Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2007-2008 Advanced Micro Devices, Inc.
4 * Author: Joerg Roedel <jroedel@suse.de>
5 */
6
7#define pr_fmt(fmt) "iommu: " fmt
8
9#include <linux/amba/bus.h>
10#include <linux/device.h>
11#include <linux/kernel.h>
12#include <linux/bits.h>
13#include <linux/bug.h>
14#include <linux/types.h>
15#include <linux/init.h>
16#include <linux/export.h>
17#include <linux/slab.h>
18#include <linux/errno.h>
19#include <linux/host1x_context_bus.h>
20#include <linux/iommu.h>
21#include <linux/idr.h>
22#include <linux/err.h>
23#include <linux/pci.h>
24#include <linux/pci-ats.h>
25#include <linux/bitops.h>
26#include <linux/platform_device.h>
27#include <linux/property.h>
28#include <linux/fsl/mc.h>
29#include <linux/module.h>
30#include <linux/cc_platform.h>
31#include <trace/events/iommu.h>
32#include <linux/sched/mm.h>
33
34#include "dma-iommu.h"
35
36#include "iommu-sva.h"
37
38static struct kset *iommu_group_kset;
39static DEFINE_IDA(iommu_group_ida);
40
41static unsigned int iommu_def_domain_type __read_mostly;
42static bool iommu_dma_strict __read_mostly = IS_ENABLED(CONFIG_IOMMU_DEFAULT_DMA_STRICT);
43static u32 iommu_cmd_line __read_mostly;
44
45struct iommu_group {
46 struct kobject kobj;
47 struct kobject *devices_kobj;
48 struct list_head devices;
49 struct xarray pasid_array;
50 struct mutex mutex;
51 void *iommu_data;
52 void (*iommu_data_release)(void *iommu_data);
53 char *name;
54 int id;
55 struct iommu_domain *default_domain;
56 struct iommu_domain *blocking_domain;
57 struct iommu_domain *domain;
58 struct list_head entry;
59 unsigned int owner_cnt;
60 void *owner;
61};
62
63struct group_device {
64 struct list_head list;
65 struct device *dev;
66 char *name;
67};
68
69struct iommu_group_attribute {
70 struct attribute attr;
71 ssize_t (*show)(struct iommu_group *group, char *buf);
72 ssize_t (*store)(struct iommu_group *group,
73 const char *buf, size_t count);
74};
75
76static const char * const iommu_group_resv_type_string[] = {
77 [IOMMU_RESV_DIRECT] = "direct",
78 [IOMMU_RESV_DIRECT_RELAXABLE] = "direct-relaxable",
79 [IOMMU_RESV_RESERVED] = "reserved",
80 [IOMMU_RESV_MSI] = "msi",
81 [IOMMU_RESV_SW_MSI] = "msi",
82};
83
84#define IOMMU_CMD_LINE_DMA_API BIT(0)
85#define IOMMU_CMD_LINE_STRICT BIT(1)
86
87static int iommu_bus_notifier(struct notifier_block *nb,
88 unsigned long action, void *data);
89static int iommu_alloc_default_domain(struct iommu_group *group,
90 struct device *dev);
91static struct iommu_domain *__iommu_domain_alloc(struct bus_type *bus,
92 unsigned type);
93static int __iommu_attach_device(struct iommu_domain *domain,
94 struct device *dev);
95static int __iommu_attach_group(struct iommu_domain *domain,
96 struct iommu_group *group);
97static int __iommu_group_set_domain(struct iommu_group *group,
98 struct iommu_domain *new_domain);
99static int iommu_create_device_direct_mappings(struct iommu_group *group,
100 struct device *dev);
101static struct iommu_group *iommu_group_get_for_dev(struct device *dev);
102static ssize_t iommu_group_store_type(struct iommu_group *group,
103 const char *buf, size_t count);
104
105#define IOMMU_GROUP_ATTR(_name, _mode, _show, _store) \
106struct iommu_group_attribute iommu_group_attr_##_name = \
107 __ATTR(_name, _mode, _show, _store)
108
109#define to_iommu_group_attr(_attr) \
110 container_of(_attr, struct iommu_group_attribute, attr)
111#define to_iommu_group(_kobj) \
112 container_of(_kobj, struct iommu_group, kobj)
113
114static LIST_HEAD(iommu_device_list);
115static DEFINE_SPINLOCK(iommu_device_lock);
116
117static struct bus_type * const iommu_buses[] = {
118 &platform_bus_type,
119#ifdef CONFIG_PCI
120 &pci_bus_type,
121#endif
122#ifdef CONFIG_ARM_AMBA
123 &amba_bustype,
124#endif
125#ifdef CONFIG_FSL_MC_BUS
126 &fsl_mc_bus_type,
127#endif
128#ifdef CONFIG_TEGRA_HOST1X_CONTEXT_BUS
129 &host1x_context_device_bus_type,
130#endif
131};
132
133/*
134 * Use a function instead of an array here because the domain-type is a
135 * bit-field, so an array would waste memory.
136 */
137static const char *iommu_domain_type_str(unsigned int t)
138{
139 switch (t) {
140 case IOMMU_DOMAIN_BLOCKED:
141 return "Blocked";
142 case IOMMU_DOMAIN_IDENTITY:
143 return "Passthrough";
144 case IOMMU_DOMAIN_UNMANAGED:
145 return "Unmanaged";
146 case IOMMU_DOMAIN_DMA:
147 case IOMMU_DOMAIN_DMA_FQ:
148 return "Translated";
149 default:
150 return "Unknown";
151 }
152}
153
154static int __init iommu_subsys_init(void)
155{
156 struct notifier_block *nb;
157
158 if (!(iommu_cmd_line & IOMMU_CMD_LINE_DMA_API)) {
159 if (IS_ENABLED(CONFIG_IOMMU_DEFAULT_PASSTHROUGH))
160 iommu_set_default_passthrough(false);
161 else
162 iommu_set_default_translated(false);
163
164 if (iommu_default_passthrough() && cc_platform_has(CC_ATTR_MEM_ENCRYPT)) {
165 pr_info("Memory encryption detected - Disabling default IOMMU Passthrough\n");
166 iommu_set_default_translated(false);
167 }
168 }
169
170 if (!iommu_default_passthrough() && !iommu_dma_strict)
171 iommu_def_domain_type = IOMMU_DOMAIN_DMA_FQ;
172
173 pr_info("Default domain type: %s %s\n",
174 iommu_domain_type_str(iommu_def_domain_type),
175 (iommu_cmd_line & IOMMU_CMD_LINE_DMA_API) ?
176 "(set via kernel command line)" : "");
177
178 if (!iommu_default_passthrough())
179 pr_info("DMA domain TLB invalidation policy: %s mode %s\n",
180 iommu_dma_strict ? "strict" : "lazy",
181 (iommu_cmd_line & IOMMU_CMD_LINE_STRICT) ?
182 "(set via kernel command line)" : "");
183
184 nb = kcalloc(ARRAY_SIZE(iommu_buses), sizeof(*nb), GFP_KERNEL);
185 if (!nb)
186 return -ENOMEM;
187
188 for (int i = 0; i < ARRAY_SIZE(iommu_buses); i++) {
189 nb[i].notifier_call = iommu_bus_notifier;
190 bus_register_notifier(iommu_buses[i], &nb[i]);
191 }
192
193 return 0;
194}
195subsys_initcall(iommu_subsys_init);
196
197static int remove_iommu_group(struct device *dev, void *data)
198{
199 if (dev->iommu && dev->iommu->iommu_dev == data)
200 iommu_release_device(dev);
201
202 return 0;
203}
204
205/**
206 * iommu_device_register() - Register an IOMMU hardware instance
207 * @iommu: IOMMU handle for the instance
208 * @ops: IOMMU ops to associate with the instance
209 * @hwdev: (optional) actual instance device, used for fwnode lookup
210 *
211 * Return: 0 on success, or an error.
212 */
213int iommu_device_register(struct iommu_device *iommu,
214 const struct iommu_ops *ops, struct device *hwdev)
215{
216 int err = 0;
217
218 /* We need to be able to take module references appropriately */
219 if (WARN_ON(is_module_address((unsigned long)ops) && !ops->owner))
220 return -EINVAL;
221 /*
222 * Temporarily enforce global restriction to a single driver. This was
223 * already the de-facto behaviour, since any possible combination of
224 * existing drivers would compete for at least the PCI or platform bus.
225 */
226 if (iommu_buses[0]->iommu_ops && iommu_buses[0]->iommu_ops != ops)
227 return -EBUSY;
228
229 iommu->ops = ops;
230 if (hwdev)
231 iommu->fwnode = dev_fwnode(hwdev);
232
233 spin_lock(&iommu_device_lock);
234 list_add_tail(&iommu->list, &iommu_device_list);
235 spin_unlock(&iommu_device_lock);
236
237 for (int i = 0; i < ARRAY_SIZE(iommu_buses) && !err; i++) {
238 iommu_buses[i]->iommu_ops = ops;
239 err = bus_iommu_probe(iommu_buses[i]);
240 }
241 if (err)
242 iommu_device_unregister(iommu);
243 return err;
244}
245EXPORT_SYMBOL_GPL(iommu_device_register);
246
247void iommu_device_unregister(struct iommu_device *iommu)
248{
249 for (int i = 0; i < ARRAY_SIZE(iommu_buses); i++)
250 bus_for_each_dev(iommu_buses[i], NULL, iommu, remove_iommu_group);
251
252 spin_lock(&iommu_device_lock);
253 list_del(&iommu->list);
254 spin_unlock(&iommu_device_lock);
255}
256EXPORT_SYMBOL_GPL(iommu_device_unregister);
257
258static struct dev_iommu *dev_iommu_get(struct device *dev)
259{
260 struct dev_iommu *param = dev->iommu;
261
262 if (param)
263 return param;
264
265 param = kzalloc(sizeof(*param), GFP_KERNEL);
266 if (!param)
267 return NULL;
268
269 mutex_init(¶m->lock);
270 dev->iommu = param;
271 return param;
272}
273
274static void dev_iommu_free(struct device *dev)
275{
276 struct dev_iommu *param = dev->iommu;
277
278 dev->iommu = NULL;
279 if (param->fwspec) {
280 fwnode_handle_put(param->fwspec->iommu_fwnode);
281 kfree(param->fwspec);
282 }
283 kfree(param);
284}
285
286static u32 dev_iommu_get_max_pasids(struct device *dev)
287{
288 u32 max_pasids = 0, bits = 0;
289 int ret;
290
291 if (dev_is_pci(dev)) {
292 ret = pci_max_pasids(to_pci_dev(dev));
293 if (ret > 0)
294 max_pasids = ret;
295 } else {
296 ret = device_property_read_u32(dev, "pasid-num-bits", &bits);
297 if (!ret)
298 max_pasids = 1UL << bits;
299 }
300
301 return min_t(u32, max_pasids, dev->iommu->iommu_dev->max_pasids);
302}
303
304static int __iommu_probe_device(struct device *dev, struct list_head *group_list)
305{
306 const struct iommu_ops *ops = dev->bus->iommu_ops;
307 struct iommu_device *iommu_dev;
308 struct iommu_group *group;
309 static DEFINE_MUTEX(iommu_probe_device_lock);
310 int ret;
311
312 if (!ops)
313 return -ENODEV;
314 /*
315 * Serialise to avoid races between IOMMU drivers registering in
316 * parallel and/or the "replay" calls from ACPI/OF code via client
317 * driver probe. Once the latter have been cleaned up we should
318 * probably be able to use device_lock() here to minimise the scope,
319 * but for now enforcing a simple global ordering is fine.
320 */
321 mutex_lock(&iommu_probe_device_lock);
322 if (!dev_iommu_get(dev)) {
323 ret = -ENOMEM;
324 goto err_unlock;
325 }
326
327 if (!try_module_get(ops->owner)) {
328 ret = -EINVAL;
329 goto err_free;
330 }
331
332 iommu_dev = ops->probe_device(dev);
333 if (IS_ERR(iommu_dev)) {
334 ret = PTR_ERR(iommu_dev);
335 goto out_module_put;
336 }
337
338 dev->iommu->iommu_dev = iommu_dev;
339 dev->iommu->max_pasids = dev_iommu_get_max_pasids(dev);
340
341 group = iommu_group_get_for_dev(dev);
342 if (IS_ERR(group)) {
343 ret = PTR_ERR(group);
344 goto out_release;
345 }
346
347 mutex_lock(&group->mutex);
348 if (group_list && !group->default_domain && list_empty(&group->entry))
349 list_add_tail(&group->entry, group_list);
350 mutex_unlock(&group->mutex);
351 iommu_group_put(group);
352
353 mutex_unlock(&iommu_probe_device_lock);
354 iommu_device_link(iommu_dev, dev);
355
356 return 0;
357
358out_release:
359 if (ops->release_device)
360 ops->release_device(dev);
361
362out_module_put:
363 module_put(ops->owner);
364
365err_free:
366 dev_iommu_free(dev);
367
368err_unlock:
369 mutex_unlock(&iommu_probe_device_lock);
370
371 return ret;
372}
373
374int iommu_probe_device(struct device *dev)
375{
376 const struct iommu_ops *ops;
377 struct iommu_group *group;
378 int ret;
379
380 ret = __iommu_probe_device(dev, NULL);
381 if (ret)
382 goto err_out;
383
384 group = iommu_group_get(dev);
385 if (!group) {
386 ret = -ENODEV;
387 goto err_release;
388 }
389
390 /*
391 * Try to allocate a default domain - needs support from the
392 * IOMMU driver. There are still some drivers which don't
393 * support default domains, so the return value is not yet
394 * checked.
395 */
396 mutex_lock(&group->mutex);
397 iommu_alloc_default_domain(group, dev);
398
399 /*
400 * If device joined an existing group which has been claimed, don't
401 * attach the default domain.
402 */
403 if (group->default_domain && !group->owner) {
404 ret = __iommu_attach_device(group->default_domain, dev);
405 if (ret) {
406 mutex_unlock(&group->mutex);
407 iommu_group_put(group);
408 goto err_release;
409 }
410 }
411
412 iommu_create_device_direct_mappings(group, dev);
413
414 mutex_unlock(&group->mutex);
415 iommu_group_put(group);
416
417 ops = dev_iommu_ops(dev);
418 if (ops->probe_finalize)
419 ops->probe_finalize(dev);
420
421 return 0;
422
423err_release:
424 iommu_release_device(dev);
425
426err_out:
427 return ret;
428
429}
430
431void iommu_release_device(struct device *dev)
432{
433 const struct iommu_ops *ops;
434
435 if (!dev->iommu)
436 return;
437
438 iommu_device_unlink(dev->iommu->iommu_dev, dev);
439
440 ops = dev_iommu_ops(dev);
441 if (ops->release_device)
442 ops->release_device(dev);
443
444 iommu_group_remove_device(dev);
445 module_put(ops->owner);
446 dev_iommu_free(dev);
447}
448
449static int __init iommu_set_def_domain_type(char *str)
450{
451 bool pt;
452 int ret;
453
454 ret = kstrtobool(str, &pt);
455 if (ret)
456 return ret;
457
458 if (pt)
459 iommu_set_default_passthrough(true);
460 else
461 iommu_set_default_translated(true);
462
463 return 0;
464}
465early_param("iommu.passthrough", iommu_set_def_domain_type);
466
467static int __init iommu_dma_setup(char *str)
468{
469 int ret = kstrtobool(str, &iommu_dma_strict);
470
471 if (!ret)
472 iommu_cmd_line |= IOMMU_CMD_LINE_STRICT;
473 return ret;
474}
475early_param("iommu.strict", iommu_dma_setup);
476
477void iommu_set_dma_strict(void)
478{
479 iommu_dma_strict = true;
480 if (iommu_def_domain_type == IOMMU_DOMAIN_DMA_FQ)
481 iommu_def_domain_type = IOMMU_DOMAIN_DMA;
482}
483
484static ssize_t iommu_group_attr_show(struct kobject *kobj,
485 struct attribute *__attr, char *buf)
486{
487 struct iommu_group_attribute *attr = to_iommu_group_attr(__attr);
488 struct iommu_group *group = to_iommu_group(kobj);
489 ssize_t ret = -EIO;
490
491 if (attr->show)
492 ret = attr->show(group, buf);
493 return ret;
494}
495
496static ssize_t iommu_group_attr_store(struct kobject *kobj,
497 struct attribute *__attr,
498 const char *buf, size_t count)
499{
500 struct iommu_group_attribute *attr = to_iommu_group_attr(__attr);
501 struct iommu_group *group = to_iommu_group(kobj);
502 ssize_t ret = -EIO;
503
504 if (attr->store)
505 ret = attr->store(group, buf, count);
506 return ret;
507}
508
509static const struct sysfs_ops iommu_group_sysfs_ops = {
510 .show = iommu_group_attr_show,
511 .store = iommu_group_attr_store,
512};
513
514static int iommu_group_create_file(struct iommu_group *group,
515 struct iommu_group_attribute *attr)
516{
517 return sysfs_create_file(&group->kobj, &attr->attr);
518}
519
520static void iommu_group_remove_file(struct iommu_group *group,
521 struct iommu_group_attribute *attr)
522{
523 sysfs_remove_file(&group->kobj, &attr->attr);
524}
525
526static ssize_t iommu_group_show_name(struct iommu_group *group, char *buf)
527{
528 return sprintf(buf, "%s\n", group->name);
529}
530
531/**
532 * iommu_insert_resv_region - Insert a new region in the
533 * list of reserved regions.
534 * @new: new region to insert
535 * @regions: list of regions
536 *
537 * Elements are sorted by start address and overlapping segments
538 * of the same type are merged.
539 */
540static int iommu_insert_resv_region(struct iommu_resv_region *new,
541 struct list_head *regions)
542{
543 struct iommu_resv_region *iter, *tmp, *nr, *top;
544 LIST_HEAD(stack);
545
546 nr = iommu_alloc_resv_region(new->start, new->length,
547 new->prot, new->type, GFP_KERNEL);
548 if (!nr)
549 return -ENOMEM;
550
551 /* First add the new element based on start address sorting */
552 list_for_each_entry(iter, regions, list) {
553 if (nr->start < iter->start ||
554 (nr->start == iter->start && nr->type <= iter->type))
555 break;
556 }
557 list_add_tail(&nr->list, &iter->list);
558
559 /* Merge overlapping segments of type nr->type in @regions, if any */
560 list_for_each_entry_safe(iter, tmp, regions, list) {
561 phys_addr_t top_end, iter_end = iter->start + iter->length - 1;
562
563 /* no merge needed on elements of different types than @new */
564 if (iter->type != new->type) {
565 list_move_tail(&iter->list, &stack);
566 continue;
567 }
568
569 /* look for the last stack element of same type as @iter */
570 list_for_each_entry_reverse(top, &stack, list)
571 if (top->type == iter->type)
572 goto check_overlap;
573
574 list_move_tail(&iter->list, &stack);
575 continue;
576
577check_overlap:
578 top_end = top->start + top->length - 1;
579
580 if (iter->start > top_end + 1) {
581 list_move_tail(&iter->list, &stack);
582 } else {
583 top->length = max(top_end, iter_end) - top->start + 1;
584 list_del(&iter->list);
585 kfree(iter);
586 }
587 }
588 list_splice(&stack, regions);
589 return 0;
590}
591
592static int
593iommu_insert_device_resv_regions(struct list_head *dev_resv_regions,
594 struct list_head *group_resv_regions)
595{
596 struct iommu_resv_region *entry;
597 int ret = 0;
598
599 list_for_each_entry(entry, dev_resv_regions, list) {
600 ret = iommu_insert_resv_region(entry, group_resv_regions);
601 if (ret)
602 break;
603 }
604 return ret;
605}
606
607int iommu_get_group_resv_regions(struct iommu_group *group,
608 struct list_head *head)
609{
610 struct group_device *device;
611 int ret = 0;
612
613 mutex_lock(&group->mutex);
614 list_for_each_entry(device, &group->devices, list) {
615 struct list_head dev_resv_regions;
616
617 /*
618 * Non-API groups still expose reserved_regions in sysfs,
619 * so filter out calls that get here that way.
620 */
621 if (!device->dev->iommu)
622 break;
623
624 INIT_LIST_HEAD(&dev_resv_regions);
625 iommu_get_resv_regions(device->dev, &dev_resv_regions);
626 ret = iommu_insert_device_resv_regions(&dev_resv_regions, head);
627 iommu_put_resv_regions(device->dev, &dev_resv_regions);
628 if (ret)
629 break;
630 }
631 mutex_unlock(&group->mutex);
632 return ret;
633}
634EXPORT_SYMBOL_GPL(iommu_get_group_resv_regions);
635
636static ssize_t iommu_group_show_resv_regions(struct iommu_group *group,
637 char *buf)
638{
639 struct iommu_resv_region *region, *next;
640 struct list_head group_resv_regions;
641 char *str = buf;
642
643 INIT_LIST_HEAD(&group_resv_regions);
644 iommu_get_group_resv_regions(group, &group_resv_regions);
645
646 list_for_each_entry_safe(region, next, &group_resv_regions, list) {
647 str += sprintf(str, "0x%016llx 0x%016llx %s\n",
648 (long long int)region->start,
649 (long long int)(region->start +
650 region->length - 1),
651 iommu_group_resv_type_string[region->type]);
652 kfree(region);
653 }
654
655 return (str - buf);
656}
657
658static ssize_t iommu_group_show_type(struct iommu_group *group,
659 char *buf)
660{
661 char *type = "unknown\n";
662
663 mutex_lock(&group->mutex);
664 if (group->default_domain) {
665 switch (group->default_domain->type) {
666 case IOMMU_DOMAIN_BLOCKED:
667 type = "blocked\n";
668 break;
669 case IOMMU_DOMAIN_IDENTITY:
670 type = "identity\n";
671 break;
672 case IOMMU_DOMAIN_UNMANAGED:
673 type = "unmanaged\n";
674 break;
675 case IOMMU_DOMAIN_DMA:
676 type = "DMA\n";
677 break;
678 case IOMMU_DOMAIN_DMA_FQ:
679 type = "DMA-FQ\n";
680 break;
681 }
682 }
683 mutex_unlock(&group->mutex);
684 strcpy(buf, type);
685
686 return strlen(type);
687}
688
689static IOMMU_GROUP_ATTR(name, S_IRUGO, iommu_group_show_name, NULL);
690
691static IOMMU_GROUP_ATTR(reserved_regions, 0444,
692 iommu_group_show_resv_regions, NULL);
693
694static IOMMU_GROUP_ATTR(type, 0644, iommu_group_show_type,
695 iommu_group_store_type);
696
697static void iommu_group_release(struct kobject *kobj)
698{
699 struct iommu_group *group = to_iommu_group(kobj);
700
701 pr_debug("Releasing group %d\n", group->id);
702
703 if (group->iommu_data_release)
704 group->iommu_data_release(group->iommu_data);
705
706 ida_free(&iommu_group_ida, group->id);
707
708 if (group->default_domain)
709 iommu_domain_free(group->default_domain);
710 if (group->blocking_domain)
711 iommu_domain_free(group->blocking_domain);
712
713 kfree(group->name);
714 kfree(group);
715}
716
717static struct kobj_type iommu_group_ktype = {
718 .sysfs_ops = &iommu_group_sysfs_ops,
719 .release = iommu_group_release,
720};
721
722/**
723 * iommu_group_alloc - Allocate a new group
724 *
725 * This function is called by an iommu driver to allocate a new iommu
726 * group. The iommu group represents the minimum granularity of the iommu.
727 * Upon successful return, the caller holds a reference to the supplied
728 * group in order to hold the group until devices are added. Use
729 * iommu_group_put() to release this extra reference count, allowing the
730 * group to be automatically reclaimed once it has no devices or external
731 * references.
732 */
733struct iommu_group *iommu_group_alloc(void)
734{
735 struct iommu_group *group;
736 int ret;
737
738 group = kzalloc(sizeof(*group), GFP_KERNEL);
739 if (!group)
740 return ERR_PTR(-ENOMEM);
741
742 group->kobj.kset = iommu_group_kset;
743 mutex_init(&group->mutex);
744 INIT_LIST_HEAD(&group->devices);
745 INIT_LIST_HEAD(&group->entry);
746 xa_init(&group->pasid_array);
747
748 ret = ida_alloc(&iommu_group_ida, GFP_KERNEL);
749 if (ret < 0) {
750 kfree(group);
751 return ERR_PTR(ret);
752 }
753 group->id = ret;
754
755 ret = kobject_init_and_add(&group->kobj, &iommu_group_ktype,
756 NULL, "%d", group->id);
757 if (ret) {
758 kobject_put(&group->kobj);
759 return ERR_PTR(ret);
760 }
761
762 group->devices_kobj = kobject_create_and_add("devices", &group->kobj);
763 if (!group->devices_kobj) {
764 kobject_put(&group->kobj); /* triggers .release & free */
765 return ERR_PTR(-ENOMEM);
766 }
767
768 /*
769 * The devices_kobj holds a reference on the group kobject, so
770 * as long as that exists so will the group. We can therefore
771 * use the devices_kobj for reference counting.
772 */
773 kobject_put(&group->kobj);
774
775 ret = iommu_group_create_file(group,
776 &iommu_group_attr_reserved_regions);
777 if (ret)
778 return ERR_PTR(ret);
779
780 ret = iommu_group_create_file(group, &iommu_group_attr_type);
781 if (ret)
782 return ERR_PTR(ret);
783
784 pr_debug("Allocated group %d\n", group->id);
785
786 return group;
787}
788EXPORT_SYMBOL_GPL(iommu_group_alloc);
789
790struct iommu_group *iommu_group_get_by_id(int id)
791{
792 struct kobject *group_kobj;
793 struct iommu_group *group;
794 const char *name;
795
796 if (!iommu_group_kset)
797 return NULL;
798
799 name = kasprintf(GFP_KERNEL, "%d", id);
800 if (!name)
801 return NULL;
802
803 group_kobj = kset_find_obj(iommu_group_kset, name);
804 kfree(name);
805
806 if (!group_kobj)
807 return NULL;
808
809 group = container_of(group_kobj, struct iommu_group, kobj);
810 BUG_ON(group->id != id);
811
812 kobject_get(group->devices_kobj);
813 kobject_put(&group->kobj);
814
815 return group;
816}
817EXPORT_SYMBOL_GPL(iommu_group_get_by_id);
818
819/**
820 * iommu_group_get_iommudata - retrieve iommu_data registered for a group
821 * @group: the group
822 *
823 * iommu drivers can store data in the group for use when doing iommu
824 * operations. This function provides a way to retrieve it. Caller
825 * should hold a group reference.
826 */
827void *iommu_group_get_iommudata(struct iommu_group *group)
828{
829 return group->iommu_data;
830}
831EXPORT_SYMBOL_GPL(iommu_group_get_iommudata);
832
833/**
834 * iommu_group_set_iommudata - set iommu_data for a group
835 * @group: the group
836 * @iommu_data: new data
837 * @release: release function for iommu_data
838 *
839 * iommu drivers can store data in the group for use when doing iommu
840 * operations. This function provides a way to set the data after
841 * the group has been allocated. Caller should hold a group reference.
842 */
843void iommu_group_set_iommudata(struct iommu_group *group, void *iommu_data,
844 void (*release)(void *iommu_data))
845{
846 group->iommu_data = iommu_data;
847 group->iommu_data_release = release;
848}
849EXPORT_SYMBOL_GPL(iommu_group_set_iommudata);
850
851/**
852 * iommu_group_set_name - set name for a group
853 * @group: the group
854 * @name: name
855 *
856 * Allow iommu driver to set a name for a group. When set it will
857 * appear in a name attribute file under the group in sysfs.
858 */
859int iommu_group_set_name(struct iommu_group *group, const char *name)
860{
861 int ret;
862
863 if (group->name) {
864 iommu_group_remove_file(group, &iommu_group_attr_name);
865 kfree(group->name);
866 group->name = NULL;
867 if (!name)
868 return 0;
869 }
870
871 group->name = kstrdup(name, GFP_KERNEL);
872 if (!group->name)
873 return -ENOMEM;
874
875 ret = iommu_group_create_file(group, &iommu_group_attr_name);
876 if (ret) {
877 kfree(group->name);
878 group->name = NULL;
879 return ret;
880 }
881
882 return 0;
883}
884EXPORT_SYMBOL_GPL(iommu_group_set_name);
885
886static int iommu_create_device_direct_mappings(struct iommu_group *group,
887 struct device *dev)
888{
889 struct iommu_domain *domain = group->default_domain;
890 struct iommu_resv_region *entry;
891 struct list_head mappings;
892 unsigned long pg_size;
893 int ret = 0;
894
895 if (!domain || !iommu_is_dma_domain(domain))
896 return 0;
897
898 BUG_ON(!domain->pgsize_bitmap);
899
900 pg_size = 1UL << __ffs(domain->pgsize_bitmap);
901 INIT_LIST_HEAD(&mappings);
902
903 iommu_get_resv_regions(dev, &mappings);
904
905 /* We need to consider overlapping regions for different devices */
906 list_for_each_entry(entry, &mappings, list) {
907 dma_addr_t start, end, addr;
908 size_t map_size = 0;
909
910 start = ALIGN(entry->start, pg_size);
911 end = ALIGN(entry->start + entry->length, pg_size);
912
913 if (entry->type != IOMMU_RESV_DIRECT &&
914 entry->type != IOMMU_RESV_DIRECT_RELAXABLE)
915 continue;
916
917 for (addr = start; addr <= end; addr += pg_size) {
918 phys_addr_t phys_addr;
919
920 if (addr == end)
921 goto map_end;
922
923 phys_addr = iommu_iova_to_phys(domain, addr);
924 if (!phys_addr) {
925 map_size += pg_size;
926 continue;
927 }
928
929map_end:
930 if (map_size) {
931 ret = iommu_map(domain, addr - map_size,
932 addr - map_size, map_size,
933 entry->prot);
934 if (ret)
935 goto out;
936 map_size = 0;
937 }
938 }
939
940 }
941
942 iommu_flush_iotlb_all(domain);
943
944out:
945 iommu_put_resv_regions(dev, &mappings);
946
947 return ret;
948}
949
950static bool iommu_is_attach_deferred(struct device *dev)
951{
952 const struct iommu_ops *ops = dev_iommu_ops(dev);
953
954 if (ops->is_attach_deferred)
955 return ops->is_attach_deferred(dev);
956
957 return false;
958}
959
960/**
961 * iommu_group_add_device - add a device to an iommu group
962 * @group: the group into which to add the device (reference should be held)
963 * @dev: the device
964 *
965 * This function is called by an iommu driver to add a device into a
966 * group. Adding a device increments the group reference count.
