• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2010-2012 Advanced Micro Devices, Inc.
3  * Author: Joerg Roedel <jroedel@suse.de>
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms of the GNU General Public License version 2 as published
7  * by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write to the Free Software
16  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
17  */
18 
19 #include <linux/mmu_notifier.h>
20 #include <linux/amd-iommu.h>
21 #include <linux/mm_types.h>
22 #include <linux/profile.h>
23 #include <linux/module.h>
24 #include <linux/sched.h>
25 #include <linux/iommu.h>
26 #include <linux/wait.h>
27 #include <linux/pci.h>
28 #include <linux/gfp.h>
29 
30 #include "amd_iommu_types.h"
31 #include "amd_iommu_proto.h"
32 
33 MODULE_LICENSE("GPL v2");
34 MODULE_AUTHOR("Joerg Roedel <jroedel@suse.de>");
35 
36 #define MAX_DEVICES		0x10000
37 #define PRI_QUEUE_SIZE		512
38 
39 struct pri_queue {
40 	atomic_t inflight;
41 	bool finish;
42 	int status;
43 };
44 
45 struct pasid_state {
46 	struct list_head list;			/* For global state-list */
47 	atomic_t count;				/* Reference count */
48 	unsigned mmu_notifier_count;		/* Counting nested mmu_notifier
49 						   calls */
50 	struct mm_struct *mm;			/* mm_struct for the faults */
51 	struct mmu_notifier mn;                 /* mmu_notifier handle */
52 	struct pri_queue pri[PRI_QUEUE_SIZE];	/* PRI tag states */
53 	struct device_state *device_state;	/* Link to our device_state */
54 	int pasid;				/* PASID index */
55 	bool invalid;				/* Used during setup and
56 						   teardown of the pasid */
57 	spinlock_t lock;			/* Protect pri_queues and
58 						   mmu_notifer_count */
59 	wait_queue_head_t wq;			/* To wait for count == 0 */
60 };
61 
62 struct device_state {
63 	struct list_head list;
64 	u16 devid;
65 	atomic_t count;
66 	struct pci_dev *pdev;
67 	struct pasid_state **states;
68 	struct iommu_domain *domain;
69 	int pasid_levels;
70 	int max_pasids;
71 	amd_iommu_invalid_ppr_cb inv_ppr_cb;
72 	amd_iommu_invalidate_ctx inv_ctx_cb;
73 	spinlock_t lock;
74 	wait_queue_head_t wq;
75 };
76 
77 struct fault {
78 	struct work_struct work;
79 	struct device_state *dev_state;
80 	struct pasid_state *state;
81 	struct mm_struct *mm;
82 	u64 address;
83 	u16 devid;
84 	u16 pasid;
85 	u16 tag;
86 	u16 finish;
87 	u16 flags;
88 };
89 
90 static LIST_HEAD(state_list);
91 static spinlock_t state_lock;
92 
93 static struct workqueue_struct *iommu_wq;
94 
95 static void free_pasid_states(struct device_state *dev_state);
96 
device_id(struct pci_dev * pdev)97 static u16 device_id(struct pci_dev *pdev)
98 {
99 	u16 devid;
100 
101 	devid = pdev->bus->number;
102 	devid = (devid << 8) | pdev->devfn;
103 
104 	return devid;
105 }
106 
__get_device_state(u16 devid)107 static struct device_state *__get_device_state(u16 devid)
108 {
109 	struct device_state *dev_state;
110 
111 	list_for_each_entry(dev_state, &state_list, list) {
112 		if (dev_state->devid == devid)
113 			return dev_state;
114 	}
115 
116 	return NULL;
117 }
118 
get_device_state(u16 devid)119 static struct device_state *get_device_state(u16 devid)
120 {
121 	struct device_state *dev_state;
122 	unsigned long flags;
123 
124 	spin_lock_irqsave(&state_lock, flags);
125 	dev_state = __get_device_state(devid);
126 	if (dev_state != NULL)
127 		atomic_inc(&dev_state->count);
128 	spin_unlock_irqrestore(&state_lock, flags);
129 
130 	return dev_state;
131 }
132 
free_device_state(struct device_state * dev_state)133 static void free_device_state(struct device_state *dev_state)
134 {
135 	struct iommu_group *group;
136 
137 	/*
138 	 * First detach device from domain - No more PRI requests will arrive
139 	 * from that device after it is unbound from the IOMMUv2 domain.
