xref: /openbmc/linux/drivers/iommu/intel/svm.c (revision 1d9f03c0a15fa01aa14fb295cbc1236403fceb0b)
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(sdev->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(sdev->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(sdev->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 device_domain_info *info;
283 	struct intel_svm *svm = NULL;
284 	unsigned long iflags;
285 	int ret = 0;
286 
287 	if (WARN_ON(!iommu) || !data)
288 		return -EINVAL;
289 
290 	if (data->format != IOMMU_PASID_FORMAT_INTEL_VTD)
291 		return -EINVAL;
292 
293 	/* IOMMU core ensures argsz is more than the start of the union */
294 	if (data->argsz < offsetofend(struct iommu_gpasid_bind_data, vendor.vtd))
295 		return -EINVAL;
296 
297 	/* Make sure no undefined flags are used in vendor data */
298 	if (data->vendor.vtd.flags & ~(IOMMU_SVA_VTD_GPASID_LAST - 1))
299 		return -EINVAL;
300 
301 	if (!dev_is_pci(dev))
302 		return -ENOTSUPP;
303 
304 	/* VT-d supports devices with full 20 bit PASIDs only */
305 	if (pci_max_pasids(to_pci_dev(dev)) != PASID_MAX)
306 		return -EINVAL;
307 
308 	/*
309 	 * We only check host PASID range, we have no knowledge to check
310 	 * guest PASID range.
311 	 */
312 	if (data->hpasid <= 0 || data->hpasid >= PASID_MAX)
313 		return -EINVAL;
314 
315 	info = get_domain_info(dev);
316 	if (!info)
317 		return -EINVAL;
318 
319 	dmar_domain = to_dmar_domain(domain);
320 
321 	mutex_lock(&pasid_mutex);
322 	ret = pasid_to_svm_sdev(dev, data->hpasid, &svm, &sdev);
323 	if (ret)
324 		goto out;
325 
326 	if (sdev) {
327 		/*
328 		 * Do not allow multiple bindings of the same device-PASID since
329 		 * there is only one SL page tables per PASID. We may revisit
330 		 * once sharing PGD across domains are supported.
331 		 */
332 		dev_warn_ratelimited(dev, "Already bound with PASID %u\n",
333 				     svm->pasid);
334 		ret = -EBUSY;
335 		goto out;
336 	}
337 
338 	if (!svm) {
339 		/* We come here when PASID has never been bond to a device. */
340 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
341 		if (!svm) {
342 			ret = -ENOMEM;
343 			goto out;
344 		}
345 		/* REVISIT: upper layer/VFIO can track host process that bind
346 		 * the PASID. ioasid_set = mm might be sufficient for vfio to
347 		 * check pasid VMM ownership. We can drop the following line
348 		 * once VFIO and IOASID set check is in place.
349 		 */
350 		svm->mm = get_task_mm(current);
351 		svm->pasid = data->hpasid;
352 		if (data->flags & IOMMU_SVA_GPASID_VAL) {
353 			svm->gpasid = data->gpasid;
354 			svm->flags |= SVM_FLAG_GUEST_PASID;
355 		}
356 		ioasid_set_data(data->hpasid, svm);
357 		INIT_LIST_HEAD_RCU(&svm->devs);
358 		mmput(svm->mm);
359 	}
360 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
361 	if (!sdev) {
362 		ret = -ENOMEM;
363 		goto out;
364 	}
365 	sdev->dev = dev;
366 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
367 	sdev->iommu = iommu;
368 
369 	/* Only count users if device has aux domains */
370 	if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
371 		sdev->users = 1;
372 
373 	/* Set up device context entry for PASID if not enabled already */
374 	ret = intel_iommu_enable_pasid(iommu, sdev->dev);
375 	if (ret) {
376 		dev_err_ratelimited(dev, "Failed to enable PASID capability\n");
377 		kfree(sdev);
378 		goto out;
379 	}
380 
381 	/*
382 	 * PASID table is per device for better security. Therefore, for
383 	 * each bind of a new device even with an existing PASID, we need to
384 	 * call the nested mode setup function here.
