xref: /openbmc/linux/drivers/iommu/intel/svm.c (revision b737eecd)
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 #include <asm/fpu/api.h>
23 
24 #include "pasid.h"
25 
26 static irqreturn_t prq_event_thread(int irq, void *d);
27 static void intel_svm_drain_prq(struct device *dev, u32 pasid);
28 
29 #define PRQ_ORDER 0
30 
31 int intel_svm_enable_prq(struct intel_iommu *iommu)
32 {
33 	struct page *pages;
34 	int irq, ret;
35 
36 	pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
37 	if (!pages) {
38 		pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
39 			iommu->name);
40 		return -ENOMEM;
41 	}
42 	iommu->prq = page_address(pages);
43 
44 	irq = dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED + iommu->seq_id, iommu->node, iommu);
45 	if (irq <= 0) {
46 		pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
47 		       iommu->name);
48 		ret = -EINVAL;
49 	err:
50 		free_pages((unsigned long)iommu->prq, PRQ_ORDER);
51 		iommu->prq = NULL;
52 		return ret;
53 	}
54 	iommu->pr_irq = irq;
55 
56 	snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
57 
58 	ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
59 				   iommu->prq_name, iommu);
60 	if (ret) {
61 		pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
62 		       iommu->name);
63 		dmar_free_hwirq(irq);
64 		iommu->pr_irq = 0;
65 		goto err;
66 	}
67 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
68 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
69 	dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
70 
71 	init_completion(&iommu->prq_complete);
72 
73 	return 0;
74 }
75 
76 int intel_svm_finish_prq(struct intel_iommu *iommu)
77 {
78 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
79 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
80 	dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
81 
82 	if (iommu->pr_irq) {
83 		free_irq(iommu->pr_irq, iommu);
84 		dmar_free_hwirq(iommu->pr_irq);
85 		iommu->pr_irq = 0;
86 	}
87 
88 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
89 	iommu->prq = NULL;
90 
91 	return 0;
92 }
93 
94 static inline bool intel_svm_capable(struct intel_iommu *iommu)
95 {
96 	return iommu->flags & VTD_FLAG_SVM_CAPABLE;
97 }
98 
99 void intel_svm_check(struct intel_iommu *iommu)
100 {
101 	if (!pasid_supported(iommu))
102 		return;
103 
104 	if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
105 	    !cap_fl1gp_support(iommu->cap)) {
106 		pr_err("%s SVM disabled, incompatible 1GB page capability\n",
107 		       iommu->name);
108 		return;
109 	}
110 
111 	if (cpu_feature_enabled(X86_FEATURE_LA57) &&
112 	    !cap_5lp_support(iommu->cap)) {
113 		pr_err("%s SVM disabled, incompatible paging mode\n",
114 		       iommu->name);
115 		return;
116 	}
117 
118 	iommu->flags |= VTD_FLAG_SVM_CAPABLE;
119 }
120 
121 static void __flush_svm_range_dev(struct intel_svm *svm,
122 				  struct intel_svm_dev *sdev,
123 				  unsigned long address,
124 				  unsigned long pages, int ih)
125 {
126 	struct device_domain_info *info = get_domain_info(sdev->dev);
127 
128 	if (WARN_ON(!pages))
129 		return;
130 
131 	qi_flush_piotlb(sdev->iommu, sdev->did, svm->pasid, address, pages, ih);
132 	if (info->ats_enabled)
133 		qi_flush_dev_iotlb_pasid(sdev->iommu, sdev->sid, info->pfsid,
134 					 svm->pasid, sdev->qdep, address,
135 					 order_base_2(pages));
136 }
137 
138 static void intel_flush_svm_range_dev(struct intel_svm *svm,
139 				      struct intel_svm_dev *sdev,
140 				      unsigned long address,
141 				      unsigned long pages, int ih)
142 {
143 	unsigned long shift = ilog2(__roundup_pow_of_two(pages));
144 	unsigned long align = (1ULL << (VTD_PAGE_SHIFT + shift));
145 	unsigned long start = ALIGN_DOWN(address, align);
146 	unsigned long end = ALIGN(address + (pages << VTD_PAGE_SHIFT), align);
147 
148 	while (start < end) {
149 		__flush_svm_range_dev(svm, sdev, start, align >> VTD_PAGE_SHIFT, ih);
150 		start += align;
151 	}
152 }
153 
154 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
155 				unsigned long pages, int ih)
156 {
157 	struct intel_svm_dev *sdev;
158 
159 	rcu_read_lock();
160 	list_for_each_entry_rcu(sdev, &svm->devs, list)
161 		intel_flush_svm_range_dev(svm, sdev, address, pages, ih);
162 	rcu_read_unlock();
163 }
164 
165 /* Pages have been freed at this point */
166 static void intel_invalidate_range(struct mmu_notifier *mn,
167 				   struct mm_struct *mm,
168 				   unsigned long start, unsigned long end)
169 {
170 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
171 
172 	intel_flush_svm_range(svm, start,
173 			      (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0);
174 }
175 
176 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
177 {
178 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
179 	struct intel_svm_dev *sdev;
180 
181 	/* This might end up being called from exit_mmap(), *before* the page
182 	 * tables are cleared. And __mmu_notifier_release() will delete us from
183 	 * the list of notifiers so that our invalidate_range() callback doesn't
184 	 * get called when the page tables are cleared. So we need to protect
185 	 * against hardware accessing those page tables.
