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 Notifier 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 const int vdpa_feature_bits[] = { 54 VIRTIO_F_NOTIFY_ON_EMPTY, 55 VIRTIO_RING_F_INDIRECT_DESC, 56 VIRTIO_RING_F_EVENT_IDX, 57 VIRTIO_F_ANY_LAYOUT, 58 VIRTIO_F_VERSION_1, 59 VIRTIO_NET_F_CSUM, 60 VIRTIO_NET_F_GUEST_CSUM, 61 VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, 62 VIRTIO_NET_F_GSO, 63 VIRTIO_NET_F_GUEST_TSO4, 64 VIRTIO_NET_F_GUEST_TSO6, 65 VIRTIO_NET_F_GUEST_ECN, 66 VIRTIO_NET_F_GUEST_UFO, 67 VIRTIO_NET_F_HOST_TSO4, 68 VIRTIO_NET_F_HOST_TSO6, 69 VIRTIO_NET_F_HOST_ECN, 70 VIRTIO_NET_F_HOST_UFO, 71 VIRTIO_NET_F_MRG_RXBUF, 72 VIRTIO_NET_F_MTU, 73 VIRTIO_NET_F_CTRL_RX, 74 VIRTIO_NET_F_CTRL_RX_EXTRA, 75 VIRTIO_NET_F_CTRL_VLAN, 76 VIRTIO_NET_F_CTRL_MAC_ADDR, 77 VIRTIO_NET_F_RSS, 78 VIRTIO_NET_F_MQ, 79 VIRTIO_NET_F_CTRL_VQ, 80 VIRTIO_F_IOMMU_PLATFORM, 81 VIRTIO_F_RING_PACKED, 82 VIRTIO_F_RING_RESET, 83 VIRTIO_NET_F_RSS, 84 VIRTIO_NET_F_HASH_REPORT, 85 VIRTIO_NET_F_STATUS, 86 VHOST_INVALID_FEATURE_BIT 87 }; 88 89 /** Supported device specific feature bits with SVQ */ 90 static const uint64_t vdpa_svq_device_features = 91 BIT_ULL(VIRTIO_NET_F_CSUM) | 92 BIT_ULL(VIRTIO_NET_F_GUEST_CSUM) | 93 BIT_ULL(VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) | 94 BIT_ULL(VIRTIO_NET_F_MTU) | 95 BIT_ULL(VIRTIO_NET_F_MAC) | 96 BIT_ULL(VIRTIO_NET_F_GUEST_TSO4) | 97 BIT_ULL(VIRTIO_NET_F_GUEST_TSO6) | 98 BIT_ULL(VIRTIO_NET_F_GUEST_ECN) | 99 BIT_ULL(VIRTIO_NET_F_GUEST_UFO) | 100 BIT_ULL(VIRTIO_NET_F_HOST_TSO4) | 101 BIT_ULL(VIRTIO_NET_F_HOST_TSO6) | 102 BIT_ULL(VIRTIO_NET_F_HOST_ECN) | 103 BIT_ULL(VIRTIO_NET_F_HOST_UFO) | 104 BIT_ULL(VIRTIO_NET_F_MRG_RXBUF) | 105 BIT_ULL(VIRTIO_NET_F_STATUS) | 106 BIT_ULL(VIRTIO_NET_F_CTRL_VQ) | 107 BIT_ULL(VIRTIO_NET_F_MQ) | 108 BIT_ULL(VIRTIO_F_ANY_LAYOUT) | 109 BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR) | 110 /* VHOST_F_LOG_ALL is exposed by SVQ */ 111 BIT_ULL(VHOST_F_LOG_ALL) | 112 BIT_ULL(VIRTIO_NET_F_RSC_EXT) | 113 BIT_ULL(VIRTIO_NET_F_STANDBY) | 114 BIT_ULL(VIRTIO_NET_F_SPEED_DUPLEX); 115 116 #define VHOST_VDPA_NET_CVQ_ASID 1 117 118 VHostNetState *vhost_vdpa_get_vhost_net(NetClientState *nc) 119 { 120 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 121 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 122 return s->vhost_net; 123 } 124 125 static size_t vhost_vdpa_net_cvq_cmd_len(void) 126 { 127 /* 128 * MAC_TABLE_SET is the ctrl command that produces the longer out buffer. 129 * In buffer is always 1 byte, so it should fit here 130 */ 131 return sizeof(struct virtio_net_ctrl_hdr) + 132 2 * sizeof(struct virtio_net_ctrl_mac) + 133 MAC_TABLE_ENTRIES * ETH_ALEN; 134 } 135 136 static size_t vhost_vdpa_net_cvq_cmd_page_len(void) 137 { 138 return ROUND_UP(vhost_vdpa_net_cvq_cmd_len(), qemu_real_host_page_size()); 139 } 140 141 static bool vhost_vdpa_net_valid_svq_features(uint64_t features, Error **errp) 142 { 143 uint64_t invalid_dev_features = 144 features & ~vdpa_svq_device_features & 145 /* Transport are all accepted at this point */ 146 ~MAKE_64BIT_MASK(VIRTIO_TRANSPORT_F_START, 147 VIRTIO_TRANSPORT_F_END - VIRTIO_TRANSPORT_F_START); 148 149 if (invalid_dev_features) { 150 error_setg(errp, "vdpa svq does not work with features 0x%" PRIx64, 151 invalid_dev_features); 152 return false; 153 } 154 155 return vhost_svq_valid_features(features, errp); 156 } 157 158 static int vhost_vdpa_net_check_device_id(struct vhost_net *net) 159 { 160 uint32_t device_id; 161 int ret; 162 struct vhost_dev *hdev; 163 164 hdev = (struct vhost_dev *)&net->dev; 165 ret = hdev->vhost_ops->vhost_get_device_id(hdev, &device_id); 166 if (device_id != VIRTIO_ID_NET) { 167 return -ENOTSUP; 168 } 169 return ret; 170 } 171 172 static int vhost_vdpa_add(NetClientState *ncs, void *be, 173 