lib/librte_vhost/vhost_user.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2018 Intel Corporation
   3  */
   4
   5 /* Security model
   6  * --------------
   7  * The vhost-user protocol connection is an external interface, so it must be
   8  * robust against invalid inputs.
   9  *
  10  * This is important because the vhost-user master is only one step removed
  11  * from the guest.  Malicious guests that have escaped will then launch further
  12  * attacks from the vhost-user master.
  13  *
  14  * Even in deployments where guests are trusted, a bug in the vhost-user master
  15  * can still cause invalid messages to be sent.  Such messages must not
  16  * compromise the stability of the DPDK application by causing crashes, memory
  17  * corruption, or other problematic behavior.
  18  *
  19  * Do not assume received VhostUserMsg fields contain sensible values!
  20  */
  21
  22 #include <stdint.h>
  23 #include <stdio.h>
  24 #include <stdlib.h>
  25 #include <string.h>
  26 #include <unistd.h>
  27 #include <fcntl.h>
  28 #include <sys/ioctl.h>
  29 #include <sys/mman.h>
  30 #include <sys/types.h>
  31 #include <sys/stat.h>
  32 #include <sys/syscall.h>
  33 #include <assert.h>
  34 #ifdef RTE_LIBRTE_VHOST_NUMA
  35 #include <numaif.h>
  36 #endif
  37 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
  38 #include <linux/userfaultfd.h>
  39 #endif
  40
  41 #include <rte_common.h>
  42 #include <rte_malloc.h>
  43 #include <rte_log.h>
  44
  45 #include "iotlb.h"
  46 #include "vhost.h"
  47 #include "vhost_user.h"
  48
  49 #define VIRTIO_MIN_MTU 68
  50 #define VIRTIO_MAX_MTU 65535
  51
  52 static const char *vhost_message_str[VHOST_USER_MAX] = {
  53         [VHOST_USER_NONE] = "VHOST_USER_NONE",
  54         [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
  55         [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
  56         [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
  57         [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
  58         [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
  59         [VHOST_USER_SET_LOG_BASE] = "VHOST_USER_SET_LOG_BASE",
  60         [VHOST_USER_SET_LOG_FD] = "VHOST_USER_SET_LOG_FD",
  61         [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
  62         [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
  63         [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
  64         [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
  65         [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
  66         [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
  67         [VHOST_USER_SET_VRING_ERR]  = "VHOST_USER_SET_VRING_ERR",
  68         [VHOST_USER_GET_PROTOCOL_FEATURES]  = "VHOST_USER_GET_PROTOCOL_FEATURES",
  69         [VHOST_USER_SET_PROTOCOL_FEATURES]  = "VHOST_USER_SET_PROTOCOL_FEATURES",
  70         [VHOST_USER_GET_QUEUE_NUM]  = "VHOST_USER_GET_QUEUE_NUM",
  71         [VHOST_USER_SET_VRING_ENABLE]  = "VHOST_USER_SET_VRING_ENABLE",
  72         [VHOST_USER_SEND_RARP]  = "VHOST_USER_SEND_RARP",
  73         [VHOST_USER_NET_SET_MTU]  = "VHOST_USER_NET_SET_MTU",
  74         [VHOST_USER_SET_SLAVE_REQ_FD]  = "VHOST_USER_SET_SLAVE_REQ_FD",
  75         [VHOST_USER_IOTLB_MSG]  = "VHOST_USER_IOTLB_MSG",
  76         [VHOST_USER_CRYPTO_CREATE_SESS] = "VHOST_USER_CRYPTO_CREATE_SESS",
  77         [VHOST_USER_CRYPTO_CLOSE_SESS] = "VHOST_USER_CRYPTO_CLOSE_SESS",
  78         [VHOST_USER_POSTCOPY_ADVISE]  = "VHOST_USER_POSTCOPY_ADVISE",
  79         [VHOST_USER_POSTCOPY_LISTEN]  = "VHOST_USER_POSTCOPY_LISTEN",
  80         [VHOST_USER_POSTCOPY_END]  = "VHOST_USER_POSTCOPY_END",
  81 };
  82
  83 static int send_vhost_reply(int sockfd, struct VhostUserMsg *msg);
  84 static int read_vhost_message(int sockfd, struct VhostUserMsg *msg);
  85
  86 static uint64_t
  87 get_blk_size(int fd)
  88 {
  89         struct stat stat;
  90         int ret;
  91
  92         ret = fstat(fd, &stat);
  93         return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
  94 }
  95
  96 /*
  97  * Reclaim all the outstanding zmbufs for a virtqueue.
  98  */
  99 static void
 100 drain_zmbuf_list(struct vhost_virtqueue *vq)
 101 {
 102         struct zcopy_mbuf *zmbuf, *next;
 103
 104         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
 105              zmbuf != NULL; zmbuf = next) {
 106                 next = TAILQ_NEXT(zmbuf, next);
 107
 108                 while (!mbuf_is_consumed(zmbuf->mbuf))
 109                         usleep(1000);
 110
 111                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
 112                 restore_mbuf(zmbuf->mbuf);
 113                 rte_pktmbuf_free(zmbuf->mbuf);
 114                 put_zmbuf(zmbuf);
 115                 vq->nr_zmbuf -= 1;
 116         }
 117 }
 118
 119 static void
 120 free_mem_region(struct virtio_net *dev)
 121 {
 122         uint32_t i;
 123         struct rte_vhost_mem_region *reg;
 124         struct vhost_virtqueue *vq;
 125
 126         if (!dev || !dev->mem)
 127                 return;
 128
 129         if (dev->dequeue_zero_copy) {
 130                 for (i = 0; i < dev->nr_vring; i++) {
 131                         vq = dev->virtqueue[i];
 132                         if (vq)
 133                                 drain_zmbuf_list(vq);
 134                 }
 135         }
 136
 137         for (i = 0; i < dev->mem->nregions; i++) {
 138                 reg = &dev->mem->regions[i];
 139                 if (reg->host_user_addr) {
 140                         munmap(reg->mmap_addr, reg->mmap_size);
 141                         close(reg->fd);
 142                 }
 143         }
 144 }
 145
 146 void
 147 vhost_backend_cleanup(struct virtio_net *dev)
 148 {
 149         if (dev->mem) {
 150                 free_mem_region(dev);
 151                 rte_free(dev->mem);
 152                 dev->mem = NULL;
 153         }
 154
 155         free(dev->guest_pages);
 156         dev->guest_pages = NULL;
 157
 158         if (dev->log_addr) {
 159                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 160                 dev->log_addr = 0;
 161         }
 162
 163         if (dev->slave_req_fd >= 0) {
 164                 close(dev->slave_req_fd);
 165                 dev->slave_req_fd = -1;
 166         }
 167
 168         if (dev->postcopy_ufd >= 0) {
 169                 close(dev->postcopy_ufd);
 170                 dev->postcopy_ufd = -1;
 171         }
 172
 173         dev->postcopy_listening = 0;
 174 }
 175
 176 /*
 177  * This function just returns success at the moment unless
 178  * the device hasn't been initialised.
 179  */
 180 static int
 181 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
 182                         struct VhostUserMsg *msg __rte_unused,
 183                         int main_fd __rte_unused)
 184 {
 185         return RTE_VHOST_MSG_RESULT_OK;
 186 }
 187
 188 static int
 189 vhost_user_reset_owner(struct virtio_net **pdev,
 190                         struct VhostUserMsg *msg __rte_unused,
 191                         int main_fd __rte_unused)
 192 {
 193         struct virtio_net *dev = *pdev;
 194         vhost_destroy_device_notify(dev);
 195
 196         cleanup_device(dev, 0);
 197         reset_device(dev);
 198         return RTE_VHOST_MSG_RESULT_OK;
 199 }
 200
 201 /*
 202  * The features that we support are requested.
 203  */
 204 static int
 205 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 206                         int main_fd __rte_unused)
 207 {
 208         struct virtio_net *dev = *pdev;
 209         uint64_t features = 0;
 210
 211         rte_vhost_driver_get_features(dev->ifname, &features);
 212
 213         msg->payload.u64 = features;
 214         msg->size = sizeof(msg->payload.u64);
 215         msg->fd_num = 0;
 216
 217         return RTE_VHOST_MSG_RESULT_REPLY;
 218 }
 219
 220 /*
 221  * The queue number that we support are requested.
 222  */
 223 static int
 224 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg,
 225                         int main_fd __rte_unused)
 226 {
 227         struct virtio_net *dev = *pdev;
 228         uint32_t queue_num = 0;
 229
 230         rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
 231
 232         msg->payload.u64 = (uint64_t)queue_num;
 233         msg->size = sizeof(msg->payload.u64);
 234         msg->fd_num = 0;
 235
 236         return RTE_VHOST_MSG_RESULT_REPLY;
 237 }
 238
 239 /*
 240  * We receive the negotiated features supported by us and the virtio device.
 241  */
 242 static int
 243 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg,
 244                         int main_fd __rte_unused)
 245 {
 246         struct virtio_net *dev = *pdev;
 247         uint64_t features = msg->payload.u64;
 248         uint64_t vhost_features = 0;
 249         struct rte_vdpa_device *vdpa_dev;
 250         int did = -1;
 251
 252         rte_vhost_driver_get_features(dev->ifname, &vhost_features);
 253         if (features & ~vhost_features) {
 254                 RTE_LOG(ERR, VHOST_CONFIG,
 255                         "(%d) received invalid negotiated features.\n",
 256                         dev->vid);
 257                 return RTE_VHOST_MSG_RESULT_ERR;
 258         }
 259
 260         if (dev->flags & VIRTIO_DEV_RUNNING) {
 261                 if (dev->features == features)
 262                         return RTE_VHOST_MSG_RESULT_OK;
 263
 264                 /*
 265                  * Error out if master tries to change features while device is
 266                  * in running state. The exception being VHOST_F_LOG_ALL, which
 267                  * is enabled when the live-migration starts.
