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 "system/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 } 292 293 static void vhost_vdpa_iommu_region_del(MemoryListener *listener, 294 MemoryRegionSection *section) 295 { 296 VhostVDPAShared *s = container_of(listener, VhostVDPAShared, listener); 297 298 struct vdpa_iommu *iommu; 299 300 QLIST_FOREACH(iommu, &s->iommu_list, iommu_next) 301 { 302 if (MEMORY_REGION(iommu->iommu_mr) == section->mr && 303 iommu->n.start == section->offset_within_region) { 304 memory_region_unregister_iommu_notifier(section->mr, &iommu->n); 305 QLIST_REMOVE(iommu, iommu_next); 306 g_free(iommu); 307 break; 308 } 309 } 310 } 311 312 static void vhost_vdpa_listener_region_add(MemoryListener *listener, 313 MemoryRegionSection *section) 314 { 315 DMAMap mem_region = {}; 316 VhostVDPAShared *s = container_of(listener, VhostVDPAShared, listener); 317 hwaddr iova; 318 Int128 llend, llsize; 319 void *vaddr; 320 int ret; 321 int page_size = qemu_target_page_size(); 322 int page_mask = -page_size; 323 324 if (vhost_vdpa_listener_skipped_section(section, s->iova_range.first, 325 s->iova_range.last, page_mask)) { 326 return; 327 } 328 if (memory_region_is_iommu(section->mr)) { 329 vhost_vdpa_iommu_region_add(listener, section); 330 return; 331 } 332 333 if (unlikely((section->offset_within_address_space & ~page_mask) != 334 (section->offset_within_region & ~page_mask))) { 335 trace_vhost_vdpa_listener_region_add_unaligned(s, section->mr->name, 336 section->offset_within_address_space & ~page_mask, 337 section->offset_within_region & ~page_mask); 338 return; 339 } 340 341 iova = ROUND_UP(section->offset_within_address_space, page_size); 342 llend = vhost_vdpa_section_end(section, page_mask); 343 if (int128_ge(int128_make64(iova), llend)) { 344 return; 345 } 346 347 memory_region_ref(section->mr); 348 349 /* Here we assume that memory_region_is_ram(section->mr)==true */ 350 351 vaddr = memory_region_get_ram_ptr(section->mr) + 352 section->offset_within_region + 353 (iova - section->offset_within_address_space); 354 355 trace_vhost_vdpa_listener_region_add(s, iova, int128_get64(llend), 356 vaddr, section->readonly); 357 358 llsize = int128_sub(llend, int128_make64(iova)); 359 if (s->shadow_data) { 360 int r; 361 hwaddr gpa = section->offset_within_address_space; 362 363 mem_region.size = int128_get64(llsize) - 1, 364 mem_region.perm = IOMMU_ACCESS_FLAG(true, section->readonly), 365 366 r = vhost_iova_tree_map_alloc_gpa(s->iova_tree, &mem_region, gpa); 367 if (unlikely(r != IOVA_OK)) { 368 error_report("Can't allocate a mapping (%d)", r); 369 370 if (mem_region.translated_addr == gpa) { 371 error_report("Insertion to GPA->IOVA tree failed"); 372 /* Remove the mapping from the IOVA-only tree */ 373 goto fail_map; 374 } 375 goto fail; 376 } 377 378 iova = mem_region.iova; 379 } 380 381 vhost_vdpa_iotlb_batch_begin_once(s); 382 ret = vhost_vdpa_dma_map(s, VHOST_VDPA_GUEST_PA_ASID, iova, 383 int128_get64(llsize), vaddr, section->readonly); 384 if (ret) { 385 error_report("vhost vdpa map fail!"); 386 goto fail_map; 387 } 388 389 return; 390 391 fail_map: 392 if (s->shadow_data) { 393 vhost_iova_tree_remove_gpa(s->iova_tree, mem_region); 394 } 395 396 fail: 397 /* 398 * On the initfn path, store the first error in the container so we 399 * can gracefully fail. Runtime, there's not much we can do other 400 * than throw a hardware error. 401 */ 402 error_report("vhost-vdpa: DMA mapping failed, unable to continue"); 403 return; 404 405 } 406 407 static void vhost_vdpa_listener_region_del(MemoryListener *listener, 408 MemoryRegionSection *section) 409 { 410 VhostVDPAShared *s = container_of(listener, VhostVDPAShared, listener); 411 hwaddr iova; 412 Int128 llend, llsize; 413 int ret; 414 int page_size = qemu_target_page_size(); 415 int page_mask = -page_size; 416 417 if (vhost_vdpa_listener_skipped_section(section, s->iova_range.first, 418 s->iova_range.last, page_mask)) { 419 return; 420 } 421 if (memory_region_is_iommu(section->mr)) { 422 vhost_vdpa_iommu_region_del(listener, section); 423 } 424 425 if (unlikely((section->offset_within_address_space & ~page_mask) != 426 (section->offset_within_region & ~page_mask))) { 427 trace_vhost_vdpa_listener_region_del_unaligned(s, section->mr->name, 428 section->offset_within_address_space & ~page_mask, 429 section->offset_within_region & ~page_mask); 430 return; 431 } 432 433 iova = ROUND_UP(section->offset_within_address_space, page_size); 434 llend = vhost_vdpa_section_end(section, page_mask); 435 436 trace_vhost_vdpa_listener_region_del(s, iova, 437 int128_get64(int128_sub(llend, int128_one()))); 438 439 if (int128_ge(int128_make64(iova), llend)) { 440 return; 441 } 442 443 llsize = int128_sub(llend, int128_make64(iova)); 