xref: /linux/drivers/iommu/intel/svm.c (revision ff9f3d7aefddbaa9a9b0f18f83e4319b5cd0e63e)
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/mmu_notifier.h>
9 #include <linux/sched.h>
10 #include <linux/sched/mm.h>
11 #include <linux/slab.h>
12 #include <linux/rculist.h>
13 #include <linux/pci.h>
14 #include <linux/pci-ats.h>
15 #include <linux/dmar.h>
16 #include <linux/interrupt.h>
17 #include <linux/mm_types.h>
18 #include <linux/xarray.h>
19 #include <linux/ioasid.h>
20 #include <asm/page.h>
21 #include <asm/fpu/api.h>
22 
23 #include "iommu.h"
24 #include "pasid.h"
25 #include "perf.h"
26 #include "../iommu-sva.h"
27 #include "trace.h"
28 
29 static irqreturn_t prq_event_thread(int irq, void *d);
30 static void intel_svm_drain_prq(struct device *dev, u32 pasid);
31 #define to_intel_svm_dev(handle) container_of(handle, struct intel_svm_dev, sva)
32 
33 static DEFINE_XARRAY_ALLOC(pasid_private_array);
34 static int pasid_private_add(ioasid_t pasid, void *priv)
35 {
36 	return xa_alloc(&pasid_private_array, &pasid, priv,
37 			XA_LIMIT(pasid, pasid), GFP_ATOMIC);
38 }
39 
40 static void pasid_private_remove(ioasid_t pasid)
41 {
42 	xa_erase(&pasid_private_array, pasid);
43 }
44 
45 static void *pasid_private_find(ioasid_t pasid)
46 {
47 	return xa_load(&pasid_private_array, pasid);
48 }
49 
50 static struct intel_svm_dev *
51 svm_lookup_device_by_dev(struct intel_svm *svm, struct device *dev)
52 {
53 	struct intel_svm_dev *sdev = NULL, *t;
54 
55 	rcu_read_lock();
56 	list_for_each_entry_rcu(t, &svm->devs, list) {
57 		if (t->dev == dev) {
58 			sdev = t;
59 			break;
60 		}
61 	}
62 	rcu_read_unlock();
63 
64 	return sdev;
65 }
66 
67 int intel_svm_enable_prq(struct intel_iommu *iommu)
68 {
69 	struct iopf_queue *iopfq;
70 	struct page *pages;
71 	int irq, ret;
72 
73 	pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
74 	if (!pages) {
75 		pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
76 			iommu->name);
77 		return -ENOMEM;
78 	}
79 	iommu->prq = page_address(pages);
80 
81 	irq = dmar_alloc_hwirq(IOMMU_IRQ_ID_OFFSET_PRQ + iommu->seq_id, iommu->node, iommu);
82 	if (irq <= 0) {
83 		pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
84 		       iommu->name);
85 		ret = -EINVAL;
86 		goto free_prq;
87 	}
88 	iommu->pr_irq = irq;
89 
90 	snprintf(iommu->iopfq_name, sizeof(iommu->iopfq_name),
91 		 "dmar%d-iopfq", iommu->seq_id);
92 	iopfq = iopf_queue_alloc(iommu->iopfq_name);
93 	if (!iopfq) {
94 		pr_err("IOMMU: %s: Failed to allocate iopf queue\n", iommu->name);
95 		ret = -ENOMEM;
96 		goto free_hwirq;
97 	}
98 	iommu->iopf_queue = iopfq;
99 
100 	snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
101 
102 	ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
103 				   iommu->prq_name, iommu);
104 	if (ret) {
105 		pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
106 		       iommu->name);
107 		goto free_iopfq;
108 	}
109 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
110 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
111 	dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
112 
113 	init_completion(&iommu->prq_complete);
114 
115 	return 0;
116 
117 free_iopfq:
118 	iopf_queue_free(iommu->iopf_queue);
119 	iommu->iopf_queue = NULL;
120 free_hwirq:
121 	dmar_free_hwirq(irq);
122 	iommu->pr_irq = 0;
123 free_prq:
124 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
125 	iommu->prq = NULL;
126 
127 	return ret;
128 }
129 
130 int intel_svm_finish_prq(struct intel_iommu *iommu)
131 {
132 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
133 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
134 	dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
135 
