int fd, i, j;
1340         void *addr;
1341
1342         if (msg->size != sizeof(msg->payload.inflight)) {
1343                 RTE_LOG(ERR, VHOST_CONFIG,
1344                         "invalid get_inflight_fd message size is %d\n",
1345                         msg->size);
1346                 return RTE_VHOST_MSG_RESULT_ERR;
1347         }
1348
1349         if (dev->inflight_info == NULL) {
1350                 dev->inflight_info = calloc(1,
1351                                             sizeof(struct inflight_mem_info));
1352                 if (!dev->inflight_info) {
1353                         RTE_LOG(ERR, VHOST_CONFIG,
1354                                 "failed to alloc dev inflight area\n");
1355                         return RTE_VHOST_MSG_RESULT_ERR;
1356                 }
1357         }
1358
1359         num_queues = msg->payload.inflight.num_queues;
1360         queue_size = msg->payload.inflight.queue_size;
1361
1362         RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd num_queues: %u\n",
1363                 msg->payload.inflight.num_queues);
1364         RTE_LOG(INFO, VHOST_CONFIG, "get_inflight_fd queue_size: %u\n",
1365                 msg->payload.inflight.queue_size);
1366
1367         if (vq_is_packed(dev))
1368                 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1369         else
1370                 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1371
1372         mmap_size = num_queues * pervq_inflight_size;
1373         addr = inflight_mem_alloc("vhost-inflight", mmap_size, &fd);
1374         if (!addr) {
1375                 RTE_LOG(ERR, VHOST_CONFIG,
1376                         "failed to alloc vhost inflight area\n");
1377                         msg->payload.inflight.mmap_size = 0;
1378                 return RTE_VHOST_MSG_RESULT_ERR;
1379         }
1380         memset(addr, 0, mmap_size);
1381
1382         dev->inflight_info->addr = addr;
1383         dev->inflight_info->size = msg->payload.inflight.mmap_size = mmap_size;
1384         dev->inflight_info->fd = msg->fds[0] = fd;
1385         msg->payload.inflight.mmap_offset = 0;
1386         msg->fd_num = 1;
1387
1388         if (vq_is_packed(dev)) {
1389                 for (i = 0; i < num_queues; i++) {
1390                         inflight_packed =
1391                                 (struct rte_vhost_inflight_info_packed *)addr;
1392                         inflight_packed->used_wrap_counter = 1;
1393                         inflight_packed->old_used_wrap_counter = 1;
1394                         for (j = 0; j < queue_size; j++)
1395                                 inflight_packed->desc[j].next = j + 1;
1396                         addr = (void *)((char *)addr + pervq_inflight_size);
1397                 }
1398         }
1399
1400         RTE_LOG(INFO, VHOST_CONFIG,
1401                 "send inflight mmap_size: %"PRIu64"\n",
1402                 msg->payload.inflight.mmap_size);
1403         RTE_LOG(INFO, VHOST_CONFIG,
1404                 "send inflight mmap_offset: %"PRIu64"\n",
1405                 msg->payload.inflight.mmap_offset);
1406         RTE_LOG(INFO, VHOST_CONFIG,
1407                 "send inflight fd: %d\n", msg->fds[0]);
1408
1409         return RTE_VHOST_MSG_RESULT_REPLY;
1410 }
1411
1412 static int
1413 vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
1414                            int main_fd __rte_unused)
1415 {
1416         uint64_t mmap_size, mmap_offset;
1417         uint16_t num_queues, queue_size;
1418         struct virtio_net *dev = *pdev;
1419         uint32_t pervq_inflight_size;
1420         struct vhost_virtqueue *vq;
1421         void *addr;
1422         int fd, i;
1423
1424         fd = msg->fds[0];
1425         if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
1426                 RTE_LOG(ERR, VHOST_CONFIG,
1427                         "invalid set_inflight_fd message size is %d,fd is %d\n",
1428                         msg->size, fd);
1429                 return RTE_VHOST_MSG_RESULT_ERR;
1430         }
1431
1432         mmap_size = msg->payload.inflight.mmap_size;
1433         mmap_offset = msg->payload.inflight.mmap_offset;
1434         num_queues = msg->payload.inflight.num_queues;
1435         queue_size = msg->payload.inflight.queue_size;
1436
1437         if (vq_is_packed(dev))
1438                 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1439         else
1440                 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1441
1442         RTE_LOG(INFO, VHOST_CONFIG,
1443                 "set_inflight_fd mmap_size: %"PRIu64"\n", mmap_size);
1444         RTE_LOG(INFO, VHOST_CONFIG,
1445                 "set_inflight_fd mmap_offset: %"PRIu64"\n", mmap_offset);
1446         RTE_LOG(INFO, VHOST_CONFIG,
1447                 "set_inflight_fd num_queues: %u\n", num_queues);
1448         RTE_LOG(INFO, VHOST_CONFIG,
1449                 "set_inflight_fd queue_size: %u\n", queue_size);
1450         RTE_LOG(INFO, VHOST_CONFIG,
1451                 "set_inflight_fd fd: %d\n", fd);
1452         RTE_LOG(INFO, VHOST_CONFIG,
1453                 "set_inflight_fd pervq_inflight_size: %d\n",
1454                 pervq_inflight_size);
1455
1456         if (!dev->inflight_info) {
1457                 dev->inflight_info = calloc(1,
1458                                             sizeof(struct inflight_mem_info));
1459                 if (dev->inflight_info == NULL) {
1460                         RTE_LOG(ERR, VHOST_CONFIG,
1461                                 "failed to alloc dev inflight area\n");
1462                         return RTE_VHOST_MSG_RESULT_ERR;
1463                 }
1464         }
1465
1466         if (dev->inflight_info->addr)
1467                 munmap(dev->inflight_info->addr, dev->inflight_info->size);
1468
1469         addr = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
1470                     fd, mmap_offset);
1471         if (addr == MAP_FAILED) {
1472                 RTE_LOG(ERR, VHOST_CONFIG, "failed to mmap share memory.\n");
1473                 return RTE_VHOST_MSG_RESULT_ERR;
1474         }
1475
1476         if (dev->inflight_info->fd)
1477                 close(dev->inflight_info->fd);
1478
1479         dev->inflight_info->fd = fd;
1480         dev->inflight_info->addr = addr;
1481         dev->inflight_info->size = mmap_size;
1482
1483         for (i = 0; i < num_queues; i++) {
1484                 vq = dev->virtqueue[i];
1485                 if (vq_is_packed(dev)) {
1486                         vq->inflight_packed = addr;
1487                         vq->inflight_packed->desc_num = queue_size;
1488                 } else {
1489                         vq->inflight_split = addr;
