lib/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 #include <rte_vfio.h>
  49 #include <rte_errno.h>
  50
  51 #include "iotlb.h"
  52 #include "vhost.h"
  53 #include "vhost_user.h"
  54
  55 #define VIRTIO_MIN_MTU 68
  56 #define VIRTIO_MAX_MTU 65535
  57
  58 #define INFLIGHT_ALIGNMENT      64
  59 #define INFLIGHT_VERSION        0x1
  60
  61 static const char *vhost_message_str[VHOST_USER_MAX] = {
  62         [VHOST_USER_NONE] = "VHOST_USER_NONE",
  63         [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
  64         [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
  65         [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
  66         [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
  67         [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
  68         [VHOST_USER_SET_LOG_BASE] = "VHOST_USER_SET_LOG_BASE",
  69         [VHOST_USER_SET_LOG_FD] = "VHOST_USER_SET_LOG_FD",
  70         [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
  71         [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
  72         [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
  73         [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
  74         [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
  75         [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
  76         [VHOST_USER_SET_VRING_ERR]  = "VHOST_USER_SET_VRING_ERR",
  77         [VHOST_USER_GET_PROTOCOL_FEATURES]  = "VHOST_USER_GET_PROTOCOL_FEATURES",
  78         [VHOST_USER_SET_PROTOCOL_FEATURES]  = "VHOST_USER_SET_PROTOCOL_FEATURES",
  79         [VHOST_USER_GET_QUEUE_NUM]  = "VHOST_USER_GET_QUEUE_NUM",
  80         [VHOST_USER_SET_VRING_ENABLE]  = "VHOST_USER_SET_VRING_ENABLE",
  81         [VHOST_USER_SEND_RARP]  = "VHOST_USER_SEND_RARP",
  82         [VHOST_USER_NET_SET_MTU]  = "VHOST_USER_NET_SET_MTU",
  83         [VHOST_USER_SET_SLAVE_REQ_FD]  = "VHOST_USER_SET_SLAVE_REQ_FD",
  84         [VHOST_USER_IOTLB_MSG]  = "VHOST_USER_IOTLB_MSG",
  85         [VHOST_USER_CRYPTO_CREATE_SESS] = "VHOST_USER_CRYPTO_CREATE_SESS",
  86         [VHOST_USER_CRYPTO_CLOSE_SESS] = "VHOST_USER_CRYPTO_CLOSE_SESS",
  87         [VHOST_USER_POSTCOPY_ADVISE]  = "VHOST_USER_POSTCOPY_ADVISE",
  88         [VHOST_USER_POSTCOPY_LISTEN]  = "VHOST_USER_POSTCOPY_LISTEN",
  89         [VHOST_USER_POSTCOPY_END]  = "VHOST_USER_POSTCOPY_END",
  90         [VHOST_USER_GET_INFLIGHT_FD] = "VHOST_USER_GET_INFLIGHT_FD",
  91         [VHOST_USER_SET_INFLIGHT_FD] = "VHOST_USER_SET_INFLIGHT_FD",
  92         [VHOST_USER_SET_STATUS] = "VHOST_USER_SET_STATUS",
  93         [VHOST_USER_GET_STATUS] = "VHOST_USER_GET_STATUS",
  94 };
  95
  96 static int send_vhost_reply(struct virtio_net *dev, int sockfd, struct VhostUserMsg *msg);
  97 static int read_vhost_message(struct virtio_net *dev, int sockfd, struct VhostUserMsg *msg);
  98
  99 static void
 100 close_msg_fds(struct VhostUserMsg *msg)
 101 {
 102         int i;
 103
 104         for (i = 0; i < msg->fd_num; i++) {
 105                 int fd = msg->fds[i];
 106
 107                 if (fd == -1)
 108                         continue;
 109
 110                 msg->fds[i] = -1;
 111                 close(fd);
 112         }
 113 }
 114
 115 /*
 116  * Ensure the expected number of FDs is received,
 117  * close all FDs and return an error if this is not the case.
 118  */
 119 static int
 120 validate_msg_fds(struct virtio_net *dev, struct VhostUserMsg *msg, int expected_fds)
 121 {
 122         if (msg->fd_num == expected_fds)
 123                 return 0;
 124
 125         VHOST_LOG_CONFIG(ERR, "(%s) expect %d FDs for request %s, received %d\n",
 126                 dev->ifname, expected_fds,
 127                 vhost_message_str[msg->request.master],
 128                 msg->fd_num);
 129
 130         close_msg_fds(msg);
 131
 132         return -1;
 133 }
 134
 135 static uint64_t
 136 get_blk_size(int fd)
 137 {
 138         struct stat stat;
 139         int ret;
 140
 141         ret = fstat(fd, &stat);
 142         return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
 143 }
 144
 145 static int
 146 async_dma_map(struct virtio_net *dev, struct rte_vhost_mem_region *region, bool do_map)
 147 {
 148         uint64_t host_iova;
 149         int ret = 0;
 150
 151         host_iova = rte_mem_virt2iova((void *)(uintptr_t)region->host_user_addr);
 152         if (do_map) {
 153                 /* Add mapped region into the default container of DPDK. */
 154                 ret = rte_vfio_container_dma_map(RTE_VFIO_DEFAULT_CONTAINER_FD,
 155                                                  region->host_user_addr,
 156                                                  host_iova,
 157                                                  region->size);
 158                 if (ret) {
 159                         /*
 160                          * DMA device may bind with kernel driver, in this case,
 161                          * we don't need to program IOMMU manually. However, if no
 162                          * device is bound with vfio/uio in DPDK, and vfio kernel
 163                          * module is loaded, the API will still be called and return
 164                          * with ENODEV/ENOSUP.
 165                          *
 166                          * DPDK vfio only returns ENODEV/ENOSUP in very similar
 167                          * situations(vfio either unsupported, or supported
 168                          * but no devices found). Either way, no mappings could be
 169                          * performed. We treat it as normal case in async path.
 170                          */
 171                         if (rte_errno == ENODEV || rte_errno == ENOTSUP)
 172                                 return 0;
 173
 174                         VHOST_LOG_CONFIG(ERR, "(%s) DMA engine map failed\n", dev->ifname);
 175                         /* DMA mapping errors won't stop VHST_USER_SET_MEM_TABLE. */
 176                         return 0;
 177                 }
 178
 179         } else {
 180                 /* Remove mapped region from the default container of DPDK. */
 181                 ret = rte_vfio_container_dma_unmap(RTE_VFIO_DEFAULT_CONTAINER_FD,
 182                                                    region->host_user_addr,
 183                                                    host_iova,
 184                                                    region->size);
 185                 if (ret) {
 186                         /* like DMA map, ignore the kernel driver case when unmap. */
 187                         if (rte_errno == EINVAL)
 188                                 return 0;
 189
 190                         VHOST_LOG_CONFIG(ERR, "(%s) DMA engine unmap failed\n", dev->ifname);
 191                         return ret;
 192                 }
 193         }
 194
 195         return ret;
 196 }
 197
 198 static void
 199 free_mem_region(struct virtio_net *dev)
 200 {
 201         uint32_t i;
 202         struct rte_vhost_mem_region *reg;
 203
 204         if (!dev || !dev->mem)
 205                 return;
 206
 207         for (i = 0; i < dev->mem->nregions; i++) {
 208                 reg = &dev->mem->regions[i];
 209                 if (reg->host_user_addr) {
 210                         if (dev->async_copy && rte_vfio_is_enabled("vfio"))
 211                                 async_dma_map(dev, reg, false);
 212
 213                         munmap(reg->mmap_addr, reg->mmap_size);
 214                         close(reg->fd);
 215                 }
 216         }
 217 }
 218
 219 void
 220 vhost_backend_cleanup(struct virtio_net *dev)
 221 {
 222         if (dev->mem) {
 223                 free_mem_region(dev);
 224                 rte_free(dev->mem);
 225                 dev->mem = NULL;
 226         }
 227
 228         rte_free(dev->guest_pages);
 229         dev->guest_pages = NULL;
 230
 231         if (dev->log_addr) {
 232                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 233                 dev->log_addr = 0;
 234         }
 235
 236         if (dev->inflight_info) {
 237                 if (dev->inflight_info->addr) {
 238                         munmap(dev->inflight_info->addr,
 239                                dev->inflight_info->size);
 240                         dev->inflight_info->addr = NULL;
 241                 }
 242
 243                 if (dev->inflight_info->fd >= 0) {
 244                         close(dev->inflight_info->fd);
 245                         dev->inflight_info->fd = -1;
 246                 }
 247
 248                 rte_free(dev->inflight_info);
 249                 dev->inflight_info = NULL;
 250         }
 251
 252         if (dev->slave_req_fd >= 0) {
 253                 close(dev->slave_req_fd);
 254                 dev->slave_req_fd = -1;
 255         }
 256
 257         if (dev->postcopy_ufd >= 0) {
 258                 close(dev->postcopy_ufd);
 259                 dev->postcopy_ufd = -1;
 260         }
 261
 262         dev->postcopy_listening = 0;
 263 }
 264
 265 static void
 266 vhost_user_notify_queue_state(struct virtio_net *dev, uint16_t index,
 267                               int enable)
 268 {
 269         struct rte_vdpa_device *vdpa_dev = dev->vdpa_dev;
 270         struct vhost_virtqueue *vq = dev->virtqueue[index];
 271
 272         /* Configure guest notifications on enable */
 273         if (enable && vq->notif_enable != VIRTIO_UNINITIALIZED_NOTIF)
 274                 vhost_enable_guest_notification(dev, vq, vq->notif_enable);
 275
 276         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
 277                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
 278
 279         if (dev->notify_ops->vring_state_changed)
 280                 dev->notify_ops->vring_state_changed(dev->vid,
 281                                 index, enable);
 282 }
 283
 284 /*
 285  * This function just returns success at the moment unless
 286  * the device hasn't been initialised.
 287  */
 288 static int
 289 vhost_user_set_owner(struct virtio_net **pdev,
 290                         struct VhostUserMsg *msg,
 291                         int main_fd __rte_unused)
 292 {
 293         struct virtio_net *dev = *pdev;
 294
 295         if (validate_msg_fds(dev, msg, 0) != 0)
 296                 return RTE_VHOST_MSG_RESULT_ERR;
 297
 298         return RTE_VHOST_MSG_RESULT_OK;
 299 }
 300
 301 static int
 302 vhost_user_reset_owner(struct virtio_net **pdev,
 303                         struct VhostUserMsg *msg,
 304                         int main_fd __rte_unused)
 305 {
 306         struct virtio_net *dev = *pdev;
 307
 308         if (validate_msg_fds(dev, msg, 0) != 0)
 309                 return RTE_VHOST_MSG_RESULT_ERR;
 310
 311         vhost_destroy_device_notify(dev);
 312
 313         cleanup_device(dev, 0);
 314         reset_device(dev);
 315         return RTE_VHOST_MSG_RESULT_OK;
 316 }
 317
 318 /*
 319  * The features that we support are requested.
 320  */
 321 static int
 322 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 323                         int main_fd __rte_unused)
 324 {
 325         struct virtio_net *dev = *pdev;
 326         uint64_t features = 0;
 327
 328         if (validate_msg_fds(dev, msg, 0) != 0)
 329                 return RTE_VHOST_MSG_RESULT_ERR;
 330
 331         rte_vhost_driver_get_features(dev->ifname, &features);
 332
 333         msg->payload.u64 = features;
 334         msg->size = sizeof(msg->payload.u64);
 335         msg->fd_num = 0;
 336
 337         return RTE_VHOST_MSG_RESULT_REPLY;
 338 }
 339
 340 /*
 341  * The queue number that we support are requested.
 342  */
 343 static int
 344 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
 345                         int main_fd __rte_unused)
 346 {
 347         struct virtio_net *dev = *pdev;
 348         uint32_t queue_num = 0;
 349
 350         if (validate_msg_fds(dev, msg, 0) != 0)
 351                 return RTE_VHOST_MSG_RESULT_ERR;
 352
 353         rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
 354
 355         msg->payload.u64 = (uint64_t)queue_num;
 356         msg->size = sizeof(msg->payload.u64);
 357         msg->fd_num = 0;
 358
 359         return RTE_VHOST_MSG_RESULT_REPLY;
 360 }
 361
 362 /*
 363  * We receive the negotiated features supported by us and the virtio device.
 364  */
 365 static int
 366 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 367                         int main_fd __rte_unused)
 368 {
 369         struct virtio_net *dev = *pdev;
 370         uint64_t features = msg->payload.u64;
 371         uint64_t vhost_features = 0;
 372         struct rte_vdpa_device *vdpa_dev;
 373
 374         if (validate_msg_fds(dev, msg, 0) != 0)
 375                 return RTE_VHOST_MSG_RESULT_ERR;
 376
 377         rte_vhost_driver_get_features(dev->ifname, &vhost_features);
 378         if (features & ~vhost_features) {
 379                 VHOST_LOG_CONFIG(ERR, "(%s) received invalid negotiated features.\n",
 380                         dev->ifname);
 381                 dev->flags |= VIRTIO_DEV_FEATURES_FAILED;
 382                 dev->status &= ~VIRTIO_DEVICE_STATUS_FEATURES_OK;
 383
 384                 return RTE_VHOST_MSG_RESULT_ERR;
 385         }
 386
 387         if (dev->flags & VIRTIO_DEV_RUNNING) {
 388                 if (dev->features == features)
 389                         return RTE_VHOST_MSG_RESULT_OK;
 390
 391                 /*
 392                  * Error out if master tries to change features while device is
 393                  * in running state. The exception being VHOST_F_LOG_ALL, which
 394                  * is enabled when the live-migration starts.
 395                  */
 396                 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
 397                         VHOST_LOG_CONFIG(ERR, "(%s) features changed while device is running.\n",
 398                                 dev->ifname);
 399                         return RTE_VHOST_MSG_RESULT_ERR;
 400                 }
 401
 402                 if (dev->notify_ops->features_changed)
 403                         dev->notify_ops->features_changed(dev->vid, features);
 404         }
 405
 406         dev->features = features;
 407         if (dev->features &
 408                 ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
 409                  (1ULL << VIRTIO_F_VERSION_1) |
 410                  (1ULL << VIRTIO_F_RING_PACKED))) {
 411                 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 412         } else {
 413                 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
 414         }
 415         VHOST_LOG_CONFIG(INFO, "(%s) negotiated Virtio features: 0x%" PRIx64 "\n",
 416                         dev->ifname, dev->features);
 417         VHOST_LOG_CONFIG(DEBUG, "(%s) mergeable RX buffers %s, virtio 1 %s\n",
 418                 dev->ifname,
 419                 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
 420                 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 421
 422         if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
 423             !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
 424                 /*
 425                  * Remove all but first queue pair if MQ hasn't been
 426                  * negotiated. This is safe because the device is not
 427                  * running at this stage.
 428                  */
 429                 while (dev->nr_vring > 2) {
 430                         struct vhost_virtqueue *vq;
 431
 432                         vq = dev->virtqueue[--dev->nr_vring];
 433                         if (!vq)
 434                                 continue;
 435
 436                         dev->virtqueue[dev->nr_vring] = NULL;
 437                         cleanup_vq(vq, 1);
 438                         cleanup_vq_inflight(dev, vq);
 439                         free_vq(dev, vq);
 440                 }
 441         }
 442
 443         vdpa_dev = dev->vdpa_dev;
 444         if (vdpa_dev)
 445                 vdpa_dev->ops->set_features(dev->vid);
 446
 447         dev->flags &= ~VIRTIO_DEV_FEATURES_FAILED;
 448         return RTE_VHOST_MSG_RESULT_OK;
 449 }
 450
 451 /*
 452  * The virtio device sends us the size of the descriptor ring.
