1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2010-2012 Advanced Micro Devices, Inc.
4 * Author: Joerg Roedel <jroedel@suse.de>
7 #define pr_fmt(fmt) "AMD-Vi: " fmt
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>
22 #include "amd_iommu.h"
24 MODULE_LICENSE("GPL v2");
25 MODULE_AUTHOR("Joerg Roedel <jroedel@suse.de>");
27 #define MAX_DEVICES 0x10000
28 #define PRI_QUEUE_SIZE 512
37 struct list_head list; /* For global state-list */
38 refcount_t count; /* Reference count */
39 unsigned mmu_notifier_count; /* Counting nested mmu_notifier
41 struct mm_struct *mm; /* mm_struct for the faults */
42 struct mmu_notifier mn; /* mmu_notifier handle */
43 struct pri_queue pri[PRI_QUEUE_SIZE]; /* PRI tag states */
44 struct device_state *device_state; /* Link to our device_state */
45 u32 pasid; /* PASID index */
46 bool invalid; /* Used during setup and
47 teardown of the pasid */
48 spinlock_t lock; /* Protect pri_queues and
50 wait_queue_head_t wq; /* To wait for count == 0 */
54 struct list_head list;
58 struct pasid_state **states;
59 struct iommu_domain *domain;
62 amd_iommu_invalid_ppr_cb inv_ppr_cb;
63 amd_iommu_invalidate_ctx inv_ctx_cb;
69 struct work_struct work;
70 struct device_state *dev_state;
71 struct pasid_state *state;
81 static LIST_HEAD(state_list);
82 static DEFINE_SPINLOCK(state_lock);
84 static struct workqueue_struct *iommu_wq;
86 static void free_pasid_states(struct device_state *dev_state);
88 static u16 device_id(struct pci_dev *pdev)
92 devid = pdev->bus->number;
93 devid = (devid << 8) | pdev->devfn;
98 static struct device_state *__get_device_state(u16 devid)
100 struct device_state *dev_state;
102 list_for_each_entry(dev_state, &state_list, list) {
103 if (dev_state->devid == devid)
110 static struct device_state *get_device_state(u16 devid)
112 struct device_state *dev_state;
115 spin_lock_irqsave(&state_lock, flags);
116 dev_state = __get_device_state(devid);
117 if (dev_state != NULL)
118 atomic_inc(&dev_state->count);
119 spin_unlock_irqrestore(&state_lock, flags);
124 static void free_device_state(struct device_state *dev_state)
126 struct iommu_group *group;
129 * First detach device from domain - No more PRI requests will arrive
130 * from that device after it is unbound from the IOMMUv2 domain.
132 group = iommu_group_get(&dev_state->pdev->dev);
136 iommu_detach_group(dev_state->domain, group);
138 iommu_group_put(group);
140 /* Everything is down now, free the IOMMUv2 domain */
141 iommu_domain_free(dev_state->domain);
143 /* Finally get rid of the device-state */
147 static void put_device_state(struct device_state *dev_state)
149 if (atomic_dec_and_test(&dev_state->count))
150 wake_up(&dev_state->wq);
153 /* Must be called under dev_state->lock */
154 static struct pasid_state **__get_pasid_state_ptr(struct device_state *dev_state,
155 u32 pasid, bool alloc)
157 struct pasid_state **root, **ptr;
160 level = dev_state->pasid_levels;
161 root = dev_state->states;
165 index = (pasid >> (9 * level)) & 0x1ff;
175 *ptr = (void *)get_zeroed_page(GFP_ATOMIC);
180 root = (struct pasid_state **)*ptr;
187 static int set_pasid_state(struct device_state *dev_state,
188 struct pasid_state *pasid_state,
191 struct pasid_state **ptr;
195 spin_lock_irqsave(&dev_state->lock, flags);
196 ptr = __get_pasid_state_ptr(dev_state, pasid, true);
211 spin_unlock_irqrestore(&dev_state->lock, flags);
216 static void clear_pasid_state(struct device_state *dev_state, u32 pasid)
218 struct pasid_state **ptr;
221 spin_lock_irqsave(&dev_state->lock, flags);
222 ptr = __get_pasid_state_ptr(dev_state, pasid, true);
230 spin_unlock_irqrestore(&dev_state->lock, flags);
233 static struct pasid_state *get_pasid_state(struct device_state *dev_state,
236 struct pasid_state **ptr, *ret = NULL;
239 spin_lock_irqsave(&dev_state->lock, flags);
240 ptr = __get_pasid_state_ptr(dev_state, pasid, false);
247 refcount_inc(&ret->count);
250 spin_unlock_irqrestore(&dev_state->lock, flags);
255 static void free_pasid_state(struct pasid_state *pasid_state)
260 static void put_pasid_state(struct pasid_state *pasid_state)
262 if (refcount_dec_and_test(&pasid_state->count))
263 wake_up(&pasid_state->wq);
266 static void put_pasid_state_wait(struct pasid_state *pasid_state)
268 refcount_dec(&pasid_state->count);
269 wait_event(pasid_state->wq, !refcount_read(&pasid_state->count));
270 free_pasid_state(pasid_state);
273 static void unbind_pasid(struct pasid_state *pasid_state)
275 struct iommu_domain *domain;
277 domain = pasid_state->device_state->domain;
280 * Mark pasid_state as invalid, no more faults will we added to the
281 * work queue after this is visible everywhere.
