xref: /linux/drivers/iommu/intel/svm.c (revision cea0f76a483d1270ac6f6513964e3e75193dda48)
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 <linux/ioasid.h>
21 #include <asm/page.h>
22 
23 #include "intel-pasid.h"
24 
25 static irqreturn_t prq_event_thread(int irq, void *d);
26 static void intel_svm_drain_prq(struct device *dev, int pasid);
27 
28 #define PRQ_ORDER 0
29 
30 int intel_svm_enable_prq(struct intel_iommu *iommu)
31 {
32 	struct page *pages;
33 	int irq, ret;
34 
35 	pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
36 	if (!pages) {
37 		pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
38 			iommu->name);
39 		return -ENOMEM;
40 	}
41 	iommu->prq = page_address(pages);
42 
43 	irq = dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED + iommu->seq_id, iommu->node, iommu);
44 	if (irq <= 0) {
45 		pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
46 		       iommu->name);
47 		ret = -EINVAL;
48 	err:
49 		free_pages((unsigned long)iommu->prq, PRQ_ORDER);
50 		iommu->prq = NULL;
51 		return ret;
52 	}
53 	iommu->pr_irq = irq;
54 
55 	snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
56 
57 	ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
58 				   iommu->prq_name, iommu);
59 	if (ret) {
60 		pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
61 		       iommu->name);
62 		dmar_free_hwirq(irq);
63 		iommu->pr_irq = 0;
64 		goto err;
65 	}
66 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
67 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
68 	dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
69 
70 	init_completion(&iommu->prq_complete);
71 
72 	return 0;
73 }
74 
75 int intel_svm_finish_prq(struct intel_iommu *iommu)
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, 0ULL);
80 
81 	if (iommu->pr_irq) {
82 		free_irq(iommu->pr_irq, iommu);
83 		dmar_free_hwirq(iommu->pr_irq);
84 		iommu->pr_irq = 0;
85 	}
86 
87 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
88 	iommu->prq = NULL;
89 
90 	return 0;
91 }
92 
93 static inline bool intel_svm_capable(struct intel_iommu *iommu)
94 {
95 	return iommu->flags & VTD_FLAG_SVM_CAPABLE;
96 }
97 
98 void intel_svm_check(struct intel_iommu *iommu)
99 {
100 	if (!pasid_supported(iommu))
101 		return;
102 
103 	if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
104 	    !cap_fl1gp_support(iommu->cap)) {
105 		pr_err("%s SVM disabled, incompatible 1GB page capability\n",
106 		       iommu->name);
107 		return;
108 	}
109 
110 	if (cpu_feature_enabled(X86_FEATURE_LA57) &&
111 	    !cap_5lp_support(iommu->cap)) {
112 		pr_err("%s SVM disabled, incompatible paging mode\n",
113 		       iommu->name);
114 		return;
115 	}
116 
117 	iommu->flags |= VTD_FLAG_SVM_CAPABLE;
118 }
119 
120 static void intel_flush_svm_range_dev (struct intel_svm *svm, struct intel_svm_dev *sdev,
121 				unsigned long address, unsigned long pages, int ih)
122 {
123 	struct qi_desc desc;
124 
125 	if (pages == -1) {
126 		desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
127 			QI_EIOTLB_DID(sdev->did) |
128 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
129 			QI_EIOTLB_TYPE;
130 		desc.qw1 = 0;
131 	} else {
132 		int mask = ilog2(__roundup_pow_of_two(pages));
133 
134 		desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
135 				QI_EIOTLB_DID(sdev->did) |
136 				QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID) |
137 				QI_EIOTLB_TYPE;
138 		desc.qw1 = QI_EIOTLB_ADDR(address) |
139 				QI_EIOTLB_IH(ih) |
140 				QI_EIOTLB_AM(mask);
141 	}
142 	desc.qw2 = 0;
143 	desc.qw3 = 0;
144 	qi_submit_sync(svm->iommu, &desc, 1, 0);
145 
146 	if (sdev->dev_iotlb) {
147 		desc.qw0 = QI_DEV_EIOTLB_PASID(svm->pasid) |
148 				QI_DEV_EIOTLB_SID(sdev->sid) |
149 				QI_DEV_EIOTLB_QDEP(sdev->qdep) |
