Loading...
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/mmu_notifier.h>
9#include <linux/sched.h>
10#include <linux/sched/mm.h>
11#include <linux/slab.h>
12#include <linux/intel-svm.h>
13#include <linux/rculist.h>
14#include <linux/pci.h>
15#include <linux/pci-ats.h>
16#include <linux/dmar.h>
17#include <linux/interrupt.h>
18#include <linux/mm_types.h>
19#include <linux/xarray.h>
20#include <linux/ioasid.h>
21#include <asm/page.h>
22#include <asm/fpu/api.h>
23
24#include "iommu.h"
25#include "pasid.h"
26#include "perf.h"
27#include "../iommu-sva.h"
28#include "trace.h"
29
30static irqreturn_t prq_event_thread(int irq, void *d);
31static void intel_svm_drain_prq(struct device *dev, u32 pasid);
32#define to_intel_svm_dev(handle) container_of(handle, struct intel_svm_dev, sva)
33
34static DEFINE_XARRAY_ALLOC(pasid_private_array);
35static int pasid_private_add(ioasid_t pasid, void *priv)
36{
37 return xa_alloc(&pasid_private_array, &pasid, priv,
38 XA_LIMIT(pasid, pasid), GFP_ATOMIC);
39}
40
41static void pasid_private_remove(ioasid_t pasid)
42{
43 xa_erase(&pasid_private_array, pasid);
44}
45
46static void *pasid_private_find(ioasid_t pasid)
47{
48 return xa_load(&pasid_private_array, pasid);
49}
50
51static struct intel_svm_dev *
52svm_lookup_device_by_dev(struct intel_svm *svm, struct device *dev)
53{
54 struct intel_svm_dev *sdev = NULL, *t;
55
56 rcu_read_lock();
57 list_for_each_entry_rcu(t, &svm->devs, list) {
58 if (t->dev == dev) {
59 sdev = t;
60 break;
61 }
62 }
63 rcu_read_unlock();
64
65 return sdev;
66}
67
68int intel_svm_enable_prq(struct intel_iommu *iommu)
69{
70 struct iopf_queue *iopfq;
71 struct page *pages;
72 int irq, ret;
73
74 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
75 if (!pages) {
76 pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
77 iommu->name);
78 return -ENOMEM;
79 }
80 iommu->prq = page_address(pages);
81
82 irq = dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED + iommu->seq_id, iommu->node, iommu);
83 if (irq <= 0) {
84 pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
85 iommu->name);
86 ret = -EINVAL;
87 goto free_prq;
88 }
89 iommu->pr_irq = irq;
90
91 snprintf(iommu->iopfq_name, sizeof(iommu->iopfq_name),
92 "dmar%d-iopfq", iommu->seq_id);
93 iopfq = iopf_queue_alloc(iommu->iopfq_name);
94 if (!iopfq) {
95 pr_err("IOMMU: %s: Failed to allocate iopf queue\n", iommu->name);
96 ret = -ENOMEM;
97 goto free_hwirq;
98 }
99 iommu->iopf_queue = iopfq;
100
101 snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
102
103 ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
104 iommu->prq_name, iommu);
105 if (ret) {
106 pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
107 iommu->name);
108 goto free_iopfq;
109 }
110 dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
111 dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
112 dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
113
114 init_completion(&iommu->prq_complete);
115
116 return 0;
117
118free_iopfq:
119 iopf_queue_free(iommu->iopf_queue);
120 iommu->iopf_queue = NULL;
121free_hwirq:
122 dmar_free_hwirq(irq);
123 iommu->pr_irq = 0;
124free_prq:
125 free_pages((unsigned long)iommu->prq, PRQ_ORDER);
126 iommu->prq = NULL;
127
128 return ret;
129}
130
131int intel_svm_finish_prq(struct intel_iommu *iommu)
132{
133 dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
134 dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
135 dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
136
137 if (iommu->pr_irq) {
138 free_irq(iommu->pr_irq, iommu);
139 dmar_free_hwirq(iommu->pr_irq);
140 iommu->pr_irq = 0;
141 }
142
143 if (iommu->iopf_queue) {
144 iopf_queue_free(iommu->iopf_queue);
145 iommu->iopf_queue = NULL;
146 }
147
148 free_pages((unsigned long)iommu->prq, PRQ_ORDER);
149 iommu->prq = NULL;
150
151 return 0;
152}
153
154void intel_svm_check(struct intel_iommu *iommu)
155{
156 if (!pasid_supported(iommu))
157 return;
158
159 if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
160 !cap_fl1gp_support(iommu->cap)) {