967 */
968int iommu_group_add_device(struct iommu_group *group, struct device *dev)
969{
970 int ret, i = 0;
971 struct group_device *device;
972
973 device = kzalloc(sizeof(*device), GFP_KERNEL);
974 if (!device)
975 return -ENOMEM;
976
977 device->dev = dev;
978
979 ret = sysfs_create_link(&dev->kobj, &group->kobj, "iommu_group");
980 if (ret)
981 goto err_free_device;
982
983 device->name = kasprintf(GFP_KERNEL, "%s", kobject_name(&dev->kobj));
984rename:
985 if (!device->name) {
986 ret = -ENOMEM;
987 goto err_remove_link;
988 }
989
990 ret = sysfs_create_link_nowarn(group->devices_kobj,
991 &dev->kobj, device->name);
992 if (ret) {
993 if (ret == -EEXIST && i >= 0) {
994 /*
995 * Account for the slim chance of collision
996 * and append an instance to the name.
997 */
998 kfree(device->name);
999 device->name = kasprintf(GFP_KERNEL, "%s.%d",
1000 kobject_name(&dev->kobj), i++);
1001 goto rename;
1002 }
1003 goto err_free_name;
1004 }
1005
1006 kobject_get(group->devices_kobj);
1007
1008 dev->iommu_group = group;
1009
1010 mutex_lock(&group->mutex);
1011 list_add_tail(&device->list, &group->devices);
1012 if (group->domain && !iommu_is_attach_deferred(dev))
1013 ret = __iommu_attach_device(group->domain, dev);
1014 mutex_unlock(&group->mutex);
1015 if (ret)
1016 goto err_put_group;
1017
1018 trace_add_device_to_group(group->id, dev);
1019
1020 dev_info(dev, "Adding to iommu group %d\n", group->id);
1021
1022 return 0;
1023
1024err_put_group:
1025 mutex_lock(&group->mutex);
1026 list_del(&device->list);
1027 mutex_unlock(&group->mutex);
1028 dev->iommu_group = NULL;
1029 kobject_put(group->devices_kobj);
1030 sysfs_remove_link(group->devices_kobj, device->name);
1031err_free_name:
1032 kfree(device->name);
1033err_remove_link:
1034 sysfs_remove_link(&dev->kobj, "iommu_group");
1035err_free_device:
1036 kfree(device);
1037 dev_err(dev, "Failed to add to iommu group %d: %d\n", group->id, ret);
1038 return ret;
1039}
1040EXPORT_SYMBOL_GPL(iommu_group_add_device);
1041
1042/**
1043 * iommu_group_remove_device - remove a device from it's current group
1044 * @dev: device to be removed
1045 *
1046 * This function is called by an iommu driver to remove the device from
1047 * it's current group. This decrements the iommu group reference count.
1048 */
1049void iommu_group_remove_device(struct device *dev)
1050{
1051 struct iommu_group *group = dev->iommu_group;
1052 struct group_device *tmp_device, *device = NULL;
1053
1054 if (!group)
1055 return;
1056
1057 dev_info(dev, "Removing from iommu group %d\n", group->id);
1058
1059 mutex_lock(&group->mutex);
1060 list_for_each_entry(tmp_device, &group->devices, list) {
1061 if (tmp_device->dev == dev) {
1062 device = tmp_device;
1063 list_del(&device->list);
1064 break;
1065 }
1066 }
1067 mutex_unlock(&group->mutex);
1068
1069 if (!device)
1070 return;
1071
1072 sysfs_remove_link(group->devices_kobj, device->name);
1073 sysfs_remove_link(&dev->kobj, "iommu_group");
1074
1075 trace_remove_device_from_group(group->id, dev);
1076
1077 kfree(device->name);
1078 kfree(device);
1079 dev->iommu_group = NULL;
1080 kobject_put(group->devices_kobj);
1081}
1082EXPORT_SYMBOL_GPL(iommu_group_remove_device);
1083
1084static int iommu_group_device_count(struct iommu_group *group)
1085{
1086 struct group_device *entry;
1087 int ret = 0;
1088
1089 list_for_each_entry(entry, &group->devices, list)
1090 ret++;
1091
1092 return ret;
1093}
1094
1095static int __iommu_group_for_each_dev(struct iommu_group *group, void *data,
1096 int (*fn)(struct device *, void *))
1097{
1098 struct group_device *device;
1099 int ret = 0;
1100
1101 list_for_each_entry(device, &group->devices, list) {
1102 ret = fn(device->dev, data);
1103 if (ret)
1104 break;
1105 }
1106 return ret;
1107}
1108
1109/**
1110 * iommu_group_for_each_dev - iterate over each device in the group
1111 * @group: the group
1112 * @data: caller opaque data to be passed to callback function
1113 * @fn: caller supplied callback function
1114 *
1115 * This function is called by group users to iterate over group devices.
1116 * Callers should hold a reference count to the group during callback.
1117 * The group->mutex is held across callbacks, which will block calls to
1118 * iommu_group_add/remove_device.
1119 */
1120int iommu_group_for_each_dev(struct iommu_group *group, void *data,
1121 int (*fn)(struct device *, void *))
1122{
1123 int ret;
1124
1125 mutex_lock(&group->mutex);
1126 ret = __iommu_group_for_each_dev(group, data, fn);
1127 mutex_unlock(&group->mutex);
1128
1129 return ret;
1130}
1131EXPORT_SYMBOL_GPL(iommu_group_for_each_dev);
1132
1133/**
1134 * iommu_group_get - Return the group for a device and increment reference
1135 * @dev: get the group that this device belongs to
1136 *
1137 * This function is called by iommu drivers and users to get the group
1138 * for the specified device. If found, the group is returned and the group
1139 * reference in incremented, else NULL.
1140 */
1141struct iommu_group *iommu_group_get(struct device *dev)
1142{
1143 struct iommu_group *group = dev->iommu_group;
1144
1145 if (group)
1146 kobject_get(group->devices_kobj);
1147
1148 return group;
1149}
1150EXPORT_SYMBOL_GPL(iommu_group_get);
1151
1152/**
1153 * iommu_group_ref_get - Increment reference on a group
1154 * @group: the group to use, must not be NULL
1155 *
1156 * This function is called by iommu drivers to take additional references on an
1157 * existing group. Returns the given group for convenience.
1158 */
1159struct iommu_group *iommu_group_ref_get(struct iommu_group *group)
1160{
1161 kobject_get(group->devices_kobj);
1162 return group;
1163}
1164EXPORT_SYMBOL_GPL(iommu_group_ref_get);
1165
1166/**
1167 * iommu_group_put - Decrement group reference
1168 * @group: the group to use
1169 *
1170 * This function is called by iommu drivers and users to release the
1171 * iommu group. Once the reference count is zero, the group is released.
1172 */
1173void iommu_group_put(struct iommu_group *group)
1174{
1175 if (group)
1176 kobject_put(group->devices_kobj);
1177}
1178EXPORT_SYMBOL_GPL(iommu_group_put);
1179
1180/**
1181 * iommu_register_device_fault_handler() - Register a device fault handler
1182 * @dev: the device
1183 * @handler: the fault handler
1184 * @data: private data passed as argument to the handler
1185 *
1186 * When an IOMMU fault event is received, this handler gets called with the
1187 * fault event and data as argument. The handler should return 0 on success. If
1188 * the fault is recoverable (IOMMU_FAULT_PAGE_REQ), the consumer should also
1189 * complete the fault by calling iommu_page_response() with one of the following
1190 * response code:
1191 * - IOMMU_PAGE_RESP_SUCCESS: retry the translation
1192 * - IOMMU_PAGE_RESP_INVALID: terminate the fault
1193 * - IOMMU_PAGE_RESP_FAILURE: terminate the fault and stop reporting
1194 * page faults if possible.
1195 *
1196 * Return 0 if the fault handler was installed successfully, or an error.
1197 */
1198int iommu_register_device_fault_handler(struct device *dev,
1199 iommu_dev_fault_handler_t handler,
1200 void *data)
1201{
1202 struct dev_iommu *param = dev->iommu;
1203 int ret = 0;
1204
1205 if (!param)
1206 return -EINVAL;
1207
1208 mutex_lock(¶m->lock);
1209 /* Only allow one fault handler registered for each device */
1210 if (param->fault_param) {
1211 ret = -EBUSY;
1212 goto done_unlock;
1213 }
1214
1215 get_device(dev);
1216 param->fault_param = kzalloc(sizeof(*param->fault_param), GFP_KERNEL);
1217 if (!param->fault_param) {
1218 put_device(dev);
1219 ret = -ENOMEM;
1220 goto done_unlock;
1221 }
1222 param->fault_param->handler = handler;
1223 param->fault_param->data = data;
1224 mutex_init(¶m->fault_param->lock);
1225 INIT_LIST_HEAD(¶m->fault_param->faults);
1226
1227done_unlock:
1228 mutex_unlock(¶m->lock);
1229
1230 return ret;
1231}
1232EXPORT_SYMBOL_GPL(iommu_register_device_fault_handler);
1233
1234/**
1235 * iommu_unregister_device_fault_handler() - Unregister the device fault handler
1236 * @dev: the device
1237 *
1238 * Remove the device fault handler installed with
1239 * iommu_register_device_fault_handler().
1240 *
1241 * Return 0 on success, or an error.
1242 */
1243int iommu_unregister_device_fault_handler(struct device *dev)
1244{
1245 struct dev_iommu *param = dev->iommu;
1246 int ret = 0;
1247
1248 if (!param)
1249 return -EINVAL;
1250
1251 mutex_lock(¶m->lock);
1252
1253 if (!param->fault_param)
1254 goto unlock;
1255
1256 /* we cannot unregister handler if there are pending faults */
1257 if (!list_empty(¶m->fault_param->faults)) {
1258 ret = -EBUSY;
1259 goto unlock;
1260 }
1261
1262 kfree(param->fault_param);
1263 param->fault_param = NULL;
1264 put_device(dev);
1265unlock:
1266 mutex_unlock(¶m->lock);
1267
1268 return ret;
1269}
1270EXPORT_SYMBOL_GPL(iommu_unregister_device_fault_handler);
1271
1272/**
1273 * iommu_report_device_fault() - Report fault event to device driver
1274 * @dev: the device
1275 * @evt: fault event data
1276 *
1277 * Called by IOMMU drivers when a fault is detected, typically in a threaded IRQ
1278 * handler. When this function fails and the fault is recoverable, it is the
1279 * caller's responsibility to complete the fault.
1280 *
1281 * Return 0 on success, or an error.
1282 */
1283int iommu_report_device_fault(struct device *dev, struct iommu_fault_event *evt)
1284{
1285 struct dev_iommu *param = dev->iommu;
1286 struct iommu_fault_event *evt_pending = NULL;
1287 struct iommu_fault_param *fparam;
1288 int ret = 0;
1289
1290 if (!param || !evt)
1291 return -EINVAL;
1292
1293 /* we only report device fault if there is a handler registered */
1294 mutex_lock(¶m->lock);
1295 fparam = param->fault_param;
1296 if (!fparam || !fparam->handler) {
1297 ret = -EINVAL;
1298 goto done_unlock;
1299 }
1300
1301 if (evt->fault.type == IOMMU_FAULT_PAGE_REQ &&
1302 (evt->fault.prm.flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE)) {
1303 evt_pending = kmemdup(evt, sizeof(struct iommu_fault_event),
1304 GFP_KERNEL);
1305 if (!evt_pending) {
1306 ret = -ENOMEM;
1307 goto done_unlock;
1308 }
1309 mutex_lock(&fparam->lock);
1310 list_add_tail(&evt_pending->list, &fparam->faults);
1311 mutex_unlock(&fparam->lock);
1312 }
1313
1314 ret = fparam->handler(&evt->fault, fparam->data);
1315 if (ret && evt_pending) {
1316 mutex_lock(&fparam->lock);
1317 list_del(&evt_pending->list);
1318 mutex_unlock(&fparam->lock);
1319 kfree(evt_pending);
1320 }
1321done_unlock:
1322 mutex_unlock(¶m->lock);
1323 return ret;
1324}
1325EXPORT_SYMBOL_GPL(iommu_report_device_fault);
1326
1327int iommu_page_response(struct device *dev,
1328 struct iommu_page_response *msg)
1329{
1330 bool needs_pasid;
1331 int ret = -EINVAL;
1332 struct iommu_fault_event *evt;
1333 struct iommu_fault_page_request *prm;
1334 struct dev_iommu *param = dev->iommu;
1335 const struct iommu_ops *ops = dev_iommu_ops(dev);
1336 bool has_pasid = msg->flags & IOMMU_PAGE_RESP_PASID_VALID;
1337
1338 if (!ops->page_response)
1339 return -ENODEV;
1340
1341 if (!param || !param->fault_param)
1342 return -EINVAL;
1343
1344 if (msg->version != IOMMU_PAGE_RESP_VERSION_1 ||
1345 msg->flags & ~IOMMU_PAGE_RESP_PASID_VALID)
1346 return -EINVAL;
1347
1348 /* Only send response if there is a fault report pending */
1349 mutex_lock(¶m->fault_param->lock);
1350 if (list_empty(¶m->fault_param->faults)) {
1351 dev_warn_ratelimited(dev, "no pending PRQ, drop response\n");
1352 goto done_unlock;
1353 }
1354 /*
1355 * Check if we have a matching page request pending to respond,
1356 * otherwise return -EINVAL
1357 */
1358 list_for_each_entry(evt, ¶m->fault_param->faults, list) {
1359 prm = &evt->fault.prm;
1360 if (prm->grpid != msg->grpid)
1361 continue;
1362
1363 /*
1364 * If the PASID is required, the corresponding request is
1365 * matched using the group ID, the PASID valid bit and the PASID
1366 * value. Otherwise only the group ID matches request and
1367 * response.
1368 */
1369 needs_pasid = prm->flags & IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
1370 if (needs_pasid && (!has_pasid || msg->pasid != prm->pasid))
1371 continue;
1372
1373 if (!needs_pasid && has_pasid) {
1374 /* No big deal, just clear it. */
1375 msg->flags &= ~IOMMU_PAGE_RESP_PASID_VALID;
1376 msg->pasid = 0;
1377 }
1378
1379 ret = ops->page_response(dev, evt, msg);
1380 list_del(&evt->list);
1381 kfree(evt);
1382 break;
1383 }
1384
1385done_unlock:
1386 mutex_unlock(¶m->fault_param->lock);
1387 return ret;
1388}
1389EXPORT_SYMBOL_GPL(iommu_page_response);
1390
1391/**
1392 * iommu_group_id - Return ID for a group
1393 * @group: the group to ID
1394 *
1395 * Return the unique ID for the group matching the sysfs group number.
1396 */
1397int iommu_group_id(struct iommu_group *group)
1398{
1399 return group->id;
1400}
1401EXPORT_SYMBOL_GPL(iommu_group_id);
1402
1403static struct iommu_group *get_pci_alias_group(struct pci_dev *pdev,
1404 unsigned long *devfns);
1405
1406/*
1407 * To consider a PCI device isolated, we require ACS to support Source
1408 * Validation, Request Redirection, Completer Redirection, and Upstream
1409 * Forwarding. This effectively means that devices cannot spoof their
1410 * requester ID, requests and completions cannot be redirected, and all
1411 * transactions are forwarded upstream, even as it passes through a
1412 * bridge where the target device is downstream.
1413 */
1414#define REQ_ACS_FLAGS (PCI_ACS_SV | PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_UF)
1415
1416/*
1417 * For multifunction devices which are not isolated from each other, find
1418 * all the other non-isolated functions and look for existing groups. For
1419 * each function, we also need to look for aliases to or from other devices
1420 * that may already have a group.
1421 */
1422static struct iommu_group *get_pci_function_alias_group(struct pci_dev *pdev,
1423 unsigned long *devfns)
1424{
1425 struct pci_dev *tmp = NULL;
1426 struct iommu_group *group;
1427
1428 if (!pdev->multifunction || pci_acs_enabled(pdev, REQ_ACS_FLAGS))
1429 return NULL;
1430
1431 for_each_pci_dev(tmp) {
1432 if (tmp == pdev || tmp->bus != pdev->bus ||
1433 PCI_SLOT(tmp->devfn) != PCI_SLOT(pdev->devfn) ||
1434 pci_acs_enabled(tmp, REQ_ACS_FLAGS))
1435 continue;
1436
1437 group = get_pci_alias_group(tmp, devfns);
1438 if (group) {
1439 pci_dev_put(tmp);
1440 return group;
1441 }
1442 }
1443
1444 return NULL;
1445}
1446
1447/*
1448 * Look for aliases to or from the given device for existing groups. DMA
1449 * aliases are only supported on the same bus, therefore the search
1450 * space is quite small (especially since we're really only looking at pcie
1451 * device, and therefore only expect multiple slots on the root complex or
1452 * downstream switch ports). It's conceivable though that a pair of
1453 * multifunction devices could have aliases between them that would cause a
1454 * loop. To prevent this, we use a bitmap to track where we've been.
1455 */
1456static struct iommu_group *get_pci_alias_group(struct pci_dev *pdev,
1457 unsigned long *devfns)
1458{
1459 struct pci_dev *tmp = NULL;
1460 struct iommu_group *group;
1461
1462 if (test_and_set_bit(pdev->devfn & 0xff, devfns))
1463 return NULL;
1464
1465 group = iommu_group_get(&pdev->dev);
1466 if (group)
1467 return group;
1468
1469 for_each_pci_dev(tmp) {
1470 if (tmp == pdev || tmp->bus != pdev->bus)
1471 continue;
1472
1473 /* We alias them or they alias us */
1474 if (pci_devs_are_dma_aliases(pdev, tmp)) {
1475 group = get_pci_alias_group(tmp, devfns);
1476 if (group) {
1477 pci_dev_put(tmp);
1478 return group;
1479 }
1480
1481 group = get_pci_function_alias_group(tmp, devfns);
1482 if (group) {
1483 pci_dev_put(tmp);
1484 return group;
1485 }
1486 }
1487 }
1488
1489 return NULL;
1490}
1491
1492struct group_for_pci_data {
1493 struct pci_dev *pdev;
1494 struct iommu_group *group;
1495};
1496
1497/*
1498 * DMA alias iterator callback, return the last seen device. Stop and return
1499 * the IOMMU group if we find one along the way.
1500 */
1501static int get_pci_alias_or_group(struct pci_dev *pdev, u16 alias, void *opaque)
1502{
1503 struct group_for_pci_data *data = opaque;
1504
1505 data->pdev = pdev;
1506 data->group = iommu_group_get(&pdev->dev);
1507
1508 return data->group != NULL;
1509}
1510
1511/*
1512 * Generic device_group call-back function. It just allocates one
1513 * iommu-group per device.
1514 */
1515struct iommu_group *generic_device_group(struct device *dev)
1516{
1517 return iommu_group_alloc();
1518}
1519EXPORT_SYMBOL_GPL(generic_device_group);
1520
1521/*
1522 * Use standard PCI bus topology, isolation features, and DMA alias quirks
1523 * to find or create an IOMMU group for a device.
1524 */
1525struct iommu_group *pci_device_group(struct device *dev)
1526{
1527 struct pci_dev *pdev = to_pci_dev(dev);
1528 struct group_for_pci_data data;
1529 struct pci_bus *bus;
1530 struct iommu_group *group = NULL;
1531 u64 devfns[4] = { 0 };
1532
1533 if (WARN_ON(!dev_is_pci(dev)))
1534 return ERR_PTR(-EINVAL);
1535
1536 /*
1537 * Find the upstream DMA alias for the device. A device must not
1538 * be aliased due to topology in order to have its own IOMMU group.
1539 * If we find an alias along the way that already belongs to a
1540 * group, use it.
1541 */
1542 if (pci_for_each_dma_alias(pdev, get_pci_alias_or_group, &data))
1543 return data.group;
1544
1545 pdev = data.pdev;
1546
1547 /*
1548 * Continue upstream from the point of minimum IOMMU granularity
1549 * due to aliases to the point where devices are protected from
1550 * peer-to-peer DMA by PCI ACS. Again, if we find an existing
1551 * group, use it.
1552 */
1553 for (bus = pdev->bus; !pci_is_root_bus(bus); bus = bus->parent) {
1554 if (!bus->self)
1555 continue;
1556
1557 if (pci_acs_path_enabled(bus->self, NULL, REQ_ACS_FLAGS))
1558 break;
1559
1560 pdev = bus->self;
1561
1562 group = iommu_group_get(&pdev->dev);
1563 if (group)
1564 return group;
1565 }
1566
1567 /*
1568 * Look for existing groups on device aliases. If we alias another
1569 * device or another device aliases us, use the same group.
1570 */
1571 group = get_pci_alias_group(pdev, (unsigned long *)devfns);
1572 if (group)
1573 return group;
1574
1575 /*
1576 * Look for existing groups on non-isolated functions on the same
1577 * slot and aliases of those funcions, if any. No need to clear
1578 * the search bitmap, the tested devfns are still valid.
1579 */
1580 group = get_pci_function_alias_group(pdev, (unsigned long *)devfns);
1581 if (group)
1582 return group;
1583
1584 /* No shared group found, allocate new */
1585 return iommu_group_alloc();
1586}
1587EXPORT_SYMBOL_GPL(pci_device_group);
1588
1589/* Get the IOMMU group for device on fsl-mc bus */
1590struct iommu_group *fsl_mc_device_group(struct device *dev)
1591{
1592 struct device *cont_dev = fsl_mc_cont_dev(dev);
1593 struct iommu_group *group;
1594
1595 group = iommu_group_get(cont_dev);
1596 if (!group)
1597 group = iommu_group_alloc();
1598 return group;
1599}
1600EXPORT_SYMBOL_GPL(fsl_mc_device_group);
1601
1602static int iommu_get_def_domain_type(struct device *dev)
1603{
1604 const struct iommu_ops *ops = dev_iommu_ops(dev);
1605
1606 if (dev_is_pci(dev) && to_pci_dev(dev)->untrusted)
1607 return IOMMU_DOMAIN_DMA;
1608
1609 if (ops->def_domain_type)
1610 return ops->def_domain_type(dev);
1611
1612 return 0;
1613}
1614
1615static int iommu_group_alloc_default_domain(struct bus_type *bus,
1616 struct iommu_group *group,
1617 unsigned int type)
1618{
1619 struct iommu_domain *dom;
1620
1621 dom = __iommu_domain_alloc(bus, type);
1622 if (!dom && type != IOMMU_DOMAIN_DMA) {
1623 dom = __iommu_domain_alloc(bus, IOMMU_DOMAIN_DMA);
1624 if (dom)
1625 pr_warn("Failed to allocate default IOMMU domain of type %u for group %s - Falling back to IOMMU_DOMAIN_DMA",
1626 type, group->name);
1627 }
1628
1629 if (!dom)
1630 return -ENOMEM;
1631
1632 group->default_domain = dom;
1633 if (!group->domain)
1634 group->domain = dom;
1635 return 0;
1636}
1637
1638static int iommu_alloc_default_domain(struct iommu_group *group,
1639 struct device *dev)
1640{
1641 unsigned int type;
1642
1643 if (group->default_domain)
1644 return 0;
1645
1646 type = iommu_get_def_domain_type(dev) ? : iommu_def_domain_type;
1647
1648 return iommu_group_alloc_default_domain(dev->bus, group, type);
1649}
1650
1651/**
1652 * iommu_group_get_for_dev - Find or create the IOMMU group for a device
1653 * @dev: target device
1654 *
1655 * This function is intended to be called by IOMMU drivers and extended to
1656 * support common, bus-defined algorithms when determining or creating the
1657 * IOMMU group for a device. On success, the caller will hold a reference
1658 * to the returned IOMMU group, which will already include the provided
1659 * device. The reference should be released with iommu_group_put().
1660 */
1661static struct iommu_group *iommu_group_get_for_dev(struct device *dev)
1662{
1663 const struct iommu_ops *ops = dev_iommu_ops(dev);
1664 struct iommu_group *group;
1665 int ret;
1666
1667 group = iommu_group_get(dev);
1668 if (group)
1669 return group;
1670
1671 group = ops->device_group(dev);
1672 if (WARN_ON_ONCE(group == NULL))
1673 return ERR_PTR(-EINVAL);
1674
1675 if (IS_ERR(group))
1676 return group;
1677
1678 ret = iommu_group_add_device(group, dev);
1679 if (ret)
1680 goto out_put_group;
1681
1682 return group;
1683
1684out_put_group:
1685 iommu_group_put(group);
1686
1687 return ERR_PTR(ret);
1688}
1689
1690struct iommu_domain *iommu_group_default_domain(struct iommu_group *group)
1691{
1692 return group->default_domain;
1693}
1694
1695static int probe_iommu_group(struct device *dev, void *data)
1696{
1697 struct list_head *group_list = data;
1698 struct iommu_group *group;
1699 int ret;
1700
1701 /* Device is probed already if in a group */
1702 group = iommu_group_get(dev);
1703 if (group) {
1704 iommu_group_put(group);
1705 return 0;
1706 }
1707
1708 ret = __iommu_probe_device(dev, group_list);
1709 if (ret == -ENODEV)
1710 ret = 0;
1711
1712 return ret;
1713}
1714
1715static int iommu_bus_notifier(struct notifier_block *nb,
1716 unsigned long action, void *data)
1717{
1718 struct device *dev = data;
1719
1720 if (action == BUS_NOTIFY_ADD_DEVICE) {
1721 int ret;
1722
1723 ret = iommu_probe_device(dev);
1724 return (ret) ? NOTIFY_DONE : NOTIFY_OK;
1725 } else if (action == BUS_NOTIFY_REMOVED_DEVICE) {
1726 iommu_release_device(dev);
1727 return NOTIFY_OK;
1728 }
1729
1730 return 0;
1731}
1732
1733struct __group_domain_type {
1734 struct device *dev;
1735 unsigned int type;
1736};
1737
1738static int probe_get_default_domain_type(struct device *dev, void *data)
1739{
1740 struct __group_domain_type *gtype = data;
1741 unsigned int type = iommu_get_def_domain_type(dev);
1742
1743 if (type) {
1744 if (gtype->type && gtype->type != type) {
1745 dev_warn(dev, "Device needs domain type %s, but device %s in the same iommu group requires type %s - using default\n",
1746 iommu_domain_type_str(type),
1747 dev_name(gtype->dev),
1748 iommu_domain_type_str(gtype->type));
1749 gtype->type = 0;
1750 }
1751
1752 if (!gtype->dev) {
1753 gtype->dev = dev;
1754 gtype->type = type;
1755 }
1756 }
1757
1758 return 0;
1759}
1760
1761static void probe_alloc_default_domain(struct bus_type *bus,
1762 struct iommu_group *group)
1763{
1764 struct __group_domain_type gtype;
1765
1766 memset(>ype, 0, sizeof(gtype));
1767
1768 /* Ask for default domain requirements of all devices in the group */
1769 __iommu_group_for_each_dev(group, >ype,
1770 probe_get_default_domain_type);
1771
1772 if (!gtype.type)
1773 gtype.type = iommu_def_domain_type;
1774
1775 iommu_group_alloc_default_domain(bus, group, gtype.type);
1776
1777}
1778
1779static int iommu_group_do_dma_attach(struct device *dev, void *data)
1780{
1781 struct iommu_domain *domain = data;
1782 int ret = 0;
1783
1784 if (!iommu_is_attach_deferred(dev))
1785 ret = __iommu_attach_device(domain, dev);
1786
1787 return ret;
1788}
1789
1790static int __iommu_group_dma_attach(struct iommu_group *group)
1791{
1792 return __iommu_group_for_each_dev(group, group->default_domain,
1793 iommu_group_do_dma_attach);
1794}
1795
1796static int iommu_group_do_probe_finalize(struct device *dev, void *data)
1797{
1798 const struct iommu_ops *ops = dev_iommu_ops(dev);
1799
1800 if (ops->probe_finalize)
1801 ops->probe_finalize(dev);
1802
1803 return 0;
1804}
1805
1806static void __iommu_group_dma_finalize(struct iommu_group *group)
1807{
1808 __iommu_group_for_each_dev(group, group->default_domain,
1809 iommu_group_do_probe_finalize);
1810}
1811
1812static int iommu_do_create_direct_mappings(struct device *dev, void *data)
1813{
1814 struct iommu_group *group = data;
1815
1816 iommu_create_device_direct_mappings(group, dev);
1817
1818 return 0;
1819}
1820
1821static int iommu_group_create_direct_mappings(struct iommu_group *group)
1822{
1823 return __iommu_group_for_each_dev(group, group,
1824 iommu_do_create_direct_mappings);
1825}
1826
1827int bus_iommu_probe(struct bus_type *bus)
1828{
1829 struct iommu_group *group, *next;
1830 LIST_HEAD(group_list);
1831 int ret;
1832
1833 /*
1834 * This code-path does not allocate the default domain when
1835 * creating the iommu group, so do it after the groups are
1836 * created.