int queue_pair_index, int nvqs) 174 { 175 VhostNetOptions options; 176 struct vhost_net *net = NULL; 177 VhostVDPAState *s; 178 int ret; 179 180 options.backend_type = VHOST_BACKEND_TYPE_VDPA; 181 assert(ncs->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 182 s = DO_UPCAST(VhostVDPAState, nc, ncs); 183 options.net_backend = ncs; 184 options.opaque = be; 185 options.busyloop_timeout = 0; 186 options.nvqs = nvqs; 187 188 net = vhost_net_init(&options); 189 if (!net) { 190 error_report("failed to init vhost_net for queue"); 191 goto err_init; 192 } 193 s->vhost_net = net; 194 ret = vhost_vdpa_net_check_device_id(net); 195 if (ret) { 196 goto err_check; 197 } 198 return 0; 199 err_check: 200 vhost_net_cleanup(net); 201 g_free(net); 202 err_init: 203 return -1; 204 } 205 206 static void vhost_vdpa_cleanup(NetClientState *nc) 207 { 208 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 209 210 munmap(s->cvq_cmd_out_buffer, vhost_vdpa_net_cvq_cmd_page_len()); 211 munmap(s->status, vhost_vdpa_net_cvq_cmd_page_len()); 212 if (s->vhost_net) { 213 vhost_net_cleanup(s->vhost_net); 214 g_free(s->vhost_net); 215 s->vhost_net = NULL; 216 } 217 if (s->vhost_vdpa.device_fd >= 0) { 218 qemu_close(s->vhost_vdpa.device_fd); 219 s->vhost_vdpa.device_fd = -1; 220 } 221 } 222 223 static bool vhost_vdpa_has_vnet_hdr(NetClientState *nc) 224 { 225 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 226 227 return true; 228 } 229 230 static bool vhost_vdpa_has_ufo(NetClientState *nc) 231 { 232 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 233 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 234 uint64_t features = 0; 235 features |= (1ULL << VIRTIO_NET_F_HOST_UFO); 236 features = vhost_net_get_features(s->vhost_net, features); 237 return !!(features & (1ULL << VIRTIO_NET_F_HOST_UFO)); 238 239 } 240 241 static bool vhost_vdpa_check_peer_type(NetClientState *nc, ObjectClass *oc, 242 Error **errp) 243 { 244 const char *driver = object_class_get_name(oc); 245 246 if (!g_str_has_prefix(driver, "virtio-net-")) { 247 error_setg(errp, "vhost-vdpa requires frontend driver virtio-net-*"); 248 return false; 249 } 250 251 return true; 252 } 253 254 /** Dummy receive in case qemu falls back to userland tap networking */ 255 static ssize_t vhost_vdpa_receive(NetClientState *nc, const uint8_t *buf, 256 size_t size) 257 { 258 return size; 259 } 260 261 /** From any vdpa net client, get the netclient of the first queue pair */ 262 static VhostVDPAState *vhost_vdpa_net_first_nc_vdpa(VhostVDPAState *s) 263 { 264 NICState *nic = qemu_get_nic(s->nc.peer); 265 NetClientState *nc0 = qemu_get_peer(nic->ncs, 0); 266 267 return DO_UPCAST(VhostVDPAState, nc, nc0); 268 } 269 270 static void vhost_vdpa_net_log_global_enable(VhostVDPAState *s, bool enable) 271 { 272 struct vhost_vdpa *v = &s->vhost_vdpa; 273 VirtIONet *n; 274 VirtIODevice *vdev; 275 int data_queue_pairs, cvq, r; 276 277 /* We are only called on the first data vqs and only if x-svq is not set */ 278 if (s->vhost_vdpa.shadow_vqs_enabled == enable) { 279 return; 280 } 281 282 vdev = v->dev->vdev; 283 n = VIRTIO_NET(vdev); 284 if (!n->vhost_started) { 285 return; 286 } 287 288 data_queue_pairs = n->multiqueue ? n->max_queue_pairs : 1; 289 cvq = virtio_vdev_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) ? 290 n->max_ncs - n->max_queue_pairs : 0; 291 /* 292 * TODO: vhost_net_stop does suspend, get_base and reset. We can be smarter 293 * in the future and resume the device if read-only operations between 294 * suspend and reset goes wrong. 