1490                         vq->inflight_split->desc_num = queue_size;
1491                 }
1492                 addr = (void *)((char *)addr + pervq_inflight_size);
1493         }
1494
1495         return RTE_VHOST_MSG_RESULT_OK;
1496 }
1497
1498 static int
1499 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1500                         int main_fd __rte_unused)
1501 {
1502         struct virtio_net *dev = *pdev;
1503         struct vhost_vring_file file;
1504         struct vhost_virtqueue *vq;
1505
1506         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1507         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1508                 file.fd = VIRTIO_INVALID_EVENTFD;
1509         else
1510                 file.fd = msg->fds[0];
1511         RTE_LOG(INFO, VHOST_CONFIG,
1512                 "vring call idx:%d file:%d\n", file.index, file.fd);
1513
1514         vq = dev->virtqueue[file.index];
1515         if (vq->callfd >= 0)
1516                 close(vq->callfd);
1517
1518         vq->callfd = file.fd;
1519
1520         return RTE_VHOST_MSG_RESULT_OK;
1521 }
1522
1523 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1524                         struct VhostUserMsg *msg,
1525                         int main_fd __rte_unused)
1526 {
1527         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1528                 close(msg->fds[0]);
1529         RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1530
1531         return RTE_VHOST_MSG_RESULT_OK;
1532 }
1533
1534 static int
1535 resubmit_desc_compare(const void *a, const void *b)
1536 {
1537         const struct rte_vhost_resubmit_desc *desc0 = a;
1538         const struct rte_vhost_resubmit_desc *desc1 = b;
1539
1540         if (desc1->counter > desc0->counter)
1541                 return 1;
1542
1543         return -1;
1544 }
1545
1546 static int
1547 vhost_check_queue_inflights_split(struct virtio_net *dev,
1548                                   struct vhost_virtqueue *vq)
1549 {
1550         uint16_t i;
1551         uint16_t resubmit_num = 0, last_io, num;
1552         struct vring_used *used = vq->used;
1553         struct rte_vhost_resubmit_info *resubmit;
1554         struct rte_vhost_inflight_info_split *inflight_split;
1555
1556         if (!(dev->protocol_features &
1557             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1558                 return RTE_VHOST_MSG_RESULT_OK;
1559
1560         if ((!vq->inflight_split))
1561                 return RTE_VHOST_MSG_RESULT_ERR;
1562
1563         if (!vq->inflight_split->version) {
1564                 vq->inflight_split->version = INFLIGHT_VERSION;
1565                 return RTE_VHOST_MSG_RESULT_OK;
1566         }
1567
1568         if (vq->resubmit_inflight)
1569                 return RTE_VHOST_MSG_RESULT_OK;
1570
1571         inflight_split = vq->inflight_split;
1572         vq->global_counter = 0;
1573         last_io = inflight_split->last_inflight_io;
1574
1575         if (inflight_split->used_idx != used->idx) {
1576                 inflight_split->desc[last_io].inflight = 0;
1577                 rte_smp_mb();
1578                 inflight_split->used_idx = used->idx;
1579         }
1580
1581         for (i = 0; i < inflight_split->desc_num; i++) {
1582                 if (inflight_split->desc[i].inflight == 1)
1583                         resubmit_num++;
1584         }
1585
1586         vq->last_avail_idx += resubmit_num;
1587
1588         if (resubmit_num) {
1589                 resubmit  = calloc(1, sizeof(struct rte_vhost_resubmit_info));
1590                 if (!resubmit) {
1591                         RTE_LOG(ERR, VHOST_CONFIG,
1592                                 "failed to allocate memory for resubmit info.\n");
1593                         return RTE_VHOST_MSG_RESULT_ERR;
1594                 }
1595
1596                 resubmit->resubmit_list = calloc(resubmit_num,
1597                         sizeof(struct rte_vhost_resubmit_desc));
1598                 if (!resubmit->resubmit_list) {
1599                         RTE_LOG(ERR, VHOST_CONFIG,
1600                                 "failed to allocate memory for inflight desc.\n");
1601                         free(resubmit);
1602                         return RTE_VHOST_MSG_RESULT_ERR;
1603                 }
1604
1605                 num = 0;
1606                 for (i = 0; i < vq->inflight_split->desc_num; i++) {
1607                         if (vq->inflight_split->desc[i].inflight == 1) {
1608                                 resubmit->resubmit_list[num].index = i;
1609                                 resubmit->resubmit_list[num].counter =
1610                                         inflight_split->desc[i].counter;
1611                                 num++;
1612                         }
1613                 }
1614                 resubmit->resubmit_num = num;
1615
1616                 if (resubmit->resubmit_num > 1)
1617                         qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1618                               sizeof(struct rte_vhost_resubmit_desc),
1619                               resubmit_desc_compare);
1620
1621                 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1622                 vq->resubmit_inflight = resubmit;
1623         }
1624
1625         return RTE_VHOST_MSG_RESULT_OK;
1626 }
1627
1628 static int
1629 vhost_check_queue_inflights_packed(struct virtio_net *dev,
1630                                    struct vhost_virtqueue *vq)
1631 {
1632         uint16_t i;
1633         uint16_t resubmit_num = 0, old_used_idx, num;
1634         struct rte_vhost_resubmit_info *resubmit;
1635         struct rte_vhost_inflight_info_packed *inflight_packed;
1636
1637         if (!(dev->protocol_features &
1638             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1639                 return RTE_VHOST_MSG_RESULT_OK;
1640
1641         if ((!vq->inflight_packed))
1642                 return RTE_VHOST_MSG_RESULT_ERR;
1643
1644         if (!vq->inflight_packed->version) {
1645                 vq->inflight_packed->version = INFLIGHT_VERSION;
1646                 return RTE_VHOST_MSG_RESULT_OK;
1647         }
1648
1649         if (vq->resubmit_inflight)
1650                 return RTE_VHOST_MSG_RESULT_OK;
1651
1652         inflight_packed = vq->inflight_packed;
1653         vq->global_counter = 0;
1654         old_used_idx = inflight_packed->old_used_idx;
1655
1656         if (inflight_packed->used_idx != old_used_idx) {
1657                 if (inflight_packed->desc[old_used_idx].inflight == 0) {
1658                         inflight_packed->old_used_idx =
1659                                 inflight_packed->used_idx;
1660                         inflight_packed->old_used_wrap_counter =
