Loading...
Note: File does not exist in v5.9.
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright © 2015 Intel Corporation.
4 *
5 * Authors: David Woodhouse <dwmw2@infradead.org>
6 */
7
8#include <linux/intel-iommu.h>
9#include <linux/mmu_notifier.h>
10#include <linux/sched.h>
11#include <linux/sched/mm.h>
12#include <linux/slab.h>
13#include <linux/intel-svm.h>
14#include <linux/rculist.h>
15#include <linux/pci.h>
16#include <linux/pci-ats.h>
17#include <linux/dmar.h>
18#include <linux/interrupt.h>
19#include <linux/mm_types.h>
20#include <asm/page.h>
21
22#include "intel-pasid.h"
23
24static irqreturn_t prq_event_thread(int irq, void *d);
25
26int intel_svm_init(struct intel_iommu *iommu)
27{
28 if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
29 !cap_fl1gp_support(iommu->cap))
30 return -EINVAL;
31
32 if (cpu_feature_enabled(X86_FEATURE_LA57) &&
33 !cap_5lp_support(iommu->cap))
34 return -EINVAL;
35
36 return 0;
37}
38
39#define PRQ_ORDER 0
40
41int intel_svm_enable_prq(struct intel_iommu *iommu)
42{
43 struct page *pages;
44 int irq, ret;
45
46 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
47 if (!pages) {
48 pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
49 iommu->name);
50 return -ENOMEM;
51 }
52 iommu->prq = page_address(pages);
53
54 irq = dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED + iommu->seq_id, iommu->node, iommu);
55 if (irq <= 0) {
56 pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
57 iommu->name);
58 ret = -EINVAL;
59 err:
60 free_pages((unsigned long)iommu->prq, PRQ_ORDER);
61 iommu->prq = NULL;
62 return ret;
63 }
64 iommu->pr_irq = irq;
65
66 snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
67
68 ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
69 iommu->prq_name, iommu);
70 if (ret) {
71 pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
72 iommu->name);
73 dmar_free_hwirq(irq);
74 iommu->pr_irq = 0;
75 goto err;
76 }
77 dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
78 dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
79 dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
80
81 return 0;
82}
83
84int intel_svm_finish_prq(struct intel_iommu *iommu)
85{
86 dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
87 dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
88 dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
89
90 if (iommu->pr_irq) {
91 free_irq(iommu->pr_irq, iommu);
92 dmar_free_hwirq(iommu->pr_irq);
93 iommu->pr_irq = 0;
94 }
95
96 free_pages((unsigned long)iommu->prq, PRQ_ORDER);
97 iommu->prq = NULL;
98
99 return 0;
100}
101
102static void intel_flush_svm_range_dev (struct intel_svm *svm, struct intel_svm_dev *sdev,
103 unsigned long address, unsigned long pages, int ih)
104{
105 struct qi_desc desc;
106
107 /*
108 * Do PASID granu IOTLB invalidation if page selective capability is
109 * not available.
110 */
111 if (pages == -1 || !cap_pgsel_inv(svm->iommu->cap)) {
112 desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
113 QI_EIOTLB_DID(sdev->did) |
114 QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
115 QI_EIOTLB_TYPE;
116 desc.qw1 = 0;
117 } else {
118 int mask = ilog2(__roundup_pow_of_two(pages));
119
120 desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
121 QI_EIOTLB_DID(sdev->did) |
122 QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID) |
123 QI_EIOTLB_TYPE;
124 desc.qw1 = QI_EIOTLB_ADDR(address) |
125 QI_EIOTLB_IH(ih) |
126 QI_EIOTLB_AM(mask);
127 }
128 desc.qw2 = 0;
129 desc.qw3 = 0;
130 qi_submit_sync(&desc, svm->iommu);
131
132 if (sdev->dev_iotlb) {
133 desc.qw0 = QI_DEV_EIOTLB_PASID(svm->pasid) |
134 QI_DEV_EIOTLB_SID(sdev->sid) |
135 QI_DEV_EIOTLB_QDEP(sdev->qdep) |
136 QI_DEIOTLB_TYPE;
