xref: /openbmc/linux/drivers/iommu/intel/svm.c (revision 36fe4655)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright © 2015 Intel Corporation.
4  *
5  * Authors: David Woodhouse <dwmw2@infradead.org>
6  */
7 
8 #include <linux/intel-iommu.h>
9 #include <linux/mmu_notifier.h>
10 #include <linux/sched.h>
11 #include <linux/sched/mm.h>
12 #include <linux/slab.h>
13 #include <linux/intel-svm.h>
14 #include <linux/rculist.h>
15 #include <linux/pci.h>
16 #include <linux/pci-ats.h>
17 #include <linux/dmar.h>
18 #include <linux/interrupt.h>
19 #include <linux/mm_types.h>
20 #include <linux/ioasid.h>
21 #include <asm/page.h>
22 #include <asm/fpu/api.h>
23 
24 #include "pasid.h"
25 
26 static irqreturn_t prq_event_thread(int irq, void *d);
27 static void intel_svm_drain_prq(struct device *dev, u32 pasid);
28 
29 #define PRQ_ORDER 0
30 
31 int intel_svm_enable_prq(struct intel_iommu *iommu)
32 {
33 	struct page *pages;
34 	int irq, ret;
35 
36 	pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
37 	if (!pages) {
38 		pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
39 			iommu->name);
40 		return -ENOMEM;
41 	}
42 	iommu->prq = page_address(pages);
43 
44 	irq = dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED + iommu->seq_id, iommu->node, iommu);
45 	if (irq <= 0) {
46 		pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
47 		       iommu->name);
48 		ret = -EINVAL;
49 	err:
50 		free_pages((unsigned long)iommu->prq, PRQ_ORDER);
51 		iommu->prq = NULL;
52 		return ret;
53 	}
54 	iommu->pr_irq = irq;
55 
56 	snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
57 
58 	ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
59 				   iommu->prq_name, iommu);
60 	if (ret) {
61 		pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
62 		       iommu->name);
63 		dmar_free_hwirq(irq);
64 		iommu->pr_irq = 0;
65 		goto err;
66 	}
67 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
68 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
69 	dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
70 
71 	init_completion(&iommu->prq_complete);
72 
73 	return 0;
74 }
75 
76 int intel_svm_finish_prq(struct intel_iommu *iommu)
77 {
78 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
79 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
80 	dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
81 
82 	if (iommu->pr_irq) {
83 		free_irq(iommu->pr_irq, iommu);
84 		dmar_free_hwirq(iommu->pr_irq);
85 		iommu->pr_irq = 0;
86 	}
87 
88 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
89 	iommu->prq = NULL;
90 
91 	return 0;
92 }
93 
94 static inline bool intel_svm_capable(struct intel_iommu *iommu)
95 {
96 	return iommu->flags & VTD_FLAG_SVM_CAPABLE;
97 }
98 
99 void intel_svm_check(struct intel_iommu *iommu)
100 {
101 	if (!pasid_supported(iommu))
102 		return;
103 
104 	if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
105 	    !cap_fl1gp_support(iommu->cap)) {
106 		pr_err("%s SVM disabled, incompatible 1GB page capability\n",
107 		       iommu->name);
108 		return;
109 	}
110 
111 	if (cpu_feature_enabled(X86_FEATURE_LA57) &&
112 	    !cap_5lp_support(iommu->cap)) {
113 		pr_err("%s SVM disabled, incompatible paging mode\n",
114 		       iommu->name);
115 		return;
116 	}
117 
118 	iommu->flags |= VTD_FLAG_SVM_CAPABLE;
119 }
120 
121 static void intel_flush_svm_range_dev (struct intel_svm *svm, struct intel_svm_dev *sdev,
122 				unsigned long address, unsigned long pages, int ih)
123 {
124 	struct qi_desc desc;
125 
126 	if (pages == -1) {
127 		desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
128 			QI_EIOTLB_DID(sdev->did) |
129 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
130 			QI_EIOTLB_TYPE;
131 		desc.qw1 = 0;
132 	} else {
133 		int mask = ilog2(__roundup_pow_of_two(pages));
134 
135 		desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
136 				QI_EIOTLB_DID(sdev->did) |
137 				QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID) |
138 				QI_EIOTLB_TYPE;
139 		desc.qw1 = QI_EIOTLB_ADDR(address) |
140 				QI_EIOTLB_IH(ih) |
141 				QI_EIOTLB_AM(mask);
142 	}
143 	desc.qw2 = 0;
144 	desc.qw3 = 0;
145 	qi_submit_sync(svm->iommu, &desc, 1, 0);
146 
147 	if (sdev->dev_iotlb) {
148 		desc.qw0 = QI_DEV_EIOTLB_PASID(svm->pasid) |
149 				QI_DEV_EIOTLB_SID(sdev->sid) |
150 				QI_DEV_EIOTLB_QDEP(sdev->qdep) |
151 				QI_DEIOTLB_TYPE;
152 		if (pages == -1) {
153 			desc.qw1 = QI_DEV_EIOTLB_ADDR(-1ULL >> 1) |
154 					QI_DEV_EIOTLB_SIZE;
155 		} else if (pages > 1) {
156 			/* The least significant zero bit indicates the size. So,
157 			 * for example, an "address" value of 0x12345f000 will
158 			 * flush from 0x123440000 to 0x12347ffff (256KiB). */
