lib/librte_vhost/vhost.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2017 Intel Corporation
   3  */
   4
   5 #include <linux/vhost.h>
   6 #include <linux/virtio_net.h>
   7 #include <stddef.h>
   8 #include <stdint.h>
   9 #include <stdlib.h>
  10 #ifdef RTE_LIBRTE_VHOST_NUMA
  11 #include <numa.h>
  12 #include <numaif.h>
  13 #endif
  14
  15 #include <rte_errno.h>
  16 #include <rte_ethdev.h>
  17 #include <rte_log.h>
  18 #include <rte_string_fns.h>
  19 #include <rte_memory.h>
  20 #include <rte_malloc.h>
  21 #include <rte_vhost.h>
  22 #include <rte_rwlock.h>
  23
  24 #include "iotlb.h"
  25 #include "vhost.h"
  26 #include "vhost_user.h"
  27
  28 struct virtio_net *vhost_devices[MAX_VHOST_DEVICE];
  29
  30 /* Called with iotlb_lock read-locked */
  31 uint64_t
  32 __vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
  33                     uint64_t iova, uint64_t *size, uint8_t perm)
  34 {
  35         uint64_t vva, tmp_size;
  36
  37         if (unlikely(!*size))
  38                 return 0;
  39
  40         tmp_size = *size;
  41
  42         vva = vhost_user_iotlb_cache_find(vq, iova, &tmp_size, perm);
  43         if (tmp_size == *size)
  44                 return vva;
  45
  46         iova += tmp_size;
  47
  48         if (!vhost_user_iotlb_pending_miss(vq, iova, perm)) {
  49                 /*
  50                  * iotlb_lock is read-locked for a full burst,
  51                  * but it only protects the iotlb cache.
  52                  * In case of IOTLB miss, we might block on the socket,
  53                  * which could cause a deadlock with QEMU if an IOTLB update
  54                  * is being handled. We can safely unlock here to avoid it.
  55                  */
  56                 vhost_user_iotlb_rd_unlock(vq);
  57
  58                 vhost_user_iotlb_pending_insert(vq, iova, perm);
  59                 if (vhost_user_iotlb_miss(dev, iova, perm)) {
  60                         RTE_LOG(ERR, VHOST_CONFIG,
  61                                 "IOTLB miss req failed for IOVA 0x%" PRIx64 "\n",
  62                                 iova);
  63                         vhost_user_iotlb_pending_remove(vq, iova, 1, perm);
  64                 }
  65
  66                 vhost_user_iotlb_rd_lock(vq);
  67         }
  68
  69         return 0;
  70 }
  71
  72 #define VHOST_LOG_PAGE  4096
  73
  74 /*
  75  * Atomically set a bit in memory.
  76  */
  77 static __rte_always_inline void
  78 vhost_set_bit(unsigned int nr, volatile uint8_t *addr)
  79 {
  80 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)
  81         /*
  82          * __sync_ built-ins are deprecated, but __atomic_ ones
  83          * are sub-optimized in older GCC versions.
  84          */
  85         __sync_fetch_and_or_1(addr, (1U << nr));
  86 #else
  87         __atomic_fetch_or(addr, (1U << nr), __ATOMIC_RELAXED);
  88 #endif
  89 }
  90
  91 static __rte_always_inline void
  92 vhost_log_page(uint8_t *log_base, uint64_t page)
  93 {
  94         vhost_set_bit(page % 8, &log_base[page / 8]);
  95 }
  96
  97 void
  98 __vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
  99 {
 100         uint64_t page;
 101
 102         if (unlikely(!dev->log_base || !len))
 103                 return;
 104
 105         if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
 106                 return;
 107
 108         /* To make sure guest memory updates are committed before logging */
 109         rte_smp_wmb();
 110
 111         page = addr / VHOST_LOG_PAGE;
 112         while (page * VHOST_LOG_PAGE < addr + len) {
 113                 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
 114                 page += 1;
 115         }
 116 }
 117
 118 void
 119 __vhost_log_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
 120                              uint64_t iova, uint64_t len)
 121 {
 122         uint64_t hva, gpa, map_len;
 123         map_len = len;
 124
 125         hva = __vhost_iova_to_vva(dev, vq, iova, &map_len, VHOST_ACCESS_RW);
 126         if (map_len != len) {
 127                 RTE_LOG(ERR, VHOST_CONFIG,
 128                         "Failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
 129                         iova);
 130                 return;
 131         }
 132
 133         gpa = hva_to_gpa(dev, hva, len);
 134         if (gpa)
 135                 __vhost_log_write(dev, gpa, len);
 136 }
 137
 138 void
 139 __vhost_log_cache_sync(struct virtio_net *dev, struct vhost_virtqueue *vq)
 140 {
 141         unsigned long *log_base;
 142         int i;
 143
 144         if (unlikely(!dev->log_base))
 145                 return;
 146
 147         rte_smp_wmb();
 148
 149         log_base = (unsigned long *)(uintptr_t)dev->log_base;
 150
 151         for (i = 0; i < vq->log_cache_nb_elem; i++) {
 152                 struct log_cache_entry *elem = vq->log_cache + i;
 153
 154 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)
 155                 /*
 156                  * '__sync' builtins are deprecated, but '__atomic' ones
 157                  * are sub-optimized in older GCC versions.
