xref: /linux/mm/khugepaged.c (revision a48b0872e69428d3d02994dcfad3519f01def7fa)
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 static struct task_struct *khugepaged_thread __read_mostly;
60 static DEFINE_MUTEX(khugepaged_mutex);
61 
62 /* default scan 8*512 pte (or vmas) every 30 second */
63 static unsigned int khugepaged_pages_to_scan __read_mostly;
64 static unsigned int khugepaged_pages_collapsed;
65 static unsigned int khugepaged_full_scans;
66 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
67 /* during fragmentation poll the hugepage allocator once every minute */
68 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
69 static unsigned long khugepaged_sleep_expire;
70 static DEFINE_SPINLOCK(khugepaged_mm_lock);
71 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
72 /*
73  * default collapse hugepages if there is at least one pte mapped like
74  * it would have happened if the vma was large enough during page
75  * fault.
76  */
77 static unsigned int khugepaged_max_ptes_none __read_mostly;
78 static unsigned int khugepaged_max_ptes_swap __read_mostly;
79 static unsigned int khugepaged_max_ptes_shared __read_mostly;
80 
81 #define MM_SLOTS_HASH_BITS 10
82 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
83 
84 static struct kmem_cache *mm_slot_cache __read_mostly;
85 
86 #define MAX_PTE_MAPPED_THP 8
87 
88 /**
89  * struct mm_slot - hash lookup from mm to mm_slot
90  * @hash: hash collision list
91  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
92  * @mm: the mm that this information is valid for
93  * @nr_pte_mapped_thp: number of pte mapped THP
94  * @pte_mapped_thp: address array corresponding pte mapped THP
95  */
96 struct mm_slot {
97 	struct hlist_node hash;
98 	struct list_head mm_node;
99 	struct mm_struct *mm;
100 
101 	/* pte-mapped THP in this mm */
102 	int nr_pte_mapped_thp;
103 	unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
104 };
105 
106 /**
107  * struct khugepaged_scan - cursor for scanning
108  * @mm_head: the head of the mm list to scan
109  * @mm_slot: the current mm_slot we are scanning
110  * @address: the next address inside that to be scanned
111  *
112  * There is only the one khugepaged_scan instance of this cursor structure.
113  */
114 struct khugepaged_scan {
115 	struct list_head mm_head;
116 	struct mm_slot *mm_slot;
117 	unsigned long address;
118 };
119 
120 static struct khugepaged_scan khugepaged_scan = {
121 	.mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
122 };
123 
124 #ifdef CONFIG_SYSFS
125 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
126 					 struct kobj_attribute *attr,
127 					 char *buf)
128 {
129 	return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
130 }
131 
132 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
133 					  struct kobj_attribute *attr,
134 					  const char *buf, size_t count)
135 {
136 	unsigned int msecs;
137 	int err;
138 
139 	err = kstrtouint(buf, 10, &msecs);
140 	if (err)
141 		return -EINVAL;
142 
143 	khugepaged_scan_sleep_millisecs = msecs;
144 	khugepaged_sleep_expire = 0;
145 	wake_up_interruptible(&khugepaged_wait);
146 
147 	return count;
148 }
149 static struct kobj_attribute scan_sleep_millisecs_attr =
150 	__ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
151 	       scan_sleep_millisecs_store);
152 
153 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
154 					  struct kobj_attribute *attr,
155 					  char *buf)
156 {
157 	return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
158 }
159 
160 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
161 					   struct kobj_attribute *attr,
162 					   const char *buf, size_t count)
163 {
164 	unsigned int msecs;
165 	int err;
166 
167 	err = kstrtouint(buf, 10, &msecs);
168 	if (err)
169 		return -EINVAL;
170 
171 	khugepaged_alloc_sleep_millisecs = msecs;
172 	khugepaged_sleep_expire = 0;
173 	wake_up_interruptible(&khugepaged_wait);
174 
175 	return count;
176 }
177 static struct kobj_attribute alloc_sleep_millisecs_attr =
178 	__ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
179 	       alloc_sleep_millisecs_store);
180 
181 static ssize_t pages_to_scan_show(struct kobject *kobj,
182 				  struct kobj_attribute *attr,
183 				  char *buf)
184 {
185 	return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
186 }
187 static ssize_t pages_to_scan_store(struct kobject *kobj,
188 				   struct kobj_attribute *attr,
189 				   const char *buf, size_t count)
190 {
191 	unsigned int pages;
192 	int err;
193 
194 	err = kstrtouint(buf, 10, &pages);
195 	if (err || !pages)
196 		return -EINVAL;
197 
198 	khugepaged_pages_to_scan = pages;
199 
200 	return count;
201 }
202 static struct kobj_attribute pages_to_scan_attr =
203 	__ATTR(pages_to_scan, 0644, pages_to_scan_show,
204 	       pages_to_scan_store);
205 
206 static ssize_t pages_collapsed_show(struct kobject *kobj,
207 				    struct kobj_attribute *attr,
208 				    char *buf)
209 {
210 	return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
211 }
212 static struct kobj_attribute pages_collapsed_attr =
213 	__ATTR_RO(pages_collapsed);
214 
215 static ssize_t full_scans_show(struct kobject *kobj,
216 			       struct kobj_attribute *attr,
217 			       char *buf)
218 {
219 	return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
220 }
221 static struct kobj_attribute full_scans_attr =
222 	__ATTR_RO(full_scans);
223 
224 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
225 				      struct kobj_attribute *attr, char *buf)
226 {
227 	return single_hugepage_flag_show(kobj, attr, buf,
228 					 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
229 }
230 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
231 				       struct kobj_attribute *attr,
232 				       const char *buf, size_t count)
233 {
234 	return single_hugepage_flag_store(kobj, attr, buf, count,
235 				 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
236 }
237 static struct kobj_attribute khugepaged_defrag_attr =
238 	__ATTR(defrag, 0644, khugepaged_defrag_show,
239 	       khugepaged_defrag_store);
240 
241 /*
242  * max_ptes_none controls if khugepaged should collapse hugepages over
243  * any unmapped ptes in turn potentially increasing the memory
244  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
245  * reduce the available free memory in the system as it
246  * runs. Increasing max_ptes_none will instead potentially reduce the
247  * free memory in the system during the khugepaged scan.
248  */
249 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
250 					     struct kobj_attribute *attr,
251 					     char *buf)
252 {
253 	return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
254 }
255 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
256 					      struct kobj_attribute *attr,
257 					      const char *buf, size_t count)
258 {
259 	int err;
260 	unsigned long max_ptes_none;
261 
262 	err = kstrtoul(buf, 10, &max_ptes_none);
263 	if (err || max_ptes_none > HPAGE_PMD_NR-1)
264 		return -EINVAL;
265 
266 	khugepaged_max_ptes_none = max_ptes_none;
267 
268 	return count;
269 }
270 static struct kobj_attribute khugepaged_max_ptes_none_attr =
271 	__ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
272 	       khugepaged_max_ptes_none_store);
273 
274 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
275 					     struct kobj_attribute *attr,
276 					     char *buf)
277 {
278 	return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
279 }
280 
281 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
282 					      struct kobj_attribute *attr,
283 					      const char *buf, size_t count)
284 {
285 	int err;
286 	unsigned long max_ptes_swap;
287 
288 	err  = kstrtoul(buf, 10, &max_ptes_swap);
289 	if (err || max_ptes_swap > HPAGE_PMD_NR-1)
290 		return -EINVAL;
291 
292 	khugepaged_max_ptes_swap = max_ptes_swap;
293 
294 	return count;
295 }
296 
297 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
298 	__ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
299 	       khugepaged_max_ptes_swap_store);
300 
301 static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
302 					       struct kobj_attribute *attr,
303 					       char *buf)
304 {
305 	return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
306 }
307 
308 static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
309 					      struct kobj_attribute *attr,
310 					      const char *buf, size_t count)
311 {
312 	int err;
313 	unsigned long max_ptes_shared;
314 
315 	err  = kstrtoul(buf, 10, &max_ptes_shared);
316 	if (err || max_ptes_shared > HPAGE_PMD_NR-1)
317 		return -EINVAL;
318 
319 	khugepaged_max_ptes_shared = max_ptes_shared;
320 
321 	return count;
322 }
323 
324 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
325 	__ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
326 	       khugepaged_max_ptes_shared_store);
327 
328 static struct attribute *khugepaged_attr[] = {
329 	&khugepaged_defrag_attr.attr,
330 	&khugepaged_max_ptes_none_attr.attr,
331 	&khugepaged_max_ptes_swap_attr.attr,
332 	&khugepaged_max_ptes_shared_attr.attr,
333 	&pages_to_scan_attr.attr,
334 	&pages_collapsed_attr.attr,
335 	&full_scans_attr.attr,
336 	&scan_sleep_millisecs_attr.attr,
337 	&alloc_sleep_millisecs_attr.attr,
338 	NULL,
339 };
340 
341 struct attribute_group khugepaged_attr_group = {
342 	.attrs = khugepaged_attr,
343 	.name = "khugepaged",
344 };
345 #endif /* CONFIG_SYSFS */
346 
347 int hugepage_madvise(struct vm_area_struct *vma,
348 		     unsigned long *vm_flags, int advice)
349 {
350 	switch (advice) {
351 	case MADV_HUGEPAGE:
352 #ifdef CONFIG_S390
353 		/*
354 		 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
355 		 * can't handle this properly after s390_enable_sie, so we simply
356 		 * ignore the madvise to prevent qemu from causing a SIGSEGV.
