xref: /openbmc/linux/drivers/iommu/amd/iommu_v2.c (revision c699ce1a)
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 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 	u32 sbdf;
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 	u32 pasid;
74 	u16 tag;
75 	u16 finish;
76 	u16 flags;
77 };
78 
79 static LIST_HEAD(state_list);
80 static DEFINE_SPINLOCK(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 struct device_state *__get_device_state(u32 sbdf)
87 {
88 	struct device_state *dev_state;
89 
90 	list_for_each_entry(dev_state, &state_list, list) {
91 		if (dev_state->sbdf == sbdf)
92 			return dev_state;
93 	}
94 
95 	return NULL;
96 }
97 
98 static struct device_state *get_device_state(u32 sbdf)
99 {
100 	struct device_state *dev_state;
101 	unsigned long flags;
102 
103 	spin_lock_irqsave(&state_lock, flags);
104 	dev_state = __get_device_state(sbdf);
105 	if (dev_state != NULL)
106 		atomic_inc(&dev_state->count);
107 	spin_unlock_irqrestore(&state_lock, flags);
108 
109 	return dev_state;
110 }
111 
112 static void free_device_state(struct device_state *dev_state)
113 {
114 	struct iommu_group *group;
115 
116 	/* Get rid of any remaining pasid states */
117 	free_pasid_states(dev_state);
118 
119 	/*
120 	 * Wait until the last reference is dropped before freeing
121 	 * the device state.
122 	 */
123 	wait_event(dev_state->wq, !atomic_read(&dev_state->count));
124 
125 	/*
126 	 * First detach device from domain - No more PRI requests will arrive
127 	 * from that device after it is unbound from the IOMMUv2 domain.
128 	 */
129 	group = iommu_group_get(&dev_state->pdev->dev);
130 	if (WARN_ON(!group))
131 		return;
132 
133 	iommu_detach_group(dev_state->domain, group);
134 
135 	iommu_group_put(group);
136 
137 	/* Everything is down now, free the IOMMUv2 domain */
138 	iommu_domain_free(dev_state->domain);
139 
140 	/* Finally get rid of the device-state */
141 	kfree(dev_state);
142 }
143 
144 static void put_device_state(struct device_state *dev_state)
145 {
146 	if (atomic_dec_and_test(&dev_state->count))
147 		wake_up(&dev_state->wq);
148 }
149 
150 /* Must be called under dev_state->lock */
151 static struct pasid_state **__get_pasid_state_ptr(struct device_state *dev_state,
152 						  u32 pasid, bool alloc)
153 {
154 	struct pasid_state **root, **ptr;
155 	int level, index;
156 
157 	level = dev_state->pasid_levels;
158 	root  = dev_state->states;
159 
160 	while (true) {
161 
162 		index = (pasid >> (9 * level)) & 0x1ff;
163 		ptr   = &root[index];
164 
165 		if (level == 0)
166 			break;
167 
168 		if (*ptr == NULL) {
169 			if (!alloc)
170 				return NULL;
171 
172 			*ptr = (void *)get_zeroed_page(GFP_ATOMIC);
173 			if (*ptr == NULL)
174 				return NULL;
175 		}
176 
177 		root   = (struct pasid_state **)*ptr;
178 		level -= 1;
179 	}
180 
181 	return ptr;
182 }
183 
184 static int set_pasid_state(struct device_state *dev_state,
185 			   struct pasid_state *pasid_state,
186 			   u32 pasid)
187 {
188 	struct pasid_state **ptr;
189 	unsigned long flags;
190 	int ret;
191 
192 	spin_lock_irqsave(&dev_state->lock, flags);
193 	ptr = __get_pasid_state_ptr(dev_state, pasid, true);
194 
195 	ret = -ENOMEM;
196 	if (ptr == NULL)
197 		goto out_unlock;
198 
199 	ret = -ENOMEM;
200 	if (*ptr != NULL)
201 		goto out_unlock;
202 
203 	*ptr = pasid_state;
204 
