xref: /openbmc/linux/mm/hmm.c (revision 5a170e9e)
1 /*
2  * Copyright 2013 Red Hat Inc.
3  *
4  * This program is free software; you can redistribute it and/or modify
5  * it under the terms of the GNU General Public License as published by
6  * the Free Software Foundation; either version 2 of the License, or
7  * (at your option) any later version.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * Authors: Jérôme Glisse <jglisse@redhat.com>
15  */
16 /*
17  * Refer to include/linux/hmm.h for information about heterogeneous memory
18  * management or HMM for short.
19  */
20 #include <linux/mm.h>
21 #include <linux/hmm.h>
22 #include <linux/init.h>
23 #include <linux/rmap.h>
24 #include <linux/swap.h>
25 #include <linux/slab.h>
26 #include <linux/sched.h>
27 #include <linux/mmzone.h>
28 #include <linux/pagemap.h>
29 #include <linux/swapops.h>
30 #include <linux/hugetlb.h>
31 #include <linux/memremap.h>
32 #include <linux/jump_label.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/memory_hotplug.h>
35 
36 #define PA_SECTION_SIZE (1UL << PA_SECTION_SHIFT)
37 
38 #if IS_ENABLED(CONFIG_HMM_MIRROR)
39 static const struct mmu_notifier_ops hmm_mmu_notifier_ops;
40 
41 /*
42  * struct hmm - HMM per mm struct
43  *
44  * @mm: mm struct this HMM struct is bound to
45  * @lock: lock protecting ranges list
46  * @ranges: list of range being snapshotted
47  * @mirrors: list of mirrors for this mm
48  * @mmu_notifier: mmu notifier to track updates to CPU page table
49  * @mirrors_sem: read/write semaphore protecting the mirrors list
50  */
51 struct hmm {
52 	struct mm_struct	*mm;
53 	spinlock_t		lock;
54 	struct list_head	ranges;
55 	struct list_head	mirrors;
56 	struct mmu_notifier	mmu_notifier;
57 	struct rw_semaphore	mirrors_sem;
58 };
59 
60 /*
61  * hmm_register - register HMM against an mm (HMM internal)
62  *
63  * @mm: mm struct to attach to
64  *
65  * This is not intended to be used directly by device drivers. It allocates an
66  * HMM struct if mm does not have one, and initializes it.
67  */
68 static struct hmm *hmm_register(struct mm_struct *mm)
69 {
70 	struct hmm *hmm = READ_ONCE(mm->hmm);
71 	bool cleanup = false;
72 
73 	/*
74 	 * The hmm struct can only be freed once the mm_struct goes away,
75 	 * hence we should always have pre-allocated an new hmm struct
76 	 * above.
77 	 */
78 	if (hmm)
79 		return hmm;
80 
81 	hmm = kmalloc(sizeof(*hmm), GFP_KERNEL);
82 	if (!hmm)
83 		return NULL;
84 	INIT_LIST_HEAD(&hmm->mirrors);
85 	init_rwsem(&hmm->mirrors_sem);
86 	hmm->mmu_notifier.ops = NULL;
87 	INIT_LIST_HEAD(&hmm->ranges);
88 	spin_lock_init(&hmm->lock);
89 	hmm->mm = mm;
90 
91 	spin_lock(&mm->page_table_lock);
92 	if (!mm->hmm)
93 		mm->hmm = hmm;
94 	else
95 		cleanup = true;
96 	spin_unlock(&mm->page_table_lock);
97 
98 	if (cleanup)
99 		goto error;
100 
101 	/*
102 	 * We should only get here if hold the mmap_sem in write mode ie on
103 	 * registration of first mirror through hmm_mirror_register()
104 	 */
105 	hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops;
106 	if (__mmu_notifier_register(&hmm->mmu_notifier, mm))
107 		goto error_mm;
108 
109 	return mm->hmm;
110 
111 error_mm:
112 	spin_lock(&mm->page_table_lock);
113 	if (mm->hmm == hmm)
114 		mm->hmm = NULL;
115 	spin_unlock(&mm->page_table_lock);
116 error:
117 	kfree(hmm);
118 	return NULL;
119 }
120 
121 void hmm_mm_destroy(struct mm_struct *mm)
122 {
123 	kfree(mm->hmm);
124 }
125 
126 static int hmm_invalidate_range(struct hmm *hmm, bool device,
127 				const struct hmm_update *update)
128 {
129 	struct hmm_mirror *mirror;
130 	struct hmm_range *range;
131 
132 	spin_lock(&hmm->lock);
133 	list_for_each_entry(range, &hmm->ranges, list) {
134 		unsigned long addr, idx, npages;
135 
136 		if (update->end < range->start || update->start >= range->end)
137 			continue;
138 
139 		range->valid = false;
140 		addr = max(update->start, range->start);
141 		idx = (addr - range->start) >> PAGE_SHIFT;
142 		npages = (min(range->end, update->end) - addr) >> PAGE_SHIFT;
143 		memset(&range->pfns[idx], 0, sizeof(*range->pfns) * npages);
144 	}
145 	spin_unlock(&hmm->lock);
146 
147 	if (!device)
148 		return 0;
149 
150 	down_read(&hmm->mirrors_sem);
151 	list_for_each_entry(mirror, &hmm->mirrors, list) {
152 		int ret;
153 
154 		ret = mirror->ops->sync_cpu_device_pagetables(mirror, update);
155 		if (!update->blockable && ret == -EAGAIN) {
156 			up_read(&hmm->mirrors_sem);
157 			return -EAGAIN;
158 		}
159 	}
160 	up_read(&hmm->mirrors_sem);
161 
162 	return 0;
163 }
164 
165 static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm)
166 {
167 	struct hmm_mirror *mirror;
168 	struct hmm *hmm = mm->hmm;
169 
170 	down_write(&hmm->mirrors_sem);
171 	mirror = list_first_entry_or_null(&hmm->mirrors, struct hmm_mirror,
172 					  list);
173 	while (mirror) {
174 		list_del_init(&mirror->list);
175 		if (mirror->ops->release) {
176 			/*
177 			 * Drop mirrors_sem so callback can wait on any pending
178 			 * work that might itself trigger mmu_notifier callback
179 			 * and thus would deadlock with us.
