Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Virtio driver for the paravirtualized IOMMU
4 *
5 * Copyright (C) 2019 Arm Limited
6 */
7
8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10#include <linux/amba/bus.h>
11#include <linux/delay.h>
12#include <linux/dma-iommu.h>
13#include <linux/freezer.h>
14#include <linux/interval_tree.h>
15#include <linux/iommu.h>
16#include <linux/module.h>
17#include <linux/of_iommu.h>
18#include <linux/of_platform.h>
19#include <linux/pci.h>
20#include <linux/platform_device.h>
21#include <linux/virtio.h>
22#include <linux/virtio_config.h>
23#include <linux/virtio_ids.h>
24#include <linux/wait.h>
25
26#include <uapi/linux/virtio_iommu.h>
27
28#define MSI_IOVA_BASE 0x8000000
29#define MSI_IOVA_LENGTH 0x100000
30
31#define VIOMMU_REQUEST_VQ 0
32#define VIOMMU_EVENT_VQ 1
33#define VIOMMU_NR_VQS 2
34
35struct viommu_dev {
36 struct iommu_device iommu;
37 struct device *dev;
38 struct virtio_device *vdev;
39
40 struct ida domain_ids;
41
42 struct virtqueue *vqs[VIOMMU_NR_VQS];
43 spinlock_t request_lock;
44 struct list_head requests;
45 void *evts;
46
47 /* Device configuration */
48 struct iommu_domain_geometry geometry;
49 u64 pgsize_bitmap;
50 u32 first_domain;
51 u32 last_domain;
52 /* Supported MAP flags */
53 u32 map_flags;
54 u32 probe_size;
55};
56
57struct viommu_mapping {
58 phys_addr_t paddr;
59 struct interval_tree_node iova;
60 u32 flags;
61};
62
63struct viommu_domain {
64 struct iommu_domain domain;
65 struct viommu_dev *viommu;
66 struct mutex mutex; /* protects viommu pointer */
67 unsigned int id;
68 u32 map_flags;
69
70 spinlock_t mappings_lock;
71 struct rb_root_cached mappings;
72
73 unsigned long nr_endpoints;
74};
75
76struct viommu_endpoint {
77 struct device *dev;
78 struct viommu_dev *viommu;
79 struct viommu_domain *vdomain;
80 struct list_head resv_regions;
81};
82
83struct viommu_request {
84 struct list_head list;
85 void *writeback;
86 unsigned int write_offset;
87 unsigned int len;
88 char buf[];
89};
90
91#define VIOMMU_FAULT_RESV_MASK 0xffffff00
92
93struct viommu_event {
94 union {
95 u32 head;
96 struct virtio_iommu_fault fault;
97 };
98};
99
100#define to_viommu_domain(domain) \
101 container_of(domain, struct viommu_domain, domain)
102
103static int viommu_get_req_errno(void *buf, size_t len)
104{
105 struct virtio_iommu_req_tail *tail = buf + len - sizeof(*tail);
106
107 switch (tail->status) {
108 case VIRTIO_IOMMU_S_OK:
109 return 0;
110 case VIRTIO_IOMMU_S_UNSUPP:
111 return -ENOSYS;
112 case VIRTIO_IOMMU_S_INVAL:
113 return -EINVAL;
114 case VIRTIO_IOMMU_S_RANGE:
115 return -ERANGE;
116 case VIRTIO_IOMMU_S_NOENT:
117 return -ENOENT;
118 case VIRTIO_IOMMU_S_FAULT:
119 return -EFAULT;
120 case VIRTIO_IOMMU_S_NOMEM:
121 return -ENOMEM;
122 case VIRTIO_IOMMU_S_IOERR:
123 case VIRTIO_IOMMU_S_DEVERR:
124 default:
125 return -EIO;
126 }
127}
128
129static void viommu_set_req_status(void *buf, size_t len, int status)
130{
131 struct virtio_iommu_req_tail *tail = buf + len - sizeof(*tail);
132
133 tail->status = status;
134}
135
136static off_t viommu_get_write_desc_offset(struct viommu_dev *viommu,
137 struct virtio_iommu_req_head *req,
138 size_t len)
139{
140 size_t tail_size = sizeof(struct virtio_iommu_req_tail);
141
142 if (req->type == VIRTIO_IOMMU_T_PROBE)
143 return len - viommu->probe_size - tail_size;
144
145 return len - tail_size;
146}
147
148/*
149 * __viommu_sync_req - Complete all in-flight requests
150 *
151 * Wait for all added requests to complete. When this function returns, all
152 * requests that were in-flight at the time of the call have completed.
153 */
154static int __viommu_sync_req(struct viommu_dev *viommu)
155{
156 unsigned int len;
157 size_t write_len;
158 struct viommu_request *req;
159 struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ];
160
161 assert_spin_locked(&viommu->request_lock);
162
163 virtqueue_kick(vq);
164
165 while (!list_empty(&viommu->requests)) {
166 len = 0;
167 req = virtqueue_get_buf(vq, &len);
168 if (!req)
169 continue;
170
171 if (!len)
172 viommu_set_req_status(req->buf, req->len,
173 VIRTIO_IOMMU_S_IOERR);
174
175 write_len = req->len - req->write_offset;
176 if (req->writeback && len == write_len)
177 memcpy(req->writeback, req->buf + req->write_offset,
178 write_len);
179
180 list_del(&req->list);
181 kfree(req);
182 }
183
184 return 0;
185}
186
187static int viommu_sync_req(struct viommu_dev *viommu)
188{
189 int ret;
190 unsigned long flags;
191
192 spin_lock_irqsave(&viommu->request_lock, flags);
193 ret = __viommu_sync_req(viommu);
194 if (ret)
195 dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret);
196 spin_unlock_irqrestore(&viommu->request_lock, flags);
197
198 return ret;
199}
200
201/*
202 * __viommu_add_request - Add one request to the queue
203 * @buf: pointer to the request buffer
204 * @len: length of the request buffer
205 * @writeback: copy data back to the buffer when the request completes.
206 *
207 * Add a request to the queue. Only synchronize the queue if it's already full.
208 * Otherwise don't kick the queue nor wait for requests to complete.
209 *
210 * When @writeback is true, data written by the device, including the request
211 * status, is copied into @buf after the request completes. This is unsafe if
212 * the caller allocates @buf on stack and drops the lock between add_req() and
213 * sync_req().
214 *
215 * Return 0 if the request was successfully added to the queue.
216 */
217static int __viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len,
218 bool writeback)
219{
220 int ret;
221 off_t write_offset;
222 struct viommu_request *req;
223 struct scatterlist top_sg, bottom_sg;
224 struct scatterlist *sg[2] = { &top_sg, &bottom_sg };
225 struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ];
226
227 assert_spin_locked(&viommu->request_lock);
228
229 write_offset = viommu_get_write_desc_offset(viommu, buf, len);
230 if (write_offset <= 0)
231 return -EINVAL;
232
233 req = kzalloc(sizeof(*req) + len, GFP_ATOMIC);
234 if (!req)
235 return -ENOMEM;
236
237 req->len = len;
238 if (writeback) {
239 req->writeback = buf + write_offset;
240 req->write_offset = write_offset;
241 }
242 memcpy(&req->buf, buf, write_offset);
243
244 sg_init_one(&top_sg, req->buf, write_offset);
245 sg_init_one(&bottom_sg, req->buf + write_offset, len - write_offset);
246
247 ret = virtqueue_add_sgs(vq, sg, 1, 1, req, GFP_ATOMIC);
248 if (ret == -ENOSPC) {
249 /* If the queue is full, sync and retry */
250 if (!__viommu_sync_req(viommu))
251 ret = virtqueue_add_sgs(vq, sg, 1, 1, req, GFP_ATOMIC);
252 }
253 if (ret)
254 goto err_free;
255
256 list_add_tail(&req->list, &viommu->requests);
257 return 0;
258
259err_free:
260 kfree(req);
261 return ret;
262}
263
264static int viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len)
265{
266 int ret;
267 unsigned long flags;
268
269 spin_lock_irqsave(&viommu->request_lock, flags);
270 ret = __viommu_add_req(viommu, buf, len, false);
271 if (ret)
272 dev_dbg(viommu->dev, "could not add request: %d\n", ret);
273 spin_unlock_irqrestore(&viommu->request_lock, flags);
274
275 return ret;
276}
277
278/*
279 * Send a request and wait for it to complete. Return the request status (as an
280 * errno)
281 */
282static int viommu_send_req_sync(struct viommu_dev *viommu, void *buf,
283 size_t len)
284{
285 int ret;
286 unsigned long flags;
287
288 spin_lock_irqsave(&viommu->request_lock, flags);
289
290 ret = __viommu_add_req(viommu, buf, len, true);
291 if (ret) {
292 dev_dbg(viommu->dev, "could not add request (%d)\n", ret);
293 goto out_unlock;
294 }
295
296 ret = __viommu_sync_req(viommu);
297 if (ret) {
298 dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret);
299 /* Fall-through (get the actual request status) */
300 }
301
302 ret = viommu_get_req_errno(buf, len);
303out_unlock:
304 spin_unlock_irqrestore(&viommu->request_lock, flags);
305 return ret;
306}
307
308/*
309 * viommu_add_mapping - add a mapping to the internal tree
310 *
311 * On success, return the new mapping. Otherwise return NULL.