205 	ret = 0;
206 
207 out_unlock:
208 	spin_unlock_irqrestore(&dev_state->lock, flags);
209 
210 	return ret;
211 }
212 
213 static void clear_pasid_state(struct device_state *dev_state, u32 pasid)
214 {
215 	struct pasid_state **ptr;
216 	unsigned long flags;
217 
218 	spin_lock_irqsave(&dev_state->lock, flags);
219 	ptr = __get_pasid_state_ptr(dev_state, pasid, true);
220 
221 	if (ptr == NULL)
222 		goto out_unlock;
223 
224 	*ptr = NULL;
225 
226 out_unlock:
227 	spin_unlock_irqrestore(&dev_state->lock, flags);
228 }
229 
230 static struct pasid_state *get_pasid_state(struct device_state *dev_state,
231 					   u32 pasid)
232 {
233 	struct pasid_state **ptr, *ret = NULL;
234 	unsigned long flags;
235 
236 	spin_lock_irqsave(&dev_state->lock, flags);
237 	ptr = __get_pasid_state_ptr(dev_state, pasid, false);
238 
239 	if (ptr == NULL)
240 		goto out_unlock;
241 
242 	ret = *ptr;
243 	if (ret)
244 		refcount_inc(&ret->count);
245 
246 out_unlock:
247 	spin_unlock_irqrestore(&dev_state->lock, flags);
248 
249 	return ret;
250 }
251 
252 static void free_pasid_state(struct pasid_state *pasid_state)
253 {
254 	kfree(pasid_state);
255 }
256 
257 static void put_pasid_state(struct pasid_state *pasid_state)
258 {
259 	if (refcount_dec_and_test(&pasid_state->count))
260 		wake_up(&pasid_state->wq);
261 }
262 
263 static void put_pasid_state_wait(struct pasid_state *pasid_state)
264 {
265 	refcount_dec(&pasid_state->count);
266 	wait_event(pasid_state->wq, !refcount_read(&pasid_state->count));
267 	free_pasid_state(pasid_state);
268 }
269 
270 static void unbind_pasid(struct pasid_state *pasid_state)
271 {
272 	struct iommu_domain *domain;
273 
274 	domain = pasid_state->device_state->domain;
275 
276 	/*
277 	 * Mark pasid_state as invalid, no more faults will we added to the
278 	 * work queue after this is visible everywhere.
279 	 */
280 	pasid_state->invalid = true;
281 
282 	/* Make sure this is visible */
283 	smp_wmb();
284 
285 	/* After this the device/pasid can't access the mm anymore */
286 	amd_iommu_domain_clear_gcr3(domain, pasid_state->pasid);
287 
288 	/* Make sure no more pending faults are in the queue */
289 	flush_workqueue(iommu_wq);
290 }
291 
292 static void free_pasid_states_level1(struct pasid_state **tbl)
293 {
294 	int i;
295 
296 	for (i = 0; i < 512; ++i) {
297 		if (tbl[i] == NULL)
298 			continue;
299 
300 		free_page((unsigned long)tbl[i]);
301 	}
302 }
303 
304 static void free_pasid_states_level2(struct pasid_state **tbl)
305 {
306 	struct pasid_state **ptr;
307 	int i;
308 
309 	for (i = 0; i < 512; ++i) {
310 		if (tbl[i] == NULL)
311 			continue;
312 
313 		ptr = (struct pasid_state **)tbl[i];
314 		free_pasid_states_level1(ptr);
315 	}
316 }
317 
318 static void free_pasid_states(struct device_state *dev_state)
319 {
320 	struct pasid_state *pasid_state;
321 	int i;
322 
323 	for (i = 0; i < dev_state->max_pasids; ++i) {
324 		pasid_state = get_pasid_state(dev_state, i);
325 		if (pasid_state == NULL)
326 			continue;
327 
328 		put_pasid_state(pasid_state);
329 
330 		/*
331 		 * This will call the mn_release function and
332 		 * unbind the PASID
333 		 */
334 		mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
335 
336 		put_pasid_state_wait(pasid_state); /* Reference taken in
337 						      amd_iommu_bind_pasid */
338 
