xref: /openbmc/linux/drivers/iommu/amd/iommu_v2.c (revision e330fb14)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2010-2012 Advanced Micro Devices, Inc.
4  * Author: Joerg Roedel <jroedel@suse.de>
5  */
6 
7 #define pr_fmt(fmt)     "AMD-Vi: " fmt
8 
9 #include <linux/refcount.h>
10 #include <linux/mmu_notifier.h>
11 #include <linux/amd-iommu.h>
12 #include <linux/mm_types.h>
13 #include <linux/profile.h>
14 #include <linux/module.h>
15 #include <linux/sched.h>
16 #include <linux/sched/mm.h>
17 #include <linux/wait.h>
18 #include <linux/pci.h>
19 #include <linux/gfp.h>
20 
21 #include "amd_iommu.h"
22 
23 MODULE_LICENSE("GPL v2");
24 MODULE_AUTHOR("Joerg Roedel <jroedel@suse.de>");
25 
26 #define MAX_DEVICES		0x10000
27 #define PRI_QUEUE_SIZE		512
28 
29 struct pri_queue {
30 	atomic_t inflight;
31 	bool finish;
32 	int status;
33 };
34 
35 struct pasid_state {
36 	struct list_head list;			/* For global state-list */
37 	refcount_t count;				/* Reference count */
38 	unsigned mmu_notifier_count;		/* Counting nested mmu_notifier
39 						   calls */
40 	struct mm_struct *mm;			/* mm_struct for the faults */
41 	struct mmu_notifier mn;                 /* mmu_notifier handle */
42 	struct pri_queue pri[PRI_QUEUE_SIZE];	/* PRI tag states */
43 	struct device_state *device_state;	/* Link to our device_state */
44 	u32 pasid;				/* PASID index */
45 	bool invalid;				/* Used during setup and
46 						   teardown of the pasid */
47 	spinlock_t lock;			/* Protect pri_queues and
48 						   mmu_notifer_count */
49 	wait_queue_head_t wq;			/* To wait for count == 0 */
50 };
51 
52 struct device_state {
53 	struct list_head list;
54 	u16 devid;
55 	atomic_t count;
56 	struct pci_dev *pdev;
57 	struct pasid_state **states;
58 	struct iommu_domain *domain;
59 	int pasid_levels;
60 	int max_pasids;
61 	amd_iommu_invalid_ppr_cb inv_ppr_cb;
62 	amd_iommu_invalidate_ctx inv_ctx_cb;
63 	spinlock_t lock;
64 	wait_queue_head_t wq;
65 };
66 
67 struct fault {
68 	struct work_struct work;
69 	struct device_state *dev_state;
70 	struct pasid_state *state;
71 	struct mm_struct *mm;
72 	u64 address;
73 	u16 devid;
74 	u32 pasid;
75 	u16 tag;
76 	u16 finish;
77 	u16 flags;
78 };
79 
80 static LIST_HEAD(state_list);
81 static DEFINE_SPINLOCK(state_lock);
82 
83 static struct workqueue_struct *iommu_wq;
84 
85 static void free_pasid_states(struct device_state *dev_state);
86 
87 static u16 device_id(struct pci_dev *pdev)
88 {
89 	u16 devid;
90 
91 	devid = pdev->bus->number;
92 	devid = (devid << 8) | pdev->devfn;
93 
94 	return devid;
95 }
96 
97 static struct device_state *__get_device_state(u16 devid)
98 {
99 	struct device_state *dev_state;
100 
101 	list_for_each_entry(dev_state, &state_list, list) {
102 		if (dev_state->devid == devid)
103 			return dev_state;
104 	}
105 
106 	return NULL;
107 }
108 
109 static struct device_state *get_device_state(u16 devid)
110 {
111 	struct device_state *dev_state;
112 	unsigned long flags;
113 
114 	spin_lock_irqsave(&state_lock, flags);
115 	dev_state = __get_device_state(devid);
116 	if (dev_state != NULL)
117 		atomic_inc(&dev_state->count);
118 	spin_unlock_irqrestore(&state_lock, flags);
119 
120 	return dev_state;
121 }
122 
123 static void free_device_state(struct device_state *dev_state)
124 {
125 	struct iommu_group *group;
126 
127 	/*
128 	 * First detach device from domain - No more PRI requests will arrive
129 	 * from that device after it is unbound from the IOMMUv2 domain.