 268                  */
 269                 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
 270                         RTE_LOG(ERR, VHOST_CONFIG,
 271                                 "(%d) features changed while device is running.\n",
 272                                 dev->vid);
 273                         return RTE_VHOST_MSG_RESULT_ERR;
 274                 }
 275
 276                 if (dev->notify_ops->features_changed)
 277                         dev->notify_ops->features_changed(dev->vid, features);
 278         }
 279
 280         dev->features = features;
 281         if (dev->features &
 282                 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
 283                 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 284         } else {
 285                 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
 286         }
 287         RTE_LOG(INFO, VHOST_CONFIG,
 288                 "negotiated Virtio features: 0x%" PRIx64 "\n", dev->features);
 289         VHOST_LOG_DEBUG(VHOST_CONFIG,
 290                 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
 291                 dev->vid,
 292                 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
 293                 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 294
 295         if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
 296             !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
 297                 /*
 298                  * Remove all but first queue pair if MQ hasn't been
 299                  * negotiated. This is safe because the device is not
 300                  * running at this stage.
 301                  */
 302                 while (dev->nr_vring > 2) {
 303                         struct vhost_virtqueue *vq;
 304
 305                         vq = dev->virtqueue[--dev->nr_vring];
 306                         if (!vq)
 307                                 continue;
 308
 309                         dev->virtqueue[dev->nr_vring] = NULL;
 310                         cleanup_vq(vq, 1);
 311                         free_vq(dev, vq);
 312                 }
 313         }
 314
 315         did = dev->vdpa_dev_id;
 316         vdpa_dev = rte_vdpa_get_device(did);
 317         if (vdpa_dev && vdpa_dev->ops->set_features)
 318                 vdpa_dev->ops->set_features(dev->vid);
 319
 320         return RTE_VHOST_MSG_RESULT_OK;
 321 }
 322
 323 /*
 324  * The virtio device sends us the size of the descriptor ring.
 325  */
 326 static int
 327 vhost_user_set_vring_num(struct virtio_net **pdev,
 328                         struct VhostUserMsg *msg,
 329                         int main_fd __rte_unused)
 330 {
 331         struct virtio_net *dev = *pdev;
 332         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 333
 334         vq->size = msg->payload.state.num;
 335
 336         /* VIRTIO 1.0, 2.4 Virtqueues says:
 337          *
 338          *   Queue Size value is always a power of 2. The maximum Queue Size
 339          *   value is 32768.
 340          */
 341         if ((vq->size & (vq->size - 1)) || vq->size > 32768) {
 342                 RTE_LOG(ERR, VHOST_CONFIG,
 343                         "invalid virtqueue size %u\n", vq->size);
 344                 return RTE_VHOST_MSG_RESULT_ERR;
 345         }
 346
 347         if (dev->dequeue_zero_copy) {
 348                 vq->nr_zmbuf = 0;
 349                 vq->last_zmbuf_idx = 0;
 350                 vq->zmbuf_size = vq->size;
 351                 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
 352                                          sizeof(struct zcopy_mbuf), 0);
 353                 if (vq->zmbufs == NULL) {
 354                         RTE_LOG(WARNING, VHOST_CONFIG,
 355                                 "failed to allocate mem for zero copy; "
 356                                 "zero copy is force disabled\n");
 357                         dev->dequeue_zero_copy = 0;
 358                 }
 359                 TAILQ_INIT(&vq->zmbuf_list);
 360         }
 361
 362         if (vq_is_packed(dev)) {
 363                 vq->shadow_used_packed = rte_malloc(NULL,
 364                                 vq->size *
 365                                 sizeof(struct vring_used_elem_packed),
 366                                 RTE_CACHE_LINE_SIZE);
 367                 if (!vq->shadow_used_packed) {
 368                         RTE_LOG(ERR, VHOST_CONFIG,
 369                                         "failed to allocate memory for shadow used ring.\n");
 370                         return RTE_VHOST_MSG_RESULT_ERR;
 371                 }
 372
 373         } else {
 374                 vq->shadow_used_split = rte_malloc(NULL,
 375                                 vq->size * sizeof(struct vring_used_elem),
 376                                 RTE_CACHE_LINE_SIZE);
 377                 if (!vq->shadow_used_split) {
 378                         RTE_LOG(ERR, VHOST_CONFIG,
 379                                         "failed to allocate memory for shadow used ring.\n");
 380                         return RTE_VHOST_MSG_RESULT_ERR;
 381                 }
 382         }
 383
 384         vq->batch_copy_elems = rte_malloc(NULL,
 385                                 vq->size * sizeof(struct batch_copy_elem),
 386                                 RTE_CACHE_LINE_SIZE);
 387         if (!vq->batch_copy_elems) {
 388                 RTE_LOG(ERR, VHOST_CONFIG,
 389                         "failed to allocate memory for batching copy.\n");
 390                 return RTE_VHOST_MSG_RESULT_ERR;
 391         }
 392
 393         return RTE_VHOST_MSG_RESULT_OK;
 394 }
 395
 396 /*
 397  * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
 398  * same numa node as the memory of vring descriptor.
 399  */
 400 #ifdef RTE_LIBRTE_VHOST_NUMA
 401 static struct virtio_net*
 402 numa_realloc(struct virtio_net *dev, int index)
 403 {
 404         int oldnode, newnode;
 405         struct virtio_net *old_dev;
 406         struct vhost_virtqueue *old_vq, *vq;
 407         struct zcopy_mbuf *new_zmbuf;
 408         struct vring_used_elem *new_shadow_used_split;
 409         struct vring_used_elem_packed *new_shadow_used_packed;
 410         struct batch_copy_elem *new_batch_copy_elems;
 411         int ret;
 412
 413         old_dev = dev;
 414         vq = old_vq = dev->virtqueue[index];
 415
 416         ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
 417                             MPOL_F_NODE | MPOL_F_ADDR);
 418
 419         /* check if we need to reallocate vq */
 420         ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
 421                              MPOL_F_NODE | MPOL_F_ADDR);
 422         if (ret) {
 423                 RTE_LOG(ERR, VHOST_CONFIG,
 424                         "Unable to get vq numa information.\n");
 425                 return dev;
 426         }
 427         if (oldnode != newnode) {
 428                 RTE_LOG(INFO, VHOST_CONFIG,
 429                         "reallocate vq from %d to %d node\n", oldnode, newnode);
 430                 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
 431                 if (!vq)
 432                         return dev;
 433
 434                 memcpy(vq, old_vq, sizeof(*vq));
 435                 TAILQ_INIT(&vq->zmbuf_list);
 436
 437                 if (dev->dequeue_zero_copy) {
 438                         new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
 439                                         sizeof(struct zcopy_mbuf), 0, newnode);
 440                         if (new_zmbuf) {
 441                                 rte_free(vq->zmbufs);
 442                                 vq->zmbufs = new_zmbuf;
 443                         }
 444                 }
 445
 446                 if (vq_is_packed(dev)) {
 447                         new_shadow_used_packed = rte_malloc_socket(NULL,
 448                                         vq->size *
 449                                         sizeof(struct vring_used_elem_packed),
 450                                         RTE_CACHE_LINE_SIZE,
 451                                         newnode);
 452                         if (new_shadow_used_packed) {
 453                                 rte_free(vq->shadow_used_packed);
 454                                 vq->shadow_used_packed = new_shadow_used_packed;
 455                         }
 456                 } else {
 457                         new_shadow_used_split = rte_malloc_socket(NULL,
 458                                         vq->size *
 459                                         sizeof(struct vring_used_elem),
 460                                         RTE_CACHE_LINE_SIZE,
 461                                         newnode);
 462                         if (new_shadow_used_split) {
 463                                 rte_free(vq->shadow_used_split);
 464                                 vq->shadow_used_split = new_shadow_used_split;
 465                         }
 466                 }
 467
 468                 new_batch_copy_elems = rte_malloc_socket(NULL,
 469                         vq->size * sizeof(struct batch_copy_elem),
 470                         RTE_CACHE_LINE_SIZE,
 471                         newnode);
 472                 if (new_batch_copy_elems) {
 473                         rte_free(vq->batch_copy_elems);
 474                         vq->batch_copy_elems = new_batch_copy_elems;
 475                 }
 476
 477                 rte_free(old_vq);
 478         }
 479
 480         /* check if we need to reallocate dev */
 481         ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
 482                             MPOL_F_NODE | MPOL_F_ADDR);
 483         if (ret) {
 484                 RTE_LOG(ERR, VHOST_CONFIG,
 485                         "Unable to get dev numa information.\n");
 486                 goto out;
 487         }
 488         if (oldnode != newnode) {
 489                 RTE_LOG(INFO, VHOST_CONFIG,
 490                         "reallocate dev from %d to %d node\n",
 491                         oldnode, newnode);
 492                 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
 493                 if (!dev) {
 494                         dev = old_dev;
 495                         goto out;
 496                 }
 497
 498                 memcpy(dev, old_dev, sizeof(*dev));
 499                 rte_free(old_dev);
 500         }
 501
 502 out:
 503         dev->virtqueue[index] = vq;
 504         vhost_devices[dev->vid] = dev;
 505
 506         if (old_vq != vq)
 507                 vhost_user_iotlb_init(dev, index);
 508
 509         return dev;
 510 }
 511 #else
 512 static struct virtio_net*
 513 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 514 {
 515         return dev;
 516 }
 517 #endif
 518
 519 /* Converts QEMU virtual address to Vhost virtual address. */
 520 static uint64_t
 521 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
 522 {
 523         struct rte_vhost_mem_region *r;
 524         uint32_t i;
 525
 526         if (unlikely(!dev || !dev->mem))
 527                 goto out_error;
 528
 529         /* Find the region where the address lives. */
 530         for (i = 0; i < dev->mem->nregions; i++) {
 531                 r = &dev->mem->regions[i];
 532
 533                 if (qva >= r->guest_user_addr &&
 534                     qva <  r->guest_user_addr + r->size) {
 535
 536                         if (unlikely(*len > r->guest_user_addr + r->size - qva))
 537                                 *len = r->guest_user_addr + r->size - qva;
 538
 539                         return qva - r->guest_user_addr +
 540                                r->host_user_addr;
 541                 }
 542         }
 543 out_error:
 544         *len = 0;
 545
 546         return 0;
 547 }
 548
 549
 550 /*
 551  * Converts ring address to Vhost virtual address.
