1 /* 2 * vhost-vdpa 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 <linux/vhost.h> 14 #include <linux/vfio.h> 15 #include <sys/eventfd.h> 16 #include <sys/ioctl.h> 17 #include "exec/target_page.h" 18 #include "hw/virtio/vhost.h" 19 #include "hw/virtio/vhost-backend.h" 20 #include "hw/virtio/virtio-net.h" 21 #include "hw/virtio/vhost-shadow-virtqueue.h" 22 #include "hw/virtio/vhost-vdpa.h" 23 #include "exec/address-spaces.h" 24 #include "migration/blocker.h" 25 #include "qemu/cutils.h" 26 #include "qemu/main-loop.h" 27 #include "trace.h" 28 #include "qapi/error.h" 29 30 /* 31 * Return one past the end of the end of section. Be careful with uint64_t 32 * conversions! 33 */ 34 static Int128 vhost_vdpa_section_end(const MemoryRegionSection *section, 35 int page_mask) 36 { 37 Int128 llend = int128_make64(section->offset_within_address_space); 38 llend = int128_add(llend, section->size); 39 llend = int128_and(llend, int128_exts64(page_mask)); 40 41 return llend; 42 } 43 44 static bool vhost_vdpa_listener_skipped_section(MemoryRegionSection *section, 45 uint64_t iova_min, 46 uint64_t iova_max, 47 int page_mask) 48 { 49 Int128 llend; 50 bool is_ram = memory_region_is_ram(section->mr); 51 bool is_iommu = memory_region_is_iommu(section->mr); 52 bool is_protected = memory_region_is_protected(section->mr); 53 54 /* vhost-vDPA doesn't allow MMIO to be mapped */ 55 bool is_ram_device = memory_region_is_ram_device(section->mr); 56 57 if ((!is_ram && !is_iommu) || is_protected || is_ram_device) { 58 trace_vhost_vdpa_skipped_memory_section(is_ram, is_iommu, is_protected, 59 is_ram_device, iova_min, 60 iova_max, page_mask); 61 return true; 62 } 63 64 if (section->offset_within_address_space < iova_min) { 65 error_report("RAM section out of device range (min=0x%" PRIx64 66 ", addr=0x%" HWADDR_PRIx ")", 67 iova_min, section->offset_within_address_space); 68 return true; 69 } 70 /* 71 * While using vIOMMU, sometimes the section will be larger than iova_max, 72 * but the memory that actually maps is smaller, so move the check to 73 * function vhost_vdpa_iommu_map_notify(). That function will use the actual 74 * size that maps to the kernel 75 */ 76 77 if (!is_iommu) { 78 llend = vhost_vdpa_section_end(section, page_mask); 79 if (int128_gt(llend, int128_make64(iova_max))) { 80 error_report("RAM section out of device range (max=0x%" PRIx64 81 ", end addr=0x%" PRIx64 ")", 82 iova_max, int128_get64(llend)); 83 return true; 84 } 85 } 86 87 return false; 88 } 89 90 /* 91 * The caller must set asid = 0 if the device does not support asid. 92 * This is not an ABI break since it is set to 0 by the initializer anyway. 93 */ 94 int vhost_vdpa_dma_map(VhostVDPAShared *s, uint32_t asid, hwaddr iova, 95 hwaddr size, void *vaddr, bool readonly) 96 { 97 struct vhost_msg_v2 msg = {}; 98 int fd = s->device_fd; 99 int ret = 0; 100 101 msg.type = VHOST_IOTLB_MSG_V2; 102 msg.asid = asid; 103 msg.iotlb.iova = iova; 104 msg.iotlb.size = size; 105 msg.iotlb.uaddr = (uint64_t)(uintptr_t)vaddr; 106 msg.iotlb.perm = readonly ? VHOST_ACCESS_RO : VHOST_ACCESS_RW; 107 msg.iotlb.type = VHOST_IOTLB_UPDATE; 108 109 trace_vhost_vdpa_dma_map(s, fd, msg.type, msg.asid, msg.iotlb.iova, 110 msg.iotlb.size, msg.iotlb.uaddr, msg.iotlb.perm, 111 msg.iotlb.type); 112 113 if (write(fd, &msg, sizeof(msg)) != sizeof(msg)) { 114 error_report("failed to write, fd=%d, errno=%d (%s)", 115 fd, errno, strerror(errno)); 116 return -EIO ; 117 } 118 119 return ret; 120 } 121 122 /* 123 * The caller must set asid = 0 if the device does not support asid. 124 * This is not an ABI break since it is set to 0 by the initializer anyway. 125 */ 126 int vhost_vdpa_dma_unmap(VhostVDPAShared *s, uint32_t asid, hwaddr iova, 127 hwaddr size) 128 { 129 struct vhost_msg_v2 msg = {}; 130 int fd = s->device_fd; 131 int ret = 0; 132 133 msg.type = VHOST_IOTLB_MSG_V2; 134 msg.asid = asid; 135 msg.iotlb.iova = iova; 136 msg.iotlb.size = size; 137 msg.iotlb.type = VHOST_IOTLB_INVALIDATE; 138 139 trace_vhost_vdpa_dma_unmap(s, fd, msg.type, msg.asid, msg.iotlb.iova, 140 msg.iotlb.size, msg.iotlb.type); 141 142 if (write(fd, &msg, sizeof(msg)) != sizeof(msg)) { 143 error_report("failed to write, fd=%d, errno=%d (%s)", 144 fd, errno, strerror(errno)); 145 return -EIO ; 146 } 147 148 return ret; 149 } 150 151 static void vhost_vdpa_listener_begin_batch(VhostVDPAShared *s) 152 { 153 int fd = s->device_fd; 154 struct vhost_msg_v2 msg = { 155 .type = VHOST_IOTLB_MSG_V2, 156 .iotlb.type = VHOST_IOTLB_BATCH_BEGIN, 157 }; 158 159 trace_vhost_vdpa_listener_begin_batch(s, fd, msg.type, msg.iotlb.type); 160 if (write(fd, &msg, sizeof(msg)) != sizeof(msg)) { 161 error_report("failed to write, fd=%d, errno=%d (%s)", 162 fd, errno, strerror(errno)); 163 } 164 } 165 166 static void vhost_vdpa_iotlb_batch_begin_once(VhostVDPAShared *s) 167 { 168 if (s->backend_cap & (0x1ULL << VHOST_BACKEND_F_IOTLB_BATCH) && 169 !s->iotlb_batch_begin_sent) { 170 vhost_vdpa_listener_begin_batch(s); 171 } 172 173 s->iotlb_batch_begin_sent = true; 174 } 175 176 static void vhost_vdpa_listener_commit(MemoryListener *listener) 177 { 178 VhostVDPAShared *s = container_of(listener, VhostVDPAShared, listener); 179 struct vhost_msg_v2 msg = {}; 180 int fd = s->device_fd; 181 182 if (!