xref: /linux/drivers/iommu/intel/svm.c (revision 37cb28ec7d3a36a5bace7063a3dba633ab110f8b)
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/xarray.h>
21 #include <linux/ioasid.h>
22 #include <asm/page.h>
23 #include <asm/fpu/api.h>
24 #include <trace/events/intel_iommu.h>
25 
26 #include "pasid.h"
27 #include "perf.h"
28 #include "../iommu-sva-lib.h"
29 
30 static irqreturn_t prq_event_thread(int irq, void *d);
31 static void intel_svm_drain_prq(struct device *dev, u32 pasid);
32 #define to_intel_svm_dev(handle) container_of(handle, struct intel_svm_dev, sva)
33 
34 static DEFINE_XARRAY_ALLOC(pasid_private_array);
35 static int pasid_private_add(ioasid_t pasid, void *priv)
36 {
37 	return xa_alloc(&pasid_private_array, &pasid, priv,
38 			XA_LIMIT(pasid, pasid), GFP_ATOMIC);
39 }
40 
41 static void pasid_private_remove(ioasid_t pasid)
42 {
43 	xa_erase(&pasid_private_array, pasid);
44 }
45 
46 static void *pasid_private_find(ioasid_t pasid)
47 {
48 	return xa_load(&pasid_private_array, pasid);
49 }
50 
51 static struct intel_svm_dev *
52 svm_lookup_device_by_sid(struct intel_svm *svm, u16 sid)
53 {
54 	struct intel_svm_dev *sdev = NULL, *t;
55 
56 	rcu_read_lock();
57 	list_for_each_entry_rcu(t, &svm->devs, list) {
58 		if (t->sid == sid) {
59 			sdev = t;
60 			break;
61 		}
62 	}
63 	rcu_read_unlock();
64 
65 	return sdev;
66 }
67 
68 static struct intel_svm_dev *
69 svm_lookup_device_by_dev(struct intel_svm *svm, struct device *dev)
70 {
71 	struct intel_svm_dev *sdev = NULL, *t;
72 
73 	rcu_read_lock();
74 	list_for_each_entry_rcu(t, &svm->devs, list) {
75 		if (t->dev == dev) {
76 			sdev = t;
77 			break;
78 		}
79 	}
80 	rcu_read_unlock();
81 
82 	return sdev;
83 }
84 
85 int intel_svm_enable_prq(struct intel_iommu *iommu)
86 {
87 	struct iopf_queue *iopfq;
88 	struct page *pages;
89 	int irq, ret;
90 
91 	pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
92 	if (!pages) {
93 		pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
94 			iommu->name);
95 		return -ENOMEM;
96 	}
97 	iommu->prq = page_address(pages);
98 
99 	irq = dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED + iommu->seq_id, iommu->node, iommu);
100 	if (irq <= 0) {
101 		pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
102 		       iommu->name);
103 		ret = -EINVAL;
104 		goto free_prq;
105 	}
106 	iommu->pr_irq = irq;
107 
108 	snprintf(iommu->iopfq_name, sizeof(iommu->iopfq_name),
109 		 "dmar%d-iopfq", iommu->seq_id);
110 	iopfq = iopf_queue_alloc(iommu->iopfq_name);
111 	if (!iopfq) {
112 		pr_err("IOMMU: %s: Failed to allocate iopf queue\n", iommu->name);
113 		ret = -ENOMEM;
114 		goto free_hwirq;
115 	}
116 	iommu->iopf_queue = iopfq;
117 
118 	snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
119 
120 	ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
121 				   iommu->prq_name, iommu);
122 	if (ret) {
123 		pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
124 		       iommu->name);
125 		goto free_iopfq;
126 	}
127 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
128 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
129 	dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
130 
131 	init_completion(&iommu->prq_complete);
132 
133 	return 0;
134 
135 free_iopfq:
136 	iopf_queue_free(iommu->iopf_queue);
137 	iommu->iopf_queue = NULL;
138 free_hwirq:
139 	dmar_free_hwirq(irq);
140 	iommu->pr_irq = 0;
141 free_prq:
142 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
143 	iommu->prq = NULL;
144 
145 	return ret;
146 }
147 
148 int intel_svm_finish_prq(struct intel_iommu *iommu)
149 {
150 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
151 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
152 	dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
153 
154 	if (iommu->pr_irq) {
155 		free_irq(iommu->pr_irq, iommu);
156 		dmar_free_hwirq(iommu->pr_irq);
157 		iommu->pr_irq = 0;
158 	}
159 
160 	if (iommu->iopf_queue) {
161 		iopf_queue_free(iommu->iopf_queue);
162 		iommu->iopf_queue = NULL;
163 	}
164 
165 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
166 	iommu->prq = NULL;
167 
168 	return 0;
169 }
170 
171 static inline bool intel_svm_capable(struct intel_iommu *iommu)
172 {
173 	return iommu->flags & VTD_FLAG_SVM_CAPABLE;
174 }
175 
176 void intel_svm_check(struct intel_iommu *iommu)
177 {