140 	 */
141 	group = iommu_group_get(&dev_state->pdev->dev);
142 	if (WARN_ON(!group))
143 		return;
144 
145 	iommu_detach_group(dev_state->domain, group);
146 
147 	iommu_group_put(group);
148 
149 	/* Everything is down now, free the IOMMUv2 domain */
150 	iommu_domain_free(dev_state->domain);
151 
152 	/* Finally get rid of the device-state */
153 	kfree(dev_state);
154 }
155 
put_device_state(struct device_state * dev_state)156 static void put_device_state(struct device_state *dev_state)
157 {
158 	if (atomic_dec_and_test(&dev_state->count))
159 		wake_up(&dev_state->wq);
160 }
161 
162 /* Must be called under dev_state->lock */
__get_pasid_state_ptr(struct device_state * dev_state,int pasid,bool alloc)163 static struct pasid_state **__get_pasid_state_ptr(struct device_state *dev_state,
164 						  int pasid, bool alloc)
165 {
166 	struct pasid_state **root, **ptr;
167 	int level, index;
168 
169 	level = dev_state->pasid_levels;
170 	root  = dev_state->states;
171 
172 	while (true) {
173 
174 		index = (pasid >> (9 * level)) & 0x1ff;
175 		ptr   = &root[index];
176 
177 		if (level == 0)
178 			break;
179 
180 		if (*ptr == NULL) {
181 			if (!alloc)
182 				return NULL;
183 
184 			*ptr = (void *)get_zeroed_page(GFP_ATOMIC);
185 			if (*ptr == NULL)
186 				return NULL;
187 		}
188 
189 		root   = (struct pasid_state **)*ptr;
190 		level -= 1;
191 	}
192 
193 	return ptr;
194 }
195 
set_pasid_state(struct device_state * dev_state,struct pasid_state * pasid_state,int pasid)196 static int set_pasid_state(struct device_state *dev_state,
197 			   struct pasid_state *pasid_state,
198 			   int pasid)
199 {
200 	struct pasid_state **ptr;
201 	unsigned long flags;
202 	int ret;
203 
204 	spin_lock_irqsave(&dev_state->lock, flags);
205 	ptr = __get_pasid_state_ptr(dev_state, pasid, true);
206 
207 	ret = -ENOMEM;
208 	if (ptr == NULL)
209 		goto out_unlock;
210 
211 	ret = -ENOMEM;
212 	if (*ptr != NULL)
213 		goto out_unlock;
214 
215 	*ptr = pasid_state;
216 
217 	ret = 0;
218 
219 out_unlock:
220 	spin_unlock_irqrestore(&dev_state->lock, flags);
221 
222 	return ret;
223 }
224 
clear_pasid_state(struct device_state * dev_state,int pasid)225 static void clear_pasid_state(struct device_state *dev_state, int pasid)
226 {
227 	struct pasid_state **ptr;
228 	unsigned long flags;
229 
230 	spin_lock_irqsave(&dev_state->lock, flags);
231 	ptr = __get_pasid_state_ptr(dev_state, pasid, true);
232 
233 	if (ptr == NULL)
234 		goto out_unlock;
235 
236 	*ptr = NULL;
237 
238 out_unlock:
239 	spin_unlock_irqrestore(&dev_state->lock, flags);
240 }
241 
get_pasid_state(struct device_state * dev_state,int pasid)242 static struct pasid_state *get_pasid_state(struct device_state *dev_state,
243 					   int pasid)
244 {
245 	struct pasid_state **ptr, *ret = NULL;
246 	unsigned long flags;
247 
248 	spin_lock_irqsave(&dev_state->lock, flags);
249 	ptr = __get_pasid_state_ptr(dev_state, pasid, false);
250 
251 	if (ptr == NULL)
252 		goto out_unlock;
253 
254 	ret = *ptr;
255 	if (ret)
256 		atomic_inc(&ret->count);
257 
258 out_unlock:
259 	spin_unlock_irqrestore(&dev_state->lock, flags);
260 
261 	return ret;
262 }
263 
free_pasid_state(struct pasid_state * pasid_state)264 static void free_pasid_state(struct pasid_state *pasid_state)
265 {
266 	kfree(pasid_state);
267 }
268 
put_pasid_state(struct pasid_state * pasid_state)269 static void put_pasid_state(struct pasid_state *pasid_state)
270 {
271 	if (atomic_dec_and_test(&pasid_state->count))
272 		wake_up(&pasid_state->wq);
273 }
274 
put_pasid_state_wait(struct pasid_state * pasid_state)275 static void put_pasid_state_wait(struct pasid_state *pasid_state)
276 {
277 	atomic_dec(&pasid_state->count);
278 	wait_event(pasid_state->wq, !atomic_read(&pasid_state->count));
279 	free_pasid_state(pasid_state);
280 }
281 
unbind_pasid(struct pasid_state * pasid_state)282 static void unbind_pasid(struct pasid_state *pasid_state)
283 {
284 	struct iommu_domain *domain;
285 
286 	domain = pasid_state->device_state->domain;
287 
288 	/*
289 	 * Mark pasid_state as invalid, no more faults will we added to the
290 	 * work queue after this is visible everywhere.
