2 * Copyright (C) 2010-2012 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <jroedel@suse.de>
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.
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.
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
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>
30 #include "amd_iommu_types.h"
31 #include "amd_iommu_proto.h"
33 MODULE_LICENSE("GPL v2");
34 MODULE_AUTHOR("Joerg Roedel <jroedel@suse.de>");
36 #define MAX_DEVICES 0x10000
37 #define PRI_QUEUE_SIZE 512
46 struct list_head list; /* For global state-list */
47 atomic_t count; /* Reference count */
48 unsigned mmu_notifier_count; /* Counting nested mmu_notifier
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
59 wait_queue_head_t wq; /* To wait for count == 0 */
63 struct list_head list;
67 struct pasid_state **states;
68 struct iommu_domain *domain;
71 amd_iommu_invalid_ppr_cb inv_ppr_cb;
72 amd_iommu_invalidate_ctx inv_ctx_cb;
78 struct work_struct work;
79 struct device_state *dev_state;
80 struct pasid_state *state;
90 static LIST_HEAD(state_list);
91 static spinlock_t state_lock;
93 static struct workqueue_struct *iommu_wq;
95 static void free_pasid_states(struct device_state *dev_state);
97 static u16 device_id(struct pci_dev *pdev)
101 devid = pdev->bus->number;
102 devid = (devid << 8) | pdev->devfn;
107 static struct device_state *__get_device_state(u16 devid)
109 struct device_state *dev_state;
111 list_for_each_entry(dev_state, &state_list, list) {
112 if (dev_state->devid == devid)
119 static struct device_state *get_device_state(u16 devid)
121 struct device_state *dev_state;
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);
133 static void free_device_state(struct device_state *dev_state)
136 * First detach device from domain - No more PRI requests will arrive
137 * from that device after it is unbound from the IOMMUv2 domain.
139 iommu_detach_device(dev_state->domain, &dev_state->pdev->dev);
141 /* Everything is down now, free the IOMMUv2 domain */
142 iommu_domain_free(dev_state->domain);
144 /* Finally get rid of the device-state */
148 static void put_device_state(struct device_state *dev_state)
150 if (atomic_dec_and_test(&dev_state->count))
151 wake_up(&dev_state->wq);
154 /* Must be called under dev_state->lock */
155 static struct pasid_state **__get_pasid_state_ptr(struct device_state *dev_state,
156 int pasid, bool alloc)
158 struct pasid_state **root, **ptr;
161 level = dev_state->pasid_levels;
162 root = dev_state->states;
166 index = (pasid >> (9 * level)) & 0x1ff;
176 *ptr = (void *)get_zeroed_page(GFP_ATOMIC);
181 root = (struct pasid_state **)*ptr;
188 static int set_pasid_state(struct device_state *dev_state,
189 struct pasid_state *pasid_state,
192 struct pasid_state **ptr;
196 spin_lock_irqsave(&dev_state->lock, flags);
197 ptr = __get_pasid_state_ptr(dev_state, pasid, true);
212 spin_unlock_irqrestore(&dev_state->lock, flags);
217 static void clear_pasid_state(struct device_state *dev_state, int pasid)
219 struct pasid_state **ptr;
222 spin_lock_irqsave(&dev_state->lock, flags);
223 ptr = __get_pasid_state_ptr(dev_state, pasid, true);
231 spin_unlock_irqrestore(&dev_state->lock, flags);
234 static struct pasid_state *get_pasid_state(struct device_state *dev_state,
237 struct pasid_state **ptr, *ret = NULL;
240 spin_lock_irqsave(&dev_state->lock, flags);
241 ptr = __get_pasid_state_ptr(dev_state, pasid, false);
248 atomic_inc(&ret->count);
251 spin_unlock_irqrestore(&dev_state->lock, flags);
256 static void free_pasid_state(struct pasid_state *pasid_state)
261 static void put_pasid_state(struct pasid_state *pasid_state)
263 if (atomic_dec_and_test(&pasid_state->count))
264 wake_up(&pasid_state->wq);
267 static void put_pasid_state_wait(struct pasid_state *pasid_state)
269 atomic_dec(&pasid_state->count);
270 wait_event(pasid_state->wq, !atomic_read(&pasid_state->count));
271 free_pasid_state(pasid_state);
274 static void unbind_pasid(struct pasid_state *pasid_state)
276 struct iommu_domain *domain;
278 domain = pasid_state->device_state->domain;
281 * Mark pasid_state as invalid, no more faults will we added to the
282 * work queue after this is visible everywhere.