1837 */
1838 ret = bus_for_each_dev(bus, NULL, &group_list, probe_iommu_group);
1839 if (ret)
1840 return ret;
1841
1842 list_for_each_entry_safe(group, next, &group_list, entry) {
1843 mutex_lock(&group->mutex);
1844
1845 /* Remove item from the list */
1846 list_del_init(&group->entry);
1847
1848 /* Try to allocate default domain */
1849 probe_alloc_default_domain(bus, group);
1850
1851 if (!group->default_domain) {
1852 mutex_unlock(&group->mutex);
1853 continue;
1854 }
1855
1856 iommu_group_create_direct_mappings(group);
1857
1858 ret = __iommu_group_dma_attach(group);
1859
1860 mutex_unlock(&group->mutex);
1861
1862 if (ret)
1863 break;
1864
1865 __iommu_group_dma_finalize(group);
1866 }
1867
1868 return ret;
1869}
1870
1871bool iommu_present(struct bus_type *bus)
1872{
1873 return bus->iommu_ops != NULL;
1874}
1875EXPORT_SYMBOL_GPL(iommu_present);
1876
1877/**
1878 * device_iommu_capable() - check for a general IOMMU capability
1879 * @dev: device to which the capability would be relevant, if available
1880 * @cap: IOMMU capability
1881 *
1882 * Return: true if an IOMMU is present and supports the given capability
1883 * for the given device, otherwise false.
1884 */
1885bool device_iommu_capable(struct device *dev, enum iommu_cap cap)
1886{
1887 const struct iommu_ops *ops;
1888
1889 if (!dev->iommu || !dev->iommu->iommu_dev)
1890 return false;
1891
1892 ops = dev_iommu_ops(dev);
1893 if (!ops->capable)
1894 return false;
1895
1896 return ops->capable(dev, cap);
1897}
1898EXPORT_SYMBOL_GPL(device_iommu_capable);
1899
1900/**
1901 * iommu_set_fault_handler() - set a fault handler for an iommu domain
1902 * @domain: iommu domain
1903 * @handler: fault handler
1904 * @token: user data, will be passed back to the fault handler
1905 *
1906 * This function should be used by IOMMU users which want to be notified
1907 * whenever an IOMMU fault happens.
1908 *
1909 * The fault handler itself should return 0 on success, and an appropriate
1910 * error code otherwise.
1911 */
1912void iommu_set_fault_handler(struct iommu_domain *domain,
1913 iommu_fault_handler_t handler,
1914 void *token)
1915{
1916 BUG_ON(!domain);
1917
1918 domain->handler = handler;
1919 domain->handler_token = token;
1920}
1921EXPORT_SYMBOL_GPL(iommu_set_fault_handler);
1922
1923static struct iommu_domain *__iommu_domain_alloc(struct bus_type *bus,
1924 unsigned type)
1925{
1926 struct iommu_domain *domain;
1927
1928 if (bus == NULL || bus->iommu_ops == NULL)
1929 return NULL;
1930
1931 domain = bus->iommu_ops->domain_alloc(type);
1932 if (!domain)
1933 return NULL;
1934
1935 domain->type = type;
1936 /* Assume all sizes by default; the driver may override this later */
1937 domain->pgsize_bitmap = bus->iommu_ops->pgsize_bitmap;
1938 if (!domain->ops)
1939 domain->ops = bus->iommu_ops->default_domain_ops;
1940
1941 if (iommu_is_dma_domain(domain) && iommu_get_dma_cookie(domain)) {
1942 iommu_domain_free(domain);
1943 domain = NULL;
1944 }
1945 return domain;
1946}
1947
1948struct iommu_domain *iommu_domain_alloc(struct bus_type *bus)
1949{
1950 return __iommu_domain_alloc(bus, IOMMU_DOMAIN_UNMANAGED);
1951}
1952EXPORT_SYMBOL_GPL(iommu_domain_alloc);
1953
1954void iommu_domain_free(struct iommu_domain *domain)
1955{
1956 if (domain->type == IOMMU_DOMAIN_SVA)
1957 mmdrop(domain->mm);
1958 iommu_put_dma_cookie(domain);
1959 domain->ops->free(domain);
1960}
1961EXPORT_SYMBOL_GPL(iommu_domain_free);
1962
1963/*
1964 * Put the group's domain back to the appropriate core-owned domain - either the
1965 * standard kernel-mode DMA configuration or an all-DMA-blocked domain.
1966 */
1967static void __iommu_group_set_core_domain(struct iommu_group *group)
1968{
1969 struct iommu_domain *new_domain;
1970 int ret;
1971
1972 if (group->owner)
1973 new_domain = group->blocking_domain;
1974 else
1975 new_domain = group->default_domain;
1976
1977 ret = __iommu_group_set_domain(group, new_domain);
1978 WARN(ret, "iommu driver failed to attach the default/blocking domain");
1979}
1980
1981static int __iommu_attach_device(struct iommu_domain *domain,
1982 struct device *dev)
1983{
1984 int ret;
1985
1986 if (unlikely(domain->ops->attach_dev == NULL))
1987 return -ENODEV;
1988
1989 ret = domain->ops->attach_dev(domain, dev);
1990 if (!ret)
1991 trace_attach_device_to_domain(dev);
1992 return ret;
1993}
1994
1995/**
1996 * iommu_attach_device - Attach an IOMMU domain to a device
1997 * @domain: IOMMU domain to attach
1998 * @dev: Device that will be attached
1999 *
2000 * Returns 0 on success and error code on failure
2001 *
2002 * Note that EINVAL can be treated as a soft failure, indicating
2003 * that certain configuration of the domain is incompatible with
2004 * the device. In this case attaching a different domain to the
2005 * device may succeed.
2006 */
2007int iommu_attach_device(struct iommu_domain *domain, struct device *dev)
2008{
2009 struct iommu_group *group;
2010 int ret;
2011
2012 group = iommu_group_get(dev);
2013 if (!group)
2014 return -ENODEV;
2015
2016 /*
2017 * Lock the group to make sure the device-count doesn't
2018 * change while we are attaching
2019 */
2020 mutex_lock(&group->mutex);
2021 ret = -EINVAL;
2022 if (iommu_group_device_count(group) != 1)
2023 goto out_unlock;
2024
2025 ret = __iommu_attach_group(domain, group);
2026
2027out_unlock:
2028 mutex_unlock(&group->mutex);
2029 iommu_group_put(group);
2030
2031 return ret;
2032}
2033EXPORT_SYMBOL_GPL(iommu_attach_device);
2034
2035int iommu_deferred_attach(struct device *dev, struct iommu_domain *domain)
2036{
2037 if (iommu_is_attach_deferred(dev))
2038 return __iommu_attach_device(domain, dev);
2039
2040 return 0;
2041}
2042
2043static void __iommu_detach_device(struct iommu_domain *domain,
2044 struct device *dev)
2045{
2046 if (iommu_is_attach_deferred(dev))
2047 return;
2048
2049 domain->ops->detach_dev(domain, dev);
2050 trace_detach_device_from_domain(dev);
2051}
2052
2053void iommu_detach_device(struct iommu_domain *domain, struct device *dev)
2054{
2055 struct iommu_group *group;
2056
2057 group = iommu_group_get(dev);
2058 if (!group)
2059 return;
2060
2061 mutex_lock(&group->mutex);
2062 if (WARN_ON(domain != group->domain) ||
2063 WARN_ON(iommu_group_device_count(group) != 1))
2064 goto out_unlock;
2065 __iommu_group_set_core_domain(group);
2066
2067out_unlock:
2068 mutex_unlock(&group->mutex);
2069 iommu_group_put(group);
2070}
2071EXPORT_SYMBOL_GPL(iommu_detach_device);
2072
2073struct iommu_domain *iommu_get_domain_for_dev(struct device *dev)
2074{
2075 struct iommu_domain *domain;
2076 struct iommu_group *group;
2077
2078 group = iommu_group_get(dev);
2079 if (!group)
2080 return NULL;
2081
2082 domain = group->domain;
2083
2084 iommu_group_put(group);
2085
2086 return domain;
2087}
2088EXPORT_SYMBOL_GPL(iommu_get_domain_for_dev);
2089
2090/*
2091 * For IOMMU_DOMAIN_DMA implementations which already provide their own
2092 * guarantees that the group and its default domain are valid and correct.
2093 */
2094struct iommu_domain *iommu_get_dma_domain(struct device *dev)
2095{
2096 return dev->iommu_group->default_domain;
2097}
2098
2099/*
2100 * IOMMU groups are really the natural working unit of the IOMMU, but
2101 * the IOMMU API works on domains and devices. Bridge that gap by
2102 * iterating over the devices in a group. Ideally we'd have a single
2103 * device which represents the requestor ID of the group, but we also
2104 * allow IOMMU drivers to create policy defined minimum sets, where
2105 * the physical hardware may be able to distiguish members, but we
2106 * wish to group them at a higher level (ex. untrusted multi-function
2107 * PCI devices). Thus we attach each device.
2108 */
2109static int iommu_group_do_attach_device(struct device *dev, void *data)
2110{
2111 struct iommu_domain *domain = data;
2112
2113 return __iommu_attach_device(domain, dev);
2114}
2115
2116static int __iommu_attach_group(struct iommu_domain *domain,
2117 struct iommu_group *group)
2118{
2119 int ret;
2120
2121 if (group->domain && group->domain != group->default_domain &&
2122 group->domain != group->blocking_domain)
2123 return -EBUSY;
2124
2125 ret = __iommu_group_for_each_dev(group, domain,
2126 iommu_group_do_attach_device);
2127 if (ret == 0)
2128 group->domain = domain;
2129
2130 return ret;
2131}
2132
2133/**
2134 * iommu_attach_group - Attach an IOMMU domain to an IOMMU group
2135 * @domain: IOMMU domain to attach
2136 * @group: IOMMU group that will be attached
2137 *
2138 * Returns 0 on success and error code on failure
2139 *
2140 * Note that EINVAL can be treated as a soft failure, indicating
2141 * that certain configuration of the domain is incompatible with
2142 * the group. In this case attaching a different domain to the
2143 * group may succeed.
2144 */
2145int iommu_attach_group(struct iommu_domain *domain, struct iommu_group *group)
2146{
2147 int ret;
2148
2149 mutex_lock(&group->mutex);
2150 ret = __iommu_attach_group(domain, group);
2151 mutex_unlock(&group->mutex);
2152
2153 return ret;
2154}
2155EXPORT_SYMBOL_GPL(iommu_attach_group);
2156
2157static int iommu_group_do_detach_device(struct device *dev, void *data)
2158{
2159 struct iommu_domain *domain = data;
2160
2161 __iommu_detach_device(domain, dev);
2162
2163 return 0;
2164}
2165
2166static int __iommu_group_set_domain(struct iommu_group *group,
2167 struct iommu_domain *new_domain)
2168{
2169 int ret;
2170
2171 if (group->domain == new_domain)
2172 return 0;
2173
2174 /*
2175 * New drivers should support default domains and so the detach_dev() op
2176 * will never be called. Otherwise the NULL domain represents some
2177 * platform specific behavior.
2178 */
2179 if (!new_domain) {
2180 if (WARN_ON(!group->domain->ops->detach_dev))
2181 return -EINVAL;
2182 __iommu_group_for_each_dev(group, group->domain,
2183 iommu_group_do_detach_device);
2184 group->domain = NULL;
2185 return 0;
2186 }
2187
2188 /*
2189 * Changing the domain is done by calling attach_dev() on the new
2190 * domain. This switch does not have to be atomic and DMA can be
2191 * discarded during the transition. DMA must only be able to access
2192 * either new_domain or group->domain, never something else.
2193 *
2194 * Note that this is called in error unwind paths, attaching to a
2195 * domain that has already been attached cannot fail.
2196 */
2197 ret = __iommu_group_for_each_dev(group, new_domain,
2198 iommu_group_do_attach_device);
2199 if (ret)
2200 return ret;
2201 group->domain = new_domain;
2202 return 0;
2203}
2204
2205void iommu_detach_group(struct iommu_domain *domain, struct iommu_group *group)
2206{
2207 mutex_lock(&group->mutex);
2208 __iommu_group_set_core_domain(group);
2209 mutex_unlock(&group->mutex);
2210}
2211EXPORT_SYMBOL_GPL(iommu_detach_group);
2212
2213phys_addr_t iommu_iova_to_phys(struct iommu_domain *domain, dma_addr_t iova)
2214{
2215 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2216 return iova;
2217
2218 if (domain->type == IOMMU_DOMAIN_BLOCKED)
2219 return 0;
2220
2221 return domain->ops->iova_to_phys(domain, iova);
2222}
2223EXPORT_SYMBOL_GPL(iommu_iova_to_phys);
2224
2225static size_t iommu_pgsize(struct iommu_domain *domain, unsigned long iova,
2226 phys_addr_t paddr, size_t size, size_t *count)
2227{
2228 unsigned int pgsize_idx, pgsize_idx_next;
2229 unsigned long pgsizes;
2230 size_t offset, pgsize, pgsize_next;
2231 unsigned long addr_merge = paddr | iova;
2232
2233 /* Page sizes supported by the hardware and small enough for @size */
2234 pgsizes = domain->pgsize_bitmap & GENMASK(__fls(size), 0);
2235
2236 /* Constrain the page sizes further based on the maximum alignment */
2237 if (likely(addr_merge))
2238 pgsizes &= GENMASK(__ffs(addr_merge), 0);
2239
2240 /* Make sure we have at least one suitable page size */
2241 BUG_ON(!pgsizes);
2242
2243 /* Pick the biggest page size remaining */
2244 pgsize_idx = __fls(pgsizes);
2245 pgsize = BIT(pgsize_idx);
2246 if (!count)
2247 return pgsize;
2248
2249 /* Find the next biggest support page size, if it exists */
2250 pgsizes = domain->pgsize_bitmap & ~GENMASK(pgsize_idx, 0);
2251 if (!pgsizes)
2252 goto out_set_count;
2253
2254 pgsize_idx_next = __ffs(pgsizes);
2255 pgsize_next = BIT(pgsize_idx_next);
2256
2257 /*
2258 * There's no point trying a bigger page size unless the virtual
2259 * and physical addresses are similarly offset within the larger page.
2260 */
2261 if ((iova ^ paddr) & (pgsize_next - 1))
2262 goto out_set_count;
2263
2264 /* Calculate the offset to the next page size alignment boundary */
2265 offset = pgsize_next - (addr_merge & (pgsize_next - 1));
2266
2267 /*
2268 * If size is big enough to accommodate the larger page, reduce
2269 * the number of smaller pages.
2270 */
2271 if (offset + pgsize_next <= size)
2272 size = offset;
2273
2274out_set_count:
2275 *count = size >> pgsize_idx;
2276 return pgsize;
2277}
2278
2279static int __iommu_map_pages(struct iommu_domain *domain, unsigned long iova,
2280 phys_addr_t paddr, size_t size, int prot,
2281 gfp_t gfp, size_t *mapped)
2282{
2283 const struct iommu_domain_ops *ops = domain->ops;
2284 size_t pgsize, count;
2285 int ret;
2286
2287 pgsize = iommu_pgsize(domain, iova, paddr, size, &count);
2288
2289 pr_debug("mapping: iova 0x%lx pa %pa pgsize 0x%zx count %zu\n",
2290 iova, &paddr, pgsize, count);
2291
2292 if (ops->map_pages) {
2293 ret = ops->map_pages(domain, iova, paddr, pgsize, count, prot,
2294 gfp, mapped);
2295 } else {
2296 ret = ops->map(domain, iova, paddr, pgsize, prot, gfp);
2297 *mapped = ret ? 0 : pgsize;
2298 }
2299
2300 return ret;
2301}
2302
2303static int __iommu_map(struct iommu_domain *domain, unsigned long iova,
2304 phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
2305{
2306 const struct iommu_domain_ops *ops = domain->ops;
2307 unsigned long orig_iova = iova;
2308 unsigned int min_pagesz;
2309 size_t orig_size = size;
2310 phys_addr_t orig_paddr = paddr;
2311 int ret = 0;
2312
2313 if (unlikely(!(ops->map || ops->map_pages) ||
2314 domain->pgsize_bitmap == 0UL))
2315 return -ENODEV;
2316
2317 if (unlikely(!(domain->type & __IOMMU_DOMAIN_PAGING)))
2318 return -EINVAL;
2319
2320 /* find out the minimum page size supported */
2321 min_pagesz = 1 << __ffs(domain->pgsize_bitmap);
2322
2323 /*
2324 * both the virtual address and the physical one, as well as
2325 * the size of the mapping, must be aligned (at least) to the
2326 * size of the smallest page supported by the hardware
2327 */
2328 if (!IS_ALIGNED(iova | paddr | size, min_pagesz)) {
2329 pr_err("unaligned: iova 0x%lx pa %pa size 0x%zx min_pagesz 0x%x\n",
2330 iova, &paddr, size, min_pagesz);
2331 return -EINVAL;
2332 }
2333
2334 pr_debug("map: iova 0x%lx pa %pa size 0x%zx\n", iova, &paddr, size);
2335
2336 while (size) {
2337 size_t mapped = 0;
2338
2339 ret = __iommu_map_pages(domain, iova, paddr, size, prot, gfp,
2340 &mapped);
2341 /*
2342 * Some pages may have been mapped, even if an error occurred,
2343 * so we should account for those so they can be unmapped.
2344 */
2345 size -= mapped;
2346
2347 if (ret)
2348 break;
2349
2350 iova += mapped;
2351 paddr += mapped;
2352 }
2353
2354 /* unroll mapping in case something went wrong */
2355 if (ret)
2356 iommu_unmap(domain, orig_iova, orig_size - size);
2357 else
2358 trace_map(orig_iova, orig_paddr, orig_size);
2359
2360 return ret;
2361}
2362
2363static int _iommu_map(struct iommu_domain *domain, unsigned long iova,
2364 phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
2365{
2366 const struct iommu_domain_ops *ops = domain->ops;
2367 int ret;
2368
2369 ret = __iommu_map(domain, iova, paddr, size, prot, gfp);
2370 if (ret == 0 && ops->iotlb_sync_map)
2371 ops->iotlb_sync_map(domain, iova, size);
2372
2373 return ret;
2374}
2375
2376int iommu_map(struct iommu_domain *domain, unsigned long iova,
2377 phys_addr_t paddr, size_t size, int prot)
2378{
2379 might_sleep();
2380 return _iommu_map(domain, iova, paddr, size, prot, GFP_KERNEL);
2381}
2382EXPORT_SYMBOL_GPL(iommu_map);
2383
2384int iommu_map_atomic(struct iommu_domain *domain, unsigned long iova,
2385 phys_addr_t paddr, size_t size, int prot)
2386{
2387 return _iommu_map(domain, iova, paddr, size, prot, GFP_ATOMIC);
2388}
2389EXPORT_SYMBOL_GPL(iommu_map_atomic);
2390
2391static size_t __iommu_unmap_pages(struct iommu_domain *domain,
2392 unsigned long iova, size_t size,
2393 struct iommu_iotlb_gather *iotlb_gather)
2394{
2395 const struct iommu_domain_ops *ops = domain->ops;
2396 size_t pgsize, count;
2397
2398 pgsize = iommu_pgsize(domain, iova, iova, size, &count);
2399 return ops->unmap_pages ?
2400 ops->unmap_pages(domain, iova, pgsize, count, iotlb_gather) :
2401 ops->unmap(domain, iova, pgsize, iotlb_gather);
2402}
2403
2404static size_t __iommu_unmap(struct iommu_domain *domain,
2405 unsigned long iova, size_t size,
2406 struct iommu_iotlb_gather *iotlb_gather)
2407{
2408 const struct iommu_domain_ops *ops = domain->ops;
2409 size_t unmapped_page, unmapped = 0;
2410 unsigned long orig_iova = iova;
2411 unsigned int min_pagesz;
2412
2413 if (unlikely(!(ops->unmap || ops->unmap_pages) ||
2414 domain->pgsize_bitmap == 0UL))
2415 return 0;
2416
2417 if (unlikely(!(domain->type & __IOMMU_DOMAIN_PAGING)))
2418 return 0;
2419
2420 /* find out the minimum page size supported */
2421 min_pagesz = 1 << __ffs(domain->pgsize_bitmap);
2422
2423 /*
2424 * The virtual address, as well as the size of the mapping, must be
2425 * aligned (at least) to the size of the smallest page supported
2426 * by the hardware
2427 */
2428 if (!IS_ALIGNED(iova | size, min_pagesz)) {
2429 pr_err("unaligned: iova 0x%lx size 0x%zx min_pagesz 0x%x\n",
2430 iova, size, min_pagesz);
2431 return 0;
2432 }
2433
2434 pr_debug("unmap this: iova 0x%lx size 0x%zx\n", iova, size);
2435
2436 /*
2437 * Keep iterating until we either unmap 'size' bytes (or more)
2438 * or we hit an area that isn't mapped.
2439 */
2440 while (unmapped < size) {
2441 unmapped_page = __iommu_unmap_pages(domain, iova,
2442 size - unmapped,
2443 iotlb_gather);
2444 if (!unmapped_page)
2445 break;
2446
2447 pr_debug("unmapped: iova 0x%lx size 0x%zx\n",
2448 iova, unmapped_page);
2449
2450 iova += unmapped_page;
2451 unmapped += unmapped_page;
2452 }
2453
2454 trace_unmap(orig_iova, size, unmapped);
2455 return unmapped;
2456}
2457
2458size_t iommu_unmap(struct iommu_domain *domain,
2459 unsigned long iova, size_t size)
2460{
2461 struct iommu_iotlb_gather iotlb_gather;
2462 size_t ret;
2463
2464 iommu_iotlb_gather_init(&iotlb_gather);
2465 ret = __iommu_unmap(domain, iova, size, &iotlb_gather);
2466 iommu_iotlb_sync(domain, &iotlb_gather);
2467
2468 return ret;
2469}
2470EXPORT_SYMBOL_GPL(iommu_unmap);
2471
2472size_t iommu_unmap_fast(struct iommu_domain *domain,
2473 unsigned long iova, size_t size,
2474 struct iommu_iotlb_gather *iotlb_gather)
2475{
2476 return __iommu_unmap(domain, iova, size, iotlb_gather);
2477}
2478EXPORT_SYMBOL_GPL(iommu_unmap_fast);
2479
2480static ssize_t __iommu_map_sg(struct iommu_domain *domain, unsigned long iova,
2481 struct scatterlist *sg, unsigned int nents, int prot,
2482 gfp_t gfp)
2483{
2484 const struct iommu_domain_ops *ops = domain->ops;
2485 size_t len = 0, mapped = 0;
2486 phys_addr_t start;
2487 unsigned int i = 0;
2488 int ret;
2489
2490 while (i <= nents) {
2491 phys_addr_t s_phys = sg_phys(sg);
2492
2493 if (len && s_phys != start + len) {
2494 ret = __iommu_map(domain, iova + mapped, start,
2495 len, prot, gfp);
2496
2497 if (ret)
2498 goto out_err;
2499
2500 mapped += len;
2501 len = 0;
2502 }
2503
2504 if (sg_is_dma_bus_address(sg))
2505 goto next;
2506
2507 if (len) {
2508 len += sg->length;
2509 } else {
2510 len = sg->length;
2511 start = s_phys;
2512 }
2513
2514next:
2515 if (++i < nents)
2516 sg = sg_next(sg);
2517 }
2518
2519 if (ops->iotlb_sync_map)
2520 ops->iotlb_sync_map(domain, iova, mapped);
2521 return mapped;
2522
2523out_err:
2524 /* undo mappings already done */
2525 iommu_unmap(domain, iova, mapped);
2526
2527 return ret;
2528}
2529
2530ssize_t iommu_map_sg(struct iommu_domain *domain, unsigned long iova,
2531 struct scatterlist *sg, unsigned int nents, int prot)
2532{
2533 might_sleep();
2534 return __iommu_map_sg(domain, iova, sg, nents, prot, GFP_KERNEL);
2535}
2536EXPORT_SYMBOL_GPL(iommu_map_sg);
2537
2538ssize_t iommu_map_sg_atomic(struct iommu_domain *domain, unsigned long iova,
2539 struct scatterlist *sg, unsigned int nents, int prot)
2540{
2541 return __iommu_map_sg(domain, iova, sg, nents, prot, GFP_ATOMIC);
2542}
2543
2544/**
2545 * report_iommu_fault() - report about an IOMMU fault to the IOMMU framework
2546 * @domain: the iommu domain where the fault has happened
2547 * @dev: the device where the fault has happened
2548 * @iova: the faulting address
2549 * @flags: mmu fault flags (e.g. IOMMU_FAULT_READ/IOMMU_FAULT_WRITE/...)
2550 *
2551 * This function should be called by the low-level IOMMU implementations
2552 * whenever IOMMU faults happen, to allow high-level users, that are
2553 * interested in such events, to know about them.
2554 *
2555 * This event may be useful for several possible use cases:
2556 * - mere logging of the event
2557 * - dynamic TLB/PTE loading
2558 * - if restarting of the faulting device is required
2559 *
2560 * Returns 0 on success and an appropriate error code otherwise (if dynamic
2561 * PTE/TLB loading will one day be supported, implementations will be able
2562 * to tell whether it succeeded or not according to this return value).
2563 *
2564 * Specifically, -ENOSYS is returned if a fault handler isn't installed
2565 * (though fault handlers can also return -ENOSYS, in case they want to
2566 * elicit the default behavior of the IOMMU drivers).
2567 */
2568int report_iommu_fault(struct iommu_domain *domain, struct device *dev,
2569 unsigned long iova, int flags)
2570{
2571 int ret = -ENOSYS;
2572
2573 /*
2574 * if upper layers showed interest and installed a fault handler,
2575 * invoke it.
2576 */
2577 if (domain->handler)
2578 ret = domain->handler(domain, dev, iova, flags,
2579 domain->handler_token);
2580
2581 trace_io_page_fault(dev, iova, flags);
2582 return ret;
2583}
2584EXPORT_SYMBOL_GPL(report_iommu_fault);
2585
2586static int __init iommu_init(void)
2587{
2588 iommu_group_kset = kset_create_and_add("iommu_groups",
2589 NULL, kernel_kobj);
2590 BUG_ON(!iommu_group_kset);
2591
2592 iommu_debugfs_setup();
2593
2594 return 0;
2595}
2596core_initcall(iommu_init);
2597
2598int iommu_enable_nesting(struct iommu_domain *domain)
2599{
2600 if (domain->type != IOMMU_DOMAIN_UNMANAGED)
2601 return -EINVAL;
2602 if (!domain->ops->enable_nesting)
2603 return -EINVAL;
2604 return domain->ops->enable_nesting(domain);
2605}
2606EXPORT_SYMBOL_GPL(iommu_enable_nesting);
2607
2608int iommu_set_pgtable_quirks(struct iommu_domain *domain,
2609 unsigned long quirk)
2610{
2611 if (domain->type != IOMMU_DOMAIN_UNMANAGED)
2612 return -EINVAL;
2613 if (!domain->ops->set_pgtable_quirks)
2614 return -EINVAL;
2615 return domain->ops->set_pgtable_quirks(domain, quirk);
2616}
2617EXPORT_SYMBOL_GPL(iommu_set_pgtable_quirks);
2618
2619void iommu_get_resv_regions(struct device *dev, struct list_head *list)
2620{
2621 const struct iommu_ops *ops = dev_iommu_ops(dev);
2622
2623 if (ops->get_resv_regions)
2624 ops->get_resv_regions(dev, list);
2625}
2626
2627/**
2628 * iommu_put_resv_regions - release resered regions
2629 * @dev: device for which to free reserved regions
2630 * @list: reserved region list for device
2631 *
2632 * This releases a reserved region list acquired by iommu_get_resv_regions().
2633 */
2634void iommu_put_resv_regions(struct device *dev, struct list_head *list)
2635{
2636 struct iommu_resv_region *entry, *next;
2637
2638 list_for_each_entry_safe(entry, next, list, list) {
2639 if (entry->free)
2640 entry->free(dev, entry);
2641 else
2642 kfree(entry);
2643 }
2644}
2645EXPORT_SYMBOL(iommu_put_resv_regions);
2646
2647struct iommu_resv_region *iommu_alloc_resv_region(phys_addr_t start,
2648 size_t length, int prot,
2649 enum iommu_resv_type type,
2650 gfp_t gfp)
2651{
2652 struct iommu_resv_region *region;
2653
2654 region = kzalloc(sizeof(*region), gfp);
2655 if (!region)
2656 return NULL;
2657
2658 INIT_LIST_HEAD(®ion->list);
2659 region->start = start;
2660 region->length = length;
2661 region->prot = prot;
2662 region->type = type;
2663 return region;
2664}
2665EXPORT_SYMBOL_GPL(iommu_alloc_resv_region);
2666
2667void iommu_set_default_passthrough(bool cmd_line)
2668{
2669 if (cmd_line)
2670 iommu_cmd_line |= IOMMU_CMD_LINE_DMA_API;
2671 iommu_def_domain_type = IOMMU_DOMAIN_IDENTITY;
2672}
2673
2674void iommu_set_default_translated(bool cmd_line)
2675{
2676 if (cmd_line)
2677 iommu_cmd_line |= IOMMU_CMD_LINE_DMA_API;
2678 iommu_def_domain_type = IOMMU_DOMAIN_DMA;
2679}
2680
2681bool iommu_default_passthrough(void)
2682{
2683 return iommu_def_domain_type == IOMMU_DOMAIN_IDENTITY;
2684}
2685EXPORT_SYMBOL_GPL(iommu_default_passthrough);
2686
2687const struct iommu_ops *iommu_ops_from_fwnode(struct fwnode_handle *fwnode)
2688{
2689 const struct iommu_ops *ops = NULL;
2690 struct iommu_device *iommu;
2691
2692 spin_lock(&iommu_device_lock);
2693 list_for_each_entry(iommu, &iommu_device_list, list)
2694 if (iommu->fwnode == fwnode) {
2695 ops = iommu->ops;
2696 break;
2697 }
2698 spin_unlock(&iommu_device_lock);
2699 return ops;
2700}
2701
2702int iommu_fwspec_init(struct device *dev, struct fwnode_handle *iommu_fwnode,
2703 const struct iommu_ops *ops)
2704{
2705 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
2706
2707 if (fwspec)
2708 return ops == fwspec->ops ? 0 : -EINVAL;
2709
2710 if (!dev_iommu_get(dev))
2711 return -ENOMEM;
2712
2713 /* Preallocate for the overwhelmingly common case of 1 ID */
2714 fwspec = kzalloc(struct_size(fwspec, ids, 1), GFP_KERNEL);
2715 if (!fwspec)
2716 return -ENOMEM;
2717
2718 of_node_get(to_of_node(iommu_fwnode));
2719 fwspec->iommu_fwnode = iommu_fwnode;
2720 fwspec->ops = ops;
2721 dev_iommu_fwspec_set(dev, fwspec);
2722 return 0;
2723}
2724EXPORT_SYMBOL_GPL(iommu_fwspec_init);
2725
2726void iommu_fwspec_free(struct device *dev)
2727{
2728 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
2729
2730 if (fwspec) {
2731 fwnode_handle_put(fwspec->iommu_fwnode);
2732 kfree(fwspec);
2733 dev_iommu_fwspec_set(dev, NULL);
2734 }
2735}
2736EXPORT_SYMBOL_GPL(iommu_fwspec_free);
2737
2738int iommu_fwspec_add_ids(struct device *dev, u32 *ids, int num_ids)
2739{
2740 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
2741 int i, new_num;
2742
2743 if (!fwspec)
2744 return -EINVAL;
2745
2746 new_num = fwspec->num_ids + num_ids;
2747 if (new_num > 1) {
2748 fwspec = krealloc(fwspec, struct_size(fwspec, ids, new_num),
2749 GFP_KERNEL);
2750 if (!fwspec)
2751 return -ENOMEM;
2752
2753 dev_iommu_fwspec_set(dev, fwspec);
2754 }
2755
2756 for (i = 0; i < num_ids; i++)
2757 fwspec->ids[fwspec->num_ids + i] = ids[i];
2758
2759 fwspec->num_ids = new_num;
2760 return 0;
2761}
2762EXPORT_SYMBOL_GPL(iommu_fwspec_add_ids);
2763
2764/*
2765 * Per device IOMMU features.
