xref: /openbmc/linux/mm/khugepaged.c (revision 4fa6893faeaaea4fe4440512d2a708527ef47051)
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/page_table_check.h>
20 #include <linux/swapops.h>
21 #include <linux/shmem_fs.h>
22 
23 #include <asm/tlb.h>
24 #include <asm/pgalloc.h>
25 #include "internal.h"
26 
27 enum scan_result {
28 	SCAN_FAIL,
29 	SCAN_SUCCEED,
30 	SCAN_PMD_NULL,
31 	SCAN_EXCEED_NONE_PTE,
32 	SCAN_EXCEED_SWAP_PTE,
33 	SCAN_EXCEED_SHARED_PTE,
34 	SCAN_PTE_NON_PRESENT,
35 	SCAN_PTE_UFFD_WP,
36 	SCAN_PAGE_RO,
37 	SCAN_LACK_REFERENCED_PAGE,
38 	SCAN_PAGE_NULL,
39 	SCAN_SCAN_ABORT,
40 	SCAN_PAGE_COUNT,
41 	SCAN_PAGE_LRU,
42 	SCAN_PAGE_LOCK,
43 	SCAN_PAGE_ANON,
44 	SCAN_PAGE_COMPOUND,
45 	SCAN_ANY_PROCESS,
46 	SCAN_VMA_NULL,
47 	SCAN_VMA_CHECK,
48 	SCAN_ADDRESS_RANGE,
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_RW(scan_sleep_millisecs);
151 
152 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
153 					  struct kobj_attribute *attr,
154 					  char *buf)
155 {
156 	return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
157 }
158 
159 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
160 					   struct kobj_attribute *attr,
161 					   const char *buf, size_t count)
162 {
163 	unsigned int msecs;
164 	int err;
165 
166 	err = kstrtouint(buf, 10, &msecs);
167 	if (err)
168 		return -EINVAL;
169 
170 	khugepaged_alloc_sleep_millisecs = msecs;
171 	khugepaged_sleep_expire = 0;
172 	wake_up_interruptible(&khugepaged_wait);
173 
174 	return count;
175 }
176 static struct kobj_attribute alloc_sleep_millisecs_attr =
177 	__ATTR_RW(alloc_sleep_millisecs);
178 
179 static ssize_t pages_to_scan_show(struct kobject *kobj,
180 				  struct kobj_attribute *attr,
181 				  char *buf)
182 {
183 	return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
184 }
185 static ssize_t pages_to_scan_store(struct kobject *kobj,
186 				   struct kobj_attribute *attr,
187 				   const char *buf, size_t count)
188 {
189 	unsigned int pages;
190 	int err;
191 
192 	err = kstrtouint(buf, 10, &pages);
193 	if (err || !pages)
194 		return -EINVAL;
195 
196 	khugepaged_pages_to_scan = pages;
197 
198 	return count;
199 }
200 static struct kobj_attribute pages_to_scan_attr =
201 	__ATTR_RW(pages_to_scan);
202 
203 static ssize_t pages_collapsed_show(struct kobject *kobj,
204 				    struct kobj_attribute *attr,
205 				    char *buf)
206 {
207 	return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
208 }
209 static struct kobj_attribute pages_collapsed_attr =
210 	__ATTR_RO(pages_collapsed);
211 
212 static ssize_t full_scans_show(struct kobject *kobj,
213 			       struct kobj_attribute *attr,
214 			       char *buf)
215 {
216 	return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
217 }
218 static struct kobj_attribute full_scans_attr =
219 	__ATTR_RO(full_scans);
220 
221 static ssize_t defrag_show(struct kobject *kobj,
222 			   struct kobj_attribute *attr, char *buf)
223 {
224 	return single_hugepage_flag_show(kobj, attr, buf,
225 					 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
226 }
227 static ssize_t defrag_store(struct kobject *kobj,
228 			    struct kobj_attribute *attr,
229 			    const char *buf, size_t count)
230 {
231 	return single_hugepage_flag_store(kobj, attr, buf, count,
232 				 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
233 }
234 static struct kobj_attribute khugepaged_defrag_attr =
235 	__ATTR_RW(defrag);
236 
237 /*
238  * max_ptes_none controls if khugepaged should collapse hugepages over
239  * any unmapped ptes in turn potentially increasing the memory
240  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
241  * reduce the available free memory in the system as it
242  * runs. Increasing max_ptes_none will instead potentially reduce the
243  * free memory in the system during the khugepaged scan.
244  */
245 static ssize_t max_ptes_none_show(struct kobject *kobj,
246 				  struct kobj_attribute *attr,
247 				  char *buf)
248 {
249 	return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
250 }
251 static ssize_t max_ptes_none_store(struct kobject *kobj,
252 				   struct kobj_attribute *attr,
253 				   const char *buf, size_t count)
254 {
255 	int err;
256 	unsigned long max_ptes_none;
257 
258 	err = kstrtoul(buf, 10, &max_ptes_none);
259 	if (err || max_ptes_none > HPAGE_PMD_NR - 1)
260 		return -EINVAL;
261 
262 	khugepaged_max_ptes_none = max_ptes_none;
263 
264 	return count;
265 }
266 static struct kobj_attribute khugepaged_max_ptes_none_attr =
267 	__ATTR_RW(max_ptes_none);
268 
269 static ssize_t max_ptes_swap_show(struct kobject *kobj,
270 				  struct kobj_attribute *attr,
271 				  char *buf)
272 {
273 	return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
274 }
275 
276 static ssize_t 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_RW(max_ptes_swap);
294 
295 static ssize_t max_ptes_shared_show(struct kobject *kobj,
296 				    struct kobj_attribute *attr,
297 				    char *buf)
298 {
299 	return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
300 }
301 
302 static ssize_t max_ptes_shared_store(struct kobject *kobj,
303 				     struct kobj_attribute *attr,
304 				     const char *buf, size_t count)
305 {
306 	int err;
307 	unsigned long max_ptes_shared;
308 
309 	err  = kstrtoul(buf, 10, &max_ptes_shared);
310 	if (err || max_ptes_shared > HPAGE_PMD_NR - 1)
311 		return -EINVAL;
312 
313 	khugepaged_max_ptes_shared = max_ptes_shared;
314 
315 	return count;
316 }
317 
318 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
319 	__ATTR_RW(max_ptes_shared);
320 
321 static struct attribute *khugepaged_attr[] = {
322 	&khugepaged_defrag_attr.attr,
323 	&khugepaged_max_ptes_none_attr.attr,
324 	&khugepaged_max_ptes_swap_attr.attr,
325 	&khugepaged_max_ptes_shared_attr.attr,
326 	&pages_to_scan_attr.attr,
327 	&pages_collapsed_attr.attr,
328 	&full_scans_attr.attr,
329 	&scan_sleep_millisecs_attr.attr,
330 	&alloc_sleep_millisecs_attr.attr,
331 	NULL,
332 };
333 
334 struct attribute_group khugepaged_attr_group = {
335 	.attrs = khugepaged_attr,
336 	.name = "khugepaged",
337 };
338 #endif /* CONFIG_SYSFS */
339 
340 int hugepage_madvise(struct vm_area_struct *vma,
341 		     unsigned long *vm_flags, int advice)
342 {
343 	switch (advice) {
344 	case MADV_HUGEPAGE:
345 #ifdef CONFIG_S390
346 		/*
347 		 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
348 		 * can't handle this properly after s390_enable_sie, so we simply
349 		 * ignore the madvise to prevent qemu from causing a SIGSEGV.
350 		 */
351 		if (mm_has_pgste(vma->vm_mm))
352 			return 0;
353 #endif
354 		*vm_flags &= ~VM_NOHUGEPAGE;
355 		*vm_flags |= VM_HUGEPAGE;
356 		/*
357 		 * If the vma become good for khugepaged to scan,
358 		 * register it here without waiting a page fault that
359 		 * may not happen any time soon.
360 		 */
361 		khugepaged_enter_vma(vma, *vm_flags);
362 		break;
363 	case MADV_NOHUGEPAGE:
364 		*vm_flags &= ~VM_HUGEPAGE;
365 		*vm_flags |= VM_NOHUGEPAGE;
366 		/*
367 		 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
368 		 * this vma even if we leave the mm registered in khugepaged if
369 		 * it got registered before VM_NOHUGEPAGE was set.