 158                  */
 159                 __sync_fetch_and_or(log_base + elem->offset, elem->val);
 160 #else
 161                 __atomic_fetch_or(log_base + elem->offset, elem->val,
 162                                 __ATOMIC_RELAXED);
 163 #endif
 164         }
 165
 166         rte_smp_wmb();
 167
 168         vq->log_cache_nb_elem = 0;
 169 }
 170
 171 static __rte_always_inline void
 172 vhost_log_cache_page(struct virtio_net *dev, struct vhost_virtqueue *vq,
 173                         uint64_t page)
 174 {
 175         uint32_t bit_nr = page % (sizeof(unsigned long) << 3);
 176         uint32_t offset = page / (sizeof(unsigned long) << 3);
 177         int i;
 178
 179         for (i = 0; i < vq->log_cache_nb_elem; i++) {
 180                 struct log_cache_entry *elem = vq->log_cache + i;
 181
 182                 if (elem->offset == offset) {
 183                         elem->val |= (1UL << bit_nr);
 184                         return;
 185                 }
 186         }
 187
 188         if (unlikely(i >= VHOST_LOG_CACHE_NR)) {
 189                 /*
 190                  * No more room for a new log cache entry,
 191                  * so write the dirty log map directly.
 192                  */
 193                 rte_smp_wmb();
 194                 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
 195
 196                 return;
 197         }
 198
 199         vq->log_cache[i].offset = offset;
 200         vq->log_cache[i].val = (1UL << bit_nr);
 201         vq->log_cache_nb_elem++;
 202 }
 203
 204 void
 205 __vhost_log_cache_write(struct virtio_net *dev, struct vhost_virtqueue *vq,
 206                         uint64_t addr, uint64_t len)
 207 {
 208         uint64_t page;
 209
 210         if (unlikely(!dev->log_base || !len))
 211                 return;
 212
 213         if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
 214                 return;
 215
 216         page = addr / VHOST_LOG_PAGE;
 217         while (page * VHOST_LOG_PAGE < addr + len) {
 218                 vhost_log_cache_page(dev, vq, page);
 219                 page += 1;
 220         }
 221 }
 222
 223 void
 224 __vhost_log_cache_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
 225                              uint64_t iova, uint64_t len)
 226 {
 227         uint64_t hva, gpa, map_len;
 228         map_len = len;
 229
 230         hva = __vhost_iova_to_vva(dev, vq, iova, &map_len, VHOST_ACCESS_RW);
 231         if (map_len != len) {
 232                 RTE_LOG(ERR, VHOST_CONFIG,
 233                         "Failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
 234                         iova);
 235                 return;
 236         }
 237
 238         gpa = hva_to_gpa(dev, hva, len);
 239         if (gpa)
 240                 __vhost_log_cache_write(dev, vq, gpa, len);
 241 }
 242
 243 void *
 244 vhost_alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
 245                 uint64_t desc_addr, uint64_t desc_len)
 246 {
 247         void *idesc;
 248         uint64_t src, dst;
 249         uint64_t len, remain = desc_len;
 250
 251         idesc = rte_malloc(__func__, desc_len, 0);
 252         if (unlikely(!idesc))
 253                 return NULL;
 254
 255         dst = (uint64_t)(uintptr_t)idesc;
 256
 257         while (remain) {
 258                 len = remain;
 259                 src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
 260                                 VHOST_ACCESS_RO);
 261                 if (unlikely(!src || !len)) {
 262                         rte_free(idesc);
 263                         return NULL;
 264                 }
 265
 266                 rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
 267
 268                 remain -= len;
 269                 dst += len;
 270                 desc_addr += len;
 271         }
 272
 273         return idesc;
 274 }
 275
 276 void
 277 cleanup_vq(struct vhost_virtqueue *vq, int destroy)
 278 {
 279         if ((vq->callfd >= 0) && (destroy != 0))
 280                 close(vq->callfd);
 281         if (vq->kickfd >= 0)
 282                 close(vq->kickfd);
 283 }
 284
 285 void
 286 cleanup_vq_inflight(struct virtio_net *dev, struct vhost_virtqueue *vq)
 287 {
 288         if (!(dev->protocol_features &
 289             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
 290                 return;
 291
 292         if (vq_is_packed(dev)) {
 293                 if (vq->inflight_packed)
 294                         vq->inflight_packed = NULL;
 295         } else {
 296                 if (vq->inflight_split)
 297                         vq->inflight_split = NULL;
 298         }
 299
 300         if (vq->resubmit_inflight) {
 301                 if (vq->resubmit_inflight->resubmit_list) {
 302                         free(vq->resubmit_inflight->resubmit_list);
 303                         vq->resubmit_inflight->resubmit_list = NULL;
 304                 }
 305                 free(vq->resubmit_inflight);
 306                 vq->resubmit_inflight = NULL;
 307         }
 308 }
 309
 310 /*
 311  * Unmap any memory, close any file descriptors and
 312  * free any memory owned by a device.
 313  */
 314 void
 315 cleanup_device(struct virtio_net *dev, int destroy)
 316 {
 317         uint32_t i;
 318
 319         vhost_backend_cleanup(dev);
 320
 321         for (i = 0; i < dev->nr_vring; i++) {
 322                 cleanup_vq(dev->virtqueue[i], destroy);
 323                 cleanup_vq_inflight(dev, dev->virtqueue[i]);
 324         }
 325 }
 326
 327 void
 328 free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq)
 329 {
 330         if (vq_is_packed(dev))
 331                 rte_free(vq->shadow_used_packed);
 332         else
 333                 rte_free(vq->shadow_used_split);
 334         rte_free(vq->batch_copy_elems);
 335         rte_mempool_free(vq->iotlb_pool);
 336         rte_free(vq);
 337 }
 338
 339 /*
 340  * Release virtqueues and device memory.
