xref: /linux/mm/khugepaged.c (revision 48dea9a700c8728cc31a1dd44588b97578de86ee)
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_EXCEED_SWAP_PTE,
32 	SCAN_EXCEED_SHARED_PTE,
33 	SCAN_PTE_NON_PRESENT,
34 	SCAN_PTE_UFFD_WP,
35 	SCAN_PAGE_RO,
36 	SCAN_LACK_REFERENCED_PAGE,
37 	SCAN_PAGE_NULL,
38 	SCAN_SCAN_ABORT,
39 	SCAN_PAGE_COUNT,
40 	SCAN_PAGE_LRU,
41 	SCAN_PAGE_LOCK,
42 	SCAN_PAGE_ANON,
43 	SCAN_PAGE_COMPOUND,
44 	SCAN_ANY_PROCESS,
45 	SCAN_VMA_NULL,
46 	SCAN_VMA_CHECK,
47 	SCAN_ADDRESS_RANGE,
48 	SCAN_SWAP_CACHE_PAGE,
49 	SCAN_DEL_PAGE_LRU,
50 	SCAN_ALLOC_HUGE_PAGE_FAIL,
51 	SCAN_CGROUP_CHARGE_FAIL,
52 	SCAN_TRUNCATED,
53 	SCAN_PAGE_HAS_PRIVATE,
54 };
55 
56 #define CREATE_TRACE_POINTS
57 #include <trace/events/huge_memory.h>
58 
59 /* default scan 8*512 pte (or vmas) every 30 second */
60 static unsigned int khugepaged_pages_to_scan __read_mostly;
61 static unsigned int khugepaged_pages_collapsed;
62 static unsigned int khugepaged_full_scans;
63 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
64 /* during fragmentation poll the hugepage allocator once every minute */
65 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
66 static unsigned long khugepaged_sleep_expire;
67 static DEFINE_SPINLOCK(khugepaged_mm_lock);
68 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
69 /*
70  * default collapse hugepages if there is at least one pte mapped like
71  * it would have happened if the vma was large enough during page
72  * fault.
73  */
74 static unsigned int khugepaged_max_ptes_none __read_mostly;
75 static unsigned int khugepaged_max_ptes_swap __read_mostly;
76 static unsigned int khugepaged_max_ptes_shared __read_mostly;
77 
78 #define MM_SLOTS_HASH_BITS 10
79 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
80 
81 static struct kmem_cache *mm_slot_cache __read_mostly;
82 
83 #define MAX_PTE_MAPPED_THP 8
84 
85 /**
86  * struct mm_slot - hash lookup from mm to mm_slot
87  * @hash: hash collision list
88  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
89  * @mm: the mm that this information is valid for
90  */
91 struct mm_slot {
92 	struct hlist_node hash;
93 	struct list_head mm_node;
94 	struct mm_struct *mm;
95 
96 	/* pte-mapped THP in this mm */
97 	int nr_pte_mapped_thp;
98 	unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
99 };
100 
101 /**
102  * struct khugepaged_scan - cursor for scanning
103  * @mm_head: the head of the mm list to scan
104  * @mm_slot: the current mm_slot we are scanning
105  * @address: the next address inside that to be scanned
106  *
107  * There is only the one khugepaged_scan instance of this cursor structure.
108  */
109 struct khugepaged_scan {
110 	struct list_head mm_head;
111 	struct mm_slot *mm_slot;
112 	unsigned long address;
113 };
114 
115 static struct khugepaged_scan khugepaged_scan = {
116 	.mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
117 };
118 
119 #ifdef CONFIG_SYSFS
120 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
121 					 struct kobj_attribute *attr,
122 					 char *buf)
123 {
124 	return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
125 }
126 
127 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
128 					  struct kobj_attribute *attr,
129 					  const char *buf, size_t count)
130 {
131 	unsigned long msecs;
132 	int err;
133 
134 	err = kstrtoul(buf, 10, &msecs);
135 	if (err || msecs > UINT_MAX)
136 		return -EINVAL;
137 
138 	khugepaged_scan_sleep_millisecs = msecs;
139 	khugepaged_sleep_expire = 0;
140 	wake_up_interruptible(&khugepaged_wait);
141 
142 	return count;
143 }
144 static struct kobj_attribute scan_sleep_millisecs_attr =
145 	__ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
146 	       scan_sleep_millisecs_store);
147 
148 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
149 					  struct kobj_attribute *attr,
150 					  char *buf)
151 {
152 	return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
153 }
154 
155 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
156 					   struct kobj_attribute *attr,
157 					   const char *buf, size_t count)
158 {
159 	unsigned long msecs;
160 	int err;
161 
162 	err = kstrtoul(buf, 10, &msecs);
163 	if (err || msecs > UINT_MAX)
164 		return -EINVAL;
165 
166 	khugepaged_alloc_sleep_millisecs = msecs;
167 	khugepaged_sleep_expire = 0;
168 	wake_up_interruptible(&khugepaged_wait);
169 
170 	return count;
171 }
172 static struct kobj_attribute alloc_sleep_millisecs_attr =
173 	__ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
174 	       alloc_sleep_millisecs_store);
175 
176 static ssize_t pages_to_scan_show(struct kobject *kobj,
177 				  struct kobj_attribute *attr,
178 				  char *buf)
179 {
180 	return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
181 }
182 static ssize_t pages_to_scan_store(struct kobject *kobj,
183 				   struct kobj_attribute *attr,
184 				   const char *buf, size_t count)
185 {
186 	int err;
187 	unsigned long pages;
188 
189 	err = kstrtoul(buf, 10, &pages);
190 	if (err || !pages || pages > UINT_MAX)
191 		return -EINVAL;
192 
193 	khugepaged_pages_to_scan = pages;
194 
195 	return count;
196 }
197 static struct kobj_attribute pages_to_scan_attr =
198 	__ATTR(pages_to_scan, 0644, pages_to_scan_show,
199 	       pages_to_scan_store);
200 
201 static ssize_t pages_collapsed_show(struct kobject *kobj,
202 				    struct kobj_attribute *attr,
203 				    char *buf)
204 {
205 	return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
206 }
207 static struct kobj_attribute pages_collapsed_attr =
208 	__ATTR_RO(pages_collapsed);
209 
210 static ssize_t full_scans_show(struct kobject *kobj,
211 			       struct kobj_attribute *attr,
212 			       char *buf)
213 {
214 	return sprintf(buf, "%u\n", khugepaged_full_scans);
215 }
216 static struct kobj_attribute full_scans_attr =
217 	__ATTR_RO(full_scans);
218 
219 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
220 				      struct kobj_attribute *attr, char *buf)
221 {
222 	return single_hugepage_flag_show(kobj, attr, buf,
223 				TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
224 }
225 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
226 				       struct kobj_attribute *attr,
227 				       const char *buf, size_t count)
228 {
229 	return single_hugepage_flag_store(kobj, attr, buf, count,
230 				 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
231 }
232 static struct kobj_attribute khugepaged_defrag_attr =
233 	__ATTR(defrag, 0644, khugepaged_defrag_show,
234 	       khugepaged_defrag_store);
235 
236 /*
237  * max_ptes_none controls if khugepaged should collapse hugepages over
238  * any unmapped ptes in turn potentially increasing the memory
239  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
240  * reduce the available free memory in the system as it
241  * runs. Increasing max_ptes_none will instead potentially reduce the
242  * free memory in the system during the khugepaged scan.
243  */
244 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
245 					     struct kobj_attribute *attr,
246 					     char *buf)
247 {
248 	return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
249 }
250 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
251 					      struct kobj_attribute *attr,
252 					      const char *buf, size_t count)
253 {
254 	int err;
255 	unsigned long max_ptes_none;
256 
257 	err = kstrtoul(buf, 10, &max_ptes_none);
258 	if (err || max_ptes_none > HPAGE_PMD_NR-1)
259 		return -EINVAL;
260 
261 	khugepaged_max_ptes_none = max_ptes_none;
262 
263 	return count;
264 }
265 static struct kobj_attribute khugepaged_max_ptes_none_attr =
266 	__ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
267 	       khugepaged_max_ptes_none_store);
268 
269 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
270 					     struct kobj_attribute *attr,
271 					     char *buf)
272 {
273 	return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
274 }
275 
276 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
277 					      struct kobj_attribute *attr,
278 					      const char *buf, size_t count)
279 {
280 	int err;
281 	unsigned long max_ptes_swap;
282 
283 	err  = kstrtoul(buf, 10, &max_ptes_swap);
284 	if (err || max_ptes_swap > HPAGE_PMD_NR-1)
285 		return -EINVAL;
286 
287 	khugepaged_max_ptes_swap = max_ptes_swap;
288 
289 	return count;
290 }
291 
292 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
293 	__ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
294 	       khugepaged_max_ptes_swap_store);
295 
296 static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
297 					     struct kobj_attribute *attr,
298 					     char *buf)
299 {
300 	return sprintf(buf, "%u\n", khugepaged_max_ptes_shared);
301 }
302 
303 static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
304 					      struct kobj_attribute *attr,
305 					      const char *buf, size_t count)
306 {
307 	int err;
308 	unsigned long max_ptes_shared;
309 
310 	err  = kstrtoul(buf, 10, &max_ptes_shared);
311 	if (err || max_ptes_shared > HPAGE_PMD_NR-1)
312 		return -EINVAL;
313 
314 	khugepaged_max_ptes_shared = max_ptes_shared;
315 
316 	return count;
317 }
318 
319 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
320 	__ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
321 	       khugepaged_max_ptes_shared_store);
322 
323 static struct attribute *khugepaged_attr[] = {
324 	&khugepaged_defrag_attr.attr,
325 	&khugepaged_max_ptes_none_attr.attr,
326 	&khugepaged_max_ptes_swap_attr.attr,
327 	&khugepaged_max_ptes_shared_attr.attr,
328 	&pages_to_scan_attr.attr,
329 	&pages_collapsed_attr.attr,
330 	&full_scans_attr.attr,
331 	&scan_sleep_millisecs_attr.attr,
332 	&alloc_sleep_millisecs_attr.attr,
333 	NULL,
334 };
335 
336 struct attribute_group khugepaged_attr_group = {
337 	.attrs = khugepaged_attr,
338 	.name = "khugepaged",
339 };
340 #endif /* CONFIG_SYSFS */
341 
342 int hugepage_madvise(struct vm_area_struct *vma,
343 		     unsigned long *vm_flags, int advice)
344 {
345 	switch (advice) {
346 	case MADV_HUGEPAGE:
347 #ifdef CONFIG_S390
348 		/*
349 		 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
350 		 * can't handle this properly after s390_enable_sie, so we simply
351 		 * ignore the madvise to prevent qemu from causing a SIGSEGV.