357 		 */
358 		if (mm_has_pgste(vma->vm_mm))
359 			return 0;
360 #endif
361 		*vm_flags &= ~VM_NOHUGEPAGE;
362 		*vm_flags |= VM_HUGEPAGE;
363 		/*
364 		 * If the vma become good for khugepaged to scan,
365 		 * register it here without waiting a page fault that
366 		 * may not happen any time soon.
367 		 */
368 		if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
369 				khugepaged_enter_vma_merge(vma, *vm_flags))
370 			return -ENOMEM;
371 		break;
372 	case MADV_NOHUGEPAGE:
373 		*vm_flags &= ~VM_HUGEPAGE;
374 		*vm_flags |= VM_NOHUGEPAGE;
375 		/*
376 		 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
377 		 * this vma even if we leave the mm registered in khugepaged if
378 		 * it got registered before VM_NOHUGEPAGE was set.
379 		 */
380 		break;
381 	}
382 
383 	return 0;
384 }
385 
386 int __init khugepaged_init(void)
387 {
388 	mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
389 					  sizeof(struct mm_slot),
390 					  __alignof__(struct mm_slot), 0, NULL);
391 	if (!mm_slot_cache)
392 		return -ENOMEM;
393 
394 	khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
395 	khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
396 	khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
397 	khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
398 
399 	return 0;
400 }
401 
402 void __init khugepaged_destroy(void)
403 {
404 	kmem_cache_destroy(mm_slot_cache);
405 }
406 
407 static inline struct mm_slot *alloc_mm_slot(void)
408 {
409 	if (!mm_slot_cache)	/* initialization failed */
410 		return NULL;
411 	return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
412 }
413 
414 static inline void free_mm_slot(struct mm_slot *mm_slot)
415 {
416 	kmem_cache_free(mm_slot_cache, mm_slot);
417 }
418 
419 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
420 {
421 	struct mm_slot *mm_slot;
422 
423 	hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
424 		if (mm == mm_slot->mm)
425 			return mm_slot;
426 
427 	return NULL;
428 }
429 
430 static void insert_to_mm_slots_hash(struct mm_struct *mm,
431 				    struct mm_slot *mm_slot)
432 {
433 	mm_slot->mm = mm;
434 	hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
435 }
436 
437 static inline int khugepaged_test_exit(struct mm_struct *mm)
438 {
439 	return atomic_read(&mm->mm_users) == 0;
440 }
441 
442 static bool hugepage_vma_check(struct vm_area_struct *vma,
443 			       unsigned long vm_flags)
444 {
445 	/* Explicitly disabled through madvise. */
446 	if ((vm_flags & VM_NOHUGEPAGE) ||
447 	    test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
448 		return false;
449 
450 	/* Enabled via shmem mount options or sysfs settings. */
451 	if (shmem_file(vma->vm_file) && shmem_huge_enabled(vma)) {
452 		return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
453 				HPAGE_PMD_NR);
454 	}
455 
456 	/* THP settings require madvise. */
457 	if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
458 		return false;
459 
460 	/* Read-only file mappings need to be aligned for THP to work. */
461 	if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
462 	    (vm_flags & VM_DENYWRITE)) {
463 		return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
464 				HPAGE_PMD_NR);
465 	}
466 
467 	if (!vma->anon_vma || vma->vm_ops)
468 		return false;
469 	if (vma_is_temporary_stack(vma))
470 		return false;
471 	return !(vm_flags & VM_NO_KHUGEPAGED);
472 }
473 
474 int __khugepaged_enter(struct mm_struct *mm)
475 {
476 	struct mm_slot *mm_slot;
477 	int wakeup;
478 
479 	mm_slot = alloc_mm_slot();
480 	if (!mm_slot)
481 		return -ENOMEM;
482 
483 	/* __khugepaged_exit() must not run from under us */
484 	VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
485 	if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
486 		free_mm_slot(mm_slot);
487 		return 0;
488 	}
489 
490 	spin_lock(&khugepaged_mm_lock);
491 	insert_to_mm_slots_hash(mm, mm_slot);
492 	/*
493 	 * Insert just behind the scanning cursor, to let the area settle
494 	 * down a little.
495 	 */
496 	wakeup = list_empty(&khugepaged_scan.mm_head);
497 	list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
498 	spin_unlock(&khugepaged_mm_lock);
499 
500 	mmgrab(mm);
501 	if (wakeup)
502 		wake_up_interruptible(&khugepaged_wait);
503 
504 	return 0;
505 }
506 
507 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
508 			       unsigned long vm_flags)
509 {
510 	unsigned long hstart, hend;
511 
512 	/*
513 	 * khugepaged only supports read-only files for non-shmem files.
514 	 * khugepaged does not yet work on special mappings. And
515 	 * file-private shmem THP is not supported.
516 	 */
517 	if (!hugepage_vma_check(vma, vm_flags))
518 		return 0;
519 
520 	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
521 	hend = vma->vm_end & HPAGE_PMD_MASK;
522 	if (hstart < hend)
523 		return khugepaged_enter(vma, vm_flags);
524 	return 0;
525 }
526 
527 void __khugepaged_exit(struct mm_struct *mm)
528 {
529 	struct mm_slot *mm_slot;
530 	int free = 0;
531 
532 	spin_lock(&khugepaged_mm_lock);
533 	mm_slot = get_mm_slot(mm);
534 	if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
535 		hash_del(&mm_slot->hash);
536 		list_del(&mm_slot->mm_node);
537 		free = 1;
538 	}
539 	spin_unlock(&khugepaged_mm_lock);
540 
541 	if (free) {
542 		clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
543 		free_mm_slot(mm_slot);
544 		mmdrop(mm);
545 	} else if (mm_slot) {
546 		/*
547 		 * This is required to serialize against
548 		 * khugepaged_test_exit() (which is guaranteed to run
549 		 * under mmap sem read mode). Stop here (after we
550 		 * return all pagetables will be destroyed) until
551 		 * khugepaged has finished working on the pagetables
552 		 * under the mmap_lock.
553 		 */
554 		mmap_write_lock(mm);
555 		mmap_write_unlock(mm);
556 	}
557 }
558 
559 static void release_pte_page(struct page *page)
560 {
561 	mod_node_page_state(page_pgdat(page),
562 			NR_ISOLATED_ANON + page_is_file_lru(page),
563 			-compound_nr(page));
564 	unlock_page(page);
565 	putback_lru_page(page);
566 }
567 
568 static void release_pte_pages(pte_t *pte, pte_t *_pte,
569 		struct list_head *compound_pagelist)
570 {
571 	struct page *page, *tmp;
572 
573 	while (--_pte >= pte) {
574 		pte_t pteval = *_pte;
575 
576 		page = pte_page(pteval);
577 		if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
578 				!PageCompound(page))
579 			release_pte_page(page);
580 	}
581 
582 	list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
583 		list_del(&page->lru);
584 		release_pte_page(page);
585 	}
586 }
587 
588 static bool is_refcount_suitable(struct page *page)
589 {
590 	int expected_refcount;
591 
592 	expected_refcount = total_mapcount(page);
593 	if (PageSwapCache(page))
594 		expected_refcount += compound_nr(page);
595 
596 	return page_count(page) == expected_refcount;
597 }
598 
599 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
600 					unsigned long address,
601 					pte_t *pte,
602 					struct list_head *compound_pagelist)
603 {
604 	struct page *page = NULL;
605 	pte_t *_pte;
606 	int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
607 	bool writable = false;
608 
609 	for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
610 	     _pte++, address += PAGE_SIZE) {
611 		pte_t pteval = *_pte;
612 		if (pte_none(pteval) || (pte_present(pteval) &&
613 				is_zero_pfn(pte_pfn(pteval)))) {
614 			if (!userfaultfd_armed(vma) &&
615 			    ++none_or_zero <= khugepaged_max_ptes_none) {
616 				continue;
617 			} else {
618 				result = SCAN_EXCEED_NONE_PTE;
619 				goto out;
620 			}
621 		}
622 		if (!pte_present(pteval)) {
623 			result = SCAN_PTE_NON_PRESENT;
624 			goto out;
625 		}
626 		page = vm_normal_page(vma, address, pteval);
627 		if (unlikely(!page)) {
628 			result = SCAN_PAGE_NULL;
629 			goto out;
630 		}
631 
632 		VM_BUG_ON_PAGE(!PageAnon(page), page);
633 
634 		if (page_mapcount(page) > 1 &&
635 				++shared > khugepaged_max_ptes_shared) {
636 			result = SCAN_EXCEED_SHARED_PTE;
637 			goto out;
638 		}
639 
640 		if (PageCompound(page)) {
641 			struct page *p;
642 			page = compound_head(page);
643 
644 			/*
645 			 * Check if we have dealt with the compound page
646 			 * already
647 			 */
648 			list_for_each_entry(p, compound_pagelist, lru) {
649 				if (page == p)
650 					goto next;
651 			}
652 		}
653 
654 		/*
655 		 * We can do it before isolate_lru_page because the
656 		 * page can't be freed from under us. NOTE: PG_lock
657 		 * is needed to serialize against split_huge_page
658 		 * when invoked from the VM.
659 		 */
660 		if (!trylock_page(page)) {
661 			result = SCAN_PAGE_LOCK;
662 			goto out;
663 		}
664 
665 		/*
666 		 * Check if the page has any GUP (or other external) pins.
667 		 *
668 		 * The page table that maps the page has been already unlinked
669 		 * from the page table tree and this process cannot get
670 		 * an additional pin on the page.
671 		 *
672 		 * New pins can come later if the page is shared across fork,
673 		 * but not from this process. The other process cannot write to
674 		 * the page, only trigger CoW.