130 	 */
131 	group = iommu_group_get(&dev_state->pdev->dev);
132 	if (WARN_ON(!group))
133 		return;
134 
135 	iommu_detach_group(dev_state->domain, group);
136 
137 	iommu_group_put(group);
138 
139 	/* Everything is down now, free the IOMMUv2 domain */
140 	iommu_domain_free(dev_state->domain);
141 
142 	/* Finally get rid of the device-state */
143 	kfree(dev_state);
144 }
145 
146 static void put_device_state(struct device_state *dev_state)
147 {
148 	if (atomic_dec_and_test(&dev_state->count))
149 		wake_up(&dev_state->wq);
150 }
151 
152 /* Must be called under dev_state->lock */
153 static struct pasid_state **__get_pasid_state_ptr(struct device_state *dev_state,
154 						  u32 pasid, bool alloc)
155 {
156 	struct pasid_state **root, **ptr;
157 	int level, index;
158 
159 	level = dev_state->pasid_levels;
160 	root  = dev_state->states;
161 
162 	while (true) {
163 
164 		index = (pasid >> (9 * level)) & 0x1ff;
165 		ptr   = &root[index];
166 
167 		if (level == 0)
168 			break;
169 
170 		if (*ptr == NULL) {
171 			if (!alloc)
172 				return NULL;
173 
174 			*ptr = (void *)get_zeroed_page(GFP_ATOMIC);
175 			if (*ptr == NULL)
176 				return NULL;
177 		}
178 
179 		root   = (struct pasid_state **)*ptr;
180 		level -= 1;
181 	}
182 
183 	return ptr;
184 }
185 
186 static int set_pasid_state(struct device_state *dev_state,
187 			   struct pasid_state *pasid_state,
188 			   u32 pasid)
189 {
190 	struct pasid_state **ptr;
191 	unsigned long flags;
192 	int ret;
193 
194 	spin_lock_irqsave(&dev_state->lock, flags);
195 	ptr = __get_pasid_state_ptr(dev_state, pasid, true);
196 
197 	ret = -ENOMEM;
198 	if (ptr == NULL)
199 		goto out_unlock;
200 
201 	ret = -ENOMEM;
202 	if (*ptr != NULL)
203 		goto out_unlock;
204 
205 	*ptr = pasid_state;
206 
207 	ret = 0;
208 
209 out_unlock:
210 	spin_unlock_irqrestore(&dev_state->lock, flags);
211 
212 	return ret;
213 }
214 
215 static void clear_pasid_state(struct device_state *dev_state, u32 pasid)
216 {
217 	struct pasid_state **ptr;
218 	unsigned long flags;
219 
220 	spin_lock_irqsave(&dev_state->lock, flags);
221 	ptr = __get_pasid_state_ptr(dev_state, pasid, true);
222 
223 	if (ptr == NULL)
224 		goto out_unlock;
225 
226 	*ptr = NULL;
227 
228 out_unlock:
229 	spin_unlock_irqrestore(&dev_state->lock, flags);
230 }
231 
232 static struct pasid_state *get_pasid_state(struct device_state *dev_state,
233 					   u32 pasid)
234 {
235 	struct pasid_state **ptr, *ret = NULL;
236 	unsigned long flags;
237 
238 	spin_lock_irqsave(&dev_state->lock, flags);
239 	ptr = __get_pasid_state_ptr(dev_state, pasid, false);
240 
241 	if (ptr == NULL)
242 		goto out_unlock;
243 
244 	ret = *ptr;
245 	if (ret)
246 		refcount_inc(&ret->count);
247 
248 out_unlock:
249 	spin_unlock_irqrestore(&dev_state->lock, flags);
250 
251 	return ret;
252 }
253 
254 static void free_pasid_state(struct pasid_state *pasid_state)
255 {
256 	kfree(pasid_state);
257 }
258 
259 static void put_pasid_state(struct pasid_state *pasid_state)
260 {
261 	if (refcount_dec_and_test(&pasid_state->count))
262 		wake_up(&pasid_state->wq);
263 }
264 
265 static void put_pasid_state_wait(struct pasid_state *pasid_state)
266 {
267 	refcount_dec(&pasid_state->count);
268 	wait_event(pasid_state->wq, !refcount_read(&pasid_state->count));
269 	free_pasid_state(pasid_state);
270 }
271 
272 static void unbind_pasid(struct pasid_state *pasid_state)
273 {
274 	struct iommu_domain *domain;
275 
276 	domain = pasid_state->device_state->domain;
277 
278 	/*
279 	 * Mark pasid_state as invalid, no more faults will we added to the
280 	 * work queue after this is visible everywhere.
