xref: /openbmc/linux/drivers/iommu/intel/svm.c (revision 29c37341)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright © 2015 Intel Corporation.
4  *
5  * Authors: David Woodhouse <dwmw2@infradead.org>
6  */
7 
8 #include <linux/intel-iommu.h>
9 #include <linux/mmu_notifier.h>
10 #include <linux/sched.h>
11 #include <linux/sched/mm.h>
12 #include <linux/slab.h>
13 #include <linux/intel-svm.h>
14 #include <linux/rculist.h>
15 #include <linux/pci.h>
16 #include <linux/pci-ats.h>
17 #include <linux/dmar.h>
18 #include <linux/interrupt.h>
19 #include <linux/mm_types.h>
20 #include <linux/ioasid.h>
21 #include <asm/page.h>
22 
23 #include "pasid.h"
24 
25 static irqreturn_t prq_event_thread(int irq, void *d);
26 static void intel_svm_drain_prq(struct device *dev, int pasid);
27 
28 #define PRQ_ORDER 0
29 
30 int intel_svm_enable_prq(struct intel_iommu *iommu)
31 {
32 	struct page *pages;
33 	int irq, ret;
34 
35 	pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
36 	if (!pages) {
37 		pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
38 			iommu->name);
39 		return -ENOMEM;
40 	}
41 	iommu->prq = page_address(pages);
42 
43 	irq = dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED + iommu->seq_id, iommu->node, iommu);
44 	if (irq <= 0) {
45 		pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
46 		       iommu->name);
47 		ret = -EINVAL;
48 	err:
49 		free_pages((unsigned long)iommu->prq, PRQ_ORDER);
50 		iommu->prq = NULL;
51 		return ret;
52 	}
53 	iommu->pr_irq = irq;
54 
55 	snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
56 
57 	ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
58 				   iommu->prq_name, iommu);
59 	if (ret) {
60 		pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
61 		       iommu->name);
62 		dmar_free_hwirq(irq);
63 		iommu->pr_irq = 0;
64 		goto err;
65 	}
66 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
67 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
68 	dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
69 
70 	init_completion(&iommu->prq_complete);
71 
72 	return 0;
73 }
74 
75 int intel_svm_finish_prq(struct intel_iommu *iommu)
76 {
77 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
78 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
79 	dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
80 
81 	if (iommu->pr_irq) {
82 		free_irq(iommu->pr_irq, iommu);
83 		dmar_free_hwirq(iommu->pr_irq);
84 		iommu->pr_irq = 0;
85 	}
86 
87 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
88 	iommu->prq = NULL;
89 
90 	return 0;
91 }
92 
93 static inline bool intel_svm_capable(struct intel_iommu *iommu)
94 {
95 	return iommu->flags & VTD_FLAG_SVM_CAPABLE;
96 }
97 
98 void intel_svm_check(struct intel_iommu *iommu)
99 {
100 	if (!pasid_supported(iommu))
101 		return;
102 
103 	if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
104 	    !cap_fl1gp_support(iommu->cap)) {
105 		pr_err("%s SVM disabled, incompatible 1GB page capability\n",
106 		       iommu->name);
107 		return;
108 	}
109 
110 	if (cpu_feature_enabled(X86_FEATURE_LA57) &&
111 	    !cap_5lp_support(iommu->cap)) {
112 		pr_err("%s SVM disabled, incompatible paging mode\n",
113 		       iommu->name);
114 		return;
115 	}
116 
117 	iommu->flags |= VTD_FLAG_SVM_CAPABLE;
118 }
119 
120 static void intel_flush_svm_range_dev (struct intel_svm *svm, struct intel_svm_dev *sdev,
121 				unsigned long address, unsigned long pages, int ih)
122 {
123 	struct qi_desc desc;
124 
125 	if (pages == -1) {
126 		desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
127 			QI_EIOTLB_DID(sdev->did) |
128 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
129 			QI_EIOTLB_TYPE;
130 		desc.qw1 = 0;
131 	} else {
132 		int mask = ilog2(__roundup_pow_of_two(pages));
133 
134 		desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
135 				QI_EIOTLB_DID(sdev->did) |
136 				QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID) |
137 				QI_EIOTLB_TYPE;
138 		desc.qw1 = QI_EIOTLB_ADDR(address) |
139 				QI_EIOTLB_IH(ih) |
140 				QI_EIOTLB_AM(mask);
141 	}
142 	desc.qw2 = 0;
143 	desc.qw3 = 0;
144 	qi_submit_sync(svm->iommu, &desc, 1, 0);
145 
146 	if (sdev->dev_iotlb) {
147 		desc.qw0 = QI_DEV_EIOTLB_PASID(svm->pasid) |
148 				QI_DEV_EIOTLB_SID(sdev->sid) |
149 				QI_DEV_EIOTLB_QDEP(sdev->qdep) |
150 				QI_DEIOTLB_TYPE;
151 		if (pages == -1) {
152 			desc.qw1 = QI_DEV_EIOTLB_ADDR(-1ULL >> 1) |
153 					QI_DEV_EIOTLB_SIZE;
154 		} else if (pages > 1) {
155 			/* The least significant zero bit indicates the size. So,