180 			 */
181 			up_write(&hmm->mirrors_sem);
182 			mirror->ops->release(mirror);
183 			down_write(&hmm->mirrors_sem);
184 		}
185 		mirror = list_first_entry_or_null(&hmm->mirrors,
186 						  struct hmm_mirror, list);
187 	}
188 	up_write(&hmm->mirrors_sem);
189 }
190 
191 static int hmm_invalidate_range_start(struct mmu_notifier *mn,
192 			const struct mmu_notifier_range *range)
193 {
194 	struct hmm_update update;
195 	struct hmm *hmm = range->mm->hmm;
196 
197 	VM_BUG_ON(!hmm);
198 
199 	update.start = range->start;
200 	update.end = range->end;
201 	update.event = HMM_UPDATE_INVALIDATE;
202 	update.blockable = range->blockable;
203 	return hmm_invalidate_range(hmm, true, &update);
204 }
205 
206 static void hmm_invalidate_range_end(struct mmu_notifier *mn,
207 			const struct mmu_notifier_range *range)
208 {
209 	struct hmm_update update;
210 	struct hmm *hmm = range->mm->hmm;
211 
212 	VM_BUG_ON(!hmm);
213 
214 	update.start = range->start;
215 	update.end = range->end;
216 	update.event = HMM_UPDATE_INVALIDATE;
217 	update.blockable = true;
218 	hmm_invalidate_range(hmm, false, &update);
219 }
220 
221 static const struct mmu_notifier_ops hmm_mmu_notifier_ops = {
222 	.release		= hmm_release,
223 	.invalidate_range_start	= hmm_invalidate_range_start,
224 	.invalidate_range_end	= hmm_invalidate_range_end,
225 };
226 
227 /*
228  * hmm_mirror_register() - register a mirror against an mm
229  *
230  * @mirror: new mirror struct to register
231  * @mm: mm to register against
232  *
233  * To start mirroring a process address space, the device driver must register
234  * an HMM mirror struct.
235  *
236  * THE mm->mmap_sem MUST BE HELD IN WRITE MODE !
237  */
238 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm)
239 {
240 	/* Sanity check */
241 	if (!mm || !mirror || !mirror->ops)
242 		return -EINVAL;
243 
244 again:
245 	mirror->hmm = hmm_register(mm);
246 	if (!mirror->hmm)
247 		return -ENOMEM;
248 
249 	down_write(&mirror->hmm->mirrors_sem);
250 	if (mirror->hmm->mm == NULL) {
251 		/*
252 		 * A racing hmm_mirror_unregister() is about to destroy the hmm
253 		 * struct. Try again to allocate a new one.
254 		 */
255 		up_write(&mirror->hmm->mirrors_sem);
256 		mirror->hmm = NULL;
257 		goto again;
258 	} else {
259 		list_add(&mirror->list, &mirror->hmm->mirrors);
260 		up_write(&mirror->hmm->mirrors_sem);
261 	}
262 
263 	return 0;
264 }
265 EXPORT_SYMBOL(hmm_mirror_register);
266 
267 /*
268  * hmm_mirror_unregister() - unregister a mirror
269  *
270  * @mirror: new mirror struct to register
271  *
272  * Stop mirroring a process address space, and cleanup.
273  */
274 void hmm_mirror_unregister(struct hmm_mirror *mirror)
275 {
276 	bool should_unregister = false;
277 	struct mm_struct *mm;
278 	struct hmm *hmm;
279 
280 	if (mirror->hmm == NULL)
281 		return;
282 
283 	hmm = mirror->hmm;
284 	down_write(&hmm->mirrors_sem);
285 	list_del_init(&mirror->list);
286 	should_unregister = list_empty(&hmm->mirrors);
287 	mirror->hmm = NULL;
288 	mm = hmm->mm;
289 	hmm->mm = NULL;
290 	up_write(&hmm->mirrors_sem);
291 
292 	if (!should_unregister || mm == NULL)
293 		return;
294 
295 	mmu_notifier_unregister_no_release(&hmm->mmu_notifier, mm);
296 
297 	spin_lock(&mm->page_table_lock);
298 	if (mm->hmm == hmm)
299 		mm->hmm = NULL;
300 	spin_unlock(&mm->page_table_lock);
301 
302 	kfree(hmm);
303 }
304 EXPORT_SYMBOL(hmm_mirror_unregister);
305 
306 struct hmm_vma_walk {
307 	struct hmm_range	*range;
308 	unsigned long		last;
309 	bool			fault;
310 	bool			block;
311 };
312 
313 static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr,
314 			    bool write_fault, uint64_t *pfn)
315 {
316 	unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE;
317 	struct hmm_vma_walk *hmm_vma_walk = walk->private;
318 	struct hmm_range *range = hmm_vma_walk->range;
319 	struct vm_area_struct *vma = walk->vma;
320 	vm_fault_t ret;
321 
322 	flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY;
323 	flags |= write_fault ? FAULT_FLAG_WRITE : 0;
324 	ret = handle_mm_fault(vma, addr, flags);
325 	if (ret & VM_FAULT_RETRY)
326 		return -EBUSY;
327 	if (ret & VM_FAULT_ERROR) {
328 		*pfn = range->values[HMM_PFN_ERROR];
329 		return -EFAULT;
330 	}
331 
332 	return -EAGAIN;
333 }
334 
335 static int hmm_pfns_bad(unsigned long addr,
336 			unsigned long end,
337 			struct mm_walk *walk)
338 {
339 	struct hmm_vma_walk *hmm_vma_walk = walk->private;
340 	struct hmm_range *range = hmm_vma_walk->range;
341 	uint64_t *pfns = range->pfns;
342 	unsigned long i;
343 
344 	i = (addr - range->start) >> PAGE_SHIFT;
345 	for (; addr < end; addr += PAGE_SIZE, i++)
346 		pfns[i] = range->values[HMM_PFN_ERROR];
347 
348 	return 0;
349 }
350 
351 /*
352  * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s)
353  * @start: range virtual start address (inclusive)
354  * @end: range virtual end address (exclusive)
355  * @fault: should we fault or not ?
