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
36	#include "dma-iommu.h"
37	#include "iommu-priv.h"
38
39	static struct kset *iommu_group_kset;
40	static DEFINE_IDA(iommu_group_ida);
41	static DEFINE_IDA(iommu_global_pasid_ida);
42
43	static unsigned int iommu_def_domain_type __read_mostly;
44	static bool iommu_dma_strict __read_mostly = IS_ENABLED(CONFIG_IOMMU_DEFAULT_DMA_STRICT);
45	static u32 iommu_cmd_line __read_mostly;
46
47	struct iommu_group {
48	struct kobject kobj;
49	struct kobject *devices_kobj;
50	struct list_head devices;
51	struct xarray pasid_array;
52	struct mutex mutex;
53	void *iommu_data;
54	void (*iommu_data_release)(void *iommu_data);
55	char *name;
56	int id;
57	struct iommu_domain *default_domain;
58	struct iommu_domain *blocking_domain;
59	struct iommu_domain *domain;
60	struct list_head entry;
61	unsigned int owner_cnt;
62	void *owner;
63	};
64
65	struct group_device {
66	struct list_head list;
67	struct device *dev;
68	char *name;
69	};
70
71	/* Iterate over each struct group_device in a struct iommu_group */
72	#define for_each_group_device(group, pos) \
73	list_for_each_entry(pos, &(group)->devices, list)
74
75	struct iommu_group_attribute {
76	struct attribute attr;
77	ssize_t (*show)(struct iommu_group *group, char *buf);
78	ssize_t (*store)(struct iommu_group *group,
79	const char *buf, size_t count);
80	};
81
82	static const char * const iommu_group_resv_type_string[] = {
83	[IOMMU_RESV_DIRECT] = "direct",
84	[IOMMU_RESV_DIRECT_RELAXABLE] = "direct-relaxable",
85	[IOMMU_RESV_RESERVED] = "reserved",
86	[IOMMU_RESV_MSI] = "msi",
87	[IOMMU_RESV_SW_MSI] = "msi",
88	};
89
90	#define IOMMU_CMD_LINE_DMA_API BIT(0)
91	#define IOMMU_CMD_LINE_STRICT BIT(1)
92
93	static int iommu_bus_notifier(struct notifier_block *nb,
94	unsigned long action, void *data);
95	static void iommu_release_device(struct device *dev);
96	static struct iommu_domain *
97	__iommu_group_domain_alloc(struct iommu_group *group, unsigned int type);
98	static int __iommu_attach_device(struct iommu_domain *domain,
99	struct device *dev);
100	static int __iommu_attach_group(struct iommu_domain *domain,
101	struct iommu_group *group);
102
103	enum {
104	IOMMU_SET_DOMAIN_MUST_SUCCEED = 1 << 0,
105	};
106
107	static int __iommu_device_set_domain(struct iommu_group *group,
108	struct device *dev,
109	struct iommu_domain *new_domain,
110	unsigned int flags);
111	static int __iommu_group_set_domain_internal(struct iommu_group *group,
112	struct iommu_domain *new_domain,
113	unsigned int flags);
114	static int __iommu_group_set_domain(struct iommu_group *group,
115	struct iommu_domain *new_domain)
116	{
117	return __iommu_group_set_domain_internal(group, new_domain, flags: 0);
118	}
119	static void __iommu_group_set_domain_nofail(struct iommu_group *group,
120	struct iommu_domain *new_domain)
121	{
122	WARN_ON(__iommu_group_set_domain_internal(
123	group, new_domain, IOMMU_SET_DOMAIN_MUST_SUCCEED));
124	}
125
126	static int iommu_setup_default_domain(struct iommu_group *group,
127	int target_type);
128	static int iommu_create_device_direct_mappings(struct iommu_domain *domain,
129	struct device *dev);
130	static ssize_t iommu_group_store_type(struct iommu_group *group,
131	const char *buf, size_t count);
132	static struct group_device *iommu_group_alloc_device(struct iommu_group *group,
133	struct device *dev);
134	static void __iommu_group_free_device(struct iommu_group *group,
135	struct group_device *grp_dev);
136
137	#define IOMMU_GROUP_ATTR(_name, _mode, _show, _store) \
138	struct iommu_group_attribute iommu_group_attr_##_name = \
139	__ATTR(_name, _mode, _show, _store)
140
141	#define to_iommu_group_attr(_attr) \
142	container_of(_attr, struct iommu_group_attribute, attr)
143	#define to_iommu_group(_kobj) \
144	container_of(_kobj, struct iommu_group, kobj)
145
146	static LIST_HEAD(iommu_device_list);
147	static DEFINE_SPINLOCK(iommu_device_lock);
148
149	static const struct bus_type * const iommu_buses[] = {
150	&platform_bus_type,
151	#ifdef CONFIG_PCI
152	&pci_bus_type,
153	#endif
154	#ifdef CONFIG_ARM_AMBA
155	&amba_bustype,
156	#endif
157	#ifdef CONFIG_FSL_MC_BUS
158	&fsl_mc_bus_type,
159	#endif
160	#ifdef CONFIG_TEGRA_HOST1X_CONTEXT_BUS
161	&host1x_context_device_bus_type,
162	#endif
163	#ifdef CONFIG_CDX_BUS
164	&cdx_bus_type,
165	#endif
166	};
167
168	/*
169	* Use a function instead of an array here because the domain-type is a
170	* bit-field, so an array would waste memory.
171	*/
172	static const char *iommu_domain_type_str(unsigned int t)
173	{
174	switch (t) {
175	case IOMMU_DOMAIN_BLOCKED:
176	return "Blocked";
177	case IOMMU_DOMAIN_IDENTITY:
178	return "Passthrough";
179	case IOMMU_DOMAIN_UNMANAGED:
180	return "Unmanaged";
181	case IOMMU_DOMAIN_DMA:
182	case IOMMU_DOMAIN_DMA_FQ:
183	return "Translated";
184	case IOMMU_DOMAIN_PLATFORM:
185	return "Platform";
186	default:
187	return "Unknown";
188	}
189	}
190
191	static int __init iommu_subsys_init(void)
192	{
193	struct notifier_block *nb;
194
195	if (!(iommu_cmd_line & IOMMU_CMD_LINE_DMA_API)) {
196	if (IS_ENABLED(CONFIG_IOMMU_DEFAULT_PASSTHROUGH))
197	iommu_set_default_passthrough(cmd_line: false);
198	else
199	iommu_set_default_translated(cmd_line: false);
200
201	if (iommu_default_passthrough() && cc_platform_has(attr: CC_ATTR_MEM_ENCRYPT)) {
202	pr_info("Memory encryption detected - Disabling default IOMMU Passthrough\n");
203	iommu_set_default_translated(cmd_line: false);
204	}
205	}
206
207	if (!iommu_default_passthrough() && !iommu_dma_strict)
208	iommu_def_domain_type = IOMMU_DOMAIN_DMA_FQ;
209
210	pr_info("Default domain type: %s%s\n",
211	iommu_domain_type_str(iommu_def_domain_type),
212	(iommu_cmd_line & IOMMU_CMD_LINE_DMA_API) ?
213	" (set via kernel command line)" : "");
214
215	if (!iommu_default_passthrough())
216	pr_info("DMA domain TLB invalidation policy: %s mode%s\n",
217	iommu_dma_strict ? "strict" : "lazy",
218	(iommu_cmd_line & IOMMU_CMD_LINE_STRICT) ?
219	" (set via kernel command line)" : "");
220
221	nb = kcalloc(ARRAY_SIZE(iommu_buses), size: sizeof(*nb), GFP_KERNEL);
222	if (!nb)
223	return -ENOMEM;
224
225	for (int i = 0; i < ARRAY_SIZE(iommu_buses); i++) {
226	nb[i].notifier_call = iommu_bus_notifier;
227	bus_register_notifier(bus: iommu_buses[i], nb: &nb[i]);
228	}
229
230	return 0;
231	}
232	subsys_initcall(iommu_subsys_init);
233
234	static int remove_iommu_group(struct device *dev, void *data)
235	{
236	if (dev->iommu && dev->iommu->iommu_dev == data)
237	iommu_release_device(dev);
238
239	return 0;
240	}
241
242	/**
243	* iommu_device_register() - Register an IOMMU hardware instance
244	* @iommu: IOMMU handle for the instance
245	* @ops: IOMMU ops to associate with the instance
246	* @hwdev: (optional) actual instance device, used for fwnode lookup
247	*
248	* Return: 0 on success, or an error.
249	*/
250	int iommu_device_register(struct iommu_device *iommu,
251	const struct iommu_ops *ops, struct device *hwdev)
252	{
253	int err = 0;
254
255	/* We need to be able to take module references appropriately */
256	if (WARN_ON(is_module_address((unsigned long)ops) && !ops->owner))
257	return -EINVAL;
258
259	iommu->ops = ops;
260	if (hwdev)
261	iommu->fwnode = dev_fwnode(hwdev);
262
263	spin_lock(lock: &iommu_device_lock);
264	list_add_tail(new: &iommu->list, head: &iommu_device_list);
265	spin_unlock(lock: &iommu_device_lock);
266
267	for (int i = 0; i < ARRAY_SIZE(iommu_buses) && !err; i++)
268	err = bus_iommu_probe(bus: iommu_buses[i]);
269	if (err)
270	iommu_device_unregister(iommu);
271	return err;
272	}
273	EXPORT_SYMBOL_GPL(iommu_device_register);
274
275	void iommu_device_unregister(struct iommu_device *iommu)
276	{
277	for (int i = 0; i < ARRAY_SIZE(iommu_buses); i++)
278	bus_for_each_dev(bus: iommu_buses[i], NULL, data: iommu, fn: remove_iommu_group);
279
280	spin_lock(lock: &iommu_device_lock);
281	list_del(entry: &iommu->list);
282	spin_unlock(lock: &iommu_device_lock);
283
284	/* Pairs with the alloc in generic_single_device_group() */
285	iommu_group_put(group: iommu->singleton_group);
286	iommu->singleton_group = NULL;
287	}
288	EXPORT_SYMBOL_GPL(iommu_device_unregister);
289
290	#if IS_ENABLED(CONFIG_IOMMUFD_TEST)
291	void iommu_device_unregister_bus(struct iommu_device *iommu,
292	const struct bus_type *bus,
293	struct notifier_block *nb)
294	{
295	bus_unregister_notifier(bus, nb);
296	iommu_device_unregister(iommu);
297	}
298	EXPORT_SYMBOL_GPL(iommu_device_unregister_bus);
299
300	/*
301	* Register an iommu driver against a single bus. This is only used by iommufd
302	* selftest to create a mock iommu driver. The caller must provide
303	* some memory to hold a notifier_block.
304	*/
305	int iommu_device_register_bus(struct iommu_device *iommu,
306	const struct iommu_ops *ops,
307	const struct bus_type *bus,
308	struct notifier_block *nb)
309	{
310	int err;
311
312	iommu->ops = ops;
313	nb->notifier_call = iommu_bus_notifier;
314	err = bus_register_notifier(bus, nb);
315	if (err)
316	return err;
317
318	spin_lock(lock: &iommu_device_lock);
319	list_add_tail(new: &iommu->list, head: &iommu_device_list);
320	spin_unlock(lock: &iommu_device_lock);
321
322	err = bus_iommu_probe(bus);
323	if (err) {
324	iommu_device_unregister_bus(iommu, bus, nb);
325	return err;
326	}
327	return 0;
328	}
329	EXPORT_SYMBOL_GPL(iommu_device_register_bus);
330	#endif
331
332	static struct dev_iommu *dev_iommu_get(struct device *dev)
333	{
334	struct dev_iommu *param = dev->iommu;
335
336	lockdep_assert_held(&iommu_probe_device_lock);
337
338	if (param)
339	return param;
340
341	param = kzalloc(size: sizeof(*param), GFP_KERNEL);
342	if (!param)
343	return NULL;
344
345	mutex_init(&param->lock);
346	dev->iommu = param;
347	return param;
348	}
349
350	static void dev_iommu_free(struct device *dev)
351	{
352	struct dev_iommu *param = dev->iommu;
353
354	dev->iommu = NULL;
355	if (param->fwspec) {
356	fwnode_handle_put(fwnode: param->fwspec->iommu_fwnode);
357	kfree(objp: param->fwspec);
358	}
359	kfree(objp: param);
360	}
361
362	/*
363	* Internal equivalent of device_iommu_mapped() for when we care that a device
364	* actually has API ops, and don't want false positives from VFIO-only groups.
365	*/
366	static bool dev_has_iommu(struct device *dev)
367	{
368	return dev->iommu && dev->iommu->iommu_dev;
369	}
370
371	static u32 dev_iommu_get_max_pasids(struct device *dev)
372	{
373	u32 max_pasids = 0, bits = 0;
374	int ret;
375
376	if (dev_is_pci(dev)) {
377	ret = pci_max_pasids(to_pci_dev(dev));
378	if (ret > 0)
379	max_pasids = ret;
380	} else {
381	ret = device_property_read_u32(dev, propname: "pasid-num-bits", val: &bits);
382	if (!ret)
383	max_pasids = 1UL << bits;
384	}
385
386	return min_t(u32, max_pasids, dev->iommu->iommu_dev->max_pasids);
387	}
388
389	void dev_iommu_priv_set(struct device *dev, void *priv)
390	{
391	/* FSL_PAMU does something weird */
392	if (!IS_ENABLED(CONFIG_FSL_PAMU))
393	lockdep_assert_held(&iommu_probe_device_lock);
394	dev->iommu->priv = priv;
395	}
396	EXPORT_SYMBOL_GPL(dev_iommu_priv_set);
397
398	/*
399	* Init the dev->iommu and dev->iommu_group in the struct device and get the
400	* driver probed
401	*/
402	static int iommu_init_device(struct device *dev, const struct iommu_ops *ops)
403	{
404	struct iommu_device *iommu_dev;
405	struct iommu_group *group;
406	int ret;
407
408	if (!dev_iommu_get(dev))
409	return -ENOMEM;
410
411	if (!try_module_get(module: ops->owner)) {
412	ret = -EINVAL;
413	goto err_free;
414	}
415
416	iommu_dev = ops->probe_device(dev);
417	if (IS_ERR(ptr: iommu_dev)) {
418	ret = PTR_ERR(ptr: iommu_dev);
419	goto err_module_put;
420	}
421	dev->iommu->iommu_dev = iommu_dev;
422
423	ret = iommu_device_link(iommu: iommu_dev, link: dev);
424	if (ret)
425	goto err_release;
426
427	group = ops->device_group(dev);
428	if (WARN_ON_ONCE(group == NULL))
429	group = ERR_PTR(error: -EINVAL);
430	if (IS_ERR(ptr: group)) {
431	ret = PTR_ERR(ptr: group);
432	goto err_unlink;
433	}
434	dev->iommu_group = group;
435
436	dev->iommu->max_pasids = dev_iommu_get_max_pasids(dev);
437	if (ops->is_attach_deferred)
438	dev->iommu->attach_deferred = ops->is_attach_deferred(dev);
439	return 0;
440
441	err_unlink:
442	iommu_device_unlink(iommu: iommu_dev, link: dev);
443	err_release:
444	if (ops->release_device)
445	ops->release_device(dev);
446	err_module_put:
447	module_put(module: ops->owner);
448	err_free:
449	dev->iommu->iommu_dev = NULL;
450	dev_iommu_free(dev);
451	return ret;
452	}
453
454	static void iommu_deinit_device(struct device *dev)
455	{
456	struct iommu_group *group = dev->iommu_group;
457	const struct iommu_ops *ops = dev_iommu_ops(dev);
458
459	lockdep_assert_held(&group->mutex);
460
461	iommu_device_unlink(iommu: dev->iommu->iommu_dev, link: dev);
462
463	/*
464	* release_device() must stop using any attached domain on the device.
