lib/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 <stdint.h>
   8 #include <stdlib.h>
   9 #ifdef RTE_LIBRTE_VHOST_NUMA
  10 #include <numa.h>
  11 #include <numaif.h>
  12 #endif
  13
  14 #include <rte_errno.h>
  15 #include <rte_log.h>
  16 #include <rte_memory.h>
  17 #include <rte_malloc.h>
  18 #include <rte_vhost.h>
  19
  20 #include "iotlb.h"
  21 #include "vhost.h"
  22 #include "vhost_user.h"
  23
  24 struct virtio_net *vhost_devices[RTE_MAX_VHOST_DEVICE];
  25 pthread_mutex_t vhost_dev_lock = PTHREAD_MUTEX_INITIALIZER;
  26
  27 struct vhost_vq_stats_name_off {
  28         char name[RTE_VHOST_STATS_NAME_SIZE];
  29         unsigned int offset;
  30 };
  31
  32 static const struct vhost_vq_stats_name_off vhost_vq_stat_strings[] = {
  33         {"good_packets",           offsetof(struct vhost_virtqueue, stats.packets)},
  34         {"good_bytes",             offsetof(struct vhost_virtqueue, stats.bytes)},
  35         {"multicast_packets",      offsetof(struct vhost_virtqueue, stats.multicast)},
  36         {"broadcast_packets",      offsetof(struct vhost_virtqueue, stats.broadcast)},
  37         {"undersize_packets",      offsetof(struct vhost_virtqueue, stats.size_bins[0])},
  38         {"size_64_packets",        offsetof(struct vhost_virtqueue, stats.size_bins[1])},
  39         {"size_65_127_packets",    offsetof(struct vhost_virtqueue, stats.size_bins[2])},
  40         {"size_128_255_packets",   offsetof(struct vhost_virtqueue, stats.size_bins[3])},
  41         {"size_256_511_packets",   offsetof(struct vhost_virtqueue, stats.size_bins[4])},
  42         {"size_512_1023_packets",  offsetof(struct vhost_virtqueue, stats.size_bins[5])},
  43         {"size_1024_1518_packets", offsetof(struct vhost_virtqueue, stats.size_bins[6])},
  44         {"size_1519_max_packets",  offsetof(struct vhost_virtqueue, stats.size_bins[7])},
  45         {"guest_notifications",    offsetof(struct vhost_virtqueue, stats.guest_notifications)},
  46 };
  47
  48 #define VHOST_NB_VQ_STATS RTE_DIM(vhost_vq_stat_strings)
  49
  50 /* Called with iotlb_lock read-locked */
  51 uint64_t
  52 __vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
  53                     uint64_t iova, uint64_t *size, uint8_t perm)
  54 {
  55         uint64_t vva, tmp_size;
  56
  57         if (unlikely(!*size))
  58                 return 0;
  59
  60         tmp_size = *size;
  61
  62         vva = vhost_user_iotlb_cache_find(vq, iova, &tmp_size, perm);
  63         if (tmp_size == *size)
  64                 return vva;
  65
  66         iova += tmp_size;
  67
  68         if (!vhost_user_iotlb_pending_miss(vq, iova, perm)) {
  69                 /*
  70                  * iotlb_lock is read-locked for a full burst,
  71                  * but it only protects the iotlb cache.
  72                  * In case of IOTLB miss, we might block on the socket,
  73                  * which could cause a deadlock with QEMU if an IOTLB update
  74                  * is being handled. We can safely unlock here to avoid it.
  75                  */
  76                 vhost_user_iotlb_rd_unlock(vq);
  77
  78                 vhost_user_iotlb_pending_insert(dev, vq, iova, perm);
  79                 if (vhost_user_iotlb_miss(dev, iova, perm)) {
  80                         VHOST_LOG_DATA(ERR, "(%s) IOTLB miss req failed for IOVA 0x%" PRIx64 "\n",
  81                                 dev->ifname, iova);
  82                         vhost_user_iotlb_pending_remove(vq, iova, 1, perm);
  83                 }
  84
  85                 vhost_user_iotlb_rd_lock(vq);
  86         }
  87
  88         return 0;
  89 }
  90
  91 #define VHOST_LOG_PAGE  4096
  92
  93 /*
  94  * Atomically set a bit in memory.
  95  */
  96 static __rte_always_inline void
  97 vhost_set_bit(unsigned int nr, volatile uint8_t *addr)
  98 {
  99 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)
 100         /*
 101          * __sync_ built-ins are deprecated, but __atomic_ ones
 102          * are sub-optimized in older GCC versions.
 103          */
 104         __sync_fetch_and_or_1(addr, (1U << nr));
 105 #else
 106         __atomic_fetch_or(addr, (1U << nr), __ATOMIC_RELAXED);
 107 #endif
 108 }
 109
 110 static __rte_always_inline void
 111 vhost_log_page(uint8_t *log_base, uint64_t page)
 112 {
 113         vhost_set_bit(page % 8, &log_base[page / 8]);
 114 }
 115
 116 void
 117 __vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
 118 {
 119         uint64_t page;
 120
 121         if (unlikely(!dev->log_base || !len))
 122                 return;
 123
 124         if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
 125                 return;
 126
 127         /* To make sure guest memory updates are committed before logging */
 128         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 129
 130         page = addr / VHOST_LOG_PAGE;
 131         while (page * VHOST_LOG_PAGE < addr + len) {
 132                 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
 133                 page += 1;
 134         }
 135 }
 136
 137 void
 138 __vhost_log_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
 139                              uint64_t iova, uint64_t len)
 140 {
 141         uint64_t hva, gpa, map_len;
 142         map_len = len;
 143
 144         hva = __vhost_iova_to_vva(dev, vq, iova, &map_len, VHOST_ACCESS_RW);
 145         if (map_len != len) {
 146                 VHOST_LOG_DATA(ERR,
 147                         "(%s) failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
 148                         dev->ifname, iova);
 149                 return;
 150         }
 151
 152         gpa = hva_to_gpa(dev, hva, len);
 153         if (gpa)
 154                 __vhost_log_write(dev, gpa, len);
 155 }
 156
 157 void
 158 __vhost_log_cache_sync(struct virtio_net *dev, struct vhost_virtqueue *vq)
 159 {
 160         unsigned long *log_base;
 161         int i;
 162
 163         if (unlikely(!dev->log_base))
 164                 return;
 165
 166         /* No cache, nothing to sync */
 167         if (unlikely(!vq->log_cache))
 168                 return;
 169
 170         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 171
 172         log_base = (unsigned long *)(uintptr_t)dev->log_base;
 173
 174         for (i = 0; i < vq->log_cache_nb_elem; i++) {
 175                 struct log_cache_entry *elem = vq->log_cache + i;
 176
 177 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)
 178                 /*
 179                  * '__sync' builtins are deprecated, but '__atomic' ones
 180                  * are sub-optimized in older GCC versions.
 181                  */
 182                 __sync_fetch_and_or(log_base + elem->offset, elem->val);
 183 #else
 184                 __atomic_fetch_or(log_base + elem->offset, elem->val,
 185                                 __ATOMIC_RELAXED);
 186 #endif
 187         }
 188
 189         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 190
 191         vq->log_cache_nb_elem = 0;
 192 }
 193
 194 static __rte_always_inline void
 195 vhost_log_cache_page(struct virtio_net *dev, struct vhost_virtqueue *vq,
 196                         uint64_t page)
 197 {
 198         uint32_t bit_nr = page % (sizeof(unsigned long) << 3);
 199         uint32_t offset = page / (sizeof(unsigned long) << 3);
 200         int i;
 201
 202         if (unlikely(!vq->log_cache)) {
 203                 /* No logging cache allocated, write dirty log map directly */
 204                 rte_atomic_thread_fence(__ATOMIC_RELEASE);
 205                 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
 206
 207                 return;
 208         }
 209
 210         for (i = 0; i < vq->log_cache_nb_elem; i++) {
 211                 struct log_cache_entry *elem = vq->log_cache + i;
 212
 213                 if (elem->offset == offset) {
 214                         elem->val |= (1UL << bit_nr);
 215                         return;
 216                 }
 217         }
 218
 219         if (unlikely(i >= VHOST_LOG_CACHE_NR)) {
 220                 /*
 221                  * No more room for a new log cache entry,
 222                  * so write the dirty log map directly.