339 		/* Drop reference taken in amd_iommu_bind_pasid */
340 		put_device_state(dev_state);
341 	}
342 
343 	if (dev_state->pasid_levels == 2)
344 		free_pasid_states_level2(dev_state->states);
345 	else if (dev_state->pasid_levels == 1)
346 		free_pasid_states_level1(dev_state->states);
347 	else
348 		BUG_ON(dev_state->pasid_levels != 0);
349 
350 	free_page((unsigned long)dev_state->states);
351 }
352 
353 static struct pasid_state *mn_to_state(struct mmu_notifier *mn)
354 {
355 	return container_of(mn, struct pasid_state, mn);
356 }
357 
358 static void mn_invalidate_range(struct mmu_notifier *mn,
359 				struct mm_struct *mm,
360 				unsigned long start, unsigned long end)
361 {
362 	struct pasid_state *pasid_state;
363 	struct device_state *dev_state;
364 
365 	pasid_state = mn_to_state(mn);
366 	dev_state   = pasid_state->device_state;
367 
368 	if ((start ^ (end - 1)) < PAGE_SIZE)
369 		amd_iommu_flush_page(dev_state->domain, pasid_state->pasid,
370 				     start);
371 	else
372 		amd_iommu_flush_tlb(dev_state->domain, pasid_state->pasid);
373 }
374 
375 static void mn_release(struct mmu_notifier *mn, struct mm_struct *mm)
376 {
377 	struct pasid_state *pasid_state;
378 	struct device_state *dev_state;
379 	bool run_inv_ctx_cb;
380 
381 	might_sleep();
382 
383 	pasid_state    = mn_to_state(mn);
384 	dev_state      = pasid_state->device_state;
385 	run_inv_ctx_cb = !pasid_state->invalid;
386 
387 	if (run_inv_ctx_cb && dev_state->inv_ctx_cb)
388 		dev_state->inv_ctx_cb(dev_state->pdev, pasid_state->pasid);
389 
390 	unbind_pasid(pasid_state);
391 }
392 
393 static const struct mmu_notifier_ops iommu_mn = {
394 	.release		= mn_release,
395 	.invalidate_range       = mn_invalidate_range,
396 };
397 
398 static void set_pri_tag_status(struct pasid_state *pasid_state,
399 			       u16 tag, int status)
400 {
401 	unsigned long flags;
402 
403 	spin_lock_irqsave(&pasid_state->lock, flags);
404 	pasid_state->pri[tag].status = status;
405 	spin_unlock_irqrestore(&pasid_state->lock, flags);
406 }
407 
408 static void finish_pri_tag(struct device_state *dev_state,
409 			   struct pasid_state *pasid_state,
410 			   u16 tag)
411 {
412 	unsigned long flags;
413 
414 	spin_lock_irqsave(&pasid_state->lock, flags);
415 	if (atomic_dec_and_test(&pasid_state->pri[tag].inflight) &&
416 	    pasid_state->pri[tag].finish) {
417 		amd_iommu_complete_ppr(dev_state->pdev, pasid_state->pasid,
418 				       pasid_state->pri[tag].status, tag);
419 		pasid_state->pri[tag].finish = false;
420 		pasid_state->pri[tag].status = PPR_SUCCESS;
421 	}
422 	spin_unlock_irqrestore(&pasid_state->lock, flags);
423 }
424 
425 static void handle_fault_error(struct fault *fault)
426 {
427 	int status;
428 
429 	if (!fault->dev_state->inv_ppr_cb) {
430 		set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
431 		return;
432 	}
433 
434 	status = fault->dev_state->inv_ppr_cb(fault->dev_state->pdev,
435 					      fault->pasid,
436 					      fault->address,
437 					      fault->flags);
438 	switch (status) {
439 	case AMD_IOMMU_INV_PRI_RSP_SUCCESS:
440 		set_pri_tag_status(fault->state, fault->tag, PPR_SUCCESS);
441 		break;
442 	case AMD_IOMMU_INV_PRI_RSP_INVALID:
443 		set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
444 		break;
445 	case AMD_IOMMU_INV_PRI_RSP_FAIL:
446 		set_pri_tag_status(fault->state, fault->tag, PPR_FAILURE);
