1/*
2 * Copyright (C) 2010-2012 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <jroedel@suse.de>
4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 as published
7 * by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
17 */
18
19#include <linux/mmu_notifier.h>
20#include <linux/amd-iommu.h>
21#include <linux/mm_types.h>
22#include <linux/profile.h>
23#include <linux/module.h>
24#include <linux/sched.h>
25#include <linux/iommu.h>
26#include <linux/wait.h>
27#include <linux/pci.h>
28#include <linux/gfp.h>
29
30#include "amd_iommu_types.h"
31#include "amd_iommu_proto.h"
32
33MODULE_LICENSE("GPL v2");
34MODULE_AUTHOR("Joerg Roedel <jroedel@suse.de>");
35
36#define MAX_DEVICES		0x10000
37#define PRI_QUEUE_SIZE		512
38
39struct pri_queue {
40	atomic_t inflight;
41	bool finish;
42	int status;
43};
44
45struct pasid_state {
46	struct list_head list;			/* For global state-list */
47	atomic_t count;				/* Reference count */
48	unsigned mmu_notifier_count;		/* Counting nested mmu_notifier
49						   calls */
50	struct mm_struct *mm;			/* mm_struct for the faults */
51	struct mmu_notifier mn;                 /* mmu_notifier handle */
52	struct pri_queue pri[PRI_QUEUE_SIZE];	/* PRI tag states */
53	struct device_state *device_state;	/* Link to our device_state */
54	int pasid;				/* PASID index */
55	bool invalid;				/* Used during setup and
56						   teardown of the pasid */
57	spinlock_t lock;			/* Protect pri_queues and
58						   mmu_notifer_count */
59	wait_queue_head_t wq;			/* To wait for count == 0 */
60};
61
62struct device_state {
63	struct list_head list;
64	u16 devid;
65	atomic_t count;
66	struct pci_dev *pdev;
67	struct pasid_state **states;
68	struct iommu_domain *domain;
69	int pasid_levels;
70	int max_pasids;
71	amd_iommu_invalid_ppr_cb inv_ppr_cb;
72	amd_iommu_invalidate_ctx inv_ctx_cb;
73	spinlock_t lock;
74	wait_queue_head_t wq;
75};
76
77struct fault {
78	struct work_struct work;
79	struct device_state *dev_state;
80	struct pasid_state *state;
81	struct mm_struct *mm;
82	u64 address;
83	u16 devid;
84	u16 pasid;
85	u16 tag;
86	u16 finish;
87	u16 flags;
88};
89
90static LIST_HEAD(state_list);
91static spinlock_t state_lock;
92
93static struct workqueue_struct *iommu_wq;
94
95static void free_pasid_states(struct device_state *dev_state);
96
97static u16 device_id(struct pci_dev *pdev)
98{
99	u16 devid;
100
101	devid = pdev->bus->number;
102	devid = (devid << 8) | pdev->devfn;
103
104	return devid;
105}
106
107static struct device_state *__get_device_state(u16 devid)
108{
109	struct device_state *dev_state;
110
111	list_for_each_entry(dev_state, &state_list, list) {
112		if (dev_state->devid == devid)
113			return dev_state;
114	}
115
116	return NULL;
117}
118
119static struct device_state *get_device_state(u16 devid)
120{
121	struct device_state *dev_state;
122	unsigned long flags;
123
124	spin_lock_irqsave(&state_lock, flags);
125	dev_state = __get_device_state(devid);
126	if (dev_state != NULL)
127		atomic_inc(&dev_state->count);
128	spin_unlock_irqrestore(&state_lock, flags);
129
130	return dev_state;
131}
132
133static void free_device_state(struct device_state *dev_state)
134{
135	/*
136	 * First detach device from domain - No more PRI requests will arrive
137	 * from that device after it is unbound from the IOMMUv2 domain.