186 	 *
187 	 * We do it by clearing the entry in the PASID table and then flushing
188 	 * the IOTLB and the PASID table caches. This might upset hardware;
189 	 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
190 	 * page) so that we end up taking a fault that the hardware really
191 	 * *has* to handle gracefully without affecting other processes.
192 	 */
193 	rcu_read_lock();
194 	list_for_each_entry_rcu(sdev, &svm->devs, list)
195 		intel_pasid_tear_down_entry(sdev->iommu, sdev->dev,
196 					    svm->pasid, true);
197 	rcu_read_unlock();
198 
199 }
200 
201 static const struct mmu_notifier_ops intel_mmuops = {
202 	.release = intel_mm_release,
203 	.invalidate_range = intel_invalidate_range,
204 };
205 
206 static DEFINE_MUTEX(pasid_mutex);
207 static LIST_HEAD(global_svm_list);
208 
209 #define for_each_svm_dev(sdev, svm, d)			\
210 	list_for_each_entry((sdev), &(svm)->devs, list)	\
211 		if ((d) != (sdev)->dev) {} else
212 
213 static int pasid_to_svm_sdev(struct device *dev, unsigned int pasid,
214 			     struct intel_svm **rsvm,
215 			     struct intel_svm_dev **rsdev)
216 {
217 	struct intel_svm_dev *d, *sdev = NULL;
218 	struct intel_svm *svm;
219 
220 	/* The caller should hold the pasid_mutex lock */
221 	if (WARN_ON(!mutex_is_locked(&pasid_mutex)))
222 		return -EINVAL;
223 
224 	if (pasid == INVALID_IOASID || pasid >= PASID_MAX)
225 		return -EINVAL;
226 
227 	svm = ioasid_find(NULL, pasid, NULL);
228 	if (IS_ERR(svm))
229 		return PTR_ERR(svm);
230 
231 	if (!svm)
232 		goto out;
233 
234 	/*
235 	 * If we found svm for the PASID, there must be at least one device
236 	 * bond.
237 	 */
238 	if (WARN_ON(list_empty(&svm->devs)))
239 		return -EINVAL;
240 
241 	rcu_read_lock();
242 	list_for_each_entry_rcu(d, &svm->devs, list) {
243 		if (d->dev == dev) {
244 			sdev = d;
245 			break;
246 		}
247 	}
248 	rcu_read_unlock();
249 
250 out:
251 	*rsvm = svm;
252 	*rsdev = sdev;
253 
254 	return 0;
255 }
256 
257 int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
258 			  struct iommu_gpasid_bind_data *data)
259 {
260 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
261 	struct intel_svm_dev *sdev = NULL;
262 	struct dmar_domain *dmar_domain;
263 	struct device_domain_info *info;
264 	struct intel_svm *svm = NULL;
265 	unsigned long iflags;
266 	int ret = 0;
267 
268 	if (WARN_ON(!iommu) || !data)
269 		return -EINVAL;
270 
271 	if (data->format != IOMMU_PASID_FORMAT_INTEL_VTD)
272 		return -EINVAL;
273 
274 	/* IOMMU core ensures argsz is more than the start of the union */
275 	if (data->argsz < offsetofend(struct iommu_gpasid_bind_data, vendor.vtd))
276 		return -EINVAL;
277 
278 	/* Make sure no undefined flags are used in vendor data */
279 	if (data->vendor.vtd.flags & ~(IOMMU_SVA_VTD_GPASID_LAST - 1))
280 		return -EINVAL;
281 
282 	if (!dev_is_pci(dev))
283 		return -ENOTSUPP;
284 
285 	/* VT-d supports devices with full 20 bit PASIDs only */
286 	if (pci_max_pasids(to_pci_dev(dev)) != PASID_MAX)
287 		return -EINVAL;
288 
289 	/*
290 	 * We only check host PASID range, we have no knowledge to check
291 	 * guest PASID range.
292 	 */
293 	if (data->hpasid <= 0 || data->hpasid >= PASID_MAX)
294 		return -EINVAL;
295 
296 	info = get_domain_info(dev);
297 	if (!info)
298 		return -EINVAL;
299 
300 	dmar_domain = to_dmar_domain(domain);
301 
302 	mutex_lock(&pasid_mutex);
303 	ret = pasid_to_svm_sdev(dev, data->hpasid, &svm, &sdev);
304 	if (ret)
305 		goto out;
306 
307 	if (sdev) {
308 		/*
309 		 * Do not allow multiple bindings of the same device-PASID since
310 		 * there is only one SL page tables per PASID. We may revisit
311 		 * once sharing PGD across domains are supported.
312 		 */
313 		dev_warn_ratelimited(dev, "Already bound with PASID %u\n",
314 				     svm->pasid);
315 		ret = -EBUSY;
316 		goto out;
317 	}
318 
319 	if (!svm) {
320 		/* We come here when PASID has never been bond to a device. */
321 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
322 		if (!svm) {
323 			ret = -ENOMEM;
324 			goto out;
325 		}
326 		/* REVISIT: upper layer/VFIO can track host process that bind
327 		 * the PASID. ioasid_set = mm might be sufficient for vfio to
328 		 * check pasid VMM ownership. We can drop the following line
329 		 * once VFIO and IOASID set check is in place.