447 		break;
448 	default:
449 		BUG();
450 	}
451 }
452 
453 static bool access_error(struct vm_area_struct *vma, struct fault *fault)
454 {
455 	unsigned long requested = 0;
456 
457 	if (fault->flags & PPR_FAULT_EXEC)
458 		requested |= VM_EXEC;
459 
460 	if (fault->flags & PPR_FAULT_READ)
461 		requested |= VM_READ;
462 
463 	if (fault->flags & PPR_FAULT_WRITE)
464 		requested |= VM_WRITE;
465 
466 	return (requested & ~vma->vm_flags) != 0;
467 }
468 
469 static void do_fault(struct work_struct *work)
470 {
471 	struct fault *fault = container_of(work, struct fault, work);
472 	struct vm_area_struct *vma;
473 	vm_fault_t ret = VM_FAULT_ERROR;
474 	unsigned int flags = 0;
475 	struct mm_struct *mm;
476 	u64 address;
477 
478 	mm = fault->state->mm;
479 	address = fault->address;
480 
481 	if (fault->flags & PPR_FAULT_USER)
482 		flags |= FAULT_FLAG_USER;
483 	if (fault->flags & PPR_FAULT_WRITE)
484 		flags |= FAULT_FLAG_WRITE;
485 	flags |= FAULT_FLAG_REMOTE;
486 
487 	mmap_read_lock(mm);
488 	vma = find_extend_vma(mm, address);
489 	if (!vma || address < vma->vm_start)
490 		/* failed to get a vma in the right range */
491 		goto out;
492 
493 	/* Check if we have the right permissions on the vma */
494 	if (access_error(vma, fault))
495 		goto out;
496 
497 	ret = handle_mm_fault(vma, address, flags, NULL);
498 out:
499 	mmap_read_unlock(mm);
500 
501 	if (ret & VM_FAULT_ERROR)
502 		/* failed to service fault */
503 		handle_fault_error(fault);
504 
505 	finish_pri_tag(fault->dev_state, fault->state, fault->tag);
506 
507 	put_pasid_state(fault->state);
508 
509 	kfree(fault);
510 }
511 
512 static int ppr_notifier(struct notifier_block *nb, unsigned long e, void *data)
513 {
514 	struct amd_iommu_fault *iommu_fault;
515 	struct pasid_state *pasid_state;
516 	struct device_state *dev_state;
517 	struct pci_dev *pdev = NULL;
518 	unsigned long flags;
519 	struct fault *fault;
520 	bool finish;
521 	u16 tag, devid, seg_id;
522 	int ret;
523 
524 	iommu_fault = data;
525 	tag         = iommu_fault->tag & 0x1ff;
526 	finish      = (iommu_fault->tag >> 9) & 1;
527 
528 	seg_id = PCI_SBDF_TO_SEGID(iommu_fault->sbdf);
529 	devid = PCI_SBDF_TO_DEVID(iommu_fault->sbdf);
530 	pdev = pci_get_domain_bus_and_slot(seg_id, 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(&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->sbdf);
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 	pci_dev_put(pdev);
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 	u32 sbdf;
605 	int ret;
606 
607 	might_sleep();
608 
609 	if (!amd_iommu_v2_supported())
610 		return -ENODEV;
611 
612 	sbdf      = get_pci_sbdf_id(pdev);
613 	dev_state = get_device_state(sbdf);
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 	ret = mmu_notifier_register(&pasid_state->mn, mm);
644 	if (ret)
645 		goto out_free;
646 
647 	ret = set_pasid_state(dev_state, pasid_state, pasid);
648 	if (ret)
649 		goto out_unregister;
650 
651 	ret = amd_iommu_domain_set_gcr3(dev_state->domain, pasid,
652 					__pa(pasid_state->mm->pgd));
653 	if (ret)
654 		goto out_clear_state;
655 
656 	/* Now we are ready to handle faults */
657 	pasid_state->invalid = false;
658 
659 	/*
660 	 * Drop the reference to the mm_struct here. We rely on the
661 	 * mmu_notifier release call-back to inform us when the mm
662 	 * is going away.