161 pr_err("%s SVM disabled, incompatible 1GB page capability\n",
162 iommu->name);
163 return;
164 }
165
166 if (cpu_feature_enabled(X86_FEATURE_LA57) &&
167 !cap_fl5lp_support(iommu->cap)) {
168 pr_err("%s SVM disabled, incompatible paging mode\n",
169 iommu->name);
170 return;
171 }
172
173 iommu->flags |= VTD_FLAG_SVM_CAPABLE;
174}
175
176static void __flush_svm_range_dev(struct intel_svm *svm,
177 struct intel_svm_dev *sdev,
178 unsigned long address,
179 unsigned long pages, int ih)
180{
181 struct device_domain_info *info = dev_iommu_priv_get(sdev->dev);
182
183 if (WARN_ON(!pages))
184 return;
185
186 qi_flush_piotlb(sdev->iommu, sdev->did, svm->pasid, address, pages, ih);
187 if (info->ats_enabled) {
188 qi_flush_dev_iotlb_pasid(sdev->iommu, sdev->sid, info->pfsid,
189 svm->pasid, sdev->qdep, address,
190 order_base_2(pages));
191 quirk_extra_dev_tlb_flush(info, address, order_base_2(pages),
192 svm->pasid, sdev->qdep);
193 }
194}
195
196static void intel_flush_svm_range_dev(struct intel_svm *svm,
197 struct intel_svm_dev *sdev,
198 unsigned long address,
199 unsigned long pages, int ih)
200{
201 unsigned long shift = ilog2(__roundup_pow_of_two(pages));
202 unsigned long align = (1ULL << (VTD_PAGE_SHIFT + shift));
203 unsigned long start = ALIGN_DOWN(address, align);
204 unsigned long end = ALIGN(address + (pages << VTD_PAGE_SHIFT), align);
205
206 while (start < end) {
207 __flush_svm_range_dev(svm, sdev, start, align >> VTD_PAGE_SHIFT, ih);
208 start += align;
209 }
210}
211
212static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
213 unsigned long pages, int ih)
214{
215 struct intel_svm_dev *sdev;
216
217 rcu_read_lock();
218 list_for_each_entry_rcu(sdev, &svm->devs, list)
219 intel_flush_svm_range_dev(svm, sdev, address, pages, ih);
220 rcu_read_unlock();
221}
222
223/* Pages have been freed at this point */
224static void intel_invalidate_range(struct mmu_notifier *mn,
225 struct mm_struct *mm,
226 unsigned long start, unsigned long end)
227{
228 struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
229
230 intel_flush_svm_range(svm, start,
231 (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0);
232}
233
234static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
235{
236 struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
237 struct intel_svm_dev *sdev;
238
239 /* This might end up being called from exit_mmap(), *before* the page
240 * tables are cleared. And __mmu_notifier_release() will delete us from
241 * the list of notifiers so that our invalidate_range() callback doesn't
242 * get called when the page tables are cleared. So we need to protect
243 * against hardware accessing those page tables.
244 *
245 * We do it by clearing the entry in the PASID table and then flushing
246 * the IOTLB and the PASID table caches. This might upset hardware;
247 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
248 * page) so that we end up taking a fault that the hardware really
249 * *has* to handle gracefully without affecting other processes.
250 */
251 rcu_read_lock();
252 list_for_each_entry_rcu(sdev, &svm->devs, list)
253 intel_pasid_tear_down_entry(sdev->iommu, sdev->dev,
254 svm->pasid, true);
255 rcu_read_unlock();
256
257}
258
259static const struct mmu_notifier_ops intel_mmuops = {
260 .release = intel_mm_release,
261 .invalidate_range = intel_invalidate_range,
262};
263
264static DEFINE_MUTEX(pasid_mutex);
265
266static int pasid_to_svm_sdev(struct device *dev, unsigned int pasid,
267 struct intel_svm **rsvm,
268 struct intel_svm_dev **rsdev)
269{
270 struct intel_svm_dev *sdev = NULL;
271 struct intel_svm *svm;
272
273 /* The caller should hold the pasid_mutex lock */
274 if (WARN_ON(!mutex_is_locked(&pasid_mutex)))
275 return -EINVAL;
276
277 if (pasid == INVALID_IOASID || pasid >= PASID_MAX)
278 return -EINVAL;
279
280 svm = pasid_private_find(pasid);
281 if (IS_ERR(svm))
282 return PTR_ERR(svm);
283
284 if (!svm)
285 goto out;
286
287 /*
288 * If we found svm for the PASID, there must be at least one device
289 * bond.