1661                                 inflight_packed->used_wrap_counter;
1662                         inflight_packed->old_free_head =
1663                                 inflight_packed->free_head;
1664                 } else {
1665                         inflight_packed->used_idx =
1666                                 inflight_packed->old_used_idx;
1667                         inflight_packed->used_wrap_counter =
1668                                 inflight_packed->old_used_wrap_counter;
1669                         inflight_packed->free_head =
1670                                 inflight_packed->old_free_head;
1671                 }
1672         }
1673
1674         for (i = 0; i < inflight_packed->desc_num; i++) {
1675                 if (inflight_packed->desc[i].inflight == 1)
1676                         resubmit_num++;
1677         }
1678
1679         if (resubmit_num) {
1680                 resubmit = calloc(1, sizeof(struct rte_vhost_resubmit_info));
1681                 if (resubmit == NULL) {
1682                         RTE_LOG(ERR, VHOST_CONFIG,
1683                                 "failed to allocate memory for resubmit info.\n");
1684                         return RTE_VHOST_MSG_RESULT_ERR;
1685                 }
1686
1687                 resubmit->resubmit_list = calloc(resubmit_num,
1688                         sizeof(struct rte_vhost_resubmit_desc));
1689                 if (resubmit->resubmit_list == NULL) {
1690                         RTE_LOG(ERR, VHOST_CONFIG,
1691                                 "failed to allocate memory for resubmit desc.\n");
1692                         free(resubmit);
1693                         return RTE_VHOST_MSG_RESULT_ERR;
1694                 }
1695
1696                 num = 0;
1697                 for (i = 0; i < inflight_packed->desc_num; i++) {
1698                         if (vq->inflight_packed->desc[i].inflight == 1) {
1699                                 resubmit->resubmit_list[num].index = i;
1700                                 resubmit->resubmit_list[num].counter =
1701                                         inflight_packed->desc[i].counter;
1702                                 num++;
1703                         }
1704                 }
1705                 resubmit->resubmit_num = num;
1706
1707                 if (resubmit->resubmit_num > 1)
1708                         qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1709                               sizeof(struct rte_vhost_resubmit_desc),
1710                               resubmit_desc_compare);
1711
1712                 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1713                 vq->resubmit_inflight = resubmit;
1714         }
1715
1716         return RTE_VHOST_MSG_RESULT_OK;
1717 }
1718
1719 static int
1720 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1721                         int main_fd __rte_unused)
1722 {
1723         struct virtio_net *dev = *pdev;
1724         struct vhost_vring_file file;
1725         struct vhost_virtqueue *vq;
1726
1727         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1728         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1729                 file.fd = VIRTIO_INVALID_EVENTFD;
1730         else
1731                 file.fd = msg->fds[0];
1732         RTE_LOG(INFO, VHOST_CONFIG,
1733                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1734
1735         /* Interpret ring addresses only when ring is started. */
1736         dev = translate_ring_addresses(dev, file.index);
1737         if (!dev)
1738                 return RTE_VHOST_MSG_RESULT_ERR;
1739
1740         *pdev = dev;
1741
1742         vq = dev->virtqueue[file.index];
1743
1744         /*
1745          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1746          * the ring starts already enabled. Otherwise, it is enabled via
1747          * the SET_VRING_ENABLE message.
1748          */
1749         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
1750                 vq->enabled = 1;
1751                 if (dev->notify_ops->vring_state_changed)
1752                         dev->notify_ops->vring_state_changed(
1753                                 dev->vid, file.index, 1);
1754         }
1755
1756         if (vq->kickfd >= 0)
1757                 close(vq->kickfd);
1758         vq->kickfd = file.fd;
1759
1760         if (vq_is_packed(dev)) {
1761                 if (vhost_check_queue_inflights_packed(dev, vq)) {
1762                         RTE_LOG(ERR, VHOST_CONFIG,
1763                                 "failed to inflights for vq: %d\n", file.index);
1764                         return RTE_VHOST_MSG_RESULT_ERR;
1765                 }
1766         } else {
1767                 if (vhost_check_queue_inflights_split(dev, vq)) {
1768                         RTE_LOG(ERR, VHOST_CONFIG,
1769                                 "failed to inflights for vq: %d\n", file.index);
1770                         return RTE_VHOST_MSG_RESULT_ERR;
1771                 }
1772         }
1773
1774         return RTE_VHOST_MSG_RESULT_OK;
1775 }
1776
1777 static void
1778 free_zmbufs(struct vhost_virtqueue *vq)
1779 {
1780         drain_zmbuf_list(vq);
1781
1782         rte_free(vq->zmbufs);
1783 }
1784
1785 /*
1786  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1787  */
1788 static int
1789 vhost_user_get_vring_base(struct virtio_net **pdev,
1790                         struct VhostUserMsg *msg,
1791                         int main_fd __rte_unused)
1792 {
1793         struct virtio_net *dev = *pdev;
1794         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1795         uint64_t val;
1796
1797         /* We have to stop the queue (virtio) if it is running. */
1798         vhost_destroy_device_notify(dev);
1799
1800         dev->flags &= ~VIRTIO_DEV_READY;
1801         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1802
1803         /* Here we are safe to get the indexes */
1804         if (vq_is_packed(dev)) {
1805                 /*
1806                  * Bit[0:14]: avail index
1807                  * Bit[15]: avail wrap counter
1808                  */
1809                 val = vq->last_avail_idx & 0x7fff;
1810                 val |= vq->avail_wrap_counter << 15;
1811                 msg->payload.state.num = val;
1812         } else {
1813                 msg->payload.state.num = vq->last_avail_idx;
1814         }
1815
1816         RTE_LOG(INFO, VHOST_CONFIG,
1817                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1818                 msg->payload.state.num);
1819         /*
1820          * Based on current qemu vhost-user implementation, this message is
1821          * sent and only sent in vhost_vring_stop.