444 445 if (s->shadow_data) { 446 const DMAMap *result; 447 DMAMap mem_region = { 448 .translated_addr = section->offset_within_address_space, 449 .size = int128_get64(llsize) - 1, 450 }; 451 452 result = vhost_iova_tree_find_gpa(s->iova_tree, &mem_region); 453 if (!result) { 454 /* The memory listener map wasn't mapped */ 455 return; 456 } 457 iova = result->iova; 458 vhost_iova_tree_remove_gpa(s->iova_tree, *result); 459 } 460 vhost_vdpa_iotlb_batch_begin_once(s); 461 /* 462 * The unmap ioctl doesn't accept a full 64-bit. need to check it 463 */ 464 if (int128_eq(llsize, int128_2_64())) { 465 llsize = int128_rshift(llsize, 1); 466 ret = vhost_vdpa_dma_unmap(s, VHOST_VDPA_GUEST_PA_ASID, iova, 467 int128_get64(llsize)); 468 469 if (ret) { 470 error_report("vhost_vdpa_dma_unmap(%p, 0x%" HWADDR_PRIx ", " 471 "0x%" HWADDR_PRIx ") = %d (%m)", 472 s, iova, int128_get64(llsize), ret); 473 } 474 iova += int128_get64(llsize); 475 } 476 ret = vhost_vdpa_dma_unmap(s, VHOST_VDPA_GUEST_PA_ASID, iova, 477 int128_get64(llsize)); 478 479 if (ret) { 480 error_report("vhost_vdpa_dma_unmap(%p, 0x%" HWADDR_PRIx ", " 481 "0x%" HWADDR_PRIx ") = %d (%m)", 482 s, iova, int128_get64(llsize), ret); 483 } 484 485 memory_region_unref(section->mr); 486 } 487 /* 488 * IOTLB API is used by vhost-vdpa which requires incremental updating 489 * of the mapping. So we can not use generic vhost memory listener which 490 * depends on the addnop(). 491 */ 492 static const MemoryListener vhost_vdpa_memory_listener = { 493 .name = "vhost-vdpa", 494 .commit = vhost_vdpa_listener_commit, 495 .region_add = vhost_vdpa_listener_region_add, 496 .region_del = vhost_vdpa_listener_region_del, 497 }; 498 499 static int vhost_vdpa_call(struct vhost_dev *dev, unsigned long int request, 500 void *arg) 501 { 502 struct vhost_vdpa *v = dev->opaque; 503 int fd = v->shared->device_fd; 504 int ret; 505 506 assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA); 507 508 ret = ioctl(fd, request, arg); 509 return ret < 0 ? -errno : ret; 510 } 511 512 static int vhost_vdpa_add_status(struct vhost_dev *dev, uint8_t status) 513 { 514 uint8_t s; 515 int ret; 516 517 trace_vhost_vdpa_add_status(dev, status); 518 ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_STATUS, &s); 519 if (ret < 0) { 520 return ret; 521 } 522 if ((s & status) == status) { 523 /* Don't set bits already set */ 524 return 0; 525 } 526 527 s |= status; 528 529 ret = vhost_vdpa_call(dev, VHOST_VDPA_SET_STATUS, &s); 530 if (ret < 0) { 531 return ret; 532 } 533 534 ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_STATUS, &s); 535 if (ret < 0) { 536 return ret; 537 } 538 539 if (!(s & status)) { 540 return -EIO; 541 } 542 543 return 0; 544 } 545 546 int vhost_vdpa_get_iova_range(int fd, struct vhost_vdpa_iova_range *iova_range) 547 { 548 int ret = ioctl(fd, VHOST_VDPA_GET_IOVA_RANGE, iova_range); 549 550 return ret < 0 ? -errno : 0; 551 } 552 553 /* 554 * The use of this function is for requests that only need to be 555 * applied once. Typically such request occurs at the beginning 556 * of operation, and before setting up queues. It should not be 557 * used for request that performs operation until all queues are 558 * set, which would need to check dev->vq_index_end instead. 559 */ 560 static bool vhost_vdpa_first_dev(struct vhost_dev *dev) 561 { 562 struct vhost_vdpa *v = dev->opaque; 563 564 return v->index == 0; 565 } 566 567 static bool vhost_vdpa_last_dev(struct vhost_dev *dev) 568 { 569 return dev->vq_index + dev->nvqs == dev->vq_index_end; 570 } 571 572 static int vhost_vdpa_get_dev_features(struct vhost_dev *dev, 573 uint64_t *features) 574 { 575 int ret; 576 577 ret = vhost_vdpa_call(dev, VHOST_GET_FEATURES, features); 578 trace_vhost_vdpa_get_features(dev, *features); 579 return ret; 580 } 581 582 static void vhost_vdpa_init_svq(struct vhost_dev *hdev, struct vhost_vdpa *v) 583 { 584 g_autoptr(GPtrArray) shadow_vqs = NULL; 585 586 shadow_vqs = g_ptr_array_new_full(hdev->nvqs, vhost_svq_free); 587 for (unsigned n = 0; n < hdev->nvqs; ++n) { 588 VhostShadowVirtqueue *svq; 589 590 svq = vhost_svq_new(v->shadow_vq_ops, v->shadow_vq_ops_opaque); 591 g_ptr_array_add(shadow_vqs, svq); 592 } 593 594 v->shadow_vqs = g_steal_pointer(&shadow_vqs); 595 } 596 597 static int vhost_vdpa_set_backend_cap(struct vhost_dev *dev) 598 { 599 struct vhost_vdpa *v = dev->opaque; 600 601 uint64_t features; 602 uint64_t f = 0x1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2 | 603 0x1ULL << VHOST_BACKEND_F_IOTLB_BATCH | 604 0x1ULL << VHOST_BACKEND_F_IOTLB_ASID | 605 0x1ULL << VHOST_BACKEND_F_SUSPEND; 606 int r; 607 608 if (vhost_vdpa_call(dev, VHOST_GET_BACKEND_FEATURES, &features)) { 609 return -EFAULT; 610 } 611 612 features &= f; 