675 		 */
676 		if (!is_refcount_suitable(page)) {
677 			unlock_page(page);
678 			result = SCAN_PAGE_COUNT;
679 			goto out;
680 		}
681 		if (!pte_write(pteval) && PageSwapCache(page) &&
682 				!reuse_swap_page(page, NULL)) {
683 			/*
684 			 * Page is in the swap cache and cannot be re-used.
685 			 * It cannot be collapsed into a THP.
686 			 */
687 			unlock_page(page);
688 			result = SCAN_SWAP_CACHE_PAGE;
689 			goto out;
690 		}
691 
692 		/*
693 		 * Isolate the page to avoid collapsing an hugepage
694 		 * currently in use by the VM.
695 		 */
696 		if (isolate_lru_page(page)) {
697 			unlock_page(page);
698 			result = SCAN_DEL_PAGE_LRU;
699 			goto out;
700 		}
701 		mod_node_page_state(page_pgdat(page),
702 				NR_ISOLATED_ANON + page_is_file_lru(page),
703 				compound_nr(page));
704 		VM_BUG_ON_PAGE(!PageLocked(page), page);
705 		VM_BUG_ON_PAGE(PageLRU(page), page);
706 
707 		if (PageCompound(page))
708 			list_add_tail(&page->lru, compound_pagelist);
709 next:
710 		/* There should be enough young pte to collapse the page */
711 		if (pte_young(pteval) ||
712 		    page_is_young(page) || PageReferenced(page) ||
713 		    mmu_notifier_test_young(vma->vm_mm, address))
714 			referenced++;
715 
716 		if (pte_write(pteval))
717 			writable = true;
718 	}
719 
720 	if (unlikely(!writable)) {
721 		result = SCAN_PAGE_RO;
722 	} else if (unlikely(!referenced)) {
723 		result = SCAN_LACK_REFERENCED_PAGE;
724 	} else {
725 		result = SCAN_SUCCEED;
726 		trace_mm_collapse_huge_page_isolate(page, none_or_zero,
727 						    referenced, writable, result);
728 		return 1;
729 	}
730 out:
731 	release_pte_pages(pte, _pte, compound_pagelist);
732 	trace_mm_collapse_huge_page_isolate(page, none_or_zero,
733 					    referenced, writable, result);
734 	return 0;
735 }
736 
737 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
738 				      struct vm_area_struct *vma,
739 				      unsigned long address,
740 				      spinlock_t *ptl,
741 				      struct list_head *compound_pagelist)
742 {
743 	struct page *src_page, *tmp;
744 	pte_t *_pte;
745 	for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
746 				_pte++, page++, address += PAGE_SIZE) {
747 		pte_t pteval = *_pte;
748 
749 		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
750 			clear_user_highpage(page, address);
751 			add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
752 			if (is_zero_pfn(pte_pfn(pteval))) {
753 				/*
754 				 * ptl mostly unnecessary.
755 				 */
756 				spin_lock(ptl);
757 				/*
758 				 * paravirt calls inside pte_clear here are
759 				 * superfluous.
760 				 */
761 				pte_clear(vma->vm_mm, address, _pte);
762 				spin_unlock(ptl);
763 			}
764 		} else {
765 			src_page = pte_page(pteval);
766 			copy_user_highpage(page, src_page, address, vma);
767 			if (!PageCompound(src_page))
768 				release_pte_page(src_page);
769 			/*
770 			 * ptl mostly unnecessary, but preempt has to
771 			 * be disabled to update the per-cpu stats
772 			 * inside page_remove_rmap().
773 			 */
774 			spin_lock(ptl);
775 			/*
776 			 * paravirt calls inside pte_clear here are
777 			 * superfluous.
778 			 */
779 			pte_clear(vma->vm_mm, address, _pte);
780 			page_remove_rmap(src_page, false);
781 			spin_unlock(ptl);
782 			free_page_and_swap_cache(src_page);
783 		}
784 	}
785 
786 	list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
787 		list_del(&src_page->lru);
788 		release_pte_page(src_page);
789 	}
790 }
791 
792 static void khugepaged_alloc_sleep(void)
793 {
794 	DEFINE_WAIT(wait);
795 
796 	add_wait_queue(&khugepaged_wait, &wait);
797 	freezable_schedule_timeout_interruptible(
798 		msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
799 	remove_wait_queue(&khugepaged_wait, &wait);
800 }
801 
802 static int khugepaged_node_load[MAX_NUMNODES];
803 
804 static bool khugepaged_scan_abort(int nid)
805 {
806 	int i;
807 
808 	/*
809 	 * If node_reclaim_mode is disabled, then no extra effort is made to
810 	 * allocate memory locally.
811 	 */
812 	if (!node_reclaim_enabled())
813 		return false;
814 
815 	/* If there is a count for this node already, it must be acceptable */
816 	if (khugepaged_node_load[nid])
817 		return false;
818 
819 	for (i = 0; i < MAX_NUMNODES; i++) {
820 		if (!khugepaged_node_load[i])
821 			continue;
822 		if (node_distance(nid, i) > node_reclaim_distance)
823 			return true;
824 	}
825 	return false;
826 }
827 
828 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
829 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
830 {
831 	return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
832 }
833 
834 #ifdef CONFIG_NUMA
835 static int khugepaged_find_target_node(void)
836 {
837 	static int last_khugepaged_target_node = NUMA_NO_NODE;
838 	int nid, target_node = 0, max_value = 0;
839 
840 	/* find first node with max normal pages hit */
841 	for (nid = 0; nid < MAX_NUMNODES; nid++)
842 		if (khugepaged_node_load[nid] > max_value) {
843 			max_value = khugepaged_node_load[nid];
844 			target_node = nid;
845 		}
846 
847 	/* do some balance if several nodes have the same hit record */
848 	if (target_node <= last_khugepaged_target_node)
849 		for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
850 				nid++)
851 			if (max_value == khugepaged_node_load[nid]) {
852 				target_node = nid;
853 				break;
854 			}
855 
856 	last_khugepaged_target_node = target_node;
857 	return target_node;
858 }
859 
860 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
861 {
862 	if (IS_ERR(*hpage)) {
863 		if (!*wait)
864 			return false;
865 
866 		*wait = false;
867 		*hpage = NULL;
868 		khugepaged_alloc_sleep();
869 	} else if (*hpage) {
870 		put_page(*hpage);
871 		*hpage = NULL;
872 	}
873 
874 	return true;
875 }
876 
877 static struct page *
878 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
879 {
880 	VM_BUG_ON_PAGE(*hpage, *hpage);
881 
882 	*hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
883 	if (unlikely(!*hpage)) {
884 		count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
885 		*hpage = ERR_PTR(-ENOMEM);
886 		return NULL;
887 	}
888 
889 	prep_transhuge_page(*hpage);
890 	count_vm_event(THP_COLLAPSE_ALLOC);
891 	return *hpage;
892 }
893 #else
894 static int khugepaged_find_target_node(void)
895 {
896 	return 0;
897 }
898 
899 static inline struct page *alloc_khugepaged_hugepage(void)
900 {
901 	struct page *page;
902 
903 	page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
904 			   HPAGE_PMD_ORDER);
905 	if (page)
906 		prep_transhuge_page(page);
907 	return page;
908 }
909 
910 static struct page *khugepaged_alloc_hugepage(bool *wait)
911 {
912 	struct page *hpage;
913 
914 	do {
915 		hpage = alloc_khugepaged_hugepage();
916 		if (!hpage) {
917 			count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
918 			if (!*wait)
919 				return NULL;
920 
921 			*wait = false;
922 			khugepaged_alloc_sleep();
923 		} else
924 			count_vm_event(THP_COLLAPSE_ALLOC);
925 	} while (unlikely(!hpage) && likely(khugepaged_enabled()));
926 
927 	return hpage;
928 }
929 
930 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
931 {
932 	/*
933 	 * If the hpage allocated earlier was briefly exposed in page cache
934 	 * before collapse_file() failed, it is possible that racing lookups
935 	 * have not yet completed, and would then be unpleasantly surprised by
936 	 * finding the hpage reused for the same mapping at a different offset.
937 	 * Just release the previous allocation if there is any danger of that.
938 	 */
939 	if (*hpage && page_count(*hpage) > 1) {
940 		put_page(*hpage);
941 		*hpage = NULL;
942 	}
943 
944 	if (!*hpage)
945 		*hpage = khugepaged_alloc_hugepage(wait);
946 
947 	if (unlikely(!*hpage))
948 		return false;
949 
950 	return true;
951 }
952 
953 static struct page *
954 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
955 {
956 	VM_BUG_ON(!*hpage);
957 
958 	return  *hpage;
959 }
960 #endif
961 
962 /*
963  * If mmap_lock temporarily dropped, revalidate vma
964  * before taking mmap_lock.
965  * Return 0 if succeeds, otherwise return none-zero
966  * value (scan code).
967  */
968 
969 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
970 		struct vm_area_struct **vmap)
971 {
972 	struct vm_area_struct *vma;
973 	unsigned long hstart, hend;
974 
975 	if (unlikely(khugepaged_test_exit(mm)))
976 		return SCAN_ANY_PROCESS;
977 
978 	*vmap = vma = find_vma(mm, address);
979 	if (!vma)
980 		return SCAN_VMA_NULL;
981 
982 	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
983 	hend = vma->vm_end & HPAGE_PMD_MASK;
984 	if (address < hstart || address + HPAGE_PMD_SIZE > hend)
985 		return SCAN_ADDRESS_RANGE;
986 	if (!hugepage_vma_check(vma, vma->vm_flags))
987 		return SCAN_VMA_CHECK;
988 	/* Anon VMA expected */
989 	if (!vma->anon_vma || vma->vm_ops)
990 		return SCAN_VMA_CHECK;
991 	return 0;
992 }
993 
994 /*
995  * Bring missing pages in from swap, to complete THP collapse.