150 				QI_DEIOTLB_TYPE;
151 		if (pages == -1) {
152 			desc.qw1 = QI_DEV_EIOTLB_ADDR(-1ULL >> 1) |
153 					QI_DEV_EIOTLB_SIZE;
154 		} else if (pages > 1) {
155 			/* The least significant zero bit indicates the size. So,
156 			 * for example, an "address" value of 0x12345f000 will
157 			 * flush from 0x123440000 to 0x12347ffff (256KiB). */
158 			unsigned long last = address + ((unsigned long)(pages - 1) << VTD_PAGE_SHIFT);
159 			unsigned long mask = __rounddown_pow_of_two(address ^ last);
160 
161 			desc.qw1 = QI_DEV_EIOTLB_ADDR((address & ~mask) |
162 					(mask - 1)) | QI_DEV_EIOTLB_SIZE;
163 		} else {
164 			desc.qw1 = QI_DEV_EIOTLB_ADDR(address);
165 		}
166 		desc.qw2 = 0;
167 		desc.qw3 = 0;
168 		qi_submit_sync(svm->iommu, &desc, 1, 0);
169 	}
170 }
171 
172 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
173 				unsigned long pages, int ih)
174 {
175 	struct intel_svm_dev *sdev;
176 
177 	rcu_read_lock();
178 	list_for_each_entry_rcu(sdev, &svm->devs, list)
179 		intel_flush_svm_range_dev(svm, sdev, address, pages, ih);
180 	rcu_read_unlock();
181 }
182 
183 /* Pages have been freed at this point */
184 static void intel_invalidate_range(struct mmu_notifier *mn,
185 				   struct mm_struct *mm,
186 				   unsigned long start, unsigned long end)
187 {
188 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
189 
190 	intel_flush_svm_range(svm, start,
191 			      (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0);
192 }
193 
194 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
195 {
196 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
197 	struct intel_svm_dev *sdev;
198 
199 	/* This might end up being called from exit_mmap(), *before* the page
200 	 * tables are cleared. And __mmu_notifier_release() will delete us from
201 	 * the list of notifiers so that our invalidate_range() callback doesn't
202 	 * get called when the page tables are cleared. So we need to protect
203 	 * against hardware accessing those page tables.
204 	 *
205 	 * We do it by clearing the entry in the PASID table and then flushing
206 	 * the IOTLB and the PASID table caches. This might upset hardware;
207 	 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
208 	 * page) so that we end up taking a fault that the hardware really
209 	 * *has* to handle gracefully without affecting other processes.
210 	 */
211 	rcu_read_lock();
212 	list_for_each_entry_rcu(sdev, &svm->devs, list)
213 		intel_pasid_tear_down_entry(svm->iommu, sdev->dev,
214 					    svm->pasid, true);
215 	rcu_read_unlock();
216 
217 }
218 
219 static const struct mmu_notifier_ops intel_mmuops = {
220 	.release = intel_mm_release,
221 	.invalidate_range = intel_invalidate_range,
222 };
223 
224 static DEFINE_MUTEX(pasid_mutex);
225 static LIST_HEAD(global_svm_list);
226 
227 #define for_each_svm_dev(sdev, svm, d)			\
228 	list_for_each_entry((sdev), &(svm)->devs, list)	\
229 		if ((d) != (sdev)->dev) {} else
230 
231 int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
232 			  struct iommu_gpasid_bind_data *data)
233 {
234 	struct intel_iommu *iommu = intel_svm_device_to_iommu(dev);
235 	struct dmar_domain *dmar_domain;
236 	struct intel_svm_dev *sdev;
237 	struct intel_svm *svm;
238 	int ret = 0;
239 
240 	if (WARN_ON(!iommu) || !data)
241 		return -EINVAL;
242 
243 	if (data->version != IOMMU_GPASID_BIND_VERSION_1 ||
244 	    data->format != IOMMU_PASID_FORMAT_INTEL_VTD)
245 		return -EINVAL;
246 
247 	if (!dev_is_pci(dev))
248 		return -ENOTSUPP;
249 
250 	/* VT-d supports devices with full 20 bit PASIDs only */
251 	if (pci_max_pasids(to_pci_dev(dev)) != PASID_MAX)
252 		return -EINVAL;
253 
254 	/*
255 	 * We only check host PASID range, we have no knowledge to check
256 	 * guest PASID range.