178 	if (!pasid_supported(iommu))
179 		return;
180 
181 	if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
182 	    !cap_fl1gp_support(iommu->cap)) {
183 		pr_err("%s SVM disabled, incompatible 1GB page capability\n",
184 		       iommu->name);
185 		return;
186 	}
187 
188 	if (cpu_feature_enabled(X86_FEATURE_LA57) &&
189 	    !cap_5lp_support(iommu->cap)) {
190 		pr_err("%s SVM disabled, incompatible paging mode\n",
191 		       iommu->name);
192 		return;
193 	}
194 
195 	iommu->flags |= VTD_FLAG_SVM_CAPABLE;
196 }
197 
198 static void __flush_svm_range_dev(struct intel_svm *svm,
199 				  struct intel_svm_dev *sdev,
200 				  unsigned long address,
201 				  unsigned long pages, int ih)
202 {
203 	struct device_domain_info *info = get_domain_info(sdev->dev);
204 
205 	if (WARN_ON(!pages))
206 		return;
207 
208 	qi_flush_piotlb(sdev->iommu, sdev->did, svm->pasid, address, pages, ih);
209 	if (info->ats_enabled)
210 		qi_flush_dev_iotlb_pasid(sdev->iommu, sdev->sid, info->pfsid,
211 					 svm->pasid, sdev->qdep, address,
212 					 order_base_2(pages));
213 }
214 
215 static void intel_flush_svm_range_dev(struct intel_svm *svm,
216 				      struct intel_svm_dev *sdev,
217 				      unsigned long address,
218 				      unsigned long pages, int ih)
219 {
220 	unsigned long shift = ilog2(__roundup_pow_of_two(pages));
221 	unsigned long align = (1ULL << (VTD_PAGE_SHIFT + shift));
222 	unsigned long start = ALIGN_DOWN(address, align);
223 	unsigned long end = ALIGN(address + (pages << VTD_PAGE_SHIFT), align);
224 
225 	while (start < end) {
226 		__flush_svm_range_dev(svm, sdev, start, align >> VTD_PAGE_SHIFT, ih);
227 		start += align;
228 	}
229 }
230 
231 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
232 				unsigned long pages, int ih)
233 {
234 	struct intel_svm_dev *sdev;
235 
236 	rcu_read_lock();
237 	list_for_each_entry_rcu(sdev, &svm->devs, list)
238 		intel_flush_svm_range_dev(svm, sdev, address, pages, ih);
239 	rcu_read_unlock();
240 }
241 
242 /* Pages have been freed at this point */
243 static void intel_invalidate_range(struct mmu_notifier *mn,
244 				   struct mm_struct *mm,
245 				   unsigned long start, unsigned long end)
246 {
247 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
248 
249 	intel_flush_svm_range(svm, start,
250 			      (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0);
251 }
252 
253 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
254 {
255 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
256 	struct intel_svm_dev *sdev;
257 
258 	/* This might end up being called from exit_mmap(), *before* the page
259 	 * tables are cleared. And __mmu_notifier_release() will delete us from
260 	 * the list of notifiers so that our invalidate_range() callback doesn't
261 	 * get called when the page tables are cleared. So we need to protect
262 	 * against hardware accessing those page tables.
263 	 *
264 	 * We do it by clearing the entry in the PASID table and then flushing
265 	 * the IOTLB and the PASID table caches. This might upset hardware;
266 	 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
267 	 * page) so that we end up taking a fault that the hardware really
268 	 * *has* to handle gracefully without affecting other processes.
269 	 */
270 	rcu_read_lock();
271 	list_for_each_entry_rcu(sdev, &svm->devs, list)
272 		intel_pasid_tear_down_entry(sdev->iommu, sdev->dev,
273 					    svm->pasid, true);
274 	rcu_read_unlock();
275 
276 }
277 
278 static const struct mmu_notifier_ops intel_mmuops = {
279 	.release = intel_mm_release,
280 	.invalidate_range = intel_invalidate_range,
281 };
282 
283 static DEFINE_MUTEX(pasid_mutex);
284 
285 static int pasid_to_svm_sdev(struct device *dev, unsigned int pasid,
286 			     struct intel_svm **rsvm,
287 			     struct intel_svm_dev **rsdev)
288 {
289 	struct intel_svm_dev *sdev = NULL;
290 	struct intel_svm *svm;
291 
292 	/* The caller should hold the pasid_mutex lock */
293 	if (WARN_ON(!mutex_is_locked(&pasid_mutex)))
294 		return -EINVAL;
295 
296 	if (pasid == INVALID_IOASID || pasid >= PASID_MAX)
297 		return -EINVAL;
298 
299 	svm = pasid_private_find(pasid);
300 	if (IS_ERR(svm))
301 		return PTR_ERR(svm);
302 
303 	if (!svm)
304 		goto out;
305 
306 	/*
307 	 * If we found svm for the PASID, there must be at least one device
308 	 * bond.