156 			 * for example, an "address" value of 0x12345f000 will
157 			 * flush from 0x123440000 to 0x12347ffff (256KiB). */
158 			unsigned long last = address + ((unsigned long)(pages - 1) << VTD_PAGE_SHIFT);
159 			unsigned long mask = __rounddown_pow_of_two(address ^ last);
160 
161 			desc.qw1 = QI_DEV_EIOTLB_ADDR((address & ~mask) |
162 					(mask - 1)) | QI_DEV_EIOTLB_SIZE;
163 		} else {
164 			desc.qw1 = QI_DEV_EIOTLB_ADDR(address);
165 		}
166 		desc.qw2 = 0;
167 		desc.qw3 = 0;
168 		qi_submit_sync(svm->iommu, &desc, 1, 0);
169 	}
170 }
171 
172 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
173 				unsigned long pages, int ih)
174 {
175 	struct intel_svm_dev *sdev;
176 
177 	rcu_read_lock();
178 	list_for_each_entry_rcu(sdev, &svm->devs, list)
179 		intel_flush_svm_range_dev(svm, sdev, address, pages, ih);
180 	rcu_read_unlock();
181 }
182 
183 /* Pages have been freed at this point */
184 static void intel_invalidate_range(struct mmu_notifier *mn,
185 				   struct mm_struct *mm,
186 				   unsigned long start, unsigned long end)
187 {
188 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
189 
190 	intel_flush_svm_range(svm, start,
191 			      (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0);
192 }
193 
194 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
195 {
196 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
197 	struct intel_svm_dev *sdev;
198 
199 	/* This might end up being called from exit_mmap(), *before* the page
200 	 * tables are cleared. And __mmu_notifier_release() will delete us from
201 	 * the list of notifiers so that our invalidate_range() callback doesn't
202 	 * get called when the page tables are cleared. So we need to protect
203 	 * against hardware accessing those page tables.
204 	 *
205 	 * We do it by clearing the entry in the PASID table and then flushing
206 	 * the IOTLB and the PASID table caches. This might upset hardware;
207 	 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
208 	 * page) so that we end up taking a fault that the hardware really
209 	 * *has* to handle gracefully without affecting other processes.
210 	 */
211 	rcu_read_lock();
212 	list_for_each_entry_rcu(sdev, &svm->devs, list)
213 		intel_pasid_tear_down_entry(svm->iommu, sdev->dev,
214 					    svm->pasid, true);
215 	rcu_read_unlock();
216 
217 }
218 
219 static const struct mmu_notifier_ops intel_mmuops = {
220 	.release = intel_mm_release,
221 	.invalidate_range = intel_invalidate_range,
222 };
223 
224 static DEFINE_MUTEX(pasid_mutex);
225 static LIST_HEAD(global_svm_list);
226 
227 #define for_each_svm_dev(sdev, svm, d)			\
228 	list_for_each_entry((sdev), &(svm)->devs, list)	\
229 		if ((d) != (sdev)->dev) {} else
230 
231 static int pasid_to_svm_sdev(struct device *dev, unsigned int pasid,
232 			     struct intel_svm **rsvm,
233 			     struct intel_svm_dev **rsdev)
234 {
235 	struct intel_svm_dev *d, *sdev = NULL;
236 	struct intel_svm *svm;
237 
238 	/* The caller should hold the pasid_mutex lock */
239 	if (WARN_ON(!mutex_is_locked(&pasid_mutex)))
240 		return -EINVAL;
241 
242 	if (pasid == INVALID_IOASID || pasid >= PASID_MAX)
243 		return -EINVAL;
244 
245 	svm = ioasid_find(NULL, pasid, NULL);
246 	if (IS_ERR(svm))
247 		return PTR_ERR(svm);
248 
249 	if (!svm)
250 		goto out;
251 
252 	/*
253 	 * If we found svm for the PASID, there must be at least one device
254 	 * bond.
255 	 */
256 	if (WARN_ON(list_empty(&svm->devs)))
257 		return -EINVAL;
258 
259 	rcu_read_lock();
260 	list_for_each_entry_rcu(d, &svm->devs, list) {
261 		if (d->dev == dev) {
262 			sdev = d;
263 			break;
264 		}
265 	}
266 	rcu_read_unlock();
267 
268 out:
269 	*rsvm = svm;
270 	*rsdev = sdev;
271 
272 	return 0;
273 }
274 
275 int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
276 			  struct iommu_gpasid_bind_data *data)
277 {
278 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
279 	struct intel_svm_dev *sdev = NULL;
280 	struct dmar_domain *dmar_domain;
281 	struct intel_svm *svm = NULL;
282 	int ret = 0;
283 
284 	if (WARN_ON(!iommu) || !data)
285 		return -EINVAL;
286 
287 	if (data->version != IOMMU_GPASID_BIND_VERSION_1 ||
288 	    data->format != IOMMU_PASID_FORMAT_INTEL_VTD)
289 		return -EINVAL;
290 
291 	if (!dev_is_pci(dev))
292 		return -ENOTSUPP;
293 
294 	/* VT-d supports devices with full 20 bit PASIDs only */
295 	if (pci_max_pasids(to_pci_dev(dev)) != PASID_MAX)
296 		return -EINVAL;
297 
298 	/*
299 	 * We only check host PASID range, we have no knowledge to check
300 	 * guest PASID range.
