2 * Copyright © 2015 Intel Corporation.
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms and conditions of the GNU General Public License,
6 * version 2, as published by the Free Software Foundation.
8 * This program is distributed in the hope it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
13 * Authors: David Woodhouse <dwmw2@infradead.org>
16 #include <linux/intel-iommu.h>
17 #include <linux/mmu_notifier.h>
18 #include <linux/sched.h>
19 #include <linux/sched/mm.h>
20 #include <linux/slab.h>
21 #include <linux/intel-svm.h>
22 #include <linux/rculist.h>
23 #include <linux/pci.h>
24 #include <linux/pci-ats.h>
25 #include <linux/dmar.h>
26 #include <linux/interrupt.h>
29 static irqreturn_t
prq_event_thread(int irq
, void *d
);
35 struct pasid_state_entry
{
39 int intel_svm_alloc_pasid_tables(struct intel_iommu
*iommu
)
44 /* Start at 2 because it's defined as 2^(1+PSS) */
45 iommu
->pasid_max
= 2 << ecap_pss(iommu
->ecap
);
47 /* Eventually I'm promised we will get a multi-level PASID table
48 * and it won't have to be physically contiguous. Until then,
49 * limit the size because 8MiB contiguous allocations can be hard
50 * to come by. The limit of 0x20000, which is 1MiB for each of
51 * the PASID and PASID-state tables, is somewhat arbitrary. */
52 if (iommu
->pasid_max
> 0x20000)
53 iommu
->pasid_max
= 0x20000;
55 order
= get_order(sizeof(struct pasid_entry
) * iommu
->pasid_max
);
56 pages
= alloc_pages(GFP_KERNEL
| __GFP_ZERO
, order
);
58 pr_warn("IOMMU: %s: Failed to allocate PASID table\n",
62 iommu
->pasid_table
= page_address(pages
);
63 pr_info("%s: Allocated order %d PASID table.\n", iommu
->name
, order
);
65 if (ecap_dis(iommu
->ecap
)) {
66 /* Just making it explicit... */
67 BUILD_BUG_ON(sizeof(struct pasid_entry
) != sizeof(struct pasid_state_entry
));
68 pages
= alloc_pages(GFP_KERNEL
| __GFP_ZERO
, order
);
70 iommu
->pasid_state_table
= page_address(pages
);
72 pr_warn("IOMMU: %s: Failed to allocate PASID state table\n",
76 idr_init(&iommu
->pasid_idr
);
81 int intel_svm_free_pasid_tables(struct intel_iommu
*iommu
)
83 int order
= get_order(sizeof(struct pasid_entry
) * iommu
->pasid_max
);
85 if (iommu
->pasid_table
) {
86 free_pages((unsigned long)iommu
->pasid_table
, order
);
87 iommu
->pasid_table
= NULL
;
89 if (iommu
->pasid_state_table
) {
90 free_pages((unsigned long)iommu
->pasid_state_table
, order
);
91 iommu
->pasid_state_table
= NULL
;
93 idr_destroy(&iommu
->pasid_idr
);
99 int intel_svm_enable_prq(struct intel_iommu
*iommu
)
104 pages
= alloc_pages(GFP_KERNEL
| __GFP_ZERO
, PRQ_ORDER
);
106 pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
110 iommu
->prq
= page_address(pages
);
112 irq
= dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED
+ iommu
->seq_id
, iommu
->node
, iommu
);
114 pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
118 free_pages((unsigned long)iommu
->prq
, PRQ_ORDER
);
124 snprintf(iommu
->prq_name
, sizeof(iommu
->prq_name
), "dmar%d-prq", iommu
->seq_id
);
126 ret
= request_threaded_irq(irq
, NULL
, prq_event_thread
, IRQF_ONESHOT
,
127 iommu
->prq_name
, iommu
);
129 pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
131 dmar_free_hwirq(irq
);
134 dmar_writeq(iommu
->reg
+ DMAR_PQH_REG
, 0ULL);
135 dmar_writeq(iommu
->reg
+ DMAR_PQT_REG
, 0ULL);
136 dmar_writeq(iommu
->reg
+ DMAR_PQA_REG
, virt_to_phys(iommu
->prq
) | PRQ_ORDER
);
141 int intel_svm_finish_prq(struct intel_iommu
*iommu
)
143 dmar_writeq(iommu
->reg
+ DMAR_PQH_REG
, 0ULL);
144 dmar_writeq(iommu
->reg
+ DMAR_PQT_REG
, 0ULL);
145 dmar_writeq(iommu
->reg
+ DMAR_PQA_REG
, 0ULL);
147 free_irq(iommu
->pr_irq
, iommu
);
148 dmar_free_hwirq(iommu
->pr_irq
);
151 free_pages((unsigned long)iommu