356  * @write_fault: write fault ?
357  * @walk: mm_walk structure
358  * Returns: 0 on success, -EAGAIN after page fault, or page fault error
359  *
360  * This function will be called whenever pmd_none() or pte_none() returns true,
361  * or whenever there is no page directory covering the virtual address range.
362  */
363 static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end,
364 			      bool fault, bool write_fault,
365 			      struct mm_walk *walk)
366 {
367 	struct hmm_vma_walk *hmm_vma_walk = walk->private;
368 	struct hmm_range *range = hmm_vma_walk->range;
369 	uint64_t *pfns = range->pfns;
370 	unsigned long i;
371 
372 	hmm_vma_walk->last = addr;
373 	i = (addr - range->start) >> PAGE_SHIFT;
374 	for (; addr < end; addr += PAGE_SIZE, i++) {
375 		pfns[i] = range->values[HMM_PFN_NONE];
376 		if (fault || write_fault) {
377 			int ret;
378 
379 			ret = hmm_vma_do_fault(walk, addr, write_fault,
380 					       &pfns[i]);
381 			if (ret != -EAGAIN)
382 				return ret;
383 		}
384 	}
385 
386 	return (fault || write_fault) ? -EAGAIN : 0;
387 }
388 
389 static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
390 				      uint64_t pfns, uint64_t cpu_flags,
391 				      bool *fault, bool *write_fault)
392 {
393 	struct hmm_range *range = hmm_vma_walk->range;
394 
395 	*fault = *write_fault = false;
396 	if (!hmm_vma_walk->fault)
397 		return;
398 
399 	/* We aren't ask to do anything ... */
400 	if (!(pfns & range->flags[HMM_PFN_VALID]))
401 		return;
402 	/* If this is device memory than only fault if explicitly requested */
403 	if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) {
404 		/* Do we fault on device memory ? */
405 		if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) {
406 			*write_fault = pfns & range->flags[HMM_PFN_WRITE];
407 			*fault = true;
408 		}
409 		return;
410 	}
411 
412 	/* If CPU page table is not valid then we need to fault */
413 	*fault = !(cpu_flags & range->flags[HMM_PFN_VALID]);
414 	/* Need to write fault ? */
415 	if ((pfns & range->flags[HMM_PFN_WRITE]) &&
416 	    !(cpu_flags & range->flags[HMM_PFN_WRITE])) {
417 		*write_fault = true;
418 		*fault = true;
419 	}
420 }
421 
422 static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
423 				 const uint64_t *pfns, unsigned long npages,
424 				 uint64_t cpu_flags, bool *fault,
425 				 bool *write_fault)
426 {
427 	unsigned long i;
428 
429 	if (!hmm_vma_walk->fault) {
430 		*fault = *write_fault = false;
431 		return;
432 	}
433 
434 	for (i = 0; i < npages; ++i) {
435 		hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags,
436 				   fault, write_fault);
437 		if ((*fault) || (*write_fault))
438 			return;
439 	}
440 }
441 
442 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
443 			     struct mm_walk *walk)
444 {
445 	struct hmm_vma_walk *hmm_vma_walk = walk->private;
446 	struct hmm_range *range = hmm_vma_walk->range;
447 	bool fault, write_fault;
448 	unsigned long i, npages;
449 	uint64_t *pfns;
450 
451 	i = (addr - range->start) >> PAGE_SHIFT;
452 	npages = (end - addr) >> PAGE_SHIFT;
453 	pfns = &range->pfns[i];
454 	hmm_range_need_fault(hmm_vma_walk, pfns, npages,
455 			     0, &fault, &write_fault);
456 	return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
457 }
458 
459 static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd)
460 {
461 	if (pmd_protnone(pmd))
462 		return 0;
463 	return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] |
464 				range->flags[HMM_PFN_WRITE] :
465 				range->flags[HMM_PFN_VALID];
466 }
467 
468 static int hmm_vma_handle_pmd(struct mm_walk *walk,
469 			      unsigned long addr,
470 			      unsigned long end,
471 			      uint64_t *pfns,
472 			      pmd_t pmd)
473 {
474 	struct hmm_vma_walk *hmm_vma_walk = walk->private;
475 	struct hmm_range *range = hmm_vma_walk->range;
476 	unsigned long pfn, npages, i;
477 	bool fault, write_fault;
478 	uint64_t cpu_flags;
479 
480 	npages = (end - addr) >> PAGE_SHIFT;
481 	cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
482 	hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags,
483 			     &fault, &write_fault);
484 
485 	if (pmd_protnone(pmd) || fault || write_fault)
486 		return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
487 
488 	pfn = pmd_pfn(pmd) + pte_index(addr);