 453  */
 454 static int
 455 vhost_user_set_vring_num(struct virtio_net **pdev,
 456                         struct VhostUserMsg *msg,
 457                         int main_fd __rte_unused)
 458 {
 459         struct virtio_net *dev = *pdev;
 460         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 461
 462         if (validate_msg_fds(dev, msg, 0) != 0)
 463                 return RTE_VHOST_MSG_RESULT_ERR;
 464
 465         if (msg->payload.state.num > 32768) {
 466                 VHOST_LOG_CONFIG(ERR, "(%s) invalid virtqueue size %u\n",
 467                                 dev->ifname, msg->payload.state.num);
 468                 return RTE_VHOST_MSG_RESULT_ERR;
 469         }
 470
 471         vq->size = msg->payload.state.num;
 472
 473         /* VIRTIO 1.0, 2.4 Virtqueues says:
 474          *
 475          *   Queue Size value is always a power of 2. The maximum Queue Size
 476          *   value is 32768.
 477          *
 478          * VIRTIO 1.1 2.7 Virtqueues says:
 479          *
 480          *   Packed virtqueues support up to 2^15 entries each.
 481          */
 482         if (!vq_is_packed(dev)) {
 483                 if (vq->size & (vq->size - 1)) {
 484                         VHOST_LOG_CONFIG(ERR, "(%s) invalid virtqueue size %u\n",
 485                                         dev->ifname, vq->size);
 486                         return RTE_VHOST_MSG_RESULT_ERR;
 487                 }
 488         }
 489
 490         if (vq_is_packed(dev)) {
 491                 if (vq->shadow_used_packed)
 492                         rte_free(vq->shadow_used_packed);
 493                 vq->shadow_used_packed = rte_malloc_socket(NULL,
 494                                 vq->size *
 495                                 sizeof(struct vring_used_elem_packed),
 496                                 RTE_CACHE_LINE_SIZE, vq->numa_node);
 497                 if (!vq->shadow_used_packed) {
 498                         VHOST_LOG_CONFIG(ERR,
 499                                 "(%s) failed to allocate memory for shadow used ring.\n",
 500                                 dev->ifname);
 501                         return RTE_VHOST_MSG_RESULT_ERR;
 502                 }
 503
 504         } else {
 505                 if (vq->shadow_used_split)
 506                         rte_free(vq->shadow_used_split);
 507
 508                 vq->shadow_used_split = rte_malloc_socket(NULL,
 509                                 vq->size * sizeof(struct vring_used_elem),
 510                                 RTE_CACHE_LINE_SIZE, vq->numa_node);
 511
 512                 if (!vq->shadow_used_split) {
 513                         VHOST_LOG_CONFIG(ERR,
 514                                 "(%s) failed to allocate memory for vq internal data.\n",
 515                                 dev->ifname);
 516                         return RTE_VHOST_MSG_RESULT_ERR;
 517                 }
 518         }
 519
 520         if (vq->batch_copy_elems)
 521                 rte_free(vq->batch_copy_elems);
 522         vq->batch_copy_elems = rte_malloc_socket(NULL,
 523                                 vq->size * sizeof(struct batch_copy_elem),
 524                                 RTE_CACHE_LINE_SIZE, vq->numa_node);
 525         if (!vq->batch_copy_elems) {
 526                 VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate memory for batching copy.\n",
 527                         dev->ifname);
 528                 return RTE_VHOST_MSG_RESULT_ERR;
 529         }
 530
 531         return RTE_VHOST_MSG_RESULT_OK;
 532 }
 533
 534 /*
 535  * Reallocate virtio_dev, vhost_virtqueue and related data structures to
 536  * make them on the same numa node as the memory of vring descriptor.
 537  */
 538 #ifdef RTE_LIBRTE_VHOST_NUMA
 539 static struct virtio_net*
 540 numa_realloc(struct virtio_net *dev, int index)
 541 {
 542         int node, dev_node;
 543         struct virtio_net *old_dev;
 544         struct vhost_virtqueue *vq;
 545         struct batch_copy_elem *bce;
 546         struct guest_page *gp;
 547         struct rte_vhost_memory *mem;
 548         size_t mem_size;
 549         int ret;
 550
 551         old_dev = dev;
 552         vq = dev->virtqueue[index];
 553
 554         /*
 555          * If VQ is ready, it is too late to reallocate, it certainly already
 556          * happened anyway on VHOST_USER_SET_VRING_ADRR.
 557          */
 558         if (vq->ready)
 559                 return dev;
 560
 561         ret = get_mempolicy(&node, NULL, 0, vq->desc, MPOL_F_NODE | MPOL_F_ADDR);
 562         if (ret) {
 563                 VHOST_LOG_CONFIG(ERR, "(%s) unable to get virtqueue %d numa information.\n",
 564                                 dev->ifname, index);
 565                 return dev;
 566         }
 567
 568         if (node == vq->numa_node)
 569                 goto out_dev_realloc;
 570
 571         vq = rte_realloc_socket(vq, sizeof(*vq), 0, node);
 572         if (!vq) {
 573                 VHOST_LOG_CONFIG(ERR, "(%s) failed to realloc virtqueue %d on node %d\n",
 574                                 dev->ifname, index, node);
 575                 return dev;
 576         }
 577
 578         if (vq != dev->virtqueue[index]) {
 579                 VHOST_LOG_CONFIG(INFO, "(%s) reallocated virtqueue on node %d\n",
 580                                 dev->ifname, node);
 581                 dev->virtqueue[index] = vq;
 582                 vhost_user_iotlb_init(dev, index);
 583         }
 584
 585         if (vq_is_packed(dev)) {
 586                 struct vring_used_elem_packed *sup;
 587
 588                 sup = rte_realloc_socket(vq->shadow_used_packed, vq->size * sizeof(*sup),
 589                                 RTE_CACHE_LINE_SIZE, node);
 590                 if (!sup) {
 591                         VHOST_LOG_CONFIG(ERR, "(%s) failed to realloc shadow packed on node %d\n",
 592                                         dev->ifname, node);
 593                         return dev;
 594                 }
 595                 vq->shadow_used_packed = sup;
 596         } else {
 597                 struct vring_used_elem *sus;
 598
 599                 sus = rte_realloc_socket(vq->shadow_used_split, vq->size * sizeof(*sus),
 600                                 RTE_CACHE_LINE_SIZE, node);
 601                 if (!sus) {
 602                         VHOST_LOG_CONFIG(ERR, "(%s) failed to realloc shadow split on node %d\n",
 603                                         dev->ifname, node);
 604                         return dev;
 605                 }
 606                 vq->shadow_used_split = sus;
 607         }
 608
 609         bce = rte_realloc_socket(vq->batch_copy_elems, vq->size * sizeof(*bce),
 610                         RTE_CACHE_LINE_SIZE, node);
 611         if (!bce) {
 612                 VHOST_LOG_CONFIG(ERR, "(%s) failed to realloc batch copy elem on node %d\n",
 613                                 dev->ifname, node);
 614                 return dev;
 615         }
 616         vq->batch_copy_elems = bce;
 617
 618         if (vq->log_cache) {
 619                 struct log_cache_entry *lc;
 620
 621                 lc = rte_realloc_socket(vq->log_cache, sizeof(*lc) * VHOST_LOG_CACHE_NR, 0, node);
 622                 if (!lc) {
 623                         VHOST_LOG_CONFIG(ERR, "(%s) failed to realloc log cache on node %d\n",
 624                                         dev->ifname, node);
 625                         return dev;
 626                 }
 627                 vq->log_cache = lc;
 628         }
 629
 630         if (vq->resubmit_inflight) {
 631                 struct rte_vhost_resubmit_info *ri;
 632
 633                 ri = rte_realloc_socket(vq->resubmit_inflight, sizeof(*ri), 0, node);
 634                 if (!ri) {
 635                         VHOST_LOG_CONFIG(ERR, "(%s) failed to realloc resubmit inflight on node %d\n",
 636                                         dev->ifname, node);
 637                         return dev;
 638                 }
 639                 vq->resubmit_inflight = ri;
 640
 641                 if (ri->resubmit_list) {
 642                         struct rte_vhost_resubmit_desc *rd;
 643
 644                         rd = rte_realloc_socket(ri->resubmit_list, sizeof(*rd) * ri->resubmit_num,
 645                                         0, node);
 646                         if (!rd) {
 647                                 VHOST_LOG_CONFIG(ERR, "(%s) failed to realloc resubmit list on node %d\n",
 648                                                 dev->ifname, node);
 649                                 return dev;
 650                         }
 651                         ri->resubmit_list = rd;
 652                 }
 653         }
 654
 655         vq->numa_node = node;
 656
 657 out_dev_realloc:
 658
 659         if (dev->flags & VIRTIO_DEV_RUNNING)
 660                 return dev;
 661
 662         ret = get_mempolicy(&dev_node, NULL, 0, dev, MPOL_F_NODE | MPOL_F_ADDR);
 663         if (ret) {
 664                 VHOST_LOG_CONFIG(ERR, "(%s) unable to get numa information.\n", dev->ifname);
 665                 return dev;
 666         }
 667
 668         if (dev_node == node)
 669                 return dev;
 670
 671         dev = rte_realloc_socket(old_dev, sizeof(*dev), 0, node);
 672         if (!dev) {
 673                 VHOST_LOG_CONFIG(ERR, "(%s) failed to realloc dev on node %d\n",
 674                                 dev->ifname, node);
 675                 return old_dev;
 676         }
 677
 678         VHOST_LOG_CONFIG(INFO, "(%s) reallocated device on node %d\n", dev->ifname, node);
 679         vhost_devices[dev->vid] = dev;
 680
 681         mem_size = sizeof(struct rte_vhost_memory) +
 682                 sizeof(struct rte_vhost_mem_region) * dev->mem->nregions;
 683         mem = rte_realloc_socket(dev->mem, mem_size, 0, node);
 684         if (!mem) {
 685                 VHOST_LOG_CONFIG(ERR, "(%s) failed to realloc mem table on node %d\n",
 686                                 dev->ifname, node);
 687                 return dev;
 688         }
 689         dev->mem = mem;
 690
 691         gp = rte_realloc_socket(dev->guest_pages, dev->max_guest_pages * sizeof(*gp),
 692                         RTE_CACHE_LINE_SIZE, node);
 693         if (!gp) {
 694                 VHOST_LOG_CONFIG(ERR, "(%s) failed to realloc guest pages on node %d\n",
 695                                 dev->ifname, node);
 696                 return dev;
 697         }
 698         dev->guest_pages = gp;
 699
 700         return dev;
 701 }
 702 #else
 703 static struct virtio_net*
 704 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 705 {
 706         return dev;
 707 }
 708 #endif
 709
 710 /* Converts QEMU virtual address to Vhost virtual address. */
 711 static uint64_t
 712 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
 713 {
 714         struct rte_vhost_mem_region *r;
 715         uint32_t i;
 716
 717         if (unlikely(!dev || !dev->mem))
 718                 goto out_error;
 719
 720         /* Find the region where the address lives. */
 721         for (i = 0; i < dev->mem->nregions; i++) {
 722                 r = &dev->mem->regions[i];
 723
 724                 if (qva >= r->guest_user_addr &&
 725                     qva <  r->guest_user_addr + r->size) {
 726
 727                         if (unlikely(*len > r->guest_user_addr + r->size - qva))
 728                                 *len = r->guest_user_addr + r->size - qva;
 729
 730                         return qva - r->guest_user_addr +
 731                                r->host_user_addr;
 732                 }
 733         }
 734 out_error:
 735         *len = 0;
 736
 737         return 0;
 738 }
 739
 740
 741 /*
 742  * Converts ring address to Vhost virtual address.
 743  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 744  * else it is a QEMU virtual address.
 745  */
 746 static uint64_t
 747 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 748                 uint64_t ra, uint64_t *size)
 749 {
 750         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 751                 uint64_t vva;
 752
 753                 vhost_user_iotlb_rd_lock(vq);
 754                 vva = vhost_iova_to_vva(dev, vq, ra,
 755                                         size, VHOST_ACCESS_RW);
 756                 vhost_user_iotlb_rd_unlock(vq);
 757
 758                 return vva;
 759         }
 760
 761         return qva_to_vva(dev, ra, size);
 762 }
 763
 764 static uint64_t
 765 log_addr_to_gpa(struct virtio_net *dev, struct vhost_virtqueue *vq)
 766 {
 767         uint64_t log_gpa;
 768
 769         vhost_user_iotlb_rd_lock(vq);
 770         log_gpa = translate_log_addr(dev, vq, vq->ring_addrs.log_guest_addr);
 771         vhost_user_iotlb_rd_unlock(vq);
 772
 773         return log_gpa;
 774 }
 775
 776 static struct virtio_net *
 777 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 778 {
 779         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 780         struct vhost_vring_addr *addr = &vq->ring_addrs;
 781         uint64_t len, expected_len;
 782
 783         if (addr->flags & (1 << VHOST_VRING_F_LOG)) {
 784                 vq->log_guest_addr =
 785                         log_addr_to_gpa(dev, vq);
 786                 if (vq->log_guest_addr == 0) {
 787                         VHOST_LOG_CONFIG(DEBUG, "(%s) failed to map log_guest_addr.\n",
 788                                 dev->ifname);
 789                         return dev;
 790                 }
 791         }
 792
 793         if (vq_is_packed(dev)) {
 794                 len = sizeof(struct vring_packed_desc) * vq->size;
 795                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 796                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 797                 if (vq->desc_packed == NULL ||
 798                                 len != sizeof(struct vring_packed_desc) *
 799                                 vq->size) {
 800                         VHOST_LOG_CONFIG(DEBUG, "(%s) failed to map desc_packed ring.\n",
 801                                 dev->ifname);
 802                         return dev;
 803                 }
 804
 805                 dev = numa_realloc(dev, vq_index);
 806                 vq = dev->virtqueue[vq_index];
 807                 addr = &vq->ring_addrs;
 808
 809                 len = sizeof(struct vring_packed_desc_event);
 810                 vq->driver_event = (struct vring_packed_desc_event *)
 811                                         (uintptr_t)ring_addr_to_vva(dev,
 812                                         vq, addr->avail_user_addr, &len);
 813                 if (vq->driver_event == NULL ||
 814                                 len != sizeof(struct vring_packed_desc_event)) {
 815                         VHOST_LOG_CONFIG(DEBUG, "(%s) failed to find driver area address.\n",
 816                                 dev->ifname);
 817                         return dev;
 818                 }
 819
 820                 len = sizeof(struct vring_packed_desc_event);
 821                 vq->device_event = (struct vring_packed_desc_event *)
 822                                         (uintptr_t)ring_addr_to_vva(dev,
 823                                         vq, addr->used_user_addr, &len);
 824                 if (vq->device_event == NULL ||
 825                                 len != sizeof(struct vring_packed_desc_event)) {
 826                         VHOST_LOG_CONFIG(DEBUG, "(%s) failed to find device area address.\n",
 827                                 dev->ifname);
 828                         return dev;
 829                 }
 830
 831                 vq->access_ok = true;
 832                 return dev;
 833         }
 834
 835         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 836         if (vq->desc && vq->avail && vq->used)
 837                 return dev;
 838
 839         len = sizeof(struct vring_desc) * vq->size;
 840         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 841                         vq, addr->desc_user_addr, &len);
 842         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 843                 VHOST_LOG_CONFIG(DEBUG, "(%s) failed to map desc ring.\n", dev->ifname);
 844                 return dev;
 845         }
 846
 847         dev = numa_realloc(dev, vq_index);
 848         vq = dev->virtqueue[vq_index];
 849         addr = &vq->ring_addrs;
 850
 851         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 852         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 853                 len += sizeof(uint16_t);
 854         expected_len = len;
 855         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 856                         vq, addr->avail_user_addr, &len);
 857         if (vq->avail == 0 || len != expected_len) {
 858                 VHOST_LOG_CONFIG(DEBUG, "(%s) failed to map avail ring.\n", dev->ifname);
 859                 return dev;
 860         }
 861
 862         len = sizeof(struct vring_used) +
 863                 sizeof(struct vring_used_elem) * vq->size;
 864         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 865                 len += sizeof(uint16_t);
 866         expected_len = len;
 867         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 868                         vq, addr->used_user_addr, &len);
 869         if (vq->used == 0 || len != expected_len) {
 870                 VHOST_LOG_CONFIG(DEBUG, "(%s) failed to map used ring.\n", dev->ifname);
 871                 return dev;
 872         }
 873
 874         if (vq->last_used_idx != vq->used->idx) {
 875                 VHOST_LOG_CONFIG(WARNING,
 876                         "(%s) last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 877                         "some packets maybe resent for Tx and dropped for Rx\n",
 878                         dev->ifname,
 879                         vq->last_used_idx, vq->used->idx);
 880                 vq->last_used_idx  = vq->used->idx;
 881                 vq->last_avail_idx = vq->used->idx;
 882         }
 883
 884         vq->access_ok = true;
 885
 886         VHOST_LOG_CONFIG(DEBUG, "(%s) mapped address desc: %p\n", dev->ifname, vq->desc);
 887         VHOST_LOG_CONFIG(DEBUG, "(%s) mapped address avail: %p\n", dev->ifname, vq->avail);
 888         VHOST_LOG_CONFIG(DEBUG, "(%s) mapped address used: %p\n", dev->ifname, vq->used);
 889         VHOST_LOG_CONFIG(DEBUG, "(%s) log_guest_addr: %" PRIx64 "\n",
 890                         dev->ifname, vq->log_guest_addr);
 891
 892         return dev;
 893 }
 894
 895 /*
 896  * The virtio device sends us the desc, used and avail ring addresses.