281 	 */
282 	pasid_state->invalid = true;
283 
284 	/* Make sure this is visible */
285 	smp_wmb();
286 
287 	/* After this the device/pasid can't access the mm anymore */
288 	amd_iommu_domain_clear_gcr3(domain, pasid_state->pasid);
289 
290 	/* Make sure no more pending faults are in the queue */
291 	flush_workqueue(iommu_wq);
292 }
293 
294 static void free_pasid_states_level1(struct pasid_state **tbl)
295 {
296 	int i;
297 
298 	for (i = 0; i < 512; ++i) {
299 		if (tbl[i] == NULL)
300 			continue;
301 
302 		free_page((unsigned long)tbl[i]);
303 	}
304 }
305 
306 static void free_pasid_states_level2(struct pasid_state **tbl)
307 {
308 	struct pasid_state **ptr;
309 	int i;
310 
311 	for (i = 0; i < 512; ++i) {
312 		if (tbl[i] == NULL)
313 			continue;
314 
315 		ptr = (struct pasid_state **)tbl[i];
316 		free_pasid_states_level1(ptr);
317 	}
318 }
319 
320 static void free_pasid_states(struct device_state *dev_state)
321 {
322 	struct pasid_state *pasid_state;
323 	int i;
324 
325 	for (i = 0; i < dev_state->max_pasids; ++i) {
326 		pasid_state = get_pasid_state(dev_state, i);
327 		if (pasid_state == NULL)
328 			continue;
329 
330 		put_pasid_state(pasid_state);
331 
332 		/*
333 		 * This will call the mn_release function and
334 		 * unbind the PASID
335 		 */
336 		mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
337 
338 		put_pasid_state_wait(pasid_state); /* Reference taken in
339 						      amd_iommu_bind_pasid */
340 
341 		/* Drop reference taken in amd_iommu_bind_pasid */
342 		put_device_state(dev_state);
343 	}
344 
345 	if (dev_state->pasid_levels == 2)
346 		free_pasid_states_level2(dev_state->states);
347 	else if (dev_state->pasid_levels == 1)
348 		free_pasid_states_level1(dev_state->states);
349 	else
350 		BUG_ON(dev_state->pasid_levels != 0);
351 
352 	free_page((unsigned long)dev_state->states);
353 }
354 
355 static struct pasid_state *mn_to_state(struct mmu_notifier *mn)
356 {
357 	return container_of(mn, struct pasid_state, mn);
358 }
359 
360 static void mn_invalidate_range(struct mmu_notifier *mn,
361 				struct mm_struct *mm,
362 				unsigned long start, unsigned long end)
363 {
364 	struct pasid_state *pasid_state;
365 	struct device_state *dev_state;
366 
367 	pasid_state = mn_to_state(mn);
368 	dev_state   = pasid_state->device_state;
369 
370 	if ((start ^ (end - 1)) < PAGE_SIZE)
371 		amd_iommu_flush_page(dev_state->domain, pasid_state->pasid,
372 				     start);
373 	else
374 		amd_iommu_flush_tlb(dev_state->domain, pasid_state->pasid);
375 }
376 
377 static void mn_release(struct mmu_notifier *mn, struct mm_struct *mm)
378 {
379 	struct pasid_state *pasid_state;
380 	struct device_state *dev_state;
381 	bool run_inv_ctx_cb;
382 
383 	might_sleep();
384 
385 	pasid_state    = mn_to_state(mn);
386 	dev_state      = pasid_state->device_state;
387 	run_inv_ctx_cb = !pasid_state->invalid;
388 
389 	if (run_inv_ctx_cb && dev_state->inv_ctx_cb)
390 		dev_state->inv_ctx_cb(dev_state->pdev, pasid_state->pasid);
391 
392 	unbind_pasid(pasid_state);
393 }
394 
395 static const struct mmu_notifier_ops iommu_mn = {
396 	.release		= mn_release,
397 	.invalidate_range       = mn_invalidate_range,
398 };
399 