663 	 */
664 	mmput(mm);
665 
666 	return 0;
667 
668 out_clear_state:
669 	clear_pasid_state(dev_state, pasid);
670 
671 out_unregister:
672 	mmu_notifier_unregister(&pasid_state->mn, mm);
673 	mmput(mm);
674 
675 out_free:
676 	free_pasid_state(pasid_state);
677 
678 out:
679 	put_device_state(dev_state);
680 
681 	return ret;
682 }
683 EXPORT_SYMBOL(amd_iommu_bind_pasid);
684 
685 void amd_iommu_unbind_pasid(struct pci_dev *pdev, u32 pasid)
686 {
687 	struct pasid_state *pasid_state;
688 	struct device_state *dev_state;
689 	u32 sbdf;
690 
691 	might_sleep();
692 
693 	if (!amd_iommu_v2_supported())
694 		return;
695 
696 	sbdf = get_pci_sbdf_id(pdev);
697 	dev_state = get_device_state(sbdf);
698 	if (dev_state == NULL)
699 		return;
700 
701 	if (pasid >= dev_state->max_pasids)
702 		goto out;
703 
704 	pasid_state = get_pasid_state(dev_state, pasid);
705 	if (pasid_state == NULL)
706 		goto out;
707 	/*
708 	 * Drop reference taken here. We are safe because we still hold
709 	 * the reference taken in the amd_iommu_bind_pasid function.
710 	 */
711 	put_pasid_state(pasid_state);
712 
713 	/* Clear the pasid state so that the pasid can be re-used */
714 	clear_pasid_state(dev_state, pasid_state->pasid);
715 
716 	/*
717 	 * Call mmu_notifier_unregister to drop our reference
718 	 * to pasid_state->mm
719 	 */
720 	mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
721 
722 	put_pasid_state_wait(pasid_state); /* Reference taken in
723 					      amd_iommu_bind_pasid */
724 out:
725 	/* Drop reference taken in this function */
726 	put_device_state(dev_state);
727 
728 	/* Drop reference taken in amd_iommu_bind_pasid */
729 	put_device_state(dev_state);
730 }
731 EXPORT_SYMBOL(amd_iommu_unbind_pasid);
732 
733 int amd_iommu_init_device(struct pci_dev *pdev, int pasids)
734 {
735 	struct device_state *dev_state;
736 	struct iommu_group *group;
737 	unsigned long flags;
738 	int ret, tmp;
739 	u32 sbdf;
740 
741 	might_sleep();
742 
743 	/*
744 	 * When memory encryption is active the device is likely not in a
745 	 * direct-mapped domain. Forbid using IOMMUv2 functionality for now.