159 			unsigned long last = address + ((unsigned long)(pages - 1) << VTD_PAGE_SHIFT);
160 			unsigned long mask = __rounddown_pow_of_two(address ^ last);
161 
162 			desc.qw1 = QI_DEV_EIOTLB_ADDR((address & ~mask) |
163 					(mask - 1)) | QI_DEV_EIOTLB_SIZE;
164 		} else {
165 			desc.qw1 = QI_DEV_EIOTLB_ADDR(address);
166 		}
167 		desc.qw2 = 0;
168 		desc.qw3 = 0;
169 		qi_submit_sync(svm->iommu, &desc, 1, 0);
170 	}
171 }
172 
173 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
174 				unsigned long pages, int ih)
175 {
176 	struct intel_svm_dev *sdev;
177 
178 	rcu_read_lock();
179 	list_for_each_entry_rcu(sdev, &svm->devs, list)
180 		intel_flush_svm_range_dev(svm, sdev, address, pages, ih);
181 	rcu_read_unlock();
182 }
183 
184 /* Pages have been freed at this point */
185 static void intel_invalidate_range(struct mmu_notifier *mn,
186 				   struct mm_struct *mm,
187 				   unsigned long start, unsigned long end)
188 {
189 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
190 
191 	intel_flush_svm_range(svm, start,
192 			      (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0);
193 }
194 
195 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
196 {
197 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
198 	struct intel_svm_dev *sdev;
199 
200 	/* This might end up being called from exit_mmap(), *before* the page
201 	 * tables are cleared. And __mmu_notifier_release() will delete us from
202 	 * the list of notifiers so that our invalidate_range() callback doesn't
203 	 * get called when the page tables are cleared. So we need to protect
204 	 * against hardware accessing those page tables.
205 	 *
206 	 * We do it by clearing the entry in the PASID table and then flushing
207 	 * the IOTLB and the PASID table caches. This might upset hardware;
208 	 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
209 	 * page) so that we end up taking a fault that the hardware really
210 	 * *has* to handle gracefully without affecting other processes.
211 	 */
212 	rcu_read_lock();
213 	list_for_each_entry_rcu(sdev, &svm->devs, list)
214 		intel_pasid_tear_down_entry(svm->iommu, sdev->dev,
215 					    svm->pasid, true);
216 	rcu_read_unlock();
217 
218 }
219 
220 static const struct mmu_notifier_ops intel_mmuops = {
221 	.release = intel_mm_release,
222 	.invalidate_range = intel_invalidate_range,
223 };
224 
225 static DEFINE_MUTEX(pasid_mutex);
226 static LIST_HEAD(global_svm_list);
227 
228 #define for_each_svm_dev(sdev, svm, d)			\
229 	list_for_each_entry((sdev), &(svm)->devs, list)	\
230 		if ((d) != (sdev)->dev) {} else
231 
232 static int pasid_to_svm_sdev(struct device *dev, unsigned int pasid,
233 			     struct intel_svm **rsvm,
234 			     struct intel_svm_dev **rsdev)
235 {
236 	struct intel_svm_dev *d, *sdev = NULL;
237 	struct intel_svm *svm;
238 
239 	/* The caller should hold the pasid_mutex lock */
240 	if (WARN_ON(!mutex_is_locked(&pasid_mutex)))
241 		return -EINVAL;
242 
243 	if (pasid == INVALID_IOASID || pasid >= PASID_MAX)
244 		return -EINVAL;
245 
246 	svm = ioasid_find(NULL, pasid, NULL);
247 	if (IS_ERR(svm))
248 		return PTR_ERR(svm);
249 
250 	if (!svm)
251 		goto out;
252 
253 	/*
254 	 * If we found svm for the PASID, there must be at least one device
255 	 * bond.
256 	 */
257 	if (WARN_ON(list_empty(&svm->devs)))
258 		return -EINVAL;
259 
260 	rcu_read_lock();
261 	list_for_each_entry_rcu(d, &svm->devs, list) {
262 		if (d->dev == dev) {
263 			sdev = d;
264 			break;
265 		}
266 	}
267 	rcu_read_unlock();
268 
269 out:
270 	*rsvm = svm;
271 	*rsdev = sdev;
272 
273 	return 0;
274 }
275 
276 int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
277 			  struct iommu_gpasid_bind_data *data)
278 {
279 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
280 	struct intel_svm_dev *sdev = NULL;
281 	struct dmar_domain *dmar_domain;
282 	struct device_domain_info *info;
283 	struct intel_svm *svm = NULL;
284 	int ret = 0;
285 
286 	if (WARN_ON(!iommu) || !data)
287 		return -EINVAL;
288 
289 	if (data->format != IOMMU_PASID_FORMAT_INTEL_VTD)
290 		return -EINVAL;
291 
292 	/* IOMMU core ensures argsz is more than the start of the union */
293 	if (data->argsz < offsetofend(struct iommu_gpasid_bind_data, vendor.vtd))
294 		return -EINVAL;
295 
296 	/* Make sure no undefined flags are used in vendor data */
297 	if (data->vendor.vtd.flags & ~(IOMMU_SVA_VTD_GPASID_LAST - 1))
298 		return -EINVAL;
299 
300 	if (!dev_is_pci(dev))
301 		return -ENOTSUPP;
302 
303 	/* VT-d supports devices with full 20 bit PASIDs only */
304 	if (pci_max_pasids(to_pci_dev(dev)) != PASID_MAX)
305 		return -EINVAL;
306 
307 	/*
308 	 * We only check host PASID range, we have no knowledge to check
309 	 * guest PASID range.