613 614 if (vhost_vdpa_first_dev(dev)) { 615 r = vhost_vdpa_call(dev, VHOST_SET_BACKEND_FEATURES, &features); 616 if (r) { 617 return -EFAULT; 618 } 619 } 620 621 dev->backend_cap = features; 622 v->shared->backend_cap = features; 623 624 return 0; 625 } 626 627 static int vhost_vdpa_init(struct vhost_dev *dev, void *opaque, Error **errp) 628 { 629 struct vhost_vdpa *v = opaque; 630 assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA); 631 trace_vhost_vdpa_init(dev, v->shared, opaque); 632 int ret; 633 634 v->dev = dev; 635 dev->opaque = opaque ; 636 v->shared->listener = vhost_vdpa_memory_listener; 637 638 ret = vhost_vdpa_set_backend_cap(dev); 639 if (unlikely(ret != 0)) { 640 return ret; 641 } 642 643 vhost_vdpa_init_svq(dev, v); 644 645 error_propagate(&dev->migration_blocker, v->migration_blocker); 646 if (!vhost_vdpa_first_dev(dev)) { 647 return 0; 648 } 649 650 /* 651 * If dev->shadow_vqs_enabled at initialization that means the device has 652 * been started with x-svq=on, so don't block migration 653 */ 654 if (dev->migration_blocker == NULL && !v->shadow_vqs_enabled) { 655 /* We don't have dev->features yet */ 656 uint64_t features; 657 ret = vhost_vdpa_get_dev_features(dev, &features); 658 if (unlikely(ret)) { 659 error_setg_errno(errp, -ret, "Could not get device features"); 660 return ret; 661 } 662 vhost_svq_valid_features(features, &dev->migration_blocker); 663 } 664 665 /* 666 * Similar to VFIO, we end up pinning all guest memory and have to 667 * disable discarding of RAM. 668 */ 669 ret = ram_block_discard_disable(true); 670 if (ret) { 671 error_report("Cannot set discarding of RAM broken"); 672 return ret; 673 } 674 675 vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_ACKNOWLEDGE | 676 VIRTIO_CONFIG_S_DRIVER); 677 678 return 0; 679 } 680 681 static void vhost_vdpa_host_notifier_uninit(struct vhost_dev *dev, 682 int queue_index) 683 { 684 size_t page_size = qemu_real_host_page_size(); 685 struct vhost_vdpa *v = dev->opaque; 686 VirtIODevice *vdev = dev->vdev; 687 VhostVDPAHostNotifier *n; 688 689 n = &v->notifier[queue_index]; 690 691 if (n->addr) { 692 virtio_queue_set_host_notifier_mr(vdev, queue_index, &n->mr, false); 693 object_unparent(OBJECT(&n->mr)); 694 munmap(n->addr, page_size); 695 n->addr = NULL; 696 } 697 } 698 699 static int vhost_vdpa_host_notifier_init(struct vhost_dev *dev, int queue_index) 700 { 701 size_t page_size = qemu_real_host_page_size(); 702 struct vhost_vdpa *v = dev->opaque; 703 VirtIODevice *vdev = dev->vdev; 704 VhostVDPAHostNotifier *n; 705 int fd = v->shared->device_fd; 706 void *addr; 707 char *name; 708 709 vhost_vdpa_host_notifier_uninit(dev, queue_index); 710 711 n = &v->notifier[queue_index]; 712 713 addr = mmap(NULL, page_size, PROT_WRITE, MAP_SHARED, fd, 714 queue_index * page_size); 715 if (addr == MAP_FAILED) { 716 goto err; 717 } 718 719 name = g_strdup_printf("vhost-vdpa/host-notifier@%p mmaps[%d]", 720 v, queue_index); 721 memory_region_init_ram_device_ptr(&n->mr, OBJECT(vdev), name, 722 page_size, addr); 723 g_free(name); 724 725 if (virtio_queue_set_host_notifier_mr(vdev, queue_index, &n->mr, true)) { 726 object_unparent(OBJECT(&n->mr)); 727 munmap(addr, page_size); 728 goto err; 729 } 730 n->addr = addr; 731 732 return 0; 733 734 err: 735 return -1; 736 } 737 738 static void vhost_vdpa_host_notifiers_uninit(struct vhost_dev *dev, int n) 739 { 740 int i; 741 742 /* 743 * Pack all the changes to the memory regions in a single 744 * transaction to avoid a few updating of the address space 745 * topology. 746 */ 747 memory_region_transaction_begin(); 748 749 for (i = dev->vq_index; i < dev->vq_index + n; i++) { 750 vhost_vdpa_host_notifier_uninit(dev, i); 751 } 752 753 memory_region_transaction_commit(); 754 } 755 756 static void vhost_vdpa_host_notifiers_init(struct vhost_dev *dev) 757 { 758 struct vhost_vdpa *v = dev->opaque; 759 int i; 760 761 if (v->shadow_vqs_enabled) { 762 /* FIXME SVQ is not compatible with host notifiers mr */ 763 return; 764 } 765 766 /* 767 * Pack all the changes to the memory regions in a single 768 * transaction to avoid a few updating of the address space 769 * topology. 