301 	 */
302 	if (data->hpasid <= 0 || data->hpasid >= PASID_MAX)
303 		return -EINVAL;
304 
305 	dmar_domain = to_dmar_domain(domain);
306 
307 	mutex_lock(&pasid_mutex);
308 	ret = pasid_to_svm_sdev(dev, data->hpasid, &svm, &sdev);
309 	if (ret)
310 		goto out;
311 
312 	if (sdev) {
313 		/*
314 		 * Do not allow multiple bindings of the same device-PASID since
315 		 * there is only one SL page tables per PASID. We may revisit
316 		 * once sharing PGD across domains are supported.
317 		 */
318 		dev_warn_ratelimited(dev, "Already bound with PASID %u\n",
319 				     svm->pasid);
320 		ret = -EBUSY;
321 		goto out;
322 	}
323 
324 	if (!svm) {
325 		/* We come here when PASID has never been bond to a device. */
326 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
327 		if (!svm) {
328 			ret = -ENOMEM;
329 			goto out;
330 		}
331 		/* REVISIT: upper layer/VFIO can track host process that bind
332 		 * the PASID. ioasid_set = mm might be sufficient for vfio to
333 		 * check pasid VMM ownership. We can drop the following line
334 		 * once VFIO and IOASID set check is in place.
335 		 */
336 		svm->mm = get_task_mm(current);
337 		svm->pasid = data->hpasid;
338 		if (data->flags & IOMMU_SVA_GPASID_VAL) {
339 			svm->gpasid = data->gpasid;
340 			svm->flags |= SVM_FLAG_GUEST_PASID;
341 		}
342 		ioasid_set_data(data->hpasid, svm);
343 		INIT_LIST_HEAD_RCU(&svm->devs);
344 		mmput(svm->mm);
345 	}
346 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
347 	if (!sdev) {
348 		ret = -ENOMEM;
349 		goto out;
350 	}
351 	sdev->dev = dev;
352 
353 	/* Only count users if device has aux domains */
354 	if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
355 		sdev->users = 1;
356 
357 	/* Set up device context entry for PASID if not enabled already */
358 	ret = intel_iommu_enable_pasid(iommu, sdev->dev);
359 	if (ret) {
360 		dev_err_ratelimited(dev, "Failed to enable PASID capability\n");
361 		kfree(sdev);
362 		goto out;
363 	}
364 
365 	/*
366 	 * PASID table is per device for better security. Therefore, for
367 	 * each bind of a new device even with an existing PASID, we need to
368 	 * call the nested mode setup function here.
369 	 */
370 	spin_lock(&iommu->lock);
371 	ret = intel_pasid_setup_nested(iommu, dev,
372 				       (pgd_t *)(uintptr_t)data->gpgd,
373 				       data->hpasid, &data->vtd, dmar_domain,
374 				       data->addr_width);
375 	spin_unlock(&iommu->lock);
376 	if (ret) {
377 		dev_err_ratelimited(dev, "Failed to set up PASID %llu in nested mode, Err %d\n",
378 				    data->hpasid, ret);
379 		/*
380 		 * PASID entry should be in cleared state if nested mode
381 		 * set up failed. So we only need to clear IOASID tracking
382 		 * data such that free call will succeed.
383 		 */
384 		kfree(sdev);
385 		goto out;
386 	}
387 
388 	svm->flags |= SVM_FLAG_GUEST_MODE;
389 
390 	init_rcu_head(&sdev->rcu);
391 	list_add_rcu(&sdev->list, &svm->devs);
392  out:
393 	if (!IS_ERR_OR_NULL(svm) && list_empty(&svm->devs)) {
394 		ioasid_set_data(data->hpasid, NULL);
395 		kfree(svm);
396 	}
397 
398 	mutex_unlock(&pasid_mutex);
399 	return ret;
400 }
401 
402 int intel_svm_unbind_gpasid(struct device *dev, int pasid)
403 {
404 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
405 	struct intel_svm_dev *sdev;
406 	struct intel_svm *svm;
407 	int ret;
408 
409 	if (WARN_ON(!iommu))
410 		return -EINVAL;
411 
412 	mutex_lock(&pasid_mutex);
413 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
414 	if (ret)
415 		goto out;
416 
417 	if (sdev) {
418 		if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
419 			sdev->users--;
420 		if (!sdev->users) {
421 			list_del_rcu(&sdev->list);
422 			intel_pasid_tear_down_entry(iommu, dev,
423 						    svm->pasid, false);
424 			intel_svm_drain_prq(dev, svm->pasid);
425 			kfree_rcu(sdev, rcu);
426 
427 			if (list_empty(&svm->devs)) {
428 				/*
429 				 * We do not free the IOASID here in that
430 				 * IOMMU driver did not allocate it.
431 				 * Unlike native SVM, IOASID for guest use was
432 				 * allocated prior to the bind call.
433 				 * In any case, if the free call comes before
434 				 * the unbind, IOMMU driver will get notified
435 				 * and perform cleanup.
