• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2015 Intel Corporation.
3  *
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.
7  *
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
11  * more details.
12  *
13  * Authors: David Woodhouse <dwmw2@infradead.org>
14  */
15 
16 #include <linux/intel-iommu.h>
17 #include <linux/mmu_notifier.h>
18 #include <linux/sched.h>
19 #include <linux/slab.h>
20 #include <linux/intel-svm.h>
21 #include <linux/rculist.h>
22 #include <linux/pci.h>
23 #include <linux/pci-ats.h>
24 #include <linux/dmar.h>
25 #include <linux/interrupt.h>
26 
27 static irqreturn_t prq_event_thread(int irq, void *d);
28 
29 struct pasid_entry {
30 	u64 val;
31 };
32 
33 struct pasid_state_entry {
34 	u64 val;
35 };
36 
intel_svm_alloc_pasid_tables(struct intel_iommu * iommu)37 int intel_svm_alloc_pasid_tables(struct intel_iommu *iommu)
38 {
39 	struct page *pages;
40 	int order;
41 
42 	/* Start at 2 because it's defined as 2^(1+PSS) */
43 	iommu->pasid_max = 2 << ecap_pss(iommu->ecap);
44 
45 	/* Eventually I'm promised we will get a multi-level PASID table
46 	 * and it won't have to be physically contiguous. Until then,
47 	 * limit the size because 8MiB contiguous allocations can be hard
48 	 * to come by. The limit of 0x20000, which is 1MiB for each of
49 	 * the PASID and PASID-state tables, is somewhat arbitrary. */
50 	if (iommu->pasid_max > 0x20000)
51 		iommu->pasid_max = 0x20000;
52 
53 	order = get_order(sizeof(struct pasid_entry) * iommu->pasid_max);
54 	pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
55 	if (!pages) {
56 		pr_warn("IOMMU: %s: Failed to allocate PASID table\n",
57 			iommu->name);
58 		return -ENOMEM;
59 	}
60 	iommu->pasid_table = page_address(pages);
61 	pr_info("%s: Allocated order %d PASID table.\n", iommu->name, order);
62 
63 	if (ecap_dis(iommu->ecap)) {
64 		/* Just making it explicit... */
65 		BUILD_BUG_ON(sizeof(struct pasid_entry) != sizeof(struct pasid_state_entry));
66 		pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
67 		if (pages)
68 			iommu->pasid_state_table = page_address(pages);
69 		else
70 			pr_warn("IOMMU: %s: Failed to allocate PASID state table\n",
71 				iommu->name);
72 	}
73 
74 	idr_init(&iommu->pasid_idr);
75 
76 	return 0;
77 }
78 
intel_svm_free_pasid_tables(struct intel_iommu * iommu)79 int intel_svm_free_pasid_tables(struct intel_iommu *iommu)
80 {
81 	int order = get_order(sizeof(struct pasid_entry) * iommu->pasid_max);
82 
83 	if (iommu->pasid_table) {
84 		free_pages((unsigned long)iommu->pasid_table, order);
85 		iommu->pasid_table = NULL;
86 	}
87 	if (iommu->pasid_state_table) {
88 		free_pages((unsigned long)iommu->pasid_state_table, order);
89 		iommu->pasid_state_table = NULL;
90 	}
91 	idr_destroy(&iommu->pasid_idr);
92 	return 0;
93 }
94 
95 #define PRQ_ORDER 0
96 
intel_svm_enable_prq(struct intel_iommu * iommu)97 int intel_svm_enable_prq(struct intel_iommu *iommu)
98 {
99 	struct page *pages;
100 	int irq, ret;
101 
102 	pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
103 	if (!pages) {
104 		pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
105 			iommu->name);
106 		return -ENOMEM;
107 	}
108 	iommu->prq = page_address(pages);
109 
110 	irq = dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED + iommu->seq_id, iommu->node, iommu);
111 	if (irq <= 0) {
112 		pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
113 		       iommu->name);
114 		ret = -EINVAL;
115 	err:
116 		free_pages((unsigned long)iommu->prq, PRQ_ORDER);
117 		iommu->prq = NULL;
118 		return ret;
119 	}
120 	iommu->pr_irq = irq;
121 
122 	snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
123 
124 	ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
125 				   iommu->prq_name, iommu);
126 	if (ret) {
127 		pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