138	 */
139	iommu_detach_device(dev_state->domain, &dev_state->pdev->dev);
140
141	/* Everything is down now, free the IOMMUv2 domain */
142	iommu_domain_free(dev_state->domain);
143
144	/* Finally get rid of the device-state */
145	kfree(dev_state);
146}
147
148static void put_device_state(struct device_state *dev_state)
149{
150	if (atomic_dec_and_test(&dev_state->count))
151		wake_up(&dev_state->wq);
152}
153
154/* Must be called under dev_state->lock */
155static struct pasid_state **__get_pasid_state_ptr(struct device_state *dev_state,
156						  int pasid, bool alloc)
157{
158	struct pasid_state **root, **ptr;
159	int level, index;
160
161	level = dev_state->pasid_levels;
162	root  = dev_state->states;
163
164	while (true) {
165
166		index = (pasid >> (9 * level)) & 0x1ff;
167		ptr   = &root[index];
168
169		if (level == 0)
170			break;
171
172		if (*ptr == NULL) {
173			if (!alloc)
174				return NULL;
175
176			*ptr = (void *)get_zeroed_page(GFP_ATOMIC);
177			if (*ptr == NULL)
178				return NULL;
179		}
180
181		root   = (struct pasid_state **)*ptr;
182		level -= 1;
183	}
184
185	return ptr;
186}
187
188static int set_pasid_state(struct device_state *dev_state,
189			   struct pasid_state *pasid_state,
190			   int pasid)
191{
192	struct pasid_state **ptr;
193	unsigned long flags;
194	int ret;
195
196	spin_lock_irqsave(&dev_state->lock, flags);
197	ptr = __get_pasid_state_ptr(dev_state, pasid, true);
198
199	ret = -ENOMEM;
200	if (ptr == NULL)
201		goto out_unlock;
202
203	ret = -ENOMEM;
204	if (*ptr != NULL)
205		goto out_unlock;
206
207	*ptr = pasid_state;
208
209	ret = 0;
210
211out_unlock:
212	spin_unlock_irqrestore(&dev_state->lock, flags);
213
214	return ret;
215}
216
217static void clear_pasid_state(struct device_state *dev_state, int pasid)
218{
219	struct pasid_state **ptr;
220	unsigned long flags;
221
222	spin_lock_irqsave(&dev_state->lock, flags);
223	ptr = __get_pasid_state_ptr(dev_state, pasid, true);
224
225	if (ptr == NULL)
226		goto out_unlock;
227
228	*ptr = NULL;
229
230out_unlock:
231	spin_unlock_irqrestore(&dev_state->lock, flags);
232}
233
234static struct pasid_state *get_pasid_state(struct device_state *dev_state,
235					   int pasid)
236{
237	struct pasid_state **ptr, *ret = NULL;
238	unsigned long flags;
239
240	spin_lock_irqsave(&dev_state->lock, flags);
241	ptr = __get_pasid_state_ptr(dev_state, pasid, false);
242
243	if (ptr == NULL)
244		goto out_unlock;
245
246	ret = *ptr;
247	if (ret)
248		atomic_inc(&ret->count);
249
250out_unlock:
251	spin_unlock_irqrestore(&dev_state->lock, flags);
252
253	return ret;
254}
255
256static void free_pasid_state(struct pasid_state *pasid_state)
257{
258	kfree(pasid_state);
259}
260
261static void put_pasid_state(struct pasid_state *pasid_state)
262{
263	if (atomic_dec_and_test(&pasid_state->count))
264		wake_up(&pasid_state->wq);
265}
266
267static void put_pasid_state_wait(struct pasid_state *pasid_state)
268{
269	atomic_dec(&pasid_state->count);
270	wait_event(pasid_state->wq, !atomic_read(&pasid_state->count));
271	free_pasid_state(pasid_state);
272}
273
274static void unbind_pasid(struct pasid_state *pasid_state)
275{
276	struct iommu_domain *domain;
277
278	domain = pasid_state->device_state->domain;
279
280	/*
281	 * Mark pasid_state as invalid, no more faults will we added to the
282	 * work queue after this is visible everywhere.
