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