xref: /linux/mm/khugepaged.c (revision 24ce659dcc02c21f8d6c0a7589c3320a4dfa8152)
1 // SPDX-License-Identifier: GPL-2.0
2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
3 
4 #include <linux/mm.h>
5 #include <linux/sched.h>
6 #include <linux/sched/mm.h>
7 #include <linux/sched/coredump.h>
8 #include <linux/mmu_notifier.h>
9 #include <linux/rmap.h>
10 #include <linux/swap.h>
11 #include <linux/mm_inline.h>
12 #include <linux/kthread.h>
13 #include <linux/khugepaged.h>
14 #include <linux/freezer.h>
15 #include <linux/mman.h>
16 #include <linux/hashtable.h>
17 #include <linux/userfaultfd_k.h>
18 #include <linux/page_idle.h>
19 #include <linux/swapops.h>
20 #include <linux/shmem_fs.h>
21 
22 #include <asm/tlb.h>
23 #include <asm/pgalloc.h>
24 #include "internal.h"
25 
26 enum scan_result {
27 	SCAN_FAIL,
28 	SCAN_SUCCEED,
29 	SCAN_PMD_NULL,
30 	SCAN_EXCEED_NONE_PTE,
31 	SCAN_PTE_NON_PRESENT,
32 	SCAN_PTE_UFFD_WP,
33 	SCAN_PAGE_RO,
34 	SCAN_LACK_REFERENCED_PAGE,
35 	SCAN_PAGE_NULL,
36 	SCAN_SCAN_ABORT,
37 	SCAN_PAGE_COUNT,
38 	SCAN_PAGE_LRU,
39 	SCAN_PAGE_LOCK,
40 	SCAN_PAGE_ANON,
41 	SCAN_PAGE_COMPOUND,
42 	SCAN_ANY_PROCESS,
43 	SCAN_VMA_NULL,
44 	SCAN_VMA_CHECK,
45 	SCAN_ADDRESS_RANGE,
46 	SCAN_SWAP_CACHE_PAGE,
47 	SCAN_DEL_PAGE_LRU,
48 	SCAN_ALLOC_HUGE_PAGE_FAIL,
49 	SCAN_CGROUP_CHARGE_FAIL,
50 	SCAN_EXCEED_SWAP_PTE,
51 	SCAN_TRUNCATED,
52 	SCAN_PAGE_HAS_PRIVATE,
53 };
54 
55 #define CREATE_TRACE_POINTS
56 #include <trace/events/huge_memory.h>
57 
58 /* default scan 8*512 pte (or vmas) every 30 second */
59 static unsigned int khugepaged_pages_to_scan __read_mostly;
60 static unsigned int khugepaged_pages_collapsed;
61 static unsigned int khugepaged_full_scans;
62 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
63 /* during fragmentation poll the hugepage allocator once every minute */
64 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
65 static unsigned long khugepaged_sleep_expire;
66 static DEFINE_SPINLOCK(khugepaged_mm_lock);
67 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
68 /*
69  * default collapse hugepages if there is at least one pte mapped like
70  * it would have happened if the vma was large enough during page
71  * fault.
72  */
73 static unsigned int khugepaged_max_ptes_none __read_mostly;
74 static unsigned int khugepaged_max_ptes_swap __read_mostly;
75 
76 #define MM_SLOTS_HASH_BITS 10
77 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
78 
79 static struct kmem_cache *mm_slot_cache __read_mostly;
80 
81 #define MAX_PTE_MAPPED_THP 8
82 
83 /**
84  * struct mm_slot - hash lookup from mm to mm_slot
85  * @hash: hash collision list
86  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
87  * @mm: the mm that this information is valid for
88  */
89 struct mm_slot {
90 	struct hlist_node hash;
91 	struct list_head mm_node;
92 	struct mm_struct *mm;
93 
94 	/* pte-mapped THP in this mm */
95 	int nr_pte_mapped_thp;
96 	unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
97 };
98 
99 /**
100  * struct khugepaged_scan - cursor for scanning
101  * @mm_head: the head of the mm list to scan
102  * @mm_slot: the current mm_slot we are scanning
103  * @address: the next address inside that to be scanned
104  *
105  * There is only the one khugepaged_scan instance of this cursor structure.
106  */
107 struct khugepaged_scan {
108 	struct list_head mm_head;
109 	struct mm_slot *mm_slot;
110 	unsigned long address;
111 };
112 
113 static struct khugepaged_scan khugepaged_scan = {
114 	.mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
115 };
116 
117 #ifdef CONFIG_SYSFS
118 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
119 					 struct kobj_attribute *attr,
120 					 char *buf)
121 {
122 	return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
123 }
124 
125 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
126 					  struct kobj_attribute *attr,
127 					  const char *buf, size_t count)
128 {
129 	unsigned long msecs;
130 	int err;
131 
132 	err = kstrtoul(buf, 10, &msecs);
133 	if (err || msecs > UINT_MAX)
134 		return -EINVAL;
135 
136 	khugepaged_scan_sleep_millisecs = msecs;
137 	khugepaged_sleep_expire = 0;
138 	wake_up_interruptible(&khugepaged_wait);
139 
140 	return count;
141 }
142 static struct kobj_attribute scan_sleep_millisecs_attr =
143 	__ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
144 	       scan_sleep_millisecs_store);
145 
146 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
147 					  struct kobj_attribute *attr,
148 					  char *buf)
149 {
150 	return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
151 }
152 
153 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
154 					   struct kobj_attribute *attr,
155 					   const char *buf, size_t count)
156 {
157 	unsigned long msecs;
158 	int err;
159 
160 	err = kstrtoul(buf, 10, &msecs);
161 	if (err || msecs > UINT_MAX)
162 		return -EINVAL;
163 
164 	khugepaged_alloc_sleep_millisecs = msecs;
165 	khugepaged_sleep_expire = 0;
166 	wake_up_interruptible(&khugepaged_wait);
167 
168 	return count;
169 }
170 static struct kobj_attribute alloc_sleep_millisecs_attr =
171 	__ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
172 	       alloc_sleep_millisecs_store);
173 
174 static ssize_t pages_to_scan_show(struct kobject *kobj,
175 				  struct kobj_attribute *attr,
176 				  char *buf)
177 {
178 	return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
179 }
180 static ssize_t pages_to_scan_store(struct kobject *kobj,
181 				   struct kobj_attribute *attr,
182 				   const char *buf, size_t count)
183 {
184 	int err;
185 	unsigned long pages;
186 
187 	err = kstrtoul(buf, 10, &pages);
188 	if (err || !pages || pages > UINT_MAX)
189 		return -EINVAL;
190 
191 	khugepaged_pages_to_scan = pages;
192 
193 	return count;
194 }
195 static struct kobj_attribute pages_to_scan_attr =
196 	__ATTR(pages_to_scan, 0644, pages_to_scan_show,
197 	       pages_to_scan_store);
198 
199 static ssize_t pages_collapsed_show(struct kobject *kobj,
200 				    struct kobj_attribute *attr,
201 				    char *buf)
202 {
203 	return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
204 }
205 static struct kobj_attribute pages_collapsed_attr =
206 	__ATTR_RO(pages_collapsed);
207 
208 static ssize_t full_scans_show(struct kobject *kobj,
209 			       struct kobj_attribute *attr,
210 			       char *buf)
211 {
212 	return sprintf(buf, "%u\n", khugepaged_full_scans);
213 }
214 static struct kobj_attribute full_scans_attr =
215 	__ATTR_RO(full_scans);
216 
217 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
218 				      struct kobj_attribute *attr, char *buf)
219 {
220 	return single_hugepage_flag_show(kobj, attr, buf,
221 				TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
222 }
223 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
224 				       struct kobj_attribute *attr,
225 				       const char *buf, size_t count)
226 {
227 	return single_hugepage_flag_store(kobj, attr, buf, count,
228 				 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
229 }
230 static struct kobj_attribute khugepaged_defrag_attr =
231 	__ATTR(defrag, 0644, khugepaged_defrag_show,
232 	       khugepaged_defrag_store);
233 
234 /*
235  * max_ptes_none controls if khugepaged should collapse hugepages over
236  * any unmapped ptes in turn potentially increasing the memory
237  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
238  * reduce the available free memory in the system as it
239  * runs. Increasing max_ptes_none will instead potentially reduce the
240  * free memory in the system during the khugepaged scan.