385 	 */
386 	spin_lock_irqsave(&iommu->lock, iflags);
387 	ret = intel_pasid_setup_nested(iommu, dev,
388 				       (pgd_t *)(uintptr_t)data->gpgd,
389 				       data->hpasid, &data->vendor.vtd, dmar_domain,
390 				       data->addr_width);
391 	spin_unlock_irqrestore(&iommu->lock, iflags);
392 	if (ret) {
393 		dev_err_ratelimited(dev, "Failed to set up PASID %llu in nested mode, Err %d\n",
394 				    data->hpasid, ret);
395 		/*
396 		 * PASID entry should be in cleared state if nested mode
397 		 * set up failed. So we only need to clear IOASID tracking
398 		 * data such that free call will succeed.
399 		 */
400 		kfree(sdev);
401 		goto out;
402 	}
403 
404 	svm->flags |= SVM_FLAG_GUEST_MODE;
405 
406 	init_rcu_head(&sdev->rcu);
407 	list_add_rcu(&sdev->list, &svm->devs);
408  out:
409 	if (!IS_ERR_OR_NULL(svm) && list_empty(&svm->devs)) {
410 		ioasid_set_data(data->hpasid, NULL);
411 		kfree(svm);
412 	}
413 
414 	mutex_unlock(&pasid_mutex);
415 	return ret;
416 }
417 
418 int intel_svm_unbind_gpasid(struct device *dev, u32 pasid)
419 {
420 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
421 	struct intel_svm_dev *sdev;
422 	struct intel_svm *svm;
423 	int ret;
424 
425 	if (WARN_ON(!iommu))
426 		return -EINVAL;
427 
428 	mutex_lock(&pasid_mutex);
429 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
430 	if (ret)
431 		goto out;
432 
433 	if (sdev) {
434 		if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
435 			sdev->users--;
436 		if (!sdev->users) {
437 			list_del_rcu(&sdev->list);
438 			intel_pasid_tear_down_entry(iommu, dev,
439 						    svm->pasid, false);
440 			intel_svm_drain_prq(dev, svm->pasid);
441 			kfree_rcu(sdev, rcu);
442 
443 			if (list_empty(&svm->devs)) {
444 				/*
445 				 * We do not free the IOASID here in that
446 				 * IOMMU driver did not allocate it.
447 				 * Unlike native SVM, IOASID for guest use was
448 				 * allocated prior to the bind call.
449 				 * In any case, if the free call comes before
450 				 * the unbind, IOMMU driver will get notified
451 				 * and perform cleanup.
452 				 */
453 				ioasid_set_data(pasid, NULL);
454 				kfree(svm);
455 			}
456 		}
457 	}
458 out:
459 	mutex_unlock(&pasid_mutex);
460 	return ret;
461 }
462 
463 static void _load_pasid(void *unused)
464 {
465 	update_pasid();
466 }
467 
468 static void load_pasid(struct mm_struct *mm, u32 pasid)
469 {
470 	mutex_lock(&mm->context.lock);
471 
472 	/* Synchronize with READ_ONCE in update_pasid(). */
473 	smp_store_release(&mm->pasid, pasid);
474 
475 	/* Update PASID MSR on all CPUs running the mm's tasks. */
476 	on_each_cpu_mask(mm_cpumask(mm), _load_pasid, NULL, true);
477 
478 	mutex_unlock(&mm->context.lock);
479 }
480 
481 /* Caller must hold pasid_mutex, mm reference */
482 static int
483 intel_svm_bind_mm(struct device *dev, unsigned int flags,
484 		  struct svm_dev_ops *ops,
485 		  struct mm_struct *mm, struct intel_svm_dev **sd)
486 {
487 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
488 	struct device_domain_info *info;
489 	struct intel_svm_dev *sdev;
490 	struct intel_svm *svm = NULL;
491 	unsigned long iflags;
492 	int pasid_max;
493 	int ret;
494 
495 	if (!iommu || dmar_disabled)
496 		return -EINVAL;
497 
498 	if (!intel_svm_capable(iommu))
499 		return -ENOTSUPP;
500 
501 	if (dev_is_pci(dev)) {
502 		pasid_max = pci_max_pasids(to_pci_dev(dev));
503 		if (pasid_max < 0)
504 			return -EINVAL;
505 	} else
506 		pasid_max = 1 << 20;
