xref: /openbmc/linux/drivers/iommu/intel/svm.c (revision 27ab1c1c)
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 __flush_svm_range_dev(struct intel_svm *svm,
122 				  struct intel_svm_dev *sdev,
123 				  unsigned long address,
124 				  unsigned long pages, int ih)
125 {
126 	struct qi_desc desc;
127 
128 	if (pages == -1) {
129 		desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
130 			QI_EIOTLB_DID(sdev->did) |
131 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
132 			QI_EIOTLB_TYPE;
133 		desc.qw1 = 0;
134 	} else {
135 		int mask = ilog2(__roundup_pow_of_two(pages));
136 
137 		desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
138 				QI_EIOTLB_DID(sdev->did) |
139 				QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID) |
140 				QI_EIOTLB_TYPE;
141 		desc.qw1 = QI_EIOTLB_ADDR(address) |
142 				QI_EIOTLB_IH(ih) |
143 				QI_EIOTLB_AM(mask);
144 	}
145 	desc.qw2 = 0;
146 	desc.qw3 = 0;
147 	qi_submit_sync(sdev->iommu, &desc, 1, 0);
148 
149 	if (sdev->dev_iotlb) {
150 		desc.qw0 = QI_DEV_EIOTLB_PASID(svm->pasid) |
151 				QI_DEV_EIOTLB_SID(sdev->sid) |
152 				QI_DEV_EIOTLB_QDEP(sdev->qdep) |
153 				QI_DEIOTLB_TYPE;
154 		if (pages == -1) {
155 			desc.qw1 = QI_DEV_EIOTLB_ADDR(-1ULL >> 1) |
156 					QI_DEV_EIOTLB_SIZE;
157 		} else if (pages > 1) {
158 			/* The least significant zero bit indicates the size. So,
159 			 * for example, an "address" value of 0x12345f000 will
160 			 * flush from 0x123440000 to 0x12347ffff (256KiB). */
161 			unsigned long last = address + ((unsigned long)(pages - 1) << VTD_PAGE_SHIFT);
162 			unsigned long mask = __rounddown_pow_of_two(address ^ last);
163 
164 			desc.qw1 = QI_DEV_EIOTLB_ADDR((address & ~mask) |
165 					(mask - 1)) | QI_DEV_EIOTLB_SIZE;
166 		} else {
167 			desc.qw1 = QI_DEV_EIOTLB_ADDR(address);
168 		}
169 		desc.qw2 = 0;
170 		desc.qw3 = 0;
171 		qi_submit_sync(sdev->iommu, &desc, 1, 0);
172 	}
173 }
174 
175 static void intel_flush_svm_range_dev(struct intel_svm *svm,
176 				      struct intel_svm_dev *sdev,
177 				      unsigned long address,
178 				      unsigned long pages, int ih)
179 {
180 	unsigned long shift = ilog2(__roundup_pow_of_two(pages));
181 	unsigned long align = (1ULL << (VTD_PAGE_SHIFT + shift));
182 	unsigned long start = ALIGN_DOWN(address, align);
183 	unsigned long end = ALIGN(address + (pages << VTD_PAGE_SHIFT), align);
184 
185 	while (start < end) {
186 		__flush_svm_range_dev(svm, sdev, start, align >> VTD_PAGE_SHIFT, ih);
187 		start += align;
188 	}
189 }
190 
191 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
192 				unsigned long pages, int ih)
193 {
194 	struct intel_svm_dev *sdev;
195 
196 	rcu_read_lock();
197 	list_for_each_entry_rcu(sdev, &svm->devs, list)
198 		intel_flush_svm_range_dev(svm, sdev, address, pages, ih);
199 	rcu_read_unlock();
200 }
201 
202 /* Pages have been freed at this point */
203 static void intel_invalidate_range(struct mmu_notifier *mn,
204 				   struct mm_struct *mm,
205 				   unsigned long start, unsigned long end)
206 {
207 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
208 
209 	intel_flush_svm_range(svm, start,
210 			      (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0);
211 }
212 
213 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
214 {
215 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
216 	struct intel_svm_dev *sdev;
217 
218 	/* This might end up being called from exit_mmap(), *before* the page
219 	 * tables are cleared. And __mmu_notifier_release() will delete us from
220 	 * the list of notifiers so that our invalidate_range() callback doesn't
221 	 * get called when the page tables are cleared. So we need to protect
222 	 * against hardware accessing those page tables.
223 	 *
224 	 * We do it by clearing the entry in the PASID table and then flushing
225 	 * the IOTLB and the PASID table caches. This might upset hardware;
226 	 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
227 	 * page) so that we end up taking a fault that the hardware really
228 	 * *has* to handle gracefully without affecting other processes.