283	 */
284	pasid_state->invalid = true;
285
286	/* Make sure this is visible */
287	smp_wmb();
288
289	/* After this the device/pasid can't access the mm anymore */
290	amd_iommu_domain_clear_gcr3(domain, pasid_state->pasid);
291
292	/* Make sure no more pending faults are in the queue */
293	flush_workqueue(iommu_wq);
294}
295
296static void free_pasid_states_level1(struct pasid_state **tbl)
297{
298	int i;
299
300	for (i = 0; i < 512; ++i) {
301		if (tbl[i] == NULL)
302			continue;
303
304		free_page((unsigned long)tbl[i]);
305	}
306}
307
308static void free_pasid_states_level2(struct pasid_state **tbl)
309{
310	struct pasid_state **ptr;
311	int i;
312
313	for (i = 0; i < 512; ++i) {
314		if (tbl[i] == NULL)
315			continue;
316
317		ptr = (struct pasid_state **)tbl[i];
318		free_pasid_states_level1(ptr);
319	}
320}
321
322static void free_pasid_states(struct device_state *dev_state)
323{
324	struct pasid_state *pasid_state;
325	int i;
326
327	for (i = 0; i < dev_state->max_pasids; ++i) {
328		pasid_state = get_pasid_state(dev_state, i);
329		if (pasid_state == NULL)
330			continue;
331
332		put_pasid_state(pasid_state);
333
334		/*
335		 * This will call the mn_release function and
336		 * unbind the PASID
337		 */
338		mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
339
340		put_pasid_state_wait(pasid_state); /* Reference taken in
341						      amd_iommu_bind_pasid */
342
343		/* Drop reference taken in amd_iommu_bind_pasid */
344		put_device_state(dev_state);
345	}
346
347	if (dev_state->pasid_levels == 2)
348		free_pasid_states_level2(dev_state->states);
349	else if (dev_state->pasid_levels == 1)
350		free_pasid_states_level1(dev_state->states);
351	else if (dev_state->pasid_levels != 0)
352		BUG();
353
354	free_page((unsigned long)dev_state->states);
355}
356
357static struct pasid_state *mn_to_state(struct mmu_notifier *mn)
358{
359	return container_of(mn, struct pasid_state, mn);
360}
361
362static void __mn_flush_page(struct mmu_notifier *mn,
363			    unsigned long address)
364{
365	struct pasid_state *pasid_state;
366	struct device_state *dev_state;
367
368	pasid_state = mn_to_state(mn);
369	dev_state   = pasid_state->device_state;
370
371	amd_iommu_flush_page(dev_state->domain, pasid_state->pasid, address);
372}
373
374static int mn_clear_flush_young(struct mmu_notifier *mn,
375				struct mm_struct *mm,
376				unsigned long start,
377				unsigned long end)
378{
379	for (; start < end; start += PAGE_SIZE)
380		__mn_flush_page(mn, start);
381
382	return 0;
383}
384
385static void mn_invalidate_page(struct mmu_notifier *mn,
386			       struct mm_struct *mm,
387			       unsigned long address)
388{
389	__mn_flush_page(mn, address);
390}
391
392static void mn_invalidate_range(struct mmu_notifier *mn,
393				struct mm_struct *mm,
394				unsigned long start, unsigned long end)
395{
396	struct pasid_state *pasid_state;
397	struct device_state *dev_state;
398
399	pasid_state = mn_to_state(mn);
400	dev_state   = pasid_state->device_state;
401
402	if ((start ^ (end - 1)) < PAGE_SIZE)
403		amd_iommu_flush_page(dev_state->domain, pasid_state->pasid,
404				     start);
405	else
406		amd_iommu_flush_tlb(dev_state->domain, pasid_state->pasid);
407}
408
409static void mn_release(struct mmu_notifier *mn, struct mm_struct *mm)
410{
411	struct pasid_state *pasid_state;
412	struct device_state *dev_state;
413	bool run_inv_ctx_cb;
414
415	might_sleep();
416
417	pasid_state    = mn_to_state(mn);
418	dev_state      = pasid_state->device_state;
419	run_inv_ctx_cb = !pasid_state->invalid;
420
421	if (run_inv_ctx_cb && dev_state->inv_ctx_cb)
422		dev_state->inv_ctx_cb(dev_state->pdev, pasid_state->pasid);
423
424	unbind_pasid(pasid_state);
425}
426
427static struct mmu_notifier_ops iommu_mn = {
428	.release		= mn_release,
