xref: /linux/drivers/iommu/intel/svm.c (revision c01044cc819160323f3ca4acd44fca487c4432e6)
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 intel_flush_svm_range_dev (struct intel_svm *svm, struct intel_svm_dev *sdev,
122 				unsigned long address, unsigned long pages, int ih)
123 {
124 	struct qi_desc desc;
125 
126 	if (pages == -1) {
127 		desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
128 			QI_EIOTLB_DID(sdev->did) |
129 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
130 			QI_EIOTLB_TYPE;
131 		desc.qw1 = 0;
132 	} else {
133 		int mask = ilog2(__roundup_pow_of_two(pages));
134 
135 		desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
136 				QI_EIOTLB_DID(sdev->did) |
137 				QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID) |
138 				QI_EIOTLB_TYPE;
139 		desc.qw1 = QI_EIOTLB_ADDR(address) |
140 				QI_EIOTLB_IH(ih) |
141 				QI_EIOTLB_AM(mask);
142 	}
143 	desc.qw2 = 0;
144 	desc.qw3 = 0;
145 	qi_submit_sync(svm->iommu, &desc, 1, 0);
146 
147 	if (sdev->dev_iotlb) {
148 		desc.qw0 = QI_DEV_EIOTLB_PASID(svm->pasid) |
149 				QI_DEV_EIOTLB_SID(sdev->sid) |
150 				QI_DEV_EIOTLB_QDEP(sdev->qdep) |
151 				QI_DEIOTLB_TYPE;
152 		if (pages == -1) {
153 			desc.qw1 = QI_DEV_EIOTLB_ADDR(-1ULL >> 1) |
154 					QI_DEV_EIOTLB_SIZE;
155 		} else if (pages > 1) {
156 			/* The least significant zero bit indicates the size. So,
157 			 * for example, an "address" value of 0x12345f000 will
158 			 * flush from 0x123440000 to 0x12347ffff (256KiB). */
159 			unsigned long last = address + ((unsigned long)(pages - 1) << VTD_PAGE_SHIFT);
160 			unsigned long mask = __rounddown_pow_of_two(address ^ last);
161 
162 			desc.qw1 = QI_DEV_EIOTLB_ADDR((address & ~mask) |
163 					(mask - 1)) | QI_DEV_EIOTLB_SIZE;
164 		} else {
165 			desc.qw1 = QI_DEV_EIOTLB_ADDR(address);
166 		}
167 		desc.qw2 = 0;
168 		desc.qw3 = 0;
169 		qi_submit_sync(svm->iommu, &desc, 1, 0);
170 	}
171 }
172 
173 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
174 				unsigned long pages, int ih)
175 {
176 	struct intel_svm_dev *sdev;
177 
178 	rcu_read_lock();
179 	list_for_each_entry_rcu(sdev, &svm->devs, list)
180 		intel_flush_svm_range_dev(svm, sdev, address, pages, ih);
181 	rcu_read_unlock();
182 }
183 
184 /* Pages have been freed at this point */
185 static void intel_invalidate_range(struct mmu_notifier *mn,
186 				   struct mm_struct *mm,
187 				   unsigned long start, unsigned long end)
188 {
189 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
190 
191 	intel_flush_svm_range(svm, start,
192 			      (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0);
193 }
194 
195 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
196 {
197 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
198 	struct intel_svm_dev *sdev;
199 
200 	/* This might end up being called from exit_mmap(), *before* the page
201 	 * tables are cleared. And __mmu_notifier_release() will delete us from
202 	 * the list of notifiers so that our invalidate_range() callback doesn't
203 	 * get called when the page tables are cleared. So we need to protect
204 	 * against hardware accessing those page tables.
205 	 *
206 	 * We do it by clearing the entry in the PASID table and then flushing
207 	 * the IOTLB and the PASID table caches. This might upset hardware;
208 	 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
209 	 * page) so that we end up taking a fault that the hardware really
210 	 * *has* to handle gracefully without affecting other processes.
211 	 */
212 	rcu_read_lock();
213 	list_for_each_entry_rcu(sdev, &svm->devs, list)
214 		intel_pasid_tear_down_entry(svm->iommu, sdev->dev,
215 					    svm->pasid, true);
216 	rcu_read_unlock();
217 
218 }
219 
220 static const struct mmu_notifier_ops intel_mmuops = {
221 	.release = intel_mm_release,
222 	.invalidate_range = intel_invalidate_range,
223 };
224 
225 static DEFINE_MUTEX(pasid_mutex);
226 static LIST_HEAD(global_svm_list);
227 
228 #define for_each_svm_dev(sdev, svm, d)			\
229 	list_for_each_entry((sdev), &(svm)->devs, list)	\
230 		if ((d) != (sdev)->dev) {} else
231 
232 static int pasid_to_svm_sdev(struct device *dev, unsigned int pasid,
233 			     struct intel_svm **rsvm,
234 			     struct intel_svm_dev **rsdev)
235 {
236 	struct intel_svm_dev *d, *sdev = NULL;
237 	struct intel_svm *svm;
238 
239 	/* The caller should hold the pasid_mutex lock */
240 	if (WARN_ON(!mutex_is_locked(&pasid_mutex)))
241 		return -EINVAL;
242 
243 	if (pasid == INVALID_IOASID || pasid >= PASID_MAX)
244 		return -EINVAL;
245 
246 	svm = ioasid_find(NULL, pasid, NULL);
247 	if (IS_ERR(svm))
248 		return PTR_ERR(svm);
249 
250 	if (!svm)
251 		goto out;
252 
253 	/*
254 	 * If we found svm for the PASID, there must be at least one device
255 	 * bond.
