virtio-iommu.c

   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/delay.h>
  11#include <linux/dma-map-ops.h>
  12#include <linux/freezer.h>
  13#include <linux/interval_tree.h>
  14#include <linux/iommu.h>
  15#include <linux/module.h>
  16#include <linux/of.h>
  17#include <linux/pci.h>
  18#include <linux/virtio.h>
  19#include <linux/virtio_config.h>
  20#include <linux/virtio_ids.h>
  21#include <linux/wait.h>
  22
  23#include <uapi/linux/virtio_iommu.h>
  24
  25#include "dma-iommu.h"
  26
  27#define MSI_IOVA_BASE			0x8000000
  28#define MSI_IOVA_LENGTH			0x100000
  29
  30#define VIOMMU_REQUEST_VQ		0
  31#define VIOMMU_EVENT_VQ			1
  32#define VIOMMU_NR_VQS			2
  33
  34struct viommu_dev {
  35	struct iommu_device		iommu;
  36	struct device			*dev;
  37	struct virtio_device		*vdev;
  38
  39	struct ida			domain_ids;
  40
  41	struct virtqueue		*vqs[VIOMMU_NR_VQS];
  42	spinlock_t			request_lock;
  43	struct list_head		requests;
  44	void				*evts;
  45
  46	/* Device configuration */
  47	struct iommu_domain_geometry	geometry;
  48	u64				pgsize_bitmap;
  49	u32				first_domain;
  50	u32				last_domain;
  51	u32				identity_domain_id;
  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	unsigned int			id;
  67	u32				map_flags;
  68
  69	spinlock_t			mappings_lock;
  70	struct rb_root_cached		mappings;
  71
  72	unsigned long			nr_endpoints;
  73};
  74
  75struct viommu_endpoint {
  76	struct device			*dev;
  77	struct viommu_dev		*viommu;
  78	struct viommu_domain		*vdomain;
  79	struct list_head		resv_regions;
  80};
  81
  82struct viommu_request {
  83	struct list_head		list;
  84	void				*writeback;
  85	unsigned int			write_offset;
  86	unsigned int			len;
  87	char				buf[] __counted_by(len);
  88};
  89
  90#define VIOMMU_FAULT_RESV_MASK		0xffffff00
  91
  92struct viommu_event {
  93	union {
  94		u32			head;
  95		struct virtio_iommu_fault fault;
  96	};
  97};
  98
  99static struct viommu_domain viommu_identity_domain;
 100
 101#define to_viommu_domain(domain)	\
 102	container_of(domain, struct viommu_domain, domain)
 103
 104static 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
 130static 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
 137static 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 */
 155static 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
 188static 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 */
 218static 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(struct_size(req, buf, 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
 260err_free:
 261	kfree(req);
 262	return ret;
 263}
 264
 265static 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 */
 283static 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);
 304out_unlock:
 305	spin_unlock_irqrestore(&viommu->request_lock, flags);
 306	return ret;
 307}
 308
 309static int viommu_send_attach_req(struct viommu_dev *viommu, struct device *dev,
 310				  struct virtio_iommu_req_attach *req)
 311{
 312	int ret;
 313	unsigned int i;
 314	struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
 315
 316	for (i = 0; i < fwspec->num_ids; i++) {
 317		req->endpoint = cpu_to_le32(fwspec->ids[i]);
 318		ret = viommu_send_req_sync(viommu, req, sizeof(*req));
 319		if (ret)
 320			return ret;
 321	}
 322	return 0;
 323}
 324
 325/*
 326 * viommu_add_mapping - add a mapping to the internal tree
 327 *
 328 * On success, return the new mapping. Otherwise return NULL.
 329 */
 330static int viommu_add_mapping(struct viommu_domain *vdomain, u64 iova, u64 end,
 331			      phys_addr_t paddr, u32 flags)
 332{
 333	unsigned long irqflags;
 334	struct viommu_mapping *mapping;
 335
 336	mapping = kzalloc(sizeof(*mapping), GFP_ATOMIC);
 337	if (!mapping)
 338		return -ENOMEM;
 339
 340	mapping->paddr		= paddr;
 341	mapping->iova.start	= iova;
 342	mapping->iova.last	= end;
 343	mapping->flags		= flags;
 344
 345	spin_lock_irqsave(&vdomain->mappings_lock, irqflags);
 346	interval_tree_insert(&mapping->iova, &vdomain->mappings);
 347	spin_unlock_irqrestore(&vdomain->mappings_lock, irqflags);
 348
 349	return 0;
 350}
 351
 352/*
 353 * viommu_del_mappings - remove mappings from the internal tree
 354 *
 355 * @vdomain: the domain
 356 * @iova: start of the range
 357 * @end: end of the range
 358 *
 359 * On success, returns the number of unmapped bytes
 360 */
 361static size_t viommu_del_mappings(struct viommu_domain *vdomain,
 362				  u64 iova, u64 end)
 363{
 364	size_t unmapped = 0;
 365	unsigned long flags;
 366	struct viommu_mapping *mapping = NULL;
 367	struct interval_tree_node *node, *next;
 368
 369	spin_lock_irqsave(&vdomain->mappings_lock, flags);
 370	next = interval_tree_iter_first(&vdomain->mappings, iova, end);
 371	while (next) {
 372		node = next;
 373		mapping = container_of(node, struct viommu_mapping, iova);
 374		next = interval_tree_iter_next(node, iova, end);
 375
 376		/* Trying to split a mapping? */
 377		if (mapping->iova.start < iova)
 378			break;
 379
 380		/*
 381		 * Virtio-iommu doesn't allow UNMAP to split a mapping created
 382		 * with a single MAP request, so remove the full mapping.