257 	 */
258 	if (data->hpasid <= 0 || data->hpasid >= PASID_MAX)
259 		return -EINVAL;
260 
261 	dmar_domain = to_dmar_domain(domain);
262 
263 	mutex_lock(&pasid_mutex);
264 	svm = ioasid_find(NULL, data->hpasid, NULL);
265 	if (IS_ERR(svm)) {
266 		ret = PTR_ERR(svm);
267 		goto out;
268 	}
269 
270 	if (svm) {
271 		/*
272 		 * If we found svm for the PASID, there must be at
273 		 * least one device bond, otherwise svm should be freed.
274 		 */
275 		if (WARN_ON(list_empty(&svm->devs))) {
276 			ret = -EINVAL;
277 			goto out;
278 		}
279 
280 		for_each_svm_dev(sdev, svm, dev) {
281 			/*
282 			 * For devices with aux domains, we should allow
283 			 * multiple bind calls with the same PASID and pdev.
284 			 */
285 			if (iommu_dev_feature_enabled(dev,
286 						      IOMMU_DEV_FEAT_AUX)) {
287 				sdev->users++;
288 			} else {
289 				dev_warn_ratelimited(dev,
290 						     "Already bound with PASID %u\n",
291 						     svm->pasid);
292 				ret = -EBUSY;
293 			}
294 			goto out;
295 		}
296 	} else {
297 		/* We come here when PASID has never been bond to a device. */
298 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
299 		if (!svm) {
300 			ret = -ENOMEM;
301 			goto out;
302 		}
303 		/* REVISIT: upper layer/VFIO can track host process that bind
304 		 * the PASID. ioasid_set = mm might be sufficient for vfio to
305 		 * check pasid VMM ownership. We can drop the following line
306 		 * once VFIO and IOASID set check is in place.
307 		 */
308 		svm->mm = get_task_mm(current);
309 		svm->pasid = data->hpasid;
310 		if (data->flags & IOMMU_SVA_GPASID_VAL) {
311 			svm->gpasid = data->gpasid;
312 			svm->flags |= SVM_FLAG_GUEST_PASID;
313 		}
314 		ioasid_set_data(data->hpasid, svm);
315 		INIT_LIST_HEAD_RCU(&svm->devs);
316 		mmput(svm->mm);
317 	}
318 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
319 	if (!sdev) {
320 		ret = -ENOMEM;
321 		goto out;
322 	}
323 	sdev->dev = dev;
324 
325 	/* Only count users if device has aux domains */
326 	if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
327 		sdev->users = 1;
328 
329 	/* Set up device context entry for PASID if not enabled already */
330 	ret = intel_iommu_enable_pasid(iommu, sdev->dev);
331 	if (ret) {
332 		dev_err_ratelimited(dev, "Failed to enable PASID capability\n");
333 		kfree(sdev);
334 		goto out;
335 	}
336 
337 	/*
338 	 * PASID table is per device for better security. Therefore, for
339 	 * each bind of a new device even with an existing PASID, we need to
340 	 * call the nested mode setup function here.
341 	 */
342 	spin_lock(&iommu->lock);
343 	ret = intel_pasid_setup_nested(iommu, dev,
344 				       (pgd_t *)(uintptr_t)data->gpgd,
345 				       data->hpasid, &data->vtd, dmar_domain,
346 				       data->addr_width);
347 	spin_unlock(&iommu->lock);
348 	if (ret) {
349 		dev_err_ratelimited(dev, "Failed to set up PASID %llu in nested mode, Err %d\n",
350 				    data->hpasid, ret);
351 		/*
352 		 * PASID entry should be in cleared state if nested mode
353 		 * set up failed. So we only need to clear IOASID tracking
354 		 * data such that free call will succeed.