 897  * This function then converts these to our address space.
 898  */
 899 static int
 900 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 901                         int main_fd __rte_unused)
 902 {
 903         struct virtio_net *dev = *pdev;
 904         struct vhost_virtqueue *vq;
 905         struct vhost_vring_addr *addr = &msg->payload.addr;
 906         bool access_ok;
 907
 908         if (validate_msg_fds(dev, msg, 0) != 0)
 909                 return RTE_VHOST_MSG_RESULT_ERR;
 910
 911         if (dev->mem == NULL)
 912                 return RTE_VHOST_MSG_RESULT_ERR;
 913
 914         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 915         vq = dev->virtqueue[msg->payload.addr.index];
 916
 917         access_ok = vq->access_ok;
 918
 919         /*
 920          * Rings addresses should not be interpreted as long as the ring is not
 921          * started and enabled
 922          */
 923         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 924
 925         vring_invalidate(dev, vq);
 926
 927         if ((vq->enabled && (dev->features &
 928                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) ||
 929                         access_ok) {
 930                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 931                 if (!dev)
 932                         return RTE_VHOST_MSG_RESULT_ERR;
 933
 934                 *pdev = dev;
 935         }
 936
 937         return RTE_VHOST_MSG_RESULT_OK;
 938 }
 939
 940 /*
 941  * The virtio device sends us the available ring last used index.
 942  */
 943 static int
 944 vhost_user_set_vring_base(struct virtio_net **pdev,
 945                         struct VhostUserMsg *msg,
 946                         int main_fd __rte_unused)
 947 {
 948         struct virtio_net *dev = *pdev;
 949         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 950         uint64_t val = msg->payload.state.num;
 951
 952         if (validate_msg_fds(dev, msg, 0) != 0)
 953                 return RTE_VHOST_MSG_RESULT_ERR;
 954
 955         if (vq_is_packed(dev)) {
 956                 /*
 957                  * Bit[0:14]: avail index
 958                  * Bit[15]: avail wrap counter
 959                  */
 960                 vq->last_avail_idx = val & 0x7fff;
 961                 vq->avail_wrap_counter = !!(val & (0x1 << 15));
 962                 /*
 963                  * Set used index to same value as available one, as
 964                  * their values should be the same since ring processing
 965                  * was stopped at get time.
 966                  */
 967                 vq->last_used_idx = vq->last_avail_idx;
 968                 vq->used_wrap_counter = vq->avail_wrap_counter;
 969         } else {
 970                 vq->last_used_idx = msg->payload.state.num;
 971                 vq->last_avail_idx = msg->payload.state.num;
 972         }
 973
 974         VHOST_LOG_CONFIG(INFO,
 975                 "(%s) vring base idx:%u last_used_idx:%u last_avail_idx:%u.\n",
 976                 dev->ifname, msg->payload.state.index, vq->last_used_idx,
 977                 vq->last_avail_idx);
 978
 979         return RTE_VHOST_MSG_RESULT_OK;
 980 }
 981
 982 static int
 983 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 984                    uint64_t host_phys_addr, uint64_t size)
 985 {
 986         struct guest_page *page, *last_page;
 987         struct guest_page *old_pages;
 988
 989         if (dev->nr_guest_pages == dev->max_guest_pages) {
 990                 dev->max_guest_pages *= 2;
 991                 old_pages = dev->guest_pages;
 992                 dev->guest_pages = rte_realloc(dev->guest_pages,
 993                                         dev->max_guest_pages * sizeof(*page),
 994                                         RTE_CACHE_LINE_SIZE);
 995                 if (dev->guest_pages == NULL) {
 996                         VHOST_LOG_CONFIG(ERR, "(%s) cannot realloc guest_pages\n", dev->ifname);
 997                         rte_free(old_pages);
 998                         return -1;
 999                 }
1000         }
1001
1002         if (dev->nr_guest_pages > 0) {
1003                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
1004                 /* merge if the two pages are continuous */
1005                 if (host_phys_addr == last_page->host_phys_addr +
1006                                       last_page->size) {
1007                         last_page->size += size;
1008                         return 0;
1009                 }
1010         }
1011
1012         page = &dev->guest_pages[dev->nr_guest_pages++];
1013         page->guest_phys_addr = guest_phys_addr;
1014         page->host_phys_addr  = host_phys_addr;
1015         page->size = size;
1016
1017         return 0;
1018 }
1019
1020 static int
1021 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
1022                 uint64_t page_size)
1023 {
1024         uint64_t reg_size = reg->size;
1025         uint64_t host_user_addr  = reg->host_user_addr;
1026         uint64_t guest_phys_addr = reg->guest_phys_addr;
1027         uint64_t host_phys_addr;
1028         uint64_t size;
1029
1030         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
1031         size = page_size - (guest_phys_addr & (page_size - 1));
1032         size = RTE_MIN(size, reg_size);
1033
1034         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
1035                 return -1;
1036
1037         host_user_addr  += size;
1038         guest_phys_addr += size;
1039         reg_size -= size;
1040
1041         while (reg_size > 0) {
1042                 size = RTE_MIN(reg_size, page_size);
1043                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
1044                                                   host_user_addr);
1045                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
1046                                 size) < 0)
1047                         return -1;
1048
1049                 host_user_addr  += size;
1050                 guest_phys_addr += size;
1051                 reg_size -= size;
1052         }
1053
1054         /* sort guest page array if over binary search threshold */
1055         if (dev->nr_guest_pages >= VHOST_BINARY_SEARCH_THRESH) {
1056                 qsort((void *)dev->guest_pages, dev->nr_guest_pages,
1057                         sizeof(struct guest_page), guest_page_addrcmp);
1058         }
1059
1060         return 0;
1061 }
1062
1063 #ifdef RTE_LIBRTE_VHOST_DEBUG
1064 /* TODO: enable it only in debug mode? */
1065 static void
1066 dump_guest_pages(struct virtio_net *dev)
1067 {
1068         uint32_t i;
1069         struct guest_page *page;
1070
1071         for (i = 0; i < dev->nr_guest_pages; i++) {
1072                 page = &dev->guest_pages[i];
1073
1074                 VHOST_LOG_CONFIG(INFO,
1075                         "(%s) guest physical page region %u\n"
1076                         "\t guest_phys_addr: %" PRIx64 "\n"
1077                         "\t host_phys_addr : %" PRIx64 "\n"
1078                         "\t size           : %" PRIx64 "\n",
1079                         dev->ifname, i,
1080                         page->guest_phys_addr,
1081                         page->host_phys_addr,
1082                         page->size);
1083         }
1084 }
1085 #else
1086 #define dump_guest_pages(dev)
1087 #endif
1088
1089 static bool
1090 vhost_memory_changed(struct VhostUserMemory *new,
1091                      struct rte_vhost_memory *old)
1092 {
1093         uint32_t i;
1094
1095         if (new->nregions != old->nregions)
1096                 return true;
1097
1098         for (i = 0; i < new->nregions; ++i) {
1099                 VhostUserMemoryRegion *new_r = &new->regions[i];
1100                 struct rte_vhost_mem_region *old_r = &old->regions[i];
1101
1102                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
1103                         return true;
1104                 if (new_r->memory_size != old_r->size)
1105                         return true;
1106                 if (new_r->userspace_addr != old_r->guest_user_addr)
1107                         return true;
1108         }
1109
1110         return false;
1111 }
1112
1113 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1114 static int
1115 vhost_user_postcopy_region_register(struct virtio_net *dev,
1116                 struct rte_vhost_mem_region *reg)
1117 {
1118         struct uffdio_register reg_struct;
1119
1120         /*
1121          * Let's register all the mmapped area to ensure
1122          * alignment on page boundary.
1123          */
1124         reg_struct.range.start = (uint64_t)(uintptr_t)reg->mmap_addr;
1125         reg_struct.range.len = reg->mmap_size;
1126         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1127
1128         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1129                                 &reg_struct)) {
1130                 VHOST_LOG_CONFIG(ERR, "(%s) failed to register ufd for region "
1131                                 "%" PRIx64 " - %" PRIx64 " (ufd = %d) %s\n",
1132                                 dev->ifname,
1133                                 (uint64_t)reg_struct.range.start,
1134                                 (uint64_t)reg_struct.range.start +
1135                                 (uint64_t)reg_struct.range.len - 1,
1136                                 dev->postcopy_ufd,
1137                                 strerror(errno));
1138                 return -1;
1139         }
1140
1141         VHOST_LOG_CONFIG(INFO, "(%s)\t userfaultfd registered for range : %" PRIx64 " - %" PRIx64 "\n",
1142                         dev->ifname,
1143                         (uint64_t)reg_struct.range.start,
1144                         (uint64_t)reg_struct.range.start +
1145                         (uint64_t)reg_struct.range.len - 1);
1146
1147         return 0;
1148 }
1149 #else
1150 static int
1151 vhost_user_postcopy_region_register(struct virtio_net *dev __rte_unused,
1152                 struct rte_vhost_mem_region *reg __rte_unused)
1153 {
1154         return -1;
1155 }
1156 #endif
1157
1158 static int
1159 vhost_user_postcopy_register(struct virtio_net *dev, int main_fd,
1160                 struct VhostUserMsg *msg)
1161 {
1162         struct VhostUserMemory *memory;
1163         struct rte_vhost_mem_region *reg;
1164         VhostUserMsg ack_msg;
1165         uint32_t i;
1166
1167         if (!dev->postcopy_listening)
1168                 return 0;
1169
1170         /*
1171          * We haven't a better way right now than sharing
1172          * DPDK's virtual address with Qemu, so that Qemu can
1173          * retrieve the region offset when handling userfaults.
1174          */
1175         memory = &msg->payload.memory;
1176         for (i = 0; i < memory->nregions; i++) {
1177                 reg = &dev->mem->regions[i];
1178                 memory->regions[i].userspace_addr = reg->host_user_addr;
1179         }
1180
1181         /* Send the addresses back to qemu */
1182         msg->fd_num = 0;
1183         send_vhost_reply(dev, main_fd, msg);
1184
1185         /* Wait for qemu to acknowledge it got the addresses
1186          * we've got to wait before we're allowed to generate faults.
1187          */
1188         if (read_vhost_message(dev, main_fd, &ack_msg) <= 0) {
1189                 VHOST_LOG_CONFIG(ERR, "(%s) failed to read qemu ack on postcopy set-mem-table\n",
1190                                 dev->ifname);
1191                 return -1;
1192         }
1193
1194         if (validate_msg_fds(dev, &ack_msg, 0) != 0)
1195                 return -1;
1196
1197         if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1198                 VHOST_LOG_CONFIG(ERR, "(%s) bad qemu ack on postcopy set-mem-table (%d)\n",
1199                                 dev->ifname, ack_msg.request.master);
1200                 return -1;
1201         }
1202
1203         /* Now userfault register and we can use the memory */
1204         for (i = 0; i < memory->nregions; i++) {
1205                 reg = &dev->mem->regions[i];
1206                 if (vhost_user_postcopy_region_register(dev, reg) < 0)
1207                         return -1;
1208         }
1209
1210         return 0;
1211 }
1212
1213 static int
1214 vhost_user_mmap_region(struct virtio_net *dev,
1215                 struct rte_vhost_mem_region *region,
1216                 uint64_t mmap_offset)
1217 {
1218         void *mmap_addr;
1219         uint64_t mmap_size;
1220         uint64_t alignment;
1221         int populate;
1222         int ret;
1223
1224         /* Check for memory_size + mmap_offset overflow */
1225         if (mmap_offset >= -region->size) {
1226                 VHOST_LOG_CONFIG(ERR, "(%s) mmap_offset (%#"PRIx64") and memory_size "
1227                                 "(%#"PRIx64") overflow\n",
1228                                 dev->ifname, mmap_offset, region->size);
1229                 return -1;
1230         }
1231
1232         mmap_size = region->size + mmap_offset;
1233
1234         /* mmap() without flag of MAP_ANONYMOUS, should be called with length
1235          * argument aligned with hugepagesz at older longterm version Linux,
1236          * like 2.6.32 and 3.2.72, or mmap() will fail with EINVAL.
1237          *
1238          * To avoid failure, make sure in caller to keep length aligned.
1239          */
1240         alignment = get_blk_size(region->fd);
1241         if (alignment == (uint64_t)-1) {
1242                 VHOST_LOG_CONFIG(ERR, "(%s) couldn't get hugepage size through fstat\n",
1243                                 dev->ifname);
1244                 return -1;
1245         }
1246         mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
1247         if (mmap_size == 0) {
1248                 /*
1249                  * It could happen if initial mmap_size + alignment overflows
1250                  * the sizeof uint64, which could happen if either mmap_size or
1251                  * alignment value is wrong.