 341  */
 342 static void
 343 free_device(struct virtio_net *dev)
 344 {
 345         uint32_t i;
 346
 347         for (i = 0; i < dev->nr_vring; i++)
 348                 free_vq(dev, dev->virtqueue[i]);
 349
 350         rte_free(dev);
 351 }
 352
 353 static int
 354 vring_translate_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
 355 {
 356         uint64_t req_size, size;
 357
 358         req_size = sizeof(struct vring_desc) * vq->size;
 359         size = req_size;
 360         vq->desc = (struct vring_desc *)(uintptr_t)vhost_iova_to_vva(dev, vq,
 361                                                 vq->ring_addrs.desc_user_addr,
 362                                                 &size, VHOST_ACCESS_RW);
 363         if (!vq->desc || size != req_size)
 364                 return -1;
 365
 366         req_size = sizeof(struct vring_avail);
 367         req_size += sizeof(uint16_t) * vq->size;
 368         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 369                 req_size += sizeof(uint16_t);
 370         size = req_size;
 371         vq->avail = (struct vring_avail *)(uintptr_t)vhost_iova_to_vva(dev, vq,
 372                                                 vq->ring_addrs.avail_user_addr,
 373                                                 &size, VHOST_ACCESS_RW);
 374         if (!vq->avail || size != req_size)
 375                 return -1;
 376
 377         req_size = sizeof(struct vring_used);
 378         req_size += sizeof(struct vring_used_elem) * vq->size;
 379         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 380                 req_size += sizeof(uint16_t);
 381         size = req_size;
 382         vq->used = (struct vring_used *)(uintptr_t)vhost_iova_to_vva(dev, vq,
 383                                                 vq->ring_addrs.used_user_addr,
 384                                                 &size, VHOST_ACCESS_RW);
 385         if (!vq->used || size != req_size)
 386                 return -1;
 387
 388         return 0;
 389 }
 390
 391 static int
 392 vring_translate_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
 393 {
 394         uint64_t req_size, size;
 395
 396         req_size = sizeof(struct vring_packed_desc) * vq->size;
 397         size = req_size;
 398         vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 399                 vhost_iova_to_vva(dev, vq, vq->ring_addrs.desc_user_addr,
 400                                 &size, VHOST_ACCESS_RW);
 401         if (!vq->desc_packed || size != req_size)
 402                 return -1;
 403
 404         req_size = sizeof(struct vring_packed_desc_event);
 405         size = req_size;
 406         vq->driver_event = (struct vring_packed_desc_event *)(uintptr_t)
 407                 vhost_iova_to_vva(dev, vq, vq->ring_addrs.avail_user_addr,
 408                                 &size, VHOST_ACCESS_RW);
 409         if (!vq->driver_event || size != req_size)
 410                 return -1;
 411
 412         req_size = sizeof(struct vring_packed_desc_event);
 413         size = req_size;
 414         vq->device_event = (struct vring_packed_desc_event *)(uintptr_t)
 415                 vhost_iova_to_vva(dev, vq, vq->ring_addrs.used_user_addr,
 416                                 &size, VHOST_ACCESS_RW);
 417         if (!vq->device_event || size != req_size)
 418                 return -1;
 419
 420         return 0;
 421 }
 422
 423 int
 424 vring_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
 425 {
 426
 427         if (!(dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
 428                 return -1;
 429
 430         if (vq_is_packed(dev)) {
 431                 if (vring_translate_packed(dev, vq) < 0)
 432                         return -1;
 433         } else {
 434                 if (vring_translate_split(dev, vq) < 0)
 435                         return -1;
 436         }
 437         vq->access_ok = 1;
 438
 439         return 0;
 440 }
 441
 442 void
 443 vring_invalidate(struct virtio_net *dev, struct vhost_virtqueue *vq)
 444 {
 445         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 446                 vhost_user_iotlb_wr_lock(vq);
 447
 448         vq->access_ok = 0;
 449         vq->desc = NULL;
 450         vq->avail = NULL;
 451         vq->used = NULL;
 452         vq->log_guest_addr = 0;
 453
 454         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 455                 vhost_user_iotlb_wr_unlock(vq);
 456 }
 457
 458 static void
 459 init_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
 460 {
 461         struct vhost_virtqueue *vq;
 462
 463         if (vring_idx >= VHOST_MAX_VRING) {
 464                 RTE_LOG(ERR, VHOST_CONFIG,
 465                                 "Failed not init vring, out of bound (%d)\n",
 466                                 vring_idx);
 467                 return;
 468         }
 469
 470         vq = dev->virtqueue[vring_idx];
 471
 472         memset(vq, 0, sizeof(struct vhost_virtqueue));
 473
 474         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
 475         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
 476
 477         vhost_user_iotlb_init(dev, vring_idx);
 478         /* Backends are set to -1 indicating an inactive device. */
 479         vq->backend = -1;
 480
 481         TAILQ_INIT(&vq->zmbuf_list);
 482 }
 483
 484 static void
 485 reset_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
 486 {
 487         struct vhost_virtqueue *vq;
 488         int callfd;
 489
 490         if (vring_idx >= VHOST_MAX_VRING) {
 491                 RTE_LOG(ERR, VHOST_CONFIG,
 492                                 "Failed not init vring, out of bound (%d)\n",
 493                                 vring_idx);
 494                 return;
 495         }
 496
 497         vq = dev->virtqueue[vring_idx];
 498         callfd = vq->callfd;
 499         init_vring_queue(dev, vring_idx);
 500         vq->callfd = callfd;
 501 }
 502
 503 int
 504 alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
 505 {
 506         struct vhost_virtqueue *vq;
 507
 508         vq = rte_malloc(NULL, sizeof(struct vhost_virtqueue), 0);
 509         if (vq == NULL) {
 510                 RTE_LOG(ERR, VHOST_CONFIG,
 511                         "Failed to allocate memory for vring:%u.\n", vring_idx);
 512                 return -1;
 513         }
 514
 515         dev->virtqueue[vring_idx] = vq;
 516         init_vring_queue(dev, vring_idx);
 517         rte_spinlock_init(&vq->access_lock);
 518         vq->avail_wrap_counter = 1;
 519         vq->used_wrap_counter = 1;
 520         vq->signalled_used_valid = false;
 521
 522         dev->nr_vring += 1;
 523
 524         return 0;
 525 }
 526
 527 /*
 528  * Reset some variables in device structure, while keeping few
 529  * others untouched, such as vid, ifname, nr_vring: they
 530  * should be same unless the device is removed.