(s->backend_cap & (0x1ULL << VHOST_BACKEND_F_IOTLB_BATCH))) { 183 return; 184 } 185 186 if (!s->iotlb_batch_begin_sent) { 187 return; 188 } 189 190 msg.type = VHOST_IOTLB_MSG_V2; 191 msg.iotlb.type = VHOST_IOTLB_BATCH_END; 192 193 trace_vhost_vdpa_listener_commit(s, fd, msg.type, msg.iotlb.type); 194 if (write(fd, &msg, sizeof(msg)) != sizeof(msg)) { 195 error_report("failed to write, fd=%d, errno=%d (%s)", 196 fd, errno, strerror(errno)); 197 } 198 199 s->iotlb_batch_begin_sent = false; 200 } 201 202 static void vhost_vdpa_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb) 203 { 204 struct vdpa_iommu *iommu = container_of(n, struct vdpa_iommu, n); 205 206 hwaddr iova = iotlb->iova + iommu->iommu_offset; 207 VhostVDPAShared *s = iommu->dev_shared; 208 void *vaddr; 209 int ret; 210 Int128 llend; 211 Error *local_err = NULL; 212 213 if (iotlb->target_as != &address_space_memory) { 214 error_report("Wrong target AS \"%s\", only system memory is allowed", 215 iotlb->target_as->name ? iotlb->target_as->name : "none"); 216 return; 217 } 218 RCU_READ_LOCK_GUARD(); 219 /* check if RAM section out of device range */ 220 llend = int128_add(int128_makes64(iotlb->addr_mask), int128_makes64(iova)); 221 if (int128_gt(llend, int128_make64(s->iova_range.last))) { 222 error_report("RAM section out of device range (max=0x%" PRIx64 223 ", end addr=0x%" PRIx64 ")", 224 s->iova_range.last, int128_get64(llend)); 225 return; 226 } 227 228 if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) { 229 bool read_only; 230 231 if (!memory_get_xlat_addr(iotlb, &vaddr, NULL, &read_only, NULL, 232 &local_err)) { 233 error_report_err(local_err); 234 return; 235 } 236 ret = vhost_vdpa_dma_map(s, VHOST_VDPA_GUEST_PA_ASID, iova, 237 iotlb->addr_mask + 1, vaddr, read_only); 238 if (ret) { 239 error_report("vhost_vdpa_dma_map(%p, 0x%" HWADDR_PRIx ", " 240 "0x%" HWADDR_PRIx ", %p) = %d (%m)", 241 s, iova, iotlb->addr_mask + 1, vaddr, ret); 242 } 243 } else { 244 ret = vhost_vdpa_dma_unmap(s, VHOST_VDPA_GUEST_PA_ASID, iova, 245 iotlb->addr_mask + 1); 246 if (ret) { 247 error_report("vhost_vdpa_dma_unmap(%p, 0x%" HWADDR_PRIx ", " 248 "0x%" HWADDR_PRIx ") = %d (%m)", 249 s, iova, iotlb->addr_mask + 1, ret); 250 } 251 } 252 } 253 254 static void vhost_vdpa_iommu_region_add(MemoryListener *listener, 255 MemoryRegionSection *section) 256 { 257 VhostVDPAShared *s = container_of(listener, VhostVDPAShared, listener); 258 259 struct vdpa_iommu *iommu; 260 Int128 end; 261 int iommu_idx; 262 IOMMUMemoryRegion *iommu_mr; 263 int ret; 264 265 iommu_mr = IOMMU_MEMORY_REGION(section->mr); 266 267 iommu = g_malloc0(sizeof(*iommu)); 268 end = int128_add(int128_make64(section->offset_within_region), 269 section->size); 270 end = int128_sub(end, int128_one()); 271 iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr, 272 MEMTXATTRS_UNSPECIFIED); 273 iommu->iommu_mr = iommu_mr; 274 iommu_notifier_init(&iommu->n, vhost_vdpa_iommu_map_notify, 275 IOMMU_NOTIFIER_IOTLB_EVENTS, 276 section->offset_within_region, 277 int128_get64(end), 278 iommu_idx); 279 iommu->iommu_offset = section->offset_within_address_space - 280 section->offset_within_region; 281 iommu->dev_shared = s; 282 283 ret = memory_region_register_iommu_notifier(section->mr, &iommu->n, NULL); 284 if (ret) { 285 g_free(iommu); 286 return; 287 } 288 289 QLIST_INSERT_HEAD(&s->iommu_list, iommu, iommu_next); 290 memory_region_iommu_replay(iommu->iommu_mr, &iommu->n); 291 292 return; 293 } 294 295 static void vhost_vdpa_iommu_region_del(MemoryListener *listener, 296 MemoryRegionSection *section) 297 { 298 VhostVDPAShared *s = container_of(listener, VhostVDPAShared, listener); 299 300 struct vdpa_iommu *iommu; 301 302 QLIST_FOREACH(iommu, &s->iommu_list, iommu_next) 303 { 304 if (MEMORY_REGION(iommu->iommu_mr) == section->mr && 305 iommu->n.start == section->offset_within_region) { 306 memory_region_unregister_iommu_notifier(section->mr, &iommu->n); 307 QLIST_REMOVE(iommu, iommu_next); 308 g_free(iommu); 309 break; 310 } 311 } 312 } 313 314 static void vhost_vdpa_listener_region_add(MemoryListener *listener, 315 MemoryRegionSection *section) 316 { 317 DMAMap mem_region = {}; 318 VhostVDPAShared *s = container_of(listener, VhostVDPAShared, listener); 319 hwaddr iova; 320 Int128 llend, llsize; 321 void *vaddr; 322 int ret; 323 int page_size = qemu_target_page_size(); 324 int page_mask = -page_size; 325 326 if (vhost_vdpa_listener_skipped_section(section, s->iova_range.first, 327 s->iova_range.last, page_mask)) { 328 return; 329 } 330 if (memory_region_is_iommu(section->mr)) { 331 vhost_vdpa_iommu_region_add(listener, section); 332 return; 333 } 334 335 if (unlikely((section->offset_within_address_space & ~page_mask) != 336 (section->offset_within_region & ~page_mask))) { 337 trace_vhost_vdpa_listener_region_add_unaligned(s, section->mr->name, 338 section->offset_within_address_space & ~page_mask, 339 section->offset_within_region & ~page_mask); 340 return; 341 } 342 343 iova = ROUND_UP(section->offset_within_address_space, page_size); 344 llend = vhost_vdpa_section_end(section, page_mask); 345 if (int128_ge(int128_make64(iova), llend)) { 346 return; 347 } 348 349 memory_region_ref(section->mr); 350 351 /* Here we assume that memory_region_is_ram(section->mr)==true */ 352 353 vaddr = memory_region_get_ram_ptr(section->mr) + 354 section->offset_within_region + 355 (iova - section->offset_within_address_space); 356 357 trace_vhost_vdpa_listener_region_add(s, iova, int128_get64(llend), 358 vaddr, section->readonly); 359 360 llsize = int128_sub(llend, int128_make64(iova)); 361 if (s->shadow_data) { 362 int r; 363 hwaddr gpa = section->offset_within_address_space; 364 365 mem_region.size = int128_get64(llsize) - 1, 366 mem_region.perm = IOMMU_ACCESS_FLAG(true, section->readonly), 367 368 r = vhost_iova_tree_map_alloc_gpa(s->iova_tree, &mem_region, gpa); 369 if (unlikely(r != IOVA_OK)) { 370 error_report("Can't allocate a mapping (%d)", r); 371 372 if (mem_region.translated_addr == gpa) { 373 error_report("Insertion to GPA->IOVA tree failed"); 374 /* Remove the mapping from the IOVA-only tree */ 375 goto fail_map; 376 } 377 goto fail; 378 } 379 380 iova = mem_region.iova; 381 } 382 383 vhost_vdpa_iotlb_batch_begin_once(s); 384 ret = vhost_vdpa_dma_map(s, VHOST_VDPA_GUEST_PA_ASID, iova, 385 int128_get64(llsize), vaddr, section->readonly); 386 if (ret) { 387 error_report("vhost vdpa map fail!"); 388 goto fail_map; 389 } 390 391 return; 392 393 fail_map: 394 if (s->shadow_data) { 395 vhost_iova_tree_remove_gpa(s->iova_tree, mem_region); 396 } 397 398 fail: 399 /* 400 * On the initfn path, store the first error in the container so we 401 * can gracefully fail. Runtime, there's not much we can do other 402 * than throw a hardware error. 