429	.clear_flush_young      = mn_clear_flush_young,
430	.invalidate_page        = mn_invalidate_page,
431	.invalidate_range       = mn_invalidate_range,
432};
433
434static void set_pri_tag_status(struct pasid_state *pasid_state,
435			       u16 tag, int status)
436{
437	unsigned long flags;
438
439	spin_lock_irqsave(&pasid_state->lock, flags);
440	pasid_state->pri[tag].status = status;
441	spin_unlock_irqrestore(&pasid_state->lock, flags);
442}
443
444static void finish_pri_tag(struct device_state *dev_state,
445			   struct pasid_state *pasid_state,
446			   u16 tag)
447{
448	unsigned long flags;
449
450	spin_lock_irqsave(&pasid_state->lock, flags);
451	if (atomic_dec_and_test(&pasid_state->pri[tag].inflight) &&
452	    pasid_state->pri[tag].finish) {
453		amd_iommu_complete_ppr(dev_state->pdev, pasid_state->pasid,
454				       pasid_state->pri[tag].status, tag);
455		pasid_state->pri[tag].finish = false;
456		pasid_state->pri[tag].status = PPR_SUCCESS;
457	}
458	spin_unlock_irqrestore(&pasid_state->lock, flags);
459}
460
461static void handle_fault_error(struct fault *fault)
462{
463	int status;
464
465	if (!fault->dev_state->inv_ppr_cb) {
466		set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
467		return;
468	}
469
470	status = fault->dev_state->inv_ppr_cb(fault->dev_state->pdev,
471					      fault->pasid,
472					      fault->address,
473					      fault->flags);
474	switch (status) {
475	case AMD_IOMMU_INV_PRI_RSP_SUCCESS:
476		set_pri_tag_status(fault->state, fault->tag, PPR_SUCCESS);
477		break;
478	case AMD_IOMMU_INV_PRI_RSP_INVALID:
479		set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
480		break;
481	case AMD_IOMMU_INV_PRI_RSP_FAIL:
482		set_pri_tag_status(fault->state, fault->tag, PPR_FAILURE);
483		break;
484	default:
485		BUG();
486	}
487}
488
489static void do_fault(struct work_struct *work)
490{
491	struct fault *fault = container_of(work, struct fault, work);
492	struct mm_struct *mm;
493	struct vm_area_struct *vma;
494	u64 address;
495	int ret, write;
496
497	write = !!(fault->flags & PPR_FAULT_WRITE);
498
499	mm = fault->state->mm;
500	address = fault->address;
501
502	down_read(&mm->mmap_sem);
503	vma = find_extend_vma(mm, address);
504	if (!vma || address < vma->vm_start) {
505		/* failed to get a vma in the right range */
506		up_read(&mm->mmap_sem);
507		handle_fault_error(fault);
508		goto out;
509	}
510
511	if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) {
512		/* handle_mm_fault would BUG_ON() */
513		up_read(&mm->mmap_sem);
514		handle_fault_error(fault);
515		goto out;
516	}
517
518	ret = handle_mm_fault(mm, vma, address, write);
519	if (ret & VM_FAULT_ERROR) {
520		/* failed to service fault */
521		up_read(&mm->mmap_sem);
522		handle_fault_error(fault);
523		goto out;
524	}
525
526	up_read(&mm->mmap_sem);
527
528out:
529	finish_pri_tag(fault->dev_state, fault->state, fault->tag);
530
531	put_pasid_state(fault->state);
532
533	kfree(fault);
534}
535
536static int ppr_notifier(struct notifier_block *nb, unsigned long e, void *data)
537{
538	struct amd_iommu_fault *iommu_fault;
539	struct pasid_state *pasid_state;
540	struct device_state *dev_state;
541	unsigned long flags;
542	struct fault *fault;
543	bool finish;
544	u16 tag;
545	int ret;
546
547	iommu_fault = data;
548	tag         = iommu_fault->tag & 0x1ff;
549	finish      = (iommu_fault->tag >> 9) & 1;
550
551	ret = NOTIFY_DONE;
552	dev_state = get_device_state(iommu_fault->device_id);
553	if (dev_state == NULL)
554		goto out;
555
556	pasid_state = get_pasid_state(dev_state, iommu_fault->pasid);
557	if (pasid_state == NULL || pasid_state->invalid) {
558		/* We know the device but not the PASID -> send INVALID */
559		amd_iommu_complete_ppr(dev_state->pdev, iommu_fault->pasid,
560				       PPR_INVALID, tag);
561		goto out_drop_state;
562	}
563