137 if (pages == -1) {
138 desc.qw1 = QI_DEV_EIOTLB_ADDR(-1ULL >> 1) |
139 QI_DEV_EIOTLB_SIZE;
140 } else if (pages > 1) {
141 /* The least significant zero bit indicates the size. So,
142 * for example, an "address" value of 0x12345f000 will
143 * flush from 0x123440000 to 0x12347ffff (256KiB). */
144 unsigned long last = address + ((unsigned long)(pages - 1) << VTD_PAGE_SHIFT);
145 unsigned long mask = __rounddown_pow_of_two(address ^ last);
146
147 desc.qw1 = QI_DEV_EIOTLB_ADDR((address & ~mask) |
148 (mask - 1)) | QI_DEV_EIOTLB_SIZE;
149 } else {
150 desc.qw1 = QI_DEV_EIOTLB_ADDR(address);
151 }
152 desc.qw2 = 0;
153 desc.qw3 = 0;
154 qi_submit_sync(&desc, svm->iommu);
155 }
156}
157
158static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
159 unsigned long pages, int ih)
160{
161 struct intel_svm_dev *sdev;
162
163 rcu_read_lock();
164 list_for_each_entry_rcu(sdev, &svm->devs, list)
165 intel_flush_svm_range_dev(svm, sdev, address, pages, ih);
166 rcu_read_unlock();
167}
168
169/* Pages have been freed at this point */
170static void intel_invalidate_range(struct mmu_notifier *mn,
171 struct mm_struct *mm,
172 unsigned long start, unsigned long end)
173{
174 struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
175
176 intel_flush_svm_range(svm, start,
177 (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0);
178}
179
180static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
181{
182 struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
183 struct intel_svm_dev *sdev;
184
185 /* This might end up being called from exit_mmap(), *before* the page
186 * tables are cleared. And __mmu_notifier_release() will delete us from
187 * the list of notifiers so that our invalidate_range() callback doesn't
188 * get called when the page tables are cleared. So we need to protect
189 * against hardware accessing those page tables.
190 *
191 * We do it by clearing the entry in the PASID table and then flushing
192 * the IOTLB and the PASID table caches. This might upset hardware;
193 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
194 * page) so that we end up taking a fault that the hardware really
195 * *has* to handle gracefully without affecting other processes.
196 */
197 rcu_read_lock();
198 list_for_each_entry_rcu(sdev, &svm->devs, list) {
199 intel_pasid_tear_down_entry(svm->iommu, sdev->dev, svm->pasid);
200 intel_flush_svm_range_dev(svm, sdev, 0, -1, 0);
201 }
202 rcu_read_unlock();
203
204}
205
206static const struct mmu_notifier_ops intel_mmuops = {
207 .release = intel_mm_release,
208 .invalidate_range = intel_invalidate_range,
209};
210
211static DEFINE_MUTEX(pasid_mutex);
212static LIST_HEAD(global_svm_list);
213
214int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_ops *ops)
215{
216 struct intel_iommu *iommu = intel_svm_device_to_iommu(dev);
217 struct device_domain_info *info;
218 struct intel_svm_dev *sdev;
219 struct intel_svm *svm = NULL;
220 struct mm_struct *mm = NULL;
221 int pasid_max;
222 int ret;
223
224 if (!iommu || dmar_disabled)
225 return -EINVAL;
226
227 if (dev_is_pci(dev)) {
228 pasid_max = pci_max_pasids(to_pci_dev(dev));
229 if (pasid_max < 0)
230 return -EINVAL;
231 } else
232 pasid_max = 1 << 20;
233
234 if (flags & SVM_FLAG_SUPERVISOR_MODE) {
235 if (!ecap_srs(iommu->ecap))
236 return -EINVAL;
237 } else if (pasid) {
238 mm = get_task_mm(current);
239 BUG_ON(!mm);
240 }
241
242 mutex_lock(&pasid_mutex);
243 if (pasid && !(flags & SVM_FLAG_PRIVATE_PASID)) {
244 struct intel_svm *t;
245
246 list_for_each_entry(t, &global_svm_list, list) {
247 if (t->mm != mm || (t->flags & SVM_FLAG_PRIVATE_PASID))
248 continue;
249
250 svm = t;
251 if (svm->pasid >= pasid_max) {
252 dev_warn(dev,