241  */
242 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
243 					     struct kobj_attribute *attr,
244 					     char *buf)
245 {
246 	return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
247 }
248 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
249 					      struct kobj_attribute *attr,
250 					      const char *buf, size_t count)
251 {
252 	int err;
253 	unsigned long max_ptes_none;
254 
255 	err = kstrtoul(buf, 10, &max_ptes_none);
256 	if (err || max_ptes_none > HPAGE_PMD_NR-1)
257 		return -EINVAL;
258 
259 	khugepaged_max_ptes_none = max_ptes_none;
260 
261 	return count;
262 }
263 static struct kobj_attribute khugepaged_max_ptes_none_attr =
264 	__ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
265 	       khugepaged_max_ptes_none_store);
266 
267 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
268 					     struct kobj_attribute *attr,
269 					     char *buf)
270 {
271 	return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
272 }
273 
274 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
275 					      struct kobj_attribute *attr,
276 					      const char *buf, size_t count)
277 {
278 	int err;
279 	unsigned long max_ptes_swap;
280 
281 	err  = kstrtoul(buf, 10, &max_ptes_swap);
282 	if (err || max_ptes_swap > HPAGE_PMD_NR-1)
283 		return -EINVAL;
284 
285 	khugepaged_max_ptes_swap = max_ptes_swap;
286 
287 	return count;
288 }
289 
290 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
291 	__ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
292 	       khugepaged_max_ptes_swap_store);
293 
294 static struct attribute *khugepaged_attr[] = {
295 	&khugepaged_defrag_attr.attr,
296 	&khugepaged_max_ptes_none_attr.attr,
297 	&pages_to_scan_attr.attr,
298 	&pages_collapsed_attr.attr,
299 	&full_scans_attr.attr,
300 	&scan_sleep_millisecs_attr.attr,
301 	&alloc_sleep_millisecs_attr.attr,
302 	&khugepaged_max_ptes_swap_attr.attr,
303 	NULL,
304 };
305 
306 struct attribute_group khugepaged_attr_group = {
307 	.attrs = khugepaged_attr,
308 	.name = "khugepaged",
309 };
310 #endif /* CONFIG_SYSFS */
311 
312 int hugepage_madvise(struct vm_area_struct *vma,
313 		     unsigned long *vm_flags, int advice)
314 {
315 	switch (advice) {
316 	case MADV_HUGEPAGE:
317 #ifdef CONFIG_S390
318 		/*
319 		 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
320 		 * can't handle this properly after s390_enable_sie, so we simply
321 		 * ignore the madvise to prevent qemu from causing a SIGSEGV.
322 		 */
323 		if (mm_has_pgste(vma->vm_mm))
324 			return 0;
325 #endif
326 		*vm_flags &= ~VM_NOHUGEPAGE;
327 		*vm_flags |= VM_HUGEPAGE;
328 		/*
329 		 * If the vma become good for khugepaged to scan,
330 		 * register it here without waiting a page fault that
331 		 * may not happen any time soon.
332 		 */
333 		if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
334 				khugepaged_enter_vma_merge(vma, *vm_flags))
335 			return -ENOMEM;
336 		break;
337 	case MADV_NOHUGEPAGE:
338 		*vm_flags &= ~VM_HUGEPAGE;
339 		*vm_flags |= VM_NOHUGEPAGE;
340 		/*
341 		 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
342 		 * this vma even if we leave the mm registered in khugepaged if
343 		 * it got registered before VM_NOHUGEPAGE was set.
344 		 */
345 		break;
346 	}
347 
348 	return 0;
349 }
350 
351 int __init khugepaged_init(void)
352 {
353 	mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
354 					  sizeof(struct mm_slot),
355 					  __alignof__(struct mm_slot), 0, NULL);
356 	if (!mm_slot_cache)
357 		return -ENOMEM;
358 
359 	khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
360 	khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
361 	khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
362 
363 	return 0;
364 }
365 
366 void __init khugepaged_destroy(void)
367 {
368 	kmem_cache_destroy(mm_slot_cache);
369 }
370 
371 static inline struct mm_slot *alloc_mm_slot(void)
372 {
373 	if (!mm_slot_cache)	/* initialization failed */
374 		return NULL;
375 	return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
376 }
377 
378 static inline void free_mm_slot(struct mm_slot *mm_slot)
379 {
380 	kmem_cache_free(mm_slot_cache, mm_slot);
381 }
382 
383 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
384 {
385 	struct mm_slot *mm_slot;
386 
387 	hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
388 		if (mm == mm_slot->mm)
389 			return mm_slot;
390 
391 	return NULL;
392 }
393 
394 static void insert_to_mm_slots_hash(struct mm_struct *mm,
395 				    struct mm_slot *mm_slot)
396 {
397 	mm_slot->mm = mm;
398 	hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
399 }
400 
401 static inline int khugepaged_test_exit(struct mm_struct *mm)
402 {
403 	return atomic_read(&mm->mm_users) == 0;
404 }
405 
406 static bool hugepage_vma_check(struct vm_area_struct *vma,
407 			       unsigned long vm_flags)
408 {
409 	if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
410 	    (vm_flags & VM_NOHUGEPAGE) ||
411 	    test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
412 		return false;
413 
414 	if (shmem_file(vma->vm_file) ||
415 	    (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
416 	     vma->vm_file &&
417 	     (vm_flags & VM_DENYWRITE))) {
418 		return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
419 				HPAGE_PMD_NR);
420 	}
421 	if (!vma->anon_vma || vma->vm_ops)
422 		return false;
423 	if (vma_is_temporary_stack(vma))
424 		return false;
425 	return !(vm_flags & VM_NO_KHUGEPAGED);
426 }
427 
428 int __khugepaged_enter(struct mm_struct *mm)
429 {
430 	struct mm_slot *mm_slot;
431 	int wakeup;
432 
433 	mm_slot = alloc_mm_slot();
434 	if (!mm_slot)
435 		return -ENOMEM;
436 
437 	/* __khugepaged_exit() must not run from under us */
438 	VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
439 	if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
440 		free_mm_slot(mm_slot);
441 		return 0;
442 	}
443 
444 	spin_lock(&khugepaged_mm_lock);
445 	insert_to_mm_slots_hash(mm, mm_slot);
446 	/*
447 	 * Insert just behind the scanning cursor, to let the area settle
448 	 * down a little.
449 	 */
450 	wakeup = list_empty(&khugepaged_scan.mm_head);
451 	list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
452 	spin_unlock(&khugepaged_mm_lock);
453 
454 	mmgrab(mm);
455 	if (wakeup)
456 		wake_up_interruptible(&khugepaged_wait);
457 
458 	return 0;
459 }
460 
461 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
462 			       unsigned long vm_flags)
463 {
464 	unsigned long hstart, hend;
465 
466 	/*
467 	 * khugepaged only supports read-only files for non-shmem files.
468 	 * khugepaged does not yet work on special mappings. And
469 	 * file-private shmem THP is not supported.
470 	 */
471 	if (!hugepage_vma_check(vma, vm_flags))
472 		return 0;
473 
474 	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
475 	hend = vma->vm_end & HPAGE_PMD_MASK;
476 	if (hstart < hend)
477 		return khugepaged_enter(vma, vm_flags);
478 	return 0;
479 }
480 
481 void __khugepaged_exit(struct mm_struct *mm)
482 {
483 	struct mm_slot *mm_slot;
484 	int free = 0;
485 
486 	spin_lock(&khugepaged_mm_lock);
487 	mm_slot = get_mm_slot(mm);
488 	if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
489 		hash_del(&mm_slot->hash);
490 		list_del(&mm_slot->mm_node);
491 		free = 1;
492 	}
493 	spin_unlock(&khugepaged_mm_lock);
494 
495 	if (free) {
496 		clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
497 		free_mm_slot(mm_slot);
498 		mmdrop(mm);
499 	} else if (mm_slot) {
500 		/*
501 		 * This is required to serialize against
502 		 * khugepaged_test_exit() (which is guaranteed to run
503 		 * under mmap sem read mode). Stop here (after we
504 		 * return all pagetables will be destroyed) until
505 		 * khugepaged has finished working on the pagetables
506 		 * under the mmap_sem.
