xref: /linux/mm/khugepaged.c (revision c0b2826aa02a0747c3016457b9c1917ba14d0655)
1 // SPDX-License-Identifier: GPL-2.0
2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
3 
4 #include <linux/mm.h>
5 #include <linux/sched.h>
6 #include <linux/sched/mm.h>
7 #include <linux/sched/coredump.h>
8 #include <linux/mmu_notifier.h>
9 #include <linux/rmap.h>
10 #include <linux/swap.h>
11 #include <linux/mm_inline.h>
12 #include <linux/kthread.h>
13 #include <linux/khugepaged.h>
14 #include <linux/freezer.h>
15 #include <linux/mman.h>
16 #include <linux/hashtable.h>
17 #include <linux/userfaultfd_k.h>
18 #include <linux/page_idle.h>
19 #include <linux/swapops.h>
20 #include <linux/shmem_fs.h>
21 
22 #include <asm/tlb.h>
23 #include <asm/pgalloc.h>
24 #include "internal.h"
25 
26 enum scan_result {
27 	SCAN_FAIL,
28 	SCAN_SUCCEED,
29 	SCAN_PMD_NULL,
30 	SCAN_EXCEED_NONE_PTE,
31 	SCAN_PTE_NON_PRESENT,
32 	SCAN_PAGE_RO,
33 	SCAN_LACK_REFERENCED_PAGE,
34 	SCAN_PAGE_NULL,
35 	SCAN_SCAN_ABORT,
36 	SCAN_PAGE_COUNT,
37 	SCAN_PAGE_LRU,
38 	SCAN_PAGE_LOCK,
39 	SCAN_PAGE_ANON,
40 	SCAN_PAGE_COMPOUND,
41 	SCAN_ANY_PROCESS,
42 	SCAN_VMA_NULL,
43 	SCAN_VMA_CHECK,
44 	SCAN_ADDRESS_RANGE,
45 	SCAN_SWAP_CACHE_PAGE,
46 	SCAN_DEL_PAGE_LRU,
47 	SCAN_ALLOC_HUGE_PAGE_FAIL,
48 	SCAN_CGROUP_CHARGE_FAIL,
49 	SCAN_EXCEED_SWAP_PTE,
50 	SCAN_TRUNCATED,
51 };
52 
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/huge_memory.h>
55 
56 /* default scan 8*512 pte (or vmas) every 30 second */
57 static unsigned int khugepaged_pages_to_scan __read_mostly;
58 static unsigned int khugepaged_pages_collapsed;
59 static unsigned int khugepaged_full_scans;
60 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
61 /* during fragmentation poll the hugepage allocator once every minute */
62 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
63 static unsigned long khugepaged_sleep_expire;
64 static DEFINE_SPINLOCK(khugepaged_mm_lock);
65 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
66 /*
67  * default collapse hugepages if there is at least one pte mapped like
68  * it would have happened if the vma was large enough during page
69  * fault.
70  */
71 static unsigned int khugepaged_max_ptes_none __read_mostly;
72 static unsigned int khugepaged_max_ptes_swap __read_mostly;
73 
74 #define MM_SLOTS_HASH_BITS 10
75 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
76 
77 static struct kmem_cache *mm_slot_cache __read_mostly;
78 
79 /**
80  * struct mm_slot - hash lookup from mm to mm_slot
81  * @hash: hash collision list
82  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
83  * @mm: the mm that this information is valid for
84  */
85 struct mm_slot {
86 	struct hlist_node hash;
87 	struct list_head mm_node;
88 	struct mm_struct *mm;
89 };
90 
91 /**
92  * struct khugepaged_scan - cursor for scanning
93  * @mm_head: the head of the mm list to scan
94  * @mm_slot: the current mm_slot we are scanning
95  * @address: the next address inside that to be scanned
96  *
97  * There is only the one khugepaged_scan instance of this cursor structure.
98  */
99 struct khugepaged_scan {
100 	struct list_head mm_head;
101 	struct mm_slot *mm_slot;
102 	unsigned long address;
103 };
104 
105 static struct khugepaged_scan khugepaged_scan = {
106 	.mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
107 };
108 
109 #ifdef CONFIG_SYSFS
110 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
111 					 struct kobj_attribute *attr,
112 					 char *buf)
113 {
114 	return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
115 }
116 
117 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
118 					  struct kobj_attribute *attr,
119 					  const char *buf, size_t count)
120 {
121 	unsigned long msecs;
122 	int err;
123 
124 	err = kstrtoul(buf, 10, &msecs);
125 	if (err || msecs > UINT_MAX)
126 		return -EINVAL;
127 
128 	khugepaged_scan_sleep_millisecs = msecs;
129 	khugepaged_sleep_expire = 0;
130 	wake_up_interruptible(&khugepaged_wait);
131 
132 	return count;
133 }
134 static struct kobj_attribute scan_sleep_millisecs_attr =
135 	__ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
136 	       scan_sleep_millisecs_store);
137 
138 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
139 					  struct kobj_attribute *attr,
140 					  char *buf)
141 {
142 	return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
143 }
144 
145 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
146 					   struct kobj_attribute *attr,
147 					   const char *buf, size_t count)
148 {
149 	unsigned long msecs;
150 	int err;
151 
152 	err = kstrtoul(buf, 10, &msecs);
153 	if (err || msecs > UINT_MAX)
154 		return -EINVAL;
155 
156 	khugepaged_alloc_sleep_millisecs = msecs;
157 	khugepaged_sleep_expire = 0;
158 	wake_up_interruptible(&khugepaged_wait);
159 
160 	return count;
161 }
162 static struct kobj_attribute alloc_sleep_millisecs_attr =
163 	__ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
164 	       alloc_sleep_millisecs_store);
165 
166 static ssize_t pages_to_scan_show(struct kobject *kobj,
167 				  struct kobj_attribute *attr,
168 				  char *buf)
169 {
170 	return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
171 }
172 static ssize_t pages_to_scan_store(struct kobject *kobj,
173 				   struct kobj_attribute *attr,
174 				   const char *buf, size_t count)
175 {
176 	int err;
177 	unsigned long pages;
178 
179 	err = kstrtoul(buf, 10, &pages);
180 	if (err || !pages || pages > UINT_MAX)
181 		return -EINVAL;
182 
183 	khugepaged_pages_to_scan = pages;
184 
185 	return count;
186 }
187 static struct kobj_attribute pages_to_scan_attr =
188 	__ATTR(pages_to_scan, 0644, pages_to_scan_show,
189 	       pages_to_scan_store);
190 
191 static ssize_t pages_collapsed_show(struct kobject *kobj,
192 				    struct kobj_attribute *attr,
193 				    char *buf)
194 {
195 	return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
196 }
197 static struct kobj_attribute pages_collapsed_attr =
198 	__ATTR_RO(pages_collapsed);
199 
200 static ssize_t full_scans_show(struct kobject *kobj,
201 			       struct kobj_attribute *attr,
202 			       char *buf)
203 {
204 	return sprintf(buf, "%u\n", khugepaged_full_scans);
205 }
206 static struct kobj_attribute full_scans_attr =
207 	__ATTR_RO(full_scans);
208 
209 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
210 				      struct kobj_attribute *attr, char *buf)
211 {
212 	return single_hugepage_flag_show(kobj, attr, buf,
213 				TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
214 }
215 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
216 				       struct kobj_attribute *attr,
217 				       const char *buf, size_t count)
218 {
219 	return single_hugepage_flag_store(kobj, attr, buf, count,
220 				 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
221 }
222 static struct kobj_attribute khugepaged_defrag_attr =
223 	__ATTR(defrag, 0644, khugepaged_defrag_show,
224 	       khugepaged_defrag_store);
225 
226 /*
227  * max_ptes_none controls if khugepaged should collapse hugepages over
228  * any unmapped ptes in turn potentially increasing the memory
229  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
230  * reduce the available free memory in the system as it
231  * runs. Increasing max_ptes_none will instead potentially reduce the
232  * free memory in the system during the khugepaged scan.