136 	if (iommu->pr_irq) {
137 		free_irq(iommu->pr_irq, iommu);
138 		dmar_free_hwirq(iommu->pr_irq);
139 		iommu->pr_irq = 0;
140 	}
141 
142 	if (iommu->iopf_queue) {
143 		iopf_queue_free(iommu->iopf_queue);
144 		iommu->iopf_queue = NULL;
145 	}
146 
147 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
148 	iommu->prq = NULL;
149 
150 	return 0;
151 }
152 
153 void intel_svm_check(struct intel_iommu *iommu)
154 {
155 	if (!pasid_supported(iommu))
156 		return;
157 
158 	if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
159 	    !cap_fl1gp_support(iommu->cap)) {
160 		pr_err("%s SVM disabled, incompatible 1GB page capability\n",
161 		       iommu->name);
162 		return;
163 	}
164 
165 	if (cpu_feature_enabled(X86_FEATURE_LA57) &&
166 	    !cap_fl5lp_support(iommu->cap)) {
167 		pr_err("%s SVM disabled, incompatible paging mode\n",
168 		       iommu->name);
169 		return;
170 	}
171 
172 	iommu->flags |= VTD_FLAG_SVM_CAPABLE;
173 }
174 
175 static void __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 	struct device_domain_info *info = dev_iommu_priv_get(sdev->dev);
181 
182 	if (WARN_ON(!pages))
183 		return;
184 
185 	qi_flush_piotlb(sdev->iommu, sdev->did, svm->pasid, address, pages, ih);
186 	if (info->ats_enabled) {
187 		qi_flush_dev_iotlb_pasid(sdev->iommu, sdev->sid, info->pfsid,
188 					 svm->pasid, sdev->qdep, address,
189 					 order_base_2(pages));
190 		quirk_extra_dev_tlb_flush(info, address, order_base_2(pages),
191 					  svm->pasid, sdev->qdep);
192 	}
193 }
194 
195 static void intel_flush_svm_range_dev(struct intel_svm *svm,
196 				      struct intel_svm_dev *sdev,
197 				      unsigned long address,
198 				      unsigned long pages, int ih)
199 {
200 	unsigned long shift = ilog2(__roundup_pow_of_two(pages));
201 	unsigned long align = (1ULL << (VTD_PAGE_SHIFT + shift));
202 	unsigned long start = ALIGN_DOWN(address, align);
203 	unsigned long end = ALIGN(address + (pages << VTD_PAGE_SHIFT), align);
204 
205 	while (start < end) {
206 		__flush_svm_range_dev(svm, sdev, start, align >> VTD_PAGE_SHIFT, ih);
207 		start += align;
208 	}
209 }
210 
211 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
212 				unsigned long pages, int ih)
213 {
214 	struct intel_svm_dev *sdev;
215 
216 	rcu_read_lock();
217 	list_for_each_entry_rcu(sdev, &svm->devs, list)
218 		intel_flush_svm_range_dev(svm, sdev, address, pages, ih);
219 	rcu_read_unlock();
220 }
221 
222 /* Pages have been freed at this point */
223 static void intel_invalidate_range(struct mmu_notifier *mn,
224 				   struct mm_struct *mm,
225 				   unsigned long start, unsigned long end)
226 {
227 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
228 
229 	intel_flush_svm_range(svm, start,
230 			      (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0);
231 }
232 
233 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
234 {
235 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
236 	struct intel_svm_dev *sdev;
237 
238 	/* This might end up being called from exit_mmap(), *before* the page
239 	 * tables are cleared. And __mmu_notifier_release() will delete us from
240 	 * the list of notifiers so that our invalidate_range() callback doesn't
241 	 * get called when the page tables are cleared. So we need to protect
242 	 * against hardware accessing those page tables.
243 	 *
244 	 * We do it by clearing the entry in the PASID table and then flushing
245 	 * the IOTLB and the PASID table caches. This might upset hardware;
246 	 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
247 	 * page) so that we end up taking a fault that the hardware really
248 	 * *has* to handle gracefully without affecting other processes.