->prq
, PRQ_ORDER
);
157 static void intel_flush_svm_range_dev (struct intel_svm
*svm
, struct intel_svm_dev
*sdev
,
158 unsigned long address
, unsigned long pages
, int ih
, int gl
)
163 /* For global kernel pages we have to flush them in *all* PASIDs
164 * because that's the only option the hardware gives us. Despite
165 * the fact that they are actually only accessible through one. */
167 desc
.low
= QI_EIOTLB_PASID(svm
->pasid
) | QI_EIOTLB_DID(sdev
->did
) |
168 QI_EIOTLB_GRAN(QI_GRAN_ALL_ALL
) | QI_EIOTLB_TYPE
;
170 desc
.low
= QI_EIOTLB_PASID(svm
->pasid
) | QI_EIOTLB_DID(sdev
->did
) |
171 QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID
) | QI_EIOTLB_TYPE
;
174 int mask
= ilog2(__roundup_pow_of_two(pages
));
176 desc
.low
= QI_EIOTLB_PASID(svm
->pasid
) | QI_EIOTLB_DID(sdev
->did
) |
177 QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID
) | QI_EIOTLB_TYPE
;
178 desc
.high
= QI_EIOTLB_ADDR(address
) | QI_EIOTLB_GL(gl
) |
179 QI_EIOTLB_IH(ih
) | QI_EIOTLB_AM(mask
);
181 qi_submit_sync(&desc
, svm
->iommu
);
183 if (sdev
->dev_iotlb
) {
184 desc
.low
= QI_DEV_EIOTLB_PASID(svm
->pasid
) | QI_DEV_EIOTLB_SID(sdev
->sid
) |
185 QI_DEV_EIOTLB_QDEP(sdev
->qdep
) | QI_DEIOTLB_TYPE
;
187 desc
.high
= QI_DEV_EIOTLB_ADDR(-1ULL >> 1) | QI_DEV_EIOTLB_SIZE
;
188 } else if (pages
> 1) {
189 /* The least significant zero bit indicates the size. So,
190 * for example, an "address" value of 0x12345f000 will
191 * flush from 0x123440000 to 0x12347ffff (256KiB). */
192 unsigned long last
= address
+ ((unsigned long)(pages
- 1) << VTD_PAGE_SHIFT
);
193 unsigned long mask
= __rounddown_pow_of_two(address
^ last
);;
195 desc
.high
= QI_DEV_EIOTLB_ADDR((address
& ~mask
) | (mask
- 1)) | QI_DEV_EIOTLB_SIZE
;
197 desc
.high
= QI_DEV_EIOTLB_ADDR(address
);
199 qi_submit_sync(&desc
, svm
->iommu
);
203 static void intel_flush_svm_range(struct intel_svm
*svm
, unsigned long address
,
204 unsigned long pages
, int ih
, int gl
)
206 struct intel_svm_dev
*sdev
;
208 /* Try deferred invalidate if available */
209 if (svm
->iommu
->pasid_state_table
&&
210 !cmpxchg64(&svm
->iommu
->pasid_state_table
[svm
->pasid
].val
, 0, 1ULL << 63))
214 list_for_each_entry_rcu(sdev
, &svm
->devs
, list
)
215 intel_flush_svm_range_dev(svm
, sdev
, address
, pages
, ih
, gl
);
219 static void intel_change_pte(struct mmu_notifier
*mn
, struct mm_struct
*mm
,
220 unsigned long address
, pte_t pte
)
222 struct intel_svm
*svm
= container_of(mn
, struct intel_svm
, notifier
);
224 intel_flush_svm_range(svm
, address
, 1, 1, 0);
227 /* Pages have been freed at this point */
228 static void intel_invalidate_range(struct mmu_notifier
*mn
,
229 struct mm_struct
*mm
,
230 unsigned long start
, unsigned long end
)
232 struct intel_svm
*svm
= container_of(mn
, struct intel_svm
, notifier
);
234 intel_flush_svm_range(svm
, start
,
235 (end
- start
+ PAGE_SIZE
- 1) >> VTD_PAGE_SHIFT
, 0, 0);
239 static void intel_flush_pasid_dev(struct intel_svm
*svm
, struct intel_svm_dev
*sdev
, int pasid
)
244 desc
.low
= QI_PC_TYPE
| QI_PC_DID(sdev
->did
) | QI_PC_PASID_SEL
| QI_PC_PASID(pasid
);
246 qi_submit_sync(&desc
, svm
->iommu
);
249 static void intel_mm_release(struct mmu_notifier
*mn
, struct mm_struct
*mm
)
251 struct intel_svm
*svm
= container_of(mn
, struct intel_svm
, notifier
);
252 struct intel_svm_dev
*sdev
;
254 /* This might end up being called from exit_mmap(), *before* the page
255 * tables are cleared. And __mmu_notifier_release() will delete us from
256 * the list of notifiers so that our invalidate_range() callback doesn't
257 * get called when the page tables are cleared. So we need to protect
258 * against hardware accessing those page tables.
