xref: /openbmc/linux/drivers/iommu/intel/svm.c (revision 6c9111bc)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright © 2015 Intel Corporation.
4  *
5  * Authors: David Woodhouse <dwmw2@infradead.org>
6  */
7 
8 #include <linux/intel-iommu.h>
9 #include <linux/mmu_notifier.h>
10 #include <linux/sched.h>
11 #include <linux/sched/mm.h>
12 #include <linux/slab.h>
13 #include <linux/intel-svm.h>
14 #include <linux/rculist.h>
15 #include <linux/pci.h>
16 #include <linux/pci-ats.h>
17 #include <linux/dmar.h>
18 #include <linux/interrupt.h>
19 #include <linux/mm_types.h>
20 #include <linux/ioasid.h>
21 #include <asm/page.h>
22 #include <asm/fpu/api.h>
23 
24 #include "pasid.h"
25 
26 static irqreturn_t prq_event_thread(int irq, void *d);
27 static void intel_svm_drain_prq(struct device *dev, u32 pasid);
28 
29 #define PRQ_ORDER 0
30 
31 int intel_svm_enable_prq(struct intel_iommu *iommu)
32 {
33 	struct page *pages;
34 	int irq, ret;
35 
36 	pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
37 	if (!pages) {
38 		pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
39 			iommu->name);
40 		return -ENOMEM;
41 	}
42 	iommu->prq = page_address(pages);
43 
44 	irq = dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED + iommu->seq_id, iommu->node, iommu);
45 	if (irq <= 0) {
46 		pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
47 		       iommu->name);
48 		ret = -EINVAL;
49 	err:
50 		free_pages((unsigned long)iommu->prq, PRQ_ORDER);
51 		iommu->prq = NULL;
52 		return ret;
53 	}
54 	iommu->pr_irq = irq;
55 
56 	snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
57 
58 	ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
59 				   iommu->prq_name, iommu);
60 	if (ret) {
61 		pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
62 		       iommu->name);
63 		dmar_free_hwirq(irq);
64 		iommu->pr_irq = 0;
65 		goto err;
66 	}
67 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
68 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
69 	dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
70 
71 	init_completion(&iommu->prq_complete);
72 
73 	return 0;
74 }
75 
76 int intel_svm_finish_prq(struct intel_iommu *iommu)
77 {
78 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
79 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
80 	dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
81 
82 	if (iommu->pr_irq) {
83 		free_irq(iommu->pr_irq, iommu);
84 		dmar_free_hwirq(iommu->pr_irq);
85 		iommu->pr_irq = 0;
86 	}
87 
88 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
89 	iommu->prq = NULL;
90 
91 	return 0;
92 }
93 
94 static inline bool intel_svm_capable(struct intel_iommu *iommu)
95 {
96 	return iommu->flags & VTD_FLAG_SVM_CAPABLE;
97 }
98 
99 void intel_svm_check(struct intel_iommu *iommu)
100 {
101 	if (!pasid_supported(iommu))
102 		return;
103 
104 	if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
105 	    !cap_fl1gp_support(iommu->cap)) {
106 		pr_err("%s SVM disabled, incompatible 1GB page capability\n",
107 		       iommu->name);
108 		return;
109 	}
110 
111 	if (cpu_feature_enabled(X86_FEATURE_LA57) &&
112 	    !cap_5lp_support(iommu->cap)) {
113 		pr_err("%s SVM disabled, incompatible paging mode\n",
114 		       iommu->name);
115 		return;
116 	}
117 
118 	iommu->flags |= VTD_FLAG_SVM_CAPABLE;
119 }
120 
121 static void intel_flush_svm_range_dev (struct intel_svm *svm, struct intel_svm_dev *sdev,
122 				unsigned long address, unsigned long pages, int ih)
123 {
124 	struct qi_desc desc;
125 
126 	if (pages == -1) {
127 		desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
128 			QI_EIOTLB_DID(sdev->did) |
129 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
130 			QI_EIOTLB_TYPE;
131 		desc.qw1 = 0;
132 	} else {
133 		int mask = ilog2(__roundup_pow_of_two(pages));
134 
135 		desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
136 				QI_EIOTLB_DID(sdev->did) |
137 				QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID) |
138 				QI_EIOTLB_TYPE;
139 		desc.qw1 = QI_EIOTLB_ADDR(address) |
140 				QI_EIOTLB_IH(ih) |
141 				QI_EIOTLB_AM(mask);
142 	}
143 	desc.qw2 = 0;
144 	desc.qw3 = 0;
145 	qi_submit_sync(svm->iommu, &desc, 1, 0);
146 
147 	if (sdev->dev_iotlb) {
148 		desc.qw0 = QI_DEV_EIOTLB_PASID(svm->pasid) |
149 				QI_DEV_EIOTLB_SID(sdev->sid) |
150 				QI_DEV_EIOTLB_QDEP(sdev->qdep) |
151 				QI_DEIOTLB_TYPE;
152 		if (pages == -1) {
153 			desc.qw1 = QI_DEV_EIOTLB_ADDR(-1ULL >> 1) |
154 					QI_DEV_EIOTLB_SIZE;
155 		} else if (pages > 1) {
156 			/* The least significant zero bit indicates the size. So,
157 			 * for example, an "address" value of 0x12345f000 will
158 			 * flush from 0x123440000 to 0x12347ffff (256KiB). */