1252                  *
1253                  * mmap() kernel implementation would return an error, but
1254                  * better catch it before and provide useful info in the logs.
1255                  */
1256                 VHOST_LOG_CONFIG(ERR, "(%s) mmap size (0x%" PRIx64 ") "
1257                                 "or alignment (0x%" PRIx64 ") is invalid\n",
1258                                 dev->ifname, region->size + mmap_offset, alignment);
1259                 return -1;
1260         }
1261
1262         populate = dev->async_copy ? MAP_POPULATE : 0;
1263         mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
1264                         MAP_SHARED | populate, region->fd, 0);
1265
1266         if (mmap_addr == MAP_FAILED) {
1267                 VHOST_LOG_CONFIG(ERR, "(%s) mmap failed (%s).\n", dev->ifname, strerror(errno));
1268                 return -1;
1269         }
1270
1271         region->mmap_addr = mmap_addr;
1272         region->mmap_size = mmap_size;
1273         region->host_user_addr = (uint64_t)(uintptr_t)mmap_addr + mmap_offset;
1274
1275         if (dev->async_copy) {
1276                 if (add_guest_pages(dev, region, alignment) < 0) {
1277                         VHOST_LOG_CONFIG(ERR, "(%s) adding guest pages to region failed.\n",
1278                                         dev->ifname);
1279                         return -1;
1280                 }
1281
1282                 if (rte_vfio_is_enabled("vfio")) {
1283                         ret = async_dma_map(dev, region, true);
1284                         if (ret) {
1285                                 VHOST_LOG_CONFIG(ERR,
1286                                         "(%s) configure IOMMU for DMA engine failed\n",
1287                                         dev->ifname);
1288                                 return -1;
1289                         }
1290                 }
1291         }
1292
1293         VHOST_LOG_CONFIG(INFO,
1294                         "(%s) guest memory region size: 0x%" PRIx64 "\n"
1295                         "\t guest physical addr: 0x%" PRIx64 "\n"
1296                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
1297                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
1298                         "\t mmap addr : 0x%" PRIx64 "\n"
1299                         "\t mmap size : 0x%" PRIx64 "\n"
1300                         "\t mmap align: 0x%" PRIx64 "\n"
1301                         "\t mmap off  : 0x%" PRIx64 "\n",
1302                         dev->ifname,
1303                         region->size,
1304                         region->guest_phys_addr,
1305                         region->guest_user_addr,
1306                         region->host_user_addr,
1307                         (uint64_t)(uintptr_t)mmap_addr,
1308                         mmap_size,
1309                         alignment,
1310                         mmap_offset);
1311
1312         return 0;
1313 }
1314
1315 static int
1316 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
1317                         int main_fd)
1318 {
1319         struct virtio_net *dev = *pdev;
1320         struct VhostUserMemory *memory = &msg->payload.memory;
1321         struct rte_vhost_mem_region *reg;
1322         int numa_node = SOCKET_ID_ANY;
1323         uint64_t mmap_offset;
1324         uint32_t i;
1325         bool async_notify = false;
1326
1327         if (validate_msg_fds(dev, msg, memory->nregions) != 0)
1328                 return RTE_VHOST_MSG_RESULT_ERR;
1329
1330         if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
1331                 VHOST_LOG_CONFIG(ERR, "(%s) too many memory regions (%u)\n",
1332                                 dev->ifname, memory->nregions);
1333                 goto close_msg_fds;
1334         }
1335
1336         if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
1337                 VHOST_LOG_CONFIG(INFO, "(%s) memory regions not changed\n", dev->ifname);
1338
1339                 close_msg_fds(msg);
1340
1341                 return RTE_VHOST_MSG_RESULT_OK;
1342         }
1343
1344         if (dev->mem) {
1345                 if (dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) {
1346                         struct rte_vdpa_device *vdpa_dev = dev->vdpa_dev;
1347
1348                         if (vdpa_dev && vdpa_dev->ops->dev_close)
1349                                 vdpa_dev->ops->dev_close(dev->vid);
1350                         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1351                 }
1352
1353                 /* notify the vhost application to stop DMA transfers */
1354                 if (dev->async_copy && dev->notify_ops->vring_state_changed) {
1355                         for (i = 0; i < dev->nr_vring; i++) {
1356                                 dev->notify_ops->vring_state_changed(dev->vid,
1357                                                 i, 0);
1358                         }
1359                         async_notify = true;
1360                 }
1361
1362                 free_mem_region(dev);
1363                 rte_free(dev->mem);
1364                 dev->mem = NULL;
1365         }
1366
1367         /* Flush IOTLB cache as previous HVAs are now invalid */
1368         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1369                 for (i = 0; i < dev->nr_vring; i++)
1370                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
1371
1372         /*
1373          * If VQ 0 has already been allocated, try to allocate on the same
1374          * NUMA node. It can be reallocated later in numa_realloc().
1375          */
1376         if (dev->nr_vring > 0)
1377                 numa_node = dev->virtqueue[0]->numa_node;
1378
1379         dev->nr_guest_pages = 0;
1380         if (dev->guest_pages == NULL) {
1381                 dev->max_guest_pages = 8;
1382                 dev->guest_pages = rte_zmalloc_socket(NULL,
1383                                         dev->max_guest_pages *
1384                                         sizeof(struct guest_page),
1385                                         RTE_CACHE_LINE_SIZE,
1386                                         numa_node);
1387                 if (dev->guest_pages == NULL) {
1388                         VHOST_LOG_CONFIG(ERR,
1389                                 "(%s) failed to allocate memory for dev->guest_pages\n",
1390                                 dev->ifname);
1391                         goto close_msg_fds;
1392                 }
1393         }
1394
1395         dev->mem = rte_zmalloc_socket("vhost-mem-table", sizeof(struct rte_vhost_memory) +
1396                 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0, numa_node);
1397         if (dev->mem == NULL) {
1398                 VHOST_LOG_CONFIG(ERR,
1399                         "(%s) failed to allocate memory for dev->mem\n",
1400                         dev->ifname);
1401                 goto free_guest_pages;
1402         }
1403
1404         for (i = 0; i < memory->nregions; i++) {
1405                 reg = &dev->mem->regions[i];
1406
1407                 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
1408                 reg->guest_user_addr = memory->regions[i].userspace_addr;
1409                 reg->size            = memory->regions[i].memory_size;
1410                 reg->fd              = msg->fds[i];
1411
1412                 /*
1413                  * Assign invalid file descriptor value to avoid double
1414                  * closing on error path.
1415                  */
1416                 msg->fds[i] = -1;
1417
1418                 mmap_offset = memory->regions[i].mmap_offset;
1419
1420                 if (vhost_user_mmap_region(dev, reg, mmap_offset) < 0) {
1421                         VHOST_LOG_CONFIG(ERR, "(%s) failed to mmap region %u\n", dev->ifname, i);
1422                         goto free_mem_table;
1423                 }
1424
1425                 dev->mem->nregions++;
1426         }
1427
1428         if (vhost_user_postcopy_register(dev, main_fd, msg) < 0)
1429                 goto free_mem_table;
1430
1431         for (i = 0; i < dev->nr_vring; i++) {
1432                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1433
1434                 if (!vq)
1435                         continue;
1436
1437                 if (vq->desc || vq->avail || vq->used) {
1438                         /*
1439                          * If the memory table got updated, the ring addresses
1440                          * need to be translated again as virtual addresses have
1441                          * changed.
1442                          */
1443                         vring_invalidate(dev, vq);
1444
1445                         dev = translate_ring_addresses(dev, i);
1446                         if (!dev) {
1447                                 dev = *pdev;
1448                                 goto free_mem_table;
1449                         }
1450
1451                         *pdev = dev;
1452                 }
1453         }
1454
1455         dump_guest_pages(dev);
1456
1457         if (async_notify) {
1458                 for (i = 0; i < dev->nr_vring; i++)
1459                         dev->notify_ops->vring_state_changed(dev->vid, i, 1);
1460         }
1461
1462         return RTE_VHOST_MSG_RESULT_OK;
1463
1464 free_mem_table:
1465         free_mem_region(dev);
1466         rte_free(dev->mem);
1467         dev->mem = NULL;
1468
1469 free_guest_pages:
1470         rte_free(dev->guest_pages);
1471         dev->guest_pages = NULL;
1472 close_msg_fds:
1473         close_msg_fds(msg);
1474         return RTE_VHOST_MSG_RESULT_ERR;
1475 }
1476
1477 static bool
1478 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1479 {
1480         bool rings_ok;
1481
1482         if (!vq)
1483                 return false;
1484
1485         if (vq_is_packed(dev))
1486                 rings_ok = vq->desc_packed && vq->driver_event &&
1487                         vq->device_event;
1488         else
1489                 rings_ok = vq->desc && vq->avail && vq->used;
1490
1491         return rings_ok &&
1492                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1493                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1494                vq->enabled;
1495 }
1496
1497 #define VIRTIO_BUILTIN_NUM_VQS_TO_BE_READY 2u
1498
1499 static int
1500 virtio_is_ready(struct virtio_net *dev)
1501 {
1502         struct vhost_virtqueue *vq;
1503         uint32_t i, nr_vring = dev->nr_vring;
1504
1505         if (dev->flags & VIRTIO_DEV_READY)
1506                 return 1;
1507
1508         if (!dev->nr_vring)
1509                 return 0;
1510
1511         if (dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) {
1512                 nr_vring = VIRTIO_BUILTIN_NUM_VQS_TO_BE_READY;
1513
1514                 if (dev->nr_vring < nr_vring)
1515                         return 0;
1516         }
1517
1518         for (i = 0; i < nr_vring; i++) {
1519                 vq = dev->virtqueue[i];
1520
1521                 if (!vq_is_ready(dev, vq))
1522                         return 0;
1523         }
1524
1525         /* If supported, ensure the frontend is really done with config */
1526         if (dev->protocol_features & (1ULL << VHOST_USER_PROTOCOL_F_STATUS))
1527                 if (!(dev->status & VIRTIO_DEVICE_STATUS_DRIVER_OK))
1528                         return 0;
1529
1530         dev->flags |= VIRTIO_DEV_READY;
1531
1532         if (!(dev->flags & VIRTIO_DEV_RUNNING))
1533                 VHOST_LOG_CONFIG(INFO, "(%s) virtio is now ready for processing.\n", dev->ifname);
1534         return 1;
1535 }
1536
1537 static void *
1538 inflight_mem_alloc(struct virtio_net *dev, const char *name, size_t size, int *fd)
1539 {
1540         void *ptr;
1541         int mfd = -1;
1542         char fname[20] = "/tmp/memfd-XXXXXX";
1543
1544         *fd = -1;
1545 #ifdef MEMFD_SUPPORTED
1546         mfd = memfd_create(name, MFD_CLOEXEC);
1547 #else
1548         RTE_SET_USED(name);
1549 #endif
1550         if (mfd == -1) {
1551                 mfd = mkstemp(fname);
1552                 if (mfd == -1) {
1553                         VHOST_LOG_CONFIG(ERR, "(%s) failed to get inflight buffer fd\n",
1554                                         dev->ifname);
1555                         return NULL;
1556                 }
1557
1558                 unlink(fname);
1559         }
1560
1561         if (ftruncate(mfd, size) == -1) {
1562                 VHOST_LOG_CONFIG(ERR, "(%s) failed to alloc inflight buffer\n", dev->ifname);
1563                 close(mfd);
1564                 return NULL;
1565         }
1566
1567         ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd, 0);
1568         if (ptr == MAP_FAILED) {
1569                 VHOST_LOG_CONFIG(ERR, "(%s) failed to mmap inflight buffer\n", dev->ifname);
1570                 close(mfd);
1571                 return NULL;
1572         }
1573
1574         *fd = mfd;
1575         return ptr;
1576 }
1577
1578 static uint32_t
1579 get_pervq_shm_size_split(uint16_t queue_size)
1580 {
1581         return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_split) *
1582                                   queue_size + sizeof(uint64_t) +
1583                                   sizeof(uint16_t) * 4, INFLIGHT_ALIGNMENT);
1584 }
1585
1586 static uint32_t
1587 get_pervq_shm_size_packed(uint16_t queue_size)
1588 {
1589         return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_packed)
1590                                   * queue_size + sizeof(uint64_t) +
1591                                   sizeof(uint16_t) * 6 + sizeof(uint8_t) * 9,
1592                                   INFLIGHT_ALIGNMENT);
1593 }
1594
1595 static int
1596 vhost_user_get_inflight_fd(struct virtio_net **pdev,
1597                            VhostUserMsg *msg,
1598                            int main_fd __rte_unused)
1599 {
1600         struct rte_vhost_inflight_info_packed *inflight_packed;
1601         uint64_t pervq_inflight_size, mmap_size;
1602         uint16_t num_queues, queue_size;
1603         struct virtio_net *dev = *pdev;
1604         int fd, i, j;
1605         int numa_node = SOCKET_ID_ANY;
1606         void *addr;
1607
1608         if (msg->size != sizeof(msg->payload.inflight)) {
1609                 VHOST_LOG_CONFIG(ERR, "(%s) invalid get_inflight_fd message size is %d\n",
1610                         dev->ifname, msg->size);
1611                 return RTE_VHOST_MSG_RESULT_ERR;
1612         }
1613
1614         /*
1615          * If VQ 0 has already been allocated, try to allocate on the same
1616          * NUMA node. It can be reallocated later in numa_realloc().
