1 /* 2 * vhost-vdpa.c 3 * 4 * Copyright(c) 2017-2018 Intel Corporation. 5 * Copyright(c) 2020 Red Hat, Inc. 6 * 7 * This work is licensed under the terms of the GNU GPL, version 2 or later. 8 * See the COPYING file in the top-level directory. 9 * 10 */ 11 12 #include "qemu/osdep.h" 13 #include "clients.h" 14 #include "hw/virtio/virtio-net.h" 15 #include "net/vhost_net.h" 16 #include "net/vhost-vdpa.h" 17 #include "hw/virtio/vhost-vdpa.h" 18 #include "qemu/config-file.h" 19 #include "qemu/error-report.h" 20 #include "qemu/log.h" 21 #include "qemu/memalign.h" 22 #include "qemu/option.h" 23 #include "qapi/error.h" 24 #include <linux/vhost.h> 25 #include <sys/ioctl.h> 26 #include <err.h> 27 #include "standard-headers/linux/virtio_net.h" 28 #include "monitor/monitor.h" 29 #include "migration/misc.h" 30 #include "hw/virtio/vhost.h" 31 #include "trace.h" 32 33 /* Todo:need to add the multiqueue support here */ 34 typedef struct VhostVDPAState { 35 NetClientState nc; 36 struct vhost_vdpa vhost_vdpa; 37 NotifierWithReturn migration_state; 38 VHostNetState *vhost_net; 39 40 /* Control commands shadow buffers */ 41 void *cvq_cmd_out_buffer; 42 virtio_net_ctrl_ack *status; 43 44 /* The device always have SVQ enabled */ 45 bool always_svq; 46 47 /* The device can isolate CVQ in its own ASID */ 48 bool cvq_isolated; 49 50 bool started; 51 } VhostVDPAState; 52 53 /* 54 * The array is sorted alphabetically in ascending order, 55 * with the exception of VHOST_INVALID_FEATURE_BIT, 56 * which should always be the last entry. 57 */ 58 const int vdpa_feature_bits[] = { 59 VIRTIO_F_ANY_LAYOUT, 60 VIRTIO_F_IOMMU_PLATFORM, 61 VIRTIO_F_NOTIFY_ON_EMPTY, 62 VIRTIO_F_RING_PACKED, 63 VIRTIO_F_RING_RESET, 64 VIRTIO_F_VERSION_1, 65 VIRTIO_F_IN_ORDER, 66 VIRTIO_F_NOTIFICATION_DATA, 67 VIRTIO_NET_F_CSUM, 68 VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, 69 VIRTIO_NET_F_CTRL_MAC_ADDR, 70 VIRTIO_NET_F_CTRL_RX, 71 VIRTIO_NET_F_CTRL_RX_EXTRA, 72 VIRTIO_NET_F_CTRL_VLAN, 73 VIRTIO_NET_F_CTRL_VQ, 74 VIRTIO_NET_F_GSO, 75 VIRTIO_NET_F_GUEST_CSUM, 76 VIRTIO_NET_F_GUEST_ECN, 77 VIRTIO_NET_F_GUEST_TSO4, 78 VIRTIO_NET_F_GUEST_TSO6, 79 VIRTIO_NET_F_GUEST_UFO, 80 VIRTIO_NET_F_GUEST_USO4, 81 VIRTIO_NET_F_GUEST_USO6, 82 VIRTIO_NET_F_HASH_REPORT, 83 VIRTIO_NET_F_HOST_ECN, 84 VIRTIO_NET_F_HOST_TSO4, 85 VIRTIO_NET_F_HOST_TSO6, 86 VIRTIO_NET_F_HOST_UFO, 87 VIRTIO_NET_F_HOST_USO, 88 VIRTIO_NET_F_MQ, 89 VIRTIO_NET_F_MRG_RXBUF, 90 VIRTIO_NET_F_MTU, 91 VIRTIO_NET_F_RSC_EXT, 92 VIRTIO_NET_F_RSS, 93 VIRTIO_NET_F_STATUS, 94 VIRTIO_RING_F_EVENT_IDX, 95 VIRTIO_RING_F_INDIRECT_DESC, 96 97 /* VHOST_INVALID_FEATURE_BIT should always be the last entry */ 98 VHOST_INVALID_FEATURE_BIT 99 }; 100 101 /** Supported device specific feature bits with SVQ */ 102 static const uint64_t vdpa_svq_device_features = 103 BIT_ULL(VIRTIO_NET_F_CSUM) | 104 BIT_ULL(VIRTIO_NET_F_GUEST_CSUM) | 105 BIT_ULL(VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) | 106 BIT_ULL(VIRTIO_NET_F_MTU) | 107 BIT_ULL(VIRTIO_NET_F_MAC) | 108 BIT_ULL(VIRTIO_NET_F_GUEST_TSO4) | 109 BIT_ULL(VIRTIO_NET_F_GUEST_TSO6) | 110 BIT_ULL(VIRTIO_NET_F_GUEST_ECN) | 111 BIT_ULL(VIRTIO_NET_F_GUEST_UFO) | 112 BIT_ULL(VIRTIO_NET_F_HOST_TSO4) | 113 BIT_ULL(VIRTIO_NET_F_HOST_TSO6) | 114 BIT_ULL(VIRTIO_NET_F_HOST_ECN) | 115 BIT_ULL(VIRTIO_NET_F_HOST_UFO) | 116 BIT_ULL(VIRTIO_NET_F_MRG_RXBUF) | 117 BIT_ULL(VIRTIO_NET_F_STATUS) | 118 BIT_ULL(VIRTIO_NET_F_CTRL_VQ) | 119 BIT_ULL(VIRTIO_NET_F_CTRL_RX) | 120 BIT_ULL(VIRTIO_NET_F_CTRL_VLAN) | 121 BIT_ULL(VIRTIO_NET_F_CTRL_RX_EXTRA) | 122 BIT_ULL(VIRTIO_NET_F_MQ) | 123 BIT_ULL(VIRTIO_F_ANY_LAYOUT) | 124 BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR) | 125 /* VHOST_F_LOG_ALL is exposed by SVQ */ 126 BIT_ULL(VHOST_F_LOG_ALL) | 127 BIT_ULL(VIRTIO_NET_F_HASH_REPORT) | 128 BIT_ULL(VIRTIO_NET_F_RSS) | 129 BIT_ULL(VIRTIO_NET_F_RSC_EXT) | 130 BIT_ULL(VIRTIO_NET_F_STANDBY) | 131 BIT_ULL(VIRTIO_NET_F_SPEED_DUPLEX); 132 133 #define VHOST_VDPA_NET_CVQ_ASID 1 134 135 VHostNetState *vhost_vdpa_get_vhost_net(NetClientState *nc) 136 { 137 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 138 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 139 return s->vhost_net; 140 } 141 142 static size_t vhost_vdpa_net_cvq_cmd_len(void) 143 { 144 /* 145 * MAC_TABLE_SET is the ctrl command that produces the longer out buffer. 146 * In buffer is always 1 byte, so it should fit here 147 */ 148 return sizeof(struct virtio_net_ctrl_hdr) + 149 2 * sizeof(struct virtio_net_ctrl_mac) + 150 MAC_TABLE_ENTRIES * ETH_ALEN; 151 } 152 153 static size_t vhost_vdpa_net_cvq_cmd_page_len(void) 154 { 155 return ROUND_UP(vhost_vdpa_net_cvq_cmd_len(), qemu_real_host_page_size()); 156 } 157 158 static bool vhost_vdpa_net_valid_svq_features(uint64_t features, Error **errp) 159 { 160 uint64_t invalid_dev_features = 161 features & ~vdpa_svq_device_features & 162 /* Transport are all accepted at this point */ 163 ~MAKE_64BIT_MASK(VIRTIO_TRANSPORT_F_START, 164 VIRTIO_TRANSPORT_F_END - VIRTIO_TRANSPORT_F_START); 165 166 if (invalid_dev_features) { 167 error_setg(errp, "vdpa svq does not work with features 0x%" PRIx64, 168 invalid_dev_features); 169 return false; 170 } 171 172 return vhost_svq_valid_features(features, errp); 173 } 174 175 static int vhost_vdpa_net_check_device_id(struct vhost_net *net) 176 { 177 uint32_t device_id; 178 int ret; 179 struct vhost_dev *hdev; 180 181 hdev = (struct vhost_dev *)&net->dev; 182 ret = hdev->vhost_ops->vhost_get_device_id(hdev, &device_id); 183 if (device_id != VIRTIO_ID_NET) { 184 return -ENOTSUP; 185 } 186 return ret; 187 } 188 189 static int vhost_vdpa_add(NetClientState *ncs, void *be, 190 int queue_pair_index, int nvqs) 191 { 192 VhostNetOptions options; 193 struct vhost_net *net = NULL; 194 VhostVDPAState *s; 195 int ret; 196 197 options.backend_type = VHOST_BACKEND_TYPE_VDPA; 198 assert(ncs->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 199 s = DO_UPCAST(VhostVDPAState, nc, ncs); 200 options.net_backend = ncs; 201 options.opaque = be; 202 options.busyloop_timeout = 0; 203 options.nvqs = nvqs; 204 205 net = vhost_net_init(&options); 206 if (!net) { 207 error_report("failed to init vhost_net for queue"); 208 goto err_init; 209 } 210 s->vhost_net = net; 211 ret = vhost_vdpa_net_check_device_id(net); 212 if (ret) { 213 goto err_check; 214 } 215 return 0; 216 err_check: 217 vhost_net_cleanup(net); 218 g_free(net); 219 err_init: 220 return -1; 221 } 222 223 static void vhost_vdpa_cleanup(NetClientState *nc) 224 { 225 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 226 227 munmap(s->cvq_cmd_out_buffer, vhost_vdpa_net_cvq_cmd_page_len()); 228 munmap(s->status, vhost_vdpa_net_cvq_cmd_page_len()); 229 if (s->vhost_net) { 230 vhost_net_cleanup(s->vhost_net); 231 g_free(s->vhost_net); 232 s->vhost_net = NULL; 233 } 234 if (s->vhost_vdpa.index != 0) { 235 return; 236 } 237 qemu_close(s->vhost_vdpa.shared->device_fd); 238 g_clear_pointer(&s->vhost_vdpa.shared->iova_tree, vhost_iova_tree_delete); 239 g_free(s->vhost_vdpa.shared); 240 } 241 242 /** Dummy SetSteeringEBPF to support RSS for vhost-vdpa backend */ 243 static bool vhost_vdpa_set_steering_ebpf(NetClientState *nc, int prog_fd) 244 { 245 return true; 246 } 247 248 static bool vhost_vdpa_has_vnet_hdr(NetClientState *nc) 249 { 250 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 251 252 return true; 253 } 254 255 static bool vhost_vdpa_get_vnet_hash_supported_types(NetClientState *nc, 256 uint32_t *types) 257 { 258 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 259 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 260 uint64_t features = s->vhost_vdpa.dev->features; 261 int fd = s->vhost_vdpa.shared->device_fd; 262 struct { 263 struct vhost_vdpa_config hdr; 264 uint32_t supported_hash_types; 265 } config; 266 267 if (!virtio_has_feature(features, VIRTIO_NET_F_HASH_REPORT) && 268 !virtio_has_feature(features, VIRTIO_NET_F_RSS)) { 269 return false; 270 } 271 272 config.hdr.off = offsetof(struct virtio_net_config, supported_hash_types); 273 config.hdr.len = sizeof(config.supported_hash_types); 274 275 assert(!ioctl(fd, VHOST_VDPA_GET_CONFIG, &config)); 276 *types = le32_to_cpu(config.supported_hash_types); 277 278 return true; 279 } 280 281 static bool vhost_vdpa_has_ufo(NetClientState *nc) 282 { 283 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 284 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 285 uint64_t features = 0; 286 features |= (1ULL << VIRTIO_NET_F_HOST_UFO); 287 features = vhost_net_get_features(s->vhost_net, features); 288 return !!(features & (1ULL << VIRTIO_NET_F_HOST_UFO)); 289 290 } 291 292 /* 293 * FIXME: vhost_vdpa doesn't have an API to "set h/w endianness". But it's 294 * reasonable to assume that h/w is LE by default, because LE is what 295 * virtio 1.0 and later ask for. So, this function just says "yes, the h/w is 296 * LE". Otherwise, on a BE machine, higher-level code would mistakely think 297 * the h/w is BE and can't support VDPA for a virtio 1.0 client. 298 */ 299 static int vhost_vdpa_set_vnet_le(NetClientState *nc, bool enable) 300 { 301 return 0; 302 } 303 304 static bool vhost_vdpa_check_peer_type(NetClientState *nc, ObjectClass *oc, 305 Error **errp) 306 { 307 const char *driver = object_class_get_name(oc); 308 309 if (!g_str_has_prefix(driver, "virtio-net-")) { 310 error_setg(errp, "vhost-vdpa requires frontend driver virtio-net-*"); 311 return false; 312 } 313 314 return true; 315 } 316 317 /** Dummy receive in case qemu falls back to userland tap networking */ 318 static ssize_t vhost_vdpa_receive(NetClientState *nc, const uint8_t *buf, 319 size_t size) 320 { 321 return size; 322 } 323 324 325 /** From any vdpa net client, get the netclient of the i-th queue pair */ 326 static VhostVDPAState *vhost_vdpa_net_get_nc_vdpa(VhostVDPAState *s, int i) 327 { 328 NICState *nic = qemu_get_nic(s->nc.peer); 329 NetClientState *nc_i = qemu_get_peer(nic->ncs, i); 330 331 return DO_UPCAST(VhostVDPAState, nc, nc_i); 332 } 333 334 static VhostVDPAState *vhost_vdpa_net_first_nc_vdpa(VhostVDPAState *s) 335 { 336 return vhost_vdpa_net_get_nc_vdpa(s, 0); 337 } 338 339 static void vhost_vdpa_net_log_global_enable(VhostVDPAState *s, bool enable) 340 { 341 struct vhost_vdpa *v = &s->vhost_vdpa; 342 VirtIONet *n; 343 VirtIODevice *vdev; 344 int data_queue_pairs, cvq, r; 345 346 /* We are only called on the first data vqs and only if x-svq is not set */ 347 if (s->vhost_vdpa.shadow_vqs_enabled == enable) { 348 return; 349 } 350 351 vdev = v->dev->vdev; 352 n = VIRTIO_NET(vdev); 353 if (!n->vhost_started) { 354 return; 355 } 356 357 data_queue_pairs = n->multiqueue ? n->max_queue_pairs : 1; 358 cvq = virtio_vdev_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) ? 359 n->max_ncs - n->max_queue_pairs : 0; 360 v->shared->svq_switching = enable ? 361 SVQ_TSTATE_ENABLING : SVQ_TSTATE_DISABLING; 362 /* 363 * TODO: vhost_net_stop does suspend, get_base and reset. We can be smarter 364 * in the future and resume the device if read-only operations between 365 * suspend and reset goes wrong. 366 */ 367 vhost_net_stop(vdev, n->nic->ncs, data_queue_pairs, cvq); 368 369 /* Start will check migration setup_or_active to configure or not SVQ */ 370 r = vhost_net_start(vdev, n->nic->ncs, data_queue_pairs, cvq); 371 if (unlikely(r < 0)) { 372 error_report("unable to start vhost net: %s(%d)", g_strerror(-r), -r); 373 } 374 v->shared->svq_switching = SVQ_TSTATE_DONE; 375 } 376 377 static int vdpa_net_migration_state_notifier(NotifierWithReturn *notifier, 378 MigrationEvent *e, Error **errp) 379 { 380 VhostVDPAState *s = container_of(notifier, VhostVDPAState, migration_state); 381 382 if (e->type == MIG_EVENT_PRECOPY_SETUP) { 383 vhost_vdpa_net_log_global_enable(s, true); 384 } else if (e->type == MIG_EVENT_PRECOPY_FAILED) { 385 vhost_vdpa_net_log_global_enable(s, false); 386 } 387 return 0; 388 } 389 390 static void vhost_vdpa_net_data_start_first(VhostVDPAState *s) 391 { 392 migration_add_notifier(&s->migration_state, 393 vdpa_net_migration_state_notifier); 394 } 395 396 static int vhost_vdpa_net_data_start(NetClientState *nc) 397 { 398 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 399 struct vhost_vdpa *v = &s->vhost_vdpa; 400 401 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 402 403 if (s->always_svq || migration_is_running()) { 404 v->shadow_vqs_enabled = true; 405 } else { 406 v->shadow_vqs_enabled = false; 407 } 408 409 if (v->index == 0) { 410 v->shared->shadow_data = v->shadow_vqs_enabled; 411 vhost_vdpa_net_data_start_first(s); 412 return 0; 413 } 414 415 return 0; 416 } 417 418 static int vhost_vdpa_net_data_load(NetClientState *nc) 419 { 420 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 421 struct vhost_vdpa *v = &s->vhost_vdpa; 422 bool has_cvq = v->dev->vq_index_end % 2; 423 424 if (has_cvq) { 425 return 0; 426 } 427 428 for (int i = 0; i < v->dev->nvqs; ++i) { 429 int ret = vhost_vdpa_set_vring_ready(v, i + v->dev->vq_index); 430 if (ret < 0) { 431 return ret; 432 } 433 } 434 return 0; 435 } 436 437 static void vhost_vdpa_net_client_stop(NetClientState *nc) 438 { 439 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 440 441 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 442 443 if (s->vhost_vdpa.index == 0) { 444 migration_remove_notifier(&s->migration_state); 445 } 446 } 447 448 static NetClientInfo net_vhost_vdpa_info = { 449 .type = NET_CLIENT_DRIVER_VHOST_VDPA, 450 .size = sizeof(VhostVDPAState), 451 .receive = vhost_vdpa_receive, 452 .start = vhost_vdpa_net_data_start, 453 .load = vhost_vdpa_net_data_load, 454 .stop = vhost_vdpa_net_client_stop, 455 .cleanup = vhost_vdpa_cleanup, 456 .has_vnet_hdr = vhost_vdpa_has_vnet_hdr, 457 .get_vnet_hash_supported_types = vhost_vdpa_get_vnet_hash_supported_types, 458 .has_ufo = vhost_vdpa_has_ufo, 459 .set_vnet_le = vhost_vdpa_set_vnet_le, 460 .check_peer_type = vhost_vdpa_check_peer_type, 461 .set_steering_ebpf = vhost_vdpa_set_steering_ebpf, 462 }; 463 464 static int64_t vhost_vdpa_get_vring_group(int device_fd, unsigned vq_index, 465 Error **errp) 466 { 467 struct vhost_vring_state state = { 468 .index = vq_index, 469 }; 470 int r = ioctl(device_fd, VHOST_VDPA_GET_VRING_GROUP, &state); 471 472 if (unlikely(r < 0)) { 473 r = -errno; 474 error_setg_errno(errp, errno, "Cannot get VQ %u group", vq_index); 475 return r; 476 } 477 478 return state.num; 479 } 480 481 static int vhost_vdpa_set_address_space_id(struct vhost_vdpa *v, 482 unsigned vq_group, 483 unsigned asid_num) 484 { 