projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
iommu: Add gfp parameter to iommu_ops::map
[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{
156 int ret = 0;
157 unsigned int len;
158 size_t write_len;
159 struct viommu_request *req;
160 struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ];
161
162 assert_spin_locked(&viommu->request_lock);
163
164 virtqueue_kick(vq);
165
166 while (!list_empty(&viommu->requests)) {
167 len = 0;
168 req = virtqueue_get_buf(vq, &len);
169 if (!req)
170 continue;
171
172 if (!len)
173 viommu_set_req_status(req->buf, req->len,
174 VIRTIO_IOMMU_S_IOERR);
175
176 write_len = req->len - req->write_offset;
177 if (req->writeback && len == write_len)
178 memcpy(req->writeback, req->buf + req->write_offset,
179 write_len);
180
181 list_del(&req->list);
182 kfree(req);
183 }
184
185 return ret;
186}
187
188static int viommu_sync_req(struct viommu_dev *viommu)
189{
190 int ret;
191 unsigned long flags;
192
193 spin_lock_irqsave(&viommu->request_lock, flags);
194 ret = __viommu_sync_req(viommu);
195 if (ret)
196 dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret);
197 spin_unlock_irqrestore(&viommu->request_lock, flags);
198
199 return ret;
200}
201
202/*
203 * __viommu_add_request - Add one request to the queue
204 * @buf: pointer to the request buffer
205 * @len: length of the request buffer
206 * @writeback: copy data back to the buffer when the request completes.
207 *
208 * Add a request to the queue. Only synchronize the queue if it's already full.
209 * Otherwise don't kick the queue nor wait for requests to complete.
210 *
211 * When @writeback is true, data written by the device, including the request
212 * status, is copied into @buf after the request completes. This is unsafe if
213 * the caller allocates @buf on stack and drops the lock between add_req() and
214 * sync_req().
215 *
216 * Return 0 if the request was successfully added to the queue.
217 */
218static int __viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len,
219 bool writeback)
220{
221 int ret;
222 off_t write_offset;
223 struct viommu_request *req;
224 struct scatterlist top_sg, bottom_sg;
225 struct scatterlist *sg[2] = { &top_sg, &bottom_sg };
226 struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ];
227
228 assert_spin_locked(&viommu->request_lock);
229
230 write_offset = viommu_get_write_desc_offset(viommu, buf, len);
231 if (write_offset <= 0)
232 return -EINVAL;
233
234 req = kzalloc(sizeof(*req) + len, GFP_ATOMIC);
235 if (!req)
236 return -ENOMEM;
237
238 req->len = len;
239 if (writeback) {
240 req->writeback = buf + write_offset;
241 req->write_offset = write_offset;
242 }
243 memcpy(&req->buf, buf, write_offset);
244
245 sg_init_one(&top_sg, req->buf, write_offset);
246 sg_init_one(&bottom_sg, req->buf + write_offset, len - write_offset);
247
248 ret = virtqueue_add_sgs(vq, sg, 1, 1, req, GFP_ATOMIC);
249 if (ret == -ENOSPC) {
250 /* If the queue is full, sync and retry */
251 if (!__viommu_sync_req(viommu))
252 ret = virtqueue_add_sgs(vq, sg, 1, 1, req, GFP_ATOMIC);
253 }
254 if (ret)
255 goto err_free;
256
257 list_add_tail(&req->list, &viommu->requests);
258 return 0;
259
260err_free:
261 kfree(req);
262 return ret;
263}
264
265static int viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len)
266{
267 int ret;
268 unsigned long flags;
269
270 spin_lock_irqsave(&viommu->request_lock, flags);
271 ret = __viommu_add_req(viommu, buf, len, false);
272 if (ret)
273 dev_dbg(viommu->dev, "could not add request: %d\n", ret);
274 spin_unlock_irqrestore(&viommu->request_lock, flags);
275
276 return ret;
277}
278
279/*
280 * Send a request and wait for it to complete. Return the request status (as an
281 * errno)
282 */
283static int viommu_send_req_sync(struct viommu_dev *viommu, void *buf,
284 size_t len)
285{
286 int ret;
287 unsigned long flags;
288
289 spin_lock_irqsave(&viommu->request_lock, flags);
290
291 ret = __viommu_add_req(viommu, buf, len, true);
292 if (ret) {
293 dev_dbg(viommu->dev, "could not add request (%d)\n", ret);
294 goto out_unlock;
295 }
296
297 ret = __viommu_sync_req(viommu);
298 if (ret) {
299 dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret);
300 /* Fall-through (get the actual request status) */
301 }
302
303 ret = viommu_get_req_errno(buf, len);
304out_unlock:
305 spin_unlock_irqrestore(&viommu->request_lock, flags);
306 return ret;
307}
308
309/*
310 * viommu_add_mapping - add a mapping to the internal tree
311 *
312 * On success, return the new mapping. Otherwise return NULL.