400 static void set_pri_tag_status(struct pasid_state *pasid_state,
401 			       u16 tag, int status)
402 {
403 	unsigned long flags;
404 
405 	spin_lock_irqsave(&pasid_state->lock, flags);
406 	pasid_state->pri[tag].status = status;
407 	spin_unlock_irqrestore(&pasid_state->lock, flags);
408 }
409 
410 static void finish_pri_tag(struct device_state *dev_state,
411 			   struct pasid_state *pasid_state,
412 			   u16 tag)
413 {
414 	unsigned long flags;
415 
416 	spin_lock_irqsave(&pasid_state->lock, flags);
417 	if (atomic_dec_and_test(&pasid_state->pri[tag].inflight) &&
418 	    pasid_state->pri[tag].finish) {
419 		amd_iommu_complete_ppr(dev_state->pdev, pasid_state->pasid,
420 				       pasid_state->pri[tag].status, tag);
421 		pasid_state->pri[tag].finish = false;
422 		pasid_state->pri[tag].status = PPR_SUCCESS;
423 	}
424 	spin_unlock_irqrestore(&pasid_state->lock, flags);
425 }
426 
427 static void handle_fault_error(struct fault *fault)
428 {
429 	int status;
430 
431 	if (!fault->dev_state->inv_ppr_cb) {
432 		set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
433 		return;
434 	}
435 
436 	status = fault->dev_state->inv_ppr_cb(fault->dev_state->pdev,
437 					      fault->pasid,
438 					      fault->address,
439 					      fault->flags);
440 	switch (status) {
441 	case AMD_IOMMU_INV_PRI_RSP_SUCCESS:
442 		set_pri_tag_status(fault->state, fault->tag, PPR_SUCCESS);
443 		break;
444 	case AMD_IOMMU_INV_PRI_RSP_INVALID:
445 		set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
446 		break;
447 	case AMD_IOMMU_INV_PRI_RSP_FAIL:
448 		set_pri_tag_status(fault->state, fault->tag, PPR_FAILURE);
449 		break;
450 	default:
451 		BUG();
452 	}
453 }
454 
455 static bool access_error(struct vm_area_struct *vma, struct fault *fault)
456 {
457 	unsigned long requested = 0;
458 
459 	if (fault->flags & PPR_FAULT_EXEC)
460 		requested |= VM_EXEC;
461 
462 	if (fault->flags & PPR_FAULT_READ)
463 		requested |= VM_READ;
464 
465 	if (fault->flags & PPR_FAULT_WRITE)
466 		requested |= VM_WRITE;
467 
468 	return (requested & ~vma->vm_flags) != 0;
469 }
470 
471 static void do_fault(struct work_struct *work)
472 {
473 	struct fault *fault = container_of(work, struct fault, work);
474 	struct vm_area_struct *vma;
475 	vm_fault_t ret = VM_FAULT_ERROR;
476 	unsigned int flags = 0;
477 	struct mm_struct *mm;
478 	u64 address;
479 
480 	mm = fault->state->mm;
481 	address = fault->address;
482 
483 	if (fault->flags & PPR_FAULT_USER)
484 		flags |= FAULT_FLAG_USER;
485 	if (fault->flags & PPR_FAULT_WRITE)
486 		flags |= FAULT_FLAG_WRITE;
487 	flags |= FAULT_FLAG_REMOTE;
488 
489 	mmap_read_lock(mm);
490 	vma = find_extend_vma(mm, address);
491 	if (!vma || address < vma->vm_start)
492 		/* failed to get a vma in the right range */
493 		goto out;
494 
495 	/* Check if we have the right permissions on the vma */
496 	if (access_error(vma, fault))
497 		goto out;
498 
499 	ret = handle_mm_fault(vma, address, flags, NULL);
500 out:
501 	mmap_read_unlock(mm);
502 
503 	if (ret & VM_FAULT_ERROR)
504 		/* failed to service fault */
505 		handle_fault_error(fault);
506 
507 	finish_pri_tag(fault->dev_state, fault->state, fault->tag);
508 
509 	put_pasid_state(fault->state);