283 pasid_state->invalid = true;
285 /* Make sure this is visible */
288 /* After this the device/pasid can't access the mm anymore */
289 amd_iommu_domain_clear_gcr3(domain, pasid_state->pasid);
291 /* Make sure no more pending faults are in the queue */
292 flush_workqueue(iommu_wq);
295 static void free_pasid_states_level1(struct pasid_state **tbl)
299 for (i = 0; i < 512; ++i) {
303 free_page((unsigned long)tbl[i]);
307 static void free_pasid_states_level2(struct pasid_state **tbl)
309 struct pasid_state **ptr;
312 for (i = 0; i < 512; ++i) {
316 ptr = (struct pasid_state **)tbl[i];
317 free_pasid_states_level1(ptr);
321 static void free_pasid_states(struct device_state *dev_state)
323 struct pasid_state *pasid_state;
326 for (i = 0; i < dev_state->max_pasids; ++i) {
327 pasid_state = get_pasid_state(dev_state, i);
328 if (pasid_state == NULL)
331 put_pasid_state(pasid_state);
334 * This will call the mn_release function and
337 mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
339 put_pasid_state_wait(pasid_state); /* Reference taken in
340 amd_iommu_bind_pasid */
342 /* Drop reference taken in amd_iommu_bind_pasid */
343 put_device_state(dev_state);
346 if (dev_state->pasid_levels == 2)
347 free_pasid_states_level2(dev_state->states);
348 else if (dev_state->pasid_levels == 1)
349 free_pasid_states_level1(dev_state->states);
351 BUG_ON(dev_state->pasid_levels != 0);
353 free_page((unsigned long)dev_state->states);
356 static struct pasid_state *mn_to_state(struct mmu_notifier *mn)
358 return container_of(mn, struct pasid_state, mn);
361 static void mn_invalidate_range(struct mmu_notifier *mn,
362 struct mm_struct *mm,
363 unsigned long start, unsigned long end)
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 if ((start ^ (end - 1)) < PAGE_SIZE)
372 amd_iommu_flush_page(dev_state->domain, pasid_state->pasid,
375 amd_iommu_flush_tlb(dev_state->domain, pasid_state->pasid);
378 static void mn_release(struct mmu_notifier *mn, struct mm_struct *mm)
380 struct pasid_state *pasid_state;
381 struct device_state *dev_state;
386 pasid_state = mn_to_state(mn);
387 dev_state = pasid_state->device_state;
388 run_inv_ctx_cb = !pasid_state->invalid;
390 if (run_inv_ctx_cb && dev_state->inv_ctx_cb)
391 dev_state->inv_ctx_cb(dev_state->pdev, pasid_state->pasid);
393 unbind_pasid(pasid_state);
396 static const struct mmu_notifier_ops iommu_mn = {
397 .release = mn_release,
398 .invalidate_range = mn_invalidate_range,
401 static void set_pri_tag_status(struct pasid_state *pasid_state,
406 spin_lock_irqsave(&pasid_state->lock, flags);
407 pasid_state->pri[tag].status = status;
408 spin_unlock_irqrestore(&pasid_state->lock, flags);
411 static void finish_pri_tag(struct device_state *dev_state,
412 struct pasid_state *pasid_state,
417 spin_lock_irqsave(&pasid_state->lock, flags);
418 if (atomic_dec_and_test(&pasid_state->pri[tag].inflight) &&
419 pasid_state->pri[tag].finish) {
420 amd_iommu_complete_ppr(dev_state->pdev, pasid_state->pasid,
421 pasid_state->pri[tag].status, tag);
422 pasid_state->pri[tag].finish = false;
423 pasid_state->pri[tag].status = PPR_SUCCESS;
425 spin_unlock_irqrestore(&pasid_state->lock, flags);
428 static void handle_fault_error(struct fault *fault)