313 */
314static int viommu_add_mapping(struct viommu_domain *vdomain, unsigned long iova,
315 phys_addr_t paddr, size_t size, u32 flags)
316{
317 unsigned long irqflags;
318 struct viommu_mapping *mapping;
319
320 mapping = kzalloc(sizeof(*mapping), GFP_ATOMIC);
321 if (!mapping)
322 return -ENOMEM;
323
324 mapping->paddr = paddr;
325 mapping->iova.start = iova;
326 mapping->iova.last = iova + size - 1;
327 mapping->flags = flags;
328
329 spin_lock_irqsave(&vdomain->mappings_lock, irqflags);
330 interval_tree_insert(&mapping->iova, &vdomain->mappings);
331 spin_unlock_irqrestore(&vdomain->mappings_lock, irqflags);
332
333 return 0;
334}
335
336/*
337 * viommu_del_mappings - remove mappings from the internal tree
338 *
339 * @vdomain: the domain
340 * @iova: start of the range
341 * @size: size of the range. A size of 0 corresponds to the entire address
342 * space.
343 *
344 * On success, returns the number of unmapped bytes (>= size)
345 */
346static size_t viommu_del_mappings(struct viommu_domain *vdomain,
347 unsigned long iova, size_t size)
348{
349 size_t unmapped = 0;
350 unsigned long flags;
351 unsigned long last = iova + size - 1;
352 struct viommu_mapping *mapping = NULL;
353 struct interval_tree_node *node, *next;
354
355 spin_lock_irqsave(&vdomain->mappings_lock, flags);
356 next = interval_tree_iter_first(&vdomain->mappings, iova, last);
357 while (next) {
358 node = next;
359 mapping = container_of(node, struct viommu_mapping, iova);
360 next = interval_tree_iter_next(node, iova, last);
361
362 /* Trying to split a mapping? */
363 if (mapping->iova.start < iova)
364 break;
365
366 /*
367 * Virtio-iommu doesn't allow UNMAP to split a mapping created
368 * with a single MAP request, so remove the full mapping.
369 */
370 unmapped += mapping->iova.last - mapping->iova.start + 1;
371
372 interval_tree_remove(node, &vdomain->mappings);
373 kfree(mapping);
374 }
375 spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
376
377 return unmapped;
378}
379
380/*
381 * viommu_replay_mappings - re-send MAP requests
382 *
383 * When reattaching a domain that was previously detached from all endpoints,
384 * mappings were deleted from the device. Re-create the mappings available in
385 * the internal tree.
386 */
387static int viommu_replay_mappings(struct viommu_domain *vdomain)
388{
389 int ret = 0;
390 unsigned long flags;
391 struct viommu_mapping *mapping;
392 struct interval_tree_node *node;
393 struct virtio_iommu_req_map map;
394
395 spin_lock_irqsave(&vdomain->mappings_lock, flags);
396 node = interval_tree_iter_first(&vdomain->mappings, 0, -1UL);
397 while (node) {
398 mapping = container_of(node, struct viommu_mapping, iova);
399 map = (struct virtio_iommu_req_map) {
400 .head.type = VIRTIO_IOMMU_T_MAP,
401 .domain = cpu_to_le32(vdomain->id),
402 .virt_start = cpu_to_le64(mapping->iova.start),
403 .virt_end = cpu_to_le64(mapping->iova.last),
404 .phys_start = cpu_to_le64(mapping->paddr),
405 .flags = cpu_to_le32(mapping->flags),
406 };
407
408 ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map));
409 if (ret)
410 break;
411
412 node = interval_tree_iter_next(node, 0, -1UL);
413 }
414 spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
415
416 return ret;
417}
418
2a5a3148
JPB		 419static int viommu_add_resv_mem(struct viommu_endpoint *vdev,
420 struct virtio_iommu_probe_resv_mem *mem,
421 size_t len)
422{
423 size_t size;
424 u64 start64, end64;
425 phys_addr_t start, end;
426 struct iommu_resv_region *region = NULL;
427 unsigned long prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
428