310 	 */
311 	if (data->hpasid <= 0 || data->hpasid >= PASID_MAX)
312 		return -EINVAL;
313 
314 	info = get_domain_info(dev);
315 	if (!info)
316 		return -EINVAL;
317 
318 	dmar_domain = to_dmar_domain(domain);
319 
320 	mutex_lock(&pasid_mutex);
321 	ret = pasid_to_svm_sdev(dev, data->hpasid, &svm, &sdev);
322 	if (ret)
323 		goto out;
324 
325 	if (sdev) {
326 		/*
327 		 * Do not allow multiple bindings of the same device-PASID since
328 		 * there is only one SL page tables per PASID. We may revisit
329 		 * once sharing PGD across domains are supported.
330 		 */
331 		dev_warn_ratelimited(dev, "Already bound with PASID %u\n",
332 				     svm->pasid);
333 		ret = -EBUSY;
334 		goto out;
335 	}
336 
337 	if (!svm) {
338 		/* We come here when PASID has never been bond to a device. */
339 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
340 		if (!svm) {
341 			ret = -ENOMEM;
342 			goto out;
343 		}
344 		/* REVISIT: upper layer/VFIO can track host process that bind
345 		 * the PASID. ioasid_set = mm might be sufficient for vfio to
346 		 * check pasid VMM ownership. We can drop the following line
347 		 * once VFIO and IOASID set check is in place.
348 		 */
349 		svm->mm = get_task_mm(current);
350 		svm->pasid = data->hpasid;
351 		if (data->flags & IOMMU_SVA_GPASID_VAL) {
352 			svm->gpasid = data->gpasid;
353 			svm->flags |= SVM_FLAG_GUEST_PASID;
354 		}
355 		ioasid_set_data(data->hpasid, svm);
356 		INIT_LIST_HEAD_RCU(&svm->devs);
357 		mmput(svm->mm);
358 	}
359 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
360 	if (!sdev) {
361 		ret = -ENOMEM;
362 		goto out;
363 	}
364 	sdev->dev = dev;
365 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
366 
367 	/* Only count users if device has aux domains */
368 	if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
369 		sdev->users = 1;
370 
371 	/* Set up device context entry for PASID if not enabled already */
372 	ret = intel_iommu_enable_pasid(iommu, sdev->dev);
373 	if (ret) {
374 		dev_err_ratelimited(dev, "Failed to enable PASID capability\n");
375 		kfree(sdev);
376 		goto out;
377 	}
378 
379 	/*
380 	 * PASID table is per device for better security. Therefore, for
381 	 * each bind of a new device even with an existing PASID, we need to
382 	 * call the nested mode setup function here.
383 	 */
384 	spin_lock(&iommu->lock);
385 	ret = intel_pasid_setup_nested(iommu, dev,
386 				       (pgd_t *)(uintptr_t)data->gpgd,
387 				       data->hpasid, &data->vendor.vtd, dmar_domain,
388 				       data->addr_width);
389 	spin_unlock(&iommu->lock);
390 	if (ret) {
391 		dev_err_ratelimited(dev, "Failed to set up PASID %llu in nested mode, Err %d\n",
392 				    data->hpasid, ret);
393 		/*
394 		 * PASID entry should be in cleared state if nested mode
395 		 * set up failed. So we only need to clear IOASID tracking
396 		 * data such that free call will succeed.
397 		 */
398 		kfree(sdev);
399 		goto out;
400 	}
401 
402 	svm->flags |= SVM_FLAG_GUEST_MODE;
403 
404 	init_rcu_head(&sdev->rcu);
405 	list_add_rcu(&sdev->list, &svm->devs);
406  out:
407 	if (!IS_ERR_OR_NULL(svm) && list_empty(&svm->devs)) {
408 		ioasid_set_data(data->hpasid, NULL);
409 		kfree(svm);
410 	}
411 
412 	mutex_unlock(&pasid_mutex);
413 	return ret;
414 }
415 
416 int intel_svm_unbind_gpasid(struct device *dev, u32 pasid)
417 {
418 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
419 	struct intel_svm_dev *sdev;
420 	struct intel_svm *svm;
421 	int ret;
422 
423 	if (WARN_ON(!iommu))
424 		return -EINVAL;
425 
426 	mutex_lock(&pasid_mutex);
427 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
428 	if (ret)
429 		goto out;
430 
431 	if (sdev) {
432 		if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
433 			sdev->users--;
434 		if (!sdev->users) {
435 			list_del_rcu(&sdev->list);
436 			intel_pasid_tear_down_entry(iommu, dev,
437 						    svm->pasid, false);
438 			intel_svm_drain_prq(dev, svm->pasid);
439 			kfree_rcu(sdev, rcu);
440 
441 			if (list_empty(&svm->devs)) {
442 				/*
443 				 * We do not free the IOASID here in that
444 				 * IOMMU driver did not allocate it.
445 				 * Unlike native SVM, IOASID for guest use was
446 				 * allocated prior to the bind call.
447 				 * In any case, if the free call comes before
448 				 * the unbind, IOMMU driver will get notified
449 				 * and perform cleanup.
