xref: /linux/drivers/iommu/virtio-iommu.c (revision 3c4fc7bf4c9e66fe71abcbf93f62f4ddb89b7f15)
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(&region->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