295 */ 296 vhost_net_stop(vdev, n->nic->ncs, data_queue_pairs, cvq); 297 298 /* Start will check migration setup_or_active to configure or not SVQ */ 299 r = vhost_net_start(vdev, n->nic->ncs, data_queue_pairs, cvq); 300 if (unlikely(r < 0)) { 301 error_report("unable to start vhost net: %s(%d)", g_strerror(-r), -r); 302 } 303 } 304 305 static void vdpa_net_migration_state_notifier(Notifier *notifier, void *data) 306 { 307 MigrationState *migration = data; 308 VhostVDPAState *s = container_of(notifier, VhostVDPAState, 309 migration_state); 310 311 if (migration_in_setup(migration)) { 312 vhost_vdpa_net_log_global_enable(s, true); 313 } else if (migration_has_failed(migration)) { 314 vhost_vdpa_net_log_global_enable(s, false); 315 } 316 } 317 318 static void vhost_vdpa_net_data_start_first(VhostVDPAState *s) 319 { 320 struct vhost_vdpa *v = &s->vhost_vdpa; 321 322 add_migration_state_change_notifier(&s->migration_state); 323 if (v->shadow_vqs_enabled) { 324 v->iova_tree = vhost_iova_tree_new(v->iova_range.first, 325 v->iova_range.last); 326 } 327 } 328 329 static int vhost_vdpa_net_data_start(NetClientState *nc) 330 { 331 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 332 struct vhost_vdpa *v = &s->vhost_vdpa; 333 334 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 335 336 if (s->always_svq || 337 migration_is_setup_or_active(migrate_get_current()->state)) { 338 v->shadow_vqs_enabled = true; 339 v->shadow_data = true; 340 } else { 341 v->shadow_vqs_enabled = false; 342 v->shadow_data = false; 343 } 344 345 if (v->index == 0) { 346 vhost_vdpa_net_data_start_first(s); 347 return 0; 348 } 349 350 if (v->shadow_vqs_enabled) { 351 VhostVDPAState *s0 = vhost_vdpa_net_first_nc_vdpa(s); 352 v->iova_tree = s0->vhost_vdpa.iova_tree; 353 } 354 355 return 0; 356 } 357 358 static void vhost_vdpa_net_client_stop(NetClientState *nc) 359 { 360 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 361 struct vhost_dev *dev; 362 363 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 364 365 if (s->vhost_vdpa.index == 0) { 366 remove_migration_state_change_notifier(&s->migration_state); 367 } 368 369 dev = s->vhost_vdpa.dev; 370 if (dev->vq_index + dev->nvqs == dev->vq_index_end) { 371 g_clear_pointer(&s->vhost_vdpa.iova_tree, vhost_iova_tree_delete); 372 } 373 } 374 375 static NetClientInfo net_vhost_vdpa_info = { 376 .type = NET_CLIENT_DRIVER_VHOST_VDPA, 377 .size = sizeof(VhostVDPAState), 378 .receive = vhost_vdpa_receive, 379 .start = vhost_vdpa_net_data_start, 380 .stop = vhost_vdpa_net_client_stop, 381 .cleanup = vhost_vdpa_cleanup, 382 .has_vnet_hdr = vhost_vdpa_has_vnet_hdr, 383 .has_ufo = vhost_vdpa_has_ufo, 384 .check_peer_type = vhost_vdpa_check_peer_type, 385 }; 386 387 static int64_t vhost_vdpa_get_vring_group(int device_fd, unsigned vq_index, 388 Error **errp) 389 { 390 struct vhost_vring_state state = { 391 .index = vq_index, 392 }; 393 int r = ioctl(device_fd, VHOST_VDPA_GET_VRING_GROUP, &state); 394 395 if (unlikely(r < 0)) { 396 r = -errno; 397 error_setg_errno(errp, errno, "Cannot get VQ %u group", vq_index); 398 return r; 399 } 400 401 return state.num; 402 } 403 404 static int vhost_vdpa_set_address_space_id(struct vhost_vdpa *v, 405 unsigned vq_group, 406 unsigned asid_num) 407 { 408 struct vhost_vring_state asid = { 409 .index = vq_group, 410 .num = asid_num, 411 }; 412 int r; 413 414 r = ioctl(v->device_fd, VHOST_VDPA_SET_GROUP_ASID, &asid); 415 if (unlikely(r < 0)) { 416 error_report("Can't set vq group %u asid %u, errno=%d (%s)", 417 asid.index, asid.num, errno, g_strerror(errno)); 418 } 419 return r; 420 } 421 422 static void vhost_vdpa_cvq_unmap_buf(struct vhost_vdpa *v, void *addr) 423 { 424 VhostIOVATree *tree = v->iova_tree; 425 DMAMap needle = { 426 /* 427 * No need to specify size or to look for more translations since 428 * this contiguous chunk was allocated by us. 429 */ 430 .translated_addr = (hwaddr)(uintptr_t)addr, 431 }; 432 const DMAMap *map = vhost_iova_tree_find_iova(tree, &needle); 433 int r; 434 435 if (unlikely(!map)) { 436 error_report("Cannot locate