xref: /linux/drivers/iommu/intel/svm.c (revision d27bb0246e5356dbef4d923e72c680bf893885a8)
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-sva.h"
26 #include "trace.h"
27 
28 static irqreturn_t prq_event_thread(int irq, void *d);
29 
30 static DEFINE_XARRAY_ALLOC(pasid_private_array);
31 static int pasid_private_add(ioasid_t pasid, void *priv)
32 {
33 	return xa_alloc(&pasid_private_array, &pasid, priv,
34 			XA_LIMIT(pasid, pasid), GFP_ATOMIC);
35 }
36 
37 static void pasid_private_remove(ioasid_t pasid)
38 {
39 	xa_erase(&pasid_private_array, pasid);
40 }
41 
42 static void *pasid_private_find(ioasid_t pasid)
43 {
44 	return xa_load(&pasid_private_array, pasid);
45 }
46 
47 static struct intel_svm_dev *
48 svm_lookup_device_by_dev(struct intel_svm *svm, struct device *dev)
49 {
50 	struct intel_svm_dev *sdev = NULL, *t;
51 
52 	rcu_read_lock();
53 	list_for_each_entry_rcu(t, &svm->devs, list) {
54 		if (t->dev == dev) {
55 			sdev = t;
56 			break;
57 		}
58 	}
59 	rcu_read_unlock();
60 
61 	return sdev;
62 }
63 
64 int intel_svm_enable_prq(struct intel_iommu *iommu)
65 {
66 	struct iopf_queue *iopfq;
67 	struct page *pages;
68 	int irq, ret;
69 
70 	pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
71 	if (!pages) {
72 		pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
73 			iommu->name);
74 		return -ENOMEM;
75 	}
76 	iommu->prq = page_address(pages);
77 
78 	irq = dmar_alloc_hwirq(IOMMU_IRQ_ID_OFFSET_PRQ + iommu->seq_id, iommu->node, iommu);
79 	if (irq <= 0) {
80 		pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
81 		       iommu->name);
82 		ret = -EINVAL;
83 		goto free_prq;
84 	}
85 	iommu->pr_irq = irq;
86 
87 	snprintf(iommu->iopfq_name, sizeof(iommu->iopfq_name),
88 		 "dmar%d-iopfq", iommu->seq_id);
89 	iopfq = iopf_queue_alloc(iommu->iopfq_name);
90 	if (!iopfq) {
91 		pr_err("IOMMU: %s: Failed to allocate iopf queue\n", iommu->name);
92 		ret = -ENOMEM;
93 		goto free_hwirq;
94 	}
95 	iommu->iopf_queue = iopfq;
96 
97 	snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
98 
99 	ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
100 				   iommu->prq_name, iommu);
101 	if (ret) {
102 		pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
103 		       iommu->name);
104 		goto free_iopfq;
105 	}
106 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
107 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
108 	dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
109 
110 	init_completion(&iommu->prq_complete);
111 
112 	return 0;
113 
114 free_iopfq:
115 	iopf_queue_free(iommu->iopf_queue);
116 	iommu->iopf_queue = NULL;
117 free_hwirq:
118 	dmar_free_hwirq(irq);
119 	iommu->pr_irq = 0;
120 free_prq:
121 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
122 	iommu->prq = NULL;
123 
124 	return ret;
125 }
126 
127 int intel_svm_finish_prq(struct intel_iommu *iommu)
128 {
129 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
130 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
131 	dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
132 
133 	if (iommu->pr_irq) {
134 		free_irq(iommu->pr_irq, iommu);
135 		dmar_free_hwirq(iommu->pr_irq);
136 		iommu->pr_irq = 0;
137 	}
138 
139 	if (iommu->iopf_queue) {
140 		iopf_queue_free(iommu->iopf_queue);
141 		iommu->iopf_queue = NULL;
142 	}
143 
144 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
145 	iommu->prq = NULL;
146 
147 	return 0;
148 }
149 