2766 */
2767int iommu_dev_enable_feature(struct device *dev, enum iommu_dev_features feat)
2768{
2769 if (dev->iommu && dev->iommu->iommu_dev) {
2770 const struct iommu_ops *ops = dev->iommu->iommu_dev->ops;
2771
2772 if (ops->dev_enable_feat)
2773 return ops->dev_enable_feat(dev, feat);
2774 }
2775
2776 return -ENODEV;
2777}
2778EXPORT_SYMBOL_GPL(iommu_dev_enable_feature);
2779
2780/*
2781 * The device drivers should do the necessary cleanups before calling this.
2782 */
2783int iommu_dev_disable_feature(struct device *dev, enum iommu_dev_features feat)
2784{
2785 if (dev->iommu && dev->iommu->iommu_dev) {
2786 const struct iommu_ops *ops = dev->iommu->iommu_dev->ops;
2787
2788 if (ops->dev_disable_feat)
2789 return ops->dev_disable_feat(dev, feat);
2790 }
2791
2792 return -EBUSY;
2793}
2794EXPORT_SYMBOL_GPL(iommu_dev_disable_feature);
2795
2796/*
2797 * Changes the default domain of an iommu group that has *only* one device
2798 *
2799 * @group: The group for which the default domain should be changed
2800 * @prev_dev: The device in the group (this is used to make sure that the device
2801 * hasn't changed after the caller has called this function)
2802 * @type: The type of the new default domain that gets associated with the group
2803 *
2804 * Returns 0 on success and error code on failure
2805 *
2806 * Note:
2807 * 1. Presently, this function is called only when user requests to change the
2808 * group's default domain type through /sys/kernel/iommu_groups/<grp_id>/type
2809 * Please take a closer look if intended to use for other purposes.
2810 */
2811static int iommu_change_dev_def_domain(struct iommu_group *group,
2812 struct device *prev_dev, int type)
2813{
2814 struct iommu_domain *prev_dom;
2815 struct group_device *grp_dev;
2816 int ret, dev_def_dom;
2817 struct device *dev;
2818
2819 mutex_lock(&group->mutex);
2820
2821 if (group->default_domain != group->domain) {
2822 dev_err_ratelimited(prev_dev, "Group not assigned to default domain\n");
2823 ret = -EBUSY;
2824 goto out;
2825 }
2826
2827 /*
2828 * iommu group wasn't locked while acquiring device lock in
2829 * iommu_group_store_type(). So, make sure that the device count hasn't
2830 * changed while acquiring device lock.
2831 *
2832 * Changing default domain of an iommu group with two or more devices
2833 * isn't supported because there could be a potential deadlock. Consider
2834 * the following scenario. T1 is trying to acquire device locks of all
2835 * the devices in the group and before it could acquire all of them,
2836 * there could be another thread T2 (from different sub-system and use
2837 * case) that has already acquired some of the device locks and might be
2838 * waiting for T1 to release other device locks.
2839 */
2840 if (iommu_group_device_count(group) != 1) {
2841 dev_err_ratelimited(prev_dev, "Cannot change default domain: Group has more than one device\n");
2842 ret = -EINVAL;
2843 goto out;
2844 }
2845
2846 /* Since group has only one device */
2847 grp_dev = list_first_entry(&group->devices, struct group_device, list);
2848 dev = grp_dev->dev;
2849
2850 if (prev_dev != dev) {
2851 dev_err_ratelimited(prev_dev, "Cannot change default domain: Device has been changed\n");
2852 ret = -EBUSY;
2853 goto out;
2854 }
2855
2856 prev_dom = group->default_domain;
2857 if (!prev_dom) {
2858 ret = -EINVAL;
2859 goto out;
2860 }
2861
2862 dev_def_dom = iommu_get_def_domain_type(dev);
2863 if (!type) {
2864 /*
2865 * If the user hasn't requested any specific type of domain and
2866 * if the device supports both the domains, then default to the
2867 * domain the device was booted with
2868 */
2869 type = dev_def_dom ? : iommu_def_domain_type;
2870 } else if (dev_def_dom && type != dev_def_dom) {
2871 dev_err_ratelimited(prev_dev, "Device cannot be in %s domain\n",
2872 iommu_domain_type_str(type));
2873 ret = -EINVAL;
2874 goto out;
2875 }
2876
2877 /*
2878 * Switch to a new domain only if the requested domain type is different
2879 * from the existing default domain type
2880 */
2881 if (prev_dom->type == type) {
2882 ret = 0;
2883 goto out;
2884 }
2885
2886 /* We can bring up a flush queue without tearing down the domain */
2887 if (type == IOMMU_DOMAIN_DMA_FQ && prev_dom->type == IOMMU_DOMAIN_DMA) {
2888 ret = iommu_dma_init_fq(prev_dom);
2889 if (!ret)
2890 prev_dom->type = IOMMU_DOMAIN_DMA_FQ;
2891 goto out;
2892 }
2893
2894 /* Sets group->default_domain to the newly allocated domain */
2895 ret = iommu_group_alloc_default_domain(dev->bus, group, type);
2896 if (ret)
2897 goto out;
2898
2899 ret = iommu_create_device_direct_mappings(group, dev);
2900 if (ret)
2901 goto free_new_domain;
2902
2903 ret = __iommu_attach_device(group->default_domain, dev);
2904 if (ret)
2905 goto free_new_domain;
2906
2907 group->domain = group->default_domain;
2908
2909 /*
2910 * Release the mutex here because ops->probe_finalize() call-back of
2911 * some vendor IOMMU drivers calls arm_iommu_attach_device() which
2912 * in-turn might call back into IOMMU core code, where it tries to take
2913 * group->mutex, resulting in a deadlock.
2914 */
2915 mutex_unlock(&group->mutex);
2916
2917 /* Make sure dma_ops is appropriatley set */
2918 iommu_group_do_probe_finalize(dev, group->default_domain);
2919 iommu_domain_free(prev_dom);
2920 return 0;
2921
2922free_new_domain:
2923 iommu_domain_free(group->default_domain);
2924 group->default_domain = prev_dom;
2925 group->domain = prev_dom;
2926
2927out:
2928 mutex_unlock(&group->mutex);
2929
2930 return ret;
2931}
2932
2933/*
2934 * Changing the default domain through sysfs requires the users to unbind the
2935 * drivers from the devices in the iommu group, except for a DMA -> DMA-FQ
2936 * transition. Return failure if this isn't met.
2937 *
2938 * We need to consider the race between this and the device release path.
2939 * device_lock(dev) is used here to guarantee that the device release path
2940 * will not be entered at the same time.
2941 */
2942static ssize_t iommu_group_store_type(struct iommu_group *group,
2943 const char *buf, size_t count)
2944{
2945 struct group_device *grp_dev;
2946 struct device *dev;
2947 int ret, req_type;
2948
2949 if (!capable(CAP_SYS_ADMIN) || !capable(CAP_SYS_RAWIO))
2950 return -EACCES;
2951
2952 if (WARN_ON(!group) || !group->default_domain)
2953 return -EINVAL;
2954
2955 if (sysfs_streq(buf, "identity"))
2956 req_type = IOMMU_DOMAIN_IDENTITY;
2957 else if (sysfs_streq(buf, "DMA"))
2958 req_type = IOMMU_DOMAIN_DMA;
2959 else if (sysfs_streq(buf, "DMA-FQ"))
2960 req_type = IOMMU_DOMAIN_DMA_FQ;
2961 else if (sysfs_streq(buf, "auto"))
2962 req_type = 0;
2963 else
2964 return -EINVAL;
2965
2966 /*
2967 * Lock/Unlock the group mutex here before device lock to
2968 * 1. Make sure that the iommu group has only one device (this is a
2969 * prerequisite for step 2)
2970 * 2. Get struct *dev which is needed to lock device
2971 */
2972 mutex_lock(&group->mutex);
2973 if (iommu_group_device_count(group) != 1) {
2974 mutex_unlock(&group->mutex);
2975 pr_err_ratelimited("Cannot change default domain: Group has more than one device\n");
2976 return -EINVAL;
2977 }
2978
2979 /* Since group has only one device */
2980 grp_dev = list_first_entry(&group->devices, struct group_device, list);
2981 dev = grp_dev->dev;
2982 get_device(dev);
2983
2984 /*
2985 * Don't hold the group mutex because taking group mutex first and then
2986 * the device lock could potentially cause a deadlock as below. Assume
2987 * two threads T1 and T2. T1 is trying to change default domain of an
2988 * iommu group and T2 is trying to hot unplug a device or release [1] VF
2989 * of a PCIe device which is in the same iommu group. T1 takes group
2990 * mutex and before it could take device lock assume T2 has taken device
2991 * lock and is yet to take group mutex. Now, both the threads will be
2992 * waiting for the other thread to release lock. Below, lock order was
2993 * suggested.
2994 * device_lock(dev);
2995 * mutex_lock(&group->mutex);
2996 * iommu_change_dev_def_domain();
2997 * mutex_unlock(&group->mutex);
2998 * device_unlock(dev);
2999 *
3000 * [1] Typical device release path
3001 * device_lock() from device/driver core code
3002 * -> bus_notifier()
3003 * -> iommu_bus_notifier()
3004 * -> iommu_release_device()
3005 * -> ops->release_device() vendor driver calls back iommu core code
3006 * -> mutex_lock() from iommu core code
3007 */
3008 mutex_unlock(&group->mutex);
3009
3010 /* Check if the device in the group still has a driver bound to it */
3011 device_lock(dev);
3012 if (device_is_bound(dev) && !(req_type == IOMMU_DOMAIN_DMA_FQ &&
3013 group->default_domain->type == IOMMU_DOMAIN_DMA)) {
3014 pr_err_ratelimited("Device is still bound to driver\n");
3015 ret = -EBUSY;
3016 goto out;
3017 }
3018
3019 ret = iommu_change_dev_def_domain(group, dev, req_type);
3020 ret = ret ?: count;
3021
3022out:
3023 device_unlock(dev);
3024 put_device(dev);
3025
3026 return ret;
3027}
3028
3029static bool iommu_is_default_domain(struct iommu_group *group)
3030{
3031 if (group->domain == group->default_domain)
3032 return true;
3033
3034 /*
3035 * If the default domain was set to identity and it is still an identity
3036 * domain then we consider this a pass. This happens because of
3037 * amd_iommu_init_device() replacing the default idenytity domain with an
3038 * identity domain that has a different configuration for AMDGPU.
3039 */
3040 if (group->default_domain &&
3041 group->default_domain->type == IOMMU_DOMAIN_IDENTITY &&
3042 group->domain && group->domain->type == IOMMU_DOMAIN_IDENTITY)
3043 return true;
3044 return false;
3045}
3046
3047/**
3048 * iommu_device_use_default_domain() - Device driver wants to handle device
3049 * DMA through the kernel DMA API.
3050 * @dev: The device.
3051 *
3052 * The device driver about to bind @dev wants to do DMA through the kernel
3053 * DMA API. Return 0 if it is allowed, otherwise an error.
3054 */
3055int iommu_device_use_default_domain(struct device *dev)
3056{
3057 struct iommu_group *group = iommu_group_get(dev);
3058 int ret = 0;
3059
3060 if (!group)
3061 return 0;
3062
3063 mutex_lock(&group->mutex);
3064 if (group->owner_cnt) {
3065 if (group->owner || !iommu_is_default_domain(group) ||
3066 !xa_empty(&group->pasid_array)) {
3067 ret = -EBUSY;
3068 goto unlock_out;
3069 }
3070 }
3071
3072 group->owner_cnt++;
3073
3074unlock_out:
3075 mutex_unlock(&group->mutex);
3076 iommu_group_put(group);
3077
3078 return ret;
3079}
3080
3081/**
3082 * iommu_device_unuse_default_domain() - Device driver stops handling device
3083 * DMA through the kernel DMA API.
3084 * @dev: The device.
3085 *
3086 * The device driver doesn't want to do DMA through kernel DMA API anymore.
3087 * It must be called after iommu_device_use_default_domain().
3088 */
3089void iommu_device_unuse_default_domain(struct device *dev)
3090{
3091 struct iommu_group *group = iommu_group_get(dev);
3092
3093 if (!group)
3094 return;
3095
3096 mutex_lock(&group->mutex);
3097 if (!WARN_ON(!group->owner_cnt || !xa_empty(&group->pasid_array)))
3098 group->owner_cnt--;
3099
3100 mutex_unlock(&group->mutex);
3101 iommu_group_put(group);
3102}
3103
3104static int __iommu_group_alloc_blocking_domain(struct iommu_group *group)
3105{
3106 struct group_device *dev =
3107 list_first_entry(&group->devices, struct group_device, list);
3108
3109 if (group->blocking_domain)
3110 return 0;
3111
3112 group->blocking_domain =
3113 __iommu_domain_alloc(dev->dev->bus, IOMMU_DOMAIN_BLOCKED);
3114 if (!group->blocking_domain) {
3115 /*
3116 * For drivers that do not yet understand IOMMU_DOMAIN_BLOCKED
3117 * create an empty domain instead.
3118 */
3119 group->blocking_domain = __iommu_domain_alloc(
3120 dev->dev->bus, IOMMU_DOMAIN_UNMANAGED);
3121 if (!group->blocking_domain)
3122 return -EINVAL;
3123 }
3124 return 0;
3125}
3126
3127static int __iommu_take_dma_ownership(struct iommu_group *group, void *owner)
3128{
3129 int ret;
3130
3131 if ((group->domain && group->domain != group->default_domain) ||
3132 !xa_empty(&group->pasid_array))
3133 return -EBUSY;
3134
3135 ret = __iommu_group_alloc_blocking_domain(group);
3136 if (ret)
3137 return ret;
3138 ret = __iommu_group_set_domain(group, group->blocking_domain);
3139 if (ret)
3140 return ret;
3141
3142 group->owner = owner;
3143 group->owner_cnt++;
3144 return 0;
3145}
3146
3147/**
3148 * iommu_group_claim_dma_owner() - Set DMA ownership of a group
3149 * @group: The group.
3150 * @owner: Caller specified pointer. Used for exclusive ownership.
3151 *
3152 * This is to support backward compatibility for vfio which manages the dma
3153 * ownership in iommu_group level. New invocations on this interface should be
3154 * prohibited. Only a single owner may exist for a group.
3155 */
3156int iommu_group_claim_dma_owner(struct iommu_group *group, void *owner)
3157{
3158 int ret = 0;
3159
3160 if (WARN_ON(!owner))
3161 return -EINVAL;
3162
3163 mutex_lock(&group->mutex);
3164 if (group->owner_cnt) {
3165 ret = -EPERM;
3166 goto unlock_out;
3167 }
3168
3169 ret = __iommu_take_dma_ownership(group, owner);
3170unlock_out:
3171 mutex_unlock(&group->mutex);
3172
3173 return ret;
3174}
3175EXPORT_SYMBOL_GPL(iommu_group_claim_dma_owner);
3176
3177/**
3178 * iommu_device_claim_dma_owner() - Set DMA ownership of a device
3179 * @dev: The device.
3180 * @owner: Caller specified pointer. Used for exclusive ownership.
3181 *
3182 * Claim the DMA ownership of a device. Multiple devices in the same group may
3183 * concurrently claim ownership if they present the same owner value. Returns 0
3184 * on success and error code on failure
3185 */
3186int iommu_device_claim_dma_owner(struct device *dev, void *owner)
3187{
3188 struct iommu_group *group;
3189 int ret = 0;
3190
3191 if (WARN_ON(!owner))
3192 return -EINVAL;
3193
3194 group = iommu_group_get(dev);
3195 if (!group)
3196 return -ENODEV;
3197
3198 mutex_lock(&group->mutex);
3199 if (group->owner_cnt) {
3200 if (group->owner != owner) {
3201 ret = -EPERM;
3202 goto unlock_out;
3203 }
3204 group->owner_cnt++;
3205 goto unlock_out;
3206 }
3207
3208 ret = __iommu_take_dma_ownership(group, owner);
3209unlock_out:
3210 mutex_unlock(&group->mutex);
3211 iommu_group_put(group);
3212
3213 return ret;
3214}
3215EXPORT_SYMBOL_GPL(iommu_device_claim_dma_owner);
3216
3217static void __iommu_release_dma_ownership(struct iommu_group *group)
3218{
3219 int ret;
3220
3221 if (WARN_ON(!group->owner_cnt || !group->owner ||
3222 !xa_empty(&group->pasid_array)))
3223 return;
3224
3225 group->owner_cnt = 0;
3226 group->owner = NULL;
3227 ret = __iommu_group_set_domain(group, group->default_domain);
3228 WARN(ret, "iommu driver failed to attach the default domain");
3229}
3230
3231/**
3232 * iommu_group_release_dma_owner() - Release DMA ownership of a group
3233 * @dev: The device
3234 *
3235 * Release the DMA ownership claimed by iommu_group_claim_dma_owner().
3236 */
3237void iommu_group_release_dma_owner(struct iommu_group *group)
3238{
3239 mutex_lock(&group->mutex);
3240 __iommu_release_dma_ownership(group);
3241 mutex_unlock(&group->mutex);
3242}
3243EXPORT_SYMBOL_GPL(iommu_group_release_dma_owner);
3244
3245/**
3246 * iommu_device_release_dma_owner() - Release DMA ownership of a device
3247 * @group: The device.
3248 *
3249 * Release the DMA ownership claimed by iommu_device_claim_dma_owner().
3250 */
3251void iommu_device_release_dma_owner(struct device *dev)
3252{
3253 struct iommu_group *group = iommu_group_get(dev);
3254
3255 mutex_lock(&group->mutex);
3256 if (group->owner_cnt > 1)
3257 group->owner_cnt--;
3258 else
3259 __iommu_release_dma_ownership(group);
3260 mutex_unlock(&group->mutex);
3261 iommu_group_put(group);
3262}
3263EXPORT_SYMBOL_GPL(iommu_device_release_dma_owner);
3264
3265/**
3266 * iommu_group_dma_owner_claimed() - Query group dma ownership status
3267 * @group: The group.
3268 *
3269 * This provides status query on a given group. It is racy and only for
3270 * non-binding status reporting.
3271 */
3272bool iommu_group_dma_owner_claimed(struct iommu_group *group)
3273{
3274 unsigned int user;
3275
3276 mutex_lock(&group->mutex);
3277 user = group->owner_cnt;
3278 mutex_unlock(&group->mutex);
3279
3280 return user;
3281}
3282EXPORT_SYMBOL_GPL(iommu_group_dma_owner_claimed);
3283
3284static int __iommu_set_group_pasid(struct iommu_domain *domain,
3285 struct iommu_group *group, ioasid_t pasid)
3286{
3287 struct group_device *device;
3288 int ret = 0;
3289
3290 list_for_each_entry(device, &group->devices, list) {
3291 ret = domain->ops->set_dev_pasid(domain, device->dev, pasid);
3292 if (ret)
3293 break;
3294 }
3295
3296 return ret;
3297}
3298
3299static void __iommu_remove_group_pasid(struct iommu_group *group,
3300 ioasid_t pasid)
3301{
3302 struct group_device *device;
3303 const struct iommu_ops *ops;
3304
3305 list_for_each_entry(device, &group->devices, list) {
3306 ops = dev_iommu_ops(device->dev);
3307 ops->remove_dev_pasid(device->dev, pasid);
3308 }
3309}
3310
3311/*
3312 * iommu_attach_device_pasid() - Attach a domain to pasid of device
3313 * @domain: the iommu domain.
3314 * @dev: the attached device.
3315 * @pasid: the pasid of the device.
3316 *
3317 * Return: 0 on success, or an error.
3318 */
3319int iommu_attach_device_pasid(struct iommu_domain *domain,
3320 struct device *dev, ioasid_t pasid)
3321{
3322 struct iommu_group *group;
3323 void *curr;
3324 int ret;
3325
3326 if (!domain->ops->set_dev_pasid)
3327 return -EOPNOTSUPP;
3328
3329 group = iommu_group_get(dev);
3330 if (!group)
3331 return -ENODEV;
3332
3333 mutex_lock(&group->mutex);
3334 curr = xa_cmpxchg(&group->pasid_array, pasid, NULL, domain, GFP_KERNEL);
3335 if (curr) {
3336 ret = xa_err(curr) ? : -EBUSY;
3337 goto out_unlock;
3338 }
3339
3340 ret = __iommu_set_group_pasid(domain, group, pasid);
3341 if (ret) {
3342 __iommu_remove_group_pasid(group, pasid);
3343 xa_erase(&group->pasid_array, pasid);
3344 }
3345out_unlock:
3346 mutex_unlock(&group->mutex);
3347 iommu_group_put(group);
3348
3349 return ret;
3350}
3351EXPORT_SYMBOL_GPL(iommu_attach_device_pasid);
3352
3353/*
3354 * iommu_detach_device_pasid() - Detach the domain from pasid of device
3355 * @domain: the iommu domain.
3356 * @dev: the attached device.
3357 * @pasid: the pasid of the device.
3358 *
3359 * The @domain must have been attached to @pasid of the @dev with
3360 * iommu_attach_device_pasid().
3361 */
3362void iommu_detach_device_pasid(struct iommu_domain *domain, struct device *dev,
3363 ioasid_t pasid)
3364{
3365 struct iommu_group *group = iommu_group_get(dev);
3366
3367 mutex_lock(&group->mutex);
3368 __iommu_remove_group_pasid(group, pasid);
3369 WARN_ON(xa_erase(&group->pasid_array, pasid) != domain);
3370 mutex_unlock(&group->mutex);
3371
3372 iommu_group_put(group);
3373}
3374EXPORT_SYMBOL_GPL(iommu_detach_device_pasid);
3375
3376/*
3377 * iommu_get_domain_for_dev_pasid() - Retrieve domain for @pasid of @dev
3378 * @dev: the queried device
3379 * @pasid: the pasid of the device
3380 * @type: matched domain type, 0 for any match
3381 *
3382 * This is a variant of iommu_get_domain_for_dev(). It returns the existing
3383 * domain attached to pasid of a device. Callers must hold a lock around this
3384 * function, and both iommu_attach/detach_dev_pasid() whenever a domain of
3385 * type is being manipulated. This API does not internally resolve races with
3386 * attach/detach.
3387 *
3388 * Return: attached domain on success, NULL otherwise.
3389 */
3390struct iommu_domain *iommu_get_domain_for_dev_pasid(struct device *dev,
3391 ioasid_t pasid,
3392 unsigned int type)
3393{
3394 struct iommu_domain *domain;
3395 struct iommu_group *group;
3396
3397 group = iommu_group_get(dev);
3398 if (!group)
3399 return NULL;
3400
3401 xa_lock(&group->pasid_array);
3402 domain = xa_load(&group->pasid_array, pasid);
3403 if (type && domain && domain->type != type)
3404 domain = ERR_PTR(-EBUSY);
3405 xa_unlock(&group->pasid_array);
3406 iommu_group_put(group);
3407
3408 return domain;
3409}
3410EXPORT_SYMBOL_GPL(iommu_get_domain_for_dev_pasid);
3411
3412struct iommu_domain *iommu_sva_domain_alloc(struct device *dev,
3413 struct mm_struct *mm)
3414{
3415 const struct iommu_ops *ops = dev_iommu_ops(dev);
3416 struct iommu_domain *domain;
3417
3418 domain = ops->domain_alloc(IOMMU_DOMAIN_SVA);
3419 if (!domain)
3420 return NULL;
3421
3422 domain->type = IOMMU_DOMAIN_SVA;
3423 mmgrab(mm);
3424 domain->mm = mm;
3425 domain->iopf_handler = iommu_sva_handle_iopf;
3426 domain->fault_data = mm;
3427
3428 return domain;
3429}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2007-2008 Advanced Micro Devices, Inc.
4 * Author: Joerg Roedel <jroedel@suse.de>
5 */
6
7#define pr_fmt(fmt) "iommu: " fmt
8
9#include <linux/amba/bus.h>
10#include <linux/device.h>
11#include <linux/kernel.h>
12#include <linux/bits.h>
13#include <linux/bug.h>
14#include <linux/types.h>
15#include <linux/init.h>
16#include <linux/export.h>
17#include <linux/slab.h>
18#include <linux/errno.h>
19#include <linux/host1x_context_bus.h>
20#include <linux/iommu.h>
21#include <linux/idr.h>
22#include <linux/err.h>
23#include <linux/pci.h>
24#include <linux/pci-ats.h>
25#include <linux/bitops.h>
26#include <linux/platform_device.h>
27#include <linux/property.h>
28#include <linux/fsl/mc.h>
29#include <linux/module.h>
30#include <linux/cc_platform.h>
31#include <linux/cdx/cdx_bus.h>
32#include <trace/events/iommu.h>
33#include <linux/sched/mm.h>
34#include <linux/msi.h>
35#include <uapi/linux/iommufd.h>
36
37#include "dma-iommu.h"
38#include "iommu-priv.h"
39
40static struct kset *iommu_group_kset;
41static DEFINE_IDA(iommu_group_ida);
42static DEFINE_IDA(iommu_global_pasid_ida);
43
44static unsigned int iommu_def_domain_type __read_mostly;
45static bool iommu_dma_strict __read_mostly = IS_ENABLED(CONFIG_IOMMU_DEFAULT_DMA_STRICT);
46static u32 iommu_cmd_line __read_mostly;
47
48struct iommu_group {
49 struct kobject kobj;
50 struct kobject *devices_kobj;
51 struct list_head devices;
52 struct xarray pasid_array;
53 struct mutex mutex;
54 void *iommu_data;
55 void (*iommu_data_release)(void *iommu_data);
56 char *name;
57 int id;
58 struct iommu_domain *default_domain;
59 struct iommu_domain *blocking_domain;
60 struct iommu_domain *domain;
61 struct list_head entry;
62 unsigned int owner_cnt;
63 void *owner;
64};
65
66struct group_device {
67 struct list_head list;
68 struct device *dev;
69 char *name;
70};
71
72/* Iterate over each struct group_device in a struct iommu_group */
73#define for_each_group_device(group, pos) \
74 list_for_each_entry(pos, &(group)->devices, list)
75
76struct iommu_group_attribute {
77 struct attribute attr;
78 ssize_t (*show)(struct iommu_group *group, char *buf);
79 ssize_t (*store)(struct iommu_group *group,
80 const char *buf, size_t count);
81};
82
83static const char * const iommu_group_resv_type_string[] = {
84 [IOMMU_RESV_DIRECT] = "direct",
85 [IOMMU_RESV_DIRECT_RELAXABLE] = "direct-relaxable",
86 [IOMMU_RESV_RESERVED] = "reserved",
87 [IOMMU_RESV_MSI] = "msi",
88 [IOMMU_RESV_SW_MSI] = "msi",
89};
90
91#define IOMMU_CMD_LINE_DMA_API BIT(0)
92#define IOMMU_CMD_LINE_STRICT BIT(1)
93
94static int bus_iommu_probe(const struct bus_type *bus);
95static int iommu_bus_notifier(struct notifier_block *nb,
96 unsigned long action, void *data);
97static void iommu_release_device(struct device *dev);
98static int __iommu_attach_device(struct iommu_domain *domain,
99 struct device *dev);
100static int __iommu_attach_group(struct iommu_domain *domain,
101 struct iommu_group *group);
102static struct iommu_domain *__iommu_paging_domain_alloc_flags(struct device *dev,
103 unsigned int type,
104 unsigned int flags);
105
106enum {
107 IOMMU_SET_DOMAIN_MUST_SUCCEED = 1 << 0,
108};
109
110static int __iommu_device_set_domain(struct iommu_group *group,
111 struct device *dev,
112 struct iommu_domain *new_domain,
113 unsigned int flags);
114static int __iommu_group_set_domain_internal(struct iommu_group *group,
115 struct iommu_domain *new_domain,
116 unsigned int flags);
117static int __iommu_group_set_domain(struct iommu_group *group,
118 struct iommu_domain *new_domain)
119{
120 return __iommu_group_set_domain_internal(group, new_domain, 0);
121}
122static void __iommu_group_set_domain_nofail(struct iommu_group *group,
123 struct iommu_domain *new_domain)
124{
125 WARN_ON(__iommu_group_set_domain_internal(
126 group, new_domain, IOMMU_SET_DOMAIN_MUST_SUCCEED));
127}
128
129static int iommu_setup_default_domain(struct iommu_group *group,
130 int target_type);
131static int iommu_create_device_direct_mappings(struct iommu_domain *domain,
132 struct device *dev);
133static ssize_t iommu_group_store_type(struct iommu_group *group,
134 const char *buf, size_t count);
135static struct group_device *iommu_group_alloc_device(struct iommu_group *group,
136 struct device *dev);
137static void __iommu_group_free_device(struct iommu_group *group,
138 struct group_device *grp_dev);
139static void iommu_domain_init(struct iommu_domain *domain, unsigned int type,
140 const struct iommu_ops *ops);
141
142#define IOMMU_GROUP_ATTR(_name, _mode, _show, _store) \
143struct iommu_group_attribute iommu_group_attr_##_name = \
144 __ATTR(_name, _mode, _show, _store)
145
146#define to_iommu_group_attr(_attr) \
147 container_of(_attr, struct iommu_group_attribute, attr)
148#define to_iommu_group(_kobj) \
149 container_of(_kobj, struct iommu_group, kobj)
150
151static LIST_HEAD(iommu_device_list);
152static DEFINE_SPINLOCK(iommu_device_lock);
153
154static const struct bus_type * const iommu_buses[] = {
155 &platform_bus_type,
156#ifdef CONFIG_PCI
157 &pci_bus_type,
158#endif
159#ifdef CONFIG_ARM_AMBA
160 &amba_bustype,
161#endif
162#ifdef CONFIG_FSL_MC_BUS
163 &fsl_mc_bus_type,
164#endif
165#ifdef CONFIG_TEGRA_HOST1X_CONTEXT_BUS
166 &host1x_context_device_bus_type,
167#endif
168#ifdef CONFIG_CDX_BUS
169 &cdx_bus_type,
170#endif
171};
172
173/*
174 * Use a function instead of an array here because the domain-type is a
175 * bit-field, so an array would waste memory.