260 * We do it by clearing the entry in the PASID table and then flushing
261 * the IOTLB and the PASID table caches. This might upset hardware;
262 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
263 * page) so that we end up taking a fault that the hardware really
264 * *has* to handle gracefully without affecting other processes.
266 svm
->iommu
->pasid_table
[svm
->pasid
].val
= 0;
270 list_for_each_entry_rcu(sdev
, &svm
->devs
, list
) {
271 intel_flush_pasid_dev(svm
, sdev
, svm
->pasid
);
272 intel_flush_svm_range_dev(svm
, sdev
, 0, -1, 0, !svm
->mm
);
278 static const struct mmu_notifier_ops intel_mmuops
= {
279 .release
= intel_mm_release
,
280 .change_pte
= intel_change_pte
,
281 .invalidate_range
= intel_invalidate_range
,
284 static DEFINE_MUTEX(pasid_mutex
);
286 int intel_svm_bind_mm(struct device
*dev
, int *pasid
, int flags
, struct svm_dev_ops
*ops
)
288 struct intel_iommu
*iommu
= intel_svm_device_to_iommu(dev
);
289 struct intel_svm_dev
*sdev
;
290 struct intel_svm
*svm
= NULL
;
291 struct mm_struct
*mm
= NULL
;
295 if (WARN_ON(!iommu
|| !iommu
->pasid_table
))
298 if (dev_is_pci(dev
)) {
299 pasid_max
= pci_max_pasids(to_pci_dev(dev
));
305 if ((flags
& SVM_FLAG_SUPERVISOR_MODE
)) {
306 if (!ecap_srs(iommu
->ecap
))
309 mm
= get_task_mm(current
);
313 mutex_lock(&pasid_mutex
);
314 if (pasid
&& !(flags
& SVM_FLAG_PRIVATE_PASID
)) {
317 idr_for_each_entry(&iommu
->pasid_idr
, svm
, i
) {
319 (svm
->flags
& SVM_FLAG_PRIVATE_PASID
))
322 if (svm
->pasid
>= pasid_max
) {
324 "Limited PASID width. Cannot use existing PASID %d\n",
330 list_for_each_entry(sdev
, &svm
->devs
, list
) {
331 if (dev
== sdev
->dev
) {
332 if (sdev
->ops
!= ops
) {
345 sdev
= kzalloc(sizeof(*sdev
), GFP_KERNEL
);
352 ret
= intel_iommu_enable_pasid(iommu
, sdev
);
354 /* If they don't actually want to assign a PASID, this is
355 * just an enabling check/preparation. */
359 /* Finish the setup now we know we're keeping it */
362 init_rcu_head(&sdev
->rcu
);
365 svm
= kzalloc(sizeof(*svm
), GFP_KERNEL
);
373 if (pasid_max
> iommu
->pasid_max
)
374 pasid_max
= iommu
->pasid_max
;
376 /* Do not use PASID 0 in caching mode (virtualised IOMMU) */
377 ret
= idr_alloc(&iommu
->pasid_idr
, svm
,
378 !!cap_caching_mode(iommu
->cap
),
379 pasid_max
- 1, GFP_KERNEL
);
385 svm
->notifier
.ops
= &intel_mmuops
;
388 INIT_LIST_HEAD_RCU(&svm
->devs
);
391 ret
= mmu_notifier_register(&svm
->notifier
, mm
);
393 idr_remove(&svm
->iommu
->pasid_idr
, svm
->pasid
);
398 iommu
->pasid_table
[svm
->pasid
].val
= (u64
)__pa(mm
->pgd
) | 1;
400 iommu
->pasid_table
[svm
->pasid
].val
= (u64
)__pa(init_mm
.pgd
) | 1 | (1ULL << 11);
402 /* In caching mode, we still have to flush with PASID 0 when
403 * a PASID table entry becomes present. Not entirely clear
404 * *why* that would be the case — surely we could just issue
405 * a flush with the PASID value that we've changed? The PASID
406 * is the index into the table, after all. It's not like domain
407 * IDs in the case of the equivalent context-entry change in
408 * caching mode. And for that matter it's not entirely clear why
409 * a VMM would be in the business of caching the PASID table
410 * anyway. Surely that can be left entirely to the guest? */
411 if (cap_caching_mode(iommu
->cap
))
412 intel_flush_pasid_dev(svm
, sdev
, 0);
414 list_add_rcu(&sdev
->list
, &svm
->devs
);
420 mutex_unlock(&pasid_mutex
);
425 EXPORT_SYMBOL_GPL(intel_svm_bind_mm
);
427 int intel_svm_unbind_mm(struct device
*dev
, int pasid
)
429 struct intel_svm_dev
*sdev
;
430 struct intel_iommu