370 		 */
371 		break;
372 	}
373 
374 	return 0;
375 }
376 
377 int __init khugepaged_init(void)
378 {
379 	mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
380 					  sizeof(struct mm_slot),
381 					  __alignof__(struct mm_slot), 0, NULL);
382 	if (!mm_slot_cache)
383 		return -ENOMEM;
384 
385 	khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
386 	khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
387 	khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
388 	khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
389 
390 	return 0;
391 }
392 
393 void __init khugepaged_destroy(void)
394 {
395 	kmem_cache_destroy(mm_slot_cache);
396 }
397 
398 static inline struct mm_slot *alloc_mm_slot(void)
399 {
400 	if (!mm_slot_cache)	/* initialization failed */
401 		return NULL;
402 	return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
403 }
404 
405 static inline void free_mm_slot(struct mm_slot *mm_slot)
406 {
407 	kmem_cache_free(mm_slot_cache, mm_slot);
408 }
409 
410 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
411 {
412 	struct mm_slot *mm_slot;
413 
414 	hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
415 		if (mm == mm_slot->mm)
416 			return mm_slot;
417 
418 	return NULL;
419 }
420 
421 static void insert_to_mm_slots_hash(struct mm_struct *mm,
422 				    struct mm_slot *mm_slot)
423 {
424 	mm_slot->mm = mm;
425 	hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
426 }
427 
428 static inline int khugepaged_test_exit(struct mm_struct *mm)
429 {
430 	return atomic_read(&mm->mm_users) == 0;
431 }
432 
433 bool hugepage_vma_check(struct vm_area_struct *vma,
434 			unsigned long vm_flags)
435 {
436 	if (!transhuge_vma_enabled(vma, vm_flags))
437 		return false;
438 
439 	if (vm_flags & VM_NO_KHUGEPAGED)
440 		return false;
441 
442 	/* Don't run khugepaged against DAX vma */
443 	if (vma_is_dax(vma))
444 		return false;
445 
446 	/* Check alignment for file vma and size for both file and anon vma */
447 	if (!transhuge_vma_suitable(vma, (vma->vm_end - HPAGE_PMD_SIZE)))
448 		return false;
449 
450 	/* Enabled via shmem mount options or sysfs settings. */
451 	if (shmem_file(vma->vm_file))
452 		return shmem_huge_enabled(vma);
453 
454 	if (!khugepaged_enabled())
455 		return false;
456 
457 	/* THP settings require madvise. */
458 	if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
459 		return false;
460 
461 	/* Only regular file is valid */
462 	if (file_thp_enabled(vma))
463 		return true;
464 
465 	if (!vma->anon_vma || !vma_is_anonymous(vma))
466 		return false;
467 	if (vma_is_temporary_stack(vma))
468 		return false;
469 
470 	return true;
471 }
472 
473 void __khugepaged_enter(struct mm_struct *mm)
474 {
475 	struct mm_slot *mm_slot;
476 	int wakeup;
477 
478 	mm_slot = alloc_mm_slot();
479 	if (!mm_slot)
480 		return;
481 
482 	/* __khugepaged_exit() must not run from under us */
483 	VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
484 	if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
485 		free_mm_slot(mm_slot);
486 		return;
487 	}
488 
489 	spin_lock(&khugepaged_mm_lock);
490 	insert_to_mm_slots_hash(mm, mm_slot);
491 	/*
492 	 * Insert just behind the scanning cursor, to let the area settle
493 	 * down a little.
494 	 */
495 	wakeup = list_empty(&khugepaged_scan.mm_head);
496 	list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
497 	spin_unlock(&khugepaged_mm_lock);
498 
499 	mmgrab(mm);
500 	if (wakeup)
501 		wake_up_interruptible(&khugepaged_wait);
502 }
503 
504 void khugepaged_enter_vma(struct vm_area_struct *vma,
505 			  unsigned long vm_flags)
506 {
507 	if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) &&
508 	    khugepaged_enabled()) {
509 		if (hugepage_vma_check(vma, vm_flags))
510 			__khugepaged_enter(vma->vm_mm);
511 	}
512 }
513 
514 void __khugepaged_exit(struct mm_struct *mm)
515 {
516 	struct mm_slot *mm_slot;
517 	int free = 0;
518 
519 	spin_lock(&khugepaged_mm_lock);
520 	mm_slot = get_mm_slot(mm);
521 	if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
522 		hash_del(&mm_slot->hash);
523 		list_del(&mm_slot->mm_node);
524 		free = 1;
525 	}
526 	spin_unlock(&khugepaged_mm_lock);
527 
528 	if (free) {
529 		clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
530 		free_mm_slot(mm_slot);
531 		mmdrop(mm);
532 	} else if (mm_slot) {
533 		/*
534 		 * This is required to serialize against
535 		 * khugepaged_test_exit() (which is guaranteed to run
536 		 * under mmap sem read mode). Stop here (after we
537 		 * return all pagetables will be destroyed) until
538 		 * khugepaged has finished working on the pagetables
539 		 * under the mmap_lock.
540 		 */
541 		mmap_write_lock(mm);
542 		mmap_write_unlock(mm);
543 	}
544 }
545 
546 static void release_pte_page(struct page *page)
547 {
548 	mod_node_page_state(page_pgdat(page),
549 			NR_ISOLATED_ANON + page_is_file_lru(page),
550 			-compound_nr(page));
551 	unlock_page(page);
552 	putback_lru_page(page);
553 }
554 
555 static void release_pte_pages(pte_t *pte, pte_t *_pte,
556 		struct list_head *compound_pagelist)
557 {
558 	struct page *page, *tmp;
559 
560 	while (--_pte >= pte) {
561 		pte_t pteval = *_pte;
562 
563 		page = pte_page(pteval);
564 		if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
565 				!PageCompound(page))
566 			release_pte_page(page);
567 	}
568 
569 	list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
570 		list_del(&page->lru);
571 		release_pte_page(page);
572 	}
573 }
574 
575 static bool is_refcount_suitable(struct page *page)
576 {
577 	int expected_refcount;
578 
579 	expected_refcount = total_mapcount(page);
580 	if (PageSwapCache(page))
581 		expected_refcount += compound_nr(page);
582 
583 	return page_count(page) == expected_refcount;
584 }
585 
586 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
587 					unsigned long address,
588 					pte_t *pte,
589 					struct list_head *compound_pagelist)
590 {
591 	struct page *page = NULL;
592 	pte_t *_pte;
593 	int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
594 	bool writable = false;
595 
596 	for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
597 	     _pte++, address += PAGE_SIZE) {
598 		pte_t pteval = *_pte;
599 		if (pte_none(pteval) || (pte_present(pteval) &&
600 				is_zero_pfn(pte_pfn(pteval)))) {
601 			if (!userfaultfd_armed(vma) &&
602 			    ++none_or_zero <= khugepaged_max_ptes_none) {
603 				continue;
604 			} else {
605 				result = SCAN_EXCEED_NONE_PTE;
606 				count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
607 				goto out;
608 			}
609 		}
610 		if (!pte_present(pteval)) {
611 			result = SCAN_PTE_NON_PRESENT;
612 			goto out;
613 		}
614 		page = vm_normal_page(vma, address, pteval);
615 		if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
616 			result = SCAN_PAGE_NULL;
617 			goto out;
618 		}
619 
620 		VM_BUG_ON_PAGE(!PageAnon(page), page);
621 
622 		if (page_mapcount(page) > 1 &&
623 				++shared > khugepaged_max_ptes_shared) {
624 			result = SCAN_EXCEED_SHARED_PTE;
625 			count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
626 			goto out;
627 		}
628 
629 		if (PageCompound(page)) {
630 			struct page *p;
631 			page = compound_head(page);
632 
633 			/*
634 			 * Check if we have dealt with the compound page
635 			 * already
636 			 */
637 			list_for_each_entry(p, compound_pagelist, lru) {
638 				if (page == p)
639 					goto next;
640 			}
641 		}
642 
643 		/*
644 		 * We can do it before isolate_lru_page because the
645 		 * page can't be freed from under us. NOTE: PG_lock
646 		 * is needed to serialize against split_huge_page
647 		 * when invoked from the VM.
648 		 */
649 		if (!trylock_page(page)) {
650 			result = SCAN_PAGE_LOCK;
651 			goto out;
652 		}
653 
654 		/*
655 		 * Check if the page has any GUP (or other external) pins.
656 		 *
657 		 * The page table that maps the page has been already unlinked
658 		 * from the page table tree and this process cannot get
659 		 * an additional pin on the page.
660 		 *
661 		 * New pins can come later if the page is shared across fork,
662 		 * but not from this process. The other process cannot write to
663 		 * the page, only trigger CoW.
664 		 */
665 		if (!is_refcount_suitable(page)) {
666 			unlock_page(page);
667 			result = SCAN_PAGE_COUNT;
668 			goto out;
669 		}
670 
671 		/*
672 		 * Isolate the page to avoid collapsing an hugepage
673 		 * currently in use by the VM.
674 		 */
675 		if (isolate_lru_page(page)) {
676 			unlock_page(page);
677 			result = SCAN_DEL_PAGE_LRU;
678 			goto out;
679 		}
680 		mod_node_page_state(page_pgdat(page),
681 				NR_ISOLATED_ANON + page_is_file_lru(page),
682 				compound_nr(page));
683 		VM_BUG_ON_PAGE(!PageLocked(page), page);
684 		VM_BUG_ON_PAGE(PageLRU(page), page);
685 
686 		if (PageCompound(page))
687 			list_add_tail(&page->lru, compound_pagelist);
688 next:
689 		/* There should be enough young pte to collapse the page */
690 		if (pte_young(pteval) ||
691 		    page_is_young(page) || PageReferenced(page) ||
692 		    mmu_notifier_test_young(vma->vm_mm, address))
693 			referenced++;
694 
695 		if (pte_write(pteval))
696 			writable = true;
697 	}
698 
699 	if (unlikely(!writable)) {
700 		result = SCAN_PAGE_RO;
701 	} else if (unlikely(!referenced)) {
702 		result = SCAN_LACK_REFERENCED_PAGE;
703 	} else {
704 		result = SCAN_SUCCEED;
705 		trace_mm_collapse_huge_page_isolate(page, none_or_zero,
706 						    referenced, writable, result);
707 		return 1;
708 	}
709 out:
710 	release_pte_pages(pte, _pte, compound_pagelist);
711 	trace_mm_collapse_huge_page_isolate(page, none_or_zero,
712 					    referenced, writable, result);
713 	return 0;
714 }
715 
716 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
717 				      struct vm_area_struct *vma,
718 				      unsigned long address,
719 				      spinlock_t *ptl,
720 				      struct list_head *compound_pagelist)
721 {
722 	struct page *src_page, *tmp;
723 	pte_t *_pte;
724 	for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
725 				_pte++, page++, address += PAGE_SIZE) {
726 		pte_t pteval = *_pte;
727 
728 		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
729 			clear_user_highpage(page, address);
730 			add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
731 			if (is_zero_pfn(pte_pfn(pteval))) {
732 				/*
733 				 * ptl mostly unnecessary.