355 		 */
356 		kfree(sdev);
357 		goto out;
358 	}
359 
360 	svm->flags |= SVM_FLAG_GUEST_MODE;
361 
362 	init_rcu_head(&sdev->rcu);
363 	list_add_rcu(&sdev->list, &svm->devs);
364  out:
365 	if (!IS_ERR_OR_NULL(svm) && list_empty(&svm->devs)) {
366 		ioasid_set_data(data->hpasid, NULL);
367 		kfree(svm);
368 	}
369 
370 	mutex_unlock(&pasid_mutex);
371 	return ret;
372 }
373 
374 int intel_svm_unbind_gpasid(struct device *dev, int pasid)
375 {
376 	struct intel_iommu *iommu = intel_svm_device_to_iommu(dev);
377 	struct intel_svm_dev *sdev;
378 	struct intel_svm *svm;
379 	int ret = -EINVAL;
380 
381 	if (WARN_ON(!iommu))
382 		return -EINVAL;
383 
384 	mutex_lock(&pasid_mutex);
385 	svm = ioasid_find(NULL, pasid, NULL);
386 	if (!svm) {
387 		ret = -EINVAL;
388 		goto out;
389 	}
390 
391 	if (IS_ERR(svm)) {
392 		ret = PTR_ERR(svm);
393 		goto out;
394 	}
395 
396 	for_each_svm_dev(sdev, svm, dev) {
397 		ret = 0;
398 		if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
399 			sdev->users--;
400 		if (!sdev->users) {
401 			list_del_rcu(&sdev->list);
402 			intel_pasid_tear_down_entry(iommu, dev,
403 						    svm->pasid, false);
404 			intel_svm_drain_prq(dev, svm->pasid);
405 			kfree_rcu(sdev, rcu);
406 
407 			if (list_empty(&svm->devs)) {
408 				/*
409 				 * We do not free the IOASID here in that
410 				 * IOMMU driver did not allocate it.
411 				 * Unlike native SVM, IOASID for guest use was
412 				 * allocated prior to the bind call.
413 				 * In any case, if the free call comes before
414 				 * the unbind, IOMMU driver will get notified
415 				 * and perform cleanup.
416 				 */
417 				ioasid_set_data(pasid, NULL);
418 				kfree(svm);
419 			}
420 		}
421 		break;
422 	}
423 out:
424 	mutex_unlock(&pasid_mutex);
425 	return ret;
426 }
427 
428 /* Caller must hold pasid_mutex, mm reference */
429 static int
430 intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
431 		  struct mm_struct *mm, struct intel_svm_dev **sd)
432 {
433 	struct intel_iommu *iommu = intel_svm_device_to_iommu(dev);
434 	struct device_domain_info *info;
435 	struct intel_svm_dev *sdev;
436 	struct intel_svm *svm = NULL;
437 	int pasid_max;
438 	int ret;
439 
440 	if (!iommu || dmar_disabled)
441 		return -EINVAL;
442 
443 	if (!intel_svm_capable(iommu))
444 		return -ENOTSUPP;
445 
446 	if (dev_is_pci(dev)) {
447 		pasid_max = pci_max_pasids(to_pci_dev(dev));
448 		if (pasid_max < 0)
449 			return -EINVAL;
450 	} else
451 		pasid_max = 1 << 20;
452 
453 	/* Bind supervisor PASID shuld have mm = NULL */
454 	if (flags & SVM_FLAG_SUPERVISOR_MODE) {
455 		if (!ecap_srs(iommu->ecap) || mm) {
456 			pr_err("Supervisor PASID with user provided mm.\n");
457 			return -EINVAL;
458 		}
459 	}
460 
461 	if (!(flags & SVM_FLAG_PRIVATE_PASID)) {
462 		struct intel_svm *t;
463 
464 		list_for_each_entry(t, &global_svm_list, list) {
465 			if (t->mm != mm || (t->flags & SVM_FLAG_PRIVATE_PASID))
466 				continue;
467 
468 			svm = t;
469 			if (svm->pasid >= pasid_max) {
470 				dev_warn(dev,
471 					 "Limited PASID width. Cannot use existing PASID %d\n",
472 					 svm->pasid);
473 				ret = -ENOSPC;
474 				goto out;
475 			}
476 
477 			/* Find the matching device in svm list */
478 			for_each_svm_dev(sdev, svm, dev) {
479 				if (sdev->ops != ops) {
480 					ret = -EBUSY;
481 					goto out;
482 				}
483 				sdev->users++;
484 				goto success;
485 			}
486 
487 			break;
488 		}
489 	}
490 
491 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
492 	if (!sdev) {
493 		ret = -ENOMEM;