291 	 */
292 	pasid_state->invalid = true;
293 
294 	/* Make sure this is visible */
295 	smp_wmb();
296 
297 	/* After this the device/pasid can't access the mm anymore */
298 	amd_iommu_domain_clear_gcr3(domain, pasid_state->pasid);
299 
300 	/* Make sure no more pending faults are in the queue */
301 	flush_workqueue(iommu_wq);
302 }
303 
free_pasid_states_level1(struct pasid_state ** tbl)304 static void free_pasid_states_level1(struct pasid_state **tbl)
305 {
306 	int i;
307 
308 	for (i = 0; i < 512; ++i) {
309 		if (tbl[i] == NULL)
310 			continue;
311 
312 		free_page((unsigned long)tbl[i]);
313 	}
314 }
315 
free_pasid_states_level2(struct pasid_state ** tbl)316 static void free_pasid_states_level2(struct pasid_state **tbl)
317 {
318 	struct pasid_state **ptr;
319 	int i;
320 
321 	for (i = 0; i < 512; ++i) {
322 		if (tbl[i] == NULL)
323 			continue;
324 
325 		ptr = (struct pasid_state **)tbl[i];
326 		free_pasid_states_level1(ptr);
327 	}
328 }
329 
free_pasid_states(struct device_state * dev_state)330 static void free_pasid_states(struct device_state *dev_state)
331 {
332 	struct pasid_state *pasid_state;
333 	int i;
334 
335 	for (i = 0; i < dev_state->max_pasids; ++i) {
336 		pasid_state = get_pasid_state(dev_state, i);
337 		if (pasid_state == NULL)
338 			continue;
339 
340 		put_pasid_state(pasid_state);
341 
342 		/*
343 		 * This will call the mn_release function and
344 		 * unbind the PASID
345 		 */
346 		mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
347 
348 		put_pasid_state_wait(pasid_state); /* Reference taken in
349 						      amd_iommu_bind_pasid */
350 
351 		/* Drop reference taken in amd_iommu_bind_pasid */
352 		put_device_state(dev_state);
353 	}
354 
355 	if (dev_state->pasid_levels == 2)
356 		free_pasid_states_level2(dev_state->states);
357 	else if (dev_state->pasid_levels == 1)
358 		free_pasid_states_level1(dev_state->states);
359 	else
360 		BUG_ON(dev_state->pasid_levels != 0);
361 
362 	free_page((unsigned long)dev_state->states);
363 }
364 
mn_to_state(struct mmu_notifier * mn)365 static struct pasid_state *mn_to_state(struct mmu_notifier *mn)
366 {
367 	return container_of(mn, struct pasid_state, mn);
368 }
369 
__mn_flush_page(struct mmu_notifier * mn,unsigned long address)370 static void __mn_flush_page(struct mmu_notifier *mn,
371 			    unsigned long address)
372 {
373 	struct pasid_state *pasid_state;
374 	struct device_state *dev_state;
375 
376 	pasid_state = mn_to_state(mn);
377 	dev_state   = pasid_state->device_state;
378 
379 	amd_iommu_flush_page(dev_state->domain, pasid_state->pasid, address);
380 }
381 
mn_clear_flush_young(struct mmu_notifier * mn,struct mm_struct * mm,unsigned long start,unsigned long end)382 static int mn_clear_flush_young(struct mmu_notifier *mn,
383 				struct mm_struct *mm,
384 				unsigned long start,
385 				unsigned long end)
386 {
387 	for (; start < end; start += PAGE_SIZE)
388 		__mn_flush_page(mn, start);
389 
390 	return 0;
391 }
392 
mn_invalidate_page(struct mmu_notifier * mn,struct mm_struct * mm,unsigned long address)393 static void mn_invalidate_page(struct mmu_notifier *mn,
394 			       struct mm_struct *mm,
395 			       unsigned long address)
396 {
397 	__mn_flush_page(mn, address);
398 }
399 