734 				 */
735 				spin_lock(ptl);
736 				ptep_clear(vma->vm_mm, address, _pte);
737 				spin_unlock(ptl);
738 			}
739 		} else {
740 			src_page = pte_page(pteval);
741 			copy_user_highpage(page, src_page, address, vma);
742 			if (!PageCompound(src_page))
743 				release_pte_page(src_page);
744 			/*
745 			 * ptl mostly unnecessary, but preempt has to
746 			 * be disabled to update the per-cpu stats
747 			 * inside page_remove_rmap().
748 			 */
749 			spin_lock(ptl);
750 			ptep_clear(vma->vm_mm, address, _pte);
751 			page_remove_rmap(src_page, vma, false);
752 			spin_unlock(ptl);
753 			free_page_and_swap_cache(src_page);
754 		}
755 	}
756 
757 	list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
758 		list_del(&src_page->lru);
759 		mod_node_page_state(page_pgdat(src_page),
760 				    NR_ISOLATED_ANON + page_is_file_lru(src_page),
761 				    -compound_nr(src_page));
762 		unlock_page(src_page);
763 		free_swap_cache(src_page);
764 		putback_lru_page(src_page);
765 	}
766 }
767 
768 static void khugepaged_alloc_sleep(void)
769 {
770 	DEFINE_WAIT(wait);
771 
772 	add_wait_queue(&khugepaged_wait, &wait);
773 	freezable_schedule_timeout_interruptible(
774 		msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
775 	remove_wait_queue(&khugepaged_wait, &wait);
776 }
777 
778 static int khugepaged_node_load[MAX_NUMNODES];
779 
780 static bool khugepaged_scan_abort(int nid)
781 {
782 	int i;
783 
784 	/*
785 	 * If node_reclaim_mode is disabled, then no extra effort is made to
786 	 * allocate memory locally.
787 	 */
788 	if (!node_reclaim_enabled())
789 		return false;
790 
791 	/* If there is a count for this node already, it must be acceptable */
792 	if (khugepaged_node_load[nid])
793 		return false;
794 
795 	for (i = 0; i < MAX_NUMNODES; i++) {
796 		if (!khugepaged_node_load[i])
797 			continue;
798 		if (node_distance(nid, i) > node_reclaim_distance)
799 			return true;
800 	}
801 	return false;
802 }
803 
804 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
805 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
806 {
807 	return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
808 }
809 
810 #ifdef CONFIG_NUMA
811 static int khugepaged_find_target_node(void)
812 {
813 	static int last_khugepaged_target_node = NUMA_NO_NODE;
814 	int nid, target_node = 0, max_value = 0;
815 
816 	/* find first node with max normal pages hit */
817 	for (nid = 0; nid < MAX_NUMNODES; nid++)
818 		if (khugepaged_node_load[nid] > max_value) {
819 			max_value = khugepaged_node_load[nid];
820 			target_node = nid;
821 		}
822 
823 	/* do some balance if several nodes have the same hit record */
824 	if (target_node <= last_khugepaged_target_node)
825 		for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
826 				nid++)
827 			if (max_value == khugepaged_node_load[nid]) {
828 				target_node = nid;
829 				break;
830 			}
831 
832 	last_khugepaged_target_node = target_node;
833 	return target_node;
834 }
835 
836 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
837 {
838 	if (IS_ERR(*hpage)) {
839 		if (!*wait)
840 			return false;
841 
842 		*wait = false;
843 		*hpage = NULL;
844 		khugepaged_alloc_sleep();
845 	} else if (*hpage) {
846 		put_page(*hpage);
847 		*hpage = NULL;
848 	}
849 
850 	return true;
851 }
852 
853 static struct page *
854 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
855 {
856 	VM_BUG_ON_PAGE(*hpage, *hpage);
857 
858 	*hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
859 	if (unlikely(!*hpage)) {
860 		count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
861 		*hpage = ERR_PTR(-ENOMEM);
862 		return NULL;
863 	}
864 
865 	prep_transhuge_page(*hpage);
866 	count_vm_event(THP_COLLAPSE_ALLOC);
867 	return *hpage;
868 }
869 #else
870 static int khugepaged_find_target_node(void)
871 {
872 	return 0;
873 }
874 
875 static inline struct page *alloc_khugepaged_hugepage(void)
876 {
877 	struct page *page;
878 
879 	page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
880 			   HPAGE_PMD_ORDER);
881 	if (page)
882 		prep_transhuge_page(page);
883 	return page;
884 }
885 
886 static struct page *khugepaged_alloc_hugepage(bool *wait)
887 {
888 	struct page *hpage;
889 
890 	do {
891 		hpage = alloc_khugepaged_hugepage();
892 		if (!hpage) {
893 			count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
894 			if (!*wait)
895 				return NULL;
896 
897 			*wait = false;
898 			khugepaged_alloc_sleep();
899 		} else
900 			count_vm_event(THP_COLLAPSE_ALLOC);
901 	} while (unlikely(!hpage) && likely(khugepaged_enabled()));
902 
903 	return hpage;
904 }
905 
906 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
907 {
908 	/*
909 	 * If the hpage allocated earlier was briefly exposed in page cache
910 	 * before collapse_file() failed, it is possible that racing lookups
911 	 * have not yet completed, and would then be unpleasantly surprised by
912 	 * finding the hpage reused for the same mapping at a different offset.
913 	 * Just release the previous allocation if there is any danger of that.
914 	 */
915 	if (*hpage && page_count(*hpage) > 1) {
916 		put_page(*hpage);
917 		*hpage = NULL;
918 	}
919 
920 	if (!*hpage)
921 		*hpage = khugepaged_alloc_hugepage(wait);
922 
923 	if (unlikely(!*hpage))
924 		return false;
925 
926 	return true;
927 }
928 
929 static struct page *
930 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
931 {
932 	VM_BUG_ON(!*hpage);
933 
934 	return  *hpage;
935 }
936 #endif
937 
938 /*
939  * If mmap_lock temporarily dropped, revalidate vma
940  * before taking mmap_lock.
941  * Return 0 if succeeds, otherwise return none-zero
942  * value (scan code).
943  */
944 
945 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
946 		struct vm_area_struct **vmap)
947 {
948 	struct vm_area_struct *vma;
949 
950 	if (unlikely(khugepaged_test_exit(mm)))
951 		return SCAN_ANY_PROCESS;
952 
953 	*vmap = vma = find_vma(mm, address);
954 	if (!vma)
955 		return SCAN_VMA_NULL;
956 
957 	if (!transhuge_vma_suitable(vma, address))
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_is_anonymous(vma))
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  * Note that if false is returned, mmap_lock will be released.
973  */
974 
975 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
976 					struct vm_area_struct *vma,
977 					unsigned long haddr, pmd_t *pmd,
978 					int referenced)
979 {
980 	int swapped_in = 0;
981 	vm_fault_t ret = 0;
982 	unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
983 
984 	for (address = haddr; address < end; address += PAGE_SIZE) {
985 		struct vm_fault vmf = {
986 			.vma = vma,
987 			.address = address,
988 			.pgoff = linear_page_index(vma, haddr),
989 			.flags = FAULT_FLAG_ALLOW_RETRY,
990 			.pmd = pmd,
991 		};
992 
993 		vmf.pte = pte_offset_map(pmd, address);
994 		vmf.orig_pte = *vmf.pte;
995 		if (!is_swap_pte(vmf.orig_pte)) {
996 			pte_unmap(vmf.pte);
997 			continue;
998 		}
999 		ret = do_swap_page(&vmf);
1000 
1001 		/*
1002 		 * do_swap_page returns VM_FAULT_RETRY with released mmap_lock.
1003 		 * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because
1004 		 * we do not retry here and swap entry will remain in pagetable
1005 		 * resulting in later failure.
1006 		 */
1007 		if (ret & VM_FAULT_RETRY) {
1008 			trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1009 			return false;
1010 		}
1011 		if (ret & VM_FAULT_ERROR) {
1012 			mmap_read_unlock(mm);
1013 			trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1014 			return false;
1015 		}
1016 		swapped_in++;
1017 	}
1018 
1019 	/* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1020 	if (swapped_in)
1021 		lru_add_drain();
1022 
1023 	trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1024 	return true;
1025 }
1026 
1027 static void collapse_huge_page(struct mm_struct *mm,
1028 				   unsigned long address,
1029 				   struct page **hpage,
1030 				   int node, int referenced, int unmapped)
1031 {
1032 	LIST_HEAD(compound_pagelist);
1033 	pmd_t *pmd, _pmd;
1034 	pte_t *pte;
1035 	pgtable_t pgtable;
1036 	struct page *new_page;
1037 	spinlock_t *pmd_ptl, *pte_ptl;
1038 	int isolated = 0, result = 0;
1039 	struct vm_area_struct *vma;
1040 	struct mmu_notifier_range range;
1041 	gfp_t gfp;
1042 
1043 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1044 
1045 	/* Only allocate from the target node */
1046 	gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1047 
1048 	/*
1049 	 * Before allocating the hugepage, release the mmap_lock read lock.