290 */
291 if (WARN_ON(list_empty(&svm->devs)))
292 return -EINVAL;
293 sdev = svm_lookup_device_by_dev(svm, dev);
294
295out:
296 *rsvm = svm;
297 *rsdev = sdev;
298
299 return 0;
300}
301
302static struct iommu_sva *intel_svm_bind_mm(struct intel_iommu *iommu,
303 struct device *dev,
304 struct mm_struct *mm)
305{
306 struct device_domain_info *info = dev_iommu_priv_get(dev);
307 struct intel_svm_dev *sdev;
308 struct intel_svm *svm;
309 unsigned long sflags;
310 int ret = 0;
311
312 svm = pasid_private_find(mm->pasid);
313 if (!svm) {
314 svm = kzalloc(sizeof(*svm), GFP_KERNEL);
315 if (!svm)
316 return ERR_PTR(-ENOMEM);
317
318 svm->pasid = mm->pasid;
319 svm->mm = mm;
320 INIT_LIST_HEAD_RCU(&svm->devs);
321
322 svm->notifier.ops = &intel_mmuops;
323 ret = mmu_notifier_register(&svm->notifier, mm);
324 if (ret) {
325 kfree(svm);
326 return ERR_PTR(ret);
327 }
328
329 ret = pasid_private_add(svm->pasid, svm);
330 if (ret) {
331 mmu_notifier_unregister(&svm->notifier, mm);
332 kfree(svm);
333 return ERR_PTR(ret);
334 }
335 }
336
337 /* Find the matching device in svm list */
338 sdev = svm_lookup_device_by_dev(svm, dev);
339 if (sdev) {
340 sdev->users++;
341 goto success;
342 }
343
344 sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
345 if (!sdev) {
346 ret = -ENOMEM;
347 goto free_svm;
348 }
349
350 sdev->dev = dev;
351 sdev->iommu = iommu;
352 sdev->did = FLPT_DEFAULT_DID;
353 sdev->sid = PCI_DEVID(info->bus, info->devfn);
354 sdev->users = 1;
355 sdev->pasid = svm->pasid;
356 sdev->sva.dev = dev;
357 init_rcu_head(&sdev->rcu);
358 if (info->ats_enabled) {
359 sdev->dev_iotlb = 1;
360 sdev->qdep = info->ats_qdep;
361 if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
362 sdev->qdep = 0;
363 }
364
365 /* Setup the pasid table: */
366 sflags = cpu_feature_enabled(X86_FEATURE_LA57) ? PASID_FLAG_FL5LP : 0;
367 ret = intel_pasid_setup_first_level(iommu, dev, mm->pgd, mm->pasid,
368 FLPT_DEFAULT_DID, sflags);
369 if (ret)
370 goto free_sdev;
371
372 list_add_rcu(&sdev->list, &svm->devs);
373success:
374 return &sdev->sva;
375
376free_sdev:
377 kfree(sdev);
378free_svm:
379 if (list_empty(&svm->devs)) {
380 mmu_notifier_unregister(&svm->notifier, mm);
381 pasid_private_remove(mm->pasid);
382 kfree(svm);
383 }
384
385 return ERR_PTR(ret);
386}
387
388/* Caller must hold pasid_mutex */
389static int intel_svm_unbind_mm(struct device *dev, u32 pasid)
390{
391 struct intel_svm_dev *sdev;
392 struct intel_iommu *iommu;
393 struct intel_svm *svm;
394 struct mm_struct *mm;
395 int ret = -EINVAL;
396
397 iommu = device_to_iommu(dev, NULL, NULL);
398 if (!iommu)
399 goto out;
400
401 ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
402 if (ret)
403 goto out;
404 mm = svm->mm;
405
406 if (sdev) {
407 sdev->users--;
408 if (!sdev->users) {
409 list_del_rcu(&sdev->list);
410 /* Flush the PASID cache and IOTLB for this device.