330 		 */
331 		svm->mm = get_task_mm(current);
332 		svm->pasid = data->hpasid;
333 		if (data->flags & IOMMU_SVA_GPASID_VAL) {
334 			svm->gpasid = data->gpasid;
335 			svm->flags |= SVM_FLAG_GUEST_PASID;
336 		}
337 		ioasid_set_data(data->hpasid, svm);
338 		INIT_LIST_HEAD_RCU(&svm->devs);
339 		mmput(svm->mm);
340 	}
341 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
342 	if (!sdev) {
343 		ret = -ENOMEM;
344 		goto out;
345 	}
346 	sdev->dev = dev;
347 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
348 	sdev->iommu = iommu;
349 
350 	/* Only count users if device has aux domains */
351 	if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
352 		sdev->users = 1;
353 
354 	/* Set up device context entry for PASID if not enabled already */
355 	ret = intel_iommu_enable_pasid(iommu, sdev->dev);
356 	if (ret) {
357 		dev_err_ratelimited(dev, "Failed to enable PASID capability\n");
358 		kfree(sdev);
359 		goto out;
360 	}
361 
362 	/*
363 	 * PASID table is per device for better security. Therefore, for
364 	 * each bind of a new device even with an existing PASID, we need to
365 	 * call the nested mode setup function here.
366 	 */
367 	spin_lock_irqsave(&iommu->lock, iflags);
368 	ret = intel_pasid_setup_nested(iommu, dev,
369 				       (pgd_t *)(uintptr_t)data->gpgd,
370 				       data->hpasid, &data->vendor.vtd, dmar_domain,
371 				       data->addr_width);
372 	spin_unlock_irqrestore(&iommu->lock, iflags);
373 	if (ret) {
374 		dev_err_ratelimited(dev, "Failed to set up PASID %llu in nested mode, Err %d\n",
375 				    data->hpasid, ret);
376 		/*
377 		 * PASID entry should be in cleared state if nested mode
378 		 * set up failed. So we only need to clear IOASID tracking
379 		 * data such that free call will succeed.
380 		 */
381 		kfree(sdev);
382 		goto out;
383 	}
384 
385 	svm->flags |= SVM_FLAG_GUEST_MODE;
386 
387 	init_rcu_head(&sdev->rcu);
388 	list_add_rcu(&sdev->list, &svm->devs);
389  out:
390 	if (!IS_ERR_OR_NULL(svm) && list_empty(&svm->devs)) {
391 		ioasid_set_data(data->hpasid, NULL);
392 		kfree(svm);
393 	}
394 
395 	mutex_unlock(&pasid_mutex);
396 	return ret;
397 }
398 
399 int intel_svm_unbind_gpasid(struct device *dev, u32 pasid)
400 {
401 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
402 	struct intel_svm_dev *sdev;
403 	struct intel_svm *svm;
404 	int ret;
405 
406 	if (WARN_ON(!iommu))
407 		return -EINVAL;
408 
409 	mutex_lock(&pasid_mutex);
410 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
411 	if (ret)
412 		goto out;
413 
414 	if (sdev) {
415 		if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
416 			sdev->users--;
417 		if (!sdev->users) {
418 			list_del_rcu(&sdev->list);
419 			intel_pasid_tear_down_entry(iommu, dev,
420 						    svm->pasid, false);
421 			intel_svm_drain_prq(dev, svm->pasid);
422 			kfree_rcu(sdev, rcu);
423 
424 			if (list_empty(&svm->devs)) {
425 				/*
426 				 * We do not free the IOASID here in that
427 				 * IOMMU driver did not allocate it.
428 				 * Unlike native SVM, IOASID for guest use was
429 				 * allocated prior to the bind call.
430 				 * In any case, if the free call comes before
431 				 * the unbind, IOMMU driver will get notified
432 				 * and perform cleanup.
