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 <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[RTE_MAX_VHOST_DEVICE];
  29 pthread_mutex_t vhost_dev_lock = PTHREAD_MUTEX_INITIALIZER;
  30
  31 /* Called with iotlb_lock read-locked */
  32 uint64_t
  33 __vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
  34                     uint64_t iova, uint64_t *size, uint8_t perm)
  35 {
  36         uint64_t vva, tmp_size;
  37
  38         if (unlikely(!*size))
  39                 return 0;
  40
  41         tmp_size = *size;
  42
  43         vva = vhost_user_iotlb_cache_find(vq, iova, &tmp_size, perm);
  44         if (tmp_size == *size)
  45                 return vva;
  46
  47         iova += tmp_size;
  48
  49         if (!vhost_user_iotlb_pending_miss(vq, iova, perm)) {
  50                 /*
  51                  * iotlb_lock is read-locked for a full burst,
  52                  * but it only protects the iotlb cache.
  53                  * In case of IOTLB miss, we might block on the socket,
  54                  * which could cause a deadlock with QEMU if an IOTLB update
  55                  * is being handled. We can safely unlock here to avoid it.
  56                  */
  57                 vhost_user_iotlb_rd_unlock(vq);
  58
  59                 vhost_user_iotlb_pending_insert(dev, vq, iova, perm);
  60                 if (vhost_user_iotlb_miss(dev, iova, perm)) {
  61                         VHOST_LOG_DATA(ERR, "(%s) IOTLB miss req failed for IOVA 0x%" PRIx64 "\n",
  62                                 dev->ifname, 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_atomic_thread_fence(__ATOMIC_RELEASE);
 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                 VHOST_LOG_DATA(ERR,
 128                         "(%s) failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
 129                         dev->ifname, 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         /* No cache, nothing to sync */
 148         if (unlikely(!vq->log_cache))
 149                 return;
 150
 151         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 152
 153         log_base = (unsigned long *)(uintptr_t)dev->log_base;
 154
 155         for (i = 0; i < vq->log_cache_nb_elem; i++) {
 156                 struct log_cache_entry *elem = vq->log_cache + i;
 157
 158 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)
 159                 /*
 160                  * '__sync' builtins are deprecated, but '__atomic' ones
 161                  * are sub-optimized in older GCC versions.
 162                  */
 163                 __sync_fetch_and_or(log_base + elem->offset, elem->val);
 164 #else
 165                 __atomic_fetch_or(log_base + elem->offset, elem->val,
 166                                 __ATOMIC_RELAXED);
 167 #endif
 168         }
 169
 170         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 171
 172         vq->log_cache_nb_elem = 0;
 173 }
 174
 175 static __rte_always_inline void
 176 vhost_log_cache_page(struct virtio_net *dev, struct vhost_virtqueue *vq,
 177                         uint64_t page)
 178 {
 179         uint32_t bit_nr = page % (sizeof(unsigned long) << 3);
 180         uint32_t offset = page / (sizeof(unsigned long) << 3);
 181         int i;
 182
 183         if (unlikely(!vq->log_cache)) {
 184                 /* No logging cache allocated, write dirty log map directly */
 185                 rte_atomic_thread_fence(__ATOMIC_RELEASE);
 186                 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
 187
 188                 return;
 189         }
 190
 191         for (i = 0; i < vq->log_cache_nb_elem; i++) {
 192                 struct log_cache_entry *elem = vq->log_cache + i;
 193
 194                 if (elem->offset == offset) {
 195                         elem->val |= (1UL << bit_nr);
 196                         return;
 197                 }
 198         }
 199
 200         if (unlikely(i >= VHOST_LOG_CACHE_NR)) {
 201                 /*
 202                  * No more room for a new log cache entry,
 203                  * so write the dirty log map directly.
 204                  */
 205                 rte_atomic_thread_fence(__ATOMIC_RELEASE);
 206                 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
 207
 208                 return;
 209         }
 210
 211         vq->log_cache[i].offset = offset;
 212         vq->log_cache[i].val = (1UL << bit_nr);
 213         vq->log_cache_nb_elem++;
 214 }
 215
 216 void
 217 __vhost_log_cache_write(struct virtio_net *dev, struct vhost_virtqueue *vq,
 218                         uint64_t addr, uint64_t len)
 219 {
 220         uint64_t page;
 221
 222         if (unlikely(!dev->log_base || !len))
 223                 return;
 224
 225         if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
 226                 return;
 227
 228         page = addr / VHOST_LOG_PAGE;
 229         while (page * VHOST_LOG_PAGE < addr + len) {
 230                 vhost_log_cache_page(dev, vq, page);
 231                 page += 1;
 232         }
 233 }
 234
 235 void
 236 __vhost_log_cache_write_iova(struct virtio_net *dev, struct vhost_virtqueue *vq,
 237                              uint64_t iova, uint64_t len)
 238 {
 239         uint64_t hva, gpa, map_len;
 240         map_len = len;
 241
 242         hva = __vhost_iova_to_vva(dev, vq, iova, &map_len, VHOST_ACCESS_RW);
 243         if (map_len != len) {
 244                 VHOST_LOG_DATA(ERR,
 245                         "(%s) failed to write log for IOVA 0x%" PRIx64 ". No IOTLB entry found\n",
 246                         dev->ifname, iova);
 247                 return;
 248         }
 249
 250         gpa = hva_to_gpa(dev, hva, len);
 251         if (gpa)
 252                 __vhost_log_cache_write(dev, vq, gpa, len);
 253 }
 254
 255 void *
 256 vhost_alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
 257                 uint64_t desc_addr, uint64_t desc_len)
 258 {
 259         void *idesc;
 260         uint64_t src, dst;
 261         uint64_t len, remain = desc_len;
 262
 263         idesc = rte_malloc_socket(__func__, desc_len, 0, vq->numa_node);
 264         if (unlikely(!idesc))
 265                 return NULL;
 266
 267         dst = (uint64_t)(uintptr_t)idesc;
 268
 269         while (remain) {
 270                 len = remain;
 271                 src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
 272                                 VHOST_ACCESS_RO);
 273                 if (unlikely(!src || !len)) {
 274                         rte_free(idesc);
 275                         return NULL;
 276                 }
 277
 278                 rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
 279
 280                 remain -= len;
 281                 dst += len;
 282                 desc_addr += len;
 283         }
 284
 285         return idesc;
 286 }
 287
 288 void
 289 cleanup_vq(struct vhost_virtqueue *vq, int destroy)
 290 {
 291         if ((vq->callfd >= 0) && (destroy != 0))
 292                 close(vq->callfd);
 293         if (vq->kickfd >= 0)
 294                 close(vq->kickfd);
 295 }
 296
 297 void
 298 cleanup_vq_inflight(struct virtio_net *dev, struct vhost_virtqueue *vq)
 299 {
 300         if (!(dev->protocol_features &
 301             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
 302                 return;
 303
 304         if (vq_is_packed(dev)) {
 305                 if (vq->inflight_packed)
 306                         vq->inflight_packed = NULL;
 307         } else {
 308                 if (vq->inflight_split)
 309                         vq->inflight_split = NULL;
 310         }
 311
 312         if (vq->resubmit_inflight) {
 313                 if (vq->resubmit_inflight->resubmit_list) {
 314                         rte_free(vq->resubmit_inflight->resubmit_list);
 315                         vq->resubmit_inflight->resubmit_list = NULL;
 316                 }
 317                 rte_free(vq->resubmit_inflight);
 318                 vq->resubmit_inflight = NULL;
 319         }
 320 }
 321
 322 /*
 323  * Unmap any memory, close any file descriptors and
 324  * free any memory owned by a device.
