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(int sockfd, struct VhostUserMsg *msg);
  97 static int read_vhost_message(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 VhostUserMsg *msg, int expected_fds)
 121 {
 122         if (msg->fd_num == expected_fds)
 123                 return 0;
 124
 125         VHOST_LOG_CONFIG(ERR,
 126                 " Expect %d FDs for request %s, received %d\n",
 127                 expected_fds,
 128                 vhost_message_str[msg->request.master],
 129                 msg->fd_num);
 130
 131         close_msg_fds(msg);
 132
 133         return -1;
 134 }
 135
 136 static uint64_t
 137 get_blk_size(int fd)
 138 {
 139         struct stat stat;
 140         int ret;
 141
 142         ret = fstat(fd, &stat);
 143         return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
 144 }
 145
 146 static int
 147 async_dma_map(struct rte_vhost_mem_region *region, bool *dma_map_success, bool do_map)
 148 {
 149         uint64_t host_iova;
 150         int ret = 0;
 151
 152         host_iova = rte_mem_virt2iova((void *)(uintptr_t)region->host_user_addr);
 153         if (do_map) {
 154                 /* Add mapped region into the default container of DPDK. */
 155                 ret = rte_vfio_container_dma_map(RTE_VFIO_DEFAULT_CONTAINER_FD,
 156                                                  region->host_user_addr,
 157                                                  host_iova,
 158                                                  region->size);
 159                 *dma_map_success = ret == 0;
 160
 161                 if (ret) {
 162                         /*
 163                          * DMA device may bind with kernel driver, in this case,
 164                          * we don't need to program IOMMU manually. However, if no
 165                          * device is bound with vfio/uio in DPDK, and vfio kernel
 166                          * module is loaded, the API will still be called and return
 167                          * with ENODEV/ENOSUP.
 168                          *
 169                          * DPDK vfio only returns ENODEV/ENOSUP in very similar
 170                          * situations(vfio either unsupported, or supported
 171                          * but no devices found). Either way, no mappings could be
 172                          * performed. We treat it as normal case in async path.
 173                          */
 174                         if (rte_errno == ENODEV || rte_errno == ENOTSUP)
 175                                 return 0;
 176
 177                         VHOST_LOG_CONFIG(ERR, "DMA engine map failed\n");
 178                         return ret;
 179
 180                 }
 181
 182         } else {
 183                 /* No need to do vfio unmap if the map failed. */
 184                 if (!*dma_map_success)
 185                         return 0;
 186
 187                 /* Remove mapped region from the default container of DPDK. */
 188                 ret = rte_vfio_container_dma_unmap(RTE_VFIO_DEFAULT_CONTAINER_FD,
 189                                                    region->host_user_addr,
 190                                                    host_iova,
 191                                                    region->size);
 192                 if (ret) {
 193                         VHOST_LOG_CONFIG(ERR, "DMA engine unmap failed\n");
 194                         return ret;
 195                 }
 196                 /* Clear the flag once the unmap succeeds. */
 197                 *dma_map_success = 0;
 198         }
 199
 200         return ret;
 201 }
 202
 203 static void
 204 free_mem_region(struct virtio_net *dev)
 205 {
 206         uint32_t i;
 207         struct rte_vhost_mem_region *reg;
 208
 209         if (!dev || !dev->mem)
 210                 return;
 211
 212         for (i = 0; i < dev->mem->nregions; i++) {
 213                 reg = &dev->mem->regions[i];
 214                 if (reg->host_user_addr) {
 215                         if (dev->async_copy && rte_vfio_is_enabled("vfio"))
 216                                 async_dma_map(reg, &dev->async_map_status[i], false);
 217
 218                         munmap(reg->mmap_addr, reg->mmap_size);
 219                         close(reg->fd);
 220                 }
 221         }
 222 }
 223
 224 void
 225 vhost_backend_cleanup(struct virtio_net *dev)
 226 {
 227         if (dev->mem) {
 228                 free_mem_region(dev);
 229                 rte_free(dev->mem);
 230                 dev->mem = NULL;
 231
 232                 if (dev->async_map_status) {
 233                         rte_free(dev->async_map_status);
 234                         dev->async_map_status = NULL;
 235                 }
 236         }
 237
 238         rte_free(dev->guest_pages);
 239         dev->guest_pages = NULL;
 240
 241         if (dev->log_addr) {
 242                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 243                 dev->log_addr = 0;
 244         }
 245
 246         if (dev->inflight_info) {
 247                 if (dev->inflight_info->addr) {
 248                         munmap(dev->inflight_info->addr,
 249                                dev->inflight_info->size);
 250                         dev->inflight_info->addr = NULL;
 251                 }
 252
 253                 if (dev->inflight_info->fd >= 0) {
 254                         close(dev->inflight_info->fd);
 255                         dev->inflight_info->fd = -1;
 256                 }
 257
 258                 rte_free(dev->inflight_info);
 259                 dev->inflight_info = NULL;
 260         }
 261
 262         if (dev->slave_req_fd >= 0) {
 263                 close(dev->slave_req_fd);
 264                 dev->slave_req_fd = -1;
 265         }
 266
 267         if (dev->postcopy_ufd >= 0) {
 268                 close(dev->postcopy_ufd);
 269                 dev->postcopy_ufd = -1;
 270         }
 271
 272         dev->postcopy_listening = 0;
 273 }
 274
 275 static void
 276 vhost_user_notify_queue_state(struct virtio_net *dev, uint16_t index,
 277                               int enable)
 278 {
 279         struct rte_vdpa_device *vdpa_dev = dev->vdpa_dev;
 280         struct vhost_virtqueue *vq = dev->virtqueue[index];
 281
 282         /* Configure guest notifications on enable */
 283         if (enable && vq->notif_enable != VIRTIO_UNINITIALIZED_NOTIF)
 284                 vhost_enable_guest_notification(dev, vq, vq->notif_enable);
 285
 286         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
 287                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
 288
 289         if (dev->notify_ops->vring_state_changed)
 290                 dev->notify_ops->vring_state_changed(dev->vid,
 291                                 index, enable);
 292 }
 293
 294 /*
 295  * This function just returns success at the moment unless
 296  * the device hasn't been initialised.
 297  */
 298 static int
 299 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
 300                         struct VhostUserMsg *msg,
 301                         int main_fd __rte_unused)
 302 {
 303         if (validate_msg_fds(msg, 0) != 0)
 304                 return RTE_VHOST_MSG_RESULT_ERR;
 305
 306         return RTE_VHOST_MSG_RESULT_OK;
 307 }
 308
 309 static int
 310 vhost_user_reset_owner(struct virtio_net **pdev,
 311                         struct VhostUserMsg *msg,
 312                         int main_fd __rte_unused)
 313 {
 314         struct virtio_net *dev = *pdev;
 315
 316         if (validate_msg_fds(msg, 0) != 0)
 317                 return RTE_VHOST_MSG_RESULT_ERR;
 318
 319         vhost_destroy_device_notify(dev);
 320
 321         cleanup_device(dev, 0);
 322         reset_device(dev);
 323         return RTE_VHOST_MSG_RESULT_OK;
 324 }
 325
 326 /*
 327  * The features that we support are requested.
 328  */
 329 static int
 330 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 331                         int main_fd __rte_unused)
 332 {
 333         struct virtio_net *dev = *pdev;
 334         uint64_t features = 0;
 335
 336         if (validate_msg_fds(msg, 0) != 0)
 337                 return RTE_VHOST_MSG_RESULT_ERR;
 338
 339         rte_vhost_driver_get_features(dev->ifname, &features);
 340
 341         msg->payload.u64 = features;
 342         msg->size = sizeof(msg->payload.u64);
 343         msg->fd_num = 0;
 344
 345         return RTE_VHOST_MSG_RESULT_REPLY;
 346 }
 347
 348 /*
 349  * The queue number that we support are requested.
 350  */
 351 static int
 352 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
 353                         int main_fd __rte_unused)
 354 {
 355         struct virtio_net *dev = *pdev;
 356         uint32_t queue_num = 0;
 357
 358         if (validate_msg_fds(msg, 0) != 0)
 359                 return RTE_VHOST_MSG_RESULT_ERR;
 360
 361         rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
 362
 363         msg->payload.u64 = (uint64_t)queue_num;
 364         msg->size = sizeof(msg->payload.u64);
 365         msg->fd_num = 0;
 366
 367         return RTE_VHOST_MSG_RESULT_REPLY;
 368 }
 369
 370 /*
 371  * We receive the negotiated features supported by us and the virtio device.
 372  */
 373 static int
 374 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 375                         int main_fd __rte_unused)
 376 {
 377         struct virtio_net *dev = *pdev;
 378         uint64_t features = msg->payload.u64;
 379         uint64_t vhost_features = 0;
 380         struct rte_vdpa_device *vdpa_dev;
 381
 382         if (validate_msg_fds(msg, 0) != 0)
 383                 return RTE_VHOST_MSG_RESULT_ERR;
 384
 385         rte_vhost_driver_get_features(dev->ifname, &vhost_features);
 386         if (features & ~vhost_features) {
 387                 VHOST_LOG_CONFIG(ERR,
 388                         "(%d) received invalid negotiated features.\n",
 389                         dev->vid);
 390                 dev->flags |= VIRTIO_DEV_FEATURES_FAILED;
 391                 dev->status &= ~VIRTIO_DEVICE_STATUS_FEATURES_OK;
 392
 393                 return RTE_VHOST_MSG_RESULT_ERR;
 394         }
 395
 396         if (dev->flags & VIRTIO_DEV_RUNNING) {
 397                 if (dev->features == features)
 398                         return RTE_VHOST_MSG_RESULT_OK;
 399
 400                 /*
 401                  * Error out if master tries to change features while device is
 402                  * in running state. The exception being VHOST_F_LOG_ALL, which
 403                  * is enabled when the live-migration starts.
 404                  */
 405                 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
 406                         VHOST_LOG_CONFIG(ERR,
 407                                 "(%d) features changed while device is running.\n",
 408                                 dev->vid);
 409                         return RTE_VHOST_MSG_RESULT_ERR;
 410                 }
 411
 412                 if (dev->notify_ops->features_changed)
 413                         dev->notify_ops->features_changed(dev->vid, features);
 414         }
 415
 416         dev->features = features;
 417         if (dev->features &
 418                 ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
 419                  (1ULL << VIRTIO_F_VERSION_1) |
 420                  (1ULL << VIRTIO_F_RING_PACKED))) {
 421                 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 422         } else {
 423                 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
 424         }
 425         VHOST_LOG_CONFIG(INFO,
 426                 "negotiated Virtio features: 0x%" PRIx64 "\n", dev->features);
 427         VHOST_LOG_CONFIG(DEBUG,
 428                 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
 429                 dev->vid,
 430                 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
 431                 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 432
 433         if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
 434             !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
 435                 /*
 436                  * Remove all but first queue pair if MQ hasn't been
 437                  * negotiated. This is safe because the device is not
 438                  * running at this stage.
 439                  */
 440                 while (dev->nr_vring > 2) {
 441                         struct vhost_virtqueue *vq;
 442
 443                         vq = dev->virtqueue[--dev->nr_vring];
 444                         if (!vq)
 445                                 continue;
 446
 447                         dev->virtqueue[dev->nr_vring] = NULL;
 448                         cleanup_vq(vq, 1);
 449                         cleanup_vq_inflight(dev, vq);
 450                         free_vq(dev, vq);
 451                 }
 452         }
 453
 454         vdpa_dev = dev->vdpa_dev;
 455         if (vdpa_dev)
 456                 vdpa_dev->ops->set_features(dev->vid);
 457
 458         dev->flags &= ~VIRTIO_DEV_FEATURES_FAILED;
 459         return RTE_VHOST_MSG_RESULT_OK;
 460 }
 461
 462 /*
 463  * The virtio device sends us the size of the descriptor ring.
 464  */
 465 static int
 466 vhost_user_set_vring_num(struct virtio_net **pdev,
 467                         struct VhostUserMsg *msg,
 468                         int main_fd __rte_unused)
 469 {
 470         struct virtio_net *dev = *pdev;
 471         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 472
 473         if (validate_msg_fds(msg, 0) != 0)
 474                 return RTE_VHOST_MSG_RESULT_ERR;
 475
 476         if (msg->payload.state.num > 32768) {
 477                 VHOST_LOG_CONFIG(ERR, "invalid virtqueue size %u\n", msg->payload.state.num);
 478                 return RTE_VHOST_MSG_RESULT_ERR;
 479         }
 480
 481         vq->size = msg->payload.state.num;
 482
 483         /* VIRTIO 1.0, 2.4 Virtqueues says:
 484          *
 485          *   Queue Size value is always a power of 2. The maximum Queue Size
 486          *   value is 32768.
 487          *
 488          * VIRTIO 1.1 2.7 Virtqueues says:
 489          *
 490          *   Packed virtqueues support up to 2^15 entries each.
 491          */
 492         if (!vq_is_packed(dev)) {
 493                 if (vq->size & (vq->size - 1)) {
 494                         VHOST_LOG_CONFIG(ERR,
 495                                 "invalid virtqueue size %u\n", vq->size);
 496                         return RTE_VHOST_MSG_RESULT_ERR;
 497                 }
 498         }
 499
 500         if (vq_is_packed(dev)) {
 501                 if (vq->shadow_used_packed)
 502                         rte_free(vq->shadow_used_packed);
 503                 vq->shadow_used_packed = rte_malloc_socket(NULL,
 504                                 vq->size *
 505                                 sizeof(struct vring_used_elem_packed),
 506                                 RTE_CACHE_LINE_SIZE, vq->numa_node);
 507                 if (!vq->shadow_used_packed) {
 508                         VHOST_LOG_CONFIG(ERR,
 509                                         "failed to allocate memory for shadow used ring.\n");
 510                         return RTE_VHOST_MSG_RESULT_ERR;
 511                 }
 512
 513         } else {
 514                 if (vq->shadow_used_split)
 515                         rte_free(vq->shadow_used_split);
 516
 517                 vq->shadow_used_split = rte_malloc_socket(NULL,
 518                                 vq->size * sizeof(struct vring_used_elem),
 519                                 RTE_CACHE_LINE_SIZE, vq->numa_node);
 520
 521                 if (!vq->shadow_used_split) {
 522                         VHOST_LOG_CONFIG(ERR,
 523                                         "failed to allocate memory for vq internal data.\n");
 524                         return RTE_VHOST_MSG_RESULT_ERR;
 525                 }
 526         }
 527
 528         if (vq->batch_copy_elems)
 529                 rte_free(vq->batch_copy_elems);
 530         vq->batch_copy_elems = rte_malloc_socket(NULL,
 531                                 vq->size * sizeof(struct batch_copy_elem),
 532                                 RTE_CACHE_LINE_SIZE, vq->numa_node);
 533         if (!vq->batch_copy_elems) {
 534                 VHOST_LOG_CONFIG(ERR,
 535                         "failed to allocate memory for batching copy.\n");
 536                 return RTE_VHOST_MSG_RESULT_ERR;
 537         }
 538
 539         return RTE_VHOST_MSG_RESULT_OK;
 540 }
 541
 542 /*
 543  * Reallocate virtio_dev, vhost_virtqueue and related data structures to
 544  * make them on the same numa node as the memory of vring descriptor.