159 			unsigned long last = address + ((unsigned long)(pages - 1) << VTD_PAGE_SHIFT);
160 			unsigned long mask = __rounddown_pow_of_two(address ^ last);
161 
162 			desc.qw1 = QI_DEV_EIOTLB_ADDR((address & ~mask) |
163 					(mask - 1)) | QI_DEV_EIOTLB_SIZE;
164 		} else {
165 			desc.qw1 = QI_DEV_EIOTLB_ADDR(address);
166 		}
167 		desc.qw2 = 0;
168 		desc.qw3 = 0;
169 		qi_submit_sync(svm->iommu, &desc, 1, 0);
170 	}
171 }
172 
173 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
174 				unsigned long pages, int ih)
175 {
176 	struct intel_svm_dev *sdev;
177 
178 	rcu_read_lock();
179 	list_for_each_entry_rcu(sdev, &svm->devs, list)
180 		intel_flush_svm_range_dev(svm, sdev, address, pages, ih);
181 	rcu_read_unlock();
182 }
183 
184 /* Pages have been freed at this point */
185 static void intel_invalidate_range(struct mmu_notifier *mn,
186 				   struct mm_struct *mm,
187 				   unsigned long start, unsigned long end)
188 {
189 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
190 
191 	intel_flush_svm_range(svm, start,
192 			      (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0);
193 }
194 
195 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
196 {
197 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
198 	struct intel_svm_dev *sdev;
199 
200 	/* This might end up being called from exit_mmap(), *before* the page
201 	 * tables are cleared. And __mmu_notifier_release() will delete us from
202 	 * the list of notifiers so that our invalidate_range() callback doesn't
203 	 * get called when the page tables are cleared. So we need to protect
204 	 * against hardware accessing those page tables.
205 	 *
206 	 * We do it by clearing the entry in the PASID table and then flushing
207 	 * the IOTLB and the PASID table caches. This might upset hardware;
208 	 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
209 	 * page) so that we end up taking a fault that the hardware really
210 	 * *has* to handle gracefully without affecting other processes.
211 	 */
212 	rcu_read_lock();
213 	list_for_each_entry_rcu(sdev, &svm->devs, list)
214 		intel_pasid_tear_down_entry(svm->iommu, sdev->dev,
215 					    svm->pasid, true);
216 	rcu_read_unlock();
217 
218 }
219 
220 static const struct mmu_notifier_ops intel_mmuops = {
221 	.release = intel_mm_release,
222 	.invalidate_range = intel_invalidate_range,
223 };
224 
225 static DEFINE_MUTEX(pasid_mutex);
226 static LIST_HEAD(global_svm_list);
227 
228 #define for_each_svm_dev(sdev, svm, d)			\
229 	list_for_each_entry((sdev), &(svm)->devs, list)	\
230 		if ((d) != (sdev)->dev) {} else
231 
232 static int pasid_to_svm_sdev(struct device *dev, unsigned int pasid,
233 			     struct intel_svm **rsvm,
234 			     struct intel_svm_dev **rsdev)
235 {
236 	struct intel_svm_dev *d, *sdev = NULL;
237 	struct intel_svm *svm;
238 
239 	/* The caller should hold the pasid_mutex lock */
240 	if (WARN_ON(!mutex_is_locked(&pasid_mutex)))
241 		return -EINVAL;
242 
243 	if (pasid == INVALID_IOASID || pasid >= PASID_MAX)
244 		return -EINVAL;
245 
246 	svm = ioasid_find(NULL, pasid, NULL);
247 	if (IS_ERR(svm))
248 		return PTR_ERR(svm);
249 
250 	if (!svm)
251 		goto out;
252 
253 	/*
254 	 * If we found svm for the PASID, there must be at least one device
255 	 * bond.
256 	 */
257 	if (WARN_ON(list_empty(&svm->devs)))
258 		return -EINVAL;
259 
260 	rcu_read_lock();
261 	list_for_each_entry_rcu(d, &svm->devs, list) {
262 		if (d->dev == dev) {
263 			sdev = d;
264 			break;
265 		}
266 	}
267 	rcu_read_unlock();
268 
269 out:
270 	*rsvm = svm;
271 	*rsdev = sdev;
272 
273 	return 0;
274 }
275 
276 int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
277 			  struct iommu_gpasid_bind_data *data)
278 {
279 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
280 	struct intel_svm_dev *sdev = NULL;
281 	struct dmar_domain *dmar_domain;
282 	struct intel_svm *svm = NULL;
283 	int ret = 0;
284 
285 	if (WARN_ON(!iommu) || !data)
286 		return -EINVAL;
287 
288 	if (data->format != IOMMU_PASID_FORMAT_INTEL_VTD)
289 		return -EINVAL;
290 
291 	/* IOMMU core ensures argsz is more than the start of the union */
292 	if (data->argsz < offsetofend(struct iommu_gpasid_bind_data, vendor.vtd))
293 		return -EINVAL;
294 
295 	/* Make sure no undefined flags are used in vendor data */
296 	if (data->vendor.vtd.flags & ~(IOMMU_SVA_VTD_GPASID_LAST - 1))
297 		return -EINVAL;
298 
299 	if (!dev_is_pci(dev))
300 		return -ENOTSUPP;
301 
302 	/* VT-d supports devices with full 20 bit PASIDs only */
303 	if (pci_max_pasids(to_pci_dev(dev)) != PASID_MAX)
304 		return -EINVAL;
305 
306 	/*
307 	 * We only check host PASID range, we have no knowledge to check
308 	 * guest PASID range.