450 				 */
451 				ioasid_set_data(pasid, NULL);
452 				kfree(svm);
453 			}
454 		}
455 	}
456 out:
457 	mutex_unlock(&pasid_mutex);
458 	return ret;
459 }
460 
461 static void _load_pasid(void *unused)
462 {
463 	update_pasid();
464 }
465 
466 static void load_pasid(struct mm_struct *mm, u32 pasid)
467 {
468 	mutex_lock(&mm->context.lock);
469 
470 	/* Synchronize with READ_ONCE in update_pasid(). */
471 	smp_store_release(&mm->pasid, pasid);
472 
473 	/* Update PASID MSR on all CPUs running the mm's tasks. */
474 	on_each_cpu_mask(mm_cpumask(mm), _load_pasid, NULL, true);
475 
476 	mutex_unlock(&mm->context.lock);
477 }
478 
479 /* Caller must hold pasid_mutex, mm reference */
480 static int
481 intel_svm_bind_mm(struct device *dev, unsigned int flags,
482 		  struct svm_dev_ops *ops,
483 		  struct mm_struct *mm, struct intel_svm_dev **sd)
484 {
485 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
486 	struct device_domain_info *info;
487 	struct intel_svm_dev *sdev;
488 	struct intel_svm *svm = NULL;
489 	int pasid_max;
490 	int ret;
491 
492 	if (!iommu || dmar_disabled)
493 		return -EINVAL;
494 
495 	if (!intel_svm_capable(iommu))
496 		return -ENOTSUPP;
497 
498 	if (dev_is_pci(dev)) {
499 		pasid_max = pci_max_pasids(to_pci_dev(dev));
500 		if (pasid_max < 0)
501 			return -EINVAL;
502 	} else
503 		pasid_max = 1 << 20;
504 
505 	/* Bind supervisor PASID shuld have mm = NULL */
506 	if (flags & SVM_FLAG_SUPERVISOR_MODE) {
507 		if (!ecap_srs(iommu->ecap) || mm) {
508 			pr_err("Supervisor PASID with user provided mm.\n");
509 			return -EINVAL;
510 		}
511 	}
512 
513 	if (!(flags & SVM_FLAG_PRIVATE_PASID)) {
514 		struct intel_svm *t;
515 
516 		list_for_each_entry(t, &global_svm_list, list) {
517 			if (t->mm != mm || (t->flags & SVM_FLAG_PRIVATE_PASID))
518 				continue;
519 
520 			svm = t;
521 			if (svm->pasid >= pasid_max) {
522 				dev_warn(dev,
523 					 "Limited PASID width. Cannot use existing PASID %d\n",
524 					 svm->pasid);
525 				ret = -ENOSPC;
526 				goto out;
527 			}
528 
529 			/* Find the matching device in svm list */
530 			for_each_svm_dev(sdev, svm, dev) {
531 				if (sdev->ops != ops) {
532 					ret = -EBUSY;
533 					goto out;
534 				}
535 				sdev->users++;
536 				goto success;
537 			}
538 
539 			break;
540 		}
541 	}
542 
543 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
544 	if (!sdev) {
545 		ret = -ENOMEM;
546 		goto out;
547 	}
548 	sdev->dev = dev;
549 
550 	ret = intel_iommu_enable_pasid(iommu, dev);
551 	if (ret) {
552 		kfree(sdev);
553 		goto out;
554 	}
555 
556 	info = get_domain_info(dev);
557 	sdev->did = FLPT_DEFAULT_DID;
558 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
559 	if (info->ats_enabled) {
560 		sdev->dev_iotlb = 1;
561 		sdev->qdep = info->ats_qdep;
562 		if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
563 			sdev->qdep = 0;
564 	}
565 
566 	/* Finish the setup now we know we're keeping it */
567 	sdev->users = 1;
568 	sdev->ops = ops;
569 	init_rcu_head(&sdev->rcu);
570 
571 	if (!svm) {
572 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
573 		if (!svm) {
574 			ret = -ENOMEM;
575 			kfree(sdev);
576 			goto out;
577 		}
578 		svm->iommu = iommu;
579 
580 		if (pasid_max > intel_pasid_max_id)
581 			pasid_max = intel_pasid_max_id;
582 
583 		/* Do not use PASID 0, reserved for RID to PASID */
584 		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
585 					  pasid_max - 1, svm);
586 		if (svm->pasid == INVALID_IOASID) {
587 			kfree(svm);
588 			kfree(sdev);
589 			ret = -ENOSPC;
590 			goto out;
591 		}
592 		svm->notifier.ops = &intel_mmuops;
593 		svm->mm = mm;
594 		svm->flags = flags;
595 		INIT_LIST_HEAD_RCU(&svm->devs);
596 		INIT_LIST_HEAD(&svm->list);
597 		ret = -ENOMEM;
598 		if (mm) {
599 			ret = mmu_notifier_register(&svm->notifier, mm);
600 			if (ret) {
601 				ioasid_put(svm->pasid);
602 				kfree(svm);
603 				kfree(sdev);
604 				goto out;
605 			}
606 		}
607 
608 		spin_lock(&iommu->lock);
609 		ret = intel_pasid_setup_first_level(iommu, dev,
610 				mm ? mm->pgd : init_mm.pgd,
611 				svm->pasid, FLPT_DEFAULT_DID,
612 				(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
613 				(cpu_feature_enabled(X86_FEATURE_LA57) ?