746 	 */
747 	if (cc_platform_has(CC_ATTR_MEM_ENCRYPT))
748 		return -ENODEV;
749 
750 	if (!amd_iommu_v2_supported())
751 		return -ENODEV;
752 
753 	if (pasids <= 0 || pasids > (PASID_MASK + 1))
754 		return -EINVAL;
755 
756 	sbdf = get_pci_sbdf_id(pdev);
757 
758 	dev_state = kzalloc(sizeof(*dev_state), GFP_KERNEL);
759 	if (dev_state == NULL)
760 		return -ENOMEM;
761 
762 	spin_lock_init(&dev_state->lock);
763 	init_waitqueue_head(&dev_state->wq);
764 	dev_state->pdev  = pdev;
765 	dev_state->sbdf = sbdf;
766 
767 	tmp = pasids;
768 	for (dev_state->pasid_levels = 0; (tmp - 1) & ~0x1ff; tmp >>= 9)
769 		dev_state->pasid_levels += 1;
770 
771 	atomic_set(&dev_state->count, 1);
772 	dev_state->max_pasids = pasids;
773 
774 	ret = -ENOMEM;
775 	dev_state->states = (void *)get_zeroed_page(GFP_KERNEL);
776 	if (dev_state->states == NULL)
777 		goto out_free_dev_state;
778 
779 	dev_state->domain = iommu_domain_alloc(&pci_bus_type);
780 	if (dev_state->domain == NULL)
781 		goto out_free_states;
782 
783 	/* See iommu_is_default_domain() */
784 	dev_state->domain->type = IOMMU_DOMAIN_IDENTITY;
785 	amd_iommu_domain_direct_map(dev_state->domain);
786 
787 	ret = amd_iommu_domain_enable_v2(dev_state->domain, pasids);
788 	if (ret)
789 		goto out_free_domain;
790 
791 	group = iommu_group_get(&pdev->dev);
792 	if (!group) {
793 		ret = -EINVAL;
794 		goto out_free_domain;
795 	}
796 
797 	ret = iommu_attach_group(dev_state->domain, group);
798 	if (ret != 0)
799 		goto out_drop_group;
800 
801 	iommu_group_put(group);
802 
803 	spin_lock_irqsave(&state_lock, flags);
804 
805 	if (__get_device_state(sbdf) != NULL) {
806 		spin_unlock_irqrestore(&state_lock, flags);
807 		ret = -EBUSY;
808 		goto out_free_domain;
809 	}
810 
811 	list_add_tail(&dev_state->list, &state_list);
812 
813 	spin_unlock_irqrestore(&state_lock, flags);
814 
815 	return 0;
816 
817 out_drop_group:
818 	iommu_group_put(group);
819 
820 out_free_domain:
821 	iommu_domain_free(dev_state->domain);
822 
823 out_free_states:
824 	free_page((unsigned long)dev_state->states);
825 
826 out_free_dev_state:
827 	kfree(dev_state);
828 
829 	return ret;
830 }
831 EXPORT_SYMBOL(amd_iommu_init_device);
832 
833 void amd_iommu_free_device(struct pci_dev *pdev)
834 {
835 	struct device_state *dev_state;
836 	unsigned long flags;
837 	u32 sbdf;
838 
839 	if (!amd_iommu_v2_supported())
840 		return;
841 
842 	sbdf = get_pci_sbdf_id(pdev);
843 
844 	spin_lock_irqsave(&state_lock, flags);
845 
846 	dev_state = __get_device_state(sbdf);
847 	if (dev_state == NULL) {
848 		spin_unlock_irqrestore(&state_lock, flags);
849 		return;
850 	}
851 
852 	list_del(&dev_state->list);
853 
854 	spin_unlock_irqrestore(&state_lock, flags);
855 
856 	put_device_state(dev_state);
857 	free_device_state(dev_state);
858 }
859 EXPORT_SYMBOL(amd_iommu_free_device);
860 
861 int amd_iommu_set_invalid_ppr_cb(struct pci_dev *pdev,
862 				 amd_iommu_invalid_ppr_cb cb)
863 {
864 	struct device_state *dev_state;
865 	unsigned long flags;