mn_invalidate_range(struct mmu_notifier * mn,struct mm_struct * mm,unsigned long start,unsigned long end)400 static void mn_invalidate_range(struct mmu_notifier *mn,
401 				struct mm_struct *mm,
402 				unsigned long start, unsigned long end)
403 {
404 	struct pasid_state *pasid_state;
405 	struct device_state *dev_state;
406 
407 	pasid_state = mn_to_state(mn);
408 	dev_state   = pasid_state->device_state;
409 
410 	if ((start ^ (end - 1)) < PAGE_SIZE)
411 		amd_iommu_flush_page(dev_state->domain, pasid_state->pasid,
412 				     start);
413 	else
414 		amd_iommu_flush_tlb(dev_state->domain, pasid_state->pasid);
415 }
416 
mn_release(struct mmu_notifier * mn,struct mm_struct * mm)417 static void mn_release(struct mmu_notifier *mn, struct mm_struct *mm)
418 {
419 	struct pasid_state *pasid_state;
420 	struct device_state *dev_state;
421 	bool run_inv_ctx_cb;
422 
423 	might_sleep();
424 
425 	pasid_state    = mn_to_state(mn);
426 	dev_state      = pasid_state->device_state;
427 	run_inv_ctx_cb = !pasid_state->invalid;
428 
429 	if (run_inv_ctx_cb && dev_state->inv_ctx_cb)
430 		dev_state->inv_ctx_cb(dev_state->pdev, pasid_state->pasid);
431 
432 	unbind_pasid(pasid_state);
433 }
434 
435 static const struct mmu_notifier_ops iommu_mn = {
436 	.release		= mn_release,
437 	.clear_flush_young      = mn_clear_flush_young,
438 	.invalidate_page        = mn_invalidate_page,
439 	.invalidate_range       = mn_invalidate_range,
440 };
441 
set_pri_tag_status(struct pasid_state * pasid_state,u16 tag,int status)442 static void set_pri_tag_status(struct pasid_state *pasid_state,
443 			       u16 tag, int status)
444 {
445 	unsigned long flags;
446 
447 	spin_lock_irqsave(&pasid_state->lock, flags);
448 	pasid_state->pri[tag].status = status;
449 	spin_unlock_irqrestore(&pasid_state->lock, flags);
450 }
451 
finish_pri_tag(struct device_state * dev_state,struct pasid_state * pasid_state,u16 tag)452 static void finish_pri_tag(struct device_state *dev_state,
453 			   struct pasid_state *pasid_state,
454 			   u16 tag)
455 {
456 	unsigned long flags;
457 
458 	spin_lock_irqsave(&pasid_state->lock, flags);
459 	if (atomic_dec_and_test(&pasid_state->pri[tag].inflight) &&
460 	    pasid_state->pri[tag].finish) {
461 		amd_iommu_complete_ppr(dev_state->pdev, pasid_state->pasid,
462 				       pasid_state->pri[tag].status, tag);
463 		pasid_state->pri[tag].finish = false;
464 		pasid_state->pri[tag].status = PPR_SUCCESS;
465 	}
466 	spin_unlock_irqrestore(&pasid_state->lock, flags);
467 }
468 
handle_fault_error(struct fault * fault)469 static void handle_fault_error(struct fault *fault)
470 {
471 	int status;
472 
473 	if (!fault->dev_state->inv_ppr_cb) {
474 		set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
475 		return;
476 	}
477 
478 	status = fault->dev_state->inv_ppr_cb(fault->dev_state->pdev,
479 					      fault->pasid,
480 					      fault->address,
481 					      fault->flags);
482 	switch (status) {
483 	case AMD_IOMMU_INV_PRI_RSP_SUCCESS:
484 		set_pri_tag_status(fault->state, fault->tag, PPR_SUCCESS);
485 		break;
486 	case AMD_IOMMU_INV_PRI_RSP_INVALID:
487 		set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
488 		break;
489 	case AMD_IOMMU_INV_PRI_RSP_FAIL:
490 		set_pri_tag_status(fault->state, fault->tag, PPR_FAILURE);
491 		break;
492 	default:
493 		BUG();
494 	}
495 }
496 