 325  */
 326 void
 327 cleanup_device(struct virtio_net *dev, int destroy)
 328 {
 329         uint32_t i;
 330
 331         vhost_backend_cleanup(dev);
 332
 333         for (i = 0; i < dev->nr_vring; i++) {
 334                 cleanup_vq(dev->virtqueue[i], destroy);
 335                 cleanup_vq_inflight(dev, dev->virtqueue[i]);
 336         }
 337 }
 338
 339 static void
 340 vhost_free_async_mem(struct vhost_virtqueue *vq)
 341 {
 342         if (!vq->async)
 343                 return;
 344
 345         rte_free(vq->async->pkts_info);
 346         rte_free(vq->async->pkts_cmpl_flag);
 347
 348         rte_free(vq->async->buffers_packed);
 349         vq->async->buffers_packed = NULL;
 350         rte_free(vq->async->descs_split);
 351         vq->async->descs_split = NULL;
 352
 353         rte_free(vq->async);
 354         vq->async = NULL;
 355 }
 356
 357 void
 358 free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq)
 359 {
 360         if (vq_is_packed(dev))
 361                 rte_free(vq->shadow_used_packed);
 362         else
 363                 rte_free(vq->shadow_used_split);
 364
 365         vhost_free_async_mem(vq);
 366         rte_free(vq->batch_copy_elems);
 367         rte_mempool_free(vq->iotlb_pool);
 368         rte_free(vq->log_cache);
 369         rte_free(vq);
 370 }
 371
 372 /*
 373  * Release virtqueues and device memory.
 374  */
 375 static void
 376 free_device(struct virtio_net *dev)
 377 {
 378         uint32_t i;
 379
 380         for (i = 0; i < dev->nr_vring; i++)
 381                 free_vq(dev, dev->virtqueue[i]);
 382
 383         rte_free(dev);
 384 }
 385
 386 static __rte_always_inline int
 387 log_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
 388 {
 389         if (likely(!(vq->ring_addrs.flags & (1 << VHOST_VRING_F_LOG))))
 390                 return 0;
 391
 392         vq->log_guest_addr = translate_log_addr(dev, vq,
 393                                                 vq->ring_addrs.log_guest_addr);
 394         if (vq->log_guest_addr == 0)
 395                 return -1;
 396
 397         return 0;
 398 }
 399
 400 /*
 401  * Converts vring log address to GPA
 402  * If IOMMU is enabled, the log address is IOVA
 403  * If IOMMU not enabled, the log address is already GPA
 404  *
 405  * Caller should have iotlb_lock read-locked
 406  */
 407 uint64_t
 408 translate_log_addr(struct virtio_net *dev, struct vhost_virtqueue *vq,
 409                 uint64_t log_addr)
 410 {
 411         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 412                 const uint64_t exp_size = sizeof(uint64_t);
 413                 uint64_t hva, gpa;
 414                 uint64_t size = exp_size;
 415
 416                 hva = vhost_iova_to_vva(dev, vq, log_addr,
 417                                         &size, VHOST_ACCESS_RW);
 418
 419                 if (size != exp_size)
 420                         return 0;
 421
 422                 gpa = hva_to_gpa(dev, hva, exp_size);
 423                 if (!gpa) {
 424                         VHOST_LOG_DATA(ERR,
 425                                 "(%s) failed to find GPA for log_addr: 0x%"
 426                                 PRIx64 " hva: 0x%" PRIx64 "\n",
 427                                 dev->ifname, log_addr, hva);
 428                         return 0;
 429                 }
 430                 return gpa;
 431
 432         } else
 433                 return log_addr;
 434 }
 435
 436 /* Caller should have iotlb_lock read-locked */
 437 static int
 438 vring_translate_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
 439 {
 440         uint64_t req_size, size;
 441
 442         req_size = sizeof(struct vring_desc) * vq->size;
 443         size = req_size;
 444         vq->desc = (struct vring_desc *)(uintptr_t)vhost_iova_to_vva(dev, vq,
 445                                                 vq->ring_addrs.desc_user_addr,
 446                                                 &size, VHOST_ACCESS_RW);
 447         if (!vq->desc || size != req_size)
 448                 return -1;
 449
 450         req_size = sizeof(struct vring_avail);
 451         req_size += sizeof(uint16_t) * vq->size;
 452         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 453                 req_size += sizeof(uint16_t);
 454         size = req_size;
 455         vq->avail = (struct vring_avail *)(uintptr_t)vhost_iova_to_vva(dev, vq,
 456                                                 vq->ring_addrs.avail_user_addr,
 457                                                 &size, VHOST_ACCESS_RW);
 458         if (!vq->avail || size != req_size)
 459                 return -1;
 460
 461         req_size = sizeof(struct vring_used);
 462         req_size += sizeof(struct vring_used_elem) * vq->size;
 463         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))
 464                 req_size += sizeof(uint16_t);
 465         size = req_size;
 466         vq->used = (struct vring_used *)(uintptr_t)vhost_iova_to_vva(dev, vq,
 467                                                 vq->ring_addrs.used_user_addr,
 468                                                 &size, VHOST_ACCESS_RW);
 469         if (!vq->used || size != req_size)
 470                 return -1;
 471
 472         return 0;
 473 }
 474
 475 /* Caller should have iotlb_lock read-locked */
 476 static int
 477 vring_translate_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
 478 {
 479         uint64_t req_size, size;
 480
 481         req_size = sizeof(struct vring_packed_desc) * vq->size;
 482         size = req_size;
 483         vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 484                 vhost_iova_to_vva(dev, vq, vq->ring_addrs.desc_user_addr,
 485                                 &size, VHOST_ACCESS_RW);
 486         if (!vq->desc_packed || size != req_size)
 487                 return -1;
 488
 489         req_size = sizeof(struct vring_packed_desc_event);
 490         size = req_size;
 491         vq->driver_event = (struct vring_packed_desc_event *)(uintptr_t)
 492                 vhost_iova_to_vva(dev, vq, vq->ring_addrs.avail_user_addr,
 493                                 &size, VHOST_ACCESS_RW);
 494         if (!vq->driver_event || size != req_size)
 495                 return -1;
 496
 497         req_size = sizeof(struct vring_packed_desc_event);
 498         size = req_size;
 499         vq->device_event = (struct vring_packed_desc_event *)(uintptr_t)
 500                 vhost_iova_to_vva(dev, vq, vq->ring_addrs.used_user_addr,
 501                                 &size, VHOST_ACCESS_RW);
 502         if (!vq->device_event || size != req_size)
 503                 return -1;
 504
 505         return 0;
 506 }
 507
 508 int
 509 vring_translate(struct virtio_net *dev, struct vhost_virtqueue *vq)
 510 {
 511
 512         if (!(dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
 513                 return -1;
 514
 515         if (vq_is_packed(dev)) {
 516                 if (vring_translate_packed(dev, vq) < 0)
 517                         return -1;
 518         } else {
 519                 if (vring_translate_split(dev, vq) < 0)
 520                         return -1;
 521         }
 522
 523         if (log_translate(dev, vq) < 0)
 524                 return -1;
 525
 526         vq->access_ok = true;
 527
 528         return 0;
 529 }
 530
 531 void