 545  */
 546 #ifdef RTE_LIBRTE_VHOST_NUMA
 547 static struct virtio_net*
 548 numa_realloc(struct virtio_net *dev, int index)
 549 {
 550         int node, dev_node;
 551         struct virtio_net *old_dev;
 552         struct vhost_virtqueue *vq;
 553         struct batch_copy_elem *bce;
 554         struct guest_page *gp;
 555         struct rte_vhost_memory *mem;
 556         size_t mem_size;
 557         int ret;
 558
 559         old_dev = dev;
 560         vq = dev->virtqueue[index];
 561
 562         /*
 563          * If VQ is ready, it is too late to reallocate, it certainly already
 564          * happened anyway on VHOST_USER_SET_VRING_ADRR.
 565          */
 566         if (vq->ready)
 567                 return dev;
 568
 569         ret = get_mempolicy(&node, NULL, 0, vq->desc, MPOL_F_NODE | MPOL_F_ADDR);
 570         if (ret) {
 571                 VHOST_LOG_CONFIG(ERR, "Unable to get virtqueue %d numa information.\n", index);
 572                 return dev;
 573         }
 574
 575         if (node == vq->numa_node)
 576                 goto out_dev_realloc;
 577
 578         vq = rte_realloc_socket(vq, sizeof(*vq), 0, node);
 579         if (!vq) {
 580                 VHOST_LOG_CONFIG(ERR, "Failed to realloc virtqueue %d on node %d\n",
 581                                 index, node);
 582                 return dev;
 583         }
 584
 585         if (vq != dev->virtqueue[index]) {
 586                 VHOST_LOG_CONFIG(INFO, "reallocated virtqueue on node %d\n", node);
 587                 dev->virtqueue[index] = vq;
 588                 vhost_user_iotlb_init(dev, index);
 589         }
 590
 591         if (vq_is_packed(dev)) {
 592                 struct vring_used_elem_packed *sup;
 593
 594                 sup = rte_realloc_socket(vq->shadow_used_packed, vq->size * sizeof(*sup),
 595                                 RTE_CACHE_LINE_SIZE, node);
 596                 if (!sup) {
 597                         VHOST_LOG_CONFIG(ERR, "Failed to realloc shadow packed on node %d\n", node);
 598                         return dev;
 599                 }
 600                 vq->shadow_used_packed = sup;
 601         } else {
 602                 struct vring_used_elem *sus;
 603
 604                 sus = rte_realloc_socket(vq->shadow_used_split, vq->size * sizeof(*sus),
 605                                 RTE_CACHE_LINE_SIZE, node);
 606                 if (!sus) {
 607                         VHOST_LOG_CONFIG(ERR, "Failed to realloc shadow split on node %d\n", node);
 608                         return dev;
 609                 }
 610                 vq->shadow_used_split = sus;
 611         }
 612
 613         bce = rte_realloc_socket(vq->batch_copy_elems, vq->size * sizeof(*bce),
 614                         RTE_CACHE_LINE_SIZE, node);
 615         if (!bce) {
 616                 VHOST_LOG_CONFIG(ERR, "Failed to realloc batch copy elem on node %d\n", node);
 617                 return dev;
 618         }
 619         vq->batch_copy_elems = bce;
 620
 621         if (vq->log_cache) {
 622                 struct log_cache_entry *lc;
 623
 624                 lc = rte_realloc_socket(vq->log_cache, sizeof(*lc) * VHOST_LOG_CACHE_NR, 0, node);
 625                 if (!lc) {
 626                         VHOST_LOG_CONFIG(ERR, "Failed to realloc log cache on node %d\n", node);
 627                         return dev;
 628                 }
 629                 vq->log_cache = lc;
 630         }
 631
 632         if (vq->resubmit_inflight) {
 633                 struct rte_vhost_resubmit_info *ri;
 634
 635                 ri = rte_realloc_socket(vq->resubmit_inflight, sizeof(*ri), 0, node);
 636                 if (!ri) {
 637                         VHOST_LOG_CONFIG(ERR, "Failed to realloc resubmit inflight on node %d\n",
 638                                         node);
 639                         return dev;
 640                 }
 641                 vq->resubmit_inflight = ri;
 642
 643                 if (ri->resubmit_list) {
 644                         struct rte_vhost_resubmit_desc *rd;
 645
 646                         rd = rte_realloc_socket(ri->resubmit_list, sizeof(*rd) * ri->resubmit_num,
 647                                         0, node);
 648                         if (!rd) {
 649                                 VHOST_LOG_CONFIG(ERR, "Failed to realloc resubmit list on node %d\n",
 650                                                 node);
 651                                 return dev;
 652                         }
 653                         ri->resubmit_list = rd;
 654                 }
 655         }
 656
 657         vq->numa_node = node;
 658
 659 out_dev_realloc:
 660
 661         if (dev->flags & VIRTIO_DEV_RUNNING)
 662                 return dev;
 663
 664         ret = get_mempolicy(&dev_node, NULL, 0, dev, MPOL_F_NODE | MPOL_F_ADDR);
 665         if (ret) {
 666                 VHOST_LOG_CONFIG(ERR, "Unable to get Virtio dev %d numa information.\n", dev->vid);
 667                 return dev;
 668         }
 669
 670         if (dev_node == node)
 671                 return dev;
 672
 673         dev = rte_realloc_socket(old_dev, sizeof(*dev), 0, node);
 674         if (!dev) {
 675                 VHOST_LOG_CONFIG(ERR, "Failed to realloc dev on node %d\n", node);
 676                 return old_dev;
 677         }
 678
 679         VHOST_LOG_CONFIG(INFO, "reallocated device on node %d\n", node);
 680         vhost_devices[dev->vid] = dev;
 681
 682         mem_size = sizeof(struct rte_vhost_memory) +
 683                 sizeof(struct rte_vhost_mem_region) * dev->mem->nregions;
 684         mem = rte_realloc_socket(dev->mem, mem_size, 0, node);
 685         if (!mem) {
 686                 VHOST_LOG_CONFIG(ERR, "Failed to realloc mem table on node %d\n", node);
 687                 return dev;
 688         }
 689         dev->mem = mem;
 690
 691         if (dev->async_copy && rte_vfio_is_enabled("vfio")) {
 692                 if (dev->async_map_status == NULL) {
 693                         dev->async_map_status = rte_zmalloc_socket("async-dma-map-status",
 694                                         sizeof(bool) * dev->mem->nregions, 0, node);
 695                         if (!dev->async_map_status) {
 696                                 VHOST_LOG_CONFIG(ERR,
 697                                         "(%d) failed to realloc dma mapping status on node\n",
 698                                         dev->vid);
 699                                 return dev;
 700                         }
 701                 }
 702         }
 703
 704         gp = rte_realloc_socket(dev->guest_pages, dev->max_guest_pages * sizeof(*gp),
 705                         RTE_CACHE_LINE_SIZE, node);
 706         if (!gp) {
 707                 VHOST_LOG_CONFIG(ERR, "Failed to realloc guest pages on node %d\n", node);
 708                 return dev;
 709         }
 710         dev->guest_pages = gp;
 711
 712         return dev;
 713 }
 714 #else
 715 static struct virtio_net*
 716 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 717 {
 718         return dev;
 719 }
 720 #endif
 721
 722 /* Converts QEMU virtual address to Vhost virtual address. */
 723 static uint64_t
 724 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
 725 {
 726         struct rte_vhost_mem_region *r;
 727         uint32_t i;
 728
 729         if (unlikely(!dev || !dev->mem))
 730                 goto out_error;
 731
 732         /* Find the region where the address lives. */
 733         for (i = 0; i < dev->mem->nregions; i++) {
 734                 r = &dev->mem->regions[i];
 735
 736                 if (qva >= r->guest_user_addr &&
 737                     qva <  r->guest_user_addr + r->size) {
 738
 739                         if (unlikely(*len > r->guest_user_addr + r->size - qva))
 740                                 *len = r->guest_user_addr + r->size - qva;
 741
 742                         return qva - r->guest_user_addr +
 743                                r->host_user_addr;
 744                 }
 745         }
 746 out_error:
 747         *len = 0;
 748
 749         return 0;
 750 }
 751
 752
 753 /*
 754  * Converts ring address to Vhost virtual address.
 755  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 756  * else it is a QEMU virtual address.
 757  */
 758 static uint64_t
 759 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 760                 uint64_t ra, uint64_t *size)
 761 {
 762         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 763                 uint64_t vva;
 764
 765                 vhost_user_iotlb_rd_lock(vq);
 766                 vva = vhost_iova_to_vva(dev, vq, ra,
 767                                         size, VHOST_ACCESS_RW);
 768                 vhost_user_iotlb_rd_unlock(vq);
 769
 770                 return vva;
 771         }
 772
 773         return qva_to_vva(dev, ra, size);
 774 }
 775
 776 static uint64_t
 777 log_addr_to_gpa(struct virtio_net *dev, struct vhost_virtqueue *vq)
 778 {
 779         uint64_t log_gpa;
 780
 781         vhost_user_iotlb_rd_lock(vq);
 782         log_gpa = translate_log_addr(dev, vq, vq->ring_addrs.log_guest_addr);
 783         vhost_user_iotlb_rd_unlock(vq);
 784
 785         return log_gpa;
 786 }
 787
 788 static struct virtio_net *
 789 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 790 {
 791         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 792         struct vhost_vring_addr *addr = &vq->ring_addrs;
 793         uint64_t len, expected_len;
 794
 795         if (addr->flags & (1 << VHOST_VRING_F_LOG)) {
 796                 vq->log_guest_addr =
 797                         log_addr_to_gpa(dev, vq);
 798                 if (vq->log_guest_addr == 0) {
 799                         VHOST_LOG_CONFIG(DEBUG,
 800                                 "(%d) failed to map log_guest_addr.\n",
 801                                 dev->vid);
 802                         return dev;
 803                 }
 804         }
 805
 806         if (vq_is_packed(dev)) {
 807                 len = sizeof(struct vring_packed_desc) * vq->size;
 808                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 809                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 810                 if (vq->desc_packed == NULL ||
 811                                 len != sizeof(struct vring_packed_desc) *
 812                                 vq->size) {
 813                         VHOST_LOG_CONFIG(DEBUG,
 814                                 "(%d) failed to map desc_packed ring.\n",
 815                                 dev->vid);
 816                         return dev;
 817                 }
 818
 819                 dev = numa_realloc(dev, vq_index);
 820                 vq = dev->virtqueue[vq_index];
 821                 addr = &vq->ring_addrs;
 822
 823                 len = sizeof(struct vring_packed_desc_event);
 824                 vq->driver_event = (struct vring_packed_desc_event *)
 825                                         (uintptr_t)ring_addr_to_vva(dev,
 826                                         vq, addr->avail_user_addr, &len);
 827                 if (vq->driver_event == NULL ||
 828                                 len != sizeof(struct vring_packed_desc_event)) {
 829                         VHOST_LOG_CONFIG(DEBUG,
 830                                 "(%d) failed to find driver area address.\n",
 831                                 dev->vid);
 832                         return dev;
 833                 }
 834
 835                 len = sizeof(struct vring_packed_desc_event);
 836                 vq->device_event = (struct vring_packed_desc_event *)
 837                                         (uintptr_t)ring_addr_to_vva(dev,
 838                                         vq, addr->used_user_addr, &len);
 839                 if (vq->device_event == NULL ||
 840                                 len != sizeof(struct vring_packed_desc_event)) {
 841                         VHOST_LOG_CONFIG(DEBUG,
 842                                 "(%d) failed to find device area address.\n",
 843                                 dev->vid);
 844                         return dev;
 845                 }
 846
 847                 vq->access_ok = true;
 848                 return dev;
 849         }
 850
 851         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 852         if (vq->desc && vq->avail && vq->used)
 853                 return dev;
 854
 855         len = sizeof(struct vring_desc) * vq->size;
 856         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 857                         vq, addr->desc_user_addr, &len);
 858         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 859                 VHOST_LOG_CONFIG(DEBUG,
 860                         "(%d) failed to map desc ring.\n",
 861                         dev->vid);
 862                 return dev;
 863         }
 864
 865         dev = numa_realloc(dev, vq_index);
 866         vq = dev->virtqueue[vq_index];
 867         addr = &vq->ring_addrs;
 868
 869         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 870         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 871                 len += sizeof(uint16_t);
 872         expected_len = len;
 873         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 874                         vq, addr->avail_user_addr, &len);
 875         if (vq->avail == 0 || len != expected_len) {
 876                 VHOST_LOG_CONFIG(DEBUG,
 877                         "(%d) failed to map avail ring.\n",
 878                         dev->vid);
 879                 return dev;
 880         }
 881
 882         len = sizeof(struct vring_used) +
 883                 sizeof(struct vring_used_elem) * vq->size;
 884         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 885                 len += sizeof(uint16_t);
 886         expected_len = len;
 887         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 888                         vq, addr->used_user_addr, &len);
 889         if (vq->used == 0 || len != expected_len) {
 890                 VHOST_LOG_CONFIG(DEBUG,
 891                         "(%d) failed to map used ring.\n",
 892                         dev->vid);
 893                 return dev;
 894         }
 895
 896         if (vq->last_used_idx != vq->used->idx) {
 897                 VHOST_LOG_CONFIG(WARNING,
 898                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 899                         "some packets maybe resent for Tx and dropped for Rx\n",
 900                         vq->last_used_idx, vq->used->idx);
 901                 vq->last_used_idx  = vq->used->idx;
 902                 vq->last_avail_idx = vq->used->idx;
 903         }
 904
 905         vq->access_ok = true;
 906
 907         VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address desc: %p\n",
 908                         dev->vid, vq->desc);
 909         VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address avail: %p\n",
 910                         dev->vid, vq->avail);
 911         VHOST_LOG_CONFIG(DEBUG, "(%d) mapped address used: %p\n",
 912                         dev->vid, vq->used);
 913         VHOST_LOG_CONFIG(DEBUG, "(%d) log_guest_addr: %" PRIx64 "\n",
 914                         dev->vid, vq->log_guest_addr);
 915
 916         return dev;
 917 }
 918
 919 /*
 920  * The virtio device sends us the desc, used and avail ring addresses.