 531  */
 532 void
 533 reset_device(struct virtio_net *dev)
 534 {
 535         uint32_t i;
 536
 537         dev->features = 0;
 538         dev->protocol_features = 0;
 539         dev->flags &= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
 540
 541         for (i = 0; i < dev->nr_vring; i++)
 542                 reset_vring_queue(dev, i);
 543 }
 544
 545 /*
 546  * Invoked when there is a new vhost-user connection established (when
 547  * there is a new virtio device being attached).
 548  */
 549 int
 550 vhost_new_device(void)
 551 {
 552         struct virtio_net *dev;
 553         int i;
 554
 555         for (i = 0; i < MAX_VHOST_DEVICE; i++) {
 556                 if (vhost_devices[i] == NULL)
 557                         break;
 558         }
 559
 560         if (i == MAX_VHOST_DEVICE) {
 561                 RTE_LOG(ERR, VHOST_CONFIG,
 562                         "Failed to find a free slot for new device.\n");
 563                 return -1;
 564         }
 565
 566         dev = rte_zmalloc(NULL, sizeof(struct virtio_net), 0);
 567         if (dev == NULL) {
 568                 RTE_LOG(ERR, VHOST_CONFIG,
 569                         "Failed to allocate memory for new dev.\n");
 570                 return -1;
 571         }
 572
 573         vhost_devices[i] = dev;
 574         dev->vid = i;
 575         dev->flags = VIRTIO_DEV_BUILTIN_VIRTIO_NET;
 576         dev->slave_req_fd = -1;
 577         dev->vdpa_dev_id = -1;
 578         dev->postcopy_ufd = -1;
 579         rte_spinlock_init(&dev->slave_req_lock);
 580
 581         return i;
 582 }
 583
 584 void
 585 vhost_destroy_device_notify(struct virtio_net *dev)
 586 {
 587         struct rte_vdpa_device *vdpa_dev;
 588         int did;
 589
 590         if (dev->flags & VIRTIO_DEV_RUNNING) {
 591                 did = dev->vdpa_dev_id;
 592                 vdpa_dev = rte_vdpa_get_device(did);
 593                 if (vdpa_dev && vdpa_dev->ops->dev_close)
 594                         vdpa_dev->ops->dev_close(dev->vid);
 595                 dev->flags &= ~VIRTIO_DEV_RUNNING;
 596                 dev->notify_ops->destroy_device(dev->vid);
 597         }
 598 }
 599
 600 /*
 601  * Invoked when there is the vhost-user connection is broken (when
 602  * the virtio device is being detached).
 603  */
 604 void
 605 vhost_destroy_device(int vid)
 606 {
 607         struct virtio_net *dev = get_device(vid);
 608
 609         if (dev == NULL)
 610                 return;
 611
 612         vhost_destroy_device_notify(dev);
 613
 614         cleanup_device(dev, 1);
 615         free_device(dev);
 616
 617         vhost_devices[vid] = NULL;
 618 }
 619
 620 void
 621 vhost_attach_vdpa_device(int vid, int did)
 622 {
 623         struct virtio_net *dev = get_device(vid);
 624
 625         if (dev == NULL)
 626                 return;
 627
 628         if (rte_vdpa_get_device(did) == NULL)
 629                 return;
 630
 631         dev->vdpa_dev_id = did;
 632 }
 633
 634 void
 635 vhost_set_ifname(int vid, const char *if_name, unsigned int if_len)
 636 {
 637         struct virtio_net *dev;
 638         unsigned int len;
 639
 640         dev = get_device(vid);
 641         if (dev == NULL)
 642                 return;
 643
 644         len = if_len > sizeof(dev->ifname) ?
 645                 sizeof(dev->ifname) : if_len;
 646
 647         strncpy(dev->ifname, if_name, len);
 648         dev->ifname[sizeof(dev->ifname) - 1] = '\0';
 649 }
 650
 651 void
 652 vhost_enable_dequeue_zero_copy(int vid)
 653 {
 654         struct virtio_net *dev = get_device(vid);
 655
 656         if (dev == NULL)
 657                 return;
 658
 659         dev->dequeue_zero_copy = 1;
 660 }
 661
 662 void
 663 vhost_set_builtin_virtio_net(int vid, bool enable)
 664 {
 665         struct virtio_net *dev = get_device(vid);
 666
 667         if (dev == NULL)
 668                 return;
 669
 670         if (enable)
 671                 dev->flags |= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
 672         else
 673                 dev->flags &= ~VIRTIO_DEV_BUILTIN_VIRTIO_NET;
 674 }
 675
 676 void
 677 vhost_enable_extbuf(int vid)
 678 {
 679         struct virtio_net *dev = get_device(vid);
 680
 681         if (dev == NULL)
 682                 return;
 683
 684         dev->extbuf = 1;
 685 }
 686
 687 void
 688 vhost_enable_linearbuf(int vid)
 689 {
 690         struct virtio_net *dev = get_device(vid);
 691
 692         if (dev == NULL)
 693                 return;
 694
 695         dev->linearbuf = 1;
 696 }
 697
 698 int
 699 rte_vhost_get_mtu(int vid, uint16_t *mtu)
 700 {
 701         struct virtio_net *dev = get_device(vid);
 702
 703         if (dev == NULL || mtu == NULL)
 704                 return -ENODEV;
 705
 706         if (!(dev->flags & VIRTIO_DEV_READY))
 707                 return -EAGAIN;
 708
 709         if (!(dev->features & (1ULL << VIRTIO_NET_F_MTU)))
 710                 return -ENOTSUP;
 711
 712         *mtu = dev->mtu;
 713
 714         return 0;
 715 }
 716
 717 int
 718 rte_vhost_get_numa_node(int vid)
 719 {
 720 #ifdef RTE_LIBRTE_VHOST_NUMA
 721         struct virtio_net *dev = get_device(vid);
 722         int numa_node;
 723         int ret;
 724
 725         if (dev == NULL || numa_available() != 0)
 726                 return -1;
 727
 728         ret = get_mempolicy(&numa_node, NULL, 0, dev,
 729                             MPOL_F_NODE | MPOL_F_ADDR);
 730         if (ret < 0) {
 731                 RTE_LOG(ERR, VHOST_CONFIG,
 732                         "(%d) failed to query numa node: %s\n",
