xref: /openbmc/linux/drivers/iommu/amd/iommu_v2.c (revision 31eeb6b0)
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 #include <linux/cc_platform.h>
21 
22 #include "amd_iommu.h"
23 
24 MODULE_LICENSE("GPL v2");
25 MODULE_AUTHOR("Joerg Roedel <jroedel@suse.de>");
26 
27 #define MAX_DEVICES		0x10000
28 #define PRI_QUEUE_SIZE		512
29 
30 struct pri_queue {
31 	atomic_t inflight;
32 	bool finish;
33 	int status;
34 };
35 
36 struct pasid_state {
37 	struct list_head list;			/* For global state-list */
38 	refcount_t count;				/* Reference count */
39 	unsigned mmu_notifier_count;		/* Counting nested mmu_notifier
40 						   calls */
41 	struct mm_struct *mm;			/* mm_struct for the faults */
42 	struct mmu_notifier mn;                 /* mmu_notifier handle */
43 	struct pri_queue pri[PRI_QUEUE_SIZE];	/* PRI tag states */
44 	struct device_state *device_state;	/* Link to our device_state */
45 	u32 pasid;				/* PASID index */
46 	bool invalid;				/* Used during setup and
47 						   teardown of the pasid */
48 	spinlock_t lock;			/* Protect pri_queues and
49 						   mmu_notifer_count */
50 	wait_queue_head_t wq;			/* To wait for count == 0 */
51 };
52 
53 struct device_state {
54 	struct list_head list;
55 	u16 devid;
56 	atomic_t count;
57 	struct pci_dev *pdev;
58 	struct pasid_state **states;
59 	struct iommu_domain *domain;
60 	int pasid_levels;
61 	int max_pasids;
62 	amd_iommu_invalid_ppr_cb inv_ppr_cb;
63 	amd_iommu_invalidate_ctx inv_ctx_cb;
64 	spinlock_t lock;
65 	wait_queue_head_t wq;
66 };
67 
68 struct fault {
69 	struct work_struct work;
70 	struct device_state *dev_state;
71 	struct pasid_state *state;
72 	struct mm_struct *mm;
73 	u64 address;
74 	u16 devid;
75 	u32 pasid;
76 	u16 tag;
77 	u16 finish;
78 	u16 flags;
79 };
80 
81 static LIST_HEAD(state_list);
82 static DEFINE_SPINLOCK(state_lock);
83 
84 static struct workqueue_struct *iommu_wq;
85 
86 static void free_pasid_states(struct device_state *dev_state);
87 
88 static u16 device_id(struct pci_dev *pdev)
89 {
90 	u16 devid;
91 
92 	devid = pdev->bus->number;
93 	devid = (devid << 8) | pdev->devfn;
94 
95 	return devid;
96 }
97 
98 static struct device_state *__get_device_state(u16 devid)
99 {
100 	struct device_state *dev_state;
101 
102 	list_for_each_entry(dev_state, &state_list, list) {
103 		if (dev_state->devid == devid)
104 			return dev_state;
105 	}
106 
107 	return NULL;
108 }
109 
110 static struct device_state *get_device_state(u16 devid)
111 {
112 	struct device_state *dev_state;
113 	unsigned long flags;
114 
115 	spin_lock_irqsave(&state_lock, flags);
116 	dev_state = __get_device_state(devid);
117 	if (dev_state != NULL)
118 		atomic_inc(&dev_state->count);
119 	spin_unlock_irqrestore(&state_lock, flags);
120 
121 	return dev_state;
122 }
123 
124 static void free_device_state(struct device_state *dev_state)
125 {
126 	struct iommu_group *group;
127 
128 	/*
129 	 * First detach device from domain - No more PRI requests will arrive
130 	 * from that device after it is unbound from the IOMMUv2 domain.