128 		       iommu->name);
129 		dmar_free_hwirq(irq);
130 		iommu->pr_irq = 0;
131 		goto err;
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, virt_to_phys(iommu->prq) | PRQ_ORDER);
136 
137 	return 0;
138 }
139 
intel_svm_finish_prq(struct intel_iommu * iommu)140 int intel_svm_finish_prq(struct intel_iommu *iommu)
141 {
142 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
143 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
144 	dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
145 
146 	if (iommu->pr_irq) {
147 		free_irq(iommu->pr_irq, iommu);
148 		dmar_free_hwirq(iommu->pr_irq);
149 		iommu->pr_irq = 0;
150 	}
151 
152 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
153 	iommu->prq = NULL;
154 
155 	return 0;
156 }
157 
intel_flush_svm_range_dev(struct intel_svm * svm,struct intel_svm_dev * sdev,unsigned long address,unsigned long pages,int ih,int gl)158 static void intel_flush_svm_range_dev (struct intel_svm *svm, struct intel_svm_dev *sdev,
159 				       unsigned long address, unsigned long pages, int ih, int gl)
160 {
161 	struct qi_desc desc;
162 
163 	if (pages == -1) {
164 		/* For global kernel pages we have to flush them in *all* PASIDs
165 		 * because that's the only option the hardware gives us. Despite
166 		 * the fact that they are actually only accessible through one. */
167 		if (gl)
168 			desc.low = QI_EIOTLB_PASID(svm->pasid) | QI_EIOTLB_DID(sdev->did) |
169 				QI_EIOTLB_GRAN(QI_GRAN_ALL_ALL) | QI_EIOTLB_TYPE;
170 		else
171 			desc.low = QI_EIOTLB_PASID(svm->pasid) | QI_EIOTLB_DID(sdev->did) |
172 				QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) | QI_EIOTLB_TYPE;
173 		desc.high = 0;
174 	} else {
175 		int mask = ilog2(__roundup_pow_of_two(pages));
176 
177 		desc.low = QI_EIOTLB_PASID(svm->pasid) | QI_EIOTLB_DID(sdev->did) |
178 			QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID) | QI_EIOTLB_TYPE;
179 		desc.high = QI_EIOTLB_ADDR(address) | QI_EIOTLB_GL(gl) |
180 			QI_EIOTLB_IH(ih) | QI_EIOTLB_AM(mask);
181 	}
182 	qi_submit_sync(&desc, svm->iommu);
183 
184 	if (sdev->dev_iotlb) {
185 		desc.low = QI_DEV_EIOTLB_PASID(svm->pasid) | QI_DEV_EIOTLB_SID(sdev->sid) |
186 			QI_DEV_EIOTLB_QDEP(sdev->qdep) | QI_DEIOTLB_TYPE;
187 		if (pages == -1) {
188 			desc.high = QI_DEV_EIOTLB_ADDR(-1ULL >> 1) | QI_DEV_EIOTLB_SIZE;
189 		} else if (pages > 1) {
190 			/* The least significant zero bit indicates the size. So,
191 			 * for example, an "address" value of 0x12345f000 will
192 			 * flush from 0x123440000 to 0x12347ffff (256KiB). */
193 			unsigned long last = address + ((unsigned long)(pages - 1) << VTD_PAGE_SHIFT);
194 			unsigned long mask = __rounddown_pow_of_two(address ^ last);;
195 
196 			desc.high = QI_DEV_EIOTLB_ADDR((address & ~mask) | (mask - 1)) | QI_DEV_EIOTLB_SIZE;
197 		} else {
198 			desc.high = QI_DEV_EIOTLB_ADDR(address);
199 		}
200 		qi_submit_sync(&desc, svm->iommu);
201 	}
202 }
203 
intel_flush_svm_range(struct intel_svm * svm,unsigned long address,unsigned long pages,int ih,int gl)204 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
205 				  unsigned long pages, int ih, int gl)
206 {
207 	struct intel_svm_dev *sdev;
208 
209 	/* Try deferred invalidate if available */
210 	if (svm->iommu->pasid_state_table &&
211 	    !cmpxchg64(&svm->iommu->pasid_state_table[svm->pasid].val, 0, 1ULL << 63))
212 		return;
213 
214 	rcu_read_lock();
215 	list_for_each_entry_rcu(sdev, &svm->devs, list)
216 		intel_flush_svm_range_dev(svm, sdev, address, pages, ih, gl);
217 	rcu_read_unlock();
218 }
219 
intel_change_pte(struct mmu_notifier * mn,struct mm_struct * mm,unsigned long address,pte_t pte)220 static void intel_change_pte(struct mmu_notifier *mn, struct mm_struct *mm,
221 			     unsigned long address, pte_t pte)
222 {
223 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