 532 vring_invalidate(struct virtio_net *dev, struct vhost_virtqueue *vq)
 533 {
 534         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 535                 vhost_user_iotlb_wr_lock(vq);
 536
 537         vq->access_ok = false;
 538         vq->desc = NULL;
 539         vq->avail = NULL;
 540         vq->used = NULL;
 541         vq->log_guest_addr = 0;
 542
 543         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 544                 vhost_user_iotlb_wr_unlock(vq);
 545 }
 546
 547 static void
 548 init_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
 549 {
 550         struct vhost_virtqueue *vq;
 551         int numa_node = SOCKET_ID_ANY;
 552
 553         if (vring_idx >= VHOST_MAX_VRING) {
 554                 VHOST_LOG_CONFIG(ERR, "(%s) failed to init vring, out of bound (%d)\n",
 555                                 dev->ifname, vring_idx);
 556                 return;
 557         }
 558
 559         vq = dev->virtqueue[vring_idx];
 560         if (!vq) {
 561                 VHOST_LOG_CONFIG(ERR, "(%s) virtqueue not allocated (%d)\n",
 562                                 dev->ifname, vring_idx);
 563                 return;
 564         }
 565
 566         memset(vq, 0, sizeof(struct vhost_virtqueue));
 567
 568         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
 569         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
 570         vq->notif_enable = VIRTIO_UNINITIALIZED_NOTIF;
 571
 572 #ifdef RTE_LIBRTE_VHOST_NUMA
 573         if (get_mempolicy(&numa_node, NULL, 0, vq, MPOL_F_NODE | MPOL_F_ADDR)) {
 574                 VHOST_LOG_CONFIG(ERR, "(%s) failed to query numa node: %s\n",
 575                         dev->ifname, rte_strerror(errno));
 576                 numa_node = SOCKET_ID_ANY;
 577         }
 578 #endif
 579         vq->numa_node = numa_node;
 580
 581         vhost_user_iotlb_init(dev, vring_idx);
 582 }
 583
 584 static void
 585 reset_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
 586 {
 587         struct vhost_virtqueue *vq;
 588         int callfd;
 589
 590         if (vring_idx >= VHOST_MAX_VRING) {
 591                 VHOST_LOG_CONFIG(ERR,
 592                                 "(%s) failed to reset vring, out of bound (%d)\n",
 593                                 dev->ifname, vring_idx);
 594                 return;
 595         }
 596
 597         vq = dev->virtqueue[vring_idx];
 598         if (!vq) {
 599                 VHOST_LOG_CONFIG(ERR, "(%s) failed to reset vring, virtqueue not allocated (%d)\n",
 600                                 dev->ifname, vring_idx);
 601                 return;
 602         }
 603
 604         callfd = vq->callfd;
 605         init_vring_queue(dev, vring_idx);
 606         vq->callfd = callfd;
 607 }
 608
 609 int
 610 alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx)
 611 {
 612         struct vhost_virtqueue *vq;
 613         uint32_t i;
 614
 615         /* Also allocate holes, if any, up to requested vring index. */
 616         for (i = 0; i <= vring_idx; i++) {
 617                 if (dev->virtqueue[i])
 618                         continue;
 619
 620                 vq = rte_zmalloc(NULL, sizeof(struct vhost_virtqueue), 0);
 621                 if (vq == NULL) {
 622                         VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate memory for vring %u.\n",
 623                                         dev->ifname, i);
 624                         return -1;
 625                 }
 626
 627                 dev->virtqueue[i] = vq;
 628                 init_vring_queue(dev, i);
 629                 rte_spinlock_init(&vq->access_lock);
 630                 vq->avail_wrap_counter = 1;
 631                 vq->used_wrap_counter = 1;
 632                 vq->signalled_used_valid = false;
 633         }
 634
 635         dev->nr_vring = RTE_MAX(dev->nr_vring, vring_idx + 1);
 636
 637         return 0;
 638 }
 639
 640 /*
 641  * Reset some variables in device structure, while keeping few
 642  * others untouched, such as vid, ifname, nr_vring: they
 643  * should be same unless the device is removed.
 644  */
 645 void
 646 reset_device(struct virtio_net *dev)
 647 {
 648         uint32_t i;
 649
 650         dev->features = 0;
 651         dev->protocol_features = 0;
 652         dev->flags &= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
 653
 654         for (i = 0; i < dev->nr_vring; i++)
 655                 reset_vring_queue(dev, i);
 656 }
 657
 658 /*
 659  * Invoked when there is a new vhost-user connection established (when
 660  * there is a new virtio device being attached).
 661  */
 662 int
 663 vhost_new_device(void)
 664 {
 665         struct virtio_net *dev;
 666         int i;
 667
 668         pthread_mutex_lock(&vhost_dev_lock);
 669         for (i = 0; i < RTE_MAX_VHOST_DEVICE; i++) {
 670                 if (vhost_devices[i] == NULL)
 671                         break;
 672         }
 673
 674         if (i == RTE_MAX_VHOST_DEVICE) {
 675                 VHOST_LOG_CONFIG(ERR, "failed to find a free slot for new device.\n");
 676                 pthread_mutex_unlock(&vhost_dev_lock);
 677                 return -1;
 678         }
 679
 680         dev = rte_zmalloc(NULL, sizeof(struct virtio_net), 0);
 681         if (dev == NULL) {
 682                 VHOST_LOG_CONFIG(ERR, "failed to allocate memory for new device.\n");
 683                 pthread_mutex_unlock(&vhost_dev_lock);
 684                 return -1;
 685         }
 686
 687         vhost_devices[i] = dev;
 688         pthread_mutex_unlock(&vhost_dev_lock);
 689
 690         dev->vid = i;
 691         dev->flags = VIRTIO_DEV_BUILTIN_VIRTIO_NET;
 692         dev->slave_req_fd = -1;
 693         dev->postcopy_ufd = -1;
 694         rte_spinlock_init(&dev->slave_req_lock);
 695
 696         return i;
 697 }
 698
 699 void
 700 vhost_destroy_device_notify(struct virtio_net *dev)
 701 {
 702         struct rte_vdpa_device *vdpa_dev;
 703
 704         if (dev->flags & VIRTIO_DEV_RUNNING) {
 705                 vdpa_dev = dev->vdpa_dev;
 706                 if (vdpa_dev)
 707                         vdpa_dev->ops->dev_close(dev->vid);
 708                 dev->flags &= ~VIRTIO_DEV_RUNNING;
 709                 dev->notify_ops->destroy_device(dev->vid);
 710         }
 711 }
 712
 713 /*
 714  * Invoked when there is the vhost-user connection is broken (when
 715  * the virtio device is being detached).
 716  */
 717 void
 718 vhost_destroy_device(int vid)
 719 {
 720         struct virtio_net *dev = get_device(vid);
 721
 722         if (dev == NULL)
 723                 return;
 724
 725         vhost_destroy_device_notify(dev);
 726
 727         cleanup_device(dev, 1);
 728         free_device(dev);
 729
 730         vhost_devices[vid] = NULL;
 731 }
 732
 733 void
 734 vhost_attach_vdpa_device(int vid, struct rte_vdpa_device *vdpa_dev)
 735 {
 736         struct virtio_net *dev = get_device(vid);
 737
 738         if (dev == NULL)
 739                 return;
 740
 741         dev->vdpa_dev = vdpa_dev;
 742 }
 743
 744 void
 745 vhost_set_ifname(int vid, const char *if_name, unsigned int if_len)
 746 {
 747         struct virtio_net *dev;
 748         unsigned int len;
 749
 750         dev = get_device(vid);
 751         if (dev == NULL)
 752                 return;
 753
 754         len = if_len > sizeof(dev->ifname) ?