*iommu
;
431 struct intel_svm
*svm
;
434 mutex_lock(&pasid_mutex
);
435 iommu
= intel_svm_device_to_iommu(dev
);
436 if (!iommu
|| !iommu
->pasid_table
)
439 svm
= idr_find(&iommu
->pasid_idr
, pasid
);
443 list_for_each_entry(sdev
, &svm
->devs
, list
) {
444 if (dev
== sdev
->dev
) {
448 list_del_rcu(&sdev
->list
);
449 /* Flush the PASID cache and IOTLB for this device.
450 * Note that we do depend on the hardware *not* using
451 * the PASID any more. Just as we depend on other
452 * devices never using PASIDs that they have no right
453 * to use. We have a *shared* PASID table, because it's
454 * large and has to be physically contiguous. So it's
455 * hard to be as defensive as we might like. */
456 intel_flush_pasid_dev(svm
, sdev
, svm
->pasid
);
457 intel_flush_svm_range_dev(svm
, sdev
, 0, -1, 0, !svm
->mm
);
458 kfree_rcu(sdev
, rcu
);
460 if (list_empty(&svm
->devs
)) {
461 svm
->iommu
->pasid_table
[svm
->pasid
].val
= 0;
464 idr_remove(&svm
->iommu
->pasid_idr
, svm
->pasid
);
466 mmu_notifier_unregister(&svm
->notifier
, svm
->mm
);
468 /* We mandate that no page faults may be outstanding
469 * for the PASID when intel_svm_unbind_mm() is called.
470 * If that is not obeyed, subtle errors will happen.
471 * Let's make them less subtle... */
472 memset(svm
, 0x6b, sizeof(*svm
));
480 mutex_unlock(&pasid_mutex
);
484 EXPORT_SYMBOL_GPL(intel_svm_unbind_mm
);
486 int intel_svm_is_pasid_valid(struct device
*dev
, int pasid
)
488 struct intel_iommu
*iommu
;
489 struct intel_svm
*svm
;
492 mutex_lock(&pasid_mutex
);
493 iommu
= intel_svm_device_to_iommu(dev
);
494 if (!iommu
|| !iommu
->pasid_table
)
497 svm
= idr_find(&iommu
->pasid_idr
, pasid
);
501 /* init_mm is used in this case */
504 else if (atomic_read(&svm
->mm
->mm_users
) > 0)
510 mutex_unlock(&pasid_mutex
);
514 EXPORT_SYMBOL_GPL(intel_svm_is_pasid_valid
);
516 /* Page request queue descriptor */
517 struct page_req_dsc
{
534 #define PRQ_RING_MASK ((0x1000 << PRQ_ORDER) - 0x10)
536 static bool access_error(struct vm_area_struct
*vma
, struct page_req_dsc
*req
)
538 unsigned long requested
= 0;
541 requested
|= VM_EXEC
;
544 requested
|= VM_READ
;
547 requested
|= VM_WRITE
;
549 return (requested
& ~vma
->vm_flags
) != 0;
552 static bool is_canonical_address(u64 addr
)
554 int shift
= 64 - (__VIRTUAL_MASK_SHIFT
+ 1);
555 long saddr
= (long) addr
;
557 return (((saddr
<< shift
) >> shift
) == saddr
);
560 static irqreturn_t
prq_event_thread(int irq
, void *d
)
562 struct intel_iommu
*iommu
= d
;
563 struct intel_svm
*svm
= NULL
;
564 int head
, tail
, handled
= 0;
566 /* Clear PPR bit before reading head/tail registers, to
567 * ensure that we get a new interrupt if needed. */
568 writel(DMA_PRS_PPR
, iommu
->reg
+ DMAR_PRS_REG
);
570 tail
= dmar_readq(iommu
->reg
+ DMAR_PQT_REG
) & PRQ_RING_MASK
;
571 head
= dmar_readq(iommu
->reg
+ DMAR_PQH_REG
) & PRQ_RING_MASK
;
572 while (head
!= tail
) {
573 struct intel_svm_dev
*sdev
;
574 struct vm_area_struct
*vma
;
575 struct page_req_dsc
*req
;
582 req
= &iommu
->prq
[head
/ sizeof(*req
)];
584 result
= QI_RESP_FAILURE
;
585 address
= (u64
)req
->addr
<< VTD_PAGE_SHIFT
;
586 if (!req
->pasid_present
) {
587 pr_err("%s: Page request without PASID: %08llx %08llx\n",
588 iommu
->name
, ((unsigned long long *)req
)[0],
589 ((unsigned long long *)req
)[1]);
593 if (!svm
|| svm
->pasid
!= req
->pasid
) {
595 svm
= idr_find(&iommu
->pasid_idr
, req
->pasid
);
596 /* It *can't* go away, because the driver is not permitted
597 * to unbind the mm while any page faults are outstanding.