494 		goto out;
495 	}
496 	sdev->dev = dev;
497 
498 	ret = intel_iommu_enable_pasid(iommu, dev);
499 	if (ret) {
500 		kfree(sdev);
501 		goto out;
502 	}
503 
504 	info = get_domain_info(dev);
505 	sdev->did = FLPT_DEFAULT_DID;
506 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
507 	if (info->ats_enabled) {
508 		sdev->dev_iotlb = 1;
509 		sdev->qdep = info->ats_qdep;
510 		if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
511 			sdev->qdep = 0;
512 	}
513 
514 	/* Finish the setup now we know we're keeping it */
515 	sdev->users = 1;
516 	sdev->ops = ops;
517 	init_rcu_head(&sdev->rcu);
518 
519 	if (!svm) {
520 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
521 		if (!svm) {
522 			ret = -ENOMEM;
523 			kfree(sdev);
524 			goto out;
525 		}
526 		svm->iommu = iommu;
527 
528 		if (pasid_max > intel_pasid_max_id)
529 			pasid_max = intel_pasid_max_id;
530 
531 		/* Do not use PASID 0, reserved for RID to PASID */
532 		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
533 					  pasid_max - 1, svm);
534 		if (svm->pasid == INVALID_IOASID) {
535 			kfree(svm);
536 			kfree(sdev);
537 			ret = -ENOSPC;
538 			goto out;
539 		}
540 		svm->notifier.ops = &intel_mmuops;
541 		svm->mm = mm;
542 		svm->flags = flags;
543 		INIT_LIST_HEAD_RCU(&svm->devs);
544 		INIT_LIST_HEAD(&svm->list);
545 		ret = -ENOMEM;
546 		if (mm) {
547 			ret = mmu_notifier_register(&svm->notifier, mm);
548 			if (ret) {
549 				ioasid_free(svm->pasid);
550 				kfree(svm);
551 				kfree(sdev);
552 				goto out;
553 			}
554 		}
555 
556 		spin_lock(&iommu->lock);
557 		ret = intel_pasid_setup_first_level(iommu, dev,
558 				mm ? mm->pgd : init_mm.pgd,
559 				svm->pasid, FLPT_DEFAULT_DID,
560 				(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
561 				(cpu_feature_enabled(X86_FEATURE_LA57) ?
562 				 PASID_FLAG_FL5LP : 0));
563 		spin_unlock(&iommu->lock);
564 		if (ret) {
565 			if (mm)
566 				mmu_notifier_unregister(&svm->notifier, mm);
567 			ioasid_free(svm->pasid);
568 			kfree(svm);
569 			kfree(sdev);
570 			goto out;
571 		}
572 
573 		list_add_tail(&svm->list, &global_svm_list);
574 	} else {
575 		/*
576 		 * Binding a new device with existing PASID, need to setup
577 		 * the PASID entry.
578 		 */
579 		spin_lock(&iommu->lock);
580 		ret = intel_pasid_setup_first_level(iommu, dev,
581 						mm ? mm->pgd : init_mm.pgd,
582 						svm->pasid, FLPT_DEFAULT_DID,
583 						(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
584 						(cpu_feature_enabled(X86_FEATURE_LA57) ?
585 						PASID_FLAG_FL5LP : 0));
586 		spin_unlock(&iommu->lock);
587 		if (ret) {
588 			kfree(sdev);
589 			goto out;
590 		}
591 	}
592 	list_add_rcu(&sdev->list, &svm->devs);
593 success:
594 	sdev->pasid = svm->pasid;
595 	sdev->sva.dev = dev;
596 	if (sd)
597 		*sd = sdev;
598 	ret = 0;
599  out:
600 	return ret;
601 }
602 
603 /* Caller must hold pasid_mutex */
604 static int intel_svm_unbind_mm(struct device *dev, int pasid)
605 {
606 	struct intel_svm_dev *sdev;
607 	struct intel_iommu *iommu;
608 	struct intel_svm *svm;
609 	int ret = -EINVAL;
610 
611 	iommu = intel_svm_device_to_iommu(dev);
612 	if (!iommu)
613 		goto out;
614 
615 	svm = ioasid_find(NULL, pasid, NULL);
616 	if (!svm)
617 		goto out;
618 
619 	if (IS_ERR(svm)) {
620 		ret = PTR_ERR(svm);
621 		goto out;
622 	}
623 
624 	for_each_svm_dev(sdev, svm, dev) {
625 		ret = 0;
626 		sdev->users--;
627 		if (!sdev->users) {
628 			list_del_rcu(&sdev->list);
629 			/* Flush the PASID cache and IOTLB for this device.