 552  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 553  * else it is a QEMU virtual address.
 554  */
 555 static uint64_t
 556 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 557                 uint64_t ra, uint64_t *size)
 558 {
 559         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 560                 uint64_t vva;
 561
 562                 vva = vhost_user_iotlb_cache_find(vq, ra,
 563                                         size, VHOST_ACCESS_RW);
 564                 if (!vva)
 565                         vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
 566
 567                 return vva;
 568         }
 569
 570         return qva_to_vva(dev, ra, size);
 571 }
 572
 573 static struct virtio_net *
 574 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 575 {
 576         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 577         struct vhost_vring_addr *addr = &vq->ring_addrs;
 578         uint64_t len, expected_len;
 579
 580         if (vq_is_packed(dev)) {
 581                 len = sizeof(struct vring_packed_desc) * vq->size;
 582                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 583                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 584                 vq->log_guest_addr = 0;
 585                 if (vq->desc_packed == NULL ||
 586                                 len != sizeof(struct vring_packed_desc) *
 587                                 vq->size) {
 588                         RTE_LOG(DEBUG, VHOST_CONFIG,
 589                                 "(%d) failed to map desc_packed ring.\n",
 590                                 dev->vid);
 591                         return dev;
 592                 }
 593
 594                 dev = numa_realloc(dev, vq_index);
 595                 vq = dev->virtqueue[vq_index];
 596                 addr = &vq->ring_addrs;
 597
 598                 len = sizeof(struct vring_packed_desc_event);
 599                 vq->driver_event = (struct vring_packed_desc_event *)
 600                                         (uintptr_t)ring_addr_to_vva(dev,
 601                                         vq, addr->avail_user_addr, &len);
 602                 if (vq->driver_event == NULL ||
 603                                 len != sizeof(struct vring_packed_desc_event)) {
 604                         RTE_LOG(DEBUG, VHOST_CONFIG,
 605                                 "(%d) failed to find driver area address.\n",
 606                                 dev->vid);
 607                         return dev;
 608                 }
 609
 610                 len = sizeof(struct vring_packed_desc_event);
 611                 vq->device_event = (struct vring_packed_desc_event *)
 612                                         (uintptr_t)ring_addr_to_vva(dev,
 613                                         vq, addr->used_user_addr, &len);
 614                 if (vq->device_event == NULL ||
 615                                 len != sizeof(struct vring_packed_desc_event)) {
 616                         RTE_LOG(DEBUG, VHOST_CONFIG,
 617                                 "(%d) failed to find device area address.\n",
 618                                 dev->vid);
 619                         return dev;
 620                 }
 621
 622                 return dev;
 623         }
 624
 625         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 626         if (vq->desc && vq->avail && vq->used)
 627                 return dev;
 628
 629         len = sizeof(struct vring_desc) * vq->size;
 630         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 631                         vq, addr->desc_user_addr, &len);
 632         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 633                 RTE_LOG(DEBUG, VHOST_CONFIG,
 634                         "(%d) failed to map desc ring.\n",
 635                         dev->vid);
 636                 return dev;
 637         }
 638
 639         dev = numa_realloc(dev, vq_index);
 640         vq = dev->virtqueue[vq_index];
 641         addr = &vq->ring_addrs;
 642
 643         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 644         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 645                 len += sizeof(uint16_t);
 646         expected_len = len;
 647         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 648                         vq, addr->avail_user_addr, &len);
 649         if (vq->avail == 0 || len != expected_len) {
 650                 RTE_LOG(DEBUG, VHOST_CONFIG,
 651                         "(%d) failed to map avail ring.\n",
 652                         dev->vid);
 653                 return dev;
 654         }
 655
 656         len = sizeof(struct vring_used) +
 657                 sizeof(struct vring_used_elem) * vq->size;
 658         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 659                 len += sizeof(uint16_t);
 660         expected_len = len;
 661         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 662                         vq, addr->used_user_addr, &len);
 663         if (vq->used == 0 || len != expected_len) {
 664                 RTE_LOG(DEBUG, VHOST_CONFIG,
 665                         "(%d) failed to map used ring.\n",
 666                         dev->vid);
 667                 return dev;
 668         }
 669
 670         if (vq->last_used_idx != vq->used->idx) {
 671                 RTE_LOG(WARNING, VHOST_CONFIG,
 672                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 673                         "some packets maybe resent for Tx and dropped for Rx\n",
 674                         vq->last_used_idx, vq->used->idx);
 675                 vq->last_used_idx  = vq->used->idx;
 676                 vq->last_avail_idx = vq->used->idx;
 677         }
 678
 679         vq->log_guest_addr = addr->log_guest_addr;
 680
 681         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
 682                         dev->vid, vq->desc);
 683         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
 684                         dev->vid, vq->avail);
 685         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
 686                         dev->vid, vq->used);
 687         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
 688                         dev->vid, vq->log_guest_addr);
 689
 690         return dev;
 691 }
 692
 693 /*
 694  * The virtio device sends us the desc, used and avail ring addresses.
 695  * This function then converts these to our address space.
 696  */
 697 static int
 698 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg,
 699                         int main_fd __rte_unused)
 700 {
 701         struct virtio_net *dev = *pdev;
 702         struct vhost_virtqueue *vq;
 703         struct vhost_vring_addr *addr = &msg->payload.addr;
 704
 705         if (dev->mem == NULL)
 706                 return RTE_VHOST_MSG_RESULT_ERR;
 707
 708         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 709         vq = dev->virtqueue[msg->payload.addr.index];
 710
 711         /*
 712          * Rings addresses should not be interpreted as long as the ring is not
 713          * started and enabled
 714          */
 715         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 716
 717         vring_invalidate(dev, vq);
 718
 719         if (vq->enabled && (dev->features &
 720                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
 721                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 722                 if (!dev)
 723                         return RTE_VHOST_MSG_RESULT_ERR;
 724
 725                 *pdev = dev;
 726         }
 727
 728         return RTE_VHOST_MSG_RESULT_OK;
 729 }
 730
 731 /*
 732  * The virtio device sends us the available ring last used index.
 733  */
 734 static int
 735 vhost_user_set_vring_base(struct virtio_net **pdev,
 736                         struct VhostUserMsg *msg,
 737                         int main_fd __rte_unused)
 738 {
 739         struct virtio_net *dev = *pdev;
 740         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 741         uint64_t val = msg->payload.state.num;
 742
 743         if (vq_is_packed(dev)) {
 744                 /*
 745                  * Bit[0:14]: avail index
 746                  * Bit[15]: avail wrap counter
 747                  */
 748                 vq->last_avail_idx = val & 0x7fff;
 749                 vq->avail_wrap_counter = !!(val & (0x1 << 15));
 750                 /*
 751                  * Set used index to same value as available one, as
 752                  * their values should be the same since ring processing
 753                  * was stopped at get time.