249 	 */
250 	rcu_read_lock();
251 	list_for_each_entry_rcu(sdev, &svm->devs, list)
252 		intel_pasid_tear_down_entry(sdev->iommu, sdev->dev,
253 					    svm->pasid, true);
254 	rcu_read_unlock();
255 
256 }
257 
258 static const struct mmu_notifier_ops intel_mmuops = {
259 	.release = intel_mm_release,
260 	.invalidate_range = intel_invalidate_range,
261 };
262 
263 static DEFINE_MUTEX(pasid_mutex);
264 
265 static int pasid_to_svm_sdev(struct device *dev, unsigned int pasid,
266 			     struct intel_svm **rsvm,
267 			     struct intel_svm_dev **rsdev)
268 {
269 	struct intel_svm_dev *sdev = NULL;
270 	struct intel_svm *svm;
271 
272 	/* The caller should hold the pasid_mutex lock */
273 	if (WARN_ON(!mutex_is_locked(&pasid_mutex)))
274 		return -EINVAL;
275 
276 	if (pasid == INVALID_IOASID || pasid >= PASID_MAX)
277 		return -EINVAL;
278 
279 	svm = pasid_private_find(pasid);
280 	if (IS_ERR(svm))
281 		return PTR_ERR(svm);
282 
283 	if (!svm)
284 		goto out;
285 
286 	/*
287 	 * If we found svm for the PASID, there must be at least one device
288 	 * bond.
289 	 */
290 	if (WARN_ON(list_empty(&svm->devs)))
291 		return -EINVAL;
292 	sdev = svm_lookup_device_by_dev(svm, dev);
293 
294 out:
295 	*rsvm = svm;
296 	*rsdev = sdev;
297 
298 	return 0;
299 }
300 
301 static int intel_svm_bind_mm(struct intel_iommu *iommu, struct device *dev,
302 			     struct mm_struct *mm)
303 {
304 	struct device_domain_info *info = dev_iommu_priv_get(dev);
305 	struct intel_svm_dev *sdev;
306 	struct intel_svm *svm;
307 	unsigned long sflags;
308 	int ret = 0;
309 
310 	svm = pasid_private_find(mm->pasid);
311 	if (!svm) {
312 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
313 		if (!svm)
314 			return -ENOMEM;
315 
316 		svm->pasid = mm->pasid;
317 		svm->mm = mm;
318 		INIT_LIST_HEAD_RCU(&svm->devs);
319 
320 		svm->notifier.ops = &intel_mmuops;
321 		ret = mmu_notifier_register(&svm->notifier, mm);
322 		if (ret) {
323 			kfree(svm);
324 			return ret;
325 		}
326 
327 		ret = pasid_private_add(svm->pasid, svm);
328 		if (ret) {
329 			mmu_notifier_unregister(&svm->notifier, mm);
330 			kfree(svm);
331 			return ret;
332 		}
333 	}
334 
335 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
336 	if (!sdev) {
337 		ret = -ENOMEM;
338 		goto free_svm;
339 	}
340 
341 	sdev->dev = dev;
342 	sdev->iommu = iommu;
343 	sdev->did = FLPT_DEFAULT_DID;
344 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
345 	init_rcu_head(&sdev->rcu);
346 	if (info->ats_enabled) {
347 		sdev->qdep = info->ats_qdep;
348 		if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
349 			sdev->qdep = 0;
350 	}
351 
352 	/* Setup the pasid table: */
353 	sflags = cpu_feature_enabled(X86_FEATURE_LA57) ? PASID_FLAG_FL5LP : 0;
354 	ret = intel_pasid_setup_first_level(iommu, dev, mm->pgd, mm->pasid,
355 					    FLPT_DEFAULT_DID, sflags);
356 	if (ret)
357 		goto free_sdev;
358 
359 	list_add_rcu(&sdev->list, &svm->devs);
360 
361 	return 0;
362 
363 free_sdev:
364 	kfree(sdev);
365 free_svm:
366 	if (list_empty(&svm->devs)) {
367 		mmu_notifier_unregister(&svm->notifier, mm);
368 		pasid_private_remove(mm->pasid);
369 		kfree(svm);
370 	}
371 
372 	return ret;
373 }
374 
375 /* Caller must hold pasid_mutex */
376 static int intel_svm_unbind_mm(struct device *dev, u32 pasid)
377 {
378 	struct intel_svm_dev *sdev;
379 	struct intel_iommu *iommu;
380 	struct intel_svm *svm;
381 	struct mm_struct *mm;
382 	int ret = -EINVAL;
383 
384 	iommu = device_to_iommu(dev, NULL, NULL);
385 	if (!iommu)
386 		goto out;
387 
388 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
389 	if (ret)
390 		goto out;
391 	mm = svm->mm;
392 
393 	if (sdev) {
394 		list_del_rcu(&sdev->list);
395 		/*
396 		 * Flush the PASID cache and IOTLB for this device.