432 if (!fault->dev_state->inv_ppr_cb) {
433 set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
437 status = fault->dev_state->inv_ppr_cb(fault->dev_state->pdev,
442 case AMD_IOMMU_INV_PRI_RSP_SUCCESS:
443 set_pri_tag_status(fault->state, fault->tag, PPR_SUCCESS);
445 case AMD_IOMMU_INV_PRI_RSP_INVALID:
446 set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
448 case AMD_IOMMU_INV_PRI_RSP_FAIL:
449 set_pri_tag_status(fault->state, fault->tag, PPR_FAILURE);
456 static bool access_error(struct vm_area_struct *vma, struct fault *fault)
458 unsigned long requested = 0;
460 if (fault->flags & PPR_FAULT_EXEC)
461 requested |= VM_EXEC;
463 if (fault->flags & PPR_FAULT_READ)
464 requested |= VM_READ;
466 if (fault->flags & PPR_FAULT_WRITE)
467 requested |= VM_WRITE;
469 return (requested & ~vma->vm_flags) != 0;
472 static void do_fault(struct work_struct *work)
474 struct fault *fault = container_of(work, struct fault, work);
475 struct vm_area_struct *vma;
476 vm_fault_t ret = VM_FAULT_ERROR;
477 unsigned int flags = 0;
478 struct mm_struct *mm;
481 mm = fault->state->mm;
482 address = fault->address;
484 if (fault->flags & PPR_FAULT_USER)
485 flags |= FAULT_FLAG_USER;
486 if (fault->flags & PPR_FAULT_WRITE)
487 flags |= FAULT_FLAG_WRITE;
488 flags |= FAULT_FLAG_REMOTE;
491 vma = find_extend_vma(mm, address);
492 if (!vma || address < vma->vm_start)
493 /* failed to get a vma in the right range */
496 /* Check if we have the right permissions on the vma */
497 if (access_error(vma, fault))
500 ret = handle_mm_fault(vma, address, flags, NULL);
502 mmap_read_unlock(mm);
504 if (ret & VM_FAULT_ERROR)
505 /* failed to service fault */
506 handle_fault_error(fault);
508 finish_pri_tag(fault->dev_state, fault->state, fault->tag);
510 put_pasid_state(fault->state);
515 static int ppr_notifier(struct notifier_block *nb, unsigned long e, void *data)
517 struct amd_iommu_fault *iommu_fault;
518 struct pasid_state *pasid_state;
519 struct device_state *dev_state;
520 struct pci_dev *pdev = NULL;
528 tag = iommu_fault->tag & 0x1ff;
529 finish = (iommu_fault->tag >> 9) & 1;
531 devid = iommu_fault->device_id;
532 pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
539 /* In kdump kernel pci dev is not initialized yet -> send INVALID */
540 if (amd_iommu_is_attach_deferred(NULL, &pdev->dev)) {
541 amd_iommu_complete_ppr(pdev, iommu_fault->pasid,
546 dev_state = get_device_state(iommu_fault->device_id);
547 if (dev_state == NULL)
550 pasid_state = get_pasid_state(dev_state, iommu_fault->pasid);
551 if (pasid_state == NULL || pasid_state->invalid) {
552 /* We know the device but not the PASID -> send INVALID */
553 amd_iommu_complete_ppr(dev_state->pdev, iommu_fault->pasid,
558 spin_lock_irqsave(&pasid_state->lock, flags);
559 atomic_inc(&pasid_state->pri[tag].inflight);
561 pasid_state->pri[tag].finish = true;
562 spin_unlock_irqrestore(&pasid_state->lock, flags);
564 fault = kzalloc(sizeof(*fault), GFP_ATOMIC);
566 /* We are OOM - send success and let the device re-fault */
567 finish_pri_tag(dev_state, pasid_state, tag);
571 fault->dev_state = dev_state;
572 fault->address = iommu_fault->address;
573 fault->state = pasid_state;
575 fault->finish = finish;