309 	 */
310 	if (WARN_ON(list_empty(&svm->devs)))
311 		return -EINVAL;
312 	sdev = svm_lookup_device_by_dev(svm, dev);
313 
314 out:
315 	*rsvm = svm;
316 	*rsdev = sdev;
317 
318 	return 0;
319 }
320 
321 int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
322 			  struct iommu_gpasid_bind_data *data)
323 {
324 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
325 	struct intel_svm_dev *sdev = NULL;
326 	struct dmar_domain *dmar_domain;
327 	struct device_domain_info *info;
328 	struct intel_svm *svm = NULL;
329 	unsigned long iflags;
330 	int ret = 0;
331 
332 	if (WARN_ON(!iommu) || !data)
333 		return -EINVAL;
334 
335 	if (data->format != IOMMU_PASID_FORMAT_INTEL_VTD)
336 		return -EINVAL;
337 
338 	/* IOMMU core ensures argsz is more than the start of the union */
339 	if (data->argsz < offsetofend(struct iommu_gpasid_bind_data, vendor.vtd))
340 		return -EINVAL;
341 
342 	/* Make sure no undefined flags are used in vendor data */
343 	if (data->vendor.vtd.flags & ~(IOMMU_SVA_VTD_GPASID_LAST - 1))
344 		return -EINVAL;
345 
346 	if (!dev_is_pci(dev))
347 		return -ENOTSUPP;
348 
349 	/* VT-d supports devices with full 20 bit PASIDs only */
350 	if (pci_max_pasids(to_pci_dev(dev)) != PASID_MAX)
351 		return -EINVAL;
352 
353 	/*
354 	 * We only check host PASID range, we have no knowledge to check
355 	 * guest PASID range.
356 	 */
357 	if (data->hpasid <= 0 || data->hpasid >= PASID_MAX)
358 		return -EINVAL;
359 
360 	info = get_domain_info(dev);
361 	if (!info)
362 		return -EINVAL;
363 
364 	dmar_domain = to_dmar_domain(domain);
365 
366 	mutex_lock(&pasid_mutex);
367 	ret = pasid_to_svm_sdev(dev, data->hpasid, &svm, &sdev);
368 	if (ret)
369 		goto out;
370 
371 	if (sdev) {
372 		/*
373 		 * Do not allow multiple bindings of the same device-PASID since
374 		 * there is only one SL page tables per PASID. We may revisit
375 		 * once sharing PGD across domains are supported.
376 		 */
377 		dev_warn_ratelimited(dev, "Already bound with PASID %u\n",
378 				     svm->pasid);
379 		ret = -EBUSY;
380 		goto out;
381 	}
382 
383 	if (!svm) {
384 		/* We come here when PASID has never been bond to a device. */
385 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
386 		if (!svm) {
387 			ret = -ENOMEM;
388 			goto out;
389 		}
390 		/* REVISIT: upper layer/VFIO can track host process that bind
391 		 * the PASID. ioasid_set = mm might be sufficient for vfio to
392 		 * check pasid VMM ownership. We can drop the following line
393 		 * once VFIO and IOASID set check is in place.
394 		 */
395 		svm->mm = get_task_mm(current);
396 		svm->pasid = data->hpasid;
397 		if (data->flags & IOMMU_SVA_GPASID_VAL) {
398 			svm->gpasid = data->gpasid;
399 			svm->flags |= SVM_FLAG_GUEST_PASID;
400 		}
401 		pasid_private_add(data->hpasid, svm);
402 		INIT_LIST_HEAD_RCU(&svm->devs);
403 		mmput(svm->mm);
404 	}
405 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
406 	if (!sdev) {
407 		ret = -ENOMEM;
408 		goto out;
409 	}
410 	sdev->dev = dev;
411 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
412 	sdev->iommu = iommu;
413 
414 	/* Only count users if device has aux domains */
415 	if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
416 		sdev->users = 1;
417 
418 	/* Set up device context entry for PASID if not enabled already */
419 	ret = intel_iommu_enable_pasid(iommu, sdev->dev);
420 	if (ret) {
421 		dev_err_ratelimited(dev, "Failed to enable PASID capability\n");
422 		kfree(sdev);
423 		goto out;
424 	}
425 
426 	/*
427 	 * PASID table is per device for better security. Therefore, for
428 	 * each bind of a new device even with an existing PASID, we need to
429 	 * call the nested mode setup function here.
430 	 */
431 	spin_lock_irqsave(&iommu->lock, iflags);
432 	ret = intel_pasid_setup_nested(iommu, dev,
433 				       (pgd_t *)(uintptr_t)data->gpgd,
434 				       data->hpasid, &data->vendor.vtd, dmar_domain,
435 				       data->addr_width);
436 	spin_unlock_irqrestore(&iommu->lock, iflags);
437 	if (ret) {
438 		dev_err_ratelimited(dev, "Failed to set up PASID %llu in nested mode, Err %d\n",
439 				    data->hpasid, ret);
440 		/*
441 		 * PASID entry should be in cleared state if nested mode
442 		 * set up failed. So we only need to clear IOASID tracking
443 		 * data such that free call will succeed.