284 pasid_state->invalid = true;
286 /* Make sure this is visible */
289 /* After this the device/pasid can't access the mm anymore */
290 amd_iommu_domain_clear_gcr3(domain, pasid_state->pasid);
292 /* Make sure no more pending faults are in the queue */
293 flush_workqueue(iommu_wq);
296 static void free_pasid_states_level1(struct pasid_state **tbl)
300 for (i = 0; i < 512; ++i) {
304 free_page((unsigned long)tbl[i]);
308 static void free_pasid_states_level2(struct pasid_state **tbl)
310 struct pasid_state **ptr;
313 for (i = 0; i < 512; ++i) {
317 ptr = (struct pasid_state **)tbl[i];
318 free_pasid_states_level1(ptr);
322 static void free_pasid_states(struct device_state *dev_state)
324 struct pasid_state *pasid_state;
327 for (i = 0; i < dev_state->max_pasids; ++i) {
328 pasid_state = get_pasid_state(dev_state, i);
329 if (pasid_state == NULL)
332 put_pasid_state(pasid_state);
335 * This will call the mn_release function and
338 mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
340 put_pasid_state_wait(pasid_state); /* Reference taken in
341 amd_iommu_bind_pasid */
343 /* Drop reference taken in amd_iommu_bind_pasid */
344 put_device_state(dev_state);
347 if (dev_state->pasid_levels == 2)
348 free_pasid_states_level2(dev_state->states);
349 else if (dev_state->pasid_levels == 1)
350 free_pasid_states_level1(dev_state->states);
351 else if (dev_state->pasid_levels != 0)
354 free_page((unsigned long)dev_state->states);
357 static struct pasid_state *mn_to_state(struct mmu_notifier *mn)
359 return container_of(mn, struct pasid_state, mn);
362 static void __mn_flush_page(struct mmu_notifier *mn,
363 unsigned long address)
365 struct pasid_state *pasid_state;
366 struct device_state *dev_state;
368 pasid_state = mn_to_state(mn);
369 dev_state = pasid_state->device_state;
371 amd_iommu_flush_page(dev_state->domain, pasid_state->pasid, address);
374 static int mn_clear_flush_young(struct mmu_notifier *mn,
375 struct mm_struct *mm,
379 for (; start < end; start += PAGE_SIZE)
380 __mn_flush_page(mn, start);
385 static void mn_invalidate_page(struct mmu_notifier *mn,
386 struct mm_struct *mm,
387 unsigned long address)
389 __mn_flush_page(mn, address);
392 static void mn_invalidate_range(struct mmu_notifier *mn,
393 struct mm_struct *mm,
394 unsigned long start, unsigned long end)
396 struct pasid_state *pasid_state;
397 struct device_state *dev_state;
399 pasid_state = mn_to_state(mn);
400 dev_state = pasid_state->device_state;
402 if ((start ^ (end - 1)) < PAGE_SIZE)
403 amd_iommu_flush_page(dev_state->domain, pasid_state->pasid,
406 amd_iommu_flush_tlb(dev_state->domain, pasid_state->pasid);
409 static void mn_release(struct mmu_notifier *mn, struct mm_struct *mm)
411 struct pasid_state *pasid_state;
412 struct device_state *dev_state;
417 pasid_state = mn_to_state(mn);
418 dev_state = pasid_state->device_state;
419 run_inv_ctx_cb = !pasid_state->invalid;
421 if (run_inv_ctx_cb && dev_state->inv_ctx_cb)
422 dev_state->inv_ctx_cb(dev_state->pdev, pasid_state->pasid);
424 unbind_pasid(pasid_state);
427 static struct mmu_notifier_ops iommu_mn = {
428 .release = mn_release,
429 .clear_flush_young = mn_clear_flush_young,
430 .invalidate_page = mn_invalidate_page,
431 .invalidate_range = mn_invalidate_range,
434 static void set_pri_tag_status(struct pasid_state *pasid_state,