433 				 */
434 				ioasid_set_data(pasid, NULL);
435 				kfree(svm);
436 			}
437 		}
438 	}
439 out:
440 	mutex_unlock(&pasid_mutex);
441 	return ret;
442 }
443 
444 static void _load_pasid(void *unused)
445 {
446 	update_pasid();
447 }
448 
449 static void load_pasid(struct mm_struct *mm, u32 pasid)
450 {
451 	mutex_lock(&mm->context.lock);
452 
453 	/* Synchronize with READ_ONCE in update_pasid(). */
454 	smp_store_release(&mm->pasid, pasid);
455 
456 	/* Update PASID MSR on all CPUs running the mm's tasks. */
457 	on_each_cpu_mask(mm_cpumask(mm), _load_pasid, NULL, true);
458 
459 	mutex_unlock(&mm->context.lock);
460 }
461 
462 /* Caller must hold pasid_mutex, mm reference */
463 static int
464 intel_svm_bind_mm(struct device *dev, unsigned int flags,
465 		  struct svm_dev_ops *ops,
466 		  struct mm_struct *mm, struct intel_svm_dev **sd)
467 {
468 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
469 	struct device_domain_info *info;
470 	struct intel_svm_dev *sdev;
471 	struct intel_svm *svm = NULL;
472 	unsigned long iflags;
473 	int pasid_max;
474 	int ret;
475 
476 	if (!iommu || dmar_disabled)
477 		return -EINVAL;
478 
479 	if (!intel_svm_capable(iommu))
480 		return -ENOTSUPP;
481 
482 	if (dev_is_pci(dev)) {
483 		pasid_max = pci_max_pasids(to_pci_dev(dev));
484 		if (pasid_max < 0)
485 			return -EINVAL;
486 	} else
487 		pasid_max = 1 << 20;
488 
489 	/* Bind supervisor PASID shuld have mm = NULL */
490 	if (flags & SVM_FLAG_SUPERVISOR_MODE) {
491 		if (!ecap_srs(iommu->ecap) || mm) {
492 			pr_err("Supervisor PASID with user provided mm.\n");
493 			return -EINVAL;
494 		}
495 	}
496 
497 	if (!(flags & SVM_FLAG_PRIVATE_PASID)) {
498 		struct intel_svm *t;
499 
500 		list_for_each_entry(t, &global_svm_list, list) {
501 			if (t->mm != mm || (t->flags & SVM_FLAG_PRIVATE_PASID))
502 				continue;
503 
504 			svm = t;
505 			if (svm->pasid >= pasid_max) {
506 				dev_warn(dev,
507 					 "Limited PASID width. Cannot use existing PASID %d\n",
508 					 svm->pasid);
509 				ret = -ENOSPC;
510 				goto out;
511 			}
512 
513 			/* Find the matching device in svm list */
514 			for_each_svm_dev(sdev, svm, dev) {
515 				if (sdev->ops != ops) {
516 					ret = -EBUSY;
517 					goto out;
518 				}
519 				sdev->users++;
520 				goto success;
521 			}
522 
523 			break;
524 		}
525 	}
526 
527 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
528 	if (!sdev) {
529 		ret = -ENOMEM;
530 		goto out;
531 	}
532 	sdev->dev = dev;
533 	sdev->iommu = iommu;
534 
535 	ret = intel_iommu_enable_pasid(iommu, dev);
536 	if (ret) {
537 		kfree(sdev);
538 		goto out;
539 	}
540 
541 	info = get_domain_info(dev);
542 	sdev->did = FLPT_DEFAULT_DID;
543 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
544 	if (info->ats_enabled) {
545 		sdev->dev_iotlb = 1;
546 		sdev->qdep = info->ats_qdep;
547 		if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
548 			sdev->qdep = 0;
549 	}
550 
551 	/* Finish the setup now we know we're keeping it */
552 	sdev->users = 1;
553 	sdev->ops = ops;
554 	init_rcu_head(&sdev->rcu);
555 
556 	if (!svm) {
557 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
558 		if (!svm) {
559 			ret = -ENOMEM;
560 			kfree(sdev);
561 			goto out;
562 		}
563 
564 		if (pasid_max > intel_pasid_max_id)
565 			pasid_max = intel_pasid_max_id;
566 
567 		/* Do not use PASID 0, reserved for RID to PASID */
568 		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
569 					  pasid_max - 1, svm);
570 		if (svm->pasid == INVALID_IOASID) {
571 			kfree(svm);
572 			kfree(sdev);
573 			ret = -ENOSPC;
574 			goto out;
575 		}
576 		svm->notifier.ops = &intel_mmuops;
577 		svm->mm = mm;
578 		svm->flags = flags;
579 		INIT_LIST_HEAD_RCU(&svm->devs);
580 		INIT_LIST_HEAD(&svm->list);
581 		ret = -ENOMEM;
582 		if (mm) {
583 			ret = mmu_notifier_register(&svm->notifier, mm);
584 			if (ret) {
585 				ioasid_put(svm->pasid);
586 				kfree(svm);
587 				kfree(sdev);
588 				goto out;
589 			}
590 		}
591 
592 		spin_lock_irqsave(&iommu->lock, iflags);
593 		ret = intel_pasid_setup_first_level(iommu, dev,
594 				mm ? mm->pgd : init_mm.pgd,
595 				svm->pasid, FLPT_DEFAULT_DID,
596 				(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
597 				(cpu_feature_enabled(X86_FEATURE_LA57) ?
598 				 PASID_FLAG_FL5LP : 0));
599 		spin_unlock_irqrestore(&iommu->lock, iflags);
600 		if (ret) {
601 			if (mm)
602 				mmu_notifier_unregister(&svm->notifier, mm);
603 			ioasid_put(svm->pasid);
604 			kfree(svm);
605 			kfree(sdev);
606 			goto out;
607 		}
608 
609 		list_add_tail(&svm->list, &global_svm_list);
610 		if (mm) {
611 			/* The newly allocated pasid is loaded to the mm. */
612 			load_pasid(mm, svm->pasid);
613 		}
614 	} else {
615 		/*
616 		 * Binding a new device with existing PASID, need to setup
617 		 * the PASID entry.
