projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
iommu: Statically set module owner
[linux-2.6-block.git] / drivers / iommu / virtio-iommu.c
CommitLineData
edcd69ab
JPB		 1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Virtio driver for the paravirtualized IOMMU
4 *
ae24fb49 5 * Copyright (C) 2019 Arm Limited
edcd69ab
JPB		 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
169a126c
JPB		 32#define VIOMMU_EVENT_VQ 1
33#define VIOMMU_NR_VQS 2
edcd69ab
JPB		 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;
169a126c 45 void *evts;
edcd69ab
JPB		 46
47 /* Device configuration */
48 struct iommu_domain_geometry geometry;
49 u64 pgsize_bitmap;
ae24fb49
JPB		 50 u32 first_domain;
51 u32 last_domain;
52 /* Supported MAP flags */
53 u32 map_flags;
2a5a3148 54 u32 probe_size;
edcd69ab
JPB		 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;
ae24fb49 68 u32 map_flags;
edcd69ab
JPB		 69
70 spinlock_t mappings_lock;
71 struct rb_root_cached mappings;
72
73 unsigned long nr_endpoints;
74};
75
76struct viommu_endpoint {
2a5a3148 77 struct device *dev;
edcd69ab
JPB		 78 struct viommu_dev *viommu;
79 struct viommu_domain *vdomain;
2a5a3148 80 struct list_head resv_regions;
edcd69ab
JPB		 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
169a126c
JPB		 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
edcd69ab
JPB		 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;
ae24fb49
JPB		 120 case VIRTIO_IOMMU_S_NOMEM:
121 return -ENOMEM;
edcd69ab
JPB		 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
2a5a3148
JPB		 142 if (req->type == VIRTIO_IOMMU_T_PROBE)
143 return len - viommu->probe_size - tail_size;
144
edcd69ab
JPB		 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{
edcd69ab
JPB		 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
c1c8058d 184 return 0;
edcd69ab
JPB		 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
2a5a3148
JPB		 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);
df561f66 443 fallthrough;
2a5a3148
JPB		 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
fb3637a1 456 list_add(&region->list, &vdev->resv_regions);
2a5a3148
JPB		 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);
a4b6c2af 469 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
2a5a3148
JPB		 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
169a126c
JPB		 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
edcd69ab
JPB		 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
39b3b3c9 610static int viommu_domain_finalise(struct viommu_endpoint *vdev,
edcd69ab
JPB		 611 struct iommu_domain *domain)
612{
613 int ret;
39b3b3c9
JPB		 614 unsigned long viommu_page_size;
615 struct viommu_dev *viommu = vdev->viommu;
edcd69ab 616 struct viommu_domain *vdomain = to_viommu_domain(domain);
edcd69ab 617
39b3b3c9
JPB		 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
7062af3e
JPB		 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;
edcd69ab
JPB		 632
633 domain->pgsize_bitmap = viommu->pgsize_bitmap;
634 domain->geometry = viommu->geometry;
635
7062af3e
JPB		 636 vdomain->map_flags = viommu->map_flags;
637 vdomain->viommu = viommu;
edcd69ab 638
7062af3e 639 return 0;
edcd69ab
JPB		 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);
a4b6c2af 663 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
edcd69ab
JPB		 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 */
39b3b3c9 672 ret = viommu_domain_finalise(vdev, domain);
edcd69ab
JPB		 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,
781ca2de 727 phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
edcd69ab
JPB		 728{
729 int ret;
ae24fb49 730 u32 flags;
edcd69ab
JPB		 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
ae24fb49
JPB		 738 if (flags & ~vdomain->map_flags)
739 return -EINVAL;
740
edcd69ab
JPB		 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,
56f8af5e 765 size_t size, struct iommu_iotlb_gather *gather)
edcd69ab
JPB		 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
56f8af5e
WD		 811static void viommu_iotlb_sync(struct iommu_domain *domain,
812 struct iommu_iotlb_gather *gather)
edcd69ab
JPB		 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{
2a5a3148 821 struct iommu_resv_region *entry, *new_entry, *msi = NULL;
a4b6c2af 822 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
edcd69ab