 383		 */
 384		unmapped += mapping->iova.last - mapping->iova.start + 1;
 385
 386		interval_tree_remove(node, &vdomain->mappings);
 387		kfree(mapping);
 388	}
 389	spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
 390
 391	return unmapped;
 392}
 393
 394/*
 395 * Fill the domain with identity mappings, skipping the device's reserved
 396 * regions.
 397 */
 398static int viommu_domain_map_identity(struct viommu_endpoint *vdev,
 399				      struct viommu_domain *vdomain)
 400{
 401	int ret;
 402	struct iommu_resv_region *resv;
 403	u64 iova = vdomain->domain.geometry.aperture_start;
 404	u64 limit = vdomain->domain.geometry.aperture_end;
 405	u32 flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE;
 406	unsigned long granule = 1UL << __ffs(vdomain->domain.pgsize_bitmap);
 407
 408	iova = ALIGN(iova, granule);
 409	limit = ALIGN_DOWN(limit + 1, granule) - 1;
 410
 411	list_for_each_entry(resv, &vdev->resv_regions, list) {
 412		u64 resv_start = ALIGN_DOWN(resv->start, granule);
 413		u64 resv_end = ALIGN(resv->start + resv->length, granule) - 1;
 414
 415		if (resv_end < iova || resv_start > limit)
 416			/* No overlap */
 417			continue;
 418
 419		if (resv_start > iova) {
 420			ret = viommu_add_mapping(vdomain, iova, resv_start - 1,
 421						 (phys_addr_t)iova, flags);
 422			if (ret)
 423				goto err_unmap;
 424		}
 425
 426		if (resv_end >= limit)
 427			return 0;
 428
 429		iova = resv_end + 1;
 430	}
 431
 432	ret = viommu_add_mapping(vdomain, iova, limit, (phys_addr_t)iova,
 433				 flags);
 434	if (ret)
 435		goto err_unmap;
 436	return 0;
 437
 438err_unmap:
 439	viommu_del_mappings(vdomain, 0, iova);
 440	return ret;
 441}
 442
 443/*
 444 * viommu_replay_mappings - re-send MAP requests
 445 *
 446 * When reattaching a domain that was previously detached from all endpoints,
 447 * mappings were deleted from the device. Re-create the mappings available in
 448 * the internal tree.
 449 */
 450static int viommu_replay_mappings(struct viommu_domain *vdomain)
 451{
 452	int ret = 0;
 453	unsigned long flags;
 454	struct viommu_mapping *mapping;
 455	struct interval_tree_node *node;
 456	struct virtio_iommu_req_map map;
 457
 458	spin_lock_irqsave(&vdomain->mappings_lock, flags);
 459	node = interval_tree_iter_first(&vdomain->mappings, 0, -1UL);
 460	while (node) {
 461		mapping = container_of(node, struct viommu_mapping, iova);
 462		map = (struct virtio_iommu_req_map) {
 463			.head.type	= VIRTIO_IOMMU_T_MAP,
 464			.domain		= cpu_to_le32(vdomain->id),
 465			.virt_start	= cpu_to_le64(mapping->iova.start),
 466			.virt_end	= cpu_to_le64(mapping->iova.last),
 467			.phys_start	= cpu_to_le64(mapping->paddr),
 468			.flags		= cpu_to_le32(mapping->flags),
 469		};
 470
 471		ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map));
 472		if (ret)
 473			break;
 474
 475		node = interval_tree_iter_next(node, 0, -1UL);
 476	}
 477	spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
 478
 479	return ret;
 480}
 481
 482static int viommu_add_resv_mem(struct viommu_endpoint *vdev,
 483			       struct virtio_iommu_probe_resv_mem *mem,
 484			       size_t len)
 485{
 486	size_t size;
 487	u64 start64, end64;
 488	phys_addr_t start, end;
 489	struct iommu_resv_region *region = NULL, *next;
 490	unsigned long prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
 491
 492	start = start64 = le64_to_cpu(mem->start);
 493	end = end64 = le64_to_cpu(mem->end);
 494	size = end64 - start64 + 1;
 495
 496	/* Catch any overflow, including the unlikely end64 - start64 + 1 = 0 */
 497	if (start != start64 || end != end64 || size < end64 - start64)
 498		return -EOVERFLOW;
 499
 500	if (len < sizeof(*mem))
 501		return -EINVAL;