564	spin_lock_irqsave(&pasid_state->lock, flags);
565	atomic_inc(&pasid_state->pri[tag].inflight);
566	if (finish)
567		pasid_state->pri[tag].finish = true;
568	spin_unlock_irqrestore(&pasid_state->lock, flags);
569
570	fault = kzalloc(sizeof(*fault), GFP_ATOMIC);
571	if (fault == NULL) {
572		/* We are OOM - send success and let the device re-fault */
573		finish_pri_tag(dev_state, pasid_state, tag);
574		goto out_drop_state;
575	}
576
577	fault->dev_state = dev_state;
578	fault->address   = iommu_fault->address;
579	fault->state     = pasid_state;
580	fault->tag       = tag;
581	fault->finish    = finish;
582	fault->pasid     = iommu_fault->pasid;
583	fault->flags     = iommu_fault->flags;
584	INIT_WORK(&fault->work, do_fault);
585
586	queue_work(iommu_wq, &fault->work);
587
588	ret = NOTIFY_OK;
589
590out_drop_state:
591
592	if (ret != NOTIFY_OK && pasid_state)
593		put_pasid_state(pasid_state);
594
595	put_device_state(dev_state);
596
597out:
598	return ret;
599}
600
601static struct notifier_block ppr_nb = {
602	.notifier_call = ppr_notifier,
603};
604
605int amd_iommu_bind_pasid(struct pci_dev *pdev, int pasid,
606			 struct task_struct *task)
607{
608	struct pasid_state *pasid_state;
609	struct device_state *dev_state;
610	struct mm_struct *mm;
611	u16 devid;
612	int ret;
613
614	might_sleep();
615
616	if (!amd_iommu_v2_supported())
617		return -ENODEV;
618
619	devid     = device_id(pdev);
620	dev_state = get_device_state(devid);
621
622	if (dev_state == NULL)
623		return -EINVAL;
624
625	ret = -EINVAL;
626	if (pasid < 0 || pasid >= dev_state->max_pasids)
627		goto out;
628
629	ret = -ENOMEM;
630	pasid_state = kzalloc(sizeof(*pasid_state), GFP_KERNEL);
631	if (pasid_state == NULL)
632		goto out;
633
634
635	atomic_set(&pasid_state->count, 1);
636	init_waitqueue_head(&pasid_state->wq);
637	spin_lock_init(&pasid_state->lock);
638
639	mm                        = get_task_mm(task);
640	pasid_state->mm           = mm;
641	pasid_state->device_state = dev_state;
642	pasid_state->pasid        = pasid;
643	pasid_state->invalid      = true; /* Mark as valid only if we are
644					     done with setting up the pasid */
645	pasid_state->mn.ops       = &iommu_mn;
646
647	if (pasid_state->mm == NULL)
648		goto out_free;
649
650	mmu_notifier_register(&pasid_state->mn, mm);
651
652	ret = set_pasid_state(dev_state, pasid_state, pasid);
653	if (ret)
654		goto out_unregister;
655
656	ret = amd_iommu_domain_set_gcr3(dev_state->domain, pasid,
657					__pa(pasid_state->mm->pgd));
658	if (ret)
659		goto out_clear_state;
660
661	/* Now we are ready to handle faults */
662	pasid_state->invalid = false;
663
664	/*
665	 * Drop the reference to the mm_struct here. We rely on the
666	 * mmu_notifier release call-back to inform us when the mm
667	 * is going away.
668	 */
669	mmput(mm);
670
671	return 0;
672
673out_clear_state:
674	clear_pasid_state(dev_state, pasid);
675
676out_unregister:
677	mmu_notifier_unregister(&pasid_state->mn, mm);
678
679out_free:
680	mmput(mm);
681	free_pasid_state(pasid_state);
682
683out:
684	put_device_state(dev_state);
685
686	return ret;
687}
688EXPORT_SYMBOL(amd_iommu_bind_pasid);
689
690void amd_iommu_unbind_pasid(struct pci_dev *pdev, int pasid)
691{
692	struct pasid_state *pasid_state;
693	struct device_state *dev_state;
694	u16 devid;
695
696	might_sleep();
697
698	if (!amd_iommu_v2_supported())
699		return;
700
701	devid = device_id(pdev);
702	dev_state = get_device_state(devid);
703	if (dev_state == NULL)
704		return;
705
706	if (pasid < 0 || pasid >= dev_state->max_pasids)
707		goto out;
708
709	pasid_state = get_pasid_state(dev_state, pasid);
710	if (pasid_state == NULL)
711		goto out;
712	/*
713	 * Drop reference taken here. We are safe because we still hold
714	 * the reference taken in the amd_iommu_bind_pasid function.