770 */ 771 memory_region_transaction_begin(); 772 773 for (i = dev->vq_index; i < dev->vq_index + dev->nvqs; i++) { 774 if (vhost_vdpa_host_notifier_init(dev, i)) { 775 vhost_vdpa_host_notifiers_uninit(dev, i - dev->vq_index); 776 break; 777 } 778 } 779 780 memory_region_transaction_commit(); 781 } 782 783 static void vhost_vdpa_svq_cleanup(struct vhost_dev *dev) 784 { 785 struct vhost_vdpa *v = dev->opaque; 786 size_t idx; 787 788 for (idx = 0; idx < v->shadow_vqs->len; ++idx) { 789 vhost_svq_stop(g_ptr_array_index(v->shadow_vqs, idx)); 790 } 791 g_ptr_array_free(v->shadow_vqs, true); 792 } 793 794 static int vhost_vdpa_cleanup(struct vhost_dev *dev) 795 { 796 struct vhost_vdpa *v; 797 assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA); 798 v = dev->opaque; 799 trace_vhost_vdpa_cleanup(dev, v); 800 if (vhost_vdpa_first_dev(dev)) { 801 ram_block_discard_disable(false); 802 memory_listener_unregister(&v->shared->listener); 803 } 804 805 vhost_vdpa_host_notifiers_uninit(dev, dev->nvqs); 806 vhost_vdpa_svq_cleanup(dev); 807 808 dev->opaque = NULL; 809 810 return 0; 811 } 812 813 static int vhost_vdpa_memslots_limit(struct vhost_dev *dev) 814 { 815 trace_vhost_vdpa_memslots_limit(dev, INT_MAX); 816 return INT_MAX; 817 } 818 819 static int vhost_vdpa_set_mem_table(struct vhost_dev *dev, 820 struct vhost_memory *mem) 821 { 822 if (!vhost_vdpa_first_dev(dev)) { 823 return 0; 824 } 825 826 trace_vhost_vdpa_set_mem_table(dev, mem->nregions, mem->padding); 827 if (trace_event_get_state_backends(TRACE_VHOST_VDPA_SET_MEM_TABLE) && 828 trace_event_get_state_backends(TRACE_VHOST_VDPA_DUMP_REGIONS)) { 829 int i; 830 for (i = 0; i < mem->nregions; i++) { 831 trace_vhost_vdpa_dump_regions(dev, i, 832 mem->regions[i].guest_phys_addr, 833 mem->regions[i].memory_size, 834 mem->regions[i].userspace_addr, 835 mem->regions[i].flags_padding); 836 } 837 } 838 if (mem->padding) { 839 return -EINVAL; 840 } 841 842 return 0; 843 } 844 845 static int vhost_vdpa_set_features(struct vhost_dev *dev, 846 uint64_t features) 847 { 848 struct vhost_vdpa *v = dev->opaque; 849 int ret; 850 851 if (!vhost_vdpa_first_dev(dev)) { 852 return 0; 853 } 854 855 if (v->shadow_vqs_enabled) { 856 if ((v->acked_features ^ features) == BIT_ULL(VHOST_F_LOG_ALL)) { 857 /* 858 * QEMU is just trying to enable or disable logging. SVQ handles 859 * this sepparately, so no need to forward this. 860 */ 861 v->acked_features = features; 862 return 0; 863 } 864 865 v->acked_features = features; 866 867 /* We must not ack _F_LOG if SVQ is enabled */ 868 features &= ~BIT_ULL(VHOST_F_LOG_ALL); 869 } 870 871 trace_vhost_vdpa_set_features(dev, features); 872 ret = vhost_vdpa_call(dev, VHOST_SET_FEATURES, &features); 873 if (ret) { 874 return ret; 875 } 876 877 return vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_FEATURES_OK); 878 } 879 880 static int vhost_vdpa_get_device_id(struct vhost_dev *dev, 881 uint32_t *device_id) 882 { 883 int ret; 884 ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_DEVICE_ID, device_id); 885 trace_vhost_vdpa_get_device_id(dev, *device_id); 886 return ret; 887 } 888 889 static int vhost_vdpa_reset_device(struct vhost_dev *dev) 890 { 891 struct vhost_vdpa *v = dev->opaque; 892 int ret; 893 uint8_t status = 0; 894 895 ret = vhost_vdpa_call(dev, VHOST_VDPA_SET_STATUS, &status); 896 trace_vhost_vdpa_reset_device(dev); 897 v->suspended = false; 898 return ret; 899 } 900 901 static int vhost_vdpa_get_vq_index(struct vhost_dev *dev, int idx) 902 { 903 assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs); 904 905 trace_vhost_vdpa_get_vq_index(dev, idx, idx); 906 return idx; 907 } 908 909 static int vhost_vdpa_set_vring_enable_one(struct vhost_vdpa *v, unsigned idx, 910 int enable) 911 { 912 struct vhost_dev *dev = v->dev; 913 struct vhost_vring_state state = { 914 .index = idx, 915 .num = enable, 916 }; 917 int r = vhost_vdpa_call(dev, VHOST_VDPA_SET_VRING_ENABLE, &state); 918 919 trace_vhost_vdpa_set_vring_enable_one(dev, idx, enable, r); 920 return r; 921 } 922 923 static int vhost_vdpa_set_vring_enable(struct vhost_dev *dev, int enable) 924 { 925 struct vhost_vdpa *v = dev->opaque; 926 unsigned int i; 927 int ret; 928 929 for (i = 0; i < dev->nvqs; ++i) { 930 ret = vhost_vdpa_set_vring_enable_one(v, i, enable); 931 if (ret < 0) { 932 return ret; 933 } 934 } 935 936 return 0; 937 } 938 939 int vhost_vdpa_set_vring_ready(struct vhost_vdpa *v, unsigned idx) 940 { 941 return vhost_vdpa_set_vring_enable_one(v, idx, 1); 942 } 943 944 static int vhost_vdpa_set_config_call(struct vhost_dev *dev, 945 int fd) 946 { 947 trace_vhost_vdpa_set_config_call(dev, fd); 948 return vhost_vdpa_call(dev, VHOST_VDPA_SET_CONFIG_CALL, &fd); 949 } 950 951 static void vhost_vdpa_dump_config(struct vhost_dev *dev, const uint8_t *config, 952 uint32_t config_len) 953 { 954 g_autoptr(GString) str = g_string_sized_new(4 * 16); 955 size_t b, len; 956 957 for (b = 0; b < config_len; b += len) { 958 len = MIN(config_len - b, 16); 959 960 g_string_truncate(str, 0); 961 qemu_hexdump_line(str, config + b, len, 1, 4); 962 trace_vhost_vdpa_dump_config(dev, b, str->str); 963 } 964 } 965 966 static int vhost_vdpa_set_config(struct vhost_dev *dev, const uint8_t *data, 967 uint32_t offset, uint32_t size, 968 uint32_t flags) 969 { 970 struct vhost_vdpa_config *config; 971 int ret; 972 unsigned long config_size = offsetof(struct vhost_vdpa_config, buf); 973 974 trace_vhost_vdpa_set_config(dev, offset, size, flags); 975 config = g_malloc(size + config_size); 976 config->off = offset; 977 config->len = size; 978 memcpy(config->buf, data, size); 979 if (trace_event_get_state_backends(TRACE_VHOST_VDPA_SET_CONFIG) && 980 trace_event_get_state_backends(TRACE_VHOST_VDPA_DUMP_CONFIG)) { 981 vhost_vdpa_dump_config(dev, data, size); 982 } 983 ret = vhost_vdpa_call(dev, VHOST_VDPA_SET_CONFIG, config); 984 g_free(config); 985 return ret; 986 } 987 988 static int vhost_vdpa_get_config(struct vhost_dev *dev, uint8_t *config, 989 uint32_t config_len, Error **errp) 990 { 991 struct vhost_vdpa_config *v_config; 992 unsigned long config_size = offsetof(struct vhost_vdpa_config, buf); 993 int ret; 994 995 trace_vhost_vdpa_get_config(dev, config, config_len); 996 v_config = g_malloc(config_len + config_size); 997 v_config->len = config_len; 998 v_config->off = 0; 999 ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_CONFIG, v_config); 1000 memcpy(config, v_config->buf, config_len); 1001 g_free(v_config); 1002 if (trace_event_get_state_backends(TRACE_VHOST_VDPA_GET_CONFIG) && 1003 trace_event_get_state_backends(TRACE_VHOST_VDPA_DUMP_CONFIG)) { 1004 vhost_vdpa_dump_config(dev, config, config_len); 1005 } 1006 return ret; 1007 } 1008 1009 static int vhost_vdpa_set_dev_vring_base(struct vhost_dev *dev, 1010 struct vhost_vring_state *ring) 1011 { 1012 struct vhost_vdpa *v = dev->opaque; 1013 1014 trace_vhost_vdpa_set_dev_vring_base(dev, ring->index, ring->num, 1015 v->shadow_vqs_enabled); 1016 return vhost_vdpa_call(dev, VHOST_SET_VRING_BASE, ring); 1017 } 1018 1019 static int vhost_vdpa_set_vring_dev_kick(struct vhost_dev *dev, 1020 struct vhost_vring_file *file) 1021 { 1022 trace_vhost_vdpa_set_vring_kick(dev, file->index, file->fd); 1023 return vhost_vdpa_call(dev, VHOST_SET_VRING_KICK, file); 1024 } 1025 1026 static int vhost_vdpa_set_vring_dev_call(struct vhost_dev *dev, 1027 struct vhost_vring_file *file) 1028 { 1029 trace_vhost_vdpa_set_vring_call(dev, file->index, file->fd); 1030 return vhost_vdpa_call(dev, VHOST_SET_VRING_CALL, file); 1031 } 1032 1033 static int vhost_vdpa_set_vring_dev_addr(struct vhost_dev *dev, 1034 struct vhost_vring_addr *addr) 1035 { 1036 trace_vhost_vdpa_set_vring_addr(dev, addr->index, addr->flags, 1037 addr->desc_user_addr, addr->used_user_addr, 1038 addr->avail_user_addr, 1039 addr->log_guest_addr); 1040 1041 return vhost_vdpa_call(dev, VHOST_SET_VRING_ADDR, addr); 1042 1043 } 1044 1045 /** 1046 * Set the shadow virtqueue descriptors to the device 1047 * 1048 * @dev: The vhost device model 1049 * @svq: The shadow virtqueue 1050 * @idx: The index of the virtqueue in the vhost device 1051 * @errp: Error 1052 * 1053 * Note that this function does not rewind kick file descriptor if cannot set 1054 * call one. 1055 */ 1056 static int vhost_vdpa_svq_set_fds(struct vhost_dev *dev, 1057 VhostShadowVirtqueue *svq, unsigned idx, 1058 Error **errp) 1059 { 1060 struct vhost_vring_file file = { 1061 .index = dev->vq_index + idx, 1062 }; 1063 const EventNotifier *event_notifier = &svq->hdev_kick; 1064 int r; 1065 1066 r = event_notifier_init(&svq->hdev_kick, 0); 1067 if (r != 0) { 1068 error_setg_errno(errp, -r, "Couldn't create kick event notifier"); 1069 goto err_init_hdev_kick; 1070 } 1071 1072 r = event_notifier_init(&svq->hdev_call, 0); 1073 if (r != 0) { 1074 error_setg_errno(errp, -r, "Couldn't create call event notifier"); 1075 goto err_init_hdev_call; 1076 } 1077 1078 file.fd = event_notifier_get_fd(event_notifier); 1079 r = vhost_vdpa_set_vring_dev_kick(dev, &file); 1080 if (unlikely(r != 0)) { 1081 error_setg_errno(errp, -r, "Can't set device kick fd"); 1082 goto err_init_set_dev_fd; 1083 } 1084 1085 event_notifier = &svq->hdev_call; 1086 file.fd = event_notifier_get_fd(event_notifier); 1087 r = vhost_vdpa_set_vring_dev_call(dev, &file); 1088 if (unlikely(r != 0)) { 1089 error_setg_errno(errp, -r, "Can't set device call fd"); 1090 goto err_init_set_dev_fd; 1091 } 1092 1093 return 0; 1094 1095 err_init_set_dev_fd: 1096 event_notifier_set_handler(&svq->hdev_call, NULL); 1097 1098 err_init_hdev_call: 1099 event_notifier_cleanup(&svq->hdev_kick); 1100 1101 err_init_hdev_kick: 1102 return r; 1103 } 1104 1105 /** 1106 * Unmap a SVQ area in the device 1107 */ 1108 static void vhost_vdpa_svq_unmap_ring(struct vhost_vdpa *v, hwaddr addr) 1109 { 1110 const DMAMap needle = { 1111 .translated_addr = addr, 1112 }; 1113 const DMAMap *result = vhost_iova_tree_find_iova(v->shared->iova_tree, 1114 &needle); 1115 hwaddr size; 1116 int r; 1117 1118 if (unlikely(!result)) { 1119 error_report("Unable to find SVQ address to unmap"); 1120 return; 1121 } 1122 1123 size = ROUND_UP(result->size, qemu_real_host_page_size()); 1124 r = vhost_vdpa_dma_unmap(v->shared, v->address_space_id, result->iova, 1125 size); 1126 if (unlikely(r < 0)) { 1127 error_report("Unable to unmap SVQ vring: %s (%d)", g_strerror(-r), -r); 1128 return; 1129 } 1130 1131 vhost_iova_tree_remove(v->shared->iova_tree, *result); 1132 } 1133 1134 static void vhost_vdpa_svq_unmap_rings(struct vhost_dev *dev, 1135 const VhostShadowVirtqueue *svq) 1136 { 1137 struct vhost_vdpa *v = dev->opaque; 1138 struct vhost_vring_addr svq_addr; 1139 1140 vhost_svq_get_vring_addr(svq, &svq_addr); 1141 1142 vhost_vdpa_svq_unmap_ring(v, svq_addr.desc_user_addr); 1143 1144 vhost_vdpa_svq_unmap_ring(v, svq_addr.used_user_addr); 1145 } 1146 1147 /** 1148 * Map the SVQ area in the device 1149 * 1150 * @v: Vhost-vdpa device 1151 * @needle: The area to search iova 1152 * @taddr: The translated address (HVA) 1153 * @errorp: Error pointer 1154 */ 1155 static bool vhost_vdpa_svq_map_ring(struct vhost_vdpa *v, DMAMap *needle, 1156 hwaddr taddr, Error **errp) 1157 { 1158 int r; 1159 1160 r = vhost_iova_tree_map_alloc(v->shared->iova_tree, needle, taddr); 1161 if (unlikely(r != IOVA_OK)) { 1162 error_setg(errp, "Cannot allocate iova (%d)", r); 1163 1164 if (needle->translated_addr == taddr) { 1165 error_append_hint(errp, "Insertion to IOVA->HVA tree failed"); 1166 /* Remove the mapping from the IOVA-only tree */ 1167 vhost_iova_tree_remove(v->shared->iova_tree, *needle); 1168 } 1169 return false; 1170 } 1171 1172 r = vhost_vdpa_dma_map(v->shared, v->address_space_id, needle->iova, 1173 needle->size + 1, 1174 (void *)(uintptr_t)needle->translated_addr, 1175 needle->perm == IOMMU_RO); 1176 if (unlikely(r != 0)) { 1177 error_setg_errno(errp, -r, "Cannot map region to device"); 1178 vhost_iova_tree_remove(v->shared->iova_tree, *needle); 1179 } 1180 1181 return r == 0; 1182 } 1183 1184 /** 1185 * Map the shadow virtqueue rings in the device 1186 * 1187 * @dev: The vhost device 1188 * @svq: The shadow virtqueue 1189 * @addr: Assigned IOVA addresses 1190 * @errp: Error pointer 1191 */ 1192 static bool vhost_vdpa_svq_map_rings(struct vhost_dev *dev, 1193 const VhostShadowVirtqueue *svq, 1194 struct vhost_vring_addr *addr, 1195 Error **errp) 1196 { 1197 ERRP_GUARD(); 1198 DMAMap device_region, driver_region; 1199 struct vhost_vring_addr svq_addr; 1200 struct vhost_vdpa *v = dev->opaque; 1201 size_t device_size = vhost_svq_device_area_size(svq); 1202 size_t driver_size = vhost_svq_driver_area_size(svq); 1203 size_t avail_offset; 1204 bool ok; 1205 1206 vhost_svq_get_vring_addr(svq, &svq_addr); 1207 1208 driver_region = (DMAMap) { 1209 .size = driver_size - 1, 1210 .perm = IOMMU_RO, 1211 }; 1212 ok = vhost_vdpa_svq_map_ring(v, &driver_region, svq_addr.desc_user_addr, 1213 errp); 1214 if (unlikely(!ok)) { 1215 error_prepend(errp, "Cannot create vq driver region: "); 1216 return false; 1217 } 1218 addr->desc_user_addr = driver_region.iova; 1219 avail_offset = svq_addr.avail_user_addr - svq_addr.desc_user_addr; 1220 addr->avail_user_addr = driver_region.iova + avail_offset; 1221 1222 device_region = (DMAMap) { 1223 .size = device_size - 1, 1224 .perm = IOMMU_RW, 1225 }; 1226 ok = vhost_vdpa_svq_map_ring(v, &device_region, svq_addr.used_user_addr, 1227 errp); 1228 if (unlikely(!ok)) { 1229 error_prepend(errp, "Cannot create vq device region: "); 1230 vhost_vdpa_svq_unmap_ring(v, driver_region.translated_addr); 1231 } 1232 addr->used_user_addr = device_region.iova; 1233 1234 return ok; 1235 } 1236 1237 static bool vhost_vdpa_svq_setup(struct vhost_dev *dev, 1238 VhostShadowVirtqueue *svq, unsigned idx, 1239 Error **errp) 1240 { 1241 uint16_t vq_index = dev->vq_index + idx; 1242 struct vhost_vring_state s = { 1243 .index = vq_index, 1244 }; 1245 int r; 1246 1247 r = vhost_vdpa_set_dev_vring_base(dev, &s); 1248 if (unlikely(r)) { 1249 error_setg_errno(errp, -r, "Cannot set vring base"); 1250 return false; 1251 } 1252 1253 r = vhost_vdpa_svq_set_fds(dev, svq, idx, errp); 1254 return r == 0; 1255 } 1256 1257 static bool vhost_vdpa_svqs_start(struct vhost_dev *dev) 1258 { 1259 struct vhost_vdpa *v = dev->opaque; 1260 Error *err = NULL; 1261 unsigned i; 1262 1263 if (!v->shadow_vqs_enabled) { 1264 return true; 1265 } 1266 1267 for (i = 0; i < v->shadow_vqs->len; ++i) { 1268 VirtQueue *vq = virtio_get_queue(dev->vdev, dev->vq_index + i); 1269 VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, i); 1270 struct vhost_vring_addr addr = { 1271 .index = dev->vq_index + i, 1272 }; 1273 int r; 1274 bool ok = vhost_vdpa_svq_setup(dev, svq, i, &err); 1275 if (unlikely(!ok)) { 1276 goto err; 1277 } 1278 1279 vhost_svq_start(svq, dev->vdev, vq, v->shared->iova_tree); 1280 ok = vhost_vdpa_svq_map_rings(dev, svq, &addr, &err); 1281 