465	* If there are still other devices in the group, they are not affected
466	* by this callback.
467	*
468	* If the iommu driver provides release_domain, the core code ensures
469	* that domain is attached prior to calling release_device. Drivers can
470	* use this to enforce a translation on the idle iommu. Typically, the
471	* global static blocked_domain is a good choice.
472	*
473	* Otherwise, the iommu driver must set the device to either an identity
474	* or a blocking translation in release_device() and stop using any
475	* domain pointer, as it is going to be freed.
476	*
477	* Regardless, if a delayed attach never occurred, then the release
478	* should still avoid touching any hardware configuration either.
479	*/
480	if (!dev->iommu->attach_deferred && ops->release_domain)
481	ops->release_domain->ops->attach_dev(ops->release_domain, dev);
482
483	if (ops->release_device)
484	ops->release_device(dev);
485
486	/*
487	* If this is the last driver to use the group then we must free the
488	* domains before we do the module_put().
489	*/
490	if (list_empty(head: &group->devices)) {
491	if (group->default_domain) {
492	iommu_domain_free(domain: group->default_domain);
493	group->default_domain = NULL;
494	}
495	if (group->blocking_domain) {
496	iommu_domain_free(domain: group->blocking_domain);
497	group->blocking_domain = NULL;
498	}
499	group->domain = NULL;
500	}
501
502	/* Caller must put iommu_group */
503	dev->iommu_group = NULL;
504	module_put(module: ops->owner);
505	dev_iommu_free(dev);
506	}
507
508	DEFINE_MUTEX(iommu_probe_device_lock);
509
510	static int __iommu_probe_device(struct device *dev, struct list_head *group_list)
511	{
512	const struct iommu_ops *ops;
513	struct iommu_fwspec *fwspec;
514	struct iommu_group *group;
515	struct group_device *gdev;
516	int ret;
517
518	/*
519	* For FDT-based systems and ACPI IORT/VIOT, drivers register IOMMU
520	* instances with non-NULL fwnodes, and client devices should have been
521	* identified with a fwspec by this point. Otherwise, we can currently
522	* assume that only one of Intel, AMD, s390, PAMU or legacy SMMUv2 can
523	* be present, and that any of their registered instances has suitable
524	* ops for probing, and thus cheekily co-opt the same mechanism.
525	*/
526	fwspec = dev_iommu_fwspec_get(dev);
527	if (fwspec && fwspec->ops)
528	ops = fwspec->ops;
529	else
530	ops = iommu_ops_from_fwnode(NULL);
531
532	if (!ops)
533	return -ENODEV;
534	/*
535	* Serialise to avoid races between IOMMU drivers registering in
536	* parallel and/or the "replay" calls from ACPI/OF code via client
537	* driver probe. Once the latter have been cleaned up we should
538	* probably be able to use device_lock() here to minimise the scope,
539	* but for now enforcing a simple global ordering is fine.
540	*/
541	lockdep_assert_held(&iommu_probe_device_lock);
542
543	/* Device is probed already if in a group */
544	if (dev->iommu_group)
545	return 0;
546
547	ret = iommu_init_device(dev, ops);
548	if (ret)
549	return ret;
550
551	group = dev->iommu_group;
552	gdev = iommu_group_alloc_device(group, dev);
553	mutex_lock(&group->mutex);
554	if (IS_ERR(ptr: gdev)) {
555	ret = PTR_ERR(ptr: gdev);
556	goto err_put_group;
557	}
558
559	/*
560	* The gdev must be in the list before calling
561	* iommu_setup_default_domain()
562	*/
563	list_add_tail(new: &gdev->list, head: &group->devices);
564	WARN_ON(group->default_domain && !group->domain);
565	if (group->default_domain)
566	iommu_create_device_direct_mappings(domain: group->default_domain, dev);
567	if (group->domain) {
568	ret = __iommu_device_set_domain(group, dev, new_domain: group->domain, flags: 0);
569	if (ret)
570	goto err_remove_gdev;
571	} else if (!group->default_domain && !group_list) {
572	ret = iommu_setup_default_domain(group, target_type: 0);
573	if (ret)
574	goto err_remove_gdev;
575	} else if (!group->default_domain) {
576	/*
577	* With a group_list argument we defer the default_domain setup
578	* to the caller by providing a de-duplicated list of groups
579	* that need further setup.
580	*/
581	if (list_empty(head: &group->entry))
582	list_add_tail(new: &group->entry, head: group_list);
583	}
584	mutex_unlock(lock: &group->mutex);
585
586	if (dev_is_pci(dev))
587	iommu_dma_set_pci_32bit_workaround(dev);
588
589	return 0;
590
591	err_remove_gdev:
592	list_del(entry: &gdev->list);
593	__iommu_group_free_device(group, grp_dev: gdev);
594	err_put_group:
595	iommu_deinit_device(dev);
596	mutex_unlock(lock: &group->mutex);
597	iommu_group_put(group);
598
599	return ret;
600	}
601
602	int 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(lock: &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
620	static 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(kobj: group->devices_kobj, name: grp_dev->name);
626	sysfs_remove_link(kobj: &dev->kobj, name: "iommu_group");
627
628	trace_remove_device_from_group(group_id: 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(head: &group->devices))
636	WARN_ON(group->owner_cnt ||
637	group->domain != group->default_domain);
638
639	kfree(objp: grp_dev->name);
640	kfree(objp: grp_dev);
641	}
642
643	/* Remove the iommu_group from the struct device. */
644	static 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(entry: &device->list);
655	__iommu_group_free_device(group, grp_dev: 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(lock: &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
671	static 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
683	static int __init iommu_set_def_domain_type(char *str)
684	{
685	bool pt;
686	int ret;
687
688	ret = kstrtobool(s: str, res: &pt);
689	if (ret)
690	return ret;
691
692	if (pt)
693	iommu_set_default_passthrough(cmd_line: true);
694	else
695	iommu_set_default_translated(cmd_line: true);
696
697	return 0;
698	}
699	early_param("iommu.passthrough", iommu_set_def_domain_type);
700
701	static int __init iommu_dma_setup(char *str)
702	{
703	int ret = kstrtobool(s: str, res: &iommu_dma_strict);
704
705	if (!ret)
706	iommu_cmd_line |= IOMMU_CMD_LINE_STRICT;
707	return ret;
708	}
709	early_param("iommu.strict", iommu_dma_setup);
710
711	void 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
718	static 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
730	static 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
743	static const struct sysfs_ops iommu_group_sysfs_ops = {
744	.show = iommu_group_attr_show,
745	.store = iommu_group_attr_store,
746	};
747
748	static int iommu_group_create_file(struct iommu_group *group,
749	struct iommu_group_attribute *attr)
750	{
751	return sysfs_create_file(kobj: &group->kobj, attr: &attr->attr);
752	}
753
754	static void iommu_group_remove_file(struct iommu_group *group,
755	struct iommu_group_attribute *attr)
756	{
757	sysfs_remove_file(kobj: &group->kobj, attr: &attr->attr);
758	}
759
760	static ssize_t iommu_group_show_name(struct iommu_group *group, char *buf)
761	{
762	return sysfs_emit(buf, fmt: "%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	*/
774	static 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(start: new->start, length: new->length,
781	prot: new->prot, type: 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(new: &nr->list, head: &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(list: &iter->list, head: &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(list: &iter->list, head: &stack);
809	continue;
810
811	check_overlap:
812	top_end = top->start + top->length - 1;
813
814	if (iter->start > top_end + 1) {
815	list_move_tail(list: &iter->list, head: &stack);
816	} else {
817	top->length = max(top_end, iter_end) - top->start + 1;
818	list_del(entry: &iter->list);
819	kfree(objp: iter);
820	}
821	}
822	list_splice(list: &stack, head: regions);
823	return 0;
824	}
825
826	static int
827	iommu_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(new: entry, regions: group_resv_regions);
835	if (ret)
836	break;
837	}
838	return ret;
839	}
840
841	int 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(dev: device->dev))
856	break;
857
858	INIT_LIST_HEAD(list: &dev_resv_regions);
859	iommu_get_resv_regions(dev: device->dev, list: &dev_resv_regions);
860	ret = iommu_insert_device_resv_regions(dev_resv_regions: &dev_resv_regions, group_resv_regions: head);
861	iommu_put_resv_regions(dev: device->dev, list: &dev_resv_regions);
862	if (ret)
863	break;
864	}
865	mutex_unlock(lock: &group->mutex);
866	return ret;
867	}
868	EXPORT_SYMBOL_GPL(iommu_get_group_resv_regions);
869
870	static 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(list: &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, at: offset, fmt: "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(objp: region);
887	}
888
889	return offset;
890	}
891
892	static 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(lock: &group->mutex);
918
919	return sysfs_emit(buf, fmt: "%s\n", type);
920	}
921
922	static IOMMU_GROUP_ATTR(name, S_IRUGO, iommu_group_show_name, NULL);
923
924	static IOMMU_GROUP_ATTR(reserved_regions, 0444,
925	iommu_group_show_resv_regions, NULL);
926
927	static IOMMU_GROUP_ATTR(type, 0644, iommu_group_show_type,
928	iommu_group_store_type);
929
930	static 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, id: 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(objp: group->name);
946	kfree(objp: group);
947	}
948
949	static 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	*/
965	struct iommu_group *iommu_group_alloc(void)
966	{
967	struct iommu_group *group;
968	int ret;
969
970	group = kzalloc(size: sizeof(*group), GFP_KERNEL);
971	if (!group)
972	return ERR_PTR(error: -ENOMEM);
973
974	group->kobj.kset = iommu_group_kset;
975	mutex_init(&group->mutex);
976	INIT_LIST_HEAD(list: &group->devices);
977	INIT_LIST_HEAD(list: &group->entry);
978	xa_init(xa: &group->pasid_array);
979
980	ret = ida_alloc(ida: &iommu_group_ida, GFP_KERNEL);
981	if (ret < 0) {
982	kfree(objp: group);
983	return ERR_PTR(error: ret);
984	}
985	group->id = ret;
986
987	ret = kobject_init_and_add(kobj: &group->kobj, ktype: &iommu_group_ktype,
988	NULL, fmt: "%d", group->id);
989	if (ret) {
990	kobject_put(kobj: &group->kobj);
991	return ERR_PTR(error: ret);
992	}
993
994	group->devices_kobj = kobject_create_and_add(name: "devices", parent: &group->kobj);
995	if (!group->devices_kobj) {
996	kobject_put(kobj: &group->kobj); /* triggers .release & free */
997	return ERR_PTR(error: -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(kobj: &group->kobj);
1006
1007	ret = iommu_group_create_file(group,
1008	attr: &iommu_group_attr_reserved_regions);
1009	if (ret) {
1010	kobject_put(kobj: group->devices_kobj);
1011	return ERR_PTR(error: ret);
1012	}
1013
1014	ret = iommu_group_create_file(group, attr: &iommu_group_attr_type);
1015	if (ret) {
1016	kobject_put(kobj: group->devices_kobj);
1017	return ERR_PTR(error: ret);
1018	}
1019
1020	pr_debug("Allocated group %d\n", group->id);
1021
1022	return group;
1023	}
1024	EXPORT_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	*/
1034	void *iommu_group_get_iommudata(struct iommu_group *group)
1035	{
1036	return group->iommu_data;
1037	}
1038	EXPORT_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	*/
1050	void 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	}
1056	EXPORT_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	*/
1066	int 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, attr: &iommu_group_attr_name);
1072	kfree(objp: group->name);
1073	group->name = NULL;
1074	if (!name)
1075	return 0;
1076	}
1077
1078	group->name = kstrdup(s: name, GFP_KERNEL);
1079	if (!group->name)
1080	return -ENOMEM;
1081
1082	ret = iommu_group_create_file(group, attr: &iommu_group_attr_name);
1083	if (ret) {
1084	kfree(objp: group->name);
1085	group->name = NULL;
1086	return ret;
1087	}
1088
1089	return 0;
1090	}
1091	EXPORT_SYMBOL_GPL(iommu_group_set_name);
1092
1093	static 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(list: &mappings);
1103
1104	if (WARN_ON_ONCE(iommu_is_dma_domain(domain) && !pg_size))
1105	return -EINVAL;
1106
1107	iommu_get_resv_regions(dev, list: &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, iova: addr);
1132	if (!phys_addr) {
1133	map_size += pg_size;
1134	continue;
1135	}
1136
1137	map_end:
1138	if (map_size) {
1139	ret = iommu_map(domain, iova: addr - map_size,
1140	paddr: addr - map_size, size: map_size,
1141	prot: entry->prot, GFP_KERNEL);
1142	if (ret)
1143	goto out;
1144	map_size = 0;
1145	}
1146	}
1147
1148	}
1149
1150	if (!list_empty(head: &mappings) && iommu_is_dma_domain(domain))
1151	iommu_flush_iotlb_all(domain);
1152
1153	out:
1154	iommu_put_resv_regions(dev, list: &mappings);
1155
1156	return ret;
1157	}
1158
1159	/* This is undone by __iommu_group_free_device() */
1160	static struct group_device *iommu_group_alloc_device(struct iommu_group *group,
1161	struct device *dev)
1162	{
1163	int ret, i = 0;
1164	struct group_device *device;
1165
1166	device = kzalloc(size: sizeof(*device), GFP_KERNEL);
1167	if (!device)
1168	return ERR_PTR(error: -ENOMEM);
1169
1170	device->dev = dev;
1171
1172	ret = sysfs_create_link(kobj: &dev->kobj, target: &group->kobj, name: "iommu_group");
1173	if (ret)
1174	goto err_free_device;
1175
1176	device->name = kasprintf(GFP_KERNEL, fmt: "%s", kobject_name(kobj: &dev->kobj));
1177	rename:
1178	if (!device->name) {
1179	ret = -ENOMEM;
1180	goto err_remove_link;
1181	}
1182
1183	ret = sysfs_create_link_nowarn(kobj: group->devices_kobj,
1184	target: &dev->kobj, name: device->name);
1185	if (ret) {
1186	if (ret == -EEXIST && i >= 0) {
1187	/*
1188	* Account for the slim chance of collision
1189	* and append an instance to the name.