429 start = start64 = le64_to_cpu(mem->start);
430 end = end64 = le64_to_cpu(mem->end);
431 size = end64 - start64 + 1;
432
433 /* Catch any overflow, including the unlikely end64 - start64 + 1 = 0 */
434 if (start != start64 || end != end64 || size < end64 - start64)
435 return -EOVERFLOW;
436
437 if (len < sizeof(*mem))
438 return -EINVAL;
439
440 switch (mem->subtype) {
441 default:
442 dev_warn(vdev->dev, "unknown resv mem subtype 0x%x\n",
443 mem->subtype);
444 /* Fall-through */
445 case VIRTIO_IOMMU_RESV_MEM_T_RESERVED:
446 region = iommu_alloc_resv_region(start, size, 0,
447 IOMMU_RESV_RESERVED);
448 break;
449 case VIRTIO_IOMMU_RESV_MEM_T_MSI:
450 region = iommu_alloc_resv_region(start, size, prot,
451 IOMMU_RESV_MSI);
452 break;
453 }
454 if (!region)
455 return -ENOMEM;
456
457 list_add(&vdev->resv_regions, &region->list);
458 return 0;
459}
460
461static int viommu_probe_endpoint(struct viommu_dev *viommu, struct device *dev)
462{
463 int ret;
464 u16 type, len;
465 size_t cur = 0;
466 size_t probe_len;
467 struct virtio_iommu_req_probe *probe;
468 struct virtio_iommu_probe_property *prop;
469 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
470 struct viommu_endpoint *vdev = fwspec->iommu_priv;
471
472 if (!fwspec->num_ids)
473 return -EINVAL;
474
475 probe_len = sizeof(*probe) + viommu->probe_size +
476 sizeof(struct virtio_iommu_req_tail);
477 probe = kzalloc(probe_len, GFP_KERNEL);
478 if (!probe)
479 return -ENOMEM;
480
481 probe->head.type = VIRTIO_IOMMU_T_PROBE;
482 /*
483 * For now, assume that properties of an endpoint that outputs multiple
484 * IDs are consistent. Only probe the first one.
485 */
486 probe->endpoint = cpu_to_le32(fwspec->ids[0]);
487
488 ret = viommu_send_req_sync(viommu, probe, probe_len);
489 if (ret)
490 goto out_free;
491
492 prop = (void *)probe->properties;
493 type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK;
494
495 while (type != VIRTIO_IOMMU_PROBE_T_NONE &&
496 cur < viommu->probe_size) {
497 len = le16_to_cpu(prop->length) + sizeof(*prop);
498
499 switch (type) {
500 case VIRTIO_IOMMU_PROBE_T_RESV_MEM:
501 ret = viommu_add_resv_mem(vdev, (void *)prop, len);
502 break;
503 default:
504 dev_err(dev, "unknown viommu prop 0x%x\n", type);
505 }
506
507 if (ret)
508 dev_err(dev, "failed to parse viommu prop 0x%x\n", type);
509
510 cur += len;
511 if (cur >= viommu->probe_size)
512 break;
513
514 prop = (void *)probe->properties + cur;
515 type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK;
516 }
517
518out_free:
519 kfree(probe);
520 return ret;
521}
522
169a126c
JPB		 523static int viommu_fault_handler(struct viommu_dev *viommu,
524 struct virtio_iommu_fault *fault)
525{
526 char *reason_str;
527
528 u8 reason = fault->reason;
529 u32 flags = le32_to_cpu(fault->flags);
530 u32 endpoint = le32_to_cpu(fault->endpoint);
531 u64 address = le64_to_cpu(fault->address);
532
533 switch (reason) {
534 case VIRTIO_IOMMU_FAULT_R_DOMAIN:
535 reason_str = "domain";
536 break;
537 case VIRTIO_IOMMU_FAULT_R_MAPPING:
538 reason_str = "page";
539 break;
540 case VIRTIO_IOMMU_FAULT_R_UNKNOWN:
541 default:
542 reason_str = "unknown";
543 break;
544 }
545
546 /* TODO: find EP by ID and report_iommu_fault */
547 if (flags & VIRTIO_IOMMU_FAULT_F_ADDRESS)
548 dev_err_ratelimited(viommu->dev, "%s fault from EP %u at %#llx [%s%s%s]\n",
549 reason_str, endpoint, address,
550 flags & VIRTIO_IOMMU_FAULT_F_READ ? "R" : "",
551 flags & VIRTIO_IOMMU_FAULT_F_WRITE ? "W" : "",
552 flags & VIRTIO_IOMMU_FAULT_F_EXEC ? "X" : "");
553 else
554 dev_err_ratelimited(viommu->dev, "%s fault from EP %u\n",