1050 	 * The allocation can take potentially a long time if it involves
1051 	 * sync compaction, and we do not need to hold the mmap_lock during
1052 	 * that. We will recheck the vma after taking it again in write mode.
1053 	 */
1054 	mmap_read_unlock(mm);
1055 	new_page = khugepaged_alloc_page(hpage, gfp, node);
1056 	if (!new_page) {
1057 		result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1058 		goto out_nolock;
1059 	}
1060 
1061 	if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) {
1062 		result = SCAN_CGROUP_CHARGE_FAIL;
1063 		goto out_nolock;
1064 	}
1065 	count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1066 
1067 	mmap_read_lock(mm);
1068 	result = hugepage_vma_revalidate(mm, address, &vma);
1069 	if (result) {
1070 		mmap_read_unlock(mm);
1071 		goto out_nolock;
1072 	}
1073 
1074 	pmd = mm_find_pmd(mm, address);
1075 	if (!pmd) {
1076 		result = SCAN_PMD_NULL;
1077 		mmap_read_unlock(mm);
1078 		goto out_nolock;
1079 	}
1080 
1081 	/*
1082 	 * __collapse_huge_page_swapin will return with mmap_lock released
1083 	 * when it fails. So we jump out_nolock directly in that case.
1084 	 * Continuing to collapse causes inconsistency.
1085 	 */
1086 	if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1087 						     pmd, referenced)) {
1088 		goto out_nolock;
1089 	}
1090 
1091 	mmap_read_unlock(mm);
1092 	/*
1093 	 * Prevent all access to pagetables with the exception of
1094 	 * gup_fast later handled by the ptep_clear_flush and the VM
1095 	 * handled by the anon_vma lock + PG_lock.
1096 	 */
1097 	mmap_write_lock(mm);
1098 	result = hugepage_vma_revalidate(mm, address, &vma);
1099 	if (result)
1100 		goto out_up_write;
1101 	/* check if the pmd is still valid */
1102 	if (mm_find_pmd(mm, address) != pmd)
1103 		goto out_up_write;
1104 
1105 	anon_vma_lock_write(vma->anon_vma);
1106 
1107 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1108 				address, address + HPAGE_PMD_SIZE);
1109 	mmu_notifier_invalidate_range_start(&range);
1110 
1111 	pte = pte_offset_map(pmd, address);
1112 	pte_ptl = pte_lockptr(mm, pmd);
1113 
1114 	pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1115 	/*
1116 	 * After this gup_fast can't run anymore. This also removes
1117 	 * any huge TLB entry from the CPU so we won't allow
1118 	 * huge and small TLB entries for the same virtual address
1119 	 * to avoid the risk of CPU bugs in that area.
1120 	 */
1121 	_pmd = pmdp_collapse_flush(vma, address, pmd);
1122 	spin_unlock(pmd_ptl);
1123 	mmu_notifier_invalidate_range_end(&range);
1124 
1125 	spin_lock(pte_ptl);
1126 	isolated = __collapse_huge_page_isolate(vma, address, pte,
1127 			&compound_pagelist);
1128 	spin_unlock(pte_ptl);
1129 
1130 	if (unlikely(!isolated)) {
1131 		pte_unmap(pte);
1132 		spin_lock(pmd_ptl);
1133 		BUG_ON(!pmd_none(*pmd));
1134 		/*
1135 		 * We can only use set_pmd_at when establishing
1136 		 * hugepmds and never for establishing regular pmds that
1137 		 * points to regular pagetables. Use pmd_populate for that
1138 		 */
1139 		pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1140 		spin_unlock(pmd_ptl);
1141 		anon_vma_unlock_write(vma->anon_vma);
1142 		result = SCAN_FAIL;
1143 		goto out_up_write;
1144 	}
1145 
1146 	/*
1147 	 * All pages are isolated and locked so anon_vma rmap
1148 	 * can't run anymore.
1149 	 */
1150 	anon_vma_unlock_write(vma->anon_vma);
1151 
1152 	__collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1153 			&compound_pagelist);
1154 	pte_unmap(pte);
1155 	/*
1156 	 * spin_lock() below is not the equivalent of smp_wmb(), but
1157 	 * the smp_wmb() inside __SetPageUptodate() can be reused to
1158 	 * avoid the copy_huge_page writes to become visible after
1159 	 * the set_pmd_at() write.
1160 	 */
1161 	__SetPageUptodate(new_page);
1162 	pgtable = pmd_pgtable(_pmd);
1163 
1164 	_pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1165 	_pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1166 
1167 	spin_lock(pmd_ptl);
1168 	BUG_ON(!pmd_none(*pmd));
1169 	page_add_new_anon_rmap(new_page, vma, address);
1170 	lru_cache_add_inactive_or_unevictable(new_page, vma);
1171 	pgtable_trans_huge_deposit(mm, pmd, pgtable);
1172 	set_pmd_at(mm, address, pmd, _pmd);
1173 	update_mmu_cache_pmd(vma, address, pmd);
1174 	spin_unlock(pmd_ptl);
1175 
1176 	*hpage = NULL;
1177 
1178 	khugepaged_pages_collapsed++;
1179 	result = SCAN_SUCCEED;
1180 out_up_write:
1181 	mmap_write_unlock(mm);
1182 out_nolock:
1183 	if (!IS_ERR_OR_NULL(*hpage))
1184 		mem_cgroup_uncharge(page_folio(*hpage));
1185 	trace_mm_collapse_huge_page(mm, isolated, result);
1186 	return;
1187 }
1188 
1189 static int khugepaged_scan_pmd(struct mm_struct *mm,
1190 			       struct vm_area_struct *vma,
1191 			       unsigned long address,
1192 			       struct page **hpage)
1193 {
1194 	pmd_t *pmd;
1195 	pte_t *pte, *_pte;
1196 	int ret = 0, result = 0, referenced = 0;
1197 	int none_or_zero = 0, shared = 0;
1198 	struct page *page = NULL;
1199 	unsigned long _address;
1200 	spinlock_t *ptl;
1201 	int node = NUMA_NO_NODE, unmapped = 0;
1202 	bool writable = false;
1203 
1204 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1205 
1206 	pmd = mm_find_pmd(mm, address);
1207 	if (!pmd) {
1208 		result = SCAN_PMD_NULL;
1209 		goto out;
1210 	}
1211 
1212 	memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1213 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1214 	for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR;
1215 	     _pte++, _address += PAGE_SIZE) {
1216 		pte_t pteval = *_pte;
1217 		if (is_swap_pte(pteval)) {
1218 			if (++unmapped <= khugepaged_max_ptes_swap) {
1219 				/*
1220 				 * Always be strict with uffd-wp
1221 				 * enabled swap entries.  Please see
1222 				 * comment below for pte_uffd_wp().
1223 				 */
1224 				if (pte_swp_uffd_wp(pteval)) {
1225 					result = SCAN_PTE_UFFD_WP;
1226 					goto out_unmap;
1227 				}
1228 				continue;
1229 			} else {
1230 				result = SCAN_EXCEED_SWAP_PTE;
1231 				count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1232 				goto out_unmap;
1233 			}
1234 		}
1235 		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1236 			if (!userfaultfd_armed(vma) &&
1237 			    ++none_or_zero <= khugepaged_max_ptes_none) {
1238 				continue;
1239 			} else {
1240 				result = SCAN_EXCEED_NONE_PTE;
1241 				count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1242 				goto out_unmap;
1243 			}
1244 		}
1245 		if (pte_uffd_wp(pteval)) {
1246 			/*
1247 			 * Don't collapse the page if any of the small
1248 			 * PTEs are armed with uffd write protection.
1249 			 * Here we can also mark the new huge pmd as
1250 			 * write protected if any of the small ones is
1251 			 * marked but that could bring unknown
1252 			 * userfault messages that falls outside of
1253 			 * the registered range.  So, just be simple.
1254 			 */
1255 			result = SCAN_PTE_UFFD_WP;
1256 			goto out_unmap;
1257 		}
1258 		if (pte_write(pteval))
1259 			writable = true;
1260 
1261 		page = vm_normal_page(vma, _address, pteval);
1262 		if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
1263 			result = SCAN_PAGE_NULL;
1264 			goto out_unmap;
1265 		}
1266 
1267 		if (page_mapcount(page) > 1 &&
1268 				++shared > khugepaged_max_ptes_shared) {
1269 			result = SCAN_EXCEED_SHARED_PTE;
1270 			count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1271 			goto out_unmap;
1272 		}
1273 
1274 		page = compound_head(page);
1275 
1276 		/*
1277 		 * Record which node the original page is from and save this
1278 		 * information to khugepaged_node_load[].
1279 		 * Khugepaged will allocate hugepage from the node has the max
1280 		 * hit record.
1281 		 */
1282 		node = page_to_nid(page);
1283 		if (khugepaged_scan_abort(node)) {
1284 			result = SCAN_SCAN_ABORT;
1285 			goto out_unmap;
1286 		}
1287 		khugepaged_node_load[node]++;
1288 		if (!PageLRU(page)) {
1289 			result = SCAN_PAGE_LRU;
1290 			goto out_unmap;
1291 		}
1292 		if (PageLocked(page)) {
1293 			result = SCAN_PAGE_LOCK;
1294 			goto out_unmap;
1295 		}
1296 		if (!PageAnon(page)) {
1297 			result = SCAN_PAGE_ANON;
1298 			goto out_unmap;
1299 		}
1300 
1301 		/*
1302 		 * Check if the page has any GUP (or other external) pins.