1617          */
1618         if (dev->nr_vring > 0)
1619                 numa_node = dev->virtqueue[0]->numa_node;
1620
1621         if (dev->inflight_info == NULL) {
1622                 dev->inflight_info = rte_zmalloc_socket("inflight_info",
1623                                 sizeof(struct inflight_mem_info), 0, numa_node);
1624                 if (!dev->inflight_info) {
1625                         VHOST_LOG_CONFIG(ERR, "(%s) failed to alloc dev inflight area\n",
1626                                         dev->ifname);
1627                         return RTE_VHOST_MSG_RESULT_ERR;
1628                 }
1629                 dev->inflight_info->fd = -1;
1630         }
1631
1632         num_queues = msg->payload.inflight.num_queues;
1633         queue_size = msg->payload.inflight.queue_size;
1634
1635         VHOST_LOG_CONFIG(INFO, "(%s) get_inflight_fd num_queues: %u\n",
1636                 dev->ifname, msg->payload.inflight.num_queues);
1637         VHOST_LOG_CONFIG(INFO, "(%s) get_inflight_fd queue_size: %u\n",
1638                 dev->ifname, msg->payload.inflight.queue_size);
1639
1640         if (vq_is_packed(dev))
1641                 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1642         else
1643                 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1644
1645         mmap_size = num_queues * pervq_inflight_size;
1646         addr = inflight_mem_alloc(dev, "vhost-inflight", mmap_size, &fd);
1647         if (!addr) {
1648                 VHOST_LOG_CONFIG(ERR, "(%s) failed to alloc vhost inflight area\n", dev->ifname);
1649                         msg->payload.inflight.mmap_size = 0;
1650                 return RTE_VHOST_MSG_RESULT_ERR;
1651         }
1652         memset(addr, 0, mmap_size);
1653
1654         if (dev->inflight_info->addr) {
1655                 munmap(dev->inflight_info->addr, dev->inflight_info->size);
1656                 dev->inflight_info->addr = NULL;
1657         }
1658
1659         if (dev->inflight_info->fd >= 0) {
1660                 close(dev->inflight_info->fd);
1661                 dev->inflight_info->fd = -1;
1662         }
1663
1664         dev->inflight_info->addr = addr;
1665         dev->inflight_info->size = msg->payload.inflight.mmap_size = mmap_size;
1666         dev->inflight_info->fd = msg->fds[0] = fd;
1667         msg->payload.inflight.mmap_offset = 0;
1668         msg->fd_num = 1;
1669
1670         if (vq_is_packed(dev)) {
1671                 for (i = 0; i < num_queues; i++) {
1672                         inflight_packed =
1673                                 (struct rte_vhost_inflight_info_packed *)addr;
1674                         inflight_packed->used_wrap_counter = 1;
1675                         inflight_packed->old_used_wrap_counter = 1;
1676                         for (j = 0; j < queue_size; j++)
1677                                 inflight_packed->desc[j].next = j + 1;
1678                         addr = (void *)((char *)addr + pervq_inflight_size);
1679                 }
1680         }
1681
1682         VHOST_LOG_CONFIG(INFO, "(%s) send inflight mmap_size: %"PRIu64"\n",
1683                         dev->ifname, msg->payload.inflight.mmap_size);
1684         VHOST_LOG_CONFIG(INFO, "(%s) send inflight mmap_offset: %"PRIu64"\n",
1685                         dev->ifname, msg->payload.inflight.mmap_offset);
1686         VHOST_LOG_CONFIG(INFO, "(%s) send inflight fd: %d\n", dev->ifname, msg->fds[0]);
1687
1688         return RTE_VHOST_MSG_RESULT_REPLY;
1689 }
1690
1691 static int
1692 vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
1693                            int main_fd __rte_unused)
1694 {
1695         uint64_t mmap_size, mmap_offset;
1696         uint16_t num_queues, queue_size;
1697         struct virtio_net *dev = *pdev;
1698         uint32_t pervq_inflight_size;
1699         struct vhost_virtqueue *vq;
1700         void *addr;
1701         int fd, i;
1702         int numa_node = SOCKET_ID_ANY;
1703
1704         fd = msg->fds[0];
1705         if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
1706                 VHOST_LOG_CONFIG(ERR, "(%s) invalid set_inflight_fd message size is %d,fd is %d\n",
1707                         dev->ifname, msg->size, fd);
1708                 return RTE_VHOST_MSG_RESULT_ERR;
1709         }
1710
1711         mmap_size = msg->payload.inflight.mmap_size;
1712         mmap_offset = msg->payload.inflight.mmap_offset;
1713         num_queues = msg->payload.inflight.num_queues;
1714         queue_size = msg->payload.inflight.queue_size;
1715
1716         if (vq_is_packed(dev))
1717                 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1718         else
1719                 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1720
1721         VHOST_LOG_CONFIG(INFO, "(%s) set_inflight_fd mmap_size: %"PRIu64"\n",
1722                         dev->ifname, mmap_size);
1723         VHOST_LOG_CONFIG(INFO, "(%s) set_inflight_fd mmap_offset: %"PRIu64"\n",
1724                         dev->ifname, mmap_offset);
1725         VHOST_LOG_CONFIG(INFO, "(%s) set_inflight_fd num_queues: %u\n", dev->ifname, num_queues);
1726         VHOST_LOG_CONFIG(INFO, "(%s) set_inflight_fd queue_size: %u\n", dev->ifname, queue_size);
1727         VHOST_LOG_CONFIG(INFO, "(%s) set_inflight_fd fd: %d\n", dev->ifname, fd);
1728         VHOST_LOG_CONFIG(INFO, "(%s) set_inflight_fd pervq_inflight_size: %d\n",
1729                         dev->ifname, pervq_inflight_size);
1730
1731         /*
1732          * If VQ 0 has already been allocated, try to allocate on the same
1733          * NUMA node. It can be reallocated later in numa_realloc().
1734          */
1735         if (dev->nr_vring > 0)
1736                 numa_node = dev->virtqueue[0]->numa_node;
1737
1738         if (!dev->inflight_info) {
1739                 dev->inflight_info = rte_zmalloc_socket("inflight_info",
1740                                 sizeof(struct inflight_mem_info), 0, numa_node);
1741                 if (dev->inflight_info == NULL) {
1742                         VHOST_LOG_CONFIG(ERR, "(%s) failed to alloc dev inflight area\n",
1743                                         dev->ifname);
1744                         return RTE_VHOST_MSG_RESULT_ERR;
1745                 }
1746                 dev->inflight_info->fd = -1;
1747         }
1748
1749         if (dev->inflight_info->addr) {
1750                 munmap(dev->inflight_info->addr, dev->inflight_info->size);
1751                 dev->inflight_info->addr = NULL;
1752         }
1753
1754         addr = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
1755                     fd, mmap_offset);
1756         if (addr == MAP_FAILED) {
1757                 VHOST_LOG_CONFIG(ERR, "(%s) failed to mmap share memory.\n", dev->ifname);
1758                 return RTE_VHOST_MSG_RESULT_ERR;
1759         }
1760
1761         if (dev->inflight_info->fd >= 0) {
1762                 close(dev->inflight_info->fd);
1763                 dev->inflight_info->fd = -1;
1764         }
1765
1766         dev->inflight_info->fd = fd;
1767         dev->inflight_info->addr = addr;
1768         dev->inflight_info->size = mmap_size;
1769
1770         for (i = 0; i < num_queues; i++) {
1771                 vq = dev->virtqueue[i];
1772                 if (!vq)
1773                         continue;
1774
1775                 if (vq_is_packed(dev)) {
1776                         vq->inflight_packed = addr;
1777                         vq->inflight_packed->desc_num = queue_size;
1778                 } else {
1779                         vq->inflight_split = addr;
1780                         vq->inflight_split->desc_num = queue_size;
1781                 }
1782                 addr = (void *)((char *)addr + pervq_inflight_size);
1783         }
1784
1785         return RTE_VHOST_MSG_RESULT_OK;
1786 }
1787
1788 static int
1789 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1790                         int main_fd __rte_unused)
1791 {
1792         struct virtio_net *dev = *pdev;
1793         struct vhost_vring_file file;
1794         struct vhost_virtqueue *vq;
1795         int expected_fds;
1796
1797         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1798         if (validate_msg_fds(dev, msg, expected_fds) != 0)
1799                 return RTE_VHOST_MSG_RESULT_ERR;
1800
1801         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1802         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1803                 file.fd = VIRTIO_INVALID_EVENTFD;
1804         else
1805                 file.fd = msg->fds[0];
1806         VHOST_LOG_CONFIG(INFO, "(%s) vring call idx:%d file:%d\n",
1807                         dev->ifname, file.index, file.fd);
1808
1809         vq = dev->virtqueue[file.index];
1810
1811         if (vq->ready) {
1812                 vq->ready = false;
1813                 vhost_user_notify_queue_state(dev, file.index, 0);
1814         }
1815
1816         if (vq->callfd >= 0)
1817                 close(vq->callfd);
1818
1819         vq->callfd = file.fd;
1820
1821         return RTE_VHOST_MSG_RESULT_OK;
1822 }
1823
1824 static int vhost_user_set_vring_err(struct virtio_net **pdev,
1825                         struct VhostUserMsg *msg,
1826                         int main_fd __rte_unused)
1827 {
1828         struct virtio_net *dev = *pdev;
1829         int expected_fds;
1830
1831         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1832         if (validate_msg_fds(dev, msg, expected_fds) != 0)
1833                 return RTE_VHOST_MSG_RESULT_ERR;
1834
1835         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1836                 close(msg->fds[0]);
1837         VHOST_LOG_CONFIG(INFO, "(%s) not implemented\n", dev->ifname);
1838
1839         return RTE_VHOST_MSG_RESULT_OK;
1840 }
1841
1842 static int
1843 resubmit_desc_compare(const void *a, const void *b)
1844 {
1845         const struct rte_vhost_resubmit_desc *desc0 = a;
1846         const struct rte_vhost_resubmit_desc *desc1 = b;
1847
1848         if (desc1->counter > desc0->counter)
1849                 return 1;
1850
1851         return -1;
1852 }
1853
1854 static int
1855 vhost_check_queue_inflights_split(struct virtio_net *dev,
1856                                   struct vhost_virtqueue *vq)
1857 {
1858         uint16_t i;
1859         uint16_t resubmit_num = 0, last_io, num;
1860         struct vring_used *used = vq->used;
1861         struct rte_vhost_resubmit_info *resubmit;
1862         struct rte_vhost_inflight_info_split *inflight_split;
1863
1864         if (!(dev->protocol_features &
1865             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1866                 return RTE_VHOST_MSG_RESULT_OK;
1867
1868         /* The frontend may still not support the inflight feature
1869          * although we negotiate the protocol feature.
1870          */
1871         if ((!vq->inflight_split))
1872                 return RTE_VHOST_MSG_RESULT_OK;
1873
1874         if (!vq->inflight_split->version) {
1875                 vq->inflight_split->version = INFLIGHT_VERSION;
1876                 return RTE_VHOST_MSG_RESULT_OK;
1877         }
1878
1879         if (vq->resubmit_inflight)
1880                 return RTE_VHOST_MSG_RESULT_OK;
1881
1882         inflight_split = vq->inflight_split;
1883         vq->global_counter = 0;
1884         last_io = inflight_split->last_inflight_io;
1885
1886         if (inflight_split->used_idx != used->idx) {
1887                 inflight_split->desc[last_io].inflight = 0;
1888                 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1889                 inflight_split->used_idx = used->idx;
1890         }
1891
1892         for (i = 0; i < inflight_split->desc_num; i++) {
1893                 if (inflight_split->desc[i].inflight == 1)
1894                         resubmit_num++;
1895         }
1896
1897         vq->last_avail_idx += resubmit_num;
1898
1899         if (resubmit_num) {
1900                 resubmit = rte_zmalloc_socket("resubmit", sizeof(struct rte_vhost_resubmit_info),
1901                                 0, vq->numa_node);
1902                 if (!resubmit) {
1903                         VHOST_LOG_CONFIG(ERR,
1904                                         "(%s) failed to allocate memory for resubmit info.\n",
1905                                         dev->ifname);
1906                         return RTE_VHOST_MSG_RESULT_ERR;
1907                 }
1908
1909                 resubmit->resubmit_list = rte_zmalloc_socket("resubmit_list",
1910                                 resubmit_num * sizeof(struct rte_vhost_resubmit_desc),
1911                                 0, vq->numa_node);
1912                 if (!resubmit->resubmit_list) {
1913                         VHOST_LOG_CONFIG(ERR,
1914                                         "(%s) failed to allocate memory for inflight desc.\n",
1915                                         dev->ifname);
1916                         rte_free(resubmit);
1917                         return RTE_VHOST_MSG_RESULT_ERR;
1918                 }
1919
1920                 num = 0;
1921                 for (i = 0; i < vq->inflight_split->desc_num; i++) {
1922                         if (vq->inflight_split->desc[i].inflight == 1) {
1923                                 resubmit->resubmit_list[num].index = i;
1924                                 resubmit->resubmit_list[num].counter =
1925                                         inflight_split->desc[i].counter;
1926                                 num++;
1927                         }
1928                 }
1929                 resubmit->resubmit_num = num;
1930
1931                 if (resubmit->resubmit_num > 1)
1932                         qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1933                               sizeof(struct rte_vhost_resubmit_desc),
1934                               resubmit_desc_compare);
1935
1936                 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1937                 vq->resubmit_inflight = resubmit;
1938         }
1939
1940         return RTE_VHOST_MSG_RESULT_OK;
1941 }
1942
1943 static int
1944 vhost_check_queue_inflights_packed(struct virtio_net *dev,
1945                                    struct vhost_virtqueue *vq)
1946 {
1947         uint16_t i;
1948         uint16_t resubmit_num = 0, old_used_idx, num;
1949         struct rte_vhost_resubmit_info *resubmit;
1950         struct rte_vhost_inflight_info_packed *inflight_packed;
1951
1952         if (!(dev->protocol_features &
1953             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1954                 return RTE_VHOST_MSG_RESULT_OK;
1955
1956         /* The frontend may still not support the inflight feature
1957          * although we negotiate the protocol feature.
1958          */
1959         if ((!vq->inflight_packed))
1960                 return RTE_VHOST_MSG_RESULT_OK;
1961
1962         if (!vq->inflight_packed->version) {
1963                 vq->inflight_packed->version = INFLIGHT_VERSION;
1964                 return RTE_VHOST_MSG_RESULT_OK;
1965         }
1966
1967         if (vq->resubmit_inflight)
1968                 return RTE_VHOST_MSG_RESULT_OK;
1969
1970         inflight_packed = vq->inflight_packed;
1971         vq->global_counter = 0;
1972         old_used_idx = inflight_packed->old_used_idx;
1973
1974         if (inflight_packed->used_idx != old_used_idx) {
1975                 if (inflight_packed->desc[old_used_idx].inflight == 0) {
1976                         inflight_packed->old_used_idx =
1977                                 inflight_packed->used_idx;
1978                         inflight_packed->old_used_wrap_counter =
1979                                 inflight_packed->used_wrap_counter;
1980                         inflight_packed->old_free_head =
1981                                 inflight_packed->free_head;
1982                 } else {
1983                         inflight_packed->used_idx =
1984                                 inflight_packed->old_used_idx;
1985                         inflight_packed->used_wrap_counter =
1986                                 inflight_packed->old_used_wrap_counter;
1987                         inflight_packed->free_head =
1988                                 inflight_packed->old_free_head;
1989                 }
1990         }
1991
1992         for (i = 0; i < inflight_packed->desc_num; i++) {
1993                 if (inflight_packed->desc[i].inflight == 1)
1994                         resubmit_num++;
1995         }
1996
1997         if (resubmit_num) {
1998                 resubmit = rte_zmalloc_socket("resubmit", sizeof(struct rte_vhost_resubmit_info),
1999                                 0, vq->numa_node);
2000                 if (resubmit == NULL) {
2001                         VHOST_LOG_CONFIG(ERR,
2002                                         "(%s) failed to allocate memory for resubmit info.\n",
2003                                         dev->ifname);
2004                         return RTE_VHOST_MSG_RESULT_ERR;
2005                 }
2006
2007                 resubmit->resubmit_list = rte_zmalloc_socket("resubmit_list",
2008                                 resubmit_num * sizeof(struct rte_vhost_resubmit_desc),
2009                                 0, vq->numa_node);
2010                 if (resubmit->resubmit_list == NULL) {
2011                         VHOST_LOG_CONFIG(ERR,
2012                                         "(%s) failed to allocate memory for resubmit desc.\n",
2013                                         dev->ifname);
2014                         rte_free(resubmit);
2015                         return RTE_VHOST_MSG_RESULT_ERR;
2016                 }
2017
2018                 num = 0;
2019                 for (i = 0; i < inflight_packed->desc_num; i++) {
2020                         if (vq->inflight_packed->desc[i].inflight == 1) {
2021                                 resubmit->resubmit_list[num].index = i;
2022                                 resubmit->resubmit_list[num].counter =
2023                                         inflight_packed->desc[i].counter;
2024                                 num++;
2025                         }
2026                 }
2027                 resubmit->resubmit_num = num;
2028
2029                 if (resubmit->resubmit_num > 1)
2030                         qsort(resubmit->resubmit_list, resubmit->resubmit_num,
2031                               sizeof(struct rte_vhost_resubmit_desc),
2032                               resubmit_desc_compare);
2033
2034                 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
2035                 vq->resubmit_inflight = resubmit;
2036         }
2037
2038         return RTE_VHOST_MSG_RESULT_OK;
2039 }
2040
2041 static int
2042 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
2043                         int main_fd __rte_unused)
2044 {
2045         struct virtio_net *dev = *pdev;
2046         struct vhost_vring_file file;
2047         struct vhost_virtqueue *vq;
2048         int expected_fds;
2049
2050         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
2051         if (validate_msg_fds(dev, msg, expected_fds) != 0)
2052                 return RTE_VHOST_MSG_RESULT_ERR;
2053
2054         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
2055         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
2056                 file.fd = VIRTIO_INVALID_EVENTFD;
2057         else
2058                 file.fd = msg->fds[0];
2059         VHOST_LOG_CONFIG(INFO, "(%s) vring kick idx:%d file:%d\n",
2060                         dev->ifname, file.index, file.fd);
2061
2062         /* Interpret ring addresses only when ring is started. */
2063         dev = translate_ring_addresses(dev, file.index);
2064         if (!dev) {
2065                 if (file.fd != VIRTIO_INVALID_EVENTFD)
2066                         close(file.fd);
2067
2068                 return RTE_VHOST_MSG_RESULT_ERR;
2069         }
2070
2071         *pdev = dev;
2072
2073         vq = dev->virtqueue[file.index];
2074
2075         /*
2076          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
2077          * the ring starts already enabled. Otherwise, it is enabled via
2078          * the SET_VRING_ENABLE message.