403 */ 404 error_report("vhost-vdpa: DMA mapping failed, unable to continue"); 405 return; 406 407 } 408 409 static void vhost_vdpa_listener_region_del(MemoryListener *listener, 410 MemoryRegionSection *section) 411 { 412 VhostVDPAShared *s = container_of(listener, VhostVDPAShared, listener); 413 hwaddr iova; 414 Int128 llend, llsize; 415 int ret; 416 int page_size = qemu_target_page_size(); 417 int page_mask = -page_size; 418 419 if (vhost_vdpa_listener_skipped_section(section, s->iova_range.first, 420 s->iova_range.last, page_mask)) { 421 return; 422 } 423 if (memory_region_is_iommu(section->mr)) { 424 vhost_vdpa_iommu_region_del(listener, section); 425 } 426 427 if (unlikely((section->offset_within_address_space & ~page_mask) != 428 (section->offset_within_region & ~page_mask))) { 429 trace_vhost_vdpa_listener_region_del_unaligned(s, section->mr->name, 430 section->offset_within_address_space & ~page_mask, 431 section->offset_within_region & ~page_mask); 432 return; 433 } 434 435 iova = ROUND_UP(section->offset_within_address_space, page_size); 436 llend = vhost_vdpa_section_end(section, page_mask); 437 438 trace_vhost_vdpa_listener_region_del(s, iova, 439 int128_get64(int128_sub(llend, int128_one()))); 440 441 if (int128_ge(int128_make64(iova), llend)) { 442 return; 443 } 444 445 llsize = int128_sub(llend, int128_make64(iova)); 446 447 if (s->shadow_data) { 448 const DMAMap *result; 449 DMAMap mem_region = { 450 .translated_addr = section->offset_within_address_space, 451 .size = int128_get64(llsize) - 1, 452 }; 453 454 result = vhost_iova_tree_find_gpa(s->iova_tree, &mem_region); 455 if (!result) { 456 /* The memory listener map wasn't mapped */ 457 return; 458 } 459 iova = result->iova; 460 vhost_iova_tree_remove_gpa(s->iova_tree, *result); 461 } 462 vhost_vdpa_iotlb_batch_begin_once(s); 463 /* 464 * The unmap ioctl doesn't accept a full 64-bit. need to check it 465 */ 466 if (int128_eq(llsize, int128_2_64())) { 467 llsize = int128_rshift(llsize, 1); 468 ret = vhost_vdpa_dma_unmap(s, VHOST_VDPA_GUEST_PA_ASID, iova, 469 int128_get64(llsize)); 470 471 if (ret) { 472 error_report("vhost_vdpa_dma_unmap(%p, 0x%" HWADDR_PRIx ", " 473 "0x%" HWADDR_PRIx ") = %d (%m)", 474 s, iova, int128_get64(llsize), ret); 475 } 476 iova += int128_get64(llsize); 477 } 478 ret = vhost_vdpa_dma_unmap(s, VHOST_VDPA_GUEST_PA_ASID, iova, 479 int128_get64(llsize)); 480 481 if (ret) { 482 error_report("vhost_vdpa_dma_unmap(%p, 0x%" HWADDR_PRIx ", " 483 "0x%" HWADDR_PRIx ") = %d (%m)", 484 s, iova, int128_get64(llsize), ret); 485 } 486 487 memory_region_unref(section->mr); 488 } 489 /* 490 * IOTLB API is used by vhost-vdpa which requires incremental updating 491 * of the mapping. So we can not use generic vhost memory listener which 492 * depends on the addnop(). 493 */ 494 static const MemoryListener vhost_vdpa_memory_listener = { 495 .name = "vhost-vdpa", 496 .commit = vhost_vdpa_listener_commit, 497 .region_add = vhost_vdpa_listener_region_add, 498 .region_del = vhost_vdpa_listener_region_del, 499 }; 500 501 static int vhost_vdpa_call(struct vhost_dev *dev, unsigned long int request, 502 void *arg) 503 { 504 struct vhost_vdpa *v = dev->opaque; 505 int fd = v->shared->device_fd; 506 int ret; 507 508 assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA); 509 510 ret = ioctl(fd, request, arg); 511 return ret < 0 ? -errno : ret; 512 } 513 514 static int vhost_vdpa_add_status(struct vhost_dev *dev, uint8_t status) 515 { 516 uint8_t s; 517 int ret; 518 519 trace_vhost_vdpa_add_status(dev, status); 520 ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_STATUS, &s); 521 if (ret < 0) { 522 return ret; 523 } 524 if ((s & status) == status) { 525 /* Don't set bits already set */ 526 return 0; 527 } 528 529 s |= status; 530 531 ret = vhost_vdpa_call(dev, VHOST_VDPA_SET_STATUS, &s); 532 if (ret < 0) { 533 return ret; 534 } 535 536 ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_STATUS, &s); 537 if (ret < 0) { 538 return ret; 539 } 540 541 if (!(s & status)) { 542 return -EIO; 543 } 544 545 return 0; 546 } 547 548 int vhost_vdpa_get_iova_range(int fd, struct vhost_vdpa_iova_range *iova_range) 549 { 550 int ret = ioctl(fd, VHOST_VDPA_GET_IOVA_RANGE, iova_range); 551 552 return ret < 0 ? -errno : 0; 553 } 554 555 /* 556 * The use of this function is for requests that only need to be 557 * applied once. Typically such request occurs at the beginning 558 * of operation, and before setting up queues. It should not be 559 * used for request that performs operation until all queues are 560 * set, which would need to check dev->vq_index_end instead. 