xref: /linux/drivers/iommu/virtio-iommu.c (revision 7fc7f25419f5a6b09199ba4b5026b94ef184fa79)
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_platform.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 
34 struct 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 	/* Supported MAP flags */
52 	u32				map_flags;
53 	u32				probe_size;
54 };
55 
56 struct viommu_mapping {
57 	phys_addr_t			paddr;
58 	struct interval_tree_node	iova;
59 	u32				flags;
60 };
61 
62 struct viommu_domain {
63 	struct iommu_domain		domain;
64 	struct viommu_dev		*viommu;
65 	struct mutex			mutex; /* protects viommu pointer */
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 	bool				bypass;
74 };
75 
76 struct viommu_endpoint {
77 	struct device			*dev;
78 	struct viommu_dev		*viommu;
79 	struct viommu_domain		*vdomain;
80 	struct list_head		resv_regions;
81 };
82 
83 struct 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 
93 struct 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 
103 static 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 
129 static 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 
136 static 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  */
154 static 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 
187 static 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  */
217 static 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 
259 err_free:
260 	kfree(req);
261 	return ret;
262 }
263 
264 static 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  */
282 static 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);
303 out_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  */
313 static int viommu_add_mapping(struct viommu_domain *vdomain, u64 iova, u64 end,
314 			      phys_addr_t paddr, 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	= end;
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  * @end: end of the range
341  *
342  * On success, returns the number of unmapped bytes
343  */
344 static size_t viommu_del_mappings(struct viommu_domain *vdomain,
345 				  u64 iova, u64 end)
346 {
347 	size_t unmapped = 0;
348 	unsigned long flags;
349 	struct viommu_mapping *mapping = NULL;
350 	struct interval_tree_node *node, *next;
351 
352 	spin_lock_irqsave(&vdomain->mappings_lock, flags);
353 	next = interval_tree_iter_first(&vdomain->mappings, iova, end);
354 	while (next) {
355 		node = next;
356 		mapping = container_of(node, struct viommu_mapping, iova);
357 		next = interval_tree_iter_next(node, iova, end);
358 
359 		/* Trying to split a mapping? */
360 		if (mapping->iova.start < iova)
361 			break;
362 
363 		/*
364 		 * Virtio-iommu doesn't allow UNMAP to split a mapping created
365 		 * with a single MAP request, so remove the full mapping.
366 		 */
367 		unmapped += mapping->iova.last - mapping->iova.start + 1;
368 
369 		interval_tree_remove(node, &vdomain->mappings);
370 		kfree(mapping);
371 	}
372 	spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
373 
374 	return unmapped;
375 }
376 
377 /*
378  * Fill the domain with identity mappings, skipping the device's reserved
379  * regions.
380  */
381 static int viommu_domain_map_identity(struct viommu_endpoint *vdev,
382 				      struct viommu_domain *vdomain)
383 {
384 	int ret;
385 	struct iommu_resv_region *resv;
386 	u64 iova = vdomain->domain.geometry.aperture_start;
387 	u64 limit = vdomain->domain.geometry.aperture_end;
388 	u32 flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE;
389 	unsigned long granule = 1UL << __ffs(vdomain->domain.pgsize_bitmap);
390 
391 	iova = ALIGN(iova, granule);
392 	limit = ALIGN_DOWN(limit + 1, granule) - 1;
393 
394 	list_for_each_entry(resv, &vdev->resv_regions, list) {
395 		u64 resv_start = ALIGN_DOWN(resv->start, granule);
396 		u64 resv_end = ALIGN(resv->start + resv->length, granule) - 1;
397 
398 		if (resv_end < iova || resv_start > limit)
399 			/* No overlap */
400 			continue;
401 
402 		if (resv_start > iova) {
403 			ret = viommu_add_mapping(vdomain, iova, resv_start - 1,
404 						 (phys_addr_t)iova, flags);
405 			if (ret)
406 				goto err_unmap;
407 		}
408 
409 		if (resv_end >= limit)
410 			return 0;
411 
412 		iova = resv_end + 1;
413 	}
414 
415 	ret = viommu_add_mapping(vdomain, iova, limit, (phys_addr_t)iova,
416 				 flags);
417 	if (ret)
418 		goto err_unmap;
419 	return 0;
420 
421 err_unmap:
422 	viommu_del_mappings(vdomain, 0, iova);
423 	return ret;
424 }
425 
426 /*
427  * viommu_replay_mappings - re-send MAP requests
428  *
429  * When reattaching a domain that was previously detached from all endpoints,
430  * mappings were deleted from the device. Re-create the mappings available in
431  * the internal tree.
