xref: /openbmc/linux/drivers/iommu/intel/svm.c (revision c4f7ac64)
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 mm_struct *mm, struct intel_svm_dev **sd)
466 {
467 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
468 	struct intel_svm *svm = NULL, *t;
469 	struct device_domain_info *info;
470 	struct intel_svm_dev *sdev;
471 	unsigned long iflags;
472 	int pasid_max;
473 	int ret;
474 
475 	if (!iommu || dmar_disabled)
476 		return -EINVAL;
477 
478 	if (!intel_svm_capable(iommu))
479 		return -ENOTSUPP;
480 
481 	if (dev_is_pci(dev)) {
482 		pasid_max = pci_max_pasids(to_pci_dev(dev));
483 		if (pasid_max < 0)
484 			return -EINVAL;
485 	} else
486 		pasid_max = 1 << 20;
487 
488 	/* Bind supervisor PASID shuld have mm = NULL */
489 	if (flags & SVM_FLAG_SUPERVISOR_MODE) {
490 		if (!ecap_srs(iommu->ecap) || mm) {
491 			pr_err("Supervisor PASID with user provided mm.\n");
492 			return -EINVAL;
493 		}
494 	}
495 
496 	list_for_each_entry(t, &global_svm_list, list) {
497 		if (t->mm != mm)
498 			continue;
499 
500 		svm = t;
501 		if (svm->pasid >= pasid_max) {
502 			dev_warn(dev,
503 				 "Limited PASID width. Cannot use existing PASID %d\n",
504 				 svm->pasid);
505 			ret = -ENOSPC;
506 			goto out;
507 		}
508 
509 		/* Find the matching device in svm list */
510 		for_each_svm_dev(sdev, svm, dev) {
511 			sdev->users++;
512 			goto success;
513 		}
514 
515 		break;
516 	}
517 
518 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
519 	if (!sdev) {
520 		ret = -ENOMEM;
521 		goto out;
522 	}
523 	sdev->dev = dev;
524 	sdev->iommu = iommu;
525 
526 	ret = intel_iommu_enable_pasid(iommu, dev);
527 	if (ret) {
528 		kfree(sdev);
529 		goto out;
530 	}
531 
532 	info = get_domain_info(dev);
533 	sdev->did = FLPT_DEFAULT_DID;
534 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
535 	if (info->ats_enabled) {
536 		sdev->dev_iotlb = 1;
537 		sdev->qdep = info->ats_qdep;
538 		if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
539 			sdev->qdep = 0;
540 	}
541 
542 	/* Finish the setup now we know we're keeping it */
543 	sdev->users = 1;
544 	init_rcu_head(&sdev->rcu);
545 
546 	if (!svm) {
547 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
548 		if (!svm) {
549 			ret = -ENOMEM;
550 			kfree(sdev);
551 			goto out;
552 		}
553 
554 		if (pasid_max > intel_pasid_max_id)
555 			pasid_max = intel_pasid_max_id;
556 
557 		/* Do not use PASID 0, reserved for RID to PASID */
558 		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
559 					  pasid_max - 1, svm);
560 		if (svm->pasid == INVALID_IOASID) {
561 			kfree(svm);
562 			kfree(sdev);
563 			ret = -ENOSPC;
564 			goto out;
565 		}
566 		svm->notifier.ops = &intel_mmuops;
567 		svm->mm = mm;
568 		svm->flags = flags;
569 		INIT_LIST_HEAD_RCU(&svm->devs);
570 		INIT_LIST_HEAD(&svm->list);
571 		ret = -ENOMEM;
572 		if (mm) {
573 			ret = mmu_notifier_register(&svm->notifier, mm);
574 			if (ret) {
575 				ioasid_put(svm->pasid);
576 				kfree(svm);
577 				kfree(sdev);
578 				goto out;
579 			}
580 		}
581 
582 		spin_lock_irqsave(&iommu->lock, iflags);
583 		ret = intel_pasid_setup_first_level(iommu, dev,
584 				mm ? mm->pgd : init_mm.pgd,
585 				svm->pasid, FLPT_DEFAULT_DID,
586 				(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
587 				(cpu_feature_enabled(X86_FEATURE_LA57) ?