309 	 */
310 	if (data->hpasid <= 0 || data->hpasid >= PASID_MAX)
311 		return -EINVAL;
312 
313 	dmar_domain = to_dmar_domain(domain);
314 
315 	mutex_lock(&pasid_mutex);
316 	ret = pasid_to_svm_sdev(dev, data->hpasid, &svm, &sdev);
317 	if (ret)
318 		goto out;
319 
320 	if (sdev) {
321 		/*
322 		 * Do not allow multiple bindings of the same device-PASID since
323 		 * there is only one SL page tables per PASID. We may revisit
324 		 * once sharing PGD across domains are supported.
325 		 */
326 		dev_warn_ratelimited(dev, "Already bound with PASID %u\n",
327 				     svm->pasid);
328 		ret = -EBUSY;
329 		goto out;
330 	}
331 
332 	if (!svm) {
333 		/* We come here when PASID has never been bond to a device. */
334 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
335 		if (!svm) {
336 			ret = -ENOMEM;
337 			goto out;
338 		}
339 		/* REVISIT: upper layer/VFIO can track host process that bind
340 		 * the PASID. ioasid_set = mm might be sufficient for vfio to
341 		 * check pasid VMM ownership. We can drop the following line
342 		 * once VFIO and IOASID set check is in place.
343 		 */
344 		svm->mm = get_task_mm(current);
345 		svm->pasid = data->hpasid;
346 		if (data->flags & IOMMU_SVA_GPASID_VAL) {
347 			svm->gpasid = data->gpasid;
348 			svm->flags |= SVM_FLAG_GUEST_PASID;
349 		}
350 		ioasid_set_data(data->hpasid, svm);
351 		INIT_LIST_HEAD_RCU(&svm->devs);
352 		mmput(svm->mm);
353 	}
354 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
355 	if (!sdev) {
356 		ret = -ENOMEM;
357 		goto out;
358 	}
359 	sdev->dev = dev;
360 
361 	/* Only count users if device has aux domains */
362 	if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
363 		sdev->users = 1;
364 
365 	/* Set up device context entry for PASID if not enabled already */
366 	ret = intel_iommu_enable_pasid(iommu, sdev->dev);
367 	if (ret) {
368 		dev_err_ratelimited(dev, "Failed to enable PASID capability\n");
369 		kfree(sdev);
370 		goto out;
371 	}
372 
373 	/*
374 	 * PASID table is per device for better security. Therefore, for
375 	 * each bind of a new device even with an existing PASID, we need to
376 	 * call the nested mode setup function here.
377 	 */
378 	spin_lock(&iommu->lock);
379 	ret = intel_pasid_setup_nested(iommu, dev,
380 				       (pgd_t *)(uintptr_t)data->gpgd,
381 				       data->hpasid, &data->vendor.vtd, dmar_domain,
382 				       data->addr_width);
383 	spin_unlock(&iommu->lock);
384 	if (ret) {
385 		dev_err_ratelimited(dev, "Failed to set up PASID %llu in nested mode, Err %d\n",
386 				    data->hpasid, ret);
387 		/*
388 		 * PASID entry should be in cleared state if nested mode
389 		 * set up failed. So we only need to clear IOASID tracking
390 		 * data such that free call will succeed.
391 		 */
392 		kfree(sdev);
393 		goto out;
394 	}
395 
396 	svm->flags |= SVM_FLAG_GUEST_MODE;
397 
398 	init_rcu_head(&sdev->rcu);
399 	list_add_rcu(&sdev->list, &svm->devs);
400  out:
401 	if (!IS_ERR_OR_NULL(svm) && list_empty(&svm->devs)) {
402 		ioasid_set_data(data->hpasid, NULL);
403 		kfree(svm);
404 	}
405 
406 	mutex_unlock(&pasid_mutex);
407 	return ret;
408 }
409 
410 int intel_svm_unbind_gpasid(struct device *dev, u32 pasid)
411 {
412 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
413 	struct intel_svm_dev *sdev;
414 	struct intel_svm *svm;
415 	int ret;
416 
417 	if (WARN_ON(!iommu))
418 		return -EINVAL;
419 
420 	mutex_lock(&pasid_mutex);
421 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
422 	if (ret)
423 		goto out;
424 
425 	if (sdev) {
426 		if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
427 			sdev->users--;
428 		if (!sdev->users) {
429 			list_del_rcu(&sdev->list);
430 			intel_pasid_tear_down_entry(iommu, dev,
431 						    svm->pasid, false);
432 			intel_svm_drain_prq(dev, svm->pasid);
433 			kfree_rcu(sdev, rcu);
434 
435 			if (list_empty(&svm->devs)) {
436 				/*
437 				 * We do not free the IOASID here in that
438 				 * IOMMU driver did not allocate it.
439 				 * Unlike native SVM, IOASID for guest use was
440 				 * allocated prior to the bind call.
441 				 * In any case, if the free call comes before
442 				 * the unbind, IOMMU driver will get notified
443 				 * and perform cleanup.