access_error(struct vm_area_struct * vma,struct fault * fault)497 static bool access_error(struct vm_area_struct *vma, struct fault *fault)
498 {
499 	unsigned long requested = 0;
500 
501 	if (fault->flags & PPR_FAULT_EXEC)
502 		requested |= VM_EXEC;
503 
504 	if (fault->flags & PPR_FAULT_READ)
505 		requested |= VM_READ;
506 
507 	if (fault->flags & PPR_FAULT_WRITE)
508 		requested |= VM_WRITE;
509 
510 	return (requested & ~vma->vm_flags) != 0;
511 }
512 
do_fault(struct work_struct * work)513 static void do_fault(struct work_struct *work)
514 {
515 	struct fault *fault = container_of(work, struct fault, work);
516 	struct vm_area_struct *vma;
517 	int ret = VM_FAULT_ERROR;
518 	unsigned int flags = 0;
519 	struct mm_struct *mm;
520 	u64 address;
521 
522 	mm = fault->state->mm;
523 	address = fault->address;
524 
525 	if (fault->flags & PPR_FAULT_USER)
526 		flags |= FAULT_FLAG_USER;
527 	if (fault->flags & PPR_FAULT_WRITE)
528 		flags |= FAULT_FLAG_WRITE;
529 	flags |= FAULT_FLAG_REMOTE;
530 
531 	down_read(&mm->mmap_sem);
532 	vma = find_extend_vma(mm, address);
533 	if (!vma || address < vma->vm_start)
534 		/* failed to get a vma in the right range */
535 		goto out;
536 
537 	/* Check if we have the right permissions on the vma */
538 	if (access_error(vma, fault))
539 		goto out;
540 
541 	ret = handle_mm_fault(vma, address, flags);
542 out:
543 	up_read(&mm->mmap_sem);
544 
545 	if (ret & VM_FAULT_ERROR)
546 		/* failed to service fault */
547 		handle_fault_error(fault);
548 
549 	finish_pri_tag(fault->dev_state, fault->state, fault->tag);
550 
551 	put_pasid_state(fault->state);
552 
553 	kfree(fault);
554 }
555 
ppr_notifier(struct notifier_block * nb,unsigned long e,void * data)556 static int ppr_notifier(struct notifier_block *nb, unsigned long e, void *data)
557 {
558 	struct amd_iommu_fault *iommu_fault;
559 	struct pasid_state *pasid_state;
560 	struct device_state *dev_state;
561 	unsigned long flags;
562 	struct fault *fault;
563 	bool finish;
564 	u16 tag;
565 	int ret;
566 
567 	iommu_fault = data;
568 	tag         = iommu_fault->tag & 0x1ff;
569 	finish      = (iommu_fault->tag >> 9) & 1;
570 
571 	ret = NOTIFY_DONE;
572 	dev_state = get_device_state(iommu_fault->device_id);
573 	if (dev_state == NULL)
574 		goto out;
575 
576 	pasid_state = get_pasid_state(dev_state, iommu_fault->pasid);
577 	if (pasid_state == NULL || pasid_state->invalid) {
578 		/* We know the device but not the PASID -> send INVALID */
579 		amd_iommu_complete_ppr(dev_state->pdev, iommu_fault->pasid,
580 				       PPR_INVALID, tag);
581 		goto out_drop_state;
582 	}
583 
584 	spin_lock_irqsave(&pasid_state->lock, flags);
585 	atomic_inc(&pasid_state->pri[tag].inflight);
586 	if (finish)
587 		pasid_state->pri[tag].finish = true;
588 	spin_unlock_irqrestore(&pasid_state->lock, flags);
589 
590 	fault = kzalloc(sizeof(*fault), GFP_ATOMIC);
591 	if (fault == NULL) {
592 		/* We are OOM - send success and let the device re-fault */
593 		finish_pri_tag(dev_state, pasid_state, tag);
594 		goto out_drop_state;
595 	}
596 
597 	fault->dev_state = dev_state;
598 	fault->address   = iommu_fault->address;
599 	fault->state     = pasid_state;
600 	fault->tag       = tag;
601 	fault->finish    = finish;
602 	fault->pasid     = iommu_fault->pasid;
603 	fault->flags     = iommu_fault->flags;
604 	INIT_WORK(&fault->work, do_fault);
605 
