1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Virtio driver for the paravirtualized IOMMU 4 * 5 * Copyright (C) 2019 Arm Limited 6 */ 7 8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 9 10 #include <linux/amba/bus.h> 11 #include <linux/delay.h> 12 #include <linux/dma-iommu.h> 13 #include <linux/dma-map-ops.h> 14 #include <linux/freezer.h> 15 #include <linux/interval_tree.h> 16 #include <linux/iommu.h> 17 #include <linux/module.h> 18 #include <linux/of_platform.h> 19 #include <linux/pci.h> 20 #include <linux/platform_device.h> 21 #include <linux/virtio.h> 22 #include <linux/virtio_config.h> 23 #include <linux/virtio_ids.h> 24 #include <linux/wait.h> 25 26 #include <uapi/linux/virtio_iommu.h> 27 28 #define MSI_IOVA_BASE 0x8000000 29 #define MSI_IOVA_LENGTH 0x100000 30 31 #define VIOMMU_REQUEST_VQ 0 32 #define VIOMMU_EVENT_VQ 1 33 #define VIOMMU_NR_VQS 2 34 35 struct viommu_dev { 36 struct iommu_device iommu; 37 struct device *dev; 38 struct virtio_device *vdev; 39 40 struct ida domain_ids; 41 42 struct virtqueue *vqs[VIOMMU_NR_VQS]; 43 spinlock_t request_lock; 44 struct list_head requests; 45 void *evts; 46 47 /* Device configuration */ 48 struct iommu_domain_geometry geometry; 49 u64 pgsize_bitmap; 50 u32 first_domain; 51 u32 last_domain; 52 /* Supported MAP flags */ 53 u32 map_flags; 54 u32 probe_size; 55 }; 56 57 struct viommu_mapping { 58 phys_addr_t paddr; 59 struct interval_tree_node iova; 60 u32 flags; 61 }; 62 63 struct viommu_domain { 64 struct iommu_domain domain; 65 struct viommu_dev *viommu; 66 struct mutex mutex; /* protects viommu pointer */ 67 unsigned int id; 68 u32 map_flags; 69 70 spinlock_t mappings_lock; 71 struct rb_root_cached mappings; 72 73 unsigned long nr_endpoints; 74 bool bypass; 75 }; 76 77 struct viommu_endpoint { 78 struct device *dev; 79 struct viommu_dev *viommu; 80 struct viommu_domain *vdomain; 81 struct list_head resv_regions; 82 }; 83 84 struct viommu_request { 85 struct list_head list; 86 void *writeback; 87 unsigned int write_offset; 88 unsigned int len; 89 char buf[]; 90 }; 91 92 #define VIOMMU_FAULT_RESV_MASK 0xffffff00 93 94 struct viommu_event { 95 union { 96 u32 head; 97 struct virtio_iommu_fault fault; 98 }; 99 }; 100 101 #define to_viommu_domain(domain) \ 102 container_of(domain, struct viommu_domain, domain) 103 104 static int viommu_get_req_errno(void *buf, size_t len) 105 { 106 struct virtio_iommu_req_tail *tail = buf + len - sizeof(*tail); 107 108 switch (tail->status) { 109 case VIRTIO_IOMMU_S_OK: 110 return 0; 111 case VIRTIO_IOMMU_S_UNSUPP: 112 return -ENOSYS; 113 case VIRTIO_IOMMU_S_INVAL: 114 return -EINVAL; 115 case VIRTIO_IOMMU_S_RANGE: 116 return -ERANGE; 117 case VIRTIO_IOMMU_S_NOENT: 118 return -ENOENT; 119 case VIRTIO_IOMMU_S_FAULT: 120 return -EFAULT; 121 case VIRTIO_IOMMU_S_NOMEM: 122 return -ENOMEM; 123 case VIRTIO_IOMMU_S_IOERR: 124 case VIRTIO_IOMMU_S_DEVERR: 125 default: 126 return -EIO; 127 } 128 } 129 130 static void viommu_set_req_status(void *buf, size_t len, int status) 131 { 132 struct virtio_iommu_req_tail *tail = buf + len - sizeof(*tail); 133 134 tail->status = status; 135 } 136 137 static off_t viommu_get_write_desc_offset(struct viommu_dev *viommu, 138 struct virtio_iommu_req_head *req, 139 size_t len) 140 { 141 size_t tail_size = sizeof(struct virtio_iommu_req_tail); 142 143 if (req->type == VIRTIO_IOMMU_T_PROBE) 144 return len - viommu->probe_size - tail_size; 145 146 return len - tail_size; 147 } 148 149 /* 150 * __viommu_sync_req - Complete all in-flight requests 151 * 152 * Wait for all added requests to complete. When this function returns, all 153 * requests that were in-flight at the time of the call have completed. 154 */ 155 static int __viommu_sync_req(struct viommu_dev *viommu) 156 { 157 unsigned int len; 158 size_t write_len; 159 struct viommu_request *req; 160 struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ]; 161 162 assert_spin_locked(&viommu->request_lock); 163 164 virtqueue_kick(vq); 165 166 while (!list_empty(&viommu->requests)) { 167 len = 0; 168 req = virtqueue_get_buf(vq, &len); 169 if (!req) 170 continue; 171 172 if (!len) 173 viommu_set_req_status(req->buf, req->len, 174 VIRTIO_IOMMU_S_IOERR); 175 176 write_len = req->len - req->write_offset; 177 if (req->writeback && len == write_len) 178 memcpy(req->writeback, req->buf + req->write_offset, 