xref: /openbmc/linux/drivers/iommu/intel/svm.c (revision 93696d8f)
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 <asm/page.h>
20 #include <asm/fpu/api.h>
21 
22 #include "iommu.h"
23 #include "pasid.h"
24 #include "perf.h"
25 #include "../iommu-sva.h"
26 #include "trace.h"
27 
28 static irqreturn_t prq_event_thread(int irq, void *d);
29 
30 static DEFINE_XARRAY_ALLOC(pasid_private_array);
31 static int pasid_private_add(ioasid_t pasid, void *priv)
32 {
33 	return xa_alloc(&pasid_private_array, &pasid, priv,
34 			XA_LIMIT(pasid, pasid), GFP_ATOMIC);
35 }
36 
37 static void pasid_private_remove(ioasid_t pasid)
38 {
39 	xa_erase(&pasid_private_array, pasid);
40 }
41 
42 static void *pasid_private_find(ioasid_t pasid)
43 {
44 	return xa_load(&pasid_private_array, pasid);
45 }
46 
47 static struct intel_svm_dev *
48 svm_lookup_device_by_dev(struct intel_svm *svm, struct device *dev)
49 {
50 	struct intel_svm_dev *sdev = NULL, *t;
51 
52 	rcu_read_lock();
53 	list_for_each_entry_rcu(t, &svm->devs, list) {
54 		if (t->dev == dev) {
55 			sdev = t;
56 			break;
57 		}
58 	}
59 	rcu_read_unlock();
60 
61 	return sdev;
62 }
63 
64 int intel_svm_enable_prq(struct intel_iommu *iommu)
65 {
66 	struct iopf_queue *iopfq;
67 	struct page *pages;
68 	int irq, ret;
69 
70 	pages = alloc_pages_node(iommu->node, GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
71 	if (!pages) {
72 		pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
73 			iommu->name);
74 		return -ENOMEM;
75 	}
76 	iommu->prq = page_address(pages);
77 
78 	irq = dmar_alloc_hwirq(IOMMU_IRQ_ID_OFFSET_PRQ + iommu->seq_id, iommu->node, iommu);
79 	if (irq <= 0) {
80 		pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
81 		       iommu->name);
82 		ret = -EINVAL;
83 		goto free_prq;
84 	}
85 	iommu->pr_irq = irq;
86 
87 	snprintf(iommu->iopfq_name, sizeof(iommu->iopfq_name),
88 		 "dmar%d-iopfq", iommu->seq_id);
89 	iopfq = iopf_queue_alloc(iommu->iopfq_name);
90 	if (!iopfq) {
91 		pr_err("IOMMU: %s: Failed to allocate iopf queue\n", iommu->name);
92 		ret = -ENOMEM;
93 		goto free_hwirq;
94 	}
95 	iommu->iopf_queue = iopfq;
96 
97 	snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
98 
99 	ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
100 				   iommu->prq_name, iommu);
101 	if (ret) {
102 		pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
103 		       iommu->name);
104 		goto free_iopfq;
105 	}
106 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
107 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
108 	dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
109 
110 	init_completion(&iommu->prq_complete);
111 
112 	return 0;
113 
114 free_iopfq:
115 	iopf_queue_free(iommu->iopf_queue);
116 	iommu->iopf_queue = NULL;
117 free_hwirq:
118 	dmar_free_hwirq(irq);
119 	iommu->pr_irq = 0;
120 free_prq:
121 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
122 	iommu->prq = NULL;
123 
124 	return ret;
125 }
126 
127 int intel_svm_finish_prq(struct intel_iommu *iommu)
128 {
129 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
130 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
131 	dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
132 
133 	if (iommu->pr_irq) {
134 		free_irq(iommu->pr_irq, iommu);
135 		dmar_free_hwirq(iommu->pr_irq);
136 		iommu->pr_irq = 0;
137 	}
138 
139 	if (iommu->iopf_queue) {
140 		iopf_queue_free(iommu->iopf_queue);
141 		iommu->iopf_queue = NULL;
142 	}
143 
144 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
145 	iommu->prq = NULL;
146 
147 	return 0;
148 }
149 
150 void intel_svm_check(struct intel_iommu *iommu)
151 {
152 	if (!pasid_supported(iommu))
153 		return;
154 