1190	*/
1191	kfree(objp: device->name);
1192	device->name = kasprintf(GFP_KERNEL, fmt: "%s.%d",
1193	kobject_name(kobj: &dev->kobj), i++);
1194	goto rename;
1195	}
1196	goto err_free_name;
1197	}
1198
1199	trace_add_device_to_group(group_id: group->id, dev);
1200
1201	dev_info(dev, "Adding to iommu group %d\n", group->id);
1202
1203	return device;
1204
1205	err_free_name:
1206	kfree(objp: device->name);
1207	err_remove_link:
1208	sysfs_remove_link(kobj: &dev->kobj, name: "iommu_group");
1209	err_free_device:
1210	kfree(objp: device);
1211	dev_err(dev, "Failed to add to iommu group %d: %d\n", group->id, ret);
1212	return ERR_PTR(error: ret);
1213	}
1214
1215	/**
1216	* iommu_group_add_device - add a device to an iommu group
1217	* @group: the group into which to add the device (reference should be held)
1218	* @dev: the device
1219	*
1220	* This function is called by an iommu driver to add a device into a
1221	* group. Adding a device increments the group reference count.
1222	*/
1223	int iommu_group_add_device(struct iommu_group *group, struct device *dev)
1224	{
1225	struct group_device *gdev;
1226
1227	gdev = iommu_group_alloc_device(group, dev);
1228	if (IS_ERR(ptr: gdev))
1229	return PTR_ERR(ptr: gdev);
1230
1231	iommu_group_ref_get(group);
1232	dev->iommu_group = group;
1233
1234	mutex_lock(&group->mutex);
1235	list_add_tail(new: &gdev->list, head: &group->devices);
1236	mutex_unlock(lock: &group->mutex);
1237	return 0;
1238	}
1239	EXPORT_SYMBOL_GPL(iommu_group_add_device);
1240
1241	/**
1242	* iommu_group_remove_device - remove a device from it's current group
1243	* @dev: device to be removed
1244	*
1245	* This function is called by an iommu driver to remove the device from
1246	* it's current group. This decrements the iommu group reference count.
1247	*/
1248	void iommu_group_remove_device(struct device *dev)
1249	{
1250	struct iommu_group *group = dev->iommu_group;
1251
1252	if (!group)
1253	return;
1254
1255	dev_info(dev, "Removing from iommu group %d\n", group->id);
1256
1257	__iommu_group_remove_device(dev);
1258	}
1259	EXPORT_SYMBOL_GPL(iommu_group_remove_device);
1260
1261	#if IS_ENABLED(CONFIG_LOCKDEP) && IS_ENABLED(CONFIG_IOMMU_API)
1262	/**
1263	* iommu_group_mutex_assert - Check device group mutex lock
1264	* @dev: the device that has group param set
1265	*
1266	* This function is called by an iommu driver to check whether it holds
1267	* group mutex lock for the given device or not.
1268	*
1269	* Note that this function must be called after device group param is set.
1270	*/
1271	void iommu_group_mutex_assert(struct device *dev)
1272	{
1273	struct iommu_group *group = dev->iommu_group;
1274
1275	lockdep_assert_held(&group->mutex);
1276	}
1277	EXPORT_SYMBOL_GPL(iommu_group_mutex_assert);
1278	#endif
1279
1280	static struct device *iommu_group_first_dev(struct iommu_group *group)
1281	{
1282	lockdep_assert_held(&group->mutex);
1283	return list_first_entry(&group->devices, struct group_device, list)->dev;
1284	}
1285
1286	/**
1287	* iommu_group_for_each_dev - iterate over each device in the group
1288	* @group: the group
1289	* @data: caller opaque data to be passed to callback function
1290	* @fn: caller supplied callback function
1291	*
1292	* This function is called by group users to iterate over group devices.
1293	* Callers should hold a reference count to the group during callback.
1294	* The group->mutex is held across callbacks, which will block calls to
1295	* iommu_group_add/remove_device.
1296	*/
1297	int iommu_group_for_each_dev(struct iommu_group *group, void *data,
1298	int (*fn)(struct device *, void *))
1299	{
1300	struct group_device *device;
1301	int ret = 0;
1302
1303	mutex_lock(&group->mutex);
1304	for_each_group_device(group, device) {
1305	ret = fn(device->dev, data);
1306	if (ret)
1307	break;
1308	}
1309	mutex_unlock(lock: &group->mutex);
1310
1311	return ret;
1312	}
1313	EXPORT_SYMBOL_GPL(iommu_group_for_each_dev);
1314
1315	/**
1316	* iommu_group_get - Return the group for a device and increment reference
1317	* @dev: get the group that this device belongs to
1318	*
1319	* This function is called by iommu drivers and users to get the group
1320	* for the specified device. If found, the group is returned and the group
1321	* reference in incremented, else NULL.
1322	*/
1323	struct iommu_group *iommu_group_get(struct device *dev)
1324	{
1325	struct iommu_group *group = dev->iommu_group;
1326
1327	if (group)
1328	kobject_get(kobj: group->devices_kobj);
1329
1330	return group;
1331	}
1332	EXPORT_SYMBOL_GPL(iommu_group_get);
1333
1334	/**
1335	* iommu_group_ref_get - Increment reference on a group
1336	* @group: the group to use, must not be NULL
1337	*
1338	* This function is called by iommu drivers to take additional references on an
1339	* existing group. Returns the given group for convenience.
1340	*/
1341	struct iommu_group *iommu_group_ref_get(struct iommu_group *group)
1342	{
1343	kobject_get(kobj: group->devices_kobj);
1344	return group;
1345	}
1346	EXPORT_SYMBOL_GPL(iommu_group_ref_get);
1347
1348	/**
1349	* iommu_group_put - Decrement group reference
1350	* @group: the group to use
1351	*
1352	* This function is called by iommu drivers and users to release the
1353	* iommu group. Once the reference count is zero, the group is released.
1354	*/
1355	void iommu_group_put(struct iommu_group *group)
1356	{
1357	if (group)
1358	kobject_put(kobj: group->devices_kobj);
1359	}
1360	EXPORT_SYMBOL_GPL(iommu_group_put);
1361
1362	/**
1363	* iommu_group_id - Return ID for a group
1364	* @group: the group to ID
1365	*
1366	* Return the unique ID for the group matching the sysfs group number.
1367	*/
1368	int iommu_group_id(struct iommu_group *group)
1369	{
1370	return group->id;
1371	}
1372	EXPORT_SYMBOL_GPL(iommu_group_id);
1373
1374	static struct iommu_group *get_pci_alias_group(struct pci_dev *pdev,
1375	unsigned long *devfns);
1376
1377	/*
1378	* To consider a PCI device isolated, we require ACS to support Source
1379	* Validation, Request Redirection, Completer Redirection, and Upstream
1380	* Forwarding. This effectively means that devices cannot spoof their
1381	* requester ID, requests and completions cannot be redirected, and all
1382	* transactions are forwarded upstream, even as it passes through a
1383	* bridge where the target device is downstream.
1384	*/
1385	#define REQ_ACS_FLAGS (PCI_ACS_SV | PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_UF)
1386
1387	/*
1388	* For multifunction devices which are not isolated from each other, find
1389	* all the other non-isolated functions and look for existing groups. For
1390	* each function, we also need to look for aliases to or from other devices
1391	* that may already have a group.
1392	*/
1393	static struct iommu_group *get_pci_function_alias_group(struct pci_dev *pdev,
1394	unsigned long *devfns)
1395	{
1396	struct pci_dev *tmp = NULL;
1397	struct iommu_group *group;
1398
1399	if (!pdev->multifunction || pci_acs_enabled(pdev, REQ_ACS_FLAGS))
1400	return NULL;
1401
1402	for_each_pci_dev(tmp) {
1403	if (tmp == pdev || tmp->bus != pdev->bus ||
1404	PCI_SLOT(tmp->devfn) != PCI_SLOT(pdev->devfn) ||
1405	pci_acs_enabled(pdev: tmp, REQ_ACS_FLAGS))
1406	continue;
1407
1408	group = get_pci_alias_group(pdev: tmp, devfns);
1409	if (group) {
1410	pci_dev_put(dev: tmp);
1411	return group;
1412	}
1413	}
1414
1415	return NULL;
1416	}
1417
1418	/*
1419	* Look for aliases to or from the given device for existing groups. DMA
1420	* aliases are only supported on the same bus, therefore the search
1421	* space is quite small (especially since we're really only looking at pcie
1422	* device, and therefore only expect multiple slots on the root complex or
1423	* downstream switch ports). It's conceivable though that a pair of
1424	* multifunction devices could have aliases between them that would cause a
1425	* loop. To prevent this, we use a bitmap to track where we've been.
1426	*/
1427	static struct iommu_group *get_pci_alias_group(struct pci_dev *pdev,
1428	unsigned long *devfns)
1429	{
1430	struct pci_dev *tmp = NULL;
1431	struct iommu_group *group;
1432
1433	if (test_and_set_bit(nr: pdev->devfn & 0xff, addr: devfns))
1434	return NULL;
1435
1436	group = iommu_group_get(&pdev->dev);
1437	if (group)
1438	return group;
1439
1440	for_each_pci_dev(tmp) {
1441	if (tmp == pdev || tmp->bus != pdev->bus)
1442	continue;
1443
1444	/* We alias them or they alias us */
1445	if (pci_devs_are_dma_aliases(dev1: pdev, dev2: tmp)) {
1446	group = get_pci_alias_group(pdev: tmp, devfns);
1447	if (group) {
1448	pci_dev_put(dev: tmp);
1449	return group;
1450	}
1451
1452	group = get_pci_function_alias_group(pdev: tmp, devfns);
1453	if (group) {
1454	pci_dev_put(dev: tmp);
1455	return group;
1456	}
1457	}
1458	}
1459
1460	return NULL;
1461	}
1462
1463	struct group_for_pci_data {
1464	struct pci_dev *pdev;
1465	struct iommu_group *group;
1466	};
1467
1468	/*
1469	* DMA alias iterator callback, return the last seen device. Stop and return
1470	* the IOMMU group if we find one along the way.
1471	*/
1472	static int get_pci_alias_or_group(struct pci_dev *pdev, u16 alias, void *opaque)
1473	{
1474	struct group_for_pci_data *data = opaque;
1475
1476	data->pdev = pdev;
1477	data->group = iommu_group_get(&pdev->dev);
1478
1479	return data->group != NULL;
1480	}
1481
1482	/*
1483	* Generic device_group call-back function. It just allocates one
1484	* iommu-group per device.
1485	*/
1486	struct iommu_group *generic_device_group(struct device *dev)
1487	{
1488	return iommu_group_alloc();
1489	}
1490	EXPORT_SYMBOL_GPL(generic_device_group);
1491
1492	/*
1493	* Generic device_group call-back function. It just allocates one
1494	* iommu-group per iommu driver instance shared by every device
1495	* probed by that iommu driver.
1496	*/
1497	struct iommu_group *generic_single_device_group(struct device *dev)
1498	{
1499	struct iommu_device *iommu = dev->iommu->iommu_dev;
1500
1501	if (!iommu->singleton_group) {
1502	struct iommu_group *group;
1503
1504	group = iommu_group_alloc();
1505	if (IS_ERR(ptr: group))
1506	return group;
1507	iommu->singleton_group = group;
1508	}
1509	return iommu_group_ref_get(iommu->singleton_group);
1510	}
1511	EXPORT_SYMBOL_GPL(generic_single_device_group);
1512
1513	/*
1514	* Use standard PCI bus topology, isolation features, and DMA alias quirks
1515	* to find or create an IOMMU group for a device.
1516	*/
1517	struct iommu_group *pci_device_group(struct device *dev)
1518	{
1519	struct pci_dev *pdev = to_pci_dev(dev);
1520	struct group_for_pci_data data;
1521	struct pci_bus *bus;
1522	struct iommu_group *group = NULL;
1523	u64 devfns[4] = { 0 };
1524
1525	if (WARN_ON(!dev_is_pci(dev)))
1526	return ERR_PTR(error: -EINVAL);
1527
1528	/*
1529	* Find the upstream DMA alias for the device. A device must not
1530	* be aliased due to topology in order to have its own IOMMU group.
1531	* If we find an alias along the way that already belongs to a
1532	* group, use it.
1533	*/
1534	if (pci_for_each_dma_alias(pdev, fn: get_pci_alias_or_group, data: &data))
1535	return data.group;
1536
1537	pdev = data.pdev;
1538
1539	/*
1540	* Continue upstream from the point of minimum IOMMU granularity
1541	* due to aliases to the point where devices are protected from
1542	* peer-to-peer DMA by PCI ACS. Again, if we find an existing
1543	* group, use it.