588 				 PASID_FLAG_FL5LP : 0));
589 		spin_unlock_irqrestore(&iommu->lock, iflags);
590 		if (ret) {
591 			if (mm)
592 				mmu_notifier_unregister(&svm->notifier, mm);
593 			ioasid_put(svm->pasid);
594 			kfree(svm);
595 			kfree(sdev);
596 			goto out;
597 		}
598 
599 		list_add_tail(&svm->list, &global_svm_list);
600 		if (mm) {
601 			/* The newly allocated pasid is loaded to the mm. */
602 			load_pasid(mm, svm->pasid);
603 		}
604 	} else {
605 		/*
606 		 * Binding a new device with existing PASID, need to setup
607 		 * the PASID entry.
608 		 */
609 		spin_lock_irqsave(&iommu->lock, iflags);
610 		ret = intel_pasid_setup_first_level(iommu, dev,
611 						mm ? mm->pgd : init_mm.pgd,
612 						svm->pasid, FLPT_DEFAULT_DID,
613 						(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
614 						(cpu_feature_enabled(X86_FEATURE_LA57) ?
615 						PASID_FLAG_FL5LP : 0));
616 		spin_unlock_irqrestore(&iommu->lock, iflags);
617 		if (ret) {
618 			kfree(sdev);
619 			goto out;
620 		}
621 	}
622 	list_add_rcu(&sdev->list, &svm->devs);
623 success:
624 	sdev->pasid = svm->pasid;
625 	sdev->sva.dev = dev;
626 	if (sd)
627 		*sd = sdev;
628 	ret = 0;
629 out:
630 	return ret;
631 }
632 
633 /* Caller must hold pasid_mutex */
634 static int intel_svm_unbind_mm(struct device *dev, u32 pasid)
635 {
636 	struct intel_svm_dev *sdev;
637 	struct intel_iommu *iommu;
638 	struct intel_svm *svm;
639 	int ret = -EINVAL;
640 
641 	iommu = device_to_iommu(dev, NULL, NULL);
642 	if (!iommu)
643 		goto out;
644 
645 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
646 	if (ret)
647 		goto out;
648 
649 	if (sdev) {
650 		sdev->users--;
651 		if (!sdev->users) {
652 			list_del_rcu(&sdev->list);
653 			/* Flush the PASID cache and IOTLB for this device.
654 			 * Note that we do depend on the hardware *not* using
655 			 * the PASID any more. Just as we depend on other
656 			 * devices never using PASIDs that they have no right
657 			 * to use. We have a *shared* PASID table, because it's
658 			 * large and has to be physically contiguous. So it's
659 			 * hard to be as defensive as we might like. */
660 			intel_pasid_tear_down_entry(iommu, dev,
661 						    svm->pasid, false);
662 			intel_svm_drain_prq(dev, svm->pasid);
663 			kfree_rcu(sdev, rcu);
664 
665 			if (list_empty(&svm->devs)) {
666 				ioasid_put(svm->pasid);
667 				if (svm->mm) {
668 					mmu_notifier_unregister(&svm->notifier, svm->mm);
669 					/* Clear mm's pasid. */
670 					load_pasid(svm->mm, PASID_DISABLED);
671 				}
672 				list_del(&svm->list);
673 				/* We mandate that no page faults may be outstanding
674 				 * for the PASID when intel_svm_unbind_mm() is called.
675 				 * If that is not obeyed, subtle errors will happen.
676 				 * Let's make them less subtle... */
677 				memset(svm, 0x6b, sizeof(*svm));
678 				kfree(svm);
679 			}
680 		}
681 	}
682 out:
683 	return ret;
684 }
685 
686 /* Page request queue descriptor */
687 struct page_req_dsc {
688 	union {
689 		struct {
690 			u64 type:8;
691 			u64 pasid_present:1;
692 			u64 priv_data_present:1;
693 			u64 rsvd:6;
694 			u64 rid:16;
695 			u64 pasid:20;
696 			u64 exe_req:1;
697 			u64 pm_req:1;
698 			u64 rsvd2:10;
699 		};
700 		u64 qw_0;
701 	};
702 	union {
703 		struct {
704 			u64 rd_req:1;
705 			u64 wr_req:1;
706 			u64 lpig:1;
707 			u64 prg_index:9;
708 			u64 addr:52;
709 		};
710 		u64 qw_1;
711 	};
712 	u64 priv_data[2];
713 };
714 
715 #define PRQ_RING_MASK	((0x1000 << PRQ_ORDER) - 0x20)
716 
717 static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
718 {
719 	unsigned long requested = 0;
720 
721 	if (req->exe_req)
722 		requested |= VM_EXEC;
723 
724 	if (req->rd_req)
725 		requested |= VM_READ;
726 
727 	if (req->wr_req)
728 		requested |= VM_WRITE;
729 
730 	return (requested & ~vma->vm_flags) != 0;
731 }
732 
733 static bool is_canonical_address(u64 addr)
734 {
735 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
736 	long saddr = (long) addr;
737 
738 	return (((saddr << shift) >> shift) == saddr);
739 }
740 
741 /**
742  * intel_svm_drain_prq - Drain page requests and responses for a pasid
743  * @dev: target device
744  * @pasid: pasid for draining
745  *
746  * Drain all pending page requests and responses related to @pasid in both
747  * software and hardware. This is supposed to be called after the device
748  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
749  * and DevTLB have been invalidated.