485 struct vhost_vring_state asid = { 486 .index = vq_group, 487 .num = asid_num, 488 }; 489 int r; 490 491 trace_vhost_vdpa_set_address_space_id(v, vq_group, asid_num); 492 493 r = ioctl(v->shared->device_fd, VHOST_VDPA_SET_GROUP_ASID, &asid); 494 if (unlikely(r < 0)) { 495 error_report("Can't set vq group %u asid %u, errno=%d (%s)", 496 asid.index, asid.num, errno, g_strerror(errno)); 497 } 498 return r; 499 } 500 501 static void vhost_vdpa_cvq_unmap_buf(struct vhost_vdpa *v, void *addr) 502 { 503 VhostIOVATree *tree = v->shared->iova_tree; 504 DMAMap needle = { 505 /* 506 * No need to specify size or to look for more translations since 507 * this contiguous chunk was allocated by us. 508 */ 509 .translated_addr = (hwaddr)(uintptr_t)addr, 510 }; 511 const DMAMap *map = vhost_iova_tree_find_iova(tree, &needle); 512 int r; 513 514 if (unlikely(!map)) { 515 error_report("Cannot locate expected map"); 516 return; 517 } 518 519 r = vhost_vdpa_dma_unmap(v->shared, v->address_space_id, map->iova, 520 map->size + 1); 521 if (unlikely(r != 0)) { 522 error_report("Device cannot unmap: %s(%d)", g_strerror(r), r); 523 } 524 525 vhost_iova_tree_remove(tree, *map); 526 } 527 528 /** Map CVQ buffer. */ 529 static int vhost_vdpa_cvq_map_buf(struct vhost_vdpa *v, void *buf, size_t size, 530 bool write) 531 { 532 DMAMap map = {}; 533 hwaddr taddr = (hwaddr)(uintptr_t)buf; 534 int r; 535 536 map.size = size - 1; 537 map.perm = write ? IOMMU_RW : IOMMU_RO, 538 r = vhost_iova_tree_map_alloc(v->shared->iova_tree, &map, taddr); 539 if (unlikely(r != IOVA_OK)) { 540 error_report("Cannot map injected element"); 541 542 if (map.translated_addr == taddr) { 543 error_report("Insertion to IOVA->HVA tree failed"); 544 /* Remove the mapping from the IOVA-only tree */ 545 goto dma_map_err; 546 } 547 return r; 548 } 549 550 r = vhost_vdpa_dma_map(v->shared, v->address_space_id, map.iova, 551 vhost_vdpa_net_cvq_cmd_page_len(), buf, !write); 552 if (unlikely(r < 0)) { 553 goto dma_map_err; 554 } 555 556 return 0; 557 558 dma_map_err: 559 vhost_iova_tree_remove(v->shared->iova_tree, map); 560 return r; 561 } 562 563 static int vhost_vdpa_net_cvq_start(NetClientState *nc) 564 { 565 VhostVDPAState *s, *s0; 566 struct vhost_vdpa *v; 567 int64_t cvq_group; 568 int r; 569 Error *err = NULL; 570 571 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 572 573 s = DO_UPCAST(VhostVDPAState, nc, nc); 574 v = &s->vhost_vdpa; 575 576 s0 = vhost_vdpa_net_first_nc_vdpa(s); 577 v->shadow_vqs_enabled = s0->vhost_vdpa.shadow_vqs_enabled; 578 s->vhost_vdpa.address_space_id = VHOST_VDPA_GUEST_PA_ASID; 579 580 if (v->shared->shadow_data) { 581 /* SVQ is already configured for all virtqueues */ 582 goto out; 583 } 584 585 /* 586 * If we early return in these cases SVQ will not be enabled. The migration 587 * will be blocked as long as vhost-vdpa backends will not offer _F_LOG. 588 */ 589 if (!vhost_vdpa_net_valid_svq_features(v->dev->features, NULL)) { 590 return 0; 591 } 592 593 if (!s->cvq_isolated) { 594 return 0; 595 } 596 597 cvq_group = vhost_vdpa_get_vring_group(v->shared->device_fd, 598 v->dev->vq_index_end - 1, 599 &err); 600 if (unlikely(cvq_group < 0)) { 601 error_report_err(err); 602 return cvq_group; 603 } 604 605 r = vhost_vdpa_set_address_space_id(v, cvq_group, VHOST_VDPA_NET_CVQ_ASID); 606 if (unlikely(r < 0)) { 607 return r; 608 } 609 610 v->shadow_vqs_enabled = true; 611 s->vhost_vdpa.address_space_id = VHOST_VDPA_NET_CVQ_ASID; 612 613 out: 614 if (!s->vhost_vdpa.shadow_vqs_enabled) { 615 return 0; 616 } 617 618 r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer, 619 vhost_vdpa_net_cvq_cmd_page_len(), false); 620 if (unlikely(r < 0)) { 621 return r; 622 } 623 624 r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->status, 625 vhost_vdpa_net_cvq_cmd_page_len(), true); 626 if (unlikely(r < 0)) { 627 vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer); 628 } 629 630 return r; 631 } 632 633 static void vhost_vdpa_net_cvq_stop(NetClientState *nc) 634 { 635 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 636 637 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 638 639 if (s->vhost_vdpa.shadow_vqs_enabled) { 640 vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer); 641 vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->status); 642 } 643 644 vhost_vdpa_net_client_stop(nc); 645 } 646 647 static ssize_t vhost_vdpa_net_cvq_add(VhostVDPAState *s, 648 const struct iovec *out_sg, size_t out_num, 649 const struct iovec *in_sg, size_t in_num) 650 { 651 VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0); 652 int r; 653 654 r = vhost_svq_add(svq, out_sg, out_num, NULL, in_sg, in_num, NULL, NULL); 655 if (unlikely(r != 0)) { 656 if (unlikely(r == -ENOSPC)) { 657 qemu_log_mask(LOG_GUEST_ERROR, "%s: No space on device queue\n", 658 __func__); 659 } 660 } 661 662 return r; 663 } 664 665 /* 666 * Convenience wrapper to poll SVQ for multiple control commands. 667 * 668 * Caller should hold the BQL when invoking this function, and should take 669 * the answer before SVQ pulls by itself when BQL is released. 670 */ 671 static ssize_t vhost_vdpa_net_svq_poll(VhostVDPAState *s, size_t cmds_in_flight) 672 { 673 VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0); 674 return vhost_svq_poll(svq, cmds_in_flight); 675 } 676 677 static void vhost_vdpa_net_load_cursor_reset(VhostVDPAState *s, 678 struct iovec *out_cursor, 679 struct iovec *in_cursor) 680 { 681 /* reset the cursor of the output buffer for the device */ 682 out_cursor->iov_base = s->cvq_cmd_out_buffer; 683 out_cursor->iov_len = vhost_vdpa_net_cvq_cmd_page_len(); 684 685 /* reset the cursor of the in buffer for the device */ 686 in_cursor->iov_base = s->status; 687 in_cursor->iov_len = vhost_vdpa_net_cvq_cmd_page_len(); 688 } 689 690 /* 691 * Poll SVQ for multiple pending control commands and check the device's ack. 692 * 693 * Caller should hold the BQL when invoking this function. 694 * 695 * @s: The VhostVDPAState 696 * @len: The length of the pending status shadow buffer 697 */ 698 static ssize_t vhost_vdpa_net_svq_flush(VhostVDPAState *s, size_t len) 699 { 700 /* device uses a one-byte length ack for each control command */ 701 ssize_t dev_written = vhost_vdpa_net_svq_poll(s, len); 702 if (unlikely(dev_written != len)) { 703 return -EIO; 704 } 705 706 /* check the device's ack */ 707 for (int i = 0; i < len; ++i) { 708 if (s->status[i] != VIRTIO_NET_OK) { 709 return -EIO; 710 } 711 } 712 return 0; 713 } 714 715 static ssize_t vhost_vdpa_net_load_cmd(VhostVDPAState *s, 716 struct iovec *out_cursor, 717 struct iovec *in_cursor, uint8_t class, 718 uint8_t cmd, const struct iovec *data_sg, 719 size_t data_num) 720 { 721 const struct virtio_net_ctrl_hdr ctrl = { 722 .class = class, 723 .cmd = cmd, 724 }; 725 size_t data_size = iov_size(data_sg, data_num), cmd_size; 726 struct iovec out, in; 727 ssize_t r; 728 unsigned dummy_cursor_iov_cnt; 729 VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0); 730 731 assert(data_size < vhost_vdpa_net_cvq_cmd_page_len() - sizeof(ctrl)); 732 cmd_size = sizeof(ctrl) + data_size; 733 trace_vhost_vdpa_net_load_cmd(s, class, cmd, data_num, data_size); 734 if (vhost_svq_available_slots(svq) < 2 || 735 iov_size(out_cursor, 1) < cmd_size) { 736 /* 737 * It is time to flush all pending control commands if SVQ is full 738 * or control commands shadow buffers are full. 739 * 740 * We can poll here since we've had BQL from the time 741 * we sent the descriptor. 