715	 */
716	put_pasid_state(pasid_state);
717
718	/* Clear the pasid state so that the pasid can be re-used */
719	clear_pasid_state(dev_state, pasid_state->pasid);
720
721	/*
722	 * Call mmu_notifier_unregister to drop our reference
723	 * to pasid_state->mm
724	 */
725	mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
726
727	put_pasid_state_wait(pasid_state); /* Reference taken in
728					      amd_iommu_bind_pasid */
729out:
730	/* Drop reference taken in this function */
731	put_device_state(dev_state);
732
733	/* Drop reference taken in amd_iommu_bind_pasid */
734	put_device_state(dev_state);
735}
736EXPORT_SYMBOL(amd_iommu_unbind_pasid);
737
738int amd_iommu_init_device(struct pci_dev *pdev, int pasids)
739{
740	struct device_state *dev_state;
741	unsigned long flags;
742	int ret, tmp;
743	u16 devid;
744
745	might_sleep();
746
747	if (!amd_iommu_v2_supported())
748		return -ENODEV;
749
750	if (pasids <= 0 || pasids > (PASID_MASK + 1))
751		return -EINVAL;
752
753	devid = device_id(pdev);
754
755	dev_state = kzalloc(sizeof(*dev_state), GFP_KERNEL);
756	if (dev_state == NULL)
757		return -ENOMEM;
758
759	spin_lock_init(&dev_state->lock);
760	init_waitqueue_head(&dev_state->wq);
761	dev_state->pdev  = pdev;
762	dev_state->devid = devid;
763
764	tmp = pasids;
765	for (dev_state->pasid_levels = 0; (tmp - 1) & ~0x1ff; tmp >>= 9)
766		dev_state->pasid_levels += 1;
767
768	atomic_set(&dev_state->count, 1);
769	dev_state->max_pasids = pasids;
770
771	ret = -ENOMEM;
772	dev_state->states = (void *)get_zeroed_page(GFP_KERNEL);
773	if (dev_state->states == NULL)
774		goto out_free_dev_state;
775
776	dev_state->domain = iommu_domain_alloc(&pci_bus_type);
777	if (dev_state->domain == NULL)
778		goto out_free_states;
779
780	amd_iommu_domain_direct_map(dev_state->domain);
781
782	ret = amd_iommu_domain_enable_v2(dev_state->domain, pasids);
783	if (ret)
784		goto out_free_domain;
785
786	ret = iommu_attach_device(dev_state->domain, &pdev->dev);
787	if (ret != 0)
788		goto out_free_domain;
789
790	spin_lock_irqsave(&state_lock, flags);
791
792	if (__get_device_state(devid) != NULL) {
793		spin_unlock_irqrestore(&state_lock, flags);
794		ret = -EBUSY;
795		goto out_free_domain;
796	}
797
798	list_add_tail(&dev_state->list, &state_list);
799
800	spin_unlock_irqrestore(&state_lock, flags);
801
802	return 0;
803
804out_free_domain:
805	iommu_domain_free(dev_state->domain);
806
807out_free_states:
808	free_page((unsigned long)dev_state->states);
809
810out_free_dev_state:
811	kfree(dev_state);
812
813	return ret;
814}
815EXPORT_SYMBOL(amd_iommu_init_device);
816
817void amd_iommu_free_device(struct pci_dev *pdev)
818{
819	struct device_state *dev_state;
820	unsigned long flags;
821	u16 devid;
822
823	if (!amd_iommu_v2_supported())
824		return;
825
826	devid = device_id(pdev);
827
828	spin_lock_irqsave(&state_lock, flags);
829
830	dev_state = __get_device_state(devid);
831	if (dev_state == NULL) {
832		spin_unlock_irqrestore(&state_lock, flags);
833		return;
834	}
835
836	list_del(&dev_state->list);
837
838	spin_unlock_irqrestore(&state_lock, flags);
839
840	/* Get rid of any remaining pasid states */
841	free_pasid_states(dev_state);
842
843	put_device_state(dev_state);
844	/*
845	 * Wait until the last reference is dropped before freeing
846	 * the device state.