233  */
234 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
235 					     struct kobj_attribute *attr,
236 					     char *buf)
237 {
238 	return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
239 }
240 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
241 					      struct kobj_attribute *attr,
242 					      const char *buf, size_t count)
243 {
244 	int err;
245 	unsigned long max_ptes_none;
246 
247 	err = kstrtoul(buf, 10, &max_ptes_none);
248 	if (err || max_ptes_none > HPAGE_PMD_NR-1)
249 		return -EINVAL;
250 
251 	khugepaged_max_ptes_none = max_ptes_none;
252 
253 	return count;
254 }
255 static struct kobj_attribute khugepaged_max_ptes_none_attr =
256 	__ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
257 	       khugepaged_max_ptes_none_store);
258 
259 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
260 					     struct kobj_attribute *attr,
261 					     char *buf)
262 {
263 	return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
264 }
265 
266 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
267 					      struct kobj_attribute *attr,
268 					      const char *buf, size_t count)
269 {
270 	int err;
271 	unsigned long max_ptes_swap;
272 
273 	err  = kstrtoul(buf, 10, &max_ptes_swap);
274 	if (err || max_ptes_swap > HPAGE_PMD_NR-1)
275 		return -EINVAL;
276 
277 	khugepaged_max_ptes_swap = max_ptes_swap;
278 
279 	return count;
280 }
281 
282 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
283 	__ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
284 	       khugepaged_max_ptes_swap_store);
285 
286 static struct attribute *khugepaged_attr[] = {
287 	&khugepaged_defrag_attr.attr,
288 	&khugepaged_max_ptes_none_attr.attr,
289 	&pages_to_scan_attr.attr,
290 	&pages_collapsed_attr.attr,
291 	&full_scans_attr.attr,
292 	&scan_sleep_millisecs_attr.attr,
293 	&alloc_sleep_millisecs_attr.attr,
294 	&khugepaged_max_ptes_swap_attr.attr,
295 	NULL,
296 };
297 
298 struct attribute_group khugepaged_attr_group = {
299 	.attrs = khugepaged_attr,
300 	.name = "khugepaged",
301 };
302 #endif /* CONFIG_SYSFS */
303 
304 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
305 
306 int hugepage_madvise(struct vm_area_struct *vma,
307 		     unsigned long *vm_flags, int advice)
308 {
309 	switch (advice) {
310 	case MADV_HUGEPAGE:
311 #ifdef CONFIG_S390
312 		/*
313 		 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
314 		 * can't handle this properly after s390_enable_sie, so we simply
315 		 * ignore the madvise to prevent qemu from causing a SIGSEGV.
316 		 */
317 		if (mm_has_pgste(vma->vm_mm))
318 			return 0;
319 #endif
320 		*vm_flags &= ~VM_NOHUGEPAGE;
321 		*vm_flags |= VM_HUGEPAGE;
322 		/*
323 		 * If the vma become good for khugepaged to scan,
324 		 * register it here without waiting a page fault that
325 		 * may not happen any time soon.
326 		 */
327 		if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
328 				khugepaged_enter_vma_merge(vma, *vm_flags))
329 			return -ENOMEM;
330 		break;
331 	case MADV_NOHUGEPAGE:
332 		*vm_flags &= ~VM_HUGEPAGE;
333 		*vm_flags |= VM_NOHUGEPAGE;
334 		/*
335 		 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
336 		 * this vma even if we leave the mm registered in khugepaged if
337 		 * it got registered before VM_NOHUGEPAGE was set.
338 		 */
339 		break;
340 	}
341 
342 	return 0;
343 }
344 
345 int __init khugepaged_init(void)
346 {
347 	mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
348 					  sizeof(struct mm_slot),
349 					  __alignof__(struct mm_slot), 0, NULL);
350 	if (!mm_slot_cache)
351 		return -ENOMEM;
352 
353 	khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
354 	khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
355 	khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
356 
357 	return 0;
358 }
359 
360 void __init khugepaged_destroy(void)
361 {
362 	kmem_cache_destroy(mm_slot_cache);
363 }
364 
365 static inline struct mm_slot *alloc_mm_slot(void)
366 {
367 	if (!mm_slot_cache)	/* initialization failed */
368 		return NULL;
369 	return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
370 }
371 
372 static inline void free_mm_slot(struct mm_slot *mm_slot)
373 {
374 	kmem_cache_free(mm_slot_cache, mm_slot);
375 }
376 
377 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
378 {
379 	struct mm_slot *mm_slot;
380 
381 	hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
382 		if (mm == mm_slot->mm)
383 			return mm_slot;
384 
385 	return NULL;
386 }
387 
388 static void insert_to_mm_slots_hash(struct mm_struct *mm,
389 				    struct mm_slot *mm_slot)
390 {
391 	mm_slot->mm = mm;
392 	hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
393 }
394 
395 static inline int khugepaged_test_exit(struct mm_struct *mm)
396 {
397 	return atomic_read(&mm->mm_users) == 0;
398 }
399 
400 static bool hugepage_vma_check(struct vm_area_struct *vma,
401 			       unsigned long vm_flags)
402 {
403 	if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
404 	    (vm_flags & VM_NOHUGEPAGE) ||
405 	    test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
406 		return false;
407 	if (shmem_file(vma->vm_file)) {
408 		if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
409 			return false;
410 		return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
411 				HPAGE_PMD_NR);
412 	}
413 	if (!vma->anon_vma || vma->vm_ops)
414 		return false;
415 	if (is_vma_temporary_stack(vma))
416 		return false;
417 	return !(vm_flags & VM_NO_KHUGEPAGED);
418 }
419 
420 int __khugepaged_enter(struct mm_struct *mm)
421 {
422 	struct mm_slot *mm_slot;
423 	int wakeup;
424 
425 	mm_slot = alloc_mm_slot();
426 	if (!mm_slot)
427 		return -ENOMEM;
428 
429 	/* __khugepaged_exit() must not run from under us */
430 	VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
431 	if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
432 		free_mm_slot(mm_slot);
433 		return 0;
434 	}
435 
436 	spin_lock(&khugepaged_mm_lock);
437 	insert_to_mm_slots_hash(mm, mm_slot);
438 	/*
439 	 * Insert just behind the scanning cursor, to let the area settle
440 	 * down a little.
441 	 */
442 	wakeup = list_empty(&khugepaged_scan.mm_head);
443 	list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
444 	spin_unlock(&khugepaged_mm_lock);
445 
446 	mmgrab(mm);
447 	if (wakeup)
448 		wake_up_interruptible(&khugepaged_wait);
449 
450 	return 0;
451 }
452 
453 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
454 			       unsigned long vm_flags)
455 {
456 	unsigned long hstart, hend;
457 
458 	/*
459 	 * khugepaged does not yet work on non-shmem files or special
460 	 * mappings. And file-private shmem THP is not supported.
461 	 */
462 	if (!hugepage_vma_check(vma, vm_flags))
463 		return 0;
464 
465 	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
466 	hend = vma->vm_end & HPAGE_PMD_MASK;
467 	if (hstart < hend)
468 		return khugepaged_enter(vma, vm_flags);
469 	return 0;
470 }
471 
472 void __khugepaged_exit(struct mm_struct *mm)
473 {
474 	struct mm_slot *mm_slot;
475 	int free = 0;
476 
477 	spin_lock(&khugepaged_mm_lock);
478 	mm_slot = get_mm_slot(mm);
479 	if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
480 		hash_del(&mm_slot->hash);
481 		list_del(&mm_slot->mm_node);
482 		free = 1;
483 	}
484 	spin_unlock(&khugepaged_mm_lock);
485 
486 	if (free) {
487 		clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
488 		free_mm_slot(mm_slot);
489 		mmdrop(mm);
490 	} else if (mm_slot) {
491 		/*
492 		 * This is required to serialize against
493 		 * khugepaged_test_exit() (which is guaranteed to run
494 		 * under mmap sem read mode). Stop here (after we
495 		 * return all pagetables will be destroyed) until
496 		 * khugepaged has finished working on the pagetables
497 		 * under the mmap_sem.