 754                  */
 755                 vq->last_used_idx = vq->last_avail_idx;
 756                 vq->used_wrap_counter = vq->avail_wrap_counter;
 757         } else {
 758                 vq->last_used_idx = msg->payload.state.num;
 759                 vq->last_avail_idx = msg->payload.state.num;
 760         }
 761
 762         return RTE_VHOST_MSG_RESULT_OK;
 763 }
 764
 765 static int
 766 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 767                    uint64_t host_phys_addr, uint64_t size)
 768 {
 769         struct guest_page *page, *last_page;
 770         struct guest_page *old_pages;
 771
 772         if (dev->nr_guest_pages == dev->max_guest_pages) {
 773                 dev->max_guest_pages *= 2;
 774                 old_pages = dev->guest_pages;
 775                 dev->guest_pages = realloc(dev->guest_pages,
 776                                         dev->max_guest_pages * sizeof(*page));
 777                 if (!dev->guest_pages) {
 778                         RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
 779                         free(old_pages);
 780                         return -1;
 781                 }
 782         }
 783
 784         if (dev->nr_guest_pages > 0) {
 785                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 786                 /* merge if the two pages are continuous */
 787                 if (host_phys_addr == last_page->host_phys_addr +
 788                                       last_page->size) {
 789                         last_page->size += size;
 790                         return 0;
 791                 }
 792         }
 793
 794         page = &dev->guest_pages[dev->nr_guest_pages++];
 795         page->guest_phys_addr = guest_phys_addr;
 796         page->host_phys_addr  = host_phys_addr;
 797         page->size = size;
 798
 799         return 0;
 800 }
 801
 802 static int
 803 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 804                 uint64_t page_size)
 805 {
 806         uint64_t reg_size = reg->size;
 807         uint64_t host_user_addr  = reg->host_user_addr;
 808         uint64_t guest_phys_addr = reg->guest_phys_addr;
 809         uint64_t host_phys_addr;
 810         uint64_t size;
 811
 812         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 813         size = page_size - (guest_phys_addr & (page_size - 1));
 814         size = RTE_MIN(size, reg_size);
 815
 816         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 817                 return -1;
 818
 819         host_user_addr  += size;
 820         guest_phys_addr += size;
 821         reg_size -= size;
 822
 823         while (reg_size > 0) {
 824                 size = RTE_MIN(reg_size, page_size);
 825                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 826                                                   host_user_addr);
 827                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 828                                 size) < 0)
 829                         return -1;
 830
 831                 host_user_addr  += size;
 832                 guest_phys_addr += size;
 833                 reg_size -= size;
 834         }
 835
 836         return 0;
 837 }
 838
 839 #ifdef RTE_LIBRTE_VHOST_DEBUG
 840 /* TODO: enable it only in debug mode? */
 841 static void
 842 dump_guest_pages(struct virtio_net *dev)
 843 {
 844         uint32_t i;
 845         struct guest_page *page;
 846
 847         for (i = 0; i < dev->nr_guest_pages; i++) {
 848                 page = &dev->guest_pages[i];
 849
 850                 RTE_LOG(INFO, VHOST_CONFIG,
 851                         "guest physical page region %u\n"
 852                         "\t guest_phys_addr: %" PRIx64 "\n"
 853                         "\t host_phys_addr : %" PRIx64 "\n"
 854                         "\t size           : %" PRIx64 "\n",
 855                         i,
 856                         page->guest_phys_addr,
 857                         page->host_phys_addr,
 858                         page->size);
 859         }
 860 }
 861 #else
 862 #define dump_guest_pages(dev)
 863 #endif
 864
 865 static bool
 866 vhost_memory_changed(struct VhostUserMemory *new,
 867                      struct rte_vhost_memory *old)
 868 {
 869         uint32_t i;
 870
 871         if (new->nregions != old->nregions)
 872                 return true;
 873
 874         for (i = 0; i < new->nregions; ++i) {
 875                 VhostUserMemoryRegion *new_r = &new->regions[i];
 876                 struct rte_vhost_mem_region *old_r = &old->regions[i];
 877
 878                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
 879                         return true;
 880                 if (new_r->memory_size != old_r->size)
 881                         return true;
 882                 if (new_r->userspace_addr != old_r->guest_user_addr)
 883                         return true;
 884         }
 885
 886         return false;
 887 }
 888
 889 static int
 890 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
 891                         int main_fd)
 892 {
 893         struct virtio_net *dev = *pdev;
 894         struct VhostUserMemory *memory = &msg->payload.memory;
 895         struct rte_vhost_mem_region *reg;
 896         void *mmap_addr;
 897         uint64_t mmap_size;
 898         uint64_t mmap_offset;
 899         uint64_t alignment;
 900         uint32_t i;
 901         int populate;
 902         int fd;
 903
 904         if (memory->nregions > VHOST_MEMORY_MAX_NREGIONS) {
 905                 RTE_LOG(ERR, VHOST_CONFIG,
 906                         "too many memory regions (%u)\n", memory->nregions);
 907                 return RTE_VHOST_MSG_RESULT_ERR;
 908         }
 909
 910         if (dev->mem && !vhost_memory_changed(memory, dev->mem)) {
 911                 RTE_LOG(INFO, VHOST_CONFIG,
 912                         "(%d) memory regions not changed\n", dev->vid);
 913
 914                 for (i = 0; i < memory->nregions; i++)
 915                         close(msg->fds[i]);
 916
 917                 return RTE_VHOST_MSG_RESULT_OK;
 918         }
 919
 920         if (dev->mem) {
 921                 free_mem_region(dev);
 922                 rte_free(dev->mem);
 923                 dev->mem = NULL;
 924         }
 925
 926         /* Flush IOTLB cache as previous HVAs are now invalid */
 927         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 928                 for (i = 0; i < dev->nr_vring; i++)
 929                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
 930
 931         dev->nr_guest_pages = 0;
 932         if (!dev->guest_pages) {
 933                 dev->max_guest_pages = 8;
 934                 dev->guest_pages = malloc(dev->max_guest_pages *
 935                                                 sizeof(struct guest_page));
 936                 if (dev->guest_pages == NULL) {
 937                         RTE_LOG(ERR, VHOST_CONFIG,
 938                                 "(%d) failed to allocate memory "
 939                                 "for dev->guest_pages\n",
 940                                 dev->vid);
 941                         return RTE_VHOST_MSG_RESULT_ERR;
 942                 }
 943         }
 944
 945         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
 946                 sizeof(struct rte_vhost_mem_region) * memory->nregions, 0);
 947         if (dev->mem == NULL) {
 948                 RTE_LOG(ERR, VHOST_CONFIG,
 949                         "(%d) failed to allocate memory for dev->mem\n",
 950                         dev->vid);
 951                 return RTE_VHOST_MSG_RESULT_ERR;
 952         }
 953         dev->mem->nregions = memory->nregions;
 954
 955         for (i = 0; i < memory->nregions; i++) {
 956                 fd  = msg->fds[i];
 957                 reg = &dev->mem->regions[i];
 958
 959                 reg->guest_phys_addr = memory->regions[i].guest_phys_addr;
 960                 reg->guest_user_addr = memory->regions[i].userspace_addr;
 961                 reg->size            = memory->regions[i].memory_size;
 962                 reg->fd              = fd;
 963
 964                 mmap_offset = memory->regions[i].mmap_offset;
 965
 966                 /* Check for memory_size + mmap_offset overflow */
 967                 if (mmap_offset >= -reg->size) {
 968                         RTE_LOG(ERR, VHOST_CONFIG,
 969                                 "mmap_offset (%#"PRIx64") and memory_size "
 970                                 "(%#"PRIx64") overflow\n",
 971                                 mmap_offset, reg->size);
 972                         goto err_mmap;
 973                 }
 974
 975                 mmap_size = reg->size + mmap_offset;
 976
 977                 /* mmap() without flag of MAP_ANONYMOUS, should be called
 978                  * with length argument aligned with hugepagesz at older
 979                  * longterm version Linux, like 2.6.32 and 3.2.72, or
 980                  * mmap() will fail with EINVAL.
 981                  *
 982                  * to avoid failure, make sure in caller to keep length
 983                  * aligned.
 984                  */
 985                 alignment = get_blk_size(fd);
 986                 if (alignment == (uint64_t)-1) {
 987                         RTE_LOG(ERR, VHOST_CONFIG,
 988                                 "couldn't get hugepage size through fstat\n");
 989                         goto err_mmap;
 990                 }
 991                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
 992
 993                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
 994                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
 995                                  MAP_SHARED | populate, fd, 0);
 996
 997                 if (mmap_addr == MAP_FAILED) {
 998                         RTE_LOG(ERR, VHOST_CONFIG,
 999                                 "mmap region %u failed.\n", i);
1000                         goto err_mmap;
1001                 }
1002
1003                 reg->mmap_addr = mmap_addr;
1004                 reg->mmap_size = mmap_size;
1005                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
1006                                       mmap_offset;
1007
1008                 if (dev->dequeue_zero_copy)
1009                         if (add_guest_pages(dev, reg, alignment) < 0) {
1010                                 RTE_LOG(ERR, VHOST_CONFIG,
1011                                         "adding guest pages to region %u failed.\n",
1012                                         i);
1013                                 goto err_mmap;
1014                         }
1015
1016                 RTE_LOG(INFO, VHOST_CONFIG,
1017                         "guest memory region %u, size: 0x%" PRIx64 "\n"
1018                         "\t guest physical addr: 0x%" PRIx64 "\n"
1019                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
1020                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
1021                         "\t mmap addr : 0x%" PRIx64 "\n"
1022                         "\t mmap size : 0x%" PRIx64 "\n"
1023                         "\t mmap align: 0x%" PRIx64 "\n"
1024                         "\t mmap off  : 0x%" PRIx64 "\n",
1025                         i, reg->size,
1026                         reg->guest_phys_addr,
1027                         reg->guest_user_addr,
1028                         reg->host_user_addr,
1029                         (uint64_t)(uintptr_t)mmap_addr,
1030                         mmap_size,
1031                         alignment,
1032                         mmap_offset);
1033
1034                 if (dev->postcopy_listening) {
1035                         /*
1036                          * We haven't a better way right now than sharing
1037                          * DPDK's virtual address with Qemu, so that Qemu can
1038                          * retrieve the region offset when handling userfaults.
1039                          */
1040                         memory->regions[i].userspace_addr =
1041                                 reg->host_user_addr;
1042                 }
1043         }
1044         if (dev->postcopy_listening) {
1045                 /* Send the addresses back to qemu */
1046                 msg->fd_num = 0;
1047                 send_vhost_reply(main_fd, msg);
1048
1049                 /* Wait for qemu to acknolwedge it's got the addresses
1050                  * we've got to wait before we're allowed to generate faults.
1051                  */
1052                 VhostUserMsg ack_msg;
1053                 if (read_vhost_message(main_fd, &ack_msg) <= 0) {
1054                         RTE_LOG(ERR, VHOST_CONFIG,
1055                                 "Failed to read qemu ack on postcopy set-mem-table\n");
1056                         goto err_mmap;
1057                 }
1058                 if (ack_msg.request.master != VHOST_USER_SET_MEM_TABLE) {
1059                         RTE_LOG(ERR, VHOST_CONFIG,
1060                                 "Bad qemu ack on postcopy set-mem-table (%d)\n",
1061                                 ack_msg.request.master);
1062                         goto err_mmap;
1063                 }
1064
1065                 /* Now userfault register and we can use the memory */
1066                 for (i = 0; i < memory->nregions; i++) {
1067 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1068                         reg = &dev->mem->regions[i];
1069                         struct uffdio_register reg_struct;
1070
1071                         /*
1072                          * Let's register all the mmap'ed area to ensure
1073                          * alignment on page boundary.
1074                          */
1075                         reg_struct.range.start =
1076                                 (uint64_t)(uintptr_t)reg->mmap_addr;
1077                         reg_struct.range.len = reg->mmap_size;
1078                         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
1079
1080                         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER,
1081                                                 &reg_struct)) {
1082                                 RTE_LOG(ERR, VHOST_CONFIG,
1083                                         "Failed to register ufd for region %d: (ufd = %d) %s\n",
1084                                         i, dev->postcopy_ufd,
1085                                         strerror(errno));
1086                                 goto err_mmap;
1087                         }
1088                         RTE_LOG(INFO, VHOST_CONFIG,
1089                                 "\t userfaultfd registered for range : %llx - %llx\n",
1090                                 reg_struct.range.start,
1091                                 reg_struct.range.start +
1092                                 reg_struct.range.len - 1);
1093 #else
1094                         goto err_mmap;
1095 #endif
1096                 }
1097         }
1098
1099         for (i = 0; i < dev->nr_vring; i++) {
1100                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1101
1102                 if (vq->desc || vq->avail || vq->used) {
1103                         /*
1104                          * If the memory table got updated, the ring addresses
1105                          * need to be translated again as virtual addresses have
1106                          * changed.