176 */
177static const char *iommu_domain_type_str(unsigned int t)
178{
179 switch (t) {
180 case IOMMU_DOMAIN_BLOCKED:
181 return "Blocked";
182 case IOMMU_DOMAIN_IDENTITY:
183 return "Passthrough";
184 case IOMMU_DOMAIN_UNMANAGED:
185 return "Unmanaged";
186 case IOMMU_DOMAIN_DMA:
187 case IOMMU_DOMAIN_DMA_FQ:
188 return "Translated";
189 case IOMMU_DOMAIN_PLATFORM:
190 return "Platform";
191 default:
192 return "Unknown";
193 }
194}
195
196static int __init iommu_subsys_init(void)
197{
198 struct notifier_block *nb;
199
200 if (!(iommu_cmd_line & IOMMU_CMD_LINE_DMA_API)) {
201 if (IS_ENABLED(CONFIG_IOMMU_DEFAULT_PASSTHROUGH))
202 iommu_set_default_passthrough(false);
203 else
204 iommu_set_default_translated(false);
205
206 if (iommu_default_passthrough() && cc_platform_has(CC_ATTR_MEM_ENCRYPT)) {
207 pr_info("Memory encryption detected - Disabling default IOMMU Passthrough\n");
208 iommu_set_default_translated(false);
209 }
210 }
211
212 if (!iommu_default_passthrough() && !iommu_dma_strict)
213 iommu_def_domain_type = IOMMU_DOMAIN_DMA_FQ;
214
215 pr_info("Default domain type: %s%s\n",
216 iommu_domain_type_str(iommu_def_domain_type),
217 (iommu_cmd_line & IOMMU_CMD_LINE_DMA_API) ?
218 " (set via kernel command line)" : "");
219
220 if (!iommu_default_passthrough())
221 pr_info("DMA domain TLB invalidation policy: %s mode%s\n",
222 iommu_dma_strict ? "strict" : "lazy",
223 (iommu_cmd_line & IOMMU_CMD_LINE_STRICT) ?
224 " (set via kernel command line)" : "");
225
226 nb = kcalloc(ARRAY_SIZE(iommu_buses), sizeof(*nb), GFP_KERNEL);
227 if (!nb)
228 return -ENOMEM;
229
230 for (int i = 0; i < ARRAY_SIZE(iommu_buses); i++) {
231 nb[i].notifier_call = iommu_bus_notifier;
232 bus_register_notifier(iommu_buses[i], &nb[i]);
233 }
234
235 return 0;
236}
237subsys_initcall(iommu_subsys_init);
238
239static int remove_iommu_group(struct device *dev, void *data)
240{
241 if (dev->iommu && dev->iommu->iommu_dev == data)
242 iommu_release_device(dev);
243
244 return 0;
245}
246
247/**
248 * iommu_device_register() - Register an IOMMU hardware instance
249 * @iommu: IOMMU handle for the instance
250 * @ops: IOMMU ops to associate with the instance
251 * @hwdev: (optional) actual instance device, used for fwnode lookup
252 *
253 * Return: 0 on success, or an error.
254 */
255int iommu_device_register(struct iommu_device *iommu,
256 const struct iommu_ops *ops, struct device *hwdev)
257{
258 int err = 0;
259
260 /* We need to be able to take module references appropriately */
261 if (WARN_ON(is_module_address((unsigned long)ops) && !ops->owner))
262 return -EINVAL;
263
264 iommu->ops = ops;
265 if (hwdev)
266 iommu->fwnode = dev_fwnode(hwdev);
267
268 spin_lock(&iommu_device_lock);
269 list_add_tail(&iommu->list, &iommu_device_list);
270 spin_unlock(&iommu_device_lock);
271
272 for (int i = 0; i < ARRAY_SIZE(iommu_buses) && !err; i++)
273 err = bus_iommu_probe(iommu_buses[i]);
274 if (err)
275 iommu_device_unregister(iommu);
276 return err;
277}
278EXPORT_SYMBOL_GPL(iommu_device_register);
279
280void iommu_device_unregister(struct iommu_device *iommu)
281{
282 for (int i = 0; i < ARRAY_SIZE(iommu_buses); i++)
283 bus_for_each_dev(iommu_buses[i], NULL, iommu, remove_iommu_group);
284
285 spin_lock(&iommu_device_lock);
286 list_del(&iommu->list);
287 spin_unlock(&iommu_device_lock);
288
289 /* Pairs with the alloc in generic_single_device_group() */
290 iommu_group_put(iommu->singleton_group);
291 iommu->singleton_group = NULL;
292}
293EXPORT_SYMBOL_GPL(iommu_device_unregister);
294
295#if IS_ENABLED(CONFIG_IOMMUFD_TEST)
296void iommu_device_unregister_bus(struct iommu_device *iommu,
297 const struct bus_type *bus,
298 struct notifier_block *nb)
299{
300 bus_unregister_notifier(bus, nb);
301 iommu_device_unregister(iommu);
302}
303EXPORT_SYMBOL_GPL(iommu_device_unregister_bus);
304
305/*
306 * Register an iommu driver against a single bus. This is only used by iommufd
307 * selftest to create a mock iommu driver. The caller must provide
308 * some memory to hold a notifier_block.
309 */
310int iommu_device_register_bus(struct iommu_device *iommu,
311 const struct iommu_ops *ops,
312 const struct bus_type *bus,
313 struct notifier_block *nb)
314{
315 int err;
316
317 iommu->ops = ops;
318 nb->notifier_call = iommu_bus_notifier;
319 err = bus_register_notifier(bus, nb);
320 if (err)
321 return err;
322
323 spin_lock(&iommu_device_lock);
324 list_add_tail(&iommu->list, &iommu_device_list);
325 spin_unlock(&iommu_device_lock);
326
327 err = bus_iommu_probe(bus);
328 if (err) {
329 iommu_device_unregister_bus(iommu, bus, nb);
330 return err;
331 }
332 return 0;
333}
334EXPORT_SYMBOL_GPL(iommu_device_register_bus);
335#endif
336
337static struct dev_iommu *dev_iommu_get(struct device *dev)
338{
339 struct dev_iommu *param = dev->iommu;
340
341 lockdep_assert_held(&iommu_probe_device_lock);
342
343 if (param)
344 return param;
345
346 param = kzalloc(sizeof(*param), GFP_KERNEL);
347 if (!param)
348 return NULL;
349
350 mutex_init(¶m->lock);
351 dev->iommu = param;
352 return param;
353}
354
355static void dev_iommu_free(struct device *dev)
356{
357 struct dev_iommu *param = dev->iommu;
358
359 dev->iommu = NULL;
360 if (param->fwspec) {
361 fwnode_handle_put(param->fwspec->iommu_fwnode);
362 kfree(param->fwspec);
363 }
364 kfree(param);
365}
366
367/*
368 * Internal equivalent of device_iommu_mapped() for when we care that a device
369 * actually has API ops, and don't want false positives from VFIO-only groups.
370 */
371static bool dev_has_iommu(struct device *dev)
372{
373 return dev->iommu && dev->iommu->iommu_dev;
374}
375
376static u32 dev_iommu_get_max_pasids(struct device *dev)
377{
378 u32 max_pasids = 0, bits = 0;
379 int ret;
380
381 if (dev_is_pci(dev)) {
382 ret = pci_max_pasids(to_pci_dev(dev));
383 if (ret > 0)
384 max_pasids = ret;
385 } else {
386 ret = device_property_read_u32(dev, "pasid-num-bits", &bits);
387 if (!ret)
388 max_pasids = 1UL << bits;
389 }
390
391 return min_t(u32, max_pasids, dev->iommu->iommu_dev->max_pasids);
392}
393
394void dev_iommu_priv_set(struct device *dev, void *priv)
395{
396 /* FSL_PAMU does something weird */
397 if (!IS_ENABLED(CONFIG_FSL_PAMU))
398 lockdep_assert_held(&iommu_probe_device_lock);
399 dev->iommu->priv = priv;
400}
401EXPORT_SYMBOL_GPL(dev_iommu_priv_set);
402
403/*
404 * Init the dev->iommu and dev->iommu_group in the struct device and get the
405 * driver probed
406 */
407static int iommu_init_device(struct device *dev, const struct iommu_ops *ops)
408{
409 struct iommu_device *iommu_dev;
410 struct iommu_group *group;
411 int ret;
412
413 if (!dev_iommu_get(dev))
414 return -ENOMEM;
415
416 if (!try_module_get(ops->owner)) {
417 ret = -EINVAL;
418 goto err_free;
419 }
420
421 iommu_dev = ops->probe_device(dev);
422 if (IS_ERR(iommu_dev)) {
423 ret = PTR_ERR(iommu_dev);
424 goto err_module_put;
425 }
426 dev->iommu->iommu_dev = iommu_dev;
427
428 ret = iommu_device_link(iommu_dev, dev);
429 if (ret)
430 goto err_release;
431
432 group = ops->device_group(dev);
433 if (WARN_ON_ONCE(group == NULL))
434 group = ERR_PTR(-EINVAL);
435 if (IS_ERR(group)) {
436 ret = PTR_ERR(group);
437 goto err_unlink;
438 }
439 dev->iommu_group = group;
440
441 dev->iommu->max_pasids = dev_iommu_get_max_pasids(dev);
442 if (ops->is_attach_deferred)
443 dev->iommu->attach_deferred = ops->is_attach_deferred(dev);
444 return 0;
445
446err_unlink:
447 iommu_device_unlink(iommu_dev, dev);
448err_release:
449 if (ops->release_device)
450 ops->release_device(dev);
451err_module_put:
452 module_put(ops->owner);
453err_free:
454 dev->iommu->iommu_dev = NULL;
455 dev_iommu_free(dev);
456 return ret;
457}
458
459static void iommu_deinit_device(struct device *dev)
460{
461 struct iommu_group *group = dev->iommu_group;
462 const struct iommu_ops *ops = dev_iommu_ops(dev);
463
464 lockdep_assert_held(&group->mutex);
465
466 iommu_device_unlink(dev->iommu->iommu_dev, dev);
467
468 /*
469 * release_device() must stop using any attached domain on the device.
470 * If there are still other devices in the group, they are not affected
471 * by this callback.
472 *
473 * If the iommu driver provides release_domain, the core code ensures
474 * that domain is attached prior to calling release_device. Drivers can
475 * use this to enforce a translation on the idle iommu. Typically, the
476 * global static blocked_domain is a good choice.
477 *
478 * Otherwise, the iommu driver must set the device to either an identity
479 * or a blocking translation in release_device() and stop using any
480 * domain pointer, as it is going to be freed.
481 *
482 * Regardless, if a delayed attach never occurred, then the release
483 * should still avoid touching any hardware configuration either.
484 */
485 if (!dev->iommu->attach_deferred && ops->release_domain)
486 ops->release_domain->ops->attach_dev(ops->release_domain, dev);
487
488 if (ops->release_device)
489 ops->release_device(dev);
490
491 /*
492 * If this is the last driver to use the group then we must free the
493 * domains before we do the module_put().
494 */
495 if (list_empty(&group->devices)) {
496 if (group->default_domain) {
497 iommu_domain_free(group->default_domain);
498 group->default_domain = NULL;
499 }
500 if (group->blocking_domain) {
501 iommu_domain_free(group->blocking_domain);
502 group->blocking_domain = NULL;
503 }
504 group->domain = NULL;
505 }
506
507 /* Caller must put iommu_group */
508 dev->iommu_group = NULL;
509 module_put(ops->owner);
510 dev_iommu_free(dev);
511}
512
513DEFINE_MUTEX(iommu_probe_device_lock);
514
515static int __iommu_probe_device(struct device *dev, struct list_head *group_list)
516{
517 const struct iommu_ops *ops;
518 struct iommu_group *group;
519 struct group_device *gdev;
520 int ret;
521
522 /*
523 * For FDT-based systems and ACPI IORT/VIOT, drivers register IOMMU
524 * instances with non-NULL fwnodes, and client devices should have been
525 * identified with a fwspec by this point. Otherwise, we can currently
526 * assume that only one of Intel, AMD, s390, PAMU or legacy SMMUv2 can
527 * be present, and that any of their registered instances has suitable
528 * ops for probing, and thus cheekily co-opt the same mechanism.
529 */
530 ops = iommu_fwspec_ops(dev_iommu_fwspec_get(dev));
531 if (!ops)
532 return -ENODEV;
533 /*
534 * Serialise to avoid races between IOMMU drivers registering in
535 * parallel and/or the "replay" calls from ACPI/OF code via client
536 * driver probe. Once the latter have been cleaned up we should
537 * probably be able to use device_lock() here to minimise the scope,
538 * but for now enforcing a simple global ordering is fine.
539 */
540 lockdep_assert_held(&iommu_probe_device_lock);
541
542 /* Device is probed already if in a group */
543 if (dev->iommu_group)
544 return 0;
545
546 ret = iommu_init_device(dev, ops);
547 if (ret)
548 return ret;
549
550 group = dev->iommu_group;
551 gdev = iommu_group_alloc_device(group, dev);
552 mutex_lock(&group->mutex);
553 if (IS_ERR(gdev)) {
554 ret = PTR_ERR(gdev);
555 goto err_put_group;
556 }
557
558 /*
559 * The gdev must be in the list before calling
560 * iommu_setup_default_domain()
561 */
562 list_add_tail(&gdev->list, &group->devices);
563 WARN_ON(group->default_domain && !group->domain);
564 if (group->default_domain)
565 iommu_create_device_direct_mappings(group->default_domain, dev);
566 if (group->domain) {
567 ret = __iommu_device_set_domain(group, dev, group->domain, 0);
568 if (ret)
569 goto err_remove_gdev;
570 } else if (!group->default_domain && !group_list) {
571 ret = iommu_setup_default_domain(group, 0);
572 if (ret)
573 goto err_remove_gdev;
574 } else if (!group->default_domain) {
575 /*
576 * With a group_list argument we defer the default_domain setup
577 * to the caller by providing a de-duplicated list of groups
578 * that need further setup.
579 */
580 if (list_empty(&group->entry))
581 list_add_tail(&group->entry, group_list);
582 }
583
584 if (group->default_domain)
585 iommu_setup_dma_ops(dev);
586
587 mutex_unlock(&group->mutex);
588
589 return 0;
590
591err_remove_gdev:
592 list_del(&gdev->list);
593 __iommu_group_free_device(group, gdev);
594err_put_group:
595 iommu_deinit_device(dev);
596 mutex_unlock(&group->mutex);
597 iommu_group_put(group);
598
599 return ret;
600}
601
602int iommu_probe_device(struct device *dev)
603{
604 const struct iommu_ops *ops;
605 int ret;
606
607 mutex_lock(&iommu_probe_device_lock);
608 ret = __iommu_probe_device(dev, NULL);
609 mutex_unlock(&iommu_probe_device_lock);
610 if (ret)
611 return ret;
612
613 ops = dev_iommu_ops(dev);
614 if (ops->probe_finalize)
615 ops->probe_finalize(dev);
616
617 return 0;
618}
619
620static void __iommu_group_free_device(struct iommu_group *group,
621 struct group_device *grp_dev)
622{
623 struct device *dev = grp_dev->dev;
624
625 sysfs_remove_link(group->devices_kobj, grp_dev->name);
626 sysfs_remove_link(&dev->kobj, "iommu_group");
627
628 trace_remove_device_from_group(group->id, dev);
629
630 /*
631 * If the group has become empty then ownership must have been
632 * released, and the current domain must be set back to NULL or
633 * the default domain.
634 */
635 if (list_empty(&group->devices))
636 WARN_ON(group->owner_cnt ||
637 group->domain != group->default_domain);
638
639 kfree(grp_dev->name);
640 kfree(grp_dev);
641}
642
643/* Remove the iommu_group from the struct device. */
644static void __iommu_group_remove_device(struct device *dev)
645{
646 struct iommu_group *group = dev->iommu_group;
647 struct group_device *device;
648
649 mutex_lock(&group->mutex);
650 for_each_group_device(group, device) {
651 if (device->dev != dev)
652 continue;
653
654 list_del(&device->list);
655 __iommu_group_free_device(group, device);
656 if (dev_has_iommu(dev))
657 iommu_deinit_device(dev);
658 else
659 dev->iommu_group = NULL;
660 break;
661 }
662 mutex_unlock(&group->mutex);
663
664 /*
665 * Pairs with the get in iommu_init_device() or
666 * iommu_group_add_device()
667 */
668 iommu_group_put(group);
669}
670
671static void iommu_release_device(struct device *dev)
672{
673 struct iommu_group *group = dev->iommu_group;
674
675 if (group)
676 __iommu_group_remove_device(dev);
677
678 /* Free any fwspec if no iommu_driver was ever attached */
679 if (dev->iommu)
680 dev_iommu_free(dev);
681}
682
683static int __init iommu_set_def_domain_type(char *str)
684{
685 bool pt;
686 int ret;
687
688 ret = kstrtobool(str, &pt);
689 if (ret)
690 return ret;
691
692 if (pt)
693 iommu_set_default_passthrough(true);
694 else
695 iommu_set_default_translated(true);
696
697 return 0;
698}
699early_param("iommu.passthrough", iommu_set_def_domain_type);
700
701static int __init iommu_dma_setup(char *str)
702{
703 int ret = kstrtobool(str, &iommu_dma_strict);
704
705 if (!ret)
706 iommu_cmd_line |= IOMMU_CMD_LINE_STRICT;
707 return ret;
708}
709early_param("iommu.strict", iommu_dma_setup);
710
711void iommu_set_dma_strict(void)
712{
713 iommu_dma_strict = true;
714 if (iommu_def_domain_type == IOMMU_DOMAIN_DMA_FQ)
715 iommu_def_domain_type = IOMMU_DOMAIN_DMA;
716}
717
718static ssize_t iommu_group_attr_show(struct kobject *kobj,
719 struct attribute *__attr, char *buf)
720{
721 struct iommu_group_attribute *attr = to_iommu_group_attr(__attr);
722 struct iommu_group *group = to_iommu_group(kobj);
723 ssize_t ret = -EIO;
724
725 if (attr->show)
726 ret = attr->show(group, buf);
727 return ret;
728}
729
730static ssize_t iommu_group_attr_store(struct kobject *kobj,
731 struct attribute *__attr,
732 const char *buf, size_t count)
733{
734 struct iommu_group_attribute *attr = to_iommu_group_attr(__attr);
735 struct iommu_group *group = to_iommu_group(kobj);
736 ssize_t ret = -EIO;
737
738 if (attr->store)
739 ret = attr->store(group, buf, count);
740 return ret;
741}
742
743static const struct sysfs_ops iommu_group_sysfs_ops = {
744 .show = iommu_group_attr_show,
745 .store = iommu_group_attr_store,
746};
747
748static int iommu_group_create_file(struct iommu_group *group,
749 struct iommu_group_attribute *attr)
750{
751 return sysfs_create_file(&group->kobj, &attr->attr);
752}
753
754static void iommu_group_remove_file(struct iommu_group *group,
755 struct iommu_group_attribute *attr)
756{
757 sysfs_remove_file(&group->kobj, &attr->attr);
758}
759
760static ssize_t iommu_group_show_name(struct iommu_group *group, char *buf)
761{
762 return sysfs_emit(buf, "%s\n", group->name);
763}
764
765/**
766 * iommu_insert_resv_region - Insert a new region in the
767 * list of reserved regions.
768 * @new: new region to insert
769 * @regions: list of regions
770 *
771 * Elements are sorted by start address and overlapping segments
772 * of the same type are merged.
773 */
774static int iommu_insert_resv_region(struct iommu_resv_region *new,
775 struct list_head *regions)
776{
777 struct iommu_resv_region *iter, *tmp, *nr, *top;
778 LIST_HEAD(stack);
779
780 nr = iommu_alloc_resv_region(new->start, new->length,
781 new->prot, new->type, GFP_KERNEL);
782 if (!nr)
783 return -ENOMEM;
784
785 /* First add the new element based on start address sorting */
786 list_for_each_entry(iter, regions, list) {
787 if (nr->start < iter->start ||
788 (nr->start == iter->start && nr->type <= iter->type))
789 break;
790 }
791 list_add_tail(&nr->list, &iter->list);
792
793 /* Merge overlapping segments of type nr->type in @regions, if any */
794 list_for_each_entry_safe(iter, tmp, regions, list) {
795 phys_addr_t top_end, iter_end = iter->start + iter->length - 1;
796
797 /* no merge needed on elements of different types than @new */
798 if (iter->type != new->type) {
799 list_move_tail(&iter->list, &stack);
800 continue;
801 }
802
803 /* look for the last stack element of same type as @iter */
804 list_for_each_entry_reverse(top, &stack, list)
805 if (top->type == iter->type)
806 goto check_overlap;
807
808 list_move_tail(&iter->list, &stack);
809 continue;
810
811check_overlap:
812 top_end = top->start + top->length - 1;
813
814 if (iter->start > top_end + 1) {
815 list_move_tail(&iter->list, &stack);
816 } else {
817 top->length = max(top_end, iter_end) - top->start + 1;
818 list_del(&iter->list);
819 kfree(iter);
820 }
821 }
822 list_splice(&stack, regions);
823 return 0;
824}
825
826static int
827iommu_insert_device_resv_regions(struct list_head *dev_resv_regions,
828 struct list_head *group_resv_regions)
829{
830 struct iommu_resv_region *entry;
831 int ret = 0;
832
833 list_for_each_entry(entry, dev_resv_regions, list) {
834 ret = iommu_insert_resv_region(entry, group_resv_regions);
835 if (ret)
836 break;
837 }
838 return ret;
839}
840
841int iommu_get_group_resv_regions(struct iommu_group *group,
842 struct list_head *head)
843{
844 struct group_device *device;
845 int ret = 0;
846
847 mutex_lock(&group->mutex);
848 for_each_group_device(group, device) {
849 struct list_head dev_resv_regions;
850
851 /*
852 * Non-API groups still expose reserved_regions in sysfs,
853 * so filter out calls that get here that way.
854 */
855 if (!dev_has_iommu(device->dev))
856 break;
857
858 INIT_LIST_HEAD(&dev_resv_regions);
859 iommu_get_resv_regions(device->dev, &dev_resv_regions);
860 ret = iommu_insert_device_resv_regions(&dev_resv_regions, head);
861 iommu_put_resv_regions(device->dev, &dev_resv_regions);
862 if (ret)
863 break;
864 }
865 mutex_unlock(&group->mutex);
866 return ret;
867}
868EXPORT_SYMBOL_GPL(iommu_get_group_resv_regions);
869
870static ssize_t iommu_group_show_resv_regions(struct iommu_group *group,
871 char *buf)
872{
873 struct iommu_resv_region *region, *next;
874 struct list_head group_resv_regions;
875 int offset = 0;
876
877 INIT_LIST_HEAD(&group_resv_regions);
878 iommu_get_group_resv_regions(group, &group_resv_regions);
879
880 list_for_each_entry_safe(region, next, &group_resv_regions, list) {
881 offset += sysfs_emit_at(buf, offset, "0x%016llx 0x%016llx %s\n",
882 (long long)region->start,
883 (long long)(region->start +
884 region->length - 1),
885 iommu_group_resv_type_string[region->type]);
886 kfree(region);
887 }
888
889 return offset;
890}
891
892static ssize_t iommu_group_show_type(struct iommu_group *group,
893 char *buf)
894{
895 char *type = "unknown";
896
897 mutex_lock(&group->mutex);
898 if (group->default_domain) {
899 switch (group->default_domain->type) {
900 case IOMMU_DOMAIN_BLOCKED:
901 type = "blocked";
902 break;
903 case IOMMU_DOMAIN_IDENTITY:
904 type = "identity";
905 break;
906 case IOMMU_DOMAIN_UNMANAGED:
907 type = "unmanaged";
908 break;
909 case IOMMU_DOMAIN_DMA:
910 type = "DMA";
911 break;
912 case IOMMU_DOMAIN_DMA_FQ:
913 type = "DMA-FQ";
914 break;
915 }
916 }
917 mutex_unlock(&group->mutex);
918
919 return sysfs_emit(buf, "%s\n", type);
920}
921
922static IOMMU_GROUP_ATTR(name, S_IRUGO, iommu_group_show_name, NULL);
923
924static IOMMU_GROUP_ATTR(reserved_regions, 0444,
925 iommu_group_show_resv_regions, NULL);
926
927static IOMMU_GROUP_ATTR(type, 0644, iommu_group_show_type,
928 iommu_group_store_type);
929
930static void iommu_group_release(struct kobject *kobj)
931{
932 struct iommu_group *group = to_iommu_group(kobj);
933
934 pr_debug("Releasing group %d\n", group->id);
935
936 if (group->iommu_data_release)
937 group->iommu_data_release(group->iommu_data);
938
939 ida_free(&iommu_group_ida, group->id);
940
941 /* Domains are free'd by iommu_deinit_device() */
942 WARN_ON(group->default_domain);
943 WARN_ON(group->blocking_domain);
944
945 kfree(group->name);
946 kfree(group);
947}
948
949static const struct kobj_type iommu_group_ktype = {
950 .sysfs_ops = &iommu_group_sysfs_ops,
951 .release = iommu_group_release,
952};
953
954/**
955 * iommu_group_alloc - Allocate a new group
956 *
957 * This function is called by an iommu driver to allocate a new iommu
958 * group. The iommu group represents the minimum granularity of the iommu.
959 * Upon successful return, the caller holds a reference to the supplied
960 * group in order to hold the group until devices are added. Use
961 * iommu_group_put() to release this extra reference count, allowing the
962 * group to be automatically reclaimed once it has no devices or external
963 * references.
964 */
965struct iommu_group *iommu_group_alloc(void)
966{
967 struct iommu_group *group;
968 int ret;
969
970 group = kzalloc(sizeof(*group), GFP_KERNEL);
971 if (!group)
972 return ERR_PTR(-ENOMEM);
973
974 group->kobj.kset = iommu_group_kset;
975 mutex_init(&group->mutex);
976 INIT_LIST_HEAD(&group->devices);
977 INIT_LIST_HEAD(&group->entry);
978 xa_init(&group->pasid_array);
979
980 ret = ida_alloc(&iommu_group_ida, GFP_KERNEL);
981 if (ret < 0) {
982 kfree(group);
983 return ERR_PTR(ret);
984 }
985 group->id = ret;
986
987 ret = kobject_init_and_add(&group->kobj, &iommu_group_ktype,
988 NULL, "%d", group->id);
989 if (ret) {
990 kobject_put(&group->kobj);
991 return ERR_PTR(ret);
992 }
993
994 group->devices_kobj = kobject_create_and_add("devices", &group->kobj);
995 if (!group->devices_kobj) {
996 kobject_put(&group->kobj); /* triggers .release & free */
997 return ERR_PTR(-ENOMEM);
998 }
999
1000 /*
1001 * The devices_kobj holds a reference on the group kobject, so
1002 * as long as that exists so will the group. We can therefore
1003 * use the devices_kobj for reference counting.
1004 */
1005 kobject_put(&group->kobj);
1006
1007 ret = iommu_group_create_file(group,
1008 &iommu_group_attr_reserved_regions);
1009 if (ret) {
1010 kobject_put(group->devices_kobj);
1011 return ERR_PTR(ret);
1012 }
1013
1014 ret = iommu_group_create_file(group, &iommu_group_attr_type);
1015 if (ret) {
1016 kobject_put(group->devices_kobj);
1017 return ERR_PTR(ret);
1018 }
1019
1020 pr_debug("Allocated group %d\n", group->id);
1021
1022 return group;
1023}
1024EXPORT_SYMBOL_GPL(iommu_group_alloc);
1025
1026/**
1027 * iommu_group_get_iommudata - retrieve iommu_data registered for a group
1028 * @group: the group
1029 *
1030 * iommu drivers can store data in the group for use when doing iommu
1031 * operations. This function provides a way to retrieve it. Caller
1032 * should hold a group reference.
1033 */
1034void *iommu_group_get_iommudata(struct iommu_group *group)
1035{
1036 return group->iommu_data;
1037}
1038EXPORT_SYMBOL_GPL(iommu_group_get_iommudata);
1039
1040/**
1041 * iommu_group_set_iommudata - set iommu_data for a group
1042 * @group: the group
1043 * @iommu_data: new data
1044 * @release: release function for iommu_data
1045 *
1046 * iommu drivers can store data in the group for use when doing iommu
1047 * operations. This function provides a way to set the data after
1048 * the group has been allocated. Caller should hold a group reference.