 755                 sizeof(dev->ifname) : if_len;
 756
 757         strncpy(dev->ifname, if_name, len);
 758         dev->ifname[sizeof(dev->ifname) - 1] = '\0';
 759 }
 760
 761 void
 762 vhost_setup_virtio_net(int vid, bool enable, bool compliant_ol_flags)
 763 {
 764         struct virtio_net *dev = get_device(vid);
 765
 766         if (dev == NULL)
 767                 return;
 768
 769         if (enable)
 770                 dev->flags |= VIRTIO_DEV_BUILTIN_VIRTIO_NET;
 771         else
 772                 dev->flags &= ~VIRTIO_DEV_BUILTIN_VIRTIO_NET;
 773         if (!compliant_ol_flags)
 774                 dev->flags |= VIRTIO_DEV_LEGACY_OL_FLAGS;
 775         else
 776                 dev->flags &= ~VIRTIO_DEV_LEGACY_OL_FLAGS;
 777 }
 778
 779 void
 780 vhost_enable_extbuf(int vid)
 781 {
 782         struct virtio_net *dev = get_device(vid);
 783
 784         if (dev == NULL)
 785                 return;
 786
 787         dev->extbuf = 1;
 788 }
 789
 790 void
 791 vhost_enable_linearbuf(int vid)
 792 {
 793         struct virtio_net *dev = get_device(vid);
 794
 795         if (dev == NULL)
 796                 return;
 797
 798         dev->linearbuf = 1;
 799 }
 800
 801 int
 802 rte_vhost_get_mtu(int vid, uint16_t *mtu)
 803 {
 804         struct virtio_net *dev = get_device(vid);
 805
 806         if (dev == NULL || mtu == NULL)
 807                 return -ENODEV;
 808
 809         if (!(dev->flags & VIRTIO_DEV_READY))
 810                 return -EAGAIN;
 811
 812         if (!(dev->features & (1ULL << VIRTIO_NET_F_MTU)))
 813                 return -ENOTSUP;
 814
 815         *mtu = dev->mtu;
 816
 817         return 0;
 818 }
 819
 820 int
 821 rte_vhost_get_numa_node(int vid)
 822 {
 823 #ifdef RTE_LIBRTE_VHOST_NUMA
 824         struct virtio_net *dev = get_device(vid);
 825         int numa_node;
 826         int ret;
 827
 828         if (dev == NULL || numa_available() != 0)
 829                 return -1;
 830
 831         ret = get_mempolicy(&numa_node, NULL, 0, dev,
 832                             MPOL_F_NODE | MPOL_F_ADDR);
 833         if (ret < 0) {
 834                 VHOST_LOG_CONFIG(ERR, "(%s) failed to query numa node: %s\n",
 835                         dev->ifname, rte_strerror(errno));
 836                 return -1;
 837         }
 838
 839         return numa_node;
 840 #else
 841         RTE_SET_USED(vid);
 842         return -1;
 843 #endif
 844 }
 845
 846 uint32_t
 847 rte_vhost_get_queue_num(int vid)
 848 {
 849         struct virtio_net *dev = get_device(vid);
 850
 851         if (dev == NULL)
 852                 return 0;
 853
 854         return dev->nr_vring / 2;
 855 }
 856
 857 uint16_t
 858 rte_vhost_get_vring_num(int vid)
 859 {
 860         struct virtio_net *dev = get_device(vid);
 861
 862         if (dev == NULL)
 863                 return 0;
 864
 865         return dev->nr_vring;
 866 }
 867
 868 int
 869 rte_vhost_get_ifname(int vid, char *buf, size_t len)
 870 {
 871         struct virtio_net *dev = get_device(vid);
 872
 873         if (dev == NULL || buf == NULL)
 874                 return -1;
 875
 876         len = RTE_MIN(len, sizeof(dev->ifname));
 877
 878         strncpy(buf, dev->ifname, len);
 879         buf[len - 1] = '\0';
 880
 881         return 0;
 882 }
 883
 884 int
 885 rte_vhost_get_negotiated_features(int vid, uint64_t *features)
 886 {
 887         struct virtio_net *dev;
 888
 889         dev = get_device(vid);
 890         if (dev == NULL || features == NULL)
 891                 return -1;
 892
 893         *features = dev->features;
 894         return 0;
 895 }
 896
 897 int
 898 rte_vhost_get_negotiated_protocol_features(int vid,
 899                                            uint64_t *protocol_features)
 900 {
 901         struct virtio_net *dev;
 902
 903         dev = get_device(vid);
 904         if (dev == NULL || protocol_features == NULL)
 905                 return -1;
 906
 907         *protocol_features = dev->protocol_features;
 908         return 0;
 909 }
 910
 911 int
 912 rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
 913 {
 914         struct virtio_net *dev;
 915         struct rte_vhost_memory *m;
 916         size_t size;
 917
 918         dev = get_device(vid);
 919         if (dev == NULL || mem == NULL)
 920                 return -1;
 921
 922         size = dev->mem->nregions * sizeof(struct rte_vhost_mem_region);
 923         m = malloc(sizeof(struct rte_vhost_memory) + size);
 924         if (!m)
 925                 return -1;
 926
 927         m->nregions = dev->mem->nregions;
 928         memcpy(m->regions, dev->mem->regions, size);
 929         *mem = m;
 930
 931         return 0;
 932 }
 933
 934 int
 935 rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
 936                           struct rte_vhost_vring *vring)
 937 {
 938         struct virtio_net *dev;
 939         struct vhost_virtqueue *vq;
 940
 941         dev = get_device(vid);
 942         if (dev == NULL || vring == NULL)
 943                 return -1;
 944
 945         if (vring_idx >= VHOST_MAX_VRING)
 946                 return -1;
 947
 948         vq = dev->virtqueue[vring_idx];
 949         if (!vq)
 950                 return -1;
 951
 952         if (vq_is_packed(dev)) {
 953                 vring->desc_packed = vq->desc_packed;
 954                 vring->driver_event = vq->driver_event;
 955                 vring->device_event = vq->device_event;
 956         } else {
 957                 vring->desc = vq->desc;
 958                 vring->avail = vq->avail;
 959                 vring->used = vq->used;
 960         }
 961         vring->log_guest_addr  = vq->log_guest_addr;
 962
 963         vring->callfd  = vq->callfd;
 964         vring->kickfd  = vq->kickfd;
 965         vring->size    = vq->size;
 966
 967         return 0;
 968 }
 969
 970 int
 971 rte_vhost_get_vhost_ring_inflight(int vid, uint16_t vring_idx,
 972                                   struct rte_vhost_ring_inflight *vring)
 973 {
 974         struct virtio_net *dev;
 975         struct vhost_virtqueue *vq;
 976
 977         dev = get_device(vid);
 978         if (unlikely(!dev))
 979                 return -1;
 980
 981         if (vring_idx >= VHOST_MAX_VRING)
 982                 return -1;
 983
 984         vq = dev->virtqueue[vring_idx];
 985         if (unlikely(!vq))
 986                 return -1;
 987
 988         if (vq_is_packed(dev)) {
 989                 if (unlikely(!vq->inflight_packed))
 990                         return -1;
 991
 992                 vring->inflight_packed = vq->inflight_packed;
 993         } else {
 994                 if (unlikely(!vq->inflight_split))
 995                         return -1;
 996
 997                 vring->inflight_split = vq->inflight_split;
 998         }
 999
1000         vring->resubmit_inflight = vq->resubmit_inflight;
1001
1002         return 0;
1003 }
1004
1005 int
1006 rte_vhost_set_inflight_desc_split(int vid, uint16_t vring_idx,
1007                                   uint16_t idx)
1008 {
1009         struct vhost_virtqueue *vq;
1010         struct virtio_net *dev;
1011
1012         dev = get_device(vid);
1013         if (unlikely(!dev))
1014                 return -1;
1015
1016         if (unlikely(!(dev->protocol_features &
1017             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1018                 return 0;
1019
1020         if (unlikely(vq_is_packed(dev)))
1021                 return -1;
1022
1023         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1024                 return -1;
1025
1026         vq = dev->virtqueue[vring_idx];
1027         if (unlikely(!vq))