179 write_len); 180 181 list_del(&req->list); 182 kfree(req); 183 } 184 185 return 0; 186 } 187 188 static int viommu_sync_req(struct viommu_dev *viommu) 189 { 190 int ret; 191 unsigned long flags; 192 193 spin_lock_irqsave(&viommu->request_lock, flags); 194 ret = __viommu_sync_req(viommu); 195 if (ret) 196 dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret); 197 spin_unlock_irqrestore(&viommu->request_lock, flags); 198 199 return ret; 200 } 201 202 /* 203 * __viommu_add_request - Add one request to the queue 204 * @buf: pointer to the request buffer 205 * @len: length of the request buffer 206 * @writeback: copy data back to the buffer when the request completes. 207 * 208 * Add a request to the queue. Only synchronize the queue if it's already full. 209 * Otherwise don't kick the queue nor wait for requests to complete. 210 * 211 * When @writeback is true, data written by the device, including the request 212 * status, is copied into @buf after the request completes. This is unsafe if 213 * the caller allocates @buf on stack and drops the lock between add_req() and 214 * sync_req(). 215 * 216 * Return 0 if the request was successfully added to the queue. 217 */ 218 static int __viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len, 219 bool writeback) 220 { 221 int ret; 222 off_t write_offset; 223 struct viommu_request *req; 224 struct scatterlist top_sg, bottom_sg; 225 struct scatterlist *sg[2] = { &top_sg, &bottom_sg }; 226 struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ]; 227 228 assert_spin_locked(&viommu->request_lock); 229 230 write_offset = viommu_get_write_desc_offset(viommu, buf, len); 231 if (write_offset <= 0) 232 return -EINVAL; 233 234 req = kzalloc(sizeof(*req) + len, GFP_ATOMIC); 235 if (!req) 236 return -ENOMEM; 237 238 req->len = len; 239 if (writeback) { 240 req->writeback = buf + write_offset; 241 req->write_offset = write_offset; 242 } 243 memcpy(&req->buf, buf, write_offset); 244 245 sg_init_one(&top_sg, req->buf, write_offset); 246 sg_init_one(&bottom_sg, req->buf + write_offset, len - write_offset); 247 248 ret = virtqueue_add_sgs(vq, sg, 1, 1, req, GFP_ATOMIC); 249 if (ret == -ENOSPC) { 250 /* If the queue is full, sync and retry */ 251 if (!__viommu_sync_req(viommu)) 252 ret = virtqueue_add_sgs(vq, sg, 1, 1, req, GFP_ATOMIC); 253 } 254 if (ret) 255 goto err_free; 256 257 list_add_tail(&req->list, &viommu->requests); 258 return 0; 259 260 err_free: 261 kfree(req); 262 return ret; 263 } 264 265 static int viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len) 266 { 267 int ret; 268 unsigned long flags; 269 270 spin_lock_irqsave(&viommu->request_lock, flags); 271 ret = __viommu_add_req(viommu, buf, len, false); 272 if (ret) 273 dev_dbg(viommu->dev, "could not add request: %d\n", ret); 274 spin_unlock_irqrestore(&viommu->request_lock, flags); 275 276 return ret; 277 } 278 279 /* 280 * Send a request and wait for it to complete. Return the request status (as an 281 * errno) 282 */ 283 static int viommu_send_req_sync(struct viommu_dev *viommu, void *buf, 284 size_t len) 285 { 286 int ret; 287 unsigned long flags; 288 289 spin_lock_irqsave(&viommu->request_lock, flags); 290 291 ret = __viommu_add_req(viommu, buf, len, true); 292 if (ret) { 293 dev_dbg(viommu->dev, "could not add request (%d)\n", ret); 294 goto out_unlock; 295 } 296 297 ret = __viommu_sync_req(viommu); 298 if (ret) { 299 dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret); 300 /* Fall-through (get the actual request status) */ 301 } 302 303 ret = viommu_get_req_errno(buf, len); 304 out_unlock: 305 spin_unlock_irqrestore(&viommu->request_lock, flags); 306 return ret; 307 } 308 309 /* 310 * viommu_add_mapping - add a mapping to the internal tree 311 * 312 * On success, return the new mapping. Otherwise return NULL. 313 */ 314 static int viommu_add_mapping(struct viommu_domain *vdomain, u64 iova, u64 end, 315 phys_addr_t paddr, u32 flags) 316 { 317 unsigned long irqflags; 318 struct viommu_mapping *mapping; 319 320 mapping = kzalloc(sizeof(*mapping), GFP_ATOMIC); 321 if (!mapping) 322 return -ENOMEM; 323 324 mapping->paddr = paddr; 325 mapping->iova.start = iova; 326 mapping->iova.last = end; 327 mapping->flags = flags; 328 329 spin_lock_irqsave(&vdomain->mappings_lock, irqflags); 330 interval_tree_insert(&mapping->iova, &vdomain->mappings); 