150 void intel_svm_check(struct intel_iommu *iommu)
151 {
152 	if (!pasid_supported(iommu))
153 		return;
154 
155 	if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
156 	    !cap_fl1gp_support(iommu->cap)) {
157 		pr_err("%s SVM disabled, incompatible 1GB page capability\n",
158 		       iommu->name);
159 		return;
160 	}
161 
162 	if (cpu_feature_enabled(X86_FEATURE_LA57) &&
163 	    !cap_fl5lp_support(iommu->cap)) {
164 		pr_err("%s SVM disabled, incompatible paging mode\n",
165 		       iommu->name);
166 		return;
167 	}
168 
169 	iommu->flags |= VTD_FLAG_SVM_CAPABLE;
170 }
171 
172 static void __flush_svm_range_dev(struct intel_svm *svm,
173 				  struct intel_svm_dev *sdev,
174 				  unsigned long address,
175 				  unsigned long pages, int ih)
176 {
177 	struct device_domain_info *info = dev_iommu_priv_get(sdev->dev);
178 
179 	if (WARN_ON(!pages))
180 		return;
181 
182 	qi_flush_piotlb(sdev->iommu, sdev->did, svm->pasid, address, pages, ih);
183 	if (info->ats_enabled) {
184 		qi_flush_dev_iotlb_pasid(sdev->iommu, sdev->sid, info->pfsid,
185 					 svm->pasid, sdev->qdep, address,
186 					 order_base_2(pages));
187 		quirk_extra_dev_tlb_flush(info, address, order_base_2(pages),
188 					  svm->pasid, sdev->qdep);
189 	}
190 }
191 
192 static void intel_flush_svm_range_dev(struct intel_svm *svm,
193 				      struct intel_svm_dev *sdev,
194 				      unsigned long address,
195 				      unsigned long pages, int ih)
196 {
197 	unsigned long shift = ilog2(__roundup_pow_of_two(pages));
198 	unsigned long align = (1ULL << (VTD_PAGE_SHIFT + shift));
199 	unsigned long start = ALIGN_DOWN(address, align);
200 	unsigned long end = ALIGN(address + (pages << VTD_PAGE_SHIFT), align);
201 
202 	while (start < end) {
203 		__flush_svm_range_dev(svm, sdev, start, align >> VTD_PAGE_SHIFT, ih);
204 		start += align;
205 	}
206 }
207 
208 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
209 				unsigned long pages, int ih)
210 {
211 	struct intel_svm_dev *sdev;
212 
213 	rcu_read_lock();
214 	list_for_each_entry_rcu(sdev, &svm->devs, list)
215 		intel_flush_svm_range_dev(svm, sdev, address, pages, ih);
216 	rcu_read_unlock();
217 }
218 
219 /* Pages have been freed at this point */
220 static void intel_arch_invalidate_secondary_tlbs(struct mmu_notifier *mn,
221 					struct mm_struct *mm,
222 					unsigned long start, unsigned long end)
223 {
224 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
225 
226 	intel_flush_svm_range(svm, start,
227 			      (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0);
228 }
229 
230 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
231 {
232 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
233 	struct intel_svm_dev *sdev;
234 
235 	/* This might end up being called from exit_mmap(), *before* the page
236 	 * tables are cleared. And __mmu_notifier_release() will delete us from
237 	 * the list of notifiers so that our invalidate_range() callback doesn't
238 	 * get called when the page tables are cleared. So we need to protect
239 	 * against hardware accessing those page tables.
240 	 *
241 	 * We do it by clearing the entry in the PASID table and then flushing
242 	 * the IOTLB and the PASID table caches. This might upset hardware;
243 	 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
244 	 * page) so that we end up taking a fault that the hardware really
245 	 * *has* to handle gracefully without affecting other processes.