507 		 */
508 		down_write(&mm->mmap_sem);
509 		up_write(&mm->mmap_sem);
510 	}
511 }
512 
513 static void release_pte_page(struct page *page)
514 {
515 	dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_lru(page));
516 	unlock_page(page);
517 	putback_lru_page(page);
518 }
519 
520 static void release_pte_pages(pte_t *pte, pte_t *_pte)
521 {
522 	while (--_pte >= pte) {
523 		pte_t pteval = *_pte;
524 		if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
525 			release_pte_page(pte_page(pteval));
526 	}
527 }
528 
529 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
530 					unsigned long address,
531 					pte_t *pte)
532 {
533 	struct page *page = NULL;
534 	pte_t *_pte;
535 	int none_or_zero = 0, result = 0, referenced = 0;
536 	bool writable = false;
537 
538 	for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
539 	     _pte++, address += PAGE_SIZE) {
540 		pte_t pteval = *_pte;
541 		if (pte_none(pteval) || (pte_present(pteval) &&
542 				is_zero_pfn(pte_pfn(pteval)))) {
543 			if (!userfaultfd_armed(vma) &&
544 			    ++none_or_zero <= khugepaged_max_ptes_none) {
545 				continue;
546 			} else {
547 				result = SCAN_EXCEED_NONE_PTE;
548 				goto out;
549 			}
550 		}
551 		if (!pte_present(pteval)) {
552 			result = SCAN_PTE_NON_PRESENT;
553 			goto out;
554 		}
555 		page = vm_normal_page(vma, address, pteval);
556 		if (unlikely(!page)) {
557 			result = SCAN_PAGE_NULL;
558 			goto out;
559 		}
560 
561 		/* TODO: teach khugepaged to collapse THP mapped with pte */
562 		if (PageCompound(page)) {
563 			result = SCAN_PAGE_COMPOUND;
564 			goto out;
565 		}
566 
567 		VM_BUG_ON_PAGE(!PageAnon(page), page);
568 
569 		/*
570 		 * We can do it before isolate_lru_page because the
571 		 * page can't be freed from under us. NOTE: PG_lock
572 		 * is needed to serialize against split_huge_page
573 		 * when invoked from the VM.
574 		 */
575 		if (!trylock_page(page)) {
576 			result = SCAN_PAGE_LOCK;
577 			goto out;
578 		}
579 
580 		/*
581 		 * cannot use mapcount: can't collapse if there's a gup pin.
582 		 * The page must only be referenced by the scanned process
583 		 * and page swap cache.
584 		 */
585 		if (page_count(page) != 1 + PageSwapCache(page)) {
586 			unlock_page(page);
587 			result = SCAN_PAGE_COUNT;
588 			goto out;
589 		}
590 		if (pte_write(pteval)) {
591 			writable = true;
592 		} else {
593 			if (PageSwapCache(page) &&
594 			    !reuse_swap_page(page, NULL)) {
595 				unlock_page(page);
596 				result = SCAN_SWAP_CACHE_PAGE;
597 				goto out;
598 			}
599 			/*
600 			 * Page is not in the swap cache. It can be collapsed
601 			 * into a THP.
602 			 */
603 		}
604 
605 		/*
606 		 * Isolate the page to avoid collapsing an hugepage
607 		 * currently in use by the VM.
608 		 */
609 		if (isolate_lru_page(page)) {
610 			unlock_page(page);
611 			result = SCAN_DEL_PAGE_LRU;
612 			goto out;
613 		}
614 		inc_node_page_state(page,
615 				NR_ISOLATED_ANON + page_is_file_lru(page));
616 		VM_BUG_ON_PAGE(!PageLocked(page), page);
617 		VM_BUG_ON_PAGE(PageLRU(page), page);
618 
619 		/* There should be enough young pte to collapse the page */
620 		if (pte_young(pteval) ||
621 		    page_is_young(page) || PageReferenced(page) ||
622 		    mmu_notifier_test_young(vma->vm_mm, address))
623 			referenced++;
624 	}
625 	if (likely(writable)) {
626 		if (likely(referenced)) {
627 			result = SCAN_SUCCEED;
628 			trace_mm_collapse_huge_page_isolate(page, none_or_zero,
629 							    referenced, writable, result);
630 			return 1;
631 		}
632 	} else {
633 		result = SCAN_PAGE_RO;
634 	}
635 
636 out:
637 	release_pte_pages(pte, _pte);
638 	trace_mm_collapse_huge_page_isolate(page, none_or_zero,
639 					    referenced, writable, result);
640 	return 0;
641 }
642 
643 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
644 				      struct vm_area_struct *vma,
645 				      unsigned long address,
646 				      spinlock_t *ptl)
647 {
648 	pte_t *_pte;
649 	for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
650 				_pte++, page++, address += PAGE_SIZE) {
651 		pte_t pteval = *_pte;
652 		struct page *src_page;
653 
654 		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
655 			clear_user_highpage(page, address);
656 			add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
657 			if (is_zero_pfn(pte_pfn(pteval))) {
658 				/*
659 				 * ptl mostly unnecessary.
660 				 */
661 				spin_lock(ptl);
662 				/*
663 				 * paravirt calls inside pte_clear here are
664 				 * superfluous.
665 				 */
666 				pte_clear(vma->vm_mm, address, _pte);
667 				spin_unlock(ptl);
668 			}
669 		} else {
670 			src_page = pte_page(pteval);
671 			copy_user_highpage(page, src_page, address, vma);
672 			VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
673 			release_pte_page(src_page);
674 			/*
675 			 * ptl mostly unnecessary, but preempt has to
676 			 * be disabled to update the per-cpu stats
677 			 * inside page_remove_rmap().
678 			 */
679 			spin_lock(ptl);
680 			/*
681 			 * paravirt calls inside pte_clear here are
682 			 * superfluous.
683 			 */
684 			pte_clear(vma->vm_mm, address, _pte);
685 			page_remove_rmap(src_page, false);
686 			spin_unlock(ptl);
687 			free_page_and_swap_cache(src_page);
688 		}
689 	}
690 }
691 
692 static void khugepaged_alloc_sleep(void)
693 {
694 	DEFINE_WAIT(wait);
695 
696 	add_wait_queue(&khugepaged_wait, &wait);
697 	freezable_schedule_timeout_interruptible(
698 		msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
699 	remove_wait_queue(&khugepaged_wait, &wait);
700 }
701 
702 static int khugepaged_node_load[MAX_NUMNODES];
703 
704 static bool khugepaged_scan_abort(int nid)
705 {
706 	int i;
707 
708 	/*
709 	 * If node_reclaim_mode is disabled, then no extra effort is made to
710 	 * allocate memory locally.
711 	 */
712 	if (!node_reclaim_mode)
713 		return false;
714 
715 	/* If there is a count for this node already, it must be acceptable */
716 	if (khugepaged_node_load[nid])
717 		return false;
718 
719 	for (i = 0; i < MAX_NUMNODES; i++) {
720 		if (!khugepaged_node_load[i])
721 			continue;
722 		if (node_distance(nid, i) > node_reclaim_distance)
723 			return true;
724 	}
725 	return false;
726 }
727 
728 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
729 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
730 {
731 	return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
732 }
733 
734 #ifdef CONFIG_NUMA
735 static int khugepaged_find_target_node(void)
736 {
737 	static int last_khugepaged_target_node = NUMA_NO_NODE;
738 	int nid, target_node = 0, max_value = 0;
739 
740 	/* find first node with max normal pages hit */
741 	for (nid = 0; nid < MAX_NUMNODES; nid++)
742 		if (khugepaged_node_load[nid] > max_value) {
743 			max_value = khugepaged_node_load[nid];
744 			target_node = nid;
745 		}
746 
747 	/* do some balance if several nodes have the same hit record */
748 	if (target_node <= last_khugepaged_target_node)
749 		for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
750 				nid++)
751 			if (max_value == khugepaged_node_load[nid]) {
752 				target_node = nid;
753 				break;
754 			}
755 
756 	last_khugepaged_target_node = target_node;
757 	return target_node;
758 }
759 
760 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
761 {
762 	if (IS_ERR(*hpage)) {
763 		if (!*wait)
764 			return false;
765 
766 		*wait = false;
767 		*hpage = NULL;
768 		khugepaged_alloc_sleep();
769 	} else if (*hpage) {
770 		put_page(*hpage);
771 		*hpage = NULL;
772 	}
773 
774 	return true;
775 }
776 
777 static struct page *
778 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
779 {
780 	VM_BUG_ON_PAGE(*hpage, *hpage);
781 
782 	*hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
783 	if (unlikely(!*hpage)) {
784 		count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
785 		*hpage = ERR_PTR(-ENOMEM);
786 		return NULL;
787 	}
788 
789 	prep_transhuge_page(*hpage);
790 	count_vm_event(THP_COLLAPSE_ALLOC);
791 	return *hpage;
792 }
793 #else
794 static int khugepaged_find_target_node(void)
795 {
796 	return 0;
797 }
798 
799 static inline struct page *alloc_khugepaged_hugepage(void)
800 {
801 	struct page *page;
802 
803 	page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
804 			   HPAGE_PMD_ORDER);
805 	if (page)
806 		prep_transhuge_page(page);
807 	return page;
808 }
809 
810 static struct page *khugepaged_alloc_hugepage(bool *wait)
811 {
812 	struct page *hpage;
813 
814 	do {
815 		hpage = alloc_khugepaged_hugepage();
816 		if (!hpage) {
817 			count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
818 			if (!*wait)
819 				return NULL;
820 
821 			*wait = false;
822 			khugepaged_alloc_sleep();
823 		} else
824 			count_vm_event(THP_COLLAPSE_ALLOC);
825 	} while (unlikely(!hpage) && likely(khugepaged_enabled()));
826 
827 	return hpage;
828 }
829 
830 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
831 {
832 	if (!*hpage)
833 		*hpage = khugepaged_alloc_hugepage(wait);
834 
835 	if (unlikely(!*hpage))
836 		return false;
837 
838 	return true;
839 }
840 
841 static struct page *
842 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
843 {
844 	VM_BUG_ON(!*hpage);
845 
846 	return  *hpage;
847 }
848 #endif
849 
850 /*
851  * If mmap_sem temporarily dropped, revalidate vma
852  * before taking mmap_sem.