 223                  */
 224                 rte_atomic_thread_fence(__ATOMIC_RELEASE);
 225                 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
 226
 227                 return;
 228         }
 229
 230         vq->log_cache[i].offset = offset;
 231         vq->log_cache[i].val = (1UL << bit_nr);
 232         vq->log_cache_nb_elem++;
 233 }
 234
 235 void
 236 __vhost_log_cache_write(struct virtio_net *dev, struct vhost_virtqueue *vq,
 237                         uint64_t addr, uint64_t len)
 238 {
 239         uint64_t page;
 240
 241         if (unlikely(!dev->log_base || !len))
 242                 return;
 243
 244         if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
 245                 return;
 246
 247         page = addr / VHOST_LOG_PAGE;
 248         while (page * VHOST_LOG_PAGE < addr + len) {
 249                 vhost_log_cache_page(dev, vq, page);
 250                 page += 1;
 251         }
 252 }
 253
 254 void
 255 __vhost_log_cache_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
 256                              uint64_t iova, uint64_t len)
 257 {
 258         uint64_t hva, gpa, map_len;
 259         map_len = len;
 260
 261         hva = __vhost_iova_to_vva(dev, vq, iova, &map_len, VHOST_ACCESS_RW);
 262         if (map_len != len) {
 263                 VHOST_LOG_DATA(ERR,
 264                         "(%s) failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
 265                         dev->ifname, iova);
 266                 return;
 267         }
 268
 269         gpa = hva_to_gpa(dev, hva, len);
 270         if (gpa)
 271                 __vhost_log_cache_write(dev, vq, gpa, len);
 272 }
 273
 274 void *
 275 vhost_alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
 276                 uint64_t desc_addr, uint64_t desc_len)
 277 {
 278         void *idesc;
 279         uint64_t src, dst;
 280         uint64_t len, remain = desc_len;
 281
 282         idesc = rte_malloc_socket(__func__, desc_len, 0, vq->numa_node);
 283         if (unlikely(!idesc))
 284                 return NULL;
 285
 286         dst = (uint64_t)(uintptr_t)idesc;
 287
 288         while (remain) {
 289                 len = remain;
 290                 src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
 291                                 VHOST_ACCESS_RO);
 292                 if (unlikely(!src || !len)) {
 293                         rte_free(idesc);
 294                         return NULL;
 295                 }
 296
 297                 rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
 298
 299                 remain -= len;
 300                 dst += len;
 301                 desc_addr += len;
 302         }
 303
 304         return idesc;
 305 }
 306
 307 void
 308 cleanup_vq(struct vhost_virtqueue *vq, int destroy)
 309 {
 310         if ((vq->callfd >= 0) && (destroy != 0))
 311                 close(vq->callfd);
 312         if (vq->kickfd >= 0)
 313                 close(vq->kickfd);
 314 }
 315
 316 void
 317 cleanup_vq_inflight(struct virtio_net *dev, struct vhost_virtqueue *vq)
 318 {
 319         if (!(dev->protocol_features &
 320             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
 321                 return;
 322
 323         if (vq_is_packed(dev)) {
 324                 if (vq->inflight_packed)
 325                         vq->inflight_packed = NULL;
 326         } else {
 327                 if (vq->inflight_split)
 328                         vq->inflight_split = NULL;
 329         }
 330
 331         if (vq->resubmit_inflight) {
 332                 if (vq->resubmit_inflight->resubmit_list) {
 333                         rte_free(vq->resubmit_inflight->resubmit_list);
 334                         vq->resubmit_inflight->resubmit_list = NULL;
 335                 }
 336                 rte_free(vq->resubmit_inflight);
 337                 vq->resubmit_inflight = NULL;
 338         }
 339 }
 340
 341 /*
 342  * Unmap any memory, close any file descriptors and
 343  * free any memory owned by a device.
 344  */
 345 void
 346 cleanup_device(struct virtio_net *dev, int destroy)
 347 {
 348         uint32_t i;
 349
 350         vhost_backend_cleanup(dev);
 351
 352         for (i = 0; i < dev->nr_vring; i++) {
 353                 cleanup_vq(dev->virtqueue[i], destroy);
 354                 cleanup_vq_inflight(dev, dev->virtqueue[i]);
 355         }
 356 }
 357
 358 static void
 359 vhost_free_async_mem(struct vhost_virtqueue *vq)
 360 {
 361         if (!vq->async)
 362                 return;
 363
 364         rte_free(vq->async->pkts_info);
 365         rte_free(vq->async->pkts_cmpl_flag);
 366
 367         rte_free(vq->async->buffers_packed);
 368         vq->async->buffers_packed = NULL;
 369         rte_free(vq->async->descs_split);
 370         vq->async->descs_split = NULL;
 371
 372         rte_free(vq->async);
 373         vq->async = NULL;
 374 }
 375
 376 void
 377 free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq)
 378 {
 379         if (vq_is_packed(dev))
 380                 rte_free(vq->shadow_used_packed);
 381         else
 382                 rte_free(vq->shadow_used_split);
 383
 384         vhost_free_async_mem(vq);
 385         rte_free(vq->batch_copy_elems);
 386         rte_mempool_free(vq->iotlb_pool);
 387         rte_free(vq->log_cache);
 388         rte_free(vq);
 389 }
 390
 391 /*
 392  * Release virtqueues and device memory.
 393  */
 394 static void
 395 free_device(struct virtio_net *dev)
 396 {
 397         uint32_t i;
 398
 399         for (i = 0; i < dev->nr_vring; i++)
 400                 free_vq(dev, dev->virtqueue[i]);
 401
 402         rte_free(dev);
 403 }
 404
 405 static __rte_always_inline int
 406 log_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
 407 {
 408         if (likely(!(vq->ring_addrs.flags & (1 << VHOST_VRING_F_LOG))))
 409                 return 0;
 410
 411         vq->log_guest_addr = translate_log_addr(dev, vq,
 412                                                 vq->ring_addrs.log_guest_addr);
 413         if (vq->log_guest_addr == 0)
 414                 return -1;
 415
 416         return 0;
 417 }
 418
 419 /*
 420  * Converts vring log address to GPA
 421  * If IOMMU is enabled, the log address is IOVA
 422  * If IOMMU not enabled, the log address is already GPA
 423  *
 424  * Caller should have iotlb_lock read-locked
 425  */
 426 uint64_t
 427 translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
 428                 uint64_t log_addr)
 429 {
 430         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 431                 const uint64_t exp_size = sizeof(uint64_t);
 432                 uint64_t hva, gpa;
 433                 uint64_t size = exp_size;
 434
 435                 hva = vhost_iova_to_vva(dev, vq, log_addr,
 436                                         &size, VHOST_ACCESS_RW);
 437
 438                 if (size != exp_size)
 439                         return 0;
 440
 441                 gpa = hva_to_gpa(dev, hva, exp_size);
 442                 if (!gpa) {
 443                         VHOST_LOG_DATA(ERR,
 444                                 "(%s) failed to find GPA for log_addr: 0x%"
 445                                 PRIx64 " hva: 0x%" PRIx64 "\n",
 446                                 dev->ifname, log_addr, hva);
 447                         return 0;
 448                 }
 449                 return gpa;
 450
 451         } else
 452                 return log_addr;
 453 }
 454
 455 /* Caller should have iotlb_lock read-locked */
 456 static int
 457 vring_translate_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
 458 {
 459         uint64_t req_size, size;
 460
 461         req_size = sizeof(struct vring_desc) * vq->size;
 462         size = req_size;
 463         vq->desc = (struct vring_desc *)(uintptr_t)vhost_iova_to_vva(dev, vq,
 464                                                 vq->ring_addrs.desc_user_addr,
 465                                                 &size, VHOST_ACCESS_RW);
 466         if (!vq->desc || size != req_size)
 467                 return -1;
 468
 469         req_size = sizeof(struct vring_avail);
 470         req_size += sizeof(uint16_t) * vq->size;
 471         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 472                 req_size += sizeof(uint16_t);
 473         size = req_size;
 474         vq->avail = (struct vring_avail *)(uintptr_t)vhost_iova_to_vva(dev, vq,
 475                                                 vq->ring_addrs.avail_user_addr,
 476                                                 &size, VHOST_ACCESS_RW);
 477         if (!vq->avail || size != req_size)
 478                 return -1;
 479
 480         req_size = sizeof(struct vring_used);
 481         req_size += sizeof(struct vring_used_elem) * vq->size;
 482         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 483                 req_size += sizeof(uint16_t);