507 
508 	/* Bind supervisor PASID shuld have mm = NULL */
509 	if (flags & SVM_FLAG_SUPERVISOR_MODE) {
510 		if (!ecap_srs(iommu->ecap) || mm) {
511 			pr_err("Supervisor PASID with user provided mm.\n");
512 			return -EINVAL;
513 		}
514 	}
515 
516 	if (!(flags & SVM_FLAG_PRIVATE_PASID)) {
517 		struct intel_svm *t;
518 
519 		list_for_each_entry(t, &global_svm_list, list) {
520 			if (t->mm != mm || (t->flags & SVM_FLAG_PRIVATE_PASID))
521 				continue;
522 
523 			svm = t;
524 			if (svm->pasid >= pasid_max) {
525 				dev_warn(dev,
526 					 "Limited PASID width. Cannot use existing PASID %d\n",
527 					 svm->pasid);
528 				ret = -ENOSPC;
529 				goto out;
530 			}
531 
532 			/* Find the matching device in svm list */
533 			for_each_svm_dev(sdev, svm, dev) {
534 				if (sdev->ops != ops) {
535 					ret = -EBUSY;
536 					goto out;
537 				}
538 				sdev->users++;
539 				goto success;
540 			}
541 
542 			break;
543 		}
544 	}
545 
546 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
547 	if (!sdev) {
548 		ret = -ENOMEM;
549 		goto out;
550 	}
551 	sdev->dev = dev;
552 	sdev->iommu = iommu;
553 
554 	ret = intel_iommu_enable_pasid(iommu, dev);
555 	if (ret) {
556 		kfree(sdev);
557 		goto out;
558 	}
559 
560 	info = get_domain_info(dev);
561 	sdev->did = FLPT_DEFAULT_DID;
562 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
563 	if (info->ats_enabled) {
564 		sdev->dev_iotlb = 1;
565 		sdev->qdep = info->ats_qdep;
566 		if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
567 			sdev->qdep = 0;
568 	}
569 
570 	/* Finish the setup now we know we're keeping it */
571 	sdev->users = 1;
572 	sdev->ops = ops;
573 	init_rcu_head(&sdev->rcu);
574 
575 	if (!svm) {
576 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
577 		if (!svm) {
578 			ret = -ENOMEM;
579 			kfree(sdev);
580 			goto out;
581 		}
582 
583 		if (pasid_max > intel_pasid_max_id)
584 			pasid_max = intel_pasid_max_id;
585 
586 		/* Do not use PASID 0, reserved for RID to PASID */
587 		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
588 					  pasid_max - 1, svm);
589 		if (svm->pasid == INVALID_IOASID) {
590 			kfree(svm);
591 			kfree(sdev);
592 			ret = -ENOSPC;
593 			goto out;
594 		}
595 		svm->notifier.ops = &intel_mmuops;
596 		svm->mm = mm;
597 		svm->flags = flags;
598 		INIT_LIST_HEAD_RCU(&svm->devs);
599 		INIT_LIST_HEAD(&svm->list);
600 		ret = -ENOMEM;
601 		if (mm) {
602 			ret = mmu_notifier_register(&svm->notifier, mm);
603 			if (ret) {
604 				ioasid_put(svm->pasid);
605 				kfree(svm);
606 				kfree(sdev);
607 				goto out;
608 			}
609 		}
610 
611 		spin_lock_irqsave(&iommu->lock, iflags);
612 		ret = intel_pasid_setup_first_level(iommu, dev,
613 				mm ? mm->pgd : init_mm.pgd,
614 				svm->pasid, FLPT_DEFAULT_DID,
615 				(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
616 				(cpu_feature_enabled(X86_FEATURE_LA57) ?
617 				 PASID_FLAG_FL5LP : 0));
618 		spin_unlock_irqrestore(&iommu->lock, iflags);
619 		if (ret) {
620 			if (mm)
621 				mmu_notifier_unregister(&svm->notifier, mm);
622 			ioasid_put(svm->pasid);
623 			kfree(svm);
624 			kfree(sdev);
625 			goto out;
626 		}
627 
628 		list_add_tail(&svm->list, &global_svm_list);
629 		if (mm) {
630 			/* The newly allocated pasid is loaded to the mm. */
631 			load_pasid(mm, svm->pasid);
632 		}
633 	} else {
634 		/*
635 		 * Binding a new device with existing PASID, need to setup
636 		 * the PASID entry.