1544	*/
1545	for (bus = pdev->bus; !pci_is_root_bus(pbus: bus); bus = bus->parent) {
1546	if (!bus->self)
1547	continue;
1548
1549	if (pci_acs_path_enabled(start: bus->self, NULL, REQ_ACS_FLAGS))
1550	break;
1551
1552	pdev = bus->self;
1553
1554	group = iommu_group_get(&pdev->dev);
1555	if (group)
1556	return group;
1557	}
1558
1559	/*
1560	* Look for existing groups on device aliases. If we alias another
1561	* device or another device aliases us, use the same group.
1562	*/
1563	group = get_pci_alias_group(pdev, devfns: (unsigned long *)devfns);
1564	if (group)
1565	return group;
1566
1567	/*
1568	* Look for existing groups on non-isolated functions on the same
1569	* slot and aliases of those funcions, if any. No need to clear
1570	* the search bitmap, the tested devfns are still valid.
1571	*/
1572	group = get_pci_function_alias_group(pdev, devfns: (unsigned long *)devfns);
1573	if (group)
1574	return group;
1575
1576	/* No shared group found, allocate new */
1577	return iommu_group_alloc();
1578	}
1579	EXPORT_SYMBOL_GPL(pci_device_group);
1580
1581	/* Get the IOMMU group for device on fsl-mc bus */
1582	struct iommu_group *fsl_mc_device_group(struct device *dev)
1583	{
1584	struct device *cont_dev = fsl_mc_cont_dev(dev);
1585	struct iommu_group *group;
1586
1587	group = iommu_group_get(cont_dev);
1588	if (!group)
1589	group = iommu_group_alloc();
1590	return group;
1591	}
1592	EXPORT_SYMBOL_GPL(fsl_mc_device_group);
1593
1594	static struct iommu_domain *
1595	__iommu_group_alloc_default_domain(struct iommu_group *group, int req_type)
1596	{
1597	if (group->default_domain && group->default_domain->type == req_type)
1598	return group->default_domain;
1599	return __iommu_group_domain_alloc(group, type: req_type);
1600	}
1601
1602	/*
1603	* req_type of 0 means "auto" which means to select a domain based on
1604	* iommu_def_domain_type or what the driver actually supports.
1605	*/
1606	static struct iommu_domain *
1607	iommu_group_alloc_default_domain(struct iommu_group *group, int req_type)
1608	{
1609	const struct iommu_ops *ops = dev_iommu_ops(dev: iommu_group_first_dev(group));
1610	struct iommu_domain *dom;
1611
1612	lockdep_assert_held(&group->mutex);
1613
1614	/*
1615	* Allow legacy drivers to specify the domain that will be the default
1616	* domain. This should always be either an IDENTITY/BLOCKED/PLATFORM
1617	* domain. Do not use in new drivers.
1618	*/
1619	if (ops->default_domain) {
1620	if (req_type != ops->default_domain->type)
1621	return ERR_PTR(error: -EINVAL);
1622	return ops->default_domain;
1623	}
1624
1625	if (req_type)
1626	return __iommu_group_alloc_default_domain(group, req_type);
1627
1628	/* The driver gave no guidance on what type to use, try the default */
1629	dom = __iommu_group_alloc_default_domain(group, req_type: iommu_def_domain_type);
1630	if (!IS_ERR(ptr: dom))
1631	return dom;
1632
1633	/* Otherwise IDENTITY and DMA_FQ defaults will try DMA */
1634	if (iommu_def_domain_type == IOMMU_DOMAIN_DMA)
1635	return ERR_PTR(error: -EINVAL);
1636	dom = __iommu_group_alloc_default_domain(group, IOMMU_DOMAIN_DMA);
1637	if (IS_ERR(ptr: dom))
1638	return dom;
1639
1640	pr_warn("Failed to allocate default IOMMU domain of type %u for group %s - Falling back to IOMMU_DOMAIN_DMA",
1641	iommu_def_domain_type, group->name);
1642	return dom;
1643	}
1644
1645	struct iommu_domain *iommu_group_default_domain(struct iommu_group *group)
1646	{
1647	return group->default_domain;
1648	}
1649
1650	static int probe_iommu_group(struct device *dev, void *data)
1651	{
1652	struct list_head *group_list = data;
1653	int ret;
1654
1655	mutex_lock(&iommu_probe_device_lock);
1656	ret = __iommu_probe_device(dev, group_list);
1657	mutex_unlock(lock: &iommu_probe_device_lock);
1658	if (ret == -ENODEV)
1659	ret = 0;
1660
1661	return ret;
1662	}
1663
1664	static int iommu_bus_notifier(struct notifier_block *nb,
1665	unsigned long action, void *data)
1666	{
1667	struct device *dev = data;
1668
1669	if (action == BUS_NOTIFY_ADD_DEVICE) {
1670	int ret;
1671
1672	ret = iommu_probe_device(dev);
1673	return (ret) ? NOTIFY_DONE : NOTIFY_OK;
1674	} else if (action == BUS_NOTIFY_REMOVED_DEVICE) {
1675	iommu_release_device(dev);
1676	return NOTIFY_OK;
1677	}
1678
1679	return 0;
1680	}
1681
1682	/*
1683	* Combine the driver's chosen def_domain_type across all the devices in a
1684	* group. Drivers must give a consistent result.
1685	*/
1686	static int iommu_get_def_domain_type(struct iommu_group *group,
1687	struct device *dev, int cur_type)
1688	{
1689	const struct iommu_ops *ops = dev_iommu_ops(dev);
1690	int type;
1691
1692	if (ops->default_domain) {
1693	/*
1694	* Drivers that declare a global static default_domain will
1695	* always choose that.
1696	*/
1697	type = ops->default_domain->type;
1698	} else {
1699	if (ops->def_domain_type)
1700	type = ops->def_domain_type(dev);
1701	else
1702	return cur_type;
1703	}
1704	if (!type || cur_type == type)
1705	return cur_type;
1706	if (!cur_type)
1707	return type;
1708
1709	dev_err_ratelimited(
1710	dev,
1711	"IOMMU driver error, requesting conflicting def_domain_type, %s and %s, for devices in group %u.\n",
1712	iommu_domain_type_str(cur_type), iommu_domain_type_str(type),
1713	group->id);
1714
1715	/*
1716	* Try to recover, drivers are allowed to force IDENITY or DMA, IDENTITY
1717	* takes precedence.
1718	*/
1719	if (type == IOMMU_DOMAIN_IDENTITY)
1720	return type;
1721	return cur_type;
1722	}
1723
1724	/*
1725	* A target_type of 0 will select the best domain type. 0 can be returned in
1726	* this case meaning the global default should be used.
1727	*/
1728	static int iommu_get_default_domain_type(struct iommu_group *group,
1729	int target_type)
1730	{
1731	struct device *untrusted = NULL;
1732	struct group_device *gdev;
1733	int driver_type = 0;
1734
1735	lockdep_assert_held(&group->mutex);
1736
1737	/*
1738	* ARM32 drivers supporting CONFIG_ARM_DMA_USE_IOMMU can declare an
1739	* identity_domain and it will automatically become their default
1740	* domain. Later on ARM_DMA_USE_IOMMU will install its UNMANAGED domain.
1741	* Override the selection to IDENTITY.
1742	*/
1743	if (IS_ENABLED(CONFIG_ARM_DMA_USE_IOMMU)) {
1744	static_assert(!(IS_ENABLED(CONFIG_ARM_DMA_USE_IOMMU) &&
1745	IS_ENABLED(CONFIG_IOMMU_DMA)));
1746	driver_type = IOMMU_DOMAIN_IDENTITY;
1747	}
1748
1749	for_each_group_device(group, gdev) {
1750	driver_type = iommu_get_def_domain_type(group, dev: gdev->dev,
1751	cur_type: driver_type);
1752
1753	if (dev_is_pci(gdev->dev) && to_pci_dev(gdev->dev)->untrusted) {
1754	/*
1755	* No ARM32 using systems will set untrusted, it cannot
1756	* work.
1757	*/
1758	if (WARN_ON(IS_ENABLED(CONFIG_ARM_DMA_USE_IOMMU)))
1759	return -1;
1760	untrusted = gdev->dev;
1761	}
1762	}
1763
1764	/*
1765	* If the common dma ops are not selected in kconfig then we cannot use
1766	* IOMMU_DOMAIN_DMA at all. Force IDENTITY if nothing else has been
1767	* selected.
1768	*/
1769	if (!IS_ENABLED(CONFIG_IOMMU_DMA)) {
1770	if (WARN_ON(driver_type == IOMMU_DOMAIN_DMA))
1771	return -1;
1772	if (!driver_type)
1773	driver_type = IOMMU_DOMAIN_IDENTITY;
1774	}
1775
1776	if (untrusted) {
1777	if (driver_type && driver_type != IOMMU_DOMAIN_DMA) {
1778	dev_err_ratelimited(
1779	untrusted,
1780	"Device is not trusted, but driver is overriding group %u to %s, refusing to probe.\n",
1781	group->id, iommu_domain_type_str(driver_type));
1782	return -1;
1783	}
1784	driver_type = IOMMU_DOMAIN_DMA;
1785	}
1786
1787	if (target_type) {
1788	if (driver_type && target_type != driver_type)
1789	return -1;
1790	return target_type;
1791	}
1792	return driver_type;
1793	}
1794
1795	static void iommu_group_do_probe_finalize(struct device *dev)
1796	{
1797	const struct iommu_ops *ops = dev_iommu_ops(dev);
1798
1799	if (ops->probe_finalize)
1800	ops->probe_finalize(dev);
1801	}
1802
1803	int bus_iommu_probe(const struct bus_type *bus)
1804	{
1805	struct iommu_group *group, *next;
1806	LIST_HEAD(group_list);
1807	int ret;
1808
1809	ret = bus_for_each_dev(bus, NULL, data: &group_list, fn: probe_iommu_group);
1810	if (ret)
1811	return ret;
1812
1813	list_for_each_entry_safe(group, next, &group_list, entry) {
1814	struct group_device *gdev;
1815
1816	mutex_lock(&group->mutex);
1817
1818	/* Remove item from the list */
1819	list_del_init(entry: &group->entry);
1820
1821	/*
1822	* We go to the trouble of deferred default domain creation so
1823	* that the cross-group default domain type and the setup of the
1824	* IOMMU_RESV_DIRECT will work correctly in non-hotpug scenarios.
1825	*/
1826	ret = iommu_setup_default_domain(group, target_type: 0);
1827	if (ret) {
1828	mutex_unlock(lock: &group->mutex);
1829	return ret;
1830	}
1831	mutex_unlock(lock: &group->mutex);
1832
1833	/*
1834	* FIXME: Mis-locked because the ops->probe_finalize() call-back
1835	* of some IOMMU drivers calls arm_iommu_attach_device() which
1836	* in-turn might call back into IOMMU core code, where it tries
1837	* to take group->mutex, resulting in a deadlock.
1838	*/
1839	for_each_group_device(group, gdev)
1840	iommu_group_do_probe_finalize(dev: gdev->dev);
1841	}
1842
1843	return 0;
1844	}
1845
1846	/**
1847	* iommu_present() - make platform-specific assumptions about an IOMMU
1848	* @bus: bus to check
1849	*
1850	* Do not use this function. You want device_iommu_mapped() instead.
1851	*
1852	* Return: true if some IOMMU is present and aware of devices on the given bus;
1853	* in general it may not be the only IOMMU, and it may not have anything to do
1854	* with whatever device you are ultimately interested in.
1855	*/
1856	bool iommu_present(const struct bus_type *bus)
1857	{
1858	bool ret = false;
1859
1860	for (int i = 0; i < ARRAY_SIZE(iommu_buses); i++) {
1861	if (iommu_buses[i] == bus) {
1862	spin_lock(lock: &iommu_device_lock);
1863	ret = !list_empty(head: &iommu_device_list);
1864	spin_unlock(lock: &iommu_device_lock);
1865	}
1866	}
1867	return ret;
1868	}
1869	EXPORT_SYMBOL_GPL(iommu_present);
1870
1871	/**
1872	* device_iommu_capable() - check for a general IOMMU capability
1873	* @dev: device to which the capability would be relevant, if available
1874	* @cap: IOMMU capability
1875	*
1876	* Return: true if an IOMMU is present and supports the given capability
1877	* for the given device, otherwise false.
1878	*/
1879	bool device_iommu_capable(struct device *dev, enum iommu_cap cap)
1880	{
1881	const struct iommu_ops *ops;
1882
1883	if (!dev_has_iommu(dev))
1884	return false;
1885
1886	ops = dev_iommu_ops(dev);
1887	if (!ops->capable)
1888	return false;
1889
1890	return ops->capable(dev, cap);
1891	}
1892	EXPORT_SYMBOL_GPL(device_iommu_capable);
1893
1894	/**
1895	* iommu_group_has_isolated_msi() - Compute msi_device_has_isolated_msi()
1896	* for a group
1897	* @group: Group to query
1898	*
1899	* IOMMU groups should not have differing values of
1900	* msi_device_has_isolated_msi() for devices in a group. However nothing
1901	* directly prevents this, so ensure mistakes don't result in isolation failures
1902	* by checking that all the devices are the same.
1903	*/
1904	bool iommu_group_has_isolated_msi(struct iommu_group *group)
1905	{
1906	struct group_device *group_dev;
1907	bool ret = true;
1908
1909	mutex_lock(&group->mutex);
1910	for_each_group_device(group, group_dev)
1911	ret &= msi_device_has_isolated_msi(dev: group_dev->dev);
1912	mutex_unlock(lock: &group->mutex);
1913	return ret;
1914	}
1915	EXPORT_SYMBOL_GPL(iommu_group_has_isolated_msi);
1916
1917	/**
1918	* iommu_set_fault_handler() - set a fault handler for an iommu domain
1919	* @domain: iommu domain
1920	* @handler: fault handler
1921	* @token: user data, will be passed back to the fault handler
1922	*
1923	* This function should be used by IOMMU users which want to be notified
1924	* whenever an IOMMU fault happens.
1925	*
1926	* The fault handler itself should return 0 on success, and an appropriate
1927	* error code otherwise.