996  * Only done if khugepaged_scan_pmd believes it is worthwhile.
997  *
998  * Called and returns without pte mapped or spinlocks held,
999  * but with mmap_lock held to protect against vma changes.
1000  */
1001 
1002 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
1003 					struct vm_area_struct *vma,
1004 					unsigned long haddr, pmd_t *pmd,
1005 					int referenced)
1006 {
1007 	int swapped_in = 0;
1008 	vm_fault_t ret = 0;
1009 	unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
1010 
1011 	for (address = haddr; address < end; address += PAGE_SIZE) {
1012 		struct vm_fault vmf = {
1013 			.vma = vma,
1014 			.address = address,
1015 			.pgoff = linear_page_index(vma, haddr),
1016 			.flags = FAULT_FLAG_ALLOW_RETRY,
1017 			.pmd = pmd,
1018 		};
1019 
1020 		vmf.pte = pte_offset_map(pmd, address);
1021 		vmf.orig_pte = *vmf.pte;
1022 		if (!is_swap_pte(vmf.orig_pte)) {
1023 			pte_unmap(vmf.pte);
1024 			continue;
1025 		}
1026 		swapped_in++;
1027 		ret = do_swap_page(&vmf);
1028 
1029 		/* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1030 		if (ret & VM_FAULT_RETRY) {
1031 			mmap_read_lock(mm);
1032 			if (hugepage_vma_revalidate(mm, haddr, &vma)) {
1033 				/* vma is no longer available, don't continue to swapin */
1034 				trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1035 				return false;
1036 			}
1037 			/* check if the pmd is still valid */
1038 			if (mm_find_pmd(mm, haddr) != pmd) {
1039 				trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1040 				return false;
1041 			}
1042 		}
1043 		if (ret & VM_FAULT_ERROR) {
1044 			trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1045 			return false;
1046 		}
1047 	}
1048 
1049 	/* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1050 	if (swapped_in)
1051 		lru_add_drain();
1052 
1053 	trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1054 	return true;
1055 }
1056 
1057 static void collapse_huge_page(struct mm_struct *mm,
1058 				   unsigned long address,
1059 				   struct page **hpage,
1060 				   int node, int referenced, int unmapped)
1061 {
1062 	LIST_HEAD(compound_pagelist);
1063 	pmd_t *pmd, _pmd;
1064 	pte_t *pte;
1065 	pgtable_t pgtable;
1066 	struct page *new_page;
1067 	spinlock_t *pmd_ptl, *pte_ptl;
1068 	int isolated = 0, result = 0;
1069 	struct vm_area_struct *vma;
1070 	struct mmu_notifier_range range;
1071 	gfp_t gfp;
1072 
1073 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1074 
1075 	/* Only allocate from the target node */
1076 	gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1077 
1078 	/*
1079 	 * Before allocating the hugepage, release the mmap_lock read lock.
1080 	 * The allocation can take potentially a long time if it involves
1081 	 * sync compaction, and we do not need to hold the mmap_lock during
1082 	 * that. We will recheck the vma after taking it again in write mode.
1083 	 */
1084 	mmap_read_unlock(mm);
1085 	new_page = khugepaged_alloc_page(hpage, gfp, node);
1086 	if (!new_page) {
1087 		result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1088 		goto out_nolock;
1089 	}
1090 
1091 	if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1092 		result = SCAN_CGROUP_CHARGE_FAIL;
1093 		goto out_nolock;
1094 	}
1095 	count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1096 
1097 	mmap_read_lock(mm);
1098 	result = hugepage_vma_revalidate(mm, address, &vma);
1099 	if (result) {
1100 		mmap_read_unlock(mm);
1101 		goto out_nolock;
1102 	}
1103 
1104 	pmd = mm_find_pmd(mm, address);
1105 	if (!pmd) {
1106 		result = SCAN_PMD_NULL;
1107 		mmap_read_unlock(mm);
1108 		goto out_nolock;
1109 	}
1110 
1111 	/*
1112 	 * __collapse_huge_page_swapin always returns with mmap_lock locked.
1113 	 * If it fails, we release mmap_lock and jump out_nolock.
1114 	 * Continuing to collapse causes inconsistency.
1115 	 */
1116 	if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1117 						     pmd, referenced)) {
1118 		mmap_read_unlock(mm);
1119 		goto out_nolock;
1120 	}
1121 
1122 	mmap_read_unlock(mm);
1123 	/*
1124 	 * Prevent all access to pagetables with the exception of
1125 	 * gup_fast later handled by the ptep_clear_flush and the VM
1126 	 * handled by the anon_vma lock + PG_lock.
1127 	 */
1128 	mmap_write_lock(mm);
1129 	result = hugepage_vma_revalidate(mm, address, &vma);
1130 	if (result)
1131 		goto out_up_write;
1132 	/* check if the pmd is still valid */
1133 	if (mm_find_pmd(mm, address) != pmd)
1134 		goto out_up_write;
1135 
1136 	anon_vma_lock_write(vma->anon_vma);
1137 
1138 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1139 				address, address + HPAGE_PMD_SIZE);
1140 	mmu_notifier_invalidate_range_start(&range);
1141 
1142 	pte = pte_offset_map(pmd, address);
1143 	pte_ptl = pte_lockptr(mm, pmd);
1144 
1145 	pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1146 	/*
1147 	 * After this gup_fast can't run anymore. This also removes
1148 	 * any huge TLB entry from the CPU so we won't allow
1149 	 * huge and small TLB entries for the same virtual address
1150 	 * to avoid the risk of CPU bugs in that area.
1151 	 */
1152 	_pmd = pmdp_collapse_flush(vma, address, pmd);
1153 	spin_unlock(pmd_ptl);
1154 	mmu_notifier_invalidate_range_end(&range);
1155 
1156 	spin_lock(pte_ptl);
1157 	isolated = __collapse_huge_page_isolate(vma, address, pte,
1158 			&compound_pagelist);
1159 	spin_unlock(pte_ptl);
1160 
1161 	if (unlikely(!isolated)) {
1162 		pte_unmap(pte);
1163 		spin_lock(pmd_ptl);
1164 		BUG_ON(!pmd_none(*pmd));
1165 		/*
1166 		 * We can only use set_pmd_at when establishing
1167 		 * hugepmds and never for establishing regular pmds that
1168 		 * points to regular pagetables. Use pmd_populate for that
1169 		 */
1170 		pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1171 		spin_unlock(pmd_ptl);
1172 		anon_vma_unlock_write(vma->anon_vma);
1173 		result = SCAN_FAIL;
1174 		goto out_up_write;
1175 	}
1176 
1177 	/*
1178 	 * All pages are isolated and locked so anon_vma rmap
1179 	 * can't run anymore.
1180 	 */
1181 	anon_vma_unlock_write(vma->anon_vma);
1182 
1183 	__collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1184 			&compound_pagelist);
1185 	pte_unmap(pte);
1186 	/*
1187 	 * spin_lock() below is not the equivalent of smp_wmb(), but
1188 	 * the smp_wmb() inside __SetPageUptodate() can be reused to
1189 	 * avoid the copy_huge_page writes to become visible after
1190 	 * the set_pmd_at() write.
1191 	 */
1192 	__SetPageUptodate(new_page);
1193 	pgtable = pmd_pgtable(_pmd);
1194 
1195 	_pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1196 	_pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1197 
1198 	spin_lock(pmd_ptl);
1199 	BUG_ON(!pmd_none(*pmd));
1200 	page_add_new_anon_rmap(new_page, vma, address, true);
1201 	lru_cache_add_inactive_or_unevictable(new_page, vma);
1202 	pgtable_trans_huge_deposit(mm, pmd, pgtable);
1203 	set_pmd_at(mm, address, pmd, _pmd);
1204 	update_mmu_cache_pmd(vma, address, pmd);
1205 	spin_unlock(pmd_ptl);
1206 
1207 	*hpage = NULL;
1208 
1209 	khugepaged_pages_collapsed++;
1210 	result = SCAN_SUCCEED;
1211 out_up_write:
1212 	mmap_write_unlock(mm);
1213 out_nolock:
1214 	if (!IS_ERR_OR_NULL(*hpage))
1215 		mem_cgroup_uncharge(*hpage);
1216 	trace_mm_collapse_huge_page(mm, isolated, result);
1217 	return;
1218 }
1219 
1220 static int khugepaged_scan_pmd(struct mm_struct *mm,
1221 			       struct vm_area_struct *vma,
1222 			       unsigned long address,
1223 			       struct page **hpage)
1224 {
1225 	pmd_t *pmd;
1226 	pte_t *pte, *_pte;
1227 	int ret = 0, result = 0, referenced = 0;
1228 	int none_or_zero = 0, shared = 0;
1229 	struct page *page = NULL;
1230 	unsigned long _address;
1231 	spinlock_t *ptl;
1232 	int node = NUMA_NO_NODE, unmapped = 0;
1233 	bool writable = false;
1234 
1235 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1236 
1237 	pmd = mm_find_pmd(mm, address);
1238 	if (!pmd) {
1239 		result = SCAN_PMD_NULL;
1240 		goto out;
1241 	}
1242 
1243 	memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1244 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1245 	for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1246 	     _pte++, _address += PAGE_SIZE) {
1247 		pte_t pteval = *_pte;
1248 		if (is_swap_pte(pteval)) {
1249 			if (++unmapped <= khugepaged_max_ptes_swap) {
1250 				/*
1251 				 * Always be strict with uffd-wp
1252 				 * enabled swap entries.  Please see
1253 				 * comment below for pte_uffd_wp().