1028                 return -1;
1029
1030         if (unlikely(!vq->inflight_split))
1031                 return -1;
1032
1033         if (unlikely(idx >= vq->size))
1034                 return -1;
1035
1036         vq->inflight_split->desc[idx].counter = vq->global_counter++;
1037         vq->inflight_split->desc[idx].inflight = 1;
1038         return 0;
1039 }
1040
1041 int
1042 rte_vhost_set_inflight_desc_packed(int vid, uint16_t vring_idx,
1043                                    uint16_t head, uint16_t last,
1044                                    uint16_t *inflight_entry)
1045 {
1046         struct rte_vhost_inflight_info_packed *inflight_info;
1047         struct virtio_net *dev;
1048         struct vhost_virtqueue *vq;
1049         struct vring_packed_desc *desc;
1050         uint16_t old_free_head, free_head;
1051
1052         dev = get_device(vid);
1053         if (unlikely(!dev))
1054                 return -1;
1055
1056         if (unlikely(!(dev->protocol_features &
1057             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1058                 return 0;
1059
1060         if (unlikely(!vq_is_packed(dev)))
1061                 return -1;
1062
1063         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1064                 return -1;
1065
1066         vq = dev->virtqueue[vring_idx];
1067         if (unlikely(!vq))
1068                 return -1;
1069
1070         inflight_info = vq->inflight_packed;
1071         if (unlikely(!inflight_info))
1072                 return -1;
1073
1074         if (unlikely(head >= vq->size))
1075                 return -1;
1076
1077         desc = vq->desc_packed;
1078         old_free_head = inflight_info->old_free_head;
1079         if (unlikely(old_free_head >= vq->size))
1080                 return -1;
1081
1082         free_head = old_free_head;
1083
1084         /* init header descriptor */
1085         inflight_info->desc[old_free_head].num = 0;
1086         inflight_info->desc[old_free_head].counter = vq->global_counter++;
1087         inflight_info->desc[old_free_head].inflight = 1;
1088
1089         /* save desc entry in flight entry */
1090         while (head != ((last + 1) % vq->size)) {
1091                 inflight_info->desc[old_free_head].num++;
1092                 inflight_info->desc[free_head].addr = desc[head].addr;
1093                 inflight_info->desc[free_head].len = desc[head].len;
1094                 inflight_info->desc[free_head].flags = desc[head].flags;
1095                 inflight_info->desc[free_head].id = desc[head].id;
1096
1097                 inflight_info->desc[old_free_head].last = free_head;
1098                 free_head = inflight_info->desc[free_head].next;
1099                 inflight_info->free_head = free_head;
1100                 head = (head + 1) % vq->size;
1101         }
1102
1103         inflight_info->old_free_head = free_head;
1104         *inflight_entry = old_free_head;
1105
1106         return 0;
1107 }
1108
1109 int
1110 rte_vhost_clr_inflight_desc_split(int vid, uint16_t vring_idx,
1111                                   uint16_t last_used_idx, uint16_t idx)
1112 {
1113         struct virtio_net *dev;
1114         struct vhost_virtqueue *vq;
1115
1116         dev = get_device(vid);
1117         if (unlikely(!dev))
1118                 return -1;
1119
1120         if (unlikely(!(dev->protocol_features &
1121             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1122                 return 0;
1123
1124         if (unlikely(vq_is_packed(dev)))
1125                 return -1;
1126
1127         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1128                 return -1;
1129
1130         vq = dev->virtqueue[vring_idx];
1131         if (unlikely(!vq))
1132                 return -1;
1133
1134         if (unlikely(!vq->inflight_split))
1135                 return -1;
1136
1137         if (unlikely(idx >= vq->size))
1138                 return -1;
1139
1140         rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1141
1142         vq->inflight_split->desc[idx].inflight = 0;
1143
1144         rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1145
1146         vq->inflight_split->used_idx = last_used_idx;
1147         return 0;
1148 }
1149
1150 int
1151 rte_vhost_clr_inflight_desc_packed(int vid, uint16_t vring_idx,
1152                                    uint16_t head)
1153 {
1154         struct rte_vhost_inflight_info_packed *inflight_info;
1155         struct virtio_net *dev;
1156         struct vhost_virtqueue *vq;
1157
1158         dev = get_device(vid);
1159         if (unlikely(!dev))
1160                 return -1;
1161
1162         if (unlikely(!(dev->protocol_features &
1163             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1164                 return 0;
1165
1166         if (unlikely(!vq_is_packed(dev)))
1167                 return -1;
1168
1169         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1170                 return -1;
1171
1172         vq = dev->virtqueue[vring_idx];
1173         if (unlikely(!vq))
1174                 return -1;
1175
1176         inflight_info = vq->inflight_packed;
1177         if (unlikely(!inflight_info))
1178                 return -1;
1179
1180         if (unlikely(head >= vq->size))
1181                 return -1;
1182
1183         rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1184
1185         inflight_info->desc[head].inflight = 0;
1186
1187         rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
1188
1189         inflight_info->old_free_head = inflight_info->free_head;
1190         inflight_info->old_used_idx = inflight_info->used_idx;
1191         inflight_info->old_used_wrap_counter = inflight_info->used_wrap_counter;
1192
1193         return 0;
1194 }
1195
1196 int
1197 rte_vhost_set_last_inflight_io_split(int vid, uint16_t vring_idx,
1198                                      uint16_t idx)
1199 {
1200         struct virtio_net *dev;
1201         struct vhost_virtqueue *vq;
1202
1203         dev = get_device(vid);
1204         if (unlikely(!dev))
1205                 return -1;
1206
1207         if (unlikely(!(dev->protocol_features &
1208             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1209                 return 0;
1210
1211         if (unlikely(vq_is_packed(dev)))
1212                 return -1;
1213
1214         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1215                 return -1;
1216
1217         vq = dev->virtqueue[vring_idx];
1218         if (unlikely(!vq))
1219                 return -1;
1220
1221         if (unlikely(!vq->inflight_split))
1222                 return -1;
1223
1224         if (unlikely(idx >= vq->size))
1225                 return -1;
1226
1227         vq->inflight_split->last_inflight_io = idx;
1228         return 0;
1229 }
1230
1231 int
1232 rte_vhost_set_last_inflight_io_packed(int vid, uint16_t vring_idx,
1233                                       uint16_t head)
1234 {
1235         struct rte_vhost_inflight_info_packed *inflight_info;
1236         struct virtio_net *dev;
1237         struct vhost_virtqueue *vq;
1238         uint16_t last;
1239
1240         dev = get_device(vid);
1241         if (unlikely(!dev))
1242                 return -1;
1243
1244         if (unlikely(!(dev->protocol_features &
1245             (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))))
1246                 return 0;
1247
1248         if (unlikely(!vq_is_packed(dev)))
1249                 return -1;
1250
1251         if (unlikely(vring_idx >= VHOST_MAX_VRING))
1252                 return -1;
1253
1254         vq = dev->virtqueue[vring_idx];