 502
 503	switch (mem->subtype) {
 504	default:
 505		dev_warn(vdev->dev, "unknown resv mem subtype 0x%x\n",
 506			 mem->subtype);
 507		fallthrough;
 508	case VIRTIO_IOMMU_RESV_MEM_T_RESERVED:
 509		region = iommu_alloc_resv_region(start, size, 0,
 510						 IOMMU_RESV_RESERVED,
 511						 GFP_KERNEL);
 512		break;
 513	case VIRTIO_IOMMU_RESV_MEM_T_MSI:
 514		region = iommu_alloc_resv_region(start, size, prot,
 515						 IOMMU_RESV_MSI,
 516						 GFP_KERNEL);
 517		break;
 518	}
 519	if (!region)
 520		return -ENOMEM;
 521
 522	/* Keep the list sorted */
 523	list_for_each_entry(next, &vdev->resv_regions, list) {
 524		if (next->start > region->start)
 525			break;
 526	}
 527	list_add_tail(&region->list, &next->list);
 528	return 0;
 529}
 530
 531static int viommu_probe_endpoint(struct viommu_dev *viommu, struct device *dev)
 532{
 533	int ret;
 534	u16 type, len;
 535	size_t cur = 0;
 536	size_t probe_len;
 537	struct virtio_iommu_req_probe *probe;
 538	struct virtio_iommu_probe_property *prop;
 539	struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
 540	struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
 541
 542	if (!fwspec->num_ids)
 543		return -EINVAL;
 544
 545	probe_len = sizeof(*probe) + viommu->probe_size +
 546		    sizeof(struct virtio_iommu_req_tail);
 547	probe = kzalloc(probe_len, GFP_KERNEL);
 548	if (!probe)
 549		return -ENOMEM;
 550
 551	probe->head.type = VIRTIO_IOMMU_T_PROBE;
 552	/*
 553	 * For now, assume that properties of an endpoint that outputs multiple
 554	 * IDs are consistent. Only probe the first one.
 555	 */
 556	probe->endpoint = cpu_to_le32(fwspec->ids[0]);
 557
 558	ret = viommu_send_req_sync(viommu, probe, probe_len);
 559	if (ret)
 560		goto out_free;
 561
 562	prop = (void *)probe->properties;
 563	type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK;
 564
 565	while (type != VIRTIO_IOMMU_PROBE_T_NONE &&
 566	       cur < viommu->probe_size) {
 567		len = le16_to_cpu(prop->length) + sizeof(*prop);
 568
 569		switch (type) {
 570		case VIRTIO_IOMMU_PROBE_T_RESV_MEM:
 571			ret = viommu_add_resv_mem(vdev, (void *)prop, len);
 572			break;
 573		default:
 574			dev_err(dev, "unknown viommu prop 0x%x\n", type);
 575		}
 576
 577		if (ret)
 578			dev_err(dev, "failed to parse viommu prop 0x%x\n", type);
 579
 580		cur += len;
 581		if (cur >= viommu->probe_size)
 582			break;
 583
 584		prop = (void *)probe->properties + cur;
 585		type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK;
 586	}
 587
 588out_free:
 589	kfree(probe);
 590	return ret;
 591}
 592
 593static int viommu_fault_handler(struct viommu_dev *viommu,
 594				struct virtio_iommu_fault *fault)
 595{
 596	char *reason_str;
 597
 598	u8 reason	= fault->reason;
 599	u32 flags	= le32_to_cpu(fault->flags);
 600	u32 endpoint	= le32_to_cpu(fault->endpoint);
 601	u64 address	= le64_to_cpu(fault->address);
 602
 603	switch (reason) {
 604	case VIRTIO_IOMMU_FAULT_R_DOMAIN:
 605		reason_str = "domain";
 606		break;
 607	case VIRTIO_IOMMU_FAULT_R_MAPPING:
 608		reason_str = "page";
 609		break;
 610	case VIRTIO_IOMMU_FAULT_R_UNKNOWN:
 611	default:
 612		reason_str = "unknown";
 613		break;
 614	}
 615
 616	/* TODO: find EP by ID and report_iommu_fault */
 617	if (flags & VIRTIO_IOMMU_FAULT_F_ADDRESS)
 618		dev_err_ratelimited(viommu->dev, "%s fault from EP %u at %#llx [%s%s%s]\n",
 619				    reason_str, endpoint, address,
 620				    flags & VIRTIO_IOMMU_FAULT_F_READ ? "R" : "",
 621				    flags & VIRTIO_IOMMU_FAULT_F_WRITE ? "W" : "",
 622				    flags & VIRTIO_IOMMU_FAULT_F_EXEC ? "X" : "");
 623	else
 624		dev_err_ratelimited(viommu->dev, "%s fault from EP %u\n",
 625				    reason_str, endpoint);
 626	return 0;