JPB		 823 int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
824
2a5a3148
JPB		 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 }
edcd69ab 847
edcd69ab
JPB		 848 iommu_dma_get_resv_regions(dev, head);
849}
850
edcd69ab
JPB		 851static struct iommu_ops viommu_ops;
852static struct virtio_driver virtio_iommu_drv;
853
3a1d5384 854static int viommu_match_node(struct device *dev, const void *data)
edcd69ab
JPB		 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
21acf659 868static struct iommu_device *viommu_probe_device(struct device *dev)
edcd69ab
JPB		 869{
870 int ret;
edcd69ab
JPB		 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)
21acf659 876 return ERR_PTR(-ENODEV);
edcd69ab
JPB		 877
878 viommu = viommu_get_by_fwnode(fwspec->iommu_fwnode);
879 if (!viommu)
21acf659 880 return ERR_PTR(-ENODEV);
edcd69ab
JPB		 881
882 vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
883 if (!vdev)
21acf659 884 return ERR_PTR(-ENOMEM);
edcd69ab 885
2a5a3148 886 vdev->dev = dev;
edcd69ab 887 vdev->viommu = viommu;
2a5a3148 888 INIT_LIST_HEAD(&vdev->resv_regions);
a4b6c2af 889 dev_iommu_priv_set(dev, vdev);
edcd69ab 890
2a5a3148
JPB		 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
21acf659 898 return &viommu->iommu;
edcd69ab 899
edcd69ab 900err_free_dev:
c11738cf 901 generic_iommu_put_resv_regions(dev, &vdev->resv_regions);
edcd69ab
JPB		 902 kfree(vdev);
903
21acf659 904 return ERR_PTR(ret);
edcd69ab
JPB		 905}
906
21acf659 907static void viommu_release_device(struct device *dev)
edcd69ab 908{
edcd69ab 909 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
21acf659 910 struct viommu_endpoint *vdev;
edcd69ab
JPB		 911
912 if (!fwspec || fwspec->ops != &viommu_ops)
913 return;
914
a4b6c2af 915 vdev = dev_iommu_priv_get(dev);
edcd69ab 916
c11738cf 917 generic_iommu_put_resv_regions(dev, &vdev->resv_regions);
edcd69ab
JPB		 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,
21acf659
JR		 942 .probe_device = viommu_probe_device,
943 .release_device = viommu_release_device,
edcd69ab
JPB		 944 .device_group = viommu_device_group,
945 .get_resv_regions = viommu_get_resv_regions,
c11738cf 946 .put_resv_regions = generic_iommu_put_resv_regions,
edcd69ab 947 .of_xlate = viommu_of_xlate,
c0aec668 948 .owner = THIS_MODULE,
edcd69ab
JPB		 949};
950
951static int viommu_init_vqs(struct viommu_dev *viommu)
952{
953 struct virtio_device *vdev = dev_to_virtio(viommu->dev);
169a126c
JPB		 954 const char *names[] = { "request", "event" };
955 vq_callback_t *callbacks[] = {
956 NULL, /* No async requests */
957 viommu_event_handler,
958 };
edcd69ab 959
169a126c
JPB		 960 return virtio_find_vqs(vdev, VIOMMU_NR_VQS, viommu->vqs, callbacks,
961 names, NULL);
962}
edcd69ab 963
169a126c
JPB		 964static int viommu_fill_evtq(struct viommu_dev *viommu)
965{
966 int i, ret;
967 struct scatterlist sg[1];
968 struct viommu_event *evts;
969 struct virtqueue *vq = viommu->vqs[VIOMMU_EVENT_VQ];
970 size_t nr_evts = vq->num_free;
971
972 viommu->evts = evts = devm_kmalloc_array(viommu->dev, nr_evts,
973 sizeof(*evts), GFP_KERNEL);
974 if (!evts)
975 return -ENOMEM;
976
977 for (i = 0; i < nr_evts; i++) {
978 sg_init_one(sg, &evts[i], sizeof(*evts));
979 ret = virtqueue_add_inbuf(vq, sg, 1, &evts[i], GFP_KERNEL);
980 if (ret)
981 return ret;
982 }
edcd69ab
JPB		 983
984 return 0;
985}
986
987static int viommu_probe(struct virtio_device *vdev)
988{
989 struct device *parent_dev = vdev->dev.parent;
990 struct viommu_dev *viommu = NULL;
991 struct device *dev = &vdev->dev;
992 u64 input_start = 0;
993 u64 input_end = -1UL;
994 int ret;
995
996 if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1) ||
997 !virtio_has_feature(vdev, VIRTIO_IOMMU_F_MAP_UNMAP))
998 return -ENODEV;
999
1000 viommu = devm_kzalloc(dev, sizeof(*viommu), GFP_KERNEL);
1001 if (!viommu)
1002 return -ENOMEM;
1003
1004 spin_lock_init(&viommu->request_lock);
1005 ida_init(&viommu->domain_ids);
1006 viommu->dev = dev;
1007 viommu->vdev = vdev;
1008 INIT_LIST_HEAD(&viommu->requests);
1009
1010 ret = viommu_init_vqs(viommu);
1011 if (ret)
1012 return ret;
1013
d83c67c4
MT		 1014 virtio_cread_le(vdev, struct virtio_iommu_config, page_size_mask,
1015 &viommu->pgsize_bitmap);
edcd69ab
JPB		 1016
1017 if (!viommu->pgsize_bitmap) {
1018 ret = -EINVAL;
1019 goto err_free_vqs;
1020 }
1021
ae24fb49