637 		 */
638 		spin_lock_irqsave(&iommu->lock, iflags);
639 		ret = intel_pasid_setup_first_level(iommu, dev,
640 						mm ? mm->pgd : init_mm.pgd,
641 						svm->pasid, FLPT_DEFAULT_DID,
642 						(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
643 						(cpu_feature_enabled(X86_FEATURE_LA57) ?
644 						PASID_FLAG_FL5LP : 0));
645 		spin_unlock_irqrestore(&iommu->lock, iflags);
646 		if (ret) {
647 			kfree(sdev);
648 			goto out;
649 		}
650 	}
651 	list_add_rcu(&sdev->list, &svm->devs);
652 success:
653 	sdev->pasid = svm->pasid;
654 	sdev->sva.dev = dev;
655 	if (sd)
656 		*sd = sdev;
657 	ret = 0;
658 out:
659 	return ret;
660 }
661 
662 /* Caller must hold pasid_mutex */
663 static int intel_svm_unbind_mm(struct device *dev, u32 pasid)
664 {
665 	struct intel_svm_dev *sdev;
666 	struct intel_iommu *iommu;
667 	struct intel_svm *svm;
668 	int ret = -EINVAL;
669 
670 	iommu = device_to_iommu(dev, NULL, NULL);
671 	if (!iommu)
672 		goto out;
673 
674 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
675 	if (ret)
676 		goto out;
677 
678 	if (sdev) {
679 		sdev->users--;
680 		if (!sdev->users) {
681 			list_del_rcu(&sdev->list);
682 			/* Flush the PASID cache and IOTLB for this device.
683 			 * Note that we do depend on the hardware *not* using
684 			 * the PASID any more. Just as we depend on other
685 			 * devices never using PASIDs that they have no right
686 			 * to use. We have a *shared* PASID table, because it's
687 			 * large and has to be physically contiguous. So it's
688 			 * hard to be as defensive as we might like. */
689 			intel_pasid_tear_down_entry(iommu, dev,
690 						    svm->pasid, false);
691 			intel_svm_drain_prq(dev, svm->pasid);
692 			kfree_rcu(sdev, rcu);
693 
694 			if (list_empty(&svm->devs)) {
695 				ioasid_put(svm->pasid);
696 				if (svm->mm) {
697 					mmu_notifier_unregister(&svm->notifier, svm->mm);
698 					/* Clear mm's pasid. */
699 					load_pasid(svm->mm, PASID_DISABLED);
700 				}
701 				list_del(&svm->list);
702 				/* We mandate that no page faults may be outstanding
703 				 * for the PASID when intel_svm_unbind_mm() is called.
704 				 * If that is not obeyed, subtle errors will happen.
705 				 * Let's make them less subtle... */
706 				memset(svm, 0x6b, sizeof(*svm));
707 				kfree(svm);
708 			}
709 		}
710 	}
711 out:
712 	return ret;
713 }
714 
715 /* Page request queue descriptor */
716 struct page_req_dsc {
717 	union {
718 		struct {
719 			u64 type:8;
720 			u64 pasid_present:1;
721 			u64 priv_data_present:1;
722 			u64 rsvd:6;
723 			u64 rid:16;
724 			u64 pasid:20;
725 			u64 exe_req:1;
726 			u64 pm_req:1;
727 			u64 rsvd2:10;
728 		};
729 		u64 qw_0;
730 	};
731 	union {
732 		struct {
733 			u64 rd_req:1;
734 			u64 wr_req:1;
735 			u64 lpig:1;
736 			u64 prg_index:9;
737 			u64 addr:52;
738 		};
739 		u64 qw_1;
740 	};
741 	u64 priv_data[2];
742 };
743 
744 #define PRQ_RING_MASK	((0x1000 << PRQ_ORDER) - 0x20)
745 
746 static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
747 {
748 	unsigned long requested = 0;
749 
750 	if (req->exe_req)
751 		requested |= VM_EXEC;
752 
753 	if (req->rd_req)
754 		requested |= VM_READ;
755 
756 	if (req->wr_req)
757 		requested |= VM_WRITE;
758 
759 	return (requested & ~vma->vm_flags) != 0;
760 }
761 
762 static bool is_canonical_address(u64 addr)
763 {
764 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
765 	long saddr = (long) addr;
766 
767 	return (((saddr << shift) >> shift) == saddr);
768 }
769 
770 /**
771  * intel_svm_drain_prq - Drain page requests and responses for a pasid
772  * @dev: target device
773  * @pasid: pasid for draining
774  *
775  * Drain all pending page requests and responses related to @pasid in both
776  * software and hardware. This is supposed to be called after the device
777  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
778  * and DevTLB have been invalidated.