555 reason_str, endpoint);
556 return 0;
557}
558
559static void viommu_event_handler(struct virtqueue *vq)
560{
561 int ret;
562 unsigned int len;
563 struct scatterlist sg[1];
564 struct viommu_event *evt;
565 struct viommu_dev *viommu = vq->vdev->priv;
566
567 while ((evt = virtqueue_get_buf(vq, &len)) != NULL) {
568 if (len > sizeof(*evt)) {
569 dev_err(viommu->dev,
570 "invalid event buffer (len %u != %zu)\n",
571 len, sizeof(*evt));
572 } else if (!(evt->head & VIOMMU_FAULT_RESV_MASK)) {
573 viommu_fault_handler(viommu, &evt->fault);
574 }
575
576 sg_init_one(sg, evt, sizeof(*evt));
577 ret = virtqueue_add_inbuf(vq, sg, 1, evt, GFP_ATOMIC);
578 if (ret)
579 dev_err(viommu->dev, "could not add event buffer\n");
580 }
581
582 virtqueue_kick(vq);
583}
584
edcd69ab
JPB		 585/* IOMMU API */
586
587static struct iommu_domain *viommu_domain_alloc(unsigned type)
588{
589 struct viommu_domain *vdomain;
590
591 if (type != IOMMU_DOMAIN_UNMANAGED && type != IOMMU_DOMAIN_DMA)
592 return NULL;
593
594 vdomain = kzalloc(sizeof(*vdomain), GFP_KERNEL);
595 if (!vdomain)
596 return NULL;
597
598 mutex_init(&vdomain->mutex);
599 spin_lock_init(&vdomain->mappings_lock);
600 vdomain->mappings = RB_ROOT_CACHED;
601
602 if (type == IOMMU_DOMAIN_DMA &&
603 iommu_get_dma_cookie(&vdomain->domain)) {
604 kfree(vdomain);
605 return NULL;
606 }
607
608 return &vdomain->domain;
609}
610
611static int viommu_domain_finalise(struct viommu_dev *viommu,
612 struct iommu_domain *domain)
613{
614 int ret;
615 struct viommu_domain *vdomain = to_viommu_domain(domain);
edcd69ab
JPB		 616
617 vdomain->viommu = viommu;
ae24fb49 618 vdomain->map_flags = viommu->map_flags;
edcd69ab
JPB		 619
620 domain->pgsize_bitmap = viommu->pgsize_bitmap;
621 domain->geometry = viommu->geometry;
622
ae24fb49
JPB		 623 ret = ida_alloc_range(&viommu->domain_ids, viommu->first_domain,
624 viommu->last_domain, GFP_KERNEL);
edcd69ab
JPB		 625 if (ret >= 0)
626 vdomain->id = (unsigned int)ret;
627
628 return ret > 0 ? 0 : ret;
629}
630
631static void viommu_domain_free(struct iommu_domain *domain)
632{
633 struct viommu_domain *vdomain = to_viommu_domain(domain);
634
635 iommu_put_dma_cookie(domain);
636
637 /* Free all remaining mappings (size 2^64) */
638 viommu_del_mappings(vdomain, 0, 0);
639
640 if (vdomain->viommu)
641 ida_free(&vdomain->viommu->domain_ids, vdomain->id);
642
643 kfree(vdomain);
644}
645
646static int viommu_attach_dev(struct iommu_domain *domain, struct device *dev)
647{
648 int i;
649 int ret = 0;
650 struct virtio_iommu_req_attach req;
651 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
652 struct viommu_endpoint *vdev = fwspec->iommu_priv;
653 struct viommu_domain *vdomain = to_viommu_domain(domain);
654
655 mutex_lock(&vdomain->mutex);
656 if (!vdomain->viommu) {
657 /*
658 * Properly initialize the domain now that we know which viommu
659 * owns it.
660 */
661 ret = viommu_domain_finalise(vdev->viommu, domain);
662 } else if (vdomain->viommu != vdev->viommu) {
663 dev_err(dev, "cannot attach to foreign vIOMMU\n");
664 ret = -EXDEV;
665 }
666 mutex_unlock(&vdomain->mutex);
667
668 if (ret)
669 return ret;
670
671 /*
672 * In the virtio-iommu device, when attaching the endpoint to a new
673 * domain, it is detached from the old one and, if as as a result the
674 * old domain isn't attached to any endpoint, all mappings are removed
675 * from the old domain and it is freed.
676 *
677 * In the driver the old domain still exists, and its mappings will be
678 * recreated if it gets reattached to an endpoint. Otherwise it will be
679 * freed explicitly.