1303 		 *
1304 		 * Here the check is racy it may see total_mapcount > refcount
1305 		 * in some cases.
1306 		 * For example, one process with one forked child process.
1307 		 * The parent has the PMD split due to MADV_DONTNEED, then
1308 		 * the child is trying unmap the whole PMD, but khugepaged
1309 		 * may be scanning the parent between the child has
1310 		 * PageDoubleMap flag cleared and dec the mapcount.  So
1311 		 * khugepaged may see total_mapcount > refcount.
1312 		 *
1313 		 * But such case is ephemeral we could always retry collapse
1314 		 * later.  However it may report false positive if the page
1315 		 * has excessive GUP pins (i.e. 512).  Anyway the same check
1316 		 * will be done again later the risk seems low.
1317 		 */
1318 		if (!is_refcount_suitable(page)) {
1319 			result = SCAN_PAGE_COUNT;
1320 			goto out_unmap;
1321 		}
1322 		if (pte_young(pteval) ||
1323 		    page_is_young(page) || PageReferenced(page) ||
1324 		    mmu_notifier_test_young(vma->vm_mm, address))
1325 			referenced++;
1326 	}
1327 	if (!writable) {
1328 		result = SCAN_PAGE_RO;
1329 	} else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1330 		result = SCAN_LACK_REFERENCED_PAGE;
1331 	} else {
1332 		result = SCAN_SUCCEED;
1333 		ret = 1;
1334 	}
1335 out_unmap:
1336 	pte_unmap_unlock(pte, ptl);
1337 	if (ret) {
1338 		node = khugepaged_find_target_node();
1339 		/* collapse_huge_page will return with the mmap_lock released */
1340 		collapse_huge_page(mm, address, hpage, node,
1341 				referenced, unmapped);
1342 	}
1343 out:
1344 	trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1345 				     none_or_zero, result, unmapped);
1346 	return ret;
1347 }
1348 
1349 static void collect_mm_slot(struct mm_slot *mm_slot)
1350 {
1351 	struct mm_struct *mm = mm_slot->mm;
1352 
1353 	lockdep_assert_held(&khugepaged_mm_lock);
1354 
1355 	if (khugepaged_test_exit(mm)) {
1356 		/* free mm_slot */
1357 		hash_del(&mm_slot->hash);
1358 		list_del(&mm_slot->mm_node);
1359 
1360 		/*
1361 		 * Not strictly needed because the mm exited already.
1362 		 *
1363 		 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1364 		 */
1365 
1366 		/* khugepaged_mm_lock actually not necessary for the below */
1367 		free_mm_slot(mm_slot);
1368 		mmdrop(mm);
1369 	}
1370 }
1371 
1372 #ifdef CONFIG_SHMEM
1373 /*
1374  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1375  * khugepaged should try to collapse the page table.
1376  */
1377 static void khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1378 					  unsigned long addr)
1379 {
1380 	struct mm_slot *mm_slot;
1381 
1382 	VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1383 
1384 	spin_lock(&khugepaged_mm_lock);
1385 	mm_slot = get_mm_slot(mm);
1386 	if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1387 		mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1388 	spin_unlock(&khugepaged_mm_lock);
1389 }
1390 
1391 static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
1392 				  unsigned long addr, pmd_t *pmdp)
1393 {
1394 	spinlock_t *ptl;
1395 	pmd_t pmd;
1396 
1397 	mmap_assert_write_locked(mm);
1398 	ptl = pmd_lock(vma->vm_mm, pmdp);
1399 	pmd = pmdp_collapse_flush(vma, addr, pmdp);
1400 	spin_unlock(ptl);
1401 	mm_dec_nr_ptes(mm);
1402 	page_table_check_pte_clear_range(mm, addr, pmd);
1403 	pte_free(mm, pmd_pgtable(pmd));
1404 }
1405 
1406 /**
1407  * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1408  * address haddr.
1409  *
1410  * @mm: process address space where collapse happens
1411  * @addr: THP collapse address
1412  *
1413  * This function checks whether all the PTEs in the PMD are pointing to the
1414  * right THP. If so, retract the page table so the THP can refault in with
1415  * as pmd-mapped.
1416  */
1417 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1418 {
1419 	unsigned long haddr = addr & HPAGE_PMD_MASK;
1420 	struct vm_area_struct *vma = find_vma(mm, haddr);
1421 	struct page *hpage;
1422 	pte_t *start_pte, *pte;
1423 	pmd_t *pmd;
1424 	spinlock_t *ptl;
1425 	int count = 0;
1426 	int i;
1427 
1428 	if (!vma || !vma->vm_file ||
1429 	    !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1430 		return;
1431 
1432 	/*
1433 	 * This vm_flags may not have VM_HUGEPAGE if the page was not
1434 	 * collapsed by this mm. But we can still collapse if the page is
1435 	 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1436 	 * will not fail the vma for missing VM_HUGEPAGE
1437 	 */
1438 	if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1439 		return;
1440 
1441 	/* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */
1442 	if (userfaultfd_wp(vma))
1443 		return;
1444 
1445 	hpage = find_lock_page(vma->vm_file->f_mapping,
1446 			       linear_page_index(vma, haddr));
1447 	if (!hpage)
1448 		return;
1449 
1450 	if (!PageHead(hpage))
1451 		goto drop_hpage;
1452 
1453 	pmd = mm_find_pmd(mm, haddr);
1454 	if (!pmd)
1455 		goto drop_hpage;
1456 
1457 	start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1458 
1459 	/* step 1: check all mapped PTEs are to the right huge page */
1460 	for (i = 0, addr = haddr, pte = start_pte;
1461 	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1462 		struct page *page;
1463 
1464 		/* empty pte, skip */
1465 		if (pte_none(*pte))
1466 			continue;
1467 
1468 		/* page swapped out, abort */
1469 		if (!pte_present(*pte))
1470 			goto abort;
1471 
1472 		page = vm_normal_page(vma, addr, *pte);
1473 		if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1474 			page = NULL;
1475 		/*
1476 		 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1477 		 * page table, but the new page will not be a subpage of hpage.
1478 		 */
1479 		if (hpage + i != page)
1480 			goto abort;
1481 		count++;
1482 	}
1483 
1484 	/* step 2: adjust rmap */
1485 	for (i = 0, addr = haddr, pte = start_pte;
1486 	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1487 		struct page *page;
1488 
1489 		if (pte_none(*pte))
1490 			continue;
1491 		page = vm_normal_page(vma, addr, *pte);
1492 		if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1493 			goto abort;
1494 		page_remove_rmap(page, vma, false);
1495 	}
1496 
1497 	pte_unmap_unlock(start_pte, ptl);
1498 
1499 	/* step 3: set proper refcount and mm_counters. */
1500 	if (count) {
1501 		page_ref_sub(hpage, count);
1502 		add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1503 	}
1504 
1505 	/* step 4: collapse pmd */
1506 	collapse_and_free_pmd(mm, vma, haddr, pmd);
1507 drop_hpage:
1508 	unlock_page(hpage);
1509 	put_page(hpage);
1510 	return;
1511 
1512 abort:
1513 	pte_unmap_unlock(start_pte, ptl);
1514 	goto drop_hpage;
1515 }
1516 
1517 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1518 {
1519 	struct mm_struct *mm = mm_slot->mm;
1520 	int i;
1521 
1522 	if (likely(mm_slot->nr_pte_mapped_thp == 0))
1523 		return;
1524 
1525 	if (!mmap_write_trylock(mm))
1526 		return;
1527 
1528 	if (unlikely(khugepaged_test_exit(mm)))
1529 		goto out;
1530 
1531 	for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1532 		collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1533 
1534 out:
1535 	mm_slot->nr_pte_mapped_thp = 0;
1536 	mmap_write_unlock(mm);
1537 }
1538 
1539 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1540 {
1541 	struct vm_area_struct *vma;
1542 	struct mm_struct *mm;
1543 	unsigned long addr;
1544 	pmd_t *pmd;
1545 
1546 	i_mmap_lock_write(mapping);
1547 	vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1548 		/*
1549 		 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1550 		 * got written to. These VMAs are likely not worth investing
1551 		 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1552 		 * later.
1553 		 *
1554 		 * Note that vma->anon_vma check is racy: it can be set up after
1555 		 * the check but before we took mmap_lock by the fault path.
1556 		 * But page lock would prevent establishing any new ptes of the
1557 		 * page, so we are safe.
1558 		 *
1559 		 * An alternative would be drop the check, but check that page
1560 		 * table is clear before calling pmdp_collapse_flush() under
1561 		 * ptl. It has higher chance to recover THP for the VMA, but
1562 		 * has higher cost too.
1563 		 */
1564 		if (vma->anon_vma)
1565 			continue;
1566 		addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1567 		if (addr & ~HPAGE_PMD_MASK)
1568 			continue;
1569 		if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1570 			continue;
1571 		mm = vma->vm_mm;
1572 		pmd = mm_find_pmd(mm, addr);
1573 		if (!pmd)
1574 			continue;
1575 		/*
1576 		 * We need exclusive mmap_lock to retract page table.