1822          * TODO: cleanup the vring, it isn't usable since here.
1823          */
1824         if (vq->kickfd >= 0)
1825                 close(vq->kickfd);
1826
1827         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1828
1829         if (vq->callfd >= 0)
1830                 close(vq->callfd);
1831
1832         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1833
1834         vq->signalled_used_valid = false;
1835
1836         if (dev->dequeue_zero_copy)
1837                 free_zmbufs(vq);
1838         if (vq_is_packed(dev)) {
1839                 rte_free(vq->shadow_used_packed);
1840                 vq->shadow_used_packed = NULL;
1841         } else {
1842                 rte_free(vq->shadow_used_split);
1843                 vq->shadow_used_split = NULL;
1844         }
1845
1846         rte_free(vq->batch_copy_elems);
1847         vq->batch_copy_elems = NULL;
1848
1849         msg->size = sizeof(msg->payload.state);
1850         msg->fd_num = 0;
1851
1852         vring_invalidate(dev, vq);
1853
1854         return RTE_VHOST_MSG_RESULT_REPLY;
1855 }
1856
1857 /*
1858  * when virtio queues are ready to work, qemu will send us to
1859  * enable the virtio queue pair.
1860  */
1861 static int
1862 vhost_user_set_vring_enable(struct virtio_net **pdev,
1863                         struct VhostUserMsg *msg,
1864                         int main_fd __rte_unused)
1865 {
1866         struct virtio_net *dev = *pdev;
1867         int enable = (int)msg->payload.state.num;
1868         int index = (int)msg->payload.state.index;
1869         struct rte_vdpa_device *vdpa_dev;
1870         int did = -1;
1871
1872         RTE_LOG(INFO, VHOST_CONFIG,
1873                 "set queue enable: %d to qp idx: %d\n",
1874                 enable, index);
1875
1876         did = dev->vdpa_dev_id;
1877         vdpa_dev = rte_vdpa_get_device(did);
1878         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1879                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1880
1881         if (dev->notify_ops->vring_state_changed)
1882                 dev->notify_ops->vring_state_changed(dev->vid,
1883                                 index, enable);
1884
1885         /* On disable, rings have to be stopped being processed. */
1886         if (!enable && dev->dequeue_zero_copy)
1887                 drain_zmbuf_list(dev->virtqueue[index]);
1888
1889         dev->virtqueue[index]->enabled = enable;
1890
1891         return RTE_VHOST_MSG_RESULT_OK;
1892 }
1893
1894 static int
1895 vhost_user_get_protocol_features(struct virtio_net **pdev,
1896                         struct VhostUserMsg *msg,
1897                         int main_fd __rte_unused)
1898 {
1899         struct virtio_net *dev = *pdev;
1900         uint64_t features, protocol_features;
1901
1902         rte_vhost_driver_get_features(dev->ifname, &features);
1903         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1904
1905         /*
1906          * REPLY_ACK protocol feature is only mandatory for now
1907          * for IOMMU feature. If IOMMU is explicitly disabled by the
1908          * application, disable also REPLY_ACK feature for older buggy
1909          * Qemu versions (from v2.7.0 to v2.9.0).
1910          */
1911         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1912                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1913
1914         msg->payload.u64 = protocol_features;
1915         msg->size = sizeof(msg->payload.u64);
1916         msg->fd_num = 0;
1917
1918         return RTE_VHOST_MSG_RESULT_REPLY;
1919 }
1920
1921 static int
1922 vhost_user_set_protocol_features(struct virtio_net **pdev,
1923                         struct VhostUserMsg *msg,
1924                         int main_fd __rte_unused)
1925 {
1926         struct virtio_net *dev = *pdev;
1927         uint64_t protocol_features = msg->payload.u64;
1928         uint64_t slave_protocol_features = 0;
1929
1930         rte_vhost_driver_get_protocol_features(dev->ifname,
1931                         &slave_protocol_features);
1932         if (protocol_features & ~slave_protocol_features) {
1933                 RTE_LOG(ERR, VHOST_CONFIG,
1934                         "(%d) received invalid protocol features.\n",
1935                         dev->vid);
1936                 return RTE_VHOST_MSG_RESULT_ERR;
1937         }
1938
1939         dev->protocol_features = protocol_features;
1940         RTE_LOG(INFO, VHOST_CONFIG,
1941                 "negotiated Vhost-user protocol features: 0x%" PRIx64 "\n",
1942                 dev->protocol_features);
1943
1944         return RTE_VHOST_MSG_RESULT_OK;
1945 }
1946
1947 static int
1948 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1949                         int main_fd __rte_unused)
1950 {
1951         struct virtio_net *dev = *pdev;
1952         int fd = msg->fds[0];
1953         uint64_t size, off;
1954         void *addr;
1955
1956         if (fd < 0) {
1957                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1958                 return RTE_VHOST_MSG_RESULT_ERR;
1959         }
1960
1961         if (msg->size != sizeof(VhostUserLog)) {
1962                 RTE_LOG(ERR, VHOST_CONFIG,
1963                         "invalid log base msg size: %"PRId32" != %d\n",
1964                         msg->size, (int)sizeof(VhostUserLog));
1965                 return RTE_VHOST_MSG_RESULT_ERR;
1966         }
1967
1968         size = msg->payload.log.mmap_size;
1969         off  = msg->payload.log.mmap_offset;
1970
1971         /* Don't allow mmap_offset to point outside the mmap region */
1972         if (off > size) {
1973                 RTE_LOG(ERR, VHOST_CONFIG,
1974                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1975                         off, size);
1976                 return RTE_VHOST_MSG_RESULT_ERR;
1977         }
1978
1979         RTE_LOG(INFO, VHOST_CONFIG,
1980                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1981                 size, off);
1982
1983         /*
1984          * mmap from 0 to workaround a hugepage mmap bug: mmap will
1985          * fail when offset is not page size aligned.
1986          */
1987         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1988         close(fd);
1989         if (addr == MAP_FAILED) {
1990                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1991                 return RTE_VHOST_MSG_RESULT_ERR;
1992         }
1993
1994         /*
1995          * Free previously mapped log memory on occasionally
1996          * multiple VHOST_USER_SET_LOG_BASE.