847	 */
848	wait_event(dev_state->wq, !atomic_read(&dev_state->count));
849	free_device_state(dev_state);
850}
851EXPORT_SYMBOL(amd_iommu_free_device);
852
853int amd_iommu_set_invalid_ppr_cb(struct pci_dev *pdev,
854				 amd_iommu_invalid_ppr_cb cb)
855{
856	struct device_state *dev_state;
857	unsigned long flags;
858	u16 devid;
859	int ret;
860
861	if (!amd_iommu_v2_supported())
862		return -ENODEV;
863
864	devid = device_id(pdev);
865
866	spin_lock_irqsave(&state_lock, flags);
867
868	ret = -EINVAL;
869	dev_state = __get_device_state(devid);
870	if (dev_state == NULL)
871		goto out_unlock;
872
873	dev_state->inv_ppr_cb = cb;
874
875	ret = 0;
876
877out_unlock:
878	spin_unlock_irqrestore(&state_lock, flags);
879
880	return ret;
881}
882EXPORT_SYMBOL(amd_iommu_set_invalid_ppr_cb);
883
884int amd_iommu_set_invalidate_ctx_cb(struct pci_dev *pdev,
885				    amd_iommu_invalidate_ctx cb)
886{
887	struct device_state *dev_state;
888	unsigned long flags;
889	u16 devid;
890	int ret;
891
892	if (!amd_iommu_v2_supported())
893		return -ENODEV;
894
895	devid = device_id(pdev);
896
897	spin_lock_irqsave(&state_lock, flags);
898
899	ret = -EINVAL;
900	dev_state = __get_device_state(devid);
901	if (dev_state == NULL)
902		goto out_unlock;
903
904	dev_state->inv_ctx_cb = cb;
905
906	ret = 0;
907
908out_unlock:
909	spin_unlock_irqrestore(&state_lock, flags);
910
911	return ret;
912}
913EXPORT_SYMBOL(amd_iommu_set_invalidate_ctx_cb);
914
915static int __init amd_iommu_v2_init(void)
916{
917	int ret;
918
919	pr_info("AMD IOMMUv2 driver by Joerg Roedel <jroedel@suse.de>\n");
920
921	if (!amd_iommu_v2_supported()) {
922		pr_info("AMD IOMMUv2 functionality not available on this system\n");
923		/*
924		 * Load anyway to provide the symbols to other modules
925		 * which may use AMD IOMMUv2 optionally.
926		 */
927		return 0;
928	}
929
930	spin_lock_init(&state_lock);
931
932	ret = -ENOMEM;
933	iommu_wq = create_workqueue("amd_iommu_v2");
934	if (iommu_wq == NULL)
935		goto out;
936
937	amd_iommu_register_ppr_notifier(&ppr_nb);
938
939	return 0;
940
941out:
942	return ret;
943}
944
945static void __exit amd_iommu_v2_exit(void)
946{
947	struct device_state *dev_state;
948	int i;
949
950	if (!amd_iommu_v2_supported())
951		return;
952
953	amd_iommu_unregister_ppr_notifier(&ppr_nb);
954
955	flush_workqueue(iommu_wq);
956
957	/*
958	 * The loop below might call flush_workqueue(), so call
959	 * destroy_workqueue() after it
960	 */
961	for (i = 0; i < MAX_DEVICES; ++i) {
962		dev_state = get_device_state(i);
963
964		if (dev_state == NULL)
965			continue;
966
967		WARN_ON_ONCE(1);
968
969		put_device_state(dev_state);
970		amd_iommu_free_device(dev_state->pdev);
971	}
972
973	destroy_workqueue(iommu_wq);
974}
975
976module_init(amd_iommu_v2_init);
977module_exit(amd_iommu_v2_exit);
978