510 
511 	kfree(fault);
512 }
513 
514 static int ppr_notifier(struct notifier_block *nb, unsigned long e, void *data)
515 {
516 	struct amd_iommu_fault *iommu_fault;
517 	struct pasid_state *pasid_state;
518 	struct device_state *dev_state;
519 	struct pci_dev *pdev = NULL;
520 	unsigned long flags;
521 	struct fault *fault;
522 	bool finish;
523 	u16 tag, devid;
524 	int ret;
525 
526 	iommu_fault = data;
527 	tag         = iommu_fault->tag & 0x1ff;
528 	finish      = (iommu_fault->tag >> 9) & 1;
529 
530 	devid = iommu_fault->device_id;
531 	pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
532 					   devid & 0xff);
533 	if (!pdev)
534 		return -ENODEV;
535 
536 	ret = NOTIFY_DONE;
537 
538 	/* In kdump kernel pci dev is not initialized yet -> send INVALID */
539 	if (amd_iommu_is_attach_deferred(NULL, &pdev->dev)) {
540 		amd_iommu_complete_ppr(pdev, iommu_fault->pasid,
541 				       PPR_INVALID, tag);
542 		goto out;
543 	}
544 
545 	dev_state = get_device_state(iommu_fault->device_id);
546 	if (dev_state == NULL)
547 		goto out;
548 
549 	pasid_state = get_pasid_state(dev_state, iommu_fault->pasid);
550 	if (pasid_state == NULL || pasid_state->invalid) {
551 		/* We know the device but not the PASID -> send INVALID */
552 		amd_iommu_complete_ppr(dev_state->pdev, iommu_fault->pasid,
553 				       PPR_INVALID, tag);
554 		goto out_drop_state;
555 	}
556 
557 	spin_lock_irqsave(&pasid_state->lock, flags);
558 	atomic_inc(&pasid_state->pri[tag].inflight);
559 	if (finish)
560 		pasid_state->pri[tag].finish = true;
561 	spin_unlock_irqrestore(&pasid_state->lock, flags);
562 
563 	fault = kzalloc(sizeof(*fault), GFP_ATOMIC);
564 	if (fault == NULL) {
565 		/* We are OOM - send success and let the device re-fault */
566 		finish_pri_tag(dev_state, pasid_state, tag);
567 		goto out_drop_state;
568 	}
569 
570 	fault->dev_state = dev_state;
571 	fault->address   = iommu_fault->address;
572 	fault->state     = pasid_state;
573 	fault->tag       = tag;
574 	fault->finish    = finish;
575 	fault->pasid     = iommu_fault->pasid;
576 	fault->flags     = iommu_fault->flags;
577 	INIT_WORK(&fault->work, do_fault);
578 
579 	queue_work(iommu_wq, &fault->work);
580 
581 	ret = NOTIFY_OK;
582 
583 out_drop_state:
584 
585 	if (ret != NOTIFY_OK && pasid_state)
586 		put_pasid_state(pasid_state);
587 
588 	put_device_state(dev_state);
589 
590 out:
591 	return ret;
592 }
593 
594 static struct notifier_block ppr_nb = {
595 	.notifier_call = ppr_notifier,
596 };
597 
598 int amd_iommu_bind_pasid(struct pci_dev *pdev, u32 pasid,
599 			 struct task_struct *task)
600 {
601 	struct pasid_state *pasid_state;
602 	struct device_state *dev_state;
603 	struct mm_struct *mm;
604 	u16 devid;
605 	int ret;
606 
607 	might_sleep();
608 
609 	if (!amd_iommu_v2_supported())
610 		return -ENODEV;
611 
612 	devid     = device_id(pdev);
613 	dev_state = get_device_state(devid);
614 
615 	if (dev_state == NULL)
616 		return -EINVAL;
617 
618 	ret = -EINVAL;
619 	if (pasid >= dev_state->max_pasids)
620 		goto out;
621 
622 	ret = -ENOMEM;
623 	pasid_state = kzalloc(sizeof(*pasid_state), GFP_KERNEL);
624 	if (pasid_state == NULL)
625 		goto out;
626 
627 