312 */
313static int viommu_add_mapping(struct viommu_domain *vdomain, unsigned long iova,
314 phys_addr_t paddr, size_t size, u32 flags)
315{
316 unsigned long irqflags;
317 struct viommu_mapping *mapping;
318
319 mapping = kzalloc(sizeof(*mapping), GFP_ATOMIC);
320 if (!mapping)
321 return -ENOMEM;
322
323 mapping->paddr = paddr;
324 mapping->iova.start = iova;
325 mapping->iova.last = iova + size - 1;
326 mapping->flags = flags;
327
328 spin_lock_irqsave(&vdomain->mappings_lock, irqflags);
329 interval_tree_insert(&mapping->iova, &vdomain->mappings);
330 spin_unlock_irqrestore(&vdomain->mappings_lock, irqflags);
331
332 return 0;
333}
334
335/*
336 * viommu_del_mappings - remove mappings from the internal tree
337 *
338 * @vdomain: the domain
339 * @iova: start of the range
340 * @size: size of the range. A size of 0 corresponds to the entire address
341 * space.
342 *
343 * On success, returns the number of unmapped bytes (>= size)
344 */
345static size_t viommu_del_mappings(struct viommu_domain *vdomain,
346 unsigned long iova, size_t size)
347{
348 size_t unmapped = 0;
349 unsigned long flags;
350 unsigned long last = iova + size - 1;
351 struct viommu_mapping *mapping = NULL;
352 struct interval_tree_node *node, *next;
353
354 spin_lock_irqsave(&vdomain->mappings_lock, flags);
355 next = interval_tree_iter_first(&vdomain->mappings, iova, last);
356 while (next) {
357 node = next;
358 mapping = container_of(node, struct viommu_mapping, iova);
359 next = interval_tree_iter_next(node, iova, last);
360
361 /* Trying to split a mapping? */
362 if (mapping->iova.start < iova)
363 break;
364
365 /*
366 * Virtio-iommu doesn't allow UNMAP to split a mapping created
367 * with a single MAP request, so remove the full mapping.
368 */
369 unmapped += mapping->iova.last - mapping->iova.start + 1;
370
371 interval_tree_remove(node, &vdomain->mappings);
372 kfree(mapping);
373 }
374 spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
375
376 return unmapped;
377}
378
379/*
380 * viommu_replay_mappings - re-send MAP requests
381 *
382 * When reattaching a domain that was previously detached from all endpoints,
383 * mappings were deleted from the device. Re-create the mappings available in
384 * the internal tree.
385 */
386static int viommu_replay_mappings(struct viommu_domain *vdomain)
387{
388 int ret = 0;
389 unsigned long flags;
390 struct viommu_mapping *mapping;
391 struct interval_tree_node *node;
392 struct virtio_iommu_req_map map;
393
394 spin_lock_irqsave(&vdomain->mappings_lock, flags);
395 node = interval_tree_iter_first(&vdomain->mappings, 0, -1UL);
396 while (node) {
397 mapping = container_of(node, struct viommu_mapping, iova);
398 map = (struct virtio_iommu_req_map) {
399 .head.type = VIRTIO_IOMMU_T_MAP,
400 .domain = cpu_to_le32(vdomain->id),
401 .virt_start = cpu_to_le64(mapping->iova.start),
402 .virt_end = cpu_to_le64(mapping->iova.last),
403 .phys_start = cpu_to_le64(mapping->paddr),
404 .flags = cpu_to_le32(mapping->flags),
405 };
406
407 ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map));
408 if (ret)
409 break;
410
411 node = interval_tree_iter_next(node, 0, -1UL);
412 }
413 spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
414
415 return ret;
416}
417
418static int viommu_add_resv_mem(struct viommu_endpoint *vdev,
419 struct virtio_iommu_probe_resv_mem *mem,
420 size_t len)
421{
422 size_t size;
423 u64 start64, end64;
424 phys_addr_t start, end;
425 struct iommu_resv_region *region = NULL;
426 unsigned long prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
427
428 start = start64 = le64_to_cpu(mem->start);
429 end = end64 = le64_to_cpu(mem->end);
430 size = end64 - start64 + 1;
431
432 /* Catch any overflow, including the unlikely end64 - start64 + 1 = 0 */
433 if (start != start64 || end != end64 || size < end64 - start64)
434 return -EOVERFLOW;
435
436 if (len < sizeof(*mem))
437 return -EINVAL;
438
439 switch (mem->subtype) {
440 default:
441 dev_warn(vdev->dev, "unknown resv mem subtype 0x%x\n",
442 mem->subtype);
443 fallthrough;
444 case VIRTIO_IOMMU_RESV_MEM_T_RESERVED:
445 region = iommu_alloc_resv_region(start, size, 0,
446 IOMMU_RESV_RESERVED);
447 break;
448 case VIRTIO_IOMMU_RESV_MEM_T_MSI:
449 region = iommu_alloc_resv_region(start, size, prot,
450 IOMMU_RESV_MSI);
451 break;
452 }
453 if (!region)
454 return -ENOMEM;
455
456 list_add(®ion->list, &vdev->resv_regions);
457 return 0;
458}
459
460static int viommu_probe_endpoint(struct viommu_dev *viommu, struct device *dev)
461{
462 int ret;
463 u16 type, len;
464 size_t cur = 0;
465 size_t probe_len;
466 struct virtio_iommu_req_probe *probe;
467 struct virtio_iommu_probe_property *prop;
468 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
469 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
470
471 if (!fwspec->num_ids)
472 return -EINVAL;
473
474 probe_len = sizeof(*probe) + viommu->probe_size +
475 sizeof(struct virtio_iommu_req_tail);
476 probe = kzalloc(probe_len, GFP_KERNEL);
477 if (!probe)
478 return -ENOMEM;
479
480 probe->head.type = VIRTIO_IOMMU_T_PROBE;
481 /*
482 * For now, assume that properties of an endpoint that outputs multiple
483 * IDs are consistent. Only probe the first one.
