xref: /openbmc/linux/drivers/iommu/amd/iommu_v2.c (revision e7bae9bb)
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/mmu_notifier.h>
10 #include <linux/amd-iommu.h>
11 #include <linux/mm_types.h>
12 #include <linux/profile.h>
13 #include <linux/module.h>
14 #include <linux/sched.h>
15 #include <linux/sched/mm.h>
16 #include <linux/wait.h>
17 #include <linux/pci.h>
18 #include <linux/gfp.h>
19 
20 #include "amd_iommu.h"
21 
22 MODULE_LICENSE("GPL v2");
23 MODULE_AUTHOR("Joerg Roedel <jroedel@suse.de>");
24 
25 #define MAX_DEVICES		0x10000
26 #define PRI_QUEUE_SIZE		512
27 
28 struct pri_queue {
29 	atomic_t inflight;
30 	bool finish;
31 	int status;
32 };
33 
34 struct pasid_state {
35 	struct list_head list;			/* For global state-list */
36 	atomic_t count;				/* Reference count */
37 	unsigned mmu_notifier_count;		/* Counting nested mmu_notifier
38 						   calls */
39 	struct mm_struct *mm;			/* mm_struct for the faults */
40 	struct mmu_notifier mn;                 /* mmu_notifier handle */
41 	struct pri_queue pri[PRI_QUEUE_SIZE];	/* PRI tag states */
42 	struct device_state *device_state;	/* Link to our device_state */
43 	int pasid;				/* PASID index */
44 	bool invalid;				/* Used during setup and
45 						   teardown of the pasid */
46 	spinlock_t lock;			/* Protect pri_queues and
47 						   mmu_notifer_count */
48 	wait_queue_head_t wq;			/* To wait for count == 0 */
49 };
50 
51 struct device_state {
52 	struct list_head list;
53 	u16 devid;
54 	atomic_t count;
55 	struct pci_dev *pdev;
56 	struct pasid_state **states;
57 	struct iommu_domain *domain;
58 	int pasid_levels;
59 	int max_pasids;
60 	amd_iommu_invalid_ppr_cb inv_ppr_cb;
61 	amd_iommu_invalidate_ctx inv_ctx_cb;
62 	spinlock_t lock;
63 	wait_queue_head_t wq;
64 };
65 
66 struct fault {
67 	struct work_struct work;
68 	struct device_state *dev_state;
69 	struct pasid_state *state;
70 	struct mm_struct *mm;
71 	u64 address;
72 	u16 devid;
73 	u16 pasid;
74 	u16 tag;
75 	u16 finish;
76 	u16 flags;
77 };
78 
79 static LIST_HEAD(state_list);
80 static spinlock_t state_lock;
81 
82 static struct workqueue_struct *iommu_wq;
83 
84 static void free_pasid_states(struct device_state *dev_state);
85 
86 static u16 device_id(struct pci_dev *pdev)
87 {
88 	u16 devid;
89 
90 	devid = pdev->bus->number;
91 	devid = (devid << 8) | pdev->devfn;
92 
93 	return devid;
94 }
95 
96 static struct device_state *__get_device_state(u16 devid)
97 {
98 	struct device_state *dev_state;
99 
100 	list_for_each_entry(dev_state, &state_list, list) {
101 		if (dev_state->devid == devid)
102 			return dev_state;
103 	}
104 
105 	return NULL;
106 }
107 
108 static struct device_state *get_device_state(u16 devid)
109 {
110 	struct device_state *dev_state;
111 	unsigned long flags;
112 
113 	spin_lock_irqsave(&state_lock, flags);
114 	dev_state = __get_device_state(devid);
115 	if (dev_state != NULL)
116 		atomic_inc(&dev_state->count);
117 	spin_unlock_irqrestore(&state_lock, flags);
118 
119 	return dev_state;
120 }
121 
122 static void free_device_state(struct device_state *dev_state)
123 {
124 	struct iommu_group *group;
125 
126 	/*
127 	 * First detach device from domain - No more PRI requests will arrive
128 	 * from that device after it is unbound from the IOMMUv2 domain.