432  */
433 static int viommu_replay_mappings(struct viommu_domain *vdomain)
434 {
435 	int ret = 0;
436 	unsigned long flags;
437 	struct viommu_mapping *mapping;
438 	struct interval_tree_node *node;
439 	struct virtio_iommu_req_map map;
440 
441 	spin_lock_irqsave(&vdomain->mappings_lock, flags);
442 	node = interval_tree_iter_first(&vdomain->mappings, 0, -1UL);
443 	while (node) {
444 		mapping = container_of(node, struct viommu_mapping, iova);
445 		map = (struct virtio_iommu_req_map) {
446 			.head.type	= VIRTIO_IOMMU_T_MAP,
447 			.domain		= cpu_to_le32(vdomain->id),
448 			.virt_start	= cpu_to_le64(mapping->iova.start),
449 			.virt_end	= cpu_to_le64(mapping->iova.last),
450 			.phys_start	= cpu_to_le64(mapping->paddr),
451 			.flags		= cpu_to_le32(mapping->flags),
452 		};
453 
454 		ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map));
455 		if (ret)
456 			break;
457 
458 		node = interval_tree_iter_next(node, 0, -1UL);
459 	}
460 	spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
461 
462 	return ret;
463 }
464 
465 static int viommu_add_resv_mem(struct viommu_endpoint *vdev,
466 			       struct virtio_iommu_probe_resv_mem *mem,
467 			       size_t len)
468 {
469 	size_t size;
470 	u64 start64, end64;
471 	phys_addr_t start, end;
472 	struct iommu_resv_region *region = NULL, *next;
473 	unsigned long prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
474 
475 	start = start64 = le64_to_cpu(mem->start);
476 	end = end64 = le64_to_cpu(mem->end);
477 	size = end64 - start64 + 1;
478 
479 	/* Catch any overflow, including the unlikely end64 - start64 + 1 = 0 */
480 	if (start != start64 || end != end64 || size < end64 - start64)
481 		return -EOVERFLOW;
482 
483 	if (len < sizeof(*mem))
484 		return -EINVAL;
485 
486 	switch (mem->subtype) {
487 	default:
488 		dev_warn(vdev->dev, "unknown resv mem subtype 0x%x\n",
489 			 mem->subtype);
490 		fallthrough;
491 	case VIRTIO_IOMMU_RESV_MEM_T_RESERVED:
492 		region = iommu_alloc_resv_region(start, size, 0,
493 						 IOMMU_RESV_RESERVED,
494 						 GFP_KERNEL);
495 		break;
496 	case VIRTIO_IOMMU_RESV_MEM_T_MSI:
497 		region = iommu_alloc_resv_region(start, size, prot,
498 						 IOMMU_RESV_MSI,
499 						 GFP_KERNEL);
500 		break;
501 	}
502 	if (!region)
503 		return -ENOMEM;
504 
505 	/* Keep the list sorted */
506 	list_for_each_entry(next, &vdev->resv_regions, list) {
507 		if (next->start > region->start)
508 			break;
509 	}
510 	list_add_tail(&region->list, &next->list);
511 	return 0;
512 }
513 
514 static int viommu_probe_endpoint(struct viommu_dev *viommu, struct device *dev)
515 {
516 	int ret;
517 	u16 type, len;
518 	size_t cur = 0;
519 	size_t probe_len;
520 	struct virtio_iommu_req_probe *probe;
521 	struct virtio_iommu_probe_property *prop;
522 	struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
523 	struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
524 
525 	if (!fwspec->num_ids)
526 		return -EINVAL;
527 
528 	probe_len = sizeof(*probe) + viommu->probe_size +
529 		    sizeof(struct virtio_iommu_req_tail);
530 	probe = kzalloc(probe_len, GFP_KERNEL);
531 	if (!probe)
532 		return -ENOMEM;
533 
534 	probe->head.type = VIRTIO_IOMMU_T_PROBE;
535 	/*
536 	 * For now, assume that properties of an endpoint that outputs multiple
537 	 * IDs are consistent. Only probe the first one.