 733                         vid, rte_strerror(errno));
 734                 return -1;
 735         }
 736
 737         return numa_node;
 738 #else
 739         RTE_SET_USED(vid);
 740         return -1;
 741 #endif
 742 }
 743
 744 uint32_t
 745 rte_vhost_get_queue_num(int vid)
 746 {
 747         struct virtio_net *dev = get_device(vid);
 748
 749         if (dev == NULL)
 750                 return 0;
 751
 752         return dev->nr_vring / 2;
 753 }
 754
 755 uint16_t
 756 rte_vhost_get_vring_num(int vid)
 757 {
 758         struct virtio_net *dev = get_device(vid);
 759
 760         if (dev == NULL)
 761                 return 0;
 762
 763         return dev->nr_vring;
 764 }
 765
 766 int
 767 rte_vhost_get_ifname(int vid, char *buf, size_t len)
 768 {
 769         struct virtio_net *dev = get_device(vid);
 770
 771         if (dev == NULL || buf == NULL)
 772                 return -1;
 773
 774         len = RTE_MIN(len, sizeof(dev->ifname));
 775
 776         strncpy(buf, dev->ifname, len);
 777         buf[len - 1] = '\0';
 778
 779         return 0;
 780 }
 781
 782 int
 783 rte_vhost_get_negotiated_features(int vid, uint64_t *features)
 784 {
 785         struct virtio_net *dev;
 786
 787         dev = get_device(vid);
 788         if (dev == NULL || features == NULL)
 789                 return -1;
 790
 791         *features = dev->features;
 792         return 0;
 793 }
 794
 795 int
 796 rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
 797 {
 798         struct virtio_net *dev;
 799         struct rte_vhost_memory *m;
 800         size_t size;
 801
 802         dev = get_device(vid);
 803         if (dev == NULL || mem == NULL)
 804                 return -1;
 805
 806         size = dev->mem->nregions * sizeof(struct rte_vhost_mem_region);
 807         m = malloc(sizeof(struct rte_vhost_memory) + size);
 808         if (!m)
 809                 return -1;
 810
 811         m->nregions = dev->mem->nregions;
 812         memcpy(m->regions, dev->mem->regions, size);
 813         *mem = m;
 814
 815         return 0;
 816 }
 817
 818 int
 819 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
 820                           struct rte_vhost_vring *vring)
 821 {
 822         struct virtio_net *dev;
 823         struct vhost_virtqueue *vq;
 824
 825         dev = get_device(vid);
 826         if (dev == NULL || vring == NULL)
 827                 return -1;
 828
 829         if (vring_idx >= VHOST_MAX_VRING)
 830                 return -1;
 831
 832         vq = dev->virtqueue[vring_idx];
 833         if (!vq)
 834                 return -1;
 835
 836         if (vq_is_packed(dev)) {
 837                 vring->desc_packed = vq->desc_packed;
 838                 vring->driver_event = vq->driver_event;
 839                 vring->device_event = vq->device_event;
 840         } else {
 841                 vring->desc = vq->desc;
 842                 vring->avail = vq->avail;
 843                 vring->used = vq->used;
 844         }
 845         vring->log_guest_addr  = vq->log_guest_addr;
 846
 847         vring->callfd  = vq->callfd;
 848         vring->kickfd  = vq->kickfd;
 849         vring->size    = vq->size;
 850
 851         return 0;
 852 }
 853
 854 int
 855 rte_vhost_get_vhost_ring_inflight(int vid, uint16_t vring_idx,
 856                                   struct rte_vhost_ring_inflight *vring)
 857 {
 858         struct virtio_net *dev;
 859         struct vhost_virtqueue *vq;
 860
 861         dev = get_device(vid);
 862         if (unlikely(!dev))
 863                 return -1;
 864
 865         if (vring_idx >= VHOST_MAX_VRING)
 866                 return -1;
 867
 868         vq = dev->virtqueue[vring_idx];
 869         if (unlikely(!vq))
 870                 return -1;
 871
 872         if (vq_is_packed(dev)) {
 873                 if (unlikely(!vq->inflight_packed))
 874                         return -1;
 875
 876                 vring->inflight_packed = vq->inflight_packed;
 877         } else {
 878                 if (unlikely(!vq->inflight_split))
 879                         return -1;
 880
 881                 vring->inflight_split = vq->inflight_split;
 882         }
 883
 884         vring->resubmit_inflight = vq->resubmit_inflight;
 885
 886         return 0;
 887 }
 888
 889 int
 890 rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
 891                                   uint16_t idx)
 892 {
 893         struct vhost_virtqueue *vq;
 894         struct virtio_net *dev;
 895
 896         dev = get_device(vid);
 897         if (unlikely(!dev))
 898                 return -1;
 899
 900         if (unlikely(!(dev->protocol_features &
 901             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
 902                 return 0;
 903
 904         if (unlikely(vq_is_packed(dev)))
 905                 return -1;
 906
 907         if (unlikely(vring_idx >= VHOST_MAX_VRING))
 908                 return -1;
 909
 910         vq = dev->virtqueue[vring_idx];
 911         if (unlikely(!vq))
 912                 return -1;
 913
 914         if (unlikely(!vq->inflight_split))
 915                 return -1;
 916
 917         if (unlikely(idx >= vq->size))
 918                 return -1;
 919
 920         vq->inflight_split->desc[idx].counter = vq->global_counter++;
 921         vq->inflight_split->desc[idx].inflight = 1;
 922         return 0;
 923 }
 924
 925 int
 926 rte_vhost_set_inflight_desc_packed(int vid, uint16_t vring_idx,
 927                                    uint16_t head, uint16_t last,
 928                                    uint16_t *inflight_entry)