439 spin_lock_irqsave(&pasid_state->lock, flags);
440 pasid_state->pri[tag].status = status;
441 spin_unlock_irqrestore(&pasid_state->lock, flags);
444 static void finish_pri_tag(struct device_state *dev_state,
445 struct pasid_state *pasid_state,
450 spin_lock_irqsave(&pasid_state->lock, flags);
451 if (atomic_dec_and_test(&pasid_state->pri[tag].inflight) &&
452 pasid_state->pri[tag].finish) {
453 amd_iommu_complete_ppr(dev_state->pdev, pasid_state->pasid,
454 pasid_state->pri[tag].status, tag);
455 pasid_state->pri[tag].finish = false;
456 pasid_state->pri[tag].status = PPR_SUCCESS;
458 spin_unlock_irqrestore(&pasid_state->lock, flags);
461 static void handle_fault_error(struct fault *fault)
465 if (!fault->dev_state->inv_ppr_cb) {
466 set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
470 status = fault->dev_state->inv_ppr_cb(fault->dev_state->pdev,
475 case AMD_IOMMU_INV_PRI_RSP_SUCCESS:
476 set_pri_tag_status(fault->state, fault->tag, PPR_SUCCESS);
478 case AMD_IOMMU_INV_PRI_RSP_INVALID:
479 set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
481 case AMD_IOMMU_INV_PRI_RSP_FAIL:
482 set_pri_tag_status(fault->state, fault->tag, PPR_FAILURE);
489 static void do_fault(struct work_struct *work)
491 struct fault *fault = container_of(work, struct fault, work);
492 struct mm_struct *mm;
493 struct vm_area_struct *vma;
497 write = !!(fault->flags & PPR_FAULT_WRITE);
499 mm = fault->state->mm;
500 address = fault->address;
502 down_read(&mm->mmap_sem);
503 vma = find_extend_vma(mm, address);
504 if (!vma || address < vma->vm_start) {
505 /* failed to get a vma in the right range */
506 up_read(&mm->mmap_sem);
507 handle_fault_error(fault);
511 ret = handle_mm_fault(mm, vma, address, write);
512 if (ret & VM_FAULT_ERROR) {
513 /* failed to service fault */
514 up_read(&mm->mmap_sem);
515 handle_fault_error(fault);
519 up_read(&mm->mmap_sem);
522 finish_pri_tag(fault->dev_state, fault->state, fault->tag);
524 put_pasid_state(fault->state);
529 static int ppr_notifier(struct notifier_block *nb, unsigned long e, void *data)
531 struct amd_iommu_fault *iommu_fault;
532 struct pasid_state *pasid_state;
533 struct device_state *dev_state;
541 tag = iommu_fault->tag & 0x1ff;
542 finish = (iommu_fault->tag >> 9) & 1;
545 dev_state = get_device_state(iommu_fault->device_id);
546 if (dev_state == NULL)
549 pasid_state = get_pasid_state(dev_state, iommu_fault->pasid);
550 if (pasid_state == NULL || pasid_state->invalid) {
551 /* We know the device but not the PASID -> send INVALID */
552 amd_iommu_complete_ppr(dev_state->pdev, iommu_fault->pasid,
557 spin_lock_irqsave(&pasid_state->lock, flags);
558 atomic_inc(&pasid_state->pri[tag].inflight);
560 pasid_state->pri[tag].finish = true;
561 spin_unlock_irqrestore(&pasid_state->lock, flags);
563 fault = kzalloc(sizeof(*fault), GFP_ATOMIC);
565 /* We are OOM - send success and let the device re-fault */
566 finish_pri_tag(dev_state, pasid_state, tag);
570 fault->dev_state = dev_state;
571 fault->address = iommu_fault->address;
572 fault->state = pasid_state;
574 fault->finish = finish;
575 fault->pasid = iommu_fault->pasid;
576 fault->flags = iommu_fault->flags;
577 INIT_WORK(&fault->work, do_fault);
579 queue_work(iommu_wq, &fault->work);
585 if (ret != NOTIFY_OK && pasid_state)
586 put_pasid_state(pasid_state);