256 	 */
257 	if (WARN_ON(list_empty(&svm->devs)))
258 		return -EINVAL;
259 
260 	rcu_read_lock();
261 	list_for_each_entry_rcu(d, &svm->devs, list) {
262 		if (d->dev == dev) {
263 			sdev = d;
264 			break;
265 		}
266 	}
267 	rcu_read_unlock();
268 
269 out:
270 	*rsvm = svm;
271 	*rsdev = sdev;
272 
273 	return 0;
274 }
275 
276 int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
277 			  struct iommu_gpasid_bind_data *data)
278 {
279 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
280 	struct intel_svm_dev *sdev = NULL;
281 	struct dmar_domain *dmar_domain;
282 	struct intel_svm *svm = NULL;
283 	int ret = 0;
284 
285 	if (WARN_ON(!iommu) || !data)
286 		return -EINVAL;
287 
288 	if (data->version != IOMMU_GPASID_BIND_VERSION_1 ||
289 	    data->format != IOMMU_PASID_FORMAT_INTEL_VTD)
290 		return -EINVAL;
291 
292 	if (!dev_is_pci(dev))
293 		return -ENOTSUPP;
294 
295 	/* VT-d supports devices with full 20 bit PASIDs only */
296 	if (pci_max_pasids(to_pci_dev(dev)) != PASID_MAX)
297 		return -EINVAL;
298 
299 	/*
300 	 * We only check host PASID range, we have no knowledge to check
301 	 * guest PASID range.
302 	 */
303 	if (data->hpasid <= 0 || data->hpasid >= PASID_MAX)
304 		return -EINVAL;
305 
306 	dmar_domain = to_dmar_domain(domain);
307 
308 	mutex_lock(&pasid_mutex);
309 	ret = pasid_to_svm_sdev(dev, data->hpasid, &svm, &sdev);
310 	if (ret)
311 		goto out;
312 
313 	if (sdev) {
314 		/*
315 		 * Do not allow multiple bindings of the same device-PASID since
316 		 * there is only one SL page tables per PASID. We may revisit
317 		 * once sharing PGD across domains are supported.
318 		 */
319 		dev_warn_ratelimited(dev, "Already bound with PASID %u\n",
320 				     svm->pasid);
321 		ret = -EBUSY;
322 		goto out;
323 	}
324 
325 	if (!svm) {
326 		/* We come here when PASID has never been bond to a device. */
327 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
328 		if (!svm) {
329 			ret = -ENOMEM;
330 			goto out;
331 		}
332 		/* REVISIT: upper layer/VFIO can track host process that bind
333 		 * the PASID. ioasid_set = mm might be sufficient for vfio to
334 		 * check pasid VMM ownership. We can drop the following line
335 		 * once VFIO and IOASID set check is in place.
336 		 */
337 		svm->mm = get_task_mm(current);
338 		svm->pasid = data->hpasid;
339 		if (data->flags & IOMMU_SVA_GPASID_VAL) {
340 			svm->gpasid = data->gpasid;
341 			svm->flags |= SVM_FLAG_GUEST_PASID;
342 		}
343 		ioasid_set_data(data->hpasid, svm);
344 		INIT_LIST_HEAD_RCU(&svm->devs);
345 		mmput(svm->mm);
346 	}
347 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
348 	if (!sdev) {
349 		ret = -ENOMEM;
350 		goto out;
351 	}
352 	sdev->dev = dev;
353 
354 	/* Only count users if device has aux domains */
355 	if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
356 		sdev->users = 1;
357 
358 	/* Set up device context entry for PASID if not enabled already */
359 	ret = intel_iommu_enable_pasid(iommu, sdev->dev);
360 	if (ret) {
361 		dev_err_ratelimited(dev, "Failed to enable PASID capability\n");
362 		kfree(sdev);
363 		goto out;
364 	}
365 
366 	/*
367 	 * PASID table is per device for better security. Therefore, for
368 	 * each bind of a new device even with an existing PASID, we need to
369 	 * call the nested mode setup function here.