750  *
751  * It waits until all pending page requests for @pasid in the page fault
752  * queue are completed by the prq handling thread. Then follow the steps
753  * described in VT-d spec CH7.10 to drain all page requests and page
754  * responses pending in the hardware.
755  */
756 static void intel_svm_drain_prq(struct device *dev, u32 pasid)
757 {
758 	struct device_domain_info *info;
759 	struct dmar_domain *domain;
760 	struct intel_iommu *iommu;
761 	struct qi_desc desc[3];
762 	struct pci_dev *pdev;
763 	int head, tail;
764 	u16 sid, did;
765 	int qdep;
766 
767 	info = get_domain_info(dev);
768 	if (WARN_ON(!info || !dev_is_pci(dev)))
769 		return;
770 
771 	if (!info->pri_enabled)
772 		return;
773 
774 	iommu = info->iommu;
775 	domain = info->domain;
776 	pdev = to_pci_dev(dev);
777 	sid = PCI_DEVID(info->bus, info->devfn);
778 	did = domain->iommu_did[iommu->seq_id];
779 	qdep = pci_ats_queue_depth(pdev);
780 
781 	/*
782 	 * Check and wait until all pending page requests in the queue are
783 	 * handled by the prq handling thread.
784 	 */
785 prq_retry:
786 	reinit_completion(&iommu->prq_complete);
787 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
788 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
789 	while (head != tail) {
790 		struct page_req_dsc *req;
791 
792 		req = &iommu->prq[head / sizeof(*req)];
793 		if (!req->pasid_present || req->pasid != pasid) {
794 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
795 			continue;
796 		}
797 
798 		wait_for_completion(&iommu->prq_complete);
799 		goto prq_retry;
800 	}
801 
802 	/*
803 	 * Perform steps described in VT-d spec CH7.10 to drain page
804 	 * requests and responses in hardware.
805 	 */
806 	memset(desc, 0, sizeof(desc));
807 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
808 			QI_IWD_FENCE |
809 			QI_IWD_TYPE;
810 	desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
811 			QI_EIOTLB_DID(did) |
812 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
813 			QI_EIOTLB_TYPE;
814 	desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
815 			QI_DEV_EIOTLB_SID(sid) |
816 			QI_DEV_EIOTLB_QDEP(qdep) |
817 			QI_DEIOTLB_TYPE |
818 			QI_DEV_IOTLB_PFSID(info->pfsid);
819 qi_retry:
820 	reinit_completion(&iommu->prq_complete);
821 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
822 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
823 		wait_for_completion(&iommu->prq_complete);
824 		goto qi_retry;
825 	}
826 }
827 
828 static int prq_to_iommu_prot(struct page_req_dsc *req)
829 {
830 	int prot = 0;
831 
832 	if (req->rd_req)
833 		prot |= IOMMU_FAULT_PERM_READ;
834 	if (req->wr_req)
835 		prot |= IOMMU_FAULT_PERM_WRITE;
836 	if (req->exe_req)
837 		prot |= IOMMU_FAULT_PERM_EXEC;
838 	if (req->pm_req)
839 		prot |= IOMMU_FAULT_PERM_PRIV;
840 
841 	return prot;
842 }
843 
844 static int
845 intel_svm_prq_report(struct device *dev, struct page_req_dsc *desc)
846 {
847 	struct iommu_fault_event event;
848 
849 	if (!dev || !dev_is_pci(dev))
850 		return -ENODEV;
851 
852 	/* Fill in event data for device specific processing */
853 	memset(&event, 0, sizeof(struct iommu_fault_event));
854 	event.fault.type = IOMMU_FAULT_PAGE_REQ;
855 	event.fault.prm.addr = (u64)desc->addr << VTD_PAGE_SHIFT;
856 	event.fault.prm.pasid = desc->pasid;
857 	event.fault.prm.grpid = desc->prg_index;
858 	event.fault.prm.perm = prq_to_iommu_prot(desc);
859 
860 	if (desc->lpig)
861 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
862 	if (desc->pasid_present) {
863 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
864 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
865 	}
866 	if (desc->priv_data_present) {
867 		/*
868 		 * Set last page in group bit if private data is present,
869 		 * page response is required as it does for LPIG.