1255         if (unlikely(!vq))
1256                 return -1;
1257
1258         inflight_info = vq->inflight_packed;
1259         if (unlikely(!inflight_info))
1260                 return -1;
1261
1262         if (unlikely(head >= vq->size))
1263                 return -1;
1264
1265         last = inflight_info->desc[head].last;
1266         if (unlikely(last >= vq->size))
1267                 return -1;
1268
1269         inflight_info->desc[last].next = inflight_info->free_head;
1270         inflight_info->free_head = head;
1271         inflight_info->used_idx += inflight_info->desc[head].num;
1272         if (inflight_info->used_idx >= inflight_info->desc_num) {
1273                 inflight_info->used_idx -= inflight_info->desc_num;
1274                 inflight_info->used_wrap_counter =
1275                         !inflight_info->used_wrap_counter;
1276         }
1277
1278         return 0;
1279 }
1280
1281 int
1282 rte_vhost_vring_call(int vid, uint16_t vring_idx)
1283 {
1284         struct virtio_net *dev;
1285         struct vhost_virtqueue *vq;
1286
1287         dev = get_device(vid);
1288         if (!dev)
1289                 return -1;
1290
1291         if (vring_idx >= VHOST_MAX_VRING)
1292                 return -1;
1293
1294         vq = dev->virtqueue[vring_idx];
1295         if (!vq)
1296                 return -1;
1297
1298         if (vq_is_packed(dev))
1299                 vhost_vring_call_packed(dev, vq);
1300         else
1301                 vhost_vring_call_split(dev, vq);
1302
1303         return 0;
1304 }
1305
1306 uint16_t
1307 rte_vhost_avail_entries(int vid, uint16_t queue_id)
1308 {
1309         struct virtio_net *dev;
1310         struct vhost_virtqueue *vq;
1311         uint16_t ret = 0;
1312
1313         dev = get_device(vid);
1314         if (!dev)
1315                 return 0;
1316
1317         if (queue_id >= VHOST_MAX_VRING)
1318                 return 0;
1319
1320         vq = dev->virtqueue[queue_id];
1321         if (!vq)
1322                 return 0;
1323
1324         rte_spinlock_lock(&vq->access_lock);
1325
1326         if (unlikely(!vq->enabled || vq->avail == NULL))
1327                 goto out;
1328
1329         ret = *(volatile uint16_t *)&vq->avail->idx - vq->last_used_idx;
1330
1331 out:
1332         rte_spinlock_unlock(&vq->access_lock);
1333         return ret;
1334 }
1335
1336 static inline int
1337 vhost_enable_notify_split(struct virtio_net *dev,
1338                 struct vhost_virtqueue *vq, int enable)
1339 {
1340         if (vq->used == NULL)
1341                 return -1;
1342
1343         if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
1344                 if (enable)
1345                         vq->used->flags &= ~VRING_USED_F_NO_NOTIFY;
1346                 else
1347                         vq->used->flags |= VRING_USED_F_NO_NOTIFY;
1348         } else {
1349                 if (enable)
1350                         vhost_avail_event(vq) = vq->last_avail_idx;
1351         }
1352         return 0;
1353 }
1354
1355 static inline int
1356 vhost_enable_notify_packed(struct virtio_net *dev,
1357                 struct vhost_virtqueue *vq, int enable)
1358 {
1359         uint16_t flags;
1360
1361         if (vq->device_event == NULL)
1362                 return -1;
1363
1364         if (!enable) {
1365                 vq->device_event->flags = VRING_EVENT_F_DISABLE;
1366                 return 0;
1367         }
1368
1369         flags = VRING_EVENT_F_ENABLE;
1370         if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
1371                 flags = VRING_EVENT_F_DESC;
1372                 vq->device_event->off_wrap = vq->last_avail_idx |
1373                         vq->avail_wrap_counter << 15;
1374         }
1375
1376         rte_atomic_thread_fence(__ATOMIC_RELEASE);
1377
1378         vq->device_event->flags = flags;
1379         return 0;
1380 }
1381
1382 int
1383 vhost_enable_guest_notification(struct virtio_net *dev,
1384                 struct vhost_virtqueue *vq, int enable)
1385 {
1386         /*
1387          * If the virtqueue is not ready yet, it will be applied
1388          * when it will become ready.
1389          */
1390         if (!vq->ready)
1391                 return 0;
1392
1393         if (vq_is_packed(dev))
1394                 return vhost_enable_notify_packed(dev, vq, enable);
1395         else
1396                 return vhost_enable_notify_split(dev, vq, enable);
1397 }
1398
1399 int
1400 rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
1401 {
1402         struct virtio_net *dev = get_device(vid);
1403         struct vhost_virtqueue *vq;
1404         int ret;
1405
1406         if (!dev)
1407                 return -1;
1408
1409         if (queue_id >= VHOST_MAX_VRING)
1410                 return -1;
1411
1412         vq = dev->virtqueue[queue_id];
1413         if (!vq)
1414                 return -1;
1415
1416         rte_spinlock_lock(&vq->access_lock);
1417
1418         vq->notif_enable = enable;
1419         ret = vhost_enable_guest_notification(dev, vq, enable);
1420
1421         rte_spinlock_unlock(&vq->access_lock);
1422
1423         return ret;
1424 }
1425
1426 void
1427 rte_vhost_log_write(int vid, uint64_t addr, uint64_t len)
1428 {
1429         struct virtio_net *dev = get_device(vid);
1430
1431         if (dev == NULL)
1432                 return;
1433
1434         vhost_log_write(dev, addr, len);
1435 }
1436
1437 void
1438 rte_vhost_log_used_vring(int vid, uint16_t vring_idx,
1439                          uint64_t offset, uint64_t len)
1440 {
1441         struct virtio_net *dev;
1442         struct vhost_virtqueue *vq;
1443
1444         dev = get_device(vid);
1445         if (dev == NULL)
1446                 return;
1447
1448         if (vring_idx >= VHOST_MAX_VRING)
1449                 return;
1450         vq = dev->virtqueue[vring_idx];
1451         if (!vq)
1452                 return;
1453
1454         vhost_log_used_vring(dev, vq, offset, len);
1455 }
1456
1457 uint32_t
1458 rte_vhost_rx_queue_count(int vid, uint16_t qid)
1459 {
1460         struct virtio_net *dev;
1461         struct vhost_virtqueue *vq;
1462         uint32_t ret = 0;
1463
1464         dev = get_device(vid);
1465         if (dev == NULL)
1466                 return 0;
1467
1468         if (unlikely(qid >= dev->nr_vring || (qid & 1) == 0)) {
1469                 VHOST_LOG_DATA(ERR, "(%s) %s: invalid virtqueue idx %d.\n",
1470                         dev->ifname, __func__, qid);
1471                 return 0;
1472         }
1473
1474         vq = dev->virtqueue[qid];
1475         if (vq == NULL)
1476                 return 0;
1477
1478         rte_spinlock_lock(&vq->access_lock);
1479
1480         if (unlikely(!vq->enabled || vq->avail == NULL))
1481                 goto out;
1482
1483         ret = *((volatile uint16_t *)&vq->avail->idx) - vq->last_avail_idx;
1484
1485 out:
1486         rte_spinlock_unlock(&vq->access_lock);
1487         return ret;
1488 }
1489
1490 struct rte_vdpa_device *
1491 rte_vhost_get_vdpa_device(int vid)
1492 {
1493         struct virtio_net *dev = get_device(vid);
1494
1495         if (dev == NULL)
1496                 return NULL;
1497
1498         return dev->vdpa_dev;
1499 }
1500
1501 int
1502 rte_vhost_get_log_base(int vid, uint64_t *log_base,
1503                 uint64_t *log_size)
1504 {
1505         struct virtio_net *dev = get_device(vid);
1506
1507         if (dev == NULL || log_base == NULL || log_size == NULL)
1508                 return -1;
1509
1510         *log_base = dev->log_base;
1511         *log_size = dev->log_size;
1512
1513         return 0;
1514 }
1515
1516 int
1517 rte_vhost_get_vring_base(int vid, uint16_t queue_id,
