xref: /openbmc/linux/drivers/iommu/intel/svm.c (revision 3db55767)
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 static void intel_svm_drain_prq(struct device *dev, u32 pasid);
30 #define to_intel_svm_dev(handle) container_of(handle, struct intel_svm_dev, sva)
31 
32 static DEFINE_XARRAY_ALLOC(pasid_private_array);
33 static int pasid_private_add(ioasid_t pasid, void *priv)
34 {
35 	return xa_alloc(&pasid_private_array, &pasid, priv,
36 			XA_LIMIT(pasid, pasid), GFP_ATOMIC);
37 }
38 
39 static void pasid_private_remove(ioasid_t pasid)
40 {
41 	xa_erase(&pasid_private_array, pasid);
42 }
43 
44 static void *pasid_private_find(ioasid_t pasid)
45 {
46 	return xa_load(&pasid_private_array, pasid);
47 }
48 
49 static struct intel_svm_dev *
50 svm_lookup_device_by_dev(struct intel_svm *svm, struct device *dev)
51 {
52 	struct intel_svm_dev *sdev = NULL, *t;
53 
54 	rcu_read_lock();
55 	list_for_each_entry_rcu(t, &svm->devs, list) {
56 		if (t->dev == dev) {
57 			sdev = t;
58 			break;
59 		}
60 	}
61 	rcu_read_unlock();
62 
63 	return sdev;
64 }
65 
66 int intel_svm_enable_prq(struct intel_iommu *iommu)
67 {
68 	struct iopf_queue *iopfq;
69 	struct page *pages;
70 	int irq, ret;
71 
72 	pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
73 	if (!pages) {
74 		pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
75 			iommu->name);
76 		return -ENOMEM;
77 	}
78 	iommu->prq = page_address(pages);
79 
80 	irq = dmar_alloc_hwirq(IOMMU_IRQ_ID_OFFSET_PRQ + iommu->seq_id, iommu->node, iommu);
81 	if (irq <= 0) {
82 		pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
83 		       iommu->name);
84 		ret = -EINVAL;
85 		goto free_prq;
86 	}
87 	iommu->pr_irq = irq;
88 
89 	snprintf(iommu->iopfq_name, sizeof(iommu->iopfq_name),
90 		 "dmar%d-iopfq", iommu->seq_id);
91 	iopfq = iopf_queue_alloc(iommu->iopfq_name);
92 	if (!iopfq) {
93 		pr_err("IOMMU: %s: Failed to allocate iopf queue\n", iommu->name);
94 		ret = -ENOMEM;
95 		goto free_hwirq;
96 	}
97 	iommu->iopf_queue = iopfq;
98 
99 	snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
100 
101 	ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
102 				   iommu->prq_name, iommu);
103 	if (ret) {
104 		pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
105 		       iommu->name);
106 		goto free_iopfq;
107 	}
108 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
109 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
110 	dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
111 
112 	init_completion(&iommu->prq_complete);
113 
114 	return 0;
115 
116 free_iopfq:
117 	iopf_queue_free(iommu->iopf_queue);
118 	iommu->iopf_queue = NULL;
119 free_hwirq:
120 	dmar_free_hwirq(irq);
121 	iommu->pr_irq = 0;
122 free_prq:
123 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
124 	iommu->prq = NULL;
125 
126 	return ret;
127 }
128 
129 int intel_svm_finish_prq(struct intel_iommu *iommu)
130 {
131 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
132 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
133 	dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
134 
135 	if (iommu->pr_irq) {
136 		free_irq(iommu->pr_irq, iommu);
137 		dmar_free_hwirq(iommu->pr_irq);
138 		iommu->pr_irq = 0;
139 	}
140 
141 	if (iommu->iopf_queue) {
142 		iopf_queue_free(iommu->iopf_queue);
143 		iommu->iopf_queue = NULL;
144 	}
145 
146 	free_pages((unsigned long)iommu->prq, PRQ_ORDER);
147 	iommu->prq = NULL;
148 
149 	return 0;
150 }
151 
152 void intel_svm_check(struct intel_iommu *iommu)
153 {
154 	if (!pasid_supported(iommu))
155 		return;
156 
157 	if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
158 	    !cap_fl1gp_support(iommu->cap)) {
159 		pr_err("%s SVM disabled, incompatible 1GB page capability\n",
160 		       iommu->name);
161 		return;
162 	}
163 
164 	if (cpu_feature_enabled(X86_FEATURE_LA57) &&
165 	    !cap_fl5lp_support(iommu->cap)) {
166 		pr_err("%s SVM disabled, incompatible paging mode\n",
167 		       iommu->name);
168 		return;
169 	}