870 		 * iommu_report_device_fault() doesn't understand this vendor
871 		 * specific requirement thus we set last_page as a workaround.
872 		 */
873 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
874 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
875 		memcpy(event.fault.prm.private_data, desc->priv_data,
876 		       sizeof(desc->priv_data));
877 	}
878 
879 	return iommu_report_device_fault(dev, &event);
880 }
881 
882 static irqreturn_t prq_event_thread(int irq, void *d)
883 {
884 	struct intel_svm_dev *sdev = NULL;
885 	struct intel_iommu *iommu = d;
886 	struct intel_svm *svm = NULL;
887 	int head, tail, handled = 0;
888 	unsigned int flags = 0;
889 
890 	/* Clear PPR bit before reading head/tail registers, to
891 	 * ensure that we get a new interrupt if needed. */
892 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
893 
894 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
895 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
896 	while (head != tail) {
897 		struct vm_area_struct *vma;
898 		struct page_req_dsc *req;
899 		struct qi_desc resp;
900 		int result;
901 		vm_fault_t ret;
902 		u64 address;
903 
904 		handled = 1;
905 		req = &iommu->prq[head / sizeof(*req)];
906 		result = QI_RESP_INVALID;
907 		address = (u64)req->addr << VTD_PAGE_SHIFT;
908 		if (!req->pasid_present) {
909 			pr_err("%s: Page request without PASID: %08llx %08llx\n",
910 			       iommu->name, ((unsigned long long *)req)[0],
911 			       ((unsigned long long *)req)[1]);
912 			goto no_pasid;
913 		}
914 		/* We shall not receive page request for supervisor SVM */
915 		if (req->pm_req && (req->rd_req | req->wr_req)) {
916 			pr_err("Unexpected page request in Privilege Mode");
917 			/* No need to find the matching sdev as for bad_req */
918 			goto no_pasid;
919 		}
920 		/* DMA read with exec requeset is not supported. */
921 		if (req->exe_req && req->rd_req) {
922 			pr_err("Execution request not supported\n");
923 			goto no_pasid;
924 		}
925 		if (!svm || svm->pasid != req->pasid) {
926 			rcu_read_lock();
927 			svm = ioasid_find(NULL, req->pasid, NULL);
928 			/* It *can't* go away, because the driver is not permitted
929 			 * to unbind the mm while any page faults are outstanding.
930 			 * So we only need RCU to protect the internal idr code. */
931 			rcu_read_unlock();
932 			if (IS_ERR_OR_NULL(svm)) {
933 				pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
934 				       iommu->name, req->pasid, ((unsigned long long *)req)[0],
935 				       ((unsigned long long *)req)[1]);
936 				goto no_pasid;
937 			}
938 		}
939 
940 		if (!sdev || sdev->sid != req->rid) {
941 			struct intel_svm_dev *t;
942 
943 			sdev = NULL;
944 			rcu_read_lock();
945 			list_for_each_entry_rcu(t, &svm->devs, list) {
946 				if (t->sid == req->rid) {
947 					sdev = t;
948 					break;
949 				}
950 			}
951 			rcu_read_unlock();
952 		}
953 
954 		/* Since we're using init_mm.pgd directly, we should never take
955 		 * any faults on kernel addresses. */
956 		if (!svm->mm)
957 			goto bad_req;
958 
959 		/* If address is not canonical, return invalid response */
960 		if (!is_canonical_address(address))
961 			goto bad_req;
962 
963 		/*
964 		 * If prq is to be handled outside iommu driver via receiver of
965 		 * the fault notifiers, we skip the page response here.