352 		 */
353 		if (mm_has_pgste(vma->vm_mm))
354 			return 0;
355 #endif
356 		*vm_flags &= ~VM_NOHUGEPAGE;
357 		*vm_flags |= VM_HUGEPAGE;
358 		/*
359 		 * If the vma become good for khugepaged to scan,
360 		 * register it here without waiting a page fault that
361 		 * may not happen any time soon.
362 		 */
363 		if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
364 				khugepaged_enter_vma_merge(vma, *vm_flags))
365 			return -ENOMEM;
366 		break;
367 	case MADV_NOHUGEPAGE:
368 		*vm_flags &= ~VM_HUGEPAGE;
369 		*vm_flags |= VM_NOHUGEPAGE;
370 		/*
371 		 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
372 		 * this vma even if we leave the mm registered in khugepaged if
373 		 * it got registered before VM_NOHUGEPAGE was set.
374 		 */
375 		break;
376 	}
377 
378 	return 0;
379 }
380 
381 int __init khugepaged_init(void)
382 {
383 	mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
384 					  sizeof(struct mm_slot),
385 					  __alignof__(struct mm_slot), 0, NULL);
386 	if (!mm_slot_cache)
387 		return -ENOMEM;
388 
389 	khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
390 	khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
391 	khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
392 	khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
393 
394 	return 0;
395 }
396 
397 void __init khugepaged_destroy(void)
398 {
399 	kmem_cache_destroy(mm_slot_cache);
400 }
401 
402 static inline struct mm_slot *alloc_mm_slot(void)
403 {
404 	if (!mm_slot_cache)	/* initialization failed */
405 		return NULL;
406 	return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
407 }
408 
409 static inline void free_mm_slot(struct mm_slot *mm_slot)
410 {
411 	kmem_cache_free(mm_slot_cache, mm_slot);
412 }
413 
414 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
415 {
416 	struct mm_slot *mm_slot;
417 
418 	hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
419 		if (mm == mm_slot->mm)
420 			return mm_slot;
421 
422 	return NULL;
423 }
424 
425 static void insert_to_mm_slots_hash(struct mm_struct *mm,
426 				    struct mm_slot *mm_slot)
427 {
428 	mm_slot->mm = mm;
429 	hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
430 }
431 
432 static inline int khugepaged_test_exit(struct mm_struct *mm)
433 {
434 	return atomic_read(&mm->mm_users) == 0 || !mmget_still_valid(mm);
435 }
436 
437 static bool hugepage_vma_check(struct vm_area_struct *vma,
438 			       unsigned long vm_flags)
439 {
440 	if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
441 	    (vm_flags & VM_NOHUGEPAGE) ||
442 	    test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
443 		return false;
444 
445 	if (shmem_file(vma->vm_file) ||
446 	    (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
447 	     vma->vm_file &&
448 	     (vm_flags & VM_DENYWRITE))) {
449 		return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
450 				HPAGE_PMD_NR);
451 	}
452 	if (!vma->anon_vma || vma->vm_ops)
453 		return false;
454 	if (vma_is_temporary_stack(vma))
455 		return false;
456 	return !(vm_flags & VM_NO_KHUGEPAGED);
457 }
458 
459 int __khugepaged_enter(struct mm_struct *mm)
460 {
461 	struct mm_slot *mm_slot;
462 	int wakeup;
463 
464 	mm_slot = alloc_mm_slot();
465 	if (!mm_slot)
466 		return -ENOMEM;
467 
468 	/* __khugepaged_exit() must not run from under us */
469 	VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
470 	if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
471 		free_mm_slot(mm_slot);
472 		return 0;
473 	}
474 
475 	spin_lock(&khugepaged_mm_lock);
476 	insert_to_mm_slots_hash(mm, mm_slot);
477 	/*
478 	 * Insert just behind the scanning cursor, to let the area settle
479 	 * down a little.
480 	 */
481 	wakeup = list_empty(&khugepaged_scan.mm_head);
482 	list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
483 	spin_unlock(&khugepaged_mm_lock);
484 
485 	mmgrab(mm);
486 	if (wakeup)
487 		wake_up_interruptible(&khugepaged_wait);
488 
489 	return 0;
490 }
491 
492 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
493 			       unsigned long vm_flags)
494 {
495 	unsigned long hstart, hend;
496 
497 	/*
498 	 * khugepaged only supports read-only files for non-shmem files.
499 	 * khugepaged does not yet work on special mappings. And
500 	 * file-private shmem THP is not supported.
501 	 */
502 	if (!hugepage_vma_check(vma, vm_flags))
503 		return 0;
504 
505 	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
506 	hend = vma->vm_end & HPAGE_PMD_MASK;
507 	if (hstart < hend)
508 		return khugepaged_enter(vma, vm_flags);
509 	return 0;
510 }
511 
512 void __khugepaged_exit(struct mm_struct *mm)
513 {
514 	struct mm_slot *mm_slot;
515 	int free = 0;
516 
517 	spin_lock(&khugepaged_mm_lock);
518 	mm_slot = get_mm_slot(mm);
519 	if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
520 		hash_del(&mm_slot->hash);
521 		list_del(&mm_slot->mm_node);
522 		free = 1;
523 	}
524 	spin_unlock(&khugepaged_mm_lock);
525 
526 	if (free) {
527 		clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
528 		free_mm_slot(mm_slot);
529 		mmdrop(mm);
530 	} else if (mm_slot) {
531 		/*
532 		 * This is required to serialize against
533 		 * khugepaged_test_exit() (which is guaranteed to run
534 		 * under mmap sem read mode). Stop here (after we
535 		 * return all pagetables will be destroyed) until
536 		 * khugepaged has finished working on the pagetables
537 		 * under the mmap_lock.
538 		 */
539 		mmap_write_lock(mm);
540 		mmap_write_unlock(mm);
541 	}
542 }
543 
544 static void release_pte_page(struct page *page)
545 {
546 	mod_node_page_state(page_pgdat(page),
547 			NR_ISOLATED_ANON + page_is_file_lru(page),
548 			-compound_nr(page));
549 	unlock_page(page);
550 	putback_lru_page(page);
551 }
552 
553 static void release_pte_pages(pte_t *pte, pte_t *_pte,
554 		struct list_head *compound_pagelist)
555 {
556 	struct page *page, *tmp;
557 
558 	while (--_pte >= pte) {
559 		pte_t pteval = *_pte;
560 
561 		page = pte_page(pteval);
562 		if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
563 				!PageCompound(page))
564 			release_pte_page(page);
565 	}
566 
567 	list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
568 		list_del(&page->lru);
569 		release_pte_page(page);
570 	}
571 }
572 
573 static bool is_refcount_suitable(struct page *page)
574 {
575 	int expected_refcount;
576 
577 	expected_refcount = total_mapcount(page);
578 	if (PageSwapCache(page))
579 		expected_refcount += compound_nr(page);
580 
581 	return page_count(page) == expected_refcount;
582 }
583 
584 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
585 					unsigned long address,
586 					pte_t *pte,
587 					struct list_head *compound_pagelist)
588 {
589 	struct page *page = NULL;
590 	pte_t *_pte;
591 	int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
592 	bool writable = false;
593 
594 	for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
595 	     _pte++, address += PAGE_SIZE) {
596 		pte_t pteval = *_pte;
597 		if (pte_none(pteval) || (pte_present(pteval) &&
598 				is_zero_pfn(pte_pfn(pteval)))) {
599 			if (!userfaultfd_armed(vma) &&
600 			    ++none_or_zero <= khugepaged_max_ptes_none) {
601 				continue;
602 			} else {
603 				result = SCAN_EXCEED_NONE_PTE;
604 				goto out;
605 			}
606 		}
607 		if (!pte_present(pteval)) {
608 			result = SCAN_PTE_NON_PRESENT;
609 			goto out;
610 		}
611 		page = vm_normal_page(vma, address, pteval);
612 		if (unlikely(!page)) {
613 			result = SCAN_PAGE_NULL;
614 			goto out;
615 		}
616 
617 		VM_BUG_ON_PAGE(!PageAnon(page), page);
618 
619 		if (page_mapcount(page) > 1 &&
620 				++shared > khugepaged_max_ptes_shared) {
621 			result = SCAN_EXCEED_SHARED_PTE;
622 			goto out;
623 		}
624 
625 		if (PageCompound(page)) {
626 			struct page *p;
627 			page = compound_head(page);
628 
629 			/*
630 			 * Check if we have dealt with the compound page
631 			 * already
632 			 */
633 			list_for_each_entry(p, compound_pagelist, lru) {
634 				if (page == p)
635 					goto next;
636 			}
637 		}
638 
639 		/*
640 		 * We can do it before isolate_lru_page because the
641 		 * page can't be freed from under us. NOTE: PG_lock
642 		 * is needed to serialize against split_huge_page
643 		 * when invoked from the VM.