129 	 */
130 	group = iommu_group_get(&dev_state->pdev->dev);
131 	if (WARN_ON(!group))
132 		return;
133 
134 	iommu_detach_group(dev_state->domain, group);
135 
136 	iommu_group_put(group);
137 
138 	/* Everything is down now, free the IOMMUv2 domain */
139 	iommu_domain_free(dev_state->domain);
140 
141 	/* Finally get rid of the device-state */
142 	kfree(dev_state);
143 }
144 
145 static void put_device_state(struct device_state *dev_state)
146 {
147 	if (atomic_dec_and_test(&dev_state->count))
148 		wake_up(&dev_state->wq);
149 }
150 
151 /* Must be called under dev_state->lock */
152 static struct pasid_state **__get_pasid_state_ptr(struct device_state *dev_state,
153 						  int pasid, bool alloc)
154 {
155 	struct pasid_state **root, **ptr;
156 	int level, index;
157 
158 	level = dev_state->pasid_levels;
159 	root  = dev_state->states;
160 
161 	while (true) {
162 
163 		index = (pasid >> (9 * level)) & 0x1ff;
164 		ptr   = &root[index];
165 
166 		if (level == 0)
167 			break;
168 
169 		if (*ptr == NULL) {
170 			if (!alloc)
171 				return NULL;
172 
173 			*ptr = (void *)get_zeroed_page(GFP_ATOMIC);
174 			if (*ptr == NULL)
175 				return NULL;
176 		}
177 
178 		root   = (struct pasid_state **)*ptr;
179 		level -= 1;
180 	}
181 
182 	return ptr;
183 }
184 
185 static int set_pasid_state(struct device_state *dev_state,
186 			   struct pasid_state *pasid_state,
187 			   int pasid)
188 {
189 	struct pasid_state **ptr;
190 	unsigned long flags;
191 	int ret;
192 
193 	spin_lock_irqsave(&dev_state->lock, flags);
194 	ptr = __get_pasid_state_ptr(dev_state, pasid, true);
195 
196 	ret = -ENOMEM;
197 	if (ptr == NULL)
198 		goto out_unlock;
199 
200 	ret = -ENOMEM;
201 	if (*ptr != NULL)
202 		goto out_unlock;
203 
204 	*ptr = pasid_state;
205 
206 	ret = 0;
207 
208 out_unlock:
209 	spin_unlock_irqrestore(&dev_state->lock, flags);
210 
211 	return ret;
212 }
213 
214 static void clear_pasid_state(struct device_state *dev_state, int pasid)
215 {
216 	struct pasid_state **ptr;
217 	unsigned long flags;
218 
219 	spin_lock_irqsave(&dev_state->lock, flags);
220 	ptr = __get_pasid_state_ptr(dev_state, pasid, true);
221 
222 	if (ptr == NULL)
223 		goto out_unlock;
224 
225 	*ptr = NULL;
226 
227 out_unlock:
228 	spin_unlock_irqrestore(&dev_state->lock, flags);
229 }
230 
231 static struct pasid_state *get_pasid_state(struct device_state *dev_state,
232 					   int pasid)
233 {
234 	struct pasid_state **ptr, *ret = NULL;
235 	unsigned long flags;
236 
237 	spin_lock_irqsave(&dev_state->lock, flags);
238 	ptr = __get_pasid_state_ptr(dev_state, pasid, false);
239 
240 	if (ptr == NULL)
241 		goto out_unlock;
242 
243 	ret = *ptr;
244 	if (ret)
245 		atomic_inc(&ret->count);
246 
247 out_unlock:
248 	spin_unlock_irqrestore(&dev_state->lock, flags);
249 
250 	return ret;
251 }
252 
253 static void free_pasid_state(struct pasid_state *pasid_state)
254 {
255 	kfree(pasid_state);
256 }
257 
258 static void put_pasid_state(struct pasid_state *pasid_state)
259 {
260 	if (atomic_dec_and_test(&pasid_state->count))
261 		wake_up(&pasid_state->wq);
262 }
263 
264 static void put_pasid_state_wait(struct pasid_state *pasid_state)
265 {
266 	atomic_dec(&pasid_state->count);
267 	wait_event(pasid_state->wq, !atomic_read(&pasid_state->count));
268 	free_pasid_state(pasid_state);
269 }
270 
271 static void unbind_pasid(struct pasid_state *pasid_state)
272 {
273 	struct iommu_domain *domain;
274 
275 	domain = pasid_state->device_state->domain;
276 
277 	/*
278 	 * Mark pasid_state as invalid, no more faults will we added to the
279 	 * work queue after this is visible everywhere.