1518                 uint16_t *last_avail_idx, uint16_t *last_used_idx)
1519 {
1520         struct vhost_virtqueue *vq;
1521         struct virtio_net *dev = get_device(vid);
1522
1523         if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1524                 return -1;
1525
1526         if (queue_id >= VHOST_MAX_VRING)
1527                 return -1;
1528
1529         vq = dev->virtqueue[queue_id];
1530         if (!vq)
1531                 return -1;
1532
1533         if (vq_is_packed(dev)) {
1534                 *last_avail_idx = (vq->avail_wrap_counter << 15) |
1535                                   vq->last_avail_idx;
1536                 *last_used_idx = (vq->used_wrap_counter << 15) |
1537                                  vq->last_used_idx;
1538         } else {
1539                 *last_avail_idx = vq->last_avail_idx;
1540                 *last_used_idx = vq->last_used_idx;
1541         }
1542
1543         return 0;
1544 }
1545
1546 int
1547 rte_vhost_set_vring_base(int vid, uint16_t queue_id,
1548                 uint16_t last_avail_idx, uint16_t last_used_idx)
1549 {
1550         struct vhost_virtqueue *vq;
1551         struct virtio_net *dev = get_device(vid);
1552
1553         if (!dev)
1554                 return -1;
1555
1556         if (queue_id >= VHOST_MAX_VRING)
1557                 return -1;
1558
1559         vq = dev->virtqueue[queue_id];
1560         if (!vq)
1561                 return -1;
1562
1563         if (vq_is_packed(dev)) {
1564                 vq->last_avail_idx = last_avail_idx & 0x7fff;
1565                 vq->avail_wrap_counter = !!(last_avail_idx & (1 << 15));
1566                 vq->last_used_idx = last_used_idx & 0x7fff;
1567                 vq->used_wrap_counter = !!(last_used_idx & (1 << 15));
1568         } else {
1569                 vq->last_avail_idx = last_avail_idx;
1570                 vq->last_used_idx = last_used_idx;
1571         }
1572
1573         return 0;
1574 }
1575
1576 int
1577 rte_vhost_get_vring_base_from_inflight(int vid,
1578                                        uint16_t queue_id,
1579                                        uint16_t *last_avail_idx,
1580                                        uint16_t *last_used_idx)
1581 {
1582         struct rte_vhost_inflight_info_packed *inflight_info;
1583         struct vhost_virtqueue *vq;
1584         struct virtio_net *dev = get_device(vid);
1585
1586         if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
1587                 return -1;
1588
1589         if (queue_id >= VHOST_MAX_VRING)
1590                 return -1;
1591
1592         vq = dev->virtqueue[queue_id];
1593         if (!vq)
1594                 return -1;
1595
1596         if (!vq_is_packed(dev))
1597                 return -1;
1598
1599         inflight_info = vq->inflight_packed;
1600         if (!inflight_info)
1601                 return -1;
1602
1603         *last_avail_idx = (inflight_info->old_used_wrap_counter << 15) |
1604                           inflight_info->old_used_idx;
1605         *last_used_idx = *last_avail_idx;
1606
1607         return 0;
1608 }
1609
1610 int
1611 rte_vhost_extern_callback_register(int vid,
1612                 struct rte_vhost_user_extern_ops const * const ops, void *ctx)
1613 {
1614         struct virtio_net *dev = get_device(vid);
1615
1616         if (dev == NULL || ops == NULL)
1617                 return -1;
1618
1619         dev->extern_ops = *ops;
1620         dev->extern_data = ctx;
1621         return 0;
1622 }
1623
1624 static __rte_always_inline int
1625 async_channel_register(int vid, uint16_t queue_id)
1626 {
1627         struct virtio_net *dev = get_device(vid);
1628         struct vhost_virtqueue *vq = dev->virtqueue[queue_id];
1629         struct vhost_async *async;
1630         int node = vq->numa_node;
1631
1632         if (unlikely(vq->async)) {
1633                 VHOST_LOG_CONFIG(ERR,
1634                                 "(%s) async register failed: already registered (qid: %d)\n",
1635                                 dev->ifname, queue_id);
1636                 return -1;
1637         }
1638
1639         async = rte_zmalloc_socket(NULL, sizeof(struct vhost_async), 0, node);
1640         if (!async) {
1641                 VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async metadata (qid: %d)\n",
1642                                 dev->ifname, queue_id);
1643                 return -1;
1644         }
1645
1646         async->pkts_info = rte_malloc_socket(NULL, vq->size * sizeof(struct async_inflight_info),
1647                         RTE_CACHE_LINE_SIZE, node);
1648         if (!async->pkts_info) {
1649                 VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async_pkts_info (qid: %d)\n",
1650                                 dev->ifname, queue_id);
1651                 goto out_free_async;
1652         }
1653
1654         async->pkts_cmpl_flag = rte_zmalloc_socket(NULL, vq->size * sizeof(bool),
1655                         RTE_CACHE_LINE_SIZE, node);
1656         if (!async->pkts_cmpl_flag) {
1657                 VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async pkts_cmpl_flag (qid: %d)\n",
1658                                 dev->ifname, queue_id);
1659                 goto out_free_async;
1660         }
1661
1662         if (vq_is_packed(dev)) {
1663                 async->buffers_packed = rte_malloc_socket(NULL,
1664                                 vq->size * sizeof(struct vring_used_elem_packed),
1665                                 RTE_CACHE_LINE_SIZE, node);
1666                 if (!async->buffers_packed) {
1667                         VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async buffers (qid: %d)\n",
1668                                         dev->ifname, queue_id);
1669                         goto out_free_inflight;
1670                 }
1671         } else {
1672                 async->descs_split = rte_malloc_socket(NULL,
1673                                 vq->size * sizeof(struct vring_used_elem),
1674                                 RTE_CACHE_LINE_SIZE, node);
1675                 if (!async->descs_split) {
1676                         VHOST_LOG_CONFIG(ERR, "(%s) failed to allocate async descs (qid: %d)\n",
1677                                         dev->ifname, queue_id);
1678                         goto out_free_inflight;
1679                 }
1680         }
1681
1682         vq->async = async;
1683
1684         return 0;
1685 out_free_inflight:
1686         rte_free(async->pkts_info);
1687 out_free_async:
1688         rte_free(async);
1689
1690         return -1;
1691 }
1692
1693 int
1694 rte_vhost_async_channel_register(int vid, uint16_t queue_id)
1695 {
1696         struct vhost_virtqueue *vq;
1697         struct virtio_net *dev = get_device(vid);
1698         int ret;
1699
1700         if (dev == NULL)
1701                 return -1;
1702
1703         if (queue_id >= VHOST_MAX_VRING)
1704                 return -1;
1705
1706         vq = dev->virtqueue[queue_id];
1707
1708         if (unlikely(vq == NULL || !dev->async_copy))
1709                 return -1;
1710
1711         rte_spinlock_lock(&vq->access_lock);
1712         ret = async_channel_register(vid, queue_id);
1713         rte_spinlock_unlock(&vq->access_lock);
1714
1715         return ret;
1716 }
1717
1718 int
1719 rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id)
1720 {
1721         struct vhost_virtqueue *vq;
1722         struct virtio_net *dev = get_device(vid);
1723
1724         if (dev == NULL)
1725                 return -1;
1726
1727         if (queue_id >= VHOST_MAX_VRING)
1728                 return -1;
1729