853  * Return 0 if succeeds, otherwise return none-zero
854  * value (scan code).
855  */
856 
857 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
858 		struct vm_area_struct **vmap)
859 {
860 	struct vm_area_struct *vma;
861 	unsigned long hstart, hend;
862 
863 	if (unlikely(khugepaged_test_exit(mm)))
864 		return SCAN_ANY_PROCESS;
865 
866 	*vmap = vma = find_vma(mm, address);
867 	if (!vma)
868 		return SCAN_VMA_NULL;
869 
870 	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
871 	hend = vma->vm_end & HPAGE_PMD_MASK;
872 	if (address < hstart || address + HPAGE_PMD_SIZE > hend)
873 		return SCAN_ADDRESS_RANGE;
874 	if (!hugepage_vma_check(vma, vma->vm_flags))
875 		return SCAN_VMA_CHECK;
876 	return 0;
877 }
878 
879 /*
880  * Bring missing pages in from swap, to complete THP collapse.
881  * Only done if khugepaged_scan_pmd believes it is worthwhile.
882  *
883  * Called and returns without pte mapped or spinlocks held,
884  * but with mmap_sem held to protect against vma changes.
885  */
886 
887 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
888 					struct vm_area_struct *vma,
889 					unsigned long address, pmd_t *pmd,
890 					int referenced)
891 {
892 	int swapped_in = 0;
893 	vm_fault_t ret = 0;
894 	struct vm_fault vmf = {
895 		.vma = vma,
896 		.address = address,
897 		.flags = FAULT_FLAG_ALLOW_RETRY,
898 		.pmd = pmd,
899 		.pgoff = linear_page_index(vma, address),
900 	};
901 
902 	/* we only decide to swapin, if there is enough young ptes */
903 	if (referenced < HPAGE_PMD_NR/2) {
904 		trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
905 		return false;
906 	}
907 	vmf.pte = pte_offset_map(pmd, address);
908 	for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
909 			vmf.pte++, vmf.address += PAGE_SIZE) {
910 		vmf.orig_pte = *vmf.pte;
911 		if (!is_swap_pte(vmf.orig_pte))
912 			continue;
913 		swapped_in++;
914 		ret = do_swap_page(&vmf);
915 
916 		/* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
917 		if (ret & VM_FAULT_RETRY) {
918 			down_read(&mm->mmap_sem);
919 			if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
920 				/* vma is no longer available, don't continue to swapin */
921 				trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
922 				return false;
923 			}
924 			/* check if the pmd is still valid */
925 			if (mm_find_pmd(mm, address) != pmd) {
926 				trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
927 				return false;
928 			}
929 		}
930 		if (ret & VM_FAULT_ERROR) {
931 			trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
932 			return false;
933 		}
934 		/* pte is unmapped now, we need to map it */
935 		vmf.pte = pte_offset_map(pmd, vmf.address);
936 	}
937 	vmf.pte--;
938 	pte_unmap(vmf.pte);
939 	trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
940 	return true;
941 }
942 
943 static void collapse_huge_page(struct mm_struct *mm,
944 				   unsigned long address,
945 				   struct page **hpage,
946 				   int node, int referenced)
947 {
948 	pmd_t *pmd, _pmd;
949 	pte_t *pte;
950 	pgtable_t pgtable;
951 	struct page *new_page;
952 	spinlock_t *pmd_ptl, *pte_ptl;
953 	int isolated = 0, result = 0;
954 	struct mem_cgroup *memcg;
955 	struct vm_area_struct *vma;
956 	struct mmu_notifier_range range;
957 	gfp_t gfp;
958 
959 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
960 
961 	/* Only allocate from the target node */
962 	gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
963 
964 	/*
965 	 * Before allocating the hugepage, release the mmap_sem read lock.
966 	 * The allocation can take potentially a long time if it involves
967 	 * sync compaction, and we do not need to hold the mmap_sem during
968 	 * that. We will recheck the vma after taking it again in write mode.
969 	 */
970 	up_read(&mm->mmap_sem);
971 	new_page = khugepaged_alloc_page(hpage, gfp, node);
972 	if (!new_page) {
973 		result = SCAN_ALLOC_HUGE_PAGE_FAIL;
974 		goto out_nolock;
975 	}
976 
977 	if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
978 		result = SCAN_CGROUP_CHARGE_FAIL;
979 		goto out_nolock;
980 	}
981 
982 	down_read(&mm->mmap_sem);
983 	result = hugepage_vma_revalidate(mm, address, &vma);
984 	if (result) {
985 		mem_cgroup_cancel_charge(new_page, memcg, true);
986 		up_read(&mm->mmap_sem);
987 		goto out_nolock;
988 	}
989 
990 	pmd = mm_find_pmd(mm, address);
991 	if (!pmd) {
992 		result = SCAN_PMD_NULL;
993 		mem_cgroup_cancel_charge(new_page, memcg, true);
994 		up_read(&mm->mmap_sem);
995 		goto out_nolock;
996 	}
997 
998 	/*
999 	 * __collapse_huge_page_swapin always returns with mmap_sem locked.
1000 	 * If it fails, we release mmap_sem and jump out_nolock.
1001 	 * Continuing to collapse causes inconsistency.
1002 	 */
1003 	if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
1004 		mem_cgroup_cancel_charge(new_page, memcg, true);
1005 		up_read(&mm->mmap_sem);
1006 		goto out_nolock;
1007 	}
1008 
1009 	up_read(&mm->mmap_sem);
1010 	/*
1011 	 * Prevent all access to pagetables with the exception of
1012 	 * gup_fast later handled by the ptep_clear_flush and the VM
1013 	 * handled by the anon_vma lock + PG_lock.
1014 	 */
1015 	down_write(&mm->mmap_sem);
1016 	result = SCAN_ANY_PROCESS;
1017 	if (!mmget_still_valid(mm))
1018 		goto out;
1019 	result = hugepage_vma_revalidate(mm, address, &vma);
1020 	if (result)
1021 		goto out;
1022 	/* check if the pmd is still valid */
1023 	if (mm_find_pmd(mm, address) != pmd)
1024 		goto out;
1025 
1026 	anon_vma_lock_write(vma->anon_vma);
1027 
1028 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1029 				address, address + HPAGE_PMD_SIZE);
1030 	mmu_notifier_invalidate_range_start(&range);
1031 
1032 	pte = pte_offset_map(pmd, address);
1033 	pte_ptl = pte_lockptr(mm, pmd);
1034 
1035 	pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1036 	/*
1037 	 * After this gup_fast can't run anymore. This also removes
1038 	 * any huge TLB entry from the CPU so we won't allow
1039 	 * huge and small TLB entries for the same virtual address
1040 	 * to avoid the risk of CPU bugs in that area.