1107                          */
1108                         vring_invalidate(dev, vq);
1109
1110                         dev = translate_ring_addresses(dev, i);
1111                         if (!dev) {
1112                                 dev = *pdev;
1113                                 goto err_mmap;
1114                         }
1115
1116                         *pdev = dev;
1117                 }
1118         }
1119
1120         dump_guest_pages(dev);
1121
1122         return RTE_VHOST_MSG_RESULT_OK;
1123
1124 err_mmap:
1125         free_mem_region(dev);
1126         rte_free(dev->mem);
1127         dev->mem = NULL;
1128         return RTE_VHOST_MSG_RESULT_ERR;
1129 }
1130
1131 static bool
1132 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
1133 {
1134         bool rings_ok;
1135
1136         if (!vq)
1137                 return false;
1138
1139         if (vq_is_packed(dev))
1140                 rings_ok = !!vq->desc_packed;
1141         else
1142                 rings_ok = vq->desc && vq->avail && vq->used;
1143
1144         return rings_ok &&
1145                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
1146                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1147 }
1148
1149 static int
1150 virtio_is_ready(struct virtio_net *dev)
1151 {
1152         struct vhost_virtqueue *vq;
1153         uint32_t i;
1154
1155         if (dev->nr_vring == 0)
1156                 return 0;
1157
1158         for (i = 0; i < dev->nr_vring; i++) {
1159                 vq = dev->virtqueue[i];
1160
1161                 if (!vq_is_ready(dev, vq))
1162                         return 0;
1163         }
1164
1165         RTE_LOG(INFO, VHOST_CONFIG,
1166                 "virtio is now ready for processing.\n");
1167         return 1;
1168 }
1169
1170 static int
1171 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg,
1172                         int main_fd __rte_unused)
1173 {
1174         struct virtio_net *dev = *pdev;
1175         struct vhost_vring_file file;
1176         struct vhost_virtqueue *vq;
1177
1178         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1179         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1180                 file.fd = VIRTIO_INVALID_EVENTFD;
1181         else
1182                 file.fd = msg->fds[0];
1183         RTE_LOG(INFO, VHOST_CONFIG,
1184                 "vring call idx:%d file:%d\n", file.index, file.fd);
1185
1186         vq = dev->virtqueue[file.index];
1187         if (vq->callfd >= 0)
1188                 close(vq->callfd);
1189
1190         vq->callfd = file.fd;
1191
1192         return RTE_VHOST_MSG_RESULT_OK;
1193 }
1194
1195 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1196                         struct VhostUserMsg *msg,
1197                         int main_fd __rte_unused)
1198 {
1199         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1200                 close(msg->fds[0]);
1201         RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1202
1203         return RTE_VHOST_MSG_RESULT_OK;
1204 }
1205
1206 static int
1207 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg,
1208                         int main_fd __rte_unused)
1209 {
1210         struct virtio_net *dev = *pdev;
1211         struct vhost_vring_file file;
1212         struct vhost_virtqueue *vq;
1213
1214         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1215         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1216                 file.fd = VIRTIO_INVALID_EVENTFD;
1217         else
1218                 file.fd = msg->fds[0];
1219         RTE_LOG(INFO, VHOST_CONFIG,
1220                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1221
1222         /* Interpret ring addresses only when ring is started. */
1223         dev = translate_ring_addresses(dev, file.index);
1224         if (!dev)
1225                 return RTE_VHOST_MSG_RESULT_ERR;
1226
1227         *pdev = dev;
1228
1229         vq = dev->virtqueue[file.index];
1230
1231         /*
1232          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1233          * the ring starts already enabled. Otherwise, it is enabled via
1234          * the SET_VRING_ENABLE message.
1235          */
1236         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
1237                 vq->enabled = 1;
1238                 if (dev->notify_ops->vring_state_changed)
1239                         dev->notify_ops->vring_state_changed(
1240                                 dev->vid, file.index, 1);
1241         }
1242
1243         if (vq->kickfd >= 0)
1244                 close(vq->kickfd);
1245         vq->kickfd = file.fd;
1246
1247         return RTE_VHOST_MSG_RESULT_OK;
1248 }
1249
1250 static void
1251 free_zmbufs(struct vhost_virtqueue *vq)
1252 {
1253         drain_zmbuf_list(vq);
1254
1255         rte_free(vq->zmbufs);
1256 }
1257
1258 /*
1259  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1260  */
1261 static int
1262 vhost_user_get_vring_base(struct virtio_net **pdev,
1263                         struct VhostUserMsg *msg,
1264                         int main_fd __rte_unused)
1265 {
1266         struct virtio_net *dev = *pdev;
1267         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1268         uint64_t val;
1269
1270         /* We have to stop the queue (virtio) if it is running. */
1271         vhost_destroy_device_notify(dev);
1272
1273         dev->flags &= ~VIRTIO_DEV_READY;
1274         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1275
1276         /* Here we are safe to get the indexes */
1277         if (vq_is_packed(dev)) {
1278                 /*
1279                  * Bit[0:14]: avail index
1280                  * Bit[15]: avail wrap counter
1281                  */
1282                 val = vq->last_avail_idx & 0x7fff;
1283                 val |= vq->avail_wrap_counter << 15;
1284                 msg->payload.state.num = val;
1285         } else {
1286                 msg->payload.state.num = vq->last_avail_idx;
1287         }
1288
1289         RTE_LOG(INFO, VHOST_CONFIG,
1290                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1291                 msg->payload.state.num);
1292         /*
1293          * Based on current qemu vhost-user implementation, this message is
1294          * sent and only sent in vhost_vring_stop.
1295          * TODO: cleanup the vring, it isn't usable since here.
1296          */
1297         if (vq->kickfd >= 0)
1298                 close(vq->kickfd);
1299
1300         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1301
1302         if (vq->callfd >= 0)
1303                 close(vq->callfd);
1304
1305         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1306
1307         vq->signalled_used_valid = false;
1308
1309         if (dev->dequeue_zero_copy)
1310                 free_zmbufs(vq);
1311         if (vq_is_packed(dev)) {
1312                 rte_free(vq->shadow_used_packed);
1313                 vq->shadow_used_packed = NULL;
1314         } else {
1315                 rte_free(vq->shadow_used_split);
1316                 vq->shadow_used_split = NULL;
1317         }
1318
1319         rte_free(vq->batch_copy_elems);
1320         vq->batch_copy_elems = NULL;
1321
1322         msg->size = sizeof(msg->payload.state);
1323         msg->fd_num = 0;
1324
1325         return RTE_VHOST_MSG_RESULT_REPLY;
1326 }
1327
1328 /*
1329  * when virtio queues are ready to work, qemu will send us to
1330  * enable the virtio queue pair.
1331  */
1332 static int
1333 vhost_user_set_vring_enable(struct virtio_net **pdev,
1334                         struct VhostUserMsg *msg,
1335                         int main_fd __rte_unused)
1336 {
1337         struct virtio_net *dev = *pdev;
1338         int enable = (int)msg->payload.state.num;
1339         int index = (int)msg->payload.state.index;
1340         struct rte_vdpa_device *vdpa_dev;
1341         int did = -1;
1342
1343         RTE_LOG(INFO, VHOST_CONFIG,
1344                 "set queue enable: %d to qp idx: %d\n",
1345                 enable, index);
1346
1347         did = dev->vdpa_dev_id;
1348         vdpa_dev = rte_vdpa_get_device(did);
1349         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1350                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1351
1352         if (dev->notify_ops->vring_state_changed)
1353                 dev->notify_ops->vring_state_changed(dev->vid,
1354                                 index, enable);
1355
1356         /* On disable, rings have to be stopped being processed. */
1357         if (!enable && dev->dequeue_zero_copy)
1358                 drain_zmbuf_list(dev->virtqueue[index]);
1359
1360         dev->virtqueue[index]->enabled = enable;
1361
1362         return RTE_VHOST_MSG_RESULT_OK;
1363 }
1364
1365 static int
1366 vhost_user_get_protocol_features(struct virtio_net **pdev,
1367                         struct VhostUserMsg *msg,
1368                         int main_fd __rte_unused)
1369 {
1370         struct virtio_net *dev = *pdev;
1371         uint64_t features, protocol_features;
1372
1373         rte_vhost_driver_get_features(dev->ifname, &features);
1374         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1375
1376         /*
1377          * REPLY_ACK protocol feature is only mandatory for now
1378          * for IOMMU feature. If IOMMU is explicitly disabled by the
1379          * application, disable also REPLY_ACK feature for older buggy
1380          * Qemu versions (from v2.7.0 to v2.9.0).