131 	 */
132 	group = iommu_group_get(&dev_state->pdev->dev);
133 	if (WARN_ON(!group))
134 		return;
135 
136 	iommu_detach_group(dev_state->domain, group);
137 
138 	iommu_group_put(group);
139 
140 	/* Everything is down now, free the IOMMUv2 domain */
141 	iommu_domain_free(dev_state->domain);
142 
143 	/* Finally get rid of the device-state */
144 	kfree(dev_state);
145 }
146 
147 static void put_device_state(struct device_state *dev_state)
148 {
149 	if (atomic_dec_and_test(&dev_state->count))
150 		wake_up(&dev_state->wq);
151 }
152 
153 /* Must be called under dev_state->lock */
154 static struct pasid_state **__get_pasid_state_ptr(struct device_state *dev_state,
155 						  u32 pasid, bool alloc)
156 {
157 	struct pasid_state **root, **ptr;
158 	int level, index;
159 
160 	level = dev_state->pasid_levels;
161 	root  = dev_state->states;
162 
163 	while (true) {
164 
165 		index = (pasid >> (9 * level)) & 0x1ff;
166 		ptr   = &root[index];
167 
168 		if (level == 0)
169 			break;
170 
171 		if (*ptr == NULL) {
172 			if (!alloc)
173 				return NULL;
174 
175 			*ptr = (void *)get_zeroed_page(GFP_ATOMIC);
176 			if (*ptr == NULL)
177 				return NULL;
178 		}
179 
180 		root   = (struct pasid_state **)*ptr;
181 		level -= 1;
182 	}
183 
184 	return ptr;
185 }
186 
187 static int set_pasid_state(struct device_state *dev_state,
188 			   struct pasid_state *pasid_state,
189 			   u32 pasid)
190 {
191 	struct pasid_state **ptr;
192 	unsigned long flags;
193 	int ret;
194 
195 	spin_lock_irqsave(&dev_state->lock, flags);
196 	ptr = __get_pasid_state_ptr(dev_state, pasid, true);
197 
198 	ret = -ENOMEM;
199 	if (ptr == NULL)
200 		goto out_unlock;
201 
202 	ret = -ENOMEM;
203 	if (*ptr != NULL)
204 		goto out_unlock;
205 
206 	*ptr = pasid_state;
207 
208 	ret = 0;
209 
210 out_unlock:
211 	spin_unlock_irqrestore(&dev_state->lock, flags);
212 
213 	return ret;
214 }
215 
216 static void clear_pasid_state(struct device_state *dev_state, u32 pasid)
217 {
218 	struct pasid_state **ptr;
219 	unsigned long flags;
220 
221 	spin_lock_irqsave(&dev_state->lock, flags);
222 	ptr = __get_pasid_state_ptr(dev_state, pasid, true);
223 
224 	if (ptr == NULL)
225 		goto out_unlock;
226 
227 	*ptr = NULL;
228 
229 out_unlock:
230 	spin_unlock_irqrestore(&dev_state->lock, flags);
231 }
232 
233 static struct pasid_state *get_pasid_state(struct device_state *dev_state,
234 					   u32 pasid)
235 {
236 	struct pasid_state **ptr, *ret = NULL;
237 	unsigned long flags;
238 
239 	spin_lock_irqsave(&dev_state->lock, flags);
240 	ptr = __get_pasid_state_ptr(dev_state, pasid, false);
241 
242 	if (ptr == NULL)
243 		goto out_unlock;
244 
245 	ret = *ptr;
246 	if (ret)
247 		refcount_inc(&ret->count);
248 
249 out_unlock:
250 	spin_unlock_irqrestore(&dev_state->lock, flags);
251 
252 	return ret;
253 }
254 
255 static void free_pasid_state(struct pasid_state *pasid_state)
256 {
257 	kfree(pasid_state);
258 }
259 
260 static void put_pasid_state(struct pasid_state *pasid_state)
261 {
262 	if (refcount_dec_and_test(&pasid_state->count))
263 		wake_up(&pasid_state->wq);
264 }
265 
266 static void put_pasid_state_wait(struct pasid_state *pasid_state)
267 {
268 	refcount_dec(&pasid_state->count);
269 	wait_event(pasid_state->wq, !refcount_read(&pasid_state->count));
270 	free_pasid_state(pasid_state);
271 }
272 
273 static void unbind_pasid(struct pasid_state *pasid_state)
274 {
275 	struct iommu_domain *domain;
276 
277 	domain = pasid_state->device_state->domain;
278 
279 	/*
280 	 * Mark pasid_state as invalid, no more faults will we added to the
281 	 * work queue after this is visible everywhere.