331 spin_unlock_irqrestore(&vdomain->mappings_lock, irqflags); 332 333 return 0; 334 } 335 336 /* 337 * viommu_del_mappings - remove mappings from the internal tree 338 * 339 * @vdomain: the domain 340 * @iova: start of the range 341 * @end: end of the range 342 * 343 * On success, returns the number of unmapped bytes 344 */ 345 static size_t viommu_del_mappings(struct viommu_domain *vdomain, 346 u64 iova, u64 end) 347 { 348 size_t unmapped = 0; 349 unsigned long flags; 350 struct viommu_mapping *mapping = NULL; 351 struct interval_tree_node *node, *next; 352 353 spin_lock_irqsave(&vdomain->mappings_lock, flags); 354 next = interval_tree_iter_first(&vdomain->mappings, iova, end); 355 while (next) { 356 node = next; 357 mapping = container_of(node, struct viommu_mapping, iova); 358 next = interval_tree_iter_next(node, iova, end); 359 360 /* Trying to split a mapping? */ 361 if (mapping->iova.start < iova) 362 break; 363 364 /* 365 * Virtio-iommu doesn't allow UNMAP to split a mapping created 366 * with a single MAP request, so remove the full mapping. 367 */ 368 unmapped += mapping->iova.last - mapping->iova.start + 1; 369 370 interval_tree_remove(node, &vdomain->mappings); 371 kfree(mapping); 372 } 373 spin_unlock_irqrestore(&vdomain->mappings_lock, flags); 374 375 return unmapped; 376 } 377 378 /* 379 * Fill the domain with identity mappings, skipping the device's reserved 380 * regions. 381 */ 382 static int viommu_domain_map_identity(struct viommu_endpoint *vdev, 383 struct viommu_domain *vdomain) 384 { 385 int ret; 386 struct iommu_resv_region *resv; 387 u64 iova = vdomain->domain.geometry.aperture_start; 388 u64 limit = vdomain->domain.geometry.aperture_end; 389 u32 flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE; 390 unsigned long granule = 1UL << __ffs(vdomain->domain.pgsize_bitmap); 391 392 iova = ALIGN(iova, granule); 393 limit = ALIGN_DOWN(limit + 1, granule) - 1; 394 395 list_for_each_entry(resv, &vdev->resv_regions, list) { 396 u64 resv_start = ALIGN_DOWN(resv->start, granule); 397 u64 resv_end = ALIGN(resv->start + resv->length, granule) - 1; 398 399 if (resv_end < iova || resv_start > limit) 400 /* No overlap */ 401 continue; 402 403 if (resv_start > iova) { 404 ret = viommu_add_mapping(vdomain, iova, resv_start - 1, 405 (phys_addr_t)iova, flags); 406 if (ret) 407 goto err_unmap; 408 } 409 410 if (resv_end >= limit) 411 return 0; 412 413 iova = resv_end + 1; 414 } 415 416 ret = viommu_add_mapping(vdomain, iova, limit, (phys_addr_t)iova, 417 flags); 418 if (ret) 419 goto err_unmap; 420 return 0; 421 422 err_unmap: 423 viommu_del_mappings(vdomain, 0, iova); 424 return ret; 425 } 426 427 /* 428 * viommu_replay_mappings - re-send MAP requests 429 * 430 * When reattaching a domain that was previously detached from all endpoints, 431 * mappings were deleted from the device. Re-create the mappings available in 432 * the internal tree. 433 */ 434 static int viommu_replay_mappings(struct viommu_domain *vdomain) 435 { 436 int ret = 0; 437 unsigned long flags; 438 struct viommu_mapping *mapping; 439 struct interval_tree_node *node; 440 struct virtio_iommu_req_map map; 441 442 spin_lock_irqsave(&vdomain->mappings_lock, flags); 443 node = interval_tree_iter_first(&vdomain->mappings, 0, -1UL); 444 while (node) { 445 mapping = container_of(node, struct viommu_mapping, iova); 446 map = (struct virtio_iommu_req_map) { 447 .head.type = VIRTIO_IOMMU_T_MAP, 448 .domain = cpu_to_le32(vdomain->id), 449 .virt_start = cpu_to_le64(mapping->iova.start), 450 .virt_end = cpu_to_le64(mapping->iova.last), 451 .phys_start = cpu_to_le64(mapping->paddr), 452 .flags = cpu_to_le32(mapping->flags), 453 }; 454 455 ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map)); 456 if (ret) 457 break; 458 459 node = interval_tree_iter_next(node, 0, -1UL); 460 } 461 spin_unlock_irqrestore(&vdomain->mappings_lock, flags); 462 463 return ret; 464 } 465 466 static int viommu_add_resv_mem(struct viommu_endpoint *vdev, 467 struct virtio_iommu_probe_resv_mem *mem, 468 size_t len) 469 { 470 size_t size; 471 u64 start64, end64; 472 phys_addr_t start, end; 473 struct iommu_resv_region *region = NULL, *next; 474 unsigned long prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO; 475 476 start = start64 = le64_to_cpu(mem->start); 477 end = end64 = le64_to_cpu(mem->end); 478 