436 				 */
437 				ioasid_set_data(pasid, NULL);
438 				kfree(svm);
439 			}
440 		}
441 	}
442 out:
443 	mutex_unlock(&pasid_mutex);
444 	return ret;
445 }
446 
447 /* Caller must hold pasid_mutex, mm reference */
448 static int
449 intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
450 		  struct mm_struct *mm, struct intel_svm_dev **sd)
451 {
452 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
453 	struct device_domain_info *info;
454 	struct intel_svm_dev *sdev;
455 	struct intel_svm *svm = NULL;
456 	int pasid_max;
457 	int ret;
458 
459 	if (!iommu || dmar_disabled)
460 		return -EINVAL;
461 
462 	if (!intel_svm_capable(iommu))
463 		return -ENOTSUPP;
464 
465 	if (dev_is_pci(dev)) {
466 		pasid_max = pci_max_pasids(to_pci_dev(dev));
467 		if (pasid_max < 0)
468 			return -EINVAL;
469 	} else
470 		pasid_max = 1 << 20;
471 
472 	/* Bind supervisor PASID shuld have mm = NULL */
473 	if (flags & SVM_FLAG_SUPERVISOR_MODE) {
474 		if (!ecap_srs(iommu->ecap) || mm) {
475 			pr_err("Supervisor PASID with user provided mm.\n");
476 			return -EINVAL;
477 		}
478 	}
479 
480 	if (!(flags & SVM_FLAG_PRIVATE_PASID)) {
481 		struct intel_svm *t;
482 
483 		list_for_each_entry(t, &global_svm_list, list) {
484 			if (t->mm != mm || (t->flags & SVM_FLAG_PRIVATE_PASID))
485 				continue;
486 
487 			svm = t;
488 			if (svm->pasid >= pasid_max) {
489 				dev_warn(dev,
490 					 "Limited PASID width. Cannot use existing PASID %d\n",
491 					 svm->pasid);
492 				ret = -ENOSPC;
493 				goto out;
494 			}
495 
496 			/* Find the matching device in svm list */
497 			for_each_svm_dev(sdev, svm, dev) {
498 				if (sdev->ops != ops) {
499 					ret = -EBUSY;
500 					goto out;
501 				}
502 				sdev->users++;
503 				goto success;
504 			}
505 
506 			break;
507 		}
508 	}
509 
510 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
511 	if (!sdev) {
512 		ret = -ENOMEM;
513 		goto out;
514 	}
515 	sdev->dev = dev;
516 
517 	ret = intel_iommu_enable_pasid(iommu, dev);
518 	if (ret) {
519 		kfree(sdev);
520 		goto out;
521 	}
522 
523 	info = get_domain_info(dev);
524 	sdev->did = FLPT_DEFAULT_DID;
525 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
526 	if (info->ats_enabled) {
527 		sdev->dev_iotlb = 1;
528 		sdev->qdep = info->ats_qdep;
529 		if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
530 			sdev->qdep = 0;
531 	}
532 
533 	/* Finish the setup now we know we're keeping it */
534 	sdev->users = 1;
535 	sdev->ops = ops;
536 	init_rcu_head(&sdev->rcu);
537 
538 	if (!svm) {
539 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
540 		if (!svm) {
541 			ret = -ENOMEM;
542 			kfree(sdev);
543 			goto out;
544 		}
545 		svm->iommu = iommu;
546 
547 		if (pasid_max > intel_pasid_max_id)
548 			pasid_max = intel_pasid_max_id;
549 
550 		/* Do not use PASID 0, reserved for RID to PASID */
551 		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
552 					  pasid_max - 1, svm);
553 		if (svm->pasid == INVALID_IOASID) {
554 			kfree(svm);
555 			kfree(sdev);
556 			ret = -ENOSPC;
557 			goto out;
558 		}
559 		svm->notifier.ops = &intel_mmuops;
560 		svm->mm = mm;
561 		svm->flags = flags;
562 		INIT_LIST_HEAD_RCU(&svm->devs);
563 		INIT_LIST_HEAD(&svm->list);
564 		ret = -ENOMEM;
565 		if (mm) {
566 			ret = mmu_notifier_register(&svm->notifier, mm);
567 			if (ret) {
568 				ioasid_free(svm->pasid);
569 				kfree(svm);
570 				kfree(sdev);
571 				goto out;
572 			}
573 		}
574 
575 		spin_lock(&iommu->lock);
576 		ret = intel_pasid_setup_first_level(iommu, dev,
577 				mm ? mm->pgd : init_mm.pgd,
578 				svm->pasid, FLPT_DEFAULT_DID,
579 				(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
580 				(cpu_feature_enabled(X86_FEATURE_LA57) ?
581 				 PASID_FLAG_FL5LP : 0));
582 		spin_unlock(&iommu->lock);
583 		if (ret) {
584 			if (mm)
585 				mmu_notifier_unregister(&svm->notifier, mm);
586 			ioasid_free(svm->pasid);
587 			kfree(svm);
588 			kfree(sdev);
589 			goto out;
590 		}
591 
592 		list_add_tail(&svm->list, &global_svm_list);
593 	} else {
594 		/*
595 		 * Binding a new device with existing PASID, need to setup
596 		 * the PASID entry.