588 put_device_state(dev_state);
594 static struct notifier_block ppr_nb = {
595 .notifier_call = ppr_notifier,
598 int amd_iommu_bind_pasid(struct pci_dev *pdev, int pasid,
599 struct task_struct *task)
601 struct pasid_state *pasid_state;
602 struct device_state *dev_state;
603 struct mm_struct *mm;
609 if (!amd_iommu_v2_supported())
612 devid = device_id(pdev);
613 dev_state = get_device_state(devid);
615 if (dev_state == NULL)
619 if (pasid < 0 || pasid >= dev_state->max_pasids)
623 pasid_state = kzalloc(sizeof(*pasid_state), GFP_KERNEL);
624 if (pasid_state == NULL)
628 atomic_set(&pasid_state->count, 1);
629 init_waitqueue_head(&pasid_state->wq);
630 spin_lock_init(&pasid_state->lock);
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;
640 if (pasid_state->mm == NULL)
643 mmu_notifier_register(&pasid_state->mn, mm);
645 ret = set_pasid_state(dev_state, pasid_state, pasid);
649 ret = amd_iommu_domain_set_gcr3(dev_state->domain, pasid,
650 __pa(pasid_state->mm->pgd));
652 goto out_clear_state;
654 /* Now we are ready to handle faults */
655 pasid_state->invalid = false;
658 * Drop the reference to the mm_struct here. We rely on the
659 * mmu_notifier release call-back to inform us when the mm
667 clear_pasid_state(dev_state, pasid);
670 mmu_notifier_unregister(&pasid_state->mn, mm);
674 free_pasid_state(pasid_state);
677 put_device_state(dev_state);
681 EXPORT_SYMBOL(amd_iommu_bind_pasid);
683 void amd_iommu_unbind_pasid(struct pci_dev *pdev, int pasid)
685 struct pasid_state *pasid_state;
686 struct device_state *dev_state;
691 if (!amd_iommu_v2_supported())
694 devid = device_id(pdev);
695 dev_state = get_device_state(devid);
696 if (dev_state == NULL)
699 if (pasid < 0 || pasid >= dev_state->max_pasids)
702 pasid_state = get_pasid_state(dev_state, pasid);
703 if (pasid_state == NULL)
706 * Drop reference taken here. We are safe because we still hold
707 * the reference taken in the amd_iommu_bind_pasid function.
709 put_pasid_state(pasid_state);
711 /* Clear the pasid state so that the pasid can be re-used */
712 clear_pasid_state(dev_state, pasid_state->pasid);
715 * Call mmu_notifier_unregister to drop our reference
718 mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
720 put_pasid_state_wait(pasid_state); /* Reference taken in
721 amd_iommu_bind_pasid */
723 /* Drop reference taken in this function */
724 put_device_state(dev_state);
726 /* Drop reference taken in amd_iommu_bind_pasid */
727 put_device_state(dev_state);
729 EXPORT_SYMBOL(amd_iommu_unbind_pasid);
731 int amd_iommu_init_device(struct pci_dev *pdev, int pasids)
733 struct device_state *dev_state;
740 if (!amd_iommu_v2_supported())
743 if (pasids <= 0 || pasids > (PASID_MASK + 1))
746 devid = device_id(pdev);
748 dev_state = kzalloc(sizeof(*dev_state), GFP_KERNEL);
749 if (dev_state == NULL)
752 spin_lock_init(&dev_state->lock);
753 init_waitqueue_head(&dev_state->wq);
754 dev_state->pdev = pdev;
755 dev_state->devid = devid;
758 for (dev_state->pasid_levels = 0; (tmp - 1) & ~0x1ff; tmp >>= 9)
759 dev_state->pasid_levels += 1;
761 atomic_set(&dev_state->count, 1);
762 dev_state->max_pasids = pasids;
765 dev_state->states = (void *)get_zeroed_page(GFP_KERNEL);
766 if (dev_state->states == NULL)
767 goto out_free_dev_state;
769 dev_state->domain = iommu_domain_alloc(&pci_bus_type);