779  *
780  * It waits until all pending page requests for @pasid in the page fault
781  * queue are completed by the prq handling thread. Then follow the steps
782  * described in VT-d spec CH7.10 to drain all page requests and page
783  * responses pending in the hardware.
784  */
785 static void intel_svm_drain_prq(struct device *dev, u32 pasid)
786 {
787 	struct device_domain_info *info;
788 	struct dmar_domain *domain;
789 	struct intel_iommu *iommu;
790 	struct qi_desc desc[3];
791 	struct pci_dev *pdev;
792 	int head, tail;
793 	u16 sid, did;
794 	int qdep;
795 
796 	info = get_domain_info(dev);
797 	if (WARN_ON(!info || !dev_is_pci(dev)))
798 		return;
799 
800 	if (!info->pri_enabled)
801 		return;
802 
803 	iommu = info->iommu;
804 	domain = info->domain;
805 	pdev = to_pci_dev(dev);
806 	sid = PCI_DEVID(info->bus, info->devfn);
807 	did = domain->iommu_did[iommu->seq_id];
808 	qdep = pci_ats_queue_depth(pdev);
809 
810 	/*
811 	 * Check and wait until all pending page requests in the queue are
812 	 * handled by the prq handling thread.
813 	 */
814 prq_retry:
815 	reinit_completion(&iommu->prq_complete);
816 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
817 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
818 	while (head != tail) {
819 		struct page_req_dsc *req;
820 
821 		req = &iommu->prq[head / sizeof(*req)];
822 		if (!req->pasid_present || req->pasid != pasid) {
823 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
824 			continue;
825 		}
826 
827 		wait_for_completion(&iommu->prq_complete);
828 		goto prq_retry;
829 	}
830 
831 	/*
832 	 * Perform steps described in VT-d spec CH7.10 to drain page
833 	 * requests and responses in hardware.
834 	 */
835 	memset(desc, 0, sizeof(desc));
836 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
837 			QI_IWD_FENCE |
838 			QI_IWD_TYPE;
839 	desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
840 			QI_EIOTLB_DID(did) |
841 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
842 			QI_EIOTLB_TYPE;
843 	desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
844 			QI_DEV_EIOTLB_SID(sid) |
845 			QI_DEV_EIOTLB_QDEP(qdep) |
846 			QI_DEIOTLB_TYPE |
847 			QI_DEV_IOTLB_PFSID(info->pfsid);
848 qi_retry:
849 	reinit_completion(&iommu->prq_complete);
850 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
851 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
852 		wait_for_completion(&iommu->prq_complete);
853 		goto qi_retry;
854 	}
855 }
856 
857 static int prq_to_iommu_prot(struct page_req_dsc *req)
858 {
859 	int prot = 0;
860 
861 	if (req->rd_req)
862 		prot |= IOMMU_FAULT_PERM_READ;
863 	if (req->wr_req)
864 		prot |= IOMMU_FAULT_PERM_WRITE;
865 	if (req->exe_req)
866 		prot |= IOMMU_FAULT_PERM_EXEC;
867 	if (req->pm_req)
868 		prot |= IOMMU_FAULT_PERM_PRIV;
869 
870 	return prot;
871 }
872 
873 static int
874 intel_svm_prq_report(struct device *dev, struct page_req_dsc *desc)
875 {
876 	struct iommu_fault_event event;
877 
878 	if (!dev || !dev_is_pci(dev))
879 		return -ENODEV;
880 
881 	/* Fill in event data for device specific processing */
882 	memset(&event, 0, sizeof(struct iommu_fault_event));
883 	event.fault.type = IOMMU_FAULT_PAGE_REQ;
884 	event.fault.prm.addr = desc->addr;
885 	event.fault.prm.pasid = desc->pasid;
886 	event.fault.prm.grpid = desc->prg_index;
887 	event.fault.prm.perm = prq_to_iommu_prot(desc);
888 
889 	if (desc->lpig)
890 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
891 	if (desc->pasid_present) {
892 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
893 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
894 	}
895 	if (desc->priv_data_present) {
896 		/*
897 		 * Set last page in group bit if private data is present,
898 		 * page response is required as it does for LPIG.