170 
171 	iommu->flags |= VTD_FLAG_SVM_CAPABLE;
172 }
173 
174 static void __flush_svm_range_dev(struct intel_svm *svm,
175 				  struct intel_svm_dev *sdev,
176 				  unsigned long address,
177 				  unsigned long pages, int ih)
178 {
179 	struct device_domain_info *info = dev_iommu_priv_get(sdev->dev);
180 
181 	if (WARN_ON(!pages))
182 		return;
183 
184 	qi_flush_piotlb(sdev->iommu, sdev->did, svm->pasid, address, pages, ih);
185 	if (info->ats_enabled) {
186 		qi_flush_dev_iotlb_pasid(sdev->iommu, sdev->sid, info->pfsid,
187 					 svm->pasid, sdev->qdep, address,
188 					 order_base_2(pages));
189 		quirk_extra_dev_tlb_flush(info, address, order_base_2(pages),
190 					  svm->pasid, sdev->qdep);
191 	}
192 }
193 
194 static void intel_flush_svm_range_dev(struct intel_svm *svm,
195 				      struct intel_svm_dev *sdev,
196 				      unsigned long address,
197 				      unsigned long pages, int ih)
198 {
199 	unsigned long shift = ilog2(__roundup_pow_of_two(pages));
200 	unsigned long align = (1ULL << (VTD_PAGE_SHIFT + shift));
201 	unsigned long start = ALIGN_DOWN(address, align);
202 	unsigned long end = ALIGN(address + (pages << VTD_PAGE_SHIFT), align);
203 
204 	while (start < end) {
205 		__flush_svm_range_dev(svm, sdev, start, align >> VTD_PAGE_SHIFT, ih);
206 		start += align;
207 	}
208 }
209 
210 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
211 				unsigned long pages, int ih)
212 {
213 	struct intel_svm_dev *sdev;
214 
215 	rcu_read_lock();
216 	list_for_each_entry_rcu(sdev, &svm->devs, list)
217 		intel_flush_svm_range_dev(svm, sdev, address, pages, ih);
218 	rcu_read_unlock();
219 }
220 
221 /* Pages have been freed at this point */
222 static void intel_invalidate_range(struct mmu_notifier *mn,
223 				   struct mm_struct *mm,
224 				   unsigned long start, unsigned long end)
225 {
226 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
227 
228 	intel_flush_svm_range(svm, start,
229 			      (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0);
230 }
231 
232 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
233 {
234 	struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
235 	struct intel_svm_dev *sdev;
236 
237 	/* This might end up being called from exit_mmap(), *before* the page
238 	 * tables are cleared. And __mmu_notifier_release() will delete us from
239 	 * the list of notifiers so that our invalidate_range() callback doesn't
240 	 * get called when the page tables are cleared. So we need to protect
241 	 * against hardware accessing those page tables.
242 	 *
243 	 * We do it by clearing the entry in the PASID table and then flushing
244 	 * the IOTLB and the PASID table caches. This might upset hardware;
245 	 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
246 	 * page) so that we end up taking a fault that the hardware really
247 	 * *has* to handle gracefully without affecting other processes.
248 	 */
249 	rcu_read_lock();
250 	list_for_each_entry_rcu(sdev, &svm->devs, list)
251 		intel_pasid_tear_down_entry(sdev->iommu, sdev->dev,
252 					    svm->pasid, true);
253 	rcu_read_unlock();
254 
255 }
256 
257 static const struct mmu_notifier_ops intel_mmuops = {
258 	.release = intel_mm_release,
259 	.invalidate_range = intel_invalidate_range,
260 };
261 
262 static DEFINE_MUTEX(pasid_mutex);
263 
264 static int pasid_to_svm_sdev(struct device *dev, unsigned int pasid,
265 			     struct intel_svm **rsvm,
266 			     struct intel_svm_dev **rsdev)
267 {
268 	struct intel_svm_dev *sdev = NULL;
269 	struct intel_svm *svm;
270 
271 	/* The caller should hold the pasid_mutex lock */
272 	if (WARN_ON(!mutex_is_locked(&pasid_mutex)))
273 		return -EINVAL;
274 
275 	if (pasid == IOMMU_PASID_INVALID || pasid >= PASID_MAX)
276 		return -EINVAL;
277 
278 	svm = pasid_private_find(pasid);
279 	if (IS_ERR(svm))
280 		return PTR_ERR(svm);
281 
282 	if (!svm)
283 		goto out;
284 
285 	/*
286 	 * If we found svm for the PASID, there must be at least one device
287 	 * bond.