expected map"); 437 return; 438 } 439 440 r = vhost_vdpa_dma_unmap(v, v->address_space_id, map->iova, map->size + 1); 441 if (unlikely(r != 0)) { 442 error_report("Device cannot unmap: %s(%d)", g_strerror(r), r); 443 } 444 445 vhost_iova_tree_remove(tree, *map); 446 } 447 448 /** Map CVQ buffer. */ 449 static int vhost_vdpa_cvq_map_buf(struct vhost_vdpa *v, void *buf, size_t size, 450 bool write) 451 { 452 DMAMap map = {}; 453 int r; 454 455 map.translated_addr = (hwaddr)(uintptr_t)buf; 456 map.size = size - 1; 457 map.perm = write ? IOMMU_RW : IOMMU_RO, 458 r = vhost_iova_tree_map_alloc(v->iova_tree, &map); 459 if (unlikely(r != IOVA_OK)) { 460 error_report("Cannot map injected element"); 461 return r; 462 } 463 464 r = vhost_vdpa_dma_map(v, v->address_space_id, map.iova, 465 vhost_vdpa_net_cvq_cmd_page_len(), buf, !write); 466 if (unlikely(r < 0)) { 467 goto dma_map_err; 468 } 469 470 return 0; 471 472 dma_map_err: 473 vhost_iova_tree_remove(v->iova_tree, map); 474 return r; 475 } 476 477 static int vhost_vdpa_net_cvq_start(NetClientState *nc) 478 { 479 VhostVDPAState *s, *s0; 480 struct vhost_vdpa *v; 481 int64_t cvq_group; 482 int r; 483 Error *err = NULL; 484 485 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 486 487 s = DO_UPCAST(VhostVDPAState, nc, nc); 488 v = &s->vhost_vdpa; 489 490 s0 = vhost_vdpa_net_first_nc_vdpa(s); 491 v->shadow_data = s0->vhost_vdpa.shadow_vqs_enabled; 492 v->shadow_vqs_enabled = s->always_svq; 493 s->vhost_vdpa.address_space_id = VHOST_VDPA_GUEST_PA_ASID; 494 495 if (s->vhost_vdpa.shadow_data) { 496 /* SVQ is already configured for all virtqueues */ 497 goto out; 498 } 499 500 /* 501 * If we early return in these cases SVQ will not be enabled. The migration 502 * will be blocked as long as vhost-vdpa backends will not offer _F_LOG. 503 */ 504 if (!vhost_vdpa_net_valid_svq_features(v->dev->features, NULL)) { 505 return 0; 506 } 507 508 if (!s->cvq_isolated) { 509 return 0; 510 } 511 512 cvq_group = vhost_vdpa_get_vring_group(v->device_fd, 513 v->dev->vq_index_end - 1, 514 &err); 515 if (unlikely(cvq_group < 0)) { 516 error_report_err(err); 517 return cvq_group; 518 } 519 520 r = vhost_vdpa_set_address_space_id(v, cvq_group, VHOST_VDPA_NET_CVQ_ASID); 521 if (unlikely(r < 0)) { 522 return r; 523 } 524 525 v->shadow_vqs_enabled = true; 526 s->vhost_vdpa.address_space_id = VHOST_VDPA_NET_CVQ_ASID; 527 528 out: 529 if (!s->vhost_vdpa.shadow_vqs_enabled) { 530 return 0; 531 } 532 533 if (s0->vhost_vdpa.iova_tree) { 534 /* 535 * SVQ is already configured for all virtqueues. Reuse IOVA tree for 536 * simplicity, whether CVQ shares ASID with guest or not, because: 537 * - Memory listener need access to guest's memory addresses allocated 538 * in the IOVA tree. 539 * - There should be plenty of IOVA address space for both ASID not to 540 * worry about collisions between them. Guest's translations are 541 * still validated with virtio virtqueue_pop so there is no risk for 542 * the guest to access memory that it shouldn't. 543 * 544 * To allocate a iova tree per ASID is doable but it complicates the 545 * code and it is not worth it for the moment. 546 */ 547 v->iova_tree = s0->vhost_vdpa.iova_tree; 548 } else { 549 v->iova_tree = vhost_iova_tree_new(v->iova_range.first, 550 v->iova_range.last); 551 } 552 553 r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer, 554 vhost_vdpa_net_cvq_cmd_page_len(), false); 555 if (unlikely(r < 0)) { 556 return r; 557 } 558 559 r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->status, 560 vhost_vdpa_net_cvq_cmd_page_len(), true); 561 if (unlikely(r < 0)) { 562 vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer); 563 } 564 565 return r; 566 } 567 568 static void vhost_vdpa_net_cvq_stop(NetClientState *nc) 569 { 570 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 