444 		 */
445 		kfree(sdev);
446 		goto out;
447 	}
448 
449 	svm->flags |= SVM_FLAG_GUEST_MODE;
450 
451 	init_rcu_head(&sdev->rcu);
452 	list_add_rcu(&sdev->list, &svm->devs);
453  out:
454 	if (!IS_ERR_OR_NULL(svm) && list_empty(&svm->devs)) {
455 		pasid_private_remove(data->hpasid);
456 		kfree(svm);
457 	}
458 
459 	mutex_unlock(&pasid_mutex);
460 	return ret;
461 }
462 
463 int intel_svm_unbind_gpasid(struct device *dev, u32 pasid)
464 {
465 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
466 	struct intel_svm_dev *sdev;
467 	struct intel_svm *svm;
468 	int ret;
469 
470 	if (WARN_ON(!iommu))
471 		return -EINVAL;
472 
473 	mutex_lock(&pasid_mutex);
474 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
475 	if (ret)
476 		goto out;
477 
478 	if (sdev) {
479 		if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
480 			sdev->users--;
481 		if (!sdev->users) {
482 			list_del_rcu(&sdev->list);
483 			intel_pasid_tear_down_entry(iommu, dev,
484 						    svm->pasid, false);
485 			intel_svm_drain_prq(dev, svm->pasid);
486 			kfree_rcu(sdev, rcu);
487 
488 			if (list_empty(&svm->devs)) {
489 				/*
490 				 * We do not free the IOASID here in that
491 				 * IOMMU driver did not allocate it.
492 				 * Unlike native SVM, IOASID for guest use was
493 				 * allocated prior to the bind call.
494 				 * In any case, if the free call comes before
495 				 * the unbind, IOMMU driver will get notified
496 				 * and perform cleanup.
497 				 */
498 				pasid_private_remove(pasid);
499 				kfree(svm);
500 			}
501 		}
502 	}
503 out:
504 	mutex_unlock(&pasid_mutex);
505 	return ret;
506 }
507 
508 static void _load_pasid(void *unused)
509 {
510 	update_pasid();
511 }
512 
513 static void load_pasid(struct mm_struct *mm, u32 pasid)
514 {
515 	mutex_lock(&mm->context.lock);
516 
517 	/* Synchronize with READ_ONCE in update_pasid(). */
518 	smp_store_release(&mm->pasid, pasid);
519 
520 	/* Update PASID MSR on all CPUs running the mm's tasks. */
521 	on_each_cpu_mask(mm_cpumask(mm), _load_pasid, NULL, true);
522 
523 	mutex_unlock(&mm->context.lock);
524 }
525 
526 static int intel_svm_alloc_pasid(struct device *dev, struct mm_struct *mm,
527 				 unsigned int flags)
528 {
529 	ioasid_t max_pasid = dev_is_pci(dev) ?
530 			pci_max_pasids(to_pci_dev(dev)) : intel_pasid_max_id;
531 
532 	return iommu_sva_alloc_pasid(mm, PASID_MIN, max_pasid - 1);
533 }
534 
535 static void intel_svm_free_pasid(struct mm_struct *mm)
536 {
537 	iommu_sva_free_pasid(mm);
538 }
539 
540 static struct iommu_sva *intel_svm_bind_mm(struct intel_iommu *iommu,
541 					   struct device *dev,
542 					   struct mm_struct *mm,
543 					   unsigned int flags)
544 {
545 	struct device_domain_info *info = get_domain_info(dev);
546 	unsigned long iflags, sflags;
547 	struct intel_svm_dev *sdev;
548 	struct intel_svm *svm;
549 	int ret = 0;
550 
551 	svm = pasid_private_find(mm->pasid);
552 	if (!svm) {
553 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
554 		if (!svm)
555 			return ERR_PTR(-ENOMEM);
556 
557 		svm->pasid = mm->pasid;
558 		svm->mm = mm;
559 		svm->flags = flags;
560 		INIT_LIST_HEAD_RCU(&svm->devs);
561 
562 		if (!(flags & SVM_FLAG_SUPERVISOR_MODE)) {
563 			svm->notifier.ops = &intel_mmuops;
564 			ret = mmu_notifier_register(&svm->notifier, mm);
565 			if (ret) {
566 				kfree(svm);
567 				return ERR_PTR(ret);
568 			}
569 		}
570 
571 		ret = pasid_private_add(svm->pasid, svm);
572 		if (ret) {
573 			if (svm->notifier.ops)
574 				mmu_notifier_unregister(&svm->notifier, mm);
575 			kfree(svm);
576 			return ERR_PTR(ret);
577 		}
578 	}
579 
580 	/* Find the matching device in svm list */
581 	sdev = svm_lookup_device_by_dev(svm, dev);
582 	if (sdev) {
583 		sdev->users++;
584 		goto success;
585 	}
586 
587 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
588 	if (!sdev) {
589 		ret = -ENOMEM;
590 		goto free_svm;
591 	}
592 
593 	sdev->dev = dev;
594 	sdev->iommu = iommu;
595 	sdev->did = FLPT_DEFAULT_DID;
596 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
597 	sdev->users = 1;
598 	sdev->pasid = svm->pasid;
599 	sdev->sva.dev = dev;
600 	init_rcu_head(&sdev->rcu);
601 	if (info->ats_enabled) {
602 		sdev->dev_iotlb = 1;
603 		sdev->qdep = info->ats_qdep;
604 		if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
605 			sdev->qdep = 0;
606 	}
607 
608 	/* Setup the pasid table: */
609 	sflags = (flags & SVM_FLAG_SUPERVISOR_MODE) ?