742 */ 743 r = vhost_vdpa_net_svq_flush(s, in_cursor->iov_base - 744 (void *)s->status); 745 if (unlikely(r < 0)) { 746 return r; 747 } 748 749 vhost_vdpa_net_load_cursor_reset(s, out_cursor, in_cursor); 750 } 751 752 /* pack the CVQ command header */ 753 iov_from_buf(out_cursor, 1, 0, &ctrl, sizeof(ctrl)); 754 /* pack the CVQ command command-specific-data */ 755 iov_to_buf(data_sg, data_num, 0, 756 out_cursor->iov_base + sizeof(ctrl), data_size); 757 758 /* extract the required buffer from the cursor for output */ 759 iov_copy(&out, 1, out_cursor, 1, 0, cmd_size); 760 /* extract the required buffer from the cursor for input */ 761 iov_copy(&in, 1, in_cursor, 1, 0, sizeof(*s->status)); 762 763 r = vhost_vdpa_net_cvq_add(s, &out, 1, &in, 1); 764 if (unlikely(r < 0)) { 765 trace_vhost_vdpa_net_load_cmd_retval(s, class, cmd, r); 766 return r; 767 } 768 769 /* iterate the cursors */ 770 dummy_cursor_iov_cnt = 1; 771 iov_discard_front(&out_cursor, &dummy_cursor_iov_cnt, cmd_size); 772 dummy_cursor_iov_cnt = 1; 773 iov_discard_front(&in_cursor, &dummy_cursor_iov_cnt, sizeof(*s->status)); 774 775 return 0; 776 } 777 778 static int vhost_vdpa_net_load_mac(VhostVDPAState *s, const VirtIONet *n, 779 struct iovec *out_cursor, 780 struct iovec *in_cursor) 781 { 782 if (virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_MAC_ADDR)) { 783 const struct iovec data = { 784 .iov_base = (void *)n->mac, 785 .iov_len = sizeof(n->mac), 786 }; 787 ssize_t r = vhost_vdpa_net_load_cmd(s, out_cursor, in_cursor, 788 VIRTIO_NET_CTRL_MAC, 789 VIRTIO_NET_CTRL_MAC_ADDR_SET, 790 &data, 1); 791 if (unlikely(r < 0)) { 792 return r; 793 } 794 } 795 796 /* 797 * According to VirtIO standard, "The device MUST have an 798 * empty MAC filtering table on reset.". 799 * 800 * Therefore, there is no need to send this CVQ command if the 801 * driver also sets an empty MAC filter table, which aligns with 802 * the device's defaults. 803 * 804 * Note that the device's defaults can mismatch the driver's 805 * configuration only at live migration. 806 */ 807 if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX) || 808 n->mac_table.in_use == 0) { 809 return 0; 810 } 811 812 uint32_t uni_entries = n->mac_table.first_multi, 813 uni_macs_size = uni_entries * ETH_ALEN, 814 mul_entries = n->mac_table.in_use - uni_entries, 815 mul_macs_size = mul_entries * ETH_ALEN; 816 struct virtio_net_ctrl_mac uni = { 817 .entries = cpu_to_le32(uni_entries), 818 }; 819 struct virtio_net_ctrl_mac mul = { 820 .entries = cpu_to_le32(mul_entries), 821 }; 822 const struct iovec data[] = { 823 { 824 .iov_base = &uni, 825 .iov_len = sizeof(uni), 826 }, { 827 .iov_base = n->mac_table.macs, 828 .iov_len = uni_macs_size, 829 }, { 830 .iov_base = &mul, 831 .iov_len = sizeof(mul), 832 }, { 833 .iov_base = &n->mac_table.macs[uni_macs_size], 834 .iov_len = mul_macs_size, 835 }, 836 }; 837 ssize_t r = vhost_vdpa_net_load_cmd(s, out_cursor, in_cursor, 838 VIRTIO_NET_CTRL_MAC, 839 VIRTIO_NET_CTRL_MAC_TABLE_SET, 840 data, ARRAY_SIZE(data)); 841 if (unlikely(r < 0)) { 842 return r; 843 } 844 845 return 0; 846 } 847 848 static int vhost_vdpa_net_load_rss(VhostVDPAState *s, const VirtIONet *n, 849 struct iovec *out_cursor, 850 struct iovec *in_cursor, bool do_rss) 851 { 852 struct virtio_net_rss_config cfg = {}; 853 ssize_t r; 854 g_autofree uint16_t *table = NULL; 855 856 /* 857 * According to VirtIO standard, "Initially the device has all hash 858 * types disabled and reports only VIRTIO_NET_HASH_REPORT_NONE.". 859 * 860 * Therefore, there is no need to send this CVQ command if the 861 * driver disables the all hash types, which aligns with 862 * the device's defaults. 863 * 864 * Note that the device's defaults can mismatch the driver's 865 * configuration only at live migration. 866 */ 867 if (!n->rss_data.enabled || 868 n->rss_data.runtime_hash_types == VIRTIO_NET_HASH_REPORT_NONE) { 869 return 0; 870 } 871 872 table = g_malloc_n(n->rss_data.indirections_len, 873 sizeof(n->rss_data.indirections_table[0])); 874 cfg.hash_types = cpu_to_le32(n->rss_data.runtime_hash_types); 875 876 if (do_rss) { 877 /* 878 * According to VirtIO standard, "Number of entries in indirection_table 879 * is (indirection_table_mask + 1)". 880 */ 881 cfg.indirection_table_mask = cpu_to_le16(n->rss_data.indirections_len - 882 1); 883 cfg.unclassified_queue = cpu_to_le16(n->rss_data.default_queue); 884 for (int i = 0; i < n->rss_data.indirections_len; ++i) { 885 table[i] = cpu_to_le16(n->rss_data.indirections_table[i]); 886 } 887 cfg.max_tx_vq = cpu_to_le16(n->curr_queue_pairs); 888 } else { 889 /* 890 * According to VirtIO standard, "Field reserved MUST contain zeroes. 891 * It is defined to make the structure to match the layout of 892 * virtio_net_rss_config structure, defined in 5.1.6.5.7.". 893 * 894 * Therefore, we need to zero the fields in 895 * struct virtio_net_rss_config, which corresponds to the 896 * `reserved` field in struct virtio_net_hash_config. 897 * 898 * Note that all other fields are zeroed at their definitions, 899 * except for the `indirection_table` field, where the actual data 900 * is stored in the `table` variable to ensure compatibility 901 * with RSS case. Therefore, we need to zero the `table` variable here. 902 */ 903 table[0] = 0; 904 } 905 906 /* 907 * Considering that virtio_net_handle_rss() currently does not restore 908 * the hash key length parsed from the CVQ command sent from the guest 909 * into n->rss_data and uses the maximum key length in other code, so 910 * we also employ the maximum key length here. 911 */ 912 cfg.hash_key_length = sizeof(n->rss_data.key); 913 914 const struct iovec data[] = { 915 { 916 .iov_base = &cfg, 917 .iov_len = offsetof(struct virtio_net_rss_config, 918 indirection_table), 919 }, { 920 .iov_base = table, 921 .iov_len = n->rss_data.indirections_len * 922 sizeof(n->rss_data.indirections_table[0]), 923 }, { 924 .iov_base = &cfg.max_tx_vq, 925 .iov_len = offsetof(struct virtio_net_rss_config, hash_key_data) - 926 offsetof(struct virtio_net_rss_config, max_tx_vq), 927 }, { 928 .iov_base = (void *)n->rss_data.key, 929 .iov_len = sizeof(n->rss_data.key), 930 } 931 }; 932 933 r = vhost_vdpa_net_load_cmd(s, out_cursor, in_cursor, 934 VIRTIO_NET_CTRL_MQ, 935 do_rss ? VIRTIO_NET_CTRL_MQ_RSS_CONFIG : 936 VIRTIO_NET_CTRL_MQ_HASH_CONFIG, 937 data, ARRAY_SIZE(data)); 938 if (unlikely(r < 0)) { 939 return r; 940 } 941 942 return 0; 943 } 944 945 static int vhost_vdpa_net_load_mq(VhostVDPAState *s, 946 const VirtIONet *n, 947 struct iovec *out_cursor, 948 struct iovec *in_cursor) 949 { 950 struct virtio_net_ctrl_mq mq; 951 ssize_t r; 952 953 if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_MQ)) { 954 return 0; 955 } 956 957 trace_vhost_vdpa_net_load_mq(s, n->curr_queue_pairs); 958 959 mq.virtqueue_pairs = cpu_to_le16(n->curr_queue_pairs); 960 const struct iovec data = { 961 .iov_base = &mq, 962 .iov_len = sizeof(mq), 963 }; 964 r = vhost_vdpa_net_load_cmd(s, out_cursor, in_cursor, 965 VIRTIO_NET_CTRL_MQ, 