1041 	 */
1042 	_pmd = pmdp_collapse_flush(vma, address, pmd);
1043 	spin_unlock(pmd_ptl);
1044 	mmu_notifier_invalidate_range_end(&range);
1045 
1046 	spin_lock(pte_ptl);
1047 	isolated = __collapse_huge_page_isolate(vma, address, pte);
1048 	spin_unlock(pte_ptl);
1049 
1050 	if (unlikely(!isolated)) {
1051 		pte_unmap(pte);
1052 		spin_lock(pmd_ptl);
1053 		BUG_ON(!pmd_none(*pmd));
1054 		/*
1055 		 * We can only use set_pmd_at when establishing
1056 		 * hugepmds and never for establishing regular pmds that
1057 		 * points to regular pagetables. Use pmd_populate for that
1058 		 */
1059 		pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1060 		spin_unlock(pmd_ptl);
1061 		anon_vma_unlock_write(vma->anon_vma);
1062 		result = SCAN_FAIL;
1063 		goto out;
1064 	}
1065 
1066 	/*
1067 	 * All pages are isolated and locked so anon_vma rmap
1068 	 * can't run anymore.
1069 	 */
1070 	anon_vma_unlock_write(vma->anon_vma);
1071 
1072 	__collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1073 	pte_unmap(pte);
1074 	__SetPageUptodate(new_page);
1075 	pgtable = pmd_pgtable(_pmd);
1076 
1077 	_pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1078 	_pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1079 
1080 	/*
1081 	 * spin_lock() below is not the equivalent of smp_wmb(), so
1082 	 * this is needed to avoid the copy_huge_page writes to become
1083 	 * visible after the set_pmd_at() write.
1084 	 */
1085 	smp_wmb();
1086 
1087 	spin_lock(pmd_ptl);
1088 	BUG_ON(!pmd_none(*pmd));
1089 	page_add_new_anon_rmap(new_page, vma, address, true);
1090 	mem_cgroup_commit_charge(new_page, memcg, false, true);
1091 	count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1092 	lru_cache_add_active_or_unevictable(new_page, vma);
1093 	pgtable_trans_huge_deposit(mm, pmd, pgtable);
1094 	set_pmd_at(mm, address, pmd, _pmd);
1095 	update_mmu_cache_pmd(vma, address, pmd);
1096 	spin_unlock(pmd_ptl);
1097 
1098 	*hpage = NULL;
1099 
1100 	khugepaged_pages_collapsed++;
1101 	result = SCAN_SUCCEED;
1102 out_up_write:
1103 	up_write(&mm->mmap_sem);
1104 out_nolock:
1105 	trace_mm_collapse_huge_page(mm, isolated, result);
1106 	return;
1107 out:
1108 	mem_cgroup_cancel_charge(new_page, memcg, true);
1109 	goto out_up_write;
1110 }
1111 
1112 static int khugepaged_scan_pmd(struct mm_struct *mm,
1113 			       struct vm_area_struct *vma,
1114 			       unsigned long address,
1115 			       struct page **hpage)
1116 {
1117 	pmd_t *pmd;
1118 	pte_t *pte, *_pte;
1119 	int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1120 	struct page *page = NULL;
1121 	unsigned long _address;
1122 	spinlock_t *ptl;
1123 	int node = NUMA_NO_NODE, unmapped = 0;
1124 	bool writable = false;
1125 
1126 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1127 
1128 	pmd = mm_find_pmd(mm, address);
1129 	if (!pmd) {
1130 		result = SCAN_PMD_NULL;
1131 		goto out;
1132 	}
1133 
1134 	memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1135 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1136 	for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1137 	     _pte++, _address += PAGE_SIZE) {
1138 		pte_t pteval = *_pte;
1139 		if (is_swap_pte(pteval)) {
1140 			if (++unmapped <= khugepaged_max_ptes_swap) {
1141 				/*
1142 				 * Always be strict with uffd-wp
1143 				 * enabled swap entries.  Please see
1144 				 * comment below for pte_uffd_wp().
1145 				 */
1146 				if (pte_swp_uffd_wp(pteval)) {
1147 					result = SCAN_PTE_UFFD_WP;
1148 					goto out_unmap;
1149 				}
1150 				continue;
1151 			} else {
1152 				result = SCAN_EXCEED_SWAP_PTE;
1153 				goto out_unmap;
1154 			}
1155 		}
1156 		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1157 			if (!userfaultfd_armed(vma) &&
1158 			    ++none_or_zero <= khugepaged_max_ptes_none) {
1159 				continue;
1160 			} else {
1161 				result = SCAN_EXCEED_NONE_PTE;
1162 				goto out_unmap;
1163 			}
1164 		}
1165 		if (!pte_present(pteval)) {
1166 			result = SCAN_PTE_NON_PRESENT;
1167 			goto out_unmap;
1168 		}
1169 		if (pte_uffd_wp(pteval)) {
1170 			/*
1171 			 * Don't collapse the page if any of the small
1172 			 * PTEs are armed with uffd write protection.
1173 			 * Here we can also mark the new huge pmd as
1174 			 * write protected if any of the small ones is
1175 			 * marked but that could bring uknown
1176 			 * userfault messages that falls outside of
1177 			 * the registered range.  So, just be simple.
1178 			 */
1179 			result = SCAN_PTE_UFFD_WP;
1180 			goto out_unmap;
1181 		}
1182 		if (pte_write(pteval))
1183 			writable = true;
1184 
1185 		page = vm_normal_page(vma, _address, pteval);
1186 		if (unlikely(!page)) {
1187 			result = SCAN_PAGE_NULL;
1188 			goto out_unmap;
1189 		}
1190 
1191 		/* TODO: teach khugepaged to collapse THP mapped with pte */
1192 		if (PageCompound(page)) {
1193 			result = SCAN_PAGE_COMPOUND;
1194 			goto out_unmap;
1195 		}
1196 
1197 		/*
1198 		 * Record which node the original page is from and save this
1199 		 * information to khugepaged_node_load[].
1200 		 * Khupaged will allocate hugepage from the node has the max
1201 		 * hit record.
1202 		 */
1203 		node = page_to_nid(page);
1204 		if (khugepaged_scan_abort(node)) {
1205 			result = SCAN_SCAN_ABORT;
1206 			goto out_unmap;
1207 		}
1208 		khugepaged_node_load[node]++;
1209 		if (!PageLRU(page)) {
1210 			result = SCAN_PAGE_LRU;
1211 			goto out_unmap;
1212 		}
1213 		if (PageLocked(page)) {
1214 			result = SCAN_PAGE_LOCK;
1215 			goto out_unmap;
1216 		}
1217 		if (!PageAnon(page)) {
1218 			result = SCAN_PAGE_ANON;
1219 			goto out_unmap;
1220 		}
1221 
1222 		/*
1223 		 * cannot use mapcount: can't collapse if there's a gup pin.
1224 		 * The page must only be referenced by the scanned process
1225 		 * and page swap cache.
1226 		 */
1227 		if (page_count(page) != 1 + PageSwapCache(page)) {
1228 			result = SCAN_PAGE_COUNT;
1229 			goto out_unmap;
1230 		}
1231 		if (pte_young(pteval) ||
1232 		    page_is_young(page) || PageReferenced(page) ||
1233 		    mmu_notifier_test_young(vma->vm_mm, address))
1234 			referenced++;
1235 	}
1236 	if (writable) {
1237 		if (referenced) {
1238 			result = SCAN_SUCCEED;
1239 			ret = 1;
1240 		} else {
1241 			result = SCAN_LACK_REFERENCED_PAGE;
1242 		}
1243 	} else {
1244 		result = SCAN_PAGE_RO;
1245 	}
1246 out_unmap:
1247 	pte_unmap_unlock(pte, ptl);
1248 	if (ret) {
1249 		node = khugepaged_find_target_node();
1250 		/* collapse_huge_page will return with the mmap_sem released */
1251 		collapse_huge_page(mm, address, hpage, node, referenced);
1252 	}
1253 out:
1254 	trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1255 				     none_or_zero, result, unmapped);
1256 	return ret;
1257 }
1258 
1259 static void collect_mm_slot(struct mm_slot *mm_slot)
1260 {
1261 	struct mm_struct *mm = mm_slot->mm;
1262 
1263 	lockdep_assert_held(&khugepaged_mm_lock);
1264 
1265 	if (khugepaged_test_exit(mm)) {
1266 		/* free mm_slot */
1267 		hash_del(&mm_slot->hash);
1268 		list_del(&mm_slot->mm_node);
1269 
1270 		/*
1271 		 * Not strictly needed because the mm exited already.