644 		 */
645 		if (!trylock_page(page)) {
646 			result = SCAN_PAGE_LOCK;
647 			goto out;
648 		}
649 
650 		/*
651 		 * Check if the page has any GUP (or other external) pins.
652 		 *
653 		 * The page table that maps the page has been already unlinked
654 		 * from the page table tree and this process cannot get
655 		 * an additinal pin on the page.
656 		 *
657 		 * New pins can come later if the page is shared across fork,
658 		 * but not from this process. The other process cannot write to
659 		 * the page, only trigger CoW.
660 		 */
661 		if (!is_refcount_suitable(page)) {
662 			unlock_page(page);
663 			result = SCAN_PAGE_COUNT;
664 			goto out;
665 		}
666 		if (!pte_write(pteval) && PageSwapCache(page) &&
667 				!reuse_swap_page(page, NULL)) {
668 			/*
669 			 * Page is in the swap cache and cannot be re-used.
670 			 * It cannot be collapsed into a THP.
671 			 */
672 			unlock_page(page);
673 			result = SCAN_SWAP_CACHE_PAGE;
674 			goto out;
675 		}
676 
677 		/*
678 		 * Isolate the page to avoid collapsing an hugepage
679 		 * currently in use by the VM.
680 		 */
681 		if (isolate_lru_page(page)) {
682 			unlock_page(page);
683 			result = SCAN_DEL_PAGE_LRU;
684 			goto out;
685 		}
686 		mod_node_page_state(page_pgdat(page),
687 				NR_ISOLATED_ANON + page_is_file_lru(page),
688 				compound_nr(page));
689 		VM_BUG_ON_PAGE(!PageLocked(page), page);
690 		VM_BUG_ON_PAGE(PageLRU(page), page);
691 
692 		if (PageCompound(page))
693 			list_add_tail(&page->lru, compound_pagelist);
694 next:
695 		/* There should be enough young pte to collapse the page */
696 		if (pte_young(pteval) ||
697 		    page_is_young(page) || PageReferenced(page) ||
698 		    mmu_notifier_test_young(vma->vm_mm, address))
699 			referenced++;
700 
701 		if (pte_write(pteval))
702 			writable = true;
703 	}
704 	if (likely(writable)) {
705 		if (likely(referenced)) {
706 			result = SCAN_SUCCEED;
707 			trace_mm_collapse_huge_page_isolate(page, none_or_zero,
708 							    referenced, writable, result);
709 			return 1;
710 		}
711 	} else {
712 		result = SCAN_PAGE_RO;
713 	}
714 
715 out:
716 	release_pte_pages(pte, _pte, compound_pagelist);
717 	trace_mm_collapse_huge_page_isolate(page, none_or_zero,
718 					    referenced, writable, result);
719 	return 0;
720 }
721 
722 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
723 				      struct vm_area_struct *vma,
724 				      unsigned long address,
725 				      spinlock_t *ptl,
726 				      struct list_head *compound_pagelist)
727 {
728 	struct page *src_page, *tmp;
729 	pte_t *_pte;
730 	for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
731 				_pte++, page++, address += PAGE_SIZE) {
732 		pte_t pteval = *_pte;
733 
734 		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
735 			clear_user_highpage(page, address);
736 			add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
737 			if (is_zero_pfn(pte_pfn(pteval))) {
738 				/*
739 				 * ptl mostly unnecessary.
740 				 */
741 				spin_lock(ptl);
742 				/*
743 				 * paravirt calls inside pte_clear here are
744 				 * superfluous.
745 				 */
746 				pte_clear(vma->vm_mm, address, _pte);
747 				spin_unlock(ptl);
748 			}
749 		} else {
750 			src_page = pte_page(pteval);
751 			copy_user_highpage(page, src_page, address, vma);
752 			if (!PageCompound(src_page))
753 				release_pte_page(src_page);
754 			/*
755 			 * ptl mostly unnecessary, but preempt has to
756 			 * be disabled to update the per-cpu stats
757 			 * inside page_remove_rmap().
758 			 */
759 			spin_lock(ptl);
760 			/*
761 			 * paravirt calls inside pte_clear here are
762 			 * superfluous.
763 			 */
764 			pte_clear(vma->vm_mm, address, _pte);
765 			page_remove_rmap(src_page, false);
766 			spin_unlock(ptl);
767 			free_page_and_swap_cache(src_page);
768 		}
769 	}
770 
771 	list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
772 		list_del(&src_page->lru);
773 		release_pte_page(src_page);
774 	}
775 }
776 
777 static void khugepaged_alloc_sleep(void)
778 {
779 	DEFINE_WAIT(wait);
780 
781 	add_wait_queue(&khugepaged_wait, &wait);
782 	freezable_schedule_timeout_interruptible(
783 		msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
784 	remove_wait_queue(&khugepaged_wait, &wait);
785 }
786 
787 static int khugepaged_node_load[MAX_NUMNODES];
788 
789 static bool khugepaged_scan_abort(int nid)
790 {
791 	int i;
792 
793 	/*
794 	 * If node_reclaim_mode is disabled, then no extra effort is made to
795 	 * allocate memory locally.
796 	 */
797 	if (!node_reclaim_mode)
798 		return false;
799 
800 	/* If there is a count for this node already, it must be acceptable */
801 	if (khugepaged_node_load[nid])
802 		return false;
803 
804 	for (i = 0; i < MAX_NUMNODES; i++) {
805 		if (!khugepaged_node_load[i])
806 			continue;
807 		if (node_distance(nid, i) > node_reclaim_distance)
808 			return true;
809 	}
810 	return false;
811 }
812 
813 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
814 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
815 {
816 	return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
817 }
818 
819 #ifdef CONFIG_NUMA
820 static int khugepaged_find_target_node(void)
821 {
822 	static int last_khugepaged_target_node = NUMA_NO_NODE;
823 	int nid, target_node = 0, max_value = 0;
824 
825 	/* find first node with max normal pages hit */
826 	for (nid = 0; nid < MAX_NUMNODES; nid++)
827 		if (khugepaged_node_load[nid] > max_value) {
828 			max_value = khugepaged_node_load[nid];
829 			target_node = nid;
830 		}
831 
832 	/* do some balance if several nodes have the same hit record */
833 	if (target_node <= last_khugepaged_target_node)
834 		for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
835 				nid++)
836 			if (max_value == khugepaged_node_load[nid]) {
837 				target_node = nid;
838 				break;
839 			}
840 
841 	last_khugepaged_target_node = target_node;
842 	return target_node;
843 }
844 
845 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
846 {
847 	if (IS_ERR(*hpage)) {
848 		if (!*wait)
849 			return false;
850 
851 		*wait = false;
852 		*hpage = NULL;
853 		khugepaged_alloc_sleep();
854 	} else if (*hpage) {
855 		put_page(*hpage);
856 		*hpage = NULL;
857 	}
858 
859 	return true;
860 }
861 
862 static struct page *
863 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
864 {
865 	VM_BUG_ON_PAGE(*hpage, *hpage);
866 
867 	*hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
868 	if (unlikely(!*hpage)) {
869 		count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
870 		*hpage = ERR_PTR(-ENOMEM);
871 		return NULL;
872 	}
873 
874 	prep_transhuge_page(*hpage);
875 	count_vm_event(THP_COLLAPSE_ALLOC);
876 	return *hpage;
877 }
878 #else
879 static int khugepaged_find_target_node(void)
880 {
881 	return 0;
882 }
883 
884 static inline struct page *alloc_khugepaged_hugepage(void)
885 {
886 	struct page *page;
887 
888 	page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
889 			   HPAGE_PMD_ORDER);
890 	if (page)
891 		prep_transhuge_page(page);
892 	return page;
893 }
894 
895 static struct page *khugepaged_alloc_hugepage(bool *wait)
896 {
897 	struct page *hpage;
898 
899 	do {
900 		hpage = alloc_khugepaged_hugepage();
901 		if (!hpage) {
902 			count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
903 			if (!*wait)
904 				return NULL;
905 
906 			*wait = false;
907 			khugepaged_alloc_sleep();
908 		} else
909 			count_vm_event(THP_COLLAPSE_ALLOC);
910 	} while (unlikely(!hpage) && likely(khugepaged_enabled()));
911 
912 	return hpage;
913 }
914 
915 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
916 {
917 	if (!*hpage)
918 		*hpage = khugepaged_alloc_hugepage(wait);
919 
920 	if (unlikely(!*hpage))
921 		return false;
922 
923 	return true;
924 }
925 
926 static struct page *
927 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
928 {
929 	VM_BUG_ON(!*hpage);
930 
931 	return  *hpage;
932 }
933 #endif
934 
935 /*
936  * If mmap_lock temporarily dropped, revalidate vma
937  * before taking mmap_lock.
938  * Return 0 if succeeds, otherwise return none-zero
939  * value (scan code).