280 	 */
281 	pasid_state->invalid = true;
282 
283 	/* Make sure this is visible */
284 	smp_wmb();
285 
286 	/* After this the device/pasid can't access the mm anymore */
287 	amd_iommu_domain_clear_gcr3(domain, pasid_state->pasid);
288 
289 	/* Make sure no more pending faults are in the queue */
290 	flush_workqueue(iommu_wq);
291 }
292 
293 static void free_pasid_states_level1(struct pasid_state **tbl)
294 {
295 	int i;
296 
297 	for (i = 0; i < 512; ++i) {
298 		if (tbl[i] == NULL)
299 			continue;
300 
301 		free_page((unsigned long)tbl[i]);
302 	}
303 }
304 
305 static void free_pasid_states_level2(struct pasid_state **tbl)
306 {
307 	struct pasid_state **ptr;
308 	int i;
309 
310 	for (i = 0; i < 512; ++i) {
311 		if (tbl[i] == NULL)
312 			continue;
313 
314 		ptr = (struct pasid_state **)tbl[i];
315 		free_pasid_states_level1(ptr);
316 	}
317 }
318 
319 static void free_pasid_states(struct device_state *dev_state)
320 {
321 	struct pasid_state *pasid_state;
322 	int i;
323 
324 	for (i = 0; i < dev_state->max_pasids; ++i) {
325 		pasid_state = get_pasid_state(dev_state, i);
326 		if (pasid_state == NULL)
327 			continue;
328 
329 		put_pasid_state(pasid_state);
330 
331 		/*
332 		 * This will call the mn_release function and
333 		 * unbind the PASID
334 		 */
335 		mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
336 
337 		put_pasid_state_wait(pasid_state); /* Reference taken in
338 						      amd_iommu_bind_pasid */
339 
340 		/* Drop reference taken in amd_iommu_bind_pasid */
341 		put_device_state(dev_state);
342 	}
343 
344 	if (dev_state->pasid_levels == 2)
345 		free_pasid_states_level2(dev_state->states);
346 	else if (dev_state->pasid_levels == 1)
347 		free_pasid_states_level1(dev_state->states);
348 	else
349 		BUG_ON(dev_state->pasid_levels != 0);
350 
351 	free_page((unsigned long)dev_state->states);
352 }
353 
354 static struct pasid_state *mn_to_state(struct mmu_notifier *mn)
355 {
356 	return container_of(mn, struct pasid_state, mn);
357 }
358 
359 static void mn_invalidate_range(struct mmu_notifier *mn,
360 				struct mm_struct *mm,
361 				unsigned long start, unsigned long end)
362 {
363 	struct pasid_state *pasid_state;
364 	struct device_state *dev_state;
365 
366 	pasid_state = mn_to_state(mn);
367 	dev_state   = pasid_state->device_state;
368 
369 	if ((start ^ (end - 1)) < PAGE_SIZE)
370 		amd_iommu_flush_page(dev_state->domain, pasid_state->pasid,
371 				     start);
372 	else
373 		amd_iommu_flush_tlb(dev_state->domain, pasid_state->pasid);
374 }
375 
376 static void mn_release(struct mmu_notifier *mn, struct mm_struct *mm)
377 {
378 	struct pasid_state *pasid_state;
379 	struct device_state *dev_state;
380 	bool run_inv_ctx_cb;
381 
382 	might_sleep();
383 
384 	pasid_state    = mn_to_state(mn);
385 	dev_state      = pasid_state->device_state;
386 	run_inv_ctx_cb = !pasid_state->invalid;
387 
388 	if (run_inv_ctx_cb && dev_state->inv_ctx_cb)
389 		dev_state->inv_ctx_cb(dev_state->pdev, pasid_state->pasid);
390 
391 	unbind_pasid(pasid_state);
392 }
393 
394 static const struct mmu_notifier_ops iommu_mn = {
395 	.release		= mn_release,
396 	.invalidate_range       = mn_invalidate_range,
397 };
398 
399 static void set_pri_tag_status(struct pasid_state *pasid_state,
400 			       u16 tag, int status)
401 {
402 	unsigned long flags;
403 
404 	spin_lock_irqsave(&pasid_state->lock, flags);
405 	pasid_state->pri[tag].status = status;
406 	spin_unlock_irqrestore(&pasid_state->lock, flags);
407 }
408 
409 static void finish_pri_tag(struct device_state *dev_state,