597 		 */
598 		spin_lock(&iommu->lock);
599 		ret = intel_pasid_setup_first_level(iommu, dev,
600 						mm ? mm->pgd : init_mm.pgd,
601 						svm->pasid, FLPT_DEFAULT_DID,
602 						(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
603 						(cpu_feature_enabled(X86_FEATURE_LA57) ?
604 						PASID_FLAG_FL5LP : 0));
605 		spin_unlock(&iommu->lock);
606 		if (ret) {
607 			kfree(sdev);
608 			goto out;
609 		}
610 	}
611 	list_add_rcu(&sdev->list, &svm->devs);
612 success:
613 	sdev->pasid = svm->pasid;
614 	sdev->sva.dev = dev;
615 	if (sd)
616 		*sd = sdev;
617 	ret = 0;
618 out:
619 	return ret;
620 }
621 
622 /* Caller must hold pasid_mutex */
623 static int intel_svm_unbind_mm(struct device *dev, int pasid)
624 {
625 	struct intel_svm_dev *sdev;
626 	struct intel_iommu *iommu;
627 	struct intel_svm *svm;
628 	int ret = -EINVAL;
629 
630 	iommu = device_to_iommu(dev, NULL, NULL);
631 	if (!iommu)
632 		goto out;
633 
634 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
635 	if (ret)
636 		goto out;
637 
638 	if (sdev) {
639 		sdev->users--;
640 		if (!sdev->users) {
641 			list_del_rcu(&sdev->list);
642 			/* Flush the PASID cache and IOTLB for this device.
643 			 * Note that we do depend on the hardware *not* using
644 			 * the PASID any more. Just as we depend on other
645 			 * devices never using PASIDs that they have no right
646 			 * to use. We have a *shared* PASID table, because it's
647 			 * large and has to be physically contiguous. So it's
648 			 * hard to be as defensive as we might like. */
649 			intel_pasid_tear_down_entry(iommu, dev,
650 						    svm->pasid, false);
651 			intel_svm_drain_prq(dev, svm->pasid);
652 			kfree_rcu(sdev, rcu);
653 
654 			if (list_empty(&svm->devs)) {
655 				ioasid_free(svm->pasid);
656 				if (svm->mm)
657 					mmu_notifier_unregister(&svm->notifier, svm->mm);
658 				list_del(&svm->list);
659 				/* We mandate that no page faults may be outstanding
660 				 * for the PASID when intel_svm_unbind_mm() is called.
661 				 * If that is not obeyed, subtle errors will happen.
662 				 * Let's make them less subtle... */
663 				memset(svm, 0x6b, sizeof(*svm));
664 				kfree(svm);
665 			}
666 		}
667 	}
668 out:
669 	return ret;
670 }
671 
672 /* Page request queue descriptor */
673 struct page_req_dsc {
674 	union {
675 		struct {
676 			u64 type:8;
677 			u64 pasid_present:1;
678 			u64 priv_data_present:1;
679 			u64 rsvd:6;
680 			u64 rid:16;
681 			u64 pasid:20;
682 			u64 exe_req:1;
683 			u64 pm_req:1;
684 			u64 rsvd2:10;
685 		};
686 		u64 qw_0;
687 	};
688 	union {
689 		struct {
690 			u64 rd_req:1;
691 			u64 wr_req:1;
692 			u64 lpig:1;
693 			u64 prg_index:9;
694 			u64 addr:52;
695 		};
696 		u64 qw_1;
697 	};
698 	u64 priv_data[2];
699 };
700 
701 #define PRQ_RING_MASK	((0x1000 << PRQ_ORDER) - 0x20)
702 
703 static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
704 {
705 	unsigned long requested = 0;
706 
707 	if (req->exe_req)
708 		requested |= VM_EXEC;
709 
710 	if (req->rd_req)
711 		requested |= VM_READ;
712 
713 	if (req->wr_req)
714 		requested |= VM_WRITE;
715 
716 	return (requested & ~vma->vm_flags) != 0;
717 }
718 
719 static bool is_canonical_address(u64 addr)
720 {
721 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
722 	long saddr = (long) addr;
723 
724 	return (((saddr << shift) >> shift) == saddr);
725 }
726 
727 /**
728  * intel_svm_drain_prq - Drain page requests and responses for a pasid
729  * @dev: target device
730  * @pasid: pasid for draining
731  *
732  * Drain all pending page requests and responses related to @pasid in both
733  * software and hardware. This is supposed to be called after the device
734  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
735  * and DevTLB have been invalidated.
736  *
737  * It waits until all pending page requests for @pasid in the page fault
738  * queue are completed by the prq handling thread. Then follow the steps
739  * described in VT-d spec CH7.10 to drain all page requests and page
740  * responses pending in the hardware.