1730         vq = dev->virtqueue[queue_id];
1731
1732         if (unlikely(vq == NULL || !dev->async_copy))
1733                 return -1;
1734
1735         return async_channel_register(vid, queue_id);
1736 }
1737
1738 int
1739 rte_vhost_async_channel_unregister(int vid, uint16_t queue_id)
1740 {
1741         struct vhost_virtqueue *vq;
1742         struct virtio_net *dev = get_device(vid);
1743         int ret = -1;
1744
1745         if (dev == NULL)
1746                 return ret;
1747
1748         if (queue_id >= VHOST_MAX_VRING)
1749                 return ret;
1750
1751         vq = dev->virtqueue[queue_id];
1752
1753         if (vq == NULL)
1754                 return ret;
1755
1756         ret = 0;
1757
1758         if (!vq->async)
1759                 return ret;
1760
1761         if (!rte_spinlock_trylock(&vq->access_lock)) {
1762                 VHOST_LOG_CONFIG(ERR, "(%s) failed to unregister async channel, virtqueue busy.\n",
1763                                 dev->ifname);
1764                 return -1;
1765         }
1766
1767         if (vq->async->pkts_inflight_n) {
1768                 VHOST_LOG_CONFIG(ERR, "(%s) failed to unregister async channel.\n", dev->ifname);
1769                 VHOST_LOG_CONFIG(ERR, "(%s) inflight packets must be completed before unregistration.\n",
1770                         dev->ifname);
1771                 ret = -1;
1772                 goto out;
1773         }
1774
1775         vhost_free_async_mem(vq);
1776 out:
1777         rte_spinlock_unlock(&vq->access_lock);
1778
1779         return ret;
1780 }
1781
1782 int
1783 rte_vhost_async_channel_unregister_thread_unsafe(int vid, uint16_t queue_id)
1784 {
1785         struct vhost_virtqueue *vq;
1786         struct virtio_net *dev = get_device(vid);
1787
1788         if (dev == NULL)
1789                 return -1;
1790
1791         if (queue_id >= VHOST_MAX_VRING)
1792                 return -1;
1793
1794         vq = dev->virtqueue[queue_id];
1795
1796         if (vq == NULL)
1797                 return -1;
1798
1799         if (!vq->async)
1800                 return 0;
1801
1802         if (vq->async->pkts_inflight_n) {
1803                 VHOST_LOG_CONFIG(ERR, "(%s) failed to unregister async channel.\n", dev->ifname);
1804                 VHOST_LOG_CONFIG(ERR, "(%s) inflight packets must be completed before unregistration.\n",
1805                         dev->ifname);
1806                 return -1;
1807         }
1808
1809         vhost_free_async_mem(vq);
1810
1811         return 0;
1812 }
1813
1814 int
1815 rte_vhost_async_dma_configure(int16_t dma_id, uint16_t vchan_id)
1816 {
1817         struct rte_dma_info info;
1818         void *pkts_cmpl_flag_addr;
1819         uint16_t max_desc;
1820
1821         if (!rte_dma_is_valid(dma_id)) {
1822                 VHOST_LOG_CONFIG(ERR, "DMA %d is not found.\n", dma_id);
1823                 return -1;
1824         }
1825
1826         rte_dma_info_get(dma_id, &info);
1827         if (vchan_id >= info.max_vchans) {
1828                 VHOST_LOG_CONFIG(ERR, "Invalid DMA %d vChannel %u.\n", dma_id, vchan_id);
1829                 return -1;
1830         }
1831
1832         if (!dma_copy_track[dma_id].vchans) {
1833                 struct async_dma_vchan_info *vchans;
1834
1835                 vchans = rte_zmalloc(NULL, sizeof(struct async_dma_vchan_info) * info.max_vchans,
1836                                 RTE_CACHE_LINE_SIZE);
1837                 if (vchans == NULL) {
1838                         VHOST_LOG_CONFIG(ERR, "Failed to allocate vchans for DMA %d vChannel %u.\n",
1839                                         dma_id, vchan_id);
1840                         return -1;
1841                 }
1842
1843                 dma_copy_track[dma_id].vchans = vchans;
1844         }
1845
1846         if (dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr) {
1847                 VHOST_LOG_CONFIG(INFO, "DMA %d vChannel %u already registered.\n", dma_id,
1848                                 vchan_id);
1849                 return 0;
1850         }
1851
1852         max_desc = info.max_desc;
1853         if (!rte_is_power_of_2(max_desc))
1854                 max_desc = rte_align32pow2(max_desc);
1855
1856         pkts_cmpl_flag_addr = rte_zmalloc(NULL, sizeof(bool *) * max_desc, RTE_CACHE_LINE_SIZE);
1857         if (!pkts_cmpl_flag_addr) {
1858                 VHOST_LOG_CONFIG(ERR, "Failed to allocate pkts_cmpl_flag_addr for DMA %d "
1859                                 "vChannel %u.\n", dma_id, vchan_id);
1860
1861                 if (dma_copy_track[dma_id].nr_vchans == 0) {
1862                         rte_free(dma_copy_track[dma_id].vchans);
1863                         dma_copy_track[dma_id].vchans = NULL;
1864                 }
1865                 return -1;
1866         }
1867
1868         dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr = pkts_cmpl_flag_addr;
1869         dma_copy_track[dma_id].vchans[vchan_id].ring_size = max_desc;
1870         dma_copy_track[dma_id].vchans[vchan_id].ring_mask = max_desc - 1;
1871         dma_copy_track[dma_id].nr_vchans++;
1872
1873         return 0;
1874 }
1875
1876 int
1877 rte_vhost_async_get_inflight(int vid, uint16_t queue_id)
1878 {
1879         struct vhost_virtqueue *vq;
1880         struct virtio_net *dev = get_device(vid);
1881         int ret = -1;
1882
1883         if (dev == NULL)
1884                 return ret;
1885
1886         if (queue_id >= VHOST_MAX_VRING)
1887                 return ret;
1888
1889         vq = dev->virtqueue[queue_id];
1890
1891         if (vq == NULL)
1892                 return ret;
1893
1894         if (!vq->async)
1895                 return ret;
1896
1897         if (!rte_spinlock_trylock(&vq->access_lock)) {
1898                 VHOST_LOG_CONFIG(DEBUG,
1899                         "(%s) failed to check in-flight packets. virtqueue busy.\n",
1900                         dev->ifname);
1901                 return ret;
1902         }
1903
1904         ret = vq->async->pkts_inflight_n;
1905         rte_spinlock_unlock(&vq->access_lock);
1906
1907         return ret;
1908 }
1909
1910 int
1911 rte_vhost_get_monitor_addr(int vid, uint16_t queue_id,
1912                 struct rte_vhost_power_monitor_cond *pmc)
1913 {
1914         struct virtio_net *dev = get_device(vid);
1915         struct vhost_virtqueue *vq;
1916
1917         if (dev == NULL)
1918                 return -1;
1919         if (queue_id >= VHOST_MAX_VRING)
1920                 return -1;
1921
1922         vq = dev->virtqueue[queue_id];
1923         if (vq == NULL)
1924                 return -1;
1925
1926         if (vq_is_packed(dev)) {
1927                 struct vring_packed_desc *desc;
1928                 desc = vq->desc_packed;
1929                 pmc->addr = &desc[vq->last_avail_idx].flags;
1930                 if (vq->avail_wrap_counter)
1931                         pmc->val = VRING_DESC_F_AVAIL;
1932                 else
1933                         pmc->val = VRING_DESC_F_USED;
1934                 pmc->mask = VRING_DESC_F_AVAIL | VRING_DESC_F_USED;
1935                 pmc->size = sizeof(desc[vq->last_avail_idx].flags);
1936                 pmc->match = 1;
1937         } else {
1938                 pmc->addr = &vq->avail->idx;
1939                 pmc->val = vq->last_avail_idx & (vq->size - 1);
1940                 pmc->mask = vq->size - 1;
1941                 pmc->size = sizeof(vq->avail->idx);
1942                 pmc->match = 0;
1943         }
1944
1945         return 0;
1946 }
1947
1948 RTE_LOG_REGISTER_SUFFIX(vhost_config_log_level, config, INFO);
1949 RTE_LOG_REGISTER_SUFFIX(vhost_data_log_level, data, WARNING);