 929 {
 930         struct rte_vhost_inflight_info_packed *inflight_info;
 931         struct virtio_net *dev;
 932         struct vhost_virtqueue *vq;
 933         struct vring_packed_desc *desc;
 934         uint16_t old_free_head, free_head;
 935
 936         dev = get_device(vid);
 937         if (unlikely(!dev))
 938                 return -1;
 939
 940         if (unlikely(!(dev->protocol_features &
 941             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
 942                 return 0;
 943
 944         if (unlikely(!vq_is_packed(dev)))
 945                 return -1;
 946
 947         if (unlikely(vring_idx >= VHOST_MAX_VRING))
 948                 return -1;
 949
 950         vq = dev->virtqueue[vring_idx];
 951         if (unlikely(!vq))
 952                 return -1;
 953
 954         inflight_info = vq->inflight_packed;
 955         if (unlikely(!inflight_info))
 956                 return -1;
 957
 958         if (unlikely(head >= vq->size))
 959                 return -1;
 960
 961         desc = vq->desc_packed;
 962         old_free_head = inflight_info->old_free_head;
 963         if (unlikely(old_free_head >= vq->size))
 964                 return -1;
 965
 966         free_head = old_free_head;
 967
 968         /* init header descriptor */
 969         inflight_info->desc[old_free_head].num = 0;
 970         inflight_info->desc[old_free_head].counter = vq->global_counter++;
 971         inflight_info->desc[old_free_head].inflight = 1;
 972
 973         /* save desc entry in flight entry */
 974         while (head != ((last + 1) % vq->size)) {
 975                 inflight_info->desc[old_free_head].num++;
 976                 inflight_info->desc[free_head].addr = desc[head].addr;
 977                 inflight_info->desc[free_head].len = desc[head].len;
 978                 inflight_info->desc[free_head].flags = desc[head].flags;
 979                 inflight_info->desc[free_head].id = desc[head].id;
 980
 981                 inflight_info->desc[old_free_head].last = free_head;
 982                 free_head = inflight_info->desc[free_head].next;
 983                 inflight_info->free_head = free_head;
 984                 head = (head + 1) % vq->size;
 985         }
 986
 987         inflight_info->old_free_head = free_head;
 988         *inflight_entry = old_free_head;
 989
 990         return 0;
 991 }
 992
 993 int
 994 rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
 995                                   uint16_t last_used_idx, uint16_t idx)
 996 {
 997         struct virtio_net *dev;
 998         struct vhost_virtqueue *vq;
 999
1000         dev = get_device(vid);
1001         if (unlikely(!dev))
1002                 return -1;
1003
1004         if (unlikely(!(dev->protocol_features &
1005             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1006                 return 0;
1007
1008         if (unlikely(vq_is_packed(dev)))
1009                 return -1;
1010
1011         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1012                 return -1;
1013
1014         vq = dev->virtqueue[vring_idx];
1015         if (unlikely(!vq))
1016                 return -1;
1017
1018         if (unlikely(!vq->inflight_split))
1019                 return -1;
1020
1021         if (unlikely(idx >= vq->size))
1022                 return -1;
1023
1024         rte_smp_mb();
1025
1026         vq->inflight_split->desc[idx].inflight = 0;
1027
1028         rte_smp_mb();
1029
1030         vq->inflight_split->used_idx = last_used_idx;
1031         return 0;
1032 }
1033
1034 int
1035 rte_vhost_clr_inflight_desc_packed(int vid, uint16_t vring_idx,
1036                                    uint16_t head)
1037 {
1038         struct rte_vhost_inflight_info_packed *inflight_info;
1039         struct virtio_net *dev;
1040         struct vhost_virtqueue *vq;
1041
1042         dev = get_device(vid);
1043         if (unlikely(!dev))
1044                 return -1;
1045
1046         if (unlikely(!(dev->protocol_features &
1047             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1048                 return 0;
1049
1050         if (unlikely(!vq_is_packed(dev)))
1051                 return -1;
1052
1053         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1054                 return -1;
1055
1056         vq = dev->virtqueue[vring_idx];
1057         if (unlikely(!vq))
1058                 return -1;
1059
1060         inflight_info = vq->inflight_packed;
1061         if (unlikely(!inflight_info))
1062                 return -1;
1063
1064         if (unlikely(head >= vq->size))
1065                 return -1;
1066
1067         rte_smp_mb();
1068
1069         inflight_info->desc[head].inflight = 0;
1070
1071         rte_smp_mb();
1072
1073         inflight_info->old_free_head = inflight_info->free_head;
1074         inflight_info->old_used_idx = inflight_info->used_idx;
1075         inflight_info->old_used_wrap_counter = inflight_info->used_wrap_counter;
1076
1077         return 0;
1078 }
1079
1080 int
1081 rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
1082                                      uint16_t idx)
1083 {
1084         struct virtio_net *dev;
1085         struct vhost_virtqueue *vq;
1086
1087         dev = get_device(vid);
1088         if (unlikely(!dev))
1089                 return -1;
1090
1091         if (unlikely(!(dev->protocol_features &
1092             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1093                 return 0;
1094
1095         if (unlikely(vq_is_packed(dev)))
1096                 return -1;
1097