1049 */
1050void iommu_group_set_iommudata(struct iommu_group *group, void *iommu_data,
1051 void (*release)(void *iommu_data))
1052{
1053 group->iommu_data = iommu_data;
1054 group->iommu_data_release = release;
1055}
1056EXPORT_SYMBOL_GPL(iommu_group_set_iommudata);
1057
1058/**
1059 * iommu_group_set_name - set name for a group
1060 * @group: the group
1061 * @name: name
1062 *
1063 * Allow iommu driver to set a name for a group. When set it will
1064 * appear in a name attribute file under the group in sysfs.
1065 */
1066int iommu_group_set_name(struct iommu_group *group, const char *name)
1067{
1068 int ret;
1069
1070 if (group->name) {
1071 iommu_group_remove_file(group, &iommu_group_attr_name);
1072 kfree(group->name);
1073 group->name = NULL;
1074 if (!name)
1075 return 0;
1076 }
1077
1078 group->name = kstrdup(name, GFP_KERNEL);
1079 if (!group->name)
1080 return -ENOMEM;
1081
1082 ret = iommu_group_create_file(group, &iommu_group_attr_name);
1083 if (ret) {
1084 kfree(group->name);
1085 group->name = NULL;
1086 return ret;
1087 }
1088
1089 return 0;
1090}
1091EXPORT_SYMBOL_GPL(iommu_group_set_name);
1092
1093static int iommu_create_device_direct_mappings(struct iommu_domain *domain,
1094 struct device *dev)
1095{
1096 struct iommu_resv_region *entry;
1097 struct list_head mappings;
1098 unsigned long pg_size;
1099 int ret = 0;
1100
1101 pg_size = domain->pgsize_bitmap ? 1UL << __ffs(domain->pgsize_bitmap) : 0;
1102 INIT_LIST_HEAD(&mappings);
1103
1104 if (WARN_ON_ONCE(iommu_is_dma_domain(domain) && !pg_size))
1105 return -EINVAL;
1106
1107 iommu_get_resv_regions(dev, &mappings);
1108
1109 /* We need to consider overlapping regions for different devices */
1110 list_for_each_entry(entry, &mappings, list) {
1111 dma_addr_t start, end, addr;
1112 size_t map_size = 0;
1113
1114 if (entry->type == IOMMU_RESV_DIRECT)
1115 dev->iommu->require_direct = 1;
1116
1117 if ((entry->type != IOMMU_RESV_DIRECT &&
1118 entry->type != IOMMU_RESV_DIRECT_RELAXABLE) ||
1119 !iommu_is_dma_domain(domain))
1120 continue;
1121
1122 start = ALIGN(entry->start, pg_size);
1123 end = ALIGN(entry->start + entry->length, pg_size);
1124
1125 for (addr = start; addr <= end; addr += pg_size) {
1126 phys_addr_t phys_addr;
1127
1128 if (addr == end)
1129 goto map_end;
1130
1131 phys_addr = iommu_iova_to_phys(domain, addr);
1132 if (!phys_addr) {
1133 map_size += pg_size;
1134 continue;
1135 }
1136
1137map_end:
1138 if (map_size) {
1139 ret = iommu_map(domain, addr - map_size,
1140 addr - map_size, map_size,
1141 entry->prot, GFP_KERNEL);
1142 if (ret)
1143 goto out;
1144 map_size = 0;
1145 }
1146 }
1147
1148 }
1149out:
1150 iommu_put_resv_regions(dev, &mappings);
1151
1152 return ret;
1153}
1154
1155/* This is undone by __iommu_group_free_device() */
1156static struct group_device *iommu_group_alloc_device(struct iommu_group *group,
1157 struct device *dev)
1158{
1159 int ret, i = 0;
1160 struct group_device *device;
1161
1162 device = kzalloc(sizeof(*device), GFP_KERNEL);
1163 if (!device)
1164 return ERR_PTR(-ENOMEM);
1165
1166 device->dev = dev;
1167
1168 ret = sysfs_create_link(&dev->kobj, &group->kobj, "iommu_group");
1169 if (ret)
1170 goto err_free_device;
1171
1172 device->name = kasprintf(GFP_KERNEL, "%s", kobject_name(&dev->kobj));
1173rename:
1174 if (!device->name) {
1175 ret = -ENOMEM;
1176 goto err_remove_link;
1177 }
1178
1179 ret = sysfs_create_link_nowarn(group->devices_kobj,
1180 &dev->kobj, device->name);
1181 if (ret) {
1182 if (ret == -EEXIST && i >= 0) {
1183 /*
1184 * Account for the slim chance of collision
1185 * and append an instance to the name.
1186 */
1187 kfree(device->name);
1188 device->name = kasprintf(GFP_KERNEL, "%s.%d",
1189 kobject_name(&dev->kobj), i++);
1190 goto rename;
1191 }
1192 goto err_free_name;
1193 }
1194
1195 trace_add_device_to_group(group->id, dev);
1196
1197 dev_info(dev, "Adding to iommu group %d\n", group->id);
1198
1199 return device;
1200
1201err_free_name:
1202 kfree(device->name);
1203err_remove_link:
1204 sysfs_remove_link(&dev->kobj, "iommu_group");
1205err_free_device:
1206 kfree(device);
1207 dev_err(dev, "Failed to add to iommu group %d: %d\n", group->id, ret);
1208 return ERR_PTR(ret);
1209}
1210
1211/**
1212 * iommu_group_add_device - add a device to an iommu group
1213 * @group: the group into which to add the device (reference should be held)
1214 * @dev: the device
1215 *
1216 * This function is called by an iommu driver to add a device into a
1217 * group. Adding a device increments the group reference count.
1218 */
1219int iommu_group_add_device(struct iommu_group *group, struct device *dev)
1220{
1221 struct group_device *gdev;
1222
1223 gdev = iommu_group_alloc_device(group, dev);
1224 if (IS_ERR(gdev))
1225 return PTR_ERR(gdev);
1226
1227 iommu_group_ref_get(group);
1228 dev->iommu_group = group;
1229
1230 mutex_lock(&group->mutex);
1231 list_add_tail(&gdev->list, &group->devices);
1232 mutex_unlock(&group->mutex);
1233 return 0;
1234}
1235EXPORT_SYMBOL_GPL(iommu_group_add_device);
1236
1237/**
1238 * iommu_group_remove_device - remove a device from it's current group
1239 * @dev: device to be removed
1240 *
1241 * This function is called by an iommu driver to remove the device from
1242 * it's current group. This decrements the iommu group reference count.
1243 */
1244void iommu_group_remove_device(struct device *dev)
1245{
1246 struct iommu_group *group = dev->iommu_group;
1247
1248 if (!group)
1249 return;
1250
1251 dev_info(dev, "Removing from iommu group %d\n", group->id);
1252
1253 __iommu_group_remove_device(dev);
1254}
1255EXPORT_SYMBOL_GPL(iommu_group_remove_device);
1256
1257#if IS_ENABLED(CONFIG_LOCKDEP) && IS_ENABLED(CONFIG_IOMMU_API)
1258/**
1259 * iommu_group_mutex_assert - Check device group mutex lock
1260 * @dev: the device that has group param set
1261 *
1262 * This function is called by an iommu driver to check whether it holds
1263 * group mutex lock for the given device or not.
1264 *
1265 * Note that this function must be called after device group param is set.
1266 */
1267void iommu_group_mutex_assert(struct device *dev)
1268{
1269 struct iommu_group *group = dev->iommu_group;
1270
1271 lockdep_assert_held(&group->mutex);
1272}
1273EXPORT_SYMBOL_GPL(iommu_group_mutex_assert);
1274#endif
1275
1276static struct device *iommu_group_first_dev(struct iommu_group *group)
1277{
1278 lockdep_assert_held(&group->mutex);
1279 return list_first_entry(&group->devices, struct group_device, list)->dev;
1280}
1281
1282/**
1283 * iommu_group_for_each_dev - iterate over each device in the group
1284 * @group: the group
1285 * @data: caller opaque data to be passed to callback function
1286 * @fn: caller supplied callback function
1287 *
1288 * This function is called by group users to iterate over group devices.
1289 * Callers should hold a reference count to the group during callback.
1290 * The group->mutex is held across callbacks, which will block calls to
1291 * iommu_group_add/remove_device.
1292 */
1293int iommu_group_for_each_dev(struct iommu_group *group, void *data,
1294 int (*fn)(struct device *, void *))
1295{
1296 struct group_device *device;
1297 int ret = 0;
1298
1299 mutex_lock(&group->mutex);
1300 for_each_group_device(group, device) {
1301 ret = fn(device->dev, data);
1302 if (ret)
1303 break;
1304 }
1305 mutex_unlock(&group->mutex);
1306
1307 return ret;
1308}
1309EXPORT_SYMBOL_GPL(iommu_group_for_each_dev);
1310
1311/**
1312 * iommu_group_get - Return the group for a device and increment reference
1313 * @dev: get the group that this device belongs to
1314 *
1315 * This function is called by iommu drivers and users to get the group
1316 * for the specified device. If found, the group is returned and the group
1317 * reference in incremented, else NULL.
1318 */
1319struct iommu_group *iommu_group_get(struct device *dev)
1320{
1321 struct iommu_group *group = dev->iommu_group;
1322
1323 if (group)
1324 kobject_get(group->devices_kobj);
1325
1326 return group;
1327}
1328EXPORT_SYMBOL_GPL(iommu_group_get);
1329
1330/**
1331 * iommu_group_ref_get - Increment reference on a group
1332 * @group: the group to use, must not be NULL
1333 *
1334 * This function is called by iommu drivers to take additional references on an
1335 * existing group. Returns the given group for convenience.
1336 */
1337struct iommu_group *iommu_group_ref_get(struct iommu_group *group)
1338{
1339 kobject_get(group->devices_kobj);
1340 return group;
1341}
1342EXPORT_SYMBOL_GPL(iommu_group_ref_get);
1343
1344/**
1345 * iommu_group_put - Decrement group reference
1346 * @group: the group to use
1347 *
1348 * This function is called by iommu drivers and users to release the
1349 * iommu group. Once the reference count is zero, the group is released.
1350 */
1351void iommu_group_put(struct iommu_group *group)
1352{
1353 if (group)
1354 kobject_put(group->devices_kobj);
1355}
1356EXPORT_SYMBOL_GPL(iommu_group_put);
1357
1358/**
1359 * iommu_group_id - Return ID for a group
1360 * @group: the group to ID
1361 *
1362 * Return the unique ID for the group matching the sysfs group number.
1363 */
1364int iommu_group_id(struct iommu_group *group)
1365{
1366 return group->id;
1367}
1368EXPORT_SYMBOL_GPL(iommu_group_id);
1369
1370static struct iommu_group *get_pci_alias_group(struct pci_dev *pdev,
1371 unsigned long *devfns);
1372
1373/*
1374 * To consider a PCI device isolated, we require ACS to support Source
1375 * Validation, Request Redirection, Completer Redirection, and Upstream
1376 * Forwarding. This effectively means that devices cannot spoof their
1377 * requester ID, requests and completions cannot be redirected, and all
1378 * transactions are forwarded upstream, even as it passes through a
1379 * bridge where the target device is downstream.
1380 */
1381#define REQ_ACS_FLAGS (PCI_ACS_SV | PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_UF)
1382
1383/*
1384 * For multifunction devices which are not isolated from each other, find
1385 * all the other non-isolated functions and look for existing groups. For
1386 * each function, we also need to look for aliases to or from other devices
1387 * that may already have a group.
1388 */
1389static struct iommu_group *get_pci_function_alias_group(struct pci_dev *pdev,
1390 unsigned long *devfns)
1391{
1392 struct pci_dev *tmp = NULL;
1393 struct iommu_group *group;
1394
1395 if (!pdev->multifunction || pci_acs_enabled(pdev, REQ_ACS_FLAGS))
1396 return NULL;
1397
1398 for_each_pci_dev(tmp) {
1399 if (tmp == pdev || tmp->bus != pdev->bus ||
1400 PCI_SLOT(tmp->devfn) != PCI_SLOT(pdev->devfn) ||
1401 pci_acs_enabled(tmp, REQ_ACS_FLAGS))
1402 continue;
1403
1404 group = get_pci_alias_group(tmp, devfns);
1405 if (group) {
1406 pci_dev_put(tmp);
1407 return group;
1408 }
1409 }
1410
1411 return NULL;
1412}
1413
1414/*
1415 * Look for aliases to or from the given device for existing groups. DMA
1416 * aliases are only supported on the same bus, therefore the search
1417 * space is quite small (especially since we're really only looking at pcie
1418 * device, and therefore only expect multiple slots on the root complex or
1419 * downstream switch ports). It's conceivable though that a pair of
1420 * multifunction devices could have aliases between them that would cause a
1421 * loop. To prevent this, we use a bitmap to track where we've been.
1422 */
1423static struct iommu_group *get_pci_alias_group(struct pci_dev *pdev,
1424 unsigned long *devfns)
1425{
1426 struct pci_dev *tmp = NULL;
1427 struct iommu_group *group;
1428
1429 if (test_and_set_bit(pdev->devfn & 0xff, devfns))
1430 return NULL;
1431
1432 group = iommu_group_get(&pdev->dev);
1433 if (group)
1434 return group;
1435
1436 for_each_pci_dev(tmp) {
1437 if (tmp == pdev || tmp->bus != pdev->bus)
1438 continue;
1439
1440 /* We alias them or they alias us */
1441 if (pci_devs_are_dma_aliases(pdev, tmp)) {
1442 group = get_pci_alias_group(tmp, devfns);
1443 if (group) {
1444 pci_dev_put(tmp);
1445 return group;
1446 }
1447
1448 group = get_pci_function_alias_group(tmp, devfns);
1449 if (group) {
1450 pci_dev_put(tmp);
1451 return group;
1452 }
1453 }
1454 }
1455
1456 return NULL;
1457}
1458
1459struct group_for_pci_data {
1460 struct pci_dev *pdev;
1461 struct iommu_group *group;
1462};
1463
1464/*
1465 * DMA alias iterator callback, return the last seen device. Stop and return
1466 * the IOMMU group if we find one along the way.
1467 */
1468static int get_pci_alias_or_group(struct pci_dev *pdev, u16 alias, void *opaque)
1469{
1470 struct group_for_pci_data *data = opaque;
1471
1472 data->pdev = pdev;
1473 data->group = iommu_group_get(&pdev->dev);
1474
1475 return data->group != NULL;
1476}
1477
1478/*
1479 * Generic device_group call-back function. It just allocates one
1480 * iommu-group per device.
1481 */
1482struct iommu_group *generic_device_group(struct device *dev)
1483{
1484 return iommu_group_alloc();
1485}
1486EXPORT_SYMBOL_GPL(generic_device_group);
1487
1488/*
1489 * Generic device_group call-back function. It just allocates one
1490 * iommu-group per iommu driver instance shared by every device
1491 * probed by that iommu driver.
1492 */
1493struct iommu_group *generic_single_device_group(struct device *dev)
1494{
1495 struct iommu_device *iommu = dev->iommu->iommu_dev;
1496
1497 if (!iommu->singleton_group) {
1498 struct iommu_group *group;
1499
1500 group = iommu_group_alloc();
1501 if (IS_ERR(group))
1502 return group;
1503 iommu->singleton_group = group;
1504 }
1505 return iommu_group_ref_get(iommu->singleton_group);
1506}
1507EXPORT_SYMBOL_GPL(generic_single_device_group);
1508
1509/*
1510 * Use standard PCI bus topology, isolation features, and DMA alias quirks
1511 * to find or create an IOMMU group for a device.
1512 */
1513struct iommu_group *pci_device_group(struct device *dev)
1514{
1515 struct pci_dev *pdev = to_pci_dev(dev);
1516 struct group_for_pci_data data;
1517 struct pci_bus *bus;
1518 struct iommu_group *group = NULL;
1519 u64 devfns[4] = { 0 };
1520
1521 if (WARN_ON(!dev_is_pci(dev)))
1522 return ERR_PTR(-EINVAL);
1523
1524 /*
1525 * Find the upstream DMA alias for the device. A device must not
1526 * be aliased due to topology in order to have its own IOMMU group.
1527 * If we find an alias along the way that already belongs to a
1528 * group, use it.
1529 */
1530 if (pci_for_each_dma_alias(pdev, get_pci_alias_or_group, &data))
1531 return data.group;
1532
1533 pdev = data.pdev;
1534
1535 /*
1536 * Continue upstream from the point of minimum IOMMU granularity
1537 * due to aliases to the point where devices are protected from
1538 * peer-to-peer DMA by PCI ACS. Again, if we find an existing
1539 * group, use it.
1540 */
1541 for (bus = pdev->bus; !pci_is_root_bus(bus); bus = bus->parent) {
1542 if (!bus->self)
1543 continue;
1544
1545 if (pci_acs_path_enabled(bus->self, NULL, REQ_ACS_FLAGS))
1546 break;
1547
1548 pdev = bus->self;
1549
1550 group = iommu_group_get(&pdev->dev);
1551 if (group)
1552 return group;
1553 }
1554
1555 /*
1556 * Look for existing groups on device aliases. If we alias another
1557 * device or another device aliases us, use the same group.
1558 */
1559 group = get_pci_alias_group(pdev, (unsigned long *)devfns);
1560 if (group)
1561 return group;
1562
1563 /*
1564 * Look for existing groups on non-isolated functions on the same
1565 * slot and aliases of those funcions, if any. No need to clear
1566 * the search bitmap, the tested devfns are still valid.
1567 */
1568 group = get_pci_function_alias_group(pdev, (unsigned long *)devfns);
1569 if (group)
1570 return group;
1571
1572 /* No shared group found, allocate new */
1573 return iommu_group_alloc();
1574}
1575EXPORT_SYMBOL_GPL(pci_device_group);
1576
1577/* Get the IOMMU group for device on fsl-mc bus */
1578struct iommu_group *fsl_mc_device_group(struct device *dev)
1579{
1580 struct device *cont_dev = fsl_mc_cont_dev(dev);
1581 struct iommu_group *group;
1582
1583 group = iommu_group_get(cont_dev);
1584 if (!group)
1585 group = iommu_group_alloc();
1586 return group;
1587}
1588EXPORT_SYMBOL_GPL(fsl_mc_device_group);
1589
1590static struct iommu_domain *__iommu_alloc_identity_domain(struct device *dev)
1591{
1592 const struct iommu_ops *ops = dev_iommu_ops(dev);
1593 struct iommu_domain *domain;
1594
1595 if (ops->identity_domain)
1596 return ops->identity_domain;
1597
1598 /* Older drivers create the identity domain via ops->domain_alloc() */
1599 if (!ops->domain_alloc)
1600 return ERR_PTR(-EOPNOTSUPP);
1601
1602 domain = ops->domain_alloc(IOMMU_DOMAIN_IDENTITY);
1603 if (IS_ERR(domain))
1604 return domain;
1605 if (!domain)
1606 return ERR_PTR(-ENOMEM);
1607
1608 iommu_domain_init(domain, IOMMU_DOMAIN_IDENTITY, ops);
1609 return domain;
1610}
1611
1612static struct iommu_domain *
1613__iommu_group_alloc_default_domain(struct iommu_group *group, int req_type)
1614{
1615 struct device *dev = iommu_group_first_dev(group);
1616 struct iommu_domain *dom;
1617
1618 if (group->default_domain && group->default_domain->type == req_type)
1619 return group->default_domain;
1620
1621 /*
1622 * When allocating the DMA API domain assume that the driver is going to
1623 * use PASID and make sure the RID's domain is PASID compatible.
1624 */
1625 if (req_type & __IOMMU_DOMAIN_PAGING) {
1626 dom = __iommu_paging_domain_alloc_flags(dev, req_type,
1627 dev->iommu->max_pasids ? IOMMU_HWPT_ALLOC_PASID : 0);
1628
1629 /*
1630 * If driver does not support PASID feature then
1631 * try to allocate non-PASID domain
1632 */
1633 if (PTR_ERR(dom) == -EOPNOTSUPP)
1634 dom = __iommu_paging_domain_alloc_flags(dev, req_type, 0);
1635
1636 return dom;
1637 }
1638
1639 if (req_type == IOMMU_DOMAIN_IDENTITY)
1640 return __iommu_alloc_identity_domain(dev);
1641
1642 return ERR_PTR(-EINVAL);
1643}
1644
1645/*
1646 * req_type of 0 means "auto" which means to select a domain based on
1647 * iommu_def_domain_type or what the driver actually supports.
1648 */
1649static struct iommu_domain *
1650iommu_group_alloc_default_domain(struct iommu_group *group, int req_type)
1651{
1652 const struct iommu_ops *ops = dev_iommu_ops(iommu_group_first_dev(group));
1653 struct iommu_domain *dom;
1654
1655 lockdep_assert_held(&group->mutex);
1656
1657 /*
1658 * Allow legacy drivers to specify the domain that will be the default
1659 * domain. This should always be either an IDENTITY/BLOCKED/PLATFORM
1660 * domain. Do not use in new drivers.
1661 */
1662 if (ops->default_domain) {
1663 if (req_type != ops->default_domain->type)
1664 return ERR_PTR(-EINVAL);
1665 return ops->default_domain;
1666 }
1667
1668 if (req_type)
1669 return __iommu_group_alloc_default_domain(group, req_type);
1670
1671 /* The driver gave no guidance on what type to use, try the default */
1672 dom = __iommu_group_alloc_default_domain(group, iommu_def_domain_type);
1673 if (!IS_ERR(dom))
1674 return dom;
1675
1676 /* Otherwise IDENTITY and DMA_FQ defaults will try DMA */
1677 if (iommu_def_domain_type == IOMMU_DOMAIN_DMA)
1678 return ERR_PTR(-EINVAL);
1679 dom = __iommu_group_alloc_default_domain(group, IOMMU_DOMAIN_DMA);
1680 if (IS_ERR(dom))
1681 return dom;
1682
1683 pr_warn("Failed to allocate default IOMMU domain of type %u for group %s - Falling back to IOMMU_DOMAIN_DMA",
1684 iommu_def_domain_type, group->name);
1685 return dom;
1686}
1687
1688struct iommu_domain *iommu_group_default_domain(struct iommu_group *group)
1689{
1690 return group->default_domain;
1691}
1692
1693static int probe_iommu_group(struct device *dev, void *data)
1694{
1695 struct list_head *group_list = data;
1696 int ret;
1697
1698 mutex_lock(&iommu_probe_device_lock);
1699 ret = __iommu_probe_device(dev, group_list);
1700 mutex_unlock(&iommu_probe_device_lock);
1701 if (ret == -ENODEV)
1702 ret = 0;
1703
1704 return ret;
1705}
1706
1707static int iommu_bus_notifier(struct notifier_block *nb,
1708 unsigned long action, void *data)
1709{
1710 struct device *dev = data;
1711
1712 if (action == BUS_NOTIFY_ADD_DEVICE) {
1713 int ret;
1714
1715 ret = iommu_probe_device(dev);
1716 return (ret) ? NOTIFY_DONE : NOTIFY_OK;
1717 } else if (action == BUS_NOTIFY_REMOVED_DEVICE) {
1718 iommu_release_device(dev);
1719 return NOTIFY_OK;
1720 }
1721
1722 return 0;
1723}
1724
1725/*
1726 * Combine the driver's chosen def_domain_type across all the devices in a
1727 * group. Drivers must give a consistent result.
1728 */
1729static int iommu_get_def_domain_type(struct iommu_group *group,
1730 struct device *dev, int cur_type)
1731{
1732 const struct iommu_ops *ops = dev_iommu_ops(dev);
1733 int type;
1734
1735 if (ops->default_domain) {
1736 /*
1737 * Drivers that declare a global static default_domain will
1738 * always choose that.
1739 */
1740 type = ops->default_domain->type;
1741 } else {
1742 if (ops->def_domain_type)
1743 type = ops->def_domain_type(dev);
1744 else
1745 return cur_type;
1746 }
1747 if (!type || cur_type == type)
1748 return cur_type;
1749 if (!cur_type)
1750 return type;
1751
1752 dev_err_ratelimited(
1753 dev,
1754 "IOMMU driver error, requesting conflicting def_domain_type, %s and %s, for devices in group %u.\n",
1755 iommu_domain_type_str(cur_type), iommu_domain_type_str(type),
1756 group->id);
1757
1758 /*
1759 * Try to recover, drivers are allowed to force IDENITY or DMA, IDENTITY
1760 * takes precedence.
1761 */
1762 if (type == IOMMU_DOMAIN_IDENTITY)
1763 return type;
1764 return cur_type;
1765}
1766
1767/*
1768 * A target_type of 0 will select the best domain type. 0 can be returned in
1769 * this case meaning the global default should be used.
1770 */
1771static int iommu_get_default_domain_type(struct iommu_group *group,
1772 int target_type)
1773{
1774 struct device *untrusted = NULL;
1775 struct group_device *gdev;
1776 int driver_type = 0;
1777
1778 lockdep_assert_held(&group->mutex);
1779
1780 /*
1781 * ARM32 drivers supporting CONFIG_ARM_DMA_USE_IOMMU can declare an
1782 * identity_domain and it will automatically become their default
1783 * domain. Later on ARM_DMA_USE_IOMMU will install its UNMANAGED domain.
1784 * Override the selection to IDENTITY.
1785 */
1786 if (IS_ENABLED(CONFIG_ARM_DMA_USE_IOMMU)) {
1787 static_assert(!(IS_ENABLED(CONFIG_ARM_DMA_USE_IOMMU) &&
1788 IS_ENABLED(CONFIG_IOMMU_DMA)));
1789 driver_type = IOMMU_DOMAIN_IDENTITY;
1790 }
1791
1792 for_each_group_device(group, gdev) {
1793 driver_type = iommu_get_def_domain_type(group, gdev->dev,
1794 driver_type);
1795
1796 if (dev_is_pci(gdev->dev) && to_pci_dev(gdev->dev)->untrusted) {
1797 /*
1798 * No ARM32 using systems will set untrusted, it cannot
1799 * work.
1800 */
1801 if (WARN_ON(IS_ENABLED(CONFIG_ARM_DMA_USE_IOMMU)))
1802 return -1;
1803 untrusted = gdev->dev;
1804 }
1805 }
1806
1807 /*
1808 * If the common dma ops are not selected in kconfig then we cannot use
1809 * IOMMU_DOMAIN_DMA at all. Force IDENTITY if nothing else has been
1810 * selected.
1811 */
1812 if (!IS_ENABLED(CONFIG_IOMMU_DMA)) {
1813 if (WARN_ON(driver_type == IOMMU_DOMAIN_DMA))
1814 return -1;
1815 if (!driver_type)
1816 driver_type = IOMMU_DOMAIN_IDENTITY;
1817 }
1818
1819 if (untrusted) {
1820 if (driver_type && driver_type != IOMMU_DOMAIN_DMA) {
1821 dev_err_ratelimited(
1822 untrusted,
1823 "Device is not trusted, but driver is overriding group %u to %s, refusing to probe.\n",
1824 group->id, iommu_domain_type_str(driver_type));
1825 return -1;
1826 }
1827 driver_type = IOMMU_DOMAIN_DMA;
1828 }
1829
1830 if (target_type) {
1831 if (driver_type && target_type != driver_type)
1832 return -1;
1833 return target_type;
1834 }
1835 return driver_type;
1836}
1837
1838static void iommu_group_do_probe_finalize(struct device *dev)
1839{
1840 const struct iommu_ops *ops = dev_iommu_ops(dev);
1841
1842 if (ops->probe_finalize)
1843 ops->probe_finalize(dev);
1844}
1845
1846static int bus_iommu_probe(const struct bus_type *bus)
1847{
1848 struct iommu_group *group, *next;
1849 LIST_HEAD(group_list);
1850 int ret;
1851
1852 ret = bus_for_each_dev(bus, NULL, &group_list, probe_iommu_group);
1853 if (ret)
1854 return ret;
1855
1856 list_for_each_entry_safe(group, next, &group_list, entry) {
1857 struct group_device *gdev;
1858
1859 mutex_lock(&group->mutex);
1860
1861 /* Remove item from the list */
1862 list_del_init(&group->entry);
1863
1864 /*
1865 * We go to the trouble of deferred default domain creation so
1866 * that the cross-group default domain type and the setup of the
1867 * IOMMU_RESV_DIRECT will work correctly in non-hotpug scenarios.
1868 */
1869 ret = iommu_setup_default_domain(group, 0);
1870 if (ret) {
1871 mutex_unlock(&group->mutex);
1872 return ret;
1873 }
1874 for_each_group_device(group, gdev)
1875 iommu_setup_dma_ops(gdev->dev);
1876 mutex_unlock(&group->mutex);
1877
1878 /*
1879 * FIXME: Mis-locked because the ops->probe_finalize() call-back
1880 * of some IOMMU drivers calls arm_iommu_attach_device() which
1881 * in-turn might call back into IOMMU core code, where it tries
1882 * to take group->mutex, resulting in a deadlock.
1883 */
1884 for_each_group_device(group, gdev)
1885 iommu_group_do_probe_finalize(gdev->dev);
1886 }
1887
1888 return 0;
1889}
1890
1891/**
1892 * device_iommu_capable() - check for a general IOMMU capability
1893 * @dev: device to which the capability would be relevant, if available
1894 * @cap: IOMMU capability
1895 *
1896 * Return: true if an IOMMU is present and supports the given capability
1897 * for the given device, otherwise false.
1898 */
1899bool device_iommu_capable(struct device *dev, enum iommu_cap cap)
1900{
1901 const struct iommu_ops *ops;
1902
1903 if (!dev_has_iommu(dev))
1904 return false;
1905
1906 ops = dev_iommu_ops(dev);
1907 if (!ops->capable)
1908 return false;
1909
1910 return ops->capable(dev, cap);
1911}
1912EXPORT_SYMBOL_GPL(device_iommu_capable);
1913
1914/**
1915 * iommu_group_has_isolated_msi() - Compute msi_device_has_isolated_msi()
1916 * for a group
1917 * @group: Group to query
1918 *
1919 * IOMMU groups should not have differing values of
1920 * msi_device_has_isolated_msi() for devices in a group. However nothing
1921 * directly prevents this, so ensure mistakes don't result in isolation failures
1922 * by checking that all the devices are the same.