1272 		 *
1273 		 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1274 		 */
1275 
1276 		/* khugepaged_mm_lock actually not necessary for the below */
1277 		free_mm_slot(mm_slot);
1278 		mmdrop(mm);
1279 	}
1280 }
1281 
1282 #ifdef CONFIG_SHMEM
1283 /*
1284  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1285  * khugepaged should try to collapse the page table.
1286  */
1287 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1288 					 unsigned long addr)
1289 {
1290 	struct mm_slot *mm_slot;
1291 
1292 	VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1293 
1294 	spin_lock(&khugepaged_mm_lock);
1295 	mm_slot = get_mm_slot(mm);
1296 	if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1297 		mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1298 	spin_unlock(&khugepaged_mm_lock);
1299 	return 0;
1300 }
1301 
1302 /**
1303  * Try to collapse a pte-mapped THP for mm at address haddr.
1304  *
1305  * This function checks whether all the PTEs in the PMD are pointing to the
1306  * right THP. If so, retract the page table so the THP can refault in with
1307  * as pmd-mapped.
1308  */
1309 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1310 {
1311 	unsigned long haddr = addr & HPAGE_PMD_MASK;
1312 	struct vm_area_struct *vma = find_vma(mm, haddr);
1313 	struct page *hpage = NULL;
1314 	pte_t *start_pte, *pte;
1315 	pmd_t *pmd, _pmd;
1316 	spinlock_t *ptl;
1317 	int count = 0;
1318 	int i;
1319 
1320 	if (!vma || !vma->vm_file ||
1321 	    vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1322 		return;
1323 
1324 	/*
1325 	 * This vm_flags may not have VM_HUGEPAGE if the page was not
1326 	 * collapsed by this mm. But we can still collapse if the page is
1327 	 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1328 	 * will not fail the vma for missing VM_HUGEPAGE
1329 	 */
1330 	if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1331 		return;
1332 
1333 	pmd = mm_find_pmd(mm, haddr);
1334 	if (!pmd)
1335 		return;
1336 
1337 	start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1338 
1339 	/* step 1: check all mapped PTEs are to the right huge page */
1340 	for (i = 0, addr = haddr, pte = start_pte;
1341 	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1342 		struct page *page;
1343 
1344 		/* empty pte, skip */
1345 		if (pte_none(*pte))
1346 			continue;
1347 
1348 		/* page swapped out, abort */
1349 		if (!pte_present(*pte))
1350 			goto abort;
1351 
1352 		page = vm_normal_page(vma, addr, *pte);
1353 
1354 		if (!page || !PageCompound(page))
1355 			goto abort;
1356 
1357 		if (!hpage) {
1358 			hpage = compound_head(page);
1359 			/*
1360 			 * The mapping of the THP should not change.
1361 			 *
1362 			 * Note that uprobe, debugger, or MAP_PRIVATE may
1363 			 * change the page table, but the new page will
1364 			 * not pass PageCompound() check.
1365 			 */
1366 			if (WARN_ON(hpage->mapping != vma->vm_file->f_mapping))
1367 				goto abort;
1368 		}
1369 
1370 		/*
1371 		 * Confirm the page maps to the correct subpage.
1372 		 *
1373 		 * Note that uprobe, debugger, or MAP_PRIVATE may change
1374 		 * the page table, but the new page will not pass
1375 		 * PageCompound() check.
1376 		 */
1377 		if (WARN_ON(hpage + i != page))
1378 			goto abort;
1379 		count++;
1380 	}
1381 
1382 	/* step 2: adjust rmap */
1383 	for (i = 0, addr = haddr, pte = start_pte;
1384 	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1385 		struct page *page;
1386 
1387 		if (pte_none(*pte))
1388 			continue;
1389 		page = vm_normal_page(vma, addr, *pte);
1390 		page_remove_rmap(page, false);
1391 	}
1392 
1393 	pte_unmap_unlock(start_pte, ptl);
1394 
1395 	/* step 3: set proper refcount and mm_counters. */
1396 	if (hpage) {
1397 		page_ref_sub(hpage, count);
1398 		add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1399 	}
1400 
1401 	/* step 4: collapse pmd */
1402 	ptl = pmd_lock(vma->vm_mm, pmd);
1403 	_pmd = pmdp_collapse_flush(vma, addr, pmd);
1404 	spin_unlock(ptl);
1405 	mm_dec_nr_ptes(mm);
1406 	pte_free(mm, pmd_pgtable(_pmd));
1407 	return;
1408 
1409 abort:
1410 	pte_unmap_unlock(start_pte, ptl);
1411 }
1412 
1413 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1414 {
1415 	struct mm_struct *mm = mm_slot->mm;
1416 	int i;
1417 
1418 	if (likely(mm_slot->nr_pte_mapped_thp == 0))
1419 		return 0;
1420 
1421 	if (!down_write_trylock(&mm->mmap_sem))
1422 		return -EBUSY;
1423 
1424 	if (unlikely(khugepaged_test_exit(mm)))
1425 		goto out;
1426 
1427 	for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1428 		collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1429 
1430 out:
1431 	mm_slot->nr_pte_mapped_thp = 0;
1432 	up_write(&mm->mmap_sem);
1433 	return 0;
1434 }
1435 
1436 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1437 {
1438 	struct vm_area_struct *vma;
1439 	unsigned long addr;
1440 	pmd_t *pmd, _pmd;
1441 
1442 	i_mmap_lock_write(mapping);
1443 	vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1444 		/*
1445 		 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1446 		 * got written to. These VMAs are likely not worth investing
1447 		 * down_write(mmap_sem) as PMD-mapping is likely to be split
1448 		 * later.
1449 		 *
1450 		 * Not that vma->anon_vma check is racy: it can be set up after
1451 		 * the check but before we took mmap_sem by the fault path.
1452 		 * But page lock would prevent establishing any new ptes of the
1453 		 * page, so we are safe.
1454 		 *
1455 		 * An alternative would be drop the check, but check that page
1456 		 * table is clear before calling pmdp_collapse_flush() under
1457 		 * ptl. It has higher chance to recover THP for the VMA, but
1458 		 * has higher cost too.
1459 		 */
1460 		if (vma->anon_vma)
1461 			continue;
1462 		addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1463 		if (addr & ~HPAGE_PMD_MASK)
1464 			continue;
1465 		if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1466 			continue;
1467 		pmd = mm_find_pmd(vma->vm_mm, addr);
1468 		if (!pmd)
1469 			continue;
1470 		/*
1471 		 * We need exclusive mmap_sem to retract page table.
1472 		 *
1473 		 * We use trylock due to lock inversion: we need to acquire
1474 		 * mmap_sem while holding page lock. Fault path does it in
1475 		 * reverse order. Trylock is a way to avoid deadlock.
1476 		 */
1477 		if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1478 			spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1479 			/* assume page table is clear */
1480 			_pmd = pmdp_collapse_flush(vma, addr, pmd);
1481 			spin_unlock(ptl);
1482 			up_write(&vma->vm_mm->mmap_sem);
1483 			mm_dec_nr_ptes(vma->vm_mm);
1484 			pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1485 		} else {
1486 			/* Try again later */
1487 			khugepaged_add_pte_mapped_thp(vma->vm_mm, addr);
1488 		}
1489 	}
1490 	i_mmap_unlock_write(mapping);
1491 }
1492 
1493 /**
1494  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1495  *
1496  * Basic scheme is simple, details are more complex:
1497  *  - allocate and lock a new huge page;
1498  *  - scan page cache replacing old pages with the new one
1499  *    + swap/gup in pages if necessary;
1500  *    + fill in gaps;
1501  *    + keep old pages around in case rollback is required;
1502  *  - if replacing succeeds:
1503  *    + copy data over;
1504  *    + free old pages;
1505  *    + unlock huge page;
1506  *  - if replacing failed;
1507  *    + put all pages back and unfreeze them;
1508  *    + restore gaps in the page cache;
1509  *    + unlock and free huge page;
1510  */
1511 static void collapse_file(struct mm_struct *mm,
1512 		struct file *file, pgoff_t start,
1513 		struct page **hpage, int node)
1514 {
1515 	struct address_space *mapping = file->f_mapping;
1516 	gfp_t gfp;
1517 	struct page *new_page;
1518 	struct mem_cgroup *memcg;
1519 	pgoff_t index, end = start + HPAGE_PMD_NR;
1520 	LIST_HEAD(pagelist);
1521 	XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1522 	int nr_none = 0, result = SCAN_SUCCEED;
1523 	bool is_shmem = shmem_file(file);
1524 
1525 	VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1526 	VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1527 
1528 	/* Only allocate from the target node */
1529 	gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1530 
1531 	new_page = khugepaged_alloc_page(hpage, gfp, node);
1532 	if (!new_page) {
1533 		result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1534 		goto out;
1535 	}
1536 
1537 	if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1538 		result = SCAN_CGROUP_CHARGE_FAIL;
1539 		goto out;
1540 	}
1541 
1542 	/* This will be less messy when we use multi-index entries */
1543 	do {
1544 		xas_lock_irq(&xas);
1545 		xas_create_range(&xas);
1546 		if (!xas_error(&xas))
1547 			break;
1548 		xas_unlock_irq(&xas);
1549 		if (!xas_nomem(&xas, GFP_KERNEL)) {
1550 			mem_cgroup_cancel_charge(new_page, memcg, true);
1551 			result = SCAN_FAIL;
1552 			goto out;
1553 		}
1554 	} while (1);
1555 
1556 	__SetPageLocked(new_page);
1557 	if (is_shmem)
1558 		__SetPageSwapBacked(new_page);
1559 	new_page->index = start;
1560 	new_page->mapping = mapping;
1561 
1562 	/*
1563 	 * At this point the new_page is locked and not up-to-date.