444 				 */
445 				ioasid_set_data(pasid, NULL);
446 				kfree(svm);
447 			}
448 		}
449 	}
450 out:
451 	mutex_unlock(&pasid_mutex);
452 	return ret;
453 }
454 
455 static void _load_pasid(void *unused)
456 {
457 	update_pasid();
458 }
459 
460 static void load_pasid(struct mm_struct *mm, u32 pasid)
461 {
462 	mutex_lock(&mm->context.lock);
463 
464 	/* Synchronize with READ_ONCE in update_pasid(). */
465 	smp_store_release(&mm->pasid, pasid);
466 
467 	/* Update PASID MSR on all CPUs running the mm's tasks. */
468 	on_each_cpu_mask(mm_cpumask(mm), _load_pasid, NULL, true);
469 
470 	mutex_unlock(&mm->context.lock);
471 }
472 
473 /* Caller must hold pasid_mutex, mm reference */
474 static int
475 intel_svm_bind_mm(struct device *dev, unsigned int flags,
476 		  struct svm_dev_ops *ops,
477 		  struct mm_struct *mm, struct intel_svm_dev **sd)
478 {
479 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
480 	struct device_domain_info *info;
481 	struct intel_svm_dev *sdev;
482 	struct intel_svm *svm = NULL;
483 	int pasid_max;
484 	int ret;
485 
486 	if (!iommu || dmar_disabled)
487 		return -EINVAL;
488 
489 	if (!intel_svm_capable(iommu))
490 		return -ENOTSUPP;
491 
492 	if (dev_is_pci(dev)) {
493 		pasid_max = pci_max_pasids(to_pci_dev(dev));
494 		if (pasid_max < 0)
495 			return -EINVAL;
496 	} else
497 		pasid_max = 1 << 20;
498 
499 	/* Bind supervisor PASID shuld have mm = NULL */
500 	if (flags & SVM_FLAG_SUPERVISOR_MODE) {
501 		if (!ecap_srs(iommu->ecap) || mm) {
502 			pr_err("Supervisor PASID with user provided mm.\n");
503 			return -EINVAL;
504 		}
505 	}
506 
507 	if (!(flags & SVM_FLAG_PRIVATE_PASID)) {
508 		struct intel_svm *t;
509 
510 		list_for_each_entry(t, &global_svm_list, list) {
511 			if (t->mm != mm || (t->flags & SVM_FLAG_PRIVATE_PASID))
512 				continue;
513 
514 			svm = t;
515 			if (svm->pasid >= pasid_max) {
516 				dev_warn(dev,
517 					 "Limited PASID width. Cannot use existing PASID %d\n",
518 					 svm->pasid);
519 				ret = -ENOSPC;
520 				goto out;
521 			}
522 
523 			/* Find the matching device in svm list */
524 			for_each_svm_dev(sdev, svm, dev) {
525 				if (sdev->ops != ops) {
526 					ret = -EBUSY;
527 					goto out;
528 				}
529 				sdev->users++;
530 				goto success;
531 			}
532 
533 			break;
534 		}
535 	}
536 
537 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
538 	if (!sdev) {
539 		ret = -ENOMEM;
540 		goto out;
541 	}
542 	sdev->dev = dev;
543 
544 	ret = intel_iommu_enable_pasid(iommu, dev);
545 	if (ret) {
546 		kfree(sdev);
547 		goto out;
548 	}
549 
550 	info = get_domain_info(dev);
551 	sdev->did = FLPT_DEFAULT_DID;
552 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
553 	if (info->ats_enabled) {
554 		sdev->dev_iotlb = 1;
555 		sdev->qdep = info->ats_qdep;
556 		if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
557 			sdev->qdep = 0;
558 	}
559 
560 	/* Finish the setup now we know we're keeping it */
561 	sdev->users = 1;
562 	sdev->ops = ops;
563 	init_rcu_head(&sdev->rcu);
564 
565 	if (!svm) {
566 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
567 		if (!svm) {
568 			ret = -ENOMEM;
569 			kfree(sdev);
570 			goto out;
571 		}
572 		svm->iommu = iommu;
573 
574 		if (pasid_max > intel_pasid_max_id)
575 			pasid_max = intel_pasid_max_id;
576 
577 		/* Do not use PASID 0, reserved for RID to PASID */
578 		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
579 					  pasid_max - 1, svm);
580 		if (svm->pasid == INVALID_IOASID) {
581 			kfree(svm);
582 			kfree(sdev);
583 			ret = -ENOSPC;
584 			goto out;
585 		}
586 		svm->notifier.ops = &intel_mmuops;
587 		svm->mm = mm;
588 		svm->flags = flags;
589 		INIT_LIST_HEAD_RCU(&svm->devs);
590 		INIT_LIST_HEAD(&svm->list);
591 		ret = -ENOMEM;
592 		if (mm) {
593 			ret = mmu_notifier_register(&svm->notifier, mm);
594 			if (ret) {
595 				ioasid_free(svm->pasid);
596 				kfree(svm);
597 				kfree(sdev);
598 				goto out;
599 			}
600 		}
601 
602 		spin_lock(&iommu->lock);
603 		ret = intel_pasid_setup_first_level(iommu, dev,
604 				mm ? mm->pgd : init_mm.pgd,
605 				svm->pasid, FLPT_DEFAULT_DID,
606 				(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
607 				(cpu_feature_enabled(X86_FEATURE_LA57) ?