 627}
 628
 629static void viommu_event_handler(struct virtqueue *vq)
 630{
 631	int ret;
 632	unsigned int len;
 633	struct scatterlist sg[1];
 634	struct viommu_event *evt;
 635	struct viommu_dev *viommu = vq->vdev->priv;
 636
 637	while ((evt = virtqueue_get_buf(vq, &len)) != NULL) {
 638		if (len > sizeof(*evt)) {
 639			dev_err(viommu->dev,
 640				"invalid event buffer (len %u != %zu)\n",
 641				len, sizeof(*evt));
 642		} else if (!(evt->head & VIOMMU_FAULT_RESV_MASK)) {
 643			viommu_fault_handler(viommu, &evt->fault);
 644		}
 645
 646		sg_init_one(sg, evt, sizeof(*evt));
 647		ret = virtqueue_add_inbuf(vq, sg, 1, evt, GFP_ATOMIC);
 648		if (ret)
 649			dev_err(viommu->dev, "could not add event buffer\n");
 650	}
 651
 652	virtqueue_kick(vq);
 653}
 654
 655/* IOMMU API */
 656
 657static struct iommu_domain *viommu_domain_alloc_paging(struct device *dev)
 658{
 659	struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
 660	struct viommu_dev *viommu = vdev->viommu;
 661	unsigned long viommu_page_size;
 662	struct viommu_domain *vdomain;
 663	int ret;
 664
 665	viommu_page_size = 1UL << __ffs(viommu->pgsize_bitmap);
 666	if (viommu_page_size > PAGE_SIZE) {
 667		dev_err(vdev->dev,
 668			"granule 0x%lx larger than system page size 0x%lx\n",
 669			viommu_page_size, PAGE_SIZE);
 670		return ERR_PTR(-ENODEV);
 671	}
 672
 673	vdomain = kzalloc(sizeof(*vdomain), GFP_KERNEL);
 674	if (!vdomain)
 675		return ERR_PTR(-ENOMEM);
 676
 677	spin_lock_init(&vdomain->mappings_lock);
 678	vdomain->mappings = RB_ROOT_CACHED;
 679
 680	ret = ida_alloc_range(&viommu->domain_ids, viommu->first_domain,
 681			      viommu->last_domain, GFP_KERNEL);
 682	if (ret < 0) {
 683		kfree(vdomain);
 684		return ERR_PTR(ret);
 685	}
 686
 687	vdomain->id = (unsigned int)ret;
 688
 689	vdomain->domain.pgsize_bitmap = viommu->pgsize_bitmap;
 690	vdomain->domain.geometry = viommu->geometry;
 691
 692	vdomain->map_flags = viommu->map_flags;
 693	vdomain->viommu = viommu;
 694
 695	return &vdomain->domain;
 696}
 697
 698static void viommu_domain_free(struct iommu_domain *domain)
 699{
 700	struct viommu_domain *vdomain = to_viommu_domain(domain);
 701
 702	/* Free all remaining mappings */
 703	viommu_del_mappings(vdomain, 0, ULLONG_MAX);
 704
 705	if (vdomain->viommu)
 706		ida_free(&vdomain->viommu->domain_ids, vdomain->id);
 707
 708	kfree(vdomain);
 709}
 710
 711static struct iommu_domain *viommu_domain_alloc_identity(struct device *dev)
 712{
 713	struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
 714	struct iommu_domain *domain;
 715	int ret;
 716
 717	if (virtio_has_feature(vdev->viommu->vdev,
 718			       VIRTIO_IOMMU_F_BYPASS_CONFIG))
 719		return &viommu_identity_domain.domain;
 720
 721	domain = viommu_domain_alloc_paging(dev);
 722	if (IS_ERR(domain))
 723		return domain;
 724
 725	ret = viommu_domain_map_identity(vdev, to_viommu_domain(domain));
 726	if (ret) {
 727		viommu_domain_free(domain);
 728		return ERR_PTR(ret);
 729	}
 730	return domain;
 731}
 732
 733static int viommu_attach_dev(struct iommu_domain *domain, struct device *dev,
 734			     struct iommu_domain *old)
 735{
 736	int ret = 0;
 737	struct virtio_iommu_req_attach req;
 738	struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
 739	struct viommu_domain *vdomain = to_viommu_domain(domain);
 740
 741	if (vdomain->viommu != vdev->viommu)
 742		return -EINVAL;
 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	ret = viommu_send_attach_req(vdomain->viommu, dev, &req);
 765	if (ret)
 766		return ret;
 767
 768	if (!vdomain->nr_endpoints) {
 769		/*
 770		 * This endpoint is the first to be attached to the domain.
 771		 * Replay existing mappings (e.g. SW MSI).
