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