484 */
485 probe->endpoint = cpu_to_le32(fwspec->ids[0]);
486
487 ret = viommu_send_req_sync(viommu, probe, probe_len);
488 if (ret)
489 goto out_free;
490
491 prop = (void *)probe->properties;
492 type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK;
493
494 while (type != VIRTIO_IOMMU_PROBE_T_NONE &&
495 cur < viommu->probe_size) {
496 len = le16_to_cpu(prop->length) + sizeof(*prop);
497
498 switch (type) {
499 case VIRTIO_IOMMU_PROBE_T_RESV_MEM:
500 ret = viommu_add_resv_mem(vdev, (void *)prop, len);
501 break;
502 default:
503 dev_err(dev, "unknown viommu prop 0x%x\n", type);
504 }
505
506 if (ret)
507 dev_err(dev, "failed to parse viommu prop 0x%x\n", type);
508
509 cur += len;
510 if (cur >= viommu->probe_size)
511 break;
512
513 prop = (void *)probe->properties + cur;
514 type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK;
515 }
516
517out_free:
518 kfree(probe);
519 return ret;
520}
521
522static int viommu_fault_handler(struct viommu_dev *viommu,
523 struct virtio_iommu_fault *fault)
524{
525 char *reason_str;
526
527 u8 reason = fault->reason;
528 u32 flags = le32_to_cpu(fault->flags);
529 u32 endpoint = le32_to_cpu(fault->endpoint);
530 u64 address = le64_to_cpu(fault->address);
531
532 switch (reason) {
533 case VIRTIO_IOMMU_FAULT_R_DOMAIN:
534 reason_str = "domain";
535 break;
536 case VIRTIO_IOMMU_FAULT_R_MAPPING:
537 reason_str = "page";
538 break;
539 case VIRTIO_IOMMU_FAULT_R_UNKNOWN:
540 default:
541 reason_str = "unknown";
542 break;
543 }
544
545 /* TODO: find EP by ID and report_iommu_fault */
546 if (flags & VIRTIO_IOMMU_FAULT_F_ADDRESS)
547 dev_err_ratelimited(viommu->dev, "%s fault from EP %u at %#llx [%s%s%s]\n",
548 reason_str, endpoint, address,
549 flags & VIRTIO_IOMMU_FAULT_F_READ ? "R" : "",
550 flags & VIRTIO_IOMMU_FAULT_F_WRITE ? "W" : "",
551 flags & VIRTIO_IOMMU_FAULT_F_EXEC ? "X" : "");
552 else
553 dev_err_ratelimited(viommu->dev, "%s fault from EP %u\n",
554 reason_str, endpoint);
555 return 0;
556}
557
558static void viommu_event_handler(struct virtqueue *vq)
559{
560 int ret;
561 unsigned int len;
562 struct scatterlist sg[1];
563 struct viommu_event *evt;
564 struct viommu_dev *viommu = vq->vdev->priv;
565
566 while ((evt = virtqueue_get_buf(vq, &len)) != NULL) {
567 if (len > sizeof(*evt)) {
568 dev_err(viommu->dev,
569 "invalid event buffer (len %u != %zu)\n",
570 len, sizeof(*evt));
571 } else if (!(evt->head & VIOMMU_FAULT_RESV_MASK)) {
572 viommu_fault_handler(viommu, &evt->fault);
573 }
574
575 sg_init_one(sg, evt, sizeof(*evt));
576 ret = virtqueue_add_inbuf(vq, sg, 1, evt, GFP_ATOMIC);
577 if (ret)
578 dev_err(viommu->dev, "could not add event buffer\n");
579 }
580
581 virtqueue_kick(vq);
582}
583
584/* IOMMU API */
585
586static struct iommu_domain *viommu_domain_alloc(unsigned type)
587{
588 struct viommu_domain *vdomain;
589
590 if (type != IOMMU_DOMAIN_UNMANAGED && type != IOMMU_DOMAIN_DMA)
591 return NULL;
592
593 vdomain = kzalloc(sizeof(*vdomain), GFP_KERNEL);
594 if (!vdomain)
595 return NULL;
596
597 mutex_init(&vdomain->mutex);
598 spin_lock_init(&vdomain->mappings_lock);
599 vdomain->mappings = RB_ROOT_CACHED;
600
601 if (type == IOMMU_DOMAIN_DMA &&
602 iommu_get_dma_cookie(&vdomain->domain)) {
603 kfree(vdomain);
604 return NULL;
605 }
606
607 return &vdomain->domain;
608}
609
610static int viommu_domain_finalise(struct viommu_endpoint *vdev,
611 struct iommu_domain *domain)
612{
613 int ret;
614 unsigned long viommu_page_size;
615 struct viommu_dev *viommu = vdev->viommu;
616 struct viommu_domain *vdomain = to_viommu_domain(domain);
617
618 viommu_page_size = 1UL << __ffs(viommu->pgsize_bitmap);
619 if (viommu_page_size > PAGE_SIZE) {
620 dev_err(vdev->dev,
621 "granule 0x%lx larger than system page size 0x%lx\n",
622 viommu_page_size, PAGE_SIZE);
623 return -EINVAL;
624 }
625
626 ret = ida_alloc_range(&viommu->domain_ids, viommu->first_domain,
627 viommu->last_domain, GFP_KERNEL);
628 if (ret < 0)
629 return ret;
630
631 vdomain->id = (unsigned int)ret;
632
633 domain->pgsize_bitmap = viommu->pgsize_bitmap;
634 domain->geometry = viommu->geometry;
635
636 vdomain->map_flags = viommu->map_flags;
637 vdomain->viommu = viommu;
638
639 return 0;
640}
641
642static void viommu_domain_free(struct iommu_domain *domain)
643{
644 struct viommu_domain *vdomain = to_viommu_domain(domain);
645
646 iommu_put_dma_cookie(domain);
647
648 /* Free all remaining mappings (size 2^64) */
649 viommu_del_mappings(vdomain, 0, 0);
650
651 if (vdomain->viommu)
652 ida_free(&vdomain->viommu->domain_ids, vdomain->id);
653
654 kfree(vdomain);
655}
656
657static int viommu_attach_dev(struct iommu_domain *domain, struct device *dev)
658{
659 int i;
660 int ret = 0;
661 struct virtio_iommu_req_attach req;
662 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
663 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
664 struct viommu_domain *vdomain = to_viommu_domain(domain);
665
666 mutex_lock(&vdomain->mutex);
667 if (!vdomain->viommu) {
668 /*
669 * Properly initialize the domain now that we know which viommu
670 * owns it.
671 */
672 ret = viommu_domain_finalise(vdev, domain);
673 } else if (vdomain->viommu != vdev->viommu) {
674 dev_err(dev, "cannot attach to foreign vIOMMU\n");
675 ret = -EXDEV;
676 }
677 mutex_unlock(&vdomain->mutex);
678
679 if (ret)
680 return ret;
681
682 /*
683 * In the virtio-iommu device, when attaching the endpoint to a new
684 * domain, it is detached from the old one and, if as as a result the
685 * old domain isn't attached to any endpoint, all mappings are removed
686 * from the old domain and it is freed.
687 *
688 * In the driver the old domain still exists, and its mappings will be
689 * recreated if it gets reattached to an endpoint. Otherwise it will be
690 * freed explicitly.
691 *
692 * vdev->vdomain is protected by group->mutex
693 */
694 if (vdev->vdomain)
695 vdev->vdomain->nr_endpoints--;
696
697 req = (struct virtio_iommu_req_attach) {
698 .head.type = VIRTIO_IOMMU_T_ATTACH,
699 .domain = cpu_to_le32(vdomain->id),
700 };
701
702 for (i = 0; i < fwspec->num_ids; i++) {
703 req.endpoint = cpu_to_le32(fwspec->ids[i]);
704
705 ret = viommu_send_req_sync(vdomain->viommu, &req, sizeof(req));
706 if (ret)
707 return ret;
708 }
709
710 if (!vdomain->nr_endpoints) {
711 /*
712 * This endpoint is the first to be attached to the domain.
713 * Replay existing mappings (e.g. SW MSI).