538 	 */
539 	probe->endpoint = cpu_to_le32(fwspec->ids[0]);
540 
541 	ret = viommu_send_req_sync(viommu, probe, probe_len);
542 	if (ret)
543 		goto out_free;
544 
545 	prop = (void *)probe->properties;
546 	type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK;
547 
548 	while (type != VIRTIO_IOMMU_PROBE_T_NONE &&
549 	       cur < viommu->probe_size) {
550 		len = le16_to_cpu(prop->length) + sizeof(*prop);
551 
552 		switch (type) {
553 		case VIRTIO_IOMMU_PROBE_T_RESV_MEM:
554 			ret = viommu_add_resv_mem(vdev, (void *)prop, len);
555 			break;
556 		default:
557 			dev_err(dev, "unknown viommu prop 0x%x\n", type);
558 		}
559 
560 		if (ret)
561 			dev_err(dev, "failed to parse viommu prop 0x%x\n", type);
562 
563 		cur += len;
564 		if (cur >= viommu->probe_size)
565 			break;
566 
567 		prop = (void *)probe->properties + cur;
568 		type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK;
569 	}
570 
571 out_free:
572 	kfree(probe);
573 	return ret;
574 }
575 
576 static int viommu_fault_handler(struct viommu_dev *viommu,
577 				struct virtio_iommu_fault *fault)
578 {
579 	char *reason_str;
580 
581 	u8 reason	= fault->reason;
582 	u32 flags	= le32_to_cpu(fault->flags);
583 	u32 endpoint	= le32_to_cpu(fault->endpoint);
584 	u64 address	= le64_to_cpu(fault->address);
585 
586 	switch (reason) {
587 	case VIRTIO_IOMMU_FAULT_R_DOMAIN:
588 		reason_str = "domain";
589 		break;
590 	case VIRTIO_IOMMU_FAULT_R_MAPPING:
591 		reason_str = "page";
592 		break;
593 	case VIRTIO_IOMMU_FAULT_R_UNKNOWN:
594 	default:
595 		reason_str = "unknown";
596 		break;
597 	}
598 
599 	/* TODO: find EP by ID and report_iommu_fault */
600 	if (flags & VIRTIO_IOMMU_FAULT_F_ADDRESS)
601 		dev_err_ratelimited(viommu->dev, "%s fault from EP %u at %#llx [%s%s%s]\n",
602 				    reason_str, endpoint, address,
603 				    flags & VIRTIO_IOMMU_FAULT_F_READ ? "R" : "",
604 				    flags & VIRTIO_IOMMU_FAULT_F_WRITE ? "W" : "",
605 				    flags & VIRTIO_IOMMU_FAULT_F_EXEC ? "X" : "");
606 	else
607 		dev_err_ratelimited(viommu->dev, "%s fault from EP %u\n",
608 				    reason_str, endpoint);
609 	return 0;
610 }
611 
612 static void viommu_event_handler(struct virtqueue *vq)
613 {
614 	int ret;
615 	unsigned int len;
616 	struct scatterlist sg[1];
617 	struct viommu_event *evt;
618 	struct viommu_dev *viommu = vq->vdev->priv;
619 
620 	while ((evt = virtqueue_get_buf(vq, &len)) != NULL) {
621 		if (len > sizeof(*evt)) {
622 			dev_err(viommu->dev,
623 				"invalid event buffer (len %u != %zu)\n",
624 				len, sizeof(*evt));
625 		} else if (!(evt->head & VIOMMU_FAULT_RESV_MASK)) {
626 			viommu_fault_handler(viommu, &evt->fault);
627 		}
628 
629 		sg_init_one(sg, evt, sizeof(*evt));
630 		ret = virtqueue_add_inbuf(vq, sg, 1, evt, GFP_ATOMIC);
631 		if (ret)
632 			dev_err(viommu->dev, "could not add event buffer\n");
633 	}
634 
635 	virtqueue_kick(vq);
636 }
637 
638 /* IOMMU API */
639 
640 static struct iommu_domain *viommu_domain_alloc(unsigned type)
641 {
642 	struct viommu_domain *vdomain;
643 
644 	if (type != IOMMU_DOMAIN_UNMANAGED &&
645 	    type != IOMMU_DOMAIN_DMA &&
646 	    type != IOMMU_DOMAIN_IDENTITY)
647 		return NULL;
648 
649 	vdomain = kzalloc(sizeof(*vdomain), GFP_KERNEL);
650 	if (!vdomain)
651 		return NULL;
652 
653 	mutex_init(&vdomain->mutex);
654 	spin_lock_init(&vdomain->mappings_lock);
655 	vdomain->mappings = RB_ROOT_CACHED;
656 
657 	return &vdomain->domain;
658 }
659 
660 static int viommu_domain_finalise(struct viommu_endpoint *vdev,
661 				  struct iommu_domain *domain)