 921  * This function then converts these to our address space.
 922  */
 923 static int
 924 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 925                         int main_fd __rte_unused)
 926 {
 927         struct virtio_net *dev = *pdev;
 928         struct vhost_virtqueue *vq;
 929         struct vhost_vring_addr *addr = &msg->payload.addr;
 930         bool access_ok;
 931
 932         if (validate_msg_fds(msg, 0) != 0)
 933                 return RTE_VHOST_MSG_RESULT_ERR;
 934
 935         if (dev->mem == NULL)
 936                 return RTE_VHOST_MSG_RESULT_ERR;
 937
 938         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 939         vq = dev->virtqueue[msg->payload.addr.index];
 940
 941         access_ok = vq->access_ok;
 942
 943         /*
 944          * Rings addresses should not be interpreted as long as the ring is not
 945          * started and enabled
 946          */
 947         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 948
 949         vring_invalidate(dev, vq);
 950
 951         if ((vq->enabled && (dev->features &
 952                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) ||
 953                         access_ok) {
 954                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 955                 if (!dev)
 956                         return RTE_VHOST_MSG_RESULT_ERR;
 957
 958                 *pdev = dev;
 959         }
 960
 961         return RTE_VHOST_MSG_RESULT_OK;
 962 }
 963
 964 /*
 965  * The virtio device sends us the available ring last used index.
 966  */
 967 static int
 968 vhost_user_set_vring_base(struct virtio_net **pdev,
 969                         struct VhostUserMsg *msg,
 970                         int main_fd __rte_unused)
 971 {
 972         struct virtio_net *dev = *pdev;
 973         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 974         uint64_t val = msg->payload.state.num;
 975
 976         if (validate_msg_fds(msg, 0) != 0)
 977                 return RTE_VHOST_MSG_RESULT_ERR;
 978
 979         if (vq_is_packed(dev)) {
 980                 /*
 981                  * Bit[0:14]: avail index
 982                  * Bit[15]: avail wrap counter
 983                  */
 984                 vq->last_avail_idx = val & 0x7fff;
 985                 vq->avail_wrap_counter = !!(val & (0x1 << 15));
 986                 /*
 987                  * Set used index to same value as available one, as
 988                  * their values should be the same since ring processing
 989                  * was stopped at get time.
 990                  */
 991                 vq->last_used_idx = vq->last_avail_idx;
 992                 vq->used_wrap_counter = vq->avail_wrap_counter;
 993         } else {
 994                 vq->last_used_idx = msg->payload.state.num;
 995                 vq->last_avail_idx = msg->payload.state.num;
 996         }
 997
 998         return RTE_VHOST_MSG_RESULT_OK;
 999 }
1000
1001 static int
1002 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
1003                    uint64_t host_phys_addr, uint64_t size)
1004 {
1005         struct guest_page *page, *last_page;
1006         struct guest_page *old_pages;
1007
1008         if (dev->nr_guest_pages == dev->max_guest_pages) {
1009                 dev->max_guest_pages *= 2;
1010                 old_pages = dev->guest_pages;
1011                 dev->guest_pages = rte_realloc(dev->guest_pages,
1012                                         dev->max_guest_pages * sizeof(*page),
1013                                         RTE_CACHE_LINE_SIZE);
1014                 if (dev->guest_pages == NULL) {
1015                         VHOST_LOG_CONFIG(ERR, "cannot realloc guest_pages\n");
1016                         rte_free(old_pages);
1017                         return -1;
1018                 }
1019         }
1020
1021         if (dev->nr_guest_pages > 0) {
1022                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
1023                 /* merge if the two pages are continuous */
1024                 if (host_phys_addr == last_page->host_phys_addr +
1025                                       last_page->size) {
1026                         last_page->size += size;
1027                         return 0;
1028                 }
1029         }
1030
1031         page = &dev->guest_pages[dev->nr_guest_pages++];
1032         page->guest_phys_addr = guest_phys_addr;
1033         page->host_phys_addr  = host_phys_addr;
1034         page->size = size;
1035
1036         return 0;
1037 }
1038
1039 static int
1040 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
1041                 uint64_t page_size)
1042 {
1043         uint64_t reg_size = reg->size;
1044         uint64_t host_user_addr  = reg->host_user_addr;
1045         uint64_t guest_phys_addr = reg->guest_phys_addr;
1046         uint64_t host_phys_addr;
1047         uint64_t size;
1048
1049         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
1050         size = page_size - (guest_phys_addr & (page_size - 1));
1051         size = RTE_MIN(size, reg_size);
1052
1053         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
1054                 return -1;
1055
1056         host_user_addr  += size;
1057         guest_phys_addr += size;
1058         reg_size -= size;
1059
1060         while (reg_size > 0) {
1061                 size = RTE_MIN(reg_size, page_size);
1062                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
1063                                                   host_user_addr);
1064                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
1065                                 size) < 0)
1066                         return -1;
1067
1068                 host_user_addr  += size;
1069                 guest_phys_addr += size;
1070                 reg_size -= size;
1071         }
1072
1073         /* sort guest page array if over binary search threshold */
1074         if (dev->nr_guest_pages >= VHOST_BINARY_SEARCH_THRESH) {
1075                 qsort((void *)dev->guest_pages, dev->nr_guest_pages,
1076                         sizeof(struct guest_page), guest_page_addrcmp);
1077         }
1078
1079         return 0;
1080 }
1081
1082 #ifdef RTE_LIBRTE_VHOST_DEBUG
1083 /* TODO: enable it only in debug mode? */
1084 static void
1085 dump_guest_pages(struct virtio_net *dev)
1086 {
1087         uint32_t i;
1088         struct guest_page *page;
1089
1090         for (i = 0; i < dev->nr_guest_pages; i++) {
1091                 page = &dev->guest_pages[i];
1092
1093                 VHOST_LOG_CONFIG(INFO,
1094                         "guest physical page region %u\n"
1095                         "\t guest_phys_addr: %" PRIx64 "\n"
1096                         "\t host_phys_addr : %" PRIx64 "\n"
1097                         "\t size           : %" PRIx64 "\n",
1098                         i,
1099                         page->guest_phys_addr,
1100                         page->host_phys_addr,
1101                         page->size);
1102         }
1103 }
1104 #else
1105 #define dump_guest_pages(dev)
1106 #endif
1107
1108 static bool
1109 vhost_memory_changed(struct VhostUserMemory *new,
1110                      struct rte_vhost_memory *old)
1111 {
1112         uint32_t i;
1113
1114         if (new->nregions != old->nregions)
1115                 return true;
1116
1117         for (i = 0; i < new->nregions; ++i) {
1118                 VhostUserMemoryRegion *new_r = &new->regions[i];
1119                 struct rte_vhost_mem_region *old_r = &old->regions[i];
1120
1121                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
1122                         return true;
1123                 if (new_r->memory_size != old_r->size)
1124                         return true;
1125                 if (new_r->userspace_addr != old_r->guest_user_addr)
1126                         return true;
1127         }
1128
1129         return false;
1130 }
1131
1132 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1133 static int
1134 vhost_user_postcopy_region_register(struct virtio_net *dev,
1135                 struct rte_vhost_mem_region *reg)
1136 {
1137         struct uffdio_register reg_struct;
1138
1139         /*
1140          * Let's register all the mmap'ed area to ensure
1141          * alignment on page boundary.
1142          */
1143         reg_struct.range.start = (uint64_t)(uintptr_t)reg->mmap_addr;
1144         reg_struct.range.len = reg->mmap_size;
1145         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1146
1147         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1148                                 &reg_struct)) {
1149                 VHOST_LOG_CONFIG(ERR, "Failed to register ufd for region "
1150                                 "%" PRIx64 " - %" PRIx64 " (ufd = %d) %s\n",
1151                                 (uint64_t)reg_struct.range.start,
1152                                 (uint64_t)reg_struct.range.start +
1153                                 (uint64_t)reg_struct.range.len - 1,
1154                                 dev->postcopy_ufd,
1155                                 strerror(errno));
1156                 return -1;
1157         }
1158
1159         VHOST_LOG_CONFIG(INFO, "\t userfaultfd registered for range : %" PRIx64 " - %" PRIx64 "\n",
1160                         (uint64_t)reg_struct.range.start,
1161                         (uint64_t)reg_struct.range.start +
1162                         (uint64_t)reg_struct.range.len - 1);
1163
1164         return 0;
1165 }
1166 #else
1167 static int
1168 vhost_user_postcopy_region_register(struct virtio_net *dev __rte_unused,
1169                 struct rte_vhost_mem_region *reg __rte_unused)
1170 {
1171         return -1;
1172 }
1173 #endif
1174
1175 static int
1176 vhost_user_postcopy_register(struct virtio_net *dev, int main_fd,
1177                 struct VhostUserMsg *msg)
1178 {
1179         struct VhostUserMemory *memory;
1180         struct rte_vhost_mem_region *reg;
1181         VhostUserMsg ack_msg;
1182         uint32_t i;
1183
1184         if (!dev->postcopy_listening)
1185                 return 0;
1186
1187         /*
1188          * We haven't a better way right now than sharing
1189          * DPDK's virtual address with Qemu, so that Qemu can
1190          * retrieve the region offset when handling userfaults.
1191          */
1192         memory = &msg->payload.memory;
1193         for (i = 0; i < memory->nregions; i++) {
1194                 reg = &dev->mem->regions[i];
1195                 memory->regions[i].userspace_addr = reg->host_user_addr;
1196         }
1197
1198         /* Send the addresses back to qemu */
1199         msg->fd_num = 0;
1200         send_vhost_reply(main_fd, msg);
1201
1202         /* Wait for qemu to acknolwedge it's got the addresses
1203          * we've got to wait before we're allowed to generate faults.
1204          */
1205         if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1206                 VHOST_LOG_CONFIG(ERR,
1207                                 "Failed to read qemu ack on postcopy set-mem-table\n");
1208                 return -1;
1209         }
1210
1211         if (validate_msg_fds(&ack_msg, 0) != 0)
1212                 return -1;
1213
1214         if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1215                 VHOST_LOG_CONFIG(ERR,
1216                                 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1217                                 ack_msg.request.master);
1218                 return -1;
1219         }
1220
1221         /* Now userfault register and we can use the memory */
1222         for (i = 0; i < memory->nregions; i++) {
1223                 reg = &dev->mem->regions[i];
1224                 if (vhost_user_postcopy_region_register(dev, reg) < 0)
1225                         return -1;
1226         }
1227
1228         return 0;
1229 }
1230
1231 static int
1232 vhost_user_mmap_region(struct virtio_net *dev,
1233                 struct rte_vhost_mem_region *region,
1234                 uint32_t region_index,
1235                 uint64_t mmap_offset)
1236 {
1237         void *mmap_addr;
1238         uint64_t mmap_size;
1239         uint64_t alignment;
1240         int populate;
1241         int ret;
1242
1243         /* Check for memory_size + mmap_offset overflow */
1244         if (mmap_offset >= -region->size) {
1245                 VHOST_LOG_CONFIG(ERR,
1246                                 "mmap_offset (%#"PRIx64") and memory_size "
1247                                 "(%#"PRIx64") overflow\n",
1248                                 mmap_offset, region->size);
1249                 return -1;
1250         }
1251
1252         mmap_size = region->size + mmap_offset;
1253
1254         /* mmap() without flag of MAP_ANONYMOUS, should be called with length
1255          * argument aligned with hugepagesz at older longterm version Linux,
1256          * like 2.6.32 and 3.2.72, or mmap() will fail with EINVAL.
1257          *
1258          * To avoid failure, make sure in caller to keep length aligned.
1259          */
1260         alignment = get_blk_size(region->fd);
1261         if (alignment == (uint64_t)-1) {
1262                 VHOST_LOG_CONFIG(ERR,
1263                                 "couldn't get hugepage size through fstat\n");
1264                 return -1;
1265         }
1266         mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
1267         if (mmap_size == 0) {
1268                 /*
1269                  * It could happen if initial mmap_size + alignment overflows
1270                  * the sizeof uint64, which could happen if either mmap_size or
1271                  * alignment value is wrong.
1272                  *
1273                  * mmap() kernel implementation would return an error, but
1274                  * better catch it before and provide useful info in the logs.