 772		 */
 773		ret = viommu_replay_mappings(vdomain);
 774		if (ret)
 775			return ret;
 776	}
 777
 778	vdomain->nr_endpoints++;
 779	vdev->vdomain = vdomain;
 780
 781	return 0;
 782}
 783
 784static int viommu_attach_identity_domain(struct iommu_domain *domain,
 785					 struct device *dev,
 786					 struct iommu_domain *old)
 787{
 788	int ret = 0;
 789	struct virtio_iommu_req_attach req;
 790	struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
 791	struct viommu_domain *vdomain = to_viommu_domain(domain);
 792
 793	req = (struct virtio_iommu_req_attach) {
 794		.head.type	= VIRTIO_IOMMU_T_ATTACH,
 795		.domain		= cpu_to_le32(vdev->viommu->identity_domain_id),
 796		.flags          = cpu_to_le32(VIRTIO_IOMMU_ATTACH_F_BYPASS),
 797	};
 798
 799	ret = viommu_send_attach_req(vdev->viommu, dev, &req);
 800	if (ret)
 801		return ret;
 802
 803	if (vdev->vdomain)
 804		vdev->vdomain->nr_endpoints--;
 805	vdomain->nr_endpoints++;
 806	vdev->vdomain = vdomain;
 807	return 0;
 808}
 809
 810static struct viommu_domain viommu_identity_domain = {
 811	.domain = {
 812		.type = IOMMU_DOMAIN_IDENTITY,
 813		.ops = &(const struct iommu_domain_ops) {
 814			.attach_dev = viommu_attach_identity_domain,
 815		},
 816	},
 817};
 818
 819static void viommu_detach_dev(struct viommu_endpoint *vdev)
 820{
 821	int i;
 822	struct virtio_iommu_req_detach req;
 823	struct viommu_domain *vdomain = vdev->vdomain;
 824	struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(vdev->dev);
 825
 826	if (!vdomain)
 827		return;
 828
 829	req = (struct virtio_iommu_req_detach) {
 830		.head.type	= VIRTIO_IOMMU_T_DETACH,
 831		.domain		= cpu_to_le32(vdomain->id),
 832	};
 833
 834	for (i = 0; i < fwspec->num_ids; i++) {
 835		req.endpoint = cpu_to_le32(fwspec->ids[i]);
 836		WARN_ON(viommu_send_req_sync(vdev->viommu, &req, sizeof(req)));
 837	}
 838	vdomain->nr_endpoints--;
 839	vdev->vdomain = NULL;
 840}
 841
 842static int viommu_map_pages(struct iommu_domain *domain, unsigned long iova,
 843			    phys_addr_t paddr, size_t pgsize, size_t pgcount,
 844			    int prot, gfp_t gfp, size_t *mapped)
 845{
 846	int ret;
 847	u32 flags;
 848	size_t size = pgsize * pgcount;
 849	u64 end = iova + size - 1;
 850	struct virtio_iommu_req_map map;
 851	struct viommu_domain *vdomain = to_viommu_domain(domain);
 852
 853	flags = (prot & IOMMU_READ ? VIRTIO_IOMMU_MAP_F_READ : 0) |
 854		(prot & IOMMU_WRITE ? VIRTIO_IOMMU_MAP_F_WRITE : 0) |
 855		(prot & IOMMU_MMIO ? VIRTIO_IOMMU_MAP_F_MMIO : 0);
 856
 857	if (flags & ~vdomain->map_flags)
 858		return -EINVAL;
 859
 860	ret = viommu_add_mapping(vdomain, iova, end, paddr, flags);
 861	if (ret)
 862		return ret;
 863
 864	if (vdomain->nr_endpoints) {
 865		map = (struct virtio_iommu_req_map) {
 866			.head.type	= VIRTIO_IOMMU_T_MAP,
 867			.domain		= cpu_to_le32(vdomain->id),
 868			.virt_start	= cpu_to_le64(iova),
 869			.phys_start	= cpu_to_le64(paddr),
 870			.virt_end	= cpu_to_le64(end),
 871			.flags		= cpu_to_le32(flags),
 872		};
 873
 874		ret = viommu_add_req(vdomain->viommu, &map, sizeof(map));
 875		if (ret) {
 876			viommu_del_mappings(vdomain, iova, end);
 877			return ret;
 878		}
 879	}
 880	if (mapped)
 881		*mapped = size;
 882
 883	return 0;
 884}
 885
 886static size_t viommu_unmap_pages(struct iommu_domain *domain, unsigned long iova,
 887				 size_t pgsize, size_t pgcount,
 888				 struct iommu_iotlb_gather *gather)
 889{
 890	int ret = 0;
 891	size_t unmapped;
 892	struct virtio_iommu_req_unmap unmap;
 893	struct viommu_domain *vdomain = to_viommu_domain(domain);
 894	size_t size = pgsize * pgcount;
 895
 896	unmapped = viommu_del_mappings(vdomain, iova, iova + size - 1);
 897	if (unmapped < size)
 898		return 0;
 899
 900	/* Device already removed all mappings after detach. */
 901	if (!vdomain->nr_endpoints)
 902		return unmapped;
 903
 904	unmap = (struct virtio_iommu_req_unmap) {