940  */
941 
942 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
943 		struct vm_area_struct **vmap)
944 {
945 	struct vm_area_struct *vma;
946 	unsigned long hstart, hend;
947 
948 	if (unlikely(khugepaged_test_exit(mm)))
949 		return SCAN_ANY_PROCESS;
950 
951 	*vmap = vma = find_vma(mm, address);
952 	if (!vma)
953 		return SCAN_VMA_NULL;
954 
955 	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
956 	hend = vma->vm_end & HPAGE_PMD_MASK;
957 	if (address < hstart || address + HPAGE_PMD_SIZE > hend)
958 		return SCAN_ADDRESS_RANGE;
959 	if (!hugepage_vma_check(vma, vma->vm_flags))
960 		return SCAN_VMA_CHECK;
961 	/* Anon VMA expected */
962 	if (!vma->anon_vma || vma->vm_ops)
963 		return SCAN_VMA_CHECK;
964 	return 0;
965 }
966 
967 /*
968  * Bring missing pages in from swap, to complete THP collapse.
969  * Only done if khugepaged_scan_pmd believes it is worthwhile.
970  *
971  * Called and returns without pte mapped or spinlocks held,
972  * but with mmap_lock held to protect against vma changes.
973  */
974 
975 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
976 					struct vm_area_struct *vma,
977 					unsigned long address, pmd_t *pmd,
978 					int referenced)
979 {
980 	int swapped_in = 0;
981 	vm_fault_t ret = 0;
982 	struct vm_fault vmf = {
983 		.vma = vma,
984 		.address = address,
985 		.flags = FAULT_FLAG_ALLOW_RETRY,
986 		.pmd = pmd,
987 		.pgoff = linear_page_index(vma, address),
988 	};
989 
990 	vmf.pte = pte_offset_map(pmd, address);
991 	for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
992 			vmf.pte++, vmf.address += PAGE_SIZE) {
993 		vmf.orig_pte = *vmf.pte;
994 		if (!is_swap_pte(vmf.orig_pte))
995 			continue;
996 		swapped_in++;
997 		ret = do_swap_page(&vmf);
998 
999 		/* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1000 		if (ret & VM_FAULT_RETRY) {
1001 			mmap_read_lock(mm);
1002 			if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
1003 				/* vma is no longer available, don't continue to swapin */
1004 				trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1005 				return false;
1006 			}
1007 			/* check if the pmd is still valid */
1008 			if (mm_find_pmd(mm, address) != pmd) {
1009 				trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1010 				return false;
1011 			}
1012 		}
1013 		if (ret & VM_FAULT_ERROR) {
1014 			trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1015 			return false;
1016 		}
1017 		/* pte is unmapped now, we need to map it */
1018 		vmf.pte = pte_offset_map(pmd, vmf.address);
1019 	}
1020 	vmf.pte--;
1021 	pte_unmap(vmf.pte);
1022 
1023 	/* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1024 	if (swapped_in)
1025 		lru_add_drain();
1026 
1027 	trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1028 	return true;
1029 }
1030 
1031 static void collapse_huge_page(struct mm_struct *mm,
1032 				   unsigned long address,
1033 				   struct page **hpage,
1034 				   int node, int referenced, int unmapped)
1035 {
1036 	LIST_HEAD(compound_pagelist);
1037 	pmd_t *pmd, _pmd;
1038 	pte_t *pte;
1039 	pgtable_t pgtable;
1040 	struct page *new_page;
1041 	spinlock_t *pmd_ptl, *pte_ptl;
1042 	int isolated = 0, result = 0;
1043 	struct vm_area_struct *vma;
1044 	struct mmu_notifier_range range;
1045 	gfp_t gfp;
1046 
1047 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1048 
1049 	/* Only allocate from the target node */
1050 	gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1051 
1052 	/*
1053 	 * Before allocating the hugepage, release the mmap_lock read lock.
1054 	 * The allocation can take potentially a long time if it involves
1055 	 * sync compaction, and we do not need to hold the mmap_lock during
1056 	 * that. We will recheck the vma after taking it again in write mode.
1057 	 */
1058 	mmap_read_unlock(mm);
1059 	new_page = khugepaged_alloc_page(hpage, gfp, node);
1060 	if (!new_page) {
1061 		result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1062 		goto out_nolock;
1063 	}
1064 
1065 	if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1066 		result = SCAN_CGROUP_CHARGE_FAIL;
1067 		goto out_nolock;
1068 	}
1069 	count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1070 
1071 	mmap_read_lock(mm);
1072 	result = hugepage_vma_revalidate(mm, address, &vma);
1073 	if (result) {
1074 		mmap_read_unlock(mm);
1075 		goto out_nolock;
1076 	}
1077 
1078 	pmd = mm_find_pmd(mm, address);
1079 	if (!pmd) {
1080 		result = SCAN_PMD_NULL;
1081 		mmap_read_unlock(mm);
1082 		goto out_nolock;
1083 	}
1084 
1085 	/*
1086 	 * __collapse_huge_page_swapin always returns with mmap_lock locked.
1087 	 * If it fails, we release mmap_lock and jump out_nolock.
1088 	 * Continuing to collapse causes inconsistency.
1089 	 */
1090 	if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1091 						     pmd, referenced)) {
1092 		mmap_read_unlock(mm);
1093 		goto out_nolock;
1094 	}
1095 
1096 	mmap_read_unlock(mm);
1097 	/*
1098 	 * Prevent all access to pagetables with the exception of
1099 	 * gup_fast later handled by the ptep_clear_flush and the VM
1100 	 * handled by the anon_vma lock + PG_lock.
1101 	 */
1102 	mmap_write_lock(mm);
1103 	result = hugepage_vma_revalidate(mm, address, &vma);
1104 	if (result)
1105 		goto out;
1106 	/* check if the pmd is still valid */
1107 	if (mm_find_pmd(mm, address) != pmd)
1108 		goto out;
1109 
1110 	anon_vma_lock_write(vma->anon_vma);
1111 
1112 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1113 				address, address + HPAGE_PMD_SIZE);
1114 	mmu_notifier_invalidate_range_start(&range);
1115 
1116 	pte = pte_offset_map(pmd, address);
1117 	pte_ptl = pte_lockptr(mm, pmd);
1118 
1119 	pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1120 	/*
1121 	 * After this gup_fast can't run anymore. This also removes
1122 	 * any huge TLB entry from the CPU so we won't allow
1123 	 * huge and small TLB entries for the same virtual address
1124 	 * to avoid the risk of CPU bugs in that area.
1125 	 */
1126 	_pmd = pmdp_collapse_flush(vma, address, pmd);
1127 	spin_unlock(pmd_ptl);
1128 	mmu_notifier_invalidate_range_end(&range);
1129 
1130 	spin_lock(pte_ptl);
1131 	isolated = __collapse_huge_page_isolate(vma, address, pte,
1132 			&compound_pagelist);
1133 	spin_unlock(pte_ptl);
1134 
1135 	if (unlikely(!isolated)) {
1136 		pte_unmap(pte);
1137 		spin_lock(pmd_ptl);
1138 		BUG_ON(!pmd_none(*pmd));
1139 		/*
1140 		 * We can only use set_pmd_at when establishing
1141 		 * hugepmds and never for establishing regular pmds that
1142 		 * points to regular pagetables. Use pmd_populate for that
1143 		 */
1144 		pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1145 		spin_unlock(pmd_ptl);
1146 		anon_vma_unlock_write(vma->anon_vma);
1147 		result = SCAN_FAIL;
1148 		goto out;
1149 	}
1150 
1151 	/*
1152 	 * All pages are isolated and locked so anon_vma rmap
1153 	 * can't run anymore.
1154 	 */
1155 	anon_vma_unlock_write(vma->anon_vma);
1156 
1157 	__collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1158 			&compound_pagelist);
1159 	pte_unmap(pte);
1160 	__SetPageUptodate(new_page);
1161 	pgtable = pmd_pgtable(_pmd);
1162 
1163 	_pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1164 	_pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1165 
1166 	/*
1167 	 * spin_lock() below is not the equivalent of smp_wmb(), so
1168 	 * this is needed to avoid the copy_huge_page writes to become
1169 	 * visible after the set_pmd_at() write.
1170 	 */
1171 	smp_wmb();
1172 
1173 	spin_lock(pmd_ptl);
1174 	BUG_ON(!pmd_none(*pmd));
1175 	page_add_new_anon_rmap(new_page, vma, address, true);
1176 	lru_cache_add_inactive_or_unevictable(new_page, vma);
1177 	pgtable_trans_huge_deposit(mm, pmd, pgtable);
1178 	set_pmd_at(mm, address, pmd, _pmd);
1179 	update_mmu_cache_pmd(vma, address, pmd);
1180 	spin_unlock(pmd_ptl);
1181 
1182 	*hpage = NULL;
1183 
1184 	khugepaged_pages_collapsed++;
1185 	result = SCAN_SUCCEED;
1186 out_up_write:
1187 	mmap_write_unlock(mm);
1188 out_nolock:
1189 	if (!IS_ERR_OR_NULL(*hpage))
1190 		mem_cgroup_uncharge(*hpage);
1191 	trace_mm_collapse_huge_page(mm, isolated, result);
1192 	return;
1193 out:
1194 	goto out_up_write;
1195 }
1196 
1197 static int khugepaged_scan_pmd(struct mm_struct *mm,
1198 			       struct vm_area_struct *vma,
1199 			       unsigned long address,
1200 			       struct page **hpage)
1201 {
1202 	pmd_t *pmd;
1203 	pte_t *pte, *_pte;
1204 	int ret = 0, result = 0, referenced = 0;
1205 	int none_or_zero = 0, shared = 0;
1206 	struct page *page = NULL;
1207 	unsigned long _address;
1208 	spinlock_t *ptl;
1209 	int node = NUMA_NO_NODE, unmapped = 0;
1210 	bool writable = false;
1211 
1212 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1213 
1214 	pmd = mm_find_pmd(mm, address);
1215 	if (!pmd) {
1216 		result = SCAN_PMD_NULL;
1217 		goto out;
1218 	}
1219 
1220 	memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1221 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1222 	for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1223 	     _pte++, _address += PAGE_SIZE) {
1224 		pte_t pteval = *_pte;
1225 		if (is_swap_pte(pteval)) {
1226 			if (++unmapped <= khugepaged_max_ptes_swap) {
1227 				/*
1228 				 * Always be strict with uffd-wp
1229 				 * enabled swap entries.  Please see
1230 				 * comment below for pte_uffd_wp().