 484         size = req_size;
 485         vq->used = (struct vring_used *)(uintptr_t)vhost_iova_to_vva(dev, vq,
 486                                                 vq->ring_addrs.used_user_addr,
 487                                                 &size, VHOST_ACCESS_RW);
 488         if (!vq->used || size != req_size)
 489                 return -1;
 490
 491         return 0;
 492 }
 493
 494 /* Caller should have iotlb_lock read-locked */
 495 static int
 496 vring_translate_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
 497 {
 498         uint64_t req_size, size;
 499
 500         req_size = sizeof(struct vring_packed_desc) * vq->size;
 501         size = req_size;
 502         vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 503                 vhost_iova_to_vva(dev, vq, vq->ring_addrs.desc_user_addr,
 504                                 &size, VHOST_ACCESS_RW);
 505         if (!vq->desc_packed || size != req_size)
 506                 return -1;
 507
 508         req_size = sizeof(struct vring_packed_desc_event);
 509         size = req_size;
 510         vq->driver_event = (struct vring_packed_desc_event *)(uintptr_t)
 511                 vhost_iova_to_vva(dev, vq, vq->ring_addrs.avail_user_addr,
 512                                 &size, VHOST_ACCESS_RW);
 513         if (!vq->driver_event || size != req_size)
 514                 return -1;
 515
 516         req_size = sizeof(struct vring_packed_desc_event);
 517         size = req_size;
 518         vq->device_event = (struct vring_packed_desc_event *)(uintptr_t)
 519                 vhost_iova_to_vva(dev, vq, vq->ring_addrs.used_user_addr,
 520                                 &size, VHOST_ACCESS_RW);
 521         if (!vq->device_event || size != req_size)
 522                 return -1;
 523
 524         return 0;
 525 }
 526
 527 int
 528 vring_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
 529 {
 530
 531         if (!(dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
 532                 return -1;
 533
 534         if (vq_is_packed(dev)) {
 535                 if (vring_translate_packed(dev, vq) < 0)
 536                         return -1;
 537         } else {
 538                 if (vring_translate_split(dev, vq) < 0)
 539                         return -1;
 540         }
 541
 542         if (log_translate(dev, vq) < 0)
 543                 return -1;
 544
 545         vq->access_ok = true;
 546
 547         return 0;
 548 }
 549
 550 void
 551 vring_invalidate(struct virtio_net *dev, struct vhost_virtqueue *vq)
 552 {
 553         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 554                 vhost_user_iotlb_wr_lock(vq);
 555
 556         vq->access_ok = false;
 557         vq->desc = NULL;
 558         vq->avail = NULL;
 559         vq->used = NULL;
 560         vq->log_guest_addr = 0;
 561
 562         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 563                 vhost_user_iotlb_wr_unlock(vq);
 564 }
 565
 566 static void
 567 init_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
 568 {
 569         struct vhost_virtqueue *vq;
 570         int numa_node = SOCKET_ID_ANY;
 571
 572         if (vring_idx >= VHOST_MAX_VRING) {
 573                 VHOST_LOG_CONFIG(ERR, "(%s) failed to init vring, out of bound (%d)\n",
 574                                 dev->ifname, vring_idx);
 575                 return;
 576         }
 577
 578         vq = dev->virtqueue[vring_idx];
 579         if (!vq) {
 580                 VHOST_LOG_CONFIG(ERR, "(%s) virtqueue not allocated (%d)\n",
 581                                 dev->ifname, vring_idx);
 582                 return;
 583         }
 584
 585         memset(vq, 0, sizeof(struct vhost_virtqueue));
 586
 587         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
 588         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
 589         vq->notif_enable = VIRTIO_UNINITIALIZED_NOTIF;
 590
 591 #ifdef RTE_LIBRTE_VHOST_NUMA
 592         if (get_mempolicy(&numa_node, NULL, 0, vq, MPOL_F_NODE | MPOL_F_ADDR)) {
 593                 VHOST_LOG_CONFIG(ERR, "(%s) failed to query numa node: %s\n",
 594                         dev->ifname, rte_strerror(errno));
 595                 numa_node = SOCKET_ID_ANY;
 596         }
 597 #endif
 598         vq->numa_node = numa_node;
 599
 600         vhost_user_iotlb_init(dev, vring_idx);
 601 }
 602
 603 static void
 604 reset_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
 605 {
 606         struct vhost_virtqueue *vq;
 607         int callfd;
 608
 609         if (vring_idx >= VHOST_MAX_VRING) {
 610                 VHOST_LOG_CONFIG(ERR,
 611                                 "(%s) failed to reset vring, out of bound (%d)\n",
 612                                 dev->ifname, vring_idx);
 613                 return;
 614         }
 615
 616         vq = dev->virtqueue[vring_idx];
 617         if (!vq) {
 618                 VHOST_LOG_CONFIG(ERR, "(%s) failed to reset vring, virtqueue not allocated (%d)\n",
 619                                 dev->ifname, vring_idx);
 620                 return;
 621         }
 622
 623         callfd = vq->callfd;
 624         init_vring_queue(dev, vring_idx);
 625         vq->callfd = callfd;
 626 }
 627
 628 int
 629 alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
 630 {
 631         struct vhost_virtqueue *vq;
 632         uint32_t i;
 633
 634         /* Also allocate holes, if any, up to requested vring index. */
 635         for (i = 0; i <= vring_idx; i++) {
 636                 if (dev->virtqueue[i])
 637                         continue;
 638
 639                 vq = rte_zmalloc(NULL, sizeof(struct vhost_virtqueue), 0);
 640                 if (vq == NULL) {
 641                         VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate memory for vring %u.\n",
 642                                         dev->ifname, i);
 643                         return -1;
 644                 }
 645
 646                 dev->virtqueue[i] = vq;
 647                 init_vring_queue(dev, i);
 648                 rte_spinlock_init(&vq->access_lock);
 649                 vq->avail_wrap_counter = 1;
 650                 vq->used_wrap_counter = 1;
 651                 vq->signalled_used_valid = false;
 652         }
 653
 654         dev->nr_vring = RTE_MAX(dev->nr_vring, vring_idx + 1);
 655
 656         return 0;
 657 }
 658
 659 /*
 660  * Reset some variables in device structure, while keeping few
 661  * others untouched, such as vid, ifname, nr_vring: they
 662  * should be same unless the device is removed.
 663  */
 664 void
 665 reset_device(struct virtio_net *dev)
 666 {
 667         uint32_t i;
 668
 669         dev->features = 0;
 670         dev->protocol_features = 0;
 671         dev->flags &= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
 672
 673         for (i = 0; i < dev->nr_vring; i++)
 674                 reset_vring_queue(dev, i);
 675 }
 676
 677 /*
 678  * Invoked when there is a new vhost-user connection established (when
 679  * there is a new virtio device being attached).
 680  */
 681 int
 682 vhost_new_device(void)
 683 {
 684         struct virtio_net *dev;
 685         int i;
 686
 687         pthread_mutex_lock(&vhost_dev_lock);
 688         for (i = 0; i < RTE_MAX_VHOST_DEVICE; i++) {
 689                 if (vhost_devices[i] == NULL)
 690                         break;
 691         }
 692
 693         if (i == RTE_MAX_VHOST_DEVICE) {
 694                 VHOST_LOG_CONFIG(ERR, "failed to find a free slot for new device.\n");
 695                 pthread_mutex_unlock(&vhost_dev_lock);
 696                 return -1;
 697         }
 698
 699         dev = rte_zmalloc(NULL, sizeof(struct virtio_net), 0);
 700         if (dev == NULL) {
 701                 VHOST_LOG_CONFIG(ERR, "failed to allocate memory for new device.\n");
 702                 pthread_mutex_unlock(&vhost_dev_lock);
 703                 return -1;
 704         }
 705
 706         vhost_devices[i] = dev;
 707         pthread_mutex_unlock(&vhost_dev_lock);
 708
 709         dev->vid = i;
 710         dev->flags = VIRTIO_DEV_BUILTIN_VIRTIO_NET;
 711         dev->slave_req_fd = -1;
 712         dev->postcopy_ufd = -1;
 713         rte_spinlock_init(&dev->slave_req_lock);
 714
 715         return i;
 716 }
 717
 718 void
 719 vhost_destroy_device_notify(struct virtio_net *dev)
 720 {
 721         struct rte_vdpa_device *vdpa_dev;
 722
 723         if (dev->flags & VIRTIO_DEV_RUNNING) {
 724                 vdpa_dev = dev->vdpa_dev;
 725                 if (vdpa_dev)
 726                         vdpa_dev->ops->dev_close(dev->vid);
 727                 dev->flags &= ~VIRTIO_DEV_RUNNING;
 728                 dev->notify_ops->destroy_device(dev->vid);
 729         }
 730 }
 731
 732 /*
 733  * Invoked when there is the vhost-user connection is broken (when
 734  * the virtio device is being detached).