JPB		 1022 viommu->map_flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE;
1023 viommu->last_domain = ~0U;
edcd69ab
JPB		 1024
1025 /* Optional features */
d83c67c4
MT		 1026 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1027 struct virtio_iommu_config, input_range.start,
1028 &input_start);
edcd69ab 1029
d83c67c4
MT		 1030 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1031 struct virtio_iommu_config, input_range.end,
1032 &input_end);
edcd69ab 1033
d83c67c4
MT		 1034 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1035 struct virtio_iommu_config, domain_range.start,
1036 &viommu->first_domain);
ae24fb49 1037
d83c67c4
MT		 1038 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1039 struct virtio_iommu_config, domain_range.end,
1040 &viommu->last_domain);
edcd69ab 1041
d83c67c4
MT		 1042 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_PROBE,
1043 struct virtio_iommu_config, probe_size,
1044 &viommu->probe_size);
2a5a3148 1045
edcd69ab
JPB		 1046 viommu->geometry = (struct iommu_domain_geometry) {
1047 .aperture_start = input_start,
1048 .aperture_end = input_end,
1049 .force_aperture = true,
1050 };
1051
ae24fb49
JPB		 1052 if (virtio_has_feature(vdev, VIRTIO_IOMMU_F_MMIO))
1053 viommu->map_flags |= VIRTIO_IOMMU_MAP_F_MMIO;
1054
edcd69ab
JPB		 1055 viommu_ops.pgsize_bitmap = viommu->pgsize_bitmap;
1056
1057 virtio_device_ready(vdev);
1058
169a126c
JPB		 1059 /* Populate the event queue with buffers */
1060 ret = viommu_fill_evtq(viommu);
1061 if (ret)
1062 goto err_free_vqs;
1063
edcd69ab
JPB		 1064 ret = iommu_device_sysfs_add(&viommu->iommu, dev, NULL, "%s",
1065 virtio_bus_name(vdev));
1066 if (ret)
1067 goto err_free_vqs;
1068
1069 iommu_device_set_ops(&viommu->iommu, &viommu_ops);
1070 iommu_device_set_fwnode(&viommu->iommu, parent_dev->fwnode);
1071
1072 iommu_device_register(&viommu->iommu);
1073
1074#ifdef CONFIG_PCI
1075 if (pci_bus_type.iommu_ops != &viommu_ops) {
edcd69ab
JPB		 1076 ret = bus_set_iommu(&pci_bus_type, &viommu_ops);
1077 if (ret)
1078 goto err_unregister;
1079 }
1080#endif
1081#ifdef CONFIG_ARM_AMBA
1082 if (amba_bustype.iommu_ops != &viommu_ops) {
1083 ret = bus_set_iommu(&amba_bustype, &viommu_ops);
1084 if (ret)
1085 goto err_unregister;
1086 }
1087#endif
1088 if (platform_bus_type.iommu_ops != &viommu_ops) {
1089 ret = bus_set_iommu(&platform_bus_type, &viommu_ops);
1090 if (ret)
1091 goto err_unregister;
1092 }
1093
1094 vdev->priv = viommu;
1095
1096 dev_info(dev, "input address: %u bits\n",
1097 order_base_2(viommu->geometry.aperture_end));
1098 dev_info(dev, "page mask: %#llx\n", viommu->pgsize_bitmap);
1099
1100 return 0;
1101
1102err_unregister:
1103 iommu_device_sysfs_remove(&viommu->iommu);
1104 iommu_device_unregister(&viommu->iommu);
1105err_free_vqs:
1106 vdev->config->del_vqs(vdev);
1107
1108 return ret;
1109}
1110
1111static void viommu_remove(struct virtio_device *vdev)
1112{
1113 struct viommu_dev *viommu = vdev->priv;
1114
1115 iommu_device_sysfs_remove(&viommu->iommu);
1116 iommu_device_unregister(&viommu->iommu);
1117
1118 /* Stop all virtqueues */
1119 vdev->config->reset(vdev);
1120 vdev->config->del_vqs(vdev);
1121
1122 dev_info(&vdev->dev, "device removed\n");
1123}
1124
1125static void viommu_config_changed(struct virtio_device *vdev)
1126{
1127 dev_warn(&vdev->dev, "config changed\n");
1128}
1129
1130static unsigned int features[] = {
1131 VIRTIO_IOMMU_F_MAP_UNMAP,
edcd69ab 1132 VIRTIO_IOMMU_F_INPUT_RANGE,
ae24fb49 1133 VIRTIO_IOMMU_F_DOMAIN_RANGE,
2a5a3148 1134 VIRTIO_IOMMU_F_PROBE,
ae24fb49 1135 VIRTIO_IOMMU_F_MMIO,
edcd69ab
JPB		 1136};
1137
1138static struct virtio_device_id id_table[] = {
1139 { VIRTIO_ID_IOMMU, VIRTIO_DEV_ANY_ID },
1140 { 0 },
1141};
1142
1143static struct virtio_driver virtio_iommu_drv = {
1144 .driver.name = KBUILD_MODNAME,
1145 .driver.owner = THIS_MODULE,
1146 .id_table = id_table,
1147 .feature_table = features,
1148 .feature_table_size = ARRAY_SIZE(features),
1149 .probe = viommu_probe,
1150 .remove = viommu_remove,
1151 .config_changed = viommu_config_changed,
1152};
1153
1154module_virtio_driver(virtio_iommu_drv);
1155
1156MODULE_DESCRIPTION("Virtio IOMMU driver");
1157MODULE_AUTHOR("Jean-Philippe Brucker <jean-philippe.brucker@arm.com>");
1158MODULE_LICENSE("GPL v2");
Linux 6.x block layer and io_uring tree(s)