410 			   struct pasid_state *pasid_state,
411 			   u16 tag)
412 {
413 	unsigned long flags;
414 
415 	spin_lock_irqsave(&pasid_state->lock, flags);
416 	if (atomic_dec_and_test(&pasid_state->pri[tag].inflight) &&
417 	    pasid_state->pri[tag].finish) {
418 		amd_iommu_complete_ppr(dev_state->pdev, pasid_state->pasid,
419 				       pasid_state->pri[tag].status, tag);
420 		pasid_state->pri[tag].finish = false;
421 		pasid_state->pri[tag].status = PPR_SUCCESS;
422 	}
423 	spin_unlock_irqrestore(&pasid_state->lock, flags);
424 }
425 
426 static void handle_fault_error(struct fault *fault)
427 {
428 	int status;
429 
430 	if (!fault->dev_state->inv_ppr_cb) {
431 		set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
432 		return;
433 	}
434 
435 	status = fault->dev_state->inv_ppr_cb(fault->dev_state->pdev,
436 					      fault->pasid,
437 					      fault->address,
438 					      fault->flags);
439 	switch (status) {
440 	case AMD_IOMMU_INV_PRI_RSP_SUCCESS:
441 		set_pri_tag_status(fault->state, fault->tag, PPR_SUCCESS);
442 		break;
443 	case AMD_IOMMU_INV_PRI_RSP_INVALID:
444 		set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
445 		break;
446 	case AMD_IOMMU_INV_PRI_RSP_FAIL:
447 		set_pri_tag_status(fault->state, fault->tag, PPR_FAILURE);
448 		break;
449 	default:
450 		BUG();
451 	}
452 }
453 
454 static bool access_error(struct vm_area_struct *vma, struct fault *fault)
455 {
456 	unsigned long requested = 0;
457 
458 	if (fault->flags & PPR_FAULT_EXEC)
459 		requested |= VM_EXEC;
460 
461 	if (fault->flags & PPR_FAULT_READ)
462 		requested |= VM_READ;
463 
464 	if (fault->flags & PPR_FAULT_WRITE)
465 		requested |= VM_WRITE;
466 
467 	return (requested & ~vma->vm_flags) != 0;
468 }
469 
470 static void do_fault(struct work_struct *work)
471 {
472 	struct fault *fault = container_of(work, struct fault, work);
473 	struct vm_area_struct *vma;
474 	vm_fault_t ret = VM_FAULT_ERROR;
475 	unsigned int flags = 0;
476 	struct mm_struct *mm;
477 	u64 address;
478 
479 	mm = fault->state->mm;
480 	address = fault->address;
481 
482 	if (fault->flags & PPR_FAULT_USER)
483 		flags |= FAULT_FLAG_USER;
484 	if (fault->flags & PPR_FAULT_WRITE)
485 		flags |= FAULT_FLAG_WRITE;
486 	flags |= FAULT_FLAG_REMOTE;
487 
488 	mmap_read_lock(mm);
489 	vma = find_extend_vma(mm, address);
490 	if (!vma || address < vma->vm_start)
491 		/* failed to get a vma in the right range */
492 		goto out;
493 
494 	/* Check if we have the right permissions on the vma */
495 	if (access_error(vma, fault))
496 		goto out;
497 
498 	ret = handle_mm_fault(vma, address, flags, NULL);
499 out:
500 	mmap_read_unlock(mm);
501 
502 	if (ret & VM_FAULT_ERROR)
503 		/* failed to service fault */
504 		handle_fault_error(fault);
505 
506 	finish_pri_tag(fault->dev_state, fault->state, fault->tag);
507 
508 	put_pasid_state(fault->state);
509 
510 	kfree(fault);
511 }
512 
513 static int ppr_notifier(struct notifier_block *nb, unsigned long e, void *data)
514 {
515 	struct amd_iommu_fault *iommu_fault;
516 	struct pasid_state *pasid_state;
517 	struct device_state *dev_state;
518 	struct pci_dev *pdev = NULL;
519 	unsigned long flags;
520 	struct fault *fault;
521 	bool finish;
522 	u16 tag, devid;
523 	int ret;
524 
525 	iommu_fault = data;
526 	tag         = iommu_fault->tag & 0x1ff;
527 	finish      = (iommu_fault->tag >> 9) & 1;
528 
529 	devid = iommu_fault->device_id;
530 	pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
531 					   devid & 0xff);
532 	if (!pdev)
533 		return -ENODEV;
534 
535 	ret = NOTIFY_DONE;
536 