598 * So we only need RCU to protect the internal idr code. */
602 pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
603 iommu
->name
, req
->pasid
, ((unsigned long long *)req
)[0],
604 ((unsigned long long *)req
)[1]);
609 result
= QI_RESP_INVALID
;
610 /* Since we're using init_mm.pgd directly, we should never take
611 * any faults on kernel addresses. */
614 /* If the mm is already defunct, don't handle faults. */
615 if (!mmget_not_zero(svm
->mm
))
618 /* If address is not canonical, return invalid response */
619 if (!is_canonical_address(address
))
622 down_read(&svm
->mm
->mmap_sem
);
623 vma
= find_extend_vma(svm
->mm
, address
);
624 if (!vma
|| address
< vma
->vm_start
)
627 if (access_error(vma
, req
))
630 ret
= handle_mm_fault(vma
, address
,
631 req
->wr_req
? FAULT_FLAG_WRITE
: 0);
632 if (ret
& VM_FAULT_ERROR
)
635 result
= QI_RESP_SUCCESS
;
637 up_read(&svm
->mm
->mmap_sem
);
640 /* Accounting for major/minor faults? */
642 list_for_each_entry_rcu(sdev
, &svm
->devs
, list
) {
643 if (sdev
->sid
== PCI_DEVID(req
->bus
, req
->devfn
))
646 /* Other devices can go away, but the drivers are not permitted
647 * to unbind while any page faults might be in flight. So it's
648 * OK to drop the 'lock' here now we have it. */
651 if (WARN_ON(&sdev
->list
== &svm
->devs
))
654 if (sdev
&& sdev
->ops
&& sdev
->ops
->fault_cb
) {
655 int rwxp
= (req
->rd_req
<< 3) | (req
->wr_req
<< 2) |
656 (req
->exe_req
<< 1) | (req
->priv_req
);
657 sdev
->ops
->fault_cb(sdev
->dev
, req
->pasid
, req
->addr
, req
->private, rwxp
, result
);
659 /* We get here in the error case where the PASID lookup failed,
660 and these can be NULL. Do not use them below this point! */
665 /* Page Group Response */
666 resp
.low
= QI_PGRP_PASID(req
->pasid
) |
667 QI_PGRP_DID((req
->bus
<< 8) | req
->devfn
) |
668 QI_PGRP_PASID_P(req
->pasid_present
) |
670 resp
.high
= QI_PGRP_IDX(req
->prg_index
) |
671 QI_PGRP_PRIV(req
->private) | QI_PGRP_RESP_CODE(result
);
673 qi_submit_sync(&resp
, iommu
);
674 } else if (req
->srr
) {
675 /* Page Stream Response */
676 resp
.low
= QI_PSTRM_IDX(req
->prg_index
) |
677 QI_PSTRM_PRIV(req
->private) | QI_PSTRM_BUS(req
->bus
) |
678 QI_PSTRM_PASID(req
->pasid
) | QI_PSTRM_RESP_TYPE
;
679 resp
.high
= QI_PSTRM_ADDR(address
) | QI_PSTRM_DEVFN(req
->devfn
) |
680 QI_PSTRM_RESP_CODE(result
);
682 qi_submit_sync(&resp
, iommu
);
685 head
= (head
+ sizeof(*req
)) & PRQ_RING_MASK
;
688 dmar_writeq(iommu
->reg
+ DMAR_PQH_REG
, tail
);
690 return IRQ_RETVAL(handled
);