1923 */
1924bool iommu_group_has_isolated_msi(struct iommu_group *group)
1925{
1926 struct group_device *group_dev;
1927 bool ret = true;
1928
1929 mutex_lock(&group->mutex);
1930 for_each_group_device(group, group_dev)
1931 ret &= msi_device_has_isolated_msi(group_dev->dev);
1932 mutex_unlock(&group->mutex);
1933 return ret;
1934}
1935EXPORT_SYMBOL_GPL(iommu_group_has_isolated_msi);
1936
1937/**
1938 * iommu_set_fault_handler() - set a fault handler for an iommu domain
1939 * @domain: iommu domain
1940 * @handler: fault handler
1941 * @token: user data, will be passed back to the fault handler
1942 *
1943 * This function should be used by IOMMU users which want to be notified
1944 * whenever an IOMMU fault happens.
1945 *
1946 * The fault handler itself should return 0 on success, and an appropriate
1947 * error code otherwise.
1948 */
1949void iommu_set_fault_handler(struct iommu_domain *domain,
1950 iommu_fault_handler_t handler,
1951 void *token)
1952{
1953 BUG_ON(!domain);
1954
1955 domain->handler = handler;
1956 domain->handler_token = token;
1957}
1958EXPORT_SYMBOL_GPL(iommu_set_fault_handler);
1959
1960static void iommu_domain_init(struct iommu_domain *domain, unsigned int type,
1961 const struct iommu_ops *ops)
1962{
1963 domain->type = type;
1964 domain->owner = ops;
1965 if (!domain->ops)
1966 domain->ops = ops->default_domain_ops;
1967
1968 /*
1969 * If not already set, assume all sizes by default; the driver
1970 * may override this later
1971 */
1972 if (!domain->pgsize_bitmap)
1973 domain->pgsize_bitmap = ops->pgsize_bitmap;
1974}
1975
1976static struct iommu_domain *
1977__iommu_paging_domain_alloc_flags(struct device *dev, unsigned int type,
1978 unsigned int flags)
1979{
1980 const struct iommu_ops *ops;
1981 struct iommu_domain *domain;
1982
1983 if (!dev_has_iommu(dev))
1984 return ERR_PTR(-ENODEV);
1985
1986 ops = dev_iommu_ops(dev);
1987
1988 if (ops->domain_alloc_paging && !flags)
1989 domain = ops->domain_alloc_paging(dev);
1990 else if (ops->domain_alloc_paging_flags)
1991 domain = ops->domain_alloc_paging_flags(dev, flags, NULL);
1992 else if (ops->domain_alloc && !flags)
1993 domain = ops->domain_alloc(IOMMU_DOMAIN_UNMANAGED);
1994 else
1995 return ERR_PTR(-EOPNOTSUPP);
1996
1997 if (IS_ERR(domain))
1998 return domain;
1999 if (!domain)
2000 return ERR_PTR(-ENOMEM);
2001
2002 iommu_domain_init(domain, type, ops);
2003 return domain;
2004}
2005
2006/**
2007 * iommu_paging_domain_alloc_flags() - Allocate a paging domain
2008 * @dev: device for which the domain is allocated
2009 * @flags: Bitmap of iommufd_hwpt_alloc_flags
2010 *
2011 * Allocate a paging domain which will be managed by a kernel driver. Return
2012 * allocated domain if successful, or an ERR pointer for failure.
2013 */
2014struct iommu_domain *iommu_paging_domain_alloc_flags(struct device *dev,
2015 unsigned int flags)
2016{
2017 return __iommu_paging_domain_alloc_flags(dev,
2018 IOMMU_DOMAIN_UNMANAGED, flags);
2019}
2020EXPORT_SYMBOL_GPL(iommu_paging_domain_alloc_flags);
2021
2022void iommu_domain_free(struct iommu_domain *domain)
2023{
2024 if (domain->type == IOMMU_DOMAIN_SVA)
2025 mmdrop(domain->mm);
2026 iommu_put_dma_cookie(domain);
2027 if (domain->ops->free)
2028 domain->ops->free(domain);
2029}
2030EXPORT_SYMBOL_GPL(iommu_domain_free);
2031
2032/*
2033 * Put the group's domain back to the appropriate core-owned domain - either the
2034 * standard kernel-mode DMA configuration or an all-DMA-blocked domain.
2035 */
2036static void __iommu_group_set_core_domain(struct iommu_group *group)
2037{
2038 struct iommu_domain *new_domain;
2039
2040 if (group->owner)
2041 new_domain = group->blocking_domain;
2042 else
2043 new_domain = group->default_domain;
2044
2045 __iommu_group_set_domain_nofail(group, new_domain);
2046}
2047
2048static int __iommu_attach_device(struct iommu_domain *domain,
2049 struct device *dev)
2050{
2051 int ret;
2052
2053 if (unlikely(domain->ops->attach_dev == NULL))
2054 return -ENODEV;
2055
2056 ret = domain->ops->attach_dev(domain, dev);
2057 if (ret)
2058 return ret;
2059 dev->iommu->attach_deferred = 0;
2060 trace_attach_device_to_domain(dev);
2061 return 0;
2062}
2063
2064/**
2065 * iommu_attach_device - Attach an IOMMU domain to a device
2066 * @domain: IOMMU domain to attach
2067 * @dev: Device that will be attached
2068 *
2069 * Returns 0 on success and error code on failure
2070 *
2071 * Note that EINVAL can be treated as a soft failure, indicating
2072 * that certain configuration of the domain is incompatible with
2073 * the device. In this case attaching a different domain to the
2074 * device may succeed.
2075 */
2076int iommu_attach_device(struct iommu_domain *domain, struct device *dev)
2077{
2078 /* Caller must be a probed driver on dev */
2079 struct iommu_group *group = dev->iommu_group;
2080 int ret;
2081
2082 if (!group)
2083 return -ENODEV;
2084
2085 /*
2086 * Lock the group to make sure the device-count doesn't
2087 * change while we are attaching
2088 */
2089 mutex_lock(&group->mutex);
2090 ret = -EINVAL;
2091 if (list_count_nodes(&group->devices) != 1)
2092 goto out_unlock;
2093
2094 ret = __iommu_attach_group(domain, group);
2095
2096out_unlock:
2097 mutex_unlock(&group->mutex);
2098 return ret;
2099}
2100EXPORT_SYMBOL_GPL(iommu_attach_device);
2101
2102int iommu_deferred_attach(struct device *dev, struct iommu_domain *domain)
2103{
2104 if (dev->iommu && dev->iommu->attach_deferred)
2105 return __iommu_attach_device(domain, dev);
2106
2107 return 0;
2108}
2109
2110void iommu_detach_device(struct iommu_domain *domain, struct device *dev)
2111{
2112 /* Caller must be a probed driver on dev */
2113 struct iommu_group *group = dev->iommu_group;
2114
2115 if (!group)
2116 return;
2117
2118 mutex_lock(&group->mutex);
2119 if (WARN_ON(domain != group->domain) ||
2120 WARN_ON(list_count_nodes(&group->devices) != 1))
2121 goto out_unlock;
2122 __iommu_group_set_core_domain(group);
2123
2124out_unlock:
2125 mutex_unlock(&group->mutex);
2126}
2127EXPORT_SYMBOL_GPL(iommu_detach_device);
2128
2129struct iommu_domain *iommu_get_domain_for_dev(struct device *dev)
2130{
2131 /* Caller must be a probed driver on dev */
2132 struct iommu_group *group = dev->iommu_group;
2133
2134 if (!group)
2135 return NULL;
2136
2137 return group->domain;
2138}
2139EXPORT_SYMBOL_GPL(iommu_get_domain_for_dev);
2140
2141/*
2142 * For IOMMU_DOMAIN_DMA implementations which already provide their own
2143 * guarantees that the group and its default domain are valid and correct.
2144 */
2145struct iommu_domain *iommu_get_dma_domain(struct device *dev)
2146{
2147 return dev->iommu_group->default_domain;
2148}
2149
2150static int __iommu_attach_group(struct iommu_domain *domain,
2151 struct iommu_group *group)
2152{
2153 struct device *dev;
2154
2155 if (group->domain && group->domain != group->default_domain &&
2156 group->domain != group->blocking_domain)
2157 return -EBUSY;
2158
2159 dev = iommu_group_first_dev(group);
2160 if (!dev_has_iommu(dev) || dev_iommu_ops(dev) != domain->owner)
2161 return -EINVAL;
2162
2163 return __iommu_group_set_domain(group, domain);
2164}
2165
2166/**
2167 * iommu_attach_group - Attach an IOMMU domain to an IOMMU group
2168 * @domain: IOMMU domain to attach
2169 * @group: IOMMU group that will be attached
2170 *
2171 * Returns 0 on success and error code on failure
2172 *
2173 * Note that EINVAL can be treated as a soft failure, indicating
2174 * that certain configuration of the domain is incompatible with
2175 * the group. In this case attaching a different domain to the
2176 * group may succeed.
2177 */
2178int iommu_attach_group(struct iommu_domain *domain, struct iommu_group *group)
2179{
2180 int ret;
2181
2182 mutex_lock(&group->mutex);
2183 ret = __iommu_attach_group(domain, group);
2184 mutex_unlock(&group->mutex);
2185
2186 return ret;
2187}
2188EXPORT_SYMBOL_GPL(iommu_attach_group);
2189
2190/**
2191 * iommu_group_replace_domain - replace the domain that a group is attached to
2192 * @group: IOMMU group that will be attached to the new domain
2193 * @new_domain: new IOMMU domain to replace with
2194 *
2195 * This API allows the group to switch domains without being forced to go to
2196 * the blocking domain in-between.
2197 *
2198 * If the currently attached domain is a core domain (e.g. a default_domain),
2199 * it will act just like the iommu_attach_group().
2200 */
2201int iommu_group_replace_domain(struct iommu_group *group,
2202 struct iommu_domain *new_domain)
2203{
2204 int ret;
2205
2206 if (!new_domain)
2207 return -EINVAL;
2208
2209 mutex_lock(&group->mutex);
2210 ret = __iommu_group_set_domain(group, new_domain);
2211 mutex_unlock(&group->mutex);
2212 return ret;
2213}
2214EXPORT_SYMBOL_NS_GPL(iommu_group_replace_domain, "IOMMUFD_INTERNAL");
2215
2216static int __iommu_device_set_domain(struct iommu_group *group,
2217 struct device *dev,
2218 struct iommu_domain *new_domain,
2219 unsigned int flags)
2220{
2221 int ret;
2222
2223 /*
2224 * If the device requires IOMMU_RESV_DIRECT then we cannot allow
2225 * the blocking domain to be attached as it does not contain the
2226 * required 1:1 mapping. This test effectively excludes the device
2227 * being used with iommu_group_claim_dma_owner() which will block
2228 * vfio and iommufd as well.
2229 */
2230 if (dev->iommu->require_direct &&
2231 (new_domain->type == IOMMU_DOMAIN_BLOCKED ||
2232 new_domain == group->blocking_domain)) {
2233 dev_warn(dev,
2234 "Firmware has requested this device have a 1:1 IOMMU mapping, rejecting configuring the device without a 1:1 mapping. Contact your platform vendor.\n");
2235 return -EINVAL;
2236 }
2237
2238 if (dev->iommu->attach_deferred) {
2239 if (new_domain == group->default_domain)
2240 return 0;
2241 dev->iommu->attach_deferred = 0;
2242 }
2243
2244 ret = __iommu_attach_device(new_domain, dev);
2245 if (ret) {
2246 /*
2247 * If we have a blocking domain then try to attach that in hopes
2248 * of avoiding a UAF. Modern drivers should implement blocking
2249 * domains as global statics that cannot fail.
2250 */
2251 if ((flags & IOMMU_SET_DOMAIN_MUST_SUCCEED) &&
2252 group->blocking_domain &&
2253 group->blocking_domain != new_domain)
2254 __iommu_attach_device(group->blocking_domain, dev);
2255 return ret;
2256 }
2257 return 0;
2258}
2259
2260/*
2261 * If 0 is returned the group's domain is new_domain. If an error is returned
2262 * then the group's domain will be set back to the existing domain unless
2263 * IOMMU_SET_DOMAIN_MUST_SUCCEED, otherwise an error is returned and the group's
2264 * domains is left inconsistent. This is a driver bug to fail attach with a
2265 * previously good domain. We try to avoid a kernel UAF because of this.
2266 *
2267 * IOMMU groups are really the natural working unit of the IOMMU, but the IOMMU
2268 * API works on domains and devices. Bridge that gap by iterating over the
2269 * devices in a group. Ideally we'd have a single device which represents the
2270 * requestor ID of the group, but we also allow IOMMU drivers to create policy
2271 * defined minimum sets, where the physical hardware may be able to distiguish
2272 * members, but we wish to group them at a higher level (ex. untrusted
2273 * multi-function PCI devices). Thus we attach each device.
2274 */
2275static int __iommu_group_set_domain_internal(struct iommu_group *group,
2276 struct iommu_domain *new_domain,
2277 unsigned int flags)
2278{
2279 struct group_device *last_gdev;
2280 struct group_device *gdev;
2281 int result;
2282 int ret;
2283
2284 lockdep_assert_held(&group->mutex);
2285
2286 if (group->domain == new_domain)
2287 return 0;
2288
2289 if (WARN_ON(!new_domain))
2290 return -EINVAL;
2291
2292 /*
2293 * Changing the domain is done by calling attach_dev() on the new
2294 * domain. This switch does not have to be atomic and DMA can be
2295 * discarded during the transition. DMA must only be able to access
2296 * either new_domain or group->domain, never something else.
2297 */
2298 result = 0;
2299 for_each_group_device(group, gdev) {
2300 ret = __iommu_device_set_domain(group, gdev->dev, new_domain,
2301 flags);
2302 if (ret) {
2303 result = ret;
2304 /*
2305 * Keep trying the other devices in the group. If a
2306 * driver fails attach to an otherwise good domain, and
2307 * does not support blocking domains, it should at least
2308 * drop its reference on the current domain so we don't
2309 * UAF.
2310 */
2311 if (flags & IOMMU_SET_DOMAIN_MUST_SUCCEED)
2312 continue;
2313 goto err_revert;
2314 }
2315 }
2316 group->domain = new_domain;
2317 return result;
2318
2319err_revert:
2320 /*
2321 * This is called in error unwind paths. A well behaved driver should
2322 * always allow us to attach to a domain that was already attached.
2323 */
2324 last_gdev = gdev;
2325 for_each_group_device(group, gdev) {
2326 /*
2327 * A NULL domain can happen only for first probe, in which case
2328 * we leave group->domain as NULL and let release clean
2329 * everything up.
2330 */
2331 if (group->domain)
2332 WARN_ON(__iommu_device_set_domain(
2333 group, gdev->dev, group->domain,
2334 IOMMU_SET_DOMAIN_MUST_SUCCEED));
2335 if (gdev == last_gdev)
2336 break;
2337 }
2338 return ret;
2339}
2340
2341void iommu_detach_group(struct iommu_domain *domain, struct iommu_group *group)
2342{
2343 mutex_lock(&group->mutex);
2344 __iommu_group_set_core_domain(group);
2345 mutex_unlock(&group->mutex);
2346}
2347EXPORT_SYMBOL_GPL(iommu_detach_group);
2348
2349phys_addr_t iommu_iova_to_phys(struct iommu_domain *domain, dma_addr_t iova)
2350{
2351 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2352 return iova;
2353
2354 if (domain->type == IOMMU_DOMAIN_BLOCKED)
2355 return 0;
2356
2357 return domain->ops->iova_to_phys(domain, iova);
2358}
2359EXPORT_SYMBOL_GPL(iommu_iova_to_phys);
2360
2361static size_t iommu_pgsize(struct iommu_domain *domain, unsigned long iova,
2362 phys_addr_t paddr, size_t size, size_t *count)
2363{
2364 unsigned int pgsize_idx, pgsize_idx_next;
2365 unsigned long pgsizes;
2366 size_t offset, pgsize, pgsize_next;
2367 unsigned long addr_merge = paddr | iova;
2368
2369 /* Page sizes supported by the hardware and small enough for @size */
2370 pgsizes = domain->pgsize_bitmap & GENMASK(__fls(size), 0);
2371
2372 /* Constrain the page sizes further based on the maximum alignment */
2373 if (likely(addr_merge))
2374 pgsizes &= GENMASK(__ffs(addr_merge), 0);
2375
2376 /* Make sure we have at least one suitable page size */
2377 BUG_ON(!pgsizes);
2378
2379 /* Pick the biggest page size remaining */
2380 pgsize_idx = __fls(pgsizes);
2381 pgsize = BIT(pgsize_idx);
2382 if (!count)
2383 return pgsize;
2384
2385 /* Find the next biggest support page size, if it exists */
2386 pgsizes = domain->pgsize_bitmap & ~GENMASK(pgsize_idx, 0);
2387 if (!pgsizes)
2388 goto out_set_count;
2389
2390 pgsize_idx_next = __ffs(pgsizes);
2391 pgsize_next = BIT(pgsize_idx_next);
2392
2393 /*
2394 * There's no point trying a bigger page size unless the virtual
2395 * and physical addresses are similarly offset within the larger page.
2396 */
2397 if ((iova ^ paddr) & (pgsize_next - 1))
2398 goto out_set_count;
2399
2400 /* Calculate the offset to the next page size alignment boundary */
2401 offset = pgsize_next - (addr_merge & (pgsize_next - 1));
2402
2403 /*
2404 * If size is big enough to accommodate the larger page, reduce
2405 * the number of smaller pages.
2406 */
2407 if (offset + pgsize_next <= size)
2408 size = offset;
2409
2410out_set_count:
2411 *count = size >> pgsize_idx;
2412 return pgsize;
2413}
2414
2415static int __iommu_map(struct iommu_domain *domain, unsigned long iova,
2416 phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
2417{
2418 const struct iommu_domain_ops *ops = domain->ops;
2419 unsigned long orig_iova = iova;
2420 unsigned int min_pagesz;
2421 size_t orig_size = size;
2422 phys_addr_t orig_paddr = paddr;
2423 int ret = 0;
2424
2425 if (unlikely(!(domain->type & __IOMMU_DOMAIN_PAGING)))
2426 return -EINVAL;
2427
2428 if (WARN_ON(!ops->map_pages || domain->pgsize_bitmap == 0UL))
2429 return -ENODEV;
2430
2431 /* find out the minimum page size supported */
2432 min_pagesz = 1 << __ffs(domain->pgsize_bitmap);
2433
2434 /*
2435 * both the virtual address and the physical one, as well as
2436 * the size of the mapping, must be aligned (at least) to the
2437 * size of the smallest page supported by the hardware
2438 */
2439 if (!IS_ALIGNED(iova | paddr | size, min_pagesz)) {
2440 pr_err("unaligned: iova 0x%lx pa %pa size 0x%zx min_pagesz 0x%x\n",
2441 iova, &paddr, size, min_pagesz);
2442 return -EINVAL;
2443 }
2444
2445 pr_debug("map: iova 0x%lx pa %pa size 0x%zx\n", iova, &paddr, size);
2446
2447 while (size) {
2448 size_t pgsize, count, mapped = 0;
2449
2450 pgsize = iommu_pgsize(domain, iova, paddr, size, &count);
2451
2452 pr_debug("mapping: iova 0x%lx pa %pa pgsize 0x%zx count %zu\n",
2453 iova, &paddr, pgsize, count);
2454 ret = ops->map_pages(domain, iova, paddr, pgsize, count, prot,
2455 gfp, &mapped);
2456 /*
2457 * Some pages may have been mapped, even if an error occurred,
2458 * so we should account for those so they can be unmapped.
2459 */
2460 size -= mapped;
2461
2462 if (ret)
2463 break;
2464
2465 iova += mapped;
2466 paddr += mapped;
2467 }
2468
2469 /* unroll mapping in case something went wrong */
2470 if (ret)
2471 iommu_unmap(domain, orig_iova, orig_size - size);
2472 else
2473 trace_map(orig_iova, orig_paddr, orig_size);
2474
2475 return ret;
2476}
2477
2478int iommu_map(struct iommu_domain *domain, unsigned long iova,
2479 phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
2480{
2481 const struct iommu_domain_ops *ops = domain->ops;
2482 int ret;
2483
2484 might_sleep_if(gfpflags_allow_blocking(gfp));
2485
2486 /* Discourage passing strange GFP flags */
2487 if (WARN_ON_ONCE(gfp & (__GFP_COMP | __GFP_DMA | __GFP_DMA32 |
2488 __GFP_HIGHMEM)))
2489 return -EINVAL;
2490
2491 ret = __iommu_map(domain, iova, paddr, size, prot, gfp);
2492 if (ret == 0 && ops->iotlb_sync_map) {
2493 ret = ops->iotlb_sync_map(domain, iova, size);
2494 if (ret)
2495 goto out_err;
2496 }
2497
2498 return ret;
2499
2500out_err:
2501 /* undo mappings already done */
2502 iommu_unmap(domain, iova, size);
2503
2504 return ret;
2505}
2506EXPORT_SYMBOL_GPL(iommu_map);
2507
2508static size_t __iommu_unmap(struct iommu_domain *domain,
2509 unsigned long iova, size_t size,
2510 struct iommu_iotlb_gather *iotlb_gather)
2511{
2512 const struct iommu_domain_ops *ops = domain->ops;
2513 size_t unmapped_page, unmapped = 0;
2514 unsigned long orig_iova = iova;
2515 unsigned int min_pagesz;
2516
2517 if (unlikely(!(domain->type & __IOMMU_DOMAIN_PAGING)))
2518 return 0;
2519
2520 if (WARN_ON(!ops->unmap_pages || domain->pgsize_bitmap == 0UL))
2521 return 0;
2522
2523 /* find out the minimum page size supported */
2524 min_pagesz = 1 << __ffs(domain->pgsize_bitmap);
2525
2526 /*
2527 * The virtual address, as well as the size of the mapping, must be
2528 * aligned (at least) to the size of the smallest page supported
2529 * by the hardware
2530 */
2531 if (!IS_ALIGNED(iova | size, min_pagesz)) {
2532 pr_err("unaligned: iova 0x%lx size 0x%zx min_pagesz 0x%x\n",
2533 iova, size, min_pagesz);
2534 return 0;
2535 }
2536
2537 pr_debug("unmap this: iova 0x%lx size 0x%zx\n", iova, size);
2538
2539 /*
2540 * Keep iterating until we either unmap 'size' bytes (or more)
2541 * or we hit an area that isn't mapped.
2542 */
2543 while (unmapped < size) {
2544 size_t pgsize, count;
2545
2546 pgsize = iommu_pgsize(domain, iova, iova, size - unmapped, &count);
2547 unmapped_page = ops->unmap_pages(domain, iova, pgsize, count, iotlb_gather);
2548 if (!unmapped_page)
2549 break;
2550
2551 pr_debug("unmapped: iova 0x%lx size 0x%zx\n",
2552 iova, unmapped_page);
2553
2554 iova += unmapped_page;
2555 unmapped += unmapped_page;
2556 }
2557
2558 trace_unmap(orig_iova, size, unmapped);
2559 return unmapped;
2560}
2561
2562/**
2563 * iommu_unmap() - Remove mappings from a range of IOVA
2564 * @domain: Domain to manipulate
2565 * @iova: IO virtual address to start
2566 * @size: Length of the range starting from @iova
2567 *
2568 * iommu_unmap() will remove a translation created by iommu_map(). It cannot
2569 * subdivide a mapping created by iommu_map(), so it should be called with IOVA
2570 * ranges that match what was passed to iommu_map(). The range can aggregate
2571 * contiguous iommu_map() calls so long as no individual range is split.
2572 *
2573 * Returns: Number of bytes of IOVA unmapped. iova + res will be the point
2574 * unmapping stopped.
2575 */
2576size_t iommu_unmap(struct iommu_domain *domain,
2577 unsigned long iova, size_t size)
2578{
2579 struct iommu_iotlb_gather iotlb_gather;
2580 size_t ret;
2581
2582 iommu_iotlb_gather_init(&iotlb_gather);
2583 ret = __iommu_unmap(domain, iova, size, &iotlb_gather);
2584 iommu_iotlb_sync(domain, &iotlb_gather);
2585
2586 return ret;
2587}
2588EXPORT_SYMBOL_GPL(iommu_unmap);
2589
2590size_t iommu_unmap_fast(struct iommu_domain *domain,
2591 unsigned long iova, size_t size,
2592 struct iommu_iotlb_gather *iotlb_gather)
2593{
2594 return __iommu_unmap(domain, iova, size, iotlb_gather);
2595}
2596EXPORT_SYMBOL_GPL(iommu_unmap_fast);
2597
2598ssize_t iommu_map_sg(struct iommu_domain *domain, unsigned long iova,
2599 struct scatterlist *sg, unsigned int nents, int prot,
2600 gfp_t gfp)
2601{
2602 const struct iommu_domain_ops *ops = domain->ops;
2603 size_t len = 0, mapped = 0;
2604 phys_addr_t start;
2605 unsigned int i = 0;
2606 int ret;
2607
2608 might_sleep_if(gfpflags_allow_blocking(gfp));
2609
2610 /* Discourage passing strange GFP flags */
2611 if (WARN_ON_ONCE(gfp & (__GFP_COMP | __GFP_DMA | __GFP_DMA32 |
2612 __GFP_HIGHMEM)))
2613 return -EINVAL;
2614
2615 while (i <= nents) {
2616 phys_addr_t s_phys = sg_phys(sg);
2617
2618 if (len && s_phys != start + len) {
2619 ret = __iommu_map(domain, iova + mapped, start,
2620 len, prot, gfp);
2621
2622 if (ret)
2623 goto out_err;
2624
2625 mapped += len;
2626 len = 0;
2627 }
2628
2629 if (sg_dma_is_bus_address(sg))
2630 goto next;
2631
2632 if (len) {
2633 len += sg->length;
2634 } else {
2635 len = sg->length;
2636 start = s_phys;
2637 }
2638
2639next:
2640 if (++i < nents)
2641 sg = sg_next(sg);
2642 }
2643
2644 if (ops->iotlb_sync_map) {
2645 ret = ops->iotlb_sync_map(domain, iova, mapped);
2646 if (ret)
2647 goto out_err;
2648 }
2649 return mapped;
2650
2651out_err:
2652 /* undo mappings already done */
2653 iommu_unmap(domain, iova, mapped);
2654
2655 return ret;
2656}
2657EXPORT_SYMBOL_GPL(iommu_map_sg);
2658
2659/**
2660 * report_iommu_fault() - report about an IOMMU fault to the IOMMU framework
2661 * @domain: the iommu domain where the fault has happened
2662 * @dev: the device where the fault has happened
2663 * @iova: the faulting address
2664 * @flags: mmu fault flags (e.g. IOMMU_FAULT_READ/IOMMU_FAULT_WRITE/...)
2665 *
2666 * This function should be called by the low-level IOMMU implementations
2667 * whenever IOMMU faults happen, to allow high-level users, that are
2668 * interested in such events, to know about them.
2669 *
2670 * This event may be useful for several possible use cases:
2671 * - mere logging of the event
2672 * - dynamic TLB/PTE loading
2673 * - if restarting of the faulting device is required
2674 *
2675 * Returns 0 on success and an appropriate error code otherwise (if dynamic
2676 * PTE/TLB loading will one day be supported, implementations will be able
2677 * to tell whether it succeeded or not according to this return value).
2678 *
2679 * Specifically, -ENOSYS is returned if a fault handler isn't installed
2680 * (though fault handlers can also return -ENOSYS, in case they want to
2681 * elicit the default behavior of the IOMMU drivers).
2682 */
2683int report_iommu_fault(struct iommu_domain *domain, struct device *dev,
2684 unsigned long iova, int flags)
2685{
2686 int ret = -ENOSYS;
2687
2688 /*
2689 * if upper layers showed interest and installed a fault handler,
2690 * invoke it.
2691 */
2692 if (domain->handler)
2693 ret = domain->handler(domain, dev, iova, flags,
2694 domain->handler_token);
2695
2696 trace_io_page_fault(dev, iova, flags);
2697 return ret;
2698}
2699EXPORT_SYMBOL_GPL(report_iommu_fault);
2700
2701static int __init iommu_init(void)
2702{
2703 iommu_group_kset = kset_create_and_add("iommu_groups",
2704 NULL, kernel_kobj);
2705 BUG_ON(!iommu_group_kset);
2706
2707 iommu_debugfs_setup();
2708
2709 return 0;
2710}
2711core_initcall(iommu_init);
2712
2713int iommu_set_pgtable_quirks(struct iommu_domain *domain,
2714 unsigned long quirk)
2715{
2716 if (domain->type != IOMMU_DOMAIN_UNMANAGED)
2717 return -EINVAL;
2718 if (!domain->ops->set_pgtable_quirks)
2719 return -EINVAL;
2720 return domain->ops->set_pgtable_quirks(domain, quirk);
2721}
2722EXPORT_SYMBOL_GPL(iommu_set_pgtable_quirks);
2723
2724/**
2725 * iommu_get_resv_regions - get reserved regions
2726 * @dev: device for which to get reserved regions
2727 * @list: reserved region list for device
2728 *
2729 * This returns a list of reserved IOVA regions specific to this device.
2730 * A domain user should not map IOVA in these ranges.
2731 */
2732void iommu_get_resv_regions(struct device *dev, struct list_head *list)
2733{
2734 const struct iommu_ops *ops = dev_iommu_ops(dev);
2735
2736 if (ops->get_resv_regions)
2737 ops->get_resv_regions(dev, list);
2738}
2739EXPORT_SYMBOL_GPL(iommu_get_resv_regions);
2740
2741/**
2742 * iommu_put_resv_regions - release reserved regions
2743 * @dev: device for which to free reserved regions
2744 * @list: reserved region list for device
2745 *
2746 * This releases a reserved region list acquired by iommu_get_resv_regions().