606 	queue_work(iommu_wq, &fault->work);
607 
608 	ret = NOTIFY_OK;
609 
610 out_drop_state:
611 
612 	if (ret != NOTIFY_OK && pasid_state)
613 		put_pasid_state(pasid_state);
614 
615 	put_device_state(dev_state);
616 
617 out:
618 	return ret;
619 }
620 
621 static struct notifier_block ppr_nb = {
622 	.notifier_call = ppr_notifier,
623 };
624 
amd_iommu_bind_pasid(struct pci_dev * pdev,int pasid,struct task_struct * task)625 int amd_iommu_bind_pasid(struct pci_dev *pdev, int pasid,
626 			 struct task_struct *task)
627 {
628 	struct pasid_state *pasid_state;
629 	struct device_state *dev_state;
630 	struct mm_struct *mm;
631 	u16 devid;
632 	int ret;
633 
634 	might_sleep();
635 
636 	if (!amd_iommu_v2_supported())
637 		return -ENODEV;
638 
639 	devid     = device_id(pdev);
640 	dev_state = get_device_state(devid);
641 
642 	if (dev_state == NULL)
643 		return -EINVAL;
644 
645 	ret = -EINVAL;
646 	if (pasid < 0 || pasid >= dev_state->max_pasids)
647 		goto out;
648 
649 	ret = -ENOMEM;
650 	pasid_state = kzalloc(sizeof(*pasid_state), GFP_KERNEL);
651 	if (pasid_state == NULL)
652 		goto out;
653 
654 
655 	atomic_set(&pasid_state->count, 1);
656 	init_waitqueue_head(&pasid_state->wq);
657 	spin_lock_init(&pasid_state->lock);
658 
659 	mm                        = get_task_mm(task);
660 	pasid_state->mm           = mm;
661 	pasid_state->device_state = dev_state;
662 	pasid_state->pasid        = pasid;
663 	pasid_state->invalid      = true; /* Mark as valid only if we are
664 					     done with setting up the pasid */
665 	pasid_state->mn.ops       = &iommu_mn;
666 
667 	if (pasid_state->mm == NULL)
668 		goto out_free;
669 
670 	mmu_notifier_register(&pasid_state->mn, mm);
671 
672 	ret = set_pasid_state(dev_state, pasid_state, pasid);
673 	if (ret)
674 		goto out_unregister;
675 
676 	ret = amd_iommu_domain_set_gcr3(dev_state->domain, pasid,
677 					__pa(pasid_state->mm->pgd));
678 	if (ret)
679 		goto out_clear_state;
680 
681 	/* Now we are ready to handle faults */
682 	pasid_state->invalid = false;
683 
684 	/*
685 	 * Drop the reference to the mm_struct here. We rely on the
686 	 * mmu_notifier release call-back to inform us when the mm
687 	 * is going away.
688 	 */
689 	mmput(mm);
690 
691 	return 0;
692 
693 out_clear_state:
694 	clear_pasid_state(dev_state, pasid);
695 
696 out_unregister:
697 	mmu_notifier_unregister(&pasid_state->mn, mm);
698 	mmput(mm);
699 
700 out_free:
701 	free_pasid_state(pasid_state);
702 
703 out:
704 	put_device_state(dev_state);
705 
706 	return ret;
707 }
708 EXPORT_SYMBOL(amd_iommu_bind_pasid);
709 
amd_iommu_unbind_pasid(struct pci_dev * pdev,int pasid)710 void amd_iommu_unbind_pasid(struct pci_dev *pdev, int pasid)
711 {
712 	struct pasid_state *pasid_state;
713 	struct device_state *dev_state;
714 	u16 devid;
715 
716 	might_sleep();
717 
718 	if (!amd_iommu_v2_supported())
719 		return;
720 
721 	devid = device_id(pdev);
722 	dev_state = get_device_state(devid);
723 	if (dev_state == NULL)
724 		return;
725 
726 	if (pasid < 0 || pasid >= dev_state->max_pasids)
727 		goto out;
728 
729 	pasid_state = get_pasid_state(dev_state, pasid);
730 	if (pasid_state == NULL)
731 		goto out;
732 	/*
733 	 * Drop reference taken here. We are safe because we still hold
734 	 * the reference taken in the amd_iommu_bind_pasid function.