253 "Limited PASID width. Cannot use existing PASID %d\n",
254 svm->pasid);
255 ret = -ENOSPC;
256 goto out;
257 }
258
259 list_for_each_entry(sdev, &svm->devs, list) {
260 if (dev == sdev->dev) {
261 if (sdev->ops != ops) {
262 ret = -EBUSY;
263 goto out;
264 }
265 sdev->users++;
266 goto success;
267 }
268 }
269
270 break;
271 }
272 }
273
274 sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
275 if (!sdev) {
276 ret = -ENOMEM;
277 goto out;
278 }
279 sdev->dev = dev;
280
281 ret = intel_iommu_enable_pasid(iommu, dev);
282 if (ret || !pasid) {
283 /* If they don't actually want to assign a PASID, this is
284 * just an enabling check/preparation. */
285 kfree(sdev);
286 goto out;
287 }
288
289 info = dev->archdata.iommu;
290 if (!info || !info->pasid_supported) {
291 kfree(sdev);
292 goto out;
293 }
294
295 sdev->did = FLPT_DEFAULT_DID;
296 sdev->sid = PCI_DEVID(info->bus, info->devfn);
297 if (info->ats_enabled) {
298 sdev->dev_iotlb = 1;
299 sdev->qdep = info->ats_qdep;
300 if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
301 sdev->qdep = 0;
302 }
303
304 /* Finish the setup now we know we're keeping it */
305 sdev->users = 1;
306 sdev->ops = ops;
307 init_rcu_head(&sdev->rcu);
308
309 if (!svm) {
310 svm = kzalloc(sizeof(*svm), GFP_KERNEL);
311 if (!svm) {
312 ret = -ENOMEM;
313 kfree(sdev);
314 goto out;
315 }
316 svm->iommu = iommu;
317
318 if (pasid_max > intel_pasid_max_id)
319 pasid_max = intel_pasid_max_id;
320
321 /* Do not use PASID 0 in caching mode (virtualised IOMMU) */
322 ret = intel_pasid_alloc_id(svm,
323 !!cap_caching_mode(iommu->cap),
324 pasid_max - 1, GFP_KERNEL);
325 if (ret < 0) {
326 kfree(svm);
327 kfree(sdev);
328 goto out;
329 }
330 svm->pasid = ret;
331 svm->notifier.ops = &intel_mmuops;
332 svm->mm = mm;
333 svm->flags = flags;
334 INIT_LIST_HEAD_RCU(&svm->devs);
335 INIT_LIST_HEAD(&svm->list);
336 ret = -ENOMEM;
337 if (mm) {
338 ret = mmu_notifier_register(&svm->notifier, mm);
339 if (ret) {
340 intel_pasid_free_id(svm->pasid);
341 kfree(svm);
342 kfree(sdev);
343 goto out;
344 }
345 }
346
347 spin_lock(&iommu->lock);
348 ret = intel_pasid_setup_first_level(iommu, dev,
349 mm ? mm->pgd : init_mm.pgd,
350 svm->pasid, FLPT_DEFAULT_DID,
351 mm ? 0 : PASID_FLAG_SUPERVISOR_MODE);
352 spin_unlock(&iommu->lock);
353 if (ret) {
354 if (mm)
355 mmu_notifier_unregister(&svm->notifier, mm);
356 intel_pasid_free_id(svm->pasid);
357 kfree(svm);
358 kfree(sdev);
359 goto out;
360 }
361
362 list_add_tail(&svm->list, &global_svm_list);
363 } else {
364 /*
365 * Binding a new device with existing PASID, need to setup
366 * the PASID entry.
367 */
368 spin_lock(&iommu->lock);
369 ret = intel_pasid_setup_first_level(iommu, dev,
370 mm ? mm->pgd : init_mm.pgd,
371 svm->pasid, FLPT_DEFAULT_DID,
372 mm ? 0 : PASID_FLAG_SUPERVISOR_MODE);
373 spin_unlock(&iommu->lock);
374 if (ret) {
375 kfree(sdev);
376 goto out;
377 }
378 }
379 list_add_rcu(&sdev->list, &svm->devs);
380
381 success:
382 *pasid = svm->pasid;
383 ret = 0;
384 out:
385 mutex_unlock(&pasid_mutex);
386 if (mm)
387 mmput(mm);
388 return ret;
389}
390EXPORT_SYMBOL_GPL(intel_svm_bind_mm);
391
392int intel_svm_unbind_mm(struct device *dev, int pasid)
393{
394 struct intel_svm_dev *sdev;
395 struct intel_iommu *iommu;
396 struct intel_svm *svm;
397 int ret = -EINVAL;
398
399 mutex_lock(&pasid_mutex);
400 iommu = intel_svm_device_to_iommu(dev);
401 if (!iommu)
402 goto out;
403
404 svm = intel_pasid_lookup_id(pasid);
405 if (!svm)
406 goto out;
407
408 list_for_each_entry(sdev, &svm->devs, list) {
409 if (dev == sdev->dev) {
410 ret = 0;
411 sdev->users--;
412 if (!sdev->users) {
413 list_del_rcu(&sdev->list);
414 /* Flush the PASID cache and IOTLB for this device.