1231 				 */
1232 				if (pte_swp_uffd_wp(pteval)) {
1233 					result = SCAN_PTE_UFFD_WP;
1234 					goto out_unmap;
1235 				}
1236 				continue;
1237 			} else {
1238 				result = SCAN_EXCEED_SWAP_PTE;
1239 				goto out_unmap;
1240 			}
1241 		}
1242 		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1243 			if (!userfaultfd_armed(vma) &&
1244 			    ++none_or_zero <= khugepaged_max_ptes_none) {
1245 				continue;
1246 			} else {
1247 				result = SCAN_EXCEED_NONE_PTE;
1248 				goto out_unmap;
1249 			}
1250 		}
1251 		if (!pte_present(pteval)) {
1252 			result = SCAN_PTE_NON_PRESENT;
1253 			goto out_unmap;
1254 		}
1255 		if (pte_uffd_wp(pteval)) {
1256 			/*
1257 			 * Don't collapse the page if any of the small
1258 			 * PTEs are armed with uffd write protection.
1259 			 * Here we can also mark the new huge pmd as
1260 			 * write protected if any of the small ones is
1261 			 * marked but that could bring uknown
1262 			 * userfault messages that falls outside of
1263 			 * the registered range.  So, just be simple.
1264 			 */
1265 			result = SCAN_PTE_UFFD_WP;
1266 			goto out_unmap;
1267 		}
1268 		if (pte_write(pteval))
1269 			writable = true;
1270 
1271 		page = vm_normal_page(vma, _address, pteval);
1272 		if (unlikely(!page)) {
1273 			result = SCAN_PAGE_NULL;
1274 			goto out_unmap;
1275 		}
1276 
1277 		if (page_mapcount(page) > 1 &&
1278 				++shared > khugepaged_max_ptes_shared) {
1279 			result = SCAN_EXCEED_SHARED_PTE;
1280 			goto out_unmap;
1281 		}
1282 
1283 		page = compound_head(page);
1284 
1285 		/*
1286 		 * Record which node the original page is from and save this
1287 		 * information to khugepaged_node_load[].
1288 		 * Khupaged will allocate hugepage from the node has the max
1289 		 * hit record.
1290 		 */
1291 		node = page_to_nid(page);
1292 		if (khugepaged_scan_abort(node)) {
1293 			result = SCAN_SCAN_ABORT;
1294 			goto out_unmap;
1295 		}
1296 		khugepaged_node_load[node]++;
1297 		if (!PageLRU(page)) {
1298 			result = SCAN_PAGE_LRU;
1299 			goto out_unmap;
1300 		}
1301 		if (PageLocked(page)) {
1302 			result = SCAN_PAGE_LOCK;
1303 			goto out_unmap;
1304 		}
1305 		if (!PageAnon(page)) {
1306 			result = SCAN_PAGE_ANON;
1307 			goto out_unmap;
1308 		}
1309 
1310 		/*
1311 		 * Check if the page has any GUP (or other external) pins.
1312 		 *
1313 		 * Here the check is racy it may see totmal_mapcount > refcount
1314 		 * in some cases.
1315 		 * For example, one process with one forked child process.
1316 		 * The parent has the PMD split due to MADV_DONTNEED, then
1317 		 * the child is trying unmap the whole PMD, but khugepaged
1318 		 * may be scanning the parent between the child has
1319 		 * PageDoubleMap flag cleared and dec the mapcount.  So
1320 		 * khugepaged may see total_mapcount > refcount.
1321 		 *
1322 		 * But such case is ephemeral we could always retry collapse
1323 		 * later.  However it may report false positive if the page
1324 		 * has excessive GUP pins (i.e. 512).  Anyway the same check
1325 		 * will be done again later the risk seems low.
1326 		 */
1327 		if (!is_refcount_suitable(page)) {
1328 			result = SCAN_PAGE_COUNT;
1329 			goto out_unmap;
1330 		}
1331 		if (pte_young(pteval) ||
1332 		    page_is_young(page) || PageReferenced(page) ||
1333 		    mmu_notifier_test_young(vma->vm_mm, address))
1334 			referenced++;
1335 	}
1336 	if (!writable) {
1337 		result = SCAN_PAGE_RO;
1338 	} else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1339 		result = SCAN_LACK_REFERENCED_PAGE;
1340 	} else {
1341 		result = SCAN_SUCCEED;
1342 		ret = 1;
1343 	}
1344 out_unmap:
1345 	pte_unmap_unlock(pte, ptl);
1346 	if (ret) {
1347 		node = khugepaged_find_target_node();
1348 		/* collapse_huge_page will return with the mmap_lock released */
1349 		collapse_huge_page(mm, address, hpage, node,
1350 				referenced, unmapped);
1351 	}
1352 out:
1353 	trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1354 				     none_or_zero, result, unmapped);
1355 	return ret;
1356 }
1357 
1358 static void collect_mm_slot(struct mm_slot *mm_slot)
1359 {
1360 	struct mm_struct *mm = mm_slot->mm;
1361 
1362 	lockdep_assert_held(&khugepaged_mm_lock);
1363 
1364 	if (khugepaged_test_exit(mm)) {
1365 		/* free mm_slot */
1366 		hash_del(&mm_slot->hash);
1367 		list_del(&mm_slot->mm_node);
1368 
1369 		/*
1370 		 * Not strictly needed because the mm exited already.
1371 		 *
1372 		 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1373 		 */
1374 
1375 		/* khugepaged_mm_lock actually not necessary for the below */
1376 		free_mm_slot(mm_slot);
1377 		mmdrop(mm);
1378 	}
1379 }
1380 
1381 #ifdef CONFIG_SHMEM
1382 /*
1383  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1384  * khugepaged should try to collapse the page table.
1385  */
1386 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1387 					 unsigned long addr)
1388 {
1389 	struct mm_slot *mm_slot;
1390 
1391 	VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1392 
1393 	spin_lock(&khugepaged_mm_lock);
1394 	mm_slot = get_mm_slot(mm);
1395 	if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1396 		mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1397 	spin_unlock(&khugepaged_mm_lock);
1398 	return 0;
1399 }
1400 
1401 /**
1402  * Try to collapse a pte-mapped THP for mm at address haddr.
1403  *
1404  * This function checks whether all the PTEs in the PMD are pointing to the
1405  * right THP. If so, retract the page table so the THP can refault in with
1406  * as pmd-mapped.
1407  */
1408 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1409 {
1410 	unsigned long haddr = addr & HPAGE_PMD_MASK;
1411 	struct vm_area_struct *vma = find_vma(mm, haddr);
1412 	struct page *hpage;
1413 	pte_t *start_pte, *pte;
1414 	pmd_t *pmd, _pmd;
1415 	spinlock_t *ptl;
1416 	int count = 0;
1417 	int i;
1418 
1419 	if (!vma || !vma->vm_file ||
1420 	    vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1421 		return;
1422 
1423 	/*
1424 	 * This vm_flags may not have VM_HUGEPAGE if the page was not
1425 	 * collapsed by this mm. But we can still collapse if the page is
1426 	 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1427 	 * will not fail the vma for missing VM_HUGEPAGE
1428 	 */
1429 	if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1430 		return;
1431 
1432 	hpage = find_lock_page(vma->vm_file->f_mapping,
1433 			       linear_page_index(vma, haddr));
1434 	if (!hpage)
1435 		return;
1436 
1437 	if (!PageHead(hpage))
1438 		goto drop_hpage;
1439 
1440 	pmd = mm_find_pmd(mm, haddr);
1441 	if (!pmd)
1442 		goto drop_hpage;
1443 
1444 	start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1445 
1446 	/* step 1: check all mapped PTEs are to the right huge page */
1447 	for (i = 0, addr = haddr, pte = start_pte;
1448 	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1449 		struct page *page;
1450 
1451 		/* empty pte, skip */
1452 		if (pte_none(*pte))
1453 			continue;
1454 
1455 		/* page swapped out, abort */
1456 		if (!pte_present(*pte))
1457 			goto abort;
1458 
1459 		page = vm_normal_page(vma, addr, *pte);
1460 
1461 		/*
1462 		 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1463 		 * page table, but the new page will not be a subpage of hpage.