370 	 */
371 	spin_lock(&iommu->lock);
372 	ret = intel_pasid_setup_nested(iommu, dev,
373 				       (pgd_t *)(uintptr_t)data->gpgd,
374 				       data->hpasid, &data->vtd, dmar_domain,
375 				       data->addr_width);
376 	spin_unlock(&iommu->lock);
377 	if (ret) {
378 		dev_err_ratelimited(dev, "Failed to set up PASID %llu in nested mode, Err %d\n",
379 				    data->hpasid, ret);
380 		/*
381 		 * PASID entry should be in cleared state if nested mode
382 		 * set up failed. So we only need to clear IOASID tracking
383 		 * data such that free call will succeed.
384 		 */
385 		kfree(sdev);
386 		goto out;
387 	}
388 
389 	svm->flags |= SVM_FLAG_GUEST_MODE;
390 
391 	init_rcu_head(&sdev->rcu);
392 	list_add_rcu(&sdev->list, &svm->devs);
393  out:
394 	if (!IS_ERR_OR_NULL(svm) && list_empty(&svm->devs)) {
395 		ioasid_set_data(data->hpasid, NULL);
396 		kfree(svm);
397 	}
398 
399 	mutex_unlock(&pasid_mutex);
400 	return ret;
401 }
402 
403 int intel_svm_unbind_gpasid(struct device *dev, u32 pasid)
404 {
405 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
406 	struct intel_svm_dev *sdev;
407 	struct intel_svm *svm;
408 	int ret;
409 
410 	if (WARN_ON(!iommu))
411 		return -EINVAL;
412 
413 	mutex_lock(&pasid_mutex);
414 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
415 	if (ret)
416 		goto out;
417 
418 	if (sdev) {
419 		if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
420 			sdev->users--;
421 		if (!sdev->users) {
422 			list_del_rcu(&sdev->list);
423 			intel_pasid_tear_down_entry(iommu, dev,
424 						    svm->pasid, false);
425 			intel_svm_drain_prq(dev, svm->pasid);
426 			kfree_rcu(sdev, rcu);
427 
428 			if (list_empty(&svm->devs)) {
429 				/*
430 				 * We do not free the IOASID here in that
431 				 * IOMMU driver did not allocate it.
432 				 * Unlike native SVM, IOASID for guest use was
433 				 * allocated prior to the bind call.
434 				 * In any case, if the free call comes before
435 				 * the unbind, IOMMU driver will get notified
436 				 * and perform cleanup.
437 				 */
438 				ioasid_set_data(pasid, NULL);
439 				kfree(svm);
440 			}
441 		}
442 	}
443 out:
444 	mutex_unlock(&pasid_mutex);
445 	return ret;
446 }
447 
448 static void _load_pasid(void *unused)
449 {
450 	update_pasid();
451 }
452 
453 static void load_pasid(struct mm_struct *mm, u32 pasid)
454 {
455 	mutex_lock(&mm->context.lock);
456 
457 	/* Synchronize with READ_ONCE in update_pasid(). */
458 	smp_store_release(&mm->pasid, pasid);
459 
460 	/* Update PASID MSR on all CPUs running the mm's tasks. */
461 	on_each_cpu_mask(mm_cpumask(mm), _load_pasid, NULL, true);
462 
463 	mutex_unlock(&mm->context.lock);
464 }
465 
466 /* Caller must hold pasid_mutex, mm reference */
467 static int
468 intel_svm_bind_mm(struct device *dev, unsigned int flags,
469 		  struct svm_dev_ops *ops,
470 		  struct mm_struct *mm, struct intel_svm_dev **sd)
471 {
472 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
473 	struct device_domain_info *info;
474 	struct intel_svm_dev *sdev;
475 	struct intel_svm *svm = NULL;
476 	int pasid_max;
477 	int ret;
478 
479 	if (!iommu || dmar_disabled)
480 		return -EINVAL;
481 
482 	if (!intel_svm_capable(iommu))
483 		return -ENOTSUPP;
484 
485 	if (dev_is_pci(dev)) {
486 		pasid_max = pci_max_pasids(to_pci_dev(dev));
487 		if (pasid_max < 0)
488 			return -EINVAL;
489 	} else
490 		pasid_max = 1 << 20;
491 
492 	/* Bind supervisor PASID shuld have mm = NULL */
493 	if (flags & SVM_FLAG_SUPERVISOR_MODE) {
494 		if (!ecap_srs(iommu->ecap) || mm) {
495 			pr_err("Supervisor PASID with user provided mm.\n");
496 			return -EINVAL;
497 		}
498 	}
499 
500 	if (!(flags & SVM_FLAG_PRIVATE_PASID)) {
501 		struct intel_svm *t;
502 
503 		list_for_each_entry(t, &global_svm_list, list) {
504 			if (t->mm != mm || (t->flags & SVM_FLAG_PRIVATE_PASID))
505 				continue;
506 
507 			svm = t;
508 			if (svm->pasid >= pasid_max) {
509 				dev_warn(dev,
510 					 "Limited PASID width. Cannot use existing PASID %d\n",