1381          */
1382         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1383                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1384
1385         msg->payload.u64 = protocol_features;
1386         msg->size = sizeof(msg->payload.u64);
1387         msg->fd_num = 0;
1388
1389         return RTE_VHOST_MSG_RESULT_REPLY;
1390 }
1391
1392 static int
1393 vhost_user_set_protocol_features(struct virtio_net **pdev,
1394                         struct VhostUserMsg *msg,
1395                         int main_fd __rte_unused)
1396 {
1397         struct virtio_net *dev = *pdev;
1398         uint64_t protocol_features = msg->payload.u64;
1399         uint64_t slave_protocol_features = 0;
1400
1401         rte_vhost_driver_get_protocol_features(dev->ifname,
1402                         &slave_protocol_features);
1403         if (protocol_features & ~slave_protocol_features) {
1404                 RTE_LOG(ERR, VHOST_CONFIG,
1405                         "(%d) received invalid protocol features.\n",
1406                         dev->vid);
1407                 return RTE_VHOST_MSG_RESULT_ERR;
1408         }
1409
1410         dev->protocol_features = protocol_features;
1411         RTE_LOG(INFO, VHOST_CONFIG,
1412                 "negotiated Vhost-user protocol features: 0x%" PRIx64 "\n",
1413                 dev->protocol_features);
1414
1415         return RTE_VHOST_MSG_RESULT_OK;
1416 }
1417
1418 static int
1419 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg,
1420                         int main_fd __rte_unused)
1421 {
1422         struct virtio_net *dev = *pdev;
1423         int fd = msg->fds[0];
1424         uint64_t size, off;
1425         void *addr;
1426
1427         if (fd < 0) {
1428                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1429                 return RTE_VHOST_MSG_RESULT_ERR;
1430         }
1431
1432         if (msg->size != sizeof(VhostUserLog)) {
1433                 RTE_LOG(ERR, VHOST_CONFIG,
1434                         "invalid log base msg size: %"PRId32" != %d\n",
1435                         msg->size, (int)sizeof(VhostUserLog));
1436                 return RTE_VHOST_MSG_RESULT_ERR;
1437         }
1438
1439         size = msg->payload.log.mmap_size;
1440         off  = msg->payload.log.mmap_offset;
1441
1442         /* Don't allow mmap_offset to point outside the mmap region */
1443         if (off > size) {
1444                 RTE_LOG(ERR, VHOST_CONFIG,
1445                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1446                         off, size);
1447                 return RTE_VHOST_MSG_RESULT_ERR;
1448         }
1449
1450         RTE_LOG(INFO, VHOST_CONFIG,
1451                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1452                 size, off);
1453
1454         /*
1455          * mmap from 0 to workaround a hugepage mmap bug: mmap will
1456          * fail when offset is not page size aligned.
1457          */
1458         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1459         close(fd);
1460         if (addr == MAP_FAILED) {
1461                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1462                 return RTE_VHOST_MSG_RESULT_ERR;
1463         }
1464
1465         /*
1466          * Free previously mapped log memory on occasionally
1467          * multiple VHOST_USER_SET_LOG_BASE.
1468          */
1469         if (dev->log_addr) {
1470                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1471         }
1472         dev->log_addr = (uint64_t)(uintptr_t)addr;
1473         dev->log_base = dev->log_addr + off;
1474         dev->log_size = size;
1475
1476         /*
1477          * The spec is not clear about it (yet), but QEMU doesn't expect
1478          * any payload in the reply.
1479          */
1480         msg->size = 0;
1481         msg->fd_num = 0;
1482
1483         return RTE_VHOST_MSG_RESULT_REPLY;
1484 }
1485
1486 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1487                         struct VhostUserMsg *msg,
1488                         int main_fd __rte_unused)
1489 {
1490         close(msg->fds[0]);
1491         RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1492
1493         return RTE_VHOST_MSG_RESULT_OK;
1494 }
1495
1496 /*
1497  * An rarp packet is constructed and broadcasted to notify switches about
1498  * the new location of the migrated VM, so that packets from outside will
1499  * not be lost after migration.
1500  *
1501  * However, we don't actually "send" a rarp packet here, instead, we set
1502  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1503  */
1504 static int
1505 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg,
1506                         int main_fd __rte_unused)
1507 {
1508         struct virtio_net *dev = *pdev;
1509         uint8_t *mac = (uint8_t *)&msg->payload.u64;
1510         struct rte_vdpa_device *vdpa_dev;
1511         int did = -1;
1512
1513         RTE_LOG(DEBUG, VHOST_CONFIG,
1514                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1515                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1516         memcpy(dev->mac.addr_bytes, mac, 6);
1517
1518         /*
1519          * Set the flag to inject a RARP broadcast packet at
1520          * rte_vhost_dequeue_burst().
1521          *
1522          * rte_smp_wmb() is for making sure the mac is copied
1523          * before the flag is set.
1524          */
1525         rte_smp_wmb();
1526         rte_atomic16_set(&dev->broadcast_rarp, 1);
1527         did = dev->vdpa_dev_id;
1528         vdpa_dev = rte_vdpa_get_device(did);
1529         if (vdpa_dev && vdpa_dev->ops->migration_done)
1530                 vdpa_dev->ops->migration_done(dev->vid);
1531
1532         return RTE_VHOST_MSG_RESULT_OK;
1533 }
1534
1535 static int
1536 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg,
1537                         int main_fd __rte_unused)
1538 {
1539         struct virtio_net *dev = *pdev;
1540         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1541                         msg->payload.u64 > VIRTIO_MAX_MTU) {
1542                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1543                                 msg->payload.u64);
1544
1545                 return RTE_VHOST_MSG_RESULT_ERR;
1546         }
1547
1548         dev->mtu = msg->payload.u64;
1549
1550         return RTE_VHOST_MSG_RESULT_OK;
1551 }
1552
1553 static int
1554 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg,
1555                         int main_fd __rte_unused)
1556 {
1557         struct virtio_net *dev = *pdev;
1558         int fd = msg->fds[0];
1559
1560         if (fd < 0) {
1561                 RTE_LOG(ERR, VHOST_CONFIG,
1562                                 "Invalid file descriptor for slave channel (%d)\n",
1563                                 fd);
1564                 return RTE_VHOST_MSG_RESULT_ERR;
1565         }
1566
1567         if (dev->slave_req_fd >= 0)
1568                 close(dev->slave_req_fd);
1569
1570         dev->slave_req_fd = fd;
1571
1572         return RTE_VHOST_MSG_RESULT_OK;
1573 }
1574
1575 static int
1576 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1577 {
1578         struct vhost_vring_addr *ra;
1579         uint64_t start, end;
1580
1581         start = imsg->iova;
1582         end = start + imsg->size;
1583
1584         ra = &vq->ring_addrs;
1585         if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1586                 return 1;
1587         if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1588                 return 1;
1589         if (ra->used_user_addr >= start && ra->used_user_addr < end)
1590                 return 1;
1591
1592         return 0;
1593 }
1594
1595 static int
1596 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1597                                 struct vhost_iotlb_msg *imsg)
1598 {
1599         uint64_t istart, iend, vstart, vend;
1600
1601         istart = imsg->iova;
1602         iend = istart + imsg->size - 1;
1603
1604         vstart = (uintptr_t)vq->desc;
1605         vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1606         if (vstart <= iend && istart <= vend)
1607                 return 1;
1608
1609         vstart = (uintptr_t)vq->avail;
1610         vend = vstart + sizeof(struct vring_avail);
1611         vend += sizeof(uint16_t) * vq->size - 1;
1612         if (vstart <= iend && istart <= vend)
1613                 return 1;
1614
1615         vstart = (uintptr_t)vq->used;
1616         vend = vstart + sizeof(struct vring_used);
1617         vend += sizeof(struct vring_used_elem) * vq->size - 1;
1618         if (vstart <= iend && istart <= vend)
1619                 return 1;
1620
1621         return 0;
1622 }
1623
1624 static int
1625 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg,
1626                         int main_fd __rte_unused)
1627 {
1628         struct virtio_net *dev = *pdev;
1629         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1630         uint16_t i;
1631         uint64_t vva, len;
1632
1633         switch (imsg->type) {
1634         case VHOST_IOTLB_UPDATE:
1635                 len = imsg->size;
1636                 vva = qva_to_vva(dev, imsg->uaddr, &len);
1637                 if (!vva)
1638                         return RTE_VHOST_MSG_RESULT_ERR;
1639
1640                 for (i = 0; i < dev->nr_vring; i++) {
1641                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1642
1643                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1644                                         len, imsg->perm);
1645
1646                         if (is_vring_iotlb_update(vq, imsg))
1647                                 *pdev = dev = translate_ring_addresses(dev, i);
1648                 }
1649                 break;
1650         case VHOST_IOTLB_INVALIDATE:
1651                 for (i = 0; i < dev->nr_vring; i++) {
1652                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1653
1654                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
1655                                         imsg->size);
1656
1657                         if (is_vring_iotlb_invalidate(vq, imsg))
1658                                 vring_invalidate(dev, vq);
1659                 }
1660                 break;
1661         default:
1662                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1663                                 imsg->type);
1664                 return RTE_VHOST_MSG_RESULT_ERR;
1665         }
1666
1667         return RTE_VHOST_MSG_RESULT_OK;
1668 }
1669
1670 static int
1671 vhost_user_set_postcopy_advise(struct virtio_net **pdev,
1672                         struct VhostUserMsg *msg,
1673                         int main_fd __rte_unused)
1674 {
1675         struct virtio_net *dev = *pdev;
1676 #ifdef RTE_LIBRTE_VHOST_POSTCOPY
1677         struct uffdio_api api_struct;
1678
1679         dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1680
1681         if (dev->postcopy_ufd == -1) {
1682                 RTE_LOG(ERR, VHOST_CONFIG, "Userfaultfd not available: %s\n",
1683                         strerror(errno));