498 		 */
499 		down_write(&mm->mmap_sem);
500 		up_write(&mm->mmap_sem);
501 	}
502 }
503 
504 static void release_pte_page(struct page *page)
505 {
506 	dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
507 	unlock_page(page);
508 	putback_lru_page(page);
509 }
510 
511 static void release_pte_pages(pte_t *pte, pte_t *_pte)
512 {
513 	while (--_pte >= pte) {
514 		pte_t pteval = *_pte;
515 		if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
516 			release_pte_page(pte_page(pteval));
517 	}
518 }
519 
520 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
521 					unsigned long address,
522 					pte_t *pte)
523 {
524 	struct page *page = NULL;
525 	pte_t *_pte;
526 	int none_or_zero = 0, result = 0, referenced = 0;
527 	bool writable = false;
528 
529 	for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
530 	     _pte++, address += PAGE_SIZE) {
531 		pte_t pteval = *_pte;
532 		if (pte_none(pteval) || (pte_present(pteval) &&
533 				is_zero_pfn(pte_pfn(pteval)))) {
534 			if (!userfaultfd_armed(vma) &&
535 			    ++none_or_zero <= khugepaged_max_ptes_none) {
536 				continue;
537 			} else {
538 				result = SCAN_EXCEED_NONE_PTE;
539 				goto out;
540 			}
541 		}
542 		if (!pte_present(pteval)) {
543 			result = SCAN_PTE_NON_PRESENT;
544 			goto out;
545 		}
546 		page = vm_normal_page(vma, address, pteval);
547 		if (unlikely(!page)) {
548 			result = SCAN_PAGE_NULL;
549 			goto out;
550 		}
551 
552 		/* TODO: teach khugepaged to collapse THP mapped with pte */
553 		if (PageCompound(page)) {
554 			result = SCAN_PAGE_COMPOUND;
555 			goto out;
556 		}
557 
558 		VM_BUG_ON_PAGE(!PageAnon(page), page);
559 
560 		/*
561 		 * We can do it before isolate_lru_page because the
562 		 * page can't be freed from under us. NOTE: PG_lock
563 		 * is needed to serialize against split_huge_page
564 		 * when invoked from the VM.
565 		 */
566 		if (!trylock_page(page)) {
567 			result = SCAN_PAGE_LOCK;
568 			goto out;
569 		}
570 
571 		/*
572 		 * cannot use mapcount: can't collapse if there's a gup pin.
573 		 * The page must only be referenced by the scanned process
574 		 * and page swap cache.
575 		 */
576 		if (page_count(page) != 1 + PageSwapCache(page)) {
577 			unlock_page(page);
578 			result = SCAN_PAGE_COUNT;
579 			goto out;
580 		}
581 		if (pte_write(pteval)) {
582 			writable = true;
583 		} else {
584 			if (PageSwapCache(page) &&
585 			    !reuse_swap_page(page, NULL)) {
586 				unlock_page(page);
587 				result = SCAN_SWAP_CACHE_PAGE;
588 				goto out;
589 			}
590 			/*
591 			 * Page is not in the swap cache. It can be collapsed
592 			 * into a THP.
593 			 */
594 		}
595 
596 		/*
597 		 * Isolate the page to avoid collapsing an hugepage
598 		 * currently in use by the VM.
599 		 */
600 		if (isolate_lru_page(page)) {
601 			unlock_page(page);
602 			result = SCAN_DEL_PAGE_LRU;
603 			goto out;
604 		}
605 		inc_node_page_state(page,
606 				NR_ISOLATED_ANON + page_is_file_cache(page));
607 		VM_BUG_ON_PAGE(!PageLocked(page), page);
608 		VM_BUG_ON_PAGE(PageLRU(page), page);
609 
610 		/* There should be enough young pte to collapse the page */
611 		if (pte_young(pteval) ||
612 		    page_is_young(page) || PageReferenced(page) ||
613 		    mmu_notifier_test_young(vma->vm_mm, address))
614 			referenced++;
615 	}
616 	if (likely(writable)) {
617 		if (likely(referenced)) {
618 			result = SCAN_SUCCEED;
619 			trace_mm_collapse_huge_page_isolate(page, none_or_zero,
620 							    referenced, writable, result);
621 			return 1;
622 		}
623 	} else {
624 		result = SCAN_PAGE_RO;
625 	}
626 
627 out:
628 	release_pte_pages(pte, _pte);
629 	trace_mm_collapse_huge_page_isolate(page, none_or_zero,
630 					    referenced, writable, result);
631 	return 0;
632 }
633 
634 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
635 				      struct vm_area_struct *vma,
636 				      unsigned long address,
637 				      spinlock_t *ptl)
638 {
639 	pte_t *_pte;
640 	for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
641 				_pte++, page++, address += PAGE_SIZE) {
642 		pte_t pteval = *_pte;
643 		struct page *src_page;
644 
645 		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
646 			clear_user_highpage(page, address);
647 			add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
648 			if (is_zero_pfn(pte_pfn(pteval))) {
649 				/*
650 				 * ptl mostly unnecessary.
651 				 */
652 				spin_lock(ptl);
653 				/*
654 				 * paravirt calls inside pte_clear here are
655 				 * superfluous.
656 				 */
657 				pte_clear(vma->vm_mm, address, _pte);
658 				spin_unlock(ptl);
659 			}
660 		} else {
661 			src_page = pte_page(pteval);
662 			copy_user_highpage(page, src_page, address, vma);
663 			VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
664 			release_pte_page(src_page);
665 			/*
666 			 * ptl mostly unnecessary, but preempt has to
667 			 * be disabled to update the per-cpu stats
668 			 * inside page_remove_rmap().
669 			 */
670 			spin_lock(ptl);
671 			/*
672 			 * paravirt calls inside pte_clear here are
673 			 * superfluous.
674 			 */
675 			pte_clear(vma->vm_mm, address, _pte);
676 			page_remove_rmap(src_page, false);
677 			spin_unlock(ptl);
678 			free_page_and_swap_cache(src_page);
679 		}
680 	}
681 }
682 
683 static void khugepaged_alloc_sleep(void)
684 {
685 	DEFINE_WAIT(wait);
686 
687 	add_wait_queue(&khugepaged_wait, &wait);
688 	freezable_schedule_timeout_interruptible(
689 		msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
690 	remove_wait_queue(&khugepaged_wait, &wait);
691 }
692 
693 static int khugepaged_node_load[MAX_NUMNODES];
694 
695 static bool khugepaged_scan_abort(int nid)
696 {
697 	int i;
698 
699 	/*
700 	 * If node_reclaim_mode is disabled, then no extra effort is made to
701 	 * allocate memory locally.