1577 		 *
1578 		 * We use trylock due to lock inversion: we need to acquire
1579 		 * mmap_lock while holding page lock. Fault path does it in
1580 		 * reverse order. Trylock is a way to avoid deadlock.
1581 		 */
1582 		if (mmap_write_trylock(mm)) {
1583 			/*
1584 			 * When a vma is registered with uffd-wp, we can't
1585 			 * recycle the pmd pgtable because there can be pte
1586 			 * markers installed.  Skip it only, so the rest mm/vma
1587 			 * can still have the same file mapped hugely, however
1588 			 * it'll always mapped in small page size for uffd-wp
1589 			 * registered ranges.
1590 			 */
1591 			if (!khugepaged_test_exit(mm) && !userfaultfd_wp(vma))
1592 				collapse_and_free_pmd(mm, vma, addr, pmd);
1593 			mmap_write_unlock(mm);
1594 		} else {
1595 			/* Try again later */
1596 			khugepaged_add_pte_mapped_thp(mm, addr);
1597 		}
1598 	}
1599 	i_mmap_unlock_write(mapping);
1600 }
1601 
1602 /**
1603  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1604  *
1605  * @mm: process address space where collapse happens
1606  * @file: file that collapse on
1607  * @start: collapse start address
1608  * @hpage: new allocated huge page for collapse
1609  * @node: appointed node the new huge page allocate from
1610  *
1611  * Basic scheme is simple, details are more complex:
1612  *  - allocate and lock a new huge page;
1613  *  - scan page cache replacing old pages with the new one
1614  *    + swap/gup in pages if necessary;
1615  *    + fill in gaps;
1616  *    + keep old pages around in case rollback is required;
1617  *  - if replacing succeeds:
1618  *    + copy data over;
1619  *    + free old pages;
1620  *    + unlock huge page;
1621  *  - if replacing failed;
1622  *    + put all pages back and unfreeze them;
1623  *    + restore gaps in the page cache;
1624  *    + unlock and free huge page;
1625  */
1626 static void collapse_file(struct mm_struct *mm,
1627 		struct file *file, pgoff_t start,
1628 		struct page **hpage, int node)
1629 {
1630 	struct address_space *mapping = file->f_mapping;
1631 	gfp_t gfp;
1632 	struct page *new_page;
1633 	pgoff_t index, end = start + HPAGE_PMD_NR;
1634 	LIST_HEAD(pagelist);
1635 	XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1636 	int nr_none = 0, result = SCAN_SUCCEED;
1637 	bool is_shmem = shmem_file(file);
1638 	int nr;
1639 
1640 	VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1641 	VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1642 
1643 	/* Only allocate from the target node */
1644 	gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1645 
1646 	new_page = khugepaged_alloc_page(hpage, gfp, node);
1647 	if (!new_page) {
1648 		result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1649 		goto out;
1650 	}
1651 
1652 	if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) {
1653 		result = SCAN_CGROUP_CHARGE_FAIL;
1654 		goto out;
1655 	}
1656 	count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1657 
1658 	/*
1659 	 * Ensure we have slots for all the pages in the range.  This is
1660 	 * almost certainly a no-op because most of the pages must be present
1661 	 */
1662 	do {
1663 		xas_lock_irq(&xas);
1664 		xas_create_range(&xas);
1665 		if (!xas_error(&xas))
1666 			break;
1667 		xas_unlock_irq(&xas);
1668 		if (!xas_nomem(&xas, GFP_KERNEL)) {
1669 			result = SCAN_FAIL;
1670 			goto out;
1671 		}
1672 	} while (1);
1673 
1674 	__SetPageLocked(new_page);
1675 	if (is_shmem)
1676 		__SetPageSwapBacked(new_page);
1677 	new_page->index = start;
1678 	new_page->mapping = mapping;
1679 
1680 	/*
1681 	 * At this point the new_page is locked and not up-to-date.
1682 	 * It's safe to insert it into the page cache, because nobody would
1683 	 * be able to map it or use it in another way until we unlock it.
1684 	 */
1685 
1686 	xas_set(&xas, start);
1687 	for (index = start; index < end; index++) {
1688 		struct page *page = xas_next(&xas);
1689 
1690 		VM_BUG_ON(index != xas.xa_index);
1691 		if (is_shmem) {
1692 			if (!page) {
1693 				/*
1694 				 * Stop if extent has been truncated or
1695 				 * hole-punched, and is now completely
1696 				 * empty.
1697 				 */
1698 				if (index == start) {
1699 					if (!xas_next_entry(&xas, end - 1)) {
1700 						result = SCAN_TRUNCATED;
1701 						goto xa_locked;
1702 					}
1703 					xas_set(&xas, index);
1704 				}
1705 				if (!shmem_charge(mapping->host, 1)) {
1706 					result = SCAN_FAIL;
1707 					goto xa_locked;
1708 				}
1709 				xas_store(&xas, new_page);
1710 				nr_none++;
1711 				continue;
1712 			}
1713 
1714 			if (xa_is_value(page) || !PageUptodate(page)) {
1715 				xas_unlock_irq(&xas);
1716 				/* swap in or instantiate fallocated page */
1717 				if (shmem_getpage(mapping->host, index, &page,
1718 						  SGP_NOALLOC)) {
1719 					result = SCAN_FAIL;
1720 					goto xa_unlocked;
1721 				}
1722 			} else if (trylock_page(page)) {
1723 				get_page(page);
1724 				xas_unlock_irq(&xas);
1725 			} else {
1726 				result = SCAN_PAGE_LOCK;
1727 				goto xa_locked;
1728 			}
1729 		} else {	/* !is_shmem */
1730 			if (!page || xa_is_value(page)) {
1731 				xas_unlock_irq(&xas);
1732 				page_cache_sync_readahead(mapping, &file->f_ra,
1733 							  file, index,
1734 							  end - index);
1735 				/* drain pagevecs to help isolate_lru_page() */
1736 				lru_add_drain();
1737 				page = find_lock_page(mapping, index);
1738 				if (unlikely(page == NULL)) {
1739 					result = SCAN_FAIL;
1740 					goto xa_unlocked;
1741 				}
1742 			} else if (PageDirty(page)) {
1743 				/*
1744 				 * khugepaged only works on read-only fd,
1745 				 * so this page is dirty because it hasn't
1746 				 * been flushed since first write. There
1747 				 * won't be new dirty pages.
1748 				 *
1749 				 * Trigger async flush here and hope the
1750 				 * writeback is done when khugepaged
1751 				 * revisits this page.
1752 				 *
1753 				 * This is a one-off situation. We are not
1754 				 * forcing writeback in loop.
1755 				 */
1756 				xas_unlock_irq(&xas);
1757 				filemap_flush(mapping);
1758 				result = SCAN_FAIL;
1759 				goto xa_unlocked;
1760 			} else if (PageWriteback(page)) {
1761 				xas_unlock_irq(&xas);
1762 				result = SCAN_FAIL;
1763 				goto xa_unlocked;
1764 			} else if (trylock_page(page)) {
1765 				get_page(page);
1766 				xas_unlock_irq(&xas);
1767 			} else {
1768 				result = SCAN_PAGE_LOCK;
1769 				goto xa_locked;
1770 			}
1771 		}
1772 
1773 		/*
1774 		 * The page must be locked, so we can drop the i_pages lock
1775 		 * without racing with truncate.
1776 		 */
1777 		VM_BUG_ON_PAGE(!PageLocked(page), page);
1778 
1779 		/* make sure the page is up to date */
1780 		if (unlikely(!PageUptodate(page))) {
1781 			result = SCAN_FAIL;
1782 			goto out_unlock;
1783 		}
1784 
1785 		/*
1786 		 * If file was truncated then extended, or hole-punched, before
1787 		 * we locked the first page, then a THP might be there already.
1788 		 */
1789 		if (PageTransCompound(page)) {
1790 			result = SCAN_PAGE_COMPOUND;
1791 			goto out_unlock;
1792 		}
1793 
1794 		if (page_mapping(page) != mapping) {
1795 			result = SCAN_TRUNCATED;
1796 			goto out_unlock;
1797 		}
1798 
1799 		if (!is_shmem && (PageDirty(page) ||
1800 				  PageWriteback(page))) {
1801 			/*
1802 			 * khugepaged only works on read-only fd, so this
1803 			 * page is dirty because it hasn't been flushed
1804 			 * since first write.
1805 			 */
1806 			result = SCAN_FAIL;
1807 			goto out_unlock;
1808 		}
1809 
1810 		if (isolate_lru_page(page)) {
1811 			result = SCAN_DEL_PAGE_LRU;
1812 			goto out_unlock;
1813 		}
1814 
1815 		if (page_has_private(page) &&
1816 		    !try_to_release_page(page, GFP_KERNEL)) {
1817 			result = SCAN_PAGE_HAS_PRIVATE;
1818 			putback_lru_page(page);
1819 			goto out_unlock;
1820 		}
1821 
1822 		if (page_mapped(page))
1823 			try_to_unmap(page_folio(page),
1824 					TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH);
1825 
1826 		xas_lock_irq(&xas);
1827 		xas_set(&xas, index);
1828 
1829 		VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1830 
1831 		/*
1832 		 * The page is expected to have page_count() == 3:
1833 		 *  - we hold a pin on it;
1834 		 *  - one reference from page cache;
1835 		 *  - one from isolate_lru_page;
1836 		 */
1837 		if (!page_ref_freeze(page, 3)) {
1838 			result = SCAN_PAGE_COUNT;
1839 			xas_unlock_irq(&xas);
1840 			putback_lru_page(page);
1841 			goto out_unlock;
1842 		}
1843 
1844 		/*
1845 		 * Add the page to the list to be able to undo the collapse if
1846 		 * something go wrong.