714 */
715 ret = viommu_replay_mappings(vdomain);
716 if (ret)
717 return ret;
718 }
719
720 vdomain->nr_endpoints++;
721 vdev->vdomain = vdomain;
722
723 return 0;
724}
725
726static int viommu_map(struct iommu_domain *domain, unsigned long iova,
727 phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
728{
729 int ret;
730 u32 flags;
731 struct virtio_iommu_req_map map;
732 struct viommu_domain *vdomain = to_viommu_domain(domain);
733
734 flags = (prot & IOMMU_READ ? VIRTIO_IOMMU_MAP_F_READ : 0) |
735 (prot & IOMMU_WRITE ? VIRTIO_IOMMU_MAP_F_WRITE : 0) |
736 (prot & IOMMU_MMIO ? VIRTIO_IOMMU_MAP_F_MMIO : 0);
737
738 if (flags & ~vdomain->map_flags)
739 return -EINVAL;
740
741 ret = viommu_add_mapping(vdomain, iova, paddr, size, flags);
742 if (ret)
743 return ret;
744
745 map = (struct virtio_iommu_req_map) {
746 .head.type = VIRTIO_IOMMU_T_MAP,
747 .domain = cpu_to_le32(vdomain->id),
748 .virt_start = cpu_to_le64(iova),
749 .phys_start = cpu_to_le64(paddr),
750 .virt_end = cpu_to_le64(iova + size - 1),
751 .flags = cpu_to_le32(flags),
752 };
753
754 if (!vdomain->nr_endpoints)
755 return 0;
756
757 ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map));
758 if (ret)
759 viommu_del_mappings(vdomain, iova, size);
760
761 return ret;
762}
763
764static size_t viommu_unmap(struct iommu_domain *domain, unsigned long iova,
765 size_t size, struct iommu_iotlb_gather *gather)
766{
767 int ret = 0;
768 size_t unmapped;
769 struct virtio_iommu_req_unmap unmap;
770 struct viommu_domain *vdomain = to_viommu_domain(domain);
771
772 unmapped = viommu_del_mappings(vdomain, iova, size);
773 if (unmapped < size)
774 return 0;
775
776 /* Device already removed all mappings after detach. */
777 if (!vdomain->nr_endpoints)
778 return unmapped;
779
780 unmap = (struct virtio_iommu_req_unmap) {
781 .head.type = VIRTIO_IOMMU_T_UNMAP,
782 .domain = cpu_to_le32(vdomain->id),
783 .virt_start = cpu_to_le64(iova),
784 .virt_end = cpu_to_le64(iova + unmapped - 1),
785 };
786
787 ret = viommu_add_req(vdomain->viommu, &unmap, sizeof(unmap));
788 return ret ? 0 : unmapped;
789}
790
791static phys_addr_t viommu_iova_to_phys(struct iommu_domain *domain,
792 dma_addr_t iova)
793{
794 u64 paddr = 0;
795 unsigned long flags;
796 struct viommu_mapping *mapping;
797 struct interval_tree_node *node;
798 struct viommu_domain *vdomain = to_viommu_domain(domain);
799
800 spin_lock_irqsave(&vdomain->mappings_lock, flags);
801 node = interval_tree_iter_first(&vdomain->mappings, iova, iova);
802 if (node) {
803 mapping = container_of(node, struct viommu_mapping, iova);
804 paddr = mapping->paddr + (iova - mapping->iova.start);
805 }
806 spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
807
808 return paddr;
809}
810
811static void viommu_iotlb_sync(struct iommu_domain *domain,
812 struct iommu_iotlb_gather *gather)
813{
814 struct viommu_domain *vdomain = to_viommu_domain(domain);
815
816 viommu_sync_req(vdomain->viommu);
817}
818
819static void viommu_get_resv_regions(struct device *dev, struct list_head *head)
820{
821 struct iommu_resv_region *entry, *new_entry, *msi = NULL;
822 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
823 int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
824
825 list_for_each_entry(entry, &vdev->resv_regions, list) {
826 if (entry->type == IOMMU_RESV_MSI)
827 msi = entry;
828
829 new_entry = kmemdup(entry, sizeof(*entry), GFP_KERNEL);
830 if (!new_entry)
831 return;
832 list_add_tail(&new_entry->list, head);
833 }
834
835 /*
836 * If the device didn't register any bypass MSI window, add a
837 * software-mapped region.
838 */
839 if (!msi) {
840 msi = iommu_alloc_resv_region(MSI_IOVA_BASE, MSI_IOVA_LENGTH,
841 prot, IOMMU_RESV_SW_MSI);
842 if (!msi)
843 return;
844
845 list_add_tail(&msi->list, head);
846 }
847
848 iommu_dma_get_resv_regions(dev, head);
849}
850
851static struct iommu_ops viommu_ops;
852static struct virtio_driver virtio_iommu_drv;
853
854static int viommu_match_node(struct device *dev, const void *data)
855{
856 return dev->parent->fwnode == data;
857}
858
859static struct viommu_dev *viommu_get_by_fwnode(struct fwnode_handle *fwnode)
860{
861 struct device *dev = driver_find_device(&virtio_iommu_drv.driver, NULL,
862 fwnode, viommu_match_node);
863 put_device(dev);
864
865 return dev ? dev_to_virtio(dev)->priv : NULL;
866}
867
868static struct iommu_device *viommu_probe_device(struct device *dev)
869{
870 int ret;
871 struct viommu_endpoint *vdev;
872 struct viommu_dev *viommu = NULL;
873 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
874
875 if (!fwspec || fwspec->ops != &viommu_ops)
876 return ERR_PTR(-ENODEV);
877
878 viommu = viommu_get_by_fwnode(fwspec->iommu_fwnode);
879 if (!viommu)
880 return ERR_PTR(-ENODEV);
881
882 vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
883 if (!vdev)
884 return ERR_PTR(-ENOMEM);
885
886 vdev->dev = dev;
887 vdev->viommu = viommu;
888 INIT_LIST_HEAD(&vdev->resv_regions);
889 dev_iommu_priv_set(dev, vdev);
890
891 if (viommu->probe_size) {
892 /* Get additional information for this endpoint */
893 ret = viommu_probe_endpoint(viommu, dev);
894 if (ret)
895 goto err_free_dev;
896 }
897
898 return &viommu->iommu;
899
900err_free_dev:
901 generic_iommu_put_resv_regions(dev, &vdev->resv_regions);
902 kfree(vdev);
903
904 return ERR_PTR(ret);
905}
906
907static void viommu_release_device(struct device *dev)
908{
909 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
910 struct viommu_endpoint *vdev;
911
912 if (!fwspec || fwspec->ops != &viommu_ops)
913 return;
914
915 vdev = dev_iommu_priv_get(dev);
916
917 generic_iommu_put_resv_regions(dev, &vdev->resv_regions);
918 kfree(vdev);
919}
920
921static struct iommu_group *viommu_device_group(struct device *dev)
922{
923 if (dev_is_pci(dev))
924 return pci_device_group(dev);
925 else
926 return generic_device_group(dev);
927}
928
929static int viommu_of_xlate(struct device *dev, struct of_phandle_args *args)
930{
931 return iommu_fwspec_add_ids(dev, args->args, 1);
932}
933
934static struct iommu_ops viommu_ops = {
935 .domain_alloc = viommu_domain_alloc,
936 .domain_free = viommu_domain_free,
937 .attach_dev = viommu_attach_dev,
938 .map = viommu_map,
939 .unmap = viommu_unmap,
940 .iova_to_phys = viommu_iova_to_phys,
941 .iotlb_sync = viommu_iotlb_sync,
942 .probe_device = viommu_probe_device,
943 .release_device = viommu_release_device,
944 .device_group = viommu_device_group,
945 .get_resv_regions = viommu_get_resv_regions,
946 .put_resv_regions = generic_iommu_put_resv_regions,
947 .of_xlate = viommu_of_xlate,
948};
949
950static int viommu_init_vqs(struct viommu_dev *viommu)
951{
952 struct virtio_device *vdev = dev_to_virtio(viommu->dev);
953 const char *names[] = { "request", "event" };
954 vq_callback_t *callbacks[] = {
955 NULL, /* No async requests */
956 viommu_event_handler,
957 };
958
959 return virtio_find_vqs(vdev, VIOMMU_NR_VQS, viommu->vqs, callbacks,
960 names, NULL);
961}
962
963static int viommu_fill_evtq(struct viommu_dev *viommu)
964{
965 int i, ret;
966 struct scatterlist sg[1];
967 struct viommu_event *evts;
968 struct virtqueue *vq = viommu->vqs[VIOMMU_EVENT_VQ];
969 size_t nr_evts = vq->num_free;
970
971 viommu->evts = evts = devm_kmalloc_array(viommu->dev, nr_evts,
972 sizeof(*evts), GFP_KERNEL);
973 if (!evts)
974 return -ENOMEM;
975