571 572 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 573 574 if (s->vhost_vdpa.shadow_vqs_enabled) { 575 vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer); 576 vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->status); 577 } 578 579 vhost_vdpa_net_client_stop(nc); 580 } 581 582 static ssize_t vhost_vdpa_net_cvq_add(VhostVDPAState *s, size_t out_len, 583 size_t in_len) 584 { 585 /* Buffers for the device */ 586 const struct iovec out = { 587 .iov_base = s->cvq_cmd_out_buffer, 588 .iov_len = out_len, 589 }; 590 const struct iovec in = { 591 .iov_base = s->status, 592 .iov_len = sizeof(virtio_net_ctrl_ack), 593 }; 594 VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0); 595 int r; 596 597 r = vhost_svq_add(svq, &out, 1, &in, 1, NULL); 598 if (unlikely(r != 0)) { 599 if (unlikely(r == -ENOSPC)) { 600 qemu_log_mask(LOG_GUEST_ERROR, "%s: No space on device queue\n", 601 __func__); 602 } 603 return r; 604 } 605 606 /* 607 * We can poll here since we've had BQL from the time we sent the 608 * descriptor. Also, we need to take the answer before SVQ pulls by itself, 609 * when BQL is released 610 */ 611 return vhost_svq_poll(svq); 612 } 613 614 static ssize_t vhost_vdpa_net_load_cmd(VhostVDPAState *s, uint8_t class, 615 uint8_t cmd, const void *data, 616 size_t data_size) 617 { 618 const struct virtio_net_ctrl_hdr ctrl = { 619 .class = class, 620 .cmd = cmd, 621 }; 622 623 assert(data_size < vhost_vdpa_net_cvq_cmd_page_len() - sizeof(ctrl)); 624 625 memcpy(s->cvq_cmd_out_buffer, &ctrl, sizeof(ctrl)); 626 memcpy(s->cvq_cmd_out_buffer + sizeof(ctrl), data, data_size); 627 628 return vhost_vdpa_net_cvq_add(s, sizeof(ctrl) + data_size, 629 sizeof(virtio_net_ctrl_ack)); 630 } 631 632 static int vhost_vdpa_net_load_mac(VhostVDPAState *s, const VirtIONet *n) 633 { 634 if (virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_MAC_ADDR)) { 635 ssize_t dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MAC, 636 VIRTIO_NET_CTRL_MAC_ADDR_SET, 637 n->mac, sizeof(n->mac)); 638 if (unlikely(dev_written < 0)) { 639 return dev_written; 640 } 641 642 return *s->status != VIRTIO_NET_OK; 643 } 644 645 return 0; 646 } 647 648 static int vhost_vdpa_net_load_mq(VhostVDPAState *s, 649 const VirtIONet *n) 650 { 651 struct virtio_net_ctrl_mq mq; 652 ssize_t dev_written; 653 654 if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_MQ)) { 655 return 0; 656 } 657 658 mq.virtqueue_pairs = cpu_to_le16(n->curr_queue_pairs); 659 dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MQ, 660 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &mq, 661 sizeof(mq)); 662 if (unlikely(dev_written < 0)) { 663 return dev_written; 664 } 665 666 return *s->status != VIRTIO_NET_OK; 667 } 668 669 static int vhost_vdpa_net_load_offloads(VhostVDPAState *s, 670 const VirtIONet *n) 671 { 672 uint64_t offloads; 673 ssize_t dev_written; 674 675 if (!virtio_vdev_has_feature(&n->parent_obj, 676 VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)) { 677 return 0; 678 } 679 680 if (n->curr_guest_offloads == virtio_net_supported_guest_offloads(n)) { 681 /* 682 * According to VirtIO standard, "Upon feature negotiation 683 * corresponding offload gets enabled to preserve 684 * backward compatibility.". 685 * 686 * Therefore, there is no need to send this CVQ command if the 687 * driver also enables all supported offloads, which aligns with 688 * the device's defaults. 689 * 690 * Note that the device's defaults can mismatch the driver's 691 * configuration only at live migration. 