282 	 */
283 	pasid_state->invalid = true;
284 
285 	/* Make sure this is visible */
286 	smp_wmb();
287 
288 	/* After this the device/pasid can't access the mm anymore */
289 	amd_iommu_domain_clear_gcr3(domain, pasid_state->pasid);
290 
291 	/* Make sure no more pending faults are in the queue */
292 	flush_workqueue(iommu_wq);
293 }
294 
295 static void free_pasid_states_level1(struct pasid_state **tbl)
296 {
297 	int i;
298 
299 	for (i = 0; i < 512; ++i) {
300 		if (tbl[i] == NULL)
301 			continue;
302 
303 		free_page((unsigned long)tbl[i]);
304 	}
305 }
306 
307 static void free_pasid_states_level2(struct pasid_state **tbl)
308 {
309 	struct pasid_state **ptr;
310 	int i;
311 
312 	for (i = 0; i < 512; ++i) {
313 		if (tbl[i] == NULL)
314 			continue;
315 
316 		ptr = (struct pasid_state **)tbl[i];
317 		free_pasid_states_level1(ptr);
318 	}
319 }
320 
321 static void free_pasid_states(struct device_state *dev_state)
322 {
323 	struct pasid_state *pasid_state;
324 	int i;
325 
326 	for (i = 0; i < dev_state->max_pasids; ++i) {
327 		pasid_state = get_pasid_state(dev_state, i);
328 		if (pasid_state == NULL)
329 			continue;
330 
331 		put_pasid_state(pasid_state);
332 
333 		/*
334 		 * This will call the mn_release function and
335 		 * unbind the PASID
336 		 */
337 		mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
338 
339 		put_pasid_state_wait(pasid_state); /* Reference taken in
340 						      amd_iommu_bind_pasid */
341 
342 		/* Drop reference taken in amd_iommu_bind_pasid */
343 		put_device_state(dev_state);
344 	}
345 
346 	if (dev_state->pasid_levels == 2)
347 		free_pasid_states_level2(dev_state->states);
348 	else if (dev_state->pasid_levels == 1)
349 		free_pasid_states_level1(dev_state->states);
350 	else
351 		BUG_ON(dev_state->pasid_levels != 0);
352 
353 	free_page((unsigned long)dev_state->states);
354 }
355 
356 static struct pasid_state *mn_to_state(struct mmu_notifier *mn)
357 {
358 	return container_of(mn, struct pasid_state, mn);
359 }
360 
361 static void mn_invalidate_range(struct mmu_notifier *mn,
362 				struct mm_struct *mm,
363 				unsigned long start, unsigned long end)
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 	if ((start ^ (end - 1)) < PAGE_SIZE)
372 		amd_iommu_flush_page(dev_state->domain, pasid_state->pasid,
373 				     start);
374 	else
375 		amd_iommu_flush_tlb(dev_state->domain, pasid_state->pasid);
376 }
377 
378 static void mn_release(struct mmu_notifier *mn, struct mm_struct *mm)
379 {
380 	struct pasid_state *pasid_state;
381 	struct device_state *dev_state;
382 	bool run_inv_ctx_cb;
383 
384 	might_sleep();
385 
386 	pasid_state    = mn_to_state(mn);
387 	dev_state      = pasid_state->device_state;
388 	run_inv_ctx_cb = !pasid_state->invalid;
389 
390 	if (run_inv_ctx_cb && dev_state->inv_ctx_cb)
391 		dev_state->inv_ctx_cb(dev_state->pdev, pasid_state->pasid);
392 
393 	unbind_pasid(pasid_state);
394 }
395 
396 static const struct mmu_notifier_ops iommu_mn = {
397 	.release		= mn_release,
398 	.invalidate_range       = mn_invalidate_range,
399 };
400 
401 static void set_pri_tag_status(struct pasid_state *pasid_state,
402 			       u16 tag, int status)