246 	 */
247 	rcu_read_lock();
248 	list_for_each_entry_rcu(sdev, &svm->devs, list)
249 		intel_pasid_tear_down_entry(sdev->iommu, sdev->dev,
250 					    svm->pasid, true);
251 	rcu_read_unlock();
252 
253 }
254 
255 static const struct mmu_notifier_ops intel_mmuops = {
256 	.release = intel_mm_release,
257 	.arch_invalidate_secondary_tlbs = intel_arch_invalidate_secondary_tlbs,
258 };
259 
260 static int pasid_to_svm_sdev(struct device *dev, unsigned int pasid,
261 			     struct intel_svm **rsvm,
262 			     struct intel_svm_dev **rsdev)
263 {
264 	struct intel_svm_dev *sdev = NULL;
265 	struct intel_svm *svm;
266 
267 	if (pasid == IOMMU_PASID_INVALID || pasid >= PASID_MAX)
268 		return -EINVAL;
269 
270 	svm = pasid_private_find(pasid);
271 	if (IS_ERR(svm))
272 		return PTR_ERR(svm);
273 
274 	if (!svm)
275 		goto out;
276 
277 	/*
278 	 * If we found svm for the PASID, there must be at least one device
279 	 * bond.
280 	 */
281 	if (WARN_ON(list_empty(&svm->devs)))
282 		return -EINVAL;
283 	sdev = svm_lookup_device_by_dev(svm, dev);
284 
285 out:
286 	*rsvm = svm;
287 	*rsdev = sdev;
288 
289 	return 0;
290 }
291 
292 static int intel_svm_bind_mm(struct intel_iommu *iommu, struct device *dev,
293 			     struct mm_struct *mm)
294 {
295 	struct device_domain_info *info = dev_iommu_priv_get(dev);
296 	struct intel_svm_dev *sdev;
297 	struct intel_svm *svm;
298 	unsigned long sflags;
299 	int ret = 0;
300 
301 	svm = pasid_private_find(mm->pasid);
302 	if (!svm) {
303 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
304 		if (!svm)
305 			return -ENOMEM;
306 
307 		svm->pasid = mm->pasid;
308 		svm->mm = mm;
309 		INIT_LIST_HEAD_RCU(&svm->devs);
310 
311 		svm->notifier.ops = &intel_mmuops;
312 		ret = mmu_notifier_register(&svm->notifier, mm);
313 		if (ret) {
314 			kfree(svm);
315 			return ret;
316 		}
317 
318 		ret = pasid_private_add(svm->pasid, svm);
319 		if (ret) {
320 			mmu_notifier_unregister(&svm->notifier, mm);
321 			kfree(svm);
322 			return ret;
323 		}
324 	}
325 
326 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
327 	if (!sdev) {
328 		ret = -ENOMEM;
329 		goto free_svm;
330 	}
331 
332 	sdev->dev = dev;
333 	sdev->iommu = iommu;
334 	sdev->did = FLPT_DEFAULT_DID;
335 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
336 	init_rcu_head(&sdev->rcu);
337 	if (info->ats_enabled) {
338 		sdev->qdep = info->ats_qdep;
339 		if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
340 			sdev->qdep = 0;
341 	}
342 
343 	/* Setup the pasid table: */
344 	sflags = cpu_feature_enabled(X86_FEATURE_LA57) ? PASID_FLAG_FL5LP : 0;
345 	ret = intel_pasid_setup_first_level(iommu, dev, mm->pgd, mm->pasid,
346 					    FLPT_DEFAULT_DID, sflags);
347 	if (ret)
348 		goto free_sdev;
349 
350 	list_add_rcu(&sdev->list, &svm->devs);
351 
352 	return 0;
353 
354 free_sdev:
355 	kfree(sdev);
356 free_svm:
357 	if (list_empty(&svm->devs)) {
358 		mmu_notifier_unregister(&svm->notifier, mm);
359 		pasid_private_remove(mm->pasid);
360 		kfree(svm);
361 	}
362 
363 	return ret;
364 }
365 
366 void intel_svm_remove_dev_pasid(struct device *dev, u32 pasid)
367 {
368 	struct intel_svm_dev *sdev;
369 	struct intel_iommu *iommu;
370 	struct intel_svm *svm;
371 	struct mm_struct *mm;
372 
373 	iommu = device_to_iommu(dev, NULL, NULL);
374 	if (!iommu)
375 		return;
376 
377 	if (pasid_to_svm_sdev(dev, pasid, &svm, &sdev))
378 		return;
379 	mm = svm->mm;
380 
381 	if (sdev) {
382 		list_del_rcu(&sdev->list);
383 		kfree_rcu(sdev, rcu);
384 
385 		if (list_empty(&svm->devs)) {
386 			if (svm->notifier.ops)
387 				mmu_notifier_unregister(&svm->notifier, mm);
388 			pasid_private_remove(svm->pasid);
389 			/*
390 			 * We mandate that no page faults may be outstanding
391 			 * for the PASID when intel_svm_unbind_mm() is called.