489 	for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++)
490 		pfns[i] = hmm_pfn_from_pfn(range, pfn) | cpu_flags;
491 	hmm_vma_walk->last = end;
492 	return 0;
493 }
494 
495 static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte)
496 {
497 	if (pte_none(pte) || !pte_present(pte))
498 		return 0;
499 	return pte_write(pte) ? range->flags[HMM_PFN_VALID] |
500 				range->flags[HMM_PFN_WRITE] :
501 				range->flags[HMM_PFN_VALID];
502 }
503 
504 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
505 			      unsigned long end, pmd_t *pmdp, pte_t *ptep,
506 			      uint64_t *pfn)
507 {
508 	struct hmm_vma_walk *hmm_vma_walk = walk->private;
509 	struct hmm_range *range = hmm_vma_walk->range;
510 	struct vm_area_struct *vma = walk->vma;
511 	bool fault, write_fault;
512 	uint64_t cpu_flags;
513 	pte_t pte = *ptep;
514 	uint64_t orig_pfn = *pfn;
515 
516 	*pfn = range->values[HMM_PFN_NONE];
517 	cpu_flags = pte_to_hmm_pfn_flags(range, pte);
518 	hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
519 			   &fault, &write_fault);
520 
521 	if (pte_none(pte)) {
522 		if (fault || write_fault)
523 			goto fault;
524 		return 0;
525 	}
526 
527 	if (!pte_present(pte)) {
528 		swp_entry_t entry = pte_to_swp_entry(pte);
529 
530 		if (!non_swap_entry(entry)) {
531 			if (fault || write_fault)
532 				goto fault;
533 			return 0;
534 		}
535 
536 		/*
537 		 * This is a special swap entry, ignore migration, use
538 		 * device and report anything else as error.
539 		 */
540 		if (is_device_private_entry(entry)) {
541 			cpu_flags = range->flags[HMM_PFN_VALID] |
542 				range->flags[HMM_PFN_DEVICE_PRIVATE];
543 			cpu_flags |= is_write_device_private_entry(entry) ?
544 				range->flags[HMM_PFN_WRITE] : 0;
545 			hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
546 					   &fault, &write_fault);
547 			if (fault || write_fault)
548 				goto fault;
549 			*pfn = hmm_pfn_from_pfn(range, swp_offset(entry));
550 			*pfn |= cpu_flags;
551 			return 0;
552 		}
553 
554 		if (is_migration_entry(entry)) {
555 			if (fault || write_fault) {
556 				pte_unmap(ptep);
557 				hmm_vma_walk->last = addr;
558 				migration_entry_wait(vma->vm_mm,
559 						     pmdp, addr);
560 				return -EAGAIN;
561 			}
562 			return 0;
563 		}
564 
565 		/* Report error for everything else */
566 		*pfn = range->values[HMM_PFN_ERROR];
567 		return -EFAULT;
568 	}
569 
570 	if (fault || write_fault)
571 		goto fault;
572 
573 	*pfn = hmm_pfn_from_pfn(range, pte_pfn(pte)) | cpu_flags;
574 	return 0;
575 
576 fault:
577 	pte_unmap(ptep);
578 	/* Fault any virtual address we were asked to fault */
579 	return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
580 }
581 
582 static int hmm_vma_walk_pmd(pmd_t *pmdp,
583 			    unsigned long start,
584 			    unsigned long end,
585 			    struct mm_walk *walk)
586 {
587 	struct hmm_vma_walk *hmm_vma_walk = walk->private;
588 	struct hmm_range *range = hmm_vma_walk->range;
589 	struct vm_area_struct *vma = walk->vma;
590 	uint64_t *pfns = range->pfns;
591 	unsigned long addr = start, i;
592 	pte_t *ptep;
593 	pmd_t pmd;
594 
595 
596 again:
597 	pmd = READ_ONCE(*pmdp);
598 	if (pmd_none(pmd))
599 		return hmm_vma_walk_hole(start, end, walk);
600 
601 	if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB))
602 		return hmm_pfns_bad(start, end, walk);
603 
604 	if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
605 		bool fault, write_fault;
606 		unsigned long npages;
607 		uint64_t *pfns;
608 
609 		i = (addr - range->start) >> PAGE_SHIFT;
610 		npages = (end - addr) >> PAGE_SHIFT;
611 		pfns = &range->pfns[i];
612 
613 		hmm_range_need_fault(hmm_vma_walk, pfns, npages,
614 				     0, &fault, &write_fault);
615 		if (fault || write_fault) {
616 			hmm_vma_walk->last = addr;
617 			pmd_migration_entry_wait(vma->vm_mm, pmdp);
618 			return -EAGAIN;
619 		}
620 		return 0;
621 	} else if (!pmd_present(pmd))
622 		return hmm_pfns_bad(start, end, walk);
623 
624 	if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
625 		/*
626 		 * No need to take pmd_lock here, even if some other threads
627 		 * is splitting the huge pmd we will get that event through
628 		 * mmu_notifier callback.
629 		 *
630 		 * So just read pmd value and check again its a transparent
631 		 * huge or device mapping one and compute corresponding pfn
632 		 * values.