618 		 */
619 		spin_lock_irqsave(&iommu->lock, iflags);
620 		ret = intel_pasid_setup_first_level(iommu, dev,
621 						mm ? mm->pgd : init_mm.pgd,
622 						svm->pasid, FLPT_DEFAULT_DID,
623 						(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
624 						(cpu_feature_enabled(X86_FEATURE_LA57) ?
625 						PASID_FLAG_FL5LP : 0));
626 		spin_unlock_irqrestore(&iommu->lock, iflags);
627 		if (ret) {
628 			kfree(sdev);
629 			goto out;
630 		}
631 	}
632 	list_add_rcu(&sdev->list, &svm->devs);
633 success:
634 	sdev->pasid = svm->pasid;
635 	sdev->sva.dev = dev;
636 	if (sd)
637 		*sd = sdev;
638 	ret = 0;
639 out:
640 	return ret;
641 }
642 
643 /* Caller must hold pasid_mutex */
644 static int intel_svm_unbind_mm(struct device *dev, u32 pasid)
645 {
646 	struct intel_svm_dev *sdev;
647 	struct intel_iommu *iommu;
648 	struct intel_svm *svm;
649 	int ret = -EINVAL;
650 
651 	iommu = device_to_iommu(dev, NULL, NULL);
652 	if (!iommu)
653 		goto out;
654 
655 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
656 	if (ret)
657 		goto out;
658 
659 	if (sdev) {
660 		sdev->users--;
661 		if (!sdev->users) {
662 			list_del_rcu(&sdev->list);
663 			/* Flush the PASID cache and IOTLB for this device.
664 			 * Note that we do depend on the hardware *not* using
665 			 * the PASID any more. Just as we depend on other
666 			 * devices never using PASIDs that they have no right
667 			 * to use. We have a *shared* PASID table, because it's
668 			 * large and has to be physically contiguous. So it's
669 			 * hard to be as defensive as we might like. */
670 			intel_pasid_tear_down_entry(iommu, dev,
671 						    svm->pasid, false);
672 			intel_svm_drain_prq(dev, svm->pasid);
673 			kfree_rcu(sdev, rcu);
674 
675 			if (list_empty(&svm->devs)) {
676 				ioasid_put(svm->pasid);
677 				if (svm->mm) {
678 					mmu_notifier_unregister(&svm->notifier, svm->mm);
679 					/* Clear mm's pasid. */
680 					load_pasid(svm->mm, PASID_DISABLED);
681 				}
682 				list_del(&svm->list);
683 				/* We mandate that no page faults may be outstanding
684 				 * for the PASID when intel_svm_unbind_mm() is called.
685 				 * If that is not obeyed, subtle errors will happen.
686 				 * Let's make them less subtle... */
687 				memset(svm, 0x6b, sizeof(*svm));
688 				kfree(svm);
689 			}
690 		}
691 	}
692 out:
693 	return ret;
694 }
695 
696 /* Page request queue descriptor */
697 struct page_req_dsc {
698 	union {
699 		struct {
700 			u64 type:8;
701 			u64 pasid_present:1;
702 			u64 priv_data_present:1;
703 			u64 rsvd:6;
704 			u64 rid:16;
705 			u64 pasid:20;
706 			u64 exe_req:1;
707 			u64 pm_req:1;
708 			u64 rsvd2:10;
709 		};
710 		u64 qw_0;
711 	};
712 	union {
713 		struct {
714 			u64 rd_req:1;
715 			u64 wr_req:1;
716 			u64 lpig:1;
717 			u64 prg_index:9;
718 			u64 addr:52;
719 		};
720 		u64 qw_1;
721 	};
722 	u64 priv_data[2];
723 };
724 
725 #define PRQ_RING_MASK	((0x1000 << PRQ_ORDER) - 0x20)
726 
727 static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
728 {
729 	unsigned long requested = 0;
730 
731 	if (req->exe_req)
732 		requested |= VM_EXEC;
733 
734 	if (req->rd_req)
735 		requested |= VM_READ;
736 
737 	if (req->wr_req)
738 		requested |= VM_WRITE;
739 
740 	return (requested & ~vma->vm_flags) != 0;
741 }
742 
743 static bool is_canonical_address(u64 addr)
744 {
745 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
746 	long saddr = (long) addr;
747 
748 	return (((saddr << shift) >> shift) == saddr);
749 }
750 
751 /**
752  * intel_svm_drain_prq - Drain page requests and responses for a pasid
753  * @dev: target device
754  * @pasid: pasid for draining
755  *
756  * Drain all pending page requests and responses related to @pasid in both
757  * software and hardware. This is supposed to be called after the device
758  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
759  * and DevTLB have been invalidated.
760  *
761  * It waits until all pending page requests for @pasid in the page fault
762  * queue are completed by the prq handling thread. Then follow the steps
763  * described in VT-d spec CH7.10 to drain all page requests and page
764  * responses pending in the hardware.