692 */ 693 return 0; 694 } 695 696 offloads = cpu_to_le64(n->curr_guest_offloads); 697 dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_GUEST_OFFLOADS, 698 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, 699 &offloads, sizeof(offloads)); 700 if (unlikely(dev_written < 0)) { 701 return dev_written; 702 } 703 704 return *s->status != VIRTIO_NET_OK; 705 } 706 707 static int vhost_vdpa_net_load(NetClientState *nc) 708 { 709 VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 710 struct vhost_vdpa *v = &s->vhost_vdpa; 711 const VirtIONet *n; 712 int r; 713 714 assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 715 716 if (!v->shadow_vqs_enabled) { 717 return 0; 718 } 719 720 n = VIRTIO_NET(v->dev->vdev); 721 r = vhost_vdpa_net_load_mac(s, n); 722 if (unlikely(r < 0)) { 723 return r; 724 } 725 r = vhost_vdpa_net_load_mq(s, n); 726 if (unlikely(r)) { 727 return r; 728 } 729 r = vhost_vdpa_net_load_offloads(s, n); 730 if (unlikely(r)) { 731 return r; 732 } 733 734 return 0; 735 } 736 737 static NetClientInfo net_vhost_vdpa_cvq_info = { 738 .type = NET_CLIENT_DRIVER_VHOST_VDPA, 739 .size = sizeof(VhostVDPAState), 740 .receive = vhost_vdpa_receive, 741 .start = vhost_vdpa_net_cvq_start, 742 .load = vhost_vdpa_net_load, 743 .stop = vhost_vdpa_net_cvq_stop, 744 .cleanup = vhost_vdpa_cleanup, 745 .has_vnet_hdr = vhost_vdpa_has_vnet_hdr, 746 .has_ufo = vhost_vdpa_has_ufo, 747 .check_peer_type = vhost_vdpa_check_peer_type, 748 }; 749 750 /** 751 * Validate and copy control virtqueue commands. 752 * 753 * Following QEMU guidelines, we offer a copy of the buffers to the device to 754 * prevent TOCTOU bugs. 755 */ 756 static int vhost_vdpa_net_handle_ctrl_avail(VhostShadowVirtqueue *svq, 757 VirtQueueElement *elem, 758 void *opaque) 759 { 760 VhostVDPAState *s = opaque; 761 size_t in_len; 762 virtio_net_ctrl_ack status = VIRTIO_NET_ERR; 763 /* Out buffer sent to both the vdpa device and the device model */ 764 struct iovec out = { 765 .iov_base = s->cvq_cmd_out_buffer, 766 }; 767 /* in buffer used for device model */ 768 const struct iovec in = { 769 .iov_base = &status, 770 .iov_len = sizeof(status), 771 }; 772 ssize_t dev_written = -EINVAL; 773 774 out.iov_len = iov_to_buf(elem->out_sg, elem->out_num, 0, 775 s->cvq_cmd_out_buffer, 776 vhost_vdpa_net_cvq_cmd_len()); 777 if (*(uint8_t *)s->cvq_cmd_out_buffer == VIRTIO_NET_CTRL_ANNOUNCE) { 778 /* 779 * Guest announce capability is emulated by qemu, so don't forward to 780 * the device. 781 */ 782 dev_written = sizeof(status); 783 *s->status = VIRTIO_NET_OK; 784 } else { 785 dev_written = vhost_vdpa_net_cvq_add(s, out.iov_len, sizeof(status)); 786 if (unlikely(dev_written < 0)) { 787 goto out; 788 } 789 } 790 791 if (unlikely(dev_written < sizeof(status))) { 792 error_report("Insufficient written data (%zu)", dev_written); 793 goto out; 794 } 795 796 if (*s->status != VIRTIO_NET_OK) { 797 goto out; 798 } 799 800 status = VIRTIO_NET_ERR; 801 virtio_net_handle_ctrl_iov(svq->vdev, &in, 1, &out, 1); 802 if (status != VIRTIO_NET_OK) { 803 error_report("Bad CVQ processing in model"); 804 } 805 806 out: 807 in_len = iov_from_buf(elem->in_sg, elem->in_num, 0, &status, 808 sizeof(status)); 809 if (unlikely(in_len < sizeof(status))) { 810 error_report("Bad device CVQ written length"); 811 } 812 vhost_svq_push_elem(svq, elem, MIN(in_len, sizeof(status))); 813 g_free(elem); 814 return dev_written < 0 ? dev_written : 0; 815 } 816 817 static const VhostShadowVirtqueueOps vhost_vdpa_net_svq_ops = { 818 .avail_handler = vhost_vdpa_net_handle_ctrl_avail, 819 }; 820 821 /** 822 * Probe if CVQ is isolated 823 * 824 * @device_fd The vdpa device fd 825 * @features Features offered by the device. 826 * @cvq_index The control vq pair index 827 * 828 * Returns <0 in case of failure, 0 if false and 1 if true. 829 */ 830 static int vhost_vdpa_probe_cvq_isolation(int device_fd, uint64_t features, 831 int cvq_index, Error **errp) 832 { 833 uint64_t backend_features; 834 int64_t cvq_group; 835 uint8_t status = VIRTIO_CONFIG_S_ACKNOWLEDGE | 836 VIRTIO_CONFIG_S_DRIVER | 837 VIRTIO_CONFIG_S_FEATURES_OK; 838 int r; 839 840 ERRP_GUARD(); 841 842 r = ioctl(device_fd, VHOST_GET_BACKEND_FEATURES, &backend_features); 843 if (unlikely(r < 0)) { 844 error_setg_errno(errp, errno, "Cannot get vdpa backend_features"); 845 return r; 846 } 847 848 if (!