1254 				 */
1255 				if (pte_swp_uffd_wp(pteval)) {
1256 					result = SCAN_PTE_UFFD_WP;
1257 					goto out_unmap;
1258 				}
1259 				continue;
1260 			} else {
1261 				result = SCAN_EXCEED_SWAP_PTE;
1262 				goto out_unmap;
1263 			}
1264 		}
1265 		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1266 			if (!userfaultfd_armed(vma) &&
1267 			    ++none_or_zero <= khugepaged_max_ptes_none) {
1268 				continue;
1269 			} else {
1270 				result = SCAN_EXCEED_NONE_PTE;
1271 				goto out_unmap;
1272 			}
1273 		}
1274 		if (pte_uffd_wp(pteval)) {
1275 			/*
1276 			 * Don't collapse the page if any of the small
1277 			 * PTEs are armed with uffd write protection.
1278 			 * Here we can also mark the new huge pmd as
1279 			 * write protected if any of the small ones is
1280 			 * marked but that could bring unknown
1281 			 * userfault messages that falls outside of
1282 			 * the registered range.  So, just be simple.
1283 			 */
1284 			result = SCAN_PTE_UFFD_WP;
1285 			goto out_unmap;
1286 		}
1287 		if (pte_write(pteval))
1288 			writable = true;
1289 
1290 		page = vm_normal_page(vma, _address, pteval);
1291 		if (unlikely(!page)) {
1292 			result = SCAN_PAGE_NULL;
1293 			goto out_unmap;
1294 		}
1295 
1296 		if (page_mapcount(page) > 1 &&
1297 				++shared > khugepaged_max_ptes_shared) {
1298 			result = SCAN_EXCEED_SHARED_PTE;
1299 			goto out_unmap;
1300 		}
1301 
1302 		page = compound_head(page);
1303 
1304 		/*
1305 		 * Record which node the original page is from and save this
1306 		 * information to khugepaged_node_load[].
1307 		 * Khupaged will allocate hugepage from the node has the max
1308 		 * hit record.
1309 		 */
1310 		node = page_to_nid(page);
1311 		if (khugepaged_scan_abort(node)) {
1312 			result = SCAN_SCAN_ABORT;
1313 			goto out_unmap;
1314 		}
1315 		khugepaged_node_load[node]++;
1316 		if (!PageLRU(page)) {
1317 			result = SCAN_PAGE_LRU;
1318 			goto out_unmap;
1319 		}
1320 		if (PageLocked(page)) {
1321 			result = SCAN_PAGE_LOCK;
1322 			goto out_unmap;
1323 		}
1324 		if (!PageAnon(page)) {
1325 			result = SCAN_PAGE_ANON;
1326 			goto out_unmap;
1327 		}
1328 
1329 		/*
1330 		 * Check if the page has any GUP (or other external) pins.
1331 		 *
1332 		 * Here the check is racy it may see totmal_mapcount > refcount
1333 		 * in some cases.
1334 		 * For example, one process with one forked child process.
1335 		 * The parent has the PMD split due to MADV_DONTNEED, then
1336 		 * the child is trying unmap the whole PMD, but khugepaged
1337 		 * may be scanning the parent between the child has
1338 		 * PageDoubleMap flag cleared and dec the mapcount.  So
1339 		 * khugepaged may see total_mapcount > refcount.
1340 		 *
1341 		 * But such case is ephemeral we could always retry collapse
1342 		 * later.  However it may report false positive if the page
1343 		 * has excessive GUP pins (i.e. 512).  Anyway the same check
1344 		 * will be done again later the risk seems low.
1345 		 */
1346 		if (!is_refcount_suitable(page)) {
1347 			result = SCAN_PAGE_COUNT;
1348 			goto out_unmap;
1349 		}
1350 		if (pte_young(pteval) ||
1351 		    page_is_young(page) || PageReferenced(page) ||
1352 		    mmu_notifier_test_young(vma->vm_mm, address))
1353 			referenced++;
1354 	}
1355 	if (!writable) {
1356 		result = SCAN_PAGE_RO;
1357 	} else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1358 		result = SCAN_LACK_REFERENCED_PAGE;
1359 	} else {
1360 		result = SCAN_SUCCEED;
1361 		ret = 1;
1362 	}
1363 out_unmap:
1364 	pte_unmap_unlock(pte, ptl);
1365 	if (ret) {
1366 		node = khugepaged_find_target_node();
1367 		/* collapse_huge_page will return with the mmap_lock released */
1368 		collapse_huge_page(mm, address, hpage, node,
1369 				referenced, unmapped);
1370 	}
1371 out:
1372 	trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1373 				     none_or_zero, result, unmapped);
1374 	return ret;
1375 }
1376 
1377 static void collect_mm_slot(struct mm_slot *mm_slot)
1378 {
1379 	struct mm_struct *mm = mm_slot->mm;
1380 
1381 	lockdep_assert_held(&khugepaged_mm_lock);
1382 
1383 	if (khugepaged_test_exit(mm)) {
1384 		/* free mm_slot */
1385 		hash_del(&mm_slot->hash);
1386 		list_del(&mm_slot->mm_node);
1387 
1388 		/*
1389 		 * Not strictly needed because the mm exited already.
1390 		 *
1391 		 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1392 		 */
1393 
1394 		/* khugepaged_mm_lock actually not necessary for the below */
1395 		free_mm_slot(mm_slot);
1396 		mmdrop(mm);
1397 	}
1398 }
1399 
1400 #ifdef CONFIG_SHMEM
1401 /*
1402  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1403  * khugepaged should try to collapse the page table.
1404  */
1405 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1406 					 unsigned long addr)
1407 {
1408 	struct mm_slot *mm_slot;
1409 
1410 	VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1411 
1412 	spin_lock(&khugepaged_mm_lock);
1413 	mm_slot = get_mm_slot(mm);
1414 	if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1415 		mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1416 	spin_unlock(&khugepaged_mm_lock);
1417 	return 0;
1418 }
1419 
1420 /**
1421  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1422  * address haddr.
1423  *
1424  * @mm: process address space where collapse happens
1425  * @addr: THP collapse address
1426  *
1427  * This function checks whether all the PTEs in the PMD are pointing to the
1428  * right THP. If so, retract the page table so the THP can refault in with
1429  * as pmd-mapped.
1430  */
1431 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1432 {
1433 	unsigned long haddr = addr & HPAGE_PMD_MASK;
1434 	struct vm_area_struct *vma = find_vma(mm, haddr);
1435 	struct page *hpage;
1436 	pte_t *start_pte, *pte;
1437 	pmd_t *pmd, _pmd;
1438 	spinlock_t *ptl;
1439 	int count = 0;
1440 	int i;
1441 
1442 	if (!vma || !vma->vm_file ||
1443 	    !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1444 		return;
1445 
1446 	/*
1447 	 * This vm_flags may not have VM_HUGEPAGE if the page was not
1448 	 * collapsed by this mm. But we can still collapse if the page is
1449 	 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1450 	 * will not fail the vma for missing VM_HUGEPAGE
1451 	 */
1452 	if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1453 		return;
1454 
1455 	hpage = find_lock_page(vma->vm_file->f_mapping,
1456 			       linear_page_index(vma, haddr));
1457 	if (!hpage)
1458 		return;
1459 
1460 	if (!PageHead(hpage))
1461 		goto drop_hpage;
1462 
1463 	pmd = mm_find_pmd(mm, haddr);
1464 	if (!pmd)
1465 		goto drop_hpage;
1466 
1467 	start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1468 
1469 	/* step 1: check all mapped PTEs are to the right huge page */
1470 	for (i = 0, addr = haddr, pte = start_pte;
1471 	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1472 		struct page *page;
1473 
1474 		/* empty pte, skip */
1475 		if (pte_none(*pte))
1476 			continue;
1477 
1478 		/* page swapped out, abort */
1479 		if (!pte_present(*pte))
1480 			goto abort;
1481 
1482 		page = vm_normal_page(vma, addr, *pte);
1483 
1484 		/*
1485 		 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1486 		 * page table, but the new page will not be a subpage of hpage.
1487 		 */
1488 		if (hpage + i != page)
1489 			goto abort;
1490 		count++;
1491 	}
1492 
1493 	/* step 2: adjust rmap */
1494 	for (i = 0, addr = haddr, pte = start_pte;
1495 	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1496 		struct page *page;
1497 
1498 		if (pte_none(*pte))
1499 			continue;
1500 		page = vm_normal_page(vma, addr, *pte);
1501 		page_remove_rmap(page, false);
1502 	}
1503 
1504 	pte_unmap_unlock(start_pte, ptl);
1505 
1506 	/* step 3: set proper refcount and mm_counters. */
1507 	if (count) {
1508 		page_ref_sub(hpage, count);
1509 		add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1510 	}
1511 
1512 	/* step 4: collapse pmd */
1513 	ptl = pmd_lock(vma->vm_mm, pmd);
1514 	_pmd = pmdp_collapse_flush(vma, haddr, pmd);
1515 	spin_unlock(ptl);
1516 	mm_dec_nr_ptes(mm);
1517 	pte_free(mm, pmd_pgtable(_pmd));
1518 
1519 drop_hpage:
1520 	unlock_page(hpage);
1521 	put_page(hpage);
1522 	return;
1523 
1524 abort:
1525 	pte_unmap_unlock(start_pte, ptl);
1526 	goto drop_hpage;
1527 }
1528 
1529 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1530 {
1531 	struct mm_struct *mm = mm_slot->mm;
1532 	int i;
1533 
1534 	if (likely(mm_slot->nr_pte_mapped_thp == 0))
1535 		return;
1536 
1537 	if (!mmap_write_trylock(mm))
1538 		return;
1539 
1540 	if (unlikely(khugepaged_test_exit(mm)))
1541 		goto out;
1542 
1543 	for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1544 		collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1545 
1546 out:
1547 	mm_slot->nr_pte_mapped_thp = 0;
1548 	mmap_write_unlock(mm);
1549 }
1550 
1551 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1552 {
1553 	struct vm_area_struct *vma;
1554 	struct mm_struct *mm;
1555 	unsigned long addr;
1556 	pmd_t *pmd, _pmd;
1557 
1558 	i_mmap_lock_write(mapping);
1559 	vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1560 		/*
1561 		 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1562 		 * got written to. These VMAs are likely not worth investing
1563 		 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1564 		 * later.