680 *
681 * vdev->vdomain is protected by group->mutex
682 */
683 if (vdev->vdomain)
684 vdev->vdomain->nr_endpoints--;
685
686 req = (struct virtio_iommu_req_attach) {
687 .head.type = VIRTIO_IOMMU_T_ATTACH,
688 .domain = cpu_to_le32(vdomain->id),
689 };
690
691 for (i = 0; i < fwspec->num_ids; i++) {
692 req.endpoint = cpu_to_le32(fwspec->ids[i]);
693
694 ret = viommu_send_req_sync(vdomain->viommu, &req, sizeof(req));
695 if (ret)
696 return ret;
697 }
698
699 if (!vdomain->nr_endpoints) {
700 /*
701 * This endpoint is the first to be attached to the domain.
702 * Replay existing mappings (e.g. SW MSI).
703 */
704 ret = viommu_replay_mappings(vdomain);
705 if (ret)
706 return ret;
707 }
708
709 vdomain->nr_endpoints++;
710 vdev->vdomain = vdomain;
711
712 return 0;
713}
714
715static int viommu_map(struct iommu_domain *domain, unsigned long iova,
781ca2de 716 phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
edcd69ab
JPB		 717{
718 int ret;
ae24fb49 719 u32 flags;
edcd69ab
JPB		 720 struct virtio_iommu_req_map map;
721 struct viommu_domain *vdomain = to_viommu_domain(domain);
722
723 flags = (prot & IOMMU_READ ? VIRTIO_IOMMU_MAP_F_READ : 0) |
724 (prot & IOMMU_WRITE ? VIRTIO_IOMMU_MAP_F_WRITE : 0) |
725 (prot & IOMMU_MMIO ? VIRTIO_IOMMU_MAP_F_MMIO : 0);
726
ae24fb49
JPB		 727 if (flags & ~vdomain->map_flags)
728 return -EINVAL;
729
edcd69ab
JPB		 730 ret = viommu_add_mapping(vdomain, iova, paddr, size, flags);
731 if (ret)
732 return ret;
733
734 map = (struct virtio_iommu_req_map) {
735 .head.type = VIRTIO_IOMMU_T_MAP,
736 .domain = cpu_to_le32(vdomain->id),
737 .virt_start = cpu_to_le64(iova),
738 .phys_start = cpu_to_le64(paddr),
739 .virt_end = cpu_to_le64(iova + size - 1),
740 .flags = cpu_to_le32(flags),
741 };
742
743 if (!vdomain->nr_endpoints)
744 return 0;
745
746 ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map));
747 if (ret)
748 viommu_del_mappings(vdomain, iova, size);
749
750 return ret;
751}
752
753static size_t viommu_unmap(struct iommu_domain *domain, unsigned long iova,
56f8af5e 754 size_t size, struct iommu_iotlb_gather *gather)
edcd69ab
JPB		 755{
756 int ret = 0;
757 size_t unmapped;
758 struct virtio_iommu_req_unmap unmap;
759 struct viommu_domain *vdomain = to_viommu_domain(domain);
760
761 unmapped = viommu_del_mappings(vdomain, iova, size);
762 if (unmapped < size)
763 return 0;
764
765 /* Device already removed all mappings after detach. */
766 if (!vdomain->nr_endpoints)
767 return unmapped;
768
769 unmap = (struct virtio_iommu_req_unmap) {
770 .head.type = VIRTIO_IOMMU_T_UNMAP,
771 .domain = cpu_to_le32(vdomain->id),
772 .virt_start = cpu_to_le64(iova),
773 .virt_end = cpu_to_le64(iova + unmapped - 1),
774 };
775
776 ret = viommu_add_req(vdomain->viommu, &unmap, sizeof(unmap));
777 return ret ? 0 : unmapped;
778}
779
780static phys_addr_t viommu_iova_to_phys(struct iommu_domain *domain,
781 dma_addr_t iova)
782{
783 u64 paddr = 0;
784 unsigned long flags;
785 struct viommu_mapping *mapping;
786 struct interval_tree_node *node;
787 struct viommu_domain *vdomain = to_viommu_domain(domain);
788
789 spin_lock_irqsave(&vdomain->mappings_lock, flags);
790 node = interval_tree_iter_first(&vdomain->mappings, iova, iova);
791 if (node) {
792 mapping = container_of(node, struct viommu_mapping, iova);
793 paddr = mapping->paddr + (iova - mapping->iova.start);
794 }
795 spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
796
797 return paddr;
798}
799
56f8af5e
WD		 800static void viommu_iotlb_sync(struct iommu_domain *domain,
801 struct iommu_iotlb_gather *gather)
edcd69ab
JPB		 802{
803 struct viommu_domain *vdomain = to_viommu_domain(domain);
804
805 viommu_sync_req(vdomain->viommu);
806}
807
808static void viommu_get_resv_regions(struct device *dev, struct list_head *head)
809{
2a5a3148
JPB		 810 struct iommu_resv_region *entry, *new_entry, *msi = NULL;
811 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
812 struct viommu_endpoint *vdev = fwspec->iommu_priv;
edcd69ab
JPB		 813 int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
814
2a5a3148
JPB		 815 list_for_each_entry(entry, &vdev->resv_regions, list) {
816 if (entry->type == IOMMU_RESV_MSI)
817 msi = entry;
818
819 new_entry = kmemdup(entry, sizeof(*entry), GFP_KERNEL);
820 if (!new_entry)
821 return;
822 list_add_tail(&new_entry->list, head);
823 }
824
825 /*
826 * If the device didn't register any bypass MSI window, add a
827 * software-mapped region.