765  */
766 static void intel_svm_drain_prq(struct device *dev, u32 pasid)
767 {
768 	struct device_domain_info *info;
769 	struct dmar_domain *domain;
770 	struct intel_iommu *iommu;
771 	struct qi_desc desc[3];
772 	struct pci_dev *pdev;
773 	int head, tail;
774 	u16 sid, did;
775 	int qdep;
776 
777 	info = get_domain_info(dev);
778 	if (WARN_ON(!info || !dev_is_pci(dev)))
779 		return;
780 
781 	if (!info->pri_enabled)
782 		return;
783 
784 	iommu = info->iommu;
785 	domain = info->domain;
786 	pdev = to_pci_dev(dev);
787 	sid = PCI_DEVID(info->bus, info->devfn);
788 	did = domain->iommu_did[iommu->seq_id];
789 	qdep = pci_ats_queue_depth(pdev);
790 
791 	/*
792 	 * Check and wait until all pending page requests in the queue are
793 	 * handled by the prq handling thread.
794 	 */
795 prq_retry:
796 	reinit_completion(&iommu->prq_complete);
797 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
798 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
799 	while (head != tail) {
800 		struct page_req_dsc *req;
801 
802 		req = &iommu->prq[head / sizeof(*req)];
803 		if (!req->pasid_present || req->pasid != pasid) {
804 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
805 			continue;
806 		}
807 
808 		wait_for_completion(&iommu->prq_complete);
809 		goto prq_retry;
810 	}
811 
812 	/*
813 	 * Perform steps described in VT-d spec CH7.10 to drain page
814 	 * requests and responses in hardware.
815 	 */
816 	memset(desc, 0, sizeof(desc));
817 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
818 			QI_IWD_FENCE |
819 			QI_IWD_TYPE;
820 	desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
821 			QI_EIOTLB_DID(did) |
822 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
823 			QI_EIOTLB_TYPE;
824 	desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
825 			QI_DEV_EIOTLB_SID(sid) |
826 			QI_DEV_EIOTLB_QDEP(qdep) |
827 			QI_DEIOTLB_TYPE |
828 			QI_DEV_IOTLB_PFSID(info->pfsid);
829 qi_retry:
830 	reinit_completion(&iommu->prq_complete);
831 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
832 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
833 		wait_for_completion(&iommu->prq_complete);
834 		goto qi_retry;
835 	}
836 }
837 
838 static int prq_to_iommu_prot(struct page_req_dsc *req)
839 {
840 	int prot = 0;
841 
842 	if (req->rd_req)
843 		prot |= IOMMU_FAULT_PERM_READ;
844 	if (req->wr_req)
845 		prot |= IOMMU_FAULT_PERM_WRITE;
846 	if (req->exe_req)
847 		prot |= IOMMU_FAULT_PERM_EXEC;
848 	if (req->pm_req)
849 		prot |= IOMMU_FAULT_PERM_PRIV;
850 
851 	return prot;
852 }
853 
854 static int
855 intel_svm_prq_report(struct device *dev, struct page_req_dsc *desc)
856 {
857 	struct iommu_fault_event event;
858 
859 	if (!dev || !dev_is_pci(dev))
860 		return -ENODEV;
861 
862 	/* Fill in event data for device specific processing */
863 	memset(&event, 0, sizeof(struct iommu_fault_event));
864 	event.fault.type = IOMMU_FAULT_PAGE_REQ;
865 	event.fault.prm.addr = desc->addr;
866 	event.fault.prm.pasid = desc->pasid;
867 	event.fault.prm.grpid = desc->prg_index;
868 	event.fault.prm.perm = prq_to_iommu_prot(desc);
869 
870 	if (desc->lpig)
871 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
872 	if (desc->pasid_present) {
873 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
874 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
875 	}
876 	if (desc->priv_data_present) {
877 		/*
878 		 * Set last page in group bit if private data is present,
879 		 * page response is required as it does for LPIG.
880 		 * iommu_report_device_fault() doesn't understand this vendor
881 		 * specific requirement thus we set last_page as a workaround.
882 		 */
883 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
884 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
885 		memcpy(event.fault.prm.private_data, desc->priv_data,
886 		       sizeof(desc->priv_data));
887 	}
888 
889 	return iommu_report_device_fault(dev, &event);
890 }
891 
892 static irqreturn_t prq_event_thread(int irq, void *d)
893 {
894 	struct intel_svm_dev *sdev = NULL;
895 	struct intel_iommu *iommu = d;
896 	struct intel_svm *svm = NULL;
897 	int head, tail, handled = 0;
898 
899 	/* Clear PPR bit before reading head/tail registers, to
900 	 * ensure that we get a new interrupt if needed. */
901 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
902 
903 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
904 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
905 	while (head != tail) {
906 		struct vm_area_struct *vma;
907 		struct page_req_dsc *req;
908 		struct qi_desc resp;
909 		int result;
910 		vm_fault_t ret;
911 		u64 address;
912 
913 		handled = 1;
914 		req = &iommu->prq[head / sizeof(*req)];
915 		result = QI_RESP_INVALID;
916 		address = (u64)req->addr << VTD_PAGE_SHIFT;
917 		if (!req->pasid_present) {
918 			pr_err("%s: Page request without PASID: %08llx %08llx\n",
919 			       iommu->name, ((unsigned long long *)req)[0],
920 			       ((unsigned long long *)req)[1]);
921 			goto no_pasid;
922 		}
923 
924 		if (!svm || svm->pasid != req->pasid) {
925 			rcu_read_lock();
926 			svm = ioasid_find(NULL, req->pasid, NULL);
927 			/* It *can't* go away, because the driver is not permitted
928 			 * to unbind the mm while any page faults are outstanding.