1464 		 */
1465 		if (hpage + i != page)
1466 			goto abort;
1467 		count++;
1468 	}
1469 
1470 	/* step 2: adjust rmap */
1471 	for (i = 0, addr = haddr, pte = start_pte;
1472 	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1473 		struct page *page;
1474 
1475 		if (pte_none(*pte))
1476 			continue;
1477 		page = vm_normal_page(vma, addr, *pte);
1478 		page_remove_rmap(page, false);
1479 	}
1480 
1481 	pte_unmap_unlock(start_pte, ptl);
1482 
1483 	/* step 3: set proper refcount and mm_counters. */
1484 	if (count) {
1485 		page_ref_sub(hpage, count);
1486 		add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1487 	}
1488 
1489 	/* step 4: collapse pmd */
1490 	ptl = pmd_lock(vma->vm_mm, pmd);
1491 	_pmd = pmdp_collapse_flush(vma, haddr, pmd);
1492 	spin_unlock(ptl);
1493 	mm_dec_nr_ptes(mm);
1494 	pte_free(mm, pmd_pgtable(_pmd));
1495 
1496 drop_hpage:
1497 	unlock_page(hpage);
1498 	put_page(hpage);
1499 	return;
1500 
1501 abort:
1502 	pte_unmap_unlock(start_pte, ptl);
1503 	goto drop_hpage;
1504 }
1505 
1506 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1507 {
1508 	struct mm_struct *mm = mm_slot->mm;
1509 	int i;
1510 
1511 	if (likely(mm_slot->nr_pte_mapped_thp == 0))
1512 		return 0;
1513 
1514 	if (!mmap_write_trylock(mm))
1515 		return -EBUSY;
1516 
1517 	if (unlikely(khugepaged_test_exit(mm)))
1518 		goto out;
1519 
1520 	for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1521 		collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1522 
1523 out:
1524 	mm_slot->nr_pte_mapped_thp = 0;
1525 	mmap_write_unlock(mm);
1526 	return 0;
1527 }
1528 
1529 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1530 {
1531 	struct vm_area_struct *vma;
1532 	struct mm_struct *mm;
1533 	unsigned long addr;
1534 	pmd_t *pmd, _pmd;
1535 
1536 	i_mmap_lock_write(mapping);
1537 	vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1538 		/*
1539 		 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1540 		 * got written to. These VMAs are likely not worth investing
1541 		 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1542 		 * later.
1543 		 *
1544 		 * Not that vma->anon_vma check is racy: it can be set up after
1545 		 * the check but before we took mmap_lock by the fault path.
1546 		 * But page lock would prevent establishing any new ptes of the
1547 		 * page, so we are safe.
1548 		 *
1549 		 * An alternative would be drop the check, but check that page
1550 		 * table is clear before calling pmdp_collapse_flush() under
1551 		 * ptl. It has higher chance to recover THP for the VMA, but
1552 		 * has higher cost too.
1553 		 */
1554 		if (vma->anon_vma)
1555 			continue;
1556 		addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1557 		if (addr & ~HPAGE_PMD_MASK)
1558 			continue;
1559 		if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1560 			continue;
1561 		mm = vma->vm_mm;
1562 		pmd = mm_find_pmd(mm, addr);
1563 		if (!pmd)
1564 			continue;
1565 		/*
1566 		 * We need exclusive mmap_lock to retract page table.
1567 		 *
1568 		 * We use trylock due to lock inversion: we need to acquire
1569 		 * mmap_lock while holding page lock. Fault path does it in
1570 		 * reverse order. Trylock is a way to avoid deadlock.
1571 		 */
1572 		if (mmap_write_trylock(mm)) {
1573 			if (!khugepaged_test_exit(mm)) {
1574 				spinlock_t *ptl = pmd_lock(mm, pmd);
1575 				/* assume page table is clear */
1576 				_pmd = pmdp_collapse_flush(vma, addr, pmd);
1577 				spin_unlock(ptl);
1578 				mm_dec_nr_ptes(mm);
1579 				pte_free(mm, pmd_pgtable(_pmd));
1580 			}
1581 			mmap_write_unlock(mm);
1582 		} else {
1583 			/* Try again later */
1584 			khugepaged_add_pte_mapped_thp(mm, addr);
1585 		}
1586 	}
1587 	i_mmap_unlock_write(mapping);
1588 }
1589 
1590 /**
1591  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1592  *
1593  * Basic scheme is simple, details are more complex:
1594  *  - allocate and lock a new huge page;
1595  *  - scan page cache replacing old pages with the new one
1596  *    + swap/gup in pages if necessary;
1597  *    + fill in gaps;
1598  *    + keep old pages around in case rollback is required;
1599  *  - if replacing succeeds:
1600  *    + copy data over;
1601  *    + free old pages;
1602  *    + unlock huge page;
1603  *  - if replacing failed;
1604  *    + put all pages back and unfreeze them;
1605  *    + restore gaps in the page cache;
1606  *    + unlock and free huge page;
1607  */
1608 static void collapse_file(struct mm_struct *mm,
1609 		struct file *file, pgoff_t start,
1610 		struct page **hpage, int node)
1611 {
1612 	struct address_space *mapping = file->f_mapping;
1613 	gfp_t gfp;
1614 	struct page *new_page;
1615 	pgoff_t index, end = start + HPAGE_PMD_NR;
1616 	LIST_HEAD(pagelist);
1617 	XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1618 	int nr_none = 0, result = SCAN_SUCCEED;
1619 	bool is_shmem = shmem_file(file);
1620 
1621 	VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1622 	VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1623 
1624 	/* Only allocate from the target node */
1625 	gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1626 
1627 	new_page = khugepaged_alloc_page(hpage, gfp, node);
1628 	if (!new_page) {
1629 		result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1630 		goto out;
1631 	}
1632 
1633 	if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1634 		result = SCAN_CGROUP_CHARGE_FAIL;
1635 		goto out;
1636 	}
1637 	count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1638 
1639 	/* This will be less messy when we use multi-index entries */
1640 	do {
1641 		xas_lock_irq(&xas);
1642 		xas_create_range(&xas);
1643 		if (!xas_error(&xas))
1644 			break;
1645 		xas_unlock_irq(&xas);
1646 		if (!xas_nomem(&xas, GFP_KERNEL)) {
1647 			result = SCAN_FAIL;
1648 			goto out;
1649 		}
1650 	} while (1);
1651 
1652 	__SetPageLocked(new_page);
1653 	if (is_shmem)
1654 		__SetPageSwapBacked(new_page);
1655 	new_page->index = start;
1656 	new_page->mapping = mapping;
1657 
1658 	/*
1659 	 * At this point the new_page is locked and not up-to-date.
1660 	 * It's safe to insert it into the page cache, because nobody would
1661 	 * be able to map it or use it in another way until we unlock it.
1662 	 */
1663 
1664 	xas_set(&xas, start);
1665 	for (index = start; index < end; index++) {
1666 		struct page *page = xas_next(&xas);
1667 
1668 		VM_BUG_ON(index != xas.xa_index);
1669 		if (is_shmem) {
1670 			if (!page) {
1671 				/*
1672 				 * Stop if extent has been truncated or
1673 				 * hole-punched, and is now completely
1674 				 * empty.
1675 				 */
1676 				if (index == start) {
1677 					if (!xas_next_entry(&xas, end - 1)) {
1678 						result = SCAN_TRUNCATED;
1679 						goto xa_locked;
1680 					}
1681 					xas_set(&xas, index);
1682 				}
1683 				if (!shmem_charge(mapping->host, 1)) {
1684 					result = SCAN_FAIL;
1685 					goto xa_locked;
1686 				}
1687 				xas_store(&xas, new_page);
1688 				nr_none++;
1689 				continue;
1690 			}
1691 
1692 			if (xa_is_value(page) || !PageUptodate(page)) {
1693 				xas_unlock_irq(&xas);
1694 				/* swap in or instantiate fallocated page */
1695 				if (shmem_getpage(mapping->host, index, &page,
1696 						  SGP_NOHUGE)) {
1697 					result = SCAN_FAIL;
1698 					goto xa_unlocked;
1699 				}
1700 			} else if (trylock_page(page)) {
1701 				get_page(page);
1702 				xas_unlock_irq(&xas);
1703 			} else {
1704 				result = SCAN_PAGE_LOCK;
1705 				goto xa_locked;
1706 			}
1707 		} else {	/* !is_shmem */
1708 			if (!page || xa_is_value(page)) {
1709 				xas_unlock_irq(&xas);
1710 				page_cache_sync_readahead(mapping, &file->f_ra,
1711 							  file, index,
1712 							  PAGE_SIZE);
1713 				/* drain pagevecs to help isolate_lru_page() */
1714 				lru_add_drain();
1715 				page = find_lock_page(mapping, index);
1716 				if (unlikely(page == NULL)) {
1717 					result = SCAN_FAIL;
1718 					goto xa_unlocked;
1719 				}
1720 			} else if (PageDirty(page)) {
1721 				/*
1722 				 * khugepaged only works on read-only fd,
1723 				 * so this page is dirty because it hasn't
1724 				 * been flushed since first write. There
1725 				 * won't be new dirty pages.
1726 				 *
1727 				 * Trigger async flush here and hope the
1728 				 * writeback is done when khugepaged
1729 				 * revisits this page.
1730 				 *
1731 				 * This is a one-off situation. We are not
1732 				 * forcing writeback in loop.
1733 				 */
1734 				xas_unlock_irq(&xas);
1735 				filemap_flush(mapping);
1736 				result = SCAN_FAIL;
1737 				goto xa_unlocked;
1738 			} else if (trylock_page(page)) {
1739 				get_page(page);
1740 				xas_unlock_irq(&xas);
1741 			} else {
1742 				result = SCAN_PAGE_LOCK;
1743 				goto xa_locked;
1744 			}
1745 		}
1746 
1747 		/*
1748 		 * The page must be locked, so we can drop the i_pages lock
1749 		 * without racing with truncate.
1750 		 */
1751 		VM_BUG_ON_PAGE(!PageLocked(page), page);
1752 
1753 		/* make sure the page is up to date */
1754 		if (unlikely(!PageUptodate(page))) {
1755 			result = SCAN_FAIL;
1756 			goto out_unlock;
1757 		}
1758 
1759 		/*
1760 		 * If file was truncated then extended, or hole-punched, before
1761 		 * we locked the first page, then a THP might be there already.