1928	*/
1929	void iommu_set_fault_handler(struct iommu_domain *domain,
1930	iommu_fault_handler_t handler,
1931	void *token)
1932	{
1933	BUG_ON(!domain);
1934
1935	domain->handler = handler;
1936	domain->handler_token = token;
1937	}
1938	EXPORT_SYMBOL_GPL(iommu_set_fault_handler);
1939
1940	static struct iommu_domain *__iommu_domain_alloc(const struct iommu_ops *ops,
1941	struct device *dev,
1942	unsigned int type)
1943	{
1944	struct iommu_domain *domain;
1945	unsigned int alloc_type = type & IOMMU_DOMAIN_ALLOC_FLAGS;
1946
1947	if (alloc_type == IOMMU_DOMAIN_IDENTITY && ops->identity_domain)
1948	return ops->identity_domain;
1949	else if (alloc_type == IOMMU_DOMAIN_BLOCKED && ops->blocked_domain)
1950	return ops->blocked_domain;
1951	else if (type & __IOMMU_DOMAIN_PAGING && ops->domain_alloc_paging)
1952	domain = ops->domain_alloc_paging(dev);
1953	else if (ops->domain_alloc)
1954	domain = ops->domain_alloc(alloc_type);
1955	else
1956	return ERR_PTR(error: -EOPNOTSUPP);
1957
1958	/*
1959	* Many domain_alloc ops now return ERR_PTR, make things easier for the
1960	* driver by accepting ERR_PTR from all domain_alloc ops instead of
1961	* having two rules.
1962	*/
1963	if (IS_ERR(ptr: domain))
1964	return domain;
1965	if (!domain)
1966	return ERR_PTR(error: -ENOMEM);
1967
1968	domain->type = type;
1969	domain->owner = ops;
1970	/*
1971	* If not already set, assume all sizes by default; the driver
1972	* may override this later
1973	*/
1974	if (!domain->pgsize_bitmap)
1975	domain->pgsize_bitmap = ops->pgsize_bitmap;
1976
1977	if (!domain->ops)
1978	domain->ops = ops->default_domain_ops;
1979
1980	if (iommu_is_dma_domain(domain)) {
1981	int rc;
1982
1983	rc = iommu_get_dma_cookie(domain);
1984	if (rc) {
1985	iommu_domain_free(domain);
1986	return ERR_PTR(error: rc);
1987	}
1988	}
1989	return domain;
1990	}
1991
1992	static struct iommu_domain *
1993	__iommu_group_domain_alloc(struct iommu_group *group, unsigned int type)
1994	{
1995	struct device *dev = iommu_group_first_dev(group);
1996
1997	return __iommu_domain_alloc(ops: dev_iommu_ops(dev), dev, type);
1998	}
1999
2000	static int __iommu_domain_alloc_dev(struct device *dev, void *data)
2001	{
2002	const struct iommu_ops **ops = data;
2003
2004	if (!dev_has_iommu(dev))
2005	return 0;
2006
2007	if (WARN_ONCE(*ops && *ops != dev_iommu_ops(dev),
2008	"Multiple IOMMU drivers present for bus %s, which the public IOMMU API can't fully support yet. You will still need to disable one or more for this to work, sorry!\n",
2009	dev_bus_name(dev)))
2010	return -EBUSY;
2011
2012	*ops = dev_iommu_ops(dev);
2013	return 0;
2014	}
2015
2016	struct iommu_domain *iommu_domain_alloc(const struct bus_type *bus)
2017	{
2018	const struct iommu_ops *ops = NULL;
2019	int err = bus_for_each_dev(bus, NULL, data: &ops, fn: __iommu_domain_alloc_dev);
2020	struct iommu_domain *domain;
2021
2022	if (err || !ops)
2023	return NULL;
2024
2025	domain = __iommu_domain_alloc(ops, NULL, IOMMU_DOMAIN_UNMANAGED);
2026	if (IS_ERR(ptr: domain))
2027	return NULL;
2028	return domain;
2029	}
2030	EXPORT_SYMBOL_GPL(iommu_domain_alloc);
2031
2032	void iommu_domain_free(struct iommu_domain *domain)
2033	{
2034	if (domain->type == IOMMU_DOMAIN_SVA)
2035	mmdrop(mm: domain->mm);
2036	iommu_put_dma_cookie(domain);
2037	if (domain->ops->free)
2038	domain->ops->free(domain);
2039	}
2040	EXPORT_SYMBOL_GPL(iommu_domain_free);
2041
2042	/*
2043	* Put the group's domain back to the appropriate core-owned domain - either the
2044	* standard kernel-mode DMA configuration or an all-DMA-blocked domain.
2045	*/
2046	static void __iommu_group_set_core_domain(struct iommu_group *group)
2047	{
2048	struct iommu_domain *new_domain;
2049
2050	if (group->owner)
2051	new_domain = group->blocking_domain;
2052	else
2053	new_domain = group->default_domain;
2054
2055	__iommu_group_set_domain_nofail(group, new_domain);
2056	}
2057
2058	static int __iommu_attach_device(struct iommu_domain *domain,
2059	struct device *dev)
2060	{
2061	int ret;
2062
2063	if (unlikely(domain->ops->attach_dev == NULL))
2064	return -ENODEV;
2065
2066	ret = domain->ops->attach_dev(domain, dev);
2067	if (ret)
2068	return ret;
2069	dev->iommu->attach_deferred = 0;
2070	trace_attach_device_to_domain(dev);
2071	return 0;
2072	}
2073
2074	/**
2075	* iommu_attach_device - Attach an IOMMU domain to a device
2076	* @domain: IOMMU domain to attach
2077	* @dev: Device that will be attached
2078	*
2079	* Returns 0 on success and error code on failure
2080	*
2081	* Note that EINVAL can be treated as a soft failure, indicating
2082	* that certain configuration of the domain is incompatible with
2083	* the device. In this case attaching a different domain to the
2084	* device may succeed.
2085	*/
2086	int iommu_attach_device(struct iommu_domain *domain, struct device *dev)
2087	{
2088	/* Caller must be a probed driver on dev */
2089	struct iommu_group *group = dev->iommu_group;
2090	int ret;
2091
2092	if (!group)
2093	return -ENODEV;
2094
2095	/*
2096	* Lock the group to make sure the device-count doesn't
2097	* change while we are attaching
2098	*/
2099	mutex_lock(&group->mutex);
2100	ret = -EINVAL;
2101	if (list_count_nodes(head: &group->devices) != 1)
2102	goto out_unlock;
2103
2104	ret = __iommu_attach_group(domain, group);
2105
2106	out_unlock:
2107	mutex_unlock(lock: &group->mutex);
2108	return ret;
2109	}
2110	EXPORT_SYMBOL_GPL(iommu_attach_device);
2111
2112	int iommu_deferred_attach(struct device *dev, struct iommu_domain *domain)
2113	{
2114	if (dev->iommu && dev->iommu->attach_deferred)
2115	return __iommu_attach_device(domain, dev);
2116
2117	return 0;
2118	}
2119
2120	void iommu_detach_device(struct iommu_domain *domain, struct device *dev)
2121	{
2122	/* Caller must be a probed driver on dev */
2123	struct iommu_group *group = dev->iommu_group;
2124
2125	if (!group)
2126	return;
2127
2128	mutex_lock(&group->mutex);
2129	if (WARN_ON(domain != group->domain) ||
2130	WARN_ON(list_count_nodes(&group->devices) != 1))
2131	goto out_unlock;
2132	__iommu_group_set_core_domain(group);
2133
2134	out_unlock:
2135	mutex_unlock(lock: &group->mutex);
2136	}
2137	EXPORT_SYMBOL_GPL(iommu_detach_device);
2138
2139	struct iommu_domain *iommu_get_domain_for_dev(struct device *dev)
2140	{
2141	/* Caller must be a probed driver on dev */
2142	struct iommu_group *group = dev->iommu_group;
2143
2144	if (!group)
2145	return NULL;
2146
2147	return group->domain;
2148	}
2149	EXPORT_SYMBOL_GPL(iommu_get_domain_for_dev);
2150
2151	/*
2152	* For IOMMU_DOMAIN_DMA implementations which already provide their own
2153	* guarantees that the group and its default domain are valid and correct.
2154	*/
2155	struct iommu_domain *iommu_get_dma_domain(struct device *dev)
2156	{
2157	return dev->iommu_group->default_domain;
2158	}
2159
2160	static int __iommu_attach_group(struct iommu_domain *domain,
2161	struct iommu_group *group)
2162	{
2163	struct device *dev;
2164
2165	if (group->domain && group->domain != group->default_domain &&
2166	group->domain != group->blocking_domain)
2167	return -EBUSY;
2168
2169	dev = iommu_group_first_dev(group);
2170	if (!dev_has_iommu(dev) || dev_iommu_ops(dev) != domain->owner)
2171	return -EINVAL;
2172
2173	return __iommu_group_set_domain(group, new_domain: domain);
2174	}
2175
2176	/**
2177	* iommu_attach_group - Attach an IOMMU domain to an IOMMU group
2178	* @domain: IOMMU domain to attach
2179	* @group: IOMMU group that will be attached
2180	*
2181	* Returns 0 on success and error code on failure
2182	*
2183	* Note that EINVAL can be treated as a soft failure, indicating
2184	* that certain configuration of the domain is incompatible with
2185	* the group. In this case attaching a different domain to the
2186	* group may succeed.
2187	*/
2188	int iommu_attach_group(struct iommu_domain *domain, struct iommu_group *group)
2189	{
2190	int ret;
2191
2192	mutex_lock(&group->mutex);
2193	ret = __iommu_attach_group(domain, group);
2194	mutex_unlock(lock: &group->mutex);
2195
2196	return ret;
2197	}
2198	EXPORT_SYMBOL_GPL(iommu_attach_group);
2199
2200	/**
2201	* iommu_group_replace_domain - replace the domain that a group is attached to
2202	* @new_domain: new IOMMU domain to replace with
2203	* @group: IOMMU group that will be attached to the new domain
2204	*
2205	* This API allows the group to switch domains without being forced to go to
2206	* the blocking domain in-between.
2207	*
2208	* If the currently attached domain is a core domain (e.g. a default_domain),
2209	* it will act just like the iommu_attach_group().
2210	*/
2211	int iommu_group_replace_domain(struct iommu_group *group,
2212	struct iommu_domain *new_domain)
2213	{
2214	int ret;
2215
2216	if (!new_domain)
2217	return -EINVAL;
2218
2219	mutex_lock(&group->mutex);
2220	ret = __iommu_group_set_domain(group, new_domain);
2221	mutex_unlock(lock: &group->mutex);
2222	return ret;
2223	}
2224	EXPORT_SYMBOL_NS_GPL(iommu_group_replace_domain, IOMMUFD_INTERNAL);
2225
2226	static int __iommu_device_set_domain(struct iommu_group *group,
2227	struct device *dev,
2228	struct iommu_domain *new_domain,
2229	unsigned int flags)
2230	{
2231	int ret;
2232
2233	/*
2234	* If the device requires IOMMU_RESV_DIRECT then we cannot allow
2235	* the blocking domain to be attached as it does not contain the
2236	* required 1:1 mapping. This test effectively excludes the device
2237	* being used with iommu_group_claim_dma_owner() which will block
2238	* vfio and iommufd as well.
2239	*/
2240	if (dev->iommu->require_direct &&
2241	(new_domain->type == IOMMU_DOMAIN_BLOCKED ||
2242	new_domain == group->blocking_domain)) {
2243	dev_warn(dev,
2244	"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");
2245	return -EINVAL;
2246	}
2247
2248	if (dev->iommu->attach_deferred) {
2249	if (new_domain == group->default_domain)
2250	return 0;
2251	dev->iommu->attach_deferred = 0;
2252	}
2253
2254	ret = __iommu_attach_device(domain: new_domain, dev);
2255	if (ret) {
2256	/*
2257	* If we have a blocking domain then try to attach that in hopes
2258	* of avoiding a UAF. Modern drivers should implement blocking
2259	* domains as global statics that cannot fail.
2260	*/
2261	if ((flags & IOMMU_SET_DOMAIN_MUST_SUCCEED) &&
2262	group->blocking_domain &&
2263	group->blocking_domain != new_domain)
2264	__iommu_attach_device(domain: group->blocking_domain, dev);
2265	return ret;
2266	}
2267	return 0;
2268	}
2269
2270	/*
2271	* If 0 is returned the group's domain is new_domain. If an error is returned
2272	* then the group's domain will be set back to the existing domain unless
2273	* IOMMU_SET_DOMAIN_MUST_SUCCEED, otherwise an error is returned and the group's
2274	* domains is left inconsistent. This is a driver bug to fail attach with a
2275	* previously good domain. We try to avoid a kernel UAF because of this.
2276	*
2277	* IOMMU groups are really the natural working unit of the IOMMU, but the IOMMU
2278	* API works on domains and devices. Bridge that gap by iterating over the
2279	* devices in a group. Ideally we'd have a single device which represents the
2280	* requestor ID of the group, but we also allow IOMMU drivers to create policy
2281	* defined minimum sets, where the physical hardware may be able to distiguish
2282	* members, but we wish to group them at a higher level (ex. untrusted
2283	* multi-function PCI devices). Thus we attach each device.
2284	*/
2285	static int __iommu_group_set_domain_internal(struct iommu_group *group,
2286	struct iommu_domain *new_domain,
2287	unsigned int flags)
2288	{
2289	struct group_device *last_gdev;
2290	struct group_device *gdev;
2291	int result;
2292	int ret;
2293
2294	lockdep_assert_held(&group->mutex);
2295
2296	if (group->domain == new_domain)
2297	return 0;
2298
2299	if (WARN_ON(!new_domain))
2300	return -EINVAL;
2301
2302	/*
2303	* Changing the domain is done by calling attach_dev() on the new
2304	* domain. This switch does not have to be atomic and DMA can be
2305	* discarded during the transition. DMA must only be able to access
2306	* either new_domain or group->domain, never something else.
2307	*/
2308	result = 0;
2309	for_each_group_device(group, gdev) {
2310	ret = __iommu_device_set_domain(group, dev: gdev->dev, new_domain,
2311	flags);
2312	if (ret) {
2313	result = ret;
2314	/*
2315	* Keep trying the other devices in the group. If a
2316	* driver fails attach to an otherwise good domain, and
2317	* does not support blocking domains, it should at least
2318	* drop its reference on the current domain so we don't
2319	* UAF.