662 {
663 	int ret;
664 	unsigned long viommu_page_size;
665 	struct viommu_dev *viommu = vdev->viommu;
666 	struct viommu_domain *vdomain = to_viommu_domain(domain);
667 
668 	viommu_page_size = 1UL << __ffs(viommu->pgsize_bitmap);
669 	if (viommu_page_size > PAGE_SIZE) {
670 		dev_err(vdev->dev,
671 			"granule 0x%lx larger than system page size 0x%lx\n",
672 			viommu_page_size, PAGE_SIZE);
673 		return -EINVAL;
674 	}
675 
676 	ret = ida_alloc_range(&viommu->domain_ids, viommu->first_domain,
677 			      viommu->last_domain, GFP_KERNEL);
678 	if (ret < 0)
679 		return ret;
680 
681 	vdomain->id		= (unsigned int)ret;
682 
683 	domain->pgsize_bitmap	= viommu->pgsize_bitmap;
684 	domain->geometry	= viommu->geometry;
685 
686 	vdomain->map_flags	= viommu->map_flags;
687 	vdomain->viommu		= viommu;
688 
689 	if (domain->type == IOMMU_DOMAIN_IDENTITY) {
690 		if (virtio_has_feature(viommu->vdev,
691 				       VIRTIO_IOMMU_F_BYPASS_CONFIG)) {
692 			vdomain->bypass = true;
693 			return 0;
694 		}
695 
696 		ret = viommu_domain_map_identity(vdev, vdomain);
697 		if (ret) {
698 			ida_free(&viommu->domain_ids, vdomain->id);
699 			vdomain->viommu = NULL;
700 			return -EOPNOTSUPP;
701 		}
702 	}
703 
704 	return 0;
705 }
706 
707 static void viommu_domain_free(struct iommu_domain *domain)
708 {
709 	struct viommu_domain *vdomain = to_viommu_domain(domain);
710 
711 	/* Free all remaining mappings */
712 	viommu_del_mappings(vdomain, 0, ULLONG_MAX);
713 
714 	if (vdomain->viommu)
715 		ida_free(&vdomain->viommu->domain_ids, vdomain->id);
716 
717 	kfree(vdomain);
718 }
719 
720 static int viommu_attach_dev(struct iommu_domain *domain, struct device *dev)
721 {
722 	int i;
723 	int ret = 0;
724 	struct virtio_iommu_req_attach req;
725 	struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
726 	struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
727 	struct viommu_domain *vdomain = to_viommu_domain(domain);
728 
729 	mutex_lock(&vdomain->mutex);
730 	if (!vdomain->viommu) {
731 		/*
732 		 * Properly initialize the domain now that we know which viommu
733 		 * owns it.
734 		 */
735 		ret = viommu_domain_finalise(vdev, domain);
736 	} else if (vdomain->viommu != vdev->viommu) {
737 		dev_err(dev, "cannot attach to foreign vIOMMU\n");
738 		ret = -EXDEV;
739 	}
740 	mutex_unlock(&vdomain->mutex);
741 
742 	if (ret)
743 		return ret;
744 
745 	/*
746 	 * In the virtio-iommu device, when attaching the endpoint to a new
747 	 * domain, it is detached from the old one and, if as a result the
748 	 * old domain isn't attached to any endpoint, all mappings are removed
749 	 * from the old domain and it is freed.
750 	 *
751 	 * In the driver the old domain still exists, and its mappings will be
752 	 * recreated if it gets reattached to an endpoint. Otherwise it will be
753 	 * freed explicitly.
754 	 *
755 	 * vdev->vdomain is protected by group->mutex
756 	 */
757 	if (vdev->vdomain)
758 		vdev->vdomain->nr_endpoints--;
759 
760 	req = (struct virtio_iommu_req_attach) {
761 		.head.type	= VIRTIO_IOMMU_T_ATTACH,
762 		.domain		= cpu_to_le32(vdomain->id),
763 	};
764 
765 	if (vdomain->bypass)
766 		req.flags |= cpu_to_le32(VIRTIO_IOMMU_ATTACH_F_BYPASS);
767 
768 	for (i = 0; i < fwspec->num_ids; i++) {
769 		req.endpoint = cpu_to_le32(fwspec->ids[i]);
770 
771 		ret = viommu_send_req_sync(vdomain->viommu, &req, sizeof(req));
772 		if (ret)
773 			return ret;
774 	}
775 
776 	if (!vdomain->nr_endpoints) {
777 		/*
778 		 * This endpoint is the first to be attached to the domain.