1762 		 */
1763 		if (PageTransCompound(page)) {
1764 			result = SCAN_PAGE_COMPOUND;
1765 			goto out_unlock;
1766 		}
1767 
1768 		if (page_mapping(page) != mapping) {
1769 			result = SCAN_TRUNCATED;
1770 			goto out_unlock;
1771 		}
1772 
1773 		if (!is_shmem && PageDirty(page)) {
1774 			/*
1775 			 * khugepaged only works on read-only fd, so this
1776 			 * page is dirty because it hasn't been flushed
1777 			 * since first write.
1778 			 */
1779 			result = SCAN_FAIL;
1780 			goto out_unlock;
1781 		}
1782 
1783 		if (isolate_lru_page(page)) {
1784 			result = SCAN_DEL_PAGE_LRU;
1785 			goto out_unlock;
1786 		}
1787 
1788 		if (page_has_private(page) &&
1789 		    !try_to_release_page(page, GFP_KERNEL)) {
1790 			result = SCAN_PAGE_HAS_PRIVATE;
1791 			putback_lru_page(page);
1792 			goto out_unlock;
1793 		}
1794 
1795 		if (page_mapped(page))
1796 			unmap_mapping_pages(mapping, index, 1, false);
1797 
1798 		xas_lock_irq(&xas);
1799 		xas_set(&xas, index);
1800 
1801 		VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1802 		VM_BUG_ON_PAGE(page_mapped(page), page);
1803 
1804 		/*
1805 		 * The page is expected to have page_count() == 3:
1806 		 *  - we hold a pin on it;
1807 		 *  - one reference from page cache;
1808 		 *  - one from isolate_lru_page;
1809 		 */
1810 		if (!page_ref_freeze(page, 3)) {
1811 			result = SCAN_PAGE_COUNT;
1812 			xas_unlock_irq(&xas);
1813 			putback_lru_page(page);
1814 			goto out_unlock;
1815 		}
1816 
1817 		/*
1818 		 * Add the page to the list to be able to undo the collapse if
1819 		 * something go wrong.
1820 		 */
1821 		list_add_tail(&page->lru, &pagelist);
1822 
1823 		/* Finally, replace with the new page. */
1824 		xas_store(&xas, new_page);
1825 		continue;
1826 out_unlock:
1827 		unlock_page(page);
1828 		put_page(page);
1829 		goto xa_unlocked;
1830 	}
1831 
1832 	if (is_shmem)
1833 		__inc_node_page_state(new_page, NR_SHMEM_THPS);
1834 	else {
1835 		__inc_node_page_state(new_page, NR_FILE_THPS);
1836 		filemap_nr_thps_inc(mapping);
1837 	}
1838 
1839 	if (nr_none) {
1840 		__mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1841 		if (is_shmem)
1842 			__mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1843 	}
1844 
1845 xa_locked:
1846 	xas_unlock_irq(&xas);
1847 xa_unlocked:
1848 
1849 	if (result == SCAN_SUCCEED) {
1850 		struct page *page, *tmp;
1851 
1852 		/*
1853 		 * Replacing old pages with new one has succeeded, now we
1854 		 * need to copy the content and free the old pages.
1855 		 */
1856 		index = start;
1857 		list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1858 			while (index < page->index) {
1859 				clear_highpage(new_page + (index % HPAGE_PMD_NR));
1860 				index++;
1861 			}
1862 			copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1863 					page);
1864 			list_del(&page->lru);
1865 			page->mapping = NULL;
1866 			page_ref_unfreeze(page, 1);
1867 			ClearPageActive(page);
1868 			ClearPageUnevictable(page);
1869 			unlock_page(page);
1870 			put_page(page);
1871 			index++;
1872 		}
1873 		while (index < end) {
1874 			clear_highpage(new_page + (index % HPAGE_PMD_NR));
1875 			index++;
1876 		}
1877 
1878 		SetPageUptodate(new_page);
1879 		page_ref_add(new_page, HPAGE_PMD_NR - 1);
1880 		if (is_shmem)
1881 			set_page_dirty(new_page);
1882 		lru_cache_add(new_page);
1883 
1884 		/*
1885 		 * Remove pte page tables, so we can re-fault the page as huge.
1886 		 */
1887 		retract_page_tables(mapping, start);
1888 		*hpage = NULL;
1889 
1890 		khugepaged_pages_collapsed++;
1891 	} else {
1892 		struct page *page;
1893 
1894 		/* Something went wrong: roll back page cache changes */
1895 		xas_lock_irq(&xas);
1896 		mapping->nrpages -= nr_none;
1897 
1898 		if (is_shmem)
1899 			shmem_uncharge(mapping->host, nr_none);
1900 
1901 		xas_set(&xas, start);
1902 		xas_for_each(&xas, page, end - 1) {
1903 			page = list_first_entry_or_null(&pagelist,
1904 					struct page, lru);
1905 			if (!page || xas.xa_index < page->index) {
1906 				if (!nr_none)
1907 					break;
1908 				nr_none--;
1909 				/* Put holes back where they were */
1910 				xas_store(&xas, NULL);
1911 				continue;
1912 			}
1913 
1914 			VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1915 
1916 			/* Unfreeze the page. */
1917 			list_del(&page->lru);
1918 			page_ref_unfreeze(page, 2);
1919 			xas_store(&xas, page);
1920 			xas_pause(&xas);
1921 			xas_unlock_irq(&xas);
1922 			unlock_page(page);
1923 			putback_lru_page(page);
1924 			xas_lock_irq(&xas);
1925 		}
1926 		VM_BUG_ON(nr_none);
1927 		xas_unlock_irq(&xas);
1928 
1929 		new_page->mapping = NULL;
1930 	}
1931 
1932 	unlock_page(new_page);
1933 out:
1934 	VM_BUG_ON(!list_empty(&pagelist));
1935 	if (!IS_ERR_OR_NULL(*hpage))
1936 		mem_cgroup_uncharge(*hpage);
1937 	/* TODO: tracepoints */
1938 }
1939 
1940 static void khugepaged_scan_file(struct mm_struct *mm,
1941 		struct file *file, pgoff_t start, struct page **hpage)
1942 {
1943 	struct page *page = NULL;
1944 	struct address_space *mapping = file->f_mapping;
1945 	XA_STATE(xas, &mapping->i_pages, start);
1946 	int present, swap;
1947 	int node = NUMA_NO_NODE;
1948 	int result = SCAN_SUCCEED;
1949 
1950 	present = 0;
1951 	swap = 0;
1952 	memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1953 	rcu_read_lock();
1954 	xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1955 		if (xas_retry(&xas, page))
1956 			continue;
1957 
1958 		if (xa_is_value(page)) {
1959 			if (++swap > khugepaged_max_ptes_swap) {
1960 				result = SCAN_EXCEED_SWAP_PTE;
1961 				break;
1962 			}
1963 			continue;
1964 		}
1965 
1966 		if (PageTransCompound(page)) {
1967 			result = SCAN_PAGE_COMPOUND;
1968 			break;
1969 		}
1970 
1971 		node = page_to_nid(page);
1972 		if (khugepaged_scan_abort(node)) {
1973 			result = SCAN_SCAN_ABORT;
1974 			break;
1975 		}
1976 		khugepaged_node_load[node]++;
1977 
1978 		if (!PageLRU(page)) {
1979 			result = SCAN_PAGE_LRU;
1980 			break;
1981 		}
1982 
1983 		if (page_count(page) !=
1984 		    1 + page_mapcount(page) + page_has_private(page)) {
1985 			result = SCAN_PAGE_COUNT;
1986 			break;
1987 		}
1988 
1989 		/*
1990 		 * We probably should check if the page is referenced here, but
1991 		 * nobody would transfer pte_young() to PageReferenced() for us.
1992 		 * And rmap walk here is just too costly...
1993 		 */
1994 
1995 		present++;
1996 
1997 		if (need_resched()) {
1998 			xas_pause(&xas);
1999 			cond_resched_rcu();
2000 		}
2001 	}
2002 	rcu_read_unlock();
2003 
2004 	if (result == SCAN_SUCCEED) {
2005 		if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2006 			result = SCAN_EXCEED_NONE_PTE;
2007 		} else {
2008 			node = khugepaged_find_target_node();
2009 			collapse_file(mm, file, start, hpage, node);
2010 		}
2011 	}
2012 
2013 	/* TODO: tracepoints */
2014 }
2015 #else
2016 static void khugepaged_scan_file(struct mm_struct *mm,
2017 		struct file *file, pgoff_t start, struct page **hpage)
2018 {
2019 	BUILD_BUG();
2020 }
2021 
2022 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2023 {
2024 	return 0;
2025 }
2026 #endif
2027 
2028 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2029 					    struct page **hpage)
2030 	__releases(&khugepaged_mm_lock)
2031 	__acquires(&khugepaged_mm_lock)
2032 {
2033 	struct mm_slot *mm_slot;
2034 	struct mm_struct *mm;
2035 	struct vm_area_struct *vma;
2036 	int progress = 0;
2037 
2038 	VM_BUG_ON(!pages);
2039 	lockdep_assert_held(&khugepaged_mm_lock);
2040 
2041 	if (khugepaged_scan.mm_slot)
2042 		mm_slot = khugepaged_scan.mm_slot;
2043 	else {
2044 		mm_slot = list_entry(khugepaged_scan.mm_head.next,
2045 				     struct mm_slot, mm_node);
2046 		khugepaged_scan.address = 0;
2047 		khugepaged_scan.mm_slot = mm_slot;
2048 	}
2049 	spin_unlock(&khugepaged_mm_lock);
2050 	khugepaged_collapse_pte_mapped_thps(mm_slot);
2051 
2052 	mm = mm_slot->mm;
2053 	/*
2054 	 * Don't wait for semaphore (to avoid long wait times).  Just move to
2055 	 * the next mm on the list.