411 * Note that we do depend on the hardware *not* using
412 * the PASID any more. Just as we depend on other
413 * devices never using PASIDs that they have no right
414 * to use. We have a *shared* PASID table, because it's
415 * large and has to be physically contiguous. So it's
416 * hard to be as defensive as we might like. */
417 intel_pasid_tear_down_entry(iommu, dev,
418 svm->pasid, false);
419 intel_svm_drain_prq(dev, svm->pasid);
420 kfree_rcu(sdev, rcu);
421
422 if (list_empty(&svm->devs)) {
423 if (svm->notifier.ops)
424 mmu_notifier_unregister(&svm->notifier, mm);
425 pasid_private_remove(svm->pasid);
426 /* We mandate that no page faults may be outstanding
427 * for the PASID when intel_svm_unbind_mm() is called.
428 * If that is not obeyed, subtle errors will happen.
429 * Let's make them less subtle... */
430 memset(svm, 0x6b, sizeof(*svm));
431 kfree(svm);
432 }
433 }
434 }
435out:
436 return ret;
437}
438
439/* Page request queue descriptor */
440struct page_req_dsc {
441 union {
442 struct {
443 u64 type:8;
444 u64 pasid_present:1;
445 u64 priv_data_present:1;
446 u64 rsvd:6;
447 u64 rid:16;
448 u64 pasid:20;
449 u64 exe_req:1;
450 u64 pm_req:1;
451 u64 rsvd2:10;
452 };
453 u64 qw_0;
454 };
455 union {
456 struct {
457 u64 rd_req:1;
458 u64 wr_req:1;
459 u64 lpig:1;
460 u64 prg_index:9;
461 u64 addr:52;
462 };
463 u64 qw_1;
464 };
465 u64 priv_data[2];
466};
467
468static bool is_canonical_address(u64 addr)
469{
470 int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
471 long saddr = (long) addr;
472
473 return (((saddr << shift) >> shift) == saddr);
474}
475
476/**
477 * intel_svm_drain_prq - Drain page requests and responses for a pasid
478 * @dev: target device
479 * @pasid: pasid for draining
480 *
481 * Drain all pending page requests and responses related to @pasid in both
482 * software and hardware. This is supposed to be called after the device
483 * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
484 * and DevTLB have been invalidated.
485 *
486 * It waits until all pending page requests for @pasid in the page fault
487 * queue are completed by the prq handling thread. Then follow the steps
488 * described in VT-d spec CH7.10 to drain all page requests and page
489 * responses pending in the hardware.
490 */
491static void intel_svm_drain_prq(struct device *dev, u32 pasid)
492{
493 struct device_domain_info *info;
494 struct dmar_domain *domain;
495 struct intel_iommu *iommu;
496 struct qi_desc desc[3];
497 struct pci_dev *pdev;
498 int head, tail;
499 u16 sid, did;
500 int qdep;
501
502 info = dev_iommu_priv_get(dev);
503 if (WARN_ON(!info || !dev_is_pci(dev)))
504 return;
505
506 if (!info->pri_enabled)
507 return;
508
509 iommu = info->iommu;
510 domain = info->domain;
511 pdev = to_pci_dev(dev);
512 sid = PCI_DEVID(info->bus, info->devfn);
513 did = domain_id_iommu(domain, iommu);
514 qdep = pci_ats_queue_depth(pdev);
515
516 /*
517 * Check and wait until all pending page requests in the queue are
518 * handled by the prq handling thread.
519 */
520prq_retry:
521 reinit_completion(&iommu->prq_complete);
522 tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
523 head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
524 while (head != tail) {
525 struct page_req_dsc *req;
526
527 req = &iommu->prq[head / sizeof(*req)];
528 if (!req->pasid_present || req->pasid != pasid) {
529 head = (head + sizeof(*req)) & PRQ_RING_MASK;
530 continue;
531 }
532
533 wait_for_completion(&iommu->prq_complete);
534 goto prq_retry;
535 }
536
537 /*
538 * A work in IO page fault workqueue may try to lock pasid_mutex now.
539 * Holding pasid_mutex while waiting in iopf_queue_flush_dev() for
540 * all works in the workqueue to finish may cause deadlock.
541 *
542 * It's unnecessary to hold pasid_mutex in iopf_queue_flush_dev().
543 * Unlock it to allow the works to be handled while waiting for
544 * them to finish.
545 */
546 lockdep_assert_held(&pasid_mutex);
547 mutex_unlock(&pasid_mutex);
548 iopf_queue_flush_dev(dev);
549 mutex_lock(&pasid_mutex);
550
551 /*
552 * Perform steps described in VT-d spec CH7.10 to drain page
553 * requests and responses in hardware.