288 	 */
289 	if (WARN_ON(list_empty(&svm->devs)))
290 		return -EINVAL;
291 	sdev = svm_lookup_device_by_dev(svm, dev);
292 
293 out:
294 	*rsvm = svm;
295 	*rsdev = sdev;
296 
297 	return 0;
298 }
299 
300 static int intel_svm_bind_mm(struct intel_iommu *iommu, struct device *dev,
301 			     struct mm_struct *mm)
302 {
303 	struct device_domain_info *info = dev_iommu_priv_get(dev);
304 	struct intel_svm_dev *sdev;
305 	struct intel_svm *svm;
306 	unsigned long sflags;
307 	int ret = 0;
308 
309 	svm = pasid_private_find(mm->pasid);
310 	if (!svm) {
311 		svm = kzalloc(sizeof(*svm), GFP_KERNEL);
312 		if (!svm)
313 			return -ENOMEM;
314 
315 		svm->pasid = mm->pasid;
316 		svm->mm = mm;
317 		INIT_LIST_HEAD_RCU(&svm->devs);
318 
319 		svm->notifier.ops = &intel_mmuops;
320 		ret = mmu_notifier_register(&svm->notifier, mm);
321 		if (ret) {
322 			kfree(svm);
323 			return ret;
324 		}
325 
326 		ret = pasid_private_add(svm->pasid, svm);
327 		if (ret) {
328 			mmu_notifier_unregister(&svm->notifier, mm);
329 			kfree(svm);
330 			return ret;
331 		}
332 	}
333 
334 	sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
335 	if (!sdev) {
336 		ret = -ENOMEM;
337 		goto free_svm;
338 	}
339 
340 	sdev->dev = dev;
341 	sdev->iommu = iommu;
342 	sdev->did = FLPT_DEFAULT_DID;
343 	sdev->sid = PCI_DEVID(info->bus, info->devfn);
344 	init_rcu_head(&sdev->rcu);
345 	if (info->ats_enabled) {
346 		sdev->qdep = info->ats_qdep;
347 		if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
348 			sdev->qdep = 0;
349 	}
350 
351 	/* Setup the pasid table: */
352 	sflags = cpu_feature_enabled(X86_FEATURE_LA57) ? PASID_FLAG_FL5LP : 0;
353 	ret = intel_pasid_setup_first_level(iommu, dev, mm->pgd, mm->pasid,
354 					    FLPT_DEFAULT_DID, sflags);
355 	if (ret)
356 		goto free_sdev;
357 
358 	list_add_rcu(&sdev->list, &svm->devs);
359 
360 	return 0;
361 
362 free_sdev:
363 	kfree(sdev);
364 free_svm:
365 	if (list_empty(&svm->devs)) {
366 		mmu_notifier_unregister(&svm->notifier, mm);
367 		pasid_private_remove(mm->pasid);
368 		kfree(svm);
369 	}
370 
371 	return ret;
372 }
373 
374 /* Caller must hold pasid_mutex */
375 static int intel_svm_unbind_mm(struct device *dev, u32 pasid)
376 {
377 	struct intel_svm_dev *sdev;
378 	struct intel_iommu *iommu;
379 	struct intel_svm *svm;
380 	struct mm_struct *mm;
381 	int ret = -EINVAL;
382 
383 	iommu = device_to_iommu(dev, NULL, NULL);
384 	if (!iommu)
385 		goto out;
386 
387 	ret = pasid_to_svm_sdev(dev, pasid, &svm, &sdev);
388 	if (ret)
389 		goto out;
390 	mm = svm->mm;
391 
392 	if (sdev) {
393 		list_del_rcu(&sdev->list);
394 		/*
395 		 * Flush the PASID cache and IOTLB for this device.