1997          */
1998         if (dev->log_addr) {
1999                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
2000         }
2001         dev->log_addr = (uint64_t)(uintptr_t)addr;
2002         dev->log_base = dev->log_addr + off;
2003         dev->log_size = size;
2004
2005         /*
2006          * The spec is not clear about it (yet), but QEMU doesn't expect
2007          * any payload in the reply.
2008          */
2009         msg->size = 0;
2010         msg->fd_num = 0;
2011
2012         return RTE_VHOST_MSG_RESULT_REPLY;
2013 }
2014
2015 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
2016                         struct VhostUserMsg *msg,
2017                         int main_fd __rte_unused)
2018 {
2019         close(msg->fds[0]);
2020         RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
2021
2022         return RTE_VHOST_MSG_RESULT_OK;
2023 }
2024
2025 /*
2026  * An rarp packet is constructed and broadcasted to notify switches about
2027  * the new location of the migrated VM, so that packets from outside will
2028  * not be lost after migration.
2029  *
2030  * However, we don't actually "send" a rarp packet here, instead, we set
2031  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
2032  */
2033 static int
2034 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
2035                         int main_fd __rte_unused)
2036 {
2037         struct virtio_net *dev = *pdev;
2038         uint8_t *mac = (uint8_t *)&msg->payload.u64;
2039         struct rte_vdpa_device *vdpa_dev;
2040         int did = -1;
2041
2042         RTE_LOG(DEBUG, VHOST_CONFIG,
2043                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
2044                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
2045         memcpy(dev->mac.addr_bytes, mac, 6);
2046
2047         /*
2048          * Set the flag to inject a RARP broadcast packet at
2049          * rte_vhost_dequeue_burst().
2050          *
2051          * rte_smp_wmb() is for making sure the mac is copied
2052          * before the flag is set.
2053          */
2054         rte_smp_wmb();
2055         rte_atomic16_set(&dev->broadcast_rarp, 1);
2056         did = dev->vdpa_dev_id;
2057         vdpa_dev = rte_vdpa_get_device(did);
2058         if (vdpa_dev && vdpa_dev->ops->migration_done)
2059                 vdpa_dev->ops->migration_done(dev->vid);
2060
2061         return RTE_VHOST_MSG_RESULT_OK;
2062 }
2063
2064 static int
2065 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
2066                         int main_fd __rte_unused)
2067 {
2068         struct virtio_net *dev = *pdev;
2069         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
2070                         msg->payload.u64 > VIRTIO_MAX_MTU) {
2071                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
2072                                 msg->payload.u64);
2073
2074                 return RTE_VHOST_MSG_RESULT_ERR;
2075         }
2076
2077         dev->mtu = msg->payload.u64;
2078
2079         return RTE_VHOST_MSG_RESULT_OK;
2080 }
2081
2082 static int
2083 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
2084                         int main_fd __rte_unused)
2085 {
2086         struct virtio_net *dev = *pdev;
2087         int fd = msg->fds[0];
2088
2089         if (fd < 0) {
2090                 RTE_LOG(ERR, VHOST_CONFIG,
2091                                 "Invalid file descriptor for slave channel (%d)\n",
2092                                 fd);
2093                 return RTE_VHOST_MSG_RESULT_ERR;
2094         }
2095
2096         if (dev->slave_req_fd >= 0)
2097                 close(dev->slave_req_fd);
2098
2099         dev->slave_req_fd = fd;
2100
2101         return RTE_VHOST_MSG_RESULT_OK;
2102 }
2103
2104 static int
2105 is_vring_iotlb_split(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2106 {
2107         struct vhost_vring_addr *ra;
2108         uint64_t start, end, len;
2109
2110         start = imsg->iova;
2111         end = start + imsg->size;
2112
2113         ra = &vq->ring_addrs;
2114         len = sizeof(struct vring_desc) * vq->size;
2115         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2116                 return 1;
2117
2118         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
2119         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2120                 return 1;
2121
2122         len = sizeof(struct vring_used) +
2123                sizeof(struct vring_used_elem) * vq->size;
2124         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2125                 return 1;
2126
2127         return 0;
2128 }
2129
2130 static int
2131 is_vring_iotlb_packed(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2132 {
2133         struct vhost_vring_addr *ra;
2134         uint64_t start, end, len;
2135
2136         start = imsg->iova;
2137         end = start + imsg->size;
2138
2139         ra = &vq->ring_addrs;
2140         len = sizeof(struct vring_packed_desc) * vq->size;
2141         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2142                 return 1;
2143
2144         len = sizeof(struct vring_packed_desc_event);
2145         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2146                 return 1;
2147
2148         len = sizeof(struct vring_packed_desc_event);
2149         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2150                 return 1;
2151
2152         return 0;
2153 }
2154
2155 static int is_vring_iotlb(struct virtio_net *dev,
2156                           struct vhost_virtqueue *vq,
2157                           struct vhost_iotlb_msg *imsg)
2158 {
2159         if (vq_is_packed(dev))
2160                 return is_vring_iotlb_packed(vq, imsg);
2161         else
2162                 return is_vring_iotlb_split(vq, imsg);
2163 }
2164
2165 static int
2166 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
2167                         int main_fd __rte_unused)
2168 {
2169         struct virtio_net *dev = *pdev;
2170         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
2171         uint16_t i;
2172         uint64_t vva, len;
2173
2174         switch (imsg->type) {
2175         case VHOST_IOTLB_UPDATE:
2176                 len = imsg->size;
2177                 vva = qva_to_vva(dev, imsg->uaddr, &len);
2178                 if (!vva)
2179                         return RTE_VHOST_MSG_RESULT_ERR;
2180
2181                 for (i = 0; i < dev->nr_vring; i++) {
2182                         struct vhost_virtqueue *vq = dev->virtqueue[i];
2183
2184                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
2185                                         len, imsg->perm);
2186
2187                         if (is_vring_iotlb(dev, vq, imsg))
2188                                 *pdev = dev = translate_ring_addresses(dev, i);
2189                 }
2190                 break;
2191         case VHOST_IOTLB_INVALIDATE:
2192                 for (i = 0; i < dev->nr_vring; i++) {
2193                         struct vhost_virtqueue *vq = dev->virtqueue[i];
2194
2195                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
2196                                         imsg->size);
2197
2198                         if (is_vring_iotlb(dev, vq, imsg))
2199                                 vring_invalidate(dev, vq);
2200                 }
2201                 break;
2202         default:
2203                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
2204                                 imsg->type);
2205                 return RTE_VHOST_MSG_RESULT_ERR;
2206         }
2207
2208         return RTE_VHOST_MSG_RESULT_OK;
2209 }
2210
2211 static int
2212 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
2213                         struct VhostUserMsg *msg,
2214                         int main_fd __rte_unused)