610 			PASID_FLAG_SUPERVISOR_MODE : 0;
611 	sflags |= cpu_feature_enabled(X86_FEATURE_LA57) ? PASID_FLAG_FL5LP : 0;
612 	spin_lock_irqsave(&iommu->lock, iflags);
613 	ret = intel_pasid_setup_first_level(iommu, dev, mm->pgd, mm->pasid,
614 					    FLPT_DEFAULT_DID, sflags);
615 	spin_unlock_irqrestore(&iommu->lock, iflags);
616 
617 	if (ret)
618 		goto free_sdev;
619 
620 	/* The newly allocated pasid is loaded to the mm. */
621 	if (!(flags & SVM_FLAG_SUPERVISOR_MODE) && list_empty(&svm->devs))
622 		load_pasid(mm, svm->pasid);
623 
624 	list_add_rcu(&sdev->list, &svm->devs);
625 success:
626 	return &sdev->sva;
627 
628 free_sdev:
629 	kfree(sdev);
630 free_svm:
631 	if (list_empty(&svm->devs)) {
632 		if (svm->notifier.ops)
633 			mmu_notifier_unregister(&svm->notifier, mm);
634 		pasid_private_remove(mm->pasid);
635 		kfree(svm);
636 	}
637 
638 	return ERR_PTR(ret);
639 }
640 
641 /* Caller must hold pasid_mutex */
642 static int intel_svm_unbind_mm(struct device *dev, u32 pasid)
643 {
644 	struct intel_svm_dev *sdev;
645 	struct intel_iommu *iommu;
646 	struct intel_svm *svm;
647 	struct mm_struct *mm;
648 	int ret = -EINVAL;
649 
650 	iommu = device_to_iommu(dev, NULL, NULL);
651 	if (!iommu)
652 		goto out;
653 
654 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
655 	if (ret)
656 		goto out;
657 	mm = svm->mm;
658 
659 	if (sdev) {
660 		sdev->users--;
661 		if (!sdev->users) {
662 			list_del_rcu(&sdev->list);
663 			/* Flush the PASID cache and IOTLB for this device.
664 			 * Note that we do depend on the hardware *not* using
665 			 * the PASID any more. Just as we depend on other
666 			 * devices never using PASIDs that they have no right
667 			 * to use. We have a *shared* PASID table, because it's
668 			 * large and has to be physically contiguous. So it's
669 			 * hard to be as defensive as we might like. */
670 			intel_pasid_tear_down_entry(iommu, dev,
671 						    svm->pasid, false);
672 			intel_svm_drain_prq(dev, svm->pasid);
673 			kfree_rcu(sdev, rcu);
674 
675 			if (list_empty(&svm->devs)) {
676 				if (svm->notifier.ops) {
677 					mmu_notifier_unregister(&svm->notifier, mm);
678 					/* Clear mm's pasid. */
679 					load_pasid(mm, PASID_DISABLED);
680 				}
681 				pasid_private_remove(svm->pasid);
682 				/* We mandate that no page faults may be outstanding
683 				 * for the PASID when intel_svm_unbind_mm() is called.
684 				 * If that is not obeyed, subtle errors will happen.
685 				 * Let's make them less subtle... */
686 				memset(svm, 0x6b, sizeof(*svm));
687 				kfree(svm);
688 			}
689 		}
690 		/* Drop a PASID reference and free it if no reference. */
691 		intel_svm_free_pasid(mm);
692 	}
693 out:
694 	return ret;
695 }
696 
697 /* Page request queue descriptor */
698 struct page_req_dsc {
699 	union {
700 		struct {
701 			u64 type:8;
702 			u64 pasid_present:1;
703 			u64 priv_data_present:1;
704 			u64 rsvd:6;
705 			u64 rid:16;
706 			u64 pasid:20;
707 			u64 exe_req:1;
708 			u64 pm_req:1;
709 			u64 rsvd2:10;
710 		};
711 		u64 qw_0;
712 	};
713 	union {
714 		struct {
715 			u64 rd_req:1;
716 			u64 wr_req:1;
717 			u64 lpig:1;
718 			u64 prg_index:9;
719 			u64 addr:52;
720 		};
721 		u64 qw_1;
722 	};
723 	u64 priv_data[2];
724 };
725 
726 static bool is_canonical_address(u64 addr)
727 {
728 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
729 	long saddr = (long) addr;
730 
731 	return (((saddr << shift) >> shift) == saddr);
732 }
733 
734 /**
735  * intel_svm_drain_prq - Drain page requests and responses for a pasid
736  * @dev: target device
737  * @pasid: pasid for draining
738  *
739  * Drain all pending page requests and responses related to @pasid in both
740  * software and hardware. This is supposed to be called after the device
741  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
742  * and DevTLB have been invalidated.
743  *
744  * It waits until all pending page requests for @pasid in the page fault
745  * queue are completed by the prq handling thread. Then follow the steps
746  * described in VT-d spec CH7.10 to drain all page requests and page
747  * responses pending in the hardware.
748  */
749 static void intel_svm_drain_prq(struct device *dev, u32 pasid)
750 {
751 	struct device_domain_info *info;
752 	struct dmar_domain *domain;
753 	struct intel_iommu *iommu;
754 	struct qi_desc desc[3];
755 	struct pci_dev *pdev;
756 	int head, tail;
757 	u16 sid, did;
758 	int qdep;
759 
760 	info = get_domain_info(dev);
761 	if (WARN_ON(!info || !dev_is_pci(dev)))
762 		return;
763 
764 	if (!info->pri_enabled)
765 		return;
766 
767 	iommu = info->iommu;
768 	domain = info->domain;
769 	pdev = to_pci_dev(dev);
770 	sid = PCI_DEVID(info->bus, info->devfn);
771 	did = domain->iommu_did[iommu->seq_id];
772 	qdep = pci_ats_queue_depth(pdev);
773 
774 	/*
775 	 * Check and wait until all pending page requests in the queue are
776 	 * handled by the prq handling thread.