396 		 * Note that we do depend on the hardware *not* using
397 		 * the PASID any more. Just as we depend on other
398 		 * devices never using PASIDs that they have no right
399 		 * to use. We have a *shared* PASID table, because it's
400 		 * large and has to be physically contiguous. So it's
401 		 * hard to be as defensive as we might like.
402 		 */
403 		intel_pasid_tear_down_entry(iommu, dev, svm->pasid, false);
404 		intel_svm_drain_prq(dev, svm->pasid);
405 		kfree_rcu(sdev, rcu);
406 
407 		if (list_empty(&svm->devs)) {
408 			if (svm->notifier.ops)
409 				mmu_notifier_unregister(&svm->notifier, mm);
410 			pasid_private_remove(svm->pasid);
411 			/*
412 			 * We mandate that no page faults may be outstanding
413 			 * for the PASID when intel_svm_unbind_mm() is called.
414 			 * If that is not obeyed, subtle errors will happen.
415 			 * Let's make them less subtle...
416 			 */
417 			memset(svm, 0x6b, sizeof(*svm));
418 			kfree(svm);
419 		}
420 	}
421 out:
422 	return ret;
423 }
424 
425 /* Page request queue descriptor */
426 struct page_req_dsc {
427 	union {
428 		struct {
429 			u64 type:8;
430 			u64 pasid_present:1;
431 			u64 priv_data_present:1;
432 			u64 rsvd:6;
433 			u64 rid:16;
434 			u64 pasid:20;
435 			u64 exe_req:1;
436 			u64 pm_req:1;
437 			u64 rsvd2:10;
438 		};
439 		u64 qw_0;
440 	};
441 	union {
442 		struct {
443 			u64 rd_req:1;
444 			u64 wr_req:1;
445 			u64 lpig:1;
446 			u64 prg_index:9;
447 			u64 addr:52;
448 		};
449 		u64 qw_1;
450 	};
451 	u64 priv_data[2];
452 };
453 
454 static bool is_canonical_address(u64 addr)
455 {
456 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
457 	long saddr = (long) addr;
458 
459 	return (((saddr << shift) >> shift) == saddr);
460 }
461 
462 /**
463  * intel_svm_drain_prq - Drain page requests and responses for a pasid
464  * @dev: target device
465  * @pasid: pasid for draining
466  *
467  * Drain all pending page requests and responses related to @pasid in both
468  * software and hardware. This is supposed to be called after the device
469  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
470  * and DevTLB have been invalidated.
471  *
472  * It waits until all pending page requests for @pasid in the page fault
473  * queue are completed by the prq handling thread. Then follow the steps
474  * described in VT-d spec CH7.10 to drain all page requests and page
475  * responses pending in the hardware.
476  */
477 static void intel_svm_drain_prq(struct device *dev, u32 pasid)
478 {
479 	struct device_domain_info *info;
480 	struct dmar_domain *domain;
481 	struct intel_iommu *iommu;
482 	struct qi_desc desc[3];
483 	struct pci_dev *pdev;
484 	int head, tail;
485 	u16 sid, did;
486 	int qdep;
487 
488 	info = dev_iommu_priv_get(dev);
489 	if (WARN_ON(!info || !dev_is_pci(dev)))
490 		return;
491 
492 	if (!info->pri_enabled)
493 		return;
494 
495 	iommu = info->iommu;
496 	domain = info->domain;
497 	pdev = to_pci_dev(dev);
498 	sid = PCI_DEVID(info->bus, info->devfn);
499 	did = domain_id_iommu(domain, iommu);
500 	qdep = pci_ats_queue_depth(pdev);
501 
502 	/*
503 	 * Check and wait until all pending page requests in the queue are
504 	 * handled by the prq handling thread.
505 	 */
506 prq_retry:
507 	reinit_completion(&iommu->prq_complete);
508 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
509 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
510 	while (head != tail) {
511 		struct page_req_dsc *req;
512 
513 		req = &iommu->prq[head / sizeof(*req)];
514 		if (!req->pasid_present || req->pasid != pasid) {
515 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
516 			continue;
517 		}
518 
519 		wait_for_completion(&iommu->prq_complete);
520 		goto prq_retry;
521 	}
522 
523 	/*
524 	 * A work in IO page fault workqueue may try to lock pasid_mutex now.