608 				 PASID_FLAG_FL5LP : 0));
609 		spin_unlock(&iommu->lock);
610 		if (ret) {
611 			if (mm)
612 				mmu_notifier_unregister(&svm->notifier, mm);
613 			ioasid_free(svm->pasid);
614 			kfree(svm);
615 			kfree(sdev);
616 			goto out;
617 		}
618 
619 		list_add_tail(&svm->list, &global_svm_list);
620 		if (mm) {
621 			/* The newly allocated pasid is loaded to the mm. */
622 			load_pasid(mm, svm->pasid);
623 		}
624 	} else {
625 		/*
626 		 * Binding a new device with existing PASID, need to setup
627 		 * the PASID entry.
628 		 */
629 		spin_lock(&iommu->lock);
630 		ret = intel_pasid_setup_first_level(iommu, dev,
631 						mm ? mm->pgd : init_mm.pgd,
632 						svm->pasid, FLPT_DEFAULT_DID,
633 						(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
634 						(cpu_feature_enabled(X86_FEATURE_LA57) ?
635 						PASID_FLAG_FL5LP : 0));
636 		spin_unlock(&iommu->lock);
637 		if (ret) {
638 			kfree(sdev);
639 			goto out;
640 		}
641 	}
642 	list_add_rcu(&sdev->list, &svm->devs);
643 success:
644 	sdev->pasid = svm->pasid;
645 	sdev->sva.dev = dev;
646 	if (sd)
647 		*sd = sdev;
648 	ret = 0;
649 out:
650 	return ret;
651 }
652 
653 /* Caller must hold pasid_mutex */
654 static int intel_svm_unbind_mm(struct device *dev, u32 pasid)
655 {
656 	struct intel_svm_dev *sdev;
657 	struct intel_iommu *iommu;
658 	struct intel_svm *svm;
659 	int ret = -EINVAL;
660 
661 	iommu = device_to_iommu(dev, NULL, NULL);
662 	if (!iommu)
663 		goto out;
664 
665 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
666 	if (ret)
667 		goto out;
668 
669 	if (sdev) {
670 		sdev->users--;
671 		if (!sdev->users) {
672 			list_del_rcu(&sdev->list);
673 			/* Flush the PASID cache and IOTLB for this device.
674 			 * Note that we do depend on the hardware *not* using
675 			 * the PASID any more. Just as we depend on other
676 			 * devices never using PASIDs that they have no right
677 			 * to use. We have a *shared* PASID table, because it's
678 			 * large and has to be physically contiguous. So it's
679 			 * hard to be as defensive as we might like. */
680 			intel_pasid_tear_down_entry(iommu, dev,
681 						    svm->pasid, false);
682 			intel_svm_drain_prq(dev, svm->pasid);
683 			kfree_rcu(sdev, rcu);
684 
685 			if (list_empty(&svm->devs)) {
686 				ioasid_free(svm->pasid);
687 				if (svm->mm) {
688 					mmu_notifier_unregister(&svm->notifier, svm->mm);
689 					/* Clear mm's pasid. */
690 					load_pasid(svm->mm, PASID_DISABLED);
691 				}
692 				list_del(&svm->list);
693 				/* We mandate that no page faults may be outstanding
694 				 * for the PASID when intel_svm_unbind_mm() is called.
695 				 * If that is not obeyed, subtle errors will happen.
696 				 * Let's make them less subtle... */
697 				memset(svm, 0x6b, sizeof(*svm));
698 				kfree(svm);
699 			}
700 		}
701 	}
702 out:
703 	return ret;
704 }
705 
706 /* Page request queue descriptor */
707 struct page_req_dsc {
708 	union {
709 		struct {
710 			u64 type:8;
711 			u64 pasid_present:1;
712 			u64 priv_data_present:1;
713 			u64 rsvd:6;
714 			u64 rid:16;
715 			u64 pasid:20;
716 			u64 exe_req:1;
717 			u64 pm_req:1;
718 			u64 rsvd2:10;
719 		};
720 		u64 qw_0;
721 	};
722 	union {
723 		struct {
724 			u64 rd_req:1;
725 			u64 wr_req:1;
726 			u64 lpig:1;
727 			u64 prg_index:9;
728 			u64 addr:52;
729 		};
730 		u64 qw_1;
731 	};
732 	u64 priv_data[2];
733 };
734 
735 #define PRQ_RING_MASK	((0x1000 << PRQ_ORDER) - 0x20)
736 
737 static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
738 {
739 	unsigned long requested = 0;
740 
741 	if (req->exe_req)
742 		requested |= VM_EXEC;
743 
744 	if (req->rd_req)
745 		requested |= VM_READ;
746 
747 	if (req->wr_req)
748 		requested |= VM_WRITE;
749 
750 	return (requested & ~vma->vm_flags) != 0;
751 }
752 
753 static bool is_canonical_address(u64 addr)
754 {
755 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
756 	long saddr = (long) addr;
757 
758 	return (((saddr << shift) >> shift) == saddr);
759 }
760 
761 /**
762  * intel_svm_drain_prq - Drain page requests and responses for a pasid
763  * @dev: target device
764  * @pasid: pasid for draining
765  *
766  * Drain all pending page requests and responses related to @pasid in both
767  * software and hardware. This is supposed to be called after the device
768  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
769  * and DevTLB have been invalidated.