537 	/* In kdump kernel pci dev is not initialized yet -> send INVALID */
538 	if (amd_iommu_is_attach_deferred(NULL, &pdev->dev)) {
539 		amd_iommu_complete_ppr(pdev, iommu_fault->pasid,
540 				       PPR_INVALID, tag);
541 		goto out;
542 	}
543 
544 	dev_state = get_device_state(iommu_fault->device_id);
545 	if (dev_state == NULL)
546 		goto out;
547 
548 	pasid_state = get_pasid_state(dev_state, iommu_fault->pasid);
549 	if (pasid_state == NULL || pasid_state->invalid) {
550 		/* We know the device but not the PASID -> send INVALID */
551 		amd_iommu_complete_ppr(dev_state->pdev, iommu_fault->pasid,
552 				       PPR_INVALID, tag);
553 		goto out_drop_state;
554 	}
555 
556 	spin_lock_irqsave(&pasid_state->lock, flags);
557 	atomic_inc(&pasid_state->pri[tag].inflight);
558 	if (finish)
559 		pasid_state->pri[tag].finish = true;
560 	spin_unlock_irqrestore(&pasid_state->lock, flags);
561 
562 	fault = kzalloc(sizeof(*fault), GFP_ATOMIC);
563 	if (fault == NULL) {
564 		/* We are OOM - send success and let the device re-fault */
565 		finish_pri_tag(dev_state, pasid_state, tag);
566 		goto out_drop_state;
567 	}
568 
569 	fault->dev_state = dev_state;
570 	fault->address   = iommu_fault->address;
571 	fault->state     = pasid_state;
572 	fault->tag       = tag;
573 	fault->finish    = finish;
574 	fault->pasid     = iommu_fault->pasid;
575 	fault->flags     = iommu_fault->flags;
576 	INIT_WORK(&fault->work, do_fault);
577 
578 	queue_work(iommu_wq, &fault->work);
579 
580 	ret = NOTIFY_OK;
581 
582 out_drop_state:
583 
584 	if (ret != NOTIFY_OK && pasid_state)
585 		put_pasid_state(pasid_state);
586 
587 	put_device_state(dev_state);
588 
589 out:
590 	return ret;
591 }
592 
593 static struct notifier_block ppr_nb = {
594 	.notifier_call = ppr_notifier,
595 };
596 
597 int amd_iommu_bind_pasid(struct pci_dev *pdev, int pasid,
598 			 struct task_struct *task)
599 {
600 	struct pasid_state *pasid_state;
601 	struct device_state *dev_state;
602 	struct mm_struct *mm;
603 	u16 devid;
604 	int ret;
605 
606 	might_sleep();
607 
608 	if (!amd_iommu_v2_supported())
609 		return -ENODEV;
610 
611 	devid     = device_id(pdev);
612 	dev_state = get_device_state(devid);
613 
614 	if (dev_state == NULL)
615 		return -EINVAL;
616 
617 	ret = -EINVAL;
618 	if (pasid < 0 || pasid >= dev_state->max_pasids)
619 		goto out;
620 
621 	ret = -ENOMEM;
622 	pasid_state = kzalloc(sizeof(*pasid_state), GFP_KERNEL);
623 	if (pasid_state == NULL)
624 		goto out;
625 
626 
627 	atomic_set(&pasid_state->count, 1);
628 	init_waitqueue_head(&pasid_state->wq);
629 	spin_lock_init(&pasid_state->lock);
630 
631 	mm                        = get_task_mm(task);
632 	pasid_state->mm           = mm;
633 	pasid_state->device_state = dev_state;
634 	pasid_state->pasid        = pasid;
635 	pasid_state->invalid      = true; /* Mark as valid only if we are
636 					     done with setting up the pasid */
637 	pasid_state->mn.ops       = &iommu_mn;
638 
639 	if (pasid_state->mm == NULL)
640 		goto out_free;
641 
642 	mmu_notifier_register(&pasid_state->mn, mm);
643 
644 	ret = set_pasid_state(dev_state, pasid_state, pasid);
645 	if (ret)
646 		goto out_unregister;
647 
648 	ret = amd_iommu_domain_set_gcr3(dev_state->domain, pasid,
649 					__pa(pasid_state->mm->pgd));
650 	if (ret)
651 		goto out_clear_state;
652 
653 	/* Now we are ready to handle faults */
654 	pasid_state->invalid = false;
655 
656 	/*
657 	 * Drop the reference to the mm_struct here. We rely on the
658 	 * mmu_notifier release call-back to inform us when the mm
659 	 * is going away.