633 		 */
634 		pmd = pmd_read_atomic(pmdp);
635 		barrier();
636 		if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
637 			goto again;
638 
639 		i = (addr - range->start) >> PAGE_SHIFT;
640 		return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd);
641 	}
642 
643 	/*
644 	 * We have handled all the valid case above ie either none, migration,
645 	 * huge or transparent huge. At this point either it is a valid pmd
646 	 * entry pointing to pte directory or it is a bad pmd that will not
647 	 * recover.
648 	 */
649 	if (pmd_bad(pmd))
650 		return hmm_pfns_bad(start, end, walk);
651 
652 	ptep = pte_offset_map(pmdp, addr);
653 	i = (addr - range->start) >> PAGE_SHIFT;
654 	for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
655 		int r;
656 
657 		r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]);
658 		if (r) {
659 			/* hmm_vma_handle_pte() did unmap pte directory */
660 			hmm_vma_walk->last = addr;
661 			return r;
662 		}
663 	}
664 	pte_unmap(ptep - 1);
665 
666 	hmm_vma_walk->last = addr;
667 	return 0;
668 }
669 
670 static void hmm_pfns_clear(struct hmm_range *range,
671 			   uint64_t *pfns,
672 			   unsigned long addr,
673 			   unsigned long end)
674 {
675 	for (; addr < end; addr += PAGE_SIZE, pfns++)
676 		*pfns = range->values[HMM_PFN_NONE];
677 }
678 
679 static void hmm_pfns_special(struct hmm_range *range)
680 {
681 	unsigned long addr = range->start, i = 0;
682 
683 	for (; addr < range->end; addr += PAGE_SIZE, i++)
684 		range->pfns[i] = range->values[HMM_PFN_SPECIAL];
685 }
686 
687 /*
688  * hmm_vma_get_pfns() - snapshot CPU page table for a range of virtual addresses
689  * @range: range being snapshotted
690  * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid
691  *          vma permission, 0 success
692  *
693  * This snapshots the CPU page table for a range of virtual addresses. Snapshot
694  * validity is tracked by range struct. See hmm_vma_range_done() for further
695  * information.
696  *
697  * The range struct is initialized here. It tracks the CPU page table, but only
698  * if the function returns success (0), in which case the caller must then call
699  * hmm_vma_range_done() to stop CPU page table update tracking on this range.
700  *
701  * NOT CALLING hmm_vma_range_done() IF FUNCTION RETURNS 0 WILL LEAD TO SERIOUS
702  * MEMORY CORRUPTION ! YOU HAVE BEEN WARNED !
703  */
704 int hmm_vma_get_pfns(struct hmm_range *range)
705 {
706 	struct vm_area_struct *vma = range->vma;
707 	struct hmm_vma_walk hmm_vma_walk;
708 	struct mm_walk mm_walk;
709 	struct hmm *hmm;
710 
711 	/* Sanity check, this really should not happen ! */
712 	if (range->start < vma->vm_start || range->start >= vma->vm_end)
713 		return -EINVAL;
714 	if (range->end < vma->vm_start || range->end > vma->vm_end)
715 		return -EINVAL;
716 
717 	hmm = hmm_register(vma->vm_mm);
718 	if (!hmm)
719 		return -ENOMEM;
720 	/* Caller must have registered a mirror, via hmm_mirror_register() ! */
721 	if (!hmm->mmu_notifier.ops)
722 		return -EINVAL;
723 
724 	/* FIXME support hugetlb fs */
725 	if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
726 			vma_is_dax(vma)) {
727 		hmm_pfns_special(range);
728 		return -EINVAL;
729 	}
730 
731 	if (!(vma->vm_flags & VM_READ)) {
732 		/*
733 		 * If vma do not allow read access, then assume that it does
734 		 * not allow write access, either. Architecture that allow
735 		 * write without read access are not supported by HMM, because
736 		 * operations such has atomic access would not work.
737 		 */
738 		hmm_pfns_clear(range, range->pfns, range->start, range->end);
739 		return -EPERM;
740 	}
741 
742 	/* Initialize range to track CPU page table update */
743 	spin_lock(&hmm->lock);
744 	range->valid = true;
745 	list_add_rcu(&range->list, &hmm->ranges);
746 	spin_unlock(&hmm->lock);
747 
748 	hmm_vma_walk.fault = false;
749 	hmm_vma_walk.range = range;
750 	mm_walk.private = &hmm_vma_walk;
751 
752 	mm_walk.vma = vma;
753 	mm_walk.mm = vma->vm_mm;
754 	mm_walk.pte_entry = NULL;
755 	mm_walk.test_walk = NULL;
756 	mm_walk.hugetlb_entry = NULL;
757 	mm_walk.pmd_entry = hmm_vma_walk_pmd;
758 	mm_walk.pte_hole = hmm_vma_walk_hole;
759 
760 	walk_page_range(range->start, range->end, &mm_walk);
761 	return 0;
762 }
763 EXPORT_SYMBOL(hmm_vma_get_pfns);
764 
765 /*
766  * hmm_vma_range_done() - stop tracking change to CPU page table over a range
767  * @range: range being tracked
768  * Returns: false if range data has been invalidated, true otherwise
769  *
770  * Range struct is used to track updates to the CPU page table after a call to
771  * either hmm_vma_get_pfns() or hmm_vma_fault(). Once the device driver is done
772  * using the data,  or wants to lock updates to the data it got from those
773  * functions, it must call the hmm_vma_range_done() function, which will then
774  * stop tracking CPU page table updates.
775  *
776  * Note that device driver must still implement general CPU page table update
777  * tracking either by using hmm_mirror (see hmm_mirror_register()) or by using
778  * the mmu_notifier API directly.
779  *
780  * CPU page table update tracking done through hmm_range is only temporary and
781  * to be used while trying to duplicate CPU page table contents for a range of
782  * virtual addresses.