735 	 */
736 	put_pasid_state(pasid_state);
737 
738 	/* Clear the pasid state so that the pasid can be re-used */
739 	clear_pasid_state(dev_state, pasid_state->pasid);
740 
741 	/*
742 	 * Call mmu_notifier_unregister to drop our reference
743 	 * to pasid_state->mm
744 	 */
745 	mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
746 
747 	put_pasid_state_wait(pasid_state); /* Reference taken in
748 					      amd_iommu_bind_pasid */
749 out:
750 	/* Drop reference taken in this function */
751 	put_device_state(dev_state);
752 
753 	/* Drop reference taken in amd_iommu_bind_pasid */
754 	put_device_state(dev_state);
755 }
756 EXPORT_SYMBOL(amd_iommu_unbind_pasid);
757 
amd_iommu_init_device(struct pci_dev * pdev,int pasids)758 int amd_iommu_init_device(struct pci_dev *pdev, int pasids)
759 {
760 	struct device_state *dev_state;
761 	struct iommu_group *group;
762 	unsigned long flags;
763 	int ret, tmp;
764 	u16 devid;
765 
766 	might_sleep();
767 
768 	if (!amd_iommu_v2_supported())
769 		return -ENODEV;
770 
771 	if (pasids <= 0 || pasids > (PASID_MASK + 1))
772 		return -EINVAL;
773 
774 	devid = device_id(pdev);
775 
776 	dev_state = kzalloc(sizeof(*dev_state), GFP_KERNEL);
777 	if (dev_state == NULL)
778 		return -ENOMEM;
779 
780 	spin_lock_init(&dev_state->lock);
781 	init_waitqueue_head(&dev_state->wq);
782 	dev_state->pdev  = pdev;
783 	dev_state->devid = devid;
784 
785 	tmp = pasids;
786 	for (dev_state->pasid_levels = 0; (tmp - 1) & ~0x1ff; tmp >>= 9)
787 		dev_state->pasid_levels += 1;
788 
789 	atomic_set(&dev_state->count, 1);
790 	dev_state->max_pasids = pasids;
791 
792 	ret = -ENOMEM;
793 	dev_state->states = (void *)get_zeroed_page(GFP_KERNEL);
794 	if (dev_state->states == NULL)
795 		goto out_free_dev_state;
796 
797 	dev_state->domain = iommu_domain_alloc(&pci_bus_type);
798 	if (dev_state->domain == NULL)
799 		goto out_free_states;
800 
801 	amd_iommu_domain_direct_map(dev_state->domain);
802 
803 	ret = amd_iommu_domain_enable_v2(dev_state->domain, pasids);
804 	if (ret)
805 		goto out_free_domain;
806 
807 	group = iommu_group_get(&pdev->dev);
808 	if (!group) {
809 		ret = -EINVAL;
810 		goto out_free_domain;
811 	}
812 
813 	ret = iommu_attach_group(dev_state->domain, group);
814 	if (ret != 0)
815 		goto out_drop_group;
816 
817 	iommu_group_put(group);
818 
819 	spin_lock_irqsave(&state_lock, flags);
820 
821 	if (__get_device_state(devid) != NULL) {
822 		spin_unlock_irqrestore(&state_lock, flags);
823 		ret = -EBUSY;
824 		goto out_free_domain;
825 	}
826 
827 	list_add_tail(&dev_state->list, &state_list);
828 
829 	spin_unlock_irqrestore(&state_lock, flags);
830 
831 	return 0;
832 
833 out_drop_group:
834 	iommu_group_put(group);
835 
836 out_free_domain:
837 	iommu_domain_free(dev_state->domain);
838 
839 out_free_states:
840 	free_page((unsigned long)dev_state->states);
841 
842 out_free_dev_state:
843 	kfree(dev_state);
844 
845 	return ret;
846 }
847 EXPORT_SYMBOL(amd_iommu_init_device);
848 
amd_iommu_free_device(struct pci_dev * pdev)849 void amd_iommu_free_device(struct pci_dev *pdev)
850 {
851 	struct device_state *dev_state;
852 	unsigned long flags;
853 	u16 devid;
854 
855 	if (!amd_iommu_v2_supported())
856 		return;
857 
858 	devid = device_id(pdev);
859 
860 	spin_lock_irqsave(&state_lock, flags);
861 
862 	dev_state = __get_device_state(devid);
863 	if (dev_state == NULL) {
864 		spin_unlock_irqrestore(&state_lock, flags);
865 		return;
866 	}
867 
868 	list_del(&dev_state->list);
869 
870 	spin_unlock_irqrestore(&state_lock, flags);
871 
872 	/* Get rid of any remaining pasid states */
873 	free_pasid_states(dev_state);
874 
875 	put_device_state(dev_state);
876 	/*
877 	 * Wait until the last reference is dropped before freeing
878 	 * the device state.