 905		.head.type	= VIRTIO_IOMMU_T_UNMAP,
 906		.domain		= cpu_to_le32(vdomain->id),
 907		.virt_start	= cpu_to_le64(iova),
 908		.virt_end	= cpu_to_le64(iova + unmapped - 1),
 909	};
 910
 911	ret = viommu_add_req(vdomain->viommu, &unmap, sizeof(unmap));
 912	return ret ? 0 : unmapped;
 913}
 914
 915static phys_addr_t viommu_iova_to_phys(struct iommu_domain *domain,
 916				       dma_addr_t iova)
 917{
 918	u64 paddr = 0;
 919	unsigned long flags;
 920	struct viommu_mapping *mapping;
 921	struct interval_tree_node *node;
 922	struct viommu_domain *vdomain = to_viommu_domain(domain);
 923
 924	spin_lock_irqsave(&vdomain->mappings_lock, flags);
 925	node = interval_tree_iter_first(&vdomain->mappings, iova, iova);
 926	if (node) {
 927		mapping = container_of(node, struct viommu_mapping, iova);
 928		paddr = mapping->paddr + (iova - mapping->iova.start);
 929	}
 930	spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
 931
 932	return paddr;
 933}
 934
 935static void viommu_iotlb_sync(struct iommu_domain *domain,
 936			      struct iommu_iotlb_gather *gather)
 937{
 938	struct viommu_domain *vdomain = to_viommu_domain(domain);
 939
 940	viommu_sync_req(vdomain->viommu);
 941}
 942
 943static int viommu_iotlb_sync_map(struct iommu_domain *domain,
 944				 unsigned long iova, size_t size)
 945{
 946	struct viommu_domain *vdomain = to_viommu_domain(domain);
 947
 948	/*
 949	 * May be called before the viommu is initialized including
 950	 * while creating direct mapping
 951	 */
 952	if (!vdomain->nr_endpoints)
 953		return 0;
 954	return viommu_sync_req(vdomain->viommu);
 955}
 956
 957static void viommu_flush_iotlb_all(struct iommu_domain *domain)
 958{
 959	struct viommu_domain *vdomain = to_viommu_domain(domain);
 960
 961	/*
 962	 * May be called before the viommu is initialized including
 963	 * while creating direct mapping
 964	 */
 965	if (!vdomain->nr_endpoints)
 966		return;
 967	viommu_sync_req(vdomain->viommu);
 968}
 969
 970static void viommu_get_resv_regions(struct device *dev, struct list_head *head)
 971{
 972	struct iommu_resv_region *entry, *new_entry, *msi = NULL;
 973	struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
 974	int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
 975
 976	list_for_each_entry(entry, &vdev->resv_regions, list) {
 977		if (entry->type == IOMMU_RESV_MSI)
 978			msi = entry;
 979
 980		new_entry = kmemdup(entry, sizeof(*entry), GFP_KERNEL);
 981		if (!new_entry)
 982			return;
 983		list_add_tail(&new_entry->list, head);
 984	}
 985
 986	/*
 987	 * If the device didn't register any bypass MSI window, add a
 988	 * software-mapped region.
 989	 */
 990	if (!msi) {
 991		msi = iommu_alloc_resv_region(MSI_IOVA_BASE, MSI_IOVA_LENGTH,
 992					      prot, IOMMU_RESV_SW_MSI,
 993					      GFP_KERNEL);
 994		if (!msi)
 995			return;
 996
 997		list_add_tail(&msi->list, head);
 998	}
 999
1000	iommu_dma_get_resv_regions(dev, head);
1001}
1002
1003static const struct bus_type *virtio_bus_type;
1004
1005static int viommu_match_node(struct device *dev, const void *data)
1006{
1007	return device_match_fwnode(dev->parent, data);
1008}
1009
1010static struct viommu_dev *viommu_get_by_fwnode(struct fwnode_handle *fwnode)
1011{
1012	struct device *dev = bus_find_device(virtio_bus_type, NULL, fwnode,
1013					     viommu_match_node);
1014
1015	put_device(dev);
1016
1017	return dev ? dev_to_virtio(dev)->priv : NULL;
1018}
1019
1020static struct iommu_device *viommu_probe_device(struct device *dev)
1021{
1022	int ret;
1023	struct viommu_endpoint *vdev;
1024	struct viommu_dev *viommu = NULL;
1025	struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
1026
1027	viommu = viommu_get_by_fwnode(fwspec->iommu_fwnode);
1028	if (!viommu)
1029		return ERR_PTR(-ENODEV);
1030
1031	vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
1032	if (!vdev)
1033		return ERR_PTR(-ENOMEM);