1684                 return RTE_VHOST_MSG_RESULT_ERR;
1685         }
1686         api_struct.api = UFFD_API;
1687         api_struct.features = 0;
1688         if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
1689                 RTE_LOG(ERR, VHOST_CONFIG, "UFFDIO_API ioctl failure: %s\n",
1690                         strerror(errno));
1691                 close(dev->postcopy_ufd);
1692                 dev->postcopy_ufd = -1;
1693                 return RTE_VHOST_MSG_RESULT_ERR;
1694         }
1695         msg->fds[0] = dev->postcopy_ufd;
1696         msg->fd_num = 1;
1697
1698         return RTE_VHOST_MSG_RESULT_REPLY;
1699 #else
1700         dev->postcopy_ufd = -1;
1701         msg->fd_num = 0;
1702
1703         return RTE_VHOST_MSG_RESULT_ERR;
1704 #endif
1705 }
1706
1707 static int
1708 vhost_user_set_postcopy_listen(struct virtio_net **pdev,
1709                         struct VhostUserMsg *msg __rte_unused,
1710                         int main_fd __rte_unused)
1711 {
1712         struct virtio_net *dev = *pdev;
1713
1714         if (dev->mem && dev->mem->nregions) {
1715                 RTE_LOG(ERR, VHOST_CONFIG,
1716                         "Regions already registered at postcopy-listen\n");
1717                 return RTE_VHOST_MSG_RESULT_ERR;
1718         }
1719         dev->postcopy_listening = 1;
1720
1721         return RTE_VHOST_MSG_RESULT_OK;
1722 }
1723
1724 static int
1725 vhost_user_postcopy_end(struct virtio_net **pdev, struct VhostUserMsg *msg,
1726                         int main_fd __rte_unused)
1727 {
1728         struct virtio_net *dev = *pdev;
1729
1730         dev->postcopy_listening = 0;
1731         if (dev->postcopy_ufd >= 0) {
1732                 close(dev->postcopy_ufd);
1733                 dev->postcopy_ufd = -1;
1734         }
1735
1736         msg->payload.u64 = 0;
1737         msg->size = sizeof(msg->payload.u64);
1738         msg->fd_num = 0;
1739
1740         return RTE_VHOST_MSG_RESULT_REPLY;
1741 }
1742
1743 typedef int (*vhost_message_handler_t)(struct virtio_net **pdev,
1744                                         struct VhostUserMsg *msg,
1745                                         int main_fd);
1746 static vhost_message_handler_t vhost_message_handlers[VHOST_USER_MAX] = {
1747         [VHOST_USER_NONE] = NULL,
1748         [VHOST_USER_GET_FEATURES] = vhost_user_get_features,
1749         [VHOST_USER_SET_FEATURES] = vhost_user_set_features,
1750         [VHOST_USER_SET_OWNER] = vhost_user_set_owner,
1751         [VHOST_USER_RESET_OWNER] = vhost_user_reset_owner,
1752         [VHOST_USER_SET_MEM_TABLE] = vhost_user_set_mem_table,
1753         [VHOST_USER_SET_LOG_BASE] = vhost_user_set_log_base,
1754         [VHOST_USER_SET_LOG_FD] = vhost_user_set_log_fd,
1755         [VHOST_USER_SET_VRING_NUM] = vhost_user_set_vring_num,
1756         [VHOST_USER_SET_VRING_ADDR] = vhost_user_set_vring_addr,
1757         [VHOST_USER_SET_VRING_BASE] = vhost_user_set_vring_base,
1758         [VHOST_USER_GET_VRING_BASE] = vhost_user_get_vring_base,
1759         [VHOST_USER_SET_VRING_KICK] = vhost_user_set_vring_kick,
1760         [VHOST_USER_SET_VRING_CALL] = vhost_user_set_vring_call,
1761         [VHOST_USER_SET_VRING_ERR] = vhost_user_set_vring_err,
1762         [VHOST_USER_GET_PROTOCOL_FEATURES] = vhost_user_get_protocol_features,
1763         [VHOST_USER_SET_PROTOCOL_FEATURES] = vhost_user_set_protocol_features,
1764         [VHOST_USER_GET_QUEUE_NUM] = vhost_user_get_queue_num,
1765         [VHOST_USER_SET_VRING_ENABLE] = vhost_user_set_vring_enable,
1766         [VHOST_USER_SEND_RARP] = vhost_user_send_rarp,
1767         [VHOST_USER_NET_SET_MTU] = vhost_user_net_set_mtu,
1768         [VHOST_USER_SET_SLAVE_REQ_FD] = vhost_user_set_req_fd,
1769         [VHOST_USER_IOTLB_MSG] = vhost_user_iotlb_msg,
1770         [VHOST_USER_POSTCOPY_ADVISE] = vhost_user_set_postcopy_advise,
1771         [VHOST_USER_POSTCOPY_LISTEN] = vhost_user_set_postcopy_listen,
1772         [VHOST_USER_POSTCOPY_END] = vhost_user_postcopy_end,
1773 };
1774
1775
1776 /* return bytes# of read on success or negative val on failure. */
1777 static int
1778 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1779 {
1780         int ret;
1781
1782         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1783                 msg->fds, VHOST_MEMORY_MAX_NREGIONS, &msg->fd_num);
1784         if (ret <= 0)
1785                 return ret;
1786
1787         if (msg->size) {
1788                 if (msg->size > sizeof(msg->payload)) {
1789                         RTE_LOG(ERR, VHOST_CONFIG,
1790                                 "invalid msg size: %d\n", msg->size);
1791                         return -1;
1792                 }
1793                 ret = read(sockfd, &msg->payload, msg->size);
1794                 if (ret <= 0)
1795                         return ret;
1796                 if (ret != (int)msg->size) {
1797                         RTE_LOG(ERR, VHOST_CONFIG,
1798                                 "read control message failed\n");
1799                         return -1;
1800                 }
1801         }
1802
1803         return ret;
1804 }
1805
1806 static int
1807 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1808 {
1809         if (!msg)
1810                 return 0;
1811
1812         return send_fd_message(sockfd, (char *)msg,
1813                 VHOST_USER_HDR_SIZE + msg->size, msg->fds, msg->fd_num);
1814 }
1815
1816 static int
1817 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1818 {
1819         if (!msg)
1820                 return 0;
1821
1822         msg->flags &= ~VHOST_USER_VERSION_MASK;
1823         msg->flags &= ~VHOST_USER_NEED_REPLY;
1824         msg->flags |= VHOST_USER_VERSION;
1825         msg->flags |= VHOST_USER_REPLY_MASK;
1826
1827         return send_vhost_message(sockfd, msg);
1828 }
1829
1830 static int
1831 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg)
1832 {
1833         int ret;
1834
1835         if (msg->flags & VHOST_USER_NEED_REPLY)
1836                 rte_spinlock_lock(&dev->slave_req_lock);
1837
1838         ret = send_vhost_message(dev->slave_req_fd, msg);
1839         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1840                 rte_spinlock_unlock(&dev->slave_req_lock);
1841
1842         return ret;
1843 }
1844
1845 /*
1846  * Allocate a queue pair if it hasn't been allocated yet
1847  */
1848 static int
1849 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1850                         struct VhostUserMsg *msg)
1851 {
1852         uint16_t vring_idx;
1853
1854         switch (msg->request.master) {
1855         case VHOST_USER_SET_VRING_KICK:
1856         case VHOST_USER_SET_VRING_CALL:
1857         case VHOST_USER_SET_VRING_ERR:
1858                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1859                 break;
1860         case VHOST_USER_SET_VRING_NUM:
1861         case VHOST_USER_SET_VRING_BASE:
1862         case VHOST_USER_SET_VRING_ENABLE:
1863                 vring_idx = msg->payload.state.index;
1864                 break;
1865         case VHOST_USER_SET_VRING_ADDR:
1866                 vring_idx = msg->payload.addr.index;
1867                 break;
1868         default:
1869                 return 0;
1870         }
1871
1872         if (vring_idx >= VHOST_MAX_VRING) {
1873                 RTE_LOG(ERR, VHOST_CONFIG,
1874                         "invalid vring index: %u\n", vring_idx);
1875                 return -1;
1876         }
1877
1878         if (dev->virtqueue[vring_idx])
1879                 return 0;
1880
1881         return alloc_vring_queue(dev, vring_idx);
1882 }
1883
1884 static void
1885 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1886 {
1887         unsigned int i = 0;
1888         unsigned int vq_num = 0;
1889
1890         while (vq_num < dev->nr_vring) {
1891                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1892
1893                 if (vq) {
1894                         rte_spinlock_lock(&vq->access_lock);
1895                         vq_num++;
1896                 }
1897                 i++;
1898         }
1899 }
1900
1901 static void
1902 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1903 {
1904         unsigned int i = 0;
1905         unsigned int vq_num = 0;
1906
1907         while (vq_num < dev->nr_vring) {
1908                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1909
1910                 if (vq) {
1911                         rte_spinlock_unlock(&vq->access_lock);
1912                         vq_num++;
1913                 }
1914                 i++;
1915         }
1916 }
1917
1918 int
1919 vhost_user_msg_handler(int vid, int fd)
1920 {
1921         struct virtio_net *dev;
1922         struct VhostUserMsg msg;
1923         struct rte_vdpa_device *vdpa_dev;
1924         int did = -1;
1925         int ret;
1926         int unlock_required = 0;
1927         bool handled;
1928         int request;
1929
1930         dev = get_device(vid);
1931         if (dev == NULL)
1932                 return -1;
1933
1934         if (!dev->notify_ops) {
1935                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1936                 if (!dev->notify_ops) {
1937                         RTE_LOG(ERR, VHOST_CONFIG,
1938                                 "failed to get callback ops for driver %s\n",
1939                                 dev->ifname);
1940                         return -1;
1941                 }
1942         }
1943
1944         ret = read_vhost_message(fd, &msg);
1945         if (ret <= 0) {
1946                 if (ret < 0)
1947                         RTE_LOG(ERR, VHOST_CONFIG,
1948                                 "vhost read message failed\n");
1949                 else
1950                         RTE_LOG(INFO, VHOST_CONFIG,
1951                                 "vhost peer closed\n");
1952
1953                 return -1;
1954         }
1955
1956         ret = 0;
1957         request = msg.request.master;
1958         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX &&
1959                         vhost_message_str[request]) {
1960                 if (request != VHOST_USER_IOTLB_MSG)
1961                         RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1962                                 vhost_message_str[request]);
1963                 else
1964                         RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1965                                 vhost_message_str[request]);
1966         } else {
1967                 RTE_LOG(DEBUG, VHOST_CONFIG, "External request %d\n", request);
1968         }
1969
1970         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1971         if (ret < 0) {
1972                 RTE_LOG(ERR, VHOST_CONFIG,
1973                         "failed to alloc queue\n");
1974                 return -1;
1975         }
1976
1977         /*
1978          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1979          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1980          * and device is destroyed. destroy_device waits for queues to be
1981          * inactive, so it is safe. Otherwise taking the access_lock
1982          * would cause a dead lock.