576 fault->pasid = iommu_fault->pasid;
577 fault->flags = iommu_fault->flags;
578 INIT_WORK(&fault->work, do_fault);
580 queue_work(iommu_wq, &fault->work);
586 if (ret != NOTIFY_OK && pasid_state)
587 put_pasid_state(pasid_state);
589 put_device_state(dev_state);
595 static struct notifier_block ppr_nb = {
596 .notifier_call = ppr_notifier,
599 int amd_iommu_bind_pasid(struct pci_dev *pdev, u32 pasid,
600 struct task_struct *task)
602 struct pasid_state *pasid_state;
603 struct device_state *dev_state;
604 struct mm_struct *mm;
610 if (!amd_iommu_v2_supported())
613 devid = device_id(pdev);
614 dev_state = get_device_state(devid);
616 if (dev_state == NULL)
620 if (pasid >= dev_state->max_pasids)
624 pasid_state = kzalloc(sizeof(*pasid_state), GFP_KERNEL);
625 if (pasid_state == NULL)
629 refcount_set(&pasid_state->count, 1);
630 init_waitqueue_head(&pasid_state->wq);
631 spin_lock_init(&pasid_state->lock);
633 mm = get_task_mm(task);
634 pasid_state->mm = mm;
635 pasid_state->device_state = dev_state;
636 pasid_state->pasid = pasid;
637 pasid_state->invalid = true; /* Mark as valid only if we are
638 done with setting up the pasid */
639 pasid_state->mn.ops = &iommu_mn;
641 if (pasid_state->mm == NULL)
644 mmu_notifier_register(&pasid_state->mn, mm);
646 ret = set_pasid_state(dev_state, pasid_state, pasid);
650 ret = amd_iommu_domain_set_gcr3(dev_state->domain, pasid,
651 __pa(pasid_state->mm->pgd));
653 goto out_clear_state;
655 /* Now we are ready to handle faults */
656 pasid_state->invalid = false;
659 * Drop the reference to the mm_struct here. We rely on the
660 * mmu_notifier release call-back to inform us when the mm
668 clear_pasid_state(dev_state, pasid);
671 mmu_notifier_unregister(&pasid_state->mn, mm);
675 free_pasid_state(pasid_state);
678 put_device_state(dev_state);
682 EXPORT_SYMBOL(amd_iommu_bind_pasid);
684 void amd_iommu_unbind_pasid(struct pci_dev *pdev, u32 pasid)
686 struct pasid_state *pasid_state;
687 struct device_state *dev_state;
692 if (!amd_iommu_v2_supported())
695 devid = device_id(pdev);
696 dev_state = get_device_state(devid);
697 if (dev_state == NULL)
700 if (pasid >= dev_state->max_pasids)
703 pasid_state = get_pasid_state(dev_state, pasid);
704 if (pasid_state == NULL)
707 * Drop reference taken here. We are safe because we still hold
708 * the reference taken in the amd_iommu_bind_pasid function.
710 put_pasid_state(pasid_state);
712 /* Clear the pasid state so that the pasid can be re-used */
713 clear_pasid_state(dev_state, pasid_state->pasid);
716 * Call mmu_notifier_unregister to drop our reference
719 mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
721 put_pasid_state_wait(pasid_state); /* Reference taken in
722 amd_iommu_bind_pasid */
724 /* Drop reference taken in this function */
725 put_device_state(dev_state);
727 /* Drop reference taken in amd_iommu_bind_pasid */
728 put_device_state(dev_state);
730 EXPORT_SYMBOL(amd_iommu_unbind_pasid);
732 int amd_iommu_init_device(struct pci_dev *pdev, int pasids)
734 struct device_state *dev_state;
735 struct iommu_group *group;
743 * When memory encryption is active the device is likely not in a
744 * direct-mapped domain. Forbid using IOMMUv2 functionality for now.