628 	refcount_set(&pasid_state->count, 1);
629 	init_waitqueue_head(&pasid_state->wq);
630 	spin_lock_init(&pasid_state->lock);
631 
632 	mm                        = get_task_mm(task);
633 	pasid_state->mm           = mm;
634 	pasid_state->device_state = dev_state;
635 	pasid_state->pasid        = pasid;
636 	pasid_state->invalid      = true; /* Mark as valid only if we are
637 					     done with setting up the pasid */
638 	pasid_state->mn.ops       = &iommu_mn;
639 
640 	if (pasid_state->mm == NULL)
641 		goto out_free;
642 
643 	mmu_notifier_register(&pasid_state->mn, mm);
644 
645 	ret = set_pasid_state(dev_state, pasid_state, pasid);
646 	if (ret)
647 		goto out_unregister;
648 
649 	ret = amd_iommu_domain_set_gcr3(dev_state->domain, pasid,
650 					__pa(pasid_state->mm->pgd));
651 	if (ret)
652 		goto out_clear_state;
653 
654 	/* Now we are ready to handle faults */
655 	pasid_state->invalid = false;
656 
657 	/*
658 	 * Drop the reference to the mm_struct here. We rely on the
659 	 * mmu_notifier release call-back to inform us when the mm
660 	 * is going away.
661 	 */
662 	mmput(mm);
663 
664 	return 0;
665 
666 out_clear_state:
667 	clear_pasid_state(dev_state, pasid);
668 
669 out_unregister:
670 	mmu_notifier_unregister(&pasid_state->mn, mm);
671 	mmput(mm);
672 
673 out_free:
674 	free_pasid_state(pasid_state);
675 
676 out:
677 	put_device_state(dev_state);
678 
679 	return ret;
680 }
681 EXPORT_SYMBOL(amd_iommu_bind_pasid);
682 
683 void amd_iommu_unbind_pasid(struct pci_dev *pdev, u32 pasid)
684 {
685 	struct pasid_state *pasid_state;
686 	struct device_state *dev_state;
687 	u16 devid;
688 
689 	might_sleep();
690 
691 	if (!amd_iommu_v2_supported())
692 		return;
693 
694 	devid = device_id(pdev);
695 	dev_state = get_device_state(devid);
696 	if (dev_state == NULL)
697 		return;
698 
699 	if (pasid >= dev_state->max_pasids)
700 		goto out;
701 
702 	pasid_state = get_pasid_state(dev_state, pasid);
703 	if (pasid_state == NULL)
704 		goto out;
705 	/*
706 	 * Drop reference taken here. We are safe because we still hold
707 	 * the reference taken in the amd_iommu_bind_pasid function.
708 	 */
709 	put_pasid_state(pasid_state);
710 
711 	/* Clear the pasid state so that the pasid can be re-used */
712 	clear_pasid_state(dev_state, pasid_state->pasid);
713 
714 	/*
715 	 * Call mmu_notifier_unregister to drop our reference
716 	 * to pasid_state->mm
717 	 */
718 	mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
719 
720 	put_pasid_state_wait(pasid_state); /* Reference taken in
721 					      amd_iommu_bind_pasid */
722 out:
723 	/* Drop reference taken in this function */
724 	put_device_state(dev_state);
725 
726 	/* Drop reference taken in amd_iommu_bind_pasid */
727 	put_device_state(dev_state);
728 }
729 EXPORT_SYMBOL(amd_iommu_unbind_pasid);
730 
731 int amd_iommu_init_device(struct pci_dev *pdev, int pasids)
732 {
733 	struct device_state *dev_state;
734 	struct iommu_group *group;
735 	unsigned long flags;
736 	int ret, tmp;
737 	u16 devid;
738 
739 	might_sleep();
740 
741 	/*
742 	 * When memory encryption is active the device is likely not in a
743 	 * direct-mapped domain. Forbid using IOMMUv2 functionality for now.