1275                  */
1276                 VHOST_LOG_CONFIG(ERR, "mmap size (0x%" PRIx64 ") "
1277                                 "or alignment (0x%" PRIx64 ") is invalid\n",
1278                                 region->size + mmap_offset, alignment);
1279                 return -1;
1280         }
1281
1282         populate = dev->async_copy ? MAP_POPULATE : 0;
1283         mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
1284                         MAP_SHARED | populate, region->fd, 0);
1285
1286         if (mmap_addr == MAP_FAILED) {
1287                 VHOST_LOG_CONFIG(ERR, "mmap failed (%s).\n", strerror(errno));
1288                 return -1;
1289         }
1290
1291         region->mmap_addr = mmap_addr;
1292         region->mmap_size = mmap_size;
1293         region->host_user_addr = (uint64_t)(uintptr_t)mmap_addr + mmap_offset;
1294
1295         if (dev->async_copy) {
1296                 if (add_guest_pages(dev, region, alignment) < 0) {
1297                         VHOST_LOG_CONFIG(ERR,
1298                                         "adding guest pages to region failed.\n");
1299                         return -1;
1300                 }
1301
1302                 if (rte_vfio_is_enabled("vfio")) {
1303                         ret = async_dma_map(region, &dev->async_map_status[region_index], true);
1304                         if (ret) {
1305                                 VHOST_LOG_CONFIG(ERR, "Configure IOMMU for DMA "
1306                                                         "engine failed\n");
1307                                 return -1;
1308                         }
1309                 }
1310         }
1311
1312         VHOST_LOG_CONFIG(INFO,
1313                         "guest memory region size: 0x%" PRIx64 "\n"
1314                         "\t guest physical addr: 0x%" PRIx64 "\n"
1315                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
1316                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
1317                         "\t mmap addr : 0x%" PRIx64 "\n"
1318                         "\t mmap size : 0x%" PRIx64 "\n"
1319                         "\t mmap align: 0x%" PRIx64 "\n"
1320                         "\t mmap off  : 0x%" PRIx64 "\n",
1321                         region->size,
1322                         region->guest_phys_addr,
1323                         region->guest_user_addr,
1324                         region->host_user_addr,
1325                         (uint64_t)(uintptr_t)mmap_addr,
1326                         mmap_size,
1327                         alignment,
1328                         mmap_offset);
1329
1330         return 0;
1331 }
1332
1333 static int
1334 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
1335                         int main_fd)
1336 {
1337         struct virtio_net *dev = *pdev;
1338         struct VhostUserMemory *memory = &msg->payload.memory;
1339         struct rte_vhost_mem_region *reg;
1340         int numa_node = SOCKET_ID_ANY;
1341         uint64_t mmap_offset;
1342         uint32_t i;
1343         bool async_notify = false;
1344
1345         if (validate_msg_fds(msg, memory->nregions) != 0)
1346                 return RTE_VHOST_MSG_RESULT_ERR;
1347
1348         if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
1349                 VHOST_LOG_CONFIG(ERR,
1350                         "too many memory regions (%u)\n", memory->nregions);
1351                 goto close_msg_fds;
1352         }
1353
1354         if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
1355                 VHOST_LOG_CONFIG(INFO,
1356                         "(%d) memory regions not changed\n", dev->vid);
1357
1358                 close_msg_fds(msg);
1359
1360                 return RTE_VHOST_MSG_RESULT_OK;
1361         }
1362
1363         if (dev->mem) {
1364                 if (dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) {
1365                         struct rte_vdpa_device *vdpa_dev = dev->vdpa_dev;
1366
1367                         if (vdpa_dev && vdpa_dev->ops->dev_close)
1368                                 vdpa_dev->ops->dev_close(dev->vid);
1369                         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1370                 }
1371
1372                 /* notify the vhost application to stop DMA transfers */
1373                 if (dev->async_copy && dev->notify_ops->vring_state_changed) {
1374                         for (i = 0; i < dev->nr_vring; i++) {
1375                                 dev->notify_ops->vring_state_changed(dev->vid,
1376                                                 i, 0);
1377                         }
1378                         async_notify = true;
1379                 }
1380
1381                 free_mem_region(dev);
1382                 rte_free(dev->mem);
1383                 dev->mem = NULL;
1384
1385                 if (dev->async_map_status) {
1386                         rte_free(dev->async_map_status);
1387                         dev->async_map_status = NULL;
1388                 }
1389         }
1390
1391         /* Flush IOTLB cache as previous HVAs are now invalid */
1392         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1393                 for (i = 0; i < dev->nr_vring; i++)
1394                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
1395
1396         /*
1397          * If VQ 0 has already been allocated, try to allocate on the same
1398          * NUMA node. It can be reallocated later in numa_realloc().
1399          */
1400         if (dev->nr_vring > 0)
1401                 numa_node = dev->virtqueue[0]->numa_node;
1402
1403         dev->nr_guest_pages = 0;
1404         if (dev->guest_pages == NULL) {
1405                 dev->max_guest_pages = 8;
1406                 dev->guest_pages = rte_zmalloc_socket(NULL,
1407                                         dev->max_guest_pages *
1408                                         sizeof(struct guest_page),
1409                                         RTE_CACHE_LINE_SIZE,
1410                                         numa_node);
1411                 if (dev->guest_pages == NULL) {
1412                         VHOST_LOG_CONFIG(ERR,
1413                                 "(%d) failed to allocate memory "
1414                                 "for dev->guest_pages\n",
1415                                 dev->vid);
1416                         goto close_msg_fds;
1417                 }
1418         }
1419
1420         dev->mem = rte_zmalloc_socket("vhost-mem-table", sizeof(struct rte_vhost_memory) +
1421                 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0, numa_node);
1422         if (dev->mem == NULL) {
1423                 VHOST_LOG_CONFIG(ERR,
1424                         "(%d) failed to allocate memory for dev->mem\n",
1425                         dev->vid);
1426                 goto free_guest_pages;
1427         }
1428
1429         if (dev->async_copy) {
1430                 dev->async_map_status = rte_zmalloc_socket("async-dma-map-status",
1431                                         sizeof(bool) * memory->nregions, 0, numa_node);
1432                 if (!dev->async_map_status) {
1433                         VHOST_LOG_CONFIG(ERR,
1434                                 "(%d) failed to allocate memory for dma mapping status\n",
1435                                 dev->vid);
1436                         goto free_mem_table;
1437                 }
1438         }
1439
1440         for (i = 0; i < memory->nregions; i++) {
1441                 reg = &dev->mem->regions[i];
1442
1443                 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
1444                 reg->guest_user_addr = memory->regions[i].userspace_addr;
1445                 reg->size            = memory->regions[i].memory_size;
1446                 reg->fd              = msg->fds[i];
1447
1448                 /*
1449                  * Assign invalid file descriptor value to avoid double
1450                  * closing on error path.
1451                  */
1452                 msg->fds[i] = -1;
1453
1454                 mmap_offset = memory->regions[i].mmap_offset;
1455
1456                 if (vhost_user_mmap_region(dev, reg, i, mmap_offset) < 0) {
1457                         VHOST_LOG_CONFIG(ERR, "Failed to mmap region %u\n", i);
1458                         goto free_mem_table;
1459                 }
1460
1461                 dev->mem->nregions++;
1462         }
1463
1464         if (vhost_user_postcopy_register(dev, main_fd, msg) < 0)
1465                 goto free_mem_table;
1466
1467         for (i = 0; i < dev->nr_vring; i++) {
1468                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1469
1470                 if (!vq)
1471                         continue;
1472
1473                 if (vq->desc || vq->avail || vq->used) {
1474                         /*
1475                          * If the memory table got updated, the ring addresses
1476                          * need to be translated again as virtual addresses have
1477                          * changed.
1478                          */
1479                         vring_invalidate(dev, vq);
1480
1481                         dev = translate_ring_addresses(dev, i);
1482                         if (!dev) {
1483                                 dev = *pdev;
1484                                 goto free_mem_table;
1485                         }
1486
1487                         *pdev = dev;
1488                 }
1489         }
1490
1491         dump_guest_pages(dev);
1492
1493         if (async_notify) {
1494                 for (i = 0; i < dev->nr_vring; i++)
1495                         dev->notify_ops->vring_state_changed(dev->vid, i, 1);
1496         }
1497
1498         return RTE_VHOST_MSG_RESULT_OK;
1499
1500 free_mem_table:
1501         free_mem_region(dev);
1502         rte_free(dev->mem);
1503         dev->mem = NULL;
1504         if (dev->async_map_status) {
1505                 rte_free(dev->async_map_status);
1506                 dev->async_map_status = NULL;
1507         }
1508 free_guest_pages:
1509         rte_free(dev->guest_pages);
1510         dev->guest_pages = NULL;
1511 close_msg_fds:
1512         close_msg_fds(msg);
1513         return RTE_VHOST_MSG_RESULT_ERR;
1514 }
1515
1516 static bool
1517 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1518 {
1519         bool rings_ok;
1520
1521         if (!vq)
1522                 return false;
1523
1524         if (vq_is_packed(dev))
1525                 rings_ok = vq->desc_packed && vq->driver_event &&
1526                         vq->device_event;
1527         else
1528                 rings_ok = vq->desc && vq->avail && vq->used;
1529
1530         return rings_ok &&
1531                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1532                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1533                vq->enabled;
1534 }
1535
1536 #define VIRTIO_BUILTIN_NUM_VQS_TO_BE_READY 2u
1537
1538 static int
1539 virtio_is_ready(struct virtio_net *dev)
1540 {
1541         struct vhost_virtqueue *vq;
1542         uint32_t i, nr_vring = dev->nr_vring;
1543
1544         if (dev->flags & VIRTIO_DEV_READY)
1545                 return 1;
1546
1547         if (!dev->nr_vring)
1548                 return 0;
1549
1550         if (dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) {
1551                 nr_vring = VIRTIO_BUILTIN_NUM_VQS_TO_BE_READY;
1552
1553                 if (dev->nr_vring < nr_vring)
1554                         return 0;
1555         }
1556
1557         for (i = 0; i < nr_vring; i++) {
1558                 vq = dev->virtqueue[i];
1559
1560                 if (!vq_is_ready(dev, vq))
1561                         return 0;
1562         }
1563
1564         /* If supported, ensure the frontend is really done with config */
1565         if (dev->protocol_features & (1ULL << VHOST_USER_PROTOCOL_F_STATUS))
1566                 if (!(dev->status & VIRTIO_DEVICE_STATUS_DRIVER_OK))
1567                         return 0;
1568
1569         dev->flags |= VIRTIO_DEV_READY;
1570
1571         if (!(dev->flags & VIRTIO_DEV_RUNNING))
1572                 VHOST_LOG_CONFIG(INFO,
1573                         "virtio is now ready for processing.\n");
1574         return 1;
1575 }
1576
1577 static void *
1578 inflight_mem_alloc(const char *name, size_t size, int *fd)
1579 {
1580         void *ptr;
1581         int mfd = -1;
1582         char fname[20] = "/tmp/memfd-XXXXXX";
1583
1584         *fd = -1;
1585 #ifdef MEMFD_SUPPORTED
1586         mfd = memfd_create(name, MFD_CLOEXEC);
1587 #else
1588         RTE_SET_USED(name);
1589 #endif
1590         if (mfd == -1) {
1591                 mfd = mkstemp(fname);
1592                 if (mfd == -1) {
1593                         VHOST_LOG_CONFIG(ERR,
1594                                 "failed to get inflight buffer fd\n");
1595                         return NULL;
1596                 }
1597
1598                 unlink(fname);
1599         }
1600
1601         if (ftruncate(mfd, size) == -1) {
1602                 VHOST_LOG_CONFIG(ERR,
1603                         "failed to alloc inflight buffer\n");
1604                 close(mfd);
1605                 return NULL;
1606         }
1607
1608         ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd, 0);
1609         if (ptr == MAP_FAILED) {
1610                 VHOST_LOG_CONFIG(ERR,
1611                         "failed to mmap inflight buffer\n");
1612                 close(mfd);
1613                 return NULL;
1614         }
1615
1616         *fd = mfd;
1617         return ptr;
1618 }
1619
1620 static uint32_t
1621 get_pervq_shm_size_split(uint16_t queue_size)
1622 {
1623         return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_split) *
1624                                   queue_size + sizeof(uint64_t) +
1625                                   sizeof(uint16_t) * 4, INFLIGHT_ALIGNMENT);
1626 }
1627
1628 static uint32_t
1629 get_pervq_shm_size_packed(uint16_t queue_size)
1630 {
1631         return RTE_ALIGN_MUL_CEIL(sizeof(struct rte_vhost_inflight_desc_packed)
1632                                   * queue_size + sizeof(uint64_t) +
1633                                   sizeof(uint16_t) * 6 + sizeof(uint8_t) * 9,
1634                                   INFLIGHT_ALIGNMENT);
1635 }
1636
1637 static int
1638 vhost_user_get_inflight_fd(struct virtio_net **pdev,
1639                            VhostUserMsg *msg,
1640                            int main_fd __rte_unused)
1641 {
1642         struct rte_vhost_inflight_info_packed *inflight_packed;
1643         uint64_t pervq_inflight_size, mmap_size;
1644         uint16_t num_queues, queue_size;
1645         struct virtio_net *dev = *pdev;
1646         int fd, i, j;
1647         int numa_node = SOCKET_ID_ANY;
1648         void *addr;
1649
1650         if (msg->size != sizeof(msg->payload.inflight)) {
1651                 VHOST_LOG_CONFIG(ERR,
1652                         "invalid get_inflight_fd message size is %d\n",
1653                         msg->size);
1654                 return RTE_VHOST_MSG_RESULT_ERR;
1655         }
1656
1657         /*
1658          * If VQ 0 has already been allocated, try to allocate on the same
1659          * NUMA node. It can be reallocated later in numa_realloc().
1660          */
1661         if (dev->nr_vring > 0)
1662                 numa_node = dev->virtqueue[0]->numa_node;
1663
1664         if (dev->inflight_info == NULL) {
1665                 dev->inflight_info = rte_zmalloc_socket("inflight_info",
1666                                 sizeof(struct inflight_mem_info), 0, numa_node);
1667                 if (!dev->inflight_info) {
1668                         VHOST_LOG_CONFIG(ERR,
1669                                 "failed to alloc dev inflight area\n");
1670                         return RTE_VHOST_MSG_RESULT_ERR;
1671                 }
1672                 dev->inflight_info->fd = -1;
1673         }
1674
1675         num_queues = msg->payload.inflight.num_queues;
1676         queue_size = msg->payload.inflight.queue_size;
1677
1678         VHOST_LOG_CONFIG(INFO, "get_inflight_fd num_queues: %u\n",
1679                 msg->payload.inflight.num_queues);
1680         VHOST_LOG_CONFIG(INFO, "get_inflight_fd queue_size: %u\n",
1681                 msg->payload.inflight.queue_size);
1682
1683         if (vq_is_packed(dev))
1684                 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1685         else
1686                 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1687
1688         mmap_size = num_queues * pervq_inflight_size;
1689         addr = inflight_mem_alloc("vhost-inflight", mmap_size, &fd);
1690         if (!addr) {
1691                 VHOST_LOG_CONFIG(ERR,
1692                         "failed to alloc vhost inflight area\n");
1693                         msg->payload.inflight.mmap_size = 0;
1694                 return RTE_VHOST_MSG_RESULT_ERR;
1695         }
1696         memset(addr, 0, mmap_size);
1697
1698         if (dev->inflight_info->addr) {
1699                 munmap(dev->inflight_info->addr, dev->inflight_info->size);
1700                 dev->inflight_info->addr = NULL;
1701         }
1702
1703         if (dev->inflight_info->fd >= 0) {
1704                 close(dev->inflight_info->fd);
1705                 dev->inflight_info->fd = -1;
1706         }
1707
1708         dev->inflight_info->addr = addr;
1709         dev->inflight_info->size = msg->payload.inflight.mmap_size = mmap_size;
1710         dev->inflight_info->fd = msg->fds[0] = fd;
1711         msg->payload.inflight.mmap_offset = 0;
1712         msg->fd_num = 1;
1713
1714         if (vq_is_packed(dev)) {
1715                 for (i = 0; i < num_queues; i++) {
1716                         inflight_packed =
1717                                 (struct rte_vhost_inflight_info_packed *)addr;
1718                         inflight_packed->used_wrap_counter = 1;
1719                         inflight_packed->old_used_wrap_counter = 1;
1720                         for (j = 0; j < queue_size; j++)
1721                                 inflight_packed->desc[j].next = j + 1;
1722                         addr = (void *)((char *)addr + pervq_inflight_size);
1723                 }
1724         }
1725
1726         VHOST_LOG_CONFIG(INFO,
1727                 "send inflight mmap_size: %"PRIu64"\n",
1728                 msg->payload.inflight.mmap_size);
1729         VHOST_LOG_CONFIG(INFO,
1730                 "send inflight mmap_offset: %"PRIu64"\n",
1731                 msg->payload.inflight.mmap_offset);
1732         VHOST_LOG_CONFIG(INFO,
1733                 "send inflight fd: %d\n", msg->fds[0]);
1734
1735         return RTE_VHOST_MSG_RESULT_REPLY;
1736 }
1737
1738 static int
1739 vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
1740                            int main_fd __rte_unused)
1741 {
1742         uint64_t mmap_size, mmap_offset;
1743         uint16_t num_queues, queue_size;
1744         struct virtio_net *dev = *pdev;
1745         uint32_t pervq_inflight_size;
1746         struct vhost_virtqueue *vq;
1747         void *addr;
1748         int fd, i;
1749         int numa_node = SOCKET_ID_ANY;
1750
1751         fd = msg->fds[0];
1752         if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
1753                 VHOST_LOG_CONFIG(ERR,
1754                         "invalid set_inflight_fd message size is %d,fd is %d\n",
1755                         msg->size, fd);
1756                 return RTE_VHOST_MSG_RESULT_ERR;
1757         }
1758
1759         mmap_size = msg->payload.inflight.mmap_size;
1760         mmap_offset = msg->payload.inflight.mmap_offset;
1761         num_queues = msg->payload.inflight.num_queues;
1762         queue_size = msg->payload.inflight.queue_size;
1763
1764         if (vq_is_packed(dev))
1765                 pervq_inflight_size = get_pervq_shm_size_packed(queue_size);
1766         else
1767                 pervq_inflight_size = get_pervq_shm_size_split(queue_size);
1768
1769         VHOST_LOG_CONFIG(INFO,
1770                 "set_inflight_fd mmap_size: %"PRIu64"\n", mmap_size);
1771         VHOST_LOG_CONFIG(INFO,
1772                 "set_inflight_fd mmap_offset: %"PRIu64"\n", mmap_offset);
1773         VHOST_LOG_CONFIG(INFO,
1774                 "set_inflight_fd num_queues: %u\n", num_queues);
1775         VHOST_LOG_CONFIG(INFO,
1776                 "set_inflight_fd queue_size: %u\n", queue_size);
1777         VHOST_LOG_CONFIG(INFO,
1778                 "set_inflight_fd fd: %d\n", fd);
1779         VHOST_LOG_CONFIG(INFO,
1780                 "set_inflight_fd pervq_inflight_size: %d\n",
1781                 pervq_inflight_size);
1782
1783         /*
1784          * If VQ 0 has already been allocated, try to allocate on the same
1785          * NUMA node. It can be reallocated later in numa_realloc().