155 	if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
156 	    !cap_fl1gp_support(iommu->cap)) {
157 		pr_err("%s SVM disabled, incompatible 1GB page capability\n",
158 		       iommu->name);
159 		return;
160 	}
161 
162 	if (cpu_feature_enabled(X86_FEATURE_LA57) &&
163 	    !cap_fl5lp_support(iommu->cap)) {
164 		pr_err("%s SVM disabled, incompatible paging mode\n",
165 		       iommu->name);
166 		return;
167 	}
168 
169 	iommu->flags |= VTD_FLAG_SVM_CAPABLE;
170 }
171 
172 static void __flush_svm_range_dev(struct intel_svm *svm,
173 				  struct intel_svm_dev *sdev,
174 				  unsigned long address,
175 				  unsigned long pages, int ih)
176 {
177 	struct device_domain_info *info = dev_iommu_priv_get(sdev->dev);
178 
179 	if (WARN_ON(!pages))
180 		return;
181 
182 	qi_flush_piotlb(sdev->iommu, sdev->did, svm->pasid, address, pages, ih);
183 	if (info->ats_enabled) {
184 		qi_flush_dev_iotlb_pasid(sdev->iommu, sdev->sid, info->pfsid,
185 					 svm->pasid, sdev->qdep, address,
186 					 order_base_2(pages));
187 		quirk_extra_dev_tlb_flush(info, address, order_base_2(pages),
188 					  svm->pasid, sdev->qdep);
189 	}
190 }
191 
192 static void intel_flush_svm_range_dev(struct intel_svm *svm,
193 				      struct intel_svm_dev *sdev,
194 				      unsigned long address,
195 				      unsigned long pages, int ih)
196 {
197 	unsigned long shift = ilog2(__roundup_pow_of_two(pages));
198 	unsigned long align = (1ULL << (VTD_PAGE_SHIFT + shift));
199 	unsigned long start = ALIGN_DOWN(address, align);
200 	unsigned long end = ALIGN(address + (pages << VTD_PAGE_SHIFT), align);
201 
202 	while (start < end) {
203 		__flush_svm_range_dev(svm, sdev, start, align >> VTD_PAGE_SHIFT, ih);
204 		start += align;
205 	}
206 }
207 
208 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
209 				unsigned long pages, int ih)
210 {
211 	struct intel_svm_dev *sdev;
212 
213 	rcu_read_lock();
214 	list_for_each_entry_rcu(sdev, &svm->devs, list)
215 		intel_flush_svm_range_dev(svm, sdev, address, pages, ih);
216 	rcu_read_unlock();
217 }
218 
219 static void intel_flush_svm_all(struct intel_svm *svm)
220 {
221 	struct device_domain_info *info;
222 	struct intel_svm_dev *sdev;
223 
224 	rcu_read_lock();
225 	list_for_each_entry_rcu(sdev, &svm->devs, list) {
226 		info = dev_iommu_priv_get(sdev->dev);
227 
228 		qi_flush_piotlb(sdev->iommu, sdev->did, svm->pasid, 0, -1UL, 0);
229 		if (info->ats_enabled) {
230 			qi_flush_dev_iotlb_pasid(sdev->iommu, sdev->sid, info->pfsid,
231 						 svm->pasid, sdev->qdep,
232 						 0, 64 - VTD_PAGE_SHIFT);
233 			quirk_extra_dev_tlb_flush(info, 0, 64 - VTD_PAGE_SHIFT,
234 						  svm->pasid, sdev->qdep);
235 		}
236 	}
237 	rcu_read_unlock();
238 }
239 
240 /* Pages have been freed at this point */
241 static void intel_arch_invalidate_secondary_tlbs(struct mmu_notifier *mn,
242 					struct mm_struct *mm,
243 					unsigned long start, unsigned long end)
244 {
245 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
246 
247 	if (start == 0 && end == -1UL) {
248 		intel_flush_svm_all(svm);
249 		return;
250 	}
251 
252 	intel_flush_svm_range(svm, start,
253 			      (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0);
254 }
255 
256 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
257 {
258 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
259 	struct intel_svm_dev *sdev;
260 
261 	/* This might end up being called from exit_mmap(), *before* the page
262 	 * tables are cleared. And __mmu_notifier_release() will delete us from
263 	 * the list of notifiers so that our invalidate_range() callback doesn't
264 	 * get called when the page tables are cleared. So we need to protect
265 	 * against hardware accessing those page tables.