(backend_features & BIT_ULL(VHOST_BACKEND_F_IOTLB_ASID))) { 849 return 0; 850 } 851 852 r = ioctl(device_fd, VHOST_SET_FEATURES, &features); 853 if (unlikely(r)) { 854 error_setg_errno(errp, errno, "Cannot set features"); 855 } 856 857 r = ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status); 858 if (unlikely(r)) { 859 error_setg_errno(errp, -r, "Cannot set device features"); 860 goto out; 861 } 862 863 cvq_group = vhost_vdpa_get_vring_group(device_fd, cvq_index, errp); 864 if (unlikely(cvq_group < 0)) { 865 if (cvq_group != -ENOTSUP) { 866 r = cvq_group; 867 goto out; 868 } 869 870 /* 871 * The kernel report VHOST_BACKEND_F_IOTLB_ASID if the vdpa frontend 872 * support ASID even if the parent driver does not. The CVQ cannot be 873 * isolated in this case. 874 */ 875 error_free(*errp); 876 *errp = NULL; 877 r = 0; 878 goto out; 879 } 880 881 for (int i = 0; i < cvq_index; ++i) { 882 int64_t group = vhost_vdpa_get_vring_group(device_fd, i, errp); 883 if (unlikely(group < 0)) { 884 r = group; 885 goto out; 886 } 887 888 if (group == (int64_t)cvq_group) { 889 r = 0; 890 goto out; 891 } 892 } 893 894 r = 1; 895 896 out: 897 status = 0; 898 ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status); 899 return r; 900 } 901 902 static NetClientState *net_vhost_vdpa_init(NetClientState *peer, 903 const char *device, 904 const char *name, 905 int vdpa_device_fd, 906 int queue_pair_index, 907 int nvqs, 908 bool is_datapath, 909 bool svq, 910 struct vhost_vdpa_iova_range iova_range, 911 uint64_t features, 912 Error **errp) 913 { 914 NetClientState *nc = NULL; 915 VhostVDPAState *s; 916 int ret = 0; 917 assert(name); 918 int cvq_isolated; 919 920 if (is_datapath) { 921 nc = qemu_new_net_client(&net_vhost_vdpa_info, peer, device, 922 name); 923 } else { 924 cvq_isolated = vhost_vdpa_probe_cvq_isolation(vdpa_device_fd, features, 925 queue_pair_index * 2, 926 errp); 927 if (unlikely(cvq_isolated < 0)) { 928 return NULL; 929 } 930 931 nc = qemu_new_net_control_client(&net_vhost_vdpa_cvq_info, peer, 932 device, name); 933 } 934 qemu_set_info_str(nc, TYPE_VHOST_VDPA); 935 s = DO_UPCAST(VhostVDPAState, nc, nc); 936 937 s->vhost_vdpa.device_fd = vdpa_device_fd; 938 s->vhost_vdpa.index = queue_pair_index; 939 s->always_svq = svq; 940 s->migration_state.notify = vdpa_net_migration_state_notifier; 941 s->vhost_vdpa.shadow_vqs_enabled = svq; 942 s->vhost_vdpa.iova_range = iova_range; 943 s->vhost_vdpa.shadow_data = svq; 944 if (queue_pair_index == 0) { 945 vhost_vdpa_net_valid_svq_features(features, 946 &s->vhost_vdpa.migration_blocker); 947 } else if (!is_datapath) { 948 s->cvq_cmd_out_buffer = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(), 949 PROT_READ | PROT_WRITE, 950 MAP_SHARED | MAP_ANONYMOUS, -1, 0); 951 s->status = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(), 952 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS, 953 -1, 0); 954 955 s->vhost_vdpa.shadow_vq_ops = &vhost_vdpa_net_svq_ops; 956 s->vhost_vdpa.shadow_vq_ops_opaque = s; 957 s->cvq_isolated = cvq_isolated; 958 959 /* 960 * TODO: We cannot migrate devices with CVQ and no x-svq enabled as 961 * there is no way to set the device state (MAC, MQ, etc) before 962 * starting the datapath. 