899 		 * iommu_report_device_fault() doesn't understand this vendor
900 		 * specific requirement thus we set last_page as a workaround.
901 		 */
902 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
903 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
904 		memcpy(event.fault.prm.private_data, desc->priv_data,
905 		       sizeof(desc->priv_data));
906 	}
907 
908 	return iommu_report_device_fault(dev, &event);
909 }
910 
911 static irqreturn_t prq_event_thread(int irq, void *d)
912 {
913 	struct intel_svm_dev *sdev = NULL;
914 	struct intel_iommu *iommu = d;
915 	struct intel_svm *svm = NULL;
916 	int head, tail, handled = 0;
917 
918 	/* Clear PPR bit before reading head/tail registers, to
919 	 * ensure that we get a new interrupt if needed. */
920 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
921 
922 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
923 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
924 	while (head != tail) {
925 		struct vm_area_struct *vma;
926 		struct page_req_dsc *req;
927 		struct qi_desc resp;
928 		int result;
929 		vm_fault_t ret;
930 		u64 address;
931 
932 		handled = 1;
933 
934 		req = &iommu->prq[head / sizeof(*req)];
935 
936 		result = QI_RESP_FAILURE;
937 		address = (u64)req->addr << VTD_PAGE_SHIFT;
938 		if (!req->pasid_present) {
939 			pr_err("%s: Page request without PASID: %08llx %08llx\n",
940 			       iommu->name, ((unsigned long long *)req)[0],
941 			       ((unsigned long long *)req)[1]);
942 			goto no_pasid;
943 		}
944 
945 		if (!svm || svm->pasid != req->pasid) {
946 			rcu_read_lock();
947 			svm = ioasid_find(NULL, req->pasid, NULL);
948 			/* It *can't* go away, because the driver is not permitted
949 			 * to unbind the mm while any page faults are outstanding.
950 			 * So we only need RCU to protect the internal idr code. */
951 			rcu_read_unlock();
952 			if (IS_ERR_OR_NULL(svm)) {
953 				pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
954 				       iommu->name, req->pasid, ((unsigned long long *)req)[0],
955 				       ((unsigned long long *)req)[1]);
956 				goto no_pasid;
957 			}
958 		}
959 
960 		if (!sdev || sdev->sid != req->rid) {
961 			struct intel_svm_dev *t;
962 
963 			sdev = NULL;
964 			rcu_read_lock();
965 			list_for_each_entry_rcu(t, &svm->devs, list) {
966 				if (t->sid == req->rid) {
967 					sdev = t;
968 					break;
969 				}
970 			}
971 			rcu_read_unlock();
972 		}
973 
974 		result = QI_RESP_INVALID;
975 		/* Since we're using init_mm.pgd directly, we should never take
976 		 * any faults on kernel addresses. */
977 		if (!svm->mm)
978 			goto bad_req;
979 
980 		/* If address is not canonical, return invalid response */
981 		if (!is_canonical_address(address))
982 			goto bad_req;
983 
984 		/*
985 		 * If prq is to be handled outside iommu driver via receiver of
986 		 * the fault notifiers, we skip the page response here.