554 */
555 memset(desc, 0, sizeof(desc));
556 desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
557 QI_IWD_FENCE |
558 QI_IWD_TYPE;
559 desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
560 QI_EIOTLB_DID(did) |
561 QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
562 QI_EIOTLB_TYPE;
563 desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
564 QI_DEV_EIOTLB_SID(sid) |
565 QI_DEV_EIOTLB_QDEP(qdep) |
566 QI_DEIOTLB_TYPE |
567 QI_DEV_IOTLB_PFSID(info->pfsid);
568qi_retry:
569 reinit_completion(&iommu->prq_complete);
570 qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
571 if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
572 wait_for_completion(&iommu->prq_complete);
573 goto qi_retry;
574 }
575}
576
577static int prq_to_iommu_prot(struct page_req_dsc *req)
578{
579 int prot = 0;
580
581 if (req->rd_req)
582 prot |= IOMMU_FAULT_PERM_READ;
583 if (req->wr_req)
584 prot |= IOMMU_FAULT_PERM_WRITE;
585 if (req->exe_req)
586 prot |= IOMMU_FAULT_PERM_EXEC;
587 if (req->pm_req)
588 prot |= IOMMU_FAULT_PERM_PRIV;
589
590 return prot;
591}
592
593static int intel_svm_prq_report(struct intel_iommu *iommu, struct device *dev,
594 struct page_req_dsc *desc)
595{
596 struct iommu_fault_event event;
597
598 if (!dev || !dev_is_pci(dev))
599 return -ENODEV;
600
601 /* Fill in event data for device specific processing */
602 memset(&event, 0, sizeof(struct iommu_fault_event));
603 event.fault.type = IOMMU_FAULT_PAGE_REQ;
604 event.fault.prm.addr = (u64)desc->addr << VTD_PAGE_SHIFT;
605 event.fault.prm.pasid = desc->pasid;
606 event.fault.prm.grpid = desc->prg_index;
607 event.fault.prm.perm = prq_to_iommu_prot(desc);
608
609 if (desc->lpig)
610 event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
611 if (desc->pasid_present) {
612 event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
613 event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
614 }
615 if (desc->priv_data_present) {
616 /*
617 * Set last page in group bit if private data is present,
618 * page response is required as it does for LPIG.
619 * iommu_report_device_fault() doesn't understand this vendor
620 * specific requirement thus we set last_page as a workaround.
621 */
622 event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
623 event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
624 event.fault.prm.private_data[0] = desc->priv_data[0];
625 event.fault.prm.private_data[1] = desc->priv_data[1];
626 } else if (dmar_latency_enabled(iommu, DMAR_LATENCY_PRQ)) {
627 /*
628 * If the private data fields are not used by hardware, use it
629 * to monitor the prq handle latency.
630 */
631 event.fault.prm.private_data[0] = ktime_to_ns(ktime_get());
632 }
633
634 return iommu_report_device_fault(dev, &event);
635}
636
637static void handle_bad_prq_event(struct intel_iommu *iommu,
638 struct page_req_dsc *req, int result)
639{
640 struct qi_desc desc;
641
642 pr_err("%s: Invalid page request: %08llx %08llx\n",
643 iommu->name, ((unsigned long long *)req)[0],
644 ((unsigned long long *)req)[1]);
645
646 /*
647 * Per VT-d spec. v3.0 ch7.7, system software must
648 * respond with page group response if private data
649 * is present (PDP) or last page in group (LPIG) bit
650 * is set. This is an additional VT-d feature beyond
651 * PCI ATS spec.
652 */
653 if (!req->lpig && !req->priv_data_present)
654 return;
655
656 desc.qw0 = QI_PGRP_PASID(req->pasid) |
657 QI_PGRP_DID(req->rid) |
658 QI_PGRP_PASID_P(req->pasid_present) |
659 QI_PGRP_PDP(req->priv_data_present) |
660 QI_PGRP_RESP_CODE(result) |
661 QI_PGRP_RESP_TYPE;
662 desc.qw1 = QI_PGRP_IDX(req->prg_index) |
663 QI_PGRP_LPIG(req->lpig);
664
665 if (req->priv_data_present) {
666 desc.qw2 = req->priv_data[0];
667 desc.qw3 = req->priv_data[1];
668 } else {
669 desc.qw2 = 0;
670 desc.qw3 = 0;
671 }
672
673 qi_submit_sync(iommu, &desc, 1, 0);
674}
675
676static irqreturn_t prq_event_thread(int irq, void *d)
677{
678 struct intel_iommu *iommu = d;
679 struct page_req_dsc *req;
680 int head, tail, handled;
681 struct pci_dev *pdev;
682 u64 address;
683
684 /*
685 * Clear PPR bit before reading head/tail registers, to ensure that
686 * we get a new interrupt if needed.