2079          */
2080         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
2081                 vq->enabled = true;
2082         }
2083
2084         if (vq->ready) {
2085                 vq->ready = false;
2086                 vhost_user_notify_queue_state(dev, file.index, 0);
2087         }
2088
2089         if (vq->kickfd >= 0)
2090                 close(vq->kickfd);
2091         vq->kickfd = file.fd;
2092
2093         if (vq_is_packed(dev)) {
2094                 if (vhost_check_queue_inflights_packed(dev, vq)) {
2095                         VHOST_LOG_CONFIG(ERR, "(%s) failed to inflights for vq: %d\n",
2096                                         dev->ifname, file.index);
2097                         return RTE_VHOST_MSG_RESULT_ERR;
2098                 }
2099         } else {
2100                 if (vhost_check_queue_inflights_split(dev, vq)) {
2101                         VHOST_LOG_CONFIG(ERR, "(%s) failed to inflights for vq: %d\n",
2102                                         dev->ifname, file.index);
2103                         return RTE_VHOST_MSG_RESULT_ERR;
2104                 }
2105         }
2106
2107         return RTE_VHOST_MSG_RESULT_OK;
2108 }
2109
2110 /*
2111  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
2112  */
2113 static int
2114 vhost_user_get_vring_base(struct virtio_net **pdev,
2115                         struct VhostUserMsg *msg,
2116                         int main_fd __rte_unused)
2117 {
2118         struct virtio_net *dev = *pdev;
2119         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
2120         uint64_t val;
2121
2122         if (validate_msg_fds(dev, msg, 0) != 0)
2123                 return RTE_VHOST_MSG_RESULT_ERR;
2124
2125         /* We have to stop the queue (virtio) if it is running. */
2126         vhost_destroy_device_notify(dev);
2127
2128         dev->flags &= ~VIRTIO_DEV_READY;
2129         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
2130
2131         /* Here we are safe to get the indexes */
2132         if (vq_is_packed(dev)) {
2133                 /*
2134                  * Bit[0:14]: avail index
2135                  * Bit[15]: avail wrap counter
2136                  */
2137                 val = vq->last_avail_idx & 0x7fff;
2138                 val |= vq->avail_wrap_counter << 15;
2139                 msg->payload.state.num = val;
2140         } else {
2141                 msg->payload.state.num = vq->last_avail_idx;
2142         }
2143
2144         VHOST_LOG_CONFIG(INFO, "(%s) vring base idx:%d file:%d\n",
2145                         dev->ifname, msg->payload.state.index,
2146                         msg->payload.state.num);
2147         /*
2148          * Based on current qemu vhost-user implementation, this message is
2149          * sent and only sent in vhost_vring_stop.
2150          * TODO: cleanup the vring, it isn't usable since here.
2151          */
2152         if (vq->kickfd >= 0)
2153                 close(vq->kickfd);
2154
2155         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
2156
2157         if (vq->callfd >= 0)
2158                 close(vq->callfd);
2159
2160         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
2161
2162         vq->signalled_used_valid = false;
2163
2164         if (vq_is_packed(dev)) {
2165                 rte_free(vq->shadow_used_packed);
2166                 vq->shadow_used_packed = NULL;
2167         } else {
2168                 rte_free(vq->shadow_used_split);
2169                 vq->shadow_used_split = NULL;
2170         }
2171
2172         rte_free(vq->batch_copy_elems);
2173         vq->batch_copy_elems = NULL;
2174
2175         rte_free(vq->log_cache);
2176         vq->log_cache = NULL;
2177
2178         msg->size = sizeof(msg->payload.state);
2179         msg->fd_num = 0;
2180
2181         vhost_user_iotlb_flush_all(vq);
2182
2183         vring_invalidate(dev, vq);
2184
2185         return RTE_VHOST_MSG_RESULT_REPLY;
2186 }
2187
2188 /*
2189  * when virtio queues are ready to work, qemu will send us to
2190  * enable the virtio queue pair.
2191  */
2192 static int
2193 vhost_user_set_vring_enable(struct virtio_net **pdev,
2194                         struct VhostUserMsg *msg,
2195                         int main_fd __rte_unused)
2196 {
2197         struct virtio_net *dev = *pdev;
2198         bool enable = !!msg->payload.state.num;
2199         int index = (int)msg->payload.state.index;
2200
2201         if (validate_msg_fds(dev, msg, 0) != 0)
2202                 return RTE_VHOST_MSG_RESULT_ERR;
2203
2204         VHOST_LOG_CONFIG(INFO, "(%s) set queue enable: %d to qp idx: %d\n",
2205                         dev->ifname, enable, index);
2206
2207         if (enable && dev->virtqueue[index]->async) {
2208                 if (dev->virtqueue[index]->async->pkts_inflight_n) {
2209                         VHOST_LOG_CONFIG(ERR, "(%s) failed to enable vring. "
2210                         "async inflight packets must be completed first\n",
2211                         dev->ifname);
2212                         return RTE_VHOST_MSG_RESULT_ERR;
2213                 }
2214         }
2215
2216         dev->virtqueue[index]->enabled = enable;
2217
2218         return RTE_VHOST_MSG_RESULT_OK;
2219 }
2220
2221 static int
2222 vhost_user_get_protocol_features(struct virtio_net **pdev,
2223                         struct VhostUserMsg *msg,
2224                         int main_fd __rte_unused)
2225 {
2226         struct virtio_net *dev = *pdev;
2227         uint64_t features, protocol_features;
2228
2229         if (validate_msg_fds(dev, msg, 0) != 0)
2230                 return RTE_VHOST_MSG_RESULT_ERR;
2231
2232         rte_vhost_driver_get_features(dev->ifname, &features);
2233         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
2234
2235         msg->payload.u64 = protocol_features;
2236         msg->size = sizeof(msg->payload.u64);
2237         msg->fd_num = 0;
2238
2239         return RTE_VHOST_MSG_RESULT_REPLY;
2240 }
2241
2242 static int
2243 vhost_user_set_protocol_features(struct virtio_net **pdev,
2244                         struct VhostUserMsg *msg,
2245                         int main_fd __rte_unused)
2246 {
2247         struct virtio_net *dev = *pdev;
2248         uint64_t protocol_features = msg->payload.u64;
2249         uint64_t slave_protocol_features = 0;
2250
2251         if (validate_msg_fds(dev, msg, 0) != 0)
2252                 return RTE_VHOST_MSG_RESULT_ERR;
2253
2254         rte_vhost_driver_get_protocol_features(dev->ifname,
2255                         &slave_protocol_features);
2256         if (protocol_features & ~slave_protocol_features) {
2257                 VHOST_LOG_CONFIG(ERR, "(%s) received invalid protocol features.\n", dev->ifname);
2258                 return RTE_VHOST_MSG_RESULT_ERR;
2259         }
2260
2261         dev->protocol_features = protocol_features;
2262         VHOST_LOG_CONFIG(INFO, "(%s) negotiated Vhost-user protocol features: 0x%" PRIx64 "\n",
2263                 dev->ifname, dev->protocol_features);
2264
2265         return RTE_VHOST_MSG_RESULT_OK;
2266 }
2267
2268 static int
2269 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
2270                         int main_fd __rte_unused)
2271 {
2272         struct virtio_net *dev = *pdev;
2273         int fd = msg->fds[0];
2274         uint64_t size, off;
2275         void *addr;
2276         uint32_t i;
2277
2278         if (validate_msg_fds(dev, msg, 1) != 0)
2279                 return RTE_VHOST_MSG_RESULT_ERR;
2280
2281         if (fd < 0) {
2282                 VHOST_LOG_CONFIG(ERR, "(%s) invalid log fd: %d\n", dev->ifname, fd);
2283                 return RTE_VHOST_MSG_RESULT_ERR;
2284         }
2285
2286         if (msg->size != sizeof(VhostUserLog)) {
2287                 VHOST_LOG_CONFIG(ERR, "(%s) invalid log base msg size: %"PRId32" != %d\n",
2288                         dev->ifname, msg->size, (int)sizeof(VhostUserLog));
2289                 goto close_msg_fds;
2290         }
2291
2292         size = msg->payload.log.mmap_size;
2293         off  = msg->payload.log.mmap_offset;
2294
2295         /* Check for mmap size and offset overflow. */
2296         if (off >= -size) {
2297                 VHOST_LOG_CONFIG(ERR,
2298                                 "(%s) log offset %#"PRIx64" and log size %#"PRIx64" overflow\n",
2299                                 dev->ifname, off, size);
2300                 goto close_msg_fds;
2301         }
2302
2303         VHOST_LOG_CONFIG(INFO, "(%s) log mmap size: %"PRId64", offset: %"PRId64"\n",
2304                         dev->ifname, size, off);
2305
2306         /*
2307          * mmap from 0 to workaround a hugepage mmap bug: mmap will
2308          * fail when offset is not page size aligned.
2309          */
2310         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
2311         close(fd);
2312         if (addr == MAP_FAILED) {
2313                 VHOST_LOG_CONFIG(ERR, "(%s) mmap log base failed!\n", dev->ifname);
2314                 return RTE_VHOST_MSG_RESULT_ERR;
2315         }
2316
2317         /*
2318          * Free previously mapped log memory on occasionally
2319          * multiple VHOST_USER_SET_LOG_BASE.
2320          */
2321         if (dev->log_addr) {
2322                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
2323         }
2324         dev->log_addr = (uint64_t)(uintptr_t)addr;
2325         dev->log_base = dev->log_addr + off;
2326         dev->log_size = size;
2327
2328         for (i = 0; i < dev->nr_vring; i++) {
2329                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2330
2331                 rte_free(vq->log_cache);
2332                 vq->log_cache = NULL;
2333                 vq->log_cache_nb_elem = 0;
2334                 vq->log_cache = rte_malloc_socket("vq log cache",
2335                                 sizeof(struct log_cache_entry) * VHOST_LOG_CACHE_NR,
2336                                 0, vq->numa_node);
2337                 /*
2338                  * If log cache alloc fail, don't fail migration, but no
2339                  * caching will be done, which will impact performance
2340                  */
2341                 if (!vq->log_cache)
2342                         VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate VQ logging cache\n",
2343                                         dev->ifname);
2344         }
2345
2346         /*
2347          * The spec is not clear about it (yet), but QEMU doesn't expect
2348          * any payload in the reply.
2349          */
2350         msg->size = 0;
2351         msg->fd_num = 0;
2352
2353         return RTE_VHOST_MSG_RESULT_REPLY;
2354
2355 close_msg_fds:
2356         close_msg_fds(msg);
2357         return RTE_VHOST_MSG_RESULT_ERR;
2358 }
2359
2360 static int vhost_user_set_log_fd(struct virtio_net **pdev,
2361                         struct VhostUserMsg *msg,
2362                         int main_fd __rte_unused)
2363 {
2364         struct virtio_net *dev = *pdev;
2365
2366         if (validate_msg_fds(dev, msg, 1) != 0)
2367                 return RTE_VHOST_MSG_RESULT_ERR;
2368
2369         close(msg->fds[0]);
2370         VHOST_LOG_CONFIG(INFO, "(%s) not implemented.\n", dev->ifname);
2371
2372         return RTE_VHOST_MSG_RESULT_OK;
2373 }
2374
2375 /*
2376  * An rarp packet is constructed and broadcasted to notify switches about
2377  * the new location of the migrated VM, so that packets from outside will
2378  * not be lost after migration.
2379  *
2380  * However, we don't actually "send" a rarp packet here, instead, we set
2381  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
2382  */
2383 static int
2384 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
2385                         int main_fd __rte_unused)
2386 {
2387         struct virtio_net *dev = *pdev;
2388         uint8_t *mac = (uint8_t *)&msg->payload.u64;
2389         struct rte_vdpa_device *vdpa_dev;
2390
2391         if (validate_msg_fds(dev, msg, 0) != 0)
2392                 return RTE_VHOST_MSG_RESULT_ERR;
2393
2394         VHOST_LOG_CONFIG(DEBUG, "(%s) MAC: " RTE_ETHER_ADDR_PRT_FMT "\n",
2395                 dev->ifname, mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
2396         memcpy(dev->mac.addr_bytes, mac, 6);
2397
2398         /*
2399          * Set the flag to inject a RARP broadcast packet at
2400          * rte_vhost_dequeue_burst().
2401          *
2402          * __ATOMIC_RELEASE ordering is for making sure the mac is
2403          * copied before the flag is set.