511 					 svm->pasid);
512 				ret = -ENOSPC;
513 				goto out;
514 			}
515 
516 			/* Find the matching device in svm list */
517 			for_each_svm_dev(sdev, svm, dev) {
518 				if (sdev->ops != ops) {
519 					ret = -EBUSY;
520 					goto out;
521 				}
522 				sdev->users++;
523 				goto success;
524 			}
525 
526 			break;
527 		}
528 	}
529 
530 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
531 	if (!sdev) {
532 		ret = -ENOMEM;
533 		goto out;
534 	}
535 	sdev->dev = dev;
536 
537 	ret = intel_iommu_enable_pasid(iommu, dev);
538 	if (ret) {
539 		kfree(sdev);
540 		goto out;
541 	}
542 
543 	info = get_domain_info(dev);
544 	sdev->did = FLPT_DEFAULT_DID;
545 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
546 	if (info->ats_enabled) {
547 		sdev->dev_iotlb = 1;
548 		sdev->qdep = info->ats_qdep;
549 		if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
550 			sdev->qdep = 0;
551 	}
552 
553 	/* Finish the setup now we know we're keeping it */
554 	sdev->users = 1;
555 	sdev->ops = ops;
556 	init_rcu_head(&sdev->rcu);
557 
558 	if (!svm) {
559 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
560 		if (!svm) {
561 			ret = -ENOMEM;
562 			kfree(sdev);
563 			goto out;
564 		}
565 		svm->iommu = iommu;
566 
567 		if (pasid_max > intel_pasid_max_id)
568 			pasid_max = intel_pasid_max_id;
569 
570 		/* Do not use PASID 0, reserved for RID to PASID */
571 		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
572 					  pasid_max - 1, svm);
573 		if (svm->pasid == INVALID_IOASID) {
574 			kfree(svm);
575 			kfree(sdev);
576 			ret = -ENOSPC;
577 			goto out;
578 		}
579 		svm->notifier.ops = &intel_mmuops;
580 		svm->mm = mm;
581 		svm->flags = flags;
582 		INIT_LIST_HEAD_RCU(&svm->devs);
583 		INIT_LIST_HEAD(&svm->list);
584 		ret = -ENOMEM;
585 		if (mm) {
586 			ret = mmu_notifier_register(&svm->notifier, mm);
587 			if (ret) {
588 				ioasid_free(svm->pasid);
589 				kfree(svm);
590 				kfree(sdev);
591 				goto out;
592 			}
593 		}
594 
595 		spin_lock(&iommu->lock);
596 		ret = intel_pasid_setup_first_level(iommu, dev,
597 				mm ? mm->pgd : init_mm.pgd,
598 				svm->pasid, FLPT_DEFAULT_DID,
599 				(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
600 				(cpu_feature_enabled(X86_FEATURE_LA57) ?
601 				 PASID_FLAG_FL5LP : 0));
602 		spin_unlock(&iommu->lock);
603 		if (ret) {
604 			if (mm)
605 				mmu_notifier_unregister(&svm->notifier, mm);
606 			ioasid_free(svm->pasid);
607 			kfree(svm);
608 			kfree(sdev);
609 			goto out;
610 		}
611 
612 		list_add_tail(&svm->list, &global_svm_list);
613 		if (mm) {
614 			/* The newly allocated pasid is loaded to the mm. */
615 			load_pasid(mm, svm->pasid);
616 		}
617 	} else {
618 		/*
619 		 * Binding a new device with existing PASID, need to setup
620 		 * the PASID entry.
621 		 */
622 		spin_lock(&iommu->lock);
623 		ret = intel_pasid_setup_first_level(iommu, dev,
624 						mm ? mm->pgd : init_mm.pgd,
625 						svm->pasid, FLPT_DEFAULT_DID,
626 						(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
627 						(cpu_feature_enabled(X86_FEATURE_LA57) ?
628 						PASID_FLAG_FL5LP : 0));
629 		spin_unlock(&iommu->lock);
630 		if (ret) {
631 			kfree(sdev);
632 			goto out;
633 		}
634 	}
635 	list_add_rcu(&sdev->list, &svm->devs);
636 success:
637 	sdev->pasid = svm->pasid;
638 	sdev->sva.dev = dev;
639 	if (sd)
640 		*sd = sdev;
641 	ret = 0;
642 out:
643 	return ret;
644 }
645 
646 /* Caller must hold pasid_mutex */
647 static int intel_svm_unbind_mm(struct device *dev, u32 pasid)
648 {
649 	struct intel_svm_dev *sdev;
650 	struct intel_iommu *iommu;
651 	struct intel_svm *svm;
652 	int ret = -EINVAL;
653 
654 	iommu = device_to_iommu(dev, NULL, NULL);
655 	if (!iommu)
656 		goto out;
657 
658 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
659 	if (ret)
660 		goto out;
661 
662 	if (sdev) {
663 		sdev->users--;
664 		if (!sdev->users) {
665 			list_del_rcu(&sdev->list);
666 			/* Flush the PASID cache and IOTLB for this device.