392 			 * If that is not obeyed, subtle errors will happen.
393 			 * Let's make them less subtle...
394 			 */
395 			memset(svm, 0x6b, sizeof(*svm));
396 			kfree(svm);
397 		}
398 	}
399 }
400 
401 /* Page request queue descriptor */
402 struct page_req_dsc {
403 	union {
404 		struct {
405 			u64 type:8;
406 			u64 pasid_present:1;
407 			u64 priv_data_present:1;
408 			u64 rsvd:6;
409 			u64 rid:16;
410 			u64 pasid:20;
411 			u64 exe_req:1;
412 			u64 pm_req:1;
413 			u64 rsvd2:10;
414 		};
415 		u64 qw_0;
416 	};
417 	union {
418 		struct {
419 			u64 rd_req:1;
420 			u64 wr_req:1;
421 			u64 lpig:1;
422 			u64 prg_index:9;
423 			u64 addr:52;
424 		};
425 		u64 qw_1;
426 	};
427 	u64 priv_data[2];
428 };
429 
430 static bool is_canonical_address(u64 addr)
431 {
432 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
433 	long saddr = (long) addr;
434 
435 	return (((saddr << shift) >> shift) == saddr);
436 }
437 
438 /**
439  * intel_drain_pasid_prq - Drain page requests and responses for a pasid
440  * @dev: target device
441  * @pasid: pasid for draining
442  *
443  * Drain all pending page requests and responses related to @pasid in both
444  * software and hardware. This is supposed to be called after the device
445  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
446  * and DevTLB have been invalidated.
447  *
448  * It waits until all pending page requests for @pasid in the page fault
449  * queue are completed by the prq handling thread. Then follow the steps
450  * described in VT-d spec CH7.10 to drain all page requests and page
451  * responses pending in the hardware.
452  */
453 void intel_drain_pasid_prq(struct device *dev, u32 pasid)
454 {
455 	struct device_domain_info *info;
456 	struct dmar_domain *domain;
457 	struct intel_iommu *iommu;
458 	struct qi_desc desc[3];
459 	struct pci_dev *pdev;
460 	int head, tail;
461 	u16 sid, did;
462 	int qdep;
463 
464 	info = dev_iommu_priv_get(dev);
465 	if (WARN_ON(!info || !dev_is_pci(dev)))
466 		return;
467 
468 	if (!info->pri_enabled)
469 		return;
470 
471 	iommu = info->iommu;
472 	domain = info->domain;
473 	pdev = to_pci_dev(dev);
474 	sid = PCI_DEVID(info->bus, info->devfn);
475 	did = domain_id_iommu(domain, iommu);
476 	qdep = pci_ats_queue_depth(pdev);
477 
478 	/*
479 	 * Check and wait until all pending page requests in the queue are
480 	 * handled by the prq handling thread.
481 	 */
482 prq_retry:
483 	reinit_completion(&iommu->prq_complete);
484 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
485 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
486 	while (head != tail) {
487 		struct page_req_dsc *req;
488 
489 		req = &iommu->prq[head / sizeof(*req)];
490 		if (!req->pasid_present || req->pasid != pasid) {
491 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
492 			continue;
493 		}
494 
495 		wait_for_completion(&iommu->prq_complete);
496 		goto prq_retry;
497 	}
498 
499 	iopf_queue_flush_dev(dev);
500 
501 	/*
502 	 * Perform steps described in VT-d spec CH7.10 to drain page
503 	 * requests and responses in hardware.