966 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, 967 &data, 1); 968 if (unlikely(r < 0)) { 969 return r; 970 } 971 972 if (virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_RSS)) { 973 /* load the receive-side scaling state */ 974 r = vhost_vdpa_net_load_rss(s, n, out_cursor, in_cursor, true); 975 if (unlikely(r < 0)) { 976 return r; 977 } 978 } else if (virtio_vdev_has_feature(&n->parent_obj, 979 VIRTIO_NET_F_HASH_REPORT)) { 980 /* load the hash calculation state */ 981 r = vhost_vdpa_net_load_rss(s, n, out_cursor, in_cursor, false); 982 if (unlikely(r < 0)) { 983 return r; 984 } 985 } 986 987 return 0; 988 } 989 990 static int vhost_vdpa_net_load_offloads(VhostVDPAState *s, 991 const VirtIONet *n, 992 struct iovec *out_cursor, 993 struct iovec *in_cursor) 994 { 995 uint64_t offloads; 996 ssize_t r; 997 998 if (!virtio_vdev_has_feature(&n->parent_obj, 999 VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)) { 1000 return 0; 1001 } 1002 1003 if (n->curr_guest_offloads == virtio_net_supported_guest_offloads(n)) { 1004 /* 1005 * According to VirtIO standard, "Upon feature negotiation 1006 * corresponding offload gets enabled to preserve 1007 * backward compatibility.". 1008 * 1009 * Therefore, there is no need to send this CVQ command if the 1010 * driver also enables all supported offloads, which aligns with 1011 * the device's defaults. 1012 * 1013 * Note that the device's defaults can mismatch the driver's 1014 * configuration only at live migration. 1015 */ 1016 return 0; 1017 } 1018 1019 offloads = cpu_to_le64(n->curr_guest_offloads); 1020 const struct iovec data = { 1021 .iov_base = &offloads, 1022 .iov_len = sizeof(offloads), 1023 }; 1024 r = vhost_vdpa_net_load_cmd(s, out_cursor, in_cursor, 1025 VIRTIO_NET_CTRL_GUEST_OFFLOADS, 1026 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, 1027 &data, 1); 1028 if (unlikely(r < 0)) { 1029 return r; 1030 } 1031 1032 return 0; 1033 } 1034 1035 static int vhost_vdpa_net_load_rx_mode(VhostVDPAState *s, 1036 struct iovec *out_cursor, 1037 struct iovec *in_cursor, 1038 uint8_t cmd, 1039 uint8_t on) 1040 { 1041 const struct iovec data = { 1042 .iov_base = &on, 1043 .iov_len = sizeof(on), 1044 }; 1045 ssize_t r; 1046 1047 r = vhost_vdpa_net_load_cmd(s, out_cursor, in_cursor, 1048 VIRTIO_NET_CTRL_RX, cmd, &data, 1); 1049 if (unlikely(r < 0)) { 1050 return r; 1051 } 1052 1053 return 0; 1054 } 1055 1056 static int vhost_vdpa_net_load_rx(VhostVDPAState *s, 1057 const VirtIONet *n, 1058 struct iovec *out_cursor, 1059 struct iovec *in_cursor) 1060 { 1061 ssize_t r; 1062 1063 if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX)) { 1064 return 0; 1065 } 1066 1067 /* 1068 * According to virtio_net_reset(), device turns promiscuous mode 1069 * on by default. 1070 * 1071 * Additionally, according to VirtIO standard, "Since there are 1072 * no guarantees, it can use a hash filter or silently switch to 1073 * allmulti or promiscuous mode if it is given too many addresses.". 1074 * QEMU marks `n->mac_table.uni_overflow` if guest sets too many 1075 * non-multicast MAC addresses, indicating that promiscuous mode 1076 * should be enabled. 1077 * 1078 * Therefore, QEMU should only send this CVQ command if the 1079 * `n->mac_table.uni_overflow` is not marked and `n->promisc` is off, 1080 * which sets promiscuous mode on, different from the device's defaults. 1081 * 1082 * Note that the device's defaults can mismatch the driver's 1083 * configuration only at live migration. 1084 */ 1085 if (!n->mac_table.uni_overflow && !n->promisc) { 1086 r = vhost_vdpa_net_load_rx_mode(s, out_cursor, in_cursor, 1087 VIRTIO_NET_CTRL_RX_PROMISC, 0); 1088 if (unlikely(r < 0)) { 1089 return r; 1090 } 1091 } 1092 1093 /* 1094 * According to virtio_net_reset(), device turns all-multicast mode 1095 * off by default. 1096 * 1097 * According to VirtIO standard, "Since there are no guarantees, 1098 * it can use a hash filter or silently switch to allmulti or 1099 * promiscuous mode if it is given too many addresses.". QEMU marks 1100 * `n->mac_table.multi_overflow` if guest sets too many 1101 * non-multicast MAC addresses. 1102 * 1103 * Therefore, QEMU should only send this CVQ command if the 1104 * `n->mac_table.multi_overflow` is marked or `n->allmulti` is on, 1105 * which sets all-multicast mode on, different from the device's defaults. 1106 * 1107 * Note that the device's defaults can mismatch the driver's 1108 * configuration only at live migration. 1109 */ 1110 if (n->mac_table.multi_overflow || n->allmulti) { 1111 r = vhost_vdpa_net_load_rx_mode(s, out_cursor, in_cursor, 1112 VIRTIO_NET_CTRL_RX_ALLMULTI, 1); 1113 if (unlikely(r < 0)) { 1114 return r; 1115 } 1116 } 1117 1118 if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX_EXTRA)) { 1119 return 0; 1120 } 1121 1122 /* 1123 * According to virtio_net_reset(), device turns all-unicast mode 1124 * off by default. 1125 * 1126 * Therefore, QEMU should only send this CVQ command if the driver 1127 * sets all-unicast mode on, different from the device's defaults. 1128 * 1129 * Note that the device's defaults can mismatch the driver's 1130 * configuration only at live migration. 1131 */ 1132 if (n->alluni) { 1133 r = vhost_vdpa_net_load_rx_mode(s, out_cursor, in_cursor, 1134 VIRTIO_NET_CTRL_RX_ALLUNI, 1); 1135 if (r < 0) { 1136 return r; 1137 } 1138 } 1139 1140 /* 1141 * According to virtio_net_reset(), device turns non-multicast mode 1142 * off by default. 1143 * 1144 * Therefore, QEMU should only send this CVQ command if the driver 1145 * sets non-multicast mode on, different from the device's defaults. 1146 * 1147 * Note that the device's defaults can mismatch the driver's 1148 * configuration only at live migration. 1149 */ 1150 if (n->nomulti) { 1151 r = vhost_vdpa_net_load_rx_mode(s, out_cursor, in_cursor, 1152 VIRTIO_NET_CTRL_RX_NOMULTI, 1); 1153 if (r < 0) { 1154 return r; 1155 } 1156 } 1157 1158 /* 1159 * According to virtio_net_reset(), device turns non-unicast mode 1160 * off by default. 1161 * 1162 * Therefore, QEMU should only send this CVQ command if the driver 1163 * sets non-unicast mode on, different from the device's defaults. 1164 * 1165 * Note that the device's defaults can mismatch the driver's 1166 * configuration only at live migration. 1167 */ 1168 if (n->nouni) { 1169 r = vhost_vdpa_net_load_rx_mode(s, out_cursor, in_cursor, 1170 VIRTIO_NET_CTRL_RX_NOUNI, 1); 1171 if (r < 0) { 1172 return r; 1173 } 1174 } 1175 1176 /* 1177 * According to virtio_net_reset(), device turns non-broadcast mode 1178 * off by default. 1179 * 1180 * Therefore, QEMU should only send this CVQ command if the driver 1181 * sets non-broadcast mode on, different from the device's defaults. 1182 * 1183 * Note that the device's defaults can mismatch the driver's 1184 * configuration only at live migration. 