746 if (cc_platform_has(CC_ATTR_MEM_ENCRYPT))
749 if (!amd_iommu_v2_supported())
752 if (pasids <= 0 || pasids > (PASID_MASK + 1))
755 devid = device_id(pdev);
757 dev_state = kzalloc(sizeof(*dev_state), GFP_KERNEL);
758 if (dev_state == NULL)
761 spin_lock_init(&dev_state->lock);
762 init_waitqueue_head(&dev_state->wq);
763 dev_state->pdev = pdev;
764 dev_state->devid = devid;
767 for (dev_state->pasid_levels = 0; (tmp - 1) & ~0x1ff; tmp >>= 9)
768 dev_state->pasid_levels += 1;
770 atomic_set(&dev_state->count, 1);
771 dev_state->max_pasids = pasids;
774 dev_state->states = (void *)get_zeroed_page(GFP_KERNEL);
775 if (dev_state->states == NULL)
776 goto out_free_dev_state;
778 dev_state->domain = iommu_domain_alloc(&pci_bus_type);
779 if (dev_state->domain == NULL)
780 goto out_free_states;
782 amd_iommu_domain_direct_map(dev_state->domain);
784 ret = amd_iommu_domain_enable_v2(dev_state->domain, pasids);
786 goto out_free_domain;
788 group = iommu_group_get(&pdev->dev);
791 goto out_free_domain;
794 ret = iommu_attach_group(dev_state->domain, group);
798 iommu_group_put(group);
800 spin_lock_irqsave(&state_lock, flags);
802 if (__get_device_state(devid) != NULL) {
803 spin_unlock_irqrestore(&state_lock, flags);
805 goto out_free_domain;
808 list_add_tail(&dev_state->list, &state_list);
810 spin_unlock_irqrestore(&state_lock, flags);
815 iommu_group_put(group);
818 iommu_domain_free(dev_state->domain);
821 free_page((unsigned long)dev_state->states);
828 EXPORT_SYMBOL(amd_iommu_init_device);
830 void amd_iommu_free_device(struct pci_dev *pdev)
832 struct device_state *dev_state;
836 if (!amd_iommu_v2_supported())
839 devid = device_id(pdev);
841 spin_lock_irqsave(&state_lock, flags);
843 dev_state = __get_device_state(devid);
844 if (dev_state == NULL) {
845 spin_unlock_irqrestore(&state_lock, flags);
849 list_del(&dev_state->list);
851 spin_unlock_irqrestore(&state_lock, flags);
853 /* Get rid of any remaining pasid states */
854 free_pasid_states(dev_state);
856 put_device_state(dev_state);
858 * Wait until the last reference is dropped before freeing
861 wait_event(dev_state->wq, !atomic_read(&dev_state->count));
862 free_device_state(dev_state);
864 EXPORT_SYMBOL(amd_iommu_free_device);
866 int amd_iommu_set_invalid_ppr_cb(struct pci_dev *pdev,
867 amd_iommu_invalid_ppr_cb cb)
869 struct device_state *dev_state;
874 if (!amd_iommu_v2_supported())
877 devid = device_id(pdev);
879 spin_lock_irqsave(&state_lock, flags);
882 dev_state = __get_device_state(devid);
883 if (dev_state == NULL)
886 dev_state->inv_ppr_cb = cb;
891 spin_unlock_irqrestore(&state_lock, flags);
895 EXPORT_SYMBOL(amd_iommu_set_invalid_ppr_cb);
897 int amd_iommu_set_invalidate_ctx_cb(struct pci_dev *pdev,
898 amd_iommu_invalidate_ctx cb)
900 struct device_state *dev_state;
905 if (!amd_iommu_v2_supported())
908 devid = device_id(pdev);
910 spin_lock_irqsave(&state_lock, flags);
913 dev_state = __get_device_state(devid);
914 if (dev_state == NULL)
917 dev_state->inv_ctx_cb = cb;
922 spin_unlock_irqrestore(&state_lock, flags);
926 EXPORT_SYMBOL(amd_iommu_set_invalidate_ctx_cb);
928 static int __init amd_iommu_v2_init(void)
932 pr_info("AMD IOMMUv2 driver by Joerg Roedel <jroedel@suse.de>\n");
934 if (!amd_iommu_v2_supported()) {
935 pr_info("AMD IOMMUv2 functionality not available on this system\n");
937 * Load anyway to provide the symbols to other modules
938 * which may use AMD IOMMUv2 optionally.
944 iommu_wq = alloc_workqueue("amd_iommu_v2", WQ_MEM_RECLAIM, 0);
945 if (iommu_wq == NULL)
948 amd_iommu_register_ppr_notifier(&ppr_nb);
956 static void __exit amd_iommu_v2_exit(void)
958 struct device_state *dev_state;
961 if (!amd_iommu_v2_supported())
964 amd_iommu_unregister_ppr_notifier(&ppr_nb);
966 flush_workqueue(iommu_wq);
969 * The loop below might call flush_workqueue(), so call
970 * destroy_workqueue() after it
972 for (i = 0; i < MAX_DEVICES; ++i) {
973 dev_state = get_device_state(i);
975 if (dev_state == NULL)
980 put_device_state(dev_state);
981 amd_iommu_free_device(dev_state->pdev);
984 destroy_workqueue(iommu_wq);
987 module_init(amd_iommu_v2_init);
988 module_exit(amd_iommu_v2_exit);