783  *
784  * There are two ways to use this :
785  * again:
786  *   hmm_vma_get_pfns(range); or hmm_vma_fault(...);
787  *   trans = device_build_page_table_update_transaction(pfns);
788  *   device_page_table_lock();
789  *   if (!hmm_vma_range_done(range)) {
790  *     device_page_table_unlock();
791  *     goto again;
792  *   }
793  *   device_commit_transaction(trans);
794  *   device_page_table_unlock();
795  *
796  * Or:
797  *   hmm_vma_get_pfns(range); or hmm_vma_fault(...);
798  *   device_page_table_lock();
799  *   hmm_vma_range_done(range);
800  *   device_update_page_table(range->pfns);
801  *   device_page_table_unlock();
802  */
803 bool hmm_vma_range_done(struct hmm_range *range)
804 {
805 	unsigned long npages = (range->end - range->start) >> PAGE_SHIFT;
806 	struct hmm *hmm;
807 
808 	if (range->end <= range->start) {
809 		BUG();
810 		return false;
811 	}
812 
813 	hmm = hmm_register(range->vma->vm_mm);
814 	if (!hmm) {
815 		memset(range->pfns, 0, sizeof(*range->pfns) * npages);
816 		return false;
817 	}
818 
819 	spin_lock(&hmm->lock);
820 	list_del_rcu(&range->list);
821 	spin_unlock(&hmm->lock);
822 
823 	return range->valid;
824 }
825 EXPORT_SYMBOL(hmm_vma_range_done);
826 
827 /*
828  * hmm_vma_fault() - try to fault some address in a virtual address range
829  * @range: range being faulted
830  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
831  * Returns: 0 success, error otherwise (-EAGAIN means mmap_sem have been drop)
832  *
833  * This is similar to a regular CPU page fault except that it will not trigger
834  * any memory migration if the memory being faulted is not accessible by CPUs.
835  *
836  * On error, for one virtual address in the range, the function will mark the
837  * corresponding HMM pfn entry with an error flag.
838  *
839  * Expected use pattern:
840  * retry:
841  *   down_read(&mm->mmap_sem);
842  *   // Find vma and address device wants to fault, initialize hmm_pfn_t
843  *   // array accordingly
844  *   ret = hmm_vma_fault(range, write, block);
845  *   switch (ret) {
846  *   case -EAGAIN:
847  *     hmm_vma_range_done(range);
848  *     // You might want to rate limit or yield to play nicely, you may
849  *     // also commit any valid pfn in the array assuming that you are
850  *     // getting true from hmm_vma_range_monitor_end()
851  *     goto retry;
852  *   case 0:
853  *     break;
854  *   case -ENOMEM:
855  *   case -EINVAL:
856  *   case -EPERM:
857  *   default:
858  *     // Handle error !
859  *     up_read(&mm->mmap_sem)
860  *     return;
861  *   }
862  *   // Take device driver lock that serialize device page table update
863  *   driver_lock_device_page_table_update();
864  *   hmm_vma_range_done(range);
865  *   // Commit pfns we got from hmm_vma_fault()
866  *   driver_unlock_device_page_table_update();
867  *   up_read(&mm->mmap_sem)
868  *
869  * YOU MUST CALL hmm_vma_range_done() AFTER THIS FUNCTION RETURN SUCCESS (0)
870  * BEFORE FREEING THE range struct OR YOU WILL HAVE SERIOUS MEMORY CORRUPTION !
871  *
872  * YOU HAVE BEEN WARNED !
873  */
874 int hmm_vma_fault(struct hmm_range *range, bool block)
875 {
876 	struct vm_area_struct *vma = range->vma;
877 	unsigned long start = range->start;
878 	struct hmm_vma_walk hmm_vma_walk;
879 	struct mm_walk mm_walk;
880 	struct hmm *hmm;
881 	int ret;
882 
883 	/* Sanity check, this really should not happen ! */
884 	if (range->start < vma->vm_start || range->start >= vma->vm_end)
885 		return -EINVAL;
886 	if (range->end < vma->vm_start || range->end > vma->vm_end)
887 		return -EINVAL;
888 
889 	hmm = hmm_register(vma->vm_mm);
890 	if (!hmm) {
891 		hmm_pfns_clear(range, range->pfns, range->start, range->end);
892 		return -ENOMEM;
893 	}
894 	/* Caller must have registered a mirror using hmm_mirror_register() */
895 	if (!hmm->mmu_notifier.ops)
896 		return -EINVAL;
897 
898 	/* FIXME support hugetlb fs */
899 	if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
900 			vma_is_dax(vma)) {
901 		hmm_pfns_special(range);
902 		return -EINVAL;
903 	}
904 
905 	if (!(vma->vm_flags & VM_READ)) {
906 		/*
907 		 * If vma do not allow read access, then assume that it does
908 		 * not allow write access, either. Architecture that allow
909 		 * write without read access are not supported by HMM, because
910 		 * operations such has atomic access would not work.