266 	 *
267 	 * We do it by clearing the entry in the PASID table and then flushing
268 	 * the IOTLB and the PASID table caches. This might upset hardware;
269 	 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
270 	 * page) so that we end up taking a fault that the hardware really
271 	 * *has* to handle gracefully without affecting other processes.
272 	 */
273 	rcu_read_lock();
274 	list_for_each_entry_rcu(sdev, &svm->devs, list)
275 		intel_pasid_tear_down_entry(sdev->iommu, sdev->dev,
276 					    svm->pasid, true);
277 	rcu_read_unlock();
278 
279 }
280 
281 static const struct mmu_notifier_ops intel_mmuops = {
282 	.release = intel_mm_release,
283 	.arch_invalidate_secondary_tlbs = intel_arch_invalidate_secondary_tlbs,
284 };
285 
286 static int pasid_to_svm_sdev(struct device *dev, unsigned int pasid,
287 			     struct intel_svm **rsvm,
288 			     struct intel_svm_dev **rsdev)
289 {
290 	struct intel_svm_dev *sdev = NULL;
291 	struct intel_svm *svm;
292 
293 	if (pasid == IOMMU_PASID_INVALID || pasid >= PASID_MAX)
294 		return -EINVAL;
295 
296 	svm = pasid_private_find(pasid);
297 	if (IS_ERR(svm))
298 		return PTR_ERR(svm);
299 
300 	if (!svm)
301 		goto out;
302 
303 	/*
304 	 * If we found svm for the PASID, there must be at least one device
305 	 * bond.
306 	 */
307 	if (WARN_ON(list_empty(&svm->devs)))
308 		return -EINVAL;
309 	sdev = svm_lookup_device_by_dev(svm, dev);
310 
311 out:
312 	*rsvm = svm;
313 	*rsdev = sdev;
314 
315 	return 0;
316 }
317 
318 static int intel_svm_bind_mm(struct intel_iommu *iommu, struct device *dev,
319 			     struct mm_struct *mm)
320 {
321 	struct device_domain_info *info = dev_iommu_priv_get(dev);
322 	struct intel_svm_dev *sdev;
323 	struct intel_svm *svm;
324 	unsigned long sflags;
325 	int ret = 0;
326 
327 	svm = pasid_private_find(mm->pasid);
328 	if (!svm) {
329 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
330 		if (!svm)
331 			return -ENOMEM;
332 
333 		svm->pasid = mm->pasid;
334 		svm->mm = mm;
335 		INIT_LIST_HEAD_RCU(&svm->devs);
336 
337 		svm->notifier.ops = &intel_mmuops;
338 		ret = mmu_notifier_register(&svm->notifier, mm);
339 		if (ret) {
340 			kfree(svm);
341 			return ret;
342 		}
343 
344 		ret = pasid_private_add(svm->pasid, svm);
345 		if (ret) {
346 			mmu_notifier_unregister(&svm->notifier, mm);
347 			kfree(svm);
348 			return ret;
349 		}
350 	}
351 
352 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
353 	if (!sdev) {
354 		ret = -ENOMEM;
355 		goto free_svm;
356 	}
357 
358 	sdev->dev = dev;
359 	sdev->iommu = iommu;
360 	sdev->did = FLPT_DEFAULT_DID;
361 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
362 	init_rcu_head(&sdev->rcu);
363 	if (info->ats_enabled) {
364 		sdev->qdep = info->ats_qdep;
365 		if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
366 			sdev->qdep = 0;
367 	}
368 
369 	/* Setup the pasid table: */
370 	sflags = cpu_feature_enabled(X86_FEATURE_LA57) ? PASID_FLAG_FL5LP : 0;
371 	ret = intel_pasid_setup_first_level(iommu, dev, mm->pgd, mm->pasid,
372 					    FLPT_DEFAULT_DID, sflags);
373 	if (ret)
374 		goto free_sdev;
375 
376 	list_add_rcu(&sdev->list, &svm->devs);
377 
378 	return 0;
379 
380 free_sdev:
381 	kfree(sdev);
382 free_svm:
383 	if (list_empty(&svm->devs)) {