702 	 */
703 	if (!node_reclaim_mode)
704 		return false;
705 
706 	/* If there is a count for this node already, it must be acceptable */
707 	if (khugepaged_node_load[nid])
708 		return false;
709 
710 	for (i = 0; i < MAX_NUMNODES; i++) {
711 		if (!khugepaged_node_load[i])
712 			continue;
713 		if (node_distance(nid, i) > RECLAIM_DISTANCE)
714 			return true;
715 	}
716 	return false;
717 }
718 
719 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
720 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
721 {
722 	return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
723 }
724 
725 #ifdef CONFIG_NUMA
726 static int khugepaged_find_target_node(void)
727 {
728 	static int last_khugepaged_target_node = NUMA_NO_NODE;
729 	int nid, target_node = 0, max_value = 0;
730 
731 	/* find first node with max normal pages hit */
732 	for (nid = 0; nid < MAX_NUMNODES; nid++)
733 		if (khugepaged_node_load[nid] > max_value) {
734 			max_value = khugepaged_node_load[nid];
735 			target_node = nid;
736 		}
737 
738 	/* do some balance if several nodes have the same hit record */
739 	if (target_node <= last_khugepaged_target_node)
740 		for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
741 				nid++)
742 			if (max_value == khugepaged_node_load[nid]) {
743 				target_node = nid;
744 				break;
745 			}
746 
747 	last_khugepaged_target_node = target_node;
748 	return target_node;
749 }
750 
751 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
752 {
753 	if (IS_ERR(*hpage)) {
754 		if (!*wait)
755 			return false;
756 
757 		*wait = false;
758 		*hpage = NULL;
759 		khugepaged_alloc_sleep();
760 	} else if (*hpage) {
761 		put_page(*hpage);
762 		*hpage = NULL;
763 	}
764 
765 	return true;
766 }
767 
768 static struct page *
769 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
770 {
771 	VM_BUG_ON_PAGE(*hpage, *hpage);
772 
773 	*hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
774 	if (unlikely(!*hpage)) {
775 		count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
776 		*hpage = ERR_PTR(-ENOMEM);
777 		return NULL;
778 	}
779 
780 	prep_transhuge_page(*hpage);
781 	count_vm_event(THP_COLLAPSE_ALLOC);
782 	return *hpage;
783 }
784 #else
785 static int khugepaged_find_target_node(void)
786 {
787 	return 0;
788 }
789 
790 static inline struct page *alloc_khugepaged_hugepage(void)
791 {
792 	struct page *page;
793 
794 	page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
795 			   HPAGE_PMD_ORDER);
796 	if (page)
797 		prep_transhuge_page(page);
798 	return page;
799 }
800 
801 static struct page *khugepaged_alloc_hugepage(bool *wait)
802 {
803 	struct page *hpage;
804 
805 	do {
806 		hpage = alloc_khugepaged_hugepage();
807 		if (!hpage) {
808 			count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
809 			if (!*wait)
810 				return NULL;
811 
812 			*wait = false;
813 			khugepaged_alloc_sleep();
814 		} else
815 			count_vm_event(THP_COLLAPSE_ALLOC);
816 	} while (unlikely(!hpage) && likely(khugepaged_enabled()));
817 
818 	return hpage;
819 }
820 
821 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
822 {
823 	if (!*hpage)
824 		*hpage = khugepaged_alloc_hugepage(wait);
825 
826 	if (unlikely(!*hpage))
827 		return false;
828 
829 	return true;
830 }
831 
832 static struct page *
833 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
834 {
835 	VM_BUG_ON(!*hpage);
836 
837 	return  *hpage;
838 }
839 #endif
840 
841 /*
842  * If mmap_sem temporarily dropped, revalidate vma
843  * before taking mmap_sem.
844  * Return 0 if succeeds, otherwise return none-zero
845  * value (scan code).
846  */
847 
848 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
849 		struct vm_area_struct **vmap)
850 {
851 	struct vm_area_struct *vma;
852 	unsigned long hstart, hend;
853 
854 	if (unlikely(khugepaged_test_exit(mm)))
855 		return SCAN_ANY_PROCESS;
856 
857 	*vmap = vma = find_vma(mm, address);
858 	if (!vma)
859 		return SCAN_VMA_NULL;
860 
861 	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
862 	hend = vma->vm_end & HPAGE_PMD_MASK;
863 	if (address < hstart || address + HPAGE_PMD_SIZE > hend)
864 		return SCAN_ADDRESS_RANGE;
865 	if (!hugepage_vma_check(vma, vma->vm_flags))
866 		return SCAN_VMA_CHECK;
867 	return 0;
868 }
869 
870 /*
871  * Bring missing pages in from swap, to complete THP collapse.
872  * Only done if khugepaged_scan_pmd believes it is worthwhile.
873  *
874  * Called and returns without pte mapped or spinlocks held,
875  * but with mmap_sem held to protect against vma changes.
876  */
877 
878 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
879 					struct vm_area_struct *vma,
880 					unsigned long address, pmd_t *pmd,
881 					int referenced)
882 {
883 	int swapped_in = 0;
884 	vm_fault_t ret = 0;
885 	struct vm_fault vmf = {
886 		.vma = vma,
887 		.address = address,
888 		.flags = FAULT_FLAG_ALLOW_RETRY,
889 		.pmd = pmd,
890 		.pgoff = linear_page_index(vma, address),
891 	};
892 
893 	/* we only decide to swapin, if there is enough young ptes */
894 	if (referenced < HPAGE_PMD_NR/2) {
895 		trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
896 		return false;
897 	}
898 	vmf.pte = pte_offset_map(pmd, address);
899 	for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
900 			vmf.pte++, vmf.address += PAGE_SIZE) {
901 		vmf.orig_pte = *vmf.pte;
902 		if (!is_swap_pte(vmf.orig_pte))
903 			continue;
904 		swapped_in++;
905 		ret = do_swap_page(&vmf);
906 
907 		/* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
908 		if (ret & VM_FAULT_RETRY) {
909 			down_read(&mm->mmap_sem);
910 			if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
911 				/* vma is no longer available, don't continue to swapin */
912 				trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
913 				return false;
914 			}
915 			/* check if the pmd is still valid */
916 			if (mm_find_pmd(mm, address) != pmd) {
917 				trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
918 				return false;
919 			}
920 		}
921 		if (ret & VM_FAULT_ERROR) {
922 			trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
923 			return false;
924 		}
925 		/* pte is unmapped now, we need to map it */
926 		vmf.pte = pte_offset_map(pmd, vmf.address);
927 	}
928 	vmf.pte--;
929 	pte_unmap(vmf.pte);
930 	trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
931 	return true;
932 }
933 
934 static void collapse_huge_page(struct mm_struct *mm,
935 				   unsigned long address,
936 				   struct page **hpage,
937 				   int node, int referenced)
938 {
939 	pmd_t *pmd, _pmd;
940 	pte_t *pte;
941 	pgtable_t pgtable;
942 	struct page *new_page;
943 	spinlock_t *pmd_ptl, *pte_ptl;
944 	int isolated = 0, result = 0;
945 	struct mem_cgroup *memcg;
946 	struct vm_area_struct *vma;
947 	unsigned long mmun_start;	/* For mmu_notifiers */
948 	unsigned long mmun_end;		/* For mmu_notifiers */
949 	gfp_t gfp;
950 
951 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
952 
953 	/* Only allocate from the target node */
954 	gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
955 
956 	/*
957 	 * Before allocating the hugepage, release the mmap_sem read lock.
958 	 * The allocation can take potentially a long time if it involves
959 	 * sync compaction, and we do not need to hold the mmap_sem during
960 	 * that. We will recheck the vma after taking it again in write mode.
961 	 */
962 	up_read(&mm->mmap_sem);
963 	new_page = khugepaged_alloc_page(hpage, gfp, node);
964 	if (!new_page) {
965 		result = SCAN_ALLOC_HUGE_PAGE_FAIL;
966 		goto out_nolock;
967 	}
968 
969 	if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
970 		result = SCAN_CGROUP_CHARGE_FAIL;
971 		goto out_nolock;
972 	}
973 
974 	down_read(&mm->mmap_sem);
975 	result = hugepage_vma_revalidate(mm, address, &vma);
976 	if (result) {
977 		mem_cgroup_cancel_charge(new_page, memcg, true);
978 		up_read(&mm->mmap_sem);
979 		goto out_nolock;
980 	}
981 
982 	pmd = mm_find_pmd(mm, address);
983 	if (!pmd) {
984 		result = SCAN_PMD_NULL;
985 		mem_cgroup_cancel_charge(new_page, memcg, true);
986 		up_read(&mm->mmap_sem);
987 		goto out_nolock;
988 	}
989 
990 	/*
991 	 * __collapse_huge_page_swapin always returns with mmap_sem locked.
992 	 * If it fails, we release mmap_sem and jump out_nolock.
993 	 * Continuing to collapse causes inconsistency.
994 	 */
995 	if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
996 		mem_cgroup_cancel_charge(new_page, memcg, true);
997 		up_read(&mm->mmap_sem);
998 		goto out_nolock;
999 	}
1000 
1001 	up_read(&mm->mmap_sem);
1002 	/*
1003 	 * Prevent all access to pagetables with the exception of
1004 	 * gup_fast later handled by the ptep_clear_flush and the VM
1005 	 * handled by the anon_vma lock + PG_lock.