size = end64 - start64 + 1; 479 480 /* Catch any overflow, including the unlikely end64 - start64 + 1 = 0 */ 481 if (start != start64 || end != end64 || size < end64 - start64) 482 return -EOVERFLOW; 483 484 if (len < sizeof(*mem)) 485 return -EINVAL; 486 487 switch (mem->subtype) { 488 default: 489 dev_warn(vdev->dev, "unknown resv mem subtype 0x%x\n", 490 mem->subtype); 491 fallthrough; 492 case VIRTIO_IOMMU_RESV_MEM_T_RESERVED: 493 region = iommu_alloc_resv_region(start, size, 0, 494 IOMMU_RESV_RESERVED); 495 break; 496 case VIRTIO_IOMMU_RESV_MEM_T_MSI: 497 region = iommu_alloc_resv_region(start, size, prot, 498 IOMMU_RESV_MSI); 499 break; 500 } 501 if (!region) 502 return -ENOMEM; 503 504 /* Keep the list sorted */ 505 list_for_each_entry(next, &vdev->resv_regions, list) { 506 if (next->start > region->start) 507 break; 508 } 509 list_add_tail(®ion->list, &next->list); 510 return 0; 511 } 512 513 static int viommu_probe_endpoint(struct viommu_dev *viommu, struct device *dev) 514 { 515 int ret; 516 u16 type, len; 517 size_t cur = 0; 518 size_t probe_len; 519 struct virtio_iommu_req_probe *probe; 520 struct virtio_iommu_probe_property *prop; 521 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev); 522 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev); 523 524 if (!fwspec->num_ids) 525 return -EINVAL; 526 527 probe_len = sizeof(*probe) + viommu->probe_size + 528 sizeof(struct virtio_iommu_req_tail); 529 probe = kzalloc(probe_len, GFP_KERNEL); 530 if (!probe) 531 return -ENOMEM; 532 533 probe->head.type = VIRTIO_IOMMU_T_PROBE; 534 /* 535 * For now, assume that properties of an endpoint that outputs multiple 536 * IDs are consistent. Only probe the first one. 537 */ 538 probe->endpoint = cpu_to_le32(fwspec->ids[0]); 539 540 ret = viommu_send_req_sync(viommu, probe, probe_len); 541 if (ret) 542 goto out_free; 543 544 prop = (void *)probe->properties; 545 type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK; 546 547 while (type != VIRTIO_IOMMU_PROBE_T_NONE && 548 cur < viommu->probe_size) { 549 len = le16_to_cpu(prop->length) + sizeof(*prop); 550 551 switch (type) { 552 case VIRTIO_IOMMU_PROBE_T_RESV_MEM: 553 ret = viommu_add_resv_mem(vdev, (void *)prop, len); 554 break; 555 default: 556 dev_err(dev, "unknown viommu prop 0x%x\n", type); 557 } 558 559 if (ret) 560 dev_err(dev, "failed to parse viommu prop 0x%x\n", type); 561 562 cur += len; 563 if (cur >= viommu->probe_size) 564 break; 565 566 prop = (void *)probe->properties + cur; 567 type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK; 568 } 569 570 out_free: 571 kfree(probe); 572 return ret; 573 } 574 575 static int viommu_fault_handler(struct viommu_dev *viommu, 576 struct virtio_iommu_fault *fault) 577 { 578 char *reason_str; 579 580 u8 reason = fault->reason; 581 u32 flags = le32_to_cpu(fault->flags); 582 u32 endpoint = le32_to_cpu(fault->endpoint); 583 u64 address = le64_to_cpu(fault->address); 584 585 switch (reason) { 586 case VIRTIO_IOMMU_FAULT_R_DOMAIN: 587 reason_str = "domain"; 588 break; 589 case VIRTIO_IOMMU_FAULT_R_MAPPING: 590 reason_str = "page"; 591 break; 592 case VIRTIO_IOMMU_FAULT_R_UNKNOWN: 593 default: 594 reason_str = "unknown"; 595 break; 596 } 597 598 /* TODO: find EP by ID and report_iommu_fault */ 599 if (flags & VIRTIO_IOMMU_FAULT_F_ADDRESS) 600 dev_err_ratelimited(viommu->dev, "%s fault from EP %u at %#llx [%s%s%s]\n", 601 reason_str, endpoint, address, 602 flags & VIRTIO_IOMMU_FAULT_F_READ ? "R" : "", 603 flags & VIRTIO_IOMMU_FAULT_F_WRITE ? "W" : "", 604 flags & VIRTIO_IOMMU_FAULT_F_EXEC ? "X" : ""); 605 else 606 dev_err_ratelimited(viommu->dev, "%s fault from EP %u\n", 607 reason_str, endpoint); 608 return 0; 609 } 610 611 static void viommu_event_handler(struct virtqueue *vq) 612 { 613 int ret; 614 unsigned int len; 615 struct scatterlist sg[1]; 616 struct viommu_event *evt; 617 struct viommu_dev *viommu = vq->vdev->priv; 618 619 while ((evt = virtqueue_get_buf(vq, &len)) != NULL) { 620 if (len > sizeof(*evt)) { 621 dev_err(viommu->dev, 622 "invalid event buffer (len %u != %zu)\n", 623 len, sizeof(*evt)); 624 } else if (!