403 {
404 	unsigned long flags;
405 
406 	spin_lock_irqsave(&pasid_state->lock, flags);
407 	pasid_state->pri[tag].status = status;
408 	spin_unlock_irqrestore(&pasid_state->lock, flags);
409 }
410 
411 static void finish_pri_tag(struct device_state *dev_state,
412 			   struct pasid_state *pasid_state,
413 			   u16 tag)
414 {
415 	unsigned long flags;
416 
417 	spin_lock_irqsave(&pasid_state->lock, flags);
418 	if (atomic_dec_and_test(&pasid_state->pri[tag].inflight) &&
419 	    pasid_state->pri[tag].finish) {
420 		amd_iommu_complete_ppr(dev_state->pdev, pasid_state->pasid,
421 				       pasid_state->pri[tag].status, tag);
422 		pasid_state->pri[tag].finish = false;
423 		pasid_state->pri[tag].status = PPR_SUCCESS;
424 	}
425 	spin_unlock_irqrestore(&pasid_state->lock, flags);
426 }
427 
428 static void handle_fault_error(struct fault *fault)
429 {
430 	int status;
431 
432 	if (!fault->dev_state->inv_ppr_cb) {
433 		set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
434 		return;
435 	}
436 
437 	status = fault->dev_state->inv_ppr_cb(fault->dev_state->pdev,
438 					      fault->pasid,
439 					      fault->address,
440 					      fault->flags);
441 	switch (status) {
442 	case AMD_IOMMU_INV_PRI_RSP_SUCCESS:
443 		set_pri_tag_status(fault->state, fault->tag, PPR_SUCCESS);
444 		break;
445 	case AMD_IOMMU_INV_PRI_RSP_INVALID:
446 		set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
447 		break;
448 	case AMD_IOMMU_INV_PRI_RSP_FAIL:
449 		set_pri_tag_status(fault->state, fault->tag, PPR_FAILURE);
450 		break;
451 	default:
452 		BUG();
453 	}
454 }
455 
456 static bool access_error(struct vm_area_struct *vma, struct fault *fault)
457 {
458 	unsigned long requested = 0;
459 
460 	if (fault->flags & PPR_FAULT_EXEC)
461 		requested |= VM_EXEC;
462 
463 	if (fault->flags & PPR_FAULT_READ)
464 		requested |= VM_READ;
465 
466 	if (fault->flags & PPR_FAULT_WRITE)
467 		requested |= VM_WRITE;
468 
469 	return (requested & ~vma->vm_flags) != 0;
470 }
471 
472 static void do_fault(struct work_struct *work)
473 {
474 	struct fault *fault = container_of(work, struct fault, work);
475 	struct vm_area_struct *vma;
476 	vm_fault_t ret = VM_FAULT_ERROR;
477 	unsigned int flags = 0;
478 	struct mm_struct *mm;
479 	u64 address;
480 
481 	mm = fault->state->mm;
482 	address = fault->address;
483 
484 	if (fault->flags & PPR_FAULT_USER)
485 		flags |= FAULT_FLAG_USER;
486 	if (fault->flags & PPR_FAULT_WRITE)
487 		flags |= FAULT_FLAG_WRITE;
488 	flags |= FAULT_FLAG_REMOTE;
489 
490 	mmap_read_lock(mm);
491 	vma = find_extend_vma(mm, address);
492 	if (!vma || address < vma->vm_start)
493 		/* failed to get a vma in the right range */
494 		goto out;
495 
496 	/* Check if we have the right permissions on the vma */
497 	if (access_error(vma, fault))
498 		goto out;
499 
500 	ret = handle_mm_fault(vma, address, flags, NULL);
501 out:
502 	mmap_read_unlock(mm);
503 
504 	if (ret & VM_FAULT_ERROR)
505 		/* failed to service fault */
506 		handle_fault_error(fault);
507 
508 	finish_pri_tag(fault->dev_state, fault->state, fault->tag);
509 
510 	put_pasid_state(fault->state);
511 
512 	kfree(fault);
513 }
514 