 735  */
 736 void
 737 vhost_destroy_device(int vid)
 738 {
 739         struct virtio_net *dev = get_device(vid);
 740
 741         if (dev == NULL)
 742                 return;
 743
 744         vhost_destroy_device_notify(dev);
 745
 746         cleanup_device(dev, 1);
 747         free_device(dev);
 748
 749         vhost_devices[vid] = NULL;
 750 }
 751
 752 void
 753 vhost_attach_vdpa_device(int vid, struct rte_vdpa_device *vdpa_dev)
 754 {
 755         struct virtio_net *dev = get_device(vid);
 756
 757         if (dev == NULL)
 758                 return;
 759
 760         dev->vdpa_dev = vdpa_dev;
 761 }
 762
 763 void
 764 vhost_set_ifname(int vid, const char *if_name, unsigned int if_len)
 765 {
 766         struct virtio_net *dev;
 767         unsigned int len;
 768
 769         dev = get_device(vid);
 770         if (dev == NULL)
 771                 return;
 772
 773         len = if_len > sizeof(dev->ifname) ?
 774                 sizeof(dev->ifname) : if_len;
 775
 776         strncpy(dev->ifname, if_name, len);
 777         dev->ifname[sizeof(dev->ifname) - 1] = '\0';
 778 }
 779
 780 void
 781 vhost_setup_virtio_net(int vid, bool enable, bool compliant_ol_flags, bool stats_enabled)
 782 {
 783         struct virtio_net *dev = get_device(vid);
 784
 785         if (dev == NULL)
 786                 return;
 787
 788         if (enable)
 789                 dev->flags |= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
 790         else
 791                 dev->flags &= ~VIRTIO_DEV_BUILTIN_VIRTIO_NET;
 792         if (!compliant_ol_flags)
 793                 dev->flags |= VIRTIO_DEV_LEGACY_OL_FLAGS;
 794         else
 795                 dev->flags &= ~VIRTIO_DEV_LEGACY_OL_FLAGS;
 796         if (stats_enabled)
 797                 dev->flags |= VIRTIO_DEV_STATS_ENABLED;
 798         else
 799                 dev->flags &= ~VIRTIO_DEV_STATS_ENABLED;
 800 }
 801
 802 void
 803 vhost_enable_extbuf(int vid)
 804 {
 805         struct virtio_net *dev = get_device(vid);
 806
 807         if (dev == NULL)
 808                 return;
 809
 810         dev->extbuf = 1;
 811 }
 812
 813 void
 814 vhost_enable_linearbuf(int vid)
 815 {
 816         struct virtio_net *dev = get_device(vid);
 817
 818         if (dev == NULL)
 819                 return;
 820
 821         dev->linearbuf = 1;
 822 }
 823
 824 int
 825 rte_vhost_get_mtu(int vid, uint16_t *mtu)
 826 {
 827         struct virtio_net *dev = get_device(vid);
 828
 829         if (dev == NULL || mtu == NULL)
 830                 return -ENODEV;
 831
 832         if (!(dev->flags & VIRTIO_DEV_READY))
 833                 return -EAGAIN;
 834
 835         if (!(dev->features & (1ULL << VIRTIO_NET_F_MTU)))
 836                 return -ENOTSUP;
 837
 838         *mtu = dev->mtu;
 839
 840         return 0;
 841 }
 842
 843 int
 844 rte_vhost_get_numa_node(int vid)
 845 {
 846 #ifdef RTE_LIBRTE_VHOST_NUMA
 847         struct virtio_net *dev = get_device(vid);
 848         int numa_node;
 849         int ret;
 850
 851         if (dev == NULL || numa_available() != 0)
 852                 return -1;
 853
 854         ret = get_mempolicy(&numa_node, NULL, 0, dev,
 855                             MPOL_F_NODE | MPOL_F_ADDR);
 856         if (ret < 0) {
 857                 VHOST_LOG_CONFIG(ERR, "(%s) failed to query numa node: %s\n",
 858                         dev->ifname, rte_strerror(errno));
 859                 return -1;
 860         }
 861
 862         return numa_node;
 863 #else
 864         RTE_SET_USED(vid);
 865         return -1;
 866 #endif
 867 }
 868
 869 uint32_t
 870 rte_vhost_get_queue_num(int vid)
 871 {
 872         struct virtio_net *dev = get_device(vid);
 873
 874         if (dev == NULL)
 875                 return 0;
 876
 877         return dev->nr_vring / 2;
 878 }
 879
 880 uint16_t
 881 rte_vhost_get_vring_num(int vid)
 882 {
 883         struct virtio_net *dev = get_device(vid);
 884
 885         if (dev == NULL)
 886                 return 0;
 887
 888         return dev->nr_vring;
 889 }
 890
 891 int
 892 rte_vhost_get_ifname(int vid, char *buf, size_t len)
 893 {
 894         struct virtio_net *dev = get_device(vid);
 895
 896         if (dev == NULL || buf == NULL)
 897                 return -1;
 898
 899         len = RTE_MIN(len, sizeof(dev->ifname));
 900
 901         strncpy(buf, dev->ifname, len);
 902         buf[len - 1] = '\0';
 903
 904         return 0;
 905 }
 906
 907 int
 908 rte_vhost_get_negotiated_features(int vid, uint64_t *features)
 909 {
 910         struct virtio_net *dev;
 911
 912         dev = get_device(vid);
 913         if (dev == NULL || features == NULL)
 914                 return -1;
 915
 916         *features = dev->features;
 917         return 0;
 918 }
 919
 920 int
 921 rte_vhost_get_negotiated_protocol_features(int vid,
 922                                            uint64_t *protocol_features)
 923 {
 924         struct virtio_net *dev;
 925
 926         dev = get_device(vid);
 927         if (dev == NULL || protocol_features == NULL)
 928                 return -1;
 929
 930         *protocol_features = dev->protocol_features;
 931         return 0;
 932 }
 933
 934 int
 935 rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
 936 {
 937         struct virtio_net *dev;
 938         struct rte_vhost_memory *m;
 939         size_t size;
 940
 941         dev = get_device(vid);
 942         if (dev == NULL || mem == NULL)
 943                 return -1;
 944
 945         size = dev->mem->nregions * sizeof(struct rte_vhost_mem_region);
 946         m = malloc(sizeof(struct rte_vhost_memory) + size);
 947         if (!m)
 948                 return -1;
 949
 950         m->nregions = dev->mem->nregions;
 951         memcpy(m->regions, dev->mem->regions, size);
 952         *mem = m;
 953
 954         return 0;
 955 }
 956
 957 int
 958 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
 959                           struct rte_vhost_vring *vring)
 960 {
 961         struct virtio_net *dev;
 962         struct vhost_virtqueue *vq;
 963
 964         dev = get_device(vid);
 965         if (dev == NULL || vring == NULL)
 966                 return -1;
 967
 968         if (vring_idx >= VHOST_MAX_VRING)
 969                 return -1;
 970
 971         vq = dev->virtqueue[vring_idx];
 972         if (!vq)
 973                 return -1;
 974
 975         if (vq_is_packed(dev)) {
 976                 vring->desc_packed = vq->desc_packed;
 977                 vring->driver_event = vq->driver_event;
 978                 vring->device_event = vq->device_event;
 979         } else {
 980                 vring->desc = vq->desc;
 981                 vring->avail = vq->avail;
 982                 vring->used = vq->used;
 983         }
 984         vring->log_guest_addr  = vq->log_guest_addr;
 985
 986         vring->callfd  = vq->callfd;
 987         vring->kickfd  = vq->kickfd;
 988         vring->size    = vq->size;
 989
 990         return 0;
 991 }
 992
 993 int
 994 rte_vhost_get_vhost_ring_inflight(int vid, uint16_t vring_idx,
 995                                   struct rte_vhost_ring_inflight *vring)
 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 (vring_idx >= VHOST_MAX_VRING)
1005                 return -1;
1006
1007         vq = dev->virtqueue[vring_idx];
1008         if (unlikely(!vq))
1009                 return -1;
1010
1011         if (vq_is_packed(dev)) {
1012                 if (unlikely(!vq->inflight_packed))
1013                         return -1;
1014
1015                 vring->inflight_packed = vq->inflight_packed;
1016         } else {
1017                 if (unlikely(!vq->inflight_split))
1018                         return -1;
1019
1020                 vring->inflight_split = vq->inflight_split;
1021         }
1022
1023         vring->resubmit_inflight = vq->resubmit_inflight;
1024
1025         return 0;
1026 }
1027
1028 int
1029 rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
1030                                   uint16_t idx)
1031 {
1032         struct vhost_virtqueue *vq;
1033         struct virtio_net *dev;
1034
1035         dev = get_device(vid);
1036         if (unlikely(!dev))
1037                 return -1;
1038
1039         if (unlikely(!(dev->protocol_features &
1040             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1041                 return 0;
1042
1043         if (unlikely(vq_is_packed(dev)))
1044                 return -1;
1045
1046         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1047                 return -1;