879 	 */
880 	wait_event(dev_state->wq, !atomic_read(&dev_state->count));
881 	free_device_state(dev_state);
882 }
883 EXPORT_SYMBOL(amd_iommu_free_device);
884 
amd_iommu_set_invalid_ppr_cb(struct pci_dev * pdev,amd_iommu_invalid_ppr_cb cb)885 int amd_iommu_set_invalid_ppr_cb(struct pci_dev *pdev,
886 				 amd_iommu_invalid_ppr_cb cb)
887 {
888 	struct device_state *dev_state;
889 	unsigned long flags;
890 	u16 devid;
891 	int ret;
892 
893 	if (!amd_iommu_v2_supported())
894 		return -ENODEV;
895 
896 	devid = device_id(pdev);
897 
898 	spin_lock_irqsave(&state_lock, flags);
899 
900 	ret = -EINVAL;
901 	dev_state = __get_device_state(devid);
902 	if (dev_state == NULL)
903 		goto out_unlock;
904 
905 	dev_state->inv_ppr_cb = cb;
906 
907 	ret = 0;
908 
909 out_unlock:
910 	spin_unlock_irqrestore(&state_lock, flags);
911 
912 	return ret;
913 }
914 EXPORT_SYMBOL(amd_iommu_set_invalid_ppr_cb);
915 
amd_iommu_set_invalidate_ctx_cb(struct pci_dev * pdev,amd_iommu_invalidate_ctx cb)916 int amd_iommu_set_invalidate_ctx_cb(struct pci_dev *pdev,
917 				    amd_iommu_invalidate_ctx cb)
918 {
919 	struct device_state *dev_state;
920 	unsigned long flags;
921 	u16 devid;
922 	int ret;
923 
924 	if (!amd_iommu_v2_supported())
925 		return -ENODEV;
926 
927 	devid = device_id(pdev);
928 
929 	spin_lock_irqsave(&state_lock, flags);
930 
931 	ret = -EINVAL;
932 	dev_state = __get_device_state(devid);
933 	if (dev_state == NULL)
934 		goto out_unlock;
935 
936 	dev_state->inv_ctx_cb = cb;
937 
938 	ret = 0;
939 
940 out_unlock:
941 	spin_unlock_irqrestore(&state_lock, flags);
942 
943 	return ret;
944 }
945 EXPORT_SYMBOL(amd_iommu_set_invalidate_ctx_cb);
946 
amd_iommu_v2_init(void)947 static int __init amd_iommu_v2_init(void)
948 {
949 	int ret;
950 
951 	pr_info("AMD IOMMUv2 driver by Joerg Roedel <jroedel@suse.de>\n");
952 
953 	if (!amd_iommu_v2_supported()) {
954 		pr_info("AMD IOMMUv2 functionality not available on this system\n");
955 		/*
956 		 * Load anyway to provide the symbols to other modules
957 		 * which may use AMD IOMMUv2 optionally.
958 		 */
959 		return 0;
960 	}
961 
962 	spin_lock_init(&state_lock);
963 
964 	ret = -ENOMEM;
965 	iommu_wq = alloc_workqueue("amd_iommu_v2", WQ_MEM_RECLAIM, 0);
966 	if (iommu_wq == NULL)
967 		goto out;
968 
969 	amd_iommu_register_ppr_notifier(&ppr_nb);
970 
971 	return 0;
972 
973 out:
974 	return ret;
975 }
976 
amd_iommu_v2_exit(void)977 static void __exit amd_iommu_v2_exit(void)
978 {
979 	struct device_state *dev_state;
980 	int i;
981 
982 	if (!amd_iommu_v2_supported())
983 		return;
984 
985 	amd_iommu_unregister_ppr_notifier(&ppr_nb);
986 
987 	flush_workqueue(iommu_wq);
988 
989 	/*
990 	 * The loop below might call flush_workqueue(), so call
991 	 * destroy_workqueue() after it
992 	 */
993 	for (i = 0; i < MAX_DEVICES; ++i) {
994 		dev_state = get_device_state(i);
995 
996 		if (dev_state == NULL)
997 			continue;
998 
999 		WARN_ON_ONCE(1);
1000 
1001 		put_device_state(dev_state);
1002 		amd_iommu_free_device(dev_state->pdev);
1003 	}
1004 
1005 	destroy_workqueue(iommu_wq);
1006 }
1007 
1008 module_init(amd_iommu_v2_init);
1009 module_exit(amd_iommu_v2_exit);
1010