(evt->head & VIOMMU_FAULT_RESV_MASK)) { 625 viommu_fault_handler(viommu, &evt->fault); 626 } 627 628 sg_init_one(sg, evt, sizeof(*evt)); 629 ret = virtqueue_add_inbuf(vq, sg, 1, evt, GFP_ATOMIC); 630 if (ret) 631 dev_err(viommu->dev, "could not add event buffer\n"); 632 } 633 634 virtqueue_kick(vq); 635 } 636 637 /* IOMMU API */ 638 639 static struct iommu_domain *viommu_domain_alloc(unsigned type) 640 { 641 struct viommu_domain *vdomain; 642 643 if (type != IOMMU_DOMAIN_UNMANAGED && 644 type != IOMMU_DOMAIN_DMA && 645 type != IOMMU_DOMAIN_IDENTITY) 646 return NULL; 647 648 vdomain = kzalloc(sizeof(*vdomain), GFP_KERNEL); 649 if (!vdomain) 650 return NULL; 651 652 mutex_init(&vdomain->mutex); 653 spin_lock_init(&vdomain->mappings_lock); 654 vdomain->mappings = RB_ROOT_CACHED; 655 656 return &vdomain->domain; 657 } 658 659 static int viommu_domain_finalise(struct viommu_endpoint *vdev, 660 struct iommu_domain *domain) 661 { 662 int ret; 663 unsigned long viommu_page_size; 664 struct viommu_dev *viommu = vdev->viommu; 665 struct viommu_domain *vdomain = to_viommu_domain(domain); 666 667 viommu_page_size = 1UL << __ffs(viommu->pgsize_bitmap); 668 if (viommu_page_size > PAGE_SIZE) { 669 dev_err(vdev->dev, 670 "granule 0x%lx larger than system page size 0x%lx\n", 671 viommu_page_size, PAGE_SIZE); 672 return -EINVAL; 673 } 674 675 ret = ida_alloc_range(&viommu->domain_ids, viommu->first_domain, 676 viommu->last_domain, GFP_KERNEL); 677 if (ret < 0) 678 return ret; 679 680 vdomain->id = (unsigned int)ret; 681 682 domain->pgsize_bitmap = viommu->pgsize_bitmap; 683 domain->geometry = viommu->geometry; 684 685 vdomain->map_flags = viommu->map_flags; 686 vdomain->viommu = viommu; 687 688 if (domain->type == IOMMU_DOMAIN_IDENTITY) { 689 if (virtio_has_feature(viommu->vdev, 690 VIRTIO_IOMMU_F_BYPASS_CONFIG)) { 691 vdomain->bypass = true; 692 return 0; 693 } 694 695 ret = viommu_domain_map_identity(vdev, vdomain); 696 if (ret) { 697 ida_free(&viommu->domain_ids, vdomain->id); 698 vdomain->viommu = NULL; 699 return -EOPNOTSUPP; 700 } 701 } 702 703 return 0; 704 } 705 706 static void viommu_domain_free(struct iommu_domain *domain) 707 { 708 struct viommu_domain *vdomain = to_viommu_domain(domain); 709 710 /* Free all remaining mappings */ 711 viommu_del_mappings(vdomain, 0, ULLONG_MAX); 712 713 if (vdomain->viommu) 714 ida_free(&vdomain->viommu->domain_ids, vdomain->id); 715 716 kfree(vdomain); 717 } 718 719 static int viommu_attach_dev(struct iommu_domain *domain, struct device *dev) 720 { 721 int i; 722 int ret = 0; 723 struct virtio_iommu_req_attach req; 724 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev); 725 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev); 726 struct viommu_domain *vdomain = to_viommu_domain(domain); 727 728 mutex_lock(&vdomain->mutex); 729 if (!vdomain->viommu) { 730 /* 731 * Properly initialize the domain now that we know which viommu 732 * owns it. 733 */ 734 ret = viommu_domain_finalise(vdev, domain); 735 } else if (vdomain->viommu != vdev->viommu) { 736 dev_err(dev, "cannot attach to foreign vIOMMU\n"); 737 ret = -EXDEV; 738 } 739 mutex_unlock(&vdomain->mutex); 740 741 if (ret) 742 return ret; 743 744 /* 745 * In the virtio-iommu device, when attaching the endpoint to a new 746 * domain, it is detached from the old one and, if as a result the 747 * old domain isn't attached to any endpoint, all mappings are removed 748 * from the old domain and it is freed. 749 * 750 * In the driver the old domain still exists, and its mappings will be 751 * recreated if it gets reattached to an endpoint. Otherwise it will be 752 * freed explicitly. 753 * 754 * vdev->vdomain is protected by group->mutex 755 */ 756 if (vdev->vdomain) 757 vdev->vdomain->nr_endpoints--; 758 759 req = (struct virtio_iommu_req_attach) { 760 .head.type = VIRTIO_IOMMU_T_ATTACH, 761 .domain = cpu_to_le32(vdomain->id), 762 }; 763 764 if (vdomain->bypass) 765 req.flags |= cpu_to_le32(VIRTIO_IOMMU_ATTACH_F_BYPASS); 766 767 for (i = 0; i < fwspec->num_ids; i++) { 768 req.endpoint = cpu_to_le32(fwspec->ids[i]); 769 770 ret = viommu_send_req_sync(vdomain->viommu, &req, sizeof(req)); 771 if (ret) 772 return ret; 773 } 774 775 if (!vdomain->nr_endpoints) { 776 /* 777 * This endpoint is the first to be attached to the domain. 778 * Replay existing mappings (e.g. SW MSI). 