2056 	 */
2057 	vma = NULL;
2058 	if (unlikely(!mmap_read_trylock(mm)))
2059 		goto breakouterloop_mmap_lock;
2060 	if (likely(!khugepaged_test_exit(mm)))
2061 		vma = find_vma(mm, khugepaged_scan.address);
2062 
2063 	progress++;
2064 	for (; vma; vma = vma->vm_next) {
2065 		unsigned long hstart, hend;
2066 
2067 		cond_resched();
2068 		if (unlikely(khugepaged_test_exit(mm))) {
2069 			progress++;
2070 			break;
2071 		}
2072 		if (!hugepage_vma_check(vma, vma->vm_flags)) {
2073 skip:
2074 			progress++;
2075 			continue;
2076 		}
2077 		hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2078 		hend = vma->vm_end & HPAGE_PMD_MASK;
2079 		if (hstart >= hend)
2080 			goto skip;
2081 		if (khugepaged_scan.address > hend)
2082 			goto skip;
2083 		if (khugepaged_scan.address < hstart)
2084 			khugepaged_scan.address = hstart;
2085 		VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2086 		if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2087 			goto skip;
2088 
2089 		while (khugepaged_scan.address < hend) {
2090 			int ret;
2091 			cond_resched();
2092 			if (unlikely(khugepaged_test_exit(mm)))
2093 				goto breakouterloop;
2094 
2095 			VM_BUG_ON(khugepaged_scan.address < hstart ||
2096 				  khugepaged_scan.address + HPAGE_PMD_SIZE >
2097 				  hend);
2098 			if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2099 				struct file *file = get_file(vma->vm_file);
2100 				pgoff_t pgoff = linear_page_index(vma,
2101 						khugepaged_scan.address);
2102 
2103 				mmap_read_unlock(mm);
2104 				ret = 1;
2105 				khugepaged_scan_file(mm, file, pgoff, hpage);
2106 				fput(file);
2107 			} else {
2108 				ret = khugepaged_scan_pmd(mm, vma,
2109 						khugepaged_scan.address,
2110 						hpage);
2111 			}
2112 			/* move to next address */
2113 			khugepaged_scan.address += HPAGE_PMD_SIZE;
2114 			progress += HPAGE_PMD_NR;
2115 			if (ret)
2116 				/* we released mmap_lock so break loop */
2117 				goto breakouterloop_mmap_lock;
2118 			if (progress >= pages)
2119 				goto breakouterloop;
2120 		}
2121 	}
2122 breakouterloop:
2123 	mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2124 breakouterloop_mmap_lock:
2125 
2126 	spin_lock(&khugepaged_mm_lock);
2127 	VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2128 	/*
2129 	 * Release the current mm_slot if this mm is about to die, or
2130 	 * if we scanned all vmas of this mm.
2131 	 */
2132 	if (khugepaged_test_exit(mm) || !vma) {
2133 		/*
2134 		 * Make sure that if mm_users is reaching zero while
2135 		 * khugepaged runs here, khugepaged_exit will find
2136 		 * mm_slot not pointing to the exiting mm.
2137 		 */
2138 		if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2139 			khugepaged_scan.mm_slot = list_entry(
2140 				mm_slot->mm_node.next,
2141 				struct mm_slot, mm_node);
2142 			khugepaged_scan.address = 0;
2143 		} else {
2144 			khugepaged_scan.mm_slot = NULL;
2145 			khugepaged_full_scans++;
2146 		}
2147 
2148 		collect_mm_slot(mm_slot);
2149 	}
2150 
2151 	return progress;
2152 }
2153 
2154 static int khugepaged_has_work(void)
2155 {
2156 	return !list_empty(&khugepaged_scan.mm_head) &&
2157 		khugepaged_enabled();
2158 }
2159 
2160 static int khugepaged_wait_event(void)
2161 {
2162 	return !list_empty(&khugepaged_scan.mm_head) ||
2163 		kthread_should_stop();
2164 }
2165 
2166 static void khugepaged_do_scan(void)
2167 {
2168 	struct page *hpage = NULL;
2169 	unsigned int progress = 0, pass_through_head = 0;
2170 	unsigned int pages = khugepaged_pages_to_scan;
2171 	bool wait = true;
2172 
2173 	barrier(); /* write khugepaged_pages_to_scan to local stack */
2174 
2175 	lru_add_drain_all();
2176 
2177 	while (progress < pages) {
2178 		if (!khugepaged_prealloc_page(&hpage, &wait))
2179 			break;
2180 
2181 		cond_resched();
2182 
2183 		if (unlikely(kthread_should_stop() || try_to_freeze()))
2184 			break;
2185 
2186 		spin_lock(&khugepaged_mm_lock);
2187 		if (!khugepaged_scan.mm_slot)
2188 			pass_through_head++;
2189 		if (khugepaged_has_work() &&
2190 		    pass_through_head < 2)
2191 			progress += khugepaged_scan_mm_slot(pages - progress,
2192 							    &hpage);
2193 		else
2194 			progress = pages;
2195 		spin_unlock(&khugepaged_mm_lock);
2196 	}
2197 
2198 	if (!IS_ERR_OR_NULL(hpage))
2199 		put_page(hpage);
2200 }
2201 
2202 static bool khugepaged_should_wakeup(void)
2203 {
2204 	return kthread_should_stop() ||
2205 	       time_after_eq(jiffies, khugepaged_sleep_expire);
2206 }
2207 
2208 static void khugepaged_wait_work(void)
2209 {
2210 	if (khugepaged_has_work()) {
2211 		const unsigned long scan_sleep_jiffies =
2212 			msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2213 
2214 		if (!scan_sleep_jiffies)
2215 			return;
2216 
2217 		khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2218 		wait_event_freezable_timeout(khugepaged_wait,
2219 					     khugepaged_should_wakeup(),
2220 					     scan_sleep_jiffies);
2221 		return;
2222 	}
2223 
2224 	if (khugepaged_enabled())
2225 		wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2226 }
2227 
2228 static int khugepaged(void *none)
2229 {
2230 	struct mm_slot *mm_slot;
2231 
2232 	set_freezable();
2233 	set_user_nice(current, MAX_NICE);
2234 
2235 	while (!kthread_should_stop()) {
2236 		khugepaged_do_scan();
2237 		khugepaged_wait_work();
2238 	}
2239 
2240 	spin_lock(&khugepaged_mm_lock);
2241 	mm_slot = khugepaged_scan.mm_slot;
2242 	khugepaged_scan.mm_slot = NULL;
2243 	if (mm_slot)
2244 		collect_mm_slot(mm_slot);
2245 	spin_unlock(&khugepaged_mm_lock);
2246 	return 0;
2247 }
2248 
2249 static void set_recommended_min_free_kbytes(void)
2250 {
2251 	struct zone *zone;
2252 	int nr_zones = 0;
2253 	unsigned long recommended_min;
2254 
2255 	for_each_populated_zone(zone) {
2256 		/*
2257 		 * We don't need to worry about fragmentation of
2258 		 * ZONE_MOVABLE since it only has movable pages.
2259 		 */
2260 		if (zone_idx(zone) > gfp_zone(GFP_USER))
2261 			continue;
2262 
2263 		nr_zones++;
2264 	}
2265 
2266 	/* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2267 	recommended_min = pageblock_nr_pages * nr_zones * 2;
2268 
2269 	/*
2270 	 * Make sure that on average at least two pageblocks are almost free
2271 	 * of another type, one for a migratetype to fall back to and a
2272 	 * second to avoid subsequent fallbacks of other types There are 3
2273 	 * MIGRATE_TYPES we care about.
2274 	 */
2275 	recommended_min += pageblock_nr_pages * nr_zones *
2276 			   MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2277 
2278 	/* don't ever allow to reserve more than 5% of the lowmem */
2279 	recommended_min = min(recommended_min,
2280 			      (unsigned long) nr_free_buffer_pages() / 20);
2281 	recommended_min <<= (PAGE_SHIFT-10);
2282 
2283 	if (recommended_min > min_free_kbytes) {
2284 		if (user_min_free_kbytes >= 0)
2285 			pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2286 				min_free_kbytes, recommended_min);
2287 
2288 		min_free_kbytes = recommended_min;
2289 	}
2290 	setup_per_zone_wmarks();
2291 }
2292 
2293 int start_stop_khugepaged(void)
2294 {
2295 	static struct task_struct *khugepaged_thread __read_mostly;
2296 	static DEFINE_MUTEX(khugepaged_mutex);
2297 	int err = 0;
2298 
2299 	mutex_lock(&khugepaged_mutex);
2300 	if (khugepaged_enabled()) {
2301 		if (!khugepaged_thread)
2302 			khugepaged_thread = kthread_run(khugepaged, NULL,
2303 							"khugepaged");
2304 		if (IS_ERR(khugepaged_thread)) {
2305 			pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2306 			err = PTR_ERR(khugepaged_thread);
2307 			khugepaged_thread = NULL;
2308 			goto fail;
2309 		}
2310 
2311 		if (!list_empty(&khugepaged_scan.mm_head))
2312 			wake_up_interruptible(&khugepaged_wait);
2313 
2314 		set_recommended_min_free_kbytes();
2315 	} else if (khugepaged_thread) {
2316 		kthread_stop(khugepaged_thread);
2317 		khugepaged_thread = NULL;
2318 	}
2319 fail:
2320 	mutex_unlock(&khugepaged_mutex);
2321 	return err;
2322 }
2323