966 		 */
967 		if (svm->flags & SVM_FLAG_GUEST_MODE) {
968 			if (sdev && !intel_svm_prq_report(sdev->dev, req))
969 				goto prq_advance;
970 			else
971 				goto bad_req;
972 		}
973 
974 		/* If the mm is already defunct, don't handle faults. */
975 		if (!mmget_not_zero(svm->mm))
976 			goto bad_req;
977 
978 		mmap_read_lock(svm->mm);
979 		vma = find_extend_vma(svm->mm, address);
980 		if (!vma || address < vma->vm_start)
981 			goto invalid;
982 
983 		if (access_error(vma, req))
984 			goto invalid;
985 
986 		flags = FAULT_FLAG_USER | FAULT_FLAG_REMOTE;
987 		if (req->wr_req)
988 			flags |= FAULT_FLAG_WRITE;
989 
990 		ret = handle_mm_fault(vma, address, flags, NULL);
991 		if (ret & VM_FAULT_ERROR)
992 			goto invalid;
993 
994 		result = QI_RESP_SUCCESS;
995 invalid:
996 		mmap_read_unlock(svm->mm);
997 		mmput(svm->mm);
998 bad_req:
999 		/* We get here in the error case where the PASID lookup failed,
1000 		   and these can be NULL. Do not use them below this point! */
1001 		sdev = NULL;
1002 		svm = NULL;
1003 no_pasid:
1004 		if (req->lpig || req->priv_data_present) {
1005 			/*
1006 			 * Per VT-d spec. v3.0 ch7.7, system software must
1007 			 * respond with page group response if private data
1008 			 * is present (PDP) or last page in group (LPIG) bit
1009 			 * is set. This is an additional VT-d feature beyond
1010 			 * PCI ATS spec.
1011 			 */
1012 			resp.qw0 = QI_PGRP_PASID(req->pasid) |
1013 				QI_PGRP_DID(req->rid) |
1014 				QI_PGRP_PASID_P(req->pasid_present) |
1015 				QI_PGRP_PDP(req->priv_data_present) |
1016 				QI_PGRP_RESP_CODE(result) |
1017 				QI_PGRP_RESP_TYPE;
1018 			resp.qw1 = QI_PGRP_IDX(req->prg_index) |
1019 				QI_PGRP_LPIG(req->lpig);
1020 			resp.qw2 = 0;
1021 			resp.qw3 = 0;
1022 
1023 			if (req->priv_data_present)
1024 				memcpy(&resp.qw2, req->priv_data,
1025 				       sizeof(req->priv_data));
1026 			qi_submit_sync(iommu, &resp, 1, 0);
1027 		}
1028 prq_advance:
1029 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
1030 	}
1031 
1032 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
1033 
1034 	/*
1035 	 * Clear the page request overflow bit and wake up all threads that
1036 	 * are waiting for the completion of this handling.
1037 	 */
1038 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
1039 		pr_info_ratelimited("IOMMU: %s: PRQ overflow detected\n",
1040 				    iommu->name);
1041 		head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
1042 		tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
1043 		if (head == tail) {
1044 			writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
1045 			pr_info_ratelimited("IOMMU: %s: PRQ overflow cleared",
1046 					    iommu->name);
1047 		}
1048 	}
1049 
1050 	if (!completion_done(&iommu->prq_complete))
1051 		complete(&iommu->prq_complete);
1052 
1053 	return IRQ_RETVAL(handled);
1054 }
1055 
1056 #define to_intel_svm_dev(handle) container_of(handle, struct intel_svm_dev, sva)
1057 struct iommu_sva *
1058 intel_svm_bind(struct device *dev, struct mm_struct *mm, void *drvdata)
1059 {
1060 	struct iommu_sva *sva = ERR_PTR(-EINVAL);
1061 	struct intel_svm_dev *sdev = NULL;
1062 	unsigned int flags = 0;
1063 	int ret;
1064 
1065 	/*
1066 	 * TODO: Consolidate with generic iommu-sva bind after it is merged.
1067 	 * It will require shared SVM data structures, i.e. combine io_mm
1068 	 * and intel_svm etc.