561 */ 562 static bool vhost_vdpa_first_dev(struct vhost_dev *dev) 563 { 564 struct vhost_vdpa *v = dev->opaque; 565 566 return v->index == 0; 567 } 568 569 static bool vhost_vdpa_last_dev(struct vhost_dev *dev) 570 { 571 return dev->vq_index + dev->nvqs == dev->vq_index_end; 572 } 573 574 static int vhost_vdpa_get_dev_features(struct vhost_dev *dev, 575 uint64_t *features) 576 { 577 int ret; 578 579 ret = vhost_vdpa_call(dev, VHOST_GET_FEATURES, features); 580 trace_vhost_vdpa_get_features(dev, *features); 581 return ret; 582 } 583 584 static void vhost_vdpa_init_svq(struct vhost_dev *hdev, struct vhost_vdpa *v) 585 { 586 g_autoptr(GPtrArray) shadow_vqs = NULL; 587 588 shadow_vqs = g_ptr_array_new_full(hdev->nvqs, vhost_svq_free); 589 for (unsigned n = 0; n < hdev->nvqs; ++n) { 590 VhostShadowVirtqueue *svq; 591 592 svq = vhost_svq_new(v->shadow_vq_ops, v->shadow_vq_ops_opaque); 593 g_ptr_array_add(shadow_vqs, svq); 594 } 595 596 v->shadow_vqs = g_steal_pointer(&shadow_vqs); 597 } 598 599 static int vhost_vdpa_init(struct vhost_dev *dev, void *opaque, Error **errp) 600 { 601 struct vhost_vdpa *v = opaque; 602 assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA); 603 trace_vhost_vdpa_init(dev, v->shared, opaque); 604 int ret; 605 606 v->dev = dev; 607 dev->opaque = opaque ; 608 v->shared->listener = vhost_vdpa_memory_listener; 609 vhost_vdpa_init_svq(dev, v); 610 611 error_propagate(&dev->migration_blocker, v->migration_blocker); 612 if (!vhost_vdpa_first_dev(dev)) { 613 return 0; 614 } 615 616 /* 617 * If dev->shadow_vqs_enabled at initialization that means the device has 618 * been started with x-svq=on, so don't block migration 619 */ 620 if (dev->migration_blocker == NULL && !v->shadow_vqs_enabled) { 621 /* We don't have dev->features yet */ 622 uint64_t features; 623 ret = vhost_vdpa_get_dev_features(dev, &features); 624 if (unlikely(ret)) { 625 error_setg_errno(errp, -ret, "Could not get device features"); 626 return ret; 627 } 628 vhost_svq_valid_features(features, &dev->migration_blocker); 629 } 630 631 /* 632 * Similar to VFIO, we end up pinning all guest memory and have to 633 * disable discarding of RAM. 634 */ 635 ret = ram_block_discard_disable(true); 636 if (ret) { 637 error_report("Cannot set discarding of RAM broken"); 638 return ret; 639 } 640 641 vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_ACKNOWLEDGE | 642 VIRTIO_CONFIG_S_DRIVER); 643 644 return 0; 645 } 646 647 static void vhost_vdpa_host_notifier_uninit(struct vhost_dev *dev, 648 int queue_index) 649 { 650 size_t page_size = qemu_real_host_page_size(); 651 struct vhost_vdpa *v = dev->opaque; 652 VirtIODevice *vdev = dev->vdev; 653 VhostVDPAHostNotifier *n; 654 655 n = &v->notifier[queue_index]; 656 657 if (n->addr) { 658 virtio_queue_set_host_notifier_mr(vdev, queue_index, &n->mr, false); 659 object_unparent(OBJECT(&n->mr)); 660 munmap(n->addr, page_size); 661 n->addr = NULL; 662 } 663 } 664 665 static int vhost_vdpa_host_notifier_init(struct vhost_dev *dev, int queue_index) 666 { 667 size_t page_size = qemu_real_host_page_size(); 668 struct vhost_vdpa *v = dev->opaque; 669 VirtIODevice *vdev = dev->vdev; 670 VhostVDPAHostNotifier *n; 671 int fd = v->shared->device_fd; 672 void *addr; 673 char *name; 674 675 vhost_vdpa_host_notifier_uninit(dev, queue_index); 676 677 n = &v->notifier[queue_index]; 678 679 addr = mmap(NULL, page_size, PROT_WRITE, MAP_SHARED, fd, 680 queue_index * page_size); 681 if (addr == MAP_FAILED) { 682 goto err; 683 } 684 685 name = g_strdup_printf("vhost-vdpa/host-notifier@%p mmaps[%d]", 686 v, queue_index); 687 memory_region_init_ram_device_ptr(&n->mr, OBJECT(vdev), name, 688 page_size, addr); 689 g_free(name); 690 691 if (virtio_queue_set_host_notifier_mr(vdev, queue_index, &n->mr, true)) { 692 object_unparent(OBJECT(&n->mr)); 693 munmap(addr, page_size); 694 goto err; 695 } 696 n->addr = addr; 697 698 return 0; 699 700 err: 701 return -1; 702 } 703 704 static void vhost_vdpa_host_notifiers_uninit(struct vhost_dev *dev, int n) 705 { 706 int i; 707 708 /* 709 * Pack all the changes to the memory regions in a single 710 * transaction to avoid a few updating of the address space 711 * topology. 712 */ 713 memory_region_transaction_begin(); 714 715 for (i = dev->vq_index; i < dev->vq_index + n; i++) { 716 vhost_vdpa_host_notifier_uninit(dev, i); 717 } 718 719 memory_region_transaction_commit(); 720 } 721 722 static void vhost_vdpa_host_notifiers_init(struct vhost_dev *dev) 723 { 724 struct vhost_vdpa *v = dev->opaque; 725 int i; 726 727 if (v->shadow_vqs_enabled) { 728 /* FIXME SVQ is not compatible with host notifiers mr */ 729 return; 730 } 731 732 /* 733 * Pack all the changes to the memory regions in a single 734 * transaction to avoid a few updating of the address space 735 * topology. 736 */ 737 memory_region_transaction_begin(); 738 739 for (i = dev->vq_index; i < dev->vq_index + dev->nvqs; i++) { 740 if (vhost_vdpa_host_notifier_init(dev, i)) { 741 vhost_vdpa_host_notifiers_uninit(dev, i - dev->vq_index); 742 break; 743 } 744 } 745 746 memory_region_transaction_commit(); 747 } 748 749 static void vhost_vdpa_svq_cleanup(struct vhost_dev *dev) 750 { 751 struct vhost_vdpa *v = dev->opaque; 752 size_t idx; 753 754 for (idx = 0; idx < v->shadow_vqs->len; ++idx) { 755 vhost_svq_stop(g_ptr_array_index(v->shadow_vqs, idx)); 756 } 757 g_ptr_array_free(v->shadow_vqs, true); 758 } 759 760 static int vhost_vdpa_cleanup(struct vhost_dev *dev) 761 { 762 struct vhost_vdpa *v; 763 assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA); 764 v = dev->opaque; 765 trace_vhost_vdpa_cleanup(dev, v); 766 if (vhost_vdpa_first_dev(dev)) { 767 ram_block_discard_disable(false); 768 memory_listener_unregister(&v->shared->listener); 769 } 770 771 vhost_vdpa_host_notifiers_uninit(dev, dev->nvqs); 772 vhost_vdpa_svq_cleanup(dev); 773 774 dev->opaque = NULL; 775 776 return 0; 777 } 778 779 static int vhost_vdpa_memslots_limit(struct vhost_dev *dev) 780 { 781 trace_vhost_vdpa_memslots_limit(dev, INT_MAX); 782 return INT_MAX; 783 } 784 785 static int vhost_vdpa_set_mem_table(struct vhost_dev *dev, 786 struct vhost_memory *mem) 787 { 788 if (!vhost_vdpa_first_dev(dev)) { 789 return 0; 790 } 791 792 trace_vhost_vdpa_set_mem_table(dev, mem->nregions, mem->padding); 793 if (trace_event_get_state_backends(TRACE_VHOST_VDPA_SET_MEM_TABLE) && 794 trace_event_get_state_backends(TRACE_VHOST_VDPA_DUMP_REGIONS)) { 795 int i; 796 for (i = 0; i < mem->nregions; i++) { 797 trace_vhost_vdpa_dump_regions(dev, i, 798 mem->regions[i].guest_phys_addr, 799 mem->regions[i].memory_size, 800 mem->regions[i].userspace_addr, 801 mem->regions[i].flags_padding); 802 } 803 } 804 if (mem->padding) { 805 return -EINVAL; 806 } 807 808 return 0; 809 } 810 811 static int vhost_vdpa_set_features(struct vhost_dev *dev, 812 uint64_t features) 813 { 814 struct vhost_vdpa *v = dev->opaque; 815 int ret; 816 817 if (!vhost_vdpa_first_dev(dev)) { 818 return 0; 819 } 820 821 if (v->shadow_vqs_enabled) { 822 if ((v->acked_features ^ features) == BIT_ULL(VHOST_F_LOG_ALL)) { 823 /* 824 * QEMU is just trying to enable or disable logging. SVQ handles 825 * this sepparately, so no need to forward this. 826 */ 827 v->acked_features = features; 828 return 0; 829 } 830 831 v->acked_features = features; 832 833 /* We must not ack _F_LOG if SVQ is enabled */ 834 features &= ~BIT_ULL(VHOST_F_LOG_ALL); 835 } 836 837 trace_vhost_vdpa_set_features(dev, features); 838 ret = vhost_vdpa_call(dev, VHOST_SET_FEATURES, &features); 839 if (ret) { 840 return ret; 841 } 842 843 return vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_FEATURES_OK); 844 } 845 846 static int vhost_vdpa_set_backend_cap(struct vhost_dev *dev) 847 { 848 struct vhost_vdpa *v = dev->opaque; 849 850 uint64_t features; 851 uint64_t f = 0x1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2 | 852 0x1ULL << VHOST_BACKEND_F_IOTLB_BATCH | 853 0x1ULL << VHOST_BACKEND_F_IOTLB_ASID | 854 0x1ULL << VHOST_BACKEND_F_SUSPEND; 855 int r; 856 857 if (vhost_vdpa_call(dev, VHOST_GET_BACKEND_FEATURES, &features)) { 858 return -EFAULT; 859 } 860 861 features &= f; 862 863 if (vhost_vdpa_first_dev(dev)) { 864 r = vhost_vdpa_call(dev, VHOST_SET_BACKEND_FEATURES, &features); 865 if (r) { 866 return -EFAULT; 867 } 868 } 869 870 dev->backend_cap = features; 871 v->shared->backend_cap = features; 872 873 return 0; 874 } 875 876 static int vhost_vdpa_get_device_id(struct vhost_dev *dev, 877 uint32_t *device_id) 878 { 879 int ret; 880 ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_DEVICE_ID, device_id); 881 trace_vhost_vdpa_get_device_id(dev, *device_id); 882 return ret; 883 } 884 885 static int vhost_vdpa_reset_device(struct vhost_dev *dev) 886 { 887 struct vhost_vdpa *v = dev->opaque; 888 int ret; 889 uint8_t status = 0; 890 891 ret = vhost_vdpa_call(dev, VHOST_VDPA_SET_STATUS, &status); 892 trace_vhost_vdpa_reset_device(dev); 893 v->suspended = false; 894 return ret; 895 } 896 897 static int vhost_vdpa_get_vq_index(struct vhost_dev *dev, int idx) 898 { 899 assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs); 900 901 trace_vhost_vdpa_get_vq_index(dev, idx, idx); 902 return idx; 903 } 904 905 static int vhost_vdpa_set_vring_enable_one(struct vhost_vdpa *v, unsigned idx, 906 int enable) 907 { 908 struct vhost_dev *dev = v->dev; 909 struct vhost_vring_state state = { 910 .index = idx, 911 .num = enable, 912 }; 913 int r = vhost_vdpa_call(dev, VHOST_VDPA_SET_VRING_ENABLE, &state); 914 915 trace_vhost_vdpa_set_vring_enable_one(dev, idx, enable, r); 916 return r; 917 } 918 919 static int vhost_vdpa_set_vring_enable(struct vhost_dev *dev, int enable) 920 { 921 struct vhost_vdpa *v = dev->opaque; 922 unsigned int i; 923 int ret; 924 925 for (i = 0; i < dev->nvqs; ++i) { 926 ret = vhost_vdpa_set_vring_enable_one(v, i, enable); 927 if (ret < 0) { 928 return ret; 929 } 930 } 931 932 return 0; 933 } 934 935 int vhost_vdpa_set_vring_ready(struct vhost_vdpa *v, unsigned idx) 936 { 937 return vhost_vdpa_set_vring_enable_one(v, idx, 1); 938 } 939 940 static int vhost_vdpa_set_config_call(struct vhost_dev *dev, 941 int fd) 942 { 943 trace_vhost_vdpa_set_config_call(dev, fd); 944 return vhost_vdpa_call(dev, VHOST_VDPA_SET_CONFIG_CALL, &fd); 945 } 946 947 static void vhost_vdpa_dump_config(struct vhost_dev *dev, const uint8_t *config, 948 uint32_t config_len) 949 { 950 g_autoptr(GString) str = g_string_sized_new(4 * 16); 951 size_t b, len; 952 953 for (b = 0; b < config_len; b += len) { 954 len = MIN(config_len - b, 16); 955 956 g_string_truncate(str, 0); 957 qemu_hexdump_line(str, config + b, len, 1, 4); 958 trace_vhost_vdpa_dump_config(dev, b, str->str); 959 } 960 } 961 962 static int vhost_vdpa_set_config(struct vhost_dev *dev, const uint8_t *data, 963 uint32_t offset, uint32_t size, 964 uint32_t flags) 965 { 966 struct vhost_vdpa_config *config; 967 int ret; 968 unsigned long config_size = offsetof(struct vhost_vdpa_config, buf); 969 970 trace_vhost_vdpa_set_config(dev, offset, size, flags); 971 config = g_malloc(size + config_size); 972 config->off = offset; 973 config->len = size; 974 memcpy(config->buf, data, size); 975 if (trace_event_get_state_backends(TRACE_VHOST_VDPA_SET_CONFIG) && 976 trace_event_get_state_backends(TRACE_VHOST_VDPA_DUMP_CONFIG)) { 977 vhost_vdpa_dump_config(dev, data, size); 978 } 979 ret = vhost_vdpa_call(dev, VHOST_VDPA_SET_CONFIG, config); 980 g_free(config); 981 return ret; 982 } 983 984 static int vhost_vdpa_get_config(struct vhost_dev *dev, uint8_t *config, 985 uint32_t config_len, Error **errp) 986 { 987 struct vhost_vdpa_config *v_config; 988 unsigned long config_size = offsetof(struct vhost_vdpa_config, buf); 989 int ret; 990 991 trace_vhost_vdpa_get_config(dev, config, config_len); 992 v_config = g_malloc(config_len + config_size); 993 v_config->len = config_len; 994 v_config->off = 0; 995 ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_CONFIG, v_config); 996 memcpy(config, v_config->buf, config_len); 997 g_free(v_config); 998 if (trace_event_get_state_backends(TRACE_VHOST_VDPA_GET_CONFIG) && 999 trace_event_get_state_backends(TRACE_VHOST_VDPA_DUMP_CONFIG)) { 1000 vhost_vdpa_dump_config(dev, config, config_len); 1001 } 1002 return ret; 1003 } 1004 1005 static int vhost_vdpa_set_dev_vring_base(struct vhost_dev *dev, 1006 struct vhost_vring_state *ring) 1007 { 1008 struct vhost_vdpa *v = dev->opaque; 1009 1010 trace_vhost_vdpa_set_dev_vring_base(dev, ring->index, ring->num, 1011 v->shadow_vqs_enabled); 1012 return