if (unlikely(!ok)) { 1282 goto err_map; 1283 } 1284 1285 /* Override vring GPA set by vhost subsystem */ 1286 r = vhost_vdpa_set_vring_dev_addr(dev, &addr); 1287 if (unlikely(r != 0)) { 1288 error_setg_errno(&err, -r, "Cannot set device address"); 1289 goto err_set_addr; 1290 } 1291 } 1292 1293 return true; 1294 1295 err_set_addr: 1296 vhost_vdpa_svq_unmap_rings(dev, g_ptr_array_index(v->shadow_vqs, i)); 1297 1298 err_map: 1299 vhost_svq_stop(g_ptr_array_index(v->shadow_vqs, i)); 1300 1301 err: 1302 error_reportf_err(err, "Cannot setup SVQ %u: ", i); 1303 for (unsigned j = 0; j < i; ++j) { 1304 VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, j); 1305 vhost_vdpa_svq_unmap_rings(dev, svq); 1306 vhost_svq_stop(svq); 1307 } 1308 1309 return false; 1310 } 1311 1312 static void vhost_vdpa_svqs_stop(struct vhost_dev *dev) 1313 { 1314 struct vhost_vdpa *v = dev->opaque; 1315 1316 if (!v->shadow_vqs_enabled) { 1317 return; 1318 } 1319 1320 for (unsigned i = 0; i < v->shadow_vqs->len; ++i) { 1321 VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, i); 1322 1323 vhost_svq_stop(svq); 1324 vhost_vdpa_svq_unmap_rings(dev, svq); 1325 1326 event_notifier_cleanup(&svq->hdev_kick); 1327 event_notifier_cleanup(&svq->hdev_call); 1328 } 1329 } 1330 1331 static void vhost_vdpa_suspend(struct vhost_dev *dev) 1332 { 1333 struct vhost_vdpa *v = dev->opaque; 1334 int r; 1335 1336 if (!vhost_vdpa_first_dev(dev)) { 1337 return; 1338 } 1339 1340 if (dev->backend_cap & BIT_ULL(VHOST_BACKEND_F_SUSPEND)) { 1341 trace_vhost_vdpa_suspend(dev); 1342 r = ioctl(v->shared->device_fd, VHOST_VDPA_SUSPEND); 1343 if (unlikely(r)) { 1344 error_report("Cannot suspend: %s(%d)", g_strerror(errno), errno); 1345 } else { 1346 v->suspended = true; 1347 return; 1348 } 1349 } 1350 1351 vhost_vdpa_reset_device(dev); 1352 } 1353 1354 static int vhost_vdpa_dev_start(struct vhost_dev *dev, bool started) 1355 { 1356 struct vhost_vdpa *v = dev->opaque; 1357 bool ok; 1358 trace_vhost_vdpa_dev_start(dev, started); 1359 1360 if (started) { 1361 vhost_vdpa_host_notifiers_init(dev); 1362 ok = vhost_vdpa_svqs_start(dev); 1363 if (unlikely(!ok)) { 1364 return -1; 1365 } 1366 } else { 1367 vhost_vdpa_suspend(dev); 1368 vhost_vdpa_svqs_stop(dev); 1369 vhost_vdpa_host_notifiers_uninit(dev, dev->nvqs); 1370 } 1371 1372 if (!vhost_vdpa_last_dev(dev)) { 1373 return 0; 1374 } 1375 1376 if (started) { 1377 if (vhost_dev_has_iommu(dev) && (v->shadow_vqs_enabled)) { 1378 error_report("SVQ can not work while IOMMU enable, please disable" 1379 "IOMMU and try again"); 1380 return -1; 1381 } 1382 memory_listener_register(&v->shared->listener, dev->vdev->dma_as); 1383 1384 return vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_DRIVER_OK); 1385 } 1386 1387 return 0; 1388 } 1389 1390 static void vhost_vdpa_reset_status(struct vhost_dev *dev) 1391 { 1392 struct vhost_vdpa *v = dev->opaque; 1393 1394 if (!vhost_vdpa_last_dev(dev)) { 1395 return; 1396 } 1397 1398 vhost_vdpa_reset_device(dev); 1399 vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_ACKNOWLEDGE | 1400 VIRTIO_CONFIG_S_DRIVER); 1401 memory_listener_unregister(&v->shared->listener); 1402 } 1403 1404 static int vhost_vdpa_set_log_base(struct vhost_dev *dev, uint64_t base, 1405 struct vhost_log *log) 1406 { 1407 struct vhost_vdpa *v = dev->opaque; 1408 if (v->shadow_vqs_enabled || !vhost_vdpa_first_dev(dev)) { 1409 return 0; 1410 } 1411 1412 trace_vhost_vdpa_set_log_base(dev, base, log->size, log->refcnt, log->fd, 1413 log->log); 1414 return vhost_vdpa_call(dev, VHOST_SET_LOG_BASE, &base); 1415 } 1416 1417 static int vhost_vdpa_set_vring_addr(struct vhost_dev *dev, 1418 struct vhost_vring_addr *addr) 1419 { 1420 struct vhost_vdpa *v = dev->opaque; 1421 1422 if (v->shadow_vqs_enabled) { 1423 /* 1424 * Device vring addr was set at device start. SVQ base is handled by 1425 * VirtQueue code. 1426 */ 1427 return 0; 1428 } 1429 1430 return vhost_vdpa_set_vring_dev_addr(dev, addr); 1431 } 1432 1433 static int vhost_vdpa_set_vring_num(struct vhost_dev *dev, 1434 struct vhost_vring_state *ring) 1435 { 1436 trace_vhost_vdpa_set_vring_num(dev, ring->index, ring->num); 1437 return vhost_vdpa_call(dev, VHOST_SET_VRING_NUM, ring); 1438 } 1439 1440 static int vhost_vdpa_set_vring_base(struct vhost_dev *dev, 1441 struct vhost_vring_state *ring) 1442 { 1443 struct vhost_vdpa *v = dev->opaque; 1444 1445 if (v->shadow_vqs_enabled) { 1446 /* 1447 * Device vring base was set at device start. SVQ base is handled by 1448 * VirtQueue code. 1449 */ 1450 return 0; 1451 } 1452 1453 return vhost_vdpa_set_dev_vring_base(dev, ring); 1454 } 1455 1456 static int vhost_vdpa_get_vring_base(struct vhost_dev *dev, 1457 struct vhost_vring_state *ring) 1458 { 1459 struct vhost_vdpa *v = dev->opaque; 1460 int ret; 1461 1462 if (v->shadow_vqs_enabled) { 1463 ring->num = virtio_queue_get_last_avail_idx(dev->vdev, ring->index); 1464 trace_vhost_vdpa_get_vring_base(dev, ring->index, ring->num, true); 1465 return 0; 1466 } 1467 1468 if (!v->suspended) { 1469 /* 1470 * Cannot trust in value returned by device, let vhost recover used 1471 * idx from guest. 