987 		 */
988 		if (svm->flags & SVM_FLAG_GUEST_MODE) {
989 			if (sdev && !intel_svm_prq_report(sdev->dev, req))
990 				goto prq_advance;
991 			else
992 				goto bad_req;
993 		}
994 
995 		/* If the mm is already defunct, don't handle faults. */
996 		if (!mmget_not_zero(svm->mm))
997 			goto bad_req;
998 
999 		mmap_read_lock(svm->mm);
1000 		vma = find_extend_vma(svm->mm, address);
1001 		if (!vma || address < vma->vm_start)
1002 			goto invalid;
1003 
1004 		if (access_error(vma, req))
1005 			goto invalid;
1006 
1007 		ret = handle_mm_fault(vma, address,
1008 				      req->wr_req ? FAULT_FLAG_WRITE : 0,
1009 				      NULL);
1010 		if (ret & VM_FAULT_ERROR)
1011 			goto invalid;
1012 
1013 		result = QI_RESP_SUCCESS;
1014 invalid:
1015 		mmap_read_unlock(svm->mm);
1016 		mmput(svm->mm);
1017 bad_req:
1018 		WARN_ON(!sdev);
1019 		if (sdev && sdev->ops && sdev->ops->fault_cb) {
1020 			int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
1021 				(req->exe_req << 1) | (req->pm_req);
1022 			sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr,
1023 					    req->priv_data, rwxp, result);
1024 		}
1025 		/* We get here in the error case where the PASID lookup failed,
1026 		   and these can be NULL. Do not use them below this point! */
1027 		sdev = NULL;
1028 		svm = NULL;
1029 no_pasid:
1030 		if (req->lpig || req->priv_data_present) {
1031 			/*
1032 			 * Per VT-d spec. v3.0 ch7.7, system software must
1033 			 * respond with page group response if private data
1034 			 * is present (PDP) or last page in group (LPIG) bit
1035 			 * is set. This is an additional VT-d feature beyond
1036 			 * PCI ATS spec.
1037 			 */
1038 			resp.qw0 = QI_PGRP_PASID(req->pasid) |
1039 				QI_PGRP_DID(req->rid) |
1040 				QI_PGRP_PASID_P(req->pasid_present) |
1041 				QI_PGRP_PDP(req->priv_data_present) |
1042 				QI_PGRP_RESP_CODE(result) |
1043 				QI_PGRP_RESP_TYPE;
1044 			resp.qw1 = QI_PGRP_IDX(req->prg_index) |
1045 				QI_PGRP_LPIG(req->lpig);
1046 
1047 			if (req->priv_data_present)
1048 				memcpy(&resp.qw2, req->priv_data,
1049 				       sizeof(req->priv_data));
1050 			resp.qw2 = 0;
1051 			resp.qw3 = 0;
1052 			qi_submit_sync(iommu, &resp, 1, 0);
1053 		}
1054 prq_advance:
1055 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
1056 	}
1057 
1058 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
1059 
1060 	/*
1061 	 * Clear the page request overflow bit and wake up all threads that
1062 	 * are waiting for the completion of this handling.
1063 	 */
1064 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO)
1065 		writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
1066 
1067 	if (!completion_done(&iommu->prq_complete))
1068 		complete(&iommu->prq_complete);
1069 
1070 	return IRQ_RETVAL(handled);
1071 }
1072 
1073 #define to_intel_svm_dev(handle) container_of(handle, struct intel_svm_dev, sva)
1074 struct iommu_sva *
1075 intel_svm_bind(struct device *dev, struct mm_struct *mm, void *drvdata)
1076 {
1077 	struct iommu_sva *sva = ERR_PTR(-EINVAL);
1078 	struct intel_svm_dev *sdev = NULL;
1079 	unsigned int flags = 0;
1080 	int ret;
1081 
1082 	/*
1083 	 * TODO: Consolidate with generic iommu-sva bind after it is merged.
1084 	 * It will require shared SVM data structures, i.e. combine io_mm
1085 	 * and intel_svm etc.