2320	*/
2321	if (flags & IOMMU_SET_DOMAIN_MUST_SUCCEED)
2322	continue;
2323	goto err_revert;
2324	}
2325	}
2326	group->domain = new_domain;
2327	return result;
2328
2329	err_revert:
2330	/*
2331	* This is called in error unwind paths. A well behaved driver should
2332	* always allow us to attach to a domain that was already attached.
2333	*/
2334	last_gdev = gdev;
2335	for_each_group_device(group, gdev) {
2336	/*
2337	* A NULL domain can happen only for first probe, in which case
2338	* we leave group->domain as NULL and let release clean
2339	* everything up.
2340	*/
2341	if (group->domain)
2342	WARN_ON(__iommu_device_set_domain(
2343	group, gdev->dev, group->domain,
2344	IOMMU_SET_DOMAIN_MUST_SUCCEED));
2345	if (gdev == last_gdev)
2346	break;
2347	}
2348	return ret;
2349	}
2350
2351	void iommu_detach_group(struct iommu_domain *domain, struct iommu_group *group)
2352	{
2353	mutex_lock(&group->mutex);
2354	__iommu_group_set_core_domain(group);
2355	mutex_unlock(lock: &group->mutex);
2356	}
2357	EXPORT_SYMBOL_GPL(iommu_detach_group);
2358
2359	phys_addr_t iommu_iova_to_phys(struct iommu_domain *domain, dma_addr_t iova)
2360	{
2361	if (domain->type == IOMMU_DOMAIN_IDENTITY)
2362	return iova;
2363
2364	if (domain->type == IOMMU_DOMAIN_BLOCKED)
2365	return 0;
2366
2367	return domain->ops->iova_to_phys(domain, iova);
2368	}
2369	EXPORT_SYMBOL_GPL(iommu_iova_to_phys);
2370
2371	static size_t iommu_pgsize(struct iommu_domain *domain, unsigned long iova,
2372	phys_addr_t paddr, size_t size, size_t *count)
2373	{
2374	unsigned int pgsize_idx, pgsize_idx_next;
2375	unsigned long pgsizes;
2376	size_t offset, pgsize, pgsize_next;
2377	unsigned long addr_merge = paddr | iova;
2378
2379	/* Page sizes supported by the hardware and small enough for @size */
2380	pgsizes = domain->pgsize_bitmap & GENMASK(__fls(size), 0);
2381
2382	/* Constrain the page sizes further based on the maximum alignment */
2383	if (likely(addr_merge))
2384	pgsizes &= GENMASK(__ffs(addr_merge), 0);
2385
2386	/* Make sure we have at least one suitable page size */
2387	BUG_ON(!pgsizes);
2388
2389	/* Pick the biggest page size remaining */
2390	pgsize_idx = __fls(word: pgsizes);
2391	pgsize = BIT(pgsize_idx);
2392	if (!count)
2393	return pgsize;
2394
2395	/* Find the next biggest support page size, if it exists */
2396	pgsizes = domain->pgsize_bitmap & ~GENMASK(pgsize_idx, 0);
2397	if (!pgsizes)
2398	goto out_set_count;
2399
2400	pgsize_idx_next = __ffs(pgsizes);
2401	pgsize_next = BIT(pgsize_idx_next);
2402
2403	/*
2404	* There's no point trying a bigger page size unless the virtual
2405	* and physical addresses are similarly offset within the larger page.
2406	*/
2407	if ((iova ^ paddr) & (pgsize_next - 1))
2408	goto out_set_count;
2409
2410	/* Calculate the offset to the next page size alignment boundary */
2411	offset = pgsize_next - (addr_merge & (pgsize_next - 1));
2412
2413	/*
2414	* If size is big enough to accommodate the larger page, reduce
2415	* the number of smaller pages.
2416	*/
2417	if (offset + pgsize_next <= size)
2418	size = offset;
2419
2420	out_set_count:
2421	*count = size >> pgsize_idx;
2422	return pgsize;
2423	}
2424
2425	static int __iommu_map(struct iommu_domain *domain, unsigned long iova,
2426	phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
2427	{
2428	const struct iommu_domain_ops *ops = domain->ops;
2429	unsigned long orig_iova = iova;
2430	unsigned int min_pagesz;
2431	size_t orig_size = size;
2432	phys_addr_t orig_paddr = paddr;
2433	int ret = 0;
2434
2435	if (unlikely(!(domain->type & __IOMMU_DOMAIN_PAGING)))
2436	return -EINVAL;
2437
2438	if (WARN_ON(!ops->map_pages || domain->pgsize_bitmap == 0UL))
2439	return -ENODEV;
2440
2441	/* find out the minimum page size supported */
2442	min_pagesz = 1 << __ffs(domain->pgsize_bitmap);
2443
2444	/*
2445	* both the virtual address and the physical one, as well as
2446	* the size of the mapping, must be aligned (at least) to the
2447	* size of the smallest page supported by the hardware
2448	*/
2449	if (!IS_ALIGNED(iova | paddr | size, min_pagesz)) {
2450	pr_err("unaligned: iova 0x%lx pa %pa size 0x%zx min_pagesz 0x%x\n",
2451	iova, &paddr, size, min_pagesz);
2452	return -EINVAL;
2453	}
2454
2455	pr_debug("map: iova 0x%lx pa %pa size 0x%zx\n", iova, &paddr, size);
2456
2457	while (size) {
2458	size_t pgsize, count, mapped = 0;
2459
2460	pgsize = iommu_pgsize(domain, iova, paddr, size, count: &count);
2461
2462	pr_debug("mapping: iova 0x%lx pa %pa pgsize 0x%zx count %zu\n",
2463	iova, &paddr, pgsize, count);
2464	ret = ops->map_pages(domain, iova, paddr, pgsize, count, prot,
2465	gfp, &mapped);
2466	/*
2467	* Some pages may have been mapped, even if an error occurred,
2468	* so we should account for those so they can be unmapped.
2469	*/
2470	size -= mapped;
2471
2472	if (ret)
2473	break;
2474
2475	iova += mapped;
2476	paddr += mapped;
2477	}
2478
2479	/* unroll mapping in case something went wrong */
2480	if (ret)
2481	iommu_unmap(domain, iova: orig_iova, size: orig_size - size);
2482	else
2483	trace_map(iova: orig_iova, paddr: orig_paddr, size: orig_size);
2484
2485	return ret;
2486	}
2487
2488	int iommu_map(struct iommu_domain *domain, unsigned long iova,
2489	phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
2490	{
2491	const struct iommu_domain_ops *ops = domain->ops;
2492	int ret;
2493
2494	might_sleep_if(gfpflags_allow_blocking(gfp));
2495
2496	/* Discourage passing strange GFP flags */
2497	if (WARN_ON_ONCE(gfp & (__GFP_COMP | __GFP_DMA | __GFP_DMA32 |
2498	__GFP_HIGHMEM)))
2499	return -EINVAL;
2500
2501	ret = __iommu_map(domain, iova, paddr, size, prot, gfp);
2502	if (ret == 0 && ops->iotlb_sync_map) {
2503	ret = ops->iotlb_sync_map(domain, iova, size);
2504	if (ret)
2505	goto out_err;
2506	}
2507
2508	return ret;
2509
2510	out_err:
2511	/* undo mappings already done */
2512	iommu_unmap(domain, iova, size);
2513
2514	return ret;
2515	}
2516	EXPORT_SYMBOL_GPL(iommu_map);
2517
2518	static size_t __iommu_unmap(struct iommu_domain *domain,
2519	unsigned long iova, size_t size,
2520	struct iommu_iotlb_gather *iotlb_gather)
2521	{
2522	const struct iommu_domain_ops *ops = domain->ops;
2523	size_t unmapped_page, unmapped = 0;
2524	unsigned long orig_iova = iova;
2525	unsigned int min_pagesz;
2526
2527	if (unlikely(!(domain->type & __IOMMU_DOMAIN_PAGING)))
2528	return 0;
2529
2530	if (WARN_ON(!ops->unmap_pages || domain->pgsize_bitmap == 0UL))
2531	return 0;
2532
2533	/* find out the minimum page size supported */
2534	min_pagesz = 1 << __ffs(domain->pgsize_bitmap);
2535
2536	/*
2537	* The virtual address, as well as the size of the mapping, must be
2538	* aligned (at least) to the size of the smallest page supported
2539	* by the hardware
2540	*/
2541	if (!IS_ALIGNED(iova | size, min_pagesz)) {
2542	pr_err("unaligned: iova 0x%lx size 0x%zx min_pagesz 0x%x\n",
2543	iova, size, min_pagesz);
2544	return 0;
2545	}
2546
2547	pr_debug("unmap this: iova 0x%lx size 0x%zx\n", iova, size);
2548
2549	/*
2550	* Keep iterating until we either unmap 'size' bytes (or more)
2551	* or we hit an area that isn't mapped.
2552	*/
2553	while (unmapped < size) {
2554	size_t pgsize, count;
2555
2556	pgsize = iommu_pgsize(domain, iova, paddr: iova, size: size - unmapped, count: &count);
2557	unmapped_page = ops->unmap_pages(domain, iova, pgsize, count, iotlb_gather);
2558	if (!unmapped_page)
2559	break;
2560
2561	pr_debug("unmapped: iova 0x%lx size 0x%zx\n",
2562	iova, unmapped_page);
2563
2564	iova += unmapped_page;
2565	unmapped += unmapped_page;
2566	}
2567
2568	trace_unmap(iova: orig_iova, size, unmapped_size: unmapped);
2569	return unmapped;
2570	}
2571
2572	size_t iommu_unmap(struct iommu_domain *domain,
2573	unsigned long iova, size_t size)
2574	{
2575	struct iommu_iotlb_gather iotlb_gather;
2576	size_t ret;
2577
2578	iommu_iotlb_gather_init(gather: &iotlb_gather);
2579	ret = __iommu_unmap(domain, iova, size, iotlb_gather: &iotlb_gather);
2580	iommu_iotlb_sync(domain, iotlb_gather: &iotlb_gather);
2581
2582	return ret;
2583	}
2584	EXPORT_SYMBOL_GPL(iommu_unmap);
2585
2586	size_t iommu_unmap_fast(struct iommu_domain *domain,
2587	unsigned long iova, size_t size,
2588	struct iommu_iotlb_gather *iotlb_gather)
2589	{
2590	return __iommu_unmap(domain, iova, size, iotlb_gather);
2591	}
2592	EXPORT_SYMBOL_GPL(iommu_unmap_fast);
2593
2594	ssize_t iommu_map_sg(struct iommu_domain *domain, unsigned long iova,
2595	struct scatterlist *sg, unsigned int nents, int prot,
2596	gfp_t gfp)
2597	{
2598	const struct iommu_domain_ops *ops = domain->ops;
2599	size_t len = 0, mapped = 0;
2600	phys_addr_t start;
2601	unsigned int i = 0;
2602	int ret;
2603
2604	might_sleep_if(gfpflags_allow_blocking(gfp));
2605
2606	/* Discourage passing strange GFP flags */
2607	if (WARN_ON_ONCE(gfp & (__GFP_COMP | __GFP_DMA | __GFP_DMA32 |
2608	__GFP_HIGHMEM)))
2609	return -EINVAL;
2610
2611	while (i <= nents) {
2612	phys_addr_t s_phys = sg_phys(sg);
2613
2614	if (len && s_phys != start + len) {
2615	ret = __iommu_map(domain, iova: iova + mapped, paddr: start,
2616	size: len, prot, gfp);
2617
2618	if (ret)
2619	goto out_err;
2620
2621	mapped += len;
2622	len = 0;
2623	}
2624
2625	if (sg_dma_is_bus_address(sg))
2626	goto next;
2627
2628	if (len) {
2629	len += sg->length;
2630	} else {
2631	len = sg->length;
2632	start = s_phys;
2633	}
2634
2635	next:
2636	if (++i < nents)
2637	sg = sg_next(sg);
2638	}
2639
2640	if (ops->iotlb_sync_map) {
2641	ret = ops->iotlb_sync_map(domain, iova, mapped);
2642	if (ret)
2643	goto out_err;
2644	}
2645	return mapped;
2646
2647	out_err:
2648	/* undo mappings already done */
2649	iommu_unmap(domain, iova, mapped);
2650
2651	return ret;
2652	}
2653	EXPORT_SYMBOL_GPL(iommu_map_sg);
2654
2655	/**
2656	* report_iommu_fault() - report about an IOMMU fault to the IOMMU framework
2657	* @domain: the iommu domain where the fault has happened
2658	* @dev: the device where the fault has happened
2659	* @iova: the faulting address
2660	* @flags: mmu fault flags (e.g. IOMMU_FAULT_READ/IOMMU_FAULT_WRITE/...)
2661	*
2662	* This function should be called by the low-level IOMMU implementations
2663	* whenever IOMMU faults happen, to allow high-level users, that are
2664	* interested in such events, to know about them.
2665	*
2666	* This event may be useful for several possible use cases:
2667	* - mere logging of the event
2668	* - dynamic TLB/PTE loading
2669	* - if restarting of the faulting device is required
2670	*
2671	* Returns 0 on success and an appropriate error code otherwise (if dynamic
2672	* PTE/TLB loading will one day be supported, implementations will be able
2673	* to tell whether it succeeded or not according to this return value).
2674	*
2675	* Specifically, -ENOSYS is returned if a fault handler isn't installed
2676	* (though fault handlers can also return -ENOSYS, in case they want to
2677	* elicit the default behavior of the IOMMU drivers).
2678	*/
2679	int report_iommu_fault(struct iommu_domain *domain, struct device *dev,
2680	unsigned long iova, int flags)
2681	{
2682	int ret = -ENOSYS;
2683
2684	/*
2685	* if upper layers showed interest and installed a fault handler,
2686	* invoke it.