1048
1049         vq = dev->virtqueue[vring_idx];
1050         if (unlikely(!vq))
1051                 return -1;
1052
1053         if (unlikely(!vq->inflight_split))
1054                 return -1;
1055
1056         if (unlikely(idx >= vq->size))
1057                 return -1;
1058
1059         vq->inflight_split->desc[idx].counter = vq->global_counter++;
1060         vq->inflight_split->desc[idx].inflight = 1;
1061         return 0;
1062 }
1063
1064 int
1065 rte_vhost_set_inflight_desc_packed(int vid, uint16_t vring_idx,
1066                                    uint16_t head, uint16_t last,
1067                                    uint16_t *inflight_entry)
1068 {
1069         struct rte_vhost_inflight_info_packed *inflight_info;
1070         struct virtio_net *dev;
1071         struct vhost_virtqueue *vq;
1072         struct vring_packed_desc *desc;
1073         uint16_t old_free_head, free_head;
1074
1075         dev = get_device(vid);
1076         if (unlikely(!dev))
1077                 return -1;
1078
1079         if (unlikely(!(dev->protocol_features &
1080             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1081                 return 0;
1082
1083         if (unlikely(!vq_is_packed(dev)))
1084                 return -1;
1085
1086         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1087                 return -1;
1088
1089         vq = dev->virtqueue[vring_idx];
1090         if (unlikely(!vq))
1091                 return -1;
1092
1093         inflight_info = vq->inflight_packed;
1094         if (unlikely(!inflight_info))
1095                 return -1;
1096
1097         if (unlikely(head >= vq->size))
1098                 return -1;
1099
1100         desc = vq->desc_packed;
1101         old_free_head = inflight_info->old_free_head;
1102         if (unlikely(old_free_head >= vq->size))
1103                 return -1;
1104
1105         free_head = old_free_head;
1106
1107         /* init header descriptor */
1108         inflight_info->desc[old_free_head].num = 0;
1109         inflight_info->desc[old_free_head].counter = vq->global_counter++;
1110         inflight_info->desc[old_free_head].inflight = 1;
1111
1112         /* save desc entry in flight entry */
1113         while (head != ((last + 1) % vq->size)) {
1114                 inflight_info->desc[old_free_head].num++;
1115                 inflight_info->desc[free_head].addr = desc[head].addr;
1116                 inflight_info->desc[free_head].len = desc[head].len;
1117                 inflight_info->desc[free_head].flags = desc[head].flags;
1118                 inflight_info->desc[free_head].id = desc[head].id;
1119
1120                 inflight_info->desc[old_free_head].last = free_head;
1121                 free_head = inflight_info->desc[free_head].next;
1122                 inflight_info->free_head = free_head;
1123                 head = (head + 1) % vq->size;
1124         }
1125
1126         inflight_info->old_free_head = free_head;
1127         *inflight_entry = old_free_head;
1128
1129         return 0;
1130 }
1131
1132 int
1133 rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
1134                                   uint16_t last_used_idx, uint16_t idx)
1135 {
1136         struct virtio_net *dev;
1137         struct vhost_virtqueue *vq;
1138
1139         dev = get_device(vid);
1140         if (unlikely(!dev))
1141                 return -1;
1142
1143         if (unlikely(!(dev->protocol_features &
1144             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1145                 return 0;
1146
1147         if (unlikely(vq_is_packed(dev)))
1148                 return -1;
1149
1150         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1151                 return -1;
1152
1153         vq = dev->virtqueue[vring_idx];
1154         if (unlikely(!vq))
1155                 return -1;
1156
1157         if (unlikely(!vq->inflight_split))
1158                 return -1;
1159
1160         if (unlikely(idx >= vq->size))
1161                 return -1;
1162
1163         rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1164
1165         vq->inflight_split->desc[idx].inflight = 0;
1166
1167         rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1168
1169         vq->inflight_split->used_idx = last_used_idx;
1170         return 0;
1171 }
1172
1173 int
1174 rte_vhost_clr_inflight_desc_packed(int vid, uint16_t vring_idx,
1175                                    uint16_t head)
1176 {
1177         struct rte_vhost_inflight_info_packed *inflight_info;
1178         struct virtio_net *dev;
1179         struct vhost_virtqueue *vq;
1180
1181         dev = get_device(vid);
1182         if (unlikely(!dev))
1183                 return -1;
1184
1185         if (unlikely(!(dev->protocol_features &
1186             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1187                 return 0;
1188
1189         if (unlikely(!vq_is_packed(dev)))
1190                 return -1;
1191
1192         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1193                 return -1;
1194
1195         vq = dev->virtqueue[vring_idx];
1196         if (unlikely(!vq))
1197                 return -1;
1198
1199         inflight_info = vq->inflight_packed;
1200         if (unlikely(!inflight_info))
1201                 return -1;
1202
1203         if (unlikely(head >= vq->size))
1204                 return -1;
1205
1206         rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1207
1208         inflight_info->desc[head].inflight = 0;
1209
1210         rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1211
1212         inflight_info->old_free_head = inflight_info->free_head;
1213         inflight_info->old_used_idx = inflight_info->used_idx;
1214         inflight_info->old_used_wrap_counter = inflight_info->used_wrap_counter;
1215
1216         return 0;
1217 }
1218
1219 int
1220 rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
1221                                      uint16_t idx)
1222 {
1223         struct virtio_net *dev;
1224         struct vhost_virtqueue *vq;
1225
1226         dev = get_device(vid);
1227         if (unlikely(!dev))
1228                 return -1;
1229
1230         if (unlikely(!(dev->protocol_features &
1231             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1232                 return 0;
1233
1234         if (unlikely(vq_is_packed(dev)))
1235                 return -1;
1236
1237         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1238                 return -1;
1239
1240         vq = dev->virtqueue[vring_idx];
1241         if (unlikely(!vq))
1242                 return -1;
1243
1244         if (unlikely(!vq->inflight_split))
1245                 return -1;
1246
1247         if (unlikely(idx >= vq->size))
1248                 return -1;
1249
1250         vq->inflight_split->last_inflight_io = idx;
1251         return 0;
1252 }
1253
1254 int
1255 rte_vhost_set_last_inflight_io_packed(int vid, uint16_t vring_idx,
1256                                       uint16_t head)
1257 {
1258         struct rte_vhost_inflight_info_packed *inflight_info;
1259         struct virtio_net *dev;
1260         struct vhost_virtqueue *vq;
1261         uint16_t last;
1262
1263         dev = get_device(vid);
1264         if (unlikely(!dev))
1265                 return -1;
1266
1267         if (unlikely(!(dev->protocol_features &
1268             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1269                 return 0;
1270
1271         if (unlikely(!vq_is_packed(dev)))
1272                 return -1;
1273
1274         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1275                 return -1;
1276
1277         vq = dev->virtqueue[vring_idx];
1278         if (unlikely(!vq))
1279                 return -1;
1280
1281         inflight_info = vq->inflight_packed;
1282         if (unlikely(!inflight_info))
1283                 return -1;
1284
1285         if (unlikely(head >= vq->size))
1286                 return -1;
1287
1288         last = inflight_info->desc[head].last;
1289         if (unlikely(last >= vq->size))
1290                 return -1;
1291
1292         inflight_info->desc[last].next = inflight_info->free_head;
1293         inflight_info->free_head = head;
1294         inflight_info->used_idx += inflight_info->desc[head].num;
1295         if (inflight_info->used_idx >= inflight_info->desc_num) {
1296                 inflight_info->used_idx -= inflight_info->desc_num;
1297                 inflight_info->used_wrap_counter =