2404          */
2405         __atomic_store_n(&dev->broadcast_rarp, 1, __ATOMIC_RELEASE);
2406         vdpa_dev = dev->vdpa_dev;
2407         if (vdpa_dev && vdpa_dev->ops->migration_done)
2408                 vdpa_dev->ops->migration_done(dev->vid);
2409
2410         return RTE_VHOST_MSG_RESULT_OK;
2411 }
2412
2413 static int
2414 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
2415                         int main_fd __rte_unused)
2416 {
2417         struct virtio_net *dev = *pdev;
2418
2419         if (validate_msg_fds(dev, msg, 0) != 0)
2420                 return RTE_VHOST_MSG_RESULT_ERR;
2421
2422         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
2423                         msg->payload.u64 > VIRTIO_MAX_MTU) {
2424                 VHOST_LOG_CONFIG(ERR, "(%s) invalid MTU size (%"PRIu64")\n",
2425                                 dev->ifname, msg->payload.u64);
2426
2427                 return RTE_VHOST_MSG_RESULT_ERR;
2428         }
2429
2430         dev->mtu = msg->payload.u64;
2431
2432         return RTE_VHOST_MSG_RESULT_OK;
2433 }
2434
2435 static int
2436 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
2437                         int main_fd __rte_unused)
2438 {
2439         struct virtio_net *dev = *pdev;
2440         int fd = msg->fds[0];
2441
2442         if (validate_msg_fds(dev, msg, 1) != 0)
2443                 return RTE_VHOST_MSG_RESULT_ERR;
2444
2445         if (fd < 0) {
2446                 VHOST_LOG_CONFIG(ERR, "(%s) invalid file descriptor for slave channel (%d)\n",
2447                                 dev->ifname, fd);
2448                 return RTE_VHOST_MSG_RESULT_ERR;
2449         }
2450
2451         if (dev->slave_req_fd >= 0)
2452                 close(dev->slave_req_fd);
2453
2454         dev->slave_req_fd = fd;
2455
2456         return RTE_VHOST_MSG_RESULT_OK;
2457 }
2458
2459 static int
2460 is_vring_iotlb_split(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2461 {
2462         struct vhost_vring_addr *ra;
2463         uint64_t start, end, len;
2464
2465         start = imsg->iova;
2466         end = start + imsg->size;
2467
2468         ra = &vq->ring_addrs;
2469         len = sizeof(struct vring_desc) * vq->size;
2470         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2471                 return 1;
2472
2473         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
2474         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2475                 return 1;
2476
2477         len = sizeof(struct vring_used) +
2478                sizeof(struct vring_used_elem) * vq->size;
2479         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2480                 return 1;
2481
2482         if (ra->flags & (1 << VHOST_VRING_F_LOG)) {
2483                 len = sizeof(uint64_t);
2484                 if (ra->log_guest_addr < end &&
2485                     (ra->log_guest_addr + len) > start)
2486                         return 1;
2487         }
2488
2489         return 0;
2490 }
2491
2492 static int
2493 is_vring_iotlb_packed(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2494 {
2495         struct vhost_vring_addr *ra;
2496         uint64_t start, end, len;
2497
2498         start = imsg->iova;
2499         end = start + imsg->size;
2500
2501         ra = &vq->ring_addrs;
2502         len = sizeof(struct vring_packed_desc) * vq->size;
2503         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2504                 return 1;
2505
2506         len = sizeof(struct vring_packed_desc_event);
2507         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2508                 return 1;
2509
2510         len = sizeof(struct vring_packed_desc_event);
2511         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2512                 return 1;
2513
2514         if (ra->flags & (1 << VHOST_VRING_F_LOG)) {
2515                 len = sizeof(uint64_t);
2516                 if (ra->log_guest_addr < end &&
2517                     (ra->log_guest_addr + len) > start)
2518                         return 1;
2519         }
2520
2521         return 0;
2522 }
2523
2524 static int is_vring_iotlb(struct virtio_net *dev,
2525                           struct vhost_virtqueue *vq,
2526                           struct vhost_iotlb_msg *imsg)
2527 {
2528         if (vq_is_packed(dev))
2529                 return is_vring_iotlb_packed(vq, imsg);
2530         else
2531                 return is_vring_iotlb_split(vq, imsg);
2532 }
2533
2534 static int
2535 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
2536                         int main_fd __rte_unused)
2537 {
2538         struct virtio_net *dev = *pdev;
2539         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
2540         uint16_t i;
2541         uint64_t vva, len;
2542
2543         if (validate_msg_fds(dev, msg, 0) != 0)
2544                 return RTE_VHOST_MSG_RESULT_ERR;
2545
2546         switch (imsg->type) {
2547         case VHOST_IOTLB_UPDATE:
2548                 len = imsg->size;
2549                 vva = qva_to_vva(dev, imsg->uaddr, &len);
2550                 if (!vva)
2551                         return RTE_VHOST_MSG_RESULT_ERR;
2552
2553                 for (i = 0; i < dev->nr_vring; i++) {
2554                         struct vhost_virtqueue *vq = dev->virtqueue[i];
2555
2556                         if (!vq)
2557                                 continue;
2558
2559                         vhost_user_iotlb_cache_insert(dev, vq, imsg->iova, vva,
2560                                         len, imsg->perm);
2561
2562                         if (is_vring_iotlb(dev, vq, imsg))
2563                                 *pdev = dev = translate_ring_addresses(dev, i);
2564                 }
2565                 break;
2566         case VHOST_IOTLB_INVALIDATE:
2567                 for (i = 0; i < dev->nr_vring; i++) {
2568                         struct vhost_virtqueue *vq = dev->virtqueue[i];
2569
2570                         if (!vq)
2571                                 continue;
2572
2573                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
2574                                         imsg->size);
2575
2576                         if (is_vring_iotlb(dev, vq, imsg))
2577                                 vring_invalidate(dev, vq);
2578                 }
2579                 break;
2580         default:
2581                 VHOST_LOG_CONFIG(ERR, "(%s) invalid IOTLB message type (%d)\n",
2582                                 dev->ifname, imsg->type);
2583                 return RTE_VHOST_MSG_RESULT_ERR;
2584         }
2585
2586         return RTE_VHOST_MSG_RESULT_OK;
2587 }
2588
2589 static int
2590 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
2591                         struct VhostUserMsg *msg,
2592                         int main_fd __rte_unused)
2593 {
2594         struct virtio_net *dev = *pdev;
2595 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
2596         struct uffdio_api api_struct;
2597
2598         if (validate_msg_fds(dev, msg, 0) != 0)
2599                 return RTE_VHOST_MSG_RESULT_ERR;
2600
2601         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
2602
2603         if (dev->postcopy_ufd == -1) {
2604                 VHOST_LOG_CONFIG(ERR, "(%s) userfaultfd not available: %s\n",
2605                         dev->ifname, strerror(errno));
2606                 return RTE_VHOST_MSG_RESULT_ERR;
2607         }
2608         api_struct.api = UFFD_API;
2609         api_struct.features = 0;
2610         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
2611                 VHOST_LOG_CONFIG(ERR, "(%s) UFFDIO_API ioctl failure: %s\n",
2612                         dev->ifname, strerror(errno));
2613                 close(dev->postcopy_ufd);
2614                 dev->postcopy_ufd = -1;
2615                 return RTE_VHOST_MSG_RESULT_ERR;
2616         }
2617         msg->fds[0] = dev->postcopy_ufd;
2618         msg->fd_num = 1;
2619
2620         return RTE_VHOST_MSG_RESULT_REPLY;
2621 #else
2622         dev->postcopy_ufd = -1;
2623         msg->fd_num = 0;
2624
2625         return RTE_VHOST_MSG_RESULT_ERR;
2626 #endif
2627 }
2628
2629 static int
2630 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
2631                         struct VhostUserMsg *msg __rte_unused,
2632                         int main_fd __rte_unused)
2633 {
2634         struct virtio_net *dev = *pdev;
2635
2636         if (validate_msg_fds(dev, msg, 0) != 0)
2637                 return RTE_VHOST_MSG_RESULT_ERR;
2638
2639         if (dev->mem && dev->mem->nregions) {
2640                 VHOST_LOG_CONFIG(ERR, "(%s) regions already registered at postcopy-listen\n",
2641                                 dev->ifname);
2642                 return RTE_VHOST_MSG_RESULT_ERR;
2643         }
2644         dev->postcopy_listening = 1;
2645
2646         return RTE_VHOST_MSG_RESULT_OK;
2647 }
2648
2649 static int
2650 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
2651                         int main_fd __rte_unused)
2652 {
2653         struct virtio_net *dev = *pdev;
2654
2655         if (validate_msg_fds(dev, msg, 0) != 0)
2656                 return RTE_VHOST_MSG_RESULT_ERR;
2657
2658         dev->postcopy_listening = 0;
2659         if (dev->postcopy_ufd >= 0) {
2660                 close(dev->postcopy_ufd);
2661                 dev->postcopy_ufd = -1;
2662         }
2663
2664         msg->payload.u64 = 0;
2665         msg->size = sizeof(msg->payload.u64);
2666         msg->fd_num = 0;
2667
2668         return RTE_VHOST_MSG_RESULT_REPLY;
2669 }
2670
2671 static int
2672 vhost_user_get_status(struct virtio_net **pdev, struct VhostUserMsg *msg,
2673                       int main_fd __rte_unused)
2674 {
2675         struct virtio_net *dev = *pdev;
2676
2677         if (validate_msg_fds(dev, msg, 0) != 0)
2678                 return RTE_VHOST_MSG_RESULT_ERR;
2679
2680         msg->payload.u64 = dev->status;
2681         msg->size = sizeof(msg->payload.u64);
2682         msg->fd_num = 0;
2683
2684         return RTE_VHOST_MSG_RESULT_REPLY;
2685 }
2686
2687 static int
2688 vhost_user_set_status(struct virtio_net **pdev, struct VhostUserMsg *msg,
2689                         int main_fd __rte_unused)
2690 {
2691         struct virtio_net *dev = *pdev;
2692
2693         if (validate_msg_fds(dev, msg, 0) != 0)
2694                 return RTE_VHOST_MSG_RESULT_ERR;
2695
2696         /* As per Virtio specification, the device status is 8bits long */
2697         if (msg->payload.u64 > UINT8_MAX) {
2698                 VHOST_LOG_CONFIG(ERR, "(%s) invalid VHOST_USER_SET_STATUS payload 0x%" PRIx64 "\n",
2699                                 dev->ifname, msg->payload.u64);
2700                 return RTE_VHOST_MSG_RESULT_ERR;
2701         }
2702
2703         dev->status = msg->payload.u64;
2704
2705         if ((dev->status & VIRTIO_DEVICE_STATUS_FEATURES_OK) &&
2706             (dev->flags & VIRTIO_DEV_FEATURES_FAILED)) {
2707                 VHOST_LOG_CONFIG(ERR,
2708                                 "(%s) FEATURES_OK bit is set but feature negotiation failed\n",
2709                                 dev->ifname);
2710                 /*
2711                  * Clear the bit to let the driver know about the feature
2712                  * negotiation failure
2713                  */
2714                 dev->status &= ~VIRTIO_DEVICE_STATUS_FEATURES_OK;
2715         }
2716
2717         VHOST_LOG_CONFIG(INFO, "(%s) new device status(0x%08x):\n"
2718                         "\t-RESET: %u\n"
2719                         "\t-ACKNOWLEDGE: %u\n"
2720                         "\t-DRIVER: %u\n"
2721                         "\t-FEATURES_OK: %u\n"
2722                         "\t-DRIVER_OK: %u\n"
2723                         "\t-DEVICE_NEED_RESET: %u\n"
2724                         "\t-FAILED: %u\n",
2725                         dev->ifname,
2726                         dev->status,
2727                         (dev->status == VIRTIO_DEVICE_STATUS_RESET),
2728                         !!(dev->status & VIRTIO_DEVICE_STATUS_ACK),
2729                         !!(dev->status & VIRTIO_DEVICE_STATUS_DRIVER),
2730                         !!(dev->status & VIRTIO_DEVICE_STATUS_FEATURES_OK),
2731                         !!(dev->status & VIRTIO_DEVICE_STATUS_DRIVER_OK),
2732                         !!(dev->status & VIRTIO_DEVICE_STATUS_DEV_NEED_RESET),
2733                         !!(dev->status & VIRTIO_DEVICE_STATUS_FAILED));
2734
2735         return RTE_VHOST_MSG_RESULT_OK;
2736 }
2737
2738 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
2739                                         struct VhostUserMsg *msg,
2740                                         int main_fd);
2741 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
2742         [VHOST_USER_NONE] = NULL,
2743         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
2744         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
2745         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
2746         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
2747         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
2748         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
2749         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
2750         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
2751         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
2752         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
2753         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
2754         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
2755         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
2756         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
2757         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
2758         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
2759         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
2760         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
2761         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
2762         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
2763         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
2764         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
2765         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
2766         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
2767         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
2768         [VHOST_USER_GET_INFLIGHT_FD] = vhost_user_get_inflight_fd,
2769         [VHOST_USER_SET_INFLIGHT_FD] = vhost_user_set_inflight_fd,
2770         [VHOST_USER_SET_STATUS] = vhost_user_set_status,
2771         [VHOST_USER_GET_STATUS] = vhost_user_get_status,
2772 };
2773
2774 /* return bytes# of read on success or negative val on failure. */
2775 static int
2776 read_vhost_message(struct virtio_net *dev, int sockfd, struct VhostUserMsg *msg)
2777 {
2778         int ret;
2779
2780         ret = read_fd_message(dev->ifname, sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
2781                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
2782         if (ret <= 0) {
2783                 return ret;
2784         } else if (ret != VHOST_USER_HDR_SIZE) {
2785                 VHOST_LOG_CONFIG(ERR, "(%s) Unexpected header size read\n", dev->ifname);
2786                 close_msg_fds(msg);
2787                 return -1;
2788         }
2789
2790         if (msg->size) {
2791                 if (msg->size > sizeof(msg->payload)) {
2792                         VHOST_LOG_CONFIG(ERR, "(%s) invalid msg size: %d\n",
2793                                         dev->ifname, msg->size);
2794                         return -1;
2795                 }
2796                 ret = read(sockfd, &msg->payload, msg->size);
2797                 if (ret <= 0)
2798                         return ret;
2799                 if (ret != (int)msg->size) {
2800                         VHOST_LOG_CONFIG(ERR, "(%s) read control message failed\n", dev->ifname);
2801                         return -1;
2802                 }
2803         }
2804
2805         return ret;
2806 }
2807
2808 static int
2809 send_vhost_message(struct virtio_net *dev, int sockfd, struct VhostUserMsg *msg)
2810 {
2811         if (!msg)
2812                 return 0;
2813
2814         return send_fd_message(dev->ifname, sockfd, (char *)msg,
2815                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
2816 }
2817
2818 static int
2819 send_vhost_reply(struct virtio_net *dev, int sockfd, struct VhostUserMsg *msg)
2820 {
2821         if (!msg)
2822                 return 0;
2823
2824         msg->flags &= ~VHOST_USER_VERSION_MASK;
2825         msg->flags &= ~VHOST_USER_NEED_REPLY;
2826         msg->flags |= VHOST_USER_VERSION;
2827         msg->flags |= VHOST_USER_REPLY_MASK;
2828
2829         return send_vhost_message(dev, sockfd, msg);
2830 }
2831
2832 static int
2833 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
2834 {
2835         int ret;
2836
2837         if (msg->flags & VHOST_USER_NEED_REPLY)
2838                 rte_spinlock_lock(&dev->slave_req_lock);
2839
2840         ret = send_vhost_message(dev, dev->slave_req_fd, msg);
2841         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
2842                 rte_spinlock_unlock(&dev->slave_req_lock);
2843
2844         return ret;
2845 }
2846
2847 /*
2848  * Allocate a queue pair if it hasn't been allocated yet
2849  */
2850 static int
2851 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
2852                         struct VhostUserMsg *msg)
2853 {
2854         uint32_t vring_idx;
2855
2856         switch (msg->request.master) {
2857         case VHOST_USER_SET_VRING_KICK:
2858         case VHOST_USER_SET_VRING_CALL:
2859         case VHOST_USER_SET_VRING_ERR:
2860                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
2861                 break;
2862         case VHOST_USER_SET_VRING_NUM:
2863         case VHOST_USER_SET_VRING_BASE:
2864         case VHOST_USER_GET_VRING_BASE:
2865         case VHOST_USER_SET_VRING_ENABLE:
2866                 vring_idx = msg->payload.state.index;
2867                 break;
2868         case VHOST_USER_SET_VRING_ADDR:
2869                 vring_idx = msg->payload.addr.index;
2870                 break;
2871         default:
2872                 return 0;
2873         }
2874
2875         if (vring_idx >= VHOST_MAX_VRING) {
2876                 VHOST_LOG_CONFIG(ERR, "(%s) invalid vring index: %u\n", dev->ifname, vring_idx);
2877                 return -1;
2878         }
2879
2880         if (dev->virtqueue[vring_idx])
2881                 return 0;
2882
2883         return alloc_vring_queue(dev, vring_idx);
2884 }
2885
2886 static void
2887 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
2888 {
2889         unsigned int i = 0;
2890         unsigned int vq_num = 0;
2891
2892         while (vq_num < dev->nr_vring) {
2893                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2894
2895                 if (vq) {
2896                         rte_spinlock_lock(&vq->access_lock);
2897                         vq_num++;
2898                 }
2899                 i++;
2900         }
2901 }
2902
2903 static void
2904 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
2905 {
2906         unsigned int i = 0;
2907         unsigned int vq_num = 0;
2908
2909         while (vq_num < dev->nr_vring) {
2910                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2911
2912                 if (vq) {
2913                         rte_spinlock_unlock(&vq->access_lock);
2914                         vq_num++;
2915                 }
2916                 i++;
2917         }
2918 }
2919
2920 int
2921 vhost_user_msg_handler(int vid, int fd)
2922 {
2923         struct virtio_net *dev;
2924         struct VhostUserMsg msg;
2925         struct rte_vdpa_device *vdpa_dev;
2926         int ret;
2927         int unlock_required = 0;
2928         bool handled;
2929         int request;
2930         uint32_t i;
2931
2932         dev = get_device(vid);
2933         if (dev == NULL)
2934                 return -1;
2935
2936         if (!dev->notify_ops) {
2937                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
2938                 if (!dev->notify_ops) {
2939                         VHOST_LOG_CONFIG(ERR, "(%s) failed to get callback ops for driver\n",
2940                                 dev->ifname);
2941                         return -1;
2942                 }
2943         }
2944
2945         ret = read_vhost_message(dev, fd, &msg);
2946         if (ret <= 0) {
2947                 if (ret < 0)
2948                         VHOST_LOG_CONFIG(ERR, "(%s) vhost read message failed\n", dev->ifname);
2949                 else
2950                         VHOST_LOG_CONFIG(INFO, "(%s) vhost peer closed\n", dev->ifname);
2951
2952                 return -1;
2953         }
2954
2955         ret = 0;
2956         request = msg.request.master;
2957         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
2958                         vhost_message_str[request]) {
2959                 if (request != VHOST_USER_IOTLB_MSG)
2960                         VHOST_LOG_CONFIG(INFO, "(%s) read message %s\n",
2961                                 dev->ifname, vhost_message_str[request]);
2962                 else
2963                         VHOST_LOG_CONFIG(DEBUG, "(%s) read message %s\n",
2964                                 dev->ifname, vhost_message_str[request]);
2965         } else {
2966                 VHOST_LOG_CONFIG(DEBUG, "(%s) external request %d\n", dev->ifname, request);
2967         }
2968
2969         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
2970         if (ret < 0) {
2971                 VHOST_LOG_CONFIG(ERR, "(%s) failed to alloc queue\n", dev->ifname);
2972                 return -1;
2973         }
2974
2975         /*
2976          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
2977          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
2978          * and device is destroyed. destroy_device waits for queues to be
2979          * inactive, so it is safe. Otherwise taking the access_lock
2980          * would cause a dead lock.