779 */ 780 ret = viommu_replay_mappings(vdomain); 781 if (ret) 782 return ret; 783 } 784 785 vdomain->nr_endpoints++; 786 vdev->vdomain = vdomain; 787 788 return 0; 789 } 790 791 static int viommu_map_pages(struct iommu_domain *domain, unsigned long iova, 792 phys_addr_t paddr, size_t pgsize, size_t pgcount, 793 int prot, gfp_t gfp, size_t *mapped) 794 { 795 int ret; 796 u32 flags; 797 size_t size = pgsize * pgcount; 798 u64 end = iova + size - 1; 799 struct virtio_iommu_req_map map; 800 struct viommu_domain *vdomain = to_viommu_domain(domain); 801 802 flags = (prot & IOMMU_READ ? VIRTIO_IOMMU_MAP_F_READ : 0) | 803 (prot & IOMMU_WRITE ? VIRTIO_IOMMU_MAP_F_WRITE : 0) | 804 (prot & IOMMU_MMIO ? VIRTIO_IOMMU_MAP_F_MMIO : 0); 805 806 if (flags & ~vdomain->map_flags) 807 return -EINVAL; 808 809 ret = viommu_add_mapping(vdomain, iova, end, paddr, flags); 810 if (ret) 811 return ret; 812 813 map = (struct virtio_iommu_req_map) { 814 .head.type = VIRTIO_IOMMU_T_MAP, 815 .domain = cpu_to_le32(vdomain->id), 816 .virt_start = cpu_to_le64(iova), 817 .phys_start = cpu_to_le64(paddr), 818 .virt_end = cpu_to_le64(end), 819 .flags = cpu_to_le32(flags), 820 }; 821 822 if (!vdomain->nr_endpoints) 823 return 0; 824 825 ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map)); 826 if (ret) 827 viommu_del_mappings(vdomain, iova, end); 828 else if (mapped) 829 *mapped = size; 830 831 return ret; 832 } 833 834 static size_t viommu_unmap_pages(struct iommu_domain *domain, unsigned long iova, 835 size_t pgsize, size_t pgcount, 836 struct iommu_iotlb_gather *gather) 837 { 838 int ret = 0; 839 size_t unmapped; 840 struct virtio_iommu_req_unmap unmap; 841 struct viommu_domain *vdomain = to_viommu_domain(domain); 842 size_t size = pgsize * pgcount; 843 844 unmapped = viommu_del_mappings(vdomain, iova, iova + size - 1); 845 if (unmapped < size) 846 return 0; 847 848 /* Device already removed all mappings after detach. */ 849 if (!vdomain->nr_endpoints) 850 return unmapped; 851 852 unmap = (struct virtio_iommu_req_unmap) { 853 .head.type = VIRTIO_IOMMU_T_UNMAP, 854 .domain = cpu_to_le32(vdomain->id), 855 .virt_start = cpu_to_le64(iova), 856 .virt_end = cpu_to_le64(iova + unmapped - 1), 857 }; 858 859 ret = viommu_add_req(vdomain->viommu, &unmap, sizeof(unmap)); 860 return ret ? 0 : unmapped; 861 } 862 863 static phys_addr_t viommu_iova_to_phys(struct iommu_domain *domain, 864 dma_addr_t iova) 865 { 866 u64 paddr = 0; 867 unsigned long flags; 868 struct viommu_mapping *mapping; 869 struct interval_tree_node *node; 870 struct viommu_domain *vdomain = to_viommu_domain(domain); 871 872 spin_lock_irqsave(&vdomain->mappings_lock, flags); 873 node = interval_tree_iter_first(&vdomain->mappings, iova, iova); 874 if (node) { 875 mapping = container_of(node, struct viommu_mapping, iova); 876 paddr = mapping->paddr + (iova - mapping->iova.start); 877 } 878 spin_unlock_irqrestore(&vdomain->mappings_lock, flags); 879 880 return paddr; 881 } 882 883 static void viommu_iotlb_sync(struct iommu_domain *domain, 884 struct iommu_iotlb_gather *gather) 885 { 886 struct viommu_domain *vdomain = to_viommu_domain(domain); 887 888 viommu_sync_req(vdomain->viommu); 889 } 890 891 static void viommu_get_resv_regions(struct device *dev, struct list_head *head) 892 { 893 struct iommu_resv_region *entry, *new_entry, *msi = NULL; 894 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev); 895 int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO; 896 897 list_for_each_entry(entry, &vdev->resv_regions, list) { 898 if (entry->type == IOMMU_RESV_MSI) 899 msi = entry; 900 901 new_entry = kmemdup(entry, sizeof(*entry), GFP_KERNEL); 902 if (!new_entry) 903 return; 904 list_add_tail(&new_entry->list, head); 905 } 906 907 /* 908 * If the device didn't register any bypass MSI window, add a 909 * software-mapped region. 910 */ 911 if (!msi) { 912 msi = iommu_alloc_resv_region(MSI_IOVA_BASE, MSI_IOVA_LENGTH, 913 prot, IOMMU_RESV_SW_MSI); 914 if (!msi) 915 return; 916 917 list_add_tail(&msi->list, head); 918 } 919 920 iommu_dma_get_resv_regions(dev, head); 921 } 922 923 static struct iommu_ops viommu_ops; 924 static struct virtio_driver virtio_iommu_drv; 925 926 static int viommu_match_node(struct device *dev, const void *data) 927 { 928 return dev->parent->fwnode == data; 929 } 930 931 static struct viommu_dev *viommu_get_by_fwnode(struct fwnode_handle *fwnode) 932 { 933 struct device *dev = driver_find_device(&virtio_iommu_drv.driver, NULL, 934 fwnode, viommu_match_node); 935 put_device(dev); 936 937 return dev ? dev_to_virtio(dev)->priv : NULL; 938 } 939 940 