687 */
688 writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
689
690 tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
691 head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
692 handled = (head != tail);
693 while (head != tail) {
694 req = &iommu->prq[head / sizeof(*req)];
695 address = (u64)req->addr << VTD_PAGE_SHIFT;
696
697 if (unlikely(!req->pasid_present)) {
698 pr_err("IOMMU: %s: Page request without PASID\n",
699 iommu->name);
700bad_req:
701 handle_bad_prq_event(iommu, req, QI_RESP_INVALID);
702 goto prq_advance;
703 }
704
705 if (unlikely(!is_canonical_address(address))) {
706 pr_err("IOMMU: %s: Address is not canonical\n",
707 iommu->name);
708 goto bad_req;
709 }
710
711 if (unlikely(req->pm_req && (req->rd_req | req->wr_req))) {
712 pr_err("IOMMU: %s: Page request in Privilege Mode\n",
713 iommu->name);
714 goto bad_req;
715 }
716
717 if (unlikely(req->exe_req && req->rd_req)) {
718 pr_err("IOMMU: %s: Execution request not supported\n",
719 iommu->name);
720 goto bad_req;
721 }
722
723 /* Drop Stop Marker message. No need for a response. */
724 if (unlikely(req->lpig && !req->rd_req && !req->wr_req))
725 goto prq_advance;
726
727 pdev = pci_get_domain_bus_and_slot(iommu->segment,
728 PCI_BUS_NUM(req->rid),
729 req->rid & 0xff);
730 /*
731 * If prq is to be handled outside iommu driver via receiver of
732 * the fault notifiers, we skip the page response here.
733 */
734 if (!pdev)
735 goto bad_req;
736
737 if (intel_svm_prq_report(iommu, &pdev->dev, req))
738 handle_bad_prq_event(iommu, req, QI_RESP_INVALID);
739 else
740 trace_prq_report(iommu, &pdev->dev, req->qw_0, req->qw_1,
741 req->priv_data[0], req->priv_data[1],
742 iommu->prq_seq_number++);
743 pci_dev_put(pdev);
744prq_advance:
745 head = (head + sizeof(*req)) & PRQ_RING_MASK;
746 }
747
748 dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
749
750 /*
751 * Clear the page request overflow bit and wake up all threads that
752 * are waiting for the completion of this handling.
753 */
754 if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
755 pr_info_ratelimited("IOMMU: %s: PRQ overflow detected\n",
756 iommu->name);
757 head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
758 tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
759 if (head == tail) {
760 iopf_queue_discard_partial(iommu->iopf_queue);
761 writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
762 pr_info_ratelimited("IOMMU: %s: PRQ overflow cleared",
763 iommu->name);
764 }
765 }
766
767 if (!completion_done(&iommu->prq_complete))
768 complete(&iommu->prq_complete);
769
770 return IRQ_RETVAL(handled);
771}
772
773int intel_svm_page_response(struct device *dev,
774 struct iommu_fault_event *evt,
775 struct iommu_page_response *msg)
776{
777 struct iommu_fault_page_request *prm;
778 struct intel_iommu *iommu;
779 bool private_present;
780 bool pasid_present;
781 bool last_page;
782 u8 bus, devfn;
783 int ret = 0;
784 u16 sid;
785
786 if (!dev || !dev_is_pci(dev))
787 return -ENODEV;
788
789 iommu = device_to_iommu(dev, &bus, &devfn);
790 if (!iommu)
791 return -ENODEV;
792
793 if (!msg || !evt)
794 return -EINVAL;
795
796 prm = &evt->fault.prm;
797 sid = PCI_DEVID(bus, devfn);
798 pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
799 private_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
800 last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
801
802 if (!pasid_present) {
803 ret = -EINVAL;
804 goto out;
805 }
806
807 if (prm->pasid == 0 || prm->pasid >= PASID_MAX) {
808 ret = -EINVAL;
809 goto out;
810 }
811
812 /*
813 * Per VT-d spec. v3.0 ch7.7, system software must respond
814 * with page group response if private data is present (PDP)
815 * or last page in group (LPIG) bit is set. This is an
816 * additional VT-d requirement beyond PCI ATS spec.