224 
225 	intel_flush_svm_range(svm, address, 1, 1, 0);
226 }
227 
intel_invalidate_page(struct mmu_notifier * mn,struct mm_struct * mm,unsigned long address)228 static void intel_invalidate_page(struct mmu_notifier *mn, struct mm_struct *mm,
229 				  unsigned long address)
230 {
231 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
232 
233 	intel_flush_svm_range(svm, address, 1, 1, 0);
234 }
235 
236 /* Pages have been freed at this point */
intel_invalidate_range(struct mmu_notifier * mn,struct mm_struct * mm,unsigned long start,unsigned long end)237 static void intel_invalidate_range(struct mmu_notifier *mn,
238 				   struct mm_struct *mm,
239 				   unsigned long start, unsigned long end)
240 {
241 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
242 
243 	intel_flush_svm_range(svm, start,
244 			      (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0, 0);
245 }
246 
247 
intel_flush_pasid_dev(struct intel_svm * svm,struct intel_svm_dev * sdev,int pasid)248 static void intel_flush_pasid_dev(struct intel_svm *svm, struct intel_svm_dev *sdev, int pasid)
249 {
250 	struct qi_desc desc;
251 
252 	desc.high = 0;
253 	desc.low = QI_PC_TYPE | QI_PC_DID(sdev->did) | QI_PC_PASID_SEL | QI_PC_PASID(pasid);
254 
255 	qi_submit_sync(&desc, svm->iommu);
256 }
257 
intel_mm_release(struct mmu_notifier * mn,struct mm_struct * mm)258 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
259 {
260 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
261 	struct intel_svm_dev *sdev;
262 
263 	/* This might end up being called from exit_mmap(), *before* the page
264 	 * tables are cleared. And __mmu_notifier_release() will delete us from
265 	 * the list of notifiers so that our invalidate_range() callback doesn't
266 	 * get called when the page tables are cleared. So we need to protect
267 	 * against hardware accessing those page tables.
268 	 *
269 	 * We do it by clearing the entry in the PASID table and then flushing
270 	 * the IOTLB and the PASID table caches. This might upset hardware;
271 	 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
272 	 * page) so that we end up taking a fault that the hardware really
273 	 * *has* to handle gracefully without affecting other processes.
274 	 */
275 	svm->iommu->pasid_table[svm->pasid].val = 0;
276 	wmb();
277 
278 	rcu_read_lock();
279 	list_for_each_entry_rcu(sdev, &svm->devs, list) {
280 		intel_flush_pasid_dev(svm, sdev, svm->pasid);
281 		intel_flush_svm_range_dev(svm, sdev, 0, -1, 0, !svm->mm);
282 	}
283 	rcu_read_unlock();
284 
285 }
286 
287 static const struct mmu_notifier_ops intel_mmuops = {
288 	.release = intel_mm_release,
289 	.change_pte = intel_change_pte,
290 	.invalidate_page = intel_invalidate_page,
291 	.invalidate_range = intel_invalidate_range,
292 };
293 
294 static DEFINE_MUTEX(pasid_mutex);
295 
intel_svm_bind_mm(struct device * dev,int * pasid,int flags,struct svm_dev_ops * ops)296 int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_ops *ops)
297 {
298 	struct intel_iommu *iommu = intel_svm_device_to_iommu(dev);
299 	struct intel_svm_dev *sdev;
300 	struct intel_svm *svm = NULL;
301 	struct mm_struct *mm = NULL;
302 	int pasid_max;
303 	int ret;
304 
305 	if (WARN_ON(!iommu))
306 		return -EINVAL;
307 
308 	if (dev_is_pci(dev)) {
309 		pasid_max = pci_max_pasids(to_pci_dev(dev));
310 		if (pasid_max < 0)
311 			return -EINVAL;
312 	} else
313 		pasid_max = 1 << 20;
314 
315 	if ((flags & SVM_FLAG_SUPERVISOR_MODE)) {
316 		if (!ecap_srs(iommu->ecap))
317 			return -EINVAL;
318 	} else if (pasid) {
319 		mm = get_task_mm(current);
320 		BUG_ON(!mm);
321 	}
322 
323 	mutex_lock(&pasid_mutex);
324 	if (pasid && !(flags & SVM_FLAG_PRIVATE_PASID)) {
325 		int i;
326 
327 		idr_for_each_entry(&iommu->pasid_idr, svm, i) {
328 			if (svm->mm != mm ||