1565 		 *
1566 		 * Not that vma->anon_vma check is racy: it can be set up after
1567 		 * the check but before we took mmap_lock by the fault path.
1568 		 * But page lock would prevent establishing any new ptes of the
1569 		 * page, so we are safe.
1570 		 *
1571 		 * An alternative would be drop the check, but check that page
1572 		 * table is clear before calling pmdp_collapse_flush() under
1573 		 * ptl. It has higher chance to recover THP for the VMA, but
1574 		 * has higher cost too.
1575 		 */
1576 		if (vma->anon_vma)
1577 			continue;
1578 		addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1579 		if (addr & ~HPAGE_PMD_MASK)
1580 			continue;
1581 		if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1582 			continue;
1583 		mm = vma->vm_mm;
1584 		pmd = mm_find_pmd(mm, addr);
1585 		if (!pmd)
1586 			continue;
1587 		/*
1588 		 * We need exclusive mmap_lock to retract page table.
1589 		 *
1590 		 * We use trylock due to lock inversion: we need to acquire
1591 		 * mmap_lock while holding page lock. Fault path does it in
1592 		 * reverse order. Trylock is a way to avoid deadlock.
1593 		 */
1594 		if (mmap_write_trylock(mm)) {
1595 			if (!khugepaged_test_exit(mm)) {
1596 				spinlock_t *ptl = pmd_lock(mm, pmd);
1597 				/* assume page table is clear */
1598 				_pmd = pmdp_collapse_flush(vma, addr, pmd);
1599 				spin_unlock(ptl);
1600 				mm_dec_nr_ptes(mm);
1601 				pte_free(mm, pmd_pgtable(_pmd));
1602 			}
1603 			mmap_write_unlock(mm);
1604 		} else {
1605 			/* Try again later */
1606 			khugepaged_add_pte_mapped_thp(mm, addr);
1607 		}
1608 	}
1609 	i_mmap_unlock_write(mapping);
1610 }
1611 
1612 /**
1613  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1614  *
1615  * @mm: process address space where collapse happens
1616  * @file: file that collapse on
1617  * @start: collapse start address
1618  * @hpage: new allocated huge page for collapse
1619  * @node: appointed node the new huge page allocate from
1620  *
1621  * Basic scheme is simple, details are more complex:
1622  *  - allocate and lock a new huge page;
1623  *  - scan page cache replacing old pages with the new one
1624  *    + swap/gup in pages if necessary;
1625  *    + fill in gaps;
1626  *    + keep old pages around in case rollback is required;
1627  *  - if replacing succeeds:
1628  *    + copy data over;
1629  *    + free old pages;
1630  *    + unlock huge page;
1631  *  - if replacing failed;
1632  *    + put all pages back and unfreeze them;
1633  *    + restore gaps in the page cache;
1634  *    + unlock and free huge page;
1635  */
1636 static void collapse_file(struct mm_struct *mm,
1637 		struct file *file, pgoff_t start,
1638 		struct page **hpage, int node)
1639 {
1640 	struct address_space *mapping = file->f_mapping;
1641 	gfp_t gfp;
1642 	struct page *new_page;
1643 	pgoff_t index, end = start + HPAGE_PMD_NR;
1644 	LIST_HEAD(pagelist);
1645 	XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1646 	int nr_none = 0, result = SCAN_SUCCEED;
1647 	bool is_shmem = shmem_file(file);
1648 	int nr;
1649 
1650 	VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1651 	VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1652 
1653 	/* Only allocate from the target node */
1654 	gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1655 
1656 	new_page = khugepaged_alloc_page(hpage, gfp, node);
1657 	if (!new_page) {
1658 		result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1659 		goto out;
1660 	}
1661 
1662 	if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1663 		result = SCAN_CGROUP_CHARGE_FAIL;
1664 		goto out;
1665 	}
1666 	count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1667 
1668 	/* This will be less messy when we use multi-index entries */
1669 	do {
1670 		xas_lock_irq(&xas);
1671 		xas_create_range(&xas);
1672 		if (!xas_error(&xas))
1673 			break;
1674 		xas_unlock_irq(&xas);
1675 		if (!xas_nomem(&xas, GFP_KERNEL)) {
1676 			result = SCAN_FAIL;
1677 			goto out;
1678 		}
1679 	} while (1);
1680 
1681 	__SetPageLocked(new_page);
1682 	if (is_shmem)
1683 		__SetPageSwapBacked(new_page);
1684 	new_page->index = start;
1685 	new_page->mapping = mapping;
1686 
1687 	/*
1688 	 * At this point the new_page is locked and not up-to-date.
1689 	 * It's safe to insert it into the page cache, because nobody would
1690 	 * be able to map it or use it in another way until we unlock it.
1691 	 */
1692 
1693 	xas_set(&xas, start);
1694 	for (index = start; index < end; index++) {
1695 		struct page *page = xas_next(&xas);
1696 
1697 		VM_BUG_ON(index != xas.xa_index);
1698 		if (is_shmem) {
1699 			if (!page) {
1700 				/*
1701 				 * Stop if extent has been truncated or
1702 				 * hole-punched, and is now completely
1703 				 * empty.
1704 				 */
1705 				if (index == start) {
1706 					if (!xas_next_entry(&xas, end - 1)) {
1707 						result = SCAN_TRUNCATED;
1708 						goto xa_locked;
1709 					}
1710 					xas_set(&xas, index);
1711 				}
1712 				if (!shmem_charge(mapping->host, 1)) {
1713 					result = SCAN_FAIL;
1714 					goto xa_locked;
1715 				}
1716 				xas_store(&xas, new_page);
1717 				nr_none++;
1718 				continue;
1719 			}
1720 
1721 			if (xa_is_value(page) || !PageUptodate(page)) {
1722 				xas_unlock_irq(&xas);
1723 				/* swap in or instantiate fallocated page */
1724 				if (shmem_getpage(mapping->host, index, &page,
1725 						  SGP_NOHUGE)) {
1726 					result = SCAN_FAIL;
1727 					goto xa_unlocked;
1728 				}
1729 			} else if (trylock_page(page)) {
1730 				get_page(page);
1731 				xas_unlock_irq(&xas);
1732 			} else {
1733 				result = SCAN_PAGE_LOCK;
1734 				goto xa_locked;
1735 			}
1736 		} else {	/* !is_shmem */
1737 			if (!page || xa_is_value(page)) {
1738 				xas_unlock_irq(&xas);
1739 				page_cache_sync_readahead(mapping, &file->f_ra,
1740 							  file, index,
1741 							  end - index);
1742 				/* drain pagevecs to help isolate_lru_page() */
1743 				lru_add_drain();
1744 				page = find_lock_page(mapping, index);
1745 				if (unlikely(page == NULL)) {
1746 					result = SCAN_FAIL;
1747 					goto xa_unlocked;
1748 				}
1749 			} else if (PageDirty(page)) {
1750 				/*
1751 				 * khugepaged only works on read-only fd,
1752 				 * so this page is dirty because it hasn't
1753 				 * been flushed since first write. There
1754 				 * won't be new dirty pages.
1755 				 *
1756 				 * Trigger async flush here and hope the
1757 				 * writeback is done when khugepaged
1758 				 * revisits this page.
1759 				 *
1760 				 * This is a one-off situation. We are not
1761 				 * forcing writeback in loop.
1762 				 */
1763 				xas_unlock_irq(&xas);
1764 				filemap_flush(mapping);
1765 				result = SCAN_FAIL;
1766 				goto xa_unlocked;
1767 			} else if (trylock_page(page)) {
1768 				get_page(page);
1769 				xas_unlock_irq(&xas);
1770 			} else {
1771 				result = SCAN_PAGE_LOCK;
1772 				goto xa_locked;
1773 			}
1774 		}
1775 
1776 		/*
1777 		 * The page must be locked, so we can drop the i_pages lock
1778 		 * without racing with truncate.
1779 		 */
1780 		VM_BUG_ON_PAGE(!PageLocked(page), page);
1781 
1782 		/* make sure the page is up to date */
1783 		if (unlikely(!PageUptodate(page))) {
1784 			result = SCAN_FAIL;
1785 			goto out_unlock;
1786 		}
1787 
1788 		/*
1789 		 * If file was truncated then extended, or hole-punched, before
1790 		 * we locked the first page, then a THP might be there already.