660 	 */
661 	mmput(mm);
662 
663 	return 0;
664 
665 out_clear_state:
666 	clear_pasid_state(dev_state, pasid);
667 
668 out_unregister:
669 	mmu_notifier_unregister(&pasid_state->mn, mm);
670 	mmput(mm);
671 
672 out_free:
673 	free_pasid_state(pasid_state);
674 
675 out:
676 	put_device_state(dev_state);
677 
678 	return ret;
679 }
680 EXPORT_SYMBOL(amd_iommu_bind_pasid);
681 
682 void amd_iommu_unbind_pasid(struct pci_dev *pdev, int pasid)
683 {
684 	struct pasid_state *pasid_state;
685 	struct device_state *dev_state;
686 	u16 devid;
687 
688 	might_sleep();
689 
690 	if (!amd_iommu_v2_supported())
691 		return;
692 
693 	devid = device_id(pdev);
694 	dev_state = get_device_state(devid);
695 	if (dev_state == NULL)
696 		return;
697 
698 	if (pasid < 0 || pasid >= dev_state->max_pasids)
699 		goto out;
700 
701 	pasid_state = get_pasid_state(dev_state, pasid);
702 	if (pasid_state == NULL)
703 		goto out;
704 	/*
705 	 * Drop reference taken here. We are safe because we still hold
706 	 * the reference taken in the amd_iommu_bind_pasid function.
707 	 */
708 	put_pasid_state(pasid_state);
709 
710 	/* Clear the pasid state so that the pasid can be re-used */
711 	clear_pasid_state(dev_state, pasid_state->pasid);
712 
713 	/*
714 	 * Call mmu_notifier_unregister to drop our reference
715 	 * to pasid_state->mm
716 	 */
717 	mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
718 
719 	put_pasid_state_wait(pasid_state); /* Reference taken in
720 					      amd_iommu_bind_pasid */
721 out:
722 	/* Drop reference taken in this function */
723 	put_device_state(dev_state);
724 
725 	/* Drop reference taken in amd_iommu_bind_pasid */
726 	put_device_state(dev_state);
727 }
728 EXPORT_SYMBOL(amd_iommu_unbind_pasid);
729 
730 int amd_iommu_init_device(struct pci_dev *pdev, int pasids)
731 {
732 	struct device_state *dev_state;
733 	struct iommu_group *group;
734 	unsigned long flags;
735 	int ret, tmp;
736 	u16 devid;
737 
738 	might_sleep();
739 
740 	/*
741 	 * When memory encryption is active the device is likely not in a
742 	 * direct-mapped domain. Forbid using IOMMUv2 functionality for now.