329 			    (svm->flags & SVM_FLAG_PRIVATE_PASID))
330 				continue;
331 
332 			if (svm->pasid >= pasid_max) {
333 				dev_warn(dev,
334 					 "Limited PASID width. Cannot use existing PASID %d\n",
335 					 svm->pasid);
336 				ret = -ENOSPC;
337 				goto out;
338 			}
339 
340 			list_for_each_entry(sdev, &svm->devs, list) {
341 				if (dev == sdev->dev) {
342 					if (sdev->ops != ops) {
343 						ret = -EBUSY;
344 						goto out;
345 					}
346 					sdev->users++;
347 					goto success;
348 				}
349 			}
350 
351 			break;
352 		}
353 	}
354 
355 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
356 	if (!sdev) {
357 		ret = -ENOMEM;
358 		goto out;
359 	}
360 	sdev->dev = dev;
361 
362 	ret = intel_iommu_enable_pasid(iommu, sdev);
363 	if (ret || !pasid) {
364 		/* If they don't actually want to assign a PASID, this is
365 		 * just an enabling check/preparation. */
366 		kfree(sdev);
367 		goto out;
368 	}
369 	/* Finish the setup now we know we're keeping it */
370 	sdev->users = 1;
371 	sdev->ops = ops;
372 	init_rcu_head(&sdev->rcu);
373 
374 	if (!svm) {
375 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
376 		if (!svm) {
377 			ret = -ENOMEM;
378 			kfree(sdev);
379 			goto out;
380 		}
381 		svm->iommu = iommu;
382 
383 		if (pasid_max > iommu->pasid_max)
384 			pasid_max = iommu->pasid_max;
385 
386 		/* Do not use PASID 0 in caching mode (virtualised IOMMU) */
387 		ret = idr_alloc(&iommu->pasid_idr, svm,
388 				!!cap_caching_mode(iommu->cap),
389 				pasid_max - 1, GFP_KERNEL);
390 		if (ret < 0) {
391 			kfree(svm);
392 			kfree(sdev);
393 			goto out;
394 		}
395 		svm->pasid = ret;
396 		svm->notifier.ops = &intel_mmuops;
397 		svm->mm = mm;
398 		svm->flags = flags;
399 		INIT_LIST_HEAD_RCU(&svm->devs);
400 		ret = -ENOMEM;
401 		if (mm) {
402 			ret = mmu_notifier_register(&svm->notifier, mm);
403 			if (ret) {
404 				idr_remove(&svm->iommu->pasid_idr, svm->pasid);
405 				kfree(svm);
406 				kfree(sdev);
407 				goto out;
408 			}
409 			iommu->pasid_table[svm->pasid].val = (u64)__pa(mm->pgd) | 1;
410 		} else
411 			iommu->pasid_table[svm->pasid].val = (u64)__pa(init_mm.pgd) | 1 | (1ULL << 11);
412 		wmb();
413 		/* In caching mode, we still have to flush with PASID 0 when
414 		 * a PASID table entry becomes present. Not entirely clear
415 		 * *why* that would be the case — surely we could just issue
416 		 * a flush with the PASID value that we've changed? The PASID
417 		 * is the index into the table, after all. It's not like domain
418 		 * IDs in the case of the equivalent context-entry change in
419 		 * caching mode. And for that matter it's not entirely clear why
420 		 * a VMM would be in the business of caching the PASID table
421 		 * anyway. Surely that can be left entirely to the guest? */
422 		if (cap_caching_mode(iommu->cap))
423 			intel_flush_pasid_dev(svm, sdev, 0);
424 	}
425 	list_add_rcu(&sdev->list, &svm->devs);
426 
427  success:
428 	*pasid = svm->pasid;
429 	ret = 0;
430  out:
431 	mutex_unlock(&pasid_mutex);
432 	if (mm)
433 		mmput(mm);
434 	return ret;
435 }
436 EXPORT_SYMBOL_GPL(intel_svm_bind_mm);
437 
intel_svm_unbind_mm(struct device * dev,int pasid)438 int intel_svm_unbind_mm(struct device *dev, int pasid)
439 {
440 	struct intel_svm_dev *sdev;
441 	struct intel_iommu *iommu;
442 	struct intel_svm *svm;
443 	int ret = -EINVAL;
444 
445 	mutex_lock(&pasid_mutex);
446 	iommu = intel_svm_device_to_iommu(dev);
447 	if (!iommu || !iommu->pasid_table)
448 		goto out;
449 
450 	svm = idr_find(&iommu->pasid_idr, pasid);
451 	if (!svm)
452 		goto out;
453 
454 	list_for_each_entry(sdev, &svm->devs, list) {
455 		if (dev == sdev->dev) {
456 			ret = 0;
457 			sdev->users--;
458 			if (!sdev->users) {
459 				list_del_rcu(&sdev->list);
460 				/* Flush the PASID cache and IOTLB for this device.