976 for (i = 0; i < nr_evts; i++) {
977 sg_init_one(sg, &evts[i], sizeof(*evts));
978 ret = virtqueue_add_inbuf(vq, sg, 1, &evts[i], GFP_KERNEL);
979 if (ret)
980 return ret;
981 }
982
983 return 0;
984}
985
986static int viommu_probe(struct virtio_device *vdev)
987{
988 struct device *parent_dev = vdev->dev.parent;
989 struct viommu_dev *viommu = NULL;
990 struct device *dev = &vdev->dev;
991 u64 input_start = 0;
992 u64 input_end = -1UL;
993 int ret;
994
995 if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1) ||
996 !virtio_has_feature(vdev, VIRTIO_IOMMU_F_MAP_UNMAP))
997 return -ENODEV;
998
999 viommu = devm_kzalloc(dev, sizeof(*viommu), GFP_KERNEL);
1000 if (!viommu)
1001 return -ENOMEM;
1002
1003 spin_lock_init(&viommu->request_lock);
1004 ida_init(&viommu->domain_ids);
1005 viommu->dev = dev;
1006 viommu->vdev = vdev;
1007 INIT_LIST_HEAD(&viommu->requests);
1008
1009 ret = viommu_init_vqs(viommu);
1010 if (ret)
1011 return ret;
1012
1013 virtio_cread_le(vdev, struct virtio_iommu_config, page_size_mask,
1014 &viommu->pgsize_bitmap);
1015
1016 if (!viommu->pgsize_bitmap) {
1017 ret = -EINVAL;
1018 goto err_free_vqs;
1019 }
1020
1021 viommu->map_flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE;
1022 viommu->last_domain = ~0U;
1023
1024 /* Optional features */
1025 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1026 struct virtio_iommu_config, input_range.start,
1027 &input_start);
1028
1029 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1030 struct virtio_iommu_config, input_range.end,
1031 &input_end);
1032
1033 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1034 struct virtio_iommu_config, domain_range.start,
1035 &viommu->first_domain);
1036
1037 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1038 struct virtio_iommu_config, domain_range.end,
1039 &viommu->last_domain);
1040
1041 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_PROBE,
1042 struct virtio_iommu_config, probe_size,
1043 &viommu->probe_size);
1044
1045 viommu->geometry = (struct iommu_domain_geometry) {
1046 .aperture_start = input_start,
1047 .aperture_end = input_end,
1048 .force_aperture = true,
1049 };
1050
1051 if (virtio_has_feature(vdev, VIRTIO_IOMMU_F_MMIO))
1052 viommu->map_flags |= VIRTIO_IOMMU_MAP_F_MMIO;
1053
1054 viommu_ops.pgsize_bitmap = viommu->pgsize_bitmap;
1055
1056 virtio_device_ready(vdev);
1057
1058 /* Populate the event queue with buffers */
1059 ret = viommu_fill_evtq(viommu);
1060 if (ret)
1061 goto err_free_vqs;
1062
1063 ret = iommu_device_sysfs_add(&viommu->iommu, dev, NULL, "%s",
1064 virtio_bus_name(vdev));
1065 if (ret)
1066 goto err_free_vqs;
1067
1068 iommu_device_set_ops(&viommu->iommu, &viommu_ops);
1069 iommu_device_set_fwnode(&viommu->iommu, parent_dev->fwnode);
1070
1071 iommu_device_register(&viommu->iommu);
1072
1073#ifdef CONFIG_PCI
1074 if (pci_bus_type.iommu_ops != &viommu_ops) {
1075 ret = bus_set_iommu(&pci_bus_type, &viommu_ops);
1076 if (ret)
1077 goto err_unregister;
1078 }
1079#endif
1080#ifdef CONFIG_ARM_AMBA
1081 if (amba_bustype.iommu_ops != &viommu_ops) {
1082 ret = bus_set_iommu(&amba_bustype, &viommu_ops);
1083 if (ret)
1084 goto err_unregister;
1085 }
1086#endif
1087 if (platform_bus_type.iommu_ops != &viommu_ops) {
1088 ret = bus_set_iommu(&platform_bus_type, &viommu_ops);
1089 if (ret)
1090 goto err_unregister;
1091 }
1092
1093 vdev->priv = viommu;
1094
1095 dev_info(dev, "input address: %u bits\n",
1096 order_base_2(viommu->geometry.aperture_end));
1097 dev_info(dev, "page mask: %#llx\n", viommu->pgsize_bitmap);
1098
1099 return 0;
1100
1101err_unregister:
1102 iommu_device_sysfs_remove(&viommu->iommu);
1103 iommu_device_unregister(&viommu->iommu);
1104err_free_vqs:
1105 vdev->config->del_vqs(vdev);
1106
1107 return ret;
1108}
1109
1110static void viommu_remove(struct virtio_device *vdev)
1111{
1112 struct viommu_dev *viommu = vdev->priv;
1113
1114 iommu_device_sysfs_remove(&viommu->iommu);
1115 iommu_device_unregister(&viommu->iommu);
1116
1117 /* Stop all virtqueues */
1118 vdev->config->reset(vdev);
1119 vdev->config->del_vqs(vdev);
1120
1121 dev_info(&vdev->dev, "device removed\n");
1122}
1123
1124static void viommu_config_changed(struct virtio_device *vdev)
1125{
1126 dev_warn(&vdev->dev, "config changed\n");
1127}
1128
1129static unsigned int features[] = {
1130 VIRTIO_IOMMU_F_MAP_UNMAP,
1131 VIRTIO_IOMMU_F_INPUT_RANGE,
1132 VIRTIO_IOMMU_F_DOMAIN_RANGE,
1133 VIRTIO_IOMMU_F_PROBE,
1134 VIRTIO_IOMMU_F_MMIO,
1135};
1136
1137static struct virtio_device_id id_table[] = {
1138 { VIRTIO_ID_IOMMU, VIRTIO_DEV_ANY_ID },
1139 { 0 },
1140};
1141
1142static struct virtio_driver virtio_iommu_drv = {
1143 .driver.name = KBUILD_MODNAME,
1144 .driver.owner = THIS_MODULE,
1145 .id_table = id_table,
1146 .feature_table = features,
1147 .feature_table_size = ARRAY_SIZE(features),
1148 .probe = viommu_probe,
1149 .remove = viommu_remove,
1150 .config_changed = viommu_config_changed,
1151};
1152
1153module_virtio_driver(virtio_iommu_drv);
1154
1155MODULE_DESCRIPTION("Virtio IOMMU driver");
1156MODULE_AUTHOR("Jean-Philippe Brucker <jean-philippe.brucker@arm.com>");
1157MODULE_LICENSE("GPL v2");
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 /* Supported MAP flags */
52 u32 map_flags;
53 u32 probe_size;
54};
55
56struct viommu_mapping {
57 phys_addr_t paddr;
58 struct interval_tree_node iova;
59 u32 flags;
60};
61
62struct 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
76struct viommu_endpoint {
77 struct device *dev;
78 struct viommu_dev *viommu;
79 struct viommu_domain *vdomain;
80 struct list_head resv_regions;
81};
82
83struct viommu_request {
84 struct list_head list;
85 void *writeback;
86 unsigned int write_offset;
87 unsigned int len;
88 char buf[] __counted_by(len);
89};
90
91#define VIOMMU_FAULT_RESV_MASK 0xffffff00
92
93struct viommu_event {
94 union {
95 u32 head;
96 struct virtio_iommu_fault fault;
97 };
98};
99
100#define to_viommu_domain(domain) \
101 container_of(domain, struct viommu_domain, domain)
102
103static int viommu_get_req_errno(void *buf, size_t len)
104{
105 struct virtio_iommu_req_tail *tail = buf + len - sizeof(*tail);
106
107 switch (tail->status) {
108 case VIRTIO_IOMMU_S_OK:
109 return 0;
110 case VIRTIO_IOMMU_S_UNSUPP:
111 return -ENOSYS;
112 case VIRTIO_IOMMU_S_INVAL:
113 return -EINVAL;
114 case VIRTIO_IOMMU_S_RANGE:
115 return -ERANGE;
116 case VIRTIO_IOMMU_S_NOENT:
117 return -ENOENT;
118 case VIRTIO_IOMMU_S_FAULT:
119 return -EFAULT;
120 case VIRTIO_IOMMU_S_NOMEM:
121 return -ENOMEM;
122 case VIRTIO_IOMMU_S_IOERR:
123 case VIRTIO_IOMMU_S_DEVERR:
124 default:
125 return -EIO;
126 }
127}
128
129static void viommu_set_req_status(void *buf, size_t len, int status)
130{
131 struct virtio_iommu_req_tail *tail = buf + len - sizeof(*tail);
132
133 tail->status = status;
134}
135
136static off_t viommu_get_write_desc_offset(struct viommu_dev *viommu,
137 struct virtio_iommu_req_head *req,
138 size_t len)
139{
140 size_t tail_size = sizeof(struct virtio_iommu_req_tail);
141
142 if (req->type == VIRTIO_IOMMU_T_PROBE)
143 return len - viommu->probe_size - tail_size;
144
145 return len - tail_size;
146}
147
148/*
149 * __viommu_sync_req - Complete all in-flight requests
150 *
151 * Wait for all added requests to complete. When this function returns, all
152 * requests that were in-flight at the time of the call have completed.