1086 	 */
1087 	if (drvdata)
1088 		flags = *(unsigned int *)drvdata;
1089 	mutex_lock(&pasid_mutex);
1090 	ret = intel_svm_bind_mm(dev, flags, NULL, mm, &sdev);
1091 	if (ret)
1092 		sva = ERR_PTR(ret);
1093 	else if (sdev)
1094 		sva = &sdev->sva;
1095 	else
1096 		WARN(!sdev, "SVM bind succeeded with no sdev!\n");
1097 
1098 	mutex_unlock(&pasid_mutex);
1099 
1100 	return sva;
1101 }
1102 
1103 void intel_svm_unbind(struct iommu_sva *sva)
1104 {
1105 	struct intel_svm_dev *sdev;
1106 
1107 	mutex_lock(&pasid_mutex);
1108 	sdev = to_intel_svm_dev(sva);
1109 	intel_svm_unbind_mm(sdev->dev, sdev->pasid);
1110 	mutex_unlock(&pasid_mutex);
1111 }
1112 
1113 u32 intel_svm_get_pasid(struct iommu_sva *sva)
1114 {
1115 	struct intel_svm_dev *sdev;
1116 	u32 pasid;
1117 
1118 	mutex_lock(&pasid_mutex);
1119 	sdev = to_intel_svm_dev(sva);
1120 	pasid = sdev->pasid;
1121 	mutex_unlock(&pasid_mutex);
1122 
1123 	return pasid;
1124 }
1125 
1126 int intel_svm_page_response(struct device *dev,
1127 			    struct iommu_fault_event *evt,
1128 			    struct iommu_page_response *msg)
1129 {
1130 	struct iommu_fault_page_request *prm;
1131 	struct intel_svm_dev *sdev = NULL;
1132 	struct intel_svm *svm = NULL;
1133 	struct intel_iommu *iommu;
1134 	bool private_present;
1135 	bool pasid_present;
1136 	bool last_page;
1137 	u8 bus, devfn;
1138 	int ret = 0;
1139 	u16 sid;
1140 
1141 	if (!dev || !dev_is_pci(dev))
1142 		return -ENODEV;
1143 
1144 	iommu = device_to_iommu(dev, &bus, &devfn);
1145 	if (!iommu)
1146 		return -ENODEV;
1147 
1148 	if (!msg || !evt)
1149 		return -EINVAL;
1150 
1151 	mutex_lock(&pasid_mutex);
1152 
1153 	prm = &evt->fault.prm;
1154 	sid = PCI_DEVID(bus, devfn);
1155 	pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
1156 	private_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
1157 	last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
1158 
1159 	if (!pasid_present) {
1160 		ret = -EINVAL;
1161 		goto out;
1162 	}
1163 
1164 	if (prm->pasid == 0 || prm->pasid >= PASID_MAX) {
1165 		ret = -EINVAL;
1166 		goto out;
1167 	}
1168 
1169 	ret = pasid_to_svm_sdev(dev, prm->pasid, &svm, &sdev);
1170 	if (ret || !sdev) {
1171 		ret = -ENODEV;
1172 		goto out;
1173 	}
1174 
1175 	/*
1176 	 * For responses from userspace, need to make sure that the
1177 	 * pasid has been bound to its mm.
1178 	 */
1179 	if (svm->flags & SVM_FLAG_GUEST_MODE) {
1180 		struct mm_struct *mm;
1181 
1182 		mm = get_task_mm(current);
1183 		if (!mm) {
1184 			ret = -EINVAL;
1185 			goto out;
1186 		}
1187 
1188 		if (mm != svm->mm) {
1189 			ret = -ENODEV;
1190 			mmput(mm);
1191 			goto out;
1192 		}
1193 
1194 		mmput(mm);
1195 	}
1196 
1197 	/*
1198 	 * Per VT-d spec. v3.0 ch7.7, system software must respond
1199 	 * with page group response if private data is present (PDP)
1200 	 * or last page in group (LPIG) bit is set. This is an
1201 	 * additional VT-d requirement beyond PCI ATS spec.
1202 	 */
1203 	if (last_page || private_present) {
1204 		struct qi_desc desc;
1205 
1206 		desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
1207 				QI_PGRP_PASID_P(pasid_present) |
1208 				QI_PGRP_PDP(private_present) |
1209 				QI_PGRP_RESP_CODE(msg->code) |
1210 				QI_PGRP_RESP_TYPE;
1211 		desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
1212 		desc.qw2 = 0;
1213 		desc.qw3 = 0;
1214 		if (private_present)
1215 			memcpy(&desc.qw2, prm->private_data,
1216 			       sizeof(prm->private_data));
1217 
1218 		qi_submit_sync(iommu, &desc, 1, 0);
1219 	}
1220 out:
1221 	mutex_unlock(&pasid_mutex);
1222 	return ret;
1223 }
1224