2687	*/
2688	if (domain->handler)
2689	ret = domain->handler(domain, dev, iova, flags,
2690	domain->handler_token);
2691
2692	trace_io_page_fault(dev, iova, flags);
2693	return ret;
2694	}
2695	EXPORT_SYMBOL_GPL(report_iommu_fault);
2696
2697	static int __init iommu_init(void)
2698	{
2699	iommu_group_kset = kset_create_and_add(name: "iommu_groups",
2700	NULL, parent_kobj: kernel_kobj);
2701	BUG_ON(!iommu_group_kset);
2702
2703	iommu_debugfs_setup();
2704
2705	return 0;
2706	}
2707	core_initcall(iommu_init);
2708
2709	int iommu_enable_nesting(struct iommu_domain *domain)
2710	{
2711	if (domain->type != IOMMU_DOMAIN_UNMANAGED)
2712	return -EINVAL;
2713	if (!domain->ops->enable_nesting)
2714	return -EINVAL;
2715	return domain->ops->enable_nesting(domain);
2716	}
2717	EXPORT_SYMBOL_GPL(iommu_enable_nesting);
2718
2719	int iommu_set_pgtable_quirks(struct iommu_domain *domain,
2720	unsigned long quirk)
2721	{
2722	if (domain->type != IOMMU_DOMAIN_UNMANAGED)
2723	return -EINVAL;
2724	if (!domain->ops->set_pgtable_quirks)
2725	return -EINVAL;
2726	return domain->ops->set_pgtable_quirks(domain, quirk);
2727	}
2728	EXPORT_SYMBOL_GPL(iommu_set_pgtable_quirks);
2729
2730	/**
2731	* iommu_get_resv_regions - get reserved regions
2732	* @dev: device for which to get reserved regions
2733	* @list: reserved region list for device
2734	*
2735	* This returns a list of reserved IOVA regions specific to this device.
2736	* A domain user should not map IOVA in these ranges.
2737	*/
2738	void iommu_get_resv_regions(struct device *dev, struct list_head *list)
2739	{
2740	const struct iommu_ops *ops = dev_iommu_ops(dev);
2741
2742	if (ops->get_resv_regions)
2743	ops->get_resv_regions(dev, list);
2744	}
2745	EXPORT_SYMBOL_GPL(iommu_get_resv_regions);
2746
2747	/**
2748	* iommu_put_resv_regions - release reserved regions
2749	* @dev: device for which to free reserved regions
2750	* @list: reserved region list for device
2751	*
2752	* This releases a reserved region list acquired by iommu_get_resv_regions().
2753	*/
2754	void iommu_put_resv_regions(struct device *dev, struct list_head *list)
2755	{
2756	struct iommu_resv_region *entry, *next;
2757
2758	list_for_each_entry_safe(entry, next, list, list) {
2759	if (entry->free)
2760	entry->free(dev, entry);
2761	else
2762	kfree(objp: entry);
2763	}
2764	}
2765	EXPORT_SYMBOL(iommu_put_resv_regions);
2766
2767	struct iommu_resv_region *iommu_alloc_resv_region(phys_addr_t start,
2768	size_t length, int prot,
2769	enum iommu_resv_type type,
2770	gfp_t gfp)
2771	{
2772	struct iommu_resv_region *region;
2773
2774	region = kzalloc(size: sizeof(*region), flags: gfp);
2775	if (!region)
2776	return NULL;
2777
2778	INIT_LIST_HEAD(list: &region->list);
2779	region->start = start;
2780	region->length = length;
2781	region->prot = prot;
2782	region->type = type;
2783	return region;
2784	}
2785	EXPORT_SYMBOL_GPL(iommu_alloc_resv_region);
2786
2787	void iommu_set_default_passthrough(bool cmd_line)
2788	{
2789	if (cmd_line)
2790	iommu_cmd_line |= IOMMU_CMD_LINE_DMA_API;
2791	iommu_def_domain_type = IOMMU_DOMAIN_IDENTITY;
2792	}
2793
2794	void iommu_set_default_translated(bool cmd_line)
2795	{
2796	if (cmd_line)
2797	iommu_cmd_line |= IOMMU_CMD_LINE_DMA_API;
2798	iommu_def_domain_type = IOMMU_DOMAIN_DMA;
2799	}
2800
2801	bool iommu_default_passthrough(void)
2802	{
2803	return iommu_def_domain_type == IOMMU_DOMAIN_IDENTITY;
2804	}
2805	EXPORT_SYMBOL_GPL(iommu_default_passthrough);
2806
2807	const struct iommu_ops *iommu_ops_from_fwnode(const struct fwnode_handle *fwnode)
2808	{
2809	const struct iommu_ops *ops = NULL;
2810	struct iommu_device *iommu;
2811
2812	spin_lock(lock: &iommu_device_lock);
2813	list_for_each_entry(iommu, &iommu_device_list, list)
2814	if (iommu->fwnode == fwnode) {
2815	ops = iommu->ops;
2816	break;
2817	}
2818	spin_unlock(lock: &iommu_device_lock);
2819	return ops;
2820	}
2821
2822	int iommu_fwspec_init(struct device *dev, struct fwnode_handle *iommu_fwnode,
2823	const struct iommu_ops *ops)
2824	{
2825	struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
2826
2827	if (fwspec)
2828	return ops == fwspec->ops ? 0 : -EINVAL;
2829
2830	if (!dev_iommu_get(dev))
2831	return -ENOMEM;
2832
2833	/* Preallocate for the overwhelmingly common case of 1 ID */
2834	fwspec = kzalloc(struct_size(fwspec, ids, 1), GFP_KERNEL);
2835	if (!fwspec)
2836	return -ENOMEM;
2837
2838	of_node_get(to_of_node(iommu_fwnode));
2839	fwspec->iommu_fwnode = iommu_fwnode;
2840	fwspec->ops = ops;
2841	dev_iommu_fwspec_set(dev, fwspec);
2842	return 0;
2843	}
2844	EXPORT_SYMBOL_GPL(iommu_fwspec_init);
2845
2846	void iommu_fwspec_free(struct device *dev)
2847	{
2848	struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
2849
2850	if (fwspec) {
2851	fwnode_handle_put(fwnode: fwspec->iommu_fwnode);
2852	kfree(objp: fwspec);
2853	dev_iommu_fwspec_set(dev, NULL);
2854	}
2855	}
2856	EXPORT_SYMBOL_GPL(iommu_fwspec_free);
2857
2858	int iommu_fwspec_add_ids(struct device *dev, const u32 *ids, int num_ids)
2859	{
2860	struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
2861	int i, new_num;
2862
2863	if (!fwspec)
2864	return -EINVAL;
2865
2866	new_num = fwspec->num_ids + num_ids;
2867	if (new_num > 1) {
2868	fwspec = krealloc(objp: fwspec, struct_size(fwspec, ids, new_num),
2869	GFP_KERNEL);
2870	if (!fwspec)
2871	return -ENOMEM;
2872
2873	dev_iommu_fwspec_set(dev, fwspec);
2874	}
2875
2876	for (i = 0; i < num_ids; i++)
2877	fwspec->ids[fwspec->num_ids + i] = ids[i];
2878
2879	fwspec->num_ids = new_num;
2880	return 0;
2881	}
2882	EXPORT_SYMBOL_GPL(iommu_fwspec_add_ids);
2883
2884	/*
2885	* Per device IOMMU features.
2886	*/
2887	int iommu_dev_enable_feature(struct device *dev, enum iommu_dev_features feat)
2888	{
2889	if (dev_has_iommu(dev)) {
2890	const struct iommu_ops *ops = dev_iommu_ops(dev);
2891
2892	if (ops->dev_enable_feat)
2893	return ops->dev_enable_feat(dev, feat);
2894	}
2895
2896	return -ENODEV;
2897	}
2898	EXPORT_SYMBOL_GPL(iommu_dev_enable_feature);
2899
2900	/*
2901	* The device drivers should do the necessary cleanups before calling this.
2902	*/
2903	int iommu_dev_disable_feature(struct device *dev, enum iommu_dev_features feat)
2904	{
2905	if (dev_has_iommu(dev)) {
2906	const struct iommu_ops *ops = dev_iommu_ops(dev);
2907
2908	if (ops->dev_disable_feat)
2909	return ops->dev_disable_feat(dev, feat);
2910	}
2911
2912	return -EBUSY;
2913	}
2914	EXPORT_SYMBOL_GPL(iommu_dev_disable_feature);
2915
2916	/**
2917	* iommu_setup_default_domain - Set the default_domain for the group
2918	* @group: Group to change
2919	* @target_type: Domain type to set as the default_domain
2920	*
2921	* Allocate a default domain and set it as the current domain on the group. If
2922	* the group already has a default domain it will be changed to the target_type.
2923	* When target_type is 0 the default domain is selected based on driver and
2924	* system preferences.
2925	*/
2926	static int iommu_setup_default_domain(struct iommu_group *group,
2927	int target_type)
2928	{
2929	struct iommu_domain *old_dom = group->default_domain;
2930	struct group_device *gdev;
2931	struct iommu_domain *dom;
2932	bool direct_failed;
2933	int req_type;
2934	int ret;
2935
2936	lockdep_assert_held(&group->mutex);
2937
2938	req_type = iommu_get_default_domain_type(group, target_type);
2939	if (req_type < 0)
2940	return -EINVAL;
2941
2942	dom = iommu_group_alloc_default_domain(group, req_type);
2943	if (IS_ERR(ptr: dom))
2944	return PTR_ERR(ptr: dom);
2945
2946	if (group->default_domain == dom)
2947	return 0;
2948
2949	/*
2950	* IOMMU_RESV_DIRECT and IOMMU_RESV_DIRECT_RELAXABLE regions must be
2951	* mapped before their device is attached, in order to guarantee
2952	* continuity with any FW activity
2953	*/
2954	direct_failed = false;
2955	for_each_group_device(group, gdev) {
2956	if (iommu_create_device_direct_mappings(domain: dom, dev: gdev->dev)) {
2957	direct_failed = true;
2958	dev_warn_once(
2959	gdev->dev->iommu->iommu_dev->dev,
2960	"IOMMU driver was not able to establish FW requested direct mapping.");
2961	}
2962	}
2963
2964	/* We must set default_domain early for __iommu_device_set_domain */
2965	group->default_domain = dom;
2966	if (!group->domain) {
2967	/*
2968	* Drivers are not allowed to fail the first domain attach.
2969	* The only way to recover from this is to fail attaching the
2970	* iommu driver and call ops->release_device. Put the domain
2971	* in group->default_domain so it is freed after.
2972	*/
2973	ret = __iommu_group_set_domain_internal(
2974	group, new_domain: dom, flags: IOMMU_SET_DOMAIN_MUST_SUCCEED);
2975	if (WARN_ON(ret))
2976	goto out_free_old;
2977	} else {
2978	ret = __iommu_group_set_domain(group, new_domain: dom);
2979	if (ret)
2980	goto err_restore_def_domain;
2981	}
2982
2983	/*
2984	* Drivers are supposed to allow mappings to be installed in a domain
2985	* before device attachment, but some don't. Hack around this defect by
2986	* trying again after attaching. If this happens it means the device
2987	* will not continuously have the IOMMU_RESV_DIRECT map.
2988	*/
2989	if (direct_failed) {
2990	for_each_group_device(group, gdev) {
2991	ret = iommu_create_device_direct_mappings(domain: dom, dev: gdev->dev);
2992	if (ret)
2993	goto err_restore_domain;
2994	}
2995	}
2996
2997	out_free_old:
2998	if (old_dom)
2999	iommu_domain_free(old_dom);
3000	return ret;
3001
3002	err_restore_domain:
3003	if (old_dom)
3004	__iommu_group_set_domain_internal(
3005	group, new_domain: old_dom, flags: IOMMU_SET_DOMAIN_MUST_SUCCEED);
3006	err_restore_def_domain:
3007	if (old_dom) {
3008	iommu_domain_free(dom);
3009	group->default_domain = old_dom;
3010	}
3011	return ret;
3012	}
3013
3014	/*
3015	* Changing the default domain through sysfs requires the users to unbind the
3016	* drivers from the devices in the iommu group, except for a DMA -> DMA-FQ
3017	* transition. Return failure if this isn't met.
3018	*
3019	* We need to consider the race between this and the device release path.
3020	* group->mutex is used here to guarantee that the device release path
3021	* will not be entered at the same time.
3022	*/
3023	static ssize_t iommu_group_store_type(struct iommu_group *group,
3024	const char *buf, size_t count)
3025	{
3026	struct group_device *gdev;
3027	int ret, req_type;
3028
3029	if (!capable(CAP_SYS_ADMIN) || !capable(CAP_SYS_RAWIO))
3030	return -EACCES;
3031
3032	if (WARN_ON(!group) || !group->default_domain)
3033	return -EINVAL;
3034
3035	if (sysfs_streq(s1: buf, s2: "identity"))
3036	req_type = IOMMU_DOMAIN_IDENTITY;
3037	else if (sysfs_streq(s1: buf, s2: "DMA"))
3038	req_type = IOMMU_DOMAIN_DMA;
3039	else if (sysfs_streq(s1: buf, s2: "DMA-FQ"))
3040	req_type = IOMMU_DOMAIN_DMA_FQ;
3041	else if (sysfs_streq(s1: buf, s2: "auto"))
3042	req_type = 0;
3043	else
3044	return -EINVAL;
3045
3046	mutex_lock(&group->mutex);
3047	/* We can bring up a flush queue without tearing down the domain. */
3048	if (req_type == IOMMU_DOMAIN_DMA_FQ &&
3049	group->default_domain->type == IOMMU_DOMAIN_DMA) {
3050	ret = iommu_dma_init_fq(domain: group->default_domain);
3051	if (ret)
3052	goto out_unlock;
3053
3054	group->default_domain->type = IOMMU_DOMAIN_DMA_FQ;
3055	ret = count;
3056	goto out_unlock;
3057	}
3058
3059	/* Otherwise, ensure that device exists and no driver is bound. */
3060	if (list_empty(head: &group->devices) || group->owner_cnt) {
3061	ret = -EPERM;
3062	goto out_unlock;
3063	}
3064
3065	ret = iommu_setup_default_domain(group, target_type: req_type);
3066	if (ret)
3067	goto out_unlock;
3068
3069	/*
3070	* Release the mutex here because ops->probe_finalize() call-back of
3071	* some vendor IOMMU drivers calls arm_iommu_attach_device() which
3072	* in-turn might call back into IOMMU core code, where it tries to take
3073	* group->mutex, resulting in a deadlock.
3074	*/
3075	mutex_unlock(lock: &group->mutex);
3076
3077	/* Make sure dma_ops is appropriatley set */
3078	for_each_group_device(group, gdev)
3079	iommu_group_do_probe_finalize(dev: gdev->dev);
3080	return count;
3081
3082	out_unlock:
3083	mutex_unlock(lock: &group->mutex);
3084	return ret ?: count;
3085	}
3086
3087	/**
3088	* iommu_device_use_default_domain() - Device driver wants to handle device
3089	* DMA through the kernel DMA API.