614 				 PASID_FLAG_FL5LP : 0));
615 		spin_unlock(&iommu->lock);
616 		if (ret) {
617 			if (mm)
618 				mmu_notifier_unregister(&svm->notifier, mm);
619 			ioasid_put(svm->pasid);
620 			kfree(svm);
621 			kfree(sdev);
622 			goto out;
623 		}
624 
625 		list_add_tail(&svm->list, &global_svm_list);
626 		if (mm) {
627 			/* The newly allocated pasid is loaded to the mm. */
628 			load_pasid(mm, svm->pasid);
629 		}
630 	} else {
631 		/*
632 		 * Binding a new device with existing PASID, need to setup
633 		 * the PASID entry.
634 		 */
635 		spin_lock(&iommu->lock);
636 		ret = intel_pasid_setup_first_level(iommu, dev,
637 						mm ? mm->pgd : init_mm.pgd,
638 						svm->pasid, FLPT_DEFAULT_DID,
639 						(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
640 						(cpu_feature_enabled(X86_FEATURE_LA57) ?
641 						PASID_FLAG_FL5LP : 0));
642 		spin_unlock(&iommu->lock);
643 		if (ret) {
644 			kfree(sdev);
645 			goto out;
646 		}
647 	}
648 	list_add_rcu(&sdev->list, &svm->devs);
649 success:
650 	sdev->pasid = svm->pasid;
651 	sdev->sva.dev = dev;
652 	if (sd)
653 		*sd = sdev;
654 	ret = 0;
655 out:
656 	return ret;
657 }
658 
659 /* Caller must hold pasid_mutex */
660 static int intel_svm_unbind_mm(struct device *dev, u32 pasid)
661 {
662 	struct intel_svm_dev *sdev;
663 	struct intel_iommu *iommu;
664 	struct intel_svm *svm;
665 	int ret = -EINVAL;
666 
667 	iommu = device_to_iommu(dev, NULL, NULL);
668 	if (!iommu)
669 		goto out;
670 
671 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
672 	if (ret)
673 		goto out;
674 
675 	if (sdev) {
676 		sdev->users--;
677 		if (!sdev->users) {
678 			list_del_rcu(&sdev->list);
679 			/* Flush the PASID cache and IOTLB for this device.
680 			 * Note that we do depend on the hardware *not* using
681 			 * the PASID any more. Just as we depend on other
682 			 * devices never using PASIDs that they have no right
683 			 * to use. We have a *shared* PASID table, because it's
684 			 * large and has to be physically contiguous. So it's
685 			 * hard to be as defensive as we might like. */
686 			intel_pasid_tear_down_entry(iommu, dev,
687 						    svm->pasid, false);
688 			intel_svm_drain_prq(dev, svm->pasid);
689 			kfree_rcu(sdev, rcu);
690 
691 			if (list_empty(&svm->devs)) {
692 				ioasid_put(svm->pasid);
693 				if (svm->mm) {
694 					mmu_notifier_unregister(&svm->notifier, svm->mm);
695 					/* Clear mm's pasid. */
696 					load_pasid(svm->mm, PASID_DISABLED);
697 				}
698 				list_del(&svm->list);
699 				/* We mandate that no page faults may be outstanding
700 				 * for the PASID when intel_svm_unbind_mm() is called.
701 				 * If that is not obeyed, subtle errors will happen.
702 				 * Let's make them less subtle... */
703 				memset(svm, 0x6b, sizeof(*svm));
704 				kfree(svm);
705 			}
706 		}
707 	}
708 out:
709 	return ret;
710 }
711 
712 /* Page request queue descriptor */
713 struct page_req_dsc {
714 	union {
715 		struct {
716 			u64 type:8;
717 			u64 pasid_present:1;
718 			u64 priv_data_present:1;
719 			u64 rsvd:6;
720 			u64 rid:16;
721 			u64 pasid:20;
722 			u64 exe_req:1;
723 			u64 pm_req:1;
724 			u64 rsvd2:10;
725 		};
726 		u64 qw_0;
727 	};
728 	union {
729 		struct {
730 			u64 rd_req:1;
731 			u64 wr_req:1;
732 			u64 lpig:1;
733 			u64 prg_index:9;
734 			u64 addr:52;
735 		};
736 		u64 qw_1;
737 	};
738 	u64 priv_data[2];
739 };
740 
741 #define PRQ_RING_MASK	((0x1000 << PRQ_ORDER) - 0x20)
742 
743 static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
744 {
745 	unsigned long requested = 0;
746 
747 	if (req->exe_req)
748 		requested |= VM_EXEC;
749 
750 	if (req->rd_req)
751 		requested |= VM_READ;
752 
753 	if (req->wr_req)
754 		requested |= VM_WRITE;
755 
756 	return (requested & ~vma->vm_flags) != 0;
757 }
758 
759 static bool is_canonical_address(u64 addr)
760 {
761 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
762 	long saddr = (long) addr;
763 
764 	return (((saddr << shift) >> shift) == saddr);
765 }
766 
767 /**
768  * intel_svm_drain_prq - Drain page requests and responses for a pasid
769  * @dev: target device
770  * @pasid: pasid for draining
771  *
772  * Drain all pending page requests and responses related to @pasid in both
773  * software and hardware. This is supposed to be called after the device
774  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
775  * and DevTLB have been invalidated.