1006 	 */
1007 	down_write(&mm->mmap_sem);
1008 	result = hugepage_vma_revalidate(mm, address, &vma);
1009 	if (result)
1010 		goto out;
1011 	/* check if the pmd is still valid */
1012 	if (mm_find_pmd(mm, address) != pmd)
1013 		goto out;
1014 
1015 	anon_vma_lock_write(vma->anon_vma);
1016 
1017 	pte = pte_offset_map(pmd, address);
1018 	pte_ptl = pte_lockptr(mm, pmd);
1019 
1020 	mmun_start = address;
1021 	mmun_end   = address + HPAGE_PMD_SIZE;
1022 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1023 	pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1024 	/*
1025 	 * After this gup_fast can't run anymore. This also removes
1026 	 * any huge TLB entry from the CPU so we won't allow
1027 	 * huge and small TLB entries for the same virtual address
1028 	 * to avoid the risk of CPU bugs in that area.
1029 	 */
1030 	_pmd = pmdp_collapse_flush(vma, address, pmd);
1031 	spin_unlock(pmd_ptl);
1032 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1033 
1034 	spin_lock(pte_ptl);
1035 	isolated = __collapse_huge_page_isolate(vma, address, pte);
1036 	spin_unlock(pte_ptl);
1037 
1038 	if (unlikely(!isolated)) {
1039 		pte_unmap(pte);
1040 		spin_lock(pmd_ptl);
1041 		BUG_ON(!pmd_none(*pmd));
1042 		/*
1043 		 * We can only use set_pmd_at when establishing
1044 		 * hugepmds and never for establishing regular pmds that
1045 		 * points to regular pagetables. Use pmd_populate for that
1046 		 */
1047 		pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1048 		spin_unlock(pmd_ptl);
1049 		anon_vma_unlock_write(vma->anon_vma);
1050 		result = SCAN_FAIL;
1051 		goto out;
1052 	}
1053 
1054 	/*
1055 	 * All pages are isolated and locked so anon_vma rmap
1056 	 * can't run anymore.
1057 	 */
1058 	anon_vma_unlock_write(vma->anon_vma);
1059 
1060 	__collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1061 	pte_unmap(pte);
1062 	__SetPageUptodate(new_page);
1063 	pgtable = pmd_pgtable(_pmd);
1064 
1065 	_pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1066 	_pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1067 
1068 	/*
1069 	 * spin_lock() below is not the equivalent of smp_wmb(), so
1070 	 * this is needed to avoid the copy_huge_page writes to become
1071 	 * visible after the set_pmd_at() write.
1072 	 */
1073 	smp_wmb();
1074 
1075 	spin_lock(pmd_ptl);
1076 	BUG_ON(!pmd_none(*pmd));
1077 	page_add_new_anon_rmap(new_page, vma, address, true);
1078 	mem_cgroup_commit_charge(new_page, memcg, false, true);
1079 	lru_cache_add_active_or_unevictable(new_page, vma);
1080 	pgtable_trans_huge_deposit(mm, pmd, pgtable);
1081 	set_pmd_at(mm, address, pmd, _pmd);
1082 	update_mmu_cache_pmd(vma, address, pmd);
1083 	spin_unlock(pmd_ptl);
1084 
1085 	*hpage = NULL;
1086 
1087 	khugepaged_pages_collapsed++;
1088 	result = SCAN_SUCCEED;
1089 out_up_write:
1090 	up_write(&mm->mmap_sem);
1091 out_nolock:
1092 	trace_mm_collapse_huge_page(mm, isolated, result);
1093 	return;
1094 out:
1095 	mem_cgroup_cancel_charge(new_page, memcg, true);
1096 	goto out_up_write;
1097 }
1098 
1099 static int khugepaged_scan_pmd(struct mm_struct *mm,
1100 			       struct vm_area_struct *vma,
1101 			       unsigned long address,
1102 			       struct page **hpage)
1103 {
1104 	pmd_t *pmd;
1105 	pte_t *pte, *_pte;
1106 	int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1107 	struct page *page = NULL;
1108 	unsigned long _address;
1109 	spinlock_t *ptl;
1110 	int node = NUMA_NO_NODE, unmapped = 0;
1111 	bool writable = false;
1112 
1113 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1114 
1115 	pmd = mm_find_pmd(mm, address);
1116 	if (!pmd) {
1117 		result = SCAN_PMD_NULL;
1118 		goto out;
1119 	}
1120 
1121 	memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1122 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1123 	for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1124 	     _pte++, _address += PAGE_SIZE) {
1125 		pte_t pteval = *_pte;
1126 		if (is_swap_pte(pteval)) {
1127 			if (++unmapped <= khugepaged_max_ptes_swap) {
1128 				continue;
1129 			} else {
1130 				result = SCAN_EXCEED_SWAP_PTE;
1131 				goto out_unmap;
1132 			}
1133 		}
1134 		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1135 			if (!userfaultfd_armed(vma) &&
1136 			    ++none_or_zero <= khugepaged_max_ptes_none) {
1137 				continue;
1138 			} else {
1139 				result = SCAN_EXCEED_NONE_PTE;
1140 				goto out_unmap;
1141 			}
1142 		}
1143 		if (!pte_present(pteval)) {
1144 			result = SCAN_PTE_NON_PRESENT;
1145 			goto out_unmap;
1146 		}
1147 		if (pte_write(pteval))
1148 			writable = true;
1149 
1150 		page = vm_normal_page(vma, _address, pteval);
1151 		if (unlikely(!page)) {
1152 			result = SCAN_PAGE_NULL;
1153 			goto out_unmap;
1154 		}
1155 
1156 		/* TODO: teach khugepaged to collapse THP mapped with pte */
1157 		if (PageCompound(page)) {
1158 			result = SCAN_PAGE_COMPOUND;
1159 			goto out_unmap;
1160 		}
1161 
1162 		/*
1163 		 * Record which node the original page is from and save this
1164 		 * information to khugepaged_node_load[].
1165 		 * Khupaged will allocate hugepage from the node has the max
1166 		 * hit record.
1167 		 */
1168 		node = page_to_nid(page);
1169 		if (khugepaged_scan_abort(node)) {
1170 			result = SCAN_SCAN_ABORT;
1171 			goto out_unmap;
1172 		}
1173 		khugepaged_node_load[node]++;
1174 		if (!PageLRU(page)) {
1175 			result = SCAN_PAGE_LRU;
1176 			goto out_unmap;
1177 		}
1178 		if (PageLocked(page)) {
1179 			result = SCAN_PAGE_LOCK;
1180 			goto out_unmap;
1181 		}
1182 		if (!PageAnon(page)) {
1183 			result = SCAN_PAGE_ANON;
1184 			goto out_unmap;
1185 		}
1186 
1187 		/*
1188 		 * cannot use mapcount: can't collapse if there's a gup pin.
1189 		 * The page must only be referenced by the scanned process
1190 		 * and page swap cache.
1191 		 */
1192 		if (page_count(page) != 1 + PageSwapCache(page)) {
1193 			result = SCAN_PAGE_COUNT;
1194 			goto out_unmap;
1195 		}
1196 		if (pte_young(pteval) ||
1197 		    page_is_young(page) || PageReferenced(page) ||
1198 		    mmu_notifier_test_young(vma->vm_mm, address))
1199 			referenced++;
1200 	}
1201 	if (writable) {
1202 		if (referenced) {
1203 			result = SCAN_SUCCEED;
1204 			ret = 1;
1205 		} else {
1206 			result = SCAN_LACK_REFERENCED_PAGE;
1207 		}
1208 	} else {
1209 		result = SCAN_PAGE_RO;
1210 	}
1211 out_unmap:
1212 	pte_unmap_unlock(pte, ptl);
1213 	if (ret) {
1214 		node = khugepaged_find_target_node();
1215 		/* collapse_huge_page will return with the mmap_sem released */
1216 		collapse_huge_page(mm, address, hpage, node, referenced);
1217 	}
1218 out:
1219 	trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1220 				     none_or_zero, result, unmapped);
1221 	return ret;
1222 }
1223 
1224 static void collect_mm_slot(struct mm_slot *mm_slot)
1225 {
1226 	struct mm_struct *mm = mm_slot->mm;
1227 
1228 	VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1229 
1230 	if (khugepaged_test_exit(mm)) {
1231 		/* free mm_slot */
1232 		hash_del(&mm_slot->hash);
1233 		list_del(&mm_slot->mm_node);
1234 
1235 		/*
1236 		 * Not strictly needed because the mm exited already.