911 		 */
912 		hmm_pfns_clear(range, range->pfns, range->start, range->end);
913 		return -EPERM;
914 	}
915 
916 	/* Initialize range to track CPU page table update */
917 	spin_lock(&hmm->lock);
918 	range->valid = true;
919 	list_add_rcu(&range->list, &hmm->ranges);
920 	spin_unlock(&hmm->lock);
921 
922 	hmm_vma_walk.fault = true;
923 	hmm_vma_walk.block = block;
924 	hmm_vma_walk.range = range;
925 	mm_walk.private = &hmm_vma_walk;
926 	hmm_vma_walk.last = range->start;
927 
928 	mm_walk.vma = vma;
929 	mm_walk.mm = vma->vm_mm;
930 	mm_walk.pte_entry = NULL;
931 	mm_walk.test_walk = NULL;
932 	mm_walk.hugetlb_entry = NULL;
933 	mm_walk.pmd_entry = hmm_vma_walk_pmd;
934 	mm_walk.pte_hole = hmm_vma_walk_hole;
935 
936 	do {
937 		ret = walk_page_range(start, range->end, &mm_walk);
938 		start = hmm_vma_walk.last;
939 	} while (ret == -EAGAIN);
940 
941 	if (ret) {
942 		unsigned long i;
943 
944 		i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
945 		hmm_pfns_clear(range, &range->pfns[i], hmm_vma_walk.last,
946 			       range->end);
947 		hmm_vma_range_done(range);
948 	}
949 	return ret;
950 }
951 EXPORT_SYMBOL(hmm_vma_fault);
952 #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
953 
954 
955 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) ||  IS_ENABLED(CONFIG_DEVICE_PUBLIC)
956 struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
957 				       unsigned long addr)
958 {
959 	struct page *page;
960 
961 	page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
962 	if (!page)
963 		return NULL;
964 	lock_page(page);
965 	return page;
966 }
967 EXPORT_SYMBOL(hmm_vma_alloc_locked_page);
968 
969 
970 static void hmm_devmem_ref_release(struct percpu_ref *ref)
971 {
972 	struct hmm_devmem *devmem;
973 
974 	devmem = container_of(ref, struct hmm_devmem, ref);
975 	complete(&devmem->completion);
976 }
977 
978 static void hmm_devmem_ref_exit(void *data)
979 {
980 	struct percpu_ref *ref = data;
981 	struct hmm_devmem *devmem;
982 
983 	devmem = container_of(ref, struct hmm_devmem, ref);
984 	wait_for_completion(&devmem->completion);
985 	percpu_ref_exit(ref);
986 }
987 
988 static void hmm_devmem_ref_kill(struct percpu_ref *ref)
989 {
990 	percpu_ref_kill(ref);
991 }
992 
993 static int hmm_devmem_fault(struct vm_area_struct *vma,
994 			    unsigned long addr,
995 			    const struct page *page,
996 			    unsigned int flags,
997 			    pmd_t *pmdp)
998 {
999 	struct hmm_devmem *devmem = page->pgmap->data;
1000 
1001 	return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp);
1002 }
1003 
1004 static void hmm_devmem_free(struct page *page, void *data)
1005 {
1006 	struct hmm_devmem *devmem = data;
1007 
1008 	page->mapping = NULL;
1009 
1010 	devmem->ops->free(devmem, page);
1011 }
1012 
1013 /*
1014  * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory
1015  *
1016  * @ops: memory event device driver callback (see struct hmm_devmem_ops)
1017  * @device: device struct to bind the resource too
1018  * @size: size in bytes of the device memory to add
1019  * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise
1020  *
1021  * This function first finds an empty range of physical address big enough to
1022  * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which
1023  * in turn allocates struct pages. It does not do anything beyond that; all
1024  * events affecting the memory will go through the various callbacks provided
1025  * by hmm_devmem_ops struct.
1026  *
1027  * Device driver should call this function during device initialization and
1028  * is then responsible of memory management. HMM only provides helpers.
1029  */
1030 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
1031 				  struct device *device,
1032 				  unsigned long size)
1033 {
1034 	struct hmm_devmem *devmem;
1035 	resource_size_t addr;
1036 	void *result;
1037 	int ret;
1038 
1039 	dev_pagemap_get_ops();
1040 
1041 	devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL);
1042 	if (!devmem)
1043 		return ERR_PTR(-ENOMEM);
1044 
1045 	init_completion(&devmem->completion);
1046 	devmem->pfn_first = -1UL;
1047 	devmem->pfn_last = -1UL;
1048 	devmem->resource = NULL;
1049 	devmem->device = device;
1050 	devmem->ops = ops;
1051 
1052 	ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1053 			      0, GFP_KERNEL);
1054 	if (ret)
1055 		return ERR_PTR(ret);
1056 
1057 	ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit, &devmem->ref);
1058 	if (ret)
1059 		return ERR_PTR(ret);
1060 
1061 	size = ALIGN(size, PA_SECTION_SIZE);
1062 	addr = min((unsigned long)iomem_resource.end,
1063 		   (1UL << MAX_PHYSMEM_BITS) - 1);
1064 	addr = addr - size + 1UL;
1065 
1066 	/*
1067 	 * FIXME add a new helper to quickly walk resource tree and find free
1068 	 * range
1069 	 *
1070 	 * FIXME what about ioport_resource resource ?