515 static int ppr_notifier(struct notifier_block *nb, unsigned long e, void *data)
516 {
517 	struct amd_iommu_fault *iommu_fault;
518 	struct pasid_state *pasid_state;
519 	struct device_state *dev_state;
520 	struct pci_dev *pdev = NULL;
521 	unsigned long flags;
522 	struct fault *fault;
523 	bool finish;
524 	u16 tag, devid;
525 	int ret;
526 
527 	iommu_fault = data;
528 	tag         = iommu_fault->tag & 0x1ff;
529 	finish      = (iommu_fault->tag >> 9) & 1;
530 
531 	devid = iommu_fault->device_id;
532 	pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
533 					   devid & 0xff);
534 	if (!pdev)
535 		return -ENODEV;
536 
537 	ret = NOTIFY_DONE;
538 
539 	/* In kdump kernel pci dev is not initialized yet -> send INVALID */
540 	if (amd_iommu_is_attach_deferred(NULL, &pdev->dev)) {
541 		amd_iommu_complete_ppr(pdev, iommu_fault->pasid,
542 				       PPR_INVALID, tag);
543 		goto out;
544 	}
545 
546 	dev_state = get_device_state(iommu_fault->device_id);
547 	if (dev_state == NULL)
548 		goto out;
549 
550 	pasid_state = get_pasid_state(dev_state, iommu_fault->pasid);
551 	if (pasid_state == NULL || pasid_state->invalid) {
552 		/* We know the device but not the PASID -> send INVALID */
553 		amd_iommu_complete_ppr(dev_state->pdev, iommu_fault->pasid,
554 				       PPR_INVALID, tag);
555 		goto out_drop_state;
556 	}
557 
558 	spin_lock_irqsave(&pasid_state->lock, flags);
559 	atomic_inc(&pasid_state->pri[tag].inflight);
560 	if (finish)
561 		pasid_state->pri[tag].finish = true;
562 	spin_unlock_irqrestore(&pasid_state->lock, flags);
563 
564 	fault = kzalloc(sizeof(*fault), GFP_ATOMIC);
565 	if (fault == NULL) {
566 		/* We are OOM - send success and let the device re-fault */
567 		finish_pri_tag(dev_state, pasid_state, tag);
568 		goto out_drop_state;
569 	}
570 
571 	fault->dev_state = dev_state;
572 	fault->address   = iommu_fault->address;
573 	fault->state     = pasid_state;
574 	fault->tag       = tag;
575 	fault->finish    = finish;
576 	fault->pasid     = iommu_fault->pasid;
577 	fault->flags     = iommu_fault->flags;
578 	INIT_WORK(&fault->work, do_fault);
579 
580 	queue_work(iommu_wq, &fault->work);
581 
582 	ret = NOTIFY_OK;
583 
584 out_drop_state:
585 
586 	if (ret != NOTIFY_OK && pasid_state)
587 		put_pasid_state(pasid_state);
588 
589 	put_device_state(dev_state);
590 
591 out:
592 	return ret;
593 }
594 
595 static struct notifier_block ppr_nb = {
596 	.notifier_call = ppr_notifier,
597 };
598 
599 int amd_iommu_bind_pasid(struct pci_dev *pdev, u32 pasid,
600 			 struct task_struct *task)
601 {
602 	struct pasid_state *pasid_state;
603 	struct device_state *dev_state;
604 	struct mm_struct *mm;
605 	u16 devid;
606 	int ret;
607 
608 	might_sleep();
609 
610 	if (!amd_iommu_v2_supported())
611 		return -ENODEV;
612 
613 	devid     = device_id(pdev);
614 	dev_state = get_device_state(devid);
615 
616 	if (dev_state == NULL)
617 		return -EINVAL;
618 
619 	ret = -EINVAL;
620 	if (pasid >= dev_state->max_pasids)
621 		goto out;
622 
623 	ret = -ENOMEM;
624 	pasid_state = kzalloc(sizeof(*pasid_state), GFP_KERNEL);
625 	if (pasid_state == NULL)
626 		goto out;
627 
628 
629 	refcount_set(&pasid_state->count, 1);