777 	 */
778 prq_retry:
779 	reinit_completion(&iommu->prq_complete);
780 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
781 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
782 	while (head != tail) {
783 		struct page_req_dsc *req;
784 
785 		req = &iommu->prq[head / sizeof(*req)];
786 		if (!req->pasid_present || req->pasid != pasid) {
787 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
788 			continue;
789 		}
790 
791 		wait_for_completion(&iommu->prq_complete);
792 		goto prq_retry;
793 	}
794 
795 	iopf_queue_flush_dev(dev);
796 
797 	/*
798 	 * Perform steps described in VT-d spec CH7.10 to drain page
799 	 * requests and responses in hardware.
800 	 */
801 	memset(desc, 0, sizeof(desc));
802 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
803 			QI_IWD_FENCE |
804 			QI_IWD_TYPE;
805 	desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
806 			QI_EIOTLB_DID(did) |
807 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
808 			QI_EIOTLB_TYPE;
809 	desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
810 			QI_DEV_EIOTLB_SID(sid) |
811 			QI_DEV_EIOTLB_QDEP(qdep) |
812 			QI_DEIOTLB_TYPE |
813 			QI_DEV_IOTLB_PFSID(info->pfsid);
814 qi_retry:
815 	reinit_completion(&iommu->prq_complete);
816 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
817 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
818 		wait_for_completion(&iommu->prq_complete);
819 		goto qi_retry;
820 	}
821 }
822 
823 static int prq_to_iommu_prot(struct page_req_dsc *req)
824 {
825 	int prot = 0;
826 
827 	if (req->rd_req)
828 		prot |= IOMMU_FAULT_PERM_READ;
829 	if (req->wr_req)
830 		prot |= IOMMU_FAULT_PERM_WRITE;
831 	if (req->exe_req)
832 		prot |= IOMMU_FAULT_PERM_EXEC;
833 	if (req->pm_req)
834 		prot |= IOMMU_FAULT_PERM_PRIV;
835 
836 	return prot;
837 }
838 
839 static int intel_svm_prq_report(struct intel_iommu *iommu, struct device *dev,
840 				struct page_req_dsc *desc)
841 {
842 	struct iommu_fault_event event;
843 
844 	if (!dev || !dev_is_pci(dev))
845 		return -ENODEV;
846 
847 	/* Fill in event data for device specific processing */
848 	memset(&event, 0, sizeof(struct iommu_fault_event));
849 	event.fault.type = IOMMU_FAULT_PAGE_REQ;
850 	event.fault.prm.addr = (u64)desc->addr << VTD_PAGE_SHIFT;
851 	event.fault.prm.pasid = desc->pasid;
852 	event.fault.prm.grpid = desc->prg_index;
853 	event.fault.prm.perm = prq_to_iommu_prot(desc);
854 
855 	if (desc->lpig)
856 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
857 	if (desc->pasid_present) {
858 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
859 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
860 	}
861 	if (desc->priv_data_present) {
862 		/*
863 		 * Set last page in group bit if private data is present,
864 		 * page response is required as it does for LPIG.
865 		 * iommu_report_device_fault() doesn't understand this vendor
866 		 * specific requirement thus we set last_page as a workaround.
867 		 */
868 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
869 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
870 		event.fault.prm.private_data[0] = desc->priv_data[0];
871 		event.fault.prm.private_data[1] = desc->priv_data[1];
872 	} else if (dmar_latency_enabled(iommu, DMAR_LATENCY_PRQ)) {
873 		/*
874 		 * If the private data fields are not used by hardware, use it
875 		 * to monitor the prq handle latency.
876 		 */
877 		event.fault.prm.private_data[0] = ktime_to_ns(ktime_get());
878 	}
879 
880 	return iommu_report_device_fault(dev, &event);
881 }
882 
883 static void handle_bad_prq_event(struct intel_iommu *iommu,
884 				 struct page_req_dsc *req, int result)
885 {
886 	struct qi_desc desc;
887 
888 	pr_err("%s: Invalid page request: %08llx %08llx\n",
889 	       iommu->name, ((unsigned long long *)req)[0],
890 	       ((unsigned long long *)req)[1]);
891 
892 	/*
893 	 * Per VT-d spec. v3.0 ch7.7, system software must
894 	 * respond with page group response if private data
895 	 * is present (PDP) or last page in group (LPIG) bit
896 	 * is set. This is an additional VT-d feature beyond
897 	 * PCI ATS spec.