929 			 * So we only need RCU to protect the internal idr code. */
930 			rcu_read_unlock();
931 			if (IS_ERR_OR_NULL(svm)) {
932 				pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
933 				       iommu->name, req->pasid, ((unsigned long long *)req)[0],
934 				       ((unsigned long long *)req)[1]);
935 				goto no_pasid;
936 			}
937 		}
938 
939 		if (!sdev || sdev->sid != req->rid) {
940 			struct intel_svm_dev *t;
941 
942 			sdev = NULL;
943 			rcu_read_lock();
944 			list_for_each_entry_rcu(t, &svm->devs, list) {
945 				if (t->sid == req->rid) {
946 					sdev = t;
947 					break;
948 				}
949 			}
950 			rcu_read_unlock();
951 		}
952 
953 		/* Since we're using init_mm.pgd directly, we should never take
954 		 * any faults on kernel addresses. */
955 		if (!svm->mm)
956 			goto bad_req;
957 
958 		/* If address is not canonical, return invalid response */
959 		if (!is_canonical_address(address))
960 			goto bad_req;
961 
962 		/*
963 		 * If prq is to be handled outside iommu driver via receiver of
964 		 * the fault notifiers, we skip the page response here.
965 		 */
966 		if (svm->flags & SVM_FLAG_GUEST_MODE) {
967 			if (sdev && !intel_svm_prq_report(sdev->dev, req))
968 				goto prq_advance;
969 			else
970 				goto bad_req;
971 		}
972 
973 		/* If the mm is already defunct, don't handle faults. */
974 		if (!mmget_not_zero(svm->mm))
975 			goto bad_req;
976 
977 		mmap_read_lock(svm->mm);
978 		vma = find_extend_vma(svm->mm, address);
979 		if (!vma || address < vma->vm_start)
980 			goto invalid;
981 
982 		if (access_error(vma, req))
983 			goto invalid;
984 
985 		ret = handle_mm_fault(vma, address,
986 				      req->wr_req ? FAULT_FLAG_WRITE : 0,
987 				      NULL);
988 		if (ret & VM_FAULT_ERROR)
989 			goto invalid;
990 
991 		result = QI_RESP_SUCCESS;
992 invalid:
993 		mmap_read_unlock(svm->mm);
994 		mmput(svm->mm);
995 bad_req:
996 		WARN_ON(!sdev);
997 		if (sdev && sdev->ops && sdev->ops->fault_cb) {
998 			int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
999 				(req->exe_req << 1) | (req->pm_req);
1000 			sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr,
1001 					    req->priv_data, rwxp, result);
1002 		}
1003 		/* We get here in the error case where the PASID lookup failed,
1004 		   and these can be NULL. Do not use them below this point! */
1005 		sdev = NULL;
1006 		svm = NULL;
1007 no_pasid:
1008 		if (req->lpig || req->priv_data_present) {
1009 			/*
1010 			 * Per VT-d spec. v3.0 ch7.7, system software must
1011 			 * respond with page group response if private data
1012 			 * is present (PDP) or last page in group (LPIG) bit
1013 			 * is set. This is an additional VT-d feature beyond
1014 			 * PCI ATS spec.
1015 			 */
1016 			resp.qw0 = QI_PGRP_PASID(req->pasid) |
1017 				QI_PGRP_DID(req->rid) |
1018 				QI_PGRP_PASID_P(req->pasid_present) |
1019 				QI_PGRP_PDP(req->priv_data_present) |
1020 				QI_PGRP_RESP_CODE(result) |
1021 				QI_PGRP_RESP_TYPE;
1022 			resp.qw1 = QI_PGRP_IDX(req->prg_index) |
1023 				QI_PGRP_LPIG(req->lpig);
1024 
1025 			if (req->priv_data_present)
1026 				memcpy(&resp.qw2, req->priv_data,
1027 				       sizeof(req->priv_data));
1028 			resp.qw2 = 0;
1029 			resp.qw3 = 0;
1030 			qi_submit_sync(iommu, &resp, 1, 0);
1031 		}
1032 prq_advance:
1033 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
1034 	}
1035 
1036 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
1037 
1038 	/*
1039 	 * Clear the page request overflow bit and wake up all threads that
1040 	 * are waiting for the completion of this handling.
1041 	 */
1042 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
1043 		pr_info_ratelimited("IOMMU: %s: PRQ overflow detected\n",
1044 				    iommu->name);
1045 		head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
1046 		tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
1047 		if (head == tail) {
1048 			writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
1049 			pr_info_ratelimited("IOMMU: %s: PRQ overflow cleared",
1050 					    iommu->name);
1051 		}
1052 	}
1053 
1054 	if (!completion_done(&iommu->prq_complete))
1055 		complete(&iommu->prq_complete);
1056 
1057 	return IRQ_RETVAL(handled);
1058 }
1059 
1060 #define to_intel_svm_dev(handle) container_of(handle, struct intel_svm_dev, sva)
1061 struct iommu_sva *
1062 intel_svm_bind(struct device *dev, struct mm_struct *mm, void *drvdata)
1063 {
1064 	struct iommu_sva *sva = ERR_PTR(-EINVAL);
1065 	struct intel_svm_dev *sdev = NULL;
1066 	unsigned int flags = 0;
1067 	int ret;
1068 
1069 	/*
1070 	 * TODO: Consolidate with generic iommu-sva bind after it is merged.