static struct iommu_device *viommu_probe_device(struct device *dev) 941 { 942 int ret; 943 struct viommu_endpoint *vdev; 944 struct viommu_dev *viommu = NULL; 945 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev); 946 947 if (!fwspec || fwspec->ops != &viommu_ops) 948 return ERR_PTR(-ENODEV); 949 950 viommu = viommu_get_by_fwnode(fwspec->iommu_fwnode); 951 if (!viommu) 952 return ERR_PTR(-ENODEV); 953 954 vdev = kzalloc(sizeof(*vdev), GFP_KERNEL); 955 if (!vdev) 956 return ERR_PTR(-ENOMEM); 957 958 vdev->dev = dev; 959 vdev->viommu = viommu; 960 INIT_LIST_HEAD(&vdev->resv_regions); 961 dev_iommu_priv_set(dev, vdev); 962 963 if (viommu->probe_size) { 964 /* Get additional information for this endpoint */ 965 ret = viommu_probe_endpoint(viommu, dev); 966 if (ret) 967 goto err_free_dev; 968 } 969 970 return &viommu->iommu; 971 972 err_free_dev: 973 iommu_put_resv_regions(dev, &vdev->resv_regions); 974 kfree(vdev); 975 976 return ERR_PTR(ret); 977 } 978 979 static void viommu_probe_finalize(struct device *dev) 980 { 981 #ifndef CONFIG_ARCH_HAS_SETUP_DMA_OPS 982 /* First clear the DMA ops in case we're switching from a DMA domain */ 983 set_dma_ops(dev, NULL); 984 iommu_setup_dma_ops(dev, 0, U64_MAX); 985 #endif 986 } 987 988 static void viommu_release_device(struct device *dev) 989 { 990 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev); 991 992 iommu_put_resv_regions(dev, &vdev->resv_regions); 993 kfree(vdev); 994 } 995 996 static struct iommu_group *viommu_device_group(struct device *dev) 997 { 998 if (dev_is_pci(dev)) 999 return pci_device_group(dev); 1000 else 1001 return generic_device_group(dev); 1002 } 1003 1004 static int viommu_of_xlate(struct device *dev, struct of_phandle_args *args) 1005 { 1006 return iommu_fwspec_add_ids(dev, args->args, 1); 1007 } 1008 1009 static bool viommu_capable(enum iommu_cap cap) 1010 { 1011 switch (cap) { 1012 case IOMMU_CAP_CACHE_COHERENCY: 1013 return true; 1014 default: 1015 return false; 1016 } 1017 } 1018 1019 static struct iommu_ops viommu_ops = { 1020 .capable = viommu_capable, 1021 .domain_alloc = viommu_domain_alloc, 1022 .probe_device = viommu_probe_device, 1023 .probe_finalize = viommu_probe_finalize, 1024 .release_device = viommu_release_device, 1025 .device_group = viommu_device_group, 1026 .get_resv_regions = viommu_get_resv_regions, 1027 .of_xlate = viommu_of_xlate, 1028 .owner = THIS_MODULE, 1029 .default_domain_ops = &(const struct iommu_domain_ops) { 1030 .attach_dev = viommu_attach_dev, 1031 .map_pages = viommu_map_pages, 1032 .unmap_pages = viommu_unmap_pages, 1033 .iova_to_phys = viommu_iova_to_phys, 1034 .iotlb_sync = viommu_iotlb_sync, 1035 .free = viommu_domain_free, 1036 } 1037 }; 1038 1039 static int viommu_init_vqs(struct viommu_dev *viommu) 1040 { 1041 struct virtio_device *vdev = dev_to_virtio(viommu->dev); 1042 const char *names[] = { "request", "event" }; 1043 vq_callback_t *callbacks[] = { 1044 NULL, /* No async requests */ 1045 viommu_event_handler, 1046 }; 1047 1048 return virtio_find_vqs(vdev, VIOMMU_NR_VQS, viommu->vqs, callbacks, 1049 names, NULL); 1050 } 1051 1052 static int viommu_fill_evtq(struct viommu_dev *viommu) 1053 { 1054 int i, ret; 1055 struct scatterlist sg[1]; 1056 struct viommu_event *evts; 1057 struct virtqueue *vq = viommu->vqs[VIOMMU_EVENT_VQ]; 1058 size_t nr_evts = vq->num_free; 1059 1060 viommu->evts = evts = devm_kmalloc_array(viommu->dev, nr_evts, 1061 sizeof(*evts), GFP_KERNEL); 1062 if (!evts) 1063 return -ENOMEM; 1064 1065 for (i = 0; i < nr_evts; i++) { 1066 sg_init_one(sg, &evts[i], sizeof(*evts)); 1067 ret = virtqueue_add_inbuf(vq, sg, 1, &evts[i], GFP_KERNEL); 1068 if (ret) 1069 return ret; 1070 } 1071 1072 return 0; 1073 } 1074 1075 static int viommu_probe(struct virtio_device *vdev) 1076 { 1077 struct device *parent_dev = vdev->dev.parent; 1078 struct viommu_dev *viommu = NULL; 1079 struct device *dev = &vdev->dev; 1080 u64 input_start = 0; 1081 u64 input_end = -1UL; 1082 int ret; 1083 1084 if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1) || 1085 !virtio_has_feature(vdev, VIRTIO_IOMMU_F_MAP_UNMAP)) 1086 return -ENODEV; 1087 1088 viommu = devm_kzalloc(dev, sizeof(*viommu), GFP_KERNEL); 1089 if (!viommu) 1090 return -ENOMEM; 1091 1092 spin_lock_init(&viommu->request_lock); 1093 ida_init(&viommu->domain_ids); 1094 viommu->dev = dev; 1095 