898 	 */
899 	if (!req->lpig && !req->priv_data_present)
900 		return;
901 
902 	desc.qw0 = QI_PGRP_PASID(req->pasid) |
903 			QI_PGRP_DID(req->rid) |
904 			QI_PGRP_PASID_P(req->pasid_present) |
905 			QI_PGRP_PDP(req->priv_data_present) |
906 			QI_PGRP_RESP_CODE(result) |
907 			QI_PGRP_RESP_TYPE;
908 	desc.qw1 = QI_PGRP_IDX(req->prg_index) |
909 			QI_PGRP_LPIG(req->lpig);
910 
911 	if (req->priv_data_present) {
912 		desc.qw2 = req->priv_data[0];
913 		desc.qw3 = req->priv_data[1];
914 	} else {
915 		desc.qw2 = 0;
916 		desc.qw3 = 0;
917 	}
918 
919 	qi_submit_sync(iommu, &desc, 1, 0);
920 }
921 
922 static irqreturn_t prq_event_thread(int irq, void *d)
923 {
924 	struct intel_svm_dev *sdev = NULL;
925 	struct intel_iommu *iommu = d;
926 	struct intel_svm *svm = NULL;
927 	struct page_req_dsc *req;
928 	int head, tail, handled;
929 	u64 address;
930 
931 	/*
932 	 * Clear PPR bit before reading head/tail registers, to ensure that
933 	 * we get a new interrupt if needed.
934 	 */
935 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
936 
937 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
938 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
939 	handled = (head != tail);
940 	while (head != tail) {
941 		req = &iommu->prq[head / sizeof(*req)];
942 		address = (u64)req->addr << VTD_PAGE_SHIFT;
943 
944 		if (unlikely(!req->pasid_present)) {
945 			pr_err("IOMMU: %s: Page request without PASID\n",
946 			       iommu->name);
947 bad_req:
948 			svm = NULL;
949 			sdev = NULL;
950 			handle_bad_prq_event(iommu, req, QI_RESP_INVALID);
951 			goto prq_advance;
952 		}
953 
954 		if (unlikely(!is_canonical_address(address))) {
955 			pr_err("IOMMU: %s: Address is not canonical\n",
956 			       iommu->name);
957 			goto bad_req;
958 		}
959 
960 		if (unlikely(req->pm_req && (req->rd_req | req->wr_req))) {
961 			pr_err("IOMMU: %s: Page request in Privilege Mode\n",
962 			       iommu->name);
963 			goto bad_req;
964 		}
965 
966 		if (unlikely(req->exe_req && req->rd_req)) {
967 			pr_err("IOMMU: %s: Execution request not supported\n",
968 			       iommu->name);
969 			goto bad_req;
970 		}
971 
972 		if (!svm || svm->pasid != req->pasid) {
973 			/*
974 			 * It can't go away, because the driver is not permitted
975 			 * to unbind the mm while any page faults are outstanding.
976 			 */
977 			svm = pasid_private_find(req->pasid);
978 			if (IS_ERR_OR_NULL(svm) || (svm->flags & SVM_FLAG_SUPERVISOR_MODE))
979 				goto bad_req;
980 		}
981 
982 		if (!sdev || sdev->sid != req->rid) {
983 			sdev = svm_lookup_device_by_sid(svm, req->rid);
984 			if (!sdev)
985 				goto bad_req;
986 		}
987 
988 		sdev->prq_seq_number++;
989 
990 		/*
991 		 * If prq is to be handled outside iommu driver via receiver of
992 		 * the fault notifiers, we skip the page response here.
993 		 */
994 		if (intel_svm_prq_report(iommu, sdev->dev, req))
995 			handle_bad_prq_event(iommu, req, QI_RESP_INVALID);
996 
997 		trace_prq_report(iommu, sdev->dev, req->qw_0, req->qw_1,
998 				 req->priv_data[0], req->priv_data[1],
999 				 sdev->prq_seq_number);
1000 prq_advance:
1001 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
1002 	}
1003 
1004 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
1005 
1006 	/*
1007 	 * Clear the page request overflow bit and wake up all threads that
1008 	 * are waiting for the completion of this handling.