1791 		 */
1792 		if (PageTransCompound(page)) {
1793 			result = SCAN_PAGE_COMPOUND;
1794 			goto out_unlock;
1795 		}
1796 
1797 		if (page_mapping(page) != mapping) {
1798 			result = SCAN_TRUNCATED;
1799 			goto out_unlock;
1800 		}
1801 
1802 		if (!is_shmem && PageDirty(page)) {
1803 			/*
1804 			 * khugepaged only works on read-only fd, so this
1805 			 * page is dirty because it hasn't been flushed
1806 			 * since first write.
1807 			 */
1808 			result = SCAN_FAIL;
1809 			goto out_unlock;
1810 		}
1811 
1812 		if (isolate_lru_page(page)) {
1813 			result = SCAN_DEL_PAGE_LRU;
1814 			goto out_unlock;
1815 		}
1816 
1817 		if (page_has_private(page) &&
1818 		    !try_to_release_page(page, GFP_KERNEL)) {
1819 			result = SCAN_PAGE_HAS_PRIVATE;
1820 			putback_lru_page(page);
1821 			goto out_unlock;
1822 		}
1823 
1824 		if (page_mapped(page))
1825 			unmap_mapping_pages(mapping, index, 1, false);
1826 
1827 		xas_lock_irq(&xas);
1828 		xas_set(&xas, index);
1829 
1830 		VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1831 		VM_BUG_ON_PAGE(page_mapped(page), page);
1832 
1833 		/*
1834 		 * The page is expected to have page_count() == 3:
1835 		 *  - we hold a pin on it;
1836 		 *  - one reference from page cache;
1837 		 *  - one from isolate_lru_page;
1838 		 */
1839 		if (!page_ref_freeze(page, 3)) {
1840 			result = SCAN_PAGE_COUNT;
1841 			xas_unlock_irq(&xas);
1842 			putback_lru_page(page);
1843 			goto out_unlock;
1844 		}
1845 
1846 		/*
1847 		 * Add the page to the list to be able to undo the collapse if
1848 		 * something go wrong.
1849 		 */
1850 		list_add_tail(&page->lru, &pagelist);
1851 
1852 		/* Finally, replace with the new page. */
1853 		xas_store(&xas, new_page);
1854 		continue;
1855 out_unlock:
1856 		unlock_page(page);
1857 		put_page(page);
1858 		goto xa_unlocked;
1859 	}
1860 	nr = thp_nr_pages(new_page);
1861 
1862 	if (is_shmem)
1863 		__mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
1864 	else {
1865 		__mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
1866 		filemap_nr_thps_inc(mapping);
1867 	}
1868 
1869 	if (nr_none) {
1870 		__mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1871 		if (is_shmem)
1872 			__mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1873 	}
1874 
1875 xa_locked:
1876 	xas_unlock_irq(&xas);
1877 xa_unlocked:
1878 
1879 	if (result == SCAN_SUCCEED) {
1880 		struct page *page, *tmp;
1881 
1882 		/*
1883 		 * Replacing old pages with new one has succeeded, now we
1884 		 * need to copy the content and free the old pages.
1885 		 */
1886 		index = start;
1887 		list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1888 			while (index < page->index) {
1889 				clear_highpage(new_page + (index % HPAGE_PMD_NR));
1890 				index++;
1891 			}
1892 			copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1893 					page);
1894 			list_del(&page->lru);
1895 			page->mapping = NULL;
1896 			page_ref_unfreeze(page, 1);
1897 			ClearPageActive(page);
1898 			ClearPageUnevictable(page);
1899 			unlock_page(page);
1900 			put_page(page);
1901 			index++;
1902 		}
1903 		while (index < end) {
1904 			clear_highpage(new_page + (index % HPAGE_PMD_NR));
1905 			index++;
1906 		}
1907 
1908 		SetPageUptodate(new_page);
1909 		page_ref_add(new_page, HPAGE_PMD_NR - 1);
1910 		if (is_shmem)
1911 			set_page_dirty(new_page);
1912 		lru_cache_add(new_page);
1913 
1914 		/*
1915 		 * Remove pte page tables, so we can re-fault the page as huge.
1916 		 */
1917 		retract_page_tables(mapping, start);
1918 		*hpage = NULL;
1919 
1920 		khugepaged_pages_collapsed++;
1921 	} else {
1922 		struct page *page;
1923 
1924 		/* Something went wrong: roll back page cache changes */
1925 		xas_lock_irq(&xas);
1926 		mapping->nrpages -= nr_none;
1927 
1928 		if (is_shmem)
1929 			shmem_uncharge(mapping->host, nr_none);
1930 
1931 		xas_set(&xas, start);
1932 		xas_for_each(&xas, page, end - 1) {
1933 			page = list_first_entry_or_null(&pagelist,
1934 					struct page, lru);
1935 			if (!page || xas.xa_index < page->index) {
1936 				if (!nr_none)
1937 					break;
1938 				nr_none--;
1939 				/* Put holes back where they were */
1940 				xas_store(&xas, NULL);
1941 				continue;
1942 			}
1943 
1944 			VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1945 
1946 			/* Unfreeze the page. */
1947 			list_del(&page->lru);
1948 			page_ref_unfreeze(page, 2);
1949 			xas_store(&xas, page);
1950 			xas_pause(&xas);
1951 			xas_unlock_irq(&xas);
1952 			unlock_page(page);
1953 			putback_lru_page(page);
1954 			xas_lock_irq(&xas);
1955 		}
1956 		VM_BUG_ON(nr_none);
1957 		xas_unlock_irq(&xas);
1958 
1959 		new_page->mapping = NULL;
1960 	}
1961 
1962 	unlock_page(new_page);
1963 out:
1964 	VM_BUG_ON(!list_empty(&pagelist));
1965 	if (!IS_ERR_OR_NULL(*hpage))
1966 		mem_cgroup_uncharge(*hpage);
1967 	/* TODO: tracepoints */
1968 }
1969 
1970 static void khugepaged_scan_file(struct mm_struct *mm,
1971 		struct file *file, pgoff_t start, struct page **hpage)
1972 {
1973 	struct page *page = NULL;
1974 	struct address_space *mapping = file->f_mapping;
1975 	XA_STATE(xas, &mapping->i_pages, start);
1976 	int present, swap;
1977 	int node = NUMA_NO_NODE;
1978 	int result = SCAN_SUCCEED;
1979 
1980 	present = 0;
1981 	swap = 0;
1982 	memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1983 	rcu_read_lock();
1984 	xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1985 		if (xas_retry(&xas, page))
1986 			continue;
1987 
1988 		if (xa_is_value(page)) {
1989 			if (++swap > khugepaged_max_ptes_swap) {
1990 				result = SCAN_EXCEED_SWAP_PTE;
1991 				break;
1992 			}
1993 			continue;
1994 		}
1995 
1996 		if (PageTransCompound(page)) {
1997 			result = SCAN_PAGE_COMPOUND;
1998 			break;
1999 		}
2000 
2001 		node = page_to_nid(page);
2002 		if (khugepaged_scan_abort(node)) {
2003 			result = SCAN_SCAN_ABORT;
2004 			break;
2005 		}
2006 		khugepaged_node_load[node]++;
2007 
2008 		if (!PageLRU(page)) {
2009 			result = SCAN_PAGE_LRU;
2010 			break;
2011 		}
2012 
2013 		if (page_count(page) !=
2014 		    1 + page_mapcount(page) + page_has_private(page)) {
2015 			result = SCAN_PAGE_COUNT;
2016 			break;
2017 		}
2018 
2019 		/*
2020 		 * We probably should check if the page is referenced here, but
2021 		 * nobody would transfer pte_young() to PageReferenced() for us.
2022 		 * And rmap walk here is just too costly...
2023 		 */
2024 
2025 		present++;
2026 
2027 		if (need_resched()) {
2028 			xas_pause(&xas);
2029 			cond_resched_rcu();
2030 		}
2031 	}
2032 	rcu_read_unlock();
2033 
2034 	if (result == SCAN_SUCCEED) {
2035 		if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2036 			result = SCAN_EXCEED_NONE_PTE;
2037 		} else {
2038 			node = khugepaged_find_target_node();
2039 			collapse_file(mm, file, start, hpage, node);
2040 		}
2041 	}
2042 
2043 	/* TODO: tracepoints */
2044 }
2045 #else
2046 static void khugepaged_scan_file(struct mm_struct *mm,
2047 		struct file *file, pgoff_t start, struct page **hpage)
2048 {
2049 	BUILD_BUG();
2050 }
2051 
2052 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2053 {
2054 }
2055 #endif
2056 
2057 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2058 					    struct page **hpage)
2059 	__releases(&khugepaged_mm_lock)
2060 	__acquires(&khugepaged_mm_lock)
2061 {
2062 	struct mm_slot *mm_slot;
2063 	struct mm_struct *mm;
2064 	struct vm_area_struct *vma;
2065 	int progress = 0;
2066 
2067 	VM_BUG_ON(!pages);
2068 	lockdep_assert_held(&khugepaged_mm_lock);
2069 
2070 	if (khugepaged_scan.mm_slot)
2071 		mm_slot = khugepaged_scan.mm_slot;
2072 	else {
2073 		mm_slot = list_entry(khugepaged_scan.mm_head.next,
2074 				     struct mm_slot, mm_node);
2075 		khugepaged_scan.address = 0;
2076 		khugepaged_scan.mm_slot = mm_slot;
2077 	}
2078 	spin_unlock(&khugepaged_mm_lock);
2079 	khugepaged_collapse_pte_mapped_thps(mm_slot);
2080 
2081 	mm = mm_slot->mm;
2082 	/*
2083 	 * Don't wait for semaphore (to avoid long wait times).  Just move to
2084 	 * the next mm on the list.