viommu->vdev = vdev; 1096 INIT_LIST_HEAD(&viommu->requests); 1097 1098 ret = viommu_init_vqs(viommu); 1099 if (ret) 1100 return ret; 1101 1102 virtio_cread_le(vdev, struct virtio_iommu_config, page_size_mask, 1103 &viommu->pgsize_bitmap); 1104 1105 if (!viommu->pgsize_bitmap) { 1106 ret = -EINVAL; 1107 goto err_free_vqs; 1108 } 1109 1110 viommu->map_flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE; 1111 viommu->last_domain = ~0U; 1112 1113 /* Optional features */ 1114 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE, 1115 struct virtio_iommu_config, input_range.start, 1116 &input_start); 1117 1118 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE, 1119 struct virtio_iommu_config, input_range.end, 1120 &input_end); 1121 1122 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE, 1123 struct virtio_iommu_config, domain_range.start, 1124 &viommu->first_domain); 1125 1126 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE, 1127 struct virtio_iommu_config, domain_range.end, 1128 &viommu->last_domain); 1129 1130 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_PROBE, 1131 struct virtio_iommu_config, probe_size, 1132 &viommu->probe_size); 1133 1134 viommu->geometry = (struct iommu_domain_geometry) { 1135 .aperture_start = input_start, 1136 .aperture_end = input_end, 1137 .force_aperture = true, 1138 }; 1139 1140 if (virtio_has_feature(vdev, VIRTIO_IOMMU_F_MMIO)) 1141 viommu->map_flags |= VIRTIO_IOMMU_MAP_F_MMIO; 1142 1143 viommu_ops.pgsize_bitmap = viommu->pgsize_bitmap; 1144 1145 virtio_device_ready(vdev); 1146 1147 /* Populate the event queue with buffers */ 1148 ret = viommu_fill_evtq(viommu); 1149 if (ret) 1150 goto err_free_vqs; 1151 1152 ret = iommu_device_sysfs_add(&viommu->iommu, dev, NULL, "%s", 1153 virtio_bus_name(vdev)); 1154 if (ret) 1155 goto err_free_vqs; 1156 1157 iommu_device_register(&viommu->iommu, &viommu_ops, parent_dev); 1158 1159 #ifdef CONFIG_PCI 1160 if (pci_bus_type.iommu_ops != &viommu_ops) { 1161 ret = bus_set_iommu(&pci_bus_type, &viommu_ops); 1162 if (ret) 1163 goto err_unregister; 1164 } 1165 #endif 1166 #ifdef CONFIG_ARM_AMBA 1167 if (amba_bustype.iommu_ops != &viommu_ops) { 1168 ret = bus_set_iommu(&amba_bustype, &viommu_ops); 1169 if (ret) 1170 goto err_unregister; 1171 } 1172 #endif 1173 if (platform_bus_type.iommu_ops != &viommu_ops) { 1174 ret = bus_set_iommu(&platform_bus_type, &viommu_ops); 1175 if (ret) 1176 goto err_unregister; 1177 } 1178 1179 vdev->priv = viommu; 1180 1181 dev_info(dev, "input address: %u bits\n", 1182 order_base_2(viommu->geometry.aperture_end)); 1183 dev_info(dev, "page mask: %#llx\n", viommu->pgsize_bitmap); 1184 1185 return 0; 1186 1187 err_unregister: 1188 iommu_device_sysfs_remove(&viommu->iommu); 1189 iommu_device_unregister(&viommu->iommu); 1190 err_free_vqs: 1191 vdev->config->del_vqs(vdev); 1192 1193 return ret; 1194 } 1195 1196 static void viommu_remove(struct virtio_device *vdev) 1197 { 1198 struct viommu_dev *viommu = vdev->priv; 1199 1200 iommu_device_sysfs_remove(&viommu->iommu); 1201 iommu_device_unregister(&viommu->iommu); 1202 1203 /* Stop all virtqueues */ 1204 virtio_reset_device(vdev); 1205 vdev->config->del_vqs(vdev); 1206 1207 dev_info(&vdev->dev, "device removed\n"); 1208 } 1209 1210 static void viommu_config_changed(struct virtio_device *vdev) 1211 { 1212 dev_warn(&vdev->dev, "config changed\n"); 1213 } 1214 1215 static unsigned int features[] = { 1216 VIRTIO_IOMMU_F_MAP_UNMAP, 1217 VIRTIO_IOMMU_F_INPUT_RANGE, 1218 VIRTIO_IOMMU_F_DOMAIN_RANGE, 1219 VIRTIO_IOMMU_F_PROBE, 1220 VIRTIO_IOMMU_F_MMIO, 1221 VIRTIO_IOMMU_F_BYPASS_CONFIG, 1222 }; 1223 1224 static struct virtio_device_id id_table[] = { 1225 { VIRTIO_ID_IOMMU, VIRTIO_DEV_ANY_ID }, 1226 { 0 }, 1227 }; 1228 MODULE_DEVICE_TABLE(virtio, id_table); 1229 1230 static struct virtio_driver virtio_iommu_drv = { 1231 .driver.name = KBUILD_MODNAME, 1232 .driver.owner = THIS_MODULE, 1233 .id_table = id_table, 1234 .feature_table = features, 1235 .feature_table_size = ARRAY_SIZE(features), 1236 .probe = viommu_probe, 1237 .remove = viommu_remove, 1238 .config_changed = viommu_config_changed, 1239 }; 1240 1241 module_virtio_driver(virtio_iommu_drv); 1242 1243 MODULE_DESCRIPTION("Virtio IOMMU driver"); 1244 MODULE_AUTHOR("Jean-Philippe Brucker <jean-philippe.brucker@arm.com>"); 1245 MODULE_LICENSE("GPL v2"); 1246