776  *
777  * It waits until all pending page requests for @pasid in the page fault
778  * queue are completed by the prq handling thread. Then follow the steps
779  * described in VT-d spec CH7.10 to drain all page requests and page
780  * responses pending in the hardware.
781  */
782 static void intel_svm_drain_prq(struct device *dev, u32 pasid)
783 {
784 	struct device_domain_info *info;
785 	struct dmar_domain *domain;
786 	struct intel_iommu *iommu;
787 	struct qi_desc desc[3];
788 	struct pci_dev *pdev;
789 	int head, tail;
790 	u16 sid, did;
791 	int qdep;
792 
793 	info = get_domain_info(dev);
794 	if (WARN_ON(!info || !dev_is_pci(dev)))
795 		return;
796 
797 	if (!info->pri_enabled)
798 		return;
799 
800 	iommu = info->iommu;
801 	domain = info->domain;
802 	pdev = to_pci_dev(dev);
803 	sid = PCI_DEVID(info->bus, info->devfn);
804 	did = domain->iommu_did[iommu->seq_id];
805 	qdep = pci_ats_queue_depth(pdev);
806 
807 	/*
808 	 * Check and wait until all pending page requests in the queue are
809 	 * handled by the prq handling thread.
810 	 */
811 prq_retry:
812 	reinit_completion(&iommu->prq_complete);
813 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
814 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
815 	while (head != tail) {
816 		struct page_req_dsc *req;
817 
818 		req = &iommu->prq[head / sizeof(*req)];
819 		if (!req->pasid_present || req->pasid != pasid) {
820 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
821 			continue;
822 		}
823 
824 		wait_for_completion(&iommu->prq_complete);
825 		goto prq_retry;
826 	}
827 
828 	/*
829 	 * Perform steps described in VT-d spec CH7.10 to drain page
830 	 * requests and responses in hardware.
831 	 */
832 	memset(desc, 0, sizeof(desc));
833 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
834 			QI_IWD_FENCE |
835 			QI_IWD_TYPE;
836 	desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
837 			QI_EIOTLB_DID(did) |
838 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
839 			QI_EIOTLB_TYPE;
840 	desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
841 			QI_DEV_EIOTLB_SID(sid) |
842 			QI_DEV_EIOTLB_QDEP(qdep) |
843 			QI_DEIOTLB_TYPE |
844 			QI_DEV_IOTLB_PFSID(info->pfsid);
845 qi_retry:
846 	reinit_completion(&iommu->prq_complete);
847 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
848 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
849 		wait_for_completion(&iommu->prq_complete);
850 		goto qi_retry;
851 	}
852 }
853 
854 static int prq_to_iommu_prot(struct page_req_dsc *req)
855 {
856 	int prot = 0;
857 
858 	if (req->rd_req)
859 		prot |= IOMMU_FAULT_PERM_READ;
860 	if (req->wr_req)
861 		prot |= IOMMU_FAULT_PERM_WRITE;
862 	if (req->exe_req)
863 		prot |= IOMMU_FAULT_PERM_EXEC;
864 	if (req->pm_req)
865 		prot |= IOMMU_FAULT_PERM_PRIV;
866 
867 	return prot;
868 }
869 
870 static int
871 intel_svm_prq_report(struct device *dev, struct page_req_dsc *desc)
872 {
873 	struct iommu_fault_event event;
874 
875 	if (!dev || !dev_is_pci(dev))
876 		return -ENODEV;
877 
878 	/* Fill in event data for device specific processing */
879 	memset(&event, 0, sizeof(struct iommu_fault_event));
880 	event.fault.type = IOMMU_FAULT_PAGE_REQ;
881 	event.fault.prm.addr = desc->addr;
882 	event.fault.prm.pasid = desc->pasid;
883 	event.fault.prm.grpid = desc->prg_index;
884 	event.fault.prm.perm = prq_to_iommu_prot(desc);
885 
886 	if (desc->lpig)
887 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
888 	if (desc->pasid_present) {
889 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
890 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
891 	}
892 	if (desc->priv_data_present) {
893 		/*
894 		 * Set last page in group bit if private data is present,
895 		 * page response is required as it does for LPIG.
896 		 * iommu_report_device_fault() doesn't understand this vendor
897 		 * specific requirement thus we set last_page as a workaround.
898 		 */
899 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
900 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
901 		memcpy(event.fault.prm.private_data, desc->priv_data,
902 		       sizeof(desc->priv_data));
903 	}
904 
905 	return iommu_report_device_fault(dev, &event);
906 }
907 
908 static irqreturn_t prq_event_thread(int irq, void *d)
909 {
910 	struct intel_svm_dev *sdev = NULL;
911 	struct intel_iommu *iommu = d;
912 	struct intel_svm *svm = NULL;
913 	int head, tail, handled = 0;
914 
915 	/* Clear PPR bit before reading head/tail registers, to
916 	 * ensure that we get a new interrupt if needed. */
917 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
918 
919 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
920 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
921 	while (head != tail) {
922 		struct vm_area_struct *vma;
923 		struct page_req_dsc *req;
924 		struct qi_desc resp;
925 		int result;
926 		vm_fault_t ret;
927 		u64 address;
928 
929 		handled = 1;
930 
931 		req = &iommu->prq[head / sizeof(*req)];
932 
933 		result = QI_RESP_FAILURE;
934 		address = (u64)req->addr << VTD_PAGE_SHIFT;
935 		if (!req->pasid_present) {
936 			pr_err("%s: Page request without PASID: %08llx %08llx\n",
937 			       iommu->name, ((unsigned long long *)req)[0],
938 			       ((unsigned long long *)req)[1]);
939 			goto no_pasid;
940 		}
941 
942 		if (!svm || svm->pasid != req->pasid) {
943 			rcu_read_lock();
944 			svm = ioasid_find(NULL, req->pasid, NULL);
945 			/* It *can't* go away, because the driver is not permitted
946 			 * to unbind the mm while any page faults are outstanding.