504 	 */
505 	memset(desc, 0, sizeof(desc));
506 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
507 			QI_IWD_FENCE |
508 			QI_IWD_TYPE;
509 	desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
510 			QI_EIOTLB_DID(did) |
511 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
512 			QI_EIOTLB_TYPE;
513 	desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
514 			QI_DEV_EIOTLB_SID(sid) |
515 			QI_DEV_EIOTLB_QDEP(qdep) |
516 			QI_DEIOTLB_TYPE |
517 			QI_DEV_IOTLB_PFSID(info->pfsid);
518 qi_retry:
519 	reinit_completion(&iommu->prq_complete);
520 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
521 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
522 		wait_for_completion(&iommu->prq_complete);
523 		goto qi_retry;
524 	}
525 }
526 
527 static int prq_to_iommu_prot(struct page_req_dsc *req)
528 {
529 	int prot = 0;
530 
531 	if (req->rd_req)
532 		prot |= IOMMU_FAULT_PERM_READ;
533 	if (req->wr_req)
534 		prot |= IOMMU_FAULT_PERM_WRITE;
535 	if (req->exe_req)
536 		prot |= IOMMU_FAULT_PERM_EXEC;
537 	if (req->pm_req)
538 		prot |= IOMMU_FAULT_PERM_PRIV;
539 
540 	return prot;
541 }
542 
543 static int intel_svm_prq_report(struct intel_iommu *iommu, struct device *dev,
544 				struct page_req_dsc *desc)
545 {
546 	struct iommu_fault_event event;
547 
548 	if (!dev || !dev_is_pci(dev))
549 		return -ENODEV;
550 
551 	/* Fill in event data for device specific processing */
552 	memset(&event, 0, sizeof(struct iommu_fault_event));
553 	event.fault.type = IOMMU_FAULT_PAGE_REQ;
554 	event.fault.prm.addr = (u64)desc->addr << VTD_PAGE_SHIFT;
555 	event.fault.prm.pasid = desc->pasid;
556 	event.fault.prm.grpid = desc->prg_index;
557 	event.fault.prm.perm = prq_to_iommu_prot(desc);
558 
559 	if (desc->lpig)
560 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
561 	if (desc->pasid_present) {
562 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
563 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
564 	}
565 	if (desc->priv_data_present) {
566 		/*
567 		 * Set last page in group bit if private data is present,
568 		 * page response is required as it does for LPIG.
569 		 * iommu_report_device_fault() doesn't understand this vendor
570 		 * specific requirement thus we set last_page as a workaround.
571 		 */
572 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
573 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
574 		event.fault.prm.private_data[0] = desc->priv_data[0];
575 		event.fault.prm.private_data[1] = desc->priv_data[1];
576 	} else if (dmar_latency_enabled(iommu, DMAR_LATENCY_PRQ)) {
577 		/*
578 		 * If the private data fields are not used by hardware, use it
579 		 * to monitor the prq handle latency.
580 		 */
581 		event.fault.prm.private_data[0] = ktime_to_ns(ktime_get());
582 	}
583 
584 	return iommu_report_device_fault(dev, &event);
585 }
586 
587 static void handle_bad_prq_event(struct intel_iommu *iommu,
588 				 struct page_req_dsc *req, int result)
589 {
590 	struct qi_desc desc;
591 
592 	pr_err("%s: Invalid page request: %08llx %08llx\n",
593 	       iommu->name, ((unsigned long long *)req)[0],
594 	       ((unsigned long long *)req)[1]);
595 
596 	/*
597 	 * Per VT-d spec. v3.0 ch7.7, system software must
598 	 * respond with page group response if private data
599 	 * is present (PDP) or last page in group (LPIG) bit
600 	 * is set. This is an additional VT-d feature beyond
601 	 * PCI ATS spec.