779 		 * Replay existing mappings (e.g. SW MSI).
780 		 */
781 		ret = viommu_replay_mappings(vdomain);
782 		if (ret)
783 			return ret;
784 	}
785 
786 	vdomain->nr_endpoints++;
787 	vdev->vdomain = vdomain;
788 
789 	return 0;
790 }
791 
792 static int viommu_map_pages(struct iommu_domain *domain, unsigned long iova,
793 			    phys_addr_t paddr, size_t pgsize, size_t pgcount,
794 			    int prot, gfp_t gfp, size_t *mapped)
795 {
796 	int ret;
797 	u32 flags;
798 	size_t size = pgsize * pgcount;
799 	u64 end = iova + size - 1;
800 	struct virtio_iommu_req_map map;
801 	struct viommu_domain *vdomain = to_viommu_domain(domain);
802 
803 	flags = (prot & IOMMU_READ ? VIRTIO_IOMMU_MAP_F_READ : 0) |
804 		(prot & IOMMU_WRITE ? VIRTIO_IOMMU_MAP_F_WRITE : 0) |
805 		(prot & IOMMU_MMIO ? VIRTIO_IOMMU_MAP_F_MMIO : 0);
806 
807 	if (flags & ~vdomain->map_flags)
808 		return -EINVAL;
809 
810 	ret = viommu_add_mapping(vdomain, iova, end, paddr, flags);
811 	if (ret)
812 		return ret;
813 
814 	map = (struct virtio_iommu_req_map) {
815 		.head.type	= VIRTIO_IOMMU_T_MAP,
816 		.domain		= cpu_to_le32(vdomain->id),
817 		.virt_start	= cpu_to_le64(iova),
818 		.phys_start	= cpu_to_le64(paddr),
819 		.virt_end	= cpu_to_le64(end),
820 		.flags		= cpu_to_le32(flags),
821 	};
822 
823 	if (!vdomain->nr_endpoints)
824 		return 0;
825 
826 	ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map));
827 	if (ret)
828 		viommu_del_mappings(vdomain, iova, end);
829 	else if (mapped)
830 		*mapped = size;
831 
832 	return ret;
833 }
834 
835 static size_t viommu_unmap_pages(struct iommu_domain *domain, unsigned long iova,
836 				 size_t pgsize, size_t pgcount,
837 				 struct iommu_iotlb_gather *gather)
838 {
839 	int ret = 0;
840 	size_t unmapped;
841 	struct virtio_iommu_req_unmap unmap;
842 	struct viommu_domain *vdomain = to_viommu_domain(domain);
843 	size_t size = pgsize * pgcount;
844 
845 	unmapped = viommu_del_mappings(vdomain, iova, iova + size - 1);
846 	if (unmapped < size)
847 		return 0;
848 
849 	/* Device already removed all mappings after detach. */
850 	if (!vdomain->nr_endpoints)
851 		return unmapped;
852 
853 	unmap = (struct virtio_iommu_req_unmap) {
854 		.head.type	= VIRTIO_IOMMU_T_UNMAP,
855 		.domain		= cpu_to_le32(vdomain->id),
856 		.virt_start	= cpu_to_le64(iova),
857 		.virt_end	= cpu_to_le64(iova + unmapped - 1),
858 	};
859 
860 	ret = viommu_add_req(vdomain->viommu, &unmap, sizeof(unmap));
861 	return ret ? 0 : unmapped;
862 }
863 
864 static phys_addr_t viommu_iova_to_phys(struct iommu_domain *domain,
865 				       dma_addr_t iova)
866 {
867 	u64 paddr = 0;
868 	unsigned long flags;
869 	struct viommu_mapping *mapping;
870 	struct interval_tree_node *node;
871 	struct viommu_domain *vdomain = to_viommu_domain(domain);
872 
873 	spin_lock_irqsave(&vdomain->mappings_lock, flags);
874 	node = interval_tree_iter_first(&vdomain->mappings, iova, iova);
875 	if (node) {
876 		mapping = container_of(node, struct viommu_mapping, iova);
877 		paddr = mapping->paddr + (iova - mapping->iova.start);
878 	}
879 	spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
880 
881 	return paddr;
882 }
883 
884 static void viommu_iotlb_sync(struct iommu_domain *domain,
885 			      struct iommu_iotlb_gather *gather)
886 {
887 	struct viommu_domain *vdomain = to_viommu_domain(domain);
888 
889 	viommu_sync_req(vdomain->viommu);
890 }
891 
892 static void viommu_get_resv_regions(struct device *dev, struct list_head *head)
893 {
894 	struct iommu_resv_region *entry, *new_entry, *msi = NULL;
895 	struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
896 	int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
897 
898 	list_for_each_entry(entry, &vdev->resv_regions, list) {
899 		if (entry->type == IOMMU_RESV_MSI)
900 			msi = entry;
901 
902 		new_entry = kmemdup(entry, sizeof(*entry), GFP_KERNEL);
903 		if (!new_entry)
904 			return;
905 		list_add_tail(&new_entry->list, head);
906 	}
907 
908 	/*
909 	 * If the device didn't register any bypass MSI window, add a
910 	 * software-mapped region.