1847 		 */
1848 		list_add_tail(&page->lru, &pagelist);
1849 
1850 		/* Finally, replace with the new page. */
1851 		xas_store(&xas, new_page);
1852 		continue;
1853 out_unlock:
1854 		unlock_page(page);
1855 		put_page(page);
1856 		goto xa_unlocked;
1857 	}
1858 	nr = thp_nr_pages(new_page);
1859 
1860 	if (is_shmem)
1861 		__mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
1862 	else {
1863 		__mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
1864 		filemap_nr_thps_inc(mapping);
1865 		/*
1866 		 * Paired with smp_mb() in do_dentry_open() to ensure
1867 		 * i_writecount is up to date and the update to nr_thps is
1868 		 * visible. Ensures the page cache will be truncated if the
1869 		 * file is opened writable.
1870 		 */
1871 		smp_mb();
1872 		if (inode_is_open_for_write(mapping->host)) {
1873 			result = SCAN_FAIL;
1874 			__mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);
1875 			filemap_nr_thps_dec(mapping);
1876 			goto xa_locked;
1877 		}
1878 	}
1879 
1880 	if (nr_none) {
1881 		__mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1882 		/* nr_none is always 0 for non-shmem. */
1883 		__mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1884 	}
1885 
1886 	/* Join all the small entries into a single multi-index entry */
1887 	xas_set_order(&xas, start, HPAGE_PMD_ORDER);
1888 	xas_store(&xas, new_page);
1889 xa_locked:
1890 	xas_unlock_irq(&xas);
1891 xa_unlocked:
1892 
1893 	/*
1894 	 * If collapse is successful, flush must be done now before copying.
1895 	 * If collapse is unsuccessful, does flush actually need to be done?
1896 	 * Do it anyway, to clear the state.
1897 	 */
1898 	try_to_unmap_flush();
1899 
1900 	if (result == SCAN_SUCCEED) {
1901 		struct page *page, *tmp;
1902 
1903 		/*
1904 		 * Replacing old pages with new one has succeeded, now we
1905 		 * need to copy the content and free the old pages.
1906 		 */
1907 		index = start;
1908 		list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1909 			while (index < page->index) {
1910 				clear_highpage(new_page + (index % HPAGE_PMD_NR));
1911 				index++;
1912 			}
1913 			copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1914 					page);
1915 			list_del(&page->lru);
1916 			page->mapping = NULL;
1917 			page_ref_unfreeze(page, 1);
1918 			ClearPageActive(page);
1919 			ClearPageUnevictable(page);
1920 			unlock_page(page);
1921 			put_page(page);
1922 			index++;
1923 		}
1924 		while (index < end) {
1925 			clear_highpage(new_page + (index % HPAGE_PMD_NR));
1926 			index++;
1927 		}
1928 
1929 		SetPageUptodate(new_page);
1930 		page_ref_add(new_page, HPAGE_PMD_NR - 1);
1931 		if (is_shmem)
1932 			set_page_dirty(new_page);
1933 		lru_cache_add(new_page);
1934 
1935 		/*
1936 		 * Remove pte page tables, so we can re-fault the page as huge.
1937 		 */
1938 		retract_page_tables(mapping, start);
1939 		*hpage = NULL;
1940 
1941 		khugepaged_pages_collapsed++;
1942 	} else {
1943 		struct page *page;
1944 
1945 		/* Something went wrong: roll back page cache changes */
1946 		xas_lock_irq(&xas);
1947 		if (nr_none) {
1948 			mapping->nrpages -= nr_none;
1949 			shmem_uncharge(mapping->host, nr_none);
1950 		}
1951 
1952 		xas_set(&xas, start);
1953 		xas_for_each(&xas, page, end - 1) {
1954 			page = list_first_entry_or_null(&pagelist,
1955 					struct page, lru);
1956 			if (!page || xas.xa_index < page->index) {
1957 				if (!nr_none)
1958 					break;
1959 				nr_none--;
1960 				/* Put holes back where they were */
1961 				xas_store(&xas, NULL);
1962 				continue;
1963 			}
1964 
1965 			VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1966 
1967 			/* Unfreeze the page. */
1968 			list_del(&page->lru);
1969 			page_ref_unfreeze(page, 2);
1970 			xas_store(&xas, page);
1971 			xas_pause(&xas);
1972 			xas_unlock_irq(&xas);
1973 			unlock_page(page);
1974 			putback_lru_page(page);
1975 			xas_lock_irq(&xas);
1976 		}
1977 		VM_BUG_ON(nr_none);
1978 		xas_unlock_irq(&xas);
1979 
1980 		new_page->mapping = NULL;
1981 	}
1982 
1983 	unlock_page(new_page);
1984 out:
1985 	VM_BUG_ON(!list_empty(&pagelist));
1986 	if (!IS_ERR_OR_NULL(*hpage))
1987 		mem_cgroup_uncharge(page_folio(*hpage));
1988 	/* TODO: tracepoints */
1989 }
1990 
1991 static void khugepaged_scan_file(struct mm_struct *mm,
1992 		struct file *file, pgoff_t start, struct page **hpage)
1993 {
1994 	struct page *page = NULL;
1995 	struct address_space *mapping = file->f_mapping;
1996 	XA_STATE(xas, &mapping->i_pages, start);
1997 	int present, swap;
1998 	int node = NUMA_NO_NODE;
1999 	int result = SCAN_SUCCEED;
2000 
2001 	present = 0;
2002 	swap = 0;
2003 	memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2004 	rcu_read_lock();
2005 	xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2006 		if (xas_retry(&xas, page))
2007 			continue;
2008 
2009 		if (xa_is_value(page)) {
2010 			if (++swap > khugepaged_max_ptes_swap) {
2011 				result = SCAN_EXCEED_SWAP_PTE;
2012 				count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
2013 				break;
2014 			}
2015 			continue;
2016 		}
2017 
2018 		/*
2019 		 * XXX: khugepaged should compact smaller compound pages
2020 		 * into a PMD sized page
2021 		 */
2022 		if (PageTransCompound(page)) {
2023 			result = SCAN_PAGE_COMPOUND;
2024 			break;
2025 		}
2026 
2027 		node = page_to_nid(page);
2028 		if (khugepaged_scan_abort(node)) {
2029 			result = SCAN_SCAN_ABORT;
2030 			break;
2031 		}
2032 		khugepaged_node_load[node]++;
2033 
2034 		if (!PageLRU(page)) {
2035 			result = SCAN_PAGE_LRU;
2036 			break;
2037 		}
2038 
2039 		if (page_count(page) !=
2040 		    1 + page_mapcount(page) + page_has_private(page)) {
2041 			result = SCAN_PAGE_COUNT;
2042 			break;
2043 		}
2044 
2045 		/*
2046 		 * We probably should check if the page is referenced here, but
2047 		 * nobody would transfer pte_young() to PageReferenced() for us.
2048 		 * And rmap walk here is just too costly...
2049 		 */
2050 
2051 		present++;
2052 
2053 		if (need_resched()) {
2054 			xas_pause(&xas);
2055 			cond_resched_rcu();
2056 		}
2057 	}
2058 	rcu_read_unlock();
2059 
2060 	if (result == SCAN_SUCCEED) {
2061 		if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2062 			result = SCAN_EXCEED_NONE_PTE;
2063 			count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2064 		} else {
2065 			node = khugepaged_find_target_node();
2066 			collapse_file(mm, file, start, hpage, node);
2067 		}
2068 	}
2069 
2070 	/* TODO: tracepoints */
2071 }
2072 #else
2073 static void khugepaged_scan_file(struct mm_struct *mm,
2074 		struct file *file, pgoff_t start, struct page **hpage)
2075 {
2076 	BUILD_BUG();
2077 }
2078 
2079 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2080 {
2081 }
2082 #endif
2083 
2084 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2085 					    struct page **hpage)
2086 	__releases(&khugepaged_mm_lock)
2087 	__acquires(&khugepaged_mm_lock)
2088 {
2089 	struct mm_slot *mm_slot;
2090 	struct mm_struct *mm;
2091 	struct vm_area_struct *vma;
2092 	int progress = 0;
2093 
2094 	VM_BUG_ON(!pages);
2095 	lockdep_assert_held(&khugepaged_mm_lock);
2096 
2097 	if (khugepaged_scan.mm_slot)
2098 		mm_slot = khugepaged_scan.mm_slot;
2099 	else {
2100 		mm_slot = list_entry(khugepaged_scan.mm_head.next,
2101 				     struct mm_slot, mm_node);
2102 		khugepaged_scan.address = 0;
2103 		khugepaged_scan.mm_slot = mm_slot;
2104 	}
2105 	spin_unlock(&khugepaged_mm_lock);
2106 	khugepaged_collapse_pte_mapped_thps(mm_slot);
2107 
2108 	mm = mm_slot->mm;
2109 	/*
2110 	 * Don't wait for semaphore (to avoid long wait times).  Just move to
2111 	 * the next mm on the list.