770  *
771  * It waits until all pending page requests for @pasid in the page fault
772  * queue are completed by the prq handling thread. Then follow the steps
773  * described in VT-d spec CH7.10 to drain all page requests and page
774  * responses pending in the hardware.
775  */
776 static void intel_svm_drain_prq(struct device *dev, u32 pasid)
777 {
778 	struct device_domain_info *info;
779 	struct dmar_domain *domain;
780 	struct intel_iommu *iommu;
781 	struct qi_desc desc[3];
782 	struct pci_dev *pdev;
783 	int head, tail;
784 	u16 sid, did;
785 	int qdep;
786 
787 	info = get_domain_info(dev);
788 	if (WARN_ON(!info || !dev_is_pci(dev)))
789 		return;
790 
791 	if (!info->pri_enabled)
792 		return;
793 
794 	iommu = info->iommu;
795 	domain = info->domain;
796 	pdev = to_pci_dev(dev);
797 	sid = PCI_DEVID(info->bus, info->devfn);
798 	did = domain->iommu_did[iommu->seq_id];
799 	qdep = pci_ats_queue_depth(pdev);
800 
801 	/*
802 	 * Check and wait until all pending page requests in the queue are
803 	 * handled by the prq handling thread.
804 	 */
805 prq_retry:
806 	reinit_completion(&iommu->prq_complete);
807 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
808 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
809 	while (head != tail) {
810 		struct page_req_dsc *req;
811 
812 		req = &iommu->prq[head / sizeof(*req)];
813 		if (!req->pasid_present || req->pasid != pasid) {
814 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
815 			continue;
816 		}
817 
818 		wait_for_completion(&iommu->prq_complete);
819 		goto prq_retry;
820 	}
821 
822 	/*
823 	 * Perform steps described in VT-d spec CH7.10 to drain page
824 	 * requests and responses in hardware.
825 	 */
826 	memset(desc, 0, sizeof(desc));
827 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
828 			QI_IWD_FENCE |
829 			QI_IWD_TYPE;
830 	desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
831 			QI_EIOTLB_DID(did) |
832 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
833 			QI_EIOTLB_TYPE;
834 	desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
835 			QI_DEV_EIOTLB_SID(sid) |
836 			QI_DEV_EIOTLB_QDEP(qdep) |
837 			QI_DEIOTLB_TYPE |
838 			QI_DEV_IOTLB_PFSID(info->pfsid);
839 qi_retry:
840 	reinit_completion(&iommu->prq_complete);
841 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
842 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
843 		wait_for_completion(&iommu->prq_complete);
844 		goto qi_retry;
845 	}
846 }
847 
848 static int prq_to_iommu_prot(struct page_req_dsc *req)
849 {
850 	int prot = 0;
851 
852 	if (req->rd_req)
853 		prot |= IOMMU_FAULT_PERM_READ;
854 	if (req->wr_req)
855 		prot |= IOMMU_FAULT_PERM_WRITE;
856 	if (req->exe_req)
857 		prot |= IOMMU_FAULT_PERM_EXEC;
858 	if (req->pm_req)
859 		prot |= IOMMU_FAULT_PERM_PRIV;
860 
861 	return prot;
862 }
863 
864 static int
865 intel_svm_prq_report(struct device *dev, struct page_req_dsc *desc)
866 {
867 	struct iommu_fault_event event;
868 
869 	if (!dev || !dev_is_pci(dev))
870 		return -ENODEV;
871 
872 	/* Fill in event data for device specific processing */
873 	memset(&event, 0, sizeof(struct iommu_fault_event));
874 	event.fault.type = IOMMU_FAULT_PAGE_REQ;
875 	event.fault.prm.addr = desc->addr;
876 	event.fault.prm.pasid = desc->pasid;
877 	event.fault.prm.grpid = desc->prg_index;
878 	event.fault.prm.perm = prq_to_iommu_prot(desc);
879 
880 	if (desc->lpig)
881 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
882 	if (desc->pasid_present) {
883 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
884 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
885 	}
886 	if (desc->priv_data_present) {
887 		/*
888 		 * Set last page in group bit if private data is present,
889 		 * page response is required as it does for LPIG.
890 		 * iommu_report_device_fault() doesn't understand this vendor
891 		 * specific requirement thus we set last_page as a workaround.
892 		 */
893 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
894 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
895 		memcpy(event.fault.prm.private_data, desc->priv_data,
896 		       sizeof(desc->priv_data));
897 	}
898 
899 	return iommu_report_device_fault(dev, &event);
900 }
901 
902 static irqreturn_t prq_event_thread(int irq, void *d)
903 {
904 	struct intel_svm_dev *sdev = NULL;
905 	struct intel_iommu *iommu = d;
906 	struct intel_svm *svm = NULL;
907 	int head, tail, handled = 0;
908 
909 	/* Clear PPR bit before reading head/tail registers, to
910 	 * ensure that we get a new interrupt if needed. */
911 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
912 
913 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
914 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
915 	while (head != tail) {
916 		struct vm_area_struct *vma;
917 		struct page_req_dsc *req;
918 		struct qi_desc resp;
919 		int result;
920 		vm_fault_t ret;
921 		u64 address;
922 
923 		handled = 1;
924 
925 		req = &iommu->prq[head / sizeof(*req)];
926 
927 		result = QI_RESP_FAILURE;
928 		address = (u64)req->addr << VTD_PAGE_SHIFT;
929 		if (!req->pasid_present) {
930 			pr_err("%s: Page request without PASID: %08llx %08llx\n",
931 			       iommu->name, ((unsigned long long *)req)[0],
932 			       ((unsigned long long *)req)[1]);
933 			goto no_pasid;
934 		}
935 
936 		if (!svm || svm->pasid != req->pasid) {
937 			rcu_read_lock();
938 			svm = ioasid_find(NULL, req->pasid, NULL);
939 			/* It *can't* go away, because the driver is not permitted
940 			 * to unbind the mm while any page faults are outstanding.