1298                         !inflight_info->used_wrap_counter;
1299         }
1300
1301         return 0;
1302 }
1303
1304 int
1305 rte_vhost_vring_call(int vid, uint16_t vring_idx)
1306 {
1307         struct virtio_net *dev;
1308         struct vhost_virtqueue *vq;
1309
1310         dev = get_device(vid);
1311         if (!dev)
1312                 return -1;
1313
1314         if (vring_idx >= VHOST_MAX_VRING)
1315                 return -1;
1316
1317         vq = dev->virtqueue[vring_idx];
1318         if (!vq)
1319                 return -1;
1320
1321         rte_spinlock_lock(&vq->access_lock);
1322
1323         if (vq_is_packed(dev))
1324                 vhost_vring_call_packed(dev, vq);
1325         else
1326                 vhost_vring_call_split(dev, vq);
1327
1328         rte_spinlock_unlock(&vq->access_lock);
1329
1330         return 0;
1331 }
1332
1333 uint16_t
1334 rte_vhost_avail_entries(int vid, uint16_t queue_id)
1335 {
1336         struct virtio_net *dev;
1337         struct vhost_virtqueue *vq;
1338         uint16_t ret = 0;
1339
1340         dev = get_device(vid);
1341         if (!dev)
1342                 return 0;
1343
1344         if (queue_id >= VHOST_MAX_VRING)
1345                 return 0;
1346
1347         vq = dev->virtqueue[queue_id];
1348         if (!vq)
1349                 return 0;
1350
1351         rte_spinlock_lock(&vq->access_lock);
1352
1353         if (unlikely(!vq->enabled || vq->avail == NULL))
1354                 goto out;
1355
1356         ret = *(volatile uint16_t *)&vq->avail->idx - vq->last_used_idx;
1357
1358 out:
1359         rte_spinlock_unlock(&vq->access_lock);
1360         return ret;
1361 }
1362
1363 static inline int
1364 vhost_enable_notify_split(struct virtio_net *dev,
1365                 struct vhost_virtqueue *vq, int enable)
1366 {
1367         if (vq->used == NULL)
1368                 return -1;
1369
1370         if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
1371                 if (enable)
1372                         vq->used->flags &= ~VRING_USED_F_NO_NOTIFY;
1373                 else
1374                         vq->used->flags |= VRING_USED_F_NO_NOTIFY;
1375         } else {
1376                 if (enable)
1377                         vhost_avail_event(vq) = vq->last_avail_idx;
1378         }
1379         return 0;
1380 }
1381
1382 static inline int
1383 vhost_enable_notify_packed(struct virtio_net *dev,
1384                 struct vhost_virtqueue *vq, int enable)
1385 {
1386         uint16_t flags;
1387
1388         if (vq->device_event == NULL)
1389                 return -1;
1390
1391         if (!enable) {
1392                 vq->device_event->flags = VRING_EVENT_F_DISABLE;
1393                 return 0;
1394         }
1395
1396         flags = VRING_EVENT_F_ENABLE;
1397         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
1398                 flags = VRING_EVENT_F_DESC;
1399                 vq->device_event->off_wrap = vq->last_avail_idx |
1400                         vq->avail_wrap_counter << 15;
1401         }
1402
1403         rte_atomic_thread_fence(__ATOMIC_RELEASE);
1404
1405         vq->device_event->flags = flags;
1406         return 0;
1407 }
1408
1409 int
1410 vhost_enable_guest_notification(struct virtio_net *dev,
1411                 struct vhost_virtqueue *vq, int enable)
1412 {
1413         /*
1414          * If the virtqueue is not ready yet, it will be applied
1415          * when it will become ready.
1416          */
1417         if (!vq->ready)
1418                 return 0;
1419
1420         if (vq_is_packed(dev))
1421                 return vhost_enable_notify_packed(dev, vq, enable);
1422         else
1423                 return vhost_enable_notify_split(dev, vq, enable);
1424 }
1425
1426 int
1427 rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
1428 {
1429         struct virtio_net *dev = get_device(vid);
1430         struct vhost_virtqueue *vq;
1431         int ret;
1432
1433         if (!dev)
1434                 return -1;
1435
1436         if (queue_id >= VHOST_MAX_VRING)
1437                 return -1;
1438
1439         vq = dev->virtqueue[queue_id];
1440         if (!vq)
1441                 return -1;
1442
1443         rte_spinlock_lock(&vq->access_lock);
1444
1445         vq->notif_enable = enable;
1446         ret = vhost_enable_guest_notification(dev, vq, enable);
1447
1448         rte_spinlock_unlock(&vq->access_lock);
1449
1450         return ret;
1451 }
1452
1453 void
1454 rte_vhost_log_write(int vid, uint64_t addr, uint64_t len)
1455 {
1456         struct virtio_net *dev = get_device(vid);
1457
1458         if (dev == NULL)
1459                 return;
1460
1461         vhost_log_write(dev, addr, len);
1462 }
1463
1464 void
1465 rte_vhost_log_used_vring(int vid, uint16_t vring_idx,
1466                          uint64_t offset, uint64_t len)
1467 {
1468         struct virtio_net *dev;
1469         struct vhost_virtqueue *vq;
1470
1471         dev = get_device(vid);
1472         if (dev == NULL)
1473                 return;
1474
1475         if (vring_idx >= VHOST_MAX_VRING)
1476                 return;
1477         vq = dev->virtqueue[vring_idx];
1478         if (!vq)
1479                 return;
1480
1481         vhost_log_used_vring(dev, vq, offset, len);
1482 }
1483
1484 uint32_t
1485 rte_vhost_rx_queue_count(int vid, uint16_t qid)
1486 {
1487         struct virtio_net *dev;
1488         struct vhost_virtqueue *vq;
1489         uint32_t ret = 0;
1490
1491         dev = get_device(vid);
1492         if (dev == NULL)
1493                 return 0;
1494
1495         if (unlikely(qid >= dev->nr_vring || (qid & 1) == 0)) {
1496                 VHOST_LOG_DATA(ERR, "(%s) %s: invalid virtqueue idx %d.\n",
1497                         dev->ifname, __func__, qid);
1498                 return 0;
1499         }
1500
1501         vq = dev->virtqueue[qid];
1502         if (vq == NULL)
1503                 return 0;
1504
1505         rte_spinlock_lock(&vq->access_lock);
1506
1507         if (unlikely(!vq->enabled || vq->avail == NULL))
1508                 goto out;
1509
1510         ret = *((volatile uint16_t *)&vq->avail->idx) - vq->last_avail_idx;
1511
1512 out:
1513         rte_spinlock_unlock(&vq->access_lock);
1514         return ret;
1515 }
1516
1517 struct rte_vdpa_device *
1518 rte_vhost_get_vdpa_device(int vid)
1519 {
1520         struct virtio_net *dev = get_device(vid);
1521
1522         if (dev == NULL)
1523                 return NULL;
1524
1525         return dev->vdpa_dev;
1526 }
1527
1528 int
1529 rte_vhost_get_log_base(int vid, uint64_t *log_base,
1530                 uint64_t *log_size)
1531 {
1532         struct virtio_net *dev = get_device(vid);
1533
1534         if (dev == NULL || log_base == NULL || log_size == NULL)
1535                 return -1;
1536
1537         *log_base = dev->log_base;
1538         *log_size = dev->log_size;
1539
1540         return 0;
1541 }
1542
1543 int
1544 rte_vhost_get_vring_base(int vid, uint16_t queue_id,
1545                 uint16_t *last_avail_idx, uint16_t *last_used_idx)
1546 {
1547         struct vhost_virtqueue *vq;
1548         struct virtio_net *dev = get_device(vid);
1549
1550         if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1551                 return -1;
1552
1553         if (queue_id >= VHOST_MAX_VRING)
1554                 return -1;
1555
1556         vq = dev->virtqueue[queue_id];
1557         if (!vq)
1558                 return -1;
1559
1560         if (vq_is_packed(dev)) {
1561                 *last_avail_idx = (vq->avail_wrap_counter << 15) |
1562                                   vq->last_avail_idx;
1563                 *last_used_idx = (vq->used_wrap_counter << 15) |
1564                                  vq->last_used_idx;
1565         } else {
1566                 *last_avail_idx = vq->last_avail_idx;
1567                 *last_used_idx = vq->last_used_idx;
1568         }
1569
1570         return 0;
1571 }
1572
1573 int
1574 rte_vhost_set_vring_base(int vid, uint16_t queue_id,
1575                 uint16_t last_avail_idx, uint16_t last_used_idx)
1576 {
1577         struct vhost_virtqueue *vq;
1578         struct virtio_net *dev = get_device(vid);
1579
1580         if (!dev)
1581                 return -1;
1582
1583         if (queue_id >= VHOST_MAX_VRING)
1584                 return -1;
1585
1586         vq = dev->virtqueue[queue_id];
1587         if (!vq)
1588                 return -1;
1589
1590         if (vq_is_packed(dev)) {