667 			 * Note that we do depend on the hardware *not* using
668 			 * the PASID any more. Just as we depend on other
669 			 * devices never using PASIDs that they have no right
670 			 * to use. We have a *shared* PASID table, because it's
671 			 * large and has to be physically contiguous. So it's
672 			 * hard to be as defensive as we might like. */
673 			intel_pasid_tear_down_entry(iommu, dev,
674 						    svm->pasid, false);
675 			intel_svm_drain_prq(dev, svm->pasid);
676 			kfree_rcu(sdev, rcu);
677 
678 			if (list_empty(&svm->devs)) {
679 				ioasid_free(svm->pasid);
680 				if (svm->mm) {
681 					mmu_notifier_unregister(&svm->notifier, svm->mm);
682 					/* Clear mm's pasid. */
683 					load_pasid(svm->mm, PASID_DISABLED);
684 				}
685 				list_del(&svm->list);
686 				/* We mandate that no page faults may be outstanding
687 				 * for the PASID when intel_svm_unbind_mm() is called.
688 				 * If that is not obeyed, subtle errors will happen.
689 				 * Let's make them less subtle... */
690 				memset(svm, 0x6b, sizeof(*svm));
691 				kfree(svm);
692 			}
693 		}
694 	}
695 out:
696 	return ret;
697 }
698 
699 /* Page request queue descriptor */
700 struct page_req_dsc {
701 	union {
702 		struct {
703 			u64 type:8;
704 			u64 pasid_present:1;
705 			u64 priv_data_present:1;
706 			u64 rsvd:6;
707 			u64 rid:16;
708 			u64 pasid:20;
709 			u64 exe_req:1;
710 			u64 pm_req:1;
711 			u64 rsvd2:10;
712 		};
713 		u64 qw_0;
714 	};
715 	union {
716 		struct {
717 			u64 rd_req:1;
718 			u64 wr_req:1;
719 			u64 lpig:1;
720 			u64 prg_index:9;
721 			u64 addr:52;
722 		};
723 		u64 qw_1;
724 	};
725 	u64 priv_data[2];
726 };
727 
728 #define PRQ_RING_MASK	((0x1000 << PRQ_ORDER) - 0x20)
729 
730 static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
731 {
732 	unsigned long requested = 0;
733 
734 	if (req->exe_req)
735 		requested |= VM_EXEC;
736 
737 	if (req->rd_req)
738 		requested |= VM_READ;
739 
740 	if (req->wr_req)
741 		requested |= VM_WRITE;
742 
743 	return (requested & ~vma->vm_flags) != 0;
744 }
745 
746 static bool is_canonical_address(u64 addr)
747 {
748 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
749 	long saddr = (long) addr;
750 
751 	return (((saddr << shift) >> shift) == saddr);
752 }
753 
754 /**
755  * intel_svm_drain_prq - Drain page requests and responses for a pasid
756  * @dev: target device
757  * @pasid: pasid for draining
758  *
759  * Drain all pending page requests and responses related to @pasid in both
760  * software and hardware. This is supposed to be called after the device
761  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
762  * and DevTLB have been invalidated.
763  *
764  * It waits until all pending page requests for @pasid in the page fault
765  * queue are completed by the prq handling thread. Then follow the steps
766  * described in VT-d spec CH7.10 to drain all page requests and page
767  * responses pending in the hardware.
768  */
769 static void intel_svm_drain_prq(struct device *dev, u32 pasid)
770 {
771 	struct device_domain_info *info;
772 	struct dmar_domain *domain;
773 	struct intel_iommu *iommu;
774 	struct qi_desc desc[3];
775 	struct pci_dev *pdev;
776 	int head, tail;
777 	u16 sid, did;
778 	int qdep;
779 
780 	info = get_domain_info(dev);
781 	if (WARN_ON(!info || !dev_is_pci(dev)))
782 		return;
783 
784 	if (!info->pri_enabled)
785 		return;
786 
787 	iommu = info->iommu;
788 	domain = info->domain;
789 	pdev = to_pci_dev(dev);
790 	sid = PCI_DEVID(info->bus, info->devfn);
791 	did = domain->iommu_did[iommu->seq_id];
792 	qdep = pci_ats_queue_depth(pdev);
793 
794 	/*
795 	 * Check and wait until all pending page requests in the queue are
796 	 * handled by the prq handling thread.