vhost_vdpa_call(dev, VHOST_SET_VRING_BASE, ring); 1013 } 1014 1015 static int vhost_vdpa_set_vring_dev_kick(struct vhost_dev *dev, 1016 struct vhost_vring_file *file) 1017 { 1018 trace_vhost_vdpa_set_vring_kick(dev, file->index, file->fd); 1019 return vhost_vdpa_call(dev, VHOST_SET_VRING_KICK, file); 1020 } 1021 1022 static int vhost_vdpa_set_vring_dev_call(struct vhost_dev *dev, 1023 struct vhost_vring_file *file) 1024 { 1025 trace_vhost_vdpa_set_vring_call(dev, file->index, file->fd); 1026 return vhost_vdpa_call(dev, VHOST_SET_VRING_CALL, file); 1027 } 1028 1029 static int vhost_vdpa_set_vring_dev_addr(struct vhost_dev *dev, 1030 struct vhost_vring_addr *addr) 1031 { 1032 trace_vhost_vdpa_set_vring_addr(dev, addr->index, addr->flags, 1033 addr->desc_user_addr, addr->used_user_addr, 1034 addr->avail_user_addr, 1035 addr->log_guest_addr); 1036 1037 return vhost_vdpa_call(dev, VHOST_SET_VRING_ADDR, addr); 1038 1039 } 1040 1041 /** 1042 * Set the shadow virtqueue descriptors to the device 1043 * 1044 * @dev: The vhost device model 1045 * @svq: The shadow virtqueue 1046 * @idx: The index of the virtqueue in the vhost device 1047 * @errp: Error 1048 * 1049 * Note that this function does not rewind kick file descriptor if cannot set 1050 * call one. 1051 */ 1052 static int vhost_vdpa_svq_set_fds(struct vhost_dev *dev, 1053 VhostShadowVirtqueue *svq, unsigned idx, 1054 Error **errp) 1055 { 1056 struct vhost_vring_file file = { 1057 .index = dev->vq_index + idx, 1058 }; 1059 const EventNotifier *event_notifier = &svq->hdev_kick; 1060 int r; 1061 1062 r = event_notifier_init(&svq->hdev_kick, 0); 1063 if (r != 0) { 1064 error_setg_errno(errp, -r, "Couldn't create kick event notifier"); 1065 goto err_init_hdev_kick; 1066 } 1067 1068 r = event_notifier_init(&svq->hdev_call, 0); 1069 if (r != 0) { 1070 error_setg_errno(errp, -r, "Couldn't create call event notifier"); 1071 goto err_init_hdev_call; 1072 } 1073 1074 file.fd = event_notifier_get_fd(event_notifier); 1075 r = vhost_vdpa_set_vring_dev_kick(dev, &file); 1076 if (unlikely(r != 0)) { 1077 error_setg_errno(errp, -r, "Can't set device kick fd"); 1078 goto err_init_set_dev_fd; 1079 } 1080 1081 event_notifier = &svq->hdev_call; 1082 file.fd = event_notifier_get_fd(event_notifier); 1083 r = vhost_vdpa_set_vring_dev_call(dev, &file); 1084 if (unlikely(r != 0)) { 1085 error_setg_errno(errp, -r, "Can't set device call fd"); 1086 goto err_init_set_dev_fd; 1087 } 1088 1089 return 0; 1090 1091 err_init_set_dev_fd: 1092 event_notifier_set_handler(&svq->hdev_call, NULL); 1093 1094 err_init_hdev_call: 1095 event_notifier_cleanup(&svq->hdev_kick); 1096 1097 err_init_hdev_kick: 1098 return r; 1099 } 1100 1101 /** 1102 * Unmap a SVQ area in the device 1103 */ 1104 static void vhost_vdpa_svq_unmap_ring(struct vhost_vdpa *v, hwaddr addr) 1105 { 1106 const DMAMap needle = { 1107 .translated_addr = addr, 1108 }; 1109 const DMAMap *result = vhost_iova_tree_find_iova(v->shared->iova_tree, 1110 &needle); 1111 hwaddr size; 1112 int r; 1113 1114 if (unlikely(!result)) { 1115 error_report("Unable to find SVQ address to unmap"); 1116 return; 1117 } 1118 1119 size = ROUND_UP(result->size, qemu_real_host_page_size()); 1120 r = vhost_vdpa_dma_unmap(v->shared, v->address_space_id, result->iova, 1121 size); 1122 if (unlikely(r < 0)) { 1123 error_report("Unable to unmap SVQ vring: %s (%d)", g_strerror(-r), -r); 1124 return; 1125 } 1126 1127 vhost_iova_tree_remove(v->shared->iova_tree, *result); 1128 } 1129 1130 static void vhost_vdpa_svq_unmap_rings(struct vhost_dev *dev, 1131 const VhostShadowVirtqueue *svq) 1132 { 1133 struct vhost_vdpa *v = dev->opaque; 1134 struct vhost_vring_addr svq_addr; 1135 1136 vhost_svq_get_vring_addr(svq, &svq_addr); 1137 1138 vhost_vdpa_svq_unmap_ring(v, svq_addr.desc_user_addr); 1139 1140 vhost_vdpa_svq_unmap_ring(v, svq_addr.used_user_addr); 1141 } 1142 1143 /** 1144 * Map the SVQ area in the device 1145 * 1146 * @v: Vhost-vdpa device 1147 * @needle: The area to search iova 1148 * @taddr: The translated address (HVA) 1149 * @errorp: Error pointer 1150 */ 1151 static bool vhost_vdpa_svq_map_ring(struct vhost_vdpa *v, DMAMap *needle, 1152 hwaddr taddr, Error **errp) 1153 { 1154 int r; 1155 1156 r = vhost_iova_tree_map_alloc(v->shared->iova_tree, needle, taddr); 1157 if (unlikely(r != IOVA_OK)) { 1158 error_setg(errp, "Cannot allocate iova (%d)", r); 1159 1160 if (needle->translated_addr == taddr) { 1161 error_append_hint(errp, "Insertion to IOVA->HVA tree failed"); 1162 /* Remove the mapping from the IOVA-only tree */ 1163 vhost_iova_tree_remove(v->shared->iova_tree, *needle); 1164 } 1165 return false; 1166 } 1167 1168 r = vhost_vdpa_dma_map(v->shared, v->address_space_id, needle->iova, 1169 needle->size + 1, 1170 (void *)(uintptr_t)needle->translated_addr, 1171 needle->perm == IOMMU_RO); 1172 if (unlikely(r != 0)) { 1173 error_setg_errno(errp, -r, "Cannot map region to device"); 1174 vhost_iova_tree_remove(v->shared->iova_tree, *needle); 1175 } 1176 1177 return r == 0; 1178 } 1179 1180 /** 1181 * Map the shadow virtqueue rings in the device 1182 * 1183 * @dev: The vhost device 1184 * @svq: The shadow virtqueue 1185 * @addr: Assigned IOVA addresses 1186 * @errp: Error pointer 1187 */ 1188 static bool vhost_vdpa_svq_map_rings(struct vhost_dev *dev, 1189 const VhostShadowVirtqueue *svq, 1190 struct vhost_vring_addr *addr, 1191 Error **errp) 1192 { 1193 ERRP_GUARD(); 1194 DMAMap device_region, driver_region; 1195 struct vhost_vring_addr svq_addr; 1196 struct vhost_vdpa *v = dev->opaque; 1197 size_t device_size = vhost_svq_device_area_size(svq); 1198 size_t driver_size = vhost_svq_driver_area_size(svq); 1199 size_t avail_offset; 1200 bool ok; 1201 1202 vhost_svq_get_vring_addr(svq, &svq_addr); 1203 1204 driver_region = (DMAMap) { 1205 .size = driver_size - 1, 1206 .perm = IOMMU_RO, 1207 }; 1208 ok = vhost_vdpa_svq_map_ring(v, &driver_region, svq_addr.desc_user_addr, 1209 errp); 1210 if (unlikely(!ok)) { 1211 error_prepend(errp, "Cannot create vq driver region: "); 1212 return false; 1213 } 1214 addr->desc_user_addr = driver_region.iova; 1215 avail_offset = svq_addr.avail_user_addr - svq_addr.desc_user_addr; 1216 addr->avail_user_addr = driver_region.iova + avail_offset; 1217 1218 device_region = (DMAMap) { 1219 .size = device_size - 1, 1220 .perm = IOMMU_RW, 1221 }; 1222 ok = vhost_vdpa_svq_map_ring(v, &device_region, svq_addr.used_user_addr, 1223 errp); 1224 if (unlikely(!ok)) { 1225 error_prepend(errp, "Cannot create vq device region: "); 1226 vhost_vdpa_svq_unmap_ring(v, driver_region.translated_addr); 1227 } 1228 addr->used_user_addr = device_region.iova; 1229 1230 return ok; 1231 } 1232 1233 static bool vhost_vdpa_svq_setup(struct vhost_dev *dev, 1234 VhostShadowVirtqueue *svq, unsigned idx, 1235 Error **errp) 1236 { 1237 uint16_t vq_index = dev->vq_index + idx; 1238 struct vhost_vring_state s = { 1239 .index = vq_index, 1240 }; 1241 int r; 1242 1243 r = vhost_vdpa_set_dev_vring_base(dev, &s); 1244 if (unlikely(r)) { 1245 error_setg_errno(errp, -r, "Cannot set vring base"); 1246 return false; 1247 } 1248 1249 r = vhost_vdpa_svq_set_fds(dev, svq, idx, errp); 1250 return r == 0; 1251 } 1252 1253 static bool vhost_vdpa_svqs_start(struct vhost_dev *dev) 1254 { 1255 struct vhost_vdpa *v = dev->opaque; 1256 Error *err = NULL; 1257 unsigned i; 1258 1259 if (!v->shadow_vqs_enabled) { 1260 return true; 1261 } 1262 1263 for (i = 0; i < v->shadow_vqs->len; ++i) { 1264 VirtQueue *vq = virtio_get_queue(dev->vdev, dev->vq_index + i); 1265 VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, i); 1266 struct vhost_vring_addr addr = { 1267 .index = dev->vq_index + i, 1268 }; 1269 int r; 1270 bool ok = vhost_vdpa_svq_setup(dev, svq, i, &err); 1271 if (unlikely(!ok)) { 1272 goto err; 1273 } 1274 1275 vhost_svq_start(svq, dev->vdev, vq, v->shared->iova_tree); 1276 ok = vhost_vdpa_svq_map_rings(dev, svq, &addr, &err); 1277 if (unlikely(!ok)) { 1278 goto err_map; 1279 } 1280 1281 /* Override vring GPA set by vhost subsystem */ 1282 r = vhost_vdpa_set_vring_dev_addr(dev, &addr); 1283 if (unlikely(r != 0)) { 1284 error_setg_errno(&err, -r, "Cannot set device address"); 1285 goto err_set_addr; 1286 } 1287 } 1288 1289 return true; 1290 1291 err_set_addr: 1292 vhost_vdpa_svq_unmap_rings(dev, g_ptr_array_index(v->shadow_vqs, i)); 1293 1294 err_map: 1295 vhost_svq_stop(g_ptr_array_index(v->shadow_vqs, i)); 1296 1297 err: 1298 error_reportf_err(err, "Cannot setup SVQ %u: ", i); 1299 for (unsigned j = 0; j < i; ++j) { 1300 VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, j); 1301 vhost_vdpa_svq_unmap_rings(dev, svq); 1302 vhost_svq_stop(svq); 1303 } 1304 1305 return false; 1306 } 1307 1308 static void vhost_vdpa_svqs_stop(struct vhost_dev *dev) 1309 { 1310 struct vhost_vdpa *v = dev->opaque; 1311 1312 if (!v->shadow_vqs_enabled) { 1313 return; 1314 } 1315 1316 for (unsigned i = 0; i < v->shadow_vqs->len; ++i) { 1317 VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, i); 1318 1319 vhost_svq_stop(svq); 1320 vhost_vdpa_svq_unmap_rings(dev, svq); 1321 1322 event_notifier_cleanup(&svq->hdev_kick); 1323 event_notifier_cleanup(&svq->hdev_call); 1324 } 1325 } 1326 1327 static void vhost_vdpa_suspend(struct vhost_dev *dev) 1328 { 1329 struct vhost_vdpa *v = dev->opaque; 1330 int r; 1331 1332 if (!vhost_vdpa_first_dev(dev)) { 1333 return; 1334 } 1335 1336 if (dev->backend_cap & BIT_ULL(VHOST_BACKEND_F_SUSPEND)) { 1337 trace_vhost_vdpa_suspend(dev); 1338 r = ioctl(v->shared->device_fd, VHOST_VDPA_SUSPEND); 1339 if (unlikely(r)) { 1340 error_report("Cannot suspend: %s(%d)", g_strerror(errno), errno); 1341 } else { 1342 v->suspended = true; 1343 return; 1344 } 1345 } 1346 1347 vhost_vdpa_reset_device(dev); 1348 } 1349 1350 static int vhost_vdpa_dev_start(struct vhost_dev *dev, bool started) 1351 { 1352 struct vhost_vdpa *v = dev->opaque; 1353 bool ok; 1354 trace_vhost_vdpa_dev_start(dev, started); 1355 1356 if (started) { 1357 vhost_vdpa_host_notifiers_init(dev); 1358 ok = vhost_vdpa_svqs_start(dev); 1359 if (unlikely(!ok)) { 1360 return -1; 1361 } 1362 } else { 1363 vhost_vdpa_suspend(dev); 1364 vhost_vdpa_svqs_stop(dev); 1365 vhost_vdpa_host_notifiers_uninit(dev, dev->nvqs); 1366 } 1367 1368 if (!vhost_vdpa_last_dev(dev)) { 1369 return 0; 1370 } 1371 1372 if (started) { 1373 if (vhost_dev_has_iommu(dev) && (v->shadow_vqs_enabled)) { 1374 error_report("SVQ can not work while IOMMU enable, please disable" 1375 "IOMMU and try again"); 1376 return -1; 1377 } 1378 memory_listener_register(&v->shared->listener, dev->vdev->dma_as); 1379 1380 return vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_DRIVER_OK); 1381 } 1382 1383 return 0; 1384 } 1385 1386 static void vhost_vdpa_reset_status(struct vhost_dev *dev) 1387 { 1388 struct vhost_vdpa *v = dev->opaque; 1389 1390 if (!vhost_vdpa_last_dev(dev)) { 1391 return; 1392 } 1393 1394 vhost_vdpa_reset_device(dev); 1395 vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_ACKNOWLEDGE | 1396 VIRTIO_CONFIG_S_DRIVER); 1397 memory_listener_unregister(&v->shared->listener); 1398 } 1399 1400 static int vhost_vdpa_set_log_base(struct vhost_dev *dev, uint64_t base, 1401 struct vhost_log *log) 1402 { 1403 struct vhost_vdpa *v = dev->opaque; 1404 if (v->shadow_vqs_enabled || !vhost_vdpa_first_dev(dev)) { 1405 return 0; 1406 } 1407 1408 trace_vhost_vdpa_set_log_base(dev, base, log->size, log->refcnt, log->fd, 1409 log->log); 1410 return vhost_vdpa_call(dev, VHOST_SET_LOG_BASE, &base); 1411 } 1412 1413 static int vhost_vdpa_set_vring_addr(struct vhost_dev *dev, 1414 struct vhost_vring_addr *addr) 1415 { 1416 struct vhost_vdpa *v = dev->opaque; 1417 1418 if (v->shadow_vqs_enabled) { 1419 /* 1420 * Device vring addr was set at device start. SVQ base is handled by 1421 * VirtQueue code. 1422 */ 1423 return 0; 1424 } 1425 1426 return vhost_vdpa_set_vring_dev_addr(dev, addr); 1427 } 1428 1429 static int vhost_vdpa_set_vring_num(struct vhost_dev *dev, 1430 struct vhost_vring_state *ring) 1431 { 1432 trace_vhost_vdpa_set_vring_num(dev, ring->index, ring->num); 1433 return vhost_vdpa_call(dev, VHOST_SET_VRING_NUM, ring); 1434 } 1435 1436 static int vhost_vdpa_set_vring_base(struct vhost_dev *dev, 1437 struct vhost_vring_state *ring) 1438 { 1439 struct vhost_vdpa *v = dev->opaque; 1440 1441 if (v->shadow_vqs_enabled) { 1442 /* 1443 * Device vring base was set at device start. SVQ base is handled by 1444 * VirtQueue code. 1445 */ 1446 return 0; 1447 } 1448 1449 return vhost_vdpa_set_dev_vring_base(dev, ring); 1450 } 1451 1452 static int vhost_vdpa_get_vring_base(struct vhost_dev *dev, 1453 struct vhost_vring_state *ring) 1454 { 1455 struct vhost_vdpa *v = dev->opaque; 1456 int ret; 1457 1458 if (v->shadow_vqs_enabled) { 1459 ring->num = virtio_queue_get_last_avail_idx(dev->vdev, ring->index); 1460 trace_vhost_vdpa_get_vring_base(dev, ring->index, ring->num, true); 1461 return 0; 1462 } 1463 1464 if (!v->suspended) { 1465 /* 1466 * Cannot trust in value returned by device, let vhost recover used 1467 * idx from guest. 1468 */ 1469 return -1; 1470 } 1471 1472 ret = vhost_vdpa_call(dev, VHOST_GET_VRING_BASE, ring); 1473 trace_vhost_vdpa_get_vring_base(dev, ring->index, ring->num, false); 1474 return ret; 1475 } 1476 1477 static int vhost_vdpa_set_vring_kick(struct vhost_dev *dev, 1478 struct vhost_vring_file *file) 1479 { 1480 struct vhost_vdpa *v = dev->opaque; 1481 int vdpa_idx = file->index - dev->vq_index; 1482 1483 if (v->shadow_vqs_enabled) { 1484 VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, vdpa_idx); 1485 vhost_svq_set_svq_kick_fd(svq, file->fd); 1486 return 0; 1487 } else { 1488 return vhost_vdpa_set_vring_dev_kick(dev, file); 1489 } 1490 } 1491 1492 static int vhost_vdpa_set_vring_call(struct vhost_dev *dev, 1493 struct vhost_vring_file *file) 1494 { 1495 struct vhost_vdpa *v = dev->opaque; 1496 int vdpa_idx = file->index - dev->vq_index; 1497 VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, vdpa_idx); 1498 1499 /* Remember last call fd because we can switch to SVQ anytime. */ 1500 vhost_svq_set_svq_call_fd(svq, file->fd); 1501 /* 1502 * When SVQ is transitioning to off, shadow_vqs_enabled has 1503 * not been set back to false yet, but the underlying call fd 1504 * will have to switch back to the guest notifier to signal the 1505 * passthrough virtqueues. In other situations, SVQ's own call 1506 * fd shall be used to signal the device model. 1507 */ 1508 if (v->shadow_vqs_enabled && 1509 v->shared->svq_switching != SVQ_TSTATE_DISABLING) { 1510 return 0; 1511 } 1512 1513 return vhost_vdpa_set_vring_dev_call(dev, file); 1514 } 1515 1516 static int vhost_vdpa_get_features(struct vhost_dev *dev, 1517 uint64_t *features) 1518 { 1519 int ret = vhost_vdpa_get_dev_features(dev, features); 1520 1521 if (ret == 0) { 1522 /* Add SVQ logging capabilities */ 1523 *features |= BIT_ULL(VHOST_F_LOG_ALL); 1524 } 1525 1526 return ret; 1527 } 1528 1529 static int vhost_vdpa_set_owner(struct vhost_dev *dev) 1530 { 1531 if (!vhost_vdpa_first_dev(dev)) { 1532 return 0; 1533 } 1534 1535 trace_vhost_vdpa_set_owner(dev); 1536 return vhost_vdpa_call(dev, VHOST_SET_OWNER, NULL); 1537 } 1538 1539 static int vhost_vdpa_vq_get_addr(struct vhost_dev *dev, 1540 struct vhost_vring_addr *addr, struct vhost_virtqueue *vq) 1541 { 1542 assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA); 1543 addr->desc_user_addr = (uint64_t)(unsigned long)vq->desc_phys; 1544 addr->avail_user_addr = (uint64_t)(unsigned long)vq->avail_phys; 1545 addr->used_user_addr = (uint64_t)(unsigned long)vq->used_phys; 1546 trace_vhost_vdpa_vq_get_addr(dev, vq, addr->desc_user_addr, 1547 addr->avail_user_addr, addr->used_user_addr); 1548 return 0; 1549 } 1550 1551 static bool vhost_vdpa_force_iommu(struct vhost_dev *dev) 1552 { 1553 return true; 1554 } 1555 1556 const VhostOps vdpa_ops = { 1557 .backend_type = VHOST_BACKEND_TYPE_VDPA, 1558 .vhost_backend_init = vhost_vdpa_init, 1559 .vhost_backend_cleanup = vhost_vdpa_cleanup, 1560 .vhost_set_log_base = vhost_vdpa_set_log_base, 1561 .vhost_set_vring_addr = vhost_vdpa_set_vring_addr, 1562 .vhost_set_vring_num = vhost_vdpa_set_vring_num, 1563 .vhost_set_vring_base = vhost_vdpa_set_vring_base, 1564 .vhost_get_vring_base = vhost_vdpa_get_vring_base, 1565 .vhost_set_vring_kick = vhost_vdpa_set_vring_kick, 1566 .vhost_set_vring_call = vhost_vdpa_set_vring_call, 1567 .vhost_get_features = vhost_vdpa_get_features, 1568 .vhost_set_backend_cap = vhost_vdpa_set_backend_cap, 1569 .vhost_set_owner = vhost_vdpa_set_owner, 1570 .vhost_set_vring_endian = NULL, 1571 .vhost_backend_memslots_limit = vhost_vdpa_memslots_limit, 1572 .vhost_set_mem_table = vhost_vdpa_set_mem_table, 1573 .vhost_set_features = vhost_vdpa_set_features, 1574 .vhost_reset_device = vhost_vdpa_reset_device, 1575 .vhost_get_vq_index = vhost_vdpa_get_vq_index, 1576 .vhost_set_vring_enable = vhost_vdpa_set_vring_enable, 1577 .vhost_get_config = vhost_vdpa_get_config, 1578 .vhost_set_config = vhost_vdpa_set_config, 1579 .vhost_requires_shm_log = NULL, 1580 .vhost_migration_done = NULL, 1581 .vhost_net_set_mtu = NULL, 1582 .vhost_set_iotlb_callback = NULL, 1583 .vhost_send_device_iotlb_msg = NULL, 1584 .vhost_dev_start = vhost_vdpa_dev_start, 1585 .vhost_get_device_id = vhost_vdpa_get_device_id, 1586 .vhost_vq_get_addr = vhost_vdpa_vq_get_addr, 1587 .vhost_force_iommu = vhost_vdpa_force_iommu, 1588 .vhost_set_config_call = vhost_vdpa_set_config_call, 1589 .vhost_reset_status = vhost_vdpa_reset_status, 1590 }; 1591