911 	 */
912 	if (!msi) {
913 		msi = iommu_alloc_resv_region(MSI_IOVA_BASE, MSI_IOVA_LENGTH,
914 					      prot, IOMMU_RESV_SW_MSI,
915 					      GFP_KERNEL);
916 		if (!msi)
917 			return;
918 
919 		list_add_tail(&msi->list, head);
920 	}
921 
922 	iommu_dma_get_resv_regions(dev, head);
923 }
924 
925 static struct iommu_ops viommu_ops;
926 static struct virtio_driver virtio_iommu_drv;
927 
928 static int viommu_match_node(struct device *dev, const void *data)
929 {
930 	return device_match_fwnode(dev->parent, data);
931 }
932 
933 static struct viommu_dev *viommu_get_by_fwnode(struct fwnode_handle *fwnode)
934 {
935 	struct device *dev = driver_find_device(&virtio_iommu_drv.driver, NULL,
936 						fwnode, viommu_match_node);
937 	put_device(dev);
938 
939 	return dev ? dev_to_virtio(dev)->priv : NULL;
940 }
941 
942 static struct iommu_device *viommu_probe_device(struct device *dev)
943 {
944 	int ret;
945 	struct viommu_endpoint *vdev;
946 	struct viommu_dev *viommu = NULL;
947 	struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
948 
949 	if (!fwspec || fwspec->ops != &viommu_ops)
950 		return ERR_PTR(-ENODEV);
951 
952 	viommu = viommu_get_by_fwnode(fwspec->iommu_fwnode);
953 	if (!viommu)
954 		return ERR_PTR(-ENODEV);
955 
956 	vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
957 	if (!vdev)
958 		return ERR_PTR(-ENOMEM);
959 
960 	vdev->dev = dev;
961 	vdev->viommu = viommu;
962 	INIT_LIST_HEAD(&vdev->resv_regions);
963 	dev_iommu_priv_set(dev, vdev);
964 
965 	if (viommu->probe_size) {
966 		/* Get additional information for this endpoint */
967 		ret = viommu_probe_endpoint(viommu, dev);
968 		if (ret)
969 			goto err_free_dev;
970 	}
971 
972 	return &viommu->iommu;
973 
974 err_free_dev:
975 	iommu_put_resv_regions(dev, &vdev->resv_regions);
976 	kfree(vdev);
977 
978 	return ERR_PTR(ret);
979 }
980 
981 static void viommu_probe_finalize(struct device *dev)
982 {
983 #ifndef CONFIG_ARCH_HAS_SETUP_DMA_OPS
984 	/* First clear the DMA ops in case we're switching from a DMA domain */
985 	set_dma_ops(dev, NULL);
986 	iommu_setup_dma_ops(dev, 0, U64_MAX);
987 #endif
988 }
989 
990 static void viommu_release_device(struct device *dev)
991 {
992 	struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
993 
994 	iommu_put_resv_regions(dev, &vdev->resv_regions);
995 	kfree(vdev);
996 }
997 
998 static struct iommu_group *viommu_device_group(struct device *dev)
999 {
1000 	if (dev_is_pci(dev))
1001 		return pci_device_group(dev);
1002 	else
1003 		return generic_device_group(dev);
1004 }
1005 
1006 static int viommu_of_xlate(struct device *dev, struct of_phandle_args *args)
1007 {
1008 	return iommu_fwspec_add_ids(dev, args->args, 1);
1009 }
1010 
1011 static bool viommu_capable(struct device *dev, enum iommu_cap cap)
1012 {
1013 	switch (cap) {
1014 	case IOMMU_CAP_CACHE_COHERENCY:
1015 		return true;
1016 	default:
1017 		return false;
1018 	}
1019 }
1020 
1021 static struct iommu_ops viommu_ops = {
1022 	.capable		= viommu_capable,
1023 	.domain_alloc		= viommu_domain_alloc,
1024 	.probe_device		= viommu_probe_device,