525 	 * Holding pasid_mutex while waiting in iopf_queue_flush_dev() for
526 	 * all works in the workqueue to finish may cause deadlock.
527 	 *
528 	 * It's unnecessary to hold pasid_mutex in iopf_queue_flush_dev().
529 	 * Unlock it to allow the works to be handled while waiting for
530 	 * them to finish.
531 	 */
532 	lockdep_assert_held(&pasid_mutex);
533 	mutex_unlock(&pasid_mutex);
534 	iopf_queue_flush_dev(dev);
535 	mutex_lock(&pasid_mutex);
536 
537 	/*
538 	 * Perform steps described in VT-d spec CH7.10 to drain page
539 	 * requests and responses in hardware.
540 	 */
541 	memset(desc, 0, sizeof(desc));
542 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
543 			QI_IWD_FENCE |
544 			QI_IWD_TYPE;
545 	desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
546 			QI_EIOTLB_DID(did) |
547 			QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
548 			QI_EIOTLB_TYPE;
549 	desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
550 			QI_DEV_EIOTLB_SID(sid) |
551 			QI_DEV_EIOTLB_QDEP(qdep) |
552 			QI_DEIOTLB_TYPE |
553 			QI_DEV_IOTLB_PFSID(info->pfsid);
554 qi_retry:
555 	reinit_completion(&iommu->prq_complete);
556 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
557 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
558 		wait_for_completion(&iommu->prq_complete);
559 		goto qi_retry;
560 	}
561 }
562 
563 static int prq_to_iommu_prot(struct page_req_dsc *req)
564 {
565 	int prot = 0;
566 
567 	if (req->rd_req)
568 		prot |= IOMMU_FAULT_PERM_READ;
569 	if (req->wr_req)
570 		prot |= IOMMU_FAULT_PERM_WRITE;
571 	if (req->exe_req)
572 		prot |= IOMMU_FAULT_PERM_EXEC;
573 	if (req->pm_req)
574 		prot |= IOMMU_FAULT_PERM_PRIV;
575 
576 	return prot;
577 }
578 
579 static int intel_svm_prq_report(struct intel_iommu *iommu, struct device *dev,
580 				struct page_req_dsc *desc)
581 {
582 	struct iommu_fault_event event;
583 
584 	if (!dev || !dev_is_pci(dev))
585 		return -ENODEV;
586 
587 	/* Fill in event data for device specific processing */
588 	memset(&event, 0, sizeof(struct iommu_fault_event));
589 	event.fault.type = IOMMU_FAULT_PAGE_REQ;
590 	event.fault.prm.addr = (u64)desc->addr << VTD_PAGE_SHIFT;
591 	event.fault.prm.pasid = desc->pasid;
592 	event.fault.prm.grpid = desc->prg_index;
593 	event.fault.prm.perm = prq_to_iommu_prot(desc);
594 
595 	if (desc->lpig)
596 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
597 	if (desc->pasid_present) {
598 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
599 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
600 	}
601 	if (desc->priv_data_present) {
602 		/*
603 		 * Set last page in group bit if private data is present,
604 		 * page response is required as it does for LPIG.
605 		 * iommu_report_device_fault() doesn't understand this vendor
606 		 * specific requirement thus we set last_page as a workaround.
607 		 */
608 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
609 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
610 		event.fault.prm.private_data[0] = desc->priv_data[0];
611 		event.fault.prm.private_data[1] = desc->priv_data[1];
612 	} else if (dmar_latency_enabled(iommu, DMAR_LATENCY_PRQ)) {
613 		/*
614 		 * If the private data fields are not used by hardware, use it
615 		 * to monitor the prq handle latency.
616 		 */
617 		event.fault.prm.private_data[0] = ktime_to_ns(ktime_get());
618 	}
619 
620 	return iommu_report_device_fault(dev, &event);
621 }
622 
623 static void handle_bad_prq_event(struct intel_iommu *iommu,
624 				 struct page_req_dsc *req, int result)
625 {
626 	struct qi_desc desc;
627 
628 	pr_err("%s: Invalid page request: %08llx %08llx\n",
629 	       iommu->name, ((unsigned long long *)req)[0],
630 	       ((unsigned long long *)req)[1]);
631 
632 	/*
633 	 * Per VT-d spec. v3.0 ch7.7, system software must
634 	 * respond with page group response if private data
635 	 * is present (PDP) or last page in group (LPIG) bit
636 	 * is set. This is an additional VT-d feature beyond
637 	 * PCI ATS spec.