941 			 * So we only need RCU to protect the internal idr code. */
942 			rcu_read_unlock();
943 			if (IS_ERR_OR_NULL(svm)) {
944 				pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
945 				       iommu->name, req->pasid, ((unsigned long long *)req)[0],
946 				       ((unsigned long long *)req)[1]);
947 				goto no_pasid;
948 			}
949 		}
950 
951 		if (!sdev || sdev->sid != req->rid) {
952 			struct intel_svm_dev *t;
953 
954 			sdev = NULL;
955 			rcu_read_lock();
956 			list_for_each_entry_rcu(t, &svm->devs, list) {
957 				if (t->sid == req->rid) {
958 					sdev = t;
959 					break;
960 				}
961 			}
962 			rcu_read_unlock();
963 		}
964 
965 		result = QI_RESP_INVALID;
966 		/* Since we're using init_mm.pgd directly, we should never take
967 		 * any faults on kernel addresses. */
968 		if (!svm->mm)
969 			goto bad_req;
970 
971 		/* If address is not canonical, return invalid response */
972 		if (!is_canonical_address(address))
973 			goto bad_req;
974 
975 		/*
976 		 * If prq is to be handled outside iommu driver via receiver of
977 		 * the fault notifiers, we skip the page response here.
978 		 */
979 		if (svm->flags & SVM_FLAG_GUEST_MODE) {
980 			if (sdev && !intel_svm_prq_report(sdev->dev, req))
981 				goto prq_advance;
982 			else
983 				goto bad_req;
984 		}
985 
986 		/* If the mm is already defunct, don't handle faults. */
987 		if (!mmget_not_zero(svm->mm))
988 			goto bad_req;
989 
990 		mmap_read_lock(svm->mm);
991 		vma = find_extend_vma(svm->mm, address);
992 		if (!vma || address < vma->vm_start)
993 			goto invalid;
994 
995 		if (access_error(vma, req))
996 			goto invalid;
997 
998 		ret = handle_mm_fault(vma, address,
999 				      req->wr_req ? FAULT_FLAG_WRITE : 0,
1000 				      NULL);
1001 		if (ret & VM_FAULT_ERROR)
1002 			goto invalid;
1003 
1004 		result = QI_RESP_SUCCESS;
1005 invalid:
1006 		mmap_read_unlock(svm->mm);
1007 		mmput(svm->mm);
1008 bad_req:
1009 		WARN_ON(!sdev);
1010 		if (sdev && sdev->ops && sdev->ops->fault_cb) {
1011 			int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
1012 				(req->exe_req << 1) | (req->pm_req);
1013 			sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr,
1014 					    req->priv_data, rwxp, result);
1015 		}
1016 		/* We get here in the error case where the PASID lookup failed,
1017 		   and these can be NULL. Do not use them below this point! */
1018 		sdev = NULL;
1019 		svm = NULL;
1020 no_pasid:
1021 		if (req->lpig || req->priv_data_present) {
1022 			/*
1023 			 * Per VT-d spec. v3.0 ch7.7, system software must
1024 			 * respond with page group response if private data
1025 			 * is present (PDP) or last page in group (LPIG) bit
1026 			 * is set. This is an additional VT-d feature beyond
1027 			 * PCI ATS spec.
1028 			 */
1029 			resp.qw0 = QI_PGRP_PASID(req->pasid) |
1030 				QI_PGRP_DID(req->rid) |
1031 				QI_PGRP_PASID_P(req->pasid_present) |
1032 				QI_PGRP_PDP(req->pasid_present) |
1033 				QI_PGRP_RESP_CODE(result) |
1034 				QI_PGRP_RESP_TYPE;
1035 			resp.qw1 = QI_PGRP_IDX(req->prg_index) |
1036 				QI_PGRP_LPIG(req->lpig);
1037 
1038 			if (req->priv_data_present)
1039 				memcpy(&resp.qw2, req->priv_data,
1040 				       sizeof(req->priv_data));
1041 			resp.qw2 = 0;
1042 			resp.qw3 = 0;
1043 			qi_submit_sync(iommu, &resp, 1, 0);
1044 		}
1045 prq_advance:
1046 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
1047 	}
1048 
1049 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
1050 
1051 	/*
1052 	 * Clear the page request overflow bit and wake up all threads that
1053 	 * are waiting for the completion of this handling.
1054 	 */
1055 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO)
1056 		writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
1057 
1058 	if (!completion_done(&iommu->prq_complete))
1059 		complete(&iommu->prq_complete);
1060 
1061 	return IRQ_RETVAL(handled);
1062 }
1063 
1064 #define to_intel_svm_dev(handle) container_of(handle, struct intel_svm_dev, sva)
1065 struct iommu_sva *
1066 intel_svm_bind(struct device *dev, struct mm_struct *mm, void *drvdata)
1067 {
1068 	struct iommu_sva *sva = ERR_PTR(-EINVAL);
1069 	struct intel_svm_dev *sdev = NULL;
1070 	unsigned int flags = 0;
1071 	int ret;
1072 
1073 	/*
1074 	 * TODO: Consolidate with generic iommu-sva bind after it is merged.