744 	 */
745 	if (mem_encrypt_active())
746 		return -ENODEV;
747 
748 	if (!amd_iommu_v2_supported())
749 		return -ENODEV;
750 
751 	if (pasids <= 0 || pasids > (PASID_MASK + 1))
752 		return -EINVAL;
753 
754 	devid = device_id(pdev);
755 
756 	dev_state = kzalloc(sizeof(*dev_state), GFP_KERNEL);
757 	if (dev_state == NULL)
758 		return -ENOMEM;
759 
760 	spin_lock_init(&dev_state->lock);
761 	init_waitqueue_head(&dev_state->wq);
762 	dev_state->pdev  = pdev;
763 	dev_state->devid = devid;
764 
765 	tmp = pasids;
766 	for (dev_state->pasid_levels = 0; (tmp - 1) & ~0x1ff; tmp >>= 9)
767 		dev_state->pasid_levels += 1;
768 
769 	atomic_set(&dev_state->count, 1);
770 	dev_state->max_pasids = pasids;
771 
772 	ret = -ENOMEM;
773 	dev_state->states = (void *)get_zeroed_page(GFP_KERNEL);
774 	if (dev_state->states == NULL)
775 		goto out_free_dev_state;
776 
777 	dev_state->domain = iommu_domain_alloc(&pci_bus_type);
778 	if (dev_state->domain == NULL)
779 		goto out_free_states;
780 
781 	amd_iommu_domain_direct_map(dev_state->domain);
782 
783 	ret = amd_iommu_domain_enable_v2(dev_state->domain, pasids);
784 	if (ret)
785 		goto out_free_domain;
786 
787 	group = iommu_group_get(&pdev->dev);
788 	if (!group) {
789 		ret = -EINVAL;
790 		goto out_free_domain;
791 	}
792 
793 	ret = iommu_attach_group(dev_state->domain, group);
794 	if (ret != 0)
795 		goto out_drop_group;
796 
797 	iommu_group_put(group);
798 
799 	spin_lock_irqsave(&state_lock, flags);
800 
801 	if (__get_device_state(devid) != NULL) {
802 		spin_unlock_irqrestore(&state_lock, flags);
803 		ret = -EBUSY;
804 		goto out_free_domain;
805 	}
806 
807 	list_add_tail(&dev_state->list, &state_list);
808 
809 	spin_unlock_irqrestore(&state_lock, flags);
810 
811 	return 0;
812 
813 out_drop_group:
814 	iommu_group_put(group);
815 
816 out_free_domain:
817 	iommu_domain_free(dev_state->domain);
818 
819 out_free_states:
820 	free_page((unsigned long)dev_state->states);
821 
822 out_free_dev_state:
823 	kfree(dev_state);
824 
825 	return ret;
826 }
827 EXPORT_SYMBOL(amd_iommu_init_device);
828 
829 void amd_iommu_free_device(struct pci_dev *pdev)
830 {
831 	struct device_state *dev_state;
832 	unsigned long flags;
833 	u16 devid;
834 
835 	if (!amd_iommu_v2_supported())
836 		return;
837 
838 	devid = device_id(pdev);
839 
840 	spin_lock_irqsave(&state_lock, flags);
841 
842 	dev_state = __get_device_state(devid);
843 	if (dev_state == NULL) {
844 		spin_unlock_irqrestore(&state_lock, flags);
845 		return;
846 	}
847 
848 	list_del(&dev_state->list);
849 
850 	spin_unlock_irqrestore(&state_lock, flags);
851 
852 	/* Get rid of any remaining pasid states */
853 	free_pasid_states(dev_state);
854 
855 	put_device_state(dev_state);
856 	/*
857 	 * Wait until the last reference is dropped before freeing
858 	 * the device state.