1034
1035	vdev->dev = dev;
1036	vdev->viommu = viommu;
1037	INIT_LIST_HEAD(&vdev->resv_regions);
1038	dev_iommu_priv_set(dev, vdev);
1039
1040	if (viommu->probe_size) {
1041		/* Get additional information for this endpoint */
1042		ret = viommu_probe_endpoint(viommu, dev);
1043		if (ret)
1044			goto err_free_dev;
1045	}
1046
1047	return &viommu->iommu;
1048
1049err_free_dev:
1050	iommu_put_resv_regions(dev, &vdev->resv_regions);
1051	kfree(vdev);
1052
1053	return ERR_PTR(ret);
1054}
1055
1056static void viommu_release_device(struct device *dev)
1057{
1058	struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
1059
1060	viommu_detach_dev(vdev);
1061	iommu_put_resv_regions(dev, &vdev->resv_regions);
1062	kfree(vdev);
1063}
1064
1065static struct iommu_group *viommu_device_group(struct device *dev)
1066{
1067	if (dev_is_pci(dev))
1068		return pci_device_group(dev);
1069	else
1070		return generic_device_group(dev);
1071}
1072
1073static int viommu_of_xlate(struct device *dev,
1074			   const struct of_phandle_args *args)
1075{
1076	return iommu_fwspec_add_ids(dev, args->args, 1);
1077}
1078
1079static bool viommu_capable(struct device *dev, enum iommu_cap cap)
1080{
1081	switch (cap) {
1082	case IOMMU_CAP_CACHE_COHERENCY:
1083		return true;
1084	case IOMMU_CAP_DEFERRED_FLUSH:
1085		return true;
1086	default:
1087		return false;
1088	}
1089}
1090
1091static const struct iommu_ops viommu_ops = {
1092	.capable		= viommu_capable,
1093	.domain_alloc_identity	= viommu_domain_alloc_identity,
1094	.domain_alloc_paging	= viommu_domain_alloc_paging,
1095	.probe_device		= viommu_probe_device,
1096	.release_device		= viommu_release_device,
1097	.device_group		= viommu_device_group,
1098	.get_resv_regions	= viommu_get_resv_regions,
1099	.of_xlate		= viommu_of_xlate,
1100	.owner			= THIS_MODULE,
1101	.default_domain_ops = &(const struct iommu_domain_ops) {
1102		.attach_dev		= viommu_attach_dev,
1103		.map_pages		= viommu_map_pages,
1104		.unmap_pages		= viommu_unmap_pages,
1105		.iova_to_phys		= viommu_iova_to_phys,
1106		.flush_iotlb_all	= viommu_flush_iotlb_all,
1107		.iotlb_sync		= viommu_iotlb_sync,
1108		.iotlb_sync_map		= viommu_iotlb_sync_map,
1109		.free			= viommu_domain_free,
1110	}
1111};
1112
1113static int viommu_init_vqs(struct viommu_dev *viommu)
1114{
1115	struct virtio_device *vdev = dev_to_virtio(viommu->dev);
1116	struct virtqueue_info vqs_info[] = {
1117		{ "request" },
1118		{ "event", viommu_event_handler },
1119	};
1120
1121	return virtio_find_vqs(vdev, VIOMMU_NR_VQS, viommu->vqs,
1122			       vqs_info, NULL);
1123}
1124
1125static int viommu_fill_evtq(struct viommu_dev *viommu)
1126{
1127	int i, ret;
1128	struct scatterlist sg[1];
1129	struct viommu_event *evts;
1130	struct virtqueue *vq = viommu->vqs[VIOMMU_EVENT_VQ];
1131	size_t nr_evts = vq->num_free;
1132
1133	viommu->evts = evts = devm_kmalloc_array(viommu->dev, nr_evts,
1134						 sizeof(*evts), GFP_KERNEL);
1135	if (!evts)
1136		return -ENOMEM;
1137
1138	for (i = 0; i < nr_evts; i++) {
1139		sg_init_one(sg, &evts[i], sizeof(*evts));
1140		ret = virtqueue_add_inbuf(vq, sg, 1, &evts[i], GFP_KERNEL);
1141		if (ret)
1142			return ret;
1143	}
1144
1145	return 0;
1146}
1147
1148static int viommu_probe(struct virtio_device *vdev)
1149{
1150	struct device *parent_dev = vdev->dev.parent;
1151	struct viommu_dev *viommu = NULL;
1152	struct device *dev = &vdev->dev;
1153	u64 input_start = 0;
1154	u64 input_end = -1UL;
1155	int ret;
1156
1157	if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1) ||
1158	    !virtio_has_feature(vdev, VIRTIO_IOMMU_F_MAP_UNMAP))
1159		return -ENODEV;
1160
1161	viommu = devm_kzalloc(dev, sizeof(*viommu), GFP_KERNEL);
1162	if (!viommu)
1163		return -ENOMEM;
1164
1165	/* Borrow this for easy lookups later */
1166	virtio_bus_type = dev->bus;
1167
1168	spin_lock_init(&viommu->request_lock);
1169	ida_init(&viommu->domain_ids);