828 */
829 if (!msi) {
830 msi = iommu_alloc_resv_region(MSI_IOVA_BASE, MSI_IOVA_LENGTH,
831 prot, IOMMU_RESV_SW_MSI);
832 if (!msi)
833 return;
834
835 list_add_tail(&msi->list, head);
836 }
edcd69ab 837
edcd69ab
JPB		 838 iommu_dma_get_resv_regions(dev, head);
839}
840
841static void viommu_put_resv_regions(struct device *dev, struct list_head *head)
842{
843 struct iommu_resv_region *entry, *next;
844
845 list_for_each_entry_safe(entry, next, head, list)
846 kfree(entry);
847}
848
849static struct iommu_ops viommu_ops;
850static struct virtio_driver virtio_iommu_drv;
851
3a1d5384 852static int viommu_match_node(struct device *dev, const void *data)
edcd69ab
JPB		 853{
854 return dev->parent->fwnode == data;
855}
856
857static struct viommu_dev *viommu_get_by_fwnode(struct fwnode_handle *fwnode)
858{
859 struct device *dev = driver_find_device(&virtio_iommu_drv.driver, NULL,
860 fwnode, viommu_match_node);
861 put_device(dev);
862
863 return dev ? dev_to_virtio(dev)->priv : NULL;
864}
865
866static int viommu_add_device(struct device *dev)
867{
868 int ret;
869 struct iommu_group *group;
870 struct viommu_endpoint *vdev;
871 struct viommu_dev *viommu = NULL;
872 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
873
874 if (!fwspec || fwspec->ops != &viommu_ops)
875 return -ENODEV;
876
877 viommu = viommu_get_by_fwnode(fwspec->iommu_fwnode);
878 if (!viommu)
879 return -ENODEV;
880
881 vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
882 if (!vdev)
883 return -ENOMEM;
884
2a5a3148 885 vdev->dev = dev;
edcd69ab 886 vdev->viommu = viommu;
2a5a3148 887 INIT_LIST_HEAD(&vdev->resv_regions);
edcd69ab
JPB		 888 fwspec->iommu_priv = vdev;
889
2a5a3148
JPB		 890 if (viommu->probe_size) {
891 /* Get additional information for this endpoint */
892 ret = viommu_probe_endpoint(viommu, dev);
893 if (ret)
894 goto err_free_dev;
895 }
896
edcd69ab
JPB		 897 ret = iommu_device_link(&viommu->iommu, dev);
898 if (ret)
899 goto err_free_dev;
900
901 /*
902 * Last step creates a default domain and attaches to it. Everything
903 * must be ready.
904 */
905 group = iommu_group_get_for_dev(dev);
906 if (IS_ERR(group)) {
907 ret = PTR_ERR(group);
908 goto err_unlink_dev;
909 }
910
911 iommu_group_put(group);
912
913 return PTR_ERR_OR_ZERO(group);
914
915err_unlink_dev:
916 iommu_device_unlink(&viommu->iommu, dev);
917err_free_dev:
2a5a3148 918 viommu_put_resv_regions(dev, &vdev->resv_regions);
edcd69ab
JPB		 919 kfree(vdev);
920
921 return ret;
922}
923
924static void viommu_remove_device(struct device *dev)
925{
926 struct viommu_endpoint *vdev;
927 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
928
929 if (!fwspec || fwspec->ops != &viommu_ops)
930 return;
931
932 vdev = fwspec->iommu_priv;
933
934 iommu_group_remove_device(dev);
935 iommu_device_unlink(&vdev->viommu->iommu, dev);
2a5a3148 936 viommu_put_resv_regions(dev, &vdev->resv_regions);
edcd69ab
JPB		 937 kfree(vdev);
938}
939
940static struct iommu_group *viommu_device_group(struct device *dev)
941{
942 if (dev_is_pci(dev))
943 return pci_device_group(dev);
944 else
945 return generic_device_group(dev);
946}
947
948static int viommu_of_xlate(struct device *dev, struct of_phandle_args *args)
949{
950 return iommu_fwspec_add_ids(dev, args->args, 1);
951}
952
953static struct iommu_ops viommu_ops = {
954 .domain_alloc = viommu_domain_alloc,