797 	 */
798 prq_retry:
799 	reinit_completion(&iommu->prq_complete);
800 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
801 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
802 	while (head != tail) {
803 		struct page_req_dsc *req;
804 
805 		req = &iommu->prq[head / sizeof(*req)];
806 		if (!req->pasid_present || req->pasid != pasid) {
807 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
808 			continue;
809 		}
810 
811 		wait_for_completion(&iommu->prq_complete);
812 		goto prq_retry;
813 	}
814 
815 	/*
816 	 * Perform steps described in VT-d spec CH7.10 to drain page
817 	 * requests and responses in hardware.
818 	 */
819 	memset(desc, 0, sizeof(desc));
820 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
821 			QI_IWD_FENCE |
822 			QI_IWD_TYPE;
823 	desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
824 			QI_EIOTLB_DID(did) |
825 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
826 			QI_EIOTLB_TYPE;
827 	desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
828 			QI_DEV_EIOTLB_SID(sid) |
829 			QI_DEV_EIOTLB_QDEP(qdep) |
830 			QI_DEIOTLB_TYPE |
831 			QI_DEV_IOTLB_PFSID(info->pfsid);
832 qi_retry:
833 	reinit_completion(&iommu->prq_complete);
834 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
835 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
836 		wait_for_completion(&iommu->prq_complete);
837 		goto qi_retry;
838 	}
839 }
840 
841 static int prq_to_iommu_prot(struct page_req_dsc *req)
842 {
843 	int prot = 0;
844 
845 	if (req->rd_req)
846 		prot |= IOMMU_FAULT_PERM_READ;
847 	if (req->wr_req)
848 		prot |= IOMMU_FAULT_PERM_WRITE;
849 	if (req->exe_req)
850 		prot |= IOMMU_FAULT_PERM_EXEC;
851 	if (req->pm_req)
852 		prot |= IOMMU_FAULT_PERM_PRIV;
853 
854 	return prot;
855 }
856 
857 static int
858 intel_svm_prq_report(struct device *dev, struct page_req_dsc *desc)
859 {
860 	struct iommu_fault_event event;
861 
862 	if (!dev || !dev_is_pci(dev))
863 		return -ENODEV;
864 
865 	/* Fill in event data for device specific processing */
866 	memset(&event, 0, sizeof(struct iommu_fault_event));
867 	event.fault.type = IOMMU_FAULT_PAGE_REQ;
868 	event.fault.prm.addr = desc->addr;
869 	event.fault.prm.pasid = desc->pasid;
870 	event.fault.prm.grpid = desc->prg_index;
871 	event.fault.prm.perm = prq_to_iommu_prot(desc);
872 
873 	if (desc->lpig)
874 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
875 	if (desc->pasid_present) {
876 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
877 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
878 	}
879 	if (desc->priv_data_present) {
880 		/*
881 		 * Set last page in group bit if private data is present,
882 		 * page response is required as it does for LPIG.
883 		 * iommu_report_device_fault() doesn't understand this vendor
884 		 * specific requirement thus we set last_page as a workaround.
885 		 */
886 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
887 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
888 		memcpy(event.fault.prm.private_data, desc->priv_data,
889 		       sizeof(desc->priv_data));
890 	}
891 
892 	return iommu_report_device_fault(dev, &event);
893 }
894 
895 static irqreturn_t prq_event_thread(int irq, void *d)
896 {
897 	struct intel_svm_dev *sdev = NULL;
898 	struct intel_iommu *iommu = d;
899 	struct intel_svm *svm = NULL;
900 	int head, tail, handled = 0;
901 
902 	/* Clear PPR bit before reading head/tail registers, to
903 	 * ensure that we get a new interrupt if needed. */
904 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
905 
906 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
907 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
908 	while (head != tail) {
909 		struct vm_area_struct *vma;
910 		struct page_req_dsc *req;
911 		struct qi_desc resp;
912 		int result;
913 		vm_fault_t ret;
914 		u64 address;
915 
916 		handled = 1;
917 
918 		req = &iommu->prq[head / sizeof(*req)];
919 
920 		result = QI_RESP_FAILURE;
921 		address = (u64)req->addr << VTD_PAGE_SHIFT;
922 		if (!req->pasid_present) {
923 			pr_err("%s: Page request without PASID: %08llx %08llx\n",
924 			       iommu->name, ((unsigned long long *)req)[0],
925 			       ((unsigned long long *)req)[1]);
926 			goto no_pasid;
927 		}
928 
929 		if (!svm || svm->pasid != req->pasid) {
930 			rcu_read_lock();
931 			svm = ioasid_find(NULL, req->pasid, NULL);
932 			/* It *can't* go away, because the driver is not permitted
933 			 * to unbind the mm while any page faults are outstanding.