3090	* @dev: The device.
3091	*
3092	* The device driver about to bind @dev wants to do DMA through the kernel
3093	* DMA API. Return 0 if it is allowed, otherwise an error.
3094	*/
3095	int iommu_device_use_default_domain(struct device *dev)
3096	{
3097	/* Caller is the driver core during the pre-probe path */
3098	struct iommu_group *group = dev->iommu_group;
3099	int ret = 0;
3100
3101	if (!group)
3102	return 0;
3103
3104	mutex_lock(&group->mutex);
3105	if (group->owner_cnt) {
3106	if (group->domain != group->default_domain || group->owner ||
3107	!xa_empty(xa: &group->pasid_array)) {
3108	ret = -EBUSY;
3109	goto unlock_out;
3110	}
3111	}
3112
3113	group->owner_cnt++;
3114
3115	unlock_out:
3116	mutex_unlock(lock: &group->mutex);
3117	return ret;
3118	}
3119
3120	/**
3121	* iommu_device_unuse_default_domain() - Device driver stops handling device
3122	* DMA through the kernel DMA API.
3123	* @dev: The device.
3124	*
3125	* The device driver doesn't want to do DMA through kernel DMA API anymore.
3126	* It must be called after iommu_device_use_default_domain().
3127	*/
3128	void iommu_device_unuse_default_domain(struct device *dev)
3129	{
3130	/* Caller is the driver core during the post-probe path */
3131	struct iommu_group *group = dev->iommu_group;
3132
3133	if (!group)
3134	return;
3135
3136	mutex_lock(&group->mutex);
3137	if (!WARN_ON(!group->owner_cnt || !xa_empty(&group->pasid_array)))
3138	group->owner_cnt--;
3139
3140	mutex_unlock(lock: &group->mutex);
3141	}
3142
3143	static int __iommu_group_alloc_blocking_domain(struct iommu_group *group)
3144	{
3145	struct iommu_domain *domain;
3146
3147	if (group->blocking_domain)
3148	return 0;
3149
3150	domain = __iommu_group_domain_alloc(group, IOMMU_DOMAIN_BLOCKED);
3151	if (IS_ERR(ptr: domain)) {
3152	/*
3153	* For drivers that do not yet understand IOMMU_DOMAIN_BLOCKED
3154	* create an empty domain instead.
3155	*/
3156	domain = __iommu_group_domain_alloc(group,
3157	IOMMU_DOMAIN_UNMANAGED);
3158	if (IS_ERR(ptr: domain))
3159	return PTR_ERR(ptr: domain);
3160	}
3161	group->blocking_domain = domain;
3162	return 0;
3163	}
3164
3165	static int __iommu_take_dma_ownership(struct iommu_group *group, void *owner)
3166	{
3167	int ret;
3168
3169	if ((group->domain && group->domain != group->default_domain) ||
3170	!xa_empty(xa: &group->pasid_array))
3171	return -EBUSY;
3172
3173	ret = __iommu_group_alloc_blocking_domain(group);
3174	if (ret)
3175	return ret;
3176	ret = __iommu_group_set_domain(group, new_domain: group->blocking_domain);
3177	if (ret)
3178	return ret;
3179
3180	group->owner = owner;
3181	group->owner_cnt++;
3182	return 0;
3183	}
3184
3185	/**
3186	* iommu_group_claim_dma_owner() - Set DMA ownership of a group
3187	* @group: The group.
3188	* @owner: Caller specified pointer. Used for exclusive ownership.
3189	*
3190	* This is to support backward compatibility for vfio which manages the dma
3191	* ownership in iommu_group level. New invocations on this interface should be
3192	* prohibited. Only a single owner may exist for a group.
3193	*/
3194	int iommu_group_claim_dma_owner(struct iommu_group *group, void *owner)
3195	{
3196	int ret = 0;
3197
3198	if (WARN_ON(!owner))
3199	return -EINVAL;
3200
3201	mutex_lock(&group->mutex);
3202	if (group->owner_cnt) {
3203	ret = -EPERM;
3204	goto unlock_out;
3205	}
3206
3207	ret = __iommu_take_dma_ownership(group, owner);
3208	unlock_out:
3209	mutex_unlock(lock: &group->mutex);
3210
3211	return ret;
3212	}
3213	EXPORT_SYMBOL_GPL(iommu_group_claim_dma_owner);
3214
3215	/**
3216	* iommu_device_claim_dma_owner() - Set DMA ownership of a device
3217	* @dev: The device.
3218	* @owner: Caller specified pointer. Used for exclusive ownership.
3219	*
3220	* Claim the DMA ownership of a device. Multiple devices in the same group may
3221	* concurrently claim ownership if they present the same owner value. Returns 0
3222	* on success and error code on failure
3223	*/
3224	int iommu_device_claim_dma_owner(struct device *dev, void *owner)
3225	{
3226	/* Caller must be a probed driver on dev */
3227	struct iommu_group *group = dev->iommu_group;
3228	int ret = 0;
3229
3230	if (WARN_ON(!owner))
3231	return -EINVAL;
3232
3233	if (!group)
3234	return -ENODEV;
3235
3236	mutex_lock(&group->mutex);
3237	if (group->owner_cnt) {
3238	if (group->owner != owner) {
3239	ret = -EPERM;
3240	goto unlock_out;
3241	}
3242	group->owner_cnt++;
3243	goto unlock_out;
3244	}
3245
3246	ret = __iommu_take_dma_ownership(group, owner);
3247	unlock_out:
3248	mutex_unlock(lock: &group->mutex);
3249	return ret;
3250	}
3251	EXPORT_SYMBOL_GPL(iommu_device_claim_dma_owner);
3252
3253	static void __iommu_release_dma_ownership(struct iommu_group *group)
3254	{
3255	if (WARN_ON(!group->owner_cnt || !group->owner ||
3256	!xa_empty(&group->pasid_array)))
3257	return;
3258
3259	group->owner_cnt = 0;
3260	group->owner = NULL;
3261	__iommu_group_set_domain_nofail(group, new_domain: group->default_domain);
3262	}
3263
3264	/**
3265	* iommu_group_release_dma_owner() - Release DMA ownership of a group
3266	* @group: The group
3267	*
3268	* Release the DMA ownership claimed by iommu_group_claim_dma_owner().
3269	*/
3270	void iommu_group_release_dma_owner(struct iommu_group *group)
3271	{
3272	mutex_lock(&group->mutex);
3273	__iommu_release_dma_ownership(group);
3274	mutex_unlock(lock: &group->mutex);
3275	}
3276	EXPORT_SYMBOL_GPL(iommu_group_release_dma_owner);
3277
3278	/**
3279	* iommu_device_release_dma_owner() - Release DMA ownership of a device
3280	* @dev: The device.
3281	*
3282	* Release the DMA ownership claimed by iommu_device_claim_dma_owner().
3283	*/
3284	void iommu_device_release_dma_owner(struct device *dev)
3285	{
3286	/* Caller must be a probed driver on dev */
3287	struct iommu_group *group = dev->iommu_group;
3288
3289	mutex_lock(&group->mutex);
3290	if (group->owner_cnt > 1)
3291	group->owner_cnt--;
3292	else
3293	__iommu_release_dma_ownership(group);
3294	mutex_unlock(lock: &group->mutex);
3295	}
3296	EXPORT_SYMBOL_GPL(iommu_device_release_dma_owner);
3297
3298	/**
3299	* iommu_group_dma_owner_claimed() - Query group dma ownership status
3300	* @group: The group.
3301	*
3302	* This provides status query on a given group. It is racy and only for
3303	* non-binding status reporting.
3304	*/
3305	bool iommu_group_dma_owner_claimed(struct iommu_group *group)
3306	{
3307	unsigned int user;
3308
3309	mutex_lock(&group->mutex);
3310	user = group->owner_cnt;
3311	mutex_unlock(lock: &group->mutex);
3312
3313	return user;
3314	}
3315	EXPORT_SYMBOL_GPL(iommu_group_dma_owner_claimed);
3316
3317	static int __iommu_set_group_pasid(struct iommu_domain *domain,
3318	struct iommu_group *group, ioasid_t pasid)
3319	{
3320	struct group_device *device;
3321	int ret = 0;
3322
3323	for_each_group_device(group, device) {
3324	ret = domain->ops->set_dev_pasid(domain, device->dev, pasid);
3325	if (ret)
3326	break;
3327	}
3328
3329	return ret;
3330	}
3331
3332	static void __iommu_remove_group_pasid(struct iommu_group *group,
3333	ioasid_t pasid)
3334	{
3335	struct group_device *device;
3336	const struct iommu_ops *ops;
3337
3338	for_each_group_device(group, device) {
3339	ops = dev_iommu_ops(dev: device->dev);
3340	ops->remove_dev_pasid(device->dev, pasid);
3341	}
3342	}
3343
3344	/*
3345	* iommu_attach_device_pasid() - Attach a domain to pasid of device
3346	* @domain: the iommu domain.
3347	* @dev: the attached device.
3348	* @pasid: the pasid of the device.
3349	*
3350	* Return: 0 on success, or an error.
3351	*/
3352	int iommu_attach_device_pasid(struct iommu_domain *domain,
3353	struct device *dev, ioasid_t pasid)
3354	{
3355	/* Caller must be a probed driver on dev */
3356	struct iommu_group *group = dev->iommu_group;
3357	struct group_device *device;
3358	void *curr;
3359	int ret;
3360
3361	if (!domain->ops->set_dev_pasid)
3362	return -EOPNOTSUPP;
3363
3364	if (!group)
3365	return -ENODEV;
3366
3367	if (!dev_has_iommu(dev) || dev_iommu_ops(dev) != domain->owner ||
3368	pasid == IOMMU_NO_PASID)
3369	return -EINVAL;
3370
3371	mutex_lock(&group->mutex);
3372	for_each_group_device(group, device) {
3373	if (pasid >= device->dev->iommu->max_pasids) {
3374	ret = -EINVAL;
3375	goto out_unlock;
3376	}
3377	}
3378
3379	curr = xa_cmpxchg(xa: &group->pasid_array, index: pasid, NULL, entry: domain, GFP_KERNEL);
3380	if (curr) {
3381	ret = xa_err(entry: curr) ? : -EBUSY;
3382	goto out_unlock;
3383	}
3384
3385	ret = __iommu_set_group_pasid(domain, group, pasid);
3386	if (ret) {
3387	__iommu_remove_group_pasid(group, pasid);
3388	xa_erase(&group->pasid_array, index: pasid);
3389	}
3390	out_unlock:
3391	mutex_unlock(lock: &group->mutex);
3392	return ret;
3393	}
3394	EXPORT_SYMBOL_GPL(iommu_attach_device_pasid);
3395
3396	/*
3397	* iommu_detach_device_pasid() - Detach the domain from pasid of device
3398	* @domain: the iommu domain.
3399	* @dev: the attached device.
3400	* @pasid: the pasid of the device.
3401	*
3402	* The @domain must have been attached to @pasid of the @dev with
3403	* iommu_attach_device_pasid().
3404	*/
3405	void iommu_detach_device_pasid(struct iommu_domain *domain, struct device *dev,
3406	ioasid_t pasid)
3407	{
3408	/* Caller must be a probed driver on dev */
3409	struct iommu_group *group = dev->iommu_group;
3410
3411	mutex_lock(&group->mutex);
3412	__iommu_remove_group_pasid(group, pasid);
3413	WARN_ON(xa_erase(&group->pasid_array, pasid) != domain);
3414	mutex_unlock(lock: &group->mutex);
3415	}
3416	EXPORT_SYMBOL_GPL(iommu_detach_device_pasid);
3417
3418	/*
3419	* iommu_get_domain_for_dev_pasid() - Retrieve domain for @pasid of @dev
3420	* @dev: the queried device
3421	* @pasid: the pasid of the device
3422	* @type: matched domain type, 0 for any match
3423	*
3424	* This is a variant of iommu_get_domain_for_dev(). It returns the existing
3425	* domain attached to pasid of a device. Callers must hold a lock around this
3426	* function, and both iommu_attach/detach_dev_pasid() whenever a domain of
3427	* type is being manipulated. This API does not internally resolve races with
3428	* attach/detach.
3429	*
3430	* Return: attached domain on success, NULL otherwise.
3431	*/
3432	struct iommu_domain *iommu_get_domain_for_dev_pasid(struct device *dev,
3433	ioasid_t pasid,
3434	unsigned int type)
3435	{
3436	/* Caller must be a probed driver on dev */
3437	struct iommu_group *group = dev->iommu_group;
3438	struct iommu_domain *domain;
3439
3440	if (!group)
3441	return NULL;
3442
3443	xa_lock(&group->pasid_array);
3444	domain = xa_load(&group->pasid_array, index: pasid);
3445	if (type && domain && domain->type != type)
3446	domain = ERR_PTR(error: -EBUSY);
3447	xa_unlock(&group->pasid_array);
3448
3449	return domain;
3450	}
3451	EXPORT_SYMBOL_GPL(iommu_get_domain_for_dev_pasid);
3452
3453	ioasid_t iommu_alloc_global_pasid(struct device *dev)
3454	{
3455	int ret;
3456
3457	/* max_pasids == 0 means that the device does not support PASID */
3458	if (!dev->iommu->max_pasids)
3459	return IOMMU_PASID_INVALID;
3460
3461	/*
3462	* max_pasids is set up by vendor driver based on number of PASID bits
3463	* supported but the IDA allocation is inclusive.
3464	*/
3465	ret = ida_alloc_range(&iommu_global_pasid_ida, IOMMU_FIRST_GLOBAL_PASID,
3466	max: dev->iommu->max_pasids - 1, GFP_KERNEL);
3467	return ret < 0 ? IOMMU_PASID_INVALID : ret;
3468	}
3469	EXPORT_SYMBOL_GPL(iommu_alloc_global_pasid);
3470
3471	void iommu_free_global_pasid(ioasid_t pasid)
3472	{
3473	if (WARN_ON(pasid == IOMMU_PASID_INVALID))
3474	return;
3475
3476	ida_free(&iommu_global_pasid_ida, id: pasid);
3477	}
3478	EXPORT_SYMBOL_GPL(iommu_free_global_pasid);
3479