1009 	 */
1010 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
1011 		pr_info_ratelimited("IOMMU: %s: PRQ overflow detected\n",
1012 				    iommu->name);
1013 		head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
1014 		tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
1015 		if (head == tail) {
1016 			iopf_queue_discard_partial(iommu->iopf_queue);
1017 			writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
1018 			pr_info_ratelimited("IOMMU: %s: PRQ overflow cleared",
1019 					    iommu->name);
1020 		}
1021 	}
1022 
1023 	if (!completion_done(&iommu->prq_complete))
1024 		complete(&iommu->prq_complete);
1025 
1026 	return IRQ_RETVAL(handled);
1027 }
1028 
1029 struct iommu_sva *intel_svm_bind(struct device *dev, struct mm_struct *mm, void *drvdata)
1030 {
1031 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
1032 	unsigned int flags = 0;
1033 	struct iommu_sva *sva;
1034 	int ret;
1035 
1036 	if (drvdata)
1037 		flags = *(unsigned int *)drvdata;
1038 
1039 	if (flags & SVM_FLAG_SUPERVISOR_MODE) {
1040 		if (!ecap_srs(iommu->ecap)) {
1041 			dev_err(dev, "%s: Supervisor PASID not supported\n",
1042 				iommu->name);
1043 			return ERR_PTR(-EOPNOTSUPP);
1044 		}
1045 
1046 		if (mm) {
1047 			dev_err(dev, "%s: Supervisor PASID with user provided mm\n",
1048 				iommu->name);
1049 			return ERR_PTR(-EINVAL);
1050 		}
1051 
1052 		mm = &init_mm;
1053 	}
1054 
1055 	mutex_lock(&pasid_mutex);
1056 	ret = intel_svm_alloc_pasid(dev, mm, flags);
1057 	if (ret) {
1058 		mutex_unlock(&pasid_mutex);
1059 		return ERR_PTR(ret);
1060 	}
1061 
1062 	sva = intel_svm_bind_mm(iommu, dev, mm, flags);
1063 	if (IS_ERR_OR_NULL(sva))
1064 		intel_svm_free_pasid(mm);
1065 	mutex_unlock(&pasid_mutex);
1066 
1067 	return sva;
1068 }
1069 
1070 void intel_svm_unbind(struct iommu_sva *sva)
1071 {
1072 	struct intel_svm_dev *sdev = to_intel_svm_dev(sva);
1073 
1074 	mutex_lock(&pasid_mutex);
1075 	intel_svm_unbind_mm(sdev->dev, sdev->pasid);
1076 	mutex_unlock(&pasid_mutex);
1077 }
1078 
1079 u32 intel_svm_get_pasid(struct iommu_sva *sva)
1080 {
1081 	struct intel_svm_dev *sdev;
1082 	u32 pasid;
1083 
1084 	mutex_lock(&pasid_mutex);
1085 	sdev = to_intel_svm_dev(sva);
1086 	pasid = sdev->pasid;
1087 	mutex_unlock(&pasid_mutex);
1088 
1089 	return pasid;
1090 }
1091 
1092 int intel_svm_page_response(struct device *dev,
1093 			    struct iommu_fault_event *evt,
1094 			    struct iommu_page_response *msg)
1095 {
1096 	struct iommu_fault_page_request *prm;
1097 	struct intel_svm_dev *sdev = NULL;
1098 	struct intel_svm *svm = NULL;
1099 	struct intel_iommu *iommu;
1100 	bool private_present;
1101 	bool pasid_present;
1102 	bool last_page;
1103 	u8 bus, devfn;
1104 	int ret = 0;
1105 	u16 sid;
1106 
1107 	if (!dev || !dev_is_pci(dev))
1108 		return -ENODEV;
1109 
1110 	iommu = device_to_iommu(dev, &bus, &devfn);
1111 	if (!iommu)
1112 		return -ENODEV;
1113 
1114 	if (!msg || !evt)
1115 		return -EINVAL;
1116 
1117 	mutex_lock(&pasid_mutex);
1118 
1119 	prm = &evt->fault.prm;
1120 	sid = PCI_DEVID(bus, devfn);
1121 	pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
1122 	private_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
1123 	last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
1124 
1125 	if (!pasid_present) {
1126 		ret = -EINVAL;
1127 		goto out;
1128 	}
1129 
1130 	if (prm->pasid == 0 || prm->pasid >= PASID_MAX) {
1131 		ret = -EINVAL;
1132 		goto out;
1133 	}
1134 
1135 	ret = pasid_to_svm_sdev(dev, prm->pasid, &svm, &sdev);
1136 	if (ret || !sdev) {
1137 		ret = -ENODEV;
1138 		goto out;
1139 	}
1140 
1141 	/*
1142 	 * For responses from userspace, need to make sure that the
1143 	 * pasid has been bound to its mm.
1144 	 */
1145 	if (svm->flags & SVM_FLAG_GUEST_MODE) {
1146 		struct mm_struct *mm;
1147 
1148 		mm = get_task_mm(current);
1149 		if (!mm) {
1150 			ret = -EINVAL;
1151 			goto out;
1152 		}
1153 
1154 		if (mm != svm->mm) {
1155 			ret = -ENODEV;
1156 			mmput(mm);
1157 			goto out;
1158 		}
1159 
1160 		mmput(mm);
1161 	}
1162 
1163 	/*
1164 	 * Per VT-d spec. v3.0 ch7.7, system software must respond
1165 	 * with page group response if private data is present (PDP)
1166 	 * or last page in group (LPIG) bit is set. This is an
1167 	 * additional VT-d requirement beyond PCI ATS spec.
1168 	 */
1169 	if (last_page || private_present) {
1170 		struct qi_desc desc;
1171 
1172 		desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
1173 				QI_PGRP_PASID_P(pasid_present) |
1174 				QI_PGRP_PDP(private_present) |
1175 				QI_PGRP_RESP_CODE(msg->code) |
1176 				QI_PGRP_RESP_TYPE;
1177 		desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
1178 		desc.qw2 = 0;
1179 		desc.qw3 = 0;
1180 
1181 		if (private_present) {
1182 			desc.qw2 = prm->private_data[0];
1183 			desc.qw3 = prm->private_data[1];
1184 		} else if (prm->private_data[0]) {
1185 			dmar_latency_update(iommu, DMAR_LATENCY_PRQ,
1186 				ktime_to_ns(ktime_get()) - prm->private_data[0]);
1187 		}
1188 
1189 		qi_submit_sync(iommu, &desc, 1, 0);
1190 	}
1191 out:
1192 	mutex_unlock(&pasid_mutex);
1193 	return ret;
1194 }
1195