384 		mmu_notifier_unregister(&svm->notifier, mm);
385 		pasid_private_remove(mm->pasid);
386 		kfree(svm);
387 	}
388 
389 	return ret;
390 }
391 
392 void intel_svm_remove_dev_pasid(struct device *dev, u32 pasid)
393 {
394 	struct intel_svm_dev *sdev;
395 	struct intel_iommu *iommu;
396 	struct intel_svm *svm;
397 	struct mm_struct *mm;
398 
399 	iommu = device_to_iommu(dev, NULL, NULL);
400 	if (!iommu)
401 		return;
402 
403 	if (pasid_to_svm_sdev(dev, pasid, &svm, &sdev))
404 		return;
405 	mm = svm->mm;
406 
407 	if (sdev) {
408 		list_del_rcu(&sdev->list);
409 		kfree_rcu(sdev, rcu);
410 
411 		if (list_empty(&svm->devs)) {
412 			if (svm->notifier.ops)
413 				mmu_notifier_unregister(&svm->notifier, mm);
414 			pasid_private_remove(svm->pasid);
415 			/*
416 			 * We mandate that no page faults may be outstanding
417 			 * for the PASID when intel_svm_unbind_mm() is called.
418 			 * If that is not obeyed, subtle errors will happen.
419 			 * Let's make them less subtle...
420 			 */
421 			memset(svm, 0x6b, sizeof(*svm));
422 			kfree(svm);
423 		}
424 	}
425 }
426 
427 /* Page request queue descriptor */
428 struct page_req_dsc {
429 	union {
430 		struct {
431 			u64 type:8;
432 			u64 pasid_present:1;
433 			u64 priv_data_present:1;
434 			u64 rsvd:6;
435 			u64 rid:16;
436 			u64 pasid:20;
437 			u64 exe_req:1;
438 			u64 pm_req:1;
439 			u64 rsvd2:10;
440 		};
441 		u64 qw_0;
442 	};
443 	union {
444 		struct {
445 			u64 rd_req:1;
446 			u64 wr_req:1;
447 			u64 lpig:1;
448 			u64 prg_index:9;
449 			u64 addr:52;
450 		};
451 		u64 qw_1;
452 	};
453 	u64 priv_data[2];
454 };
455 
456 static bool is_canonical_address(u64 addr)
457 {
458 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
459 	long saddr = (long) addr;
460 
461 	return (((saddr << shift) >> shift) == saddr);
462 }
463 
464 /**
465  * intel_drain_pasid_prq - Drain page requests and responses for a pasid
466  * @dev: target device
467  * @pasid: pasid for draining
468  *
469  * Drain all pending page requests and responses related to @pasid in both
470  * software and hardware. This is supposed to be called after the device
471  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
472  * and DevTLB have been invalidated.
473  *
474  * It waits until all pending page requests for @pasid in the page fault
475  * queue are completed by the prq handling thread. Then follow the steps
476  * described in VT-d spec CH7.10 to drain all page requests and page
477  * responses pending in the hardware.
478  */
479 void intel_drain_pasid_prq(struct device *dev, u32 pasid)
480 {
481 	struct device_domain_info *info;
482 	struct dmar_domain *domain;
483 	struct intel_iommu *iommu;
484 	struct qi_desc desc[3];
485 	struct pci_dev *pdev;
486 	int head, tail;
487 	u16 sid, did;
488 	int qdep;
489 
490 	info = dev_iommu_priv_get(dev);
491 	if (WARN_ON(!info || !dev_is_pci(dev)))
492 		return;
493 
494 	if (!info->pri_enabled)
495 		return;
496 
497 	iommu = info->iommu;
498 	domain = info->domain;
499 	pdev = to_pci_dev(dev);
500 	sid = PCI_DEVID(info->bus, info->devfn);
501 	did = domain_id_iommu(domain, iommu);
502 	qdep = pci_ats_queue_depth(pdev);
503 
504 	/*
505 	 * Check and wait until all pending page requests in the queue are
506 	 * handled by the prq handling thread.