638 	 */
639 	if (!req->lpig && !req->priv_data_present)
640 		return;
641 
642 	desc.qw0 = QI_PGRP_PASID(req->pasid) |
643 			QI_PGRP_DID(req->rid) |
644 			QI_PGRP_PASID_P(req->pasid_present) |
645 			QI_PGRP_PDP(req->priv_data_present) |
646 			QI_PGRP_RESP_CODE(result) |
647 			QI_PGRP_RESP_TYPE;
648 	desc.qw1 = QI_PGRP_IDX(req->prg_index) |
649 			QI_PGRP_LPIG(req->lpig);
650 
651 	if (req->priv_data_present) {
652 		desc.qw2 = req->priv_data[0];
653 		desc.qw3 = req->priv_data[1];
654 	} else {
655 		desc.qw2 = 0;
656 		desc.qw3 = 0;
657 	}
658 
659 	qi_submit_sync(iommu, &desc, 1, 0);
660 }
661 
662 static irqreturn_t prq_event_thread(int irq, void *d)
663 {
664 	struct intel_iommu *iommu = d;
665 	struct page_req_dsc *req;
666 	int head, tail, handled;
667 	struct pci_dev *pdev;
668 	u64 address;
669 
670 	/*
671 	 * Clear PPR bit before reading head/tail registers, to ensure that
672 	 * we get a new interrupt if needed.
673 	 */
674 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
675 
676 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
677 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
678 	handled = (head != tail);
679 	while (head != tail) {
680 		req = &iommu->prq[head / sizeof(*req)];
681 		address = (u64)req->addr << VTD_PAGE_SHIFT;
682 
683 		if (unlikely(!req->pasid_present)) {
684 			pr_err("IOMMU: %s: Page request without PASID\n",
685 			       iommu->name);
686 bad_req:
687 			handle_bad_prq_event(iommu, req, QI_RESP_INVALID);
688 			goto prq_advance;
689 		}
690 
691 		if (unlikely(!is_canonical_address(address))) {
692 			pr_err("IOMMU: %s: Address is not canonical\n",
693 			       iommu->name);
694 			goto bad_req;
695 		}
696 
697 		if (unlikely(req->pm_req && (req->rd_req | req->wr_req))) {
698 			pr_err("IOMMU: %s: Page request in Privilege Mode\n",
699 			       iommu->name);
700 			goto bad_req;
701 		}
702 
703 		if (unlikely(req->exe_req && req->rd_req)) {
704 			pr_err("IOMMU: %s: Execution request not supported\n",
705 			       iommu->name);
706 			goto bad_req;
707 		}
708 
709 		/* Drop Stop Marker message. No need for a response. */
710 		if (unlikely(req->lpig && !req->rd_req && !req->wr_req))
711 			goto prq_advance;
712 
713 		pdev = pci_get_domain_bus_and_slot(iommu->segment,
714 						   PCI_BUS_NUM(req->rid),
715 						   req->rid & 0xff);
716 		/*
717 		 * If prq is to be handled outside iommu driver via receiver of
718 		 * the fault notifiers, we skip the page response here.
719 		 */
720 		if (!pdev)
721 			goto bad_req;
722 
723 		if (intel_svm_prq_report(iommu, &pdev->dev, req))
724 			handle_bad_prq_event(iommu, req, QI_RESP_INVALID);
725 		else
726 			trace_prq_report(iommu, &pdev->dev, req->qw_0, req->qw_1,
727 					 req->priv_data[0], req->priv_data[1],
728 					 iommu->prq_seq_number++);
729 		pci_dev_put(pdev);
730 prq_advance:
731 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
732 	}
733 
734 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
735 
736 	/*
737 	 * Clear the page request overflow bit and wake up all threads that
738 	 * are waiting for the completion of this handling.