153 */
154static int __viommu_sync_req(struct viommu_dev *viommu)
155{
156 unsigned int len;
157 size_t write_len;
158 struct viommu_request *req;
159 struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ];
160
161 assert_spin_locked(&viommu->request_lock);
162
163 virtqueue_kick(vq);
164
165 while (!list_empty(&viommu->requests)) {
166 len = 0;
167 req = virtqueue_get_buf(vq, &len);
168 if (!req)
169 continue;
170
171 if (!len)
172 viommu_set_req_status(req->buf, req->len,
173 VIRTIO_IOMMU_S_IOERR);
174
175 write_len = req->len - req->write_offset;
176 if (req->writeback && len == write_len)
177 memcpy(req->writeback, req->buf + req->write_offset,
178 write_len);
179
180 list_del(&req->list);
181 kfree(req);
182 }
183
184 return 0;
185}
186
187static int viommu_sync_req(struct viommu_dev *viommu)
188{
189 int ret;
190 unsigned long flags;
191
192 spin_lock_irqsave(&viommu->request_lock, flags);
193 ret = __viommu_sync_req(viommu);
194 if (ret)
195 dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret);
196 spin_unlock_irqrestore(&viommu->request_lock, flags);
197
198 return ret;
199}
200
201/*
202 * __viommu_add_request - Add one request to the queue
203 * @buf: pointer to the request buffer
204 * @len: length of the request buffer
205 * @writeback: copy data back to the buffer when the request completes.
206 *
207 * Add a request to the queue. Only synchronize the queue if it's already full.
208 * Otherwise don't kick the queue nor wait for requests to complete.
209 *
210 * When @writeback is true, data written by the device, including the request
211 * status, is copied into @buf after the request completes. This is unsafe if
212 * the caller allocates @buf on stack and drops the lock between add_req() and
213 * sync_req().
214 *
215 * Return 0 if the request was successfully added to the queue.
216 */
217static int __viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len,
218 bool writeback)
219{
220 int ret;
221 off_t write_offset;
222 struct viommu_request *req;
223 struct scatterlist top_sg, bottom_sg;
224 struct scatterlist *sg[2] = { &top_sg, &bottom_sg };
225 struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ];
226
227 assert_spin_locked(&viommu->request_lock);
228
229 write_offset = viommu_get_write_desc_offset(viommu, buf, len);
230 if (write_offset <= 0)
231 return -EINVAL;
232
233 req = kzalloc(struct_size(req, buf, len), GFP_ATOMIC);
234 if (!req)
235 return -ENOMEM;
236
237 req->len = len;
238 if (writeback) {
239 req->writeback = buf + write_offset;
240 req->write_offset = write_offset;
241 }
242 memcpy(&req->buf, buf, write_offset);
243
244 sg_init_one(&top_sg, req->buf, write_offset);
245 sg_init_one(&bottom_sg, req->buf + write_offset, len - write_offset);
246
247 ret = virtqueue_add_sgs(vq, sg, 1, 1, req, GFP_ATOMIC);
248 if (ret == -ENOSPC) {
249 /* If the queue is full, sync and retry */
250 if (!__viommu_sync_req(viommu))
251 ret = virtqueue_add_sgs(vq, sg, 1, 1, req, GFP_ATOMIC);
252 }
253 if (ret)
254 goto err_free;
255
256 list_add_tail(&req->list, &viommu->requests);
257 return 0;
258
259err_free:
260 kfree(req);
261 return ret;
262}
263
264static int viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len)
265{
266 int ret;
267 unsigned long flags;
268
269 spin_lock_irqsave(&viommu->request_lock, flags);
270 ret = __viommu_add_req(viommu, buf, len, false);
271 if (ret)
272 dev_dbg(viommu->dev, "could not add request: %d\n", ret);
273 spin_unlock_irqrestore(&viommu->request_lock, flags);
274
275 return ret;
276}
277
278/*
279 * Send a request and wait for it to complete. Return the request status (as an
280 * errno)
281 */
282static int viommu_send_req_sync(struct viommu_dev *viommu, void *buf,
283 size_t len)
284{
285 int ret;
286 unsigned long flags;
287
288 spin_lock_irqsave(&viommu->request_lock, flags);
289
290 ret = __viommu_add_req(viommu, buf, len, true);
291 if (ret) {
292 dev_dbg(viommu->dev, "could not add request (%d)\n", ret);
293 goto out_unlock;
294 }
295
296 ret = __viommu_sync_req(viommu);
297 if (ret) {
298 dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret);
299 /* Fall-through (get the actual request status) */
300 }
301
302 ret = viommu_get_req_errno(buf, len);
303out_unlock:
304 spin_unlock_irqrestore(&viommu->request_lock, flags);
305 return ret;
306}
307
308/*
309 * viommu_add_mapping - add a mapping to the internal tree
310 *
311 * On success, return the new mapping. Otherwise return NULL.