630 			 * Note that we do depend on the hardware *not* using
631 			 * the PASID any more. Just as we depend on other
632 			 * devices never using PASIDs that they have no right
633 			 * to use. We have a *shared* PASID table, because it's
634 			 * large and has to be physically contiguous. So it's
635 			 * hard to be as defensive as we might like. */
636 			intel_pasid_tear_down_entry(iommu, dev,
637 						    svm->pasid, false);
638 			intel_svm_drain_prq(dev, svm->pasid);
639 			kfree_rcu(sdev, rcu);
640 
641 			if (list_empty(&svm->devs)) {
642 				ioasid_free(svm->pasid);
643 				if (svm->mm)
644 					mmu_notifier_unregister(&svm->notifier, svm->mm);
645 				list_del(&svm->list);
646 				/* We mandate that no page faults may be outstanding
647 				 * for the PASID when intel_svm_unbind_mm() is called.
648 				 * If that is not obeyed, subtle errors will happen.
649 				 * Let's make them less subtle... */
650 				memset(svm, 0x6b, sizeof(*svm));
651 				kfree(svm);
652 			}
653 		}
654 		break;
655 	}
656  out:
657 
658 	return ret;
659 }
660 
661 /* Page request queue descriptor */
662 struct page_req_dsc {
663 	union {
664 		struct {
665 			u64 type:8;
666 			u64 pasid_present:1;
667 			u64 priv_data_present:1;
668 			u64 rsvd:6;
669 			u64 rid:16;
670 			u64 pasid:20;
671 			u64 exe_req:1;
672 			u64 pm_req:1;
673 			u64 rsvd2:10;
674 		};
675 		u64 qw_0;
676 	};
677 	union {
678 		struct {
679 			u64 rd_req:1;
680 			u64 wr_req:1;
681 			u64 lpig:1;
682 			u64 prg_index:9;
683 			u64 addr:52;
684 		};
685 		u64 qw_1;
686 	};
687 	u64 priv_data[2];
688 };
689 
690 #define PRQ_RING_MASK	((0x1000 << PRQ_ORDER) - 0x20)
691 
692 static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
693 {
694 	unsigned long requested = 0;
695 
696 	if (req->exe_req)
697 		requested |= VM_EXEC;
698 
699 	if (req->rd_req)
700 		requested |= VM_READ;
701 
702 	if (req->wr_req)
703 		requested |= VM_WRITE;
704 
705 	return (requested & ~vma->vm_flags) != 0;
706 }
707 
708 static bool is_canonical_address(u64 addr)
709 {
710 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
711 	long saddr = (long) addr;
712 
713 	return (((saddr << shift) >> shift) == saddr);
714 }
715 
716 /**
717  * intel_svm_drain_prq - Drain page requests and responses for a pasid
718  * @dev: target device
719  * @pasid: pasid for draining
720  *
721  * Drain all pending page requests and responses related to @pasid in both
722  * software and hardware. This is supposed to be called after the device
723  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
724  * and DevTLB have been invalidated.
725  *
726  * It waits until all pending page requests for @pasid in the page fault
727  * queue are completed by the prq handling thread. Then follow the steps
728  * described in VT-d spec CH7.10 to drain all page requests and page
729  * responses pending in the hardware.
730  */
731 static void intel_svm_drain_prq(struct device *dev, int pasid)
732 {
733 	struct device_domain_info *info;
734 	struct dmar_domain *domain;
735 	struct intel_iommu *iommu;
736 	struct qi_desc desc[3];
737 	struct pci_dev *pdev;
738 	int head, tail;
739 	u16 sid, did;
740 	int qdep;
741 
742 	info = get_domain_info(dev);
743 	if (WARN_ON(!info || !dev_is_pci(dev)))
744 		return;
745 
746 	if (!info->pri_enabled)
747 		return;
748 
749 	iommu = info->iommu;
750 	domain = info->domain;
751 	pdev = to_pci_dev(dev);
752 	sid = PCI_DEVID(info->bus, info->devfn);
753 	did = domain->iommu_did[iommu->seq_id];
754 	qdep = pci_ats_queue_depth(pdev);
755 
756 	/*
757 	 * Check and wait until all pending page requests in the queue are
758 	 * handled by the prq handling thread.