2215 {
2216         struct virtio_net *dev = *pdev;
2217 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
2218         struct uffdio_api api_struct;
2219
2220         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
2221
2222         if (dev->postcopy_ufd == -1) {
2223                 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
2224                         strerror(errno));
2225                 return RTE_VHOST_MSG_RESULT_ERR;
2226         }
2227         api_struct.api = UFFD_API;
2228         api_struct.features = 0;
2229         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
2230                 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
2231                         strerror(errno));
2232                 close(dev->postcopy_ufd);
2233                 dev->postcopy_ufd = -1;
2234                 return RTE_VHOST_MSG_RESULT_ERR;
2235         }
2236         msg->fds[0] = dev->postcopy_ufd;
2237         msg->fd_num = 1;
2238
2239         return RTE_VHOST_MSG_RESULT_REPLY;
2240 #else
2241         dev->postcopy_ufd = -1;
2242         msg->fd_num = 0;
2243
2244         return RTE_VHOST_MSG_RESULT_ERR;
2245 #endif
2246 }
2247
2248 static int
2249 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
2250                         struct VhostUserMsg *msg __rte_unused,
2251                         int main_fd __rte_unused)
2252 {
2253         struct virtio_net *dev = *pdev;
2254
2255         if (dev->mem && dev->mem->nregions) {
2256                 RTE_LOG(ERR, VHOST_CONFIG,
2257                         "Regions already registered at postcopy-listen\n");
2258                 return RTE_VHOST_MSG_RESULT_ERR;
2259         }
2260         dev->postcopy_listening = 1;
2261
2262         return RTE_VHOST_MSG_RESULT_OK;
2263 }
2264
2265 static int
2266 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
2267                         int main_fd __rte_unused)
2268 {
2269         struct virtio_net *dev = *pdev;
2270
2271         dev->postcopy_listening = 0;
2272         if (dev->postcopy_ufd >= 0) {
2273                 close(dev->postcopy_ufd);
2274                 dev->postcopy_ufd = -1;
2275         }
2276
2277         msg->payload.u64 = 0;
2278         msg->size = sizeof(msg->payload.u64);
2279         msg->fd_num = 0;
2280
2281         return RTE_VHOST_MSG_RESULT_REPLY;
2282 }
2283
2284 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
2285                                         struct VhostUserMsg *msg,
2286                                         int main_fd);
2287 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
2288         [VHOST_USER_NONE] = NULL,
2289         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
2290         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
2291         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
2292         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
2293         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
2294         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
2295         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
2296         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
2297         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
2298         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
2299         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
2300         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
2301         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
2302         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
2303         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
2304         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
2305         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
2306         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
2307         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
2308         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
2309         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
2310         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
2311         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
2312         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
2313         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
2314         [VHOST_USER_GET_INFLIGHT_FD] = vhost_user_get_inflight_fd,
2315         [VHOST_USER_SET_INFLIGHT_FD] = vhost_user_set_inflight_fd,
2316 };
2317
2318 /* return bytes# of read on success or negative val on failure. */
2319 static int
2320 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
2321 {
2322         int ret;
2323
2324         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
2325                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
2326         if (ret <= 0)
2327                 return ret;
2328
2329         if (msg->size) {
2330                 if (msg->size > sizeof(msg->payload)) {
2331                         RTE_LOG(ERR, VHOST_CONFIG,
2332                                 "invalid msg size: %d\n", msg->size);
2333                         return -1;
2334                 }
2335                 ret = read(sockfd, &msg->payload, msg->size);
2336                 if (ret <= 0)
2337                         return ret;
2338                 if (ret != (int)msg->size) {
2339                         RTE_LOG(ERR, VHOST_CONFIG,
2340                                 "read control message failed\n");
2341                         return -1;
2342                 }
2343         }
2344
2345         return ret;
2346 }
2347
2348 static int
2349 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
2350 {
2351         if (!msg)
2352                 return 0;
2353
2354         return send_fd_message(sockfd, (char *)msg,
2355                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
2356 }
2357
2358 static int
2359 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
2360 {
2361         if (!msg)
2362                 return 0;
2363
2364         msg->flags &= ~VHOST_USER_VERSION_MASK;
2365         msg->flags &= ~VHOST_USER_NEED_REPLY;
2366         msg->flags |= VHOST_USER_VERSION;
2367         msg->flags |= VHOST_USER_REPLY_MASK;
2368
2369         return send_vhost_message(sockfd, msg);
2370 }
2371
2372 static int
2373 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
2374 {
2375         int ret;
2376
2377         if (msg->flags & VHOST_USER_NEED_REPLY)
2378                 rte_spinlock_lock(&dev->slave_req_lock);
2379
2380         ret = send_vhost_message(dev->slave_req_fd, msg);
2381         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
2382                 rte_spinlock_unlock(&dev->slave_req_lock);
2383
2384         return ret;
2385 }
2386
2387 /*
2388  * Allocate a queue pair if it hasn't been allocated yet
2389  */
2390 static int
2391 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
2392                         struct VhostUserMsg *msg)
2393 {
2394         uint16_t vring_idx;
2395
2396         switch (msg->request.master) {
2397         case VHOST_USER_SET_VRING_KICK:
2398         case VHOST_USER_SET_VRING_CALL:
2399         case VHOST_USER_SET_VRING_ERR:
2400                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
2401                 break;
2402         case VHOST_USER_SET_VRING_NUM:
2403         case VHOST_USER_SET_VRING_BASE:
2404         case VHOST_USER_SET_VRING_ENABLE:
2405                 vring_idx = msg->payload.state.index;
2406                 break;
2407         case VHOST_USER_SET_VRING_ADDR:
2408                 vring_idx = msg->payload.addr.index;
2409                 break;
2410         default:
2411                 return 0;
2412         }
2413
2414         if (vring_idx >= VHOST_MAX_VRING) {
2415                 RTE_LOG(ERR, VHOST_CONFIG,
2416                         "invalid vring index: %u\n", vring_idx);