415 * Note that we do depend on the hardware *not* using
416 * the PASID any more. Just as we depend on other
417 * devices never using PASIDs that they have no right
418 * to use. We have a *shared* PASID table, because it's
419 * large and has to be physically contiguous. So it's
420 * hard to be as defensive as we might like. */
421 intel_pasid_tear_down_entry(iommu, dev, svm->pasid);
422 intel_flush_svm_range_dev(svm, sdev, 0, -1, 0);
423 kfree_rcu(sdev, rcu);
424
425 if (list_empty(&svm->devs)) {
426 intel_pasid_free_id(svm->pasid);
427 if (svm->mm)
428 mmu_notifier_unregister(&svm->notifier, svm->mm);
429
430 list_del(&svm->list);
431
432 /* We mandate that no page faults may be outstanding
433 * for the PASID when intel_svm_unbind_mm() is called.
434 * If that is not obeyed, subtle errors will happen.
435 * Let's make them less subtle... */
436 memset(svm, 0x6b, sizeof(*svm));
437 kfree(svm);
438 }
439 }
440 break;
441 }
442 }
443 out:
444 mutex_unlock(&pasid_mutex);
445
446 return ret;
447}
448EXPORT_SYMBOL_GPL(intel_svm_unbind_mm);
449
450int intel_svm_is_pasid_valid(struct device *dev, int pasid)
451{
452 struct intel_iommu *iommu;
453 struct intel_svm *svm;
454 int ret = -EINVAL;
455
456 mutex_lock(&pasid_mutex);
457 iommu = intel_svm_device_to_iommu(dev);
458 if (!iommu)
459 goto out;
460
461 svm = intel_pasid_lookup_id(pasid);
462 if (!svm)
463 goto out;
464
465 /* init_mm is used in this case */
466 if (!svm->mm)
467 ret = 1;
468 else if (atomic_read(&svm->mm->mm_users) > 0)
469 ret = 1;
470 else
471 ret = 0;
472
473 out:
474 mutex_unlock(&pasid_mutex);
475
476 return ret;
477}
478EXPORT_SYMBOL_GPL(intel_svm_is_pasid_valid);
479
480/* Page request queue descriptor */
481struct page_req_dsc {
482 union {
483 struct {
484 u64 type:8;
485 u64 pasid_present:1;
486 u64 priv_data_present:1;
487 u64 rsvd:6;
488 u64 rid:16;
489 u64 pasid:20;
490 u64 exe_req:1;
491 u64 pm_req:1;
492 u64 rsvd2:10;
493 };
494 u64 qw_0;
495 };
496 union {
497 struct {
498 u64 rd_req:1;
499 u64 wr_req:1;
500 u64 lpig:1;
501 u64 prg_index:9;
502 u64 addr:52;
503 };
504 u64 qw_1;
505 };
506 u64 priv_data[2];
507};
508
509#define PRQ_RING_MASK ((0x1000 << PRQ_ORDER) - 0x10)
510
511static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
512{
513 unsigned long requested = 0;
514
515 if (req->exe_req)
516 requested |= VM_EXEC;
517
518 if (req->rd_req)
519 requested |= VM_READ;
520
521 if (req->wr_req)
522 requested |= VM_WRITE;
523
524 return (requested & ~vma->vm_flags) != 0;
525}
526
527static bool is_canonical_address(u64 addr)
528{
529 int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
530 long saddr = (long) addr;
531
532 return (((saddr << shift) >> shift) == saddr);
533}
534
535static irqreturn_t prq_event_thread(int irq, void *d)
536{
537 struct intel_iommu *iommu = d;
538 struct intel_svm *svm = NULL;
539 int head, tail, handled = 0;
540
541 /* Clear PPR bit before reading head/tail registers, to
542 * ensure that we get a new interrupt if needed. */
543 writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
544
545 tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
546 head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
547 while (head != tail) {
548 struct intel_svm_dev *sdev;
549 struct vm_area_struct *vma;
550 struct page_req_dsc *req;
551 struct qi_desc resp;
552 int result;
553 vm_fault_t ret;
554 u64 address;
555
556 handled = 1;
557
558 req = &iommu->prq[head / sizeof(*req)];
559
560 result = QI_RESP_FAILURE;
561 address = (u64)req->addr << VTD_PAGE_SHIFT;
562 if (!req->pasid_present) {
563 pr_err("%s: Page request without PASID: %08llx %08llx\n",
564 iommu->name, ((unsigned long long *)req)[0],
565 ((unsigned long long *)req)[1]);
566 goto no_pasid;
567 }
568
569 if (!svm || svm->pasid != req->pasid) {
570 rcu_read_lock();
571 svm = intel_pasid_lookup_id(req->pasid);
572 /* It *can't* go away, because the driver is not permitted
573 * to unbind the mm while any page faults are outstanding.