759 	 */
760 prq_retry:
761 	reinit_completion(&iommu->prq_complete);
762 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
763 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
764 	while (head != tail) {
765 		struct page_req_dsc *req;
766 
767 		req = &iommu->prq[head / sizeof(*req)];
768 		if (!req->pasid_present || req->pasid != pasid) {
769 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
770 			continue;
771 		}
772 
773 		wait_for_completion(&iommu->prq_complete);
774 		goto prq_retry;
775 	}
776 
777 	/*
778 	 * Perform steps described in VT-d spec CH7.10 to drain page
779 	 * requests and responses in hardware.
780 	 */
781 	memset(desc, 0, sizeof(desc));
782 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
783 			QI_IWD_FENCE |
784 			QI_IWD_TYPE;
785 	desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
786 			QI_EIOTLB_DID(did) |
787 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
788 			QI_EIOTLB_TYPE;
789 	desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
790 			QI_DEV_EIOTLB_SID(sid) |
791 			QI_DEV_EIOTLB_QDEP(qdep) |
792 			QI_DEIOTLB_TYPE |
793 			QI_DEV_IOTLB_PFSID(info->pfsid);
794 qi_retry:
795 	reinit_completion(&iommu->prq_complete);
796 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
797 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
798 		wait_for_completion(&iommu->prq_complete);
799 		goto qi_retry;
800 	}
801 }
802 
803 static irqreturn_t prq_event_thread(int irq, void *d)
804 {
805 	struct intel_iommu *iommu = d;
806 	struct intel_svm *svm = NULL;
807 	int head, tail, handled = 0;
808 
809 	/* Clear PPR bit before reading head/tail registers, to
810 	 * ensure that we get a new interrupt if needed. */
811 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
812 
813 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
814 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
815 	while (head != tail) {
816 		struct intel_svm_dev *sdev;
817 		struct vm_area_struct *vma;
818 		struct page_req_dsc *req;
819 		struct qi_desc resp;
820 		int result;
821 		vm_fault_t ret;
822 		u64 address;
823 
824 		handled = 1;
825 
826 		req = &iommu->prq[head / sizeof(*req)];
827 
828 		result = QI_RESP_FAILURE;
829 		address = (u64)req->addr << VTD_PAGE_SHIFT;
830 		if (!req->pasid_present) {
831 			pr_err("%s: Page request without PASID: %08llx %08llx\n",
832 			       iommu->name, ((unsigned long long *)req)[0],
833 			       ((unsigned long long *)req)[1]);
834 			goto no_pasid;
835 		}
836 
837 		if (!svm || svm->pasid != req->pasid) {
838 			rcu_read_lock();
839 			svm = ioasid_find(NULL, req->pasid, NULL);
840 			/* It *can't* go away, because the driver is not permitted
841 			 * to unbind the mm while any page faults are outstanding.
842 			 * So we only need RCU to protect the internal idr code. */
843 			rcu_read_unlock();
844 			if (IS_ERR_OR_NULL(svm)) {
845 				pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
846 				       iommu->name, req->pasid, ((unsigned long long *)req)[0],
847 				       ((unsigned long long *)req)[1]);
848 				goto no_pasid;
849 			}
850 		}
851 
852 		result = QI_RESP_INVALID;
853 		/* Since we're using init_mm.pgd directly, we should never take
854 		 * any faults on kernel addresses. */
855 		if (!svm->mm)
856 			goto bad_req;
857 
858 		/* If address is not canonical, return invalid response */
859 		if (!is_canonical_address(address))
860 			goto bad_req;
861 
862 		/* If the mm is already defunct, don't handle faults. */
863 		if (!mmget_not_zero(svm->mm))
864 			goto bad_req;
865 
866 		mmap_read_lock(svm->mm);
867 		vma = find_extend_vma(svm->mm, address);
868 		if (!vma || address < vma->vm_start)
869 			goto invalid;
870 
871 		if (access_error(vma, req))
872 			goto invalid;
873 
874 		ret = handle_mm_fault(vma, address,
875 				      req->wr_req ? FAULT_FLAG_WRITE : 0);
876 		if (ret & VM_FAULT_ERROR)
877 			goto invalid;
878 
879 		result = QI_RESP_SUCCESS;
880 	invalid:
881 		mmap_read_unlock(svm->mm);
882 		mmput(svm->mm);
883 	bad_req:
884 		/* Accounting for major/minor faults? */
885 		rcu_read_lock();
886 		list_for_each_entry_rcu(sdev, &svm->devs, list) {
887 			if (sdev->sid == req->rid)
888 				break;
889 		}
890 		/* Other devices can go away, but the drivers are not permitted
891 		 * to unbind while any page faults might be in flight. So it's
892 		 * OK to drop the 'lock' here now we have it. */
893 		rcu_read_unlock();
894 
895 		if (WARN_ON(&sdev->list == &svm->devs))
896 			sdev = NULL;
897 
898 		if (sdev && sdev->ops && sdev->ops->fault_cb) {
899 			int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
900 				(req->exe_req << 1) | (req->pm_req);
901 			sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr,
902 					    req->priv_data, rwxp, result);
903 		}
904 		/* We get here in the error case where the PASID lookup failed,
905 		   and these can be NULL. Do not use them below this point! */
906 		sdev = NULL;
907 		svm = NULL;
908 	no_pasid:
909 		if (req->lpig || req->priv_data_present) {
910 			/*
911 			 * Per VT-d spec. v3.0 ch7.7, system software must
912 			 * respond with page group response if private data
913 			 * is present (PDP) or last page in group (LPIG) bit
914 			 * is set. This is an additional VT-d feature beyond
915 			 * PCI ATS spec.