1237 		 *
1238 		 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1239 		 */
1240 
1241 		/* khugepaged_mm_lock actually not necessary for the below */
1242 		free_mm_slot(mm_slot);
1243 		mmdrop(mm);
1244 	}
1245 }
1246 
1247 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1248 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1249 {
1250 	struct vm_area_struct *vma;
1251 	unsigned long addr;
1252 	pmd_t *pmd, _pmd;
1253 
1254 	i_mmap_lock_write(mapping);
1255 	vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1256 		/* probably overkill */
1257 		if (vma->anon_vma)
1258 			continue;
1259 		addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1260 		if (addr & ~HPAGE_PMD_MASK)
1261 			continue;
1262 		if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1263 			continue;
1264 		pmd = mm_find_pmd(vma->vm_mm, addr);
1265 		if (!pmd)
1266 			continue;
1267 		/*
1268 		 * We need exclusive mmap_sem to retract page table.
1269 		 * If trylock fails we would end up with pte-mapped THP after
1270 		 * re-fault. Not ideal, but it's more important to not disturb
1271 		 * the system too much.
1272 		 */
1273 		if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1274 			spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1275 			/* assume page table is clear */
1276 			_pmd = pmdp_collapse_flush(vma, addr, pmd);
1277 			spin_unlock(ptl);
1278 			up_write(&vma->vm_mm->mmap_sem);
1279 			mm_dec_nr_ptes(vma->vm_mm);
1280 			pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1281 		}
1282 	}
1283 	i_mmap_unlock_write(mapping);
1284 }
1285 
1286 /**
1287  * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
1288  *
1289  * Basic scheme is simple, details are more complex:
1290  *  - allocate and freeze a new huge page;
1291  *  - scan page cache replacing old pages with the new one
1292  *    + swap in pages if necessary;
1293  *    + fill in gaps;
1294  *    + keep old pages around in case rollback is required;
1295  *  - if replacing succeeds:
1296  *    + copy data over;
1297  *    + free old pages;
1298  *    + unfreeze huge page;
1299  *  - if replacing failed;
1300  *    + put all pages back and unfreeze them;
1301  *    + restore gaps in the page cache;
1302  *    + free huge page;
1303  */
1304 static void collapse_shmem(struct mm_struct *mm,
1305 		struct address_space *mapping, pgoff_t start,
1306 		struct page **hpage, int node)
1307 {
1308 	gfp_t gfp;
1309 	struct page *new_page;
1310 	struct mem_cgroup *memcg;
1311 	pgoff_t index, end = start + HPAGE_PMD_NR;
1312 	LIST_HEAD(pagelist);
1313 	XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1314 	int nr_none = 0, result = SCAN_SUCCEED;
1315 
1316 	VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1317 
1318 	/* Only allocate from the target node */
1319 	gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1320 
1321 	new_page = khugepaged_alloc_page(hpage, gfp, node);
1322 	if (!new_page) {
1323 		result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1324 		goto out;
1325 	}
1326 
1327 	if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1328 		result = SCAN_CGROUP_CHARGE_FAIL;
1329 		goto out;
1330 	}
1331 
1332 	new_page->index = start;
1333 	new_page->mapping = mapping;
1334 	__SetPageSwapBacked(new_page);
1335 	__SetPageLocked(new_page);
1336 	BUG_ON(!page_ref_freeze(new_page, 1));
1337 
1338 	/*
1339 	 * At this point the new_page is 'frozen' (page_count() is zero),
1340 	 * locked and not up-to-date. It's safe to insert it into the page
1341 	 * cache, because nobody would be able to map it or use it in other
1342 	 * way until we unfreeze it.
1343 	 */
1344 
1345 	/* This will be less messy when we use multi-index entries */
1346 	do {
1347 		xas_lock_irq(&xas);
1348 		xas_create_range(&xas);
1349 		if (!xas_error(&xas))
1350 			break;
1351 		xas_unlock_irq(&xas);
1352 		if (!xas_nomem(&xas, GFP_KERNEL))
1353 			goto out;
1354 	} while (1);
1355 
1356 	xas_set(&xas, start);
1357 	for (index = start; index < end; index++) {
1358 		struct page *page = xas_next(&xas);
1359 
1360 		VM_BUG_ON(index != xas.xa_index);
1361 		if (!page) {
1362 			if (!shmem_charge(mapping->host, 1)) {
1363 				result = SCAN_FAIL;
1364 				break;
1365 			}
1366 			xas_store(&xas, new_page + (index % HPAGE_PMD_NR));
1367 			nr_none++;
1368 			continue;
1369 		}
1370 
1371 		if (xa_is_value(page) || !PageUptodate(page)) {
1372 			xas_unlock_irq(&xas);
1373 			/* swap in or instantiate fallocated page */
1374 			if (shmem_getpage(mapping->host, index, &page,
1375 						SGP_NOHUGE)) {
1376 				result = SCAN_FAIL;
1377 				goto xa_unlocked;
1378 			}
1379 			xas_lock_irq(&xas);
1380 			xas_set(&xas, index);
1381 		} else if (trylock_page(page)) {
1382 			get_page(page);
1383 		} else {
1384 			result = SCAN_PAGE_LOCK;
1385 			break;
1386 		}
1387 
1388 		/*
1389 		 * The page must be locked, so we can drop the i_pages lock
1390 		 * without racing with truncate.
1391 		 */
1392 		VM_BUG_ON_PAGE(!PageLocked(page), page);
1393 		VM_BUG_ON_PAGE(!PageUptodate(page), page);
1394 		VM_BUG_ON_PAGE(PageTransCompound(page), page);
1395 
1396 		if (page_mapping(page) != mapping) {
1397 			result = SCAN_TRUNCATED;
1398 			goto out_unlock;
1399 		}
1400 		xas_unlock_irq(&xas);
1401 
1402 		if (isolate_lru_page(page)) {
1403 			result = SCAN_DEL_PAGE_LRU;
1404 			goto out_isolate_failed;
1405 		}
1406 
1407 		if (page_mapped(page))
1408 			unmap_mapping_pages(mapping, index, 1, false);
1409 
1410 		xas_lock_irq(&xas);
1411 		xas_set(&xas, index);
1412 
1413 		VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1414 		VM_BUG_ON_PAGE(page_mapped(page), page);
1415 
1416 		/*
1417 		 * The page is expected to have page_count() == 3:
1418 		 *  - we hold a pin on it;
1419 		 *  - one reference from page cache;
1420 		 *  - one from isolate_lru_page;
1421 		 */
1422 		if (!page_ref_freeze(page, 3)) {
1423 			result = SCAN_PAGE_COUNT;
1424 			goto out_lru;
1425 		}
1426 
1427 		/*
1428 		 * Add the page to the list to be able to undo the collapse if
1429 		 * something go wrong.
1430 		 */
1431 		list_add_tail(&page->lru, &pagelist);
1432 
1433 		/* Finally, replace with the new page. */
1434 		xas_store(&xas, new_page + (index % HPAGE_PMD_NR));
1435 		continue;
1436 out_lru:
1437 		xas_unlock_irq(&xas);
1438 		putback_lru_page(page);
1439 out_isolate_failed:
1440 		unlock_page(page);
1441 		put_page(page);
1442 		goto xa_unlocked;
1443 out_unlock:
1444 		unlock_page(page);
1445 		put_page(page);
1446 		break;
1447 	}
1448 	xas_unlock_irq(&xas);
1449 
1450 xa_unlocked:
1451 	if (result == SCAN_SUCCEED) {
1452 		struct page *page, *tmp;
1453 		struct zone *zone = page_zone(new_page);
1454 
1455 		/*
1456 		 * Replacing old pages with new one has succeeded, now we
1457 		 * need to copy the content and free the old pages.