602 	 */
603 	if (!req->lpig && !req->priv_data_present)
604 		return;
605 
606 	desc.qw0 = QI_PGRP_PASID(req->pasid) |
607 			QI_PGRP_DID(req->rid) |
608 			QI_PGRP_PASID_P(req->pasid_present) |
609 			QI_PGRP_PDP(req->priv_data_present) |
610 			QI_PGRP_RESP_CODE(result) |
611 			QI_PGRP_RESP_TYPE;
612 	desc.qw1 = QI_PGRP_IDX(req->prg_index) |
613 			QI_PGRP_LPIG(req->lpig);
614 
615 	if (req->priv_data_present) {
616 		desc.qw2 = req->priv_data[0];
617 		desc.qw3 = req->priv_data[1];
618 	} else {
619 		desc.qw2 = 0;
620 		desc.qw3 = 0;
621 	}
622 
623 	qi_submit_sync(iommu, &desc, 1, 0);
624 }
625 
626 static irqreturn_t prq_event_thread(int irq, void *d)
627 {
628 	struct intel_iommu *iommu = d;
629 	struct page_req_dsc *req;
630 	int head, tail, handled;
631 	struct pci_dev *pdev;
632 	u64 address;
633 
634 	/*
635 	 * Clear PPR bit before reading head/tail registers, to ensure that
636 	 * we get a new interrupt if needed.
637 	 */
638 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
639 
640 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
641 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
642 	handled = (head != tail);
643 	while (head != tail) {
644 		req = &iommu->prq[head / sizeof(*req)];
645 		address = (u64)req->addr << VTD_PAGE_SHIFT;
646 
647 		if (unlikely(!req->pasid_present)) {
648 			pr_err("IOMMU: %s: Page request without PASID\n",
649 			       iommu->name);
650 bad_req:
651 			handle_bad_prq_event(iommu, req, QI_RESP_INVALID);
652 			goto prq_advance;
653 		}
654 
655 		if (unlikely(!is_canonical_address(address))) {
656 			pr_err("IOMMU: %s: Address is not canonical\n",
657 			       iommu->name);
658 			goto bad_req;
659 		}
660 
661 		if (unlikely(req->pm_req && (req->rd_req | req->wr_req))) {
662 			pr_err("IOMMU: %s: Page request in Privilege Mode\n",
663 			       iommu->name);
664 			goto bad_req;
665 		}
666 
667 		if (unlikely(req->exe_req && req->rd_req)) {
668 			pr_err("IOMMU: %s: Execution request not supported\n",
669 			       iommu->name);
670 			goto bad_req;
671 		}
672 
673 		/* Drop Stop Marker message. No need for a response. */
674 		if (unlikely(req->lpig && !req->rd_req && !req->wr_req))
675 			goto prq_advance;
676 
677 		pdev = pci_get_domain_bus_and_slot(iommu->segment,
678 						   PCI_BUS_NUM(req->rid),
679 						   req->rid & 0xff);
680 		/*
681 		 * If prq is to be handled outside iommu driver via receiver of
682 		 * the fault notifiers, we skip the page response here.
683 		 */
684 		if (!pdev)
685 			goto bad_req;
686 
687 		if (intel_svm_prq_report(iommu, &pdev->dev, req))
688 			handle_bad_prq_event(iommu, req, QI_RESP_INVALID);
689 		else
690 			trace_prq_report(iommu, &pdev->dev, req->qw_0, req->qw_1,
691 					 req->priv_data[0], req->priv_data[1],
692 					 iommu->prq_seq_number++);
693 		pci_dev_put(pdev);
694 prq_advance:
695 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
696 	}
697 
698 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
699 
700 	/*
701 	 * Clear the page request overflow bit and wake up all threads that
702 	 * are waiting for the completion of this handling.