741  */
742 static void intel_svm_drain_prq(struct device *dev, int pasid)
743 {
744 	struct device_domain_info *info;
745 	struct dmar_domain *domain;
746 	struct intel_iommu *iommu;
747 	struct qi_desc desc[3];
748 	struct pci_dev *pdev;
749 	int head, tail;
750 	u16 sid, did;
751 	int qdep;
752 
753 	info = get_domain_info(dev);
754 	if (WARN_ON(!info || !dev_is_pci(dev)))
755 		return;
756 
757 	if (!info->pri_enabled)
758 		return;
759 
760 	iommu = info->iommu;
761 	domain = info->domain;
762 	pdev = to_pci_dev(dev);
763 	sid = PCI_DEVID(info->bus, info->devfn);
764 	did = domain->iommu_did[iommu->seq_id];
765 	qdep = pci_ats_queue_depth(pdev);
766 
767 	/*
768 	 * Check and wait until all pending page requests in the queue are
769 	 * handled by the prq handling thread.
770 	 */
771 prq_retry:
772 	reinit_completion(&iommu->prq_complete);
773 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
774 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
775 	while (head != tail) {
776 		struct page_req_dsc *req;
777 
778 		req = &iommu->prq[head / sizeof(*req)];
779 		if (!req->pasid_present || req->pasid != pasid) {
780 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
781 			continue;
782 		}
783 
784 		wait_for_completion(&iommu->prq_complete);
785 		goto prq_retry;
786 	}
787 
788 	/*
789 	 * Perform steps described in VT-d spec CH7.10 to drain page
790 	 * requests and responses in hardware.
791 	 */
792 	memset(desc, 0, sizeof(desc));
793 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
794 			QI_IWD_FENCE |
795 			QI_IWD_TYPE;
796 	desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
797 			QI_EIOTLB_DID(did) |
798 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
799 			QI_EIOTLB_TYPE;
800 	desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
801 			QI_DEV_EIOTLB_SID(sid) |
802 			QI_DEV_EIOTLB_QDEP(qdep) |
803 			QI_DEIOTLB_TYPE |
804 			QI_DEV_IOTLB_PFSID(info->pfsid);
805 qi_retry:
806 	reinit_completion(&iommu->prq_complete);
807 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
808 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
809 		wait_for_completion(&iommu->prq_complete);
810 		goto qi_retry;
811 	}
812 }
813 
814 static int prq_to_iommu_prot(struct page_req_dsc *req)
815 {
816 	int prot = 0;
817 
818 	if (req->rd_req)
819 		prot |= IOMMU_FAULT_PERM_READ;
820 	if (req->wr_req)
821 		prot |= IOMMU_FAULT_PERM_WRITE;
822 	if (req->exe_req)
823 		prot |= IOMMU_FAULT_PERM_EXEC;
824 	if (req->pm_req)
825 		prot |= IOMMU_FAULT_PERM_PRIV;
826 
827 	return prot;
828 }
829 
830 static int
831 intel_svm_prq_report(struct device *dev, struct page_req_dsc *desc)
832 {
833 	struct iommu_fault_event event;
834 
835 	if (!dev || !dev_is_pci(dev))
836 		return -ENODEV;
837 
838 	/* Fill in event data for device specific processing */
839 	memset(&event, 0, sizeof(struct iommu_fault_event));
840 	event.fault.type = IOMMU_FAULT_PAGE_REQ;
841 	event.fault.prm.addr = desc->addr;
842 	event.fault.prm.pasid = desc->pasid;
843 	event.fault.prm.grpid = desc->prg_index;
844 	event.fault.prm.perm = prq_to_iommu_prot(desc);
845 
846 	if (desc->lpig)
847 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
848 	if (desc->pasid_present) {
849 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
850 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
851 	}
852 	if (desc->priv_data_present) {
853 		/*
854 		 * Set last page in group bit if private data is present,
855 		 * page response is required as it does for LPIG.
856 		 * iommu_report_device_fault() doesn't understand this vendor
857 		 * specific requirement thus we set last_page as a workaround.
858 		 */
859 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
860 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
861 		memcpy(event.fault.prm.private_data, desc->priv_data,
862 		       sizeof(desc->priv_data));
863 	}
864 
865 	return iommu_report_device_fault(dev, &event);
866 }
867 
868 static irqreturn_t prq_event_thread(int irq, void *d)
869 {
870 	struct intel_svm_dev *sdev = NULL;
871 	struct intel_iommu *iommu = d;
872 	struct intel_svm *svm = NULL;
873 	int head, tail, handled = 0;
874 
875 	/* Clear PPR bit before reading head/tail registers, to
876 	 * ensure that we get a new interrupt if needed. */
877 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
878 
879 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
880 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
881 	while (head != tail) {
882 		struct vm_area_struct *vma;
883 		struct page_req_dsc *req;
884 		struct qi_desc resp;
885 		int result;
886 		vm_fault_t ret;
887 		u64 address;
888 
889 		handled = 1;
890 
891 		req = &iommu->prq[head / sizeof(*req)];
892 
893 		result = QI_RESP_FAILURE;
894 		address = (u64)req->addr << VTD_PAGE_SHIFT;
895 		if (!req->pasid_present) {
896 			pr_err("%s: Page request without PASID: %08llx %08llx\n",
897 			       iommu->name, ((unsigned long long *)req)[0],
898 			       ((unsigned long long *)req)[1]);
899 			goto no_pasid;
900 		}
901 
902 		if (!svm || svm->pasid != req->pasid) {
903 			rcu_read_lock();
904 			svm = ioasid_find(NULL, req->pasid, NULL);
905 			/* It *can't* go away, because the driver is not permitted
906 			 * to unbind the mm while any page faults are outstanding.