866 	u32 sbdf;
867 	int ret;
868 
869 	if (!amd_iommu_v2_supported())
870 		return -ENODEV;
871 
872 	sbdf = get_pci_sbdf_id(pdev);
873 
874 	spin_lock_irqsave(&state_lock, flags);
875 
876 	ret = -EINVAL;
877 	dev_state = __get_device_state(sbdf);
878 	if (dev_state == NULL)
879 		goto out_unlock;
880 
881 	dev_state->inv_ppr_cb = cb;
882 
883 	ret = 0;
884 
885 out_unlock:
886 	spin_unlock_irqrestore(&state_lock, flags);
887 
888 	return ret;
889 }
890 EXPORT_SYMBOL(amd_iommu_set_invalid_ppr_cb);
891 
892 int amd_iommu_set_invalidate_ctx_cb(struct pci_dev *pdev,
893 				    amd_iommu_invalidate_ctx cb)
894 {
895 	struct device_state *dev_state;
896 	unsigned long flags;
897 	u32 sbdf;
898 	int ret;
899 
900 	if (!amd_iommu_v2_supported())
901 		return -ENODEV;
902 
903 	sbdf = get_pci_sbdf_id(pdev);
904 
905 	spin_lock_irqsave(&state_lock, flags);
906 
907 	ret = -EINVAL;
908 	dev_state = __get_device_state(sbdf);
909 	if (dev_state == NULL)
910 		goto out_unlock;
911 
912 	dev_state->inv_ctx_cb = cb;
913 
914 	ret = 0;
915 
916 out_unlock:
917 	spin_unlock_irqrestore(&state_lock, flags);
918 
919 	return ret;
920 }
921 EXPORT_SYMBOL(amd_iommu_set_invalidate_ctx_cb);
922 
923 static int __init amd_iommu_v2_init(void)
924 {
925 	int ret;
926 
927 	if (!amd_iommu_v2_supported()) {
928 		pr_info("AMD IOMMUv2 functionality not available on this system - This is not a bug.\n");
929 		/*
930 		 * Load anyway to provide the symbols to other modules
931 		 * which may use AMD IOMMUv2 optionally.
932 		 */
933 		return 0;
934 	}
935 
936 	ret = -ENOMEM;
937 	iommu_wq = alloc_workqueue("amd_iommu_v2", WQ_MEM_RECLAIM, 0);
938 	if (iommu_wq == NULL)
939 		goto out;
940 
941 	amd_iommu_register_ppr_notifier(&ppr_nb);
942 
943 	pr_info("AMD IOMMUv2 loaded and initialized\n");
944 
945 	return 0;
946 
947 out:
948 	return ret;
949 }
950 
951 static void __exit amd_iommu_v2_exit(void)
952 {
953 	struct device_state *dev_state, *next;
954 	unsigned long flags;
955 	LIST_HEAD(freelist);
956 
957 	if (!amd_iommu_v2_supported())
958 		return;
959 
960 	amd_iommu_unregister_ppr_notifier(&ppr_nb);
961 
962 	flush_workqueue(iommu_wq);
963 
964 	/*
965 	 * The loop below might call flush_workqueue(), so call
966 	 * destroy_workqueue() after it
967 	 */
968 	spin_lock_irqsave(&state_lock, flags);
969 
970 	list_for_each_entry_safe(dev_state, next, &state_list, list) {
971 		WARN_ON_ONCE(1);
972 
973 		put_device_state(dev_state);
974 		list_del(&dev_state->list);
975 		list_add_tail(&dev_state->list, &freelist);
976 	}
977 
978 	spin_unlock_irqrestore(&state_lock, flags);
979 
980 	/*
981 	 * Since free_device_state waits on the count to be zero,
982 	 * we need to free dev_state outside the spinlock.
983 	 */
984 	list_for_each_entry_safe(dev_state, next, &freelist, list) {
985 		list_del(&dev_state->list);
986 		free_device_state(dev_state);
987 	}
988 
989 	destroy_workqueue(iommu_wq);
990 }
991 
992 module_init(amd_iommu_v2_init);
993 module_exit(amd_iommu_v2_exit);
994