397 		 * Note that we do depend on the hardware *not* using
398 		 * the PASID any more. Just as we depend on other
399 		 * devices never using PASIDs that they have no right
400 		 * to use. We have a *shared* PASID table, because it's
401 		 * large and has to be physically contiguous. So it's
402 		 * hard to be as defensive as we might like.
403 		 */
404 		intel_pasid_tear_down_entry(iommu, dev, svm->pasid, false);
405 		intel_svm_drain_prq(dev, svm->pasid);
406 		kfree_rcu(sdev, rcu);
407 
408 		if (list_empty(&svm->devs)) {
409 			if (svm->notifier.ops)
410 				mmu_notifier_unregister(&svm->notifier, mm);
411 			pasid_private_remove(svm->pasid);
412 			/*
413 			 * We mandate that no page faults may be outstanding
414 			 * for the PASID when intel_svm_unbind_mm() is called.
415 			 * If that is not obeyed, subtle errors will happen.
416 			 * Let's make them less subtle...
417 			 */
418 			memset(svm, 0x6b, sizeof(*svm));
419 			kfree(svm);
420 		}
421 	}
422 out:
423 	return ret;
424 }
425 
426 /* Page request queue descriptor */
427 struct page_req_dsc {
428 	union {
429 		struct {
430 			u64 type:8;
431 			u64 pasid_present:1;
432 			u64 priv_data_present:1;
433 			u64 rsvd:6;
434 			u64 rid:16;
435 			u64 pasid:20;
436 			u64 exe_req:1;
437 			u64 pm_req:1;
438 			u64 rsvd2:10;
439 		};
440 		u64 qw_0;
441 	};
442 	union {
443 		struct {
444 			u64 rd_req:1;
445 			u64 wr_req:1;
446 			u64 lpig:1;
447 			u64 prg_index:9;
448 			u64 addr:52;
449 		};
450 		u64 qw_1;
451 	};
452 	u64 priv_data[2];
453 };
454 
455 static bool is_canonical_address(u64 addr)
456 {
457 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
458 	long saddr = (long) addr;
459 
460 	return (((saddr << shift) >> shift) == saddr);
461 }
462 
463 /**
464  * intel_svm_drain_prq - Drain page requests and responses for a pasid
465  * @dev: target device
466  * @pasid: pasid for draining
467  *
468  * Drain all pending page requests and responses related to @pasid in both
469  * software and hardware. This is supposed to be called after the device
470  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
471  * and DevTLB have been invalidated.
472  *
473  * It waits until all pending page requests for @pasid in the page fault
474  * queue are completed by the prq handling thread. Then follow the steps
475  * described in VT-d spec CH7.10 to drain all page requests and page
476  * responses pending in the hardware.
477  */
478 static void intel_svm_drain_prq(struct device *dev, u32 pasid)
479 {
480 	struct device_domain_info *info;
481 	struct dmar_domain *domain;
482 	struct intel_iommu *iommu;
483 	struct qi_desc desc[3];
484 	struct pci_dev *pdev;
485 	int head, tail;
486 	u16 sid, did;
487 	int qdep;
488 
489 	info = dev_iommu_priv_get(dev);
490 	if (WARN_ON(!info || !dev_is_pci(dev)))
491 		return;
492 
493 	if (!info->pri_enabled)
494 		return;
495 
496 	iommu = info->iommu;
497 	domain = info->domain;
498 	pdev = to_pci_dev(dev);
499 	sid = PCI_DEVID(info->bus, info->devfn);
500 	did = domain_id_iommu(domain, iommu);
501 	qdep = pci_ats_queue_depth(pdev);
502 
503 	/*
504 	 * Check and wait until all pending page requests in the queue are
505 	 * handled by the prq handling thread.
506 	 */
507 prq_retry:
508 	reinit_completion(&iommu->prq_complete);
509 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
510 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
511 	while (head != tail) {
512 		struct page_req_dsc *req;
513 
514 		req = &iommu->prq[head / sizeof(*req)];
515 		if (!req->pasid_present || req->pasid != pasid) {
516 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
517 			continue;
518 		}
519 
520 		wait_for_completion(&iommu->prq_complete);
521 		goto prq_retry;
522 	}
523 
524 	/*
525 	 * A work in IO page fault workqueue may try to lock pasid_mutex now.