1564 	 * It's safe to insert it into the page cache, because nobody would
1565 	 * be able to map it or use it in another way until we unlock it.
1566 	 */
1567 
1568 	xas_set(&xas, start);
1569 	for (index = start; index < end; index++) {
1570 		struct page *page = xas_next(&xas);
1571 
1572 		VM_BUG_ON(index != xas.xa_index);
1573 		if (is_shmem) {
1574 			if (!page) {
1575 				/*
1576 				 * Stop if extent has been truncated or
1577 				 * hole-punched, and is now completely
1578 				 * empty.
1579 				 */
1580 				if (index == start) {
1581 					if (!xas_next_entry(&xas, end - 1)) {
1582 						result = SCAN_TRUNCATED;
1583 						goto xa_locked;
1584 					}
1585 					xas_set(&xas, index);
1586 				}
1587 				if (!shmem_charge(mapping->host, 1)) {
1588 					result = SCAN_FAIL;
1589 					goto xa_locked;
1590 				}
1591 				xas_store(&xas, new_page);
1592 				nr_none++;
1593 				continue;
1594 			}
1595 
1596 			if (xa_is_value(page) || !PageUptodate(page)) {
1597 				xas_unlock_irq(&xas);
1598 				/* swap in or instantiate fallocated page */
1599 				if (shmem_getpage(mapping->host, index, &page,
1600 						  SGP_NOHUGE)) {
1601 					result = SCAN_FAIL;
1602 					goto xa_unlocked;
1603 				}
1604 			} else if (trylock_page(page)) {
1605 				get_page(page);
1606 				xas_unlock_irq(&xas);
1607 			} else {
1608 				result = SCAN_PAGE_LOCK;
1609 				goto xa_locked;
1610 			}
1611 		} else {	/* !is_shmem */
1612 			if (!page || xa_is_value(page)) {
1613 				xas_unlock_irq(&xas);
1614 				page_cache_sync_readahead(mapping, &file->f_ra,
1615 							  file, index,
1616 							  PAGE_SIZE);
1617 				/* drain pagevecs to help isolate_lru_page() */
1618 				lru_add_drain();
1619 				page = find_lock_page(mapping, index);
1620 				if (unlikely(page == NULL)) {
1621 					result = SCAN_FAIL;
1622 					goto xa_unlocked;
1623 				}
1624 			} else if (PageDirty(page)) {
1625 				/*
1626 				 * khugepaged only works on read-only fd,
1627 				 * so this page is dirty because it hasn't
1628 				 * been flushed since first write. There
1629 				 * won't be new dirty pages.
1630 				 *
1631 				 * Trigger async flush here and hope the
1632 				 * writeback is done when khugepaged
1633 				 * revisits this page.
1634 				 *
1635 				 * This is a one-off situation. We are not
1636 				 * forcing writeback in loop.
1637 				 */
1638 				xas_unlock_irq(&xas);
1639 				filemap_flush(mapping);
1640 				result = SCAN_FAIL;
1641 				goto xa_unlocked;
1642 			} else if (trylock_page(page)) {
1643 				get_page(page);
1644 				xas_unlock_irq(&xas);
1645 			} else {
1646 				result = SCAN_PAGE_LOCK;
1647 				goto xa_locked;
1648 			}
1649 		}
1650 
1651 		/*
1652 		 * The page must be locked, so we can drop the i_pages lock
1653 		 * without racing with truncate.
1654 		 */
1655 		VM_BUG_ON_PAGE(!PageLocked(page), page);
1656 
1657 		/* make sure the page is up to date */
1658 		if (unlikely(!PageUptodate(page))) {
1659 			result = SCAN_FAIL;
1660 			goto out_unlock;
1661 		}
1662 
1663 		/*
1664 		 * If file was truncated then extended, or hole-punched, before
1665 		 * we locked the first page, then a THP might be there already.
1666 		 */
1667 		if (PageTransCompound(page)) {
1668 			result = SCAN_PAGE_COMPOUND;
1669 			goto out_unlock;
1670 		}
1671 
1672 		if (page_mapping(page) != mapping) {
1673 			result = SCAN_TRUNCATED;
1674 			goto out_unlock;
1675 		}
1676 
1677 		if (!is_shmem && PageDirty(page)) {
1678 			/*
1679 			 * khugepaged only works on read-only fd, so this
1680 			 * page is dirty because it hasn't been flushed
1681 			 * since first write.
1682 			 */
1683 			result = SCAN_FAIL;
1684 			goto out_unlock;
1685 		}
1686 
1687 		if (isolate_lru_page(page)) {
1688 			result = SCAN_DEL_PAGE_LRU;
1689 			goto out_unlock;
1690 		}
1691 
1692 		if (page_has_private(page) &&
1693 		    !try_to_release_page(page, GFP_KERNEL)) {
1694 			result = SCAN_PAGE_HAS_PRIVATE;
1695 			goto out_unlock;
1696 		}
1697 
1698 		if (page_mapped(page))
1699 			unmap_mapping_pages(mapping, index, 1, false);
1700 
1701 		xas_lock_irq(&xas);
1702 		xas_set(&xas, index);
1703 
1704 		VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1705 		VM_BUG_ON_PAGE(page_mapped(page), page);
1706 
1707 		/*
1708 		 * The page is expected to have page_count() == 3:
1709 		 *  - we hold a pin on it;
1710 		 *  - one reference from page cache;
1711 		 *  - one from isolate_lru_page;
1712 		 */
1713 		if (!page_ref_freeze(page, 3)) {
1714 			result = SCAN_PAGE_COUNT;
1715 			xas_unlock_irq(&xas);
1716 			putback_lru_page(page);
1717 			goto out_unlock;
1718 		}
1719 
1720 		/*
1721 		 * Add the page to the list to be able to undo the collapse if
1722 		 * something go wrong.
1723 		 */
1724 		list_add_tail(&page->lru, &pagelist);
1725 
1726 		/* Finally, replace with the new page. */
1727 		xas_store(&xas, new_page);
1728 		continue;
1729 out_unlock:
1730 		unlock_page(page);
1731 		put_page(page);
1732 		goto xa_unlocked;
1733 	}
1734 
1735 	if (is_shmem)
1736 		__inc_node_page_state(new_page, NR_SHMEM_THPS);
1737 	else {
1738 		__inc_node_page_state(new_page, NR_FILE_THPS);
1739 		filemap_nr_thps_inc(mapping);
1740 	}
1741 
1742 	if (nr_none) {
1743 		struct zone *zone = page_zone(new_page);
1744 
1745 		__mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1746 		if (is_shmem)
1747 			__mod_node_page_state(zone->zone_pgdat,
1748 					      NR_SHMEM, nr_none);
1749 	}
1750 
1751 xa_locked:
1752 	xas_unlock_irq(&xas);
1753 xa_unlocked:
1754 
1755 	if (result == SCAN_SUCCEED) {
1756 		struct page *page, *tmp;
1757 
1758 		/*
1759 		 * Replacing old pages with new one has succeeded, now we
1760 		 * need to copy the content and free the old pages.