1185 */ 1186 if (n->nobcast) { 1187 r = vhost_vdpa_net_load_rx_mode(s, out_cursor, in_cursor, 1188 VIRTIO_NET_CTRL_RX_NOBCAST, 1); 1189 if (r < 0) { 1190 return r; 1191 } 1192 } 1193 1194 return 0; 1195 } 1196 1197 static int vhost_vdpa_net_load_single_vlan(VhostVDPAState *s, 1198 const VirtIONet *n, 1199 struct iovec *out_cursor, 1200 struct iovec *in_cursor, 1201 uint16_t vid) 1202 { 1203 const struct iovec data = { 1204 .iov_base = &vid, 1205 .iov_len = sizeof(vid), 1206 }; 1207 ssize_t r = vhost_vdpa_net_load_cmd(s, out_cursor, in_cursor, 1208 VIRTIO_NET_CTRL_VLAN, 1209 VIRTIO_NET_CTRL_VLAN_ADD, 1210 &data, 1); 1211 if (unlikely(r < 0)) { 1212 return r; 1213 } 1214 1215 return 0; 1216 } 1217 1218 static int vhost_vdpa_net_load_vlan(VhostVDPAState *s, 1219 const VirtIONet *n, 1220 struct iovec *out_cursor, 1221 struct iovec *in_cursor) 1222 { 1223 int r; 1224 1225 if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_VLAN)) { 1226 return 0; 1227 } 1228 1229 for (int i = 0; i < MAX_VLAN >> 5; i++) { 1230 for (int j = 0; n->vlans[i] && j <= 0x1f; j++) { 1231 if (n->vlans[i] & (1U << j)) { 1232 r = vhost_vdpa_net_load_single_vlan(s, n, out_cursor, 1233 in_cursor, (i << 5) + j); 1234 if (unlikely(r != 0)) { 1235 return r; 1236 } 1237 } 1238 } 1239 } 1240 1241 return 0; 1242 } 1243 1244 static int vhost_vdpa_net_cvq_load(NetClientState *nc) 1245 { 1246 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 1247 struct vhost_vdpa *v = &s->vhost_vdpa; 1248 const VirtIONet *n; 1249 int r; 1250 struct iovec out_cursor, in_cursor; 1251 1252 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 1253 1254 r = vhost_vdpa_set_vring_ready(v, v->dev->vq_index); 1255 if (unlikely(r < 0)) { 1256 return r; 1257 } 1258 1259 if (v->shadow_vqs_enabled) { 1260 n = VIRTIO_NET(v->dev->vdev); 1261 vhost_vdpa_net_load_cursor_reset(s, &out_cursor, &in_cursor); 1262 r = vhost_vdpa_net_load_mac(s, n, &out_cursor, &in_cursor); 1263 if (unlikely(r < 0)) { 1264 return r; 1265 } 1266 r = vhost_vdpa_net_load_mq(s, n, &out_cursor, &in_cursor); 1267 if (unlikely(r)) { 1268 return r; 1269 } 1270 r = vhost_vdpa_net_load_offloads(s, n, &out_cursor, &in_cursor); 1271 if (unlikely(r)) { 1272 return r; 1273 } 1274 r = vhost_vdpa_net_load_rx(s, n, &out_cursor, &in_cursor); 1275 if (unlikely(r)) { 1276 return r; 1277 } 1278 r = vhost_vdpa_net_load_vlan(s, n, &out_cursor, &in_cursor); 1279 if (unlikely(r)) { 1280 return r; 1281 } 1282 1283 /* 1284 * We need to poll and check all pending device's used buffers. 1285 * 1286 * We can poll here since we've had BQL from the time 1287 * we sent the descriptor. 1288 */ 1289 r = vhost_vdpa_net_svq_flush(s, in_cursor.iov_base - (void *)s->status); 1290 if (unlikely(r)) { 1291 return r; 1292 } 1293 } 1294 1295 for (int i = 0; i < v->dev->vq_index; ++i) { 1296 r = vhost_vdpa_set_vring_ready(v, i); 1297 if (unlikely(r < 0)) { 1298 return r; 1299 } 1300 } 1301 1302 return 0; 1303 } 1304 1305 static NetClientInfo net_vhost_vdpa_cvq_info = { 1306 .type = NET_CLIENT_DRIVER_VHOST_VDPA, 1307 .size = sizeof(VhostVDPAState), 1308 .receive = vhost_vdpa_receive, 1309 .start = vhost_vdpa_net_cvq_start, 1310 .load = vhost_vdpa_net_cvq_load, 1311 .stop = vhost_vdpa_net_cvq_stop, 1312 .cleanup = vhost_vdpa_cleanup, 1313 .has_vnet_hdr = vhost_vdpa_has_vnet_hdr, 1314 .get_vnet_hash_supported_types = vhost_vdpa_get_vnet_hash_supported_types, 1315 .has_ufo = vhost_vdpa_has_ufo, 1316 .check_peer_type = vhost_vdpa_check_peer_type, 1317 .set_steering_ebpf = vhost_vdpa_set_steering_ebpf, 1318 }; 1319 1320 /* 1321 * Forward the excessive VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command to 1322 * vdpa device. 1323 * 1324 * Considering that QEMU cannot send the entire filter table to the 1325 * vdpa device, it should send the VIRTIO_NET_CTRL_RX_PROMISC CVQ 1326 * command to enable promiscuous mode to receive all packets, 1327 * according to VirtIO standard, "Since there are no guarantees, 1328 * it can use a hash filter or silently switch to allmulti or 1329 * promiscuous mode if it is given too many addresses.". 1330 * 1331 * Since QEMU ignores MAC addresses beyond `MAC_TABLE_ENTRIES` and 1332 * marks `n->mac_table.x_overflow` accordingly, it should have 1333 * the same effect on the device model to receive 1334 * (`MAC_TABLE_ENTRIES` + 1) or more non-multicast MAC addresses. 1335 * The same applies to multicast MAC addresses. 1336 * 1337 * Therefore, QEMU can provide the device model with a fake 1338 * VIRTIO_NET_CTRL_MAC_TABLE_SET command with (`MAC_TABLE_ENTRIES` + 1) 1339 * non-multicast MAC addresses and (`MAC_TABLE_ENTRIES` + 1) multicast 1340 * MAC addresses. This ensures that the device model marks 1341 * `n->mac_table.uni_overflow` and `n->mac_table.multi_overflow`, 1342 * allowing all packets to be received, which aligns with the 1343 * state of the vdpa device. 1344 */ 1345 static int vhost_vdpa_net_excessive_mac_filter_cvq_add(VhostVDPAState *s, 1346 VirtQueueElement *elem, 1347 struct iovec *out, 1348 const struct iovec *in) 1349 { 1350 struct virtio_net_ctrl_mac mac_data, *mac_ptr; 1351 struct virtio_net_ctrl_hdr *hdr_ptr; 1352 uint32_t cursor; 1353 ssize_t r; 1354 uint8_t on = 1; 1355 1356 /* parse the non-multicast MAC address entries from CVQ command */ 1357 cursor = sizeof(*hdr_ptr); 1358 r = iov_to_buf(elem->out_sg, elem->out_num, cursor, 1359 &mac_data, sizeof(mac_data)); 1360 if (unlikely(r != sizeof(mac_data))) { 1361 /* 1362 * If the CVQ command is invalid, we should simulate the vdpa device 1363 * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command 1364 */ 1365 *s->status = VIRTIO_NET_ERR; 1366 return sizeof(*s->status); 1367 } 1368 cursor += sizeof(mac_data) + le32_to_cpu(mac_data.entries) * ETH_ALEN; 1369 1370 /* parse the multicast MAC address entries from CVQ command */ 1371 r = iov_to_buf(elem->out_sg, elem->out_num, cursor, 1372 &mac_data, sizeof(mac_data)); 1373 if (r != sizeof(mac_data)) { 1374 /* 1375 * If the CVQ command is invalid, we should simulate the vdpa device 1376 * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command 1377 */ 1378 *s->status = VIRTIO_NET_ERR; 1379 return sizeof(*s->status); 1380 } 1381 cursor += sizeof(mac_data) + le32_to_cpu(mac_data.entries) * ETH_ALEN; 1382 1383 /* validate the CVQ command */ 1384 if (iov_size(elem->out_sg, elem->out_num) != cursor) { 1385 /* 1386 * If the CVQ command is invalid, we should simulate the vdpa device 1387 * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command 1388 */ 1389 *s->status = VIRTIO_NET_ERR; 1390 return sizeof(*s->status); 1391 } 1392 1393 /* 1394 * According to VirtIO standard, "Since there are no guarantees, 1395 * it can use a hash filter or silently switch to allmulti or 1396 * promiscuous mode if it is given too many addresses.". 1397 * 1398 * Therefore, considering that QEMU is unable to send the entire 1399 * filter table to the vdpa device, it should send the 1400 * VIRTIO_NET_CTRL_RX_PROMISC CVQ command to enable promiscuous mode 1401 */ 1402 hdr_ptr = out->iov_base; 1403 out->iov_len = sizeof(*hdr_ptr) + sizeof(on); 1404 1405 hdr_ptr->class = VIRTIO_NET_CTRL_RX; 1406 hdr_ptr->cmd = VIRTIO_NET_CTRL_RX_PROMISC; 1407 iov_from_buf(out, 1, sizeof(*hdr_ptr), &on, sizeof(on)); 1408 r = vhost_vdpa_net_cvq_add(s, out, 1, in, 1); 1409 if (unlikely(r < 0)) { 1410 return r; 1411 } 1412 1413 /* 1414 * We can poll here since we've had BQL from the time 1415 * we sent the descriptor. 1416 */ 1417 r = vhost_vdpa_net_svq_poll(s, 1); 1418 if (unlikely(r < sizeof(*s->status))) { 1419 return r; 1420 } 1421 if (*s->status != VIRTIO_NET_OK) { 1422 return sizeof(*s->status); 1423 } 1424 1425 /* 1426 * QEMU should also send a fake VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ 1427 * command to the device model, including (`MAC_TABLE_ENTRIES` + 1) 1428 * non-multicast MAC addresses and (`MAC_TABLE_ENTRIES` + 1) 1429 * multicast MAC addresses. 