526 	 * Holding pasid_mutex while waiting in iopf_queue_flush_dev() for
527 	 * all works in the workqueue to finish may cause deadlock.
528 	 *
529 	 * It's unnecessary to hold pasid_mutex in iopf_queue_flush_dev().
530 	 * Unlock it to allow the works to be handled while waiting for
531 	 * them to finish.
532 	 */
533 	lockdep_assert_held(&pasid_mutex);
534 	mutex_unlock(&pasid_mutex);
535 	iopf_queue_flush_dev(dev);
536 	mutex_lock(&pasid_mutex);
537 
538 	/*
539 	 * Perform steps described in VT-d spec CH7.10 to drain page
540 	 * requests and responses in hardware.
541 	 */
542 	memset(desc, 0, sizeof(desc));
543 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
544 			QI_IWD_FENCE |
545 			QI_IWD_TYPE;
546 	desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
547 			QI_EIOTLB_DID(did) |
548 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
549 			QI_EIOTLB_TYPE;
550 	desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
551 			QI_DEV_EIOTLB_SID(sid) |
552 			QI_DEV_EIOTLB_QDEP(qdep) |
553 			QI_DEIOTLB_TYPE |
554 			QI_DEV_IOTLB_PFSID(info->pfsid);
555 qi_retry:
556 	reinit_completion(&iommu->prq_complete);
557 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
558 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
559 		wait_for_completion(&iommu->prq_complete);
560 		goto qi_retry;
561 	}
562 }
563 
564 static int prq_to_iommu_prot(struct page_req_dsc *req)
565 {
566 	int prot = 0;
567 
568 	if (req->rd_req)
569 		prot |= IOMMU_FAULT_PERM_READ;
570 	if (req->wr_req)
571 		prot |= IOMMU_FAULT_PERM_WRITE;
572 	if (req->exe_req)
573 		prot |= IOMMU_FAULT_PERM_EXEC;
574 	if (req->pm_req)
575 		prot |= IOMMU_FAULT_PERM_PRIV;
576 
577 	return prot;
578 }
579 
580 static int intel_svm_prq_report(struct intel_iommu *iommu, struct device *dev,
581 				struct page_req_dsc *desc)
582 {
583 	struct iommu_fault_event event;
584 
585 	if (!dev || !dev_is_pci(dev))
586 		return -ENODEV;
587 
588 	/* Fill in event data for device specific processing */
589 	memset(&event, 0, sizeof(struct iommu_fault_event));
590 	event.fault.type = IOMMU_FAULT_PAGE_REQ;
591 	event.fault.prm.addr = (u64)desc->addr << VTD_PAGE_SHIFT;
592 	event.fault.prm.pasid = desc->pasid;
593 	event.fault.prm.grpid = desc->prg_index;
594 	event.fault.prm.perm = prq_to_iommu_prot(desc);
595 
596 	if (desc->lpig)
597 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
598 	if (desc->pasid_present) {
599 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
600 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
601 	}
602 	if (desc->priv_data_present) {
603 		/*
604 		 * Set last page in group bit if private data is present,
605 		 * page response is required as it does for LPIG.
606 		 * iommu_report_device_fault() doesn't understand this vendor
607 		 * specific requirement thus we set last_page as a workaround.
608 		 */
609 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
610 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
611 		event.fault.prm.private_data[0] = desc->priv_data[0];
612 		event.fault.prm.private_data[1] = desc->priv_data[1];
613 	} else if (dmar_latency_enabled(iommu, DMAR_LATENCY_PRQ)) {
614 		/*
615 		 * If the private data fields are not used by hardware, use it
616 		 * to monitor the prq handle latency.
617 		 */
618 		event.fault.prm.private_data[0] = ktime_to_ns(ktime_get());
619 	}
620 
621 	return iommu_report_device_fault(dev, &event);
622 }
623 
624 static void handle_bad_prq_event(struct intel_iommu *iommu,
625 				 struct page_req_dsc *req, int result)
626 {
627 	struct qi_desc desc;
628 
629 	pr_err("%s: Invalid page request: %08llx %08llx\n",
630 	       iommu->name, ((unsigned long long *)req)[0],
631 	       ((unsigned long long *)req)[1]);
632 
633 	/*
634 	 * Per VT-d spec. v3.0 ch7.7, system software must
635 	 * respond with page group response if private data
636 	 * is present (PDP) or last page in group (LPIG) bit
637 	 * is set. This is an additional VT-d feature beyond
638 	 * PCI ATS spec.