907 			 * So we only need RCU to protect the internal idr code. */
908 			rcu_read_unlock();
909 			if (IS_ERR_OR_NULL(svm)) {
910 				pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
911 				       iommu->name, req->pasid, ((unsigned long long *)req)[0],
912 				       ((unsigned long long *)req)[1]);
913 				goto no_pasid;
914 			}
915 		}
916 
917 		if (!sdev || sdev->sid != req->rid) {
918 			struct intel_svm_dev *t;
919 
920 			sdev = NULL;
921 			rcu_read_lock();
922 			list_for_each_entry_rcu(t, &svm->devs, list) {
923 				if (t->sid == req->rid) {
924 					sdev = t;
925 					break;
926 				}
927 			}
928 			rcu_read_unlock();
929 		}
930 
931 		result = QI_RESP_INVALID;
932 		/* Since we're using init_mm.pgd directly, we should never take
933 		 * any faults on kernel addresses. */
934 		if (!svm->mm)
935 			goto bad_req;
936 
937 		/* If address is not canonical, return invalid response */
938 		if (!is_canonical_address(address))
939 			goto bad_req;
940 
941 		/*
942 		 * If prq is to be handled outside iommu driver via receiver of
943 		 * the fault notifiers, we skip the page response here.
944 		 */
945 		if (svm->flags & SVM_FLAG_GUEST_MODE) {
946 			if (sdev && !intel_svm_prq_report(sdev->dev, req))
947 				goto prq_advance;
948 			else
949 				goto bad_req;
950 		}
951 
952 		/* If the mm is already defunct, don't handle faults. */
953 		if (!mmget_not_zero(svm->mm))
954 			goto bad_req;
955 
956 		mmap_read_lock(svm->mm);
957 		vma = find_extend_vma(svm->mm, address);
958 		if (!vma || address < vma->vm_start)
959 			goto invalid;
960 
961 		if (access_error(vma, req))
962 			goto invalid;
963 
964 		ret = handle_mm_fault(vma, address,
965 				      req->wr_req ? FAULT_FLAG_WRITE : 0,
966 				      NULL);
967 		if (ret & VM_FAULT_ERROR)
968 			goto invalid;
969 
970 		result = QI_RESP_SUCCESS;
971 invalid:
972 		mmap_read_unlock(svm->mm);
973 		mmput(svm->mm);
974 bad_req:
975 		WARN_ON(!sdev);
976 		if (sdev && sdev->ops && sdev->ops->fault_cb) {
977 			int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
978 				(req->exe_req << 1) | (req->pm_req);
979 			sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr,
980 					    req->priv_data, rwxp, result);
981 		}
982 		/* We get here in the error case where the PASID lookup failed,
983 		   and these can be NULL. Do not use them below this point! */
984 		sdev = NULL;
985 		svm = NULL;
986 no_pasid:
987 		if (req->lpig || req->priv_data_present) {
988 			/*
989 			 * Per VT-d spec. v3.0 ch7.7, system software must
990 			 * respond with page group response if private data
991 			 * is present (PDP) or last page in group (LPIG) bit
992 			 * is set. This is an additional VT-d feature beyond
993 			 * PCI ATS spec.
994 			 */
995 			resp.qw0 = QI_PGRP_PASID(req->pasid) |
996 				QI_PGRP_DID(req->rid) |
997 				QI_PGRP_PASID_P(req->pasid_present) |
998 				QI_PGRP_PDP(req->pasid_present) |
999 				QI_PGRP_RESP_CODE(result) |
1000 				QI_PGRP_RESP_TYPE;
1001 			resp.qw1 = QI_PGRP_IDX(req->prg_index) |
1002 				QI_PGRP_LPIG(req->lpig);
1003 
1004 			if (req->priv_data_present)
1005 				memcpy(&resp.qw2, req->priv_data,
1006 				       sizeof(req->priv_data));
1007 			resp.qw2 = 0;
1008 			resp.qw3 = 0;
1009 			qi_submit_sync(iommu, &resp, 1, 0);
1010 		}
1011 prq_advance:
1012 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
1013 	}
1014 
1015 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
1016 
1017 	/*
1018 	 * Clear the page request overflow bit and wake up all threads that
1019 	 * are waiting for the completion of this handling.
1020 	 */
1021 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO)
1022 		writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
1023 
1024 	if (!completion_done(&iommu->prq_complete))
1025 		complete(&iommu->prq_complete);
1026 
1027 	return IRQ_RETVAL(handled);
1028 }
1029 
1030 #define to_intel_svm_dev(handle) container_of(handle, struct intel_svm_dev, sva)
1031 struct iommu_sva *
1032 intel_svm_bind(struct device *dev, struct mm_struct *mm, void *drvdata)
1033 {
1034 	struct iommu_sva *sva = ERR_PTR(-EINVAL);
1035 	struct intel_svm_dev *sdev = NULL;
1036 	int flags = 0;
1037 	int ret;
1038 
1039 	/*
1040 	 * TODO: Consolidate with generic iommu-sva bind after it is merged.