1591                 vq->last_avail_idx = last_avail_idx & 0x7fff;
1592                 vq->avail_wrap_counter = !!(last_avail_idx & (1 << 15));
1593                 vq->last_used_idx = last_used_idx & 0x7fff;
1594                 vq->used_wrap_counter = !!(last_used_idx & (1 << 15));
1595         } else {
1596                 vq->last_avail_idx = last_avail_idx;
1597                 vq->last_used_idx = last_used_idx;
1598         }
1599
1600         return 0;
1601 }
1602
1603 int
1604 rte_vhost_get_vring_base_from_inflight(int vid,
1605                                        uint16_t queue_id,
1606                                        uint16_t *last_avail_idx,
1607                                        uint16_t *last_used_idx)
1608 {
1609         struct rte_vhost_inflight_info_packed *inflight_info;
1610         struct vhost_virtqueue *vq;
1611         struct virtio_net *dev = get_device(vid);
1612
1613         if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1614                 return -1;
1615
1616         if (queue_id >= VHOST_MAX_VRING)
1617                 return -1;
1618
1619         vq = dev->virtqueue[queue_id];
1620         if (!vq)
1621                 return -1;
1622
1623         if (!vq_is_packed(dev))
1624                 return -1;
1625
1626         inflight_info = vq->inflight_packed;
1627         if (!inflight_info)
1628                 return -1;
1629
1630         *last_avail_idx = (inflight_info->old_used_wrap_counter << 15) |
1631                           inflight_info->old_used_idx;
1632         *last_used_idx = *last_avail_idx;
1633
1634         return 0;
1635 }
1636
1637 int
1638 rte_vhost_extern_callback_register(int vid,
1639                 struct rte_vhost_user_extern_ops const * const ops, void *ctx)
1640 {
1641         struct virtio_net *dev = get_device(vid);
1642
1643         if (dev == NULL || ops == NULL)
1644                 return -1;
1645
1646         dev->extern_ops = *ops;
1647         dev->extern_data = ctx;
1648         return 0;
1649 }
1650
1651 static __rte_always_inline int
1652 async_channel_register(int vid, uint16_t queue_id)
1653 {
1654         struct virtio_net *dev = get_device(vid);
1655         struct vhost_virtqueue *vq = dev->virtqueue[queue_id];
1656         struct vhost_async *async;
1657         int node = vq->numa_node;
1658
1659         if (unlikely(vq->async)) {
1660                 VHOST_LOG_CONFIG(ERR,
1661                                 "(%s) async register failed: already registered (qid: %d)\n",
1662                                 dev->ifname, queue_id);
1663                 return -1;
1664         }
1665
1666         async = rte_zmalloc_socket(NULL, sizeof(struct vhost_async), 0, node);
1667         if (!async) {
1668                 VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async metadata (qid: %d)\n",
1669                                 dev->ifname, queue_id);
1670                 return -1;
1671         }
1672
1673         async->pkts_info = rte_malloc_socket(NULL, vq->size * sizeof(struct async_inflight_info),
1674                         RTE_CACHE_LINE_SIZE, node);
1675         if (!async->pkts_info) {
1676                 VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async_pkts_info (qid: %d)\n",
1677                                 dev->ifname, queue_id);
1678                 goto out_free_async;
1679         }
1680
1681         async->pkts_cmpl_flag = rte_zmalloc_socket(NULL, vq->size * sizeof(bool),
1682                         RTE_CACHE_LINE_SIZE, node);
1683         if (!async->pkts_cmpl_flag) {
1684                 VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async pkts_cmpl_flag (qid: %d)\n",
1685                                 dev->ifname, queue_id);
1686                 goto out_free_async;
1687         }
1688
1689         if (vq_is_packed(dev)) {
1690                 async->buffers_packed = rte_malloc_socket(NULL,
1691                                 vq->size * sizeof(struct vring_used_elem_packed),
1692                                 RTE_CACHE_LINE_SIZE, node);
1693                 if (!async->buffers_packed) {
1694                         VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async buffers (qid: %d)\n",
1695                                         dev->ifname, queue_id);
1696                         goto out_free_inflight;
1697                 }
1698         } else {
1699                 async->descs_split = rte_malloc_socket(NULL,
1700                                 vq->size * sizeof(struct vring_used_elem),
1701                                 RTE_CACHE_LINE_SIZE, node);
1702                 if (!async->descs_split) {
1703                         VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async descs (qid: %d)\n",
1704                                         dev->ifname, queue_id);
1705                         goto out_free_inflight;
1706                 }
1707         }
1708
1709         vq->async = async;
1710
1711         return 0;
1712 out_free_inflight:
1713         rte_free(async->pkts_info);
1714 out_free_async:
1715         rte_free(async);
1716
1717         return -1;
1718 }
1719
1720 int
1721 rte_vhost_async_channel_register(int vid, uint16_t queue_id)
1722 {
1723         struct vhost_virtqueue *vq;
1724         struct virtio_net *dev = get_device(vid);
1725         int ret;
1726
1727         if (dev == NULL)
1728                 return -1;
1729
1730         if (queue_id >= VHOST_MAX_VRING)
1731                 return -1;
1732
1733         vq = dev->virtqueue[queue_id];
1734
1735         if (unlikely(vq == NULL || !dev->async_copy))
1736                 return -1;
1737
1738         rte_spinlock_lock(&vq->access_lock);
1739         ret = async_channel_register(vid, queue_id);
1740         rte_spinlock_unlock(&vq->access_lock);
1741
1742         return ret;
1743 }
1744
1745 int
1746 rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id)
1747 {
1748         struct vhost_virtqueue *vq;
1749         struct virtio_net *dev = get_device(vid);
1750
1751         if (dev == NULL)
1752                 return -1;
1753
1754         if (queue_id >= VHOST_MAX_VRING)
1755                 return -1;
1756
1757         vq = dev->virtqueue[queue_id];
1758
1759         if (unlikely(vq == NULL || !dev->async_copy))
1760                 return -1;
1761
1762         return async_channel_register(vid, queue_id);
1763 }
1764
1765 int
1766 rte_vhost_async_channel_unregister(int vid, uint16_t queue_id)
1767 {
1768         struct vhost_virtqueue *vq;
1769         struct virtio_net *dev = get_device(vid);
1770         int ret = -1;
1771
1772         if (dev == NULL)
1773                 return ret;
1774
1775         if (queue_id >= VHOST_MAX_VRING)
1776                 return ret;
1777
1778         vq = dev->virtqueue[queue_id];
1779
1780         if (vq == NULL)
1781                 return ret;
1782
1783         if (!rte_spinlock_trylock(&vq->access_lock)) {
1784                 VHOST_LOG_CONFIG(ERR, "(%s) failed to unregister async channel, virtqueue busy.\n",
1785                                 dev->ifname);
1786                 return ret;
1787         }
1788
1789         if (!vq->async) {
1790                 ret = 0;
1791         } else if (vq->async->pkts_inflight_n) {
1792                 VHOST_LOG_CONFIG(ERR, "(%s) failed to unregister async channel.\n", dev->ifname);
1793                 VHOST_LOG_CONFIG(ERR, "(%s) inflight packets must be completed before unregistration.\n",
1794                         dev->ifname);
1795         } else {
1796                 vhost_free_async_mem(vq);
1797                 ret = 0;
1798         }
1799
1800         rte_spinlock_unlock(&vq->access_lock);
1801
1802         return ret;
1803 }
1804
1805 int
1806 rte_vhost_async_channel_unregister_thread_unsafe(int vid, uint16_t queue_id)
1807 {
1808         struct vhost_virtqueue *vq;
1809         struct virtio_net *dev = get_device(vid);
1810
1811         if (dev == NULL)
1812                 return -1;
1813
1814         if (queue_id >= VHOST_MAX_VRING)
1815                 return -1;
1816
1817         vq = dev->virtqueue[queue_id];
1818
1819         if (vq == NULL)
1820                 return -1;
1821
1822         if (!vq->async)
1823                 return 0;
1824
1825         if (vq->async->pkts_inflight_n) {
1826                 VHOST_LOG_CONFIG(ERR, "(%s) failed to unregister async channel.\n", dev->ifname);
1827                 VHOST_LOG_CONFIG(ERR, "(%s) inflight packets must be completed before unregistration.\n",
1828                         dev->ifname);