859 	 */
860 	wait_event(dev_state->wq, !atomic_read(&dev_state->count));
861 	free_device_state(dev_state);
862 }
863 EXPORT_SYMBOL(amd_iommu_free_device);
864 
865 int amd_iommu_set_invalid_ppr_cb(struct pci_dev *pdev,
866 				 amd_iommu_invalid_ppr_cb cb)
867 {
868 	struct device_state *dev_state;
869 	unsigned long flags;
870 	u16 devid;
871 	int ret;
872 
873 	if (!amd_iommu_v2_supported())
874 		return -ENODEV;
875 
876 	devid = device_id(pdev);
877 
878 	spin_lock_irqsave(&state_lock, flags);
879 
880 	ret = -EINVAL;
881 	dev_state = __get_device_state(devid);
882 	if (dev_state == NULL)
883 		goto out_unlock;
884 
885 	dev_state->inv_ppr_cb = cb;
886 
887 	ret = 0;
888 
889 out_unlock:
890 	spin_unlock_irqrestore(&state_lock, flags);
891 
892 	return ret;
893 }
894 EXPORT_SYMBOL(amd_iommu_set_invalid_ppr_cb);
895 
896 int amd_iommu_set_invalidate_ctx_cb(struct pci_dev *pdev,
897 				    amd_iommu_invalidate_ctx cb)
898 {
899 	struct device_state *dev_state;
900 	unsigned long flags;
901 	u16 devid;
902 	int ret;
903 
904 	if (!amd_iommu_v2_supported())
905 		return -ENODEV;
906 
907 	devid = device_id(pdev);
908 
909 	spin_lock_irqsave(&state_lock, flags);
910 
911 	ret = -EINVAL;
912 	dev_state = __get_device_state(devid);
913 	if (dev_state == NULL)
914 		goto out_unlock;
915 
916 	dev_state->inv_ctx_cb = cb;
917 
918 	ret = 0;
919 
920 out_unlock:
921 	spin_unlock_irqrestore(&state_lock, flags);
922 
923 	return ret;
924 }
925 EXPORT_SYMBOL(amd_iommu_set_invalidate_ctx_cb);
926 
927 static int __init amd_iommu_v2_init(void)
928 {
929 	int ret;
930 
931 	pr_info("AMD IOMMUv2 driver by Joerg Roedel <jroedel@suse.de>\n");
932 
933 	if (!amd_iommu_v2_supported()) {
934 		pr_info("AMD IOMMUv2 functionality not available on this system\n");
935 		/*
936 		 * Load anyway to provide the symbols to other modules
937 		 * which may use AMD IOMMUv2 optionally.
938 		 */
939 		return 0;
940 	}
941 
942 	ret = -ENOMEM;
943 	iommu_wq = alloc_workqueue("amd_iommu_v2", WQ_MEM_RECLAIM, 0);
944 	if (iommu_wq == NULL)
945 		goto out;
946 
947 	amd_iommu_register_ppr_notifier(&ppr_nb);
948 
949 	return 0;
950 
951 out:
952 	return ret;
953 }
954 
955 static void __exit amd_iommu_v2_exit(void)
956 {
957 	struct device_state *dev_state;
958 	int i;
959 
960 	if (!amd_iommu_v2_supported())
961 		return;
962 
963 	amd_iommu_unregister_ppr_notifier(&ppr_nb);
964 
965 	flush_workqueue(iommu_wq);
966 
967 	/*
968 	 * The loop below might call flush_workqueue(), so call
969 	 * destroy_workqueue() after it
970 	 */
971 	for (i = 0; i < MAX_DEVICES; ++i) {
972 		dev_state = get_device_state(i);
973 
974 		if (dev_state == NULL)
975 			continue;
976 
977 		WARN_ON_ONCE(1);
978 
979 		put_device_state(dev_state);
980 		amd_iommu_free_device(dev_state->pdev);
981 	}
982 
983 	destroy_workqueue(iommu_wq);
984 }
985 
986 module_init(amd_iommu_v2_init);
987 module_exit(amd_iommu_v2_exit);
988