1025 	.probe_finalize		= viommu_probe_finalize,
1026 	.release_device		= viommu_release_device,
1027 	.device_group		= viommu_device_group,
1028 	.get_resv_regions	= viommu_get_resv_regions,
1029 	.of_xlate		= viommu_of_xlate,
1030 	.owner			= THIS_MODULE,
1031 	.default_domain_ops = &(const struct iommu_domain_ops) {
1032 		.attach_dev		= viommu_attach_dev,
1033 		.map_pages		= viommu_map_pages,
1034 		.unmap_pages		= viommu_unmap_pages,
1035 		.iova_to_phys		= viommu_iova_to_phys,
1036 		.iotlb_sync		= viommu_iotlb_sync,
1037 		.free			= viommu_domain_free,
1038 	}
1039 };
1040 
1041 static int viommu_init_vqs(struct viommu_dev *viommu)
1042 {
1043 	struct virtio_device *vdev = dev_to_virtio(viommu->dev);
1044 	const char *names[] = { "request", "event" };
1045 	vq_callback_t *callbacks[] = {
1046 		NULL, /* No async requests */
1047 		viommu_event_handler,
1048 	};
1049 
1050 	return virtio_find_vqs(vdev, VIOMMU_NR_VQS, viommu->vqs, callbacks,
1051 			       names, NULL);
1052 }
1053 
1054 static int viommu_fill_evtq(struct viommu_dev *viommu)
1055 {
1056 	int i, ret;
1057 	struct scatterlist sg[1];
1058 	struct viommu_event *evts;
1059 	struct virtqueue *vq = viommu->vqs[VIOMMU_EVENT_VQ];
1060 	size_t nr_evts = vq->num_free;
1061 
1062 	viommu->evts = evts = devm_kmalloc_array(viommu->dev, nr_evts,
1063 						 sizeof(*evts), GFP_KERNEL);
1064 	if (!evts)
1065 		return -ENOMEM;
1066 
1067 	for (i = 0; i < nr_evts; i++) {
1068 		sg_init_one(sg, &evts[i], sizeof(*evts));
1069 		ret = virtqueue_add_inbuf(vq, sg, 1, &evts[i], GFP_KERNEL);
1070 		if (ret)
1071 			return ret;
1072 	}
1073 
1074 	return 0;
1075 }
1076 
1077 static int viommu_probe(struct virtio_device *vdev)
1078 {
1079 	struct device *parent_dev = vdev->dev.parent;
1080 	struct viommu_dev *viommu = NULL;
1081 	struct device *dev = &vdev->dev;
1082 	u64 input_start = 0;
1083 	u64 input_end = -1UL;
1084 	int ret;
1085 
1086 	if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1) ||
1087 	    !virtio_has_feature(vdev, VIRTIO_IOMMU_F_MAP_UNMAP))
1088 		return -ENODEV;
1089 
1090 	viommu = devm_kzalloc(dev, sizeof(*viommu), GFP_KERNEL);
1091 	if (!viommu)
1092 		return -ENOMEM;
1093 
1094 	spin_lock_init(&viommu->request_lock);
1095 	ida_init(&viommu->domain_ids);
1096 	viommu->dev = dev;
1097 	viommu->vdev = vdev;
1098 	INIT_LIST_HEAD(&viommu->requests);
1099 
1100 	ret = viommu_init_vqs(viommu);
1101 	if (ret)
1102 		return ret;
1103 
1104 	virtio_cread_le(vdev, struct virtio_iommu_config, page_size_mask,
1105 			&viommu->pgsize_bitmap);
1106 
1107 	if (!viommu->pgsize_bitmap) {
1108 		ret = -EINVAL;
1109 		goto err_free_vqs;
1110 	}
1111 
1112 	viommu->map_flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE;
1113 	viommu->last_domain = ~0U;
1114 
1115 	/* Optional features */
1116 	virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1117 				struct virtio_iommu_config, input_range.start,
1118 				&input_start);
1119 
1120 	virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1121 				struct virtio_iommu_config, input_range.end,