2112 	 */
2113 	vma = NULL;
2114 	if (unlikely(!mmap_read_trylock(mm)))
2115 		goto breakouterloop_mmap_lock;
2116 	if (likely(!khugepaged_test_exit(mm)))
2117 		vma = find_vma(mm, khugepaged_scan.address);
2118 
2119 	progress++;
2120 	for (; vma; vma = vma->vm_next) {
2121 		unsigned long hstart, hend;
2122 
2123 		cond_resched();
2124 		if (unlikely(khugepaged_test_exit(mm))) {
2125 			progress++;
2126 			break;
2127 		}
2128 		if (!hugepage_vma_check(vma, vma->vm_flags)) {
2129 skip:
2130 			progress++;
2131 			continue;
2132 		}
2133 		hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE);
2134 		hend = round_down(vma->vm_end, HPAGE_PMD_SIZE);
2135 		if (khugepaged_scan.address > hend)
2136 			goto skip;
2137 		if (khugepaged_scan.address < hstart)
2138 			khugepaged_scan.address = hstart;
2139 		VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2140 
2141 		while (khugepaged_scan.address < hend) {
2142 			int ret;
2143 			cond_resched();
2144 			if (unlikely(khugepaged_test_exit(mm)))
2145 				goto breakouterloop;
2146 
2147 			VM_BUG_ON(khugepaged_scan.address < hstart ||
2148 				  khugepaged_scan.address + HPAGE_PMD_SIZE >
2149 				  hend);
2150 			if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2151 				struct file *file = get_file(vma->vm_file);
2152 				pgoff_t pgoff = linear_page_index(vma,
2153 						khugepaged_scan.address);
2154 
2155 				mmap_read_unlock(mm);
2156 				ret = 1;
2157 				khugepaged_scan_file(mm, file, pgoff, hpage);
2158 				fput(file);
2159 			} else {
2160 				ret = khugepaged_scan_pmd(mm, vma,
2161 						khugepaged_scan.address,
2162 						hpage);
2163 			}
2164 			/* move to next address */
2165 			khugepaged_scan.address += HPAGE_PMD_SIZE;
2166 			progress += HPAGE_PMD_NR;
2167 			if (ret)
2168 				/* we released mmap_lock so break loop */
2169 				goto breakouterloop_mmap_lock;
2170 			if (progress >= pages)
2171 				goto breakouterloop;
2172 		}
2173 	}
2174 breakouterloop:
2175 	mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2176 breakouterloop_mmap_lock:
2177 
2178 	spin_lock(&khugepaged_mm_lock);
2179 	VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2180 	/*
2181 	 * Release the current mm_slot if this mm is about to die, or
2182 	 * if we scanned all vmas of this mm.
2183 	 */
2184 	if (khugepaged_test_exit(mm) || !vma) {
2185 		/*
2186 		 * Make sure that if mm_users is reaching zero while
2187 		 * khugepaged runs here, khugepaged_exit will find
2188 		 * mm_slot not pointing to the exiting mm.
2189 		 */
2190 		if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2191 			khugepaged_scan.mm_slot = list_entry(
2192 				mm_slot->mm_node.next,
2193 				struct mm_slot, mm_node);
2194 			khugepaged_scan.address = 0;
2195 		} else {
2196 			khugepaged_scan.mm_slot = NULL;
2197 			khugepaged_full_scans++;
2198 		}
2199 
2200 		collect_mm_slot(mm_slot);
2201 	}
2202 
2203 	return progress;
2204 }
2205 
2206 static int khugepaged_has_work(void)
2207 {
2208 	return !list_empty(&khugepaged_scan.mm_head) &&
2209 		khugepaged_enabled();
2210 }
2211 
2212 static int khugepaged_wait_event(void)
2213 {
2214 	return !list_empty(&khugepaged_scan.mm_head) ||
2215 		kthread_should_stop();
2216 }
2217 
2218 static void khugepaged_do_scan(void)
2219 {
2220 	struct page *hpage = NULL;
2221 	unsigned int progress = 0, pass_through_head = 0;
2222 	unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2223 	bool wait = true;
2224 
2225 	lru_add_drain_all();
2226 
2227 	while (progress < pages) {
2228 		if (!khugepaged_prealloc_page(&hpage, &wait))
2229 			break;
2230 
2231 		cond_resched();
2232 
2233 		if (unlikely(kthread_should_stop() || try_to_freeze()))
2234 			break;
2235 
2236 		spin_lock(&khugepaged_mm_lock);
2237 		if (!khugepaged_scan.mm_slot)
2238 			pass_through_head++;
2239 		if (khugepaged_has_work() &&
2240 		    pass_through_head < 2)
2241 			progress += khugepaged_scan_mm_slot(pages - progress,
2242 							    &hpage);
2243 		else
2244 			progress = pages;
2245 		spin_unlock(&khugepaged_mm_lock);
2246 	}
2247 
2248 	if (!IS_ERR_OR_NULL(hpage))
2249 		put_page(hpage);
2250 }
2251 
2252 static bool khugepaged_should_wakeup(void)
2253 {
2254 	return kthread_should_stop() ||
2255 	       time_after_eq(jiffies, khugepaged_sleep_expire);
2256 }
2257 
2258 static void khugepaged_wait_work(void)
2259 {
2260 	if (khugepaged_has_work()) {
2261 		const unsigned long scan_sleep_jiffies =
2262 			msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2263 
2264 		if (!scan_sleep_jiffies)
2265 			return;
2266 
2267 		khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2268 		wait_event_freezable_timeout(khugepaged_wait,
2269 					     khugepaged_should_wakeup(),
2270 					     scan_sleep_jiffies);
2271 		return;
2272 	}
2273 
2274 	if (khugepaged_enabled())
2275 		wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2276 }
2277 
2278 static int khugepaged(void *none)
2279 {
2280 	struct mm_slot *mm_slot;
2281 
2282 	set_freezable();
2283 	set_user_nice(current, MAX_NICE);
2284 
2285 	while (!kthread_should_stop()) {
2286 		khugepaged_do_scan();
2287 		khugepaged_wait_work();
2288 	}
2289 
2290 	spin_lock(&khugepaged_mm_lock);
2291 	mm_slot = khugepaged_scan.mm_slot;
2292 	khugepaged_scan.mm_slot = NULL;
2293 	if (mm_slot)
2294 		collect_mm_slot(mm_slot);
2295 	spin_unlock(&khugepaged_mm_lock);
2296 	return 0;
2297 }
2298 
2299 static void set_recommended_min_free_kbytes(void)
2300 {
2301 	struct zone *zone;
2302 	int nr_zones = 0;
2303 	unsigned long recommended_min;
2304 
2305 	if (!khugepaged_enabled()) {
2306 		calculate_min_free_kbytes();
2307 		goto update_wmarks;
2308 	}
2309 
2310 	for_each_populated_zone(zone) {
2311 		/*
2312 		 * We don't need to worry about fragmentation of
2313 		 * ZONE_MOVABLE since it only has movable pages.
2314 		 */
2315 		if (zone_idx(zone) > gfp_zone(GFP_USER))
2316 			continue;
2317 
2318 		nr_zones++;
2319 	}
2320 
2321 	/* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2322 	recommended_min = pageblock_nr_pages * nr_zones * 2;
2323 
2324 	/*
2325 	 * Make sure that on average at least two pageblocks are almost free
2326 	 * of another type, one for a migratetype to fall back to and a
2327 	 * second to avoid subsequent fallbacks of other types There are 3
2328 	 * MIGRATE_TYPES we care about.
2329 	 */
2330 	recommended_min += pageblock_nr_pages * nr_zones *
2331 			   MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2332 
2333 	/* don't ever allow to reserve more than 5% of the lowmem */
2334 	recommended_min = min(recommended_min,
2335 			      (unsigned long) nr_free_buffer_pages() / 20);
2336 	recommended_min <<= (PAGE_SHIFT-10);
2337 
2338 	if (recommended_min > min_free_kbytes) {
2339 		if (user_min_free_kbytes >= 0)
2340 			pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2341 				min_free_kbytes, recommended_min);
2342 
2343 		min_free_kbytes = recommended_min;
2344 	}
2345 
2346 update_wmarks:
2347 	setup_per_zone_wmarks();
2348 }
2349 
2350 int start_stop_khugepaged(void)
2351 {
2352 	int err = 0;
2353 
2354 	mutex_lock(&khugepaged_mutex);
2355 	if (khugepaged_enabled()) {
2356 		if (!khugepaged_thread)
2357 			khugepaged_thread = kthread_run(khugepaged, NULL,
2358 							"khugepaged");
2359 		if (IS_ERR(khugepaged_thread)) {
2360 			pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2361 			err = PTR_ERR(khugepaged_thread);
2362 			khugepaged_thread = NULL;
2363 			goto fail;
2364 		}
2365 
2366 		if (!list_empty(&khugepaged_scan.mm_head))
2367 			wake_up_interruptible(&khugepaged_wait);
2368 	} else if (khugepaged_thread) {
2369 		kthread_stop(khugepaged_thread);
2370 		khugepaged_thread = NULL;
2371 	}
2372 	set_recommended_min_free_kbytes();
2373 fail:
2374 	mutex_unlock(&khugepaged_mutex);
2375 	return err;
2376 }
2377 
2378 void khugepaged_min_free_kbytes_update(void)
2379 {
2380 	mutex_lock(&khugepaged_mutex);
2381 	if (khugepaged_enabled() && khugepaged_thread)
2382 		set_recommended_min_free_kbytes();
2383 	mutex_unlock(&khugepaged_mutex);
2384 }
2385