1098         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1099                 return -1;
1100
1101         vq = dev->virtqueue[vring_idx];
1102         if (unlikely(!vq))
1103                 return -1;
1104
1105         if (unlikely(!vq->inflight_split))
1106                 return -1;
1107
1108         vq->inflight_split->last_inflight_io = idx;
1109         return 0;
1110 }
1111
1112 int
1113 rte_vhost_set_last_inflight_io_packed(int vid, uint16_t vring_idx,
1114                                       uint16_t head)
1115 {
1116         struct rte_vhost_inflight_info_packed *inflight_info;
1117         struct virtio_net *dev;
1118         struct vhost_virtqueue *vq;
1119         uint16_t last;
1120
1121         dev = get_device(vid);
1122         if (unlikely(!dev))
1123                 return -1;
1124
1125         if (unlikely(!(dev->protocol_features &
1126             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1127                 return 0;
1128
1129         if (unlikely(!vq_is_packed(dev)))
1130                 return -1;
1131
1132         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1133                 return -1;
1134
1135         vq = dev->virtqueue[vring_idx];
1136         if (unlikely(!vq))
1137                 return -1;
1138
1139         inflight_info = vq->inflight_packed;
1140         if (unlikely(!inflight_info))
1141                 return -1;
1142
1143         if (unlikely(head >= vq->size))
1144                 return -1;
1145
1146         last = inflight_info->desc[head].last;
1147         if (unlikely(last >= vq->size))
1148                 return -1;
1149
1150         inflight_info->desc[last].next = inflight_info->free_head;
1151         inflight_info->free_head = head;
1152         inflight_info->used_idx += inflight_info->desc[head].num;
1153         if (inflight_info->used_idx >= inflight_info->desc_num) {
1154                 inflight_info->used_idx -= inflight_info->desc_num;
1155                 inflight_info->used_wrap_counter =
1156                         !inflight_info->used_wrap_counter;
1157         }
1158
1159         return 0;
1160 }
1161
1162 int
1163 rte_vhost_vring_call(int vid, uint16_t vring_idx)
1164 {
1165         struct virtio_net *dev;
1166         struct vhost_virtqueue *vq;
1167
1168         dev = get_device(vid);
1169         if (!dev)
1170                 return -1;
1171
1172         if (vring_idx >= VHOST_MAX_VRING)
1173                 return -1;
1174
1175         vq = dev->virtqueue[vring_idx];
1176         if (!vq)
1177                 return -1;
1178
1179         if (vq_is_packed(dev))
1180                 vhost_vring_call_packed(dev, vq);
1181         else
1182                 vhost_vring_call_split(dev, vq);
1183
1184         return 0;
1185 }
1186
1187 uint16_t
1188 rte_vhost_avail_entries(int vid, uint16_t queue_id)
1189 {
1190         struct virtio_net *dev;
1191         struct vhost_virtqueue *vq;
1192         uint16_t ret = 0;
1193
1194         dev = get_device(vid);
1195         if (!dev)
1196                 return 0;
1197
1198         vq = dev->virtqueue[queue_id];
1199
1200         rte_spinlock_lock(&vq->access_lock);
1201
1202         if (unlikely(!vq->enabled || vq->avail == NULL))
1203                 goto out;
1204
1205         ret = *(volatile uint16_t *)&vq->avail->idx - vq->last_used_idx;
1206
1207 out:
1208         rte_spinlock_unlock(&vq->access_lock);
1209         return ret;
1210 }
1211
1212 static inline int
1213 vhost_enable_notify_split(struct virtio_net *dev,
1214                 struct vhost_virtqueue *vq, int enable)
1215 {
1216         if (vq->used == NULL)
1217                 return -1;
1218
1219         if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
1220                 if (enable)
1221                         vq->used->flags &= ~VRING_USED_F_NO_NOTIFY;
1222                 else
1223                         vq->used->flags |= VRING_USED_F_NO_NOTIFY;
1224         } else {
1225                 if (enable)
1226                         vhost_avail_event(vq) = vq->last_avail_idx;
1227         }
1228         return 0;
1229 }
1230
1231 static inline int
1232 vhost_enable_notify_packed(struct virtio_net *dev,
1233                 struct vhost_virtqueue *vq, int enable)
1234 {
1235         uint16_t flags;
1236
1237         if (vq->device_event == NULL)
1238                 return -1;
1239
1240         if (!enable) {
1241                 vq->device_event->flags = VRING_EVENT_F_DISABLE;
1242                 return 0;
1243         }
1244
1245         flags = VRING_EVENT_F_ENABLE;
1246         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
1247                 flags = VRING_EVENT_F_DESC;
1248                 vq->device_event->off_wrap = vq->last_avail_idx |
1249                         vq->avail_wrap_counter << 15;
1250         }
1251
1252         rte_smp_wmb();
1253
1254         vq->device_event->flags = flags;
1255         return 0;
1256 }
1257
1258 int
1259 rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
1260 {
1261         struct virtio_net *dev = get_device(vid);
1262         struct vhost_virtqueue *vq;
1263         int ret;
1264
1265         if (!dev)
1266                 return -1;
1267
1268         vq = dev->virtqueue[queue_id];
1269
1270         rte_spinlock_lock(&vq->access_lock);
1271
1272         if (vq_is_packed(dev))
1273                 ret = vhost_enable_notify_packed(dev, vq, enable);
1274         else
1275                 ret = vhost_enable_notify_split(dev, vq, enable);
1276
1277         rte_spinlock_unlock(&vq->access_lock);
1278
1279         return ret;
1280 }
1281
1282 void