1829                 return -1;
1830         }
1831
1832         vhost_free_async_mem(vq);
1833
1834         return 0;
1835 }
1836
1837 int
1838 rte_vhost_async_dma_configure(int16_t dma_id, uint16_t vchan_id)
1839 {
1840         struct rte_dma_info info;
1841         void *pkts_cmpl_flag_addr;
1842         uint16_t max_desc;
1843
1844         if (!rte_dma_is_valid(dma_id)) {
1845                 VHOST_LOG_CONFIG(ERR, "DMA %d is not found.\n", dma_id);
1846                 return -1;
1847         }
1848
1849         rte_dma_info_get(dma_id, &info);
1850         if (vchan_id >= info.max_vchans) {
1851                 VHOST_LOG_CONFIG(ERR, "Invalid DMA %d vChannel %u.\n", dma_id, vchan_id);
1852                 return -1;
1853         }
1854
1855         if (!dma_copy_track[dma_id].vchans) {
1856                 struct async_dma_vchan_info *vchans;
1857
1858                 vchans = rte_zmalloc(NULL, sizeof(struct async_dma_vchan_info) * info.max_vchans,
1859                                 RTE_CACHE_LINE_SIZE);
1860                 if (vchans == NULL) {
1861                         VHOST_LOG_CONFIG(ERR, "Failed to allocate vchans for DMA %d vChannel %u.\n",
1862                                         dma_id, vchan_id);
1863                         return -1;
1864                 }
1865
1866                 dma_copy_track[dma_id].vchans = vchans;
1867         }
1868
1869         if (dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr) {
1870                 VHOST_LOG_CONFIG(INFO, "DMA %d vChannel %u already registered.\n", dma_id,
1871                                 vchan_id);
1872                 return 0;
1873         }
1874
1875         max_desc = info.max_desc;
1876         if (!rte_is_power_of_2(max_desc))
1877                 max_desc = rte_align32pow2(max_desc);
1878
1879         pkts_cmpl_flag_addr = rte_zmalloc(NULL, sizeof(bool *) * max_desc, RTE_CACHE_LINE_SIZE);
1880         if (!pkts_cmpl_flag_addr) {
1881                 VHOST_LOG_CONFIG(ERR, "Failed to allocate pkts_cmpl_flag_addr for DMA %d "
1882                                 "vChannel %u.\n", dma_id, vchan_id);
1883
1884                 if (dma_copy_track[dma_id].nr_vchans == 0) {
1885                         rte_free(dma_copy_track[dma_id].vchans);
1886                         dma_copy_track[dma_id].vchans = NULL;
1887                 }
1888                 return -1;
1889         }
1890
1891         dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr = pkts_cmpl_flag_addr;
1892         dma_copy_track[dma_id].vchans[vchan_id].ring_size = max_desc;
1893         dma_copy_track[dma_id].vchans[vchan_id].ring_mask = max_desc - 1;
1894         dma_copy_track[dma_id].nr_vchans++;
1895
1896         return 0;
1897 }
1898
1899 int
1900 rte_vhost_async_get_inflight(int vid, uint16_t queue_id)
1901 {
1902         struct vhost_virtqueue *vq;
1903         struct virtio_net *dev = get_device(vid);
1904         int ret = -1;
1905
1906         if (dev == NULL)
1907                 return ret;
1908
1909         if (queue_id >= VHOST_MAX_VRING)
1910                 return ret;
1911
1912         vq = dev->virtqueue[queue_id];
1913
1914         if (vq == NULL)
1915                 return ret;
1916
1917         if (!rte_spinlock_trylock(&vq->access_lock)) {
1918                 VHOST_LOG_CONFIG(DEBUG,
1919                         "(%s) failed to check in-flight packets. virtqueue busy.\n",
1920                         dev->ifname);
1921                 return ret;
1922         }
1923
1924         if (vq->async)
1925                 ret = vq->async->pkts_inflight_n;
1926
1927         rte_spinlock_unlock(&vq->access_lock);
1928
1929         return ret;
1930 }
1931
1932 int
1933 rte_vhost_async_get_inflight_thread_unsafe(int vid, uint16_t queue_id)
1934 {
1935         struct vhost_virtqueue *vq;
1936         struct virtio_net *dev = get_device(vid);
1937         int ret = -1;
1938
1939         if (dev == NULL)
1940                 return ret;
1941
1942         if (queue_id >= VHOST_MAX_VRING)
1943                 return ret;
1944
1945         vq = dev->virtqueue[queue_id];
1946
1947         if (vq == NULL)
1948                 return ret;
1949
1950         if (!vq->async)
1951                 return ret;
1952
1953         ret = vq->async->pkts_inflight_n;
1954
1955         return ret;
1956 }
1957
1958 int
1959 rte_vhost_get_monitor_addr(int vid, uint16_t queue_id,
1960                 struct rte_vhost_power_monitor_cond *pmc)
1961 {
1962         struct virtio_net *dev = get_device(vid);
1963         struct vhost_virtqueue *vq;
1964
1965         if (dev == NULL)
1966                 return -1;
1967         if (queue_id >= VHOST_MAX_VRING)
1968                 return -1;
1969
1970         vq = dev->virtqueue[queue_id];
1971         if (vq == NULL)
1972                 return -1;
1973
1974         if (vq_is_packed(dev)) {
1975                 struct vring_packed_desc *desc;
1976                 desc = vq->desc_packed;
1977                 pmc->addr = &desc[vq->last_avail_idx].flags;
1978                 if (vq->avail_wrap_counter)
1979                         pmc->val = VRING_DESC_F_AVAIL;
1980                 else
1981                         pmc->val = VRING_DESC_F_USED;
1982                 pmc->mask = VRING_DESC_F_AVAIL | VRING_DESC_F_USED;
1983                 pmc->size = sizeof(desc[vq->last_avail_idx].flags);
1984                 pmc->match = 1;
1985         } else {
1986                 pmc->addr = &vq->avail->idx;
1987                 pmc->val = vq->last_avail_idx & (vq->size - 1);
1988                 pmc->mask = vq->size - 1;
1989                 pmc->size = sizeof(vq->avail->idx);
1990                 pmc->match = 0;
1991         }
1992
1993         return 0;
1994 }
1995
1996
1997 int
1998 rte_vhost_vring_stats_get_names(int vid, uint16_t queue_id,
1999                 struct rte_vhost_stat_name *name, unsigned int size)
2000 {
2001         struct virtio_net *dev = get_device(vid);
2002         unsigned int i;
2003
2004         if (dev == NULL)
2005                 return -1;
2006
2007         if (queue_id >= dev->nr_vring)
2008                 return -1;
2009
2010         if (!(dev->flags & VIRTIO_DEV_STATS_ENABLED))
2011                 return -1;
2012
2013         if (name == NULL || size < VHOST_NB_VQ_STATS)
2014                 return VHOST_NB_VQ_STATS;
2015
2016         for (i = 0; i < VHOST_NB_VQ_STATS; i++)
2017                 snprintf(name[i].name, sizeof(name[i].name), "%s_q%u_%s",
2018                                 (queue_id & 1) ? "rx" : "tx",
2019                                 queue_id / 2, vhost_vq_stat_strings[i].name);
2020
2021         return VHOST_NB_VQ_STATS;
2022 }
2023
2024 int
2025 rte_vhost_vring_stats_get(int vid, uint16_t queue_id,
2026                 struct rte_vhost_stat *stats, unsigned int n)
2027 {
2028         struct virtio_net *dev = get_device(vid);
2029         struct vhost_virtqueue *vq;
2030         unsigned int i;
2031
2032         if (dev == NULL)
2033                 return -1;
2034
2035         if (queue_id >= dev->nr_vring)
2036                 return -1;
2037
2038         if (!(dev->flags & VIRTIO_DEV_STATS_ENABLED))
2039                 return -1;
2040
2041         if (stats == NULL || n < VHOST_NB_VQ_STATS)
2042                 return VHOST_NB_VQ_STATS;
2043
2044         vq = dev->virtqueue[queue_id];
2045
2046         rte_spinlock_lock(&vq->access_lock);
2047         for (i = 0; i < VHOST_NB_VQ_STATS; i++) {
2048                 stats[i].value =
2049                         *(uint64_t *)(((char *)vq) + vhost_vq_stat_strings[i].offset);
2050                 stats[i].id = i;
2051         }
2052         rte_spinlock_unlock(&vq->access_lock);
2053
2054         return VHOST_NB_VQ_STATS;
2055 }
2056
2057 int rte_vhost_vring_stats_reset(int vid, uint16_t queue_id)
2058 {
2059         struct virtio_net *dev = get_device(vid);
2060         struct vhost_virtqueue *vq;
2061
2062         if (dev == NULL)
2063                 return -1;
2064
2065         if (queue_id >= dev->nr_vring)
2066                 return -1;
2067
2068         if (!(dev->flags & VIRTIO_DEV_STATS_ENABLED))
2069                 return -1;
2070
2071         vq = dev->virtqueue[queue_id];
2072
2073         rte_spinlock_lock(&vq->access_lock);
2074         memset(&vq->stats, 0, sizeof(vq->stats));
2075         rte_spinlock_unlock(&vq->access_lock);
2076
2077         return 0;
2078 }
2079
2080 RTE_LOG_REGISTER_SUFFIX(vhost_config_log_level, config, INFO);
2081 RTE_LOG_REGISTER_SUFFIX(vhost_data_log_level, data, WARNING);