229 	 */
230 	rcu_read_lock();
231 	list_for_each_entry_rcu(sdev, &svm->devs, list)
232 		intel_pasid_tear_down_entry(sdev->iommu, sdev->dev,
233 					    svm->pasid, true);
234 	rcu_read_unlock();
235 
236 }
237 
238 static const struct mmu_notifier_ops intel_mmuops = {
239 	.release = intel_mm_release,
240 	.invalidate_range = intel_invalidate_range,
241 };
242 
243 static DEFINE_MUTEX(pasid_mutex);
244 static LIST_HEAD(global_svm_list);
245 
246 #define for_each_svm_dev(sdev, svm, d)			\
247 	list_for_each_entry((sdev), &(svm)->devs, list)	\
248 		if ((d) != (sdev)->dev) {} else
249 
250 static int pasid_to_svm_sdev(struct device *dev, unsigned int pasid,
251 			     struct intel_svm **rsvm,
252 			     struct intel_svm_dev **rsdev)
253 {
254 	struct intel_svm_dev *d, *sdev = NULL;
255 	struct intel_svm *svm;
256 
257 	/* The caller should hold the pasid_mutex lock */
258 	if (WARN_ON(!mutex_is_locked(&pasid_mutex)))
259 		return -EINVAL;
260 
261 	if (pasid == INVALID_IOASID || pasid >= PASID_MAX)
262 		return -EINVAL;
263 
264 	svm = ioasid_find(NULL, pasid, NULL);
265 	if (IS_ERR(svm))
266 		return PTR_ERR(svm);
267 
268 	if (!svm)
269 		goto out;
270 
271 	/*
272 	 * If we found svm for the PASID, there must be at least one device
273 	 * bond.
274 	 */
275 	if (WARN_ON(list_empty(&svm->devs)))
276 		return -EINVAL;
277 
278 	rcu_read_lock();
279 	list_for_each_entry_rcu(d, &svm->devs, list) {
280 		if (d->dev == dev) {
281 			sdev = d;
282 			break;
283 		}
284 	}
285 	rcu_read_unlock();
286 
287 out:
288 	*rsvm = svm;
289 	*rsdev = sdev;
290 
291 	return 0;
292 }
293 
294 int intel_svm_bind_gpasid(struct iommu_domain *domain, struct device *dev,
295 			  struct iommu_gpasid_bind_data *data)
296 {
297 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
298 	struct intel_svm_dev *sdev = NULL;
299 	struct dmar_domain *dmar_domain;
300 	struct device_domain_info *info;
301 	struct intel_svm *svm = NULL;
302 	unsigned long iflags;
303 	int ret = 0;
304 
305 	if (WARN_ON(!iommu) || !data)
306 		return -EINVAL;
307 
308 	if (data->format != IOMMU_PASID_FORMAT_INTEL_VTD)
309 		return -EINVAL;
310 
311 	/* IOMMU core ensures argsz is more than the start of the union */
312 	if (data->argsz < offsetofend(struct iommu_gpasid_bind_data, vendor.vtd))
313 		return -EINVAL;
314 
315 	/* Make sure no undefined flags are used in vendor data */
316 	if (data->vendor.vtd.flags & ~(IOMMU_SVA_VTD_GPASID_LAST - 1))
317 		return -EINVAL;
318 
319 	if (!dev_is_pci(dev))
320 		return -ENOTSUPP;
321 
322 	/* VT-d supports devices with full 20 bit PASIDs only */
323 	if (pci_max_pasids(to_pci_dev(dev)) != PASID_MAX)
324 		return -EINVAL;
325 
326 	/*
327 	 * We only check host PASID range, we have no knowledge to check
328 	 * guest PASID range.
329 	 */
330 	if (data->hpasid <= 0 || data->hpasid >= PASID_MAX)
331 		return -EINVAL;
332 
333 	info = get_domain_info(dev);
334 	if (!info)
335 		return -EINVAL;
336 
337 	dmar_domain = to_dmar_domain(domain);
338 
339 	mutex_lock(&pasid_mutex);
340 	ret = pasid_to_svm_sdev(dev, data->hpasid, &svm, &sdev);
341 	if (ret)
342 		goto out;
343 
344 	if (sdev) {
345 		/*
346 		 * Do not allow multiple bindings of the same device-PASID since
347 		 * there is only one SL page tables per PASID. We may revisit
348 		 * once sharing PGD across domains are supported.
349 		 */
350 		dev_warn_ratelimited(dev, "Already bound with PASID %u\n",
351 				     svm->pasid);
352 		ret = -EBUSY;
353 		goto out;
354 	}
355 
356 	if (!svm) {
357 		/* We come here when PASID has never been bond to a device. */
358 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
359 		if (!svm) {
360 			ret = -ENOMEM;
361 			goto out;
362 		}
363 		/* REVISIT: upper layer/VFIO can track host process that bind
364 		 * the PASID. ioasid_set = mm might be sufficient for vfio to
365 		 * check pasid VMM ownership. We can drop the following line
366 		 * once VFIO and IOASID set check is in place.