817 */
818 if (last_page || private_present) {
819 struct qi_desc desc;
820
821 desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
822 QI_PGRP_PASID_P(pasid_present) |
823 QI_PGRP_PDP(private_present) |
824 QI_PGRP_RESP_CODE(msg->code) |
825 QI_PGRP_RESP_TYPE;
826 desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
827 desc.qw2 = 0;
828 desc.qw3 = 0;
829
830 if (private_present) {
831 desc.qw2 = prm->private_data[0];
832 desc.qw3 = prm->private_data[1];
833 } else if (prm->private_data[0]) {
834 dmar_latency_update(iommu, DMAR_LATENCY_PRQ,
835 ktime_to_ns(ktime_get()) - prm->private_data[0]);
836 }
837
838 qi_submit_sync(iommu, &desc, 1, 0);
839 }
840out:
841 return ret;
842}
843
844void intel_svm_remove_dev_pasid(struct device *dev, ioasid_t pasid)
845{
846 mutex_lock(&pasid_mutex);
847 intel_svm_unbind_mm(dev, pasid);
848 mutex_unlock(&pasid_mutex);
849}
850
851static int intel_svm_set_dev_pasid(struct iommu_domain *domain,
852 struct device *dev, ioasid_t pasid)
853{
854 struct device_domain_info *info = dev_iommu_priv_get(dev);
855 struct intel_iommu *iommu = info->iommu;
856 struct mm_struct *mm = domain->mm;
857 struct iommu_sva *sva;
858 int ret = 0;
859
860 mutex_lock(&pasid_mutex);
861 sva = intel_svm_bind_mm(iommu, dev, mm);
862 if (IS_ERR(sva))
863 ret = PTR_ERR(sva);
864 mutex_unlock(&pasid_mutex);
865
866 return ret;
867}
868
869static void intel_svm_domain_free(struct iommu_domain *domain)
870{
871 kfree(to_dmar_domain(domain));
872}
873
874static const struct iommu_domain_ops intel_svm_domain_ops = {
875 .set_dev_pasid = intel_svm_set_dev_pasid,
876 .free = intel_svm_domain_free
877};
878
879struct iommu_domain *intel_svm_domain_alloc(void)
880{
881 struct dmar_domain *domain;
882
883 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
884 if (!domain)
885 return NULL;
886 domain->domain.ops = &intel_svm_domain_ops;
887
888 return &domain->domain;
889}
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/mmu_notifier.h>
9#include <linux/sched.h>
10#include <linux/sched/mm.h>
11#include <linux/slab.h>
12#include <linux/rculist.h>
13#include <linux/pci.h>
14#include <linux/pci-ats.h>
15#include <linux/dmar.h>
16#include <linux/interrupt.h>
17#include <linux/mm_types.h>
18#include <linux/xarray.h>
19#include <asm/page.h>
20#include <asm/fpu/api.h>
21
22#include "iommu.h"
23#include "pasid.h"
24#include "perf.h"
25#include "../iommu-pages.h"
26#include "trace.h"
27
28void intel_svm_check(struct intel_iommu *iommu)
29{
30 if (!pasid_supported(iommu))
31 return;
32
33 if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
34 !cap_fl1gp_support(iommu->cap)) {
35 pr_err("%s SVM disabled, incompatible 1GB page capability\n",
36 iommu->name);
37 return;
38 }
39
40 if (cpu_feature_enabled(X86_FEATURE_LA57) &&
41 !cap_fl5lp_support(iommu->cap)) {
42 pr_err("%s SVM disabled, incompatible paging mode\n",
43 iommu->name);
44 return;
45 }
46
47 iommu->flags |= VTD_FLAG_SVM_CAPABLE;
48}
49
50/* Pages have been freed at this point */
51static void intel_arch_invalidate_secondary_tlbs(struct mmu_notifier *mn,
52 struct mm_struct *mm,
53 unsigned long start, unsigned long end)
54{
55 struct dmar_domain *domain = container_of(mn, struct dmar_domain, notifier);
56
57 if (start == 0 && end == ULONG_MAX) {
58 cache_tag_flush_all(domain);
59 return;
60 }
61
62 /*
63 * The mm_types defines vm_end as the first byte after the end address,
64 * different from IOMMU subsystem using the last address of an address
65 * range.