743 	 */
744 	if (mem_encrypt_active())
745 		return -ENODEV;
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 	group = iommu_group_get(&pdev->dev);
787 	if (!group) {
788 		ret = -EINVAL;
789 		goto out_free_domain;
790 	}
791 
792 	ret = iommu_attach_group(dev_state->domain, group);
793 	if (ret != 0)
794 		goto out_drop_group;
795 
796 	iommu_group_put(group);
797 
798 	spin_lock_irqsave(&state_lock, flags);
799 
800 	if (__get_device_state(devid) != NULL) {
801 		spin_unlock_irqrestore(&state_lock, flags);
802 		ret = -EBUSY;
803 		goto out_free_domain;
804 	}
805 
806 	list_add_tail(&dev_state->list, &state_list);
807 
808 	spin_unlock_irqrestore(&state_lock, flags);
809 
810 	return 0;
811 
812 out_drop_group:
813 	iommu_group_put(group);
814 
815 out_free_domain:
816 	iommu_domain_free(dev_state->domain);
817 
818 out_free_states:
819 	free_page((unsigned long)dev_state->states);
820 
821 out_free_dev_state:
822 	kfree(dev_state);
823 
824 	return ret;
825 }
826 EXPORT_SYMBOL(amd_iommu_init_device);
827 
828 void amd_iommu_free_device(struct pci_dev *pdev)
829 {
830 	struct device_state *dev_state;
831 	unsigned long flags;
832 	u16 devid;
833 
834 	if (!amd_iommu_v2_supported())
835 		return;
836 
837 	devid = device_id(pdev);
838 
839 	spin_lock_irqsave(&state_lock, flags);
840 
841 	dev_state = __get_device_state(devid);
842 	if (dev_state == NULL) {
843 		spin_unlock_irqrestore(&state_lock, flags);
844 		return;
845 	}
846 
847 	list_del(&dev_state->list);
848 
849 	spin_unlock_irqrestore(&state_lock, flags);
850 
851 	/* Get rid of any remaining pasid states */
852 	free_pasid_states(dev_state);
853 
854 	put_device_state(dev_state);
855 	/*
856 	 * Wait until the last reference is dropped before freeing
857 	 * the device state.
858 	 */
859 	wait_event(dev_state->wq, !atomic_read(&dev_state->count));
860 	free_device_state(dev_state);
861 }
862 EXPORT_SYMBOL(amd_iommu_free_device);
863 
864 int amd_iommu_set_invalid_ppr_cb(struct pci_dev *pdev,
865 				 amd_iommu_invalid_ppr_cb cb)
866 {
867 	struct device_state *dev_state;
868 	unsigned long flags;
869 	u16 devid;
870 	int ret;
871 
872 	if (!amd_iommu_v2_supported())
873 		return -ENODEV;
874 
875 	devid = device_id(pdev);
876 
877 	spin_lock_irqsave(&state_lock, flags);
878 
879 	ret = -EINVAL;
880 	dev_state = __get_device_state(devid);
881 	if (dev_state == NULL)
882 		goto out_unlock;
883 
884 	dev_state->inv_ppr_cb = cb;
885 
886 	ret = 0;
887 
888 out_unlock:
889 	spin_unlock_irqrestore(&state_lock, flags);
890 
891 	return ret;
892 }
893 EXPORT_SYMBOL(amd_iommu_set_invalid_ppr_cb);
894 
895 int amd_iommu_set_invalidate_ctx_cb(struct pci_dev *pdev,
896 				    amd_iommu_invalidate_ctx cb)
897 {
898 	struct device_state *dev_state;
899 	unsigned long flags;
900 	u16 devid;
901 	int ret;
902 
903 	if (!amd_iommu_v2_supported())
904 		return -ENODEV;
905 
906 	devid = device_id(pdev);
907 
908 	spin_lock_irqsave(&state_lock, flags);
909 
910 	ret = -EINVAL;
911 	dev_state = __get_device_state(devid);
912 	if (dev_state == NULL)
913 		goto out_unlock;
914 
915 	dev_state->inv_ctx_cb = cb;
916 
917 	ret = 0;
918 
919 out_unlock:
920 	spin_unlock_irqrestore(&state_lock, flags);
921 
922 	return ret;
923 }
924 EXPORT_SYMBOL(amd_iommu_set_invalidate_ctx_cb);
925 
926 static int __init amd_iommu_v2_init(void)
927 {
928 	int ret;
929 
930 	pr_info("AMD IOMMUv2 driver by Joerg Roedel <jroedel@suse.de>\n");
931 
932 	if (!amd_iommu_v2_supported()) {
933 		pr_info("AMD IOMMUv2 functionality not available on this system\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 	spin_lock_init(&state_lock);
942 
943 	ret = -ENOMEM;
944 	iommu_wq = alloc_workqueue("amd_iommu_v2", WQ_MEM_RECLAIM, 0);
945 	if (iommu_wq == NULL)
946 		goto out;
947 
948 	amd_iommu_register_ppr_notifier(&ppr_nb);
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