1458 		 */
1459 		list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1460 			copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1461 					page);
1462 			list_del(&page->lru);
1463 			unlock_page(page);
1464 			page_ref_unfreeze(page, 1);
1465 			page->mapping = NULL;
1466 			ClearPageActive(page);
1467 			ClearPageUnevictable(page);
1468 			put_page(page);
1469 		}
1470 
1471 		local_irq_disable();
1472 		__inc_node_page_state(new_page, NR_SHMEM_THPS);
1473 		if (nr_none) {
1474 			__mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1475 			__mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none);
1476 		}
1477 		local_irq_enable();
1478 
1479 		/*
1480 		 * Remove pte page tables, so we can re-fault
1481 		 * the page as huge.
1482 		 */
1483 		retract_page_tables(mapping, start);
1484 
1485 		/* Everything is ready, let's unfreeze the new_page */
1486 		set_page_dirty(new_page);
1487 		SetPageUptodate(new_page);
1488 		page_ref_unfreeze(new_page, HPAGE_PMD_NR);
1489 		mem_cgroup_commit_charge(new_page, memcg, false, true);
1490 		lru_cache_add_anon(new_page);
1491 		unlock_page(new_page);
1492 
1493 		*hpage = NULL;
1494 
1495 		khugepaged_pages_collapsed++;
1496 	} else {
1497 		struct page *page;
1498 		/* Something went wrong: roll back page cache changes */
1499 		shmem_uncharge(mapping->host, nr_none);
1500 		xas_lock_irq(&xas);
1501 		xas_set(&xas, start);
1502 		xas_for_each(&xas, page, end - 1) {
1503 			page = list_first_entry_or_null(&pagelist,
1504 					struct page, lru);
1505 			if (!page || xas.xa_index < page->index) {
1506 				if (!nr_none)
1507 					break;
1508 				nr_none--;
1509 				/* Put holes back where they were */
1510 				xas_store(&xas, NULL);
1511 				continue;
1512 			}
1513 
1514 			VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1515 
1516 			/* Unfreeze the page. */
1517 			list_del(&page->lru);
1518 			page_ref_unfreeze(page, 2);
1519 			xas_store(&xas, page);
1520 			xas_pause(&xas);
1521 			xas_unlock_irq(&xas);
1522 			putback_lru_page(page);
1523 			unlock_page(page);
1524 			xas_lock_irq(&xas);
1525 		}
1526 		VM_BUG_ON(nr_none);
1527 		xas_unlock_irq(&xas);
1528 
1529 		/* Unfreeze new_page, caller would take care about freeing it */
1530 		page_ref_unfreeze(new_page, 1);
1531 		mem_cgroup_cancel_charge(new_page, memcg, true);
1532 		unlock_page(new_page);
1533 		new_page->mapping = NULL;
1534 	}
1535 out:
1536 	VM_BUG_ON(!list_empty(&pagelist));
1537 	/* TODO: tracepoints */
1538 }
1539 
1540 static void khugepaged_scan_shmem(struct mm_struct *mm,
1541 		struct address_space *mapping,
1542 		pgoff_t start, struct page **hpage)
1543 {
1544 	struct page *page = NULL;
1545 	XA_STATE(xas, &mapping->i_pages, start);
1546 	int present, swap;
1547 	int node = NUMA_NO_NODE;
1548 	int result = SCAN_SUCCEED;
1549 
1550 	present = 0;
1551 	swap = 0;
1552 	memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1553 	rcu_read_lock();
1554 	xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1555 		if (xas_retry(&xas, page))
1556 			continue;
1557 
1558 		if (xa_is_value(page)) {
1559 			if (++swap > khugepaged_max_ptes_swap) {
1560 				result = SCAN_EXCEED_SWAP_PTE;
1561 				break;
1562 			}
1563 			continue;
1564 		}
1565 
1566 		if (PageTransCompound(page)) {
1567 			result = SCAN_PAGE_COMPOUND;
1568 			break;
1569 		}
1570 
1571 		node = page_to_nid(page);
1572 		if (khugepaged_scan_abort(node)) {
1573 			result = SCAN_SCAN_ABORT;
1574 			break;
1575 		}
1576 		khugepaged_node_load[node]++;
1577 
1578 		if (!PageLRU(page)) {
1579 			result = SCAN_PAGE_LRU;
1580 			break;
1581 		}
1582 
1583 		if (page_count(page) != 1 + page_mapcount(page)) {
1584 			result = SCAN_PAGE_COUNT;
1585 			break;
1586 		}
1587 
1588 		/*
1589 		 * We probably should check if the page is referenced here, but
1590 		 * nobody would transfer pte_young() to PageReferenced() for us.
1591 		 * And rmap walk here is just too costly...
1592 		 */
1593 
1594 		present++;
1595 
1596 		if (need_resched()) {
1597 			xas_pause(&xas);
1598 			cond_resched_rcu();
1599 		}
1600 	}
1601 	rcu_read_unlock();
1602 
1603 	if (result == SCAN_SUCCEED) {
1604 		if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1605 			result = SCAN_EXCEED_NONE_PTE;
1606 		} else {
1607 			node = khugepaged_find_target_node();
1608 			collapse_shmem(mm, mapping, start, hpage, node);
1609 		}
1610 	}
1611 
1612 	/* TODO: tracepoints */
1613 }
1614 #else
1615 static void khugepaged_scan_shmem(struct mm_struct *mm,
1616 		struct address_space *mapping,
1617 		pgoff_t start, struct page **hpage)
1618 {
1619 	BUILD_BUG();
1620 }
1621 #endif
1622 
1623 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1624 					    struct page **hpage)
1625 	__releases(&khugepaged_mm_lock)
1626 	__acquires(&khugepaged_mm_lock)
1627 {
1628 	struct mm_slot *mm_slot;
1629 	struct mm_struct *mm;
1630 	struct vm_area_struct *vma;
1631 	int progress = 0;
1632 
1633 	VM_BUG_ON(!pages);
1634 	VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1635 
1636 	if (khugepaged_scan.mm_slot)
1637 		mm_slot = khugepaged_scan.mm_slot;
1638 	else {
1639 		mm_slot = list_entry(khugepaged_scan.mm_head.next,
1640 				     struct mm_slot, mm_node);
1641 		khugepaged_scan.address = 0;
1642 		khugepaged_scan.mm_slot = mm_slot;
1643 	}
1644 	spin_unlock(&khugepaged_mm_lock);
1645 
1646 	mm = mm_slot->mm;
1647 	/*
1648 	 * Don't wait for semaphore (to avoid long wait times).  Just move to
1649 	 * the next mm on the list.
1650 	 */
1651 	vma = NULL;
1652 	if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1653 		goto breakouterloop_mmap_sem;
1654 	if (likely(!khugepaged_test_exit(mm)))
1655 		vma = find_vma(mm, khugepaged_scan.address);
1656 
1657 	progress++;
1658 	for (; vma; vma = vma->vm_next) {
1659 		unsigned long hstart, hend;
1660 
1661 		cond_resched();
1662 		if (unlikely(khugepaged_test_exit(mm))) {
1663 			progress++;
1664 			break;
1665 		}
1666 		if (!hugepage_vma_check(vma, vma->vm_flags)) {
1667 skip:
1668 			progress++;
1669 			continue;
1670 		}
1671 		hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1672 		hend = vma->vm_end & HPAGE_PMD_MASK;
1673 		if (hstart >= hend)
1674 			goto skip;
1675 		if (khugepaged_scan.address > hend)
1676 			goto skip;
1677 		if (khugepaged_scan.address < hstart)
1678 			khugepaged_scan.address = hstart;
1679 		VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1680 
1681 		while (khugepaged_scan.address < hend) {
1682 			int ret;
1683 			cond_resched();
1684 			if (unlikely(khugepaged_test_exit(mm)))
1685 				goto breakouterloop;
1686 
1687 			VM_BUG_ON(khugepaged_scan.address < hstart ||
1688 				  khugepaged_scan.address + HPAGE_PMD_SIZE >
1689 				  hend);
1690 			if (shmem_file(vma->vm_file)) {
1691 				struct file *file;
1692 				pgoff_t pgoff = linear_page_index(vma,
1693 						khugepaged_scan.address);
1694 				if (!shmem_huge_enabled(vma))
1695 					goto skip;
1696 				file = get_file(vma->vm_file);
1697 				up_read(&mm->mmap_sem);
1698 				ret = 1;
1699 				khugepaged_scan_shmem(mm, file->f_mapping,
1700 						pgoff, hpage);
1701 				fput(file);
1702 			} else {
1703 				ret = khugepaged_scan_pmd(mm, vma,
1704 						khugepaged_scan.address,
1705 						hpage);
1706 			}
1707 			/* move to next address */
1708 			khugepaged_scan.address += HPAGE_PMD_SIZE;
1709 			progress += HPAGE_PMD_NR;
1710 			if (ret)
1711 				/* we released mmap_sem so break loop */
1712 				goto breakouterloop_mmap_sem;
1713 			if (progress >= pages)
1714 				goto breakouterloop;
1715 		}
1716 	}
1717 breakouterloop:
1718 	up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1719 breakouterloop_mmap_sem:
1720 
1721 	spin_lock(&khugepaged_mm_lock);
1722 	VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1723 	/*
1724 	 * Release the current mm_slot if this mm is about to die, or
1725 	 * if we scanned all vmas of this mm.