916 			 */
917 			resp.qw0 = QI_PGRP_PASID(req->pasid) |
918 				QI_PGRP_DID(req->rid) |
919 				QI_PGRP_PASID_P(req->pasid_present) |
920 				QI_PGRP_PDP(req->pasid_present) |
921 				QI_PGRP_RESP_CODE(result) |
922 				QI_PGRP_RESP_TYPE;
923 			resp.qw1 = QI_PGRP_IDX(req->prg_index) |
924 				QI_PGRP_LPIG(req->lpig);
925 
926 			if (req->priv_data_present)
927 				memcpy(&resp.qw2, req->priv_data,
928 				       sizeof(req->priv_data));
929 			resp.qw2 = 0;
930 			resp.qw3 = 0;
931 			qi_submit_sync(iommu, &resp, 1, 0);
932 		}
933 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
934 	}
935 
936 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
937 
938 	/*
939 	 * Clear the page request overflow bit and wake up all threads that
940 	 * are waiting for the completion of this handling.
941 	 */
942 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO)
943 		writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
944 
945 	if (!completion_done(&iommu->prq_complete))
946 		complete(&iommu->prq_complete);
947 
948 	return IRQ_RETVAL(handled);
949 }
950 
951 #define to_intel_svm_dev(handle) container_of(handle, struct intel_svm_dev, sva)
952 struct iommu_sva *
953 intel_svm_bind(struct device *dev, struct mm_struct *mm, void *drvdata)
954 {
955 	struct iommu_sva *sva = ERR_PTR(-EINVAL);
956 	struct intel_svm_dev *sdev = NULL;
957 	int flags = 0;
958 	int ret;
959 
960 	/*
961 	 * TODO: Consolidate with generic iommu-sva bind after it is merged.
962 	 * It will require shared SVM data structures, i.e. combine io_mm
963 	 * and intel_svm etc.
964 	 */
965 	if (drvdata)
966 		flags = *(int *)drvdata;
967 	mutex_lock(&pasid_mutex);
968 	ret = intel_svm_bind_mm(dev, flags, NULL, mm, &sdev);
969 	if (ret)
970 		sva = ERR_PTR(ret);
971 	else if (sdev)
972 		sva = &sdev->sva;
973 	else
974 		WARN(!sdev, "SVM bind succeeded with no sdev!\n");
975 
976 	mutex_unlock(&pasid_mutex);
977 
978 	return sva;
979 }
980 
981 void intel_svm_unbind(struct iommu_sva *sva)
982 {
983 	struct intel_svm_dev *sdev;
984 
985 	mutex_lock(&pasid_mutex);
986 	sdev = to_intel_svm_dev(sva);
987 	intel_svm_unbind_mm(sdev->dev, sdev->pasid);
988 	mutex_unlock(&pasid_mutex);
989 }
990 
991 int intel_svm_get_pasid(struct iommu_sva *sva)
992 {
993 	struct intel_svm_dev *sdev;
994 	int pasid;
995 
996 	mutex_lock(&pasid_mutex);
997 	sdev = to_intel_svm_dev(sva);
998 	pasid = sdev->pasid;
999 	mutex_unlock(&pasid_mutex);
1000 
1001 	return pasid;
1002 }
1003