2417                 return -1;
2418         }
2419
2420         if (dev->virtqueue[vring_idx])
2421                 return 0;
2422
2423         return alloc_vring_queue(dev, vring_idx);
2424 }
2425
2426 static void
2427 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
2428 {
2429         unsigned int i = 0;
2430         unsigned int vq_num = 0;
2431
2432         while (vq_num < dev->nr_vring) {
2433                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2434
2435                 if (vq) {
2436                         rte_spinlock_lock(&vq->access_lock);
2437                         vq_num++;
2438                 }
2439                 i++;
2440         }
2441 }
2442
2443 static void
2444 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
2445 {
2446         unsigned int i = 0;
2447         unsigned int vq_num = 0;
2448
2449         while (vq_num < dev->nr_vring) {
2450                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2451
2452                 if (vq) {
2453                         rte_spinlock_unlock(&vq->access_lock);
2454                         vq_num++;
2455                 }
2456                 i++;
2457         }
2458 }
2459
2460 int
2461 vhost_user_msg_handler(int vid, int fd)
2462 {
2463         struct virtio_net *dev;
2464         struct VhostUserMsg msg;
2465         struct rte_vdpa_device *vdpa_dev;
2466         int did = -1;
2467         int ret;
2468         int unlock_required = 0;
2469         bool handled;
2470         int request;
2471
2472         dev = get_device(vid);
2473         if (dev == NULL)
2474                 return -1;
2475
2476         if (!dev->notify_ops) {
2477                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
2478                 if (!dev->notify_ops) {
2479                         RTE_LOG(ERR, VHOST_CONFIG,
2480                                 "failed to get callback ops for driver %s\n",
2481                                 dev->ifname);
2482                         return -1;
2483                 }
2484         }
2485
2486         ret = read_vhost_message(fd, &msg);
2487         if (ret <= 0) {
2488                 if (ret < 0)
2489                         RTE_LOG(ERR, VHOST_CONFIG,
2490                                 "vhost read message failed\n");
2491                 else
2492                         RTE_LOG(INFO, VHOST_CONFIG,
2493                                 "vhost peer closed\n");
2494
2495                 return -1;
2496         }
2497
2498         ret = 0;
2499         request = msg.request.master;
2500         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
2501                         vhost_message_str[request]) {
2502                 if (request != VHOST_USER_IOTLB_MSG)
2503                         RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
2504                                 vhost_message_str[request]);
2505                 else
2506                         RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
2507                                 vhost_message_str[request]);
2508         } else {
2509                 RTE_LOG(DEBUG, VHOST_CONFIG, "External request %d\n", request);
2510         }
2511
2512         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
2513         if (ret < 0) {
2514                 RTE_LOG(ERR, VHOST_CONFIG,
2515                         "failed to alloc queue\n");
2516                 return -1;
2517         }
2518
2519         /*
2520          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
2521          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
2522          * and device is destroyed. destroy_device waits for queues to be
2523          * inactive, so it is safe. Otherwise taking the access_lock
2524          * would cause a dead lock.
2525          */
2526         switch (request) {
2527         case VHOST_USER_SET_FEATURES:
2528         case VHOST_USER_SET_PROTOCOL_FEATURES:
2529         case VHOST_USER_SET_OWNER:
2530         case VHOST_USER_SET_MEM_TABLE:
2531         case VHOST_USER_SET_LOG_BASE:
2532         case VHOST_USER_SET_LOG_FD:
2533         case VHOST_USER_SET_VRING_NUM:
2534         case VHOST_USER_SET_VRING_ADDR:
2535         case VHOST_USER_SET_VRING_BASE:
2536         case VHOST_USER_SET_VRING_KICK:
2537         case VHOST_USER_SET_VRING_CALL:
2538         case VHOST_USER_SET_VRING_ERR:
2539         case VHOST_USER_SET_VRING_ENABLE:
2540         case VHOST_USER_SEND_RARP:
2541         case VHOST_USER_NET_SET_MTU:
2542         case VHOST_USER_SET_SLAVE_REQ_FD:
2543                 vhost_user_lock_all_queue_pairs(dev);
2544                 unlock_required = 1;
2545                 break;
2546         default:
2547                 break;
2548
2549         }
2550
2551         handled = false;
2552         if (dev->extern_ops.pre_msg_handle) {
2553                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
2554                                 (void *)&msg);
2555                 switch (ret) {
2556                 case RTE_VHOST_MSG_RESULT_REPLY:
2557                         send_vhost_reply(fd, &msg);
2558                         /* Fall-through */
2559                 case RTE_VHOST_MSG_RESULT_ERR:
2560                 case RTE_VHOST_MSG_RESULT_OK:
2561                         handled = true;
2562                         goto skip_to_post_handle;
2563                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2564                 default:
2565                         break;
2566                 }
2567         }
2568
2569         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
2570                 if (!vhost_message_handlers[request])
2571                         goto skip_to_post_handle;
2572                 ret = vhost_message_handlers[request](&dev, &msg, fd);
2573
2574                 switch (ret) {
2575                 case RTE_VHOST_MSG_RESULT_ERR:
2576                         RTE_LOG(ERR, VHOST_CONFIG,
2577                                 "Processing %s failed.\n",
2578                                 vhost_message_str[request]);
2579                         handled = true;
2580                         break;
2581                 case RTE_VHOST_MSG_RESULT_OK:
2582                         RTE_LOG(DEBUG, VHOST_CONFIG,
2583                                 "Processing %s succeeded.\n",
2584                                 vhost_message_str[request]);
2585                         handled = true;
2586                         break;
2587                 case RTE_VHOST_MSG_RESULT_REPLY:
2588                         RTE_LOG(DEBUG, VHOST_CONFIG,
2589                                 "Processing %s succeeded and needs reply.\n",
2590                                 vhost_message_str[request]);
2591                         send_vhost_reply(fd, &msg);
2592                         handled = true;
2593                         break;
2594                 default:
2595                         break;
2596                 }
2597         }
2598
2599 skip_to_post_handle:
2600         if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2601                         dev->extern_ops.post_msg_handle) {
2602                 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
2603                                 (void *)&msg);
2604                 switch (ret) {
2605                 case RTE_VHOST_MSG_RESULT_REPLY:
2606                         send_vhost_reply(fd, &msg);
2607                         /* Fall-through */
2608                 case RTE_VHOST_MSG_RESULT_ERR:
2609                 case RTE_VHOST_MSG_RESULT_OK:
2610                         handled = true;
2611                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2612                 default:
2613                         break;
2614                 }
2615         }
2616
2617         if (unlock_required)