507 	 */
508 prq_retry:
509 	reinit_completion(&iommu->prq_complete);
510 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
511 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
512 	while (head != tail) {
513 		struct page_req_dsc *req;
514 
515 		req = &iommu->prq[head / sizeof(*req)];
516 		if (!req->pasid_present || req->pasid != pasid) {
517 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
518 			continue;
519 		}
520 
521 		wait_for_completion(&iommu->prq_complete);
522 		goto prq_retry;
523 	}
524 
525 	iopf_queue_flush_dev(dev);
526 
527 	/*
528 	 * Perform steps described in VT-d spec CH7.10 to drain page
529 	 * requests and responses in hardware.
530 	 */
531 	memset(desc, 0, sizeof(desc));
532 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
533 			QI_IWD_FENCE |
534 			QI_IWD_TYPE;
535 	desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
536 			QI_EIOTLB_DID(did) |
537 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
538 			QI_EIOTLB_TYPE;
539 	desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
540 			QI_DEV_EIOTLB_SID(sid) |
541 			QI_DEV_EIOTLB_QDEP(qdep) |
542 			QI_DEIOTLB_TYPE |
543 			QI_DEV_IOTLB_PFSID(info->pfsid);
544 qi_retry:
545 	reinit_completion(&iommu->prq_complete);
546 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
547 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
548 		wait_for_completion(&iommu->prq_complete);
549 		goto qi_retry;
550 	}
551 }
552 
553 static int prq_to_iommu_prot(struct page_req_dsc *req)
554 {
555 	int prot = 0;
556 
557 	if (req->rd_req)
558 		prot |= IOMMU_FAULT_PERM_READ;
559 	if (req->wr_req)
560 		prot |= IOMMU_FAULT_PERM_WRITE;
561 	if (req->exe_req)
562 		prot |= IOMMU_FAULT_PERM_EXEC;
563 	if (req->pm_req)
564 		prot |= IOMMU_FAULT_PERM_PRIV;
565 
566 	return prot;
567 }
568 
569 static int intel_svm_prq_report(struct intel_iommu *iommu, struct device *dev,
570 				struct page_req_dsc *desc)
571 {
572 	struct iommu_fault_event event;
573 
574 	if (!dev || !dev_is_pci(dev))
575 		return -ENODEV;
576 
577 	/* Fill in event data for device specific processing */
578 	memset(&event, 0, sizeof(struct iommu_fault_event));
579 	event.fault.type = IOMMU_FAULT_PAGE_REQ;
580 	event.fault.prm.addr = (u64)desc->addr << VTD_PAGE_SHIFT;
581 	event.fault.prm.pasid = desc->pasid;
582 	event.fault.prm.grpid = desc->prg_index;
583 	event.fault.prm.perm = prq_to_iommu_prot(desc);
584 
585 	if (desc->lpig)
586 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
587 	if (desc->pasid_present) {
588 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
589 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
590 	}
591 	if (desc->priv_data_present) {
592 		/*
593 		 * Set last page in group bit if private data is present,
594 		 * page response is required as it does for LPIG.
595 		 * iommu_report_device_fault() doesn't understand this vendor
596 		 * specific requirement thus we set last_page as a workaround.
597 		 */
598 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
599 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
600 		event.fault.prm.private_data[0] = desc->priv_data[0];
601 		event.fault.prm.private_data[1] = desc->priv_data[1];
602 	} else if (dmar_latency_enabled(iommu, DMAR_LATENCY_PRQ)) {
603 		/*
604 		 * If the private data fields are not used by hardware, use it
605 		 * to monitor the prq handle latency.