367 		 */
368 		svm->mm = get_task_mm(current);
369 		svm->pasid = data->hpasid;
370 		if (data->flags & IOMMU_SVA_GPASID_VAL) {
371 			svm->gpasid = data->gpasid;
372 			svm->flags |= SVM_FLAG_GUEST_PASID;
373 		}
374 		ioasid_set_data(data->hpasid, svm);
375 		INIT_LIST_HEAD_RCU(&svm->devs);
376 		mmput(svm->mm);
377 	}
378 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
379 	if (!sdev) {
380 		ret = -ENOMEM;
381 		goto out;
382 	}
383 	sdev->dev = dev;
384 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
385 	sdev->iommu = iommu;
386 
387 	/* Only count users if device has aux domains */
388 	if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
389 		sdev->users = 1;
390 
391 	/* Set up device context entry for PASID if not enabled already */
392 	ret = intel_iommu_enable_pasid(iommu, sdev->dev);
393 	if (ret) {
394 		dev_err_ratelimited(dev, "Failed to enable PASID capability\n");
395 		kfree(sdev);
396 		goto out;
397 	}
398 
399 	/*
400 	 * PASID table is per device for better security. Therefore, for
401 	 * each bind of a new device even with an existing PASID, we need to
402 	 * call the nested mode setup function here.
403 	 */
404 	spin_lock_irqsave(&iommu->lock, iflags);
405 	ret = intel_pasid_setup_nested(iommu, dev,
406 				       (pgd_t *)(uintptr_t)data->gpgd,
407 				       data->hpasid, &data->vendor.vtd, dmar_domain,
408 				       data->addr_width);
409 	spin_unlock_irqrestore(&iommu->lock, iflags);
410 	if (ret) {
411 		dev_err_ratelimited(dev, "Failed to set up PASID %llu in nested mode, Err %d\n",
412 				    data->hpasid, ret);
413 		/*
414 		 * PASID entry should be in cleared state if nested mode
415 		 * set up failed. So we only need to clear IOASID tracking
416 		 * data such that free call will succeed.
417 		 */
418 		kfree(sdev);
419 		goto out;
420 	}
421 
422 	svm->flags |= SVM_FLAG_GUEST_MODE;
423 
424 	init_rcu_head(&sdev->rcu);
425 	list_add_rcu(&sdev->list, &svm->devs);
426  out:
427 	if (!IS_ERR_OR_NULL(svm) && list_empty(&svm->devs)) {
428 		ioasid_set_data(data->hpasid, NULL);
429 		kfree(svm);
430 	}
431 
432 	mutex_unlock(&pasid_mutex);
433 	return ret;
434 }
435 
436 int intel_svm_unbind_gpasid(struct device *dev, u32 pasid)
437 {
438 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
439 	struct intel_svm_dev *sdev;
440 	struct intel_svm *svm;
441 	int ret;
442 
443 	if (WARN_ON(!iommu))
444 		return -EINVAL;
445 
446 	mutex_lock(&pasid_mutex);
447 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
448 	if (ret)
449 		goto out;
450 
451 	if (sdev) {
452 		if (iommu_dev_feature_enabled(dev, IOMMU_DEV_FEAT_AUX))
453 			sdev->users--;
454 		if (!sdev->users) {
455 			list_del_rcu(&sdev->list);
456 			intel_pasid_tear_down_entry(iommu, dev,
457 						    svm->pasid, false);
458 			intel_svm_drain_prq(dev, svm->pasid);
459 			kfree_rcu(sdev, rcu);
460 
461 			if (list_empty(&svm->devs)) {
462 				/*
463 				 * We do not free the IOASID here in that
464 				 * IOMMU driver did not allocate it.
465 				 * Unlike native SVM, IOASID for guest use was
466 				 * allocated prior to the bind call.
467 				 * In any case, if the free call comes before
468 				 * the unbind, IOMMU driver will get notified
469 				 * and perform cleanup.