1069 	 */
1070 	if (drvdata)
1071 		flags = *(unsigned int *)drvdata;
1072 	mutex_lock(&pasid_mutex);
1073 	ret = intel_svm_bind_mm(dev, flags, mm, &sdev);
1074 	if (ret)
1075 		sva = ERR_PTR(ret);
1076 	else if (sdev)
1077 		sva = &sdev->sva;
1078 	else
1079 		WARN(!sdev, "SVM bind succeeded with no sdev!\n");
1080 
1081 	mutex_unlock(&pasid_mutex);
1082 
1083 	return sva;
1084 }
1085 
1086 void intel_svm_unbind(struct iommu_sva *sva)
1087 {
1088 	struct intel_svm_dev *sdev;
1089 
1090 	mutex_lock(&pasid_mutex);
1091 	sdev = to_intel_svm_dev(sva);
1092 	intel_svm_unbind_mm(sdev->dev, sdev->pasid);
1093 	mutex_unlock(&pasid_mutex);
1094 }
1095 
1096 u32 intel_svm_get_pasid(struct iommu_sva *sva)
1097 {
1098 	struct intel_svm_dev *sdev;
1099 	u32 pasid;
1100 
1101 	mutex_lock(&pasid_mutex);
1102 	sdev = to_intel_svm_dev(sva);
1103 	pasid = sdev->pasid;
1104 	mutex_unlock(&pasid_mutex);
1105 
1106 	return pasid;
1107 }
1108 
1109 int intel_svm_page_response(struct device *dev,
1110 			    struct iommu_fault_event *evt,
1111 			    struct iommu_page_response *msg)
1112 {
1113 	struct iommu_fault_page_request *prm;
1114 	struct intel_svm_dev *sdev = NULL;
1115 	struct intel_svm *svm = NULL;
1116 	struct intel_iommu *iommu;
1117 	bool private_present;
1118 	bool pasid_present;
1119 	bool last_page;
1120 	u8 bus, devfn;
1121 	int ret = 0;
1122 	u16 sid;
1123 
1124 	if (!dev || !dev_is_pci(dev))
1125 		return -ENODEV;
1126 
1127 	iommu = device_to_iommu(dev, &bus, &devfn);
1128 	if (!iommu)
1129 		return -ENODEV;
1130 
1131 	if (!msg || !evt)
1132 		return -EINVAL;
1133 
1134 	mutex_lock(&pasid_mutex);
1135 
1136 	prm = &evt->fault.prm;
1137 	sid = PCI_DEVID(bus, devfn);
1138 	pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
1139 	private_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
1140 	last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
1141 
1142 	if (!pasid_present) {
1143 		ret = -EINVAL;
1144 		goto out;
1145 	}
1146 
1147 	if (prm->pasid == 0 || prm->pasid >= PASID_MAX) {
1148 		ret = -EINVAL;
1149 		goto out;
1150 	}
1151 
1152 	ret = pasid_to_svm_sdev(dev, prm->pasid, &svm, &sdev);
1153 	if (ret || !sdev) {
1154 		ret = -ENODEV;
1155 		goto out;
1156 	}
1157 
1158 	/*
1159 	 * For responses from userspace, need to make sure that the
1160 	 * pasid has been bound to its mm.
1161 	 */
1162 	if (svm->flags & SVM_FLAG_GUEST_MODE) {
1163 		struct mm_struct *mm;
1164 
1165 		mm = get_task_mm(current);
1166 		if (!mm) {
1167 			ret = -EINVAL;
1168 			goto out;
1169 		}
1170 
1171 		if (mm != svm->mm) {
1172 			ret = -ENODEV;
1173 			mmput(mm);
1174 			goto out;
1175 		}
1176 
1177 		mmput(mm);
1178 	}
1179 
1180 	/*
1181 	 * Per VT-d spec. v3.0 ch7.7, system software must respond
1182 	 * with page group response if private data is present (PDP)
1183 	 * or last page in group (LPIG) bit is set. This is an
1184 	 * additional VT-d requirement beyond PCI ATS spec.
1185 	 */
1186 	if (last_page || private_present) {
1187 		struct qi_desc desc;
1188 
1189 		desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
1190 				QI_PGRP_PASID_P(pasid_present) |
1191 				QI_PGRP_PDP(private_present) |
1192 				QI_PGRP_RESP_CODE(msg->code) |
1193 				QI_PGRP_RESP_TYPE;
1194 		desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
1195 		desc.qw2 = 0;
1196 		desc.qw3 = 0;
1197 		if (private_present)
1198 			memcpy(&desc.qw2, prm->private_data,
1199 			       sizeof(prm->private_data));
1200 
1201 		qi_submit_sync(iommu, &desc, 1, 0);
1202 	}
1203 out:
1204 	mutex_unlock(&pasid_mutex);
1205 	return ret;
1206 }
1207