1726 	 */
1727 	if (khugepaged_test_exit(mm) || !vma) {
1728 		/*
1729 		 * Make sure that if mm_users is reaching zero while
1730 		 * khugepaged runs here, khugepaged_exit will find
1731 		 * mm_slot not pointing to the exiting mm.
1732 		 */
1733 		if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1734 			khugepaged_scan.mm_slot = list_entry(
1735 				mm_slot->mm_node.next,
1736 				struct mm_slot, mm_node);
1737 			khugepaged_scan.address = 0;
1738 		} else {
1739 			khugepaged_scan.mm_slot = NULL;
1740 			khugepaged_full_scans++;
1741 		}
1742 
1743 		collect_mm_slot(mm_slot);
1744 	}
1745 
1746 	return progress;
1747 }
1748 
1749 static int khugepaged_has_work(void)
1750 {
1751 	return !list_empty(&khugepaged_scan.mm_head) &&
1752 		khugepaged_enabled();
1753 }
1754 
1755 static int khugepaged_wait_event(void)
1756 {
1757 	return !list_empty(&khugepaged_scan.mm_head) ||
1758 		kthread_should_stop();
1759 }
1760 
1761 static void khugepaged_do_scan(void)
1762 {
1763 	struct page *hpage = NULL;
1764 	unsigned int progress = 0, pass_through_head = 0;
1765 	unsigned int pages = khugepaged_pages_to_scan;
1766 	bool wait = true;
1767 
1768 	barrier(); /* write khugepaged_pages_to_scan to local stack */
1769 
1770 	while (progress < pages) {
1771 		if (!khugepaged_prealloc_page(&hpage, &wait))
1772 			break;
1773 
1774 		cond_resched();
1775 
1776 		if (unlikely(kthread_should_stop() || try_to_freeze()))
1777 			break;
1778 
1779 		spin_lock(&khugepaged_mm_lock);
1780 		if (!khugepaged_scan.mm_slot)
1781 			pass_through_head++;
1782 		if (khugepaged_has_work() &&
1783 		    pass_through_head < 2)
1784 			progress += khugepaged_scan_mm_slot(pages - progress,
1785 							    &hpage);
1786 		else
1787 			progress = pages;
1788 		spin_unlock(&khugepaged_mm_lock);
1789 	}
1790 
1791 	if (!IS_ERR_OR_NULL(hpage))
1792 		put_page(hpage);
1793 }
1794 
1795 static bool khugepaged_should_wakeup(void)
1796 {
1797 	return kthread_should_stop() ||
1798 	       time_after_eq(jiffies, khugepaged_sleep_expire);
1799 }
1800 
1801 static void khugepaged_wait_work(void)
1802 {
1803 	if (khugepaged_has_work()) {
1804 		const unsigned long scan_sleep_jiffies =
1805 			msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1806 
1807 		if (!scan_sleep_jiffies)
1808 			return;
1809 
1810 		khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1811 		wait_event_freezable_timeout(khugepaged_wait,
1812 					     khugepaged_should_wakeup(),
1813 					     scan_sleep_jiffies);
1814 		return;
1815 	}
1816 
1817 	if (khugepaged_enabled())
1818 		wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1819 }
1820 
1821 static int khugepaged(void *none)
1822 {
1823 	struct mm_slot *mm_slot;
1824 
1825 	set_freezable();
1826 	set_user_nice(current, MAX_NICE);
1827 
1828 	while (!kthread_should_stop()) {
1829 		khugepaged_do_scan();
1830 		khugepaged_wait_work();
1831 	}
1832 
1833 	spin_lock(&khugepaged_mm_lock);
1834 	mm_slot = khugepaged_scan.mm_slot;
1835 	khugepaged_scan.mm_slot = NULL;
1836 	if (mm_slot)
1837 		collect_mm_slot(mm_slot);
1838 	spin_unlock(&khugepaged_mm_lock);
1839 	return 0;
1840 }
1841 
1842 static void set_recommended_min_free_kbytes(void)
1843 {
1844 	struct zone *zone;
1845 	int nr_zones = 0;
1846 	unsigned long recommended_min;
1847 
1848 	for_each_populated_zone(zone) {
1849 		/*
1850 		 * We don't need to worry about fragmentation of
1851 		 * ZONE_MOVABLE since it only has movable pages.
1852 		 */
1853 		if (zone_idx(zone) > gfp_zone(GFP_USER))
1854 			continue;
1855 
1856 		nr_zones++;
1857 	}
1858 
1859 	/* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1860 	recommended_min = pageblock_nr_pages * nr_zones * 2;
1861 
1862 	/*
1863 	 * Make sure that on average at least two pageblocks are almost free
1864 	 * of another type, one for a migratetype to fall back to and a
1865 	 * second to avoid subsequent fallbacks of other types There are 3
1866 	 * MIGRATE_TYPES we care about.
1867 	 */
1868 	recommended_min += pageblock_nr_pages * nr_zones *
1869 			   MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1870 
1871 	/* don't ever allow to reserve more than 5% of the lowmem */
1872 	recommended_min = min(recommended_min,
1873 			      (unsigned long) nr_free_buffer_pages() / 20);
1874 	recommended_min <<= (PAGE_SHIFT-10);
1875 
1876 	if (recommended_min > min_free_kbytes) {
1877 		if (user_min_free_kbytes >= 0)
1878 			pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1879 				min_free_kbytes, recommended_min);
1880 
1881 		min_free_kbytes = recommended_min;
1882 	}
1883 	setup_per_zone_wmarks();
1884 }
1885 
1886 int start_stop_khugepaged(void)
1887 {
1888 	static struct task_struct *khugepaged_thread __read_mostly;
1889 	static DEFINE_MUTEX(khugepaged_mutex);
1890 	int err = 0;
1891 
1892 	mutex_lock(&khugepaged_mutex);
1893 	if (khugepaged_enabled()) {
1894 		if (!khugepaged_thread)
1895 			khugepaged_thread = kthread_run(khugepaged, NULL,
1896 							"khugepaged");
1897 		if (IS_ERR(khugepaged_thread)) {
1898 			pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1899 			err = PTR_ERR(khugepaged_thread);
1900 			khugepaged_thread = NULL;
1901 			goto fail;
1902 		}
1903 
1904 		if (!list_empty(&khugepaged_scan.mm_head))
1905 			wake_up_interruptible(&khugepaged_wait);
1906 
1907 		set_recommended_min_free_kbytes();
1908 	} else if (khugepaged_thread) {
1909 		kthread_stop(khugepaged_thread);
1910 		khugepaged_thread = NULL;
1911 	}
1912 fail:
1913 	mutex_unlock(&khugepaged_mutex);
1914 	return err;
1915 }
1916