963 * 964 * Migration blocker ownership now belongs to s->vhost_vdpa. 965 */ 966 if (!svq) { 967 error_setg(&s->vhost_vdpa.migration_blocker, 968 "net vdpa cannot migrate with CVQ feature"); 969 } 970 } 971 ret = vhost_vdpa_add(nc, (void *)&s->vhost_vdpa, queue_pair_index, nvqs); 972 if (ret) { 973 qemu_del_net_client(nc); 974 return NULL; 975 } 976 return nc; 977 } 978 979 static int vhost_vdpa_get_features(int fd, uint64_t *features, Error **errp) 980 { 981 int ret = ioctl(fd, VHOST_GET_FEATURES, features); 982 if (unlikely(ret < 0)) { 983 error_setg_errno(errp, errno, 984 "Fail to query features from vhost-vDPA device"); 985 } 986 return ret; 987 } 988 989 static int vhost_vdpa_get_max_queue_pairs(int fd, uint64_t features, 990 int *has_cvq, Error **errp) 991 { 992 unsigned long config_size = offsetof(struct vhost_vdpa_config, buf); 993 g_autofree struct vhost_vdpa_config *config = NULL; 994 __virtio16 *max_queue_pairs; 995 int ret; 996 997 if (features & (1 << VIRTIO_NET_F_CTRL_VQ)) { 998 *has_cvq = 1; 999 } else { 1000 *has_cvq = 0; 1001 } 1002 1003 if (features & (1 << VIRTIO_NET_F_MQ)) { 1004 config = g_malloc0(config_size + sizeof(*max_queue_pairs)); 1005 config->off = offsetof(struct virtio_net_config, max_virtqueue_pairs); 1006 config->len = sizeof(*max_queue_pairs); 1007 1008 ret = ioctl(fd, VHOST_VDPA_GET_CONFIG, config); 1009 if (ret) { 1010 error_setg(errp, "Fail to get config from vhost-vDPA device"); 1011 return -ret; 1012 } 1013 1014 max_queue_pairs = (__virtio16 *)&config->buf; 1015 1016 return lduw_le_p(max_queue_pairs); 1017 } 1018 1019 return 1; 1020 } 1021 1022 int net_init_vhost_vdpa(const Netdev *netdev, const char *name, 1023 NetClientState *peer, Error **errp) 1024 { 1025 const NetdevVhostVDPAOptions *opts; 1026 uint64_t features; 1027 int vdpa_device_fd; 1028 g_autofree NetClientState **ncs = NULL; 1029 struct vhost_vdpa_iova_range iova_range; 1030 NetClientState *nc; 1031 int queue_pairs, r, i = 0, has_cvq = 0; 1032 1033 assert(netdev->type == NET_CLIENT_DRIVER_VHOST_VDPA); 1034 opts = &netdev->u.vhost_vdpa; 1035 if (!opts->vhostdev && !opts->vhostfd) { 1036 error_setg(errp, 1037 "vhost-vdpa: neither vhostdev= nor vhostfd= was specified"); 1038 return -1; 1039 } 1040 1041 if (opts->vhostdev && opts->vhostfd) { 1042 error_setg(errp, 1043 "vhost-vdpa: vhostdev= and vhostfd= are mutually exclusive"); 1044 return -1; 1045 } 1046 1047 if (opts->vhostdev) { 1048 vdpa_device_fd = qemu_open(opts->vhostdev, O_RDWR, errp); 1049 if (vdpa_device_fd == -1) { 1050 return -errno; 1051 } 1052 } else { 1053 /* has_vhostfd */ 1054 vdpa_device_fd = monitor_fd_param(monitor_cur(), opts->vhostfd, errp); 1055 if (vdpa_device_fd == -1) { 1056 error_prepend(errp, "vhost-vdpa: unable to parse vhostfd: "); 1057 return -1; 1058 } 1059 } 1060 1061 r = vhost_vdpa_get_features(vdpa_device_fd, &features, errp); 1062 if (unlikely(r < 0)) { 1063 goto err; 1064 } 1065 1066 queue_pairs = vhost_vdpa_get_max_queue_pairs(vdpa_device_fd, features, 1067 &has_cvq, errp); 1068 if (queue_pairs < 0) { 1069 qemu_close(vdpa_device_fd); 1070 return queue_pairs; 1071 } 1072 1073 r = vhost_vdpa_get_iova_range(vdpa_device_fd, &iova_range); 1074 if (unlikely(r < 0)) { 1075 error_setg(errp, "vhost-vdpa: get iova range failed: %s", 1076 strerror(-r)); 1077 goto err; 1078 } 1079 1080 if (opts->x_svq && !vhost_vdpa_net_valid_svq_features(features, errp)) { 1081 goto err; 1082 } 1083 1084 ncs = g_malloc0(sizeof(*ncs) * queue_pairs); 1085 1086 for (i = 0; i < queue_pairs; i++) { 1087 ncs[i] = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name, 1088 vdpa_device_fd, i, 2, true, opts->x_svq, 1089 iova_range, features, errp); 1090 if (!ncs[i]) 1091 goto err; 1092 } 1093 1094 if (has_cvq) { 1095 nc = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name, 1096 vdpa_device_fd, i, 1, false, 1097 opts->x_svq, iova_range, features, errp); 1098 if (!nc) 1099 goto err; 1100 } 1101 1102 return 0; 1103 1104 err: 1105 if (i) { 1106 for (i--; i >= 0; i--) { 1107 qemu_del_net_client(ncs[i]); 1108 } 1109 } 1110 1111 qemu_close(vdpa_device_fd); 1112 1113 return -1; 1114 } 1115