1075 	 * It will require shared SVM data structures, i.e. combine io_mm
1076 	 * and intel_svm etc.
1077 	 */
1078 	if (drvdata)
1079 		flags = *(unsigned int *)drvdata;
1080 	mutex_lock(&pasid_mutex);
1081 	ret = intel_svm_bind_mm(dev, flags, NULL, mm, &sdev);
1082 	if (ret)
1083 		sva = ERR_PTR(ret);
1084 	else if (sdev)
1085 		sva = &sdev->sva;
1086 	else
1087 		WARN(!sdev, "SVM bind succeeded with no sdev!\n");
1088 
1089 	mutex_unlock(&pasid_mutex);
1090 
1091 	return sva;
1092 }
1093 
1094 void intel_svm_unbind(struct iommu_sva *sva)
1095 {
1096 	struct intel_svm_dev *sdev;
1097 
1098 	mutex_lock(&pasid_mutex);
1099 	sdev = to_intel_svm_dev(sva);
1100 	intel_svm_unbind_mm(sdev->dev, sdev->pasid);
1101 	mutex_unlock(&pasid_mutex);
1102 }
1103 
1104 u32 intel_svm_get_pasid(struct iommu_sva *sva)
1105 {
1106 	struct intel_svm_dev *sdev;
1107 	u32 pasid;
1108 
1109 	mutex_lock(&pasid_mutex);
1110 	sdev = to_intel_svm_dev(sva);
1111 	pasid = sdev->pasid;
1112 	mutex_unlock(&pasid_mutex);
1113 
1114 	return pasid;
1115 }
1116 
1117 int intel_svm_page_response(struct device *dev,
1118 			    struct iommu_fault_event *evt,
1119 			    struct iommu_page_response *msg)
1120 {
1121 	struct iommu_fault_page_request *prm;
1122 	struct intel_svm_dev *sdev = NULL;
1123 	struct intel_svm *svm = NULL;
1124 	struct intel_iommu *iommu;
1125 	bool private_present;
1126 	bool pasid_present;
1127 	bool last_page;
1128 	u8 bus, devfn;
1129 	int ret = 0;
1130 	u16 sid;
1131 
1132 	if (!dev || !dev_is_pci(dev))
1133 		return -ENODEV;
1134 
1135 	iommu = device_to_iommu(dev, &bus, &devfn);
1136 	if (!iommu)
1137 		return -ENODEV;
1138 
1139 	if (!msg || !evt)
1140 		return -EINVAL;
1141 
1142 	mutex_lock(&pasid_mutex);
1143 
1144 	prm = &evt->fault.prm;
1145 	sid = PCI_DEVID(bus, devfn);
1146 	pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
1147 	private_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
1148 	last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
1149 
1150 	if (!pasid_present) {
1151 		ret = -EINVAL;
1152 		goto out;
1153 	}
1154 
1155 	if (prm->pasid == 0 || prm->pasid >= PASID_MAX) {
1156 		ret = -EINVAL;
1157 		goto out;
1158 	}
1159 
1160 	ret = pasid_to_svm_sdev(dev, prm->pasid, &svm, &sdev);
1161 	if (ret || !sdev) {
1162 		ret = -ENODEV;
1163 		goto out;
1164 	}
1165 
1166 	/*
1167 	 * For responses from userspace, need to make sure that the
1168 	 * pasid has been bound to its mm.
1169 	 */
1170 	if (svm->flags & SVM_FLAG_GUEST_MODE) {
1171 		struct mm_struct *mm;
1172 
1173 		mm = get_task_mm(current);
1174 		if (!mm) {
1175 			ret = -EINVAL;
1176 			goto out;
1177 		}
1178 
1179 		if (mm != svm->mm) {
1180 			ret = -ENODEV;
1181 			mmput(mm);
1182 			goto out;
1183 		}
1184 
1185 		mmput(mm);
1186 	}
1187 
1188 	/*
1189 	 * Per VT-d spec. v3.0 ch7.7, system software must respond
1190 	 * with page group response if private data is present (PDP)
1191 	 * or last page in group (LPIG) bit is set. This is an
1192 	 * additional VT-d requirement beyond PCI ATS spec.
1193 	 */
1194 	if (last_page || private_present) {
1195 		struct qi_desc desc;
1196 
1197 		desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
1198 				QI_PGRP_PASID_P(pasid_present) |
1199 				QI_PGRP_PDP(private_present) |
1200 				QI_PGRP_RESP_CODE(msg->code) |
1201 				QI_PGRP_RESP_TYPE;
1202 		desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
1203 		desc.qw2 = 0;
1204 		desc.qw3 = 0;
1205 		if (private_present)
1206 			memcpy(&desc.qw2, prm->private_data,
1207 			       sizeof(prm->private_data));
1208 
1209 		qi_submit_sync(iommu, &desc, 1, 0);
1210 	}
1211 out:
1212 	mutex_unlock(&pasid_mutex);
1213 	return ret;
1214 }
1215