1430 * 1431 * By doing so, the device model can mark `n->mac_table.uni_overflow` 1432 * and `n->mac_table.multi_overflow`, enabling all packets to be 1433 * received, which aligns with the state of the vdpa device. 1434 */ 1435 cursor = 0; 1436 uint32_t fake_uni_entries = MAC_TABLE_ENTRIES + 1, 1437 fake_mul_entries = MAC_TABLE_ENTRIES + 1, 1438 fake_cvq_size = sizeof(struct virtio_net_ctrl_hdr) + 1439 sizeof(mac_data) + fake_uni_entries * ETH_ALEN + 1440 sizeof(mac_data) + fake_mul_entries * ETH_ALEN; 1441 1442 assert(fake_cvq_size < vhost_vdpa_net_cvq_cmd_page_len()); 1443 out->iov_len = fake_cvq_size; 1444 1445 /* pack the header for fake CVQ command */ 1446 hdr_ptr = out->iov_base + cursor; 1447 hdr_ptr->class = VIRTIO_NET_CTRL_MAC; 1448 hdr_ptr->cmd = VIRTIO_NET_CTRL_MAC_TABLE_SET; 1449 cursor += sizeof(*hdr_ptr); 1450 1451 /* 1452 * Pack the non-multicast MAC addresses part for fake CVQ command. 1453 * 1454 * According to virtio_net_handle_mac(), QEMU doesn't verify the MAC 1455 * addresses provided in CVQ command. Therefore, only the entries 1456 * field need to be prepared in the CVQ command. 1457 */ 1458 mac_ptr = out->iov_base + cursor; 1459 mac_ptr->entries = cpu_to_le32(fake_uni_entries); 1460 cursor += sizeof(*mac_ptr) + fake_uni_entries * ETH_ALEN; 1461 1462 /* 1463 * Pack the multicast MAC addresses part for fake CVQ command. 1464 * 1465 * According to virtio_net_handle_mac(), QEMU doesn't verify the MAC 1466 * addresses provided in CVQ command. Therefore, only the entries 1467 * field need to be prepared in the CVQ command. 1468 */ 1469 mac_ptr = out->iov_base + cursor; 1470 mac_ptr->entries = cpu_to_le32(fake_mul_entries); 1471 1472 /* 1473 * Simulating QEMU poll a vdpa device used buffer 1474 * for VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command 1475 */ 1476 return sizeof(*s->status); 1477 } 1478 1479 /** 1480 * Validate and copy control virtqueue commands. 1481 * 1482 * Following QEMU guidelines, we offer a copy of the buffers to the device to 1483 * prevent TOCTOU bugs. 1484 */ 1485 static int vhost_vdpa_net_handle_ctrl_avail(VhostShadowVirtqueue *svq, 1486 VirtQueueElement *elem, 1487 void *opaque) 1488 { 1489 VhostVDPAState *s = opaque; 1490 size_t in_len; 1491 const struct virtio_net_ctrl_hdr *ctrl; 1492 virtio_net_ctrl_ack status = VIRTIO_NET_ERR; 1493 /* Out buffer sent to both the vdpa device and the device model */ 1494 struct iovec out = { 1495 .iov_base = s->cvq_cmd_out_buffer, 1496 }; 1497 /* in buffer used for device model */ 1498 const struct iovec model_in = { 1499 .iov_base = &status, 1500 .iov_len = sizeof(status), 1501 }; 1502 /* in buffer used for vdpa device */ 1503 const struct iovec vdpa_in = { 1504 .iov_base = s->status, 1505 .iov_len = sizeof(*s->status), 1506 }; 1507 ssize_t dev_written = -EINVAL; 1508 1509 out.iov_len = iov_to_buf(elem->out_sg, elem->out_num, 0, 1510 s->cvq_cmd_out_buffer, 1511 vhost_vdpa_net_cvq_cmd_page_len()); 1512 1513 ctrl = s->cvq_cmd_out_buffer; 1514 if (ctrl->class == VIRTIO_NET_CTRL_ANNOUNCE) { 1515 /* 1516 * Guest announce capability is emulated by qemu, so don't forward to 1517 * the device. 1518 */ 1519 dev_written = sizeof(status); 1520 *s->status = VIRTIO_NET_OK; 1521 } else if (unlikely(ctrl->class == VIRTIO_NET_CTRL_MAC && 1522 ctrl->cmd == VIRTIO_NET_CTRL_MAC_TABLE_SET && 1523 iov_size(elem->out_sg, elem->out_num) > out.iov_len)) { 1524 /* 1525 * Due to the size limitation of the out buffer sent to the vdpa device, 1526 * which is determined by vhost_vdpa_net_cvq_cmd_page_len(), excessive 1527 * MAC addresses set by the driver for the filter table can cause 1528 * truncation of the CVQ command in QEMU. As a result, the vdpa device 1529 * rejects the flawed CVQ command. 1530 * 1531 * Therefore, QEMU must handle this situation instead of sending 1532 * the CVQ command directly. 1533 */ 1534 dev_written = vhost_vdpa_net_excessive_mac_filter_cvq_add(s, elem, 1535 &out, &vdpa_in); 1536 if (unlikely(dev_written < 0)) { 1537 goto out; 1538 } 1539 } else { 1540 ssize_t r; 1541 r = vhost_vdpa_net_cvq_add(s, &out, 1, &vdpa_in, 1); 1542 if (unlikely(r < 0)) { 1543 dev_written = r; 1544 goto out; 1545 } 1546 1547 /* 1548 * We can poll here since we've had BQL from the time 1549 * we sent the descriptor. 1550 */ 1551 dev_written = vhost_vdpa_net_svq_poll(s, 1); 1552 } 1553 1554 if (unlikely(dev_written < sizeof(status))) { 1555 error_report("Insufficient written data (%zu)", dev_written); 1556 goto out; 1557 } 1558 1559 if (*s->status != VIRTIO_NET_OK) { 1560 goto out; 1561 } 1562 1563 status = VIRTIO_NET_ERR; 1564 virtio_net_handle_ctrl_iov(svq->vdev, &model_in, 1, &out, 1); 1565 if (status != VIRTIO_NET_OK) { 1566 error_report("Bad CVQ processing in model"); 1567 } 1568 1569 out: 1570 in_len = iov_from_buf(elem->in_sg, elem->in_num, 0, &status, 1571 sizeof(status)); 1572 if (unlikely(in_len < sizeof(status))) { 1573 error_report("Bad device CVQ written length"); 1574 } 1575 vhost_svq_push_elem(svq, elem, MIN(in_len, sizeof(status))); 1576 /* 1577 * `elem` belongs to vhost_vdpa_net_handle_ctrl_avail() only when 1578 * the function successfully forwards the CVQ command, indicated 1579 * by a non-negative value of `dev_written`. Otherwise, it still 1580 * belongs to SVQ. 1581 * This function should only free the `elem` when it owns. 1582 */ 1583 if (dev_written >= 0) { 1584 g_free(elem); 1585 } 1586 return dev_written < 0 ? dev_written : 0; 1587 } 1588 1589 static const VhostShadowVirtqueueOps vhost_vdpa_net_svq_ops = { 1590 .avail_handler = vhost_vdpa_net_handle_ctrl_avail, 1591 }; 1592 1593 /** 1594 * Probe if CVQ is isolated 1595 * 1596 * @device_fd The vdpa device fd 1597 * @features Features offered by the device. 1598 * @cvq_index The control vq pair index 1599 * 1600 * Returns <0 in case of failure, 0 if false and 1 if true. 1601 */ 1602 static int vhost_vdpa_probe_cvq_isolation(int device_fd, uint64_t features, 1603 int cvq_index, Error **errp) 1604 { 1605 ERRP_GUARD(); 1606 uint64_t backend_features; 1607 int64_t cvq_group; 1608 uint8_t status = VIRTIO_CONFIG_S_ACKNOWLEDGE | 1609 VIRTIO_CONFIG_S_DRIVER; 1610 int r; 1611 1612 r = ioctl(device_fd, VHOST_GET_BACKEND_FEATURES, &backend_features); 1613 if (unlikely(r < 0)) { 1614 error_setg_errno(errp, errno, "Cannot get vdpa backend_features"); 1615 return r; 1616 } 1617 1618 if (!