1122 				&input_end);
1123 
1124 	virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1125 				struct virtio_iommu_config, domain_range.start,
1126 				&viommu->first_domain);
1127 
1128 	virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1129 				struct virtio_iommu_config, domain_range.end,
1130 				&viommu->last_domain);
1131 
1132 	virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_PROBE,
1133 				struct virtio_iommu_config, probe_size,
1134 				&viommu->probe_size);
1135 
1136 	viommu->geometry = (struct iommu_domain_geometry) {
1137 		.aperture_start	= input_start,
1138 		.aperture_end	= input_end,
1139 		.force_aperture	= true,
1140 	};
1141 
1142 	if (virtio_has_feature(vdev, VIRTIO_IOMMU_F_MMIO))
1143 		viommu->map_flags |= VIRTIO_IOMMU_MAP_F_MMIO;
1144 
1145 	viommu_ops.pgsize_bitmap = viommu->pgsize_bitmap;
1146 
1147 	virtio_device_ready(vdev);
1148 
1149 	/* Populate the event queue with buffers */
1150 	ret = viommu_fill_evtq(viommu);
1151 	if (ret)
1152 		goto err_free_vqs;
1153 
1154 	ret = iommu_device_sysfs_add(&viommu->iommu, dev, NULL, "%s",
1155 				     virtio_bus_name(vdev));
1156 	if (ret)
1157 		goto err_free_vqs;
1158 
1159 	iommu_device_register(&viommu->iommu, &viommu_ops, parent_dev);
1160 
1161 	vdev->priv = viommu;
1162 
1163 	dev_info(dev, "input address: %u bits\n",
1164 		 order_base_2(viommu->geometry.aperture_end));
1165 	dev_info(dev, "page mask: %#llx\n", viommu->pgsize_bitmap);
1166 
1167 	return 0;
1168 
1169 err_free_vqs:
1170 	vdev->config->del_vqs(vdev);
1171 
1172 	return ret;
1173 }
1174 
1175 static void viommu_remove(struct virtio_device *vdev)
1176 {
1177 	struct viommu_dev *viommu = vdev->priv;
1178 
1179 	iommu_device_sysfs_remove(&viommu->iommu);
1180 	iommu_device_unregister(&viommu->iommu);
1181 
1182 	/* Stop all virtqueues */
1183 	virtio_reset_device(vdev);
1184 	vdev->config->del_vqs(vdev);
1185 
1186 	dev_info(&vdev->dev, "device removed\n");
1187 }
1188 
1189 static void viommu_config_changed(struct virtio_device *vdev)
1190 {
1191 	dev_warn(&vdev->dev, "config changed\n");
1192 }
1193 
1194 static unsigned int features[] = {
1195 	VIRTIO_IOMMU_F_MAP_UNMAP,
1196 	VIRTIO_IOMMU_F_INPUT_RANGE,
1197 	VIRTIO_IOMMU_F_DOMAIN_RANGE,
1198 	VIRTIO_IOMMU_F_PROBE,
1199 	VIRTIO_IOMMU_F_MMIO,
1200 	VIRTIO_IOMMU_F_BYPASS_CONFIG,
1201 };
1202 
1203 static struct virtio_device_id id_table[] = {
1204 	{ VIRTIO_ID_IOMMU, VIRTIO_DEV_ANY_ID },
1205 	{ 0 },
1206 };
1207 MODULE_DEVICE_TABLE(virtio, id_table);
1208 
1209 static struct virtio_driver virtio_iommu_drv = {
1210 	.driver.name		= KBUILD_MODNAME,
1211 	.driver.owner		= THIS_MODULE,
1212 	.id_table		= id_table,
1213 	.feature_table		= features,
1214 	.feature_table_size	= ARRAY_SIZE(features),
1215 	.probe			= viommu_probe,
1216 	.remove			= viommu_remove,
1217 	.config_changed		= viommu_config_changed,
1218 };
1219 
1220 module_virtio_driver(virtio_iommu_drv);
1221 
1222 MODULE_DESCRIPTION("Virtio IOMMU driver");
1223 MODULE_AUTHOR("Jean-Philippe Brucker <jean-philippe.brucker@arm.com>");
1224 MODULE_LICENSE("GPL v2");
1225