312 */
313static int viommu_add_mapping(struct viommu_domain *vdomain, 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 */
344static 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 */
381static 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
421err_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 */
433static 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
465static 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(®ion->list, &next->list);
511 return 0;
512}
513
514static 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
571out_free:
572 kfree(probe);
573 return ret;
574}
575
576static 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
612static 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
640static 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
660static 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 -ENODEV;
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 ret;
701 }
702 }
703
704 return 0;
705}
706
707static 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
720static 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 ret = -EINVAL;
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
791static void viommu_detach_dev(struct viommu_endpoint *vdev)
792{
793 int i;
794 struct virtio_iommu_req_detach req;
795 struct viommu_domain *vdomain = vdev->vdomain;
796 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(vdev->dev);
797
798 if (!vdomain)
799 return;
800
801 req = (struct virtio_iommu_req_detach) {
802 .head.type = VIRTIO_IOMMU_T_DETACH,
803 .domain = cpu_to_le32(vdomain->id),
804 };
805
806 for (i = 0; i < fwspec->num_ids; i++) {
807 req.endpoint = cpu_to_le32(fwspec->ids[i]);
808 WARN_ON(viommu_send_req_sync(vdev->viommu, &req, sizeof(req)));
809 }
810 vdomain->nr_endpoints--;
811 vdev->vdomain = NULL;
812}
813
814static int viommu_map_pages(struct iommu_domain *domain, unsigned long iova,
815 phys_addr_t paddr, size_t pgsize, size_t pgcount,
816 int prot, gfp_t gfp, size_t *mapped)
817{
818 int ret;
819 u32 flags;
820 size_t size = pgsize * pgcount;
821 u64 end = iova + size - 1;
822 struct virtio_iommu_req_map map;
823 struct viommu_domain *vdomain = to_viommu_domain(domain);
824
825 flags = (prot & IOMMU_READ ? VIRTIO_IOMMU_MAP_F_READ : 0) |
826 (prot & IOMMU_WRITE ? VIRTIO_IOMMU_MAP_F_WRITE : 0) |
827 (prot & IOMMU_MMIO ? VIRTIO_IOMMU_MAP_F_MMIO : 0);
828
829 if (flags & ~vdomain->map_flags)
830 return -EINVAL;
831
832 ret = viommu_add_mapping(vdomain, iova, end, paddr, flags);
833 if (ret)
834 return ret;
835
836 if (vdomain->nr_endpoints) {
837 map = (struct virtio_iommu_req_map) {
838 .head.type = VIRTIO_IOMMU_T_MAP,
839 .domain = cpu_to_le32(vdomain->id),
840 .virt_start = cpu_to_le64(iova),
841 .phys_start = cpu_to_le64(paddr),
842 .virt_end = cpu_to_le64(end),
843 .flags = cpu_to_le32(flags),
844 };
845
846 ret = viommu_add_req(vdomain->viommu, &map, sizeof(map));
847 if (ret) {
848 viommu_del_mappings(vdomain, iova, end);
849 return ret;
850 }
851 }
852 if (mapped)
853 *mapped = size;
854
855 return 0;
856}
857
858static size_t viommu_unmap_pages(struct iommu_domain *domain, unsigned long iova,
859 size_t pgsize, size_t pgcount,
860 struct iommu_iotlb_gather *gather)
861{
862 int ret = 0;
863 size_t unmapped;
864 struct virtio_iommu_req_unmap unmap;
865 struct viommu_domain *vdomain = to_viommu_domain(domain);
866 size_t size = pgsize * pgcount;
867
868 unmapped = viommu_del_mappings(vdomain, iova, iova + size - 1);
869 if (unmapped < size)
870 return 0;
871
872 /* Device already removed all mappings after detach. */
873 if (!vdomain->nr_endpoints)
874 return unmapped;
875
876 unmap = (struct virtio_iommu_req_unmap) {
877 .head.type = VIRTIO_IOMMU_T_UNMAP,
878 .domain = cpu_to_le32(vdomain->id),
879 .virt_start = cpu_to_le64(iova),
880 .virt_end = cpu_to_le64(iova + unmapped - 1),
881 };
882
883 ret = viommu_add_req(vdomain->viommu, &unmap, sizeof(unmap));
884 return ret ? 0 : unmapped;
885}
886
887static phys_addr_t viommu_iova_to_phys(struct iommu_domain *domain,
888 dma_addr_t iova)
889{
890 u64 paddr = 0;
891 unsigned long flags;
892 struct viommu_mapping *mapping;
893 struct interval_tree_node *node;
894 struct viommu_domain *vdomain = to_viommu_domain(domain);
895
896 spin_lock_irqsave(&vdomain->mappings_lock, flags);
897 node = interval_tree_iter_first(&vdomain->mappings, iova, iova);
898 if (node) {
899 mapping = container_of(node, struct viommu_mapping, iova);
900 paddr = mapping->paddr + (iova - mapping->iova.start);
901 }
902 spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
903
904 return paddr;
905}
906
907static void viommu_iotlb_sync(struct iommu_domain *domain,
908 struct iommu_iotlb_gather *gather)
909{
910 struct viommu_domain *vdomain = to_viommu_domain(domain);
911
912 viommu_sync_req(vdomain->viommu);
913}
914
915static int viommu_iotlb_sync_map(struct iommu_domain *domain,
916 unsigned long iova, size_t size)
917{
918 struct viommu_domain *vdomain = to_viommu_domain(domain);
919
920 /*
921 * May be called before the viommu is initialized including
922 * while creating direct mapping
923 */
924 if (!vdomain->nr_endpoints)
925 return 0;
926 return viommu_sync_req(vdomain->viommu);
927}
928
929static void viommu_flush_iotlb_all(struct iommu_domain *domain)
930{
931 struct viommu_domain *vdomain = to_viommu_domain(domain);
932
933 /*
934 * May be called before the viommu is initialized including
935 * while creating direct mapping
936 */
937 if (!vdomain->nr_endpoints)
938 return;
939 viommu_sync_req(vdomain->viommu);
940}
941
942static void viommu_get_resv_regions(struct device *dev, struct list_head *head)
943{
944 struct iommu_resv_region *entry, *new_entry, *msi = NULL;
945 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
946 int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
947
948 list_for_each_entry(entry, &vdev->resv_regions, list) {
949 if (entry->type == IOMMU_RESV_MSI)
950 msi = entry;
951
952 new_entry = kmemdup(entry, sizeof(*entry), GFP_KERNEL);
953 if (!new_entry)
954 return;
955 list_add_tail(&new_entry->list, head);
956 }
957
958 /*
959 * If the device didn't register any bypass MSI window, add a
960 * software-mapped region.
961 */
962 if (!msi) {
963 msi = iommu_alloc_resv_region(MSI_IOVA_BASE, MSI_IOVA_LENGTH,
964 prot, IOMMU_RESV_SW_MSI,
965 GFP_KERNEL);
966 if (!msi)
967 return;
968
969 list_add_tail(&msi->list, head);
970 }
971
972 iommu_dma_get_resv_regions(dev, head);
973}
974
975static struct iommu_ops viommu_ops;
976static struct virtio_driver virtio_iommu_drv;
977
978static int viommu_match_node(struct device *dev, const void *data)
979{
980 return device_match_fwnode(dev->parent, data);
981}
982
983static struct viommu_dev *viommu_get_by_fwnode(struct fwnode_handle *fwnode)
984{
985 struct device *dev = driver_find_device(&virtio_iommu_drv.driver, NULL,
986 fwnode, viommu_match_node);
987 put_device(dev);
988
989 return dev ? dev_to_virtio(dev)->priv : NULL;
990}
991
992static struct iommu_device *viommu_probe_device(struct device *dev)
993{
994 int ret;
995 struct viommu_endpoint *vdev;
996 struct viommu_dev *viommu = NULL;
997 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
998
999 viommu = viommu_get_by_fwnode(fwspec->iommu_fwnode);
1000 if (!viommu)
1001 return ERR_PTR(-ENODEV);
1002
1003 vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
1004 if (!vdev)
1005 return ERR_PTR(-ENOMEM);
1006
1007 vdev->dev = dev;
1008 vdev->viommu = viommu;
1009 INIT_LIST_HEAD(&vdev->resv_regions);
1010 dev_iommu_priv_set(dev, vdev);
1011