1786          */
1787         if (dev->nr_vring > 0)
1788                 numa_node = dev->virtqueue[0]->numa_node;
1789
1790         if (!dev->inflight_info) {
1791                 dev->inflight_info = rte_zmalloc_socket("inflight_info",
1792                                 sizeof(struct inflight_mem_info), 0, numa_node);
1793                 if (dev->inflight_info == NULL) {
1794                         VHOST_LOG_CONFIG(ERR,
1795                                 "failed to alloc dev inflight area\n");
1796                         return RTE_VHOST_MSG_RESULT_ERR;
1797                 }
1798                 dev->inflight_info->fd = -1;
1799         }
1800
1801         if (dev->inflight_info->addr) {
1802                 munmap(dev->inflight_info->addr, dev->inflight_info->size);
1803                 dev->inflight_info->addr = NULL;
1804         }
1805
1806         addr = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
1807                     fd, mmap_offset);
1808         if (addr == MAP_FAILED) {
1809                 VHOST_LOG_CONFIG(ERR, "failed to mmap share memory.\n");
1810                 return RTE_VHOST_MSG_RESULT_ERR;
1811         }
1812
1813         if (dev->inflight_info->fd >= 0) {
1814                 close(dev->inflight_info->fd);
1815                 dev->inflight_info->fd = -1;
1816         }
1817
1818         dev->inflight_info->fd = fd;
1819         dev->inflight_info->addr = addr;
1820         dev->inflight_info->size = mmap_size;
1821
1822         for (i = 0; i < num_queues; i++) {
1823                 vq = dev->virtqueue[i];
1824                 if (!vq)
1825                         continue;
1826
1827                 if (vq_is_packed(dev)) {
1828                         vq->inflight_packed = addr;
1829                         vq->inflight_packed->desc_num = queue_size;
1830                 } else {
1831                         vq->inflight_split = addr;
1832                         vq->inflight_split->desc_num = queue_size;
1833                 }
1834                 addr = (void *)((char *)addr + pervq_inflight_size);
1835         }
1836
1837         return RTE_VHOST_MSG_RESULT_OK;
1838 }
1839
1840 static int
1841 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1842                         int main_fd __rte_unused)
1843 {
1844         struct virtio_net *dev = *pdev;
1845         struct vhost_vring_file file;
1846         struct vhost_virtqueue *vq;
1847         int expected_fds;
1848
1849         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1850         if (validate_msg_fds(msg, expected_fds) != 0)
1851                 return RTE_VHOST_MSG_RESULT_ERR;
1852
1853         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1854         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1855                 file.fd = VIRTIO_INVALID_EVENTFD;
1856         else
1857                 file.fd = msg->fds[0];
1858         VHOST_LOG_CONFIG(INFO,
1859                 "vring call idx:%d file:%d\n", file.index, file.fd);
1860
1861         vq = dev->virtqueue[file.index];
1862
1863         if (vq->ready) {
1864                 vq->ready = false;
1865                 vhost_user_notify_queue_state(dev, file.index, 0);
1866         }
1867
1868         if (vq->callfd >= 0)
1869                 close(vq->callfd);
1870
1871         vq->callfd = file.fd;
1872
1873         return RTE_VHOST_MSG_RESULT_OK;
1874 }
1875
1876 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1877                         struct VhostUserMsg *msg,
1878                         int main_fd __rte_unused)
1879 {
1880         int expected_fds;
1881
1882         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
1883         if (validate_msg_fds(msg, expected_fds) != 0)
1884                 return RTE_VHOST_MSG_RESULT_ERR;
1885
1886         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1887                 close(msg->fds[0]);
1888         VHOST_LOG_CONFIG(INFO, "not implemented\n");
1889
1890         return RTE_VHOST_MSG_RESULT_OK;
1891 }
1892
1893 static int
1894 resubmit_desc_compare(const void *a, const void *b)
1895 {
1896         const struct rte_vhost_resubmit_desc *desc0 = a;
1897         const struct rte_vhost_resubmit_desc *desc1 = b;
1898
1899         if (desc1->counter > desc0->counter)
1900                 return 1;
1901
1902         return -1;
1903 }
1904
1905 static int
1906 vhost_check_queue_inflights_split(struct virtio_net *dev,
1907                                   struct vhost_virtqueue *vq)
1908 {
1909         uint16_t i;
1910         uint16_t resubmit_num = 0, last_io, num;
1911         struct vring_used *used = vq->used;
1912         struct rte_vhost_resubmit_info *resubmit;
1913         struct rte_vhost_inflight_info_split *inflight_split;
1914
1915         if (!(dev->protocol_features &
1916             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
1917                 return RTE_VHOST_MSG_RESULT_OK;
1918
1919         /* The frontend may still not support the inflight feature
1920          * although we negotiate the protocol feature.
1921          */
1922         if ((!vq->inflight_split))
1923                 return RTE_VHOST_MSG_RESULT_OK;
1924
1925         if (!vq->inflight_split->version) {
1926                 vq->inflight_split->version = INFLIGHT_VERSION;
1927                 return RTE_VHOST_MSG_RESULT_OK;
1928         }
1929
1930         if (vq->resubmit_inflight)
1931                 return RTE_VHOST_MSG_RESULT_OK;
1932
1933         inflight_split = vq->inflight_split;
1934         vq->global_counter = 0;
1935         last_io = inflight_split->last_inflight_io;
1936
1937         if (inflight_split->used_idx != used->idx) {
1938                 inflight_split->desc[last_io].inflight = 0;
1939                 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1940                 inflight_split->used_idx = used->idx;
1941         }
1942
1943         for (i = 0; i < inflight_split->desc_num; i++) {
1944                 if (inflight_split->desc[i].inflight == 1)
1945                         resubmit_num++;
1946         }
1947
1948         vq->last_avail_idx += resubmit_num;
1949
1950         if (resubmit_num) {
1951                 resubmit = rte_zmalloc_socket("resubmit", sizeof(struct rte_vhost_resubmit_info),
1952                                 0, vq->numa_node);
1953                 if (!resubmit) {
1954                         VHOST_LOG_CONFIG(ERR,
1955                                 "failed to allocate memory for resubmit info.\n");
1956                         return RTE_VHOST_MSG_RESULT_ERR;
1957                 }
1958
1959                 resubmit->resubmit_list = rte_zmalloc_socket("resubmit_list",
1960                                 resubmit_num * sizeof(struct rte_vhost_resubmit_desc),
1961                                 0, vq->numa_node);
1962                 if (!resubmit->resubmit_list) {
1963                         VHOST_LOG_CONFIG(ERR,
1964                                 "failed to allocate memory for inflight desc.\n");
1965                         rte_free(resubmit);
1966                         return RTE_VHOST_MSG_RESULT_ERR;
1967                 }
1968
1969                 num = 0;
1970                 for (i = 0; i < vq->inflight_split->desc_num; i++) {
1971                         if (vq->inflight_split->desc[i].inflight == 1) {
1972                                 resubmit->resubmit_list[num].index = i;
1973                                 resubmit->resubmit_list[num].counter =
1974                                         inflight_split->desc[i].counter;
1975                                 num++;
1976                         }
1977                 }
1978                 resubmit->resubmit_num = num;
1979
1980                 if (resubmit->resubmit_num > 1)
1981                         qsort(resubmit->resubmit_list, resubmit->resubmit_num,
1982                               sizeof(struct rte_vhost_resubmit_desc),
1983                               resubmit_desc_compare);
1984
1985                 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
1986                 vq->resubmit_inflight = resubmit;
1987         }
1988
1989         return RTE_VHOST_MSG_RESULT_OK;
1990 }
1991
1992 static int
1993 vhost_check_queue_inflights_packed(struct virtio_net *dev,
1994                                    struct vhost_virtqueue *vq)
1995 {
1996         uint16_t i;
1997         uint16_t resubmit_num = 0, old_used_idx, num;
1998         struct rte_vhost_resubmit_info *resubmit;
1999         struct rte_vhost_inflight_info_packed *inflight_packed;
2000
2001         if (!(dev->protocol_features &
2002             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
2003                 return RTE_VHOST_MSG_RESULT_OK;
2004
2005         /* The frontend may still not support the inflight feature
2006          * although we negotiate the protocol feature.
2007          */
2008         if ((!vq->inflight_packed))
2009                 return RTE_VHOST_MSG_RESULT_OK;
2010
2011         if (!vq->inflight_packed->version) {
2012                 vq->inflight_packed->version = INFLIGHT_VERSION;
2013                 return RTE_VHOST_MSG_RESULT_OK;
2014         }
2015
2016         if (vq->resubmit_inflight)
2017                 return RTE_VHOST_MSG_RESULT_OK;
2018
2019         inflight_packed = vq->inflight_packed;
2020         vq->global_counter = 0;
2021         old_used_idx = inflight_packed->old_used_idx;
2022
2023         if (inflight_packed->used_idx != old_used_idx) {
2024                 if (inflight_packed->desc[old_used_idx].inflight == 0) {
2025                         inflight_packed->old_used_idx =
2026                                 inflight_packed->used_idx;
2027                         inflight_packed->old_used_wrap_counter =
2028                                 inflight_packed->used_wrap_counter;
2029                         inflight_packed->old_free_head =
2030                                 inflight_packed->free_head;
2031                 } else {
2032                         inflight_packed->used_idx =
2033                                 inflight_packed->old_used_idx;
2034                         inflight_packed->used_wrap_counter =
2035                                 inflight_packed->old_used_wrap_counter;
2036                         inflight_packed->free_head =
2037                                 inflight_packed->old_free_head;
2038                 }
2039         }
2040
2041         for (i = 0; i < inflight_packed->desc_num; i++) {
2042                 if (inflight_packed->desc[i].inflight == 1)
2043                         resubmit_num++;
2044         }
2045
2046         if (resubmit_num) {
2047                 resubmit = rte_zmalloc_socket("resubmit", sizeof(struct rte_vhost_resubmit_info),
2048                                 0, vq->numa_node);
2049                 if (resubmit == NULL) {
2050                         VHOST_LOG_CONFIG(ERR,
2051                                 "failed to allocate memory for resubmit info.\n");
2052                         return RTE_VHOST_MSG_RESULT_ERR;
2053                 }
2054
2055                 resubmit->resubmit_list = rte_zmalloc_socket("resubmit_list",
2056                                 resubmit_num * sizeof(struct rte_vhost_resubmit_desc),
2057                                 0, vq->numa_node);
2058                 if (resubmit->resubmit_list == NULL) {
2059                         VHOST_LOG_CONFIG(ERR,
2060                                 "failed to allocate memory for resubmit desc.\n");
2061                         rte_free(resubmit);
2062                         return RTE_VHOST_MSG_RESULT_ERR;
2063                 }
2064
2065                 num = 0;
2066                 for (i = 0; i < inflight_packed->desc_num; i++) {
2067                         if (vq->inflight_packed->desc[i].inflight == 1) {
2068                                 resubmit->resubmit_list[num].index = i;
2069                                 resubmit->resubmit_list[num].counter =
2070                                         inflight_packed->desc[i].counter;
2071                                 num++;
2072                         }
2073                 }
2074                 resubmit->resubmit_num = num;
2075
2076                 if (resubmit->resubmit_num > 1)
2077                         qsort(resubmit->resubmit_list, resubmit->resubmit_num,
2078                               sizeof(struct rte_vhost_resubmit_desc),
2079                               resubmit_desc_compare);
2080
2081                 vq->global_counter = resubmit->resubmit_list[0].counter + 1;
2082                 vq->resubmit_inflight = resubmit;
2083         }
2084
2085         return RTE_VHOST_MSG_RESULT_OK;
2086 }
2087
2088 static int
2089 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
2090                         int main_fd __rte_unused)
2091 {
2092         struct virtio_net *dev = *pdev;
2093         struct vhost_vring_file file;
2094         struct vhost_virtqueue *vq;
2095         int expected_fds;
2096
2097         expected_fds = (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK) ? 0 : 1;
2098         if (validate_msg_fds(msg, expected_fds) != 0)
2099                 return RTE_VHOST_MSG_RESULT_ERR;
2100
2101         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
2102         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
2103                 file.fd = VIRTIO_INVALID_EVENTFD;
2104         else
2105                 file.fd = msg->fds[0];
2106         VHOST_LOG_CONFIG(INFO,
2107                 "vring kick idx:%d file:%d\n", file.index, file.fd);
2108
2109         /* Interpret ring addresses only when ring is started. */
2110         dev = translate_ring_addresses(dev, file.index);
2111         if (!dev) {
2112                 if (file.fd != VIRTIO_INVALID_EVENTFD)
2113                         close(file.fd);
2114
2115                 return RTE_VHOST_MSG_RESULT_ERR;
2116         }
2117
2118         *pdev = dev;
2119
2120         vq = dev->virtqueue[file.index];
2121
2122         /*
2123          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
2124          * the ring starts already enabled. Otherwise, it is enabled via
2125          * the SET_VRING_ENABLE message.