606 		 */
607 		event.fault.prm.private_data[0] = ktime_to_ns(ktime_get());
608 	}
609 
610 	return iommu_report_device_fault(dev, &event);
611 }
612 
613 static void handle_bad_prq_event(struct intel_iommu *iommu,
614 				 struct page_req_dsc *req, int result)
615 {
616 	struct qi_desc desc;
617 
618 	pr_err("%s: Invalid page request: %08llx %08llx\n",
619 	       iommu->name, ((unsigned long long *)req)[0],
620 	       ((unsigned long long *)req)[1]);
621 
622 	/*
623 	 * Per VT-d spec. v3.0 ch7.7, system software must
624 	 * respond with page group response if private data
625 	 * is present (PDP) or last page in group (LPIG) bit
626 	 * is set. This is an additional VT-d feature beyond
627 	 * PCI ATS spec.
628 	 */
629 	if (!req->lpig && !req->priv_data_present)
630 		return;
631 
632 	desc.qw0 = QI_PGRP_PASID(req->pasid) |
633 			QI_PGRP_DID(req->rid) |
634 			QI_PGRP_PASID_P(req->pasid_present) |
635 			QI_PGRP_PDP(req->priv_data_present) |
636 			QI_PGRP_RESP_CODE(result) |
637 			QI_PGRP_RESP_TYPE;
638 	desc.qw1 = QI_PGRP_IDX(req->prg_index) |
639 			QI_PGRP_LPIG(req->lpig);
640 
641 	if (req->priv_data_present) {
642 		desc.qw2 = req->priv_data[0];
643 		desc.qw3 = req->priv_data[1];
644 	} else {
645 		desc.qw2 = 0;
646 		desc.qw3 = 0;
647 	}
648 
649 	qi_submit_sync(iommu, &desc, 1, 0);
650 }
651 
652 static irqreturn_t prq_event_thread(int irq, void *d)
653 {
654 	struct intel_iommu *iommu = d;
655 	struct page_req_dsc *req;
656 	int head, tail, handled;
657 	struct pci_dev *pdev;
658 	u64 address;
659 
660 	/*
661 	 * Clear PPR bit before reading head/tail registers, to ensure that
662 	 * we get a new interrupt if needed.
663 	 */
664 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
665 
666 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
667 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
668 	handled = (head != tail);
669 	while (head != tail) {
670 		req = &iommu->prq[head / sizeof(*req)];
671 		address = (u64)req->addr << VTD_PAGE_SHIFT;
672 
673 		if (unlikely(!req->pasid_present)) {
674 			pr_err("IOMMU: %s: Page request without PASID\n",
675 			       iommu->name);
676 bad_req:
677 			handle_bad_prq_event(iommu, req, QI_RESP_INVALID);
678 			goto prq_advance;
679 		}
680 
681 		if (unlikely(!is_canonical_address(address))) {
682 			pr_err("IOMMU: %s: Address is not canonical\n",
683 			       iommu->name);
684 			goto bad_req;
685 		}
686 
687 		if (unlikely(req->pm_req && (req->rd_req | req->wr_req))) {
688 			pr_err("IOMMU: %s: Page request in Privilege Mode\n",
689 			       iommu->name);
690 			goto bad_req;
691 		}
692 
693 		if (unlikely(req->exe_req && req->rd_req)) {
694 			pr_err("IOMMU: %s: Execution request not supported\n",
695 			       iommu->name);
696 			goto bad_req;
697 		}
698 
699 		/* Drop Stop Marker message. No need for a response. */
700 		if (unlikely(req->lpig && !req->rd_req && !req->wr_req))
701 			goto prq_advance;
702 
703 		pdev = pci_get_domain_bus_and_slot(iommu->segment,
704 						   PCI_BUS_NUM(req->rid),
705 						   req->rid & 0xff);
706 		/*
707 		 * If prq is to be handled outside iommu driver via receiver of
708 		 * the fault notifiers, we skip the page response here.
709 		 */
710 		if (!pdev)
711 			goto bad_req;
712 
713 		if (intel_svm_prq_report(iommu, &pdev->dev, req))
714 			handle_bad_prq_event(iommu, req, QI_RESP_INVALID);
715 		else
716 			trace_prq_report(iommu, &pdev->dev, req->qw_0, req->qw_1,
717 					 req->priv_data[0], req->priv_data[1],
718 					 iommu->prq_seq_number++);
719 		pci_dev_put(pdev);
720 prq_advance:
721 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
722 	}
723 
724 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
725 
726 	/*
727 	 * Clear the page request overflow bit and wake up all threads that
728 	 * are waiting for the completion of this handling.