639 	 */
640 	if (!req->lpig && !req->priv_data_present)
641 		return;
642 
643 	desc.qw0 = QI_PGRP_PASID(req->pasid) |
644 			QI_PGRP_DID(req->rid) |
645 			QI_PGRP_PASID_P(req->pasid_present) |
646 			QI_PGRP_PDP(req->priv_data_present) |
647 			QI_PGRP_RESP_CODE(result) |
648 			QI_PGRP_RESP_TYPE;
649 	desc.qw1 = QI_PGRP_IDX(req->prg_index) |
650 			QI_PGRP_LPIG(req->lpig);
651 
652 	if (req->priv_data_present) {
653 		desc.qw2 = req->priv_data[0];
654 		desc.qw3 = req->priv_data[1];
655 	} else {
656 		desc.qw2 = 0;
657 		desc.qw3 = 0;
658 	}
659 
660 	qi_submit_sync(iommu, &desc, 1, 0);
661 }
662 
663 static irqreturn_t prq_event_thread(int irq, void *d)
664 {
665 	struct intel_iommu *iommu = d;
666 	struct page_req_dsc *req;
667 	int head, tail, handled;
668 	struct pci_dev *pdev;
669 	u64 address;
670 
671 	/*
672 	 * Clear PPR bit before reading head/tail registers, to ensure that
673 	 * we get a new interrupt if needed.
674 	 */
675 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
676 
677 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
678 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
679 	handled = (head != tail);
680 	while (head != tail) {
681 		req = &iommu->prq[head / sizeof(*req)];
682 		address = (u64)req->addr << VTD_PAGE_SHIFT;
683 
684 		if (unlikely(!req->pasid_present)) {
685 			pr_err("IOMMU: %s: Page request without PASID\n",
686 			       iommu->name);
687 bad_req:
688 			handle_bad_prq_event(iommu, req, QI_RESP_INVALID);
689 			goto prq_advance;
690 		}
691 
692 		if (unlikely(!is_canonical_address(address))) {
693 			pr_err("IOMMU: %s: Address is not canonical\n",
694 			       iommu->name);
695 			goto bad_req;
696 		}
697 
698 		if (unlikely(req->pm_req && (req->rd_req | req->wr_req))) {
699 			pr_err("IOMMU: %s: Page request in Privilege Mode\n",
700 			       iommu->name);
701 			goto bad_req;
702 		}
703 
704 		if (unlikely(req->exe_req && req->rd_req)) {
705 			pr_err("IOMMU: %s: Execution request not supported\n",
706 			       iommu->name);
707 			goto bad_req;
708 		}
709 
710 		/* Drop Stop Marker message. No need for a response. */
711 		if (unlikely(req->lpig && !req->rd_req && !req->wr_req))
712 			goto prq_advance;
713 
714 		pdev = pci_get_domain_bus_and_slot(iommu->segment,
715 						   PCI_BUS_NUM(req->rid),
716 						   req->rid & 0xff);
717 		/*
718 		 * If prq is to be handled outside iommu driver via receiver of
719 		 * the fault notifiers, we skip the page response here.
720 		 */
721 		if (!pdev)
722 			goto bad_req;
723 
724 		if (intel_svm_prq_report(iommu, &pdev->dev, req))
725 			handle_bad_prq_event(iommu, req, QI_RESP_INVALID);
726 		else
727 			trace_prq_report(iommu, &pdev->dev, req->qw_0, req->qw_1,
728 					 req->priv_data[0], req->priv_data[1],
729 					 iommu->prq_seq_number++);
730 		pci_dev_put(pdev);
731 prq_advance:
732 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
733 	}
734 
735 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
736 
737 	/*
738 	 * Clear the page request overflow bit and wake up all threads that
739 	 * are waiting for the completion of this handling.