934 			 * So we only need RCU to protect the internal idr code. */
935 			rcu_read_unlock();
936 			if (IS_ERR_OR_NULL(svm)) {
937 				pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
938 				       iommu->name, req->pasid, ((unsigned long long *)req)[0],
939 				       ((unsigned long long *)req)[1]);
940 				goto no_pasid;
941 			}
942 		}
943 
944 		if (!sdev || sdev->sid != req->rid) {
945 			struct intel_svm_dev *t;
946 
947 			sdev = NULL;
948 			rcu_read_lock();
949 			list_for_each_entry_rcu(t, &svm->devs, list) {
950 				if (t->sid == req->rid) {
951 					sdev = t;
952 					break;
953 				}
954 			}
955 			rcu_read_unlock();
956 		}
957 
958 		result = QI_RESP_INVALID;
959 		/* Since we're using init_mm.pgd directly, we should never take
960 		 * any faults on kernel addresses. */
961 		if (!svm->mm)
962 			goto bad_req;
963 
964 		/* If address is not canonical, return invalid response */
965 		if (!is_canonical_address(address))
966 			goto bad_req;
967 
968 		/*
969 		 * If prq is to be handled outside iommu driver via receiver of
970 		 * the fault notifiers, we skip the page response here.
971 		 */
972 		if (svm->flags & SVM_FLAG_GUEST_MODE) {
973 			if (sdev && !intel_svm_prq_report(sdev->dev, req))
974 				goto prq_advance;
975 			else
976 				goto bad_req;
977 		}
978 
979 		/* If the mm is already defunct, don't handle faults. */
980 		if (!mmget_not_zero(svm->mm))
981 			goto bad_req;
982 
983 		mmap_read_lock(svm->mm);
984 		vma = find_extend_vma(svm->mm, address);
985 		if (!vma || address < vma->vm_start)
986 			goto invalid;
987 
988 		if (access_error(vma, req))
989 			goto invalid;
990 
991 		ret = handle_mm_fault(vma, address,
992 				      req->wr_req ? FAULT_FLAG_WRITE : 0,
993 				      NULL);
994 		if (ret & VM_FAULT_ERROR)
995 			goto invalid;
996 
997 		result = QI_RESP_SUCCESS;
998 invalid:
999 		mmap_read_unlock(svm->mm);
1000 		mmput(svm->mm);
1001 bad_req:
1002 		WARN_ON(!sdev);
1003 		if (sdev && sdev->ops && sdev->ops->fault_cb) {
1004 			int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
1005 				(req->exe_req << 1) | (req->pm_req);
1006 			sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr,
1007 					    req->priv_data, rwxp, result);
1008 		}
1009 		/* We get here in the error case where the PASID lookup failed,
1010 		   and these can be NULL. Do not use them below this point! */
1011 		sdev = NULL;
1012 		svm = NULL;
1013 no_pasid:
1014 		if (req->lpig || req->priv_data_present) {
1015 			/*
1016 			 * Per VT-d spec. v3.0 ch7.7, system software must
1017 			 * respond with page group response if private data
1018 			 * is present (PDP) or last page in group (LPIG) bit
1019 			 * is set. This is an additional VT-d feature beyond
1020 			 * PCI ATS spec.
1021 			 */
1022 			resp.qw0 = QI_PGRP_PASID(req->pasid) |
1023 				QI_PGRP_DID(req->rid) |
1024 				QI_PGRP_PASID_P(req->pasid_present) |
1025 				QI_PGRP_PDP(req->pasid_present) |
1026 				QI_PGRP_RESP_CODE(result) |
1027 				QI_PGRP_RESP_TYPE;
1028 			resp.qw1 = QI_PGRP_IDX(req->prg_index) |
1029 				QI_PGRP_LPIG(req->lpig);
1030 
1031 			if (req->priv_data_present)
1032 				memcpy(&resp.qw2, req->priv_data,
1033 				       sizeof(req->priv_data));
1034 			resp.qw2 = 0;
1035 			resp.qw3 = 0;
1036 			qi_submit_sync(iommu, &resp, 1, 0);
1037 		}
1038 prq_advance:
1039 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
1040 	}
1041 
1042 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
1043 
1044 	/*
1045 	 * Clear the page request overflow bit and wake up all threads that
1046 	 * are waiting for the completion of this handling.
1047 	 */
1048 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO)
1049 		writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
1050 
1051 	if (!completion_done(&iommu->prq_complete))
1052 		complete(&iommu->prq_complete);
1053 
1054 	return IRQ_RETVAL(handled);
1055 }
1056 
1057 #define to_intel_svm_dev(handle) container_of(handle, struct intel_svm_dev, sva)
1058 struct iommu_sva *
1059 intel_svm_bind(struct device *dev, struct mm_struct *mm, void *drvdata)
1060 {
1061 	struct iommu_sva *sva = ERR_PTR(-EINVAL);
1062 	struct intel_svm_dev *sdev = NULL;
1063 	unsigned int flags = 0;
1064 	int ret;
1065 
1066 	/*
1067 	 * TODO: Consolidate with generic iommu-sva bind after it is merged.
1068 	 * It will require shared SVM data structures, i.e. combine io_mm
1069 	 * and intel_svm etc.