574 * So we only need RCU to protect the internal idr code. */
575 rcu_read_unlock();
576
577 if (!svm) {
578 pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
579 iommu->name, req->pasid, ((unsigned long long *)req)[0],
580 ((unsigned long long *)req)[1]);
581 goto no_pasid;
582 }
583 }
584
585 result = QI_RESP_INVALID;
586 /* Since we're using init_mm.pgd directly, we should never take
587 * any faults on kernel addresses. */
588 if (!svm->mm)
589 goto bad_req;
590 /* If the mm is already defunct, don't handle faults. */
591 if (!mmget_not_zero(svm->mm))
592 goto bad_req;
593
594 /* If address is not canonical, return invalid response */
595 if (!is_canonical_address(address))
596 goto bad_req;
597
598 down_read(&svm->mm->mmap_sem);
599 vma = find_extend_vma(svm->mm, address);
600 if (!vma || address < vma->vm_start)
601 goto invalid;
602
603 if (access_error(vma, req))
604 goto invalid;
605
606 ret = handle_mm_fault(vma, address,
607 req->wr_req ? FAULT_FLAG_WRITE : 0);
608 if (ret & VM_FAULT_ERROR)
609 goto invalid;
610
611 result = QI_RESP_SUCCESS;
612 invalid:
613 up_read(&svm->mm->mmap_sem);
614 mmput(svm->mm);
615 bad_req:
616 /* Accounting for major/minor faults? */
617 rcu_read_lock();
618 list_for_each_entry_rcu(sdev, &svm->devs, list) {
619 if (sdev->sid == req->rid)
620 break;
621 }
622 /* Other devices can go away, but the drivers are not permitted
623 * to unbind while any page faults might be in flight. So it's
624 * OK to drop the 'lock' here now we have it. */
625 rcu_read_unlock();
626
627 if (WARN_ON(&sdev->list == &svm->devs))
628 sdev = NULL;
629
630 if (sdev && sdev->ops && sdev->ops->fault_cb) {
631 int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
632 (req->exe_req << 1) | (req->pm_req);
633 sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr,
634 req->priv_data, rwxp, result);
635 }
636 /* We get here in the error case where the PASID lookup failed,
637 and these can be NULL. Do not use them below this point! */
638 sdev = NULL;
639 svm = NULL;
640 no_pasid:
641 if (req->lpig || req->priv_data_present) {
642 /*
643 * Per VT-d spec. v3.0 ch7.7, system software must
644 * respond with page group response if private data
645 * is present (PDP) or last page in group (LPIG) bit
646 * is set. This is an additional VT-d feature beyond
647 * PCI ATS spec.
648 */
649 resp.qw0 = QI_PGRP_PASID(req->pasid) |
650 QI_PGRP_DID(req->rid) |
651 QI_PGRP_PASID_P(req->pasid_present) |
652 QI_PGRP_PDP(req->pasid_present) |
653 QI_PGRP_RESP_CODE(result) |
654 QI_PGRP_RESP_TYPE;
655 resp.qw1 = QI_PGRP_IDX(req->prg_index) |
656 QI_PGRP_LPIG(req->lpig);
657
658 if (req->priv_data_present)
659 memcpy(&resp.qw2, req->priv_data,
660 sizeof(req->priv_data));
661 }
662 resp.qw2 = 0;
663 resp.qw3 = 0;
664 qi_submit_sync(&resp, iommu);
665
666 head = (head + sizeof(*req)) & PRQ_RING_MASK;
667 }
668
669 dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
670
671 return IRQ_RETVAL(handled);
672}