955 .domain_free = viommu_domain_free,
956 .attach_dev = viommu_attach_dev,
957 .map = viommu_map,
958 .unmap = viommu_unmap,
959 .iova_to_phys = viommu_iova_to_phys,
960 .iotlb_sync = viommu_iotlb_sync,
961 .add_device = viommu_add_device,
962 .remove_device = viommu_remove_device,
963 .device_group = viommu_device_group,
964 .get_resv_regions = viommu_get_resv_regions,
965 .put_resv_regions = viommu_put_resv_regions,
966 .of_xlate = viommu_of_xlate,
967};
968
969static int viommu_init_vqs(struct viommu_dev *viommu)
970{
971 struct virtio_device *vdev = dev_to_virtio(viommu->dev);
169a126c
JPB		 972 const char *names[] = { "request", "event" };
973 vq_callback_t *callbacks[] = {
974 NULL, /* No async requests */
975 viommu_event_handler,
976 };
edcd69ab 977
169a126c
JPB		 978 return virtio_find_vqs(vdev, VIOMMU_NR_VQS, viommu->vqs, callbacks,
979 names, NULL);
980}
edcd69ab 981
169a126c
JPB		 982static int viommu_fill_evtq(struct viommu_dev *viommu)
983{
984 int i, ret;
985 struct scatterlist sg[1];
986 struct viommu_event *evts;
987 struct virtqueue *vq = viommu->vqs[VIOMMU_EVENT_VQ];
988 size_t nr_evts = vq->num_free;
989
990 viommu->evts = evts = devm_kmalloc_array(viommu->dev, nr_evts,
991 sizeof(*evts), GFP_KERNEL);
992 if (!evts)
993 return -ENOMEM;
994
995 for (i = 0; i < nr_evts; i++) {
996 sg_init_one(sg, &evts[i], sizeof(*evts));
997 ret = virtqueue_add_inbuf(vq, sg, 1, &evts[i], GFP_KERNEL);
998 if (ret)
999 return ret;
1000 }
edcd69ab
JPB		 1001
1002 return 0;
1003}
1004
1005static int viommu_probe(struct virtio_device *vdev)
1006{
1007 struct device *parent_dev = vdev->dev.parent;
1008 struct viommu_dev *viommu = NULL;
1009 struct device *dev = &vdev->dev;
1010 u64 input_start = 0;
1011 u64 input_end = -1UL;
1012 int ret;
1013
1014 if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1) ||
1015 !virtio_has_feature(vdev, VIRTIO_IOMMU_F_MAP_UNMAP))
1016 return -ENODEV;
1017
1018 viommu = devm_kzalloc(dev, sizeof(*viommu), GFP_KERNEL);
1019 if (!viommu)
1020 return -ENOMEM;
1021
1022 spin_lock_init(&viommu->request_lock);
1023 ida_init(&viommu->domain_ids);
1024 viommu->dev = dev;
1025 viommu->vdev = vdev;
1026 INIT_LIST_HEAD(&viommu->requests);
1027
1028 ret = viommu_init_vqs(viommu);
1029 if (ret)
1030 return ret;
1031
1032 virtio_cread(vdev, struct virtio_iommu_config, page_size_mask,
1033 &viommu->pgsize_bitmap);
1034
1035 if (!viommu->pgsize_bitmap) {
1036 ret = -EINVAL;
1037 goto err_free_vqs;
1038 }
1039
ae24fb49
JPB		 1040 viommu->map_flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE;
1041 viommu->last_domain = ~0U;
edcd69ab
JPB		 1042
1043 /* Optional features */
1044 virtio_cread_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1045 struct virtio_iommu_config, input_range.start,
1046 &input_start);
1047
1048 virtio_cread_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1049 struct virtio_iommu_config, input_range.end,
1050 &input_end);
1051
ae24fb49
JPB		 1052 virtio_cread_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1053 struct virtio_iommu_config, domain_range.start,
1054 &viommu->first_domain);
1055
1056 virtio_cread_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1057 struct virtio_iommu_config, domain_range.end,
1058 &viommu->last_domain);
edcd69ab 1059
2a5a3148
JPB		 1060 virtio_cread_feature(vdev, VIRTIO_IOMMU_F_PROBE,
1061 struct virtio_iommu_config, probe_size,