1071 	 * It will require shared SVM data structures, i.e. combine io_mm
1072 	 * and intel_svm etc.
1073 	 */
1074 	if (drvdata)
1075 		flags = *(unsigned int *)drvdata;
1076 	mutex_lock(&pasid_mutex);
1077 	ret = intel_svm_bind_mm(dev, flags, NULL, mm, &sdev);
1078 	if (ret)
1079 		sva = ERR_PTR(ret);
1080 	else if (sdev)
1081 		sva = &sdev->sva;
1082 	else
1083 		WARN(!sdev, "SVM bind succeeded with no sdev!\n");
1084 
1085 	mutex_unlock(&pasid_mutex);
1086 
1087 	return sva;
1088 }
1089 
1090 void intel_svm_unbind(struct iommu_sva *sva)
1091 {
1092 	struct intel_svm_dev *sdev;
1093 
1094 	mutex_lock(&pasid_mutex);
1095 	sdev = to_intel_svm_dev(sva);
1096 	intel_svm_unbind_mm(sdev->dev, sdev->pasid);
1097 	mutex_unlock(&pasid_mutex);
1098 }
1099 
1100 u32 intel_svm_get_pasid(struct iommu_sva *sva)
1101 {
1102 	struct intel_svm_dev *sdev;
1103 	u32 pasid;
1104 
1105 	mutex_lock(&pasid_mutex);
1106 	sdev = to_intel_svm_dev(sva);
1107 	pasid = sdev->pasid;
1108 	mutex_unlock(&pasid_mutex);
1109 
1110 	return pasid;
1111 }
1112 
1113 int intel_svm_page_response(struct device *dev,
1114 			    struct iommu_fault_event *evt,
1115 			    struct iommu_page_response *msg)
1116 {
1117 	struct iommu_fault_page_request *prm;
1118 	struct intel_svm_dev *sdev = NULL;
1119 	struct intel_svm *svm = NULL;
1120 	struct intel_iommu *iommu;
1121 	bool private_present;
1122 	bool pasid_present;
1123 	bool last_page;
1124 	u8 bus, devfn;
1125 	int ret = 0;
1126 	u16 sid;
1127 
1128 	if (!dev || !dev_is_pci(dev))
1129 		return -ENODEV;
1130 
1131 	iommu = device_to_iommu(dev, &bus, &devfn);
1132 	if (!iommu)
1133 		return -ENODEV;
1134 
1135 	if (!msg || !evt)
1136 		return -EINVAL;
1137 
1138 	mutex_lock(&pasid_mutex);
1139 
1140 	prm = &evt->fault.prm;
1141 	sid = PCI_DEVID(bus, devfn);
1142 	pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
1143 	private_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
1144 	last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
1145 
1146 	if (!pasid_present) {
1147 		ret = -EINVAL;
1148 		goto out;
1149 	}
1150 
1151 	if (prm->pasid == 0 || prm->pasid >= PASID_MAX) {
1152 		ret = -EINVAL;
1153 		goto out;
1154 	}
1155 
1156 	ret = pasid_to_svm_sdev(dev, prm->pasid, &svm, &sdev);
1157 	if (ret || !sdev) {
1158 		ret = -ENODEV;
1159 		goto out;
1160 	}
1161 
1162 	/*
1163 	 * For responses from userspace, need to make sure that the
1164 	 * pasid has been bound to its mm.
1165 	 */
1166 	if (svm->flags & SVM_FLAG_GUEST_MODE) {
1167 		struct mm_struct *mm;
1168 
1169 		mm = get_task_mm(current);
1170 		if (!mm) {
1171 			ret = -EINVAL;
1172 			goto out;
1173 		}
1174 
1175 		if (mm != svm->mm) {
1176 			ret = -ENODEV;
1177 			mmput(mm);
1178 			goto out;
1179 		}
1180 
1181 		mmput(mm);
1182 	}
1183 
1184 	/*
1185 	 * Per VT-d spec. v3.0 ch7.7, system software must respond
1186 	 * with page group response if private data is present (PDP)
1187 	 * or last page in group (LPIG) bit is set. This is an
1188 	 * additional VT-d requirement beyond PCI ATS spec.
1189 	 */
1190 	if (last_page || private_present) {
1191 		struct qi_desc desc;
1192 
1193 		desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
1194 				QI_PGRP_PASID_P(pasid_present) |
1195 				QI_PGRP_PDP(private_present) |
1196 				QI_PGRP_RESP_CODE(msg->code) |
1197 				QI_PGRP_RESP_TYPE;
1198 		desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
1199 		desc.qw2 = 0;
1200 		desc.qw3 = 0;
1201 		if (private_present)
1202 			memcpy(&desc.qw2, prm->private_data,
1203 			       sizeof(prm->private_data));
1204 
1205 		qi_submit_sync(iommu, &desc, 1, 0);
1206 	}
1207 out:
1208 	mutex_unlock(&pasid_mutex);
1209 	return ret;
1210 }
1211