470 				 */
471 				ioasid_set_data(pasid, NULL);
472 				kfree(svm);
473 			}
474 		}
475 	}
476 out:
477 	mutex_unlock(&pasid_mutex);
478 	return ret;
479 }
480 
481 static void _load_pasid(void *unused)
482 {
483 	update_pasid();
484 }
485 
486 static void load_pasid(struct mm_struct *mm, u32 pasid)
487 {
488 	mutex_lock(&mm->context.lock);
489 
490 	/* Synchronize with READ_ONCE in update_pasid(). */
491 	smp_store_release(&mm->pasid, pasid);
492 
493 	/* Update PASID MSR on all CPUs running the mm's tasks. */
494 	on_each_cpu_mask(mm_cpumask(mm), _load_pasid, NULL, true);
495 
496 	mutex_unlock(&mm->context.lock);
497 }
498 
499 /* Caller must hold pasid_mutex, mm reference */
500 static int
501 intel_svm_bind_mm(struct device *dev, unsigned int flags,
502 		  struct svm_dev_ops *ops,
503 		  struct mm_struct *mm, struct intel_svm_dev **sd)
504 {
505 	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
506 	struct device_domain_info *info;
507 	struct intel_svm_dev *sdev;
508 	struct intel_svm *svm = NULL;
509 	unsigned long iflags;
510 	int pasid_max;
511 	int ret;
512 
513 	if (!iommu || dmar_disabled)
514 		return -EINVAL;
515 
516 	if (!intel_svm_capable(iommu))
517 		return -ENOTSUPP;
518 
519 	if (dev_is_pci(dev)) {
520 		pasid_max = pci_max_pasids(to_pci_dev(dev));
521 		if (pasid_max < 0)
522 			return -EINVAL;
523 	} else
524 		pasid_max = 1 << 20;
525 
526 	/* Bind supervisor PASID shuld have mm = NULL */
527 	if (flags & SVM_FLAG_SUPERVISOR_MODE) {
528 		if (!ecap_srs(iommu->ecap) || mm) {
529 			pr_err("Supervisor PASID with user provided mm.\n");
530 			return -EINVAL;
531 		}
532 	}
533 
534 	if (!(flags & SVM_FLAG_PRIVATE_PASID)) {
535 		struct intel_svm *t;
536 
537 		list_for_each_entry(t, &global_svm_list, list) {
538 			if (t->mm != mm || (t->flags & SVM_FLAG_PRIVATE_PASID))
539 				continue;
540 
541 			svm = t;
542 			if (svm->pasid >= pasid_max) {
543 				dev_warn(dev,
544 					 "Limited PASID width. Cannot use existing PASID %d\n",
545 					 svm->pasid);
546 				ret = -ENOSPC;
547 				goto out;
548 			}
549 
550 			/* Find the matching device in svm list */
551 			for_each_svm_dev(sdev, svm, dev) {
552 				if (sdev->ops != ops) {
553 					ret = -EBUSY;
554 					goto out;
555 				}
556 				sdev->users++;
557 				goto success;
558 			}
559 
560 			break;
561 		}
562 	}
563 
564 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
565 	if (!sdev) {
566 		ret = -ENOMEM;
567 		goto out;
568 	}
569 	sdev->dev = dev;
570 	sdev->iommu = iommu;
571 
572 	ret = intel_iommu_enable_pasid(iommu, dev);
573 	if (ret) {
574 		kfree(sdev);
575 		goto out;
576 	}
577 
578 	info = get_domain_info(dev);
579 	sdev->did = FLPT_DEFAULT_DID;
580 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
581 	if (info->ats_enabled) {
582 		sdev->dev_iotlb = 1;
583 		sdev->qdep = info->ats_qdep;
584 		if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
585 			sdev->qdep = 0;
586 	}
587 
588 	/* Finish the setup now we know we're keeping it */
589 	sdev->users = 1;
590 	sdev->ops = ops;
591 	init_rcu_head(&sdev->rcu);
592 
593 	if (!svm) {
594 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
595 		if (!svm) {
596 			ret = -ENOMEM;
597 			kfree(sdev);
598 			goto out;
599 		}
600 
601 		if (pasid_max > intel_pasid_max_id)
602 			pasid_max = intel_pasid_max_id;
603 
604 		/* Do not use PASID 0, reserved for RID to PASID */
605 		svm->pasid = ioasid_alloc(NULL, PASID_MIN,
606 					  pasid_max - 1, svm);
607 		if (svm->pasid == INVALID_IOASID) {
608 			kfree(svm);
609 			kfree(sdev);
610 			ret = -ENOSPC;
611 			goto out;
612 		}
613 		svm->notifier.ops = &intel_mmuops;
614 		svm->mm = mm;
615 		svm->flags = flags;
616 		INIT_LIST_HEAD_RCU(&svm->devs);
617 		INIT_LIST_HEAD(&svm->list);
618 		ret = -ENOMEM;
619 		if (mm) {
620 			ret = mmu_notifier_register(&svm->notifier, mm);
621 			if (ret) {
622 				ioasid_put(svm->pasid);
623 				kfree(svm);
624 				kfree(sdev);
625 				goto out;
626 			}
627 		}
628 
629 		spin_lock_irqsave(&iommu->lock, iflags);
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_irqrestore(&iommu->lock, iflags);
637 		if (ret) {
638 			if (mm)
639 				mmu_notifier_unregister(&svm->notifier, mm);
640 			ioasid_put(svm->pasid);
641 			kfree(svm);
642 			kfree(sdev);
643 			goto out;
644 		}
645 
646 		list_add_tail(&svm->list, &global_svm_list);
647 		if (mm) {
648 			/* The newly allocated pasid is loaded to the mm. */
649 			load_pasid(mm, svm->pasid);
650 		}
651 	} else {
652 		/*
653 		 * Binding a new device with existing PASID, need to setup
654 		 * the PASID entry.
655 		 */
656 		spin_lock_irqsave(&iommu->lock, iflags);
657 		ret = intel_pasid_setup_first_level(iommu, dev,
658 						mm ? mm->pgd : init_mm.pgd,
659 						svm->pasid, FLPT_DEFAULT_DID,
660 						(mm ? 0 : PASID_FLAG_SUPERVISOR_MODE) |
661 						(cpu_feature_enabled(X86_FEATURE_LA57) ?