739 	 */
740 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
741 		pr_info_ratelimited("IOMMU: %s: PRQ overflow detected\n",
742 				    iommu->name);
743 		head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
744 		tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
745 		if (head == tail) {
746 			iopf_queue_discard_partial(iommu->iopf_queue);
747 			writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
748 			pr_info_ratelimited("IOMMU: %s: PRQ overflow cleared",
749 					    iommu->name);
750 		}
751 	}
752 
753 	if (!completion_done(&iommu->prq_complete))
754 		complete(&iommu->prq_complete);
755 
756 	return IRQ_RETVAL(handled);
757 }
758 
759 int intel_svm_page_response(struct device *dev,
760 			    struct iommu_fault_event *evt,
761 			    struct iommu_page_response *msg)
762 {
763 	struct iommu_fault_page_request *prm;
764 	struct intel_iommu *iommu;
765 	bool private_present;
766 	bool pasid_present;
767 	bool last_page;
768 	u8 bus, devfn;
769 	int ret = 0;
770 	u16 sid;
771 
772 	if (!dev || !dev_is_pci(dev))
773 		return -ENODEV;
774 
775 	iommu = device_to_iommu(dev, &bus, &devfn);
776 	if (!iommu)
777 		return -ENODEV;
778 
779 	if (!msg || !evt)
780 		return -EINVAL;
781 
782 	prm = &evt->fault.prm;
783 	sid = PCI_DEVID(bus, devfn);
784 	pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
785 	private_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PRIV_DATA;
786 	last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
787 
788 	if (!pasid_present) {
789 		ret = -EINVAL;
790 		goto out;
791 	}
792 
793 	if (prm->pasid == 0 || prm->pasid >= PASID_MAX) {
794 		ret = -EINVAL;
795 		goto out;
796 	}
797 
798 	/*
799 	 * Per VT-d spec. v3.0 ch7.7, system software must respond
800 	 * with page group response if private data is present (PDP)
801 	 * or last page in group (LPIG) bit is set. This is an
802 	 * additional VT-d requirement beyond PCI ATS spec.
803 	 */
804 	if (last_page || private_present) {
805 		struct qi_desc desc;
806 
807 		desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
808 				QI_PGRP_PASID_P(pasid_present) |
809 				QI_PGRP_PDP(private_present) |
810 				QI_PGRP_RESP_CODE(msg->code) |
811 				QI_PGRP_RESP_TYPE;
812 		desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
813 		desc.qw2 = 0;
814 		desc.qw3 = 0;
815 
816 		if (private_present) {
817 			desc.qw2 = prm->private_data[0];
818 			desc.qw3 = prm->private_data[1];
819 		} else if (prm->private_data[0]) {
820 			dmar_latency_update(iommu, DMAR_LATENCY_PRQ,
821 				ktime_to_ns(ktime_get()) - prm->private_data[0]);
822 		}
823 
824 		qi_submit_sync(iommu, &desc, 1, 0);
825 	}
826 out:
827 	return ret;
828 }
829 
830 void intel_svm_remove_dev_pasid(struct device *dev, ioasid_t pasid)
831 {
832 	mutex_lock(&pasid_mutex);
833 	intel_svm_unbind_mm(dev, pasid);
834 	mutex_unlock(&pasid_mutex);
835 }
836 
837 static int intel_svm_set_dev_pasid(struct iommu_domain *domain,
838 				   struct device *dev, ioasid_t pasid)
839 {
840 	struct device_domain_info *info = dev_iommu_priv_get(dev);
841 	struct intel_iommu *iommu = info->iommu;
842 	struct mm_struct *mm = domain->mm;
843 	int ret;
844 
845 	mutex_lock(&pasid_mutex);
846 	ret = intel_svm_bind_mm(iommu, dev, mm);
847 	mutex_unlock(&pasid_mutex);
848 
849 	return ret;
850 }
851 
852 static void intel_svm_domain_free(struct iommu_domain *domain)
853 {
854 	kfree(to_dmar_domain(domain));
855 }
856 
857 static const struct iommu_domain_ops intel_svm_domain_ops = {
858 	.set_dev_pasid		= intel_svm_set_dev_pasid,
859 	.free			= intel_svm_domain_free
860 };
861 
862 struct iommu_domain *intel_svm_domain_alloc(void)
863 {
864 	struct dmar_domain *domain;
865 
866 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
867 	if (!domain)
868 		return NULL;
869 	domain->domain.ops = &intel_svm_domain_ops;
870 
871 	return &domain->domain;
872 }
873