2085 	 */
2086 	vma = NULL;
2087 	if (unlikely(!mmap_read_trylock(mm)))
2088 		goto breakouterloop_mmap_lock;
2089 	if (likely(!khugepaged_test_exit(mm)))
2090 		vma = find_vma(mm, khugepaged_scan.address);
2091 
2092 	progress++;
2093 	for (; vma; vma = vma->vm_next) {
2094 		unsigned long hstart, hend;
2095 
2096 		cond_resched();
2097 		if (unlikely(khugepaged_test_exit(mm))) {
2098 			progress++;
2099 			break;
2100 		}
2101 		if (!hugepage_vma_check(vma, vma->vm_flags)) {
2102 skip:
2103 			progress++;
2104 			continue;
2105 		}
2106 		hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2107 		hend = vma->vm_end & HPAGE_PMD_MASK;
2108 		if (hstart >= hend)
2109 			goto skip;
2110 		if (khugepaged_scan.address > hend)
2111 			goto skip;
2112 		if (khugepaged_scan.address < hstart)
2113 			khugepaged_scan.address = hstart;
2114 		VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2115 		if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2116 			goto skip;
2117 
2118 		while (khugepaged_scan.address < hend) {
2119 			int ret;
2120 			cond_resched();
2121 			if (unlikely(khugepaged_test_exit(mm)))
2122 				goto breakouterloop;
2123 
2124 			VM_BUG_ON(khugepaged_scan.address < hstart ||
2125 				  khugepaged_scan.address + HPAGE_PMD_SIZE >
2126 				  hend);
2127 			if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2128 				struct file *file = get_file(vma->vm_file);
2129 				pgoff_t pgoff = linear_page_index(vma,
2130 						khugepaged_scan.address);
2131 
2132 				mmap_read_unlock(mm);
2133 				ret = 1;
2134 				khugepaged_scan_file(mm, file, pgoff, hpage);
2135 				fput(file);
2136 			} else {
2137 				ret = khugepaged_scan_pmd(mm, vma,
2138 						khugepaged_scan.address,
2139 						hpage);
2140 			}
2141 			/* move to next address */
2142 			khugepaged_scan.address += HPAGE_PMD_SIZE;
2143 			progress += HPAGE_PMD_NR;
2144 			if (ret)
2145 				/* we released mmap_lock so break loop */
2146 				goto breakouterloop_mmap_lock;
2147 			if (progress >= pages)
2148 				goto breakouterloop;
2149 		}
2150 	}
2151 breakouterloop:
2152 	mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2153 breakouterloop_mmap_lock:
2154 
2155 	spin_lock(&khugepaged_mm_lock);
2156 	VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2157 	/*
2158 	 * Release the current mm_slot if this mm is about to die, or
2159 	 * if we scanned all vmas of this mm.
2160 	 */
2161 	if (khugepaged_test_exit(mm) || !vma) {
2162 		/*
2163 		 * Make sure that if mm_users is reaching zero while
2164 		 * khugepaged runs here, khugepaged_exit will find
2165 		 * mm_slot not pointing to the exiting mm.
2166 		 */
2167 		if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2168 			khugepaged_scan.mm_slot = list_entry(
2169 				mm_slot->mm_node.next,
2170 				struct mm_slot, mm_node);
2171 			khugepaged_scan.address = 0;
2172 		} else {
2173 			khugepaged_scan.mm_slot = NULL;
2174 			khugepaged_full_scans++;
2175 		}
2176 
2177 		collect_mm_slot(mm_slot);
2178 	}
2179 
2180 	return progress;
2181 }
2182 
2183 static int khugepaged_has_work(void)
2184 {
2185 	return !list_empty(&khugepaged_scan.mm_head) &&
2186 		khugepaged_enabled();
2187 }
2188 
2189 static int khugepaged_wait_event(void)
2190 {
2191 	return !list_empty(&khugepaged_scan.mm_head) ||
2192 		kthread_should_stop();
2193 }
2194 
2195 static void khugepaged_do_scan(void)
2196 {
2197 	struct page *hpage = NULL;
2198 	unsigned int progress = 0, pass_through_head = 0;
2199 	unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2200 	bool wait = true;
2201 
2202 	lru_add_drain_all();
2203 
2204 	while (progress < pages) {
2205 		if (!khugepaged_prealloc_page(&hpage, &wait))
2206 			break;
2207 
2208 		cond_resched();
2209 
2210 		if (unlikely(kthread_should_stop() || try_to_freeze()))
2211 			break;
2212 
2213 		spin_lock(&khugepaged_mm_lock);
2214 		if (!khugepaged_scan.mm_slot)
2215 			pass_through_head++;
2216 		if (khugepaged_has_work() &&
2217 		    pass_through_head < 2)
2218 			progress += khugepaged_scan_mm_slot(pages - progress,
2219 							    &hpage);
2220 		else
2221 			progress = pages;
2222 		spin_unlock(&khugepaged_mm_lock);
2223 	}
2224 
2225 	if (!IS_ERR_OR_NULL(hpage))
2226 		put_page(hpage);
2227 }
2228 
2229 static bool khugepaged_should_wakeup(void)
2230 {
2231 	return kthread_should_stop() ||
2232 	       time_after_eq(jiffies, khugepaged_sleep_expire);
2233 }
2234 
2235 static void khugepaged_wait_work(void)
2236 {
2237 	if (khugepaged_has_work()) {
2238 		const unsigned long scan_sleep_jiffies =
2239 			msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2240 
2241 		if (!scan_sleep_jiffies)
2242 			return;
2243 
2244 		khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2245 		wait_event_freezable_timeout(khugepaged_wait,
2246 					     khugepaged_should_wakeup(),
2247 					     scan_sleep_jiffies);
2248 		return;
2249 	}
2250 
2251 	if (khugepaged_enabled())
2252 		wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2253 }
2254 
2255 static int khugepaged(void *none)
2256 {
2257 	struct mm_slot *mm_slot;
2258 
2259 	set_freezable();
2260 	set_user_nice(current, MAX_NICE);
2261 
2262 	while (!kthread_should_stop()) {
2263 		khugepaged_do_scan();
2264 		khugepaged_wait_work();
2265 	}
2266 
2267 	spin_lock(&khugepaged_mm_lock);
2268 	mm_slot = khugepaged_scan.mm_slot;
2269 	khugepaged_scan.mm_slot = NULL;
2270 	if (mm_slot)
2271 		collect_mm_slot(mm_slot);
2272 	spin_unlock(&khugepaged_mm_lock);
2273 	return 0;
2274 }
2275 
2276 static void set_recommended_min_free_kbytes(void)
2277 {
2278 	struct zone *zone;
2279 	int nr_zones = 0;
2280 	unsigned long recommended_min;
2281 
2282 	for_each_populated_zone(zone) {
2283 		/*
2284 		 * We don't need to worry about fragmentation of
2285 		 * ZONE_MOVABLE since it only has movable pages.
2286 		 */
2287 		if (zone_idx(zone) > gfp_zone(GFP_USER))
2288 			continue;
2289 
2290 		nr_zones++;
2291 	}
2292 
2293 	/* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2294 	recommended_min = pageblock_nr_pages * nr_zones * 2;
2295 
2296 	/*
2297 	 * Make sure that on average at least two pageblocks are almost free
2298 	 * of another type, one for a migratetype to fall back to and a
2299 	 * second to avoid subsequent fallbacks of other types There are 3
2300 	 * MIGRATE_TYPES we care about.
2301 	 */
2302 	recommended_min += pageblock_nr_pages * nr_zones *
2303 			   MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2304 
2305 	/* don't ever allow to reserve more than 5% of the lowmem */
2306 	recommended_min = min(recommended_min,
2307 			      (unsigned long) nr_free_buffer_pages() / 20);
2308 	recommended_min <<= (PAGE_SHIFT-10);
2309 
2310 	if (recommended_min > min_free_kbytes) {
2311 		if (user_min_free_kbytes >= 0)
2312 			pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2313 				min_free_kbytes, recommended_min);
2314 
2315 		min_free_kbytes = recommended_min;
2316 	}
2317 	setup_per_zone_wmarks();
2318 }
2319 
2320 int start_stop_khugepaged(void)
2321 {
2322 	int err = 0;
2323 
2324 	mutex_lock(&khugepaged_mutex);
2325 	if (khugepaged_enabled()) {
2326 		if (!khugepaged_thread)
2327 			khugepaged_thread = kthread_run(khugepaged, NULL,
2328 							"khugepaged");
2329 		if (IS_ERR(khugepaged_thread)) {
2330 			pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2331 			err = PTR_ERR(khugepaged_thread);
2332 			khugepaged_thread = NULL;
2333 			goto fail;
2334 		}
2335 
2336 		if (!list_empty(&khugepaged_scan.mm_head))
2337 			wake_up_interruptible(&khugepaged_wait);
2338 
2339 		set_recommended_min_free_kbytes();
2340 	} else if (khugepaged_thread) {
2341 		kthread_stop(khugepaged_thread);
2342 		khugepaged_thread = NULL;
2343 	}
2344 fail:
2345 	mutex_unlock(&khugepaged_mutex);
2346 	return err;
2347 }
2348 
2349 void khugepaged_min_free_kbytes_update(void)
2350 {
2351 	mutex_lock(&khugepaged_mutex);
2352 	if (khugepaged_enabled() && khugepaged_thread)
2353 		set_recommended_min_free_kbytes();
2354 	mutex_unlock(&khugepaged_mutex);
2355 }
2356