740 	 */
741 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
742 		pr_info_ratelimited("IOMMU: %s: PRQ overflow detected\n",
743 				    iommu->name);
744 		head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
745 		tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
746 		if (head == tail) {
747 			iopf_queue_discard_partial(iommu->iopf_queue);
748 			writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
749 			pr_info_ratelimited("IOMMU: %s: PRQ overflow cleared",
750 					    iommu->name);
751 		}
752 	}
753 
754 	if (!completion_done(&iommu->prq_complete))
755 		complete(&iommu->prq_complete);
756 
757 	return IRQ_RETVAL(handled);
758 }
759 
760 int intel_svm_page_response(struct device *dev,
761 			    struct iommu_fault_event *evt,
762 			    struct iommu_page_response *msg)
763 {
764 	struct iommu_fault_page_request *prm;
765 	struct intel_iommu *iommu;
766 	bool private_present;
767 	bool pasid_present;
768 	bool last_page;
769 	u8 bus, devfn;
770 	int ret = 0;
771 	u16 sid;
772 
773 	if (!dev || !dev_is_pci(dev))
774 		return -ENODEV;
775 
776 	iommu = device_to_iommu(dev, &bus, &devfn);
777 	if (!iommu)
778 		return -ENODEV;
779 
780 	if (!msg || !evt)
781 		return -EINVAL;
782 
783 	prm = &evt->fault.prm;
784 	sid = PCI_DEVID(bus, devfn);
785 	pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
786 	private_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
787 	last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
788 
789 	if (!pasid_present) {
790 		ret = -EINVAL;
791 		goto out;
792 	}
793 
794 	if (prm->pasid == 0 || prm->pasid >= PASID_MAX) {
795 		ret = -EINVAL;
796 		goto out;
797 	}
798 
799 	/*
800 	 * Per VT-d spec. v3.0 ch7.7, system software must respond
801 	 * with page group response if private data is present (PDP)
802 	 * or last page in group (LPIG) bit is set. This is an
803 	 * additional VT-d requirement beyond PCI ATS spec.
804 	 */
805 	if (last_page || private_present) {
806 		struct qi_desc desc;
807 
808 		desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
809 				QI_PGRP_PASID_P(pasid_present) |
810 				QI_PGRP_PDP(private_present) |
811 				QI_PGRP_RESP_CODE(msg->code) |
812 				QI_PGRP_RESP_TYPE;
813 		desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
814 		desc.qw2 = 0;
815 		desc.qw3 = 0;
816 
817 		if (private_present) {
818 			desc.qw2 = prm->private_data[0];
819 			desc.qw3 = prm->private_data[1];
820 		} else if (prm->private_data[0]) {
821 			dmar_latency_update(iommu, DMAR_LATENCY_PRQ,
822 				ktime_to_ns(ktime_get()) - prm->private_data[0]);
823 		}
824 
825 		qi_submit_sync(iommu, &desc, 1, 0);
826 	}
827 out:
828 	return ret;
829 }
830 
831 void intel_svm_remove_dev_pasid(struct device *dev, ioasid_t pasid)
832 {
833 	mutex_lock(&pasid_mutex);
834 	intel_svm_unbind_mm(dev, pasid);
835 	mutex_unlock(&pasid_mutex);
836 }
837 
838 static int intel_svm_set_dev_pasid(struct iommu_domain *domain,
839 				   struct device *dev, ioasid_t pasid)
840 {
841 	struct device_domain_info *info = dev_iommu_priv_get(dev);
842 	struct intel_iommu *iommu = info->iommu;
843 	struct mm_struct *mm = domain->mm;
844 	int ret;
845 
846 	mutex_lock(&pasid_mutex);
847 	ret = intel_svm_bind_mm(iommu, dev, mm);
848 	mutex_unlock(&pasid_mutex);
849 
850 	return ret;
851 }
852 
853 static void intel_svm_domain_free(struct iommu_domain *domain)
854 {
855 	kfree(to_dmar_domain(domain));
856 }
857 
858 static const struct iommu_domain_ops intel_svm_domain_ops = {
859 	.set_dev_pasid		= intel_svm_set_dev_pasid,
860 	.free			= intel_svm_domain_free
861 };
862 
863 struct iommu_domain *intel_svm_domain_alloc(void)
864 {
865 	struct dmar_domain *domain;
866 
867 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
868 	if (!domain)
869 		return NULL;
870 	domain->domain.ops = &intel_svm_domain_ops;
871 
872 	return &domain->domain;
873 }
874