770 if (dev_state->domain == NULL)
771 goto out_free_states;
773 amd_iommu_domain_direct_map(dev_state->domain);
775 ret = amd_iommu_domain_enable_v2(dev_state->domain, pasids);
777 goto out_free_domain;
779 ret = iommu_attach_device(dev_state->domain, &pdev->dev);
781 goto out_free_domain;
783 spin_lock_irqsave(&state_lock, flags);
785 if (__get_device_state(devid) != NULL) {
786 spin_unlock_irqrestore(&state_lock, flags);
788 goto out_free_domain;
791 list_add_tail(&dev_state->list, &state_list);
793 spin_unlock_irqrestore(&state_lock, flags);
798 iommu_domain_free(dev_state->domain);
801 free_page((unsigned long)dev_state->states);
808 EXPORT_SYMBOL(amd_iommu_init_device);
810 void amd_iommu_free_device(struct pci_dev *pdev)
812 struct device_state *dev_state;
816 if (!amd_iommu_v2_supported())
819 devid = device_id(pdev);
821 spin_lock_irqsave(&state_lock, flags);
823 dev_state = __get_device_state(devid);
824 if (dev_state == NULL) {
825 spin_unlock_irqrestore(&state_lock, flags);
829 list_del(&dev_state->list);
831 spin_unlock_irqrestore(&state_lock, flags);
833 /* Get rid of any remaining pasid states */
834 free_pasid_states(dev_state);
836 put_device_state(dev_state);
838 * Wait until the last reference is dropped before freeing
841 wait_event(dev_state->wq, !atomic_read(&dev_state->count));
842 free_device_state(dev_state);
844 EXPORT_SYMBOL(amd_iommu_free_device);
846 int amd_iommu_set_invalid_ppr_cb(struct pci_dev *pdev,
847 amd_iommu_invalid_ppr_cb cb)
849 struct device_state *dev_state;
854 if (!amd_iommu_v2_supported())
857 devid = device_id(pdev);
859 spin_lock_irqsave(&state_lock, flags);
862 dev_state = __get_device_state(devid);
863 if (dev_state == NULL)
866 dev_state->inv_ppr_cb = cb;
871 spin_unlock_irqrestore(&state_lock, flags);
875 EXPORT_SYMBOL(amd_iommu_set_invalid_ppr_cb);
877 int amd_iommu_set_invalidate_ctx_cb(struct pci_dev *pdev,
878 amd_iommu_invalidate_ctx cb)
880 struct device_state *dev_state;
885 if (!amd_iommu_v2_supported())
888 devid = device_id(pdev);
890 spin_lock_irqsave(&state_lock, flags);
893 dev_state = __get_device_state(devid);
894 if (dev_state == NULL)
897 dev_state->inv_ctx_cb = cb;
902 spin_unlock_irqrestore(&state_lock, flags);
906 EXPORT_SYMBOL(amd_iommu_set_invalidate_ctx_cb);
908 static int __init amd_iommu_v2_init(void)
912 pr_info("AMD IOMMUv2 driver by Joerg Roedel <jroedel@suse.de>\n");
914 if (!amd_iommu_v2_supported()) {
915 pr_info("AMD IOMMUv2 functionality not available on this system\n");
917 * Load anyway to provide the symbols to other modules
918 * which may use AMD IOMMUv2 optionally.
923 spin_lock_init(&state_lock);
926 iommu_wq = create_workqueue("amd_iommu_v2");
927 if (iommu_wq == NULL)
930 amd_iommu_register_ppr_notifier(&ppr_nb);
938 static void __exit amd_iommu_v2_exit(void)
940 struct device_state *dev_state;
943 if (!amd_iommu_v2_supported())
946 amd_iommu_unregister_ppr_notifier(&ppr_nb);
948 flush_workqueue(iommu_wq);
951 * The loop below might call flush_workqueue(), so call
952 * destroy_workqueue() after it
954 for (i = 0; i < MAX_DEVICES; ++i) {
955 dev_state = get_device_state(i);
957 if (dev_state == NULL)
962 put_device_state(dev_state);
963 amd_iommu_free_device(dev_state->pdev);
966 destroy_workqueue(iommu_wq);
969 module_init(amd_iommu_v2_init);
970 module_exit(amd_iommu_v2_exit);