2126          */
2127         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
2128                 vq->enabled = true;
2129         }
2130
2131         if (vq->ready) {
2132                 vq->ready = false;
2133                 vhost_user_notify_queue_state(dev, file.index, 0);
2134         }
2135
2136         if (vq->kickfd >= 0)
2137                 close(vq->kickfd);
2138         vq->kickfd = file.fd;
2139
2140         if (vq_is_packed(dev)) {
2141                 if (vhost_check_queue_inflights_packed(dev, vq)) {
2142                         VHOST_LOG_CONFIG(ERR,
2143                                 "failed to inflights for vq: %d\n", file.index);
2144                         return RTE_VHOST_MSG_RESULT_ERR;
2145                 }
2146         } else {
2147                 if (vhost_check_queue_inflights_split(dev, vq)) {
2148                         VHOST_LOG_CONFIG(ERR,
2149                                 "failed to inflights for vq: %d\n", file.index);
2150                         return RTE_VHOST_MSG_RESULT_ERR;
2151                 }
2152         }
2153
2154         return RTE_VHOST_MSG_RESULT_OK;
2155 }
2156
2157 /*
2158  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
2159  */
2160 static int
2161 vhost_user_get_vring_base(struct virtio_net **pdev,
2162                         struct VhostUserMsg *msg,
2163                         int main_fd __rte_unused)
2164 {
2165         struct virtio_net *dev = *pdev;
2166         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
2167         uint64_t val;
2168
2169         if (validate_msg_fds(msg, 0) != 0)
2170                 return RTE_VHOST_MSG_RESULT_ERR;
2171
2172         /* We have to stop the queue (virtio) if it is running. */
2173         vhost_destroy_device_notify(dev);
2174
2175         dev->flags &= ~VIRTIO_DEV_READY;
2176         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
2177
2178         /* Here we are safe to get the indexes */
2179         if (vq_is_packed(dev)) {
2180                 /*
2181                  * Bit[0:14]: avail index
2182                  * Bit[15]: avail wrap counter
2183                  */
2184                 val = vq->last_avail_idx & 0x7fff;
2185                 val |= vq->avail_wrap_counter << 15;
2186                 msg->payload.state.num = val;
2187         } else {
2188                 msg->payload.state.num = vq->last_avail_idx;
2189         }
2190
2191         VHOST_LOG_CONFIG(INFO,
2192                 "vring base idx:%d file:%d\n", msg->payload.state.index,
2193                 msg->payload.state.num);
2194         /*
2195          * Based on current qemu vhost-user implementation, this message is
2196          * sent and only sent in vhost_vring_stop.
2197          * TODO: cleanup the vring, it isn't usable since here.
2198          */
2199         if (vq->kickfd >= 0)
2200                 close(vq->kickfd);
2201
2202         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
2203
2204         if (vq->callfd >= 0)
2205                 close(vq->callfd);
2206
2207         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
2208
2209         vq->signalled_used_valid = false;
2210
2211         if (vq_is_packed(dev)) {
2212                 rte_free(vq->shadow_used_packed);
2213                 vq->shadow_used_packed = NULL;
2214         } else {
2215                 rte_free(vq->shadow_used_split);
2216                 vq->shadow_used_split = NULL;
2217         }
2218
2219         rte_free(vq->batch_copy_elems);
2220         vq->batch_copy_elems = NULL;
2221
2222         rte_free(vq->log_cache);
2223         vq->log_cache = NULL;
2224
2225         msg->size = sizeof(msg->payload.state);
2226         msg->fd_num = 0;
2227
2228         vhost_user_iotlb_flush_all(vq);
2229
2230         vring_invalidate(dev, vq);
2231
2232         return RTE_VHOST_MSG_RESULT_REPLY;
2233 }
2234
2235 /*
2236  * when virtio queues are ready to work, qemu will send us to
2237  * enable the virtio queue pair.
2238  */
2239 static int
2240 vhost_user_set_vring_enable(struct virtio_net **pdev,
2241                         struct VhostUserMsg *msg,
2242                         int main_fd __rte_unused)
2243 {
2244         struct virtio_net *dev = *pdev;
2245         bool enable = !!msg->payload.state.num;
2246         int index = (int)msg->payload.state.index;
2247
2248         if (validate_msg_fds(msg, 0) != 0)
2249                 return RTE_VHOST_MSG_RESULT_ERR;
2250
2251         VHOST_LOG_CONFIG(INFO,
2252                 "set queue enable: %d to qp idx: %d\n",
2253                 enable, index);
2254
2255         if (enable && dev->virtqueue[index]->async_registered) {
2256                 if (dev->virtqueue[index]->async_pkts_inflight_n) {
2257                         VHOST_LOG_CONFIG(ERR, "failed to enable vring. "
2258                         "async inflight packets must be completed first\n");
2259                         return RTE_VHOST_MSG_RESULT_ERR;
2260                 }
2261         }
2262
2263         dev->virtqueue[index]->enabled = enable;
2264
2265         return RTE_VHOST_MSG_RESULT_OK;
2266 }
2267
2268 static int
2269 vhost_user_get_protocol_features(struct virtio_net **pdev,
2270                         struct VhostUserMsg *msg,
2271                         int main_fd __rte_unused)
2272 {
2273         struct virtio_net *dev = *pdev;
2274         uint64_t features, protocol_features;
2275
2276         if (validate_msg_fds(msg, 0) != 0)
2277                 return RTE_VHOST_MSG_RESULT_ERR;
2278
2279         rte_vhost_driver_get_features(dev->ifname, &features);
2280         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
2281
2282         msg->payload.u64 = protocol_features;
2283         msg->size = sizeof(msg->payload.u64);
2284         msg->fd_num = 0;
2285
2286         return RTE_VHOST_MSG_RESULT_REPLY;
2287 }
2288
2289 static int
2290 vhost_user_set_protocol_features(struct virtio_net **pdev,
2291                         struct VhostUserMsg *msg,
2292                         int main_fd __rte_unused)
2293 {
2294         struct virtio_net *dev = *pdev;
2295         uint64_t protocol_features = msg->payload.u64;
2296         uint64_t slave_protocol_features = 0;
2297
2298         if (validate_msg_fds(msg, 0) != 0)
2299                 return RTE_VHOST_MSG_RESULT_ERR;
2300
2301         rte_vhost_driver_get_protocol_features(dev->ifname,
2302                         &slave_protocol_features);
2303         if (protocol_features & ~slave_protocol_features) {
2304                 VHOST_LOG_CONFIG(ERR,
2305                         "(%d) received invalid protocol features.\n",
2306                         dev->vid);
2307                 return RTE_VHOST_MSG_RESULT_ERR;
2308         }
2309
2310         dev->protocol_features = protocol_features;
2311         VHOST_LOG_CONFIG(INFO,
2312                 "negotiated Vhost-user protocol features: 0x%" PRIx64 "\n",
2313                 dev->protocol_features);
2314
2315         return RTE_VHOST_MSG_RESULT_OK;
2316 }
2317
2318 static int
2319 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
2320                         int main_fd __rte_unused)
2321 {
2322         struct virtio_net *dev = *pdev;
2323         int fd = msg->fds[0];
2324         uint64_t size, off;
2325         void *addr;
2326         uint32_t i;
2327
2328         if (validate_msg_fds(msg, 1) != 0)
2329                 return RTE_VHOST_MSG_RESULT_ERR;
2330
2331         if (fd < 0) {
2332                 VHOST_LOG_CONFIG(ERR, "invalid log fd: %d\n", fd);
2333                 return RTE_VHOST_MSG_RESULT_ERR;
2334         }
2335
2336         if (msg->size != sizeof(VhostUserLog)) {
2337                 VHOST_LOG_CONFIG(ERR,
2338                         "invalid log base msg size: %"PRId32" != %d\n",
2339                         msg->size, (int)sizeof(VhostUserLog));
2340                 goto close_msg_fds;
2341         }
2342
2343         size = msg->payload.log.mmap_size;
2344         off  = msg->payload.log.mmap_offset;
2345
2346         /* Check for mmap size and offset overflow. */
2347         if (off >= -size) {
2348                 VHOST_LOG_CONFIG(ERR,
2349                         "log offset %#"PRIx64" and log size %#"PRIx64" overflow\n",
2350                         off, size);
2351                 goto close_msg_fds;
2352         }
2353
2354         VHOST_LOG_CONFIG(INFO,
2355                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
2356                 size, off);
2357
2358         /*
2359          * mmap from 0 to workaround a hugepage mmap bug: mmap will
2360          * fail when offset is not page size aligned.
2361          */
2362         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
2363         close(fd);
2364         if (addr == MAP_FAILED) {
2365                 VHOST_LOG_CONFIG(ERR, "mmap log base failed!\n");
2366                 return RTE_VHOST_MSG_RESULT_ERR;
2367         }
2368
2369         /*
2370          * Free previously mapped log memory on occasionally
2371          * multiple VHOST_USER_SET_LOG_BASE.
2372          */
2373         if (dev->log_addr) {
2374                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
2375         }
2376         dev->log_addr = (uint64_t)(uintptr_t)addr;
2377         dev->log_base = dev->log_addr + off;
2378         dev->log_size = size;
2379
2380         for (i = 0; i < dev->nr_vring; i++) {
2381                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2382
2383                 rte_free(vq->log_cache);
2384                 vq->log_cache = NULL;
2385                 vq->log_cache_nb_elem = 0;
2386                 vq->log_cache = rte_malloc_socket("vq log cache",
2387                                 sizeof(struct log_cache_entry) * VHOST_LOG_CACHE_NR,
2388                                 0, vq->numa_node);
2389                 /*
2390                  * If log cache alloc fail, don't fail migration, but no
2391                  * caching will be done, which will impact performance
2392                  */
2393                 if (!vq->log_cache)
2394                         VHOST_LOG_CONFIG(ERR, "Failed to allocate VQ logging cache\n");
2395         }
2396
2397         /*
2398          * The spec is not clear about it (yet), but QEMU doesn't expect
2399          * any payload in the reply.
2400          */
2401         msg->size = 0;
2402         msg->fd_num = 0;
2403
2404         return RTE_VHOST_MSG_RESULT_REPLY;
2405
2406 close_msg_fds:
2407         close_msg_fds(msg);
2408         return RTE_VHOST_MSG_RESULT_ERR;
2409 }
2410
2411 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
2412                         struct VhostUserMsg *msg,
2413                         int main_fd __rte_unused)
2414 {
2415         if (validate_msg_fds(msg, 1) != 0)
2416                 return RTE_VHOST_MSG_RESULT_ERR;
2417
2418         close(msg->fds[0]);
2419         VHOST_LOG_CONFIG(INFO, "not implemented.\n");
2420
2421         return RTE_VHOST_MSG_RESULT_OK;
2422 }
2423
2424 /*
2425  * An rarp packet is constructed and broadcasted to notify switches about
2426  * the new location of the migrated VM, so that packets from outside will
2427  * not be lost after migration.
2428  *
2429  * However, we don't actually "send" a rarp packet here, instead, we set
2430  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
2431  */
2432 static int
2433 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
2434                         int main_fd __rte_unused)
2435 {
2436         struct virtio_net *dev = *pdev;
2437         uint8_t *mac = (uint8_t *)&msg->payload.u64;
2438         struct rte_vdpa_device *vdpa_dev;
2439
2440         if (validate_msg_fds(msg, 0) != 0)
2441                 return RTE_VHOST_MSG_RESULT_ERR;
2442
2443         VHOST_LOG_CONFIG(DEBUG,
2444                 ":: mac: " RTE_ETHER_ADDR_PRT_FMT "\n",
2445                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
2446         memcpy(dev->mac.addr_bytes, mac, 6);
2447
2448         /*
2449          * Set the flag to inject a RARP broadcast packet at
2450          * rte_vhost_dequeue_burst().
2451          *
2452          * __ATOMIC_RELEASE ordering is for making sure the mac is
2453          * copied before the flag is set.