2618                 vhost_user_unlock_all_queue_pairs(dev);
2619
2620         /* If message was not handled at this stage, treat it as an error */
2621         if (!handled) {
2622                 RTE_LOG(ERR, VHOST_CONFIG,
2623                         "vhost message (req: %d) was not handled.\n", request);
2624                 ret = RTE_VHOST_MSG_RESULT_ERR;
2625         }
2626
2627         /*
2628          * If the request required a reply that was already sent,
2629          * this optional reply-ack won't be sent as the
2630          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2631          */
2632         if (msg.flags & VHOST_USER_NEED_REPLY) {
2633                 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2634                 msg.size = sizeof(msg.payload.u64);
2635                 msg.fd_num = 0;
2636                 send_vhost_reply(fd, &msg);
2637         } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2638                 RTE_LOG(ERR, VHOST_CONFIG,
2639                         "vhost message handling failed.\n");
2640                 return -1;
2641         }
2642
2643         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2644                 dev->flags |= VIRTIO_DEV_READY;
2645
2646                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2647                         if (dev->dequeue_zero_copy) {
2648                                 RTE_LOG(INFO, VHOST_CONFIG,
2649                                                 "dequeue zero copy is enabled\n");
2650                         }
2651
2652                         if (dev->notify_ops->new_device(dev->vid) == 0)
2653                                 dev->flags |= VIRTIO_DEV_RUNNING;
2654                 }
2655         }
2656
2657         did = dev->vdpa_dev_id;
2658         vdpa_dev = rte_vdpa_get_device(did);
2659         if (vdpa_dev && virtio_is_ready(dev) &&
2660                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2661                         msg.request.master == VHOST_USER_SET_VRING_CALL) {
2662                 if (vdpa_dev->ops->dev_conf)
2663                         vdpa_dev->ops->dev_conf(dev->vid);
2664                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2665         }
2666
2667         return 0;
2668 }
2669
2670 static int process_slave_message_reply(struct virtio_net *dev,
2671                                        const struct VhostUserMsg *msg)
2672 {
2673         struct VhostUserMsg msg_reply;
2674         int ret;
2675
2676         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2677                 return 0;
2678
2679         if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
2680                 ret = -1;
2681                 goto out;
2682         }
2683
2684         if (msg_reply.request.slave != msg->request.slave) {
2685                 RTE_LOG(ERR, VHOST_CONFIG,
2686                         "Received unexpected msg type (%u), expected %u\n",
2687                         msg_reply.request.slave, msg->request.slave);
2688                 ret = -1;
2689                 goto out;
2690         }
2691
2692         ret = msg_reply.payload.u64 ? -1 : 0;
2693
2694 out:
2695         rte_spinlock_unlock(&dev->slave_req_lock);
2696         return ret;
2697 }
2698
2699 int
2700 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2701 {
2702         int ret;
2703         struct VhostUserMsg msg = {
2704                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2705                 .flags = VHOST_USER_VERSION,
2706                 .size = sizeof(msg.payload.iotlb),
2707                 .payload.iotlb = {
2708                         .iova = iova,
2709                         .perm = perm,
2710                         .type = VHOST_IOTLB_MISS,
2711                 },
2712         };
2713
2714         ret = send_vhost_message(dev->slave_req_fd, &msg);
2715         if (ret < 0) {
2716                 RTE_LOG(ERR, VHOST_CONFIG,
2717                                 "Failed to send IOTLB miss message (%d)\n",
2718                                 ret);
2719                 return ret;
2720         }
2721
2722         return 0;
2723 }
2724
2725 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2726                                                     int index, int fd,
2727                                                     uint64_t offset,
2728                                                     uint64_t size)
2729 {
2730         int ret;
2731         struct VhostUserMsg msg = {
2732                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2733                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2734                 .size = sizeof(msg.payload.area),
2735                 .payload.area = {
2736                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
2737                         .size = size,
2738                         .offset = offset,
2739                 },
2740         };
2741
2742         if (fd < 0)
2743                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2744         else {
2745                 msg.fds[0] = fd;
2746                 msg.fd_num = 1;
2747         }
2748
2749         ret = send_vhost_slave_message(dev, &msg);
2750         if (ret < 0) {
2751                 RTE_LOG(ERR, VHOST_CONFIG,
2752                         "Failed to set host notifier (%d)\n", ret);
2753                 return ret;
2754         }
2755
2756         return process_slave_message_reply(dev, &msg);
2757 }
2758
2759 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2760 {
2761         struct virtio_net *dev;
2762         struct rte_vdpa_device *vdpa_dev;
2763         int vfio_device_fd, did, ret = 0;
2764         uint64_t offset, size;
2765         unsigned int i;
2766
2767         dev = get_device(vid);
2768         if (!dev)
2769                 return -ENODEV;
2770
2771         did = dev->vdpa_dev_id;
2772         if (did < 0)
2773                 return -EINVAL;
2774
2775         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2776             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2777             !(dev->protocol_features &
2778                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2779             !(dev->protocol_features &
2780                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2781             !(dev->protocol_features &
2782                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2783                 return -ENOTSUP;
2784
2785         vdpa_dev = rte_vdpa_get_device(did);
2786         if (!vdpa_dev)
2787                 return -ENODEV;
2788
2789         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2790         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2791
2792         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2793         if (vfio_device_fd < 0)
2794                 return -ENOTSUP;
2795
2796         if (enable) {
2797                 for (i = 0; i < dev->nr_vring; i++) {
2798                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2799                                         &size) < 0) {
2800                                 ret = -ENOTSUP;
2801                                 goto disable;
2802                         }
2803
2804                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2805                                         vfio_device_fd, offset, size) < 0) {
2806                                 ret = -EFAULT;
2807                                 goto disable;
2808                         }
2809                 }
2810         } else {
2811 disable:
2812                 for (i = 0; i < dev->nr_vring; i++) {
2813                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
2814                                         0, 0);
2815                 }
2816         }
2817
2818         return ret;
2819 }