1062 &viommu->probe_size);
1063
edcd69ab
JPB		 1064 viommu->geometry = (struct iommu_domain_geometry) {
1065 .aperture_start = input_start,
1066 .aperture_end = input_end,
1067 .force_aperture = true,
1068 };
1069
ae24fb49
JPB		 1070 if (virtio_has_feature(vdev, VIRTIO_IOMMU_F_MMIO))
1071 viommu->map_flags |= VIRTIO_IOMMU_MAP_F_MMIO;
1072
edcd69ab
JPB		 1073 viommu_ops.pgsize_bitmap = viommu->pgsize_bitmap;
1074
1075 virtio_device_ready(vdev);
1076
169a126c
JPB		 1077 /* Populate the event queue with buffers */
1078 ret = viommu_fill_evtq(viommu);
1079 if (ret)
1080 goto err_free_vqs;
1081
edcd69ab
JPB		 1082 ret = iommu_device_sysfs_add(&viommu->iommu, dev, NULL, "%s",
1083 virtio_bus_name(vdev));
1084 if (ret)
1085 goto err_free_vqs;
1086
1087 iommu_device_set_ops(&viommu->iommu, &viommu_ops);
1088 iommu_device_set_fwnode(&viommu->iommu, parent_dev->fwnode);
1089
1090 iommu_device_register(&viommu->iommu);
1091
1092#ifdef CONFIG_PCI
1093 if (pci_bus_type.iommu_ops != &viommu_ops) {
1094 pci_request_acs();
1095 ret = bus_set_iommu(&pci_bus_type, &viommu_ops);
1096 if (ret)
1097 goto err_unregister;
1098 }
1099#endif
1100#ifdef CONFIG_ARM_AMBA
1101 if (amba_bustype.iommu_ops != &viommu_ops) {
1102 ret = bus_set_iommu(&amba_bustype, &viommu_ops);
1103 if (ret)
1104 goto err_unregister;
1105 }
1106#endif
1107 if (platform_bus_type.iommu_ops != &viommu_ops) {
1108 ret = bus_set_iommu(&platform_bus_type, &viommu_ops);
1109 if (ret)
1110 goto err_unregister;
1111 }
1112
1113 vdev->priv = viommu;
1114
1115 dev_info(dev, "input address: %u bits\n",
1116 order_base_2(viommu->geometry.aperture_end));
1117 dev_info(dev, "page mask: %#llx\n", viommu->pgsize_bitmap);
1118
1119 return 0;
1120
1121err_unregister:
1122 iommu_device_sysfs_remove(&viommu->iommu);
1123 iommu_device_unregister(&viommu->iommu);
1124err_free_vqs:
1125 vdev->config->del_vqs(vdev);
1126
1127 return ret;
1128}
1129
1130static void viommu_remove(struct virtio_device *vdev)
1131{
1132 struct viommu_dev *viommu = vdev->priv;
1133
1134 iommu_device_sysfs_remove(&viommu->iommu);
1135 iommu_device_unregister(&viommu->iommu);
1136
1137 /* Stop all virtqueues */
1138 vdev->config->reset(vdev);
1139 vdev->config->del_vqs(vdev);
1140
1141 dev_info(&vdev->dev, "device removed\n");
1142}
1143
1144static void viommu_config_changed(struct virtio_device *vdev)
1145{
1146 dev_warn(&vdev->dev, "config changed\n");
1147}
1148
1149static unsigned int features[] = {
1150 VIRTIO_IOMMU_F_MAP_UNMAP,
edcd69ab 1151 VIRTIO_IOMMU_F_INPUT_RANGE,
ae24fb49 1152 VIRTIO_IOMMU_F_DOMAIN_RANGE,
2a5a3148 1153 VIRTIO_IOMMU_F_PROBE,
ae24fb49 1154 VIRTIO_IOMMU_F_MMIO,
edcd69ab
JPB		 1155};
1156
1157static struct virtio_device_id id_table[] = {
1158 { VIRTIO_ID_IOMMU, VIRTIO_DEV_ANY_ID },
1159 { 0 },
1160};
1161
1162static struct virtio_driver virtio_iommu_drv = {
1163 .driver.name = KBUILD_MODNAME,
1164 .driver.owner = THIS_MODULE,
1165 .id_table = id_table,
1166 .feature_table = features,
1167 .feature_table_size = ARRAY_SIZE(features),
1168 .probe = viommu_probe,
1169 .remove = viommu_remove,
1170 .config_changed = viommu_config_changed,
1171};
1172
1173module_virtio_driver(virtio_iommu_drv);
1174
1175MODULE_DESCRIPTION("Virtio IOMMU driver");
1176MODULE_AUTHOR("Jean-Philippe Brucker <jean-philippe.brucker@arm.com>");
1177MODULE_LICENSE("GPL v2");
Linux 6.x block layer and io_uring tree(s)