703 	 */
704 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
705 		pr_info_ratelimited("IOMMU: %s: PRQ overflow detected\n",
706 				    iommu->name);
707 		head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
708 		tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
709 		if (head == tail) {
710 			iopf_queue_discard_partial(iommu->iopf_queue);
711 			writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
712 			pr_info_ratelimited("IOMMU: %s: PRQ overflow cleared",
713 					    iommu->name);
714 		}
715 	}
716 
717 	if (!completion_done(&iommu->prq_complete))
718 		complete(&iommu->prq_complete);
719 
720 	return IRQ_RETVAL(handled);
721 }
722 
723 int intel_svm_page_response(struct device *dev,
724 			    struct iommu_fault_event *evt,
725 			    struct iommu_page_response *msg)
726 {
727 	struct iommu_fault_page_request *prm;
728 	struct intel_iommu *iommu;
729 	bool private_present;
730 	bool pasid_present;
731 	bool last_page;
732 	u8 bus, devfn;
733 	int ret = 0;
734 	u16 sid;
735 
736 	if (!dev || !dev_is_pci(dev))
737 		return -ENODEV;
738 
739 	iommu = device_to_iommu(dev, &bus, &devfn);
740 	if (!iommu)
741 		return -ENODEV;
742 
743 	if (!msg || !evt)
744 		return -EINVAL;
745 
746 	prm = &evt->fault.prm;
747 	sid = PCI_DEVID(bus, devfn);
748 	pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
749 	private_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
750 	last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
751 
752 	if (!pasid_present) {
753 		ret = -EINVAL;
754 		goto out;
755 	}
756 
757 	if (prm->pasid == 0 || prm->pasid >= PASID_MAX) {
758 		ret = -EINVAL;
759 		goto out;
760 	}
761 
762 	/*
763 	 * Per VT-d spec. v3.0 ch7.7, system software must respond
764 	 * with page group response if private data is present (PDP)
765 	 * or last page in group (LPIG) bit is set. This is an
766 	 * additional VT-d requirement beyond PCI ATS spec.
767 	 */
768 	if (last_page || private_present) {
769 		struct qi_desc desc;
770 
771 		desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
772 				QI_PGRP_PASID_P(pasid_present) |
773 				QI_PGRP_PDP(private_present) |
774 				QI_PGRP_RESP_CODE(msg->code) |
775 				QI_PGRP_RESP_TYPE;
776 		desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
777 		desc.qw2 = 0;
778 		desc.qw3 = 0;
779 
780 		if (private_present) {
781 			desc.qw2 = prm->private_data[0];
782 			desc.qw3 = prm->private_data[1];
783 		} else if (prm->private_data[0]) {
784 			dmar_latency_update(iommu, DMAR_LATENCY_PRQ,
785 				ktime_to_ns(ktime_get()) - prm->private_data[0]);
786 		}
787 
788 		qi_submit_sync(iommu, &desc, 1, 0);
789 	}
790 out:
791 	return ret;
792 }
793 
794 static int intel_svm_set_dev_pasid(struct iommu_domain *domain,
795 				   struct device *dev, ioasid_t pasid)
796 {
797 	struct device_domain_info *info = dev_iommu_priv_get(dev);
798 	struct intel_iommu *iommu = info->iommu;
799 	struct mm_struct *mm = domain->mm;
800 
801 	return intel_svm_bind_mm(iommu, dev, mm);
802 }
803 
804 static void intel_svm_domain_free(struct iommu_domain *domain)
805 {
806 	kfree(to_dmar_domain(domain));
807 }
808 
809 static const struct iommu_domain_ops intel_svm_domain_ops = {
810 	.set_dev_pasid		= intel_svm_set_dev_pasid,
811 	.free			= intel_svm_domain_free
812 };
813 
814 struct iommu_domain *intel_svm_domain_alloc(void)
815 {
816 	struct dmar_domain *domain;
817 
818 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
819 	if (!domain)
820 		return NULL;
821 	domain->domain.ops = &intel_svm_domain_ops;
822 
823 	return &domain->domain;
824 }
825