1070 	 */
1071 	if (drvdata)
1072 		flags = *(unsigned int *)drvdata;
1073 	mutex_lock(&pasid_mutex);
1074 	ret = intel_svm_bind_mm(dev, flags, NULL, mm, &sdev);
1075 	if (ret)
1076 		sva = ERR_PTR(ret);
1077 	else if (sdev)
1078 		sva = &sdev->sva;
1079 	else
1080 		WARN(!sdev, "SVM bind succeeded with no sdev!\n");
1081 
1082 	mutex_unlock(&pasid_mutex);
1083 
1084 	return sva;
1085 }
1086 
1087 void intel_svm_unbind(struct iommu_sva *sva)
1088 {
1089 	struct intel_svm_dev *sdev;
1090 
1091 	mutex_lock(&pasid_mutex);
1092 	sdev = to_intel_svm_dev(sva);
1093 	intel_svm_unbind_mm(sdev->dev, sdev->pasid);
1094 	mutex_unlock(&pasid_mutex);
1095 }
1096 
1097 u32 intel_svm_get_pasid(struct iommu_sva *sva)
1098 {
1099 	struct intel_svm_dev *sdev;
1100 	u32 pasid;
1101 
1102 	mutex_lock(&pasid_mutex);
1103 	sdev = to_intel_svm_dev(sva);
1104 	pasid = sdev->pasid;
1105 	mutex_unlock(&pasid_mutex);
1106 
1107 	return pasid;
1108 }
1109 
1110 int intel_svm_page_response(struct device *dev,
1111 			    struct iommu_fault_event *evt,
1112 			    struct iommu_page_response *msg)
1113 {
1114 	struct iommu_fault_page_request *prm;
1115 	struct intel_svm_dev *sdev = NULL;
1116 	struct intel_svm *svm = NULL;
1117 	struct intel_iommu *iommu;
1118 	bool private_present;
1119 	bool pasid_present;
1120 	bool last_page;
1121 	u8 bus, devfn;
1122 	int ret = 0;
1123 	u16 sid;
1124 
1125 	if (!dev || !dev_is_pci(dev))
1126 		return -ENODEV;
1127 
1128 	iommu = device_to_iommu(dev, &bus, &devfn);
1129 	if (!iommu)
1130 		return -ENODEV;
1131 
1132 	if (!msg || !evt)
1133 		return -EINVAL;
1134 
1135 	mutex_lock(&pasid_mutex);
1136 
1137 	prm = &evt->fault.prm;
1138 	sid = PCI_DEVID(bus, devfn);
1139 	pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
1140 	private_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
1141 	last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
1142 
1143 	if (!pasid_present) {
1144 		ret = -EINVAL;
1145 		goto out;
1146 	}
1147 
1148 	if (prm->pasid == 0 || prm->pasid >= PASID_MAX) {
1149 		ret = -EINVAL;
1150 		goto out;
1151 	}
1152 
1153 	ret = pasid_to_svm_sdev(dev, prm->pasid, &svm, &sdev);
1154 	if (ret || !sdev) {
1155 		ret = -ENODEV;
1156 		goto out;
1157 	}
1158 
1159 	/*
1160 	 * For responses from userspace, need to make sure that the
1161 	 * pasid has been bound to its mm.
1162 	 */
1163 	if (svm->flags & SVM_FLAG_GUEST_MODE) {
1164 		struct mm_struct *mm;
1165 
1166 		mm = get_task_mm(current);
1167 		if (!mm) {
1168 			ret = -EINVAL;
1169 			goto out;
1170 		}
1171 
1172 		if (mm != svm->mm) {
1173 			ret = -ENODEV;
1174 			mmput(mm);
1175 			goto out;
1176 		}
1177 
1178 		mmput(mm);
1179 	}
1180 
1181 	/*
1182 	 * Per VT-d spec. v3.0 ch7.7, system software must respond
1183 	 * with page group response if private data is present (PDP)
1184 	 * or last page in group (LPIG) bit is set. This is an
1185 	 * additional VT-d requirement beyond PCI ATS spec.
1186 	 */
1187 	if (last_page || private_present) {
1188 		struct qi_desc desc;
1189 
1190 		desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
1191 				QI_PGRP_PASID_P(pasid_present) |
1192 				QI_PGRP_PDP(private_present) |
1193 				QI_PGRP_RESP_CODE(msg->code) |
1194 				QI_PGRP_RESP_TYPE;
1195 		desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
1196 		desc.qw2 = 0;
1197 		desc.qw3 = 0;
1198 		if (private_present)
1199 			memcpy(&desc.qw2, prm->private_data,
1200 			       sizeof(prm->private_data));
1201 
1202 		qi_submit_sync(iommu, &desc, 1, 0);
1203 	}
1204 out:
1205 	mutex_unlock(&pasid_mutex);
1206 	return ret;
1207 }
1208