947 			 * So we only need RCU to protect the internal idr code. */
948 			rcu_read_unlock();
949 			if (IS_ERR_OR_NULL(svm)) {
950 				pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
951 				       iommu->name, req->pasid, ((unsigned long long *)req)[0],
952 				       ((unsigned long long *)req)[1]);
953 				goto no_pasid;
954 			}
955 		}
956 
957 		if (!sdev || sdev->sid != req->rid) {
958 			struct intel_svm_dev *t;
959 
960 			sdev = NULL;
961 			rcu_read_lock();
962 			list_for_each_entry_rcu(t, &svm->devs, list) {
963 				if (t->sid == req->rid) {
964 					sdev = t;
965 					break;
966 				}
967 			}
968 			rcu_read_unlock();
969 		}
970 
971 		result = QI_RESP_INVALID;
972 		/* Since we're using init_mm.pgd directly, we should never take
973 		 * any faults on kernel addresses. */
974 		if (!svm->mm)
975 			goto bad_req;
976 
977 		/* If address is not canonical, return invalid response */
978 		if (!is_canonical_address(address))
979 			goto bad_req;
980 
981 		/*
982 		 * If prq is to be handled outside iommu driver via receiver of
983 		 * the fault notifiers, we skip the page response here.
984 		 */
985 		if (svm->flags & SVM_FLAG_GUEST_MODE) {
986 			if (sdev && !intel_svm_prq_report(sdev->dev, req))
987 				goto prq_advance;
988 			else
989 				goto bad_req;
990 		}
991 
992 		/* If the mm is already defunct, don't handle faults. */
993 		if (!mmget_not_zero(svm->mm))
994 			goto bad_req;
995 
996 		mmap_read_lock(svm->mm);
997 		vma = find_extend_vma(svm->mm, address);
998 		if (!vma || address < vma->vm_start)
999 			goto invalid;
1000 
1001 		if (access_error(vma, req))
1002 			goto invalid;
1003 
1004 		ret = handle_mm_fault(vma, address,
1005 				      req->wr_req ? FAULT_FLAG_WRITE : 0,
1006 				      NULL);
1007 		if (ret & VM_FAULT_ERROR)
1008 			goto invalid;
1009 
1010 		result = QI_RESP_SUCCESS;
1011 invalid:
1012 		mmap_read_unlock(svm->mm);
1013 		mmput(svm->mm);
1014 bad_req:
1015 		WARN_ON(!sdev);
1016 		if (sdev && sdev->ops && sdev->ops->fault_cb) {
1017 			int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
1018 				(req->exe_req << 1) | (req->pm_req);
1019 			sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr,
1020 					    req->priv_data, rwxp, result);
1021 		}
1022 		/* We get here in the error case where the PASID lookup failed,
1023 		   and these can be NULL. Do not use them below this point! */
1024 		sdev = NULL;
1025 		svm = NULL;
1026 no_pasid:
1027 		if (req->lpig || req->priv_data_present) {
1028 			/*
1029 			 * Per VT-d spec. v3.0 ch7.7, system software must
1030 			 * respond with page group response if private data
1031 			 * is present (PDP) or last page in group (LPIG) bit
1032 			 * is set. This is an additional VT-d feature beyond
1033 			 * PCI ATS spec.
1034 			 */
1035 			resp.qw0 = QI_PGRP_PASID(req->pasid) |
1036 				QI_PGRP_DID(req->rid) |
1037 				QI_PGRP_PASID_P(req->pasid_present) |
1038 				QI_PGRP_PDP(req->priv_data_present) |
1039 				QI_PGRP_RESP_CODE(result) |
1040 				QI_PGRP_RESP_TYPE;
1041 			resp.qw1 = QI_PGRP_IDX(req->prg_index) |
1042 				QI_PGRP_LPIG(req->lpig);
1043 
1044 			if (req->priv_data_present)
1045 				memcpy(&resp.qw2, req->priv_data,
1046 				       sizeof(req->priv_data));
1047 			resp.qw2 = 0;
1048 			resp.qw3 = 0;
1049 			qi_submit_sync(iommu, &resp, 1, 0);
1050 		}
1051 prq_advance:
1052 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
1053 	}
1054 
1055 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
1056 
1057 	/*
1058 	 * Clear the page request overflow bit and wake up all threads that
1059 	 * are waiting for the completion of this handling.
1060 	 */
1061 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO)
1062 		writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
1063 
1064 	if (!completion_done(&iommu->prq_complete))
1065 		complete(&iommu->prq_complete);
1066 
1067 	return IRQ_RETVAL(handled);
1068 }
1069 
1070 #define to_intel_svm_dev(handle) container_of(handle, struct intel_svm_dev, sva)
1071 struct iommu_sva *
1072 intel_svm_bind(struct device *dev, struct mm_struct *mm, void *drvdata)
1073 {
1074 	struct iommu_sva *sva = ERR_PTR(-EINVAL);
1075 	struct intel_svm_dev *sdev = NULL;
1076 	unsigned int flags = 0;
1077 	int ret;
1078 
1079 	/*
1080 	 * TODO: Consolidate with generic iommu-sva bind after it is merged.