662 						PASID_FLAG_FL5LP : 0));
663 		spin_unlock_irqrestore(&iommu->lock, iflags);
664 		if (ret) {
665 			kfree(sdev);
666 			goto out;
667 		}
668 	}
669 	list_add_rcu(&sdev->list, &svm->devs);
670 success:
671 	sdev->pasid = svm->pasid;
672 	sdev->sva.dev = dev;
673 	if (sd)
674 		*sd = sdev;
675 	ret = 0;
676 out:
677 	return ret;
678 }
679 
680 /* Caller must hold pasid_mutex */
681 static int intel_svm_unbind_mm(struct device *dev, u32 pasid)
682 {
683 	struct intel_svm_dev *sdev;
684 	struct intel_iommu *iommu;
685 	struct intel_svm *svm;
686 	int ret = -EINVAL;
687 
688 	iommu = device_to_iommu(dev, NULL, NULL);
689 	if (!iommu)
690 		goto out;
691 
692 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
693 	if (ret)
694 		goto out;
695 
696 	if (sdev) {
697 		sdev->users--;
698 		if (!sdev->users) {
699 			list_del_rcu(&sdev->list);
700 			/* Flush the PASID cache and IOTLB for this device.
701 			 * Note that we do depend on the hardware *not* using
702 			 * the PASID any more. Just as we depend on other
703 			 * devices never using PASIDs that they have no right
704 			 * to use. We have a *shared* PASID table, because it's
705 			 * large and has to be physically contiguous. So it's
706 			 * hard to be as defensive as we might like. */
707 			intel_pasid_tear_down_entry(iommu, dev,
708 						    svm->pasid, false);
709 			intel_svm_drain_prq(dev, svm->pasid);
710 			kfree_rcu(sdev, rcu);
711 
712 			if (list_empty(&svm->devs)) {
713 				ioasid_put(svm->pasid);
714 				if (svm->mm) {
715 					mmu_notifier_unregister(&svm->notifier, svm->mm);
716 					/* Clear mm's pasid. */
717 					load_pasid(svm->mm, PASID_DISABLED);
718 				}
719 				list_del(&svm->list);
720 				/* We mandate that no page faults may be outstanding
721 				 * for the PASID when intel_svm_unbind_mm() is called.
722 				 * If that is not obeyed, subtle errors will happen.
723 				 * Let's make them less subtle... */
724 				memset(svm, 0x6b, sizeof(*svm));
725 				kfree(svm);
726 			}
727 		}
728 	}
729 out:
730 	return ret;
731 }
732 
733 /* Page request queue descriptor */
734 struct page_req_dsc {
735 	union {
736 		struct {
737 			u64 type:8;
738 			u64 pasid_present:1;
739 			u64 priv_data_present:1;
740 			u64 rsvd:6;
741 			u64 rid:16;
742 			u64 pasid:20;
743 			u64 exe_req:1;
744 			u64 pm_req:1;
745 			u64 rsvd2:10;
746 		};
747 		u64 qw_0;
748 	};
749 	union {
750 		struct {
751 			u64 rd_req:1;
752 			u64 wr_req:1;
753 			u64 lpig:1;
754 			u64 prg_index:9;
755 			u64 addr:52;
756 		};
757 		u64 qw_1;
758 	};
759 	u64 priv_data[2];
760 };
761 
762 #define PRQ_RING_MASK	((0x1000 << PRQ_ORDER) - 0x20)
763 
764 static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
765 {
766 	unsigned long requested = 0;
767 
768 	if (req->exe_req)
769 		requested |= VM_EXEC;
770 
771 	if (req->rd_req)
772 		requested |= VM_READ;
773 
774 	if (req->wr_req)
775 		requested |= VM_WRITE;
776 
777 	return (requested & ~vma->vm_flags) != 0;
778 }
779 
780 static bool is_canonical_address(u64 addr)
781 {
782 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
783 	long saddr = (long) addr;
784 
785 	return (((saddr << shift) >> shift) == saddr);
786 }
787 
788 /**
789  * intel_svm_drain_prq - Drain page requests and responses for a pasid
790  * @dev: target device
791  * @pasid: pasid for draining
792  *
793  * Drain all pending page requests and responses related to @pasid in both
794  * software and hardware. This is supposed to be called after the device
795  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
796  * and DevTLB have been invalidated.
797  *
798  * It waits until all pending page requests for @pasid in the page fault
799  * queue are completed by the prq handling thread. Then follow the steps
800  * described in VT-d spec CH7.10 to drain all page requests and page
801  * responses pending in the hardware.
802  */
803 static void intel_svm_drain_prq(struct device *dev, u32 pasid)
804 {
805 	struct device_domain_info *info;
806 	struct dmar_domain *domain;
807 	struct intel_iommu *iommu;
808 	struct qi_desc desc[3];
809 	struct pci_dev *pdev;
810 	int head, tail;
811 	u16 sid, did;
812 	int qdep;
813 
814 	info = get_domain_info(dev);
815 	if (WARN_ON(!info || !dev_is_pci(dev)))
816 		return;
817 
818 	if (!info->pri_enabled)
819 		return;
820 
821 	iommu = info->iommu;
822 	domain = info->domain;
823 	pdev = to_pci_dev(dev);
824 	sid = PCI_DEVID(info->bus, info->devfn);
825 	did = domain->iommu_did[iommu->seq_id];
826 	qdep = pci_ats_queue_depth(pdev);
827 
828 	/*
829 	 * Check and wait until all pending page requests in the queue are
830 	 * handled by the prq handling thread.