1472 */ 1473 return -1; 1474 } 1475 1476 ret = vhost_vdpa_call(dev, VHOST_GET_VRING_BASE, ring); 1477 trace_vhost_vdpa_get_vring_base(dev, ring->index, ring->num, false); 1478 return ret; 1479 } 1480 1481 static int vhost_vdpa_set_vring_kick(struct vhost_dev *dev, 1482 struct vhost_vring_file *file) 1483 { 1484 struct vhost_vdpa *v = dev->opaque; 1485 int vdpa_idx = file->index - dev->vq_index; 1486 1487 if (v->shadow_vqs_enabled) { 1488 VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, vdpa_idx); 1489 vhost_svq_set_svq_kick_fd(svq, file->fd); 1490 return 0; 1491 } else { 1492 return vhost_vdpa_set_vring_dev_kick(dev, file); 1493 } 1494 } 1495 1496 static int vhost_vdpa_set_vring_call(struct vhost_dev *dev, 1497 struct vhost_vring_file *file) 1498 { 1499 struct vhost_vdpa *v = dev->opaque; 1500 int vdpa_idx = file->index - dev->vq_index; 1501 VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, vdpa_idx); 1502 1503 /* Remember last call fd because we can switch to SVQ anytime. */ 1504 vhost_svq_set_svq_call_fd(svq, file->fd); 1505 /* 1506 * When SVQ is transitioning to off, shadow_vqs_enabled has 1507 * not been set back to false yet, but the underlying call fd 1508 * will have to switch back to the guest notifier to signal the 1509 * passthrough virtqueues. In other situations, SVQ's own call 1510 * fd shall be used to signal the device model. 1511 */ 1512 if (v->shadow_vqs_enabled && 1513 v->shared->svq_switching != SVQ_TSTATE_DISABLING) { 1514 return 0; 1515 } 1516 1517 return vhost_vdpa_set_vring_dev_call(dev, file); 1518 } 1519 1520 static int vhost_vdpa_get_features(struct vhost_dev *dev, 1521 uint64_t *features) 1522 { 1523 int ret = vhost_vdpa_get_dev_features(dev, features); 1524 1525 if (ret == 0) { 1526 /* Add SVQ logging capabilities */ 1527 *features |= BIT_ULL(VHOST_F_LOG_ALL); 1528 } 1529 1530 return ret; 1531 } 1532 1533 static int vhost_vdpa_set_owner(struct vhost_dev *dev) 1534 { 1535 if (!vhost_vdpa_first_dev(dev)) { 1536 return 0; 1537 } 1538 1539 trace_vhost_vdpa_set_owner(dev); 1540 return vhost_vdpa_call(dev, VHOST_SET_OWNER, NULL); 1541 } 1542 1543 static int vhost_vdpa_vq_get_addr(struct vhost_dev *dev, 1544 struct vhost_vring_addr *addr, struct vhost_virtqueue *vq) 1545 { 1546 assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA); 1547 addr->desc_user_addr = (uint64_t)(unsigned long)vq->desc_phys; 1548 addr->avail_user_addr = (uint64_t)(unsigned long)vq->avail_phys; 1549 addr->used_user_addr = (uint64_t)(unsigned long)vq->used_phys; 1550 trace_vhost_vdpa_vq_get_addr(dev, vq, addr->desc_user_addr, 1551 addr->avail_user_addr, addr->used_user_addr); 1552 return 0; 1553 } 1554 1555 static bool vhost_vdpa_force_iommu(struct vhost_dev *dev) 1556 { 1557 return true; 1558 } 1559 1560 const VhostOps vdpa_ops = { 1561 .backend_type = VHOST_BACKEND_TYPE_VDPA, 1562 .vhost_backend_init = vhost_vdpa_init, 1563 .vhost_backend_cleanup = vhost_vdpa_cleanup, 1564 .vhost_set_log_base = vhost_vdpa_set_log_base, 1565 .vhost_set_vring_addr = vhost_vdpa_set_vring_addr, 1566 .vhost_set_vring_num = vhost_vdpa_set_vring_num, 1567 .vhost_set_vring_base = vhost_vdpa_set_vring_base, 1568 .vhost_get_vring_base = vhost_vdpa_get_vring_base, 1569 .vhost_set_vring_kick = vhost_vdpa_set_vring_kick, 1570 .vhost_set_vring_call = vhost_vdpa_set_vring_call, 1571 .vhost_get_features = vhost_vdpa_get_features, 1572 .vhost_set_owner = vhost_vdpa_set_owner, 1573 .vhost_set_vring_endian = NULL, 1574 .vhost_backend_memslots_limit = vhost_vdpa_memslots_limit, 1575 .vhost_set_mem_table = vhost_vdpa_set_mem_table, 1576 .vhost_set_features = vhost_vdpa_set_features, 1577 .vhost_reset_device = vhost_vdpa_reset_device, 1578 .vhost_get_vq_index = vhost_vdpa_get_vq_index, 1579 .vhost_set_vring_enable = vhost_vdpa_set_vring_enable, 1580 .vhost_get_config = vhost_vdpa_get_config, 1581 .vhost_set_config = vhost_vdpa_set_config, 1582 .vhost_requires_shm_log = NULL, 1583 .vhost_migration_done = NULL, 1584 .vhost_net_set_mtu = NULL, 1585 .vhost_set_iotlb_callback = NULL, 1586 .vhost_send_device_iotlb_msg = NULL, 1587 .vhost_dev_start = vhost_vdpa_dev_start, 1588 .vhost_get_device_id = vhost_vdpa_get_device_id, 1589 .vhost_vq_get_addr = vhost_vdpa_vq_get_addr, 1590 .vhost_force_iommu = vhost_vdpa_force_iommu, 1591 .vhost_set_config_call = vhost_vdpa_set_config_call, 1592 .vhost_reset_status = vhost_vdpa_reset_status, 1593 }; 1594