1081 	 * It will require shared SVM data structures, i.e. combine io_mm
1082 	 * and intel_svm etc.
1083 	 */
1084 	if (drvdata)
1085 		flags = *(unsigned int *)drvdata;
1086 	mutex_lock(&pasid_mutex);
1087 	ret = intel_svm_bind_mm(dev, flags, NULL, mm, &sdev);
1088 	if (ret)
1089 		sva = ERR_PTR(ret);
1090 	else if (sdev)
1091 		sva = &sdev->sva;
1092 	else
1093 		WARN(!sdev, "SVM bind succeeded with no sdev!\n");
1094 
1095 	mutex_unlock(&pasid_mutex);
1096 
1097 	return sva;
1098 }
1099 
1100 void intel_svm_unbind(struct iommu_sva *sva)
1101 {
1102 	struct intel_svm_dev *sdev;
1103 
1104 	mutex_lock(&pasid_mutex);
1105 	sdev = to_intel_svm_dev(sva);
1106 	intel_svm_unbind_mm(sdev->dev, sdev->pasid);
1107 	mutex_unlock(&pasid_mutex);
1108 }
1109 
1110 u32 intel_svm_get_pasid(struct iommu_sva *sva)
1111 {
1112 	struct intel_svm_dev *sdev;
1113 	u32 pasid;
1114 
1115 	mutex_lock(&pasid_mutex);
1116 	sdev = to_intel_svm_dev(sva);
1117 	pasid = sdev->pasid;
1118 	mutex_unlock(&pasid_mutex);
1119 
1120 	return pasid;
1121 }
1122 
1123 int intel_svm_page_response(struct device *dev,
1124 			    struct iommu_fault_event *evt,
1125 			    struct iommu_page_response *msg)
1126 {
1127 	struct iommu_fault_page_request *prm;
1128 	struct intel_svm_dev *sdev = NULL;
1129 	struct intel_svm *svm = NULL;
1130 	struct intel_iommu *iommu;
1131 	bool private_present;
1132 	bool pasid_present;
1133 	bool last_page;
1134 	u8 bus, devfn;
1135 	int ret = 0;
1136 	u16 sid;
1137 
1138 	if (!dev || !dev_is_pci(dev))
1139 		return -ENODEV;
1140 
1141 	iommu = device_to_iommu(dev, &bus, &devfn);
1142 	if (!iommu)
1143 		return -ENODEV;
1144 
1145 	if (!msg || !evt)
1146 		return -EINVAL;
1147 
1148 	mutex_lock(&pasid_mutex);
1149 
1150 	prm = &evt->fault.prm;
1151 	sid = PCI_DEVID(bus, devfn);
1152 	pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
1153 	private_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
1154 	last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
1155 
1156 	if (!pasid_present) {
1157 		ret = -EINVAL;
1158 		goto out;
1159 	}
1160 
1161 	if (prm->pasid == 0 || prm->pasid >= PASID_MAX) {
1162 		ret = -EINVAL;
1163 		goto out;
1164 	}
1165 
1166 	ret = pasid_to_svm_sdev(dev, prm->pasid, &svm, &sdev);
1167 	if (ret || !sdev) {
1168 		ret = -ENODEV;
1169 		goto out;
1170 	}
1171 
1172 	/*
1173 	 * For responses from userspace, need to make sure that the
1174 	 * pasid has been bound to its mm.
1175 	 */
1176 	if (svm->flags & SVM_FLAG_GUEST_MODE) {
1177 		struct mm_struct *mm;
1178 
1179 		mm = get_task_mm(current);
1180 		if (!mm) {
1181 			ret = -EINVAL;
1182 			goto out;
1183 		}
1184 
1185 		if (mm != svm->mm) {
1186 			ret = -ENODEV;
1187 			mmput(mm);
1188 			goto out;
1189 		}
1190 
1191 		mmput(mm);
1192 	}
1193 
1194 	/*
1195 	 * Per VT-d spec. v3.0 ch7.7, system software must respond
1196 	 * with page group response if private data is present (PDP)
1197 	 * or last page in group (LPIG) bit is set. This is an
1198 	 * additional VT-d requirement beyond PCI ATS spec.
1199 	 */
1200 	if (last_page || private_present) {
1201 		struct qi_desc desc;
1202 
1203 		desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
1204 				QI_PGRP_PASID_P(pasid_present) |
1205 				QI_PGRP_PDP(private_present) |
1206 				QI_PGRP_RESP_CODE(msg->code) |
1207 				QI_PGRP_RESP_TYPE;
1208 		desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
1209 		desc.qw2 = 0;
1210 		desc.qw3 = 0;
1211 		if (private_present)
1212 			memcpy(&desc.qw2, prm->private_data,
1213 			       sizeof(prm->private_data));
1214 
1215 		qi_submit_sync(iommu, &desc, 1, 0);
1216 	}
1217 out:
1218 	mutex_unlock(&pasid_mutex);
1219 	return ret;
1220 }
1221