630 	init_waitqueue_head(&pasid_state->wq);
631 	spin_lock_init(&pasid_state->lock);
632 
633 	mm                        = get_task_mm(task);
634 	pasid_state->mm           = mm;
635 	pasid_state->device_state = dev_state;
636 	pasid_state->pasid        = pasid;
637 	pasid_state->invalid      = true; /* Mark as valid only if we are
638 					     done with setting up the pasid */
639 	pasid_state->mn.ops       = &iommu_mn;
640 
641 	if (pasid_state->mm == NULL)
642 		goto out_free;
643 
644 	mmu_notifier_register(&pasid_state->mn, mm);
645 
646 	ret = set_pasid_state(dev_state, pasid_state, pasid);
647 	if (ret)
648 		goto out_unregister;
649 
650 	ret = amd_iommu_domain_set_gcr3(dev_state->domain, pasid,
651 					__pa(pasid_state->mm->pgd));
652 	if (ret)
653 		goto out_clear_state;
654 
655 	/* Now we are ready to handle faults */
656 	pasid_state->invalid = false;
657 
658 	/*
659 	 * Drop the reference to the mm_struct here. We rely on the
660 	 * mmu_notifier release call-back to inform us when the mm
661 	 * is going away.
662 	 */
663 	mmput(mm);
664 
665 	return 0;
666 
667 out_clear_state:
668 	clear_pasid_state(dev_state, pasid);
669 
670 out_unregister:
671 	mmu_notifier_unregister(&pasid_state->mn, mm);
672 	mmput(mm);
673 
674 out_free:
675 	free_pasid_state(pasid_state);
676 
677 out:
678 	put_device_state(dev_state);
679 
680 	return ret;
681 }
682 EXPORT_SYMBOL(amd_iommu_bind_pasid);
683 
684 void amd_iommu_unbind_pasid(struct pci_dev *pdev, u32 pasid)
685 {
686 	struct pasid_state *pasid_state;
687 	struct device_state *dev_state;
688 	u16 devid;
689 
690 	might_sleep();
691 
692 	if (!amd_iommu_v2_supported())
693 		return;
694 
695 	devid = device_id(pdev);
696 	dev_state = get_device_state(devid);
697 	if (dev_state == NULL)
698 		return;
699 
700 	if (pasid >= dev_state->max_pasids)
701 		goto out;
702 
703 	pasid_state = get_pasid_state(dev_state, pasid);
704 	if (pasid_state == NULL)
705 		goto out;
706 	/*
707 	 * Drop reference taken here. We are safe because we still hold
708 	 * the reference taken in the amd_iommu_bind_pasid function.
709 	 */
710 	put_pasid_state(pasid_state);
711 
712 	/* Clear the pasid state so that the pasid can be re-used */
713 	clear_pasid_state(dev_state, pasid_state->pasid);
714 
715 	/*
716 	 * Call mmu_notifier_unregister to drop our reference
717 	 * to pasid_state->mm
718 	 */
719 	mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
720 
721 	put_pasid_state_wait(pasid_state); /* Reference taken in
722 					      amd_iommu_bind_pasid */
723 out:
724 	/* Drop reference taken in this function */
725 	put_device_state(dev_state);
726 
727 	/* Drop reference taken in amd_iommu_bind_pasid */
728 	put_device_state(dev_state);
729 }
730 EXPORT_SYMBOL(amd_iommu_unbind_pasid);
731 
732 int amd_iommu_init_device(struct pci_dev *pdev, int pasids)
733 {
734 	struct device_state *dev_state;
735 	struct iommu_group *group;
736 	unsigned long flags;
737 	int ret, tmp;
738 	u16 devid;
739 
740 	might_sleep();
741 
742 	/*
743 	 * When memory encryption is active the device is likely not in a
744 	 * direct-mapped domain. Forbid using IOMMUv2 functionality for now.