1283 rte_vhost_log_write(int vid, uint64_t addr, uint64_t len)
1284 {
1285         struct virtio_net *dev = get_device(vid);
1286
1287         if (dev == NULL)
1288                 return;
1289
1290         vhost_log_write(dev, addr, len);
1291 }
1292
1293 void
1294 rte_vhost_log_used_vring(int vid, uint16_t vring_idx,
1295                          uint64_t offset, uint64_t len)
1296 {
1297         struct virtio_net *dev;
1298         struct vhost_virtqueue *vq;
1299
1300         dev = get_device(vid);
1301         if (dev == NULL)
1302                 return;
1303
1304         if (vring_idx >= VHOST_MAX_VRING)
1305                 return;
1306         vq = dev->virtqueue[vring_idx];
1307         if (!vq)
1308                 return;
1309
1310         vhost_log_used_vring(dev, vq, offset, len);
1311 }
1312
1313 uint32_t
1314 rte_vhost_rx_queue_count(int vid, uint16_t qid)
1315 {
1316         struct virtio_net *dev;
1317         struct vhost_virtqueue *vq;
1318         uint32_t ret = 0;
1319
1320         dev = get_device(vid);
1321         if (dev == NULL)
1322                 return 0;
1323
1324         if (unlikely(qid >= dev->nr_vring || (qid & 1) == 0)) {
1325                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1326                         dev->vid, __func__, qid);
1327                 return 0;
1328         }
1329
1330         vq = dev->virtqueue[qid];
1331         if (vq == NULL)
1332                 return 0;
1333
1334         rte_spinlock_lock(&vq->access_lock);
1335
1336         if (unlikely(vq->enabled == 0 || vq->avail == NULL))
1337                 goto out;
1338
1339         ret = *((volatile uint16_t *)&vq->avail->idx) - vq->last_avail_idx;
1340
1341 out:
1342         rte_spinlock_unlock(&vq->access_lock);
1343         return ret;
1344 }
1345
1346 int rte_vhost_get_vdpa_device_id(int vid)
1347 {
1348         struct virtio_net *dev = get_device(vid);
1349
1350         if (dev == NULL)
1351                 return -1;
1352
1353         return dev->vdpa_dev_id;
1354 }
1355
1356 int rte_vhost_get_log_base(int vid, uint64_t *log_base,
1357                 uint64_t *log_size)
1358 {
1359         struct virtio_net *dev = get_device(vid);
1360
1361         if (dev == NULL || log_base == NULL || log_size == NULL)
1362                 return -1;
1363
1364         *log_base = dev->log_base;
1365         *log_size = dev->log_size;
1366
1367         return 0;
1368 }
1369
1370 int rte_vhost_get_vring_base(int vid, uint16_t queue_id,
1371                 uint16_t *last_avail_idx, uint16_t *last_used_idx)
1372 {
1373         struct vhost_virtqueue *vq;
1374         struct virtio_net *dev = get_device(vid);
1375
1376         if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1377                 return -1;
1378
1379         vq = dev->virtqueue[queue_id];
1380         if (!vq)
1381                 return -1;
1382
1383         if (vq_is_packed(dev)) {
1384                 *last_avail_idx = (vq->avail_wrap_counter << 15) |
1385                                   vq->last_avail_idx;
1386                 *last_used_idx = (vq->used_wrap_counter << 15) |
1387                                  vq->last_used_idx;
1388         } else {
1389                 *last_avail_idx = vq->last_avail_idx;
1390                 *last_used_idx = vq->last_used_idx;
1391         }
1392
1393         return 0;
1394 }
1395
1396 int rte_vhost_set_vring_base(int vid, uint16_t queue_id,
1397                 uint16_t last_avail_idx, uint16_t last_used_idx)
1398 {
1399         struct vhost_virtqueue *vq;
1400         struct virtio_net *dev = get_device(vid);
1401
1402         if (!dev)
1403                 return -1;
1404
1405         vq = dev->virtqueue[queue_id];
1406         if (!vq)
1407                 return -1;
1408
1409         if (vq_is_packed(dev)) {
1410                 vq->last_avail_idx = last_avail_idx & 0x7fff;
1411                 vq->avail_wrap_counter = !!(last_avail_idx & (1 << 15));
1412                 vq->last_used_idx = last_used_idx & 0x7fff;
1413                 vq->used_wrap_counter = !!(last_used_idx & (1 << 15));
1414         } else {
1415                 vq->last_avail_idx = last_avail_idx;
1416                 vq->last_used_idx = last_used_idx;
1417         }
1418
1419         return 0;
1420 }
1421
1422 int
1423 rte_vhost_get_vring_base_from_inflight(int vid,
1424                                        uint16_t queue_id,
1425                                        uint16_t *last_avail_idx,
1426                                        uint16_t *last_used_idx)
1427 {
1428         struct rte_vhost_inflight_info_packed *inflight_info;
1429         struct virtio_net *dev = get_device(vid);
1430
1431         if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1432                 return -1;
1433
1434         if (!vq_is_packed(dev))
1435                 return -1;
1436
1437         inflight_info = dev->virtqueue[queue_id]->inflight_packed;
1438         if (!inflight_info)
1439                 return -1;
1440
1441         *last_avail_idx = (inflight_info->old_used_wrap_counter << 15) |
1442                           inflight_info->old_used_idx;
1443         *last_used_idx = *last_avail_idx;
1444
1445         return 0;
1446 }
1447
1448 int rte_vhost_extern_callback_register(int vid,
1449                 struct rte_vhost_user_extern_ops const * const ops, void *ctx)
1450 {
1451         struct virtio_net *dev = get_device(vid);
1452
1453         if (dev == NULL || ops == NULL)
1454                 return -1;
1455
1456         dev->extern_ops = *ops;
1457         dev->extern_data = ctx;
1458         return 0;
1459 }