831 	 */
832 prq_retry:
833 	reinit_completion(&iommu->prq_complete);
834 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
835 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
836 	while (head != tail) {
837 		struct page_req_dsc *req;
838 
839 		req = &iommu->prq[head / sizeof(*req)];
840 		if (!req->pasid_present || req->pasid != pasid) {
841 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
842 			continue;
843 		}
844 
845 		wait_for_completion(&iommu->prq_complete);
846 		goto prq_retry;
847 	}
848 
849 	/*
850 	 * Perform steps described in VT-d spec CH7.10 to drain page
851 	 * requests and responses in hardware.
852 	 */
853 	memset(desc, 0, sizeof(desc));
854 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
855 			QI_IWD_FENCE |
856 			QI_IWD_TYPE;
857 	desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
858 			QI_EIOTLB_DID(did) |
859 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
860 			QI_EIOTLB_TYPE;
861 	desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
862 			QI_DEV_EIOTLB_SID(sid) |
863 			QI_DEV_EIOTLB_QDEP(qdep) |
864 			QI_DEIOTLB_TYPE |
865 			QI_DEV_IOTLB_PFSID(info->pfsid);
866 qi_retry:
867 	reinit_completion(&iommu->prq_complete);
868 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
869 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
870 		wait_for_completion(&iommu->prq_complete);
871 		goto qi_retry;
872 	}
873 }
874 
875 static int prq_to_iommu_prot(struct page_req_dsc *req)
876 {
877 	int prot = 0;
878 
879 	if (req->rd_req)
880 		prot |= IOMMU_FAULT_PERM_READ;
881 	if (req->wr_req)
882 		prot |= IOMMU_FAULT_PERM_WRITE;
883 	if (req->exe_req)
884 		prot |= IOMMU_FAULT_PERM_EXEC;
885 	if (req->pm_req)
886 		prot |= IOMMU_FAULT_PERM_PRIV;
887 
888 	return prot;
889 }
890 
891 static int
892 intel_svm_prq_report(struct device *dev, struct page_req_dsc *desc)
893 {
894 	struct iommu_fault_event event;
895 
896 	if (!dev || !dev_is_pci(dev))
897 		return -ENODEV;
898 
899 	/* Fill in event data for device specific processing */
900 	memset(&event, 0, sizeof(struct iommu_fault_event));
901 	event.fault.type = IOMMU_FAULT_PAGE_REQ;
902 	event.fault.prm.addr = desc->addr;
903 	event.fault.prm.pasid = desc->pasid;
904 	event.fault.prm.grpid = desc->prg_index;
905 	event.fault.prm.perm = prq_to_iommu_prot(desc);
906 
907 	if (desc->lpig)
908 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
909 	if (desc->pasid_present) {
910 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
911 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
912 	}
913 	if (desc->priv_data_present) {
914 		/*
915 		 * Set last page in group bit if private data is present,
916 		 * page response is required as it does for LPIG.
917 		 * iommu_report_device_fault() doesn't understand this vendor
918 		 * specific requirement thus we set last_page as a workaround.
919 		 */
920 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
921 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
922 		memcpy(event.fault.prm.private_data, desc->priv_data,
923 		       sizeof(desc->priv_data));
924 	}
925 
926 	return iommu_report_device_fault(dev, &event);
927 }
928 
929 static irqreturn_t prq_event_thread(int irq, void *d)
930 {
931 	struct intel_svm_dev *sdev = NULL;
932 	struct intel_iommu *iommu = d;
933 	struct intel_svm *svm = NULL;
934 	int head, tail, handled = 0;
935 
936 	/* Clear PPR bit before reading head/tail registers, to
937 	 * ensure that we get a new interrupt if needed. */
938 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
939 
940 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
941 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
942 	while (head != tail) {
943 		struct vm_area_struct *vma;
944 		struct page_req_dsc *req;
945 		struct qi_desc resp;
946 		int result;
947 		vm_fault_t ret;
948 		u64 address;
949 
950 		handled = 1;
951 
952 		req = &iommu->prq[head / sizeof(*req)];
953 
954 		result = QI_RESP_FAILURE;
955 		address = (u64)req->addr << VTD_PAGE_SHIFT;
956 		if (!req->pasid_present) {
957 			pr_err("%s: Page request without PASID: %08llx %08llx\n",
958 			       iommu->name, ((unsigned long long *)req)[0],
959 			       ((unsigned long long *)req)[1]);
960 			goto no_pasid;
961 		}
962 
963 		if (!svm || svm->pasid != req->pasid) {
964 			rcu_read_lock();
965 			svm = ioasid_find(NULL, req->pasid, NULL);
966 			/* It *can't* go away, because the driver is not permitted
967 			 * to unbind the mm while any page faults are outstanding.