1071 	 */
1072 	for (; addr > size && addr >= iomem_resource.start; addr -= size) {
1073 		ret = region_intersects(addr, size, 0, IORES_DESC_NONE);
1074 		if (ret != REGION_DISJOINT)
1075 			continue;
1076 
1077 		devmem->resource = devm_request_mem_region(device, addr, size,
1078 							   dev_name(device));
1079 		if (!devmem->resource)
1080 			return ERR_PTR(-ENOMEM);
1081 		break;
1082 	}
1083 	if (!devmem->resource)
1084 		return ERR_PTR(-ERANGE);
1085 
1086 	devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY;
1087 	devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1088 	devmem->pfn_last = devmem->pfn_first +
1089 			   (resource_size(devmem->resource) >> PAGE_SHIFT);
1090 	devmem->page_fault = hmm_devmem_fault;
1091 
1092 	devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
1093 	devmem->pagemap.res = *devmem->resource;
1094 	devmem->pagemap.page_free = hmm_devmem_free;
1095 	devmem->pagemap.altmap_valid = false;
1096 	devmem->pagemap.ref = &devmem->ref;
1097 	devmem->pagemap.data = devmem;
1098 	devmem->pagemap.kill = hmm_devmem_ref_kill;
1099 
1100 	result = devm_memremap_pages(devmem->device, &devmem->pagemap);
1101 	if (IS_ERR(result))
1102 		return result;
1103 	return devmem;
1104 }
1105 EXPORT_SYMBOL_GPL(hmm_devmem_add);
1106 
1107 struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
1108 					   struct device *device,
1109 					   struct resource *res)
1110 {
1111 	struct hmm_devmem *devmem;
1112 	void *result;
1113 	int ret;
1114 
1115 	if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY)
1116 		return ERR_PTR(-EINVAL);
1117 
1118 	dev_pagemap_get_ops();
1119 
1120 	devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL);
1121 	if (!devmem)
1122 		return ERR_PTR(-ENOMEM);
1123 
1124 	init_completion(&devmem->completion);
1125 	devmem->pfn_first = -1UL;
1126 	devmem->pfn_last = -1UL;
1127 	devmem->resource = res;
1128 	devmem->device = device;
1129 	devmem->ops = ops;
1130 
1131 	ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1132 			      0, GFP_KERNEL);
1133 	if (ret)
1134 		return ERR_PTR(ret);
1135 
1136 	ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit,
1137 			&devmem->ref);
1138 	if (ret)
1139 		return ERR_PTR(ret);
1140 
1141 	devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1142 	devmem->pfn_last = devmem->pfn_first +
1143 			   (resource_size(devmem->resource) >> PAGE_SHIFT);
1144 	devmem->page_fault = hmm_devmem_fault;
1145 
1146 	devmem->pagemap.type = MEMORY_DEVICE_PUBLIC;
1147 	devmem->pagemap.res = *devmem->resource;
1148 	devmem->pagemap.page_free = hmm_devmem_free;
1149 	devmem->pagemap.altmap_valid = false;
1150 	devmem->pagemap.ref = &devmem->ref;
1151 	devmem->pagemap.data = devmem;
1152 	devmem->pagemap.kill = hmm_devmem_ref_kill;
1153 
1154 	result = devm_memremap_pages(devmem->device, &devmem->pagemap);
1155 	if (IS_ERR(result))
1156 		return result;
1157 	return devmem;
1158 }
1159 EXPORT_SYMBOL_GPL(hmm_devmem_add_resource);
1160 
1161 /*
1162  * A device driver that wants to handle multiple devices memory through a
1163  * single fake device can use hmm_device to do so. This is purely a helper
1164  * and it is not needed to make use of any HMM functionality.
1165  */
1166 #define HMM_DEVICE_MAX 256
1167 
1168 static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX);
1169 static DEFINE_SPINLOCK(hmm_device_lock);
1170 static struct class *hmm_device_class;
1171 static dev_t hmm_device_devt;
1172 
1173 static void hmm_device_release(struct device *device)
1174 {
1175 	struct hmm_device *hmm_device;
1176 
1177 	hmm_device = container_of(device, struct hmm_device, device);
1178 	spin_lock(&hmm_device_lock);
1179 	clear_bit(hmm_device->minor, hmm_device_mask);
1180 	spin_unlock(&hmm_device_lock);
1181 
1182 	kfree(hmm_device);
1183 }
1184 
1185 struct hmm_device *hmm_device_new(void *drvdata)
1186 {
1187 	struct hmm_device *hmm_device;
1188 
1189 	hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL);
1190 	if (!hmm_device)
1191 		return ERR_PTR(-ENOMEM);
1192 
1193 	spin_lock(&hmm_device_lock);
1194 	hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX);
1195 	if (hmm_device->minor >= HMM_DEVICE_MAX) {
1196 		spin_unlock(&hmm_device_lock);
1197 		kfree(hmm_device);
1198 		return ERR_PTR(-EBUSY);
1199 	}
1200 	set_bit(hmm_device->minor, hmm_device_mask);
1201 	spin_unlock(&hmm_device_lock);
1202 
1203 	dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor);
1204 	hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt),
1205 					hmm_device->minor);
1206 	hmm_device->device.release = hmm_device_release;
1207 	dev_set_drvdata(&hmm_device->device, drvdata);
1208 	hmm_device->device.class = hmm_device_class;
1209 	device_initialize(&hmm_device->device);
1210 
1211 	return hmm_device;
1212 }
1213 EXPORT_SYMBOL(hmm_device_new);
1214 
1215 void hmm_device_put(struct hmm_device *hmm_device)
1216 {
1217 	put_device(&hmm_device->device);
1218 }
1219 EXPORT_SYMBOL(hmm_device_put);
1220 
1221 static int __init hmm_init(void)
1222 {
1223 	int ret;
1224 
1225 	ret = alloc_chrdev_region(&hmm_device_devt, 0,
1226 				  HMM_DEVICE_MAX,
1227 				  "hmm_device");
1228 	if (ret)
1229 		return ret;
1230 
1231 	hmm_device_class = class_create(THIS_MODULE, "hmm_device");
1232 	if (IS_ERR(hmm_device_class)) {
1233 		unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX);
1234 		return PTR_ERR(hmm_device_class);
1235 	}
1236 	return 0;
1237 }
1238 
1239 device_initcall(hmm_init);
1240 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */
1241