729 	 */
730 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
731 		pr_info_ratelimited("IOMMU: %s: PRQ overflow detected\n",
732 				    iommu->name);
733 		head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
734 		tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
735 		if (head == tail) {
736 			iopf_queue_discard_partial(iommu->iopf_queue);
737 			writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
738 			pr_info_ratelimited("IOMMU: %s: PRQ overflow cleared",
739 					    iommu->name);
740 		}
741 	}
742 
743 	if (!completion_done(&iommu->prq_complete))
744 		complete(&iommu->prq_complete);
745 
746 	return IRQ_RETVAL(handled);
747 }
748 
749 int intel_svm_page_response(struct device *dev,
750 			    struct iommu_fault_event *evt,
751 			    struct iommu_page_response *msg)
752 {
753 	struct iommu_fault_page_request *prm;
754 	struct intel_iommu *iommu;
755 	bool private_present;
756 	bool pasid_present;
757 	bool last_page;
758 	u8 bus, devfn;
759 	int ret = 0;
760 	u16 sid;
761 
762 	if (!dev || !dev_is_pci(dev))
763 		return -ENODEV;
764 
765 	iommu = device_to_iommu(dev, &bus, &devfn);
766 	if (!iommu)
767 		return -ENODEV;
768 
769 	if (!msg || !evt)
770 		return -EINVAL;
771 
772 	prm = &evt->fault.prm;
773 	sid = PCI_DEVID(bus, devfn);
774 	pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
775 	private_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
776 	last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
777 
778 	if (!pasid_present) {
779 		ret = -EINVAL;
780 		goto out;
781 	}
782 
783 	if (prm->pasid == 0 || prm->pasid >= PASID_MAX) {
784 		ret = -EINVAL;
785 		goto out;
786 	}
787 
788 	/*
789 	 * Per VT-d spec. v3.0 ch7.7, system software must respond
790 	 * with page group response if private data is present (PDP)
791 	 * or last page in group (LPIG) bit is set. This is an
792 	 * additional VT-d requirement beyond PCI ATS spec.
793 	 */
794 	if (last_page || private_present) {
795 		struct qi_desc desc;
796 
797 		desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
798 				QI_PGRP_PASID_P(pasid_present) |
799 				QI_PGRP_PDP(private_present) |
800 				QI_PGRP_RESP_CODE(msg->code) |
801 				QI_PGRP_RESP_TYPE;
802 		desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
803 		desc.qw2 = 0;
804 		desc.qw3 = 0;
805 
806 		if (private_present) {
807 			desc.qw2 = prm->private_data[0];
808 			desc.qw3 = prm->private_data[1];
809 		} else if (prm->private_data[0]) {
810 			dmar_latency_update(iommu, DMAR_LATENCY_PRQ,
811 				ktime_to_ns(ktime_get()) - prm->private_data[0]);
812 		}
813 
814 		qi_submit_sync(iommu, &desc, 1, 0);
815 	}
816 out:
817 	return ret;
818 }
819 
820 static int intel_svm_set_dev_pasid(struct iommu_domain *domain,
821 				   struct device *dev, ioasid_t pasid)
822 {
823 	struct device_domain_info *info = dev_iommu_priv_get(dev);
824 	struct intel_iommu *iommu = info->iommu;
825 	struct mm_struct *mm = domain->mm;
826 
827 	return intel_svm_bind_mm(iommu, dev, mm);
828 }
829 
830 static void intel_svm_domain_free(struct iommu_domain *domain)
831 {
832 	kfree(to_dmar_domain(domain));
833 }
834 
835 static const struct iommu_domain_ops intel_svm_domain_ops = {
836 	.set_dev_pasid		= intel_svm_set_dev_pasid,
837 	.free			= intel_svm_domain_free
838 };
839 
840 struct iommu_domain *intel_svm_domain_alloc(void)
841 {
842 	struct dmar_domain *domain;
843 
844 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
845 	if (!domain)
846 		return NULL;
847 	domain->domain.ops = &intel_svm_domain_ops;
848 
849 	return &domain->domain;
850 }
851