(backend_features & BIT_ULL(VHOST_BACKEND_F_IOTLB_ASID))) { 1619 return 0; 1620 } 1621 1622 r = ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status); 1623 if (unlikely(r)) { 1624 error_setg_errno(errp, -r, "Cannot set device status"); 1625 goto out; 1626 } 1627 1628 r = ioctl(device_fd, VHOST_SET_FEATURES, &features); 1629 if (unlikely(r)) { 1630 error_setg_errno(errp, -r, "Cannot set features"); 1631 goto out; 1632 } 1633 1634 status |= VIRTIO_CONFIG_S_FEATURES_OK; 1635 r = ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status); 1636 if (unlikely(r)) { 1637 error_setg_errno(errp, -r, "Cannot set device status"); 1638 goto out; 1639 } 1640 1641 cvq_group = vhost_vdpa_get_vring_group(device_fd, cvq_index, errp); 1642 if (unlikely(cvq_group < 0)) { 1643 if (cvq_group != -ENOTSUP) { 1644 r = cvq_group; 1645 goto out; 1646 } 1647 1648 /* 1649 * The kernel report VHOST_BACKEND_F_IOTLB_ASID if the vdpa frontend 1650 * support ASID even if the parent driver does not. The CVQ cannot be 1651 * isolated in this case. 1652 */ 1653 error_free(*errp); 1654 *errp = NULL; 1655 r = 0; 1656 goto out; 1657 } 1658 1659 for (int i = 0; i < cvq_index; ++i) { 1660 int64_t group = vhost_vdpa_get_vring_group(device_fd, i, errp); 1661 if (unlikely(group < 0)) { 1662 r = group; 1663 goto out; 1664 } 1665 1666 if (group == (int64_t)cvq_group) { 1667 r = 0; 1668 goto out; 1669 } 1670 } 1671 1672 r = 1; 1673 1674 out: 1675 status = 0; 1676 ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status); 1677 return r; 1678 } 1679 1680 static NetClientState *net_vhost_vdpa_init(NetClientState *peer, 1681 const char *device, 1682 const char *name, 1683 int vdpa_device_fd, 1684 int queue_pair_index, 1685 int nvqs, 1686 bool is_datapath, 1687 bool svq, 1688 struct vhost_vdpa_iova_range iova_range, 1689 uint64_t features, 1690 VhostVDPAShared *shared, 1691 Error **errp) 1692 { 1693 NetClientState *nc = NULL; 1694 VhostVDPAState *s; 1695 int ret = 0; 1696 assert(name); 1697 int cvq_isolated = 0; 1698 1699 if (is_datapath) { 1700 nc = qemu_new_net_client(&net_vhost_vdpa_info, peer, device, 1701 name); 1702 } else { 1703 cvq_isolated = vhost_vdpa_probe_cvq_isolation(vdpa_device_fd, features, 1704 queue_pair_index * 2, 1705 errp); 1706 if (unlikely(cvq_isolated < 0)) { 1707 return NULL; 1708 } 1709 1710 nc = qemu_new_net_control_client(&net_vhost_vdpa_cvq_info, peer, 1711 device, name); 1712 } 1713 qemu_set_info_str(nc, TYPE_VHOST_VDPA); 1714 s = DO_UPCAST(VhostVDPAState, nc, nc); 1715 1716 s->vhost_vdpa.index = queue_pair_index; 1717 s->always_svq = svq; 1718 s->migration_state.notify = NULL; 1719 s->vhost_vdpa.shadow_vqs_enabled = svq; 1720 if (queue_pair_index == 0) { 1721 vhost_vdpa_net_valid_svq_features(features, 1722 &s->vhost_vdpa.migration_blocker); 1723 s->vhost_vdpa.shared = g_new0(VhostVDPAShared, 1); 1724 s->vhost_vdpa.shared->device_fd = vdpa_device_fd; 1725 s->vhost_vdpa.shared->iova_range = iova_range; 1726 s->vhost_vdpa.shared->shadow_data = svq; 1727 s->vhost_vdpa.shared->iova_tree = vhost_iova_tree_new(iova_range.first, 1728 iova_range.last); 1729 } else if (!is_datapath) { 1730 s->cvq_cmd_out_buffer = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(), 1731 PROT_READ | PROT_WRITE, 1732 MAP_SHARED | MAP_ANONYMOUS, -1, 0); 1733 s->status = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(), 1734 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS, 1735 -1, 0); 1736 1737 s->vhost_vdpa.shadow_vq_ops = &vhost_vdpa_net_svq_ops; 1738 s->vhost_vdpa.shadow_vq_ops_opaque = s; 1739 s->cvq_isolated = cvq_isolated; 1740 } 1741 if (queue_pair_index != 0) { 1742 s->vhost_vdpa.shared = shared; 1743 } 1744 1745 ret = vhost_vdpa_add(nc, (void *)&s->vhost_vdpa, queue_pair_index, nvqs); 1746 if (ret) { 1747 qemu_del_net_client(nc); 1748 return NULL; 1749 } 1750 1751 return nc; 1752 } 1753 1754 static int vhost_vdpa_get_features(int fd, uint64_t *features, Error **errp) 1755 { 1756 int ret = ioctl(fd, VHOST_GET_FEATURES, features); 1757 if (unlikely(ret < 0)) { 1758 error_setg_errno(errp, errno, 1759 "Fail to query features from vhost-vDPA device"); 1760 } 1761 return ret; 1762 } 1763 1764 static int vhost_vdpa_get_max_queue_pairs(int fd, uint64_t features, 1765 int *has_cvq, Error **errp) 1766 { 1767 unsigned long config_size = offsetof(struct vhost_vdpa_config, buf); 1768 g_autofree struct vhost_vdpa_config *config = NULL; 1769 __virtio16 *max_queue_pairs; 1770 int ret; 1771 1772 if (features & (1 << VIRTIO_NET_F_CTRL_VQ)) { 1773 *has_cvq = 1; 1774 } else { 1775 *has_cvq = 0; 1776 } 1777 1778 if (features & (1 << VIRTIO_NET_F_MQ)) { 1779 config = g_malloc0(config_size + sizeof(*max_queue_pairs)); 1780 config->off = offsetof(struct virtio_net_config, max_virtqueue_pairs); 1781 config->len = sizeof(*max_queue_pairs); 1782 1783 ret = ioctl(fd, VHOST_VDPA_GET_CONFIG, config); 1784 if (ret) { 1785 error_setg(errp, "Fail to get config from vhost-vDPA device"); 1786 return -ret; 1787 } 1788 1789 max_queue_pairs = (__virtio16 *)&config->buf; 1790 1791 return lduw_le_p(max_queue_pairs); 1792 } 1793 1794 return 1; 1795 } 1796 1797 int net_init_vhost_vdpa(const Netdev *netdev, const char *name, 1798 NetClientState *peer, Error **errp) 1799 { 1800 ERRP_GUARD(); 1801 const NetdevVhostVDPAOptions *opts; 1802 uint64_t features; 1803 int vdpa_device_fd; 1804 g_autofree NetClientState **ncs = NULL; 1805 struct vhost_vdpa_iova_range iova_range; 1806 NetClientState *nc; 1807 int queue_pairs, r, i = 0, has_cvq = 0; 1808 1809 assert(netdev->type == NET_CLIENT_DRIVER_VHOST_VDPA); 1810 opts = &netdev->u.vhost_vdpa; 1811 if (!opts->vhostdev && !opts->vhostfd) { 1812 error_setg(errp, 1813 "vhost-vdpa: neither vhostdev= nor vhostfd= was specified"); 1814 return -1; 1815 } 1816 1817 if (opts->vhostdev && opts->vhostfd) { 1818 error_setg(errp, 1819 "vhost-vdpa: vhostdev= and vhostfd= are mutually exclusive"); 1820 return -1; 1821 } 1822 1823 if (opts->vhostdev) { 1824 vdpa_device_fd = qemu_open(opts->vhostdev, O_RDWR, errp); 1825 if (vdpa_device_fd == -1) { 1826 return -errno; 1827 } 1828 } else { 1829 /* has_vhostfd */ 1830 vdpa_device_fd = monitor_fd_param(monitor_cur(), opts->vhostfd, errp); 1831 if (vdpa_device_fd == -1) { 1832 error_prepend(errp, "vhost-vdpa: unable to parse vhostfd: "); 1833 return -1; 1834 } 1835 } 1836 1837 r = vhost_vdpa_get_features(vdpa_device_fd, &features, errp); 1838 if (unlikely(r < 0)) { 1839 goto err; 1840 } 1841 1842 queue_pairs = vhost_vdpa_get_max_queue_pairs(vdpa_device_fd, features, 1843 &has_cvq, errp); 1844 if (queue_pairs < 0) { 1845 qemu_close(vdpa_device_fd); 1846 return queue_pairs; 1847 } 1848 1849 r = vhost_vdpa_get_iova_range(vdpa_device_fd, &iova_range); 1850 if (unlikely(r < 0)) { 1851 error_setg(errp, "vhost-vdpa: get iova range failed: %s", 1852 strerror(-r)); 1853 goto err; 1854 } 1855 1856 if (opts->x_svq && !vhost_vdpa_net_valid_svq_features(features, errp)) { 1857 goto err; 1858 } 1859 1860 ncs = g_malloc0(sizeof(*ncs) * queue_pairs); 1861 1862 for (i = 0; i < queue_pairs; i++) { 1863 VhostVDPAShared *shared = NULL; 1864 1865 if (i) { 1866 shared = DO_UPCAST(VhostVDPAState, nc, ncs[0])->vhost_vdpa.shared; 1867 } 1868 ncs[i] = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name, 1869 vdpa_device_fd, i, 2, true, opts->x_svq, 1870 iova_range, features, shared, errp); 1871 if (!ncs[i]) 1872 goto err; 1873 } 1874 1875 if (has_cvq) { 1876 VhostVDPAState *s0 = DO_UPCAST(VhostVDPAState, nc, ncs[0]); 1877 VhostVDPAShared *shared = s0->vhost_vdpa.shared; 1878 1879 nc = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name, 1880 vdpa_device_fd, i, 1, false, 1881 opts->x_svq, iova_range, features, shared, 1882 errp); 1883 if (!nc) 1884 goto err; 1885 } 1886 1887 return 0; 1888 1889 err: 1890 if (i) { 1891 for (i--; i >= 0; i--) { 1892 qemu_del_net_client(ncs[i]); 1893 } 1894 } 1895 1896 qemu_close(vdpa_device_fd); 1897 1898 return -1; 1899 } 1900