461 				 * Note that we do depend on the hardware *not* using
462 				 * the PASID any more. Just as we depend on other
463 				 * devices never using PASIDs that they have no right
464 				 * to use. We have a *shared* PASID table, because it's
465 				 * large and has to be physically contiguous. So it's
466 				 * hard to be as defensive as we might like. */
467 				intel_flush_pasid_dev(svm, sdev, svm->pasid);
468 				intel_flush_svm_range_dev(svm, sdev, 0, -1, 0, !svm->mm);
469 				kfree_rcu(sdev, rcu);
470 
471 				if (list_empty(&svm->devs)) {
472 
473 					idr_remove(&svm->iommu->pasid_idr, svm->pasid);
474 					if (svm->mm)
475 						mmu_notifier_unregister(&svm->notifier, svm->mm);
476 
477 					/* We mandate that no page faults may be outstanding
478 					 * for the PASID when intel_svm_unbind_mm() is called.
479 					 * If that is not obeyed, subtle errors will happen.
480 					 * Let's make them less subtle... */
481 					memset(svm, 0x6b, sizeof(*svm));
482 					kfree(svm);
483 				}
484 			}
485 			break;
486 		}
487 	}
488  out:
489 	mutex_unlock(&pasid_mutex);
490 
491 	return ret;
492 }
493 EXPORT_SYMBOL_GPL(intel_svm_unbind_mm);
494 
495 /* Page request queue descriptor */
496 struct page_req_dsc {
497 	u64 srr:1;
498 	u64 bof:1;
499 	u64 pasid_present:1;
500 	u64 lpig:1;
501 	u64 pasid:20;
502 	u64 bus:8;
503 	u64 private:23;
504 	u64 prg_index:9;
505 	u64 rd_req:1;
506 	u64 wr_req:1;
507 	u64 exe_req:1;
508 	u64 priv_req:1;
509 	u64 devfn:8;
510 	u64 addr:52;
511 };
512 
513 #define PRQ_RING_MASK ((0x1000 << PRQ_ORDER) - 0x10)
514 
access_error(struct vm_area_struct * vma,struct page_req_dsc * req)515 static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
516 {
517 	unsigned long requested = 0;
518 
519 	if (req->exe_req)
520 		requested |= VM_EXEC;
521 
522 	if (req->rd_req)
523 		requested |= VM_READ;
524 
525 	if (req->wr_req)
526 		requested |= VM_WRITE;
527 
528 	return (requested & ~vma->vm_flags) != 0;
529 }
530 
prq_event_thread(int irq,void * d)531 static irqreturn_t prq_event_thread(int irq, void *d)
532 {
533 	struct intel_iommu *iommu = d;
534 	struct intel_svm *svm = NULL;
535 	int head, tail, handled = 0;
536 
537 	/* Clear PPR bit before reading head/tail registers, to
538 	 * ensure that we get a new interrupt if needed. */
539 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
540 
541 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
542 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
543 	while (head != tail) {
544 		struct intel_svm_dev *sdev;
545 		struct vm_area_struct *vma;
546 		struct page_req_dsc *req;
547 		struct qi_desc resp;
548 		int ret, result;
549 		u64 address;
550 
551 		handled = 1;
552 
553 		req = &iommu->prq[head / sizeof(*req)];
554 
555 		result = QI_RESP_FAILURE;
556 		address = (u64)req->addr << VTD_PAGE_SHIFT;
557 		if (!req->pasid_present) {
558 			pr_err("%s: Page request without PASID: %08llx %08llx\n",
559 			       iommu->name, ((unsigned long long *)req)[0],
560 			       ((unsigned long long *)req)[1]);
561 			goto bad_req;
562 		}
563 
564 		if (!svm || svm->pasid != req->pasid) {
565 			rcu_read_lock();
566 			svm = idr_find(&iommu->pasid_idr, req->pasid);
567 			/* It *can't* go away, because the driver is not permitted
568 			 * to unbind the mm while any page faults are outstanding.