1983          */
1984         switch (request) {
1985         case VHOST_USER_SET_FEATURES:
1986         case VHOST_USER_SET_PROTOCOL_FEATURES:
1987         case VHOST_USER_SET_OWNER:
1988         case VHOST_USER_SET_MEM_TABLE:
1989         case VHOST_USER_SET_LOG_BASE:
1990         case VHOST_USER_SET_LOG_FD:
1991         case VHOST_USER_SET_VRING_NUM:
1992         case VHOST_USER_SET_VRING_ADDR:
1993         case VHOST_USER_SET_VRING_BASE:
1994         case VHOST_USER_SET_VRING_KICK:
1995         case VHOST_USER_SET_VRING_CALL:
1996         case VHOST_USER_SET_VRING_ERR:
1997         case VHOST_USER_SET_VRING_ENABLE:
1998         case VHOST_USER_SEND_RARP:
1999         case VHOST_USER_NET_SET_MTU:
2000         case VHOST_USER_SET_SLAVE_REQ_FD:
2001                 vhost_user_lock_all_queue_pairs(dev);
2002                 unlock_required = 1;
2003                 break;
2004         default:
2005                 break;
2006
2007         }
2008
2009         handled = false;
2010         if (dev->extern_ops.pre_msg_handle) {
2011                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
2012                                 (void *)&msg);
2013                 switch (ret) {
2014                 case RTE_VHOST_MSG_RESULT_REPLY:
2015                         send_vhost_reply(fd, &msg);
2016                         /* Fall-through */
2017                 case RTE_VHOST_MSG_RESULT_ERR:
2018                 case RTE_VHOST_MSG_RESULT_OK:
2019                         handled = true;
2020                         goto skip_to_post_handle;
2021                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2022                 default:
2023                         break;
2024                 }
2025         }
2026
2027         if (request > VHOST_USER_NONE && request < VHOST_USER_MAX) {
2028                 if (!vhost_message_handlers[request])
2029                         goto skip_to_post_handle;
2030                 ret = vhost_message_handlers[request](&dev, &msg, fd);
2031
2032                 switch (ret) {
2033                 case RTE_VHOST_MSG_RESULT_ERR:
2034                         RTE_LOG(ERR, VHOST_CONFIG,
2035                                 "Processing %s failed.\n",
2036                                 vhost_message_str[request]);
2037                         handled = true;
2038                         break;
2039                 case RTE_VHOST_MSG_RESULT_OK:
2040                         RTE_LOG(DEBUG, VHOST_CONFIG,
2041                                 "Processing %s succeeded.\n",
2042                                 vhost_message_str[request]);
2043                         handled = true;
2044                         break;
2045                 case RTE_VHOST_MSG_RESULT_REPLY:
2046                         RTE_LOG(DEBUG, VHOST_CONFIG,
2047                                 "Processing %s succeeded and needs reply.\n",
2048                                 vhost_message_str[request]);
2049                         send_vhost_reply(fd, &msg);
2050                         handled = true;
2051                         break;
2052                 default:
2053                         break;
2054                 }
2055         }
2056
2057 skip_to_post_handle:
2058         if (ret != RTE_VHOST_MSG_RESULT_ERR &&
2059                         dev->extern_ops.post_msg_handle) {
2060                 ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
2061                                 (void *)&msg);
2062                 switch (ret) {
2063                 case RTE_VHOST_MSG_RESULT_REPLY:
2064                         send_vhost_reply(fd, &msg);
2065                         /* Fall-through */
2066                 case RTE_VHOST_MSG_RESULT_ERR:
2067                 case RTE_VHOST_MSG_RESULT_OK:
2068                         handled = true;
2069                 case RTE_VHOST_MSG_RESULT_NOT_HANDLED:
2070                 default:
2071                         break;
2072                 }
2073         }
2074
2075         if (unlock_required)
2076                 vhost_user_unlock_all_queue_pairs(dev);
2077
2078         /* If message was not handled at this stage, treat it as an error */
2079         if (!handled) {
2080                 RTE_LOG(ERR, VHOST_CONFIG,
2081                         "vhost message (req: %d) was not handled.\n", request);
2082                 ret = RTE_VHOST_MSG_RESULT_ERR;
2083         }
2084
2085         /*
2086          * If the request required a reply that was already sent,
2087          * this optional reply-ack won't be sent as the
2088          * VHOST_USER_NEED_REPLY was cleared in send_vhost_reply().
2089          */
2090         if (msg.flags & VHOST_USER_NEED_REPLY) {
2091                 msg.payload.u64 = ret == RTE_VHOST_MSG_RESULT_ERR;
2092                 msg.size = sizeof(msg.payload.u64);
2093                 msg.fd_num = 0;
2094                 send_vhost_reply(fd, &msg);
2095         } else if (ret == RTE_VHOST_MSG_RESULT_ERR) {
2096                 RTE_LOG(ERR, VHOST_CONFIG,
2097                         "vhost message handling failed.\n");
2098                 return -1;
2099         }
2100
2101         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
2102                 dev->flags |= VIRTIO_DEV_READY;
2103
2104                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
2105                         if (dev->dequeue_zero_copy) {
2106                                 RTE_LOG(INFO, VHOST_CONFIG,
2107                                                 "dequeue zero copy is enabled\n");
2108                         }
2109
2110                         if (dev->notify_ops->new_device(dev->vid) == 0)
2111                                 dev->flags |= VIRTIO_DEV_RUNNING;
2112                 }
2113         }
2114
2115         did = dev->vdpa_dev_id;
2116         vdpa_dev = rte_vdpa_get_device(did);
2117         if (vdpa_dev && virtio_is_ready(dev) &&
2118                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
2119                         msg.request.master == VHOST_USER_SET_VRING_CALL) {
2120                 if (vdpa_dev->ops->dev_conf)
2121                         vdpa_dev->ops->dev_conf(dev->vid);
2122                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
2123         }
2124
2125         return 0;
2126 }
2127
2128 static int process_slave_message_reply(struct virtio_net *dev,
2129                                        const struct VhostUserMsg *msg)
2130 {
2131         struct VhostUserMsg msg_reply;
2132         int ret;
2133
2134         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
2135                 return 0;
2136
2137         if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
2138                 ret = -1;
2139                 goto out;
2140         }
2141
2142         if (msg_reply.request.slave != msg->request.slave) {
2143                 RTE_LOG(ERR, VHOST_CONFIG,
2144                         "Received unexpected msg type (%u), expected %u\n",
2145                         msg_reply.request.slave, msg->request.slave);
2146                 ret = -1;
2147                 goto out;
2148         }
2149
2150         ret = msg_reply.payload.u64 ? -1 : 0;
2151
2152 out:
2153         rte_spinlock_unlock(&dev->slave_req_lock);
2154         return ret;
2155 }
2156
2157 int
2158 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
2159 {
2160         int ret;
2161         struct VhostUserMsg msg = {
2162                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
2163                 .flags = VHOST_USER_VERSION,
2164                 .size = sizeof(msg.payload.iotlb),
2165                 .payload.iotlb = {
2166                         .iova = iova,
2167                         .perm = perm,
2168                         .type = VHOST_IOTLB_MISS,
2169                 },
2170         };
2171
2172         ret = send_vhost_message(dev->slave_req_fd, &msg);
2173         if (ret < 0) {
2174                 RTE_LOG(ERR, VHOST_CONFIG,
2175                                 "Failed to send IOTLB miss message (%d)\n",
2176                                 ret);
2177                 return ret;
2178         }
2179
2180         return 0;
2181 }
2182
2183 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
2184                                                     int index, int fd,
2185                                                     uint64_t offset,
2186                                                     uint64_t size)
2187 {
2188         int ret;
2189         struct VhostUserMsg msg = {
2190                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
2191                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
2192                 .size = sizeof(msg.payload.area),
2193                 .payload.area = {
2194                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
2195                         .size = size,
2196                         .offset = offset,
2197                 },
2198         };
2199
2200         if (fd < 0)
2201                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
2202         else {
2203                 msg.fds[0] = fd;
2204                 msg.fd_num = 1;
2205         }
2206
2207         ret = send_vhost_slave_message(dev, &msg);
2208         if (ret < 0) {
2209                 RTE_LOG(ERR, VHOST_CONFIG,
2210                         "Failed to set host notifier (%d)\n", ret);
2211                 return ret;
2212         }
2213
2214         return process_slave_message_reply(dev, &msg);
2215 }
2216
2217 int rte_vhost_host_notifier_ctrl(int vid, bool enable)
2218 {
2219         struct virtio_net *dev;
2220         struct rte_vdpa_device *vdpa_dev;
2221         int vfio_device_fd, did, ret = 0;
2222         uint64_t offset, size;
2223         unsigned int i;
2224
2225         dev = get_device(vid);
2226         if (!dev)
2227                 return -ENODEV;
2228
2229         did = dev->vdpa_dev_id;
2230         if (did < 0)
2231                 return -EINVAL;
2232
2233         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
2234             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
2235             !(dev->protocol_features &
2236                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
2237             !(dev->protocol_features &
2238                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
2239             !(dev->protocol_features &
2240                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
2241                 return -ENOTSUP;
2242
2243         vdpa_dev = rte_vdpa_get_device(did);
2244         if (!vdpa_dev)
2245                 return -ENODEV;
2246
2247         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2248         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2249
2250         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2251         if (vfio_device_fd < 0)
2252                 return -ENOTSUP;
2253
2254         if (enable) {
2255                 for (i = 0; i < dev->nr_vring; i++) {
2256                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2257                                         &size) < 0) {
2258                                 ret = -ENOTSUP;
2259                                 goto disable;
2260                         }
2261
2262                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2263                                         vfio_device_fd, offset, size) < 0) {
2264                                 ret = -EFAULT;
2265                                 goto disable;
2266                         }
2267                 }
2268         } else {
2269 disable:
2270                 for (i = 0; i < dev->nr_vring; i++) {
2271                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
2272                                         0, 0);
2273                 }
2274         }
2275
2276         return ret;
2277 }