66 */
67 cache_tag_flush_range(domain, start, end - 1, 0);
68}
69
70static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
71{
72 struct dmar_domain *domain = container_of(mn, struct dmar_domain, notifier);
73 struct dev_pasid_info *dev_pasid;
74 struct device_domain_info *info;
75 unsigned long flags;
76
77 /* This might end up being called from exit_mmap(), *before* the page
78 * tables are cleared. And __mmu_notifier_release() will delete us from
79 * the list of notifiers so that our invalidate_range() callback doesn't
80 * get called when the page tables are cleared. So we need to protect
81 * against hardware accessing those page tables.
82 *
83 * We do it by clearing the entry in the PASID table and then flushing
84 * the IOTLB and the PASID table caches. This might upset hardware;
85 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
86 * page) so that we end up taking a fault that the hardware really
87 * *has* to handle gracefully without affecting other processes.
88 */
89 spin_lock_irqsave(&domain->lock, flags);
90 list_for_each_entry(dev_pasid, &domain->dev_pasids, link_domain) {
91 info = dev_iommu_priv_get(dev_pasid->dev);
92 intel_pasid_tear_down_entry(info->iommu, dev_pasid->dev,
93 dev_pasid->pasid, true);
94 }
95 spin_unlock_irqrestore(&domain->lock, flags);
96
97}
98
99static void intel_mm_free_notifier(struct mmu_notifier *mn)
100{
101 struct dmar_domain *domain = container_of(mn, struct dmar_domain, notifier);
102
103 kfree(domain->qi_batch);
104 kfree(domain);
105}
106
107static const struct mmu_notifier_ops intel_mmuops = {
108 .release = intel_mm_release,
109 .arch_invalidate_secondary_tlbs = intel_arch_invalidate_secondary_tlbs,
110 .free_notifier = intel_mm_free_notifier,
111};
112
113static int intel_svm_set_dev_pasid(struct iommu_domain *domain,
114 struct device *dev, ioasid_t pasid,
115 struct iommu_domain *old)
116{
117 struct device_domain_info *info = dev_iommu_priv_get(dev);
118 struct intel_iommu *iommu = info->iommu;
119 struct mm_struct *mm = domain->mm;
120 struct dev_pasid_info *dev_pasid;
121 unsigned long sflags;
122 int ret = 0;
123
124 dev_pasid = domain_add_dev_pasid(domain, dev, pasid);
125 if (IS_ERR(dev_pasid))
126 return PTR_ERR(dev_pasid);
127
128 /* Setup the pasid table: */
129 sflags = cpu_feature_enabled(X86_FEATURE_LA57) ? PASID_FLAG_FL5LP : 0;
130 ret = __domain_setup_first_level(iommu, dev, pasid,
131 FLPT_DEFAULT_DID, mm->pgd,
132 sflags, old);
133 if (ret)
134 goto out_remove_dev_pasid;
135
136 domain_remove_dev_pasid(old, dev, pasid);
137
138 return 0;
139
140out_remove_dev_pasid:
141 domain_remove_dev_pasid(domain, dev, pasid);
142 return ret;
143}
144
145static void intel_svm_domain_free(struct iommu_domain *domain)
146{
147 struct dmar_domain *dmar_domain = to_dmar_domain(domain);
148
149 /* dmar_domain free is deferred to the mmu free_notifier callback. */
150 mmu_notifier_put(&dmar_domain->notifier);
151}
152
153static const struct iommu_domain_ops intel_svm_domain_ops = {
154 .set_dev_pasid = intel_svm_set_dev_pasid,
155 .free = intel_svm_domain_free
156};
157
158struct iommu_domain *intel_svm_domain_alloc(struct device *dev,
159 struct mm_struct *mm)
160{
161 struct dmar_domain *domain;
162 int ret;
163
164 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
165 if (!domain)
166 return ERR_PTR(-ENOMEM);
167
168 domain->domain.ops = &intel_svm_domain_ops;
169 domain->use_first_level = true;
170 INIT_LIST_HEAD(&domain->dev_pasids);
171 INIT_LIST_HEAD(&domain->cache_tags);
172 spin_lock_init(&domain->cache_lock);
173 spin_lock_init(&domain->lock);
174
175 domain->notifier.ops = &intel_mmuops;
176 ret = mmu_notifier_register(&domain->notifier, mm);
177 if (ret) {
178 kfree(domain);
179 return ERR_PTR(ret);
180 }
181
182 return &domain->domain;
183}