2981          */
2982         switch (request) {
2983         case VHOST_USER_SET_FEATURES:
2984         case VHOST_USER_SET_PROTOCOL_FEATURES:
2985         case VHOST_USER_SET_OWNER:
2986         case VHOST_USER_SET_MEM_TABLE:
2987         case VHOST_USER_SET_LOG_BASE:
2988         case VHOST_USER_SET_LOG_FD:
2989         case VHOST_USER_SET_VRING_NUM:
2990         case VHOST_USER_SET_VRING_ADDR:
2991         case VHOST_USER_SET_VRING_BASE:
2992         case VHOST_USER_SET_VRING_KICK:
2993         case VHOST_USER_SET_VRING_CALL:
2994         case VHOST_USER_SET_VRING_ERR:
2995         case VHOST_USER_SET_VRING_ENABLE:
2996         case VHOST_USER_SEND_RARP:
2997         case VHOST_USER_NET_SET_MTU:
2998         case VHOST_USER_SET_SLAVE_REQ_FD:
2999                 if (!(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED)) {
3000                         vhost_user_lock_all_queue_pairs(dev);
3001                         unlock_required = 1;
3002                 }
3003                 break;
3004         default:
3005                 break;
3006
3007         }
3008
3009         handled = false;
3010         if (dev->extern_ops.pre_msg_handle) {
3011                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
3012                                 (void *)&msg);
3013                 switch (ret) {
3014                 case RTE_VHOST_MSG_RESULT_REPLY:
3015                         send_vhost_reply(dev, fd, &msg);
3016                         /* Fall-through */
3017                 case RTE_VHOST_MSG_RESULT_ERR:
3018                 case RTE_VHOST_MSG_RESULT_OK:
3019                         handled = true;
3020                         goto skip_to_post_handle;
3021                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
3022                 default:
3023                         break;
3024                 }
3025         }
3026
3027         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
3028                 if (!vhost_message_handlers[request])
3029                         goto skip_to_post_handle;
3030                 ret = vhost_message_handlers[request](&dev, &msg, fd);
3031
3032                 switch (ret) {
3033                 case RTE_VHOST_MSG_RESULT_ERR:
3034                         VHOST_LOG_CONFIG(ERR, "(%s) processing %s failed.\n",
3035                                         dev->ifname, vhost_message_str[request]);
3036                         handled = true;
3037                         break;
3038                 case RTE_VHOST_MSG_RESULT_OK:
3039                         VHOST_LOG_CONFIG(DEBUG, "(%s) processing %s succeeded.\n",
3040                                         dev->ifname, vhost_message_str[request]);
3041                         handled = true;
3042                         break;
3043                 case RTE_VHOST_MSG_RESULT_REPLY:
3044                         VHOST_LOG_CONFIG(DEBUG, "(%s) processing %s succeeded and needs reply.\n",
3045                                         dev->ifname, vhost_message_str[request]);
3046                         send_vhost_reply(dev, fd, &msg);
3047                         handled = true;
3048                         break;
3049                 default:
3050                         break;
3051                 }
3052         }
3053
3054 skip_to_post_handle:
3055         if (ret != RTE_VHOST_MSG_RESULT_ERR &&
3056                         dev->extern_ops.post_msg_handle) {
3057                 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
3058                                 (void *)&msg);
3059                 switch (ret) {
3060                 case RTE_VHOST_MSG_RESULT_REPLY:
3061                         send_vhost_reply(dev, fd, &msg);
3062                         /* Fall-through */
3063                 case RTE_VHOST_MSG_RESULT_ERR:
3064                 case RTE_VHOST_MSG_RESULT_OK:
3065                         handled = true;
3066                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
3067                 default:
3068                         break;
3069                 }
3070         }
3071
3072         /* If message was not handled at this stage, treat it as an error */
3073         if (!handled) {
3074                 VHOST_LOG_CONFIG(ERR, "(%s) vhost message (req: %d) was not handled.\n",
3075                                 dev->ifname, request);
3076                 close_msg_fds(&msg);
3077                 ret = RTE_VHOST_MSG_RESULT_ERR;
3078         }
3079
3080         /*
3081          * If the request required a reply that was already sent,
3082          * this optional reply-ack won't be sent as the
3083          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
3084          */
3085         if (msg.flags & VHOST_USER_NEED_REPLY) {
3086                 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
3087                 msg.size = sizeof(msg.payload.u64);
3088                 msg.fd_num = 0;
3089                 send_vhost_reply(dev, fd, &msg);
3090         } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
3091                 VHOST_LOG_CONFIG(ERR, "(%s) vhost message handling failed.\n", dev->ifname);
3092                 return -1;
3093         }
3094
3095         for (i = 0; i < dev->nr_vring; i++) {
3096                 struct vhost_virtqueue *vq = dev->virtqueue[i];
3097                 bool cur_ready = vq_is_ready(dev, vq);
3098
3099                 if (cur_ready != (vq && vq->ready)) {
3100                         vq->ready = cur_ready;
3101                         vhost_user_notify_queue_state(dev, i, cur_ready);
3102                 }
3103         }
3104
3105         if (unlock_required)
3106                 vhost_user_unlock_all_queue_pairs(dev);
3107
3108         if (!virtio_is_ready(dev))
3109                 goto out;
3110
3111         /*
3112          * Virtio is now ready. If not done already, it is time
3113          * to notify the application it can process the rings and
3114          * configure the vDPA device if present.
3115          */
3116
3117         if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
3118                 if (dev->notify_ops->new_device(dev->vid) == 0)
3119                         dev->flags |= VIRTIO_DEV_RUNNING;
3120         }
3121
3122         vdpa_dev = dev->vdpa_dev;
3123         if (!vdpa_dev)
3124                 goto out;
3125
3126         if (!(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED)) {
3127                 if (vdpa_dev->ops->dev_conf(dev->vid))
3128                         VHOST_LOG_CONFIG(ERR, "(%s) failed to configure vDPA device\n",
3129                                         dev->ifname);
3130                 else
3131                         dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
3132         }
3133
3134 out:
3135         return 0;
3136 }
3137
3138 static int process_slave_message_reply(struct virtio_net *dev,
3139                                        const struct VhostUserMsg *msg)
3140 {
3141         struct VhostUserMsg msg_reply;
3142         int ret;
3143
3144         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
3145                 return 0;
3146
3147         ret = read_vhost_message(dev, dev->slave_req_fd, &msg_reply);
3148         if (ret <= 0) {
3149                 if (ret < 0)
3150                         VHOST_LOG_CONFIG(ERR, "(%s) vhost read slave message reply failed\n",
3151                                         dev->ifname);
3152                 else
3153                         VHOST_LOG_CONFIG(INFO, "(%s) vhost peer closed\n", dev->ifname);
3154                 ret = -1;
3155                 goto out;
3156         }
3157
3158         ret = 0;
3159         if (msg_reply.request.slave != msg->request.slave) {
3160                 VHOST_LOG_CONFIG(ERR, "(%s) received unexpected msg type (%u), expected %u\n",
3161                                 dev->ifname, msg_reply.request.slave, msg->request.slave);
3162                 ret = -1;
3163                 goto out;
3164         }
3165
3166         ret = msg_reply.payload.u64 ? -1 : 0;
3167
3168 out:
3169         rte_spinlock_unlock(&dev->slave_req_lock);
3170         return ret;
3171 }
3172
3173 int
3174 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
3175 {
3176         int ret;
3177         struct VhostUserMsg msg = {
3178                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
3179                 .flags = VHOST_USER_VERSION,
3180                 .size = sizeof(msg.payload.iotlb),
3181                 .payload.iotlb = {
3182                         .iova = iova,
3183                         .perm = perm,
3184                         .type = VHOST_IOTLB_MISS,
3185                 },
3186         };
3187
3188         ret = send_vhost_message(dev, dev->slave_req_fd, &msg);
3189         if (ret < 0) {
3190                 VHOST_LOG_CONFIG(ERR, "(%s) failed to send IOTLB miss message (%d)\n",
3191                                 dev->ifname, ret);
3192                 return ret;
3193         }
3194
3195         return 0;
3196 }
3197
3198 static int
3199 vhost_user_slave_config_change(struct virtio_net *dev, bool need_reply)
3200 {
3201         int ret;
3202         struct VhostUserMsg msg = {
3203                 .request.slave = VHOST_USER_SLAVE_CONFIG_CHANGE_MSG,
3204                 .flags = VHOST_USER_VERSION,
3205                 .size = 0,
3206         };
3207
3208         if (need_reply)
3209                 msg.flags |= VHOST_USER_NEED_REPLY;
3210
3211         ret = send_vhost_slave_message(dev, &msg);
3212         if (ret < 0) {
3213                 VHOST_LOG_CONFIG(ERR, "(%s) failed to send config change (%d)\n",
3214                                 dev->ifname, ret);
3215                 return ret;
3216         }
3217
3218         return process_slave_message_reply(dev, &msg);
3219 }
3220
3221 int
3222 rte_vhost_slave_config_change(int vid, bool need_reply)
3223 {
3224         struct virtio_net *dev;
3225
3226         dev = get_device(vid);
3227         if (!dev)
3228                 return -ENODEV;
3229
3230         return vhost_user_slave_config_change(dev, need_reply);
3231 }
3232
3233 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
3234                                                     int index, int fd,
3235                                                     uint64_t offset,
3236                                                     uint64_t size)
3237 {
3238         int ret;
3239         struct VhostUserMsg msg = {
3240                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
3241                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
3242                 .size = sizeof(msg.payload.area),
3243                 .payload.area = {
3244                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
3245                         .size = size,
3246                         .offset = offset,
3247                 },
3248         };
3249
3250         if (fd < 0)
3251                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
3252         else {
3253                 msg.fds[0] = fd;
3254                 msg.fd_num = 1;
3255         }
3256
3257         ret = send_vhost_slave_message(dev, &msg);
3258         if (ret < 0) {
3259                 VHOST_LOG_CONFIG(ERR, "(%s) failed to set host notifier (%d)\n",
3260                                 dev->ifname, ret);
3261                 return ret;
3262         }
3263
3264         return process_slave_message_reply(dev, &msg);
3265 }
3266
3267 int rte_vhost_host_notifier_ctrl(int vid, uint16_t qid, bool enable)
3268 {
3269         struct virtio_net *dev;
3270         struct rte_vdpa_device *vdpa_dev;
3271         int vfio_device_fd, ret = 0;
3272         uint64_t offset, size;
3273         unsigned int i, q_start, q_last;
3274
3275         dev = get_device(vid);
3276         if (!dev)
3277                 return -ENODEV;
3278
3279         vdpa_dev = dev->vdpa_dev;
3280         if (vdpa_dev == NULL)
3281                 return -ENODEV;
3282
3283         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
3284             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
3285             !(dev->protocol_features &
3286                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
3287             !(dev->protocol_features &
3288                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
3289             !(dev->protocol_features &
3290                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
3291                 return -ENOTSUP;
3292
3293         if (qid == RTE_VHOST_QUEUE_ALL) {
3294                 q_start = 0;
3295                 q_last = dev->nr_vring - 1;
3296         } else {
3297                 if (qid >= dev->nr_vring)
3298                         return -EINVAL;
3299                 q_start = qid;
3300                 q_last = qid;
3301         }
3302
3303         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
3304         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
3305
3306         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
3307         if (vfio_device_fd < 0)
3308                 return -ENOTSUP;
3309
3310         if (enable) {
3311                 for (i = q_start; i <= q_last; i++) {
3312                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
3313                                         &size) < 0) {
3314                                 ret = -ENOTSUP;
3315                                 goto disable;
3316                         }
3317
3318                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
3319                                         vfio_device_fd, offset, size) < 0) {
3320                                 ret = -EFAULT;
3321                                 goto disable;
3322                         }
3323                 }
3324         } else {
3325 disable:
3326                 for (i = q_start; i <= q_last; i++) {
3327                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
3328                                         0, 0);
3329                 }
3330         }
3331
3332         return ret;
3333 }