745 	 */
746 	if (cc_platform_has(CC_ATTR_MEM_ENCRYPT))
747 		return -ENODEV;
748 
749 	if (!amd_iommu_v2_supported())
750 		return -ENODEV;
751 
752 	if (pasids <= 0 || pasids > (PASID_MASK + 1))
753 		return -EINVAL;
754 
755 	devid = device_id(pdev);
756 
757 	dev_state = kzalloc(sizeof(*dev_state), GFP_KERNEL);
758 	if (dev_state == NULL)
759 		return -ENOMEM;
760 
761 	spin_lock_init(&dev_state->lock);
762 	init_waitqueue_head(&dev_state->wq);
763 	dev_state->pdev  = pdev;
764 	dev_state->devid = devid;
765 
766 	tmp = pasids;
767 	for (dev_state->pasid_levels = 0; (tmp - 1) & ~0x1ff; tmp >>= 9)
768 		dev_state->pasid_levels += 1;
769 
770 	atomic_set(&dev_state->count, 1);
771 	dev_state->max_pasids = pasids;
772 
773 	ret = -ENOMEM;
774 	dev_state->states = (void *)get_zeroed_page(GFP_KERNEL);
775 	if (dev_state->states == NULL)
776 		goto out_free_dev_state;
777 
778 	dev_state->domain = iommu_domain_alloc(&pci_bus_type);
779 	if (dev_state->domain == NULL)
780 		goto out_free_states;
781 
782 	amd_iommu_domain_direct_map(dev_state->domain);
783 
784 	ret = amd_iommu_domain_enable_v2(dev_state->domain, pasids);
785 	if (ret)
786 		goto out_free_domain;
787 
788 	group = iommu_group_get(&pdev->dev);
789 	if (!group) {
790 		ret = -EINVAL;
791 		goto out_free_domain;
792 	}
793 
794 	ret = iommu_attach_group(dev_state->domain, group);
795 	if (ret != 0)
796 		goto out_drop_group;
797 
798 	iommu_group_put(group);
799 
800 	spin_lock_irqsave(&state_lock, flags);
801 
802 	if (__get_device_state(devid) != NULL) {
803 		spin_unlock_irqrestore(&state_lock, flags);
804 		ret = -EBUSY;
805 		goto out_free_domain;
806 	}
807 
808 	list_add_tail(&dev_state->list, &state_list);
809 
810 	spin_unlock_irqrestore(&state_lock, flags);
811 
812 	return 0;
813 
814 out_drop_group:
815 	iommu_group_put(group);
816 
817 out_free_domain:
818 	iommu_domain_free(dev_state->domain);
819 
820 out_free_states:
821 	free_page((unsigned long)dev_state->states);
822 
823 out_free_dev_state:
824 	kfree(dev_state);
825 
826 	return ret;
827 }
828 EXPORT_SYMBOL(amd_iommu_init_device);
829 
830 void amd_iommu_free_device(struct pci_dev *pdev)
831 {
832 	struct device_state *dev_state;
833 	unsigned long flags;
834 	u16 devid;
835 
836 	if (!amd_iommu_v2_supported())
837 		return;
838 
839 	devid = device_id(pdev);
840 
841 	spin_lock_irqsave(&state_lock, flags);
842 
843 	dev_state = __get_device_state(devid);
844 	if (dev_state == NULL) {
845 		spin_unlock_irqrestore(&state_lock, flags);
846 		return;
847 	}
848 
849 	list_del(&dev_state->list);
850 
851 	spin_unlock_irqrestore(&state_lock, flags);
852 
853 	/* Get rid of any remaining pasid states */
854 	free_pasid_states(dev_state);
855 
856 	put_device_state(dev_state);
857 	/*
858 	 * Wait until the last reference is dropped before freeing
859 	 * the device state.