569 			 * So we only need RCU to protect the internal idr code. */
570 			rcu_read_unlock();
571 
572 			if (!svm) {
573 				pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
574 				       iommu->name, req->pasid, ((unsigned long long *)req)[0],
575 				       ((unsigned long long *)req)[1]);
576 				goto no_pasid;
577 			}
578 		}
579 
580 		result = QI_RESP_INVALID;
581 		/* Since we're using init_mm.pgd directly, we should never take
582 		 * any faults on kernel addresses. */
583 		if (!svm->mm)
584 			goto bad_req;
585 		/* If the mm is already defunct, don't handle faults. */
586 		if (!atomic_inc_not_zero(&svm->mm->mm_users))
587 			goto bad_req;
588 		down_read(&svm->mm->mmap_sem);
589 		vma = find_extend_vma(svm->mm, address);
590 		if (!vma || address < vma->vm_start)
591 			goto invalid;
592 
593 		if (access_error(vma, req))
594 			goto invalid;
595 
596 		ret = handle_mm_fault(vma, address,
597 				      req->wr_req ? FAULT_FLAG_WRITE : 0);
598 		if (ret & VM_FAULT_ERROR)
599 			goto invalid;
600 
601 		result = QI_RESP_SUCCESS;
602 	invalid:
603 		up_read(&svm->mm->mmap_sem);
604 		mmput(svm->mm);
605 	bad_req:
606 		/* Accounting for major/minor faults? */
607 		rcu_read_lock();
608 		list_for_each_entry_rcu(sdev, &svm->devs, list) {
609 			if (sdev->sid == PCI_DEVID(req->bus, req->devfn))
610 				break;
611 		}
612 		/* Other devices can go away, but the drivers are not permitted
613 		 * to unbind while any page faults might be in flight. So it's
614 		 * OK to drop the 'lock' here now we have it. */
615 		rcu_read_unlock();
616 
617 		if (WARN_ON(&sdev->list == &svm->devs))
618 			sdev = NULL;
619 
620 		if (sdev && sdev->ops && sdev->ops->fault_cb) {
621 			int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
622 				(req->exe_req << 1) | (req->priv_req);
623 			sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr, req->private, rwxp, result);
624 		}
625 		/* We get here in the error case where the PASID lookup failed,
626 		   and these can be NULL. Do not use them below this point! */
627 		sdev = NULL;
628 		svm = NULL;
629 	no_pasid:
630 		if (req->lpig) {
631 			/* Page Group Response */
632 			resp.low = QI_PGRP_PASID(req->pasid) |
633 				QI_PGRP_DID((req->bus << 8) | req->devfn) |
634 				QI_PGRP_PASID_P(req->pasid_present) |
635 				QI_PGRP_RESP_TYPE;
636 			resp.high = QI_PGRP_IDX(req->prg_index) |
637 				QI_PGRP_PRIV(req->private) | QI_PGRP_RESP_CODE(result);
638 
639 			qi_submit_sync(&resp, iommu);
640 		} else if (req->srr) {
641 			/* Page Stream Response */
642 			resp.low = QI_PSTRM_IDX(req->prg_index) |
643 				QI_PSTRM_PRIV(req->private) | QI_PSTRM_BUS(req->bus) |
644 				QI_PSTRM_PASID(req->pasid) | QI_PSTRM_RESP_TYPE;
645 			resp.high = QI_PSTRM_ADDR(address) | QI_PSTRM_DEVFN(req->devfn) |
646 				QI_PSTRM_RESP_CODE(result);
647 
648 			qi_submit_sync(&resp, iommu);
649 		}
650 
651 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
652 	}
653 
654 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
655 
656 	return IRQ_RETVAL(handled);
657 }
658