1041 	 * It will require shared SVM data structures, i.e. combine io_mm
1042 	 * and intel_svm etc.
1043 	 */
1044 	if (drvdata)
1045 		flags = *(int *)drvdata;
1046 	mutex_lock(&pasid_mutex);
1047 	ret = intel_svm_bind_mm(dev, flags, NULL, mm, &sdev);
1048 	if (ret)
1049 		sva = ERR_PTR(ret);
1050 	else if (sdev)
1051 		sva = &sdev->sva;
1052 	else
1053 		WARN(!sdev, "SVM bind succeeded with no sdev!\n");
1054 
1055 	mutex_unlock(&pasid_mutex);
1056 
1057 	return sva;
1058 }
1059 
1060 void intel_svm_unbind(struct iommu_sva *sva)
1061 {
1062 	struct intel_svm_dev *sdev;
1063 
1064 	mutex_lock(&pasid_mutex);
1065 	sdev = to_intel_svm_dev(sva);
1066 	intel_svm_unbind_mm(sdev->dev, sdev->pasid);
1067 	mutex_unlock(&pasid_mutex);
1068 }
1069 
1070 int intel_svm_get_pasid(struct iommu_sva *sva)
1071 {
1072 	struct intel_svm_dev *sdev;
1073 	int pasid;
1074 
1075 	mutex_lock(&pasid_mutex);
1076 	sdev = to_intel_svm_dev(sva);
1077 	pasid = sdev->pasid;
1078 	mutex_unlock(&pasid_mutex);
1079 
1080 	return pasid;
1081 }
1082 
1083 int intel_svm_page_response(struct device *dev,
1084 			    struct iommu_fault_event *evt,
1085 			    struct iommu_page_response *msg)
1086 {
1087 	struct iommu_fault_page_request *prm;
1088 	struct intel_svm_dev *sdev = NULL;
1089 	struct intel_svm *svm = NULL;
1090 	struct intel_iommu *iommu;
1091 	bool private_present;
1092 	bool pasid_present;
1093 	bool last_page;
1094 	u8 bus, devfn;
1095 	int ret = 0;
1096 	u16 sid;
1097 
1098 	if (!dev || !dev_is_pci(dev))
1099 		return -ENODEV;
1100 
1101 	iommu = device_to_iommu(dev, &bus, &devfn);
1102 	if (!iommu)
1103 		return -ENODEV;
1104 
1105 	if (!msg || !evt)
1106 		return -EINVAL;
1107 
1108 	mutex_lock(&pasid_mutex);
1109 
1110 	prm = &evt->fault.prm;
1111 	sid = PCI_DEVID(bus, devfn);
1112 	pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
1113 	private_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
1114 	last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
1115 
1116 	if (!pasid_present) {
1117 		ret = -EINVAL;
1118 		goto out;
1119 	}
1120 
1121 	if (prm->pasid == 0 || prm->pasid >= PASID_MAX) {
1122 		ret = -EINVAL;
1123 		goto out;
1124 	}
1125 
1126 	ret = pasid_to_svm_sdev(dev, prm->pasid, &svm, &sdev);
1127 	if (ret || !sdev) {
1128 		ret = -ENODEV;
1129 		goto out;
1130 	}
1131 
1132 	/*
1133 	 * For responses from userspace, need to make sure that the
1134 	 * pasid has been bound to its mm.
1135 	 */
1136 	if (svm->flags & SVM_FLAG_GUEST_MODE) {
1137 		struct mm_struct *mm;
1138 
1139 		mm = get_task_mm(current);
1140 		if (!mm) {
1141 			ret = -EINVAL;
1142 			goto out;
1143 		}
1144 
1145 		if (mm != svm->mm) {
1146 			ret = -ENODEV;
1147 			mmput(mm);
1148 			goto out;
1149 		}
1150 
1151 		mmput(mm);
1152 	}
1153 
1154 	/*
1155 	 * Per VT-d spec. v3.0 ch7.7, system software must respond
1156 	 * with page group response if private data is present (PDP)
1157 	 * or last page in group (LPIG) bit is set. This is an
1158 	 * additional VT-d requirement beyond PCI ATS spec.
1159 	 */
1160 	if (last_page || private_present) {
1161 		struct qi_desc desc;
1162 
1163 		desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
1164 				QI_PGRP_PASID_P(pasid_present) |
1165 				QI_PGRP_PDP(private_present) |
1166 				QI_PGRP_RESP_CODE(msg->code) |
1167 				QI_PGRP_RESP_TYPE;
1168 		desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
1169 		desc.qw2 = 0;
1170 		desc.qw3 = 0;
1171 		if (private_present)
1172 			memcpy(&desc.qw2, prm->private_data,
1173 			       sizeof(prm->private_data));
1174 
1175 		qi_submit_sync(iommu, &desc, 1, 0);
1176 	}
1177 out:
1178 	mutex_unlock(&pasid_mutex);
1179 	return ret;
1180 }
1181