968 			 * So we only need RCU to protect the internal idr code. */
969 			rcu_read_unlock();
970 			if (IS_ERR_OR_NULL(svm)) {
971 				pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
972 				       iommu->name, req->pasid, ((unsigned long long *)req)[0],
973 				       ((unsigned long long *)req)[1]);
974 				goto no_pasid;
975 			}
976 		}
977 
978 		if (!sdev || sdev->sid != req->rid) {
979 			struct intel_svm_dev *t;
980 
981 			sdev = NULL;
982 			rcu_read_lock();
983 			list_for_each_entry_rcu(t, &svm->devs, list) {
984 				if (t->sid == req->rid) {
985 					sdev = t;
986 					break;
987 				}
988 			}
989 			rcu_read_unlock();
990 		}
991 
992 		result = QI_RESP_INVALID;
993 		/* Since we're using init_mm.pgd directly, we should never take
994 		 * any faults on kernel addresses. */
995 		if (!svm->mm)
996 			goto bad_req;
997 
998 		/* If address is not canonical, return invalid response */
999 		if (!is_canonical_address(address))
1000 			goto bad_req;
1001 
1002 		/*
1003 		 * If prq is to be handled outside iommu driver via receiver of
1004 		 * the fault notifiers, we skip the page response here.
1005 		 */
1006 		if (svm->flags & SVM_FLAG_GUEST_MODE) {
1007 			if (sdev && !intel_svm_prq_report(sdev->dev, req))
1008 				goto prq_advance;
1009 			else
1010 				goto bad_req;
1011 		}
1012 
1013 		/* If the mm is already defunct, don't handle faults. */
1014 		if (!mmget_not_zero(svm->mm))
1015 			goto bad_req;
1016 
1017 		mmap_read_lock(svm->mm);
1018 		vma = find_extend_vma(svm->mm, address);
1019 		if (!vma || address < vma->vm_start)
1020 			goto invalid;
1021 
1022 		if (access_error(vma, req))
1023 			goto invalid;
1024 
1025 		ret = handle_mm_fault(vma, address,
1026 				      req->wr_req ? FAULT_FLAG_WRITE : 0,
1027 				      NULL);
1028 		if (ret & VM_FAULT_ERROR)
1029 			goto invalid;
1030 
1031 		result = QI_RESP_SUCCESS;
1032 invalid:
1033 		mmap_read_unlock(svm->mm);
1034 		mmput(svm->mm);
1035 bad_req:
1036 		WARN_ON(!sdev);
1037 		if (sdev && sdev->ops && sdev->ops->fault_cb) {
1038 			int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
1039 				(req->exe_req << 1) | (req->pm_req);
1040 			sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr,
1041 					    req->priv_data, rwxp, result);
1042 		}
1043 		/* We get here in the error case where the PASID lookup failed,
1044 		   and these can be NULL. Do not use them below this point! */
1045 		sdev = NULL;
1046 		svm = NULL;
1047 no_pasid:
1048 		if (req->lpig || req->priv_data_present) {
1049 			/*
1050 			 * Per VT-d spec. v3.0 ch7.7, system software must
1051 			 * respond with page group response if private data
1052 			 * is present (PDP) or last page in group (LPIG) bit
1053 			 * is set. This is an additional VT-d feature beyond
1054 			 * PCI ATS spec.
1055 			 */
1056 			resp.qw0 = QI_PGRP_PASID(req->pasid) |
1057 				QI_PGRP_DID(req->rid) |
1058 				QI_PGRP_PASID_P(req->pasid_present) |
1059 				QI_PGRP_PDP(req->priv_data_present) |
1060 				QI_PGRP_RESP_CODE(result) |
1061 				QI_PGRP_RESP_TYPE;
1062 			resp.qw1 = QI_PGRP_IDX(req->prg_index) |
1063 				QI_PGRP_LPIG(req->lpig);
1064 
1065 			if (req->priv_data_present)
1066 				memcpy(&resp.qw2, req->priv_data,
1067 				       sizeof(req->priv_data));
1068 			resp.qw2 = 0;
1069 			resp.qw3 = 0;
1070 			qi_submit_sync(iommu, &resp, 1, 0);
1071 		}
1072 prq_advance:
1073 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
1074 	}
1075 
1076 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
1077 
1078 	/*
1079 	 * Clear the page request overflow bit and wake up all threads that
1080 	 * are waiting for the completion of this handling.
1081 	 */
1082 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO)
1083 		writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
1084 
1085 	if (!completion_done(&iommu->prq_complete))
1086 		complete(&iommu->prq_complete);
1087 
1088 	return IRQ_RETVAL(handled);
1089 }
1090 
1091 #define to_intel_svm_dev(handle) container_of(handle, struct intel_svm_dev, sva)
1092 struct iommu_sva *
1093 intel_svm_bind(struct device *dev, struct mm_struct *mm, void *drvdata)
1094 {
1095 	struct iommu_sva *sva = ERR_PTR(-EINVAL);
1096 	struct intel_svm_dev *sdev = NULL;
1097 	unsigned int flags = 0;
1098 	int ret;
1099 
1100 	/*
1101 	 * TODO: Consolidate with generic iommu-sva bind after it is merged.