1170	viommu->dev = dev;
1171	viommu->vdev = vdev;
1172	INIT_LIST_HEAD(&viommu->requests);
1173
1174	ret = viommu_init_vqs(viommu);
1175	if (ret)
1176		return ret;
1177
1178	virtio_cread_le(vdev, struct virtio_iommu_config, page_size_mask,
1179			&viommu->pgsize_bitmap);
1180
1181	if (!viommu->pgsize_bitmap) {
1182		ret = -EINVAL;
1183		goto err_free_vqs;
1184	}
1185
1186	viommu->map_flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE;
1187	viommu->last_domain = ~0U;
1188
1189	/* Optional features */
1190	virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1191				struct virtio_iommu_config, input_range.start,
1192				&input_start);
1193
1194	virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1195				struct virtio_iommu_config, input_range.end,
1196				&input_end);
1197
1198	virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1199				struct virtio_iommu_config, domain_range.start,
1200				&viommu->first_domain);
1201
1202	virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1203				struct virtio_iommu_config, domain_range.end,
1204				&viommu->last_domain);
1205
1206	virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_PROBE,
1207				struct virtio_iommu_config, probe_size,
1208				&viommu->probe_size);
1209
1210	viommu->geometry = (struct iommu_domain_geometry) {
1211		.aperture_start	= input_start,
1212		.aperture_end	= input_end,
1213		.force_aperture	= true,
1214	};
1215
1216	if (virtio_has_feature(vdev, VIRTIO_IOMMU_F_MMIO))
1217		viommu->map_flags |= VIRTIO_IOMMU_MAP_F_MMIO;
1218
1219	/* Reserve an ID to use as the bypass domain */
1220	if (virtio_has_feature(viommu->vdev, VIRTIO_IOMMU_F_BYPASS_CONFIG)) {
1221		viommu->identity_domain_id = viommu->first_domain;
1222		viommu->first_domain++;
1223	}
1224
1225	virtio_device_ready(vdev);
1226
1227	/* Populate the event queue with buffers */
1228	ret = viommu_fill_evtq(viommu);
1229	if (ret)
1230		goto err_free_vqs;
1231
1232	ret = iommu_device_sysfs_add(&viommu->iommu, dev, NULL, "%s",
1233				     virtio_bus_name(vdev));
1234	if (ret)
1235		goto err_free_vqs;
1236
1237	vdev->priv = viommu;
1238
1239	iommu_device_register(&viommu->iommu, &viommu_ops, parent_dev);
1240
1241	dev_info(dev, "input address: %u bits\n",
1242		 order_base_2(viommu->geometry.aperture_end));
1243	dev_info(dev, "page mask: %#llx\n", viommu->pgsize_bitmap);
1244
1245	return 0;
1246
1247err_free_vqs:
1248	vdev->config->del_vqs(vdev);
1249
1250	return ret;
1251}
1252
1253static void viommu_remove(struct virtio_device *vdev)
1254{
1255	struct viommu_dev *viommu = vdev->priv;
1256
1257	iommu_device_sysfs_remove(&viommu->iommu);
1258	iommu_device_unregister(&viommu->iommu);
1259
1260	/* Stop all virtqueues */
1261	virtio_reset_device(vdev);
1262	vdev->config->del_vqs(vdev);
1263
1264	dev_info(&vdev->dev, "device removed\n");
1265}
1266
1267static void viommu_config_changed(struct virtio_device *vdev)
1268{
1269	dev_warn(&vdev->dev, "config changed\n");
1270}
1271
1272static unsigned int features[] = {
1273	VIRTIO_IOMMU_F_MAP_UNMAP,
1274	VIRTIO_IOMMU_F_INPUT_RANGE,
1275	VIRTIO_IOMMU_F_DOMAIN_RANGE,
1276	VIRTIO_IOMMU_F_PROBE,
1277	VIRTIO_IOMMU_F_MMIO,
1278	VIRTIO_IOMMU_F_BYPASS_CONFIG,
1279};
1280
1281static struct virtio_device_id id_table[] = {
1282	{ VIRTIO_ID_IOMMU, VIRTIO_DEV_ANY_ID },
1283	{ 0 },
1284};
1285MODULE_DEVICE_TABLE(virtio, id_table);
1286
1287static struct virtio_driver virtio_iommu_drv = {
1288	.driver.name		= KBUILD_MODNAME,
1289	.id_table		= id_table,
1290	.feature_table		= features,
1291	.feature_table_size	= ARRAY_SIZE(features),
1292	.probe			= viommu_probe,
1293	.remove			= viommu_remove,
1294	.config_changed		= viommu_config_changed,
1295};
1296
1297module_virtio_driver(virtio_iommu_drv);
1298
1299MODULE_DESCRIPTION("Virtio IOMMU driver");
1300MODULE_AUTHOR("Jean-Philippe Brucker <jean-philippe.brucker@arm.com>");
1301MODULE_LICENSE("GPL v2");