1012 if (viommu->probe_size) {
1013 /* Get additional information for this endpoint */
1014 ret = viommu_probe_endpoint(viommu, dev);
1015 if (ret)
1016 goto err_free_dev;
1017 }
1018
1019 return &viommu->iommu;
1020
1021err_free_dev:
1022 iommu_put_resv_regions(dev, &vdev->resv_regions);
1023 kfree(vdev);
1024
1025 return ERR_PTR(ret);
1026}
1027
1028static void viommu_probe_finalize(struct device *dev)
1029{
1030#ifndef CONFIG_ARCH_HAS_SETUP_DMA_OPS
1031 /* First clear the DMA ops in case we're switching from a DMA domain */
1032 set_dma_ops(dev, NULL);
1033 iommu_setup_dma_ops(dev, 0, U64_MAX);
1034#endif
1035}
1036
1037static void viommu_release_device(struct device *dev)
1038{
1039 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
1040
1041 viommu_detach_dev(vdev);
1042 iommu_put_resv_regions(dev, &vdev->resv_regions);
1043 kfree(vdev);
1044}
1045
1046static struct iommu_group *viommu_device_group(struct device *dev)
1047{
1048 if (dev_is_pci(dev))
1049 return pci_device_group(dev);
1050 else
1051 return generic_device_group(dev);
1052}
1053
1054static int viommu_of_xlate(struct device *dev,
1055 const struct of_phandle_args *args)
1056{
1057 return iommu_fwspec_add_ids(dev, args->args, 1);
1058}
1059
1060static bool viommu_capable(struct device *dev, enum iommu_cap cap)
1061{
1062 switch (cap) {
1063 case IOMMU_CAP_CACHE_COHERENCY:
1064 return true;
1065 case IOMMU_CAP_DEFERRED_FLUSH:
1066 return true;
1067 default:
1068 return false;
1069 }
1070}
1071
1072static struct iommu_ops viommu_ops = {
1073 .capable = viommu_capable,
1074 .domain_alloc = viommu_domain_alloc,
1075 .probe_device = viommu_probe_device,
1076 .probe_finalize = viommu_probe_finalize,
1077 .release_device = viommu_release_device,
1078 .device_group = viommu_device_group,
1079 .get_resv_regions = viommu_get_resv_regions,
1080 .of_xlate = viommu_of_xlate,
1081 .owner = THIS_MODULE,
1082 .default_domain_ops = &(const struct iommu_domain_ops) {
1083 .attach_dev = viommu_attach_dev,
1084 .map_pages = viommu_map_pages,
1085 .unmap_pages = viommu_unmap_pages,
1086 .iova_to_phys = viommu_iova_to_phys,
1087 .flush_iotlb_all = viommu_flush_iotlb_all,
1088 .iotlb_sync = viommu_iotlb_sync,
1089 .iotlb_sync_map = viommu_iotlb_sync_map,
1090 .free = viommu_domain_free,
1091 }
1092};
1093
1094static int viommu_init_vqs(struct viommu_dev *viommu)
1095{
1096 struct virtio_device *vdev = dev_to_virtio(viommu->dev);
1097 const char *names[] = { "request", "event" };
1098 vq_callback_t *callbacks[] = {
1099 NULL, /* No async requests */
1100 viommu_event_handler,
1101 };
1102
1103 return virtio_find_vqs(vdev, VIOMMU_NR_VQS, viommu->vqs, callbacks,
1104 names, NULL);
1105}
1106
1107static int viommu_fill_evtq(struct viommu_dev *viommu)
1108{
1109 int i, ret;
1110 struct scatterlist sg[1];
1111 struct viommu_event *evts;
1112 struct virtqueue *vq = viommu->vqs[VIOMMU_EVENT_VQ];
1113 size_t nr_evts = vq->num_free;
1114
1115 viommu->evts = evts = devm_kmalloc_array(viommu->dev, nr_evts,
1116 sizeof(*evts), GFP_KERNEL);
1117 if (!evts)
1118 return -ENOMEM;
1119
1120 for (i = 0; i < nr_evts; i++) {
1121 sg_init_one(sg, &evts[i], sizeof(*evts));
1122 ret = virtqueue_add_inbuf(vq, sg, 1, &evts[i], GFP_KERNEL);
1123 if (ret)
1124 return ret;
1125 }
1126
1127 return 0;
1128}
1129
1130static int viommu_probe(struct virtio_device *vdev)
1131{
1132 struct device *parent_dev = vdev->dev.parent;
1133 struct viommu_dev *viommu = NULL;
1134 struct device *dev = &vdev->dev;
1135 u64 input_start = 0;
1136 u64 input_end = -1UL;
1137 int ret;
1138
1139 if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1) ||
1140 !virtio_has_feature(vdev, VIRTIO_IOMMU_F_MAP_UNMAP))
1141 return -ENODEV;
1142
1143 viommu = devm_kzalloc(dev, sizeof(*viommu), GFP_KERNEL);
1144 if (!viommu)
1145 return -ENOMEM;
1146
1147 spin_lock_init(&viommu->request_lock);
1148 ida_init(&viommu->domain_ids);
1149 viommu->dev = dev;
1150 viommu->vdev = vdev;
1151 INIT_LIST_HEAD(&viommu->requests);
1152
1153 ret = viommu_init_vqs(viommu);
1154 if (ret)
1155 return ret;
1156
1157 virtio_cread_le(vdev, struct virtio_iommu_config, page_size_mask,
1158 &viommu->pgsize_bitmap);
1159
1160 if (!viommu->pgsize_bitmap) {
1161 ret = -EINVAL;
1162 goto err_free_vqs;
1163 }
1164
1165 viommu->map_flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE;
1166 viommu->last_domain = ~0U;
1167
1168 /* Optional features */
1169 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1170 struct virtio_iommu_config, input_range.start,
1171 &input_start);
1172
1173 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1174 struct virtio_iommu_config, input_range.end,
1175 &input_end);
1176
1177 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1178 struct virtio_iommu_config, domain_range.start,
1179 &viommu->first_domain);
1180
1181 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1182 struct virtio_iommu_config, domain_range.end,
1183 &viommu->last_domain);
1184
1185 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_PROBE,
1186 struct virtio_iommu_config, probe_size,
1187 &viommu->probe_size);
1188
1189 viommu->geometry = (struct iommu_domain_geometry) {
1190 .aperture_start = input_start,
1191 .aperture_end = input_end,
1192 .force_aperture = true,
1193 };
1194
1195 if (virtio_has_feature(vdev, VIRTIO_IOMMU_F_MMIO))
1196 viommu->map_flags |= VIRTIO_IOMMU_MAP_F_MMIO;
1197
1198 viommu_ops.pgsize_bitmap = viommu->pgsize_bitmap;
1199
1200 virtio_device_ready(vdev);
1201
1202 /* Populate the event queue with buffers */
1203 ret = viommu_fill_evtq(viommu);
1204 if (ret)
1205 goto err_free_vqs;
1206
1207 ret = iommu_device_sysfs_add(&viommu->iommu, dev, NULL, "%s",
1208 virtio_bus_name(vdev));
1209 if (ret)
1210 goto err_free_vqs;
1211
1212 iommu_device_register(&viommu->iommu, &viommu_ops, parent_dev);
1213
1214 vdev->priv = viommu;
1215
1216 dev_info(dev, "input address: %u bits\n",
1217 order_base_2(viommu->geometry.aperture_end));
1218 dev_info(dev, "page mask: %#llx\n", viommu->pgsize_bitmap);
1219
1220 return 0;
1221
1222err_free_vqs:
1223 vdev->config->del_vqs(vdev);
1224
1225 return ret;
1226}
1227
1228static void viommu_remove(struct virtio_device *vdev)
1229{
1230 struct viommu_dev *viommu = vdev->priv;
1231
1232 iommu_device_sysfs_remove(&viommu->iommu);
1233 iommu_device_unregister(&viommu->iommu);
1234
1235 /* Stop all virtqueues */
1236 virtio_reset_device(vdev);
1237 vdev->config->del_vqs(vdev);
1238
1239 dev_info(&vdev->dev, "device removed\n");
1240}
1241
1242static void viommu_config_changed(struct virtio_device *vdev)
1243{
1244 dev_warn(&vdev->dev, "config changed\n");
1245}
1246
1247static unsigned int features[] = {
1248 VIRTIO_IOMMU_F_MAP_UNMAP,
1249 VIRTIO_IOMMU_F_INPUT_RANGE,
1250 VIRTIO_IOMMU_F_DOMAIN_RANGE,
1251 VIRTIO_IOMMU_F_PROBE,
1252 VIRTIO_IOMMU_F_MMIO,
1253 VIRTIO_IOMMU_F_BYPASS_CONFIG,
1254};
1255
1256static struct virtio_device_id id_table[] = {
1257 { VIRTIO_ID_IOMMU, VIRTIO_DEV_ANY_ID },
1258 { 0 },
1259};
1260MODULE_DEVICE_TABLE(virtio, id_table);
1261
1262static struct virtio_driver virtio_iommu_drv = {
1263 .driver.name = KBUILD_MODNAME,
1264 .driver.owner = THIS_MODULE,
1265 .id_table = id_table,
1266 .feature_table = features,
1267 .feature_table_size = ARRAY_SIZE(features),
1268 .probe = viommu_probe,
1269 .remove = viommu_remove,
1270 .config_changed = viommu_config_changed,
1271};
1272
1273module_virtio_driver(virtio_iommu_drv);
1274
1275MODULE_DESCRIPTION("Virtio IOMMU driver");
1276MODULE_AUTHOR("Jean-Philippe Brucker <jean-philippe.brucker@arm.com>");
1277MODULE_LICENSE("GPL v2");