860 	 */
861 	wait_event(dev_state->wq, !atomic_read(&dev_state->count));
862 	free_device_state(dev_state);
863 }
864 EXPORT_SYMBOL(amd_iommu_free_device);
865 
866 int amd_iommu_set_invalid_ppr_cb(struct pci_dev *pdev,
867 				 amd_iommu_invalid_ppr_cb cb)
868 {
869 	struct device_state *dev_state;
870 	unsigned long flags;
871 	u16 devid;
872 	int ret;
873 
874 	if (!amd_iommu_v2_supported())
875 		return -ENODEV;
876 
877 	devid = device_id(pdev);
878 
879 	spin_lock_irqsave(&state_lock, flags);
880 
881 	ret = -EINVAL;
882 	dev_state = __get_device_state(devid);
883 	if (dev_state == NULL)
884 		goto out_unlock;
885 
886 	dev_state->inv_ppr_cb = cb;
887 
888 	ret = 0;
889 
890 out_unlock:
891 	spin_unlock_irqrestore(&state_lock, flags);
892 
893 	return ret;
894 }
895 EXPORT_SYMBOL(amd_iommu_set_invalid_ppr_cb);
896 
897 int amd_iommu_set_invalidate_ctx_cb(struct pci_dev *pdev,
898 				    amd_iommu_invalidate_ctx cb)
899 {
900 	struct device_state *dev_state;
901 	unsigned long flags;
902 	u16 devid;
903 	int ret;
904 
905 	if (!amd_iommu_v2_supported())
906 		return -ENODEV;
907 
908 	devid = device_id(pdev);
909 
910 	spin_lock_irqsave(&state_lock, flags);
911 
912 	ret = -EINVAL;
913 	dev_state = __get_device_state(devid);
914 	if (dev_state == NULL)
915 		goto out_unlock;
916 
917 	dev_state->inv_ctx_cb = cb;
918 
919 	ret = 0;
920 
921 out_unlock:
922 	spin_unlock_irqrestore(&state_lock, flags);
923 
924 	return ret;
925 }
926 EXPORT_SYMBOL(amd_iommu_set_invalidate_ctx_cb);
927 
928 static int __init amd_iommu_v2_init(void)
929 {
930 	int ret;
931 
932 	if (!amd_iommu_v2_supported()) {
933 		pr_info("AMD IOMMUv2 functionality not available on this system - This is not a bug.\n");
934 		/*
935 		 * Load anyway to provide the symbols to other modules
936 		 * which may use AMD IOMMUv2 optionally.
937 		 */
938 		return 0;
939 	}
940 
941 	ret = -ENOMEM;
942 	iommu_wq = alloc_workqueue("amd_iommu_v2", WQ_MEM_RECLAIM, 0);
943 	if (iommu_wq == NULL)
944 		goto out;
945 
946 	amd_iommu_register_ppr_notifier(&ppr_nb);
947 
948 	pr_info("AMD IOMMUv2 loaded and initialized\n");
949 
950 	return 0;
951 
952 out:
953 	return ret;
954 }
955 
956 static void __exit amd_iommu_v2_exit(void)
957 {
958 	struct device_state *dev_state;
959 	int i;
960 
961 	if (!amd_iommu_v2_supported())
962 		return;
963 
964 	amd_iommu_unregister_ppr_notifier(&ppr_nb);
965 
966 	flush_workqueue(iommu_wq);
967 
968 	/*
969 	 * The loop below might call flush_workqueue(), so call
970 	 * destroy_workqueue() after it
971 	 */
972 	for (i = 0; i < MAX_DEVICES; ++i) {
973 		dev_state = get_device_state(i);
974 
975 		if (dev_state == NULL)
976 			continue;
977 
978 		WARN_ON_ONCE(1);
979 
980 		put_device_state(dev_state);
981 		amd_iommu_free_device(dev_state->pdev);
982 	}
983 
984 	destroy_workqueue(iommu_wq);
985 }
986 
987 module_init(amd_iommu_v2_init);
988 module_exit(amd_iommu_v2_exit);
989