1102 	 * It will require shared SVM data structures, i.e. combine io_mm
1103 	 * and intel_svm etc.
1104 	 */
1105 	if (drvdata)
1106 		flags = *(unsigned int *)drvdata;
1107 	mutex_lock(&pasid_mutex);
1108 	ret = intel_svm_bind_mm(dev, flags, NULL, mm, &sdev);
1109 	if (ret)
1110 		sva = ERR_PTR(ret);
1111 	else if (sdev)
1112 		sva = &sdev->sva;
1113 	else
1114 		WARN(!sdev, "SVM bind succeeded with no sdev!\n");
1115 
1116 	mutex_unlock(&pasid_mutex);
1117 
1118 	return sva;
1119 }
1120 
1121 void intel_svm_unbind(struct iommu_sva *sva)
1122 {
1123 	struct intel_svm_dev *sdev;
1124 
1125 	mutex_lock(&pasid_mutex);
1126 	sdev = to_intel_svm_dev(sva);
1127 	intel_svm_unbind_mm(sdev->dev, sdev->pasid);
1128 	mutex_unlock(&pasid_mutex);
1129 }
1130 
1131 u32 intel_svm_get_pasid(struct iommu_sva *sva)
1132 {
1133 	struct intel_svm_dev *sdev;
1134 	u32 pasid;
1135 
1136 	mutex_lock(&pasid_mutex);
1137 	sdev = to_intel_svm_dev(sva);
1138 	pasid = sdev->pasid;
1139 	mutex_unlock(&pasid_mutex);
1140 
1141 	return pasid;
1142 }
1143 
1144 int intel_svm_page_response(struct device *dev,
1145 			    struct iommu_fault_event *evt,
1146 			    struct iommu_page_response *msg)
1147 {
1148 	struct iommu_fault_page_request *prm;
1149 	struct intel_svm_dev *sdev = NULL;
1150 	struct intel_svm *svm = NULL;
1151 	struct intel_iommu *iommu;
1152 	bool private_present;
1153 	bool pasid_present;
1154 	bool last_page;
1155 	u8 bus, devfn;
1156 	int ret = 0;
1157 	u16 sid;
1158 
1159 	if (!dev || !dev_is_pci(dev))
1160 		return -ENODEV;
1161 
1162 	iommu = device_to_iommu(dev, &bus, &devfn);
1163 	if (!iommu)
1164 		return -ENODEV;
1165 
1166 	if (!msg || !evt)
1167 		return -EINVAL;
1168 
1169 	mutex_lock(&pasid_mutex);
1170 
1171 	prm = &evt->fault.prm;
1172 	sid = PCI_DEVID(bus, devfn);
1173 	pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
1174 	private_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
1175 	last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
1176 
1177 	if (!pasid_present) {
1178 		ret = -EINVAL;
1179 		goto out;
1180 	}
1181 
1182 	if (prm->pasid == 0 || prm->pasid >= PASID_MAX) {
1183 		ret = -EINVAL;
1184 		goto out;
1185 	}
1186 
1187 	ret = pasid_to_svm_sdev(dev, prm->pasid, &svm, &sdev);
1188 	if (ret || !sdev) {
1189 		ret = -ENODEV;
1190 		goto out;
1191 	}
1192 
1193 	/*
1194 	 * For responses from userspace, need to make sure that the
1195 	 * pasid has been bound to its mm.
1196 	 */
1197 	if (svm->flags & SVM_FLAG_GUEST_MODE) {
1198 		struct mm_struct *mm;
1199 
1200 		mm = get_task_mm(current);
1201 		if (!mm) {
1202 			ret = -EINVAL;
1203 			goto out;
1204 		}
1205 
1206 		if (mm != svm->mm) {
1207 			ret = -ENODEV;
1208 			mmput(mm);
1209 			goto out;
1210 		}
1211 
1212 		mmput(mm);
1213 	}
1214 
1215 	/*
1216 	 * Per VT-d spec. v3.0 ch7.7, system software must respond
1217 	 * with page group response if private data is present (PDP)
1218 	 * or last page in group (LPIG) bit is set. This is an
1219 	 * additional VT-d requirement beyond PCI ATS spec.
1220 	 */
1221 	if (last_page || private_present) {
1222 		struct qi_desc desc;
1223 
1224 		desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
1225 				QI_PGRP_PASID_P(pasid_present) |
1226 				QI_PGRP_PDP(private_present) |
1227 				QI_PGRP_RESP_CODE(msg->code) |
1228 				QI_PGRP_RESP_TYPE;
1229 		desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
1230 		desc.qw2 = 0;
1231 		desc.qw3 = 0;
1232 		if (private_present)
1233 			memcpy(&desc.qw2, prm->private_data,
1234 			       sizeof(prm->private_data));
1235 
1236 		qi_submit_sync(iommu, &desc, 1, 0);
1237 	}
1238 out:
1239 	mutex_unlock(&pasid_mutex);
1240 	return ret;
1241 }
1242