2747 */
2748void iommu_put_resv_regions(struct device *dev, struct list_head *list)
2749{
2750 struct iommu_resv_region *entry, *next;
2751
2752 list_for_each_entry_safe(entry, next, list, list) {
2753 if (entry->free)
2754 entry->free(dev, entry);
2755 else
2756 kfree(entry);
2757 }
2758}
2759EXPORT_SYMBOL(iommu_put_resv_regions);
2760
2761struct iommu_resv_region *iommu_alloc_resv_region(phys_addr_t start,
2762 size_t length, int prot,
2763 enum iommu_resv_type type,
2764 gfp_t gfp)
2765{
2766 struct iommu_resv_region *region;
2767
2768 region = kzalloc(sizeof(*region), gfp);
2769 if (!region)
2770 return NULL;
2771
2772 INIT_LIST_HEAD(®ion->list);
2773 region->start = start;
2774 region->length = length;
2775 region->prot = prot;
2776 region->type = type;
2777 return region;
2778}
2779EXPORT_SYMBOL_GPL(iommu_alloc_resv_region);
2780
2781void iommu_set_default_passthrough(bool cmd_line)
2782{
2783 if (cmd_line)
2784 iommu_cmd_line |= IOMMU_CMD_LINE_DMA_API;
2785 iommu_def_domain_type = IOMMU_DOMAIN_IDENTITY;
2786}
2787
2788void iommu_set_default_translated(bool cmd_line)
2789{
2790 if (cmd_line)
2791 iommu_cmd_line |= IOMMU_CMD_LINE_DMA_API;
2792 iommu_def_domain_type = IOMMU_DOMAIN_DMA;
2793}
2794
2795bool iommu_default_passthrough(void)
2796{
2797 return iommu_def_domain_type == IOMMU_DOMAIN_IDENTITY;
2798}
2799EXPORT_SYMBOL_GPL(iommu_default_passthrough);
2800
2801const struct iommu_ops *iommu_ops_from_fwnode(const struct fwnode_handle *fwnode)
2802{
2803 const struct iommu_ops *ops = NULL;
2804 struct iommu_device *iommu;
2805
2806 spin_lock(&iommu_device_lock);
2807 list_for_each_entry(iommu, &iommu_device_list, list)
2808 if (iommu->fwnode == fwnode) {
2809 ops = iommu->ops;
2810 break;
2811 }
2812 spin_unlock(&iommu_device_lock);
2813 return ops;
2814}
2815
2816int iommu_fwspec_init(struct device *dev, struct fwnode_handle *iommu_fwnode)
2817{
2818 const struct iommu_ops *ops = iommu_ops_from_fwnode(iommu_fwnode);
2819 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
2820
2821 if (!ops)
2822 return -EPROBE_DEFER;
2823
2824 if (fwspec)
2825 return ops == iommu_fwspec_ops(fwspec) ? 0 : -EINVAL;
2826
2827 if (!dev_iommu_get(dev))
2828 return -ENOMEM;
2829
2830 /* Preallocate for the overwhelmingly common case of 1 ID */
2831 fwspec = kzalloc(struct_size(fwspec, ids, 1), GFP_KERNEL);
2832 if (!fwspec)
2833 return -ENOMEM;
2834
2835 fwnode_handle_get(iommu_fwnode);
2836 fwspec->iommu_fwnode = iommu_fwnode;
2837 dev_iommu_fwspec_set(dev, fwspec);
2838 return 0;
2839}
2840EXPORT_SYMBOL_GPL(iommu_fwspec_init);
2841
2842void iommu_fwspec_free(struct device *dev)
2843{
2844 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
2845
2846 if (fwspec) {
2847 fwnode_handle_put(fwspec->iommu_fwnode);
2848 kfree(fwspec);
2849 dev_iommu_fwspec_set(dev, NULL);
2850 }
2851}
2852EXPORT_SYMBOL_GPL(iommu_fwspec_free);
2853
2854int iommu_fwspec_add_ids(struct device *dev, const u32 *ids, int num_ids)
2855{
2856 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
2857 int i, new_num;
2858
2859 if (!fwspec)
2860 return -EINVAL;
2861
2862 new_num = fwspec->num_ids + num_ids;
2863 if (new_num > 1) {
2864 fwspec = krealloc(fwspec, struct_size(fwspec, ids, new_num),
2865 GFP_KERNEL);
2866 if (!fwspec)
2867 return -ENOMEM;
2868
2869 dev_iommu_fwspec_set(dev, fwspec);
2870 }
2871
2872 for (i = 0; i < num_ids; i++)
2873 fwspec->ids[fwspec->num_ids + i] = ids[i];
2874
2875 fwspec->num_ids = new_num;
2876 return 0;
2877}
2878EXPORT_SYMBOL_GPL(iommu_fwspec_add_ids);
2879
2880/*
2881 * Per device IOMMU features.
2882 */
2883int iommu_dev_enable_feature(struct device *dev, enum iommu_dev_features feat)
2884{
2885 if (dev_has_iommu(dev)) {
2886 const struct iommu_ops *ops = dev_iommu_ops(dev);
2887
2888 if (ops->dev_enable_feat)
2889 return ops->dev_enable_feat(dev, feat);
2890 }
2891
2892 return -ENODEV;
2893}
2894EXPORT_SYMBOL_GPL(iommu_dev_enable_feature);
2895
2896/*
2897 * The device drivers should do the necessary cleanups before calling this.
2898 */
2899int iommu_dev_disable_feature(struct device *dev, enum iommu_dev_features feat)
2900{
2901 if (dev_has_iommu(dev)) {
2902 const struct iommu_ops *ops = dev_iommu_ops(dev);
2903
2904 if (ops->dev_disable_feat)
2905 return ops->dev_disable_feat(dev, feat);
2906 }
2907
2908 return -EBUSY;
2909}
2910EXPORT_SYMBOL_GPL(iommu_dev_disable_feature);
2911
2912/**
2913 * iommu_setup_default_domain - Set the default_domain for the group
2914 * @group: Group to change
2915 * @target_type: Domain type to set as the default_domain
2916 *
2917 * Allocate a default domain and set it as the current domain on the group. If
2918 * the group already has a default domain it will be changed to the target_type.
2919 * When target_type is 0 the default domain is selected based on driver and
2920 * system preferences.
2921 */
2922static int iommu_setup_default_domain(struct iommu_group *group,
2923 int target_type)
2924{
2925 struct iommu_domain *old_dom = group->default_domain;
2926 struct group_device *gdev;
2927 struct iommu_domain *dom;
2928 bool direct_failed;
2929 int req_type;
2930 int ret;
2931
2932 lockdep_assert_held(&group->mutex);
2933
2934 req_type = iommu_get_default_domain_type(group, target_type);
2935 if (req_type < 0)
2936 return -EINVAL;
2937
2938 dom = iommu_group_alloc_default_domain(group, req_type);
2939 if (IS_ERR(dom))
2940 return PTR_ERR(dom);
2941
2942 if (group->default_domain == dom)
2943 return 0;
2944
2945 if (iommu_is_dma_domain(dom)) {
2946 ret = iommu_get_dma_cookie(dom);
2947 if (ret) {
2948 iommu_domain_free(dom);
2949 return ret;
2950 }
2951 }
2952
2953 /*
2954 * IOMMU_RESV_DIRECT and IOMMU_RESV_DIRECT_RELAXABLE regions must be
2955 * mapped before their device is attached, in order to guarantee
2956 * continuity with any FW activity
2957 */
2958 direct_failed = false;
2959 for_each_group_device(group, gdev) {
2960 if (iommu_create_device_direct_mappings(dom, gdev->dev)) {
2961 direct_failed = true;
2962 dev_warn_once(
2963 gdev->dev->iommu->iommu_dev->dev,
2964 "IOMMU driver was not able to establish FW requested direct mapping.");
2965 }
2966 }
2967
2968 /* We must set default_domain early for __iommu_device_set_domain */
2969 group->default_domain = dom;
2970 if (!group->domain) {
2971 /*
2972 * Drivers are not allowed to fail the first domain attach.
2973 * The only way to recover from this is to fail attaching the
2974 * iommu driver and call ops->release_device. Put the domain
2975 * in group->default_domain so it is freed after.
2976 */
2977 ret = __iommu_group_set_domain_internal(
2978 group, dom, IOMMU_SET_DOMAIN_MUST_SUCCEED);
2979 if (WARN_ON(ret))
2980 goto out_free_old;
2981 } else {
2982 ret = __iommu_group_set_domain(group, dom);
2983 if (ret)
2984 goto err_restore_def_domain;
2985 }
2986
2987 /*
2988 * Drivers are supposed to allow mappings to be installed in a domain
2989 * before device attachment, but some don't. Hack around this defect by
2990 * trying again after attaching. If this happens it means the device
2991 * will not continuously have the IOMMU_RESV_DIRECT map.
2992 */
2993 if (direct_failed) {
2994 for_each_group_device(group, gdev) {
2995 ret = iommu_create_device_direct_mappings(dom, gdev->dev);
2996 if (ret)
2997 goto err_restore_domain;
2998 }
2999 }
3000
3001out_free_old:
3002 if (old_dom)
3003 iommu_domain_free(old_dom);
3004 return ret;
3005
3006err_restore_domain:
3007 if (old_dom)
3008 __iommu_group_set_domain_internal(
3009 group, old_dom, IOMMU_SET_DOMAIN_MUST_SUCCEED);
3010err_restore_def_domain:
3011 if (old_dom) {
3012 iommu_domain_free(dom);
3013 group->default_domain = old_dom;
3014 }
3015 return ret;
3016}
3017
3018/*
3019 * Changing the default domain through sysfs requires the users to unbind the
3020 * drivers from the devices in the iommu group, except for a DMA -> DMA-FQ
3021 * transition. Return failure if this isn't met.
3022 *
3023 * We need to consider the race between this and the device release path.
3024 * group->mutex is used here to guarantee that the device release path
3025 * will not be entered at the same time.
3026 */
3027static ssize_t iommu_group_store_type(struct iommu_group *group,
3028 const char *buf, size_t count)
3029{
3030 struct group_device *gdev;
3031 int ret, req_type;
3032
3033 if (!capable(CAP_SYS_ADMIN) || !capable(CAP_SYS_RAWIO))
3034 return -EACCES;
3035
3036 if (WARN_ON(!group) || !group->default_domain)
3037 return -EINVAL;
3038
3039 if (sysfs_streq(buf, "identity"))
3040 req_type = IOMMU_DOMAIN_IDENTITY;
3041 else if (sysfs_streq(buf, "DMA"))
3042 req_type = IOMMU_DOMAIN_DMA;
3043 else if (sysfs_streq(buf, "DMA-FQ"))
3044 req_type = IOMMU_DOMAIN_DMA_FQ;
3045 else if (sysfs_streq(buf, "auto"))
3046 req_type = 0;
3047 else
3048 return -EINVAL;
3049
3050 mutex_lock(&group->mutex);
3051 /* We can bring up a flush queue without tearing down the domain. */
3052 if (req_type == IOMMU_DOMAIN_DMA_FQ &&
3053 group->default_domain->type == IOMMU_DOMAIN_DMA) {
3054 ret = iommu_dma_init_fq(group->default_domain);
3055 if (ret)
3056 goto out_unlock;
3057
3058 group->default_domain->type = IOMMU_DOMAIN_DMA_FQ;
3059 ret = count;
3060 goto out_unlock;
3061 }
3062
3063 /* Otherwise, ensure that device exists and no driver is bound. */
3064 if (list_empty(&group->devices) || group->owner_cnt) {
3065 ret = -EPERM;
3066 goto out_unlock;
3067 }
3068
3069 ret = iommu_setup_default_domain(group, req_type);
3070 if (ret)
3071 goto out_unlock;
3072
3073 /* Make sure dma_ops is appropriatley set */
3074 for_each_group_device(group, gdev)
3075 iommu_setup_dma_ops(gdev->dev);
3076
3077out_unlock:
3078 mutex_unlock(&group->mutex);
3079 return ret ?: count;
3080}
3081
3082/**
3083 * iommu_device_use_default_domain() - Device driver wants to handle device
3084 * DMA through the kernel DMA API.
3085 * @dev: The device.
3086 *
3087 * The device driver about to bind @dev wants to do DMA through the kernel
3088 * DMA API. Return 0 if it is allowed, otherwise an error.
3089 */
3090int iommu_device_use_default_domain(struct device *dev)
3091{
3092 /* Caller is the driver core during the pre-probe path */
3093 struct iommu_group *group = dev->iommu_group;
3094 int ret = 0;
3095
3096 if (!group)
3097 return 0;
3098
3099 mutex_lock(&group->mutex);
3100 if (group->owner_cnt) {
3101 if (group->domain != group->default_domain || group->owner ||
3102 !xa_empty(&group->pasid_array)) {
3103 ret = -EBUSY;
3104 goto unlock_out;
3105 }
3106 }
3107
3108 group->owner_cnt++;
3109
3110unlock_out:
3111 mutex_unlock(&group->mutex);
3112 return ret;
3113}
3114
3115/**
3116 * iommu_device_unuse_default_domain() - Device driver stops handling device
3117 * DMA through the kernel DMA API.
3118 * @dev: The device.
3119 *
3120 * The device driver doesn't want to do DMA through kernel DMA API anymore.
3121 * It must be called after iommu_device_use_default_domain().
3122 */
3123void iommu_device_unuse_default_domain(struct device *dev)
3124{
3125 /* Caller is the driver core during the post-probe path */
3126 struct iommu_group *group = dev->iommu_group;
3127
3128 if (!group)
3129 return;
3130
3131 mutex_lock(&group->mutex);
3132 if (!WARN_ON(!group->owner_cnt || !xa_empty(&group->pasid_array)))
3133 group->owner_cnt--;
3134
3135 mutex_unlock(&group->mutex);
3136}
3137
3138static int __iommu_group_alloc_blocking_domain(struct iommu_group *group)
3139{
3140 struct device *dev = iommu_group_first_dev(group);
3141 const struct iommu_ops *ops = dev_iommu_ops(dev);
3142 struct iommu_domain *domain;
3143
3144 if (group->blocking_domain)
3145 return 0;
3146
3147 if (ops->blocked_domain) {
3148 group->blocking_domain = ops->blocked_domain;
3149 return 0;
3150 }
3151
3152 /*
3153 * For drivers that do not yet understand IOMMU_DOMAIN_BLOCKED create an
3154 * empty PAGING domain instead.
3155 */
3156 domain = iommu_paging_domain_alloc(dev);
3157 if (IS_ERR(domain))
3158 return PTR_ERR(domain);
3159 group->blocking_domain = domain;
3160 return 0;
3161}
3162
3163static int __iommu_take_dma_ownership(struct iommu_group *group, void *owner)
3164{
3165 int ret;
3166
3167 if ((group->domain && group->domain != group->default_domain) ||
3168 !xa_empty(&group->pasid_array))
3169 return -EBUSY;
3170
3171 ret = __iommu_group_alloc_blocking_domain(group);
3172 if (ret)
3173 return ret;
3174 ret = __iommu_group_set_domain(group, group->blocking_domain);
3175 if (ret)
3176 return ret;
3177
3178 group->owner = owner;
3179 group->owner_cnt++;
3180 return 0;
3181}
3182
3183/**
3184 * iommu_group_claim_dma_owner() - Set DMA ownership of a group
3185 * @group: The group.
3186 * @owner: Caller specified pointer. Used for exclusive ownership.
3187 *
3188 * This is to support backward compatibility for vfio which manages the dma
3189 * ownership in iommu_group level. New invocations on this interface should be
3190 * prohibited. Only a single owner may exist for a group.
3191 */
3192int iommu_group_claim_dma_owner(struct iommu_group *group, void *owner)
3193{
3194 int ret = 0;
3195
3196 if (WARN_ON(!owner))
3197 return -EINVAL;
3198
3199 mutex_lock(&group->mutex);
3200 if (group->owner_cnt) {
3201 ret = -EPERM;
3202 goto unlock_out;
3203 }
3204
3205 ret = __iommu_take_dma_ownership(group, owner);
3206unlock_out:
3207 mutex_unlock(&group->mutex);
3208
3209 return ret;
3210}
3211EXPORT_SYMBOL_GPL(iommu_group_claim_dma_owner);
3212
3213/**
3214 * iommu_device_claim_dma_owner() - Set DMA ownership of a device
3215 * @dev: The device.
3216 * @owner: Caller specified pointer. Used for exclusive ownership.
3217 *
3218 * Claim the DMA ownership of a device. Multiple devices in the same group may
3219 * concurrently claim ownership if they present the same owner value. Returns 0
3220 * on success and error code on failure
3221 */
3222int iommu_device_claim_dma_owner(struct device *dev, void *owner)
3223{
3224 /* Caller must be a probed driver on dev */
3225 struct iommu_group *group = dev->iommu_group;
3226 int ret = 0;
3227
3228 if (WARN_ON(!owner))
3229 return -EINVAL;
3230
3231 if (!group)
3232 return -ENODEV;
3233
3234 mutex_lock(&group->mutex);
3235 if (group->owner_cnt) {
3236 if (group->owner != owner) {
3237 ret = -EPERM;
3238 goto unlock_out;
3239 }
3240 group->owner_cnt++;
3241 goto unlock_out;
3242 }
3243
3244 ret = __iommu_take_dma_ownership(group, owner);
3245unlock_out:
3246 mutex_unlock(&group->mutex);
3247 return ret;
3248}
3249EXPORT_SYMBOL_GPL(iommu_device_claim_dma_owner);
3250
3251static void __iommu_release_dma_ownership(struct iommu_group *group)
3252{
3253 if (WARN_ON(!group->owner_cnt || !group->owner ||
3254 !xa_empty(&group->pasid_array)))
3255 return;
3256
3257 group->owner_cnt = 0;
3258 group->owner = NULL;
3259 __iommu_group_set_domain_nofail(group, group->default_domain);
3260}
3261
3262/**
3263 * iommu_group_release_dma_owner() - Release DMA ownership of a group
3264 * @group: The group
3265 *
3266 * Release the DMA ownership claimed by iommu_group_claim_dma_owner().
3267 */
3268void iommu_group_release_dma_owner(struct iommu_group *group)
3269{
3270 mutex_lock(&group->mutex);
3271 __iommu_release_dma_ownership(group);
3272 mutex_unlock(&group->mutex);
3273}
3274EXPORT_SYMBOL_GPL(iommu_group_release_dma_owner);
3275
3276/**
3277 * iommu_device_release_dma_owner() - Release DMA ownership of a device
3278 * @dev: The device.
3279 *
3280 * Release the DMA ownership claimed by iommu_device_claim_dma_owner().
3281 */
3282void iommu_device_release_dma_owner(struct device *dev)
3283{
3284 /* Caller must be a probed driver on dev */
3285 struct iommu_group *group = dev->iommu_group;
3286
3287 mutex_lock(&group->mutex);
3288 if (group->owner_cnt > 1)
3289 group->owner_cnt--;
3290 else
3291 __iommu_release_dma_ownership(group);
3292 mutex_unlock(&group->mutex);
3293}
3294EXPORT_SYMBOL_GPL(iommu_device_release_dma_owner);
3295
3296/**
3297 * iommu_group_dma_owner_claimed() - Query group dma ownership status
3298 * @group: The group.
3299 *
3300 * This provides status query on a given group. It is racy and only for
3301 * non-binding status reporting.
3302 */
3303bool iommu_group_dma_owner_claimed(struct iommu_group *group)
3304{
3305 unsigned int user;
3306
3307 mutex_lock(&group->mutex);
3308 user = group->owner_cnt;
3309 mutex_unlock(&group->mutex);
3310
3311 return user;
3312}
3313EXPORT_SYMBOL_GPL(iommu_group_dma_owner_claimed);
3314
3315static int __iommu_set_group_pasid(struct iommu_domain *domain,
3316 struct iommu_group *group, ioasid_t pasid)
3317{
3318 struct group_device *device, *last_gdev;
3319 int ret;
3320
3321 for_each_group_device(group, device) {
3322 ret = domain->ops->set_dev_pasid(domain, device->dev,
3323 pasid, NULL);
3324 if (ret)
3325 goto err_revert;
3326 }
3327
3328 return 0;
3329
3330err_revert:
3331 last_gdev = device;
3332 for_each_group_device(group, device) {
3333 const struct iommu_ops *ops = dev_iommu_ops(device->dev);
3334
3335 if (device == last_gdev)
3336 break;
3337 ops->remove_dev_pasid(device->dev, pasid, domain);
3338 }
3339 return ret;
3340}
3341
3342static void __iommu_remove_group_pasid(struct iommu_group *group,
3343 ioasid_t pasid,
3344 struct iommu_domain *domain)
3345{
3346 struct group_device *device;
3347 const struct iommu_ops *ops;
3348
3349 for_each_group_device(group, device) {
3350 ops = dev_iommu_ops(device->dev);
3351 ops->remove_dev_pasid(device->dev, pasid, domain);
3352 }
3353}
3354
3355/*
3356 * iommu_attach_device_pasid() - Attach a domain to pasid of device
3357 * @domain: the iommu domain.
3358 * @dev: the attached device.
3359 * @pasid: the pasid of the device.
3360 * @handle: the attach handle.
3361 *
3362 * Return: 0 on success, or an error.
3363 */
3364int iommu_attach_device_pasid(struct iommu_domain *domain,
3365 struct device *dev, ioasid_t pasid,
3366 struct iommu_attach_handle *handle)
3367{
3368 /* Caller must be a probed driver on dev */
3369 struct iommu_group *group = dev->iommu_group;
3370 struct group_device *device;
3371 int ret;
3372
3373 if (!domain->ops->set_dev_pasid)
3374 return -EOPNOTSUPP;
3375
3376 if (!group)
3377 return -ENODEV;
3378
3379 if (!dev_has_iommu(dev) || dev_iommu_ops(dev) != domain->owner ||
3380 pasid == IOMMU_NO_PASID)
3381 return -EINVAL;
3382
3383 mutex_lock(&group->mutex);
3384 for_each_group_device(group, device) {
3385 if (pasid >= device->dev->iommu->max_pasids) {
3386 ret = -EINVAL;
3387 goto out_unlock;
3388 }
3389 }
3390
3391 if (handle)
3392 handle->domain = domain;
3393
3394 ret = xa_insert(&group->pasid_array, pasid, handle, GFP_KERNEL);
3395 if (ret)
3396 goto out_unlock;
3397
3398 ret = __iommu_set_group_pasid(domain, group, pasid);
3399 if (ret)
3400 xa_erase(&group->pasid_array, pasid);
3401out_unlock:
3402 mutex_unlock(&group->mutex);
3403 return ret;
3404}
3405EXPORT_SYMBOL_GPL(iommu_attach_device_pasid);
3406
3407/*
3408 * iommu_detach_device_pasid() - Detach the domain from pasid of device
3409 * @domain: the iommu domain.
3410 * @dev: the attached device.
3411 * @pasid: the pasid of the device.
3412 *
3413 * The @domain must have been attached to @pasid of the @dev with
3414 * iommu_attach_device_pasid().
3415 */
3416void iommu_detach_device_pasid(struct iommu_domain *domain, struct device *dev,
3417 ioasid_t pasid)
3418{
3419 /* Caller must be a probed driver on dev */
3420 struct iommu_group *group = dev->iommu_group;
3421
3422 mutex_lock(&group->mutex);
3423 __iommu_remove_group_pasid(group, pasid, domain);
3424 xa_erase(&group->pasid_array, pasid);
3425 mutex_unlock(&group->mutex);
3426}
3427EXPORT_SYMBOL_GPL(iommu_detach_device_pasid);
3428
3429ioasid_t iommu_alloc_global_pasid(struct device *dev)
3430{
3431 int ret;
3432
3433 /* max_pasids == 0 means that the device does not support PASID */
3434 if (!dev->iommu->max_pasids)
3435 return IOMMU_PASID_INVALID;
3436
3437 /*
3438 * max_pasids is set up by vendor driver based on number of PASID bits
3439 * supported but the IDA allocation is inclusive.
3440 */
3441 ret = ida_alloc_range(&iommu_global_pasid_ida, IOMMU_FIRST_GLOBAL_PASID,
3442 dev->iommu->max_pasids - 1, GFP_KERNEL);
3443 return ret < 0 ? IOMMU_PASID_INVALID : ret;
3444}
3445EXPORT_SYMBOL_GPL(iommu_alloc_global_pasid);
3446
3447void iommu_free_global_pasid(ioasid_t pasid)
3448{
3449 if (WARN_ON(pasid == IOMMU_PASID_INVALID))
3450 return;
3451
3452 ida_free(&iommu_global_pasid_ida, pasid);
3453}
3454EXPORT_SYMBOL_GPL(iommu_free_global_pasid);
3455
3456/**
3457 * iommu_attach_handle_get - Return the attach handle
3458 * @group: the iommu group that domain was attached to
3459 * @pasid: the pasid within the group
3460 * @type: matched domain type, 0 for any match
3461 *
3462 * Return handle or ERR_PTR(-ENOENT) on none, ERR_PTR(-EBUSY) on mismatch.
3463 *
3464 * Return the attach handle to the caller. The life cycle of an iommu attach
3465 * handle is from the time when the domain is attached to the time when the
3466 * domain is detached. Callers are required to synchronize the call of
3467 * iommu_attach_handle_get() with domain attachment and detachment. The attach
3468 * handle can only be used during its life cycle.
3469 */
3470struct iommu_attach_handle *
3471iommu_attach_handle_get(struct iommu_group *group, ioasid_t pasid, unsigned int type)
3472{
3473 struct iommu_attach_handle *handle;
3474
3475 xa_lock(&group->pasid_array);
3476 handle = xa_load(&group->pasid_array, pasid);
3477 if (!handle)
3478 handle = ERR_PTR(-ENOENT);
3479 else if (type && handle->domain->type != type)
3480 handle = ERR_PTR(-EBUSY);
3481 xa_unlock(&group->pasid_array);
3482
3483 return handle;
3484}
3485EXPORT_SYMBOL_NS_GPL(iommu_attach_handle_get, "IOMMUFD_INTERNAL");
3486
3487/**
3488 * iommu_attach_group_handle - Attach an IOMMU domain to an IOMMU group
3489 * @domain: IOMMU domain to attach
3490 * @group: IOMMU group that will be attached
3491 * @handle: attach handle
3492 *
3493 * Returns 0 on success and error code on failure.
3494 *
3495 * This is a variant of iommu_attach_group(). It allows the caller to provide
3496 * an attach handle and use it when the domain is attached. This is currently
3497 * used by IOMMUFD to deliver the I/O page faults.
3498 */
3499int iommu_attach_group_handle(struct iommu_domain *domain,
3500 struct iommu_group *group,
3501 struct iommu_attach_handle *handle)
3502{
3503 int ret;
3504
3505 if (handle)
3506 handle->domain = domain;
3507
3508 mutex_lock(&group->mutex);
3509 ret = xa_insert(&group->pasid_array, IOMMU_NO_PASID, handle, GFP_KERNEL);
3510 if (ret)
3511 goto err_unlock;
3512
3513 ret = __iommu_attach_group(domain, group);
3514 if (ret)
3515 goto err_erase;
3516 mutex_unlock(&group->mutex);
3517
3518 return 0;
3519err_erase:
3520 xa_erase(&group->pasid_array, IOMMU_NO_PASID);
3521err_unlock:
3522 mutex_unlock(&group->mutex);
3523 return ret;
3524}
3525EXPORT_SYMBOL_NS_GPL(iommu_attach_group_handle, "IOMMUFD_INTERNAL");
3526
3527/**
3528 * iommu_detach_group_handle - Detach an IOMMU domain from an IOMMU group
3529 * @domain: IOMMU domain to attach
3530 * @group: IOMMU group that will be attached
3531 *
3532 * Detach the specified IOMMU domain from the specified IOMMU group.
3533 * It must be used in conjunction with iommu_attach_group_handle().
3534 */
3535void iommu_detach_group_handle(struct iommu_domain *domain,
3536 struct iommu_group *group)
3537{
3538 mutex_lock(&group->mutex);
3539 __iommu_group_set_core_domain(group);
3540 xa_erase(&group->pasid_array, IOMMU_NO_PASID);
3541 mutex_unlock(&group->mutex);
3542}
3543EXPORT_SYMBOL_NS_GPL(iommu_detach_group_handle, "IOMMUFD_INTERNAL");
3544
3545/**
3546 * iommu_replace_group_handle - replace the domain that a group is attached to
3547 * @group: IOMMU group that will be attached to the new domain
3548 * @new_domain: new IOMMU domain to replace with
3549 * @handle: attach handle
3550 *
3551 * This is a variant of iommu_group_replace_domain(). It allows the caller to
3552 * provide an attach handle for the new domain and use it when the domain is
3553 * attached.
3554 */
3555int iommu_replace_group_handle(struct iommu_group *group,
3556 struct iommu_domain *new_domain,
3557 struct iommu_attach_handle *handle)
3558{
3559 void *curr;
3560 int ret;
3561
3562 if (!new_domain)
3563 return -EINVAL;
3564
3565 mutex_lock(&group->mutex);
3566 if (handle) {
3567 ret = xa_reserve(&group->pasid_array, IOMMU_NO_PASID, GFP_KERNEL);
3568 if (ret)
3569 goto err_unlock;
3570 handle->domain = new_domain;
3571 }
3572
3573 ret = __iommu_group_set_domain(group, new_domain);
3574 if (ret)
3575 goto err_release;
3576
3577 curr = xa_store(&group->pasid_array, IOMMU_NO_PASID, handle, GFP_KERNEL);
3578 WARN_ON(xa_is_err(curr));
3579
3580 mutex_unlock(&group->mutex);
3581
3582 return 0;
3583err_release:
3584 xa_release(&group->pasid_array, IOMMU_NO_PASID);
3585err_unlock:
3586 mutex_unlock(&group->mutex);
3587 return ret;
3588}
3589EXPORT_SYMBOL_NS_GPL(iommu_replace_group_handle, "IOMMUFD_INTERNAL");