2454          */
2455         __atomic_store_n(&dev->broadcast_rarp, 1, __ATOMIC_RELEASE);
2456         vdpa_dev = dev->vdpa_dev;
2457         if (vdpa_dev && vdpa_dev->ops->migration_done)
2458                 vdpa_dev->ops->migration_done(dev->vid);
2459
2460         return RTE_VHOST_MSG_RESULT_OK;
2461 }
2462
2463 static int
2464 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
2465                         int main_fd __rte_unused)
2466 {
2467         struct virtio_net *dev = *pdev;
2468
2469         if (validate_msg_fds(msg, 0) != 0)
2470                 return RTE_VHOST_MSG_RESULT_ERR;
2471
2472         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
2473                         msg->payload.u64 > VIRTIO_MAX_MTU) {
2474                 VHOST_LOG_CONFIG(ERR, "Invalid MTU size (%"PRIu64")\n",
2475                                 msg->payload.u64);
2476
2477                 return RTE_VHOST_MSG_RESULT_ERR;
2478         }
2479
2480         dev->mtu = msg->payload.u64;
2481
2482         return RTE_VHOST_MSG_RESULT_OK;
2483 }
2484
2485 static int
2486 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
2487                         int main_fd __rte_unused)
2488 {
2489         struct virtio_net *dev = *pdev;
2490         int fd = msg->fds[0];
2491
2492         if (validate_msg_fds(msg, 1) != 0)
2493                 return RTE_VHOST_MSG_RESULT_ERR;
2494
2495         if (fd < 0) {
2496                 VHOST_LOG_CONFIG(ERR,
2497                                 "Invalid file descriptor for slave channel (%d)\n",
2498                                 fd);
2499                 return RTE_VHOST_MSG_RESULT_ERR;
2500         }
2501
2502         if (dev->slave_req_fd >= 0)
2503                 close(dev->slave_req_fd);
2504
2505         dev->slave_req_fd = fd;
2506
2507         return RTE_VHOST_MSG_RESULT_OK;
2508 }
2509
2510 static int
2511 is_vring_iotlb_split(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2512 {
2513         struct vhost_vring_addr *ra;
2514         uint64_t start, end, len;
2515
2516         start = imsg->iova;
2517         end = start + imsg->size;
2518
2519         ra = &vq->ring_addrs;
2520         len = sizeof(struct vring_desc) * vq->size;
2521         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2522                 return 1;
2523
2524         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
2525         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2526                 return 1;
2527
2528         len = sizeof(struct vring_used) +
2529                sizeof(struct vring_used_elem) * vq->size;
2530         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2531                 return 1;
2532
2533         if (ra->flags & (1 << VHOST_VRING_F_LOG)) {
2534                 len = sizeof(uint64_t);
2535                 if (ra->log_guest_addr < end &&
2536                     (ra->log_guest_addr + len) > start)
2537                         return 1;
2538         }
2539
2540         return 0;
2541 }
2542
2543 static int
2544 is_vring_iotlb_packed(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
2545 {
2546         struct vhost_vring_addr *ra;
2547         uint64_t start, end, len;
2548
2549         start = imsg->iova;
2550         end = start + imsg->size;
2551
2552         ra = &vq->ring_addrs;
2553         len = sizeof(struct vring_packed_desc) * vq->size;
2554         if (ra->desc_user_addr < end && (ra->desc_user_addr + len) > start)
2555                 return 1;
2556
2557         len = sizeof(struct vring_packed_desc_event);
2558         if (ra->avail_user_addr < end && (ra->avail_user_addr + len) > start)
2559                 return 1;
2560
2561         len = sizeof(struct vring_packed_desc_event);
2562         if (ra->used_user_addr < end && (ra->used_user_addr + len) > start)
2563                 return 1;
2564
2565         if (ra->flags & (1 << VHOST_VRING_F_LOG)) {
2566                 len = sizeof(uint64_t);
2567                 if (ra->log_guest_addr < end &&
2568                     (ra->log_guest_addr + len) > start)
2569                         return 1;
2570         }
2571
2572         return 0;
2573 }
2574
2575 static int is_vring_iotlb(struct virtio_net *dev,
2576                           struct vhost_virtqueue *vq,
2577                           struct vhost_iotlb_msg *imsg)
2578 {
2579         if (vq_is_packed(dev))
2580                 return is_vring_iotlb_packed(vq, imsg);
2581         else
2582                 return is_vring_iotlb_split(vq, imsg);
2583 }
2584
2585 static int
2586 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
2587                         int main_fd __rte_unused)
2588 {
2589         struct virtio_net *dev = *pdev;
2590         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
2591         uint16_t i;
2592         uint64_t vva, len;
2593
2594         if (validate_msg_fds(msg, 0) != 0)
2595                 return RTE_VHOST_MSG_RESULT_ERR;
2596
2597         switch (imsg->type) {
2598         case VHOST_IOTLB_UPDATE:
2599                 len = imsg->size;
2600                 vva = qva_to_vva(dev, imsg->uaddr, &len);
2601                 if (!vva)
2602                         return RTE_VHOST_MSG_RESULT_ERR;
2603
2604                 for (i = 0; i < dev->nr_vring; i++) {
2605                         struct vhost_virtqueue *vq = dev->virtqueue[i];
2606
2607                         if (!vq)
2608                                 continue;
2609
2610                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
2611                                         len, imsg->perm);
2612
2613                         if (is_vring_iotlb(dev, vq, imsg))
2614                                 *pdev = dev = translate_ring_addresses(dev, i);
2615                 }
2616                 break;
2617         case VHOST_IOTLB_INVALIDATE:
2618                 for (i = 0; i < dev->nr_vring; i++) {
2619                         struct vhost_virtqueue *vq = dev->virtqueue[i];
2620
2621                         if (!vq)
2622                                 continue;
2623
2624                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
2625                                         imsg->size);
2626
2627                         if (is_vring_iotlb(dev, vq, imsg))
2628                                 vring_invalidate(dev, vq);
2629                 }
2630                 break;
2631         default:
2632                 VHOST_LOG_CONFIG(ERR, "Invalid IOTLB message type (%d)\n",
2633                                 imsg->type);
2634                 return RTE_VHOST_MSG_RESULT_ERR;
2635         }
2636
2637         return RTE_VHOST_MSG_RESULT_OK;
2638 }
2639
2640 static int
2641 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
2642                         struct VhostUserMsg *msg,
2643                         int main_fd __rte_unused)
2644 {
2645         struct virtio_net *dev = *pdev;
2646 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
2647         struct uffdio_api api_struct;
2648
2649         if (validate_msg_fds(msg, 0) != 0)
2650                 return RTE_VHOST_MSG_RESULT_ERR;
2651
2652         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
2653
2654         if (dev->postcopy_ufd == -1) {
2655                 VHOST_LOG_CONFIG(ERR, "Userfaultfd not available: %s\n",
2656                         strerror(errno));
2657                 return RTE_VHOST_MSG_RESULT_ERR;
2658         }
2659         api_struct.api = UFFD_API;
2660         api_struct.features = 0;
2661         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
2662                 VHOST_LOG_CONFIG(ERR, "UFFDIO_API ioctl failure: %s\n",
2663                         strerror(errno));
2664                 close(dev->postcopy_ufd);
2665                 dev->postcopy_ufd = -1;
2666                 return RTE_VHOST_MSG_RESULT_ERR;
2667         }
2668         msg->fds[0] = dev->postcopy_ufd;
2669         msg->fd_num = 1;
2670
2671         return RTE_VHOST_MSG_RESULT_REPLY;
2672 #else
2673         dev->postcopy_ufd = -1;
2674         msg->fd_num = 0;
2675
2676         return RTE_VHOST_MSG_RESULT_ERR;
2677 #endif
2678 }
2679
2680 static int
2681 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
2682                         struct VhostUserMsg *msg __rte_unused,
2683                         int main_fd __rte_unused)
2684 {
2685         struct virtio_net *dev = *pdev;
2686
2687         if (validate_msg_fds(msg, 0) != 0)
2688                 return RTE_VHOST_MSG_RESULT_ERR;
2689
2690         if (dev->mem && dev->mem->nregions) {
2691                 VHOST_LOG_CONFIG(ERR,
2692                         "Regions already registered at postcopy-listen\n");
2693                 return RTE_VHOST_MSG_RESULT_ERR;
2694         }
2695         dev->postcopy_listening = 1;
2696
2697         return RTE_VHOST_MSG_RESULT_OK;
2698 }
2699
2700 static int
2701 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
2702                         int main_fd __rte_unused)
2703 {
2704         struct virtio_net *dev = *pdev;
2705
2706         if (validate_msg_fds(msg, 0) != 0)
2707                 return RTE_VHOST_MSG_RESULT_ERR;
2708
2709         dev->postcopy_listening = 0;
2710         if (dev->postcopy_ufd >= 0) {
2711                 close(dev->postcopy_ufd);
2712                 dev->postcopy_ufd = -1;
2713         }
2714
2715         msg->payload.u64 = 0;
2716         msg->size = sizeof(msg->payload.u64);
2717         msg->fd_num = 0;
2718
2719         return RTE_VHOST_MSG_RESULT_REPLY;
2720 }
2721
2722 static int
2723 vhost_user_get_status(struct virtio_net **pdev, struct VhostUserMsg *msg,
2724                       int main_fd __rte_unused)
2725 {
2726         struct virtio_net *dev = *pdev;
2727
2728         if (validate_msg_fds(msg, 0) != 0)
2729                 return RTE_VHOST_MSG_RESULT_ERR;
2730
2731         msg->payload.u64 = dev->status;
2732         msg->size = sizeof(msg->payload.u64);
2733         msg->fd_num = 0;
2734
2735         return RTE_VHOST_MSG_RESULT_REPLY;
2736 }
2737
2738 static int
2739 vhost_user_set_status(struct virtio_net **pdev, struct VhostUserMsg *msg,
2740                         int main_fd __rte_unused)
2741 {
2742         struct virtio_net *dev = *pdev;
2743
2744         if (validate_msg_fds(msg, 0) != 0)
2745                 return RTE_VHOST_MSG_RESULT_ERR;
2746
2747         /* As per Virtio specification, the device status is 8bits long */
2748         if (msg->payload.u64 > UINT8_MAX) {
2749                 VHOST_LOG_CONFIG(ERR, "Invalid VHOST_USER_SET_STATUS payload 0x%" PRIx64 "\n",
2750                                 msg->payload.u64);
2751                 return RTE_VHOST_MSG_RESULT_ERR;
2752         }
2753
2754         dev->status = msg->payload.u64;
2755
2756         if ((dev->status & VIRTIO_DEVICE_STATUS_FEATURES_OK) &&
2757             (dev->flags & VIRTIO_DEV_FEATURES_FAILED)) {
2758                 VHOST_LOG_CONFIG(ERR, "FEATURES_OK bit is set but feature negotiation failed\n");
2759                 /*
2760                  * Clear the bit to let the driver know about the feature
2761                  * negotiation failure
2762                  */
2763                 dev->status &= ~VIRTIO_DEVICE_STATUS_FEATURES_OK;
2764         }
2765
2766         VHOST_LOG_CONFIG(INFO, "New device status(0x%08x):\n"
2767                         "\t-RESET: %u\n"
2768                         "\t-ACKNOWLEDGE: %u\n"
2769                         "\t-DRIVER: %u\n"
2770                         "\t-FEATURES_OK: %u\n"
2771                         "\t-DRIVER_OK: %u\n"
2772                         "\t-DEVICE_NEED_RESET: %u\n"
2773                         "\t-FAILED: %u\n",
2774                         dev->status,
2775                         (dev->status == VIRTIO_DEVICE_STATUS_RESET),
2776                         !!(dev->status & VIRTIO_DEVICE_STATUS_ACK),
2777                         !!(dev->status & VIRTIO_DEVICE_STATUS_DRIVER),
2778                         !!(dev->status & VIRTIO_DEVICE_STATUS_FEATURES_OK),
2779                         !!(dev->status & VIRTIO_DEVICE_STATUS_DRIVER_OK),
2780                         !!(dev->status & VIRTIO_DEVICE_STATUS_DEV_NEED_RESET),
2781                         !!(dev->status & VIRTIO_DEVICE_STATUS_FAILED));
2782
2783         return RTE_VHOST_MSG_RESULT_OK;
2784 }
2785
2786 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
2787                                         struct VhostUserMsg *msg,
2788                                         int main_fd);
2789 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
2790         [VHOST_USER_NONE] = NULL,
2791         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
2792         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
2793         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
2794         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
2795         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
2796         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
2797         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
2798         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
2799         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
2800         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
2801         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
2802         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
2803         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
2804         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
2805         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
2806         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
2807         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
2808         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
2809         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
2810         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
2811         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
2812         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
2813         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
2814         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
2815         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
2816         [VHOST_USER_GET_INFLIGHT_FD] = vhost_user_get_inflight_fd,
2817         [VHOST_USER_SET_INFLIGHT_FD] = vhost_user_set_inflight_fd,
2818         [VHOST_USER_SET_STATUS] = vhost_user_set_status,
2819         [VHOST_USER_GET_STATUS] = vhost_user_get_status,
2820 };
2821
2822 /* return bytes# of read on success or negative val on failure. */
2823 static int
2824 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
2825 {
2826         int ret;
2827
2828         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
2829                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
2830         if (ret <= 0) {
2831                 return ret;
2832         } else if (ret != VHOST_USER_HDR_SIZE) {
2833                 VHOST_LOG_CONFIG(ERR, "Unexpected header size read\n");
2834                 close_msg_fds(msg);
2835                 return -1;
2836         }
2837
2838         if (msg->size) {
2839                 if (msg->size > sizeof(msg->payload)) {
2840                         VHOST_LOG_CONFIG(ERR,
2841                                 "invalid msg size: %d\n", msg->size);
2842                         return -1;
2843                 }
2844                 ret = read(sockfd, &msg->payload, msg->size);
2845                 if (ret <= 0)
2846                         return ret;
2847                 if (ret != (int)msg->size) {
2848                         VHOST_LOG_CONFIG(ERR,
2849                                 "read control message failed\n");
2850                         return -1;
2851                 }
2852         }
2853
2854         return ret;
2855 }
2856
2857 static int
2858 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
2859 {
2860         if (!msg)
2861                 return 0;
2862
2863         return send_fd_message(sockfd, (char *)msg,
2864                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
2865 }
2866
2867 static int
2868 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
2869 {
2870         if (!msg)
2871                 return 0;
2872
2873         msg->flags &= ~VHOST_USER_VERSION_MASK;
2874         msg->flags &= ~VHOST_USER_NEED_REPLY;
2875         msg->flags |= VHOST_USER_VERSION;
2876         msg->flags |= VHOST_USER_REPLY_MASK;
2877
2878         return send_vhost_message(sockfd, msg);
2879 }
2880
2881 static int
2882 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
2883 {
2884         int ret;
2885
2886         if (msg->flags & VHOST_USER_NEED_REPLY)
2887                 rte_spinlock_lock(&dev->slave_req_lock);
2888
2889         ret = send_vhost_message(dev->slave_req_fd, msg);
2890         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
2891                 rte_spinlock_unlock(&dev->slave_req_lock);
2892
2893         return ret;
2894 }
2895
2896 /*
2897  * Allocate a queue pair if it hasn't been allocated yet
2898  */
2899 static int
2900 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
2901                         struct VhostUserMsg *msg)
2902 {
2903         uint32_t vring_idx;
2904
2905         switch (msg->request.master) {
2906         case VHOST_USER_SET_VRING_KICK:
2907         case VHOST_USER_SET_VRING_CALL:
2908         case VHOST_USER_SET_VRING_ERR:
2909                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
2910                 break;
2911         case VHOST_USER_SET_VRING_NUM:
2912         case VHOST_USER_SET_VRING_BASE:
2913         case VHOST_USER_GET_VRING_BASE:
2914         case VHOST_USER_SET_VRING_ENABLE:
2915                 vring_idx = msg->payload.state.index;
2916                 break;
2917         case VHOST_USER_SET_VRING_ADDR:
2918                 vring_idx = msg->payload.addr.index;
2919                 break;
2920         default:
2921                 return 0;
2922         }
2923
2924         if (vring_idx >= VHOST_MAX_VRING) {
2925                 VHOST_LOG_CONFIG(ERR,
2926                         "invalid vring index: %u\n", vring_idx);
2927                 return -1;
2928         }
2929
2930         if (dev->virtqueue[vring_idx])
2931                 return 0;
2932
2933         return alloc_vring_queue(dev, vring_idx);
2934 }
2935
2936 static void
2937 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
2938 {
2939         unsigned int i = 0;
2940         unsigned int vq_num = 0;
2941
2942         while (vq_num < dev->nr_vring) {
2943                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2944
2945                 if (vq) {
2946                         rte_spinlock_lock(&vq->access_lock);
2947                         vq_num++;
2948                 }
2949                 i++;
2950         }
2951 }
2952
2953 static void
2954 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
2955 {
2956         unsigned int i = 0;
2957         unsigned int vq_num = 0;
2958
2959         while (vq_num < dev->nr_vring) {
2960                 struct vhost_virtqueue *vq = dev->virtqueue[i];
2961
2962                 if (vq) {
2963                         rte_spinlock_unlock(&vq->access_lock);
2964                         vq_num++;
2965                 }
2966                 i++;
2967         }
2968 }
2969
2970 int
2971 vhost_user_msg_handler(int vid, int fd)
2972 {
2973         struct virtio_net *dev;
2974         struct VhostUserMsg msg;
2975         struct rte_vdpa_device *vdpa_dev;
2976         int ret;
2977         int unlock_required = 0;
2978         bool handled;
2979         int request;
2980         uint32_t i;
2981
2982         dev = get_device(vid);
2983         if (dev == NULL)
2984                 return -1;
2985
2986         if (!dev->notify_ops) {
2987                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
2988                 if (!dev->notify_ops) {
2989                         VHOST_LOG_CONFIG(ERR,
2990                                 "failed to get callback ops for driver %s\n",
2991                                 dev->ifname);
2992                         return -1;
2993                 }
2994         }
2995
2996         ret = read_vhost_message(fd, &msg);
2997         if (ret <= 0) {
2998                 if (ret < 0)
2999                         VHOST_LOG_CONFIG(ERR,
3000                                 "vhost read message failed\n");
3001                 else
3002                         VHOST_LOG_CONFIG(INFO,
3003                                 "vhost peer closed\n");
3004
3005                 return -1;
3006         }
3007
3008         ret = 0;
3009         request = msg.request.master;
3010         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
3011                         vhost_message_str[request]) {
3012                 if (request != VHOST_USER_IOTLB_MSG)
3013                         VHOST_LOG_CONFIG(INFO, "read message %s\n",
3014                                 vhost_message_str[request]);
3015                 else
3016                         VHOST_LOG_CONFIG(DEBUG, "read message %s\n",
3017                                 vhost_message_str[request]);
3018         } else {
3019                 VHOST_LOG_CONFIG(DEBUG, "External request %d\n", request);
3020         }
3021
3022         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
3023         if (ret < 0) {
3024                 VHOST_LOG_CONFIG(ERR,
3025                         "failed to alloc queue\n");
3026                 return -1;
3027         }
3028
3029         /*
3030          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
3031          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
3032          * and device is destroyed. destroy_device waits for queues to be
3033          * inactive, so it is safe. Otherwise taking the access_lock
3034          * would cause a dead lock.
3035          */
3036         switch (request) {
3037         case VHOST_USER_SET_FEATURES:
3038         case VHOST_USER_SET_PROTOCOL_FEATURES:
3039         case VHOST_USER_SET_OWNER:
3040         case VHOST_USER_SET_MEM_TABLE:
3041         case VHOST_USER_SET_LOG_BASE:
3042         case VHOST_USER_SET_LOG_FD:
3043         case VHOST_USER_SET_VRING_NUM:
3044         case VHOST_USER_SET_VRING_ADDR:
3045         case VHOST_USER_SET_VRING_BASE:
3046         case VHOST_USER_SET_VRING_KICK:
3047         case VHOST_USER_SET_VRING_CALL:
3048         case VHOST_USER_SET_VRING_ERR:
3049         case VHOST_USER_SET_VRING_ENABLE:
3050         case VHOST_USER_SEND_RARP:
3051         case VHOST_USER_NET_SET_MTU:
3052         case VHOST_USER_SET_SLAVE_REQ_FD:
3053                 if (!(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED)) {
3054                         vhost_user_lock_all_queue_pairs(dev);
3055                         unlock_required = 1;
3056                 }
3057                 break;
3058         default:
3059                 break;
3060
3061         }
3062
3063         handled = false;
3064         if (dev->extern_ops.pre_msg_handle) {
3065                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
3066                                 (void *)&msg);
3067                 switch (ret) {
3068                 case RTE_VHOST_MSG_RESULT_REPLY:
3069                         send_vhost_reply(fd, &msg);
3070                         /* Fall-through */
3071                 case RTE_VHOST_MSG_RESULT_ERR:
3072                 case RTE_VHOST_MSG_RESULT_OK:
3073                         handled = true;
3074                         goto skip_to_post_handle;
3075                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
3076                 default:
3077                         break;
3078                 }
3079         }
3080
3081         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
3082                 if (!vhost_message_handlers[request])
3083                         goto skip_to_post_handle;
3084                 ret = vhost_message_handlers[request](&dev, &msg, fd);
3085
3086                 switch (ret) {
3087                 case RTE_VHOST_MSG_RESULT_ERR:
3088                         VHOST_LOG_CONFIG(ERR,
3089                                 "Processing %s failed.\n",
3090                                 vhost_message_str[request]);
3091                         handled = true;
3092                         break;
3093                 case RTE_VHOST_MSG_RESULT_OK:
3094                         VHOST_LOG_CONFIG(DEBUG,
3095                                 "Processing %s succeeded.\n",
3096                                 vhost_message_str[request]);
3097                         handled = true;
3098                         break;
3099                 case RTE_VHOST_MSG_RESULT_REPLY:
3100                         VHOST_LOG_CONFIG(DEBUG,
3101                                 "Processing %s succeeded and needs reply.\n",
3102                                 vhost_message_str[request]);
3103                         send_vhost_reply(fd, &msg);
3104                         handled = true;
3105                         break;
3106                 default:
3107                         break;
3108                 }
3109         }
3110
3111 skip_to_post_handle:
3112         if (ret != RTE_VHOST_MSG_RESULT_ERR &&
3113                         dev->extern_ops.post_msg_handle) {
3114                 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
3115                                 (void *)&msg);
3116                 switch (ret) {
3117                 case RTE_VHOST_MSG_RESULT_REPLY:
3118                         send_vhost_reply(fd, &msg);
3119                         /* Fall-through */
3120                 case RTE_VHOST_MSG_RESULT_ERR:
3121                 case RTE_VHOST_MSG_RESULT_OK:
3122                         handled = true;
3123                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
3124                 default:
3125                         break;
3126                 }
3127         }
3128
3129         /* If message was not handled at this stage, treat it as an error */
3130         if (!handled) {
3131                 VHOST_LOG_CONFIG(ERR,
3132                         "vhost message (req: %d) was not handled.\n", request);
3133                 close_msg_fds(&msg);
3134                 ret = RTE_VHOST_MSG_RESULT_ERR;
3135         }
3136
3137         /*
3138          * If the request required a reply that was already sent,
3139          * this optional reply-ack won't be sent as the
3140          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
3141          */
3142         if (msg.flags & VHOST_USER_NEED_REPLY) {
3143                 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
3144                 msg.size = sizeof(msg.payload.u64);
3145                 msg.fd_num = 0;
3146                 send_vhost_reply(fd, &msg);
3147         } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
3148                 VHOST_LOG_CONFIG(ERR,
3149                         "vhost message handling failed.\n");
3150                 return -1;
3151         }
3152
3153         for (i = 0; i < dev->nr_vring; i++) {
3154                 struct vhost_virtqueue *vq = dev->virtqueue[i];
3155                 bool cur_ready = vq_is_ready(dev, vq);
3156
3157                 if (cur_ready != (vq && vq->ready)) {
3158                         vq->ready = cur_ready;
3159                         vhost_user_notify_queue_state(dev, i, cur_ready);
3160                 }
3161         }
3162
3163         if (unlock_required)
3164                 vhost_user_unlock_all_queue_pairs(dev);
3165
3166         if (!virtio_is_ready(dev))
3167                 goto out;
3168
3169         /*
3170          * Virtio is now ready. If not done already, it is time
3171          * to notify the application it can process the rings and
3172          * configure the vDPA device if present.
3173          */
3174
3175         if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
3176                 if (dev->notify_ops->new_device(dev->vid) == 0)
3177                         dev->flags |= VIRTIO_DEV_RUNNING;
3178         }
3179
3180         vdpa_dev = dev->vdpa_dev;
3181         if (!vdpa_dev)
3182                 goto out;
3183
3184         if (!(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED)) {
3185                 if (vdpa_dev->ops->dev_conf(dev->vid))
3186                         VHOST_LOG_CONFIG(ERR,
3187                                          "Failed to configure vDPA device\n");
3188                 else
3189                         dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
3190         }
3191
3192 out:
3193         return 0;
3194 }
3195
3196 static int process_slave_message_reply(struct virtio_net *dev,
3197                                        const struct VhostUserMsg *msg)
3198 {
3199         struct VhostUserMsg msg_reply;
3200         int ret;
3201
3202         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
3203                 return 0;
3204
3205         ret = read_vhost_message(dev->slave_req_fd, &msg_reply);
3206         if (ret <= 0) {
3207                 if (ret < 0)
3208                         VHOST_LOG_CONFIG(ERR,
3209                                 "vhost read slave message reply failed\n");
3210                 else
3211                         VHOST_LOG_CONFIG(INFO,
3212                                 "vhost peer closed\n");
3213                 ret = -1;
3214                 goto out;
3215         }
3216
3217         ret = 0;
3218         if (msg_reply.request.slave != msg->request.slave) {
3219                 VHOST_LOG_CONFIG(ERR,
3220                         "Received unexpected msg type (%u), expected %u\n",
3221                         msg_reply.request.slave, msg->request.slave);
3222                 ret = -1;
3223                 goto out;
3224         }
3225
3226         ret = msg_reply.payload.u64 ? -1 : 0;
3227
3228 out:
3229         rte_spinlock_unlock(&dev->slave_req_lock);
3230         return ret;
3231 }
3232
3233 int
3234 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
3235 {
3236         int ret;
3237         struct VhostUserMsg msg = {
3238                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
3239                 .flags = VHOST_USER_VERSION,
3240                 .size = sizeof(msg.payload.iotlb),
3241                 .payload.iotlb = {
3242                         .iova = iova,
3243                         .perm = perm,
3244                         .type = VHOST_IOTLB_MISS,
3245                 },
3246         };
3247
3248         ret = send_vhost_message(dev->slave_req_fd, &msg);
3249         if (ret < 0) {
3250                 VHOST_LOG_CONFIG(ERR,
3251                                 "Failed to send IOTLB miss message (%d)\n",
3252                                 ret);
3253                 return ret;
3254         }
3255
3256         return 0;
3257 }
3258
3259 static int
3260 vhost_user_slave_config_change(struct virtio_net *dev, bool need_reply)
3261 {
3262         int ret;
3263         struct VhostUserMsg msg = {
3264                 .request.slave = VHOST_USER_SLAVE_CONFIG_CHANGE_MSG,
3265                 .flags = VHOST_USER_VERSION,
3266                 .size = 0,
3267         };
3268
3269         if (need_reply)
3270                 msg.flags |= VHOST_USER_NEED_REPLY;
3271
3272         ret = send_vhost_slave_message(dev, &msg);
3273         if (ret < 0) {
3274                 VHOST_LOG_CONFIG(ERR,
3275                                 "Failed to send config change (%d)\n",
3276                                 ret);
3277                 return ret;
3278         }
3279
3280         return process_slave_message_reply(dev, &msg);
3281 }
3282
3283 int
3284 rte_vhost_slave_config_change(int vid, bool need_reply)
3285 {
3286         struct virtio_net *dev;
3287
3288         dev = get_device(vid);
3289         if (!dev)
3290                 return -ENODEV;
3291
3292         return vhost_user_slave_config_change(dev, need_reply);
3293 }
3294
3295 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
3296                                                     int index, int fd,
3297                                                     uint64_t offset,
3298                                                     uint64_t size)
3299 {
3300         int ret;
3301         struct VhostUserMsg msg = {
3302                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
3303                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
3304                 .size = sizeof(msg.payload.area),
3305                 .payload.area = {
3306                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
3307                         .size = size,
3308                         .offset = offset,
3309                 },
3310         };
3311
3312         if (fd < 0)
3313                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
3314         else {
3315                 msg.fds[0] = fd;
3316                 msg.fd_num = 1;
3317         }
3318
3319         ret = send_vhost_slave_message(dev, &msg);
3320         if (ret < 0) {
3321                 VHOST_LOG_CONFIG(ERR,
3322                         "Failed to set host notifier (%d)\n", ret);
3323                 return ret;
3324         }
3325
3326         return process_slave_message_reply(dev, &msg);
3327 }
3328
3329 int rte_vhost_host_notifier_ctrl(int vid, uint16_t qid, bool enable)
3330 {
3331         struct virtio_net *dev;
3332         struct rte_vdpa_device *vdpa_dev;
3333         int vfio_device_fd, ret = 0;
3334         uint64_t offset, size;
3335         unsigned int i, q_start, q_last;
3336
3337         dev = get_device(vid);
3338         if (!dev)
3339                 return -ENODEV;
3340
3341         vdpa_dev = dev->vdpa_dev;
3342         if (vdpa_dev == NULL)
3343                 return -ENODEV;
3344
3345         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
3346             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
3347             !(dev->protocol_features &
3348                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
3349             !(dev->protocol_features &
3350                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
3351             !(dev->protocol_features &
3352                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
3353                 return -ENOTSUP;
3354
3355         if (qid == RTE_VHOST_QUEUE_ALL) {
3356                 q_start = 0;
3357                 q_last = dev->nr_vring - 1;
3358         } else {
3359                 if (qid >= dev->nr_vring)
3360                         return -EINVAL;
3361                 q_start = qid;
3362                 q_last = qid;
3363         }
3364
3365         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
3366         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
3367
3368         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
3369         if (vfio_device_fd < 0)
3370                 return -ENOTSUP;
3371
3372         if (enable) {
3373                 for (i = q_start; i <= q_last; i++) {
3374                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
3375                                         &size) < 0) {
3376                                 ret = -ENOTSUP;
3377                                 goto disable;
3378                         }
3379
3380                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
3381                                         vfio_device_fd, offset, size) < 0) {
3382                                 ret = -EFAULT;
3383                                 goto disable;
3384                         }
3385                 }
3386         } else {
3387 disable:
3388                 for (i = q_start; i <= q_last; i++) {
3389                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
3390                                         0, 0);
3391                 }
3392         }
3393
3394         return ret;
3395 }