1761 		 */
1762 		index = start;
1763 		list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1764 			while (index < page->index) {
1765 				clear_highpage(new_page + (index % HPAGE_PMD_NR));
1766 				index++;
1767 			}
1768 			copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1769 					page);
1770 			list_del(&page->lru);
1771 			page->mapping = NULL;
1772 			page_ref_unfreeze(page, 1);
1773 			ClearPageActive(page);
1774 			ClearPageUnevictable(page);
1775 			unlock_page(page);
1776 			put_page(page);
1777 			index++;
1778 		}
1779 		while (index < end) {
1780 			clear_highpage(new_page + (index % HPAGE_PMD_NR));
1781 			index++;
1782 		}
1783 
1784 		SetPageUptodate(new_page);
1785 		page_ref_add(new_page, HPAGE_PMD_NR - 1);
1786 		mem_cgroup_commit_charge(new_page, memcg, false, true);
1787 
1788 		if (is_shmem) {
1789 			set_page_dirty(new_page);
1790 			lru_cache_add_anon(new_page);
1791 		} else {
1792 			lru_cache_add_file(new_page);
1793 		}
1794 		count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1795 
1796 		/*
1797 		 * Remove pte page tables, so we can re-fault the page as huge.
1798 		 */
1799 		retract_page_tables(mapping, start);
1800 		*hpage = NULL;
1801 
1802 		khugepaged_pages_collapsed++;
1803 	} else {
1804 		struct page *page;
1805 
1806 		/* Something went wrong: roll back page cache changes */
1807 		xas_lock_irq(&xas);
1808 		mapping->nrpages -= nr_none;
1809 
1810 		if (is_shmem)
1811 			shmem_uncharge(mapping->host, nr_none);
1812 
1813 		xas_set(&xas, start);
1814 		xas_for_each(&xas, page, end - 1) {
1815 			page = list_first_entry_or_null(&pagelist,
1816 					struct page, lru);
1817 			if (!page || xas.xa_index < page->index) {
1818 				if (!nr_none)
1819 					break;
1820 				nr_none--;
1821 				/* Put holes back where they were */
1822 				xas_store(&xas, NULL);
1823 				continue;
1824 			}
1825 
1826 			VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1827 
1828 			/* Unfreeze the page. */
1829 			list_del(&page->lru);
1830 			page_ref_unfreeze(page, 2);
1831 			xas_store(&xas, page);
1832 			xas_pause(&xas);
1833 			xas_unlock_irq(&xas);
1834 			unlock_page(page);
1835 			putback_lru_page(page);
1836 			xas_lock_irq(&xas);
1837 		}
1838 		VM_BUG_ON(nr_none);
1839 		xas_unlock_irq(&xas);
1840 
1841 		mem_cgroup_cancel_charge(new_page, memcg, true);
1842 		new_page->mapping = NULL;
1843 	}
1844 
1845 	unlock_page(new_page);
1846 out:
1847 	VM_BUG_ON(!list_empty(&pagelist));
1848 	/* TODO: tracepoints */
1849 }
1850 
1851 static void khugepaged_scan_file(struct mm_struct *mm,
1852 		struct file *file, pgoff_t start, struct page **hpage)
1853 {
1854 	struct page *page = NULL;
1855 	struct address_space *mapping = file->f_mapping;
1856 	XA_STATE(xas, &mapping->i_pages, start);
1857 	int present, swap;
1858 	int node = NUMA_NO_NODE;
1859 	int result = SCAN_SUCCEED;
1860 
1861 	present = 0;
1862 	swap = 0;
1863 	memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1864 	rcu_read_lock();
1865 	xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1866 		if (xas_retry(&xas, page))
1867 			continue;
1868 
1869 		if (xa_is_value(page)) {
1870 			if (++swap > khugepaged_max_ptes_swap) {
1871 				result = SCAN_EXCEED_SWAP_PTE;
1872 				break;
1873 			}
1874 			continue;
1875 		}
1876 
1877 		if (PageTransCompound(page)) {
1878 			result = SCAN_PAGE_COMPOUND;
1879 			break;
1880 		}
1881 
1882 		node = page_to_nid(page);
1883 		if (khugepaged_scan_abort(node)) {
1884 			result = SCAN_SCAN_ABORT;
1885 			break;
1886 		}
1887 		khugepaged_node_load[node]++;
1888 
1889 		if (!PageLRU(page)) {
1890 			result = SCAN_PAGE_LRU;
1891 			break;
1892 		}
1893 
1894 		if (page_count(page) !=
1895 		    1 + page_mapcount(page) + page_has_private(page)) {
1896 			result = SCAN_PAGE_COUNT;
1897 			break;
1898 		}
1899 
1900 		/*
1901 		 * We probably should check if the page is referenced here, but
1902 		 * nobody would transfer pte_young() to PageReferenced() for us.
1903 		 * And rmap walk here is just too costly...
1904 		 */
1905 
1906 		present++;
1907 
1908 		if (need_resched()) {
1909 			xas_pause(&xas);
1910 			cond_resched_rcu();
1911 		}
1912 	}
1913 	rcu_read_unlock();
1914 
1915 	if (result == SCAN_SUCCEED) {
1916 		if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1917 			result = SCAN_EXCEED_NONE_PTE;
1918 		} else {
1919 			node = khugepaged_find_target_node();
1920 			collapse_file(mm, file, start, hpage, node);
1921 		}
1922 	}
1923 
1924 	/* TODO: tracepoints */
1925 }
1926 #else
1927 static void khugepaged_scan_file(struct mm_struct *mm,
1928 		struct file *file, pgoff_t start, struct page **hpage)
1929 {
1930 	BUILD_BUG();
1931 }
1932 
1933 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1934 {
1935 	return 0;
1936 }
1937 #endif
1938 
1939 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1940 					    struct page **hpage)
1941 	__releases(&khugepaged_mm_lock)
1942 	__acquires(&khugepaged_mm_lock)
1943 {
1944 	struct mm_slot *mm_slot;
1945 	struct mm_struct *mm;
1946 	struct vm_area_struct *vma;
1947 	int progress = 0;
1948 
1949 	VM_BUG_ON(!pages);
1950 	lockdep_assert_held(&khugepaged_mm_lock);
1951 
1952 	if (khugepaged_scan.mm_slot)
1953 		mm_slot = khugepaged_scan.mm_slot;
1954 	else {
1955 		mm_slot = list_entry(khugepaged_scan.mm_head.next,
1956 				     struct mm_slot, mm_node);
1957 		khugepaged_scan.address = 0;
1958 		khugepaged_scan.mm_slot = mm_slot;
1959 	}
1960 	spin_unlock(&khugepaged_mm_lock);
1961 	khugepaged_collapse_pte_mapped_thps(mm_slot);
1962 
1963 	mm = mm_slot->mm;
1964 	/*
1965 	 * Don't wait for semaphore (to avoid long wait times).  Just move to
1966 	 * the next mm on the list.
1967 	 */
1968 	vma = NULL;
1969 	if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1970 		goto breakouterloop_mmap_sem;
1971 	if (likely(!khugepaged_test_exit(mm)))
1972 		vma = find_vma(mm, khugepaged_scan.address);
1973 
1974 	progress++;
1975 	for (; vma; vma = vma->vm_next) {
1976 		unsigned long hstart, hend;
1977 
1978 		cond_resched();
1979 		if (unlikely(khugepaged_test_exit(mm))) {
1980 			progress++;
1981 			break;
1982 		}
1983 		if (!hugepage_vma_check(vma, vma->vm_flags)) {
1984 skip:
1985 			progress++;
1986 			continue;
1987 		}
1988 		hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1989 		hend = vma->vm_end & HPAGE_PMD_MASK;
1990 		if (hstart >= hend)
1991 			goto skip;
1992 		if (khugepaged_scan.address > hend)
1993 			goto skip;
1994 		if (khugepaged_scan.address < hstart)
1995 			khugepaged_scan.address = hstart;
1996 		VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1997 		if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
1998 			goto skip;
1999 
2000 		while (khugepaged_scan.address < hend) {
2001 			int ret;
2002 			cond_resched();
2003 			if (unlikely(khugepaged_test_exit(mm)))
2004 				goto breakouterloop;
2005 
2006 			VM_BUG_ON(khugepaged_scan.address < hstart ||
2007 				  khugepaged_scan.address + HPAGE_PMD_SIZE >
2008 				  hend);
2009 			if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2010 				struct file *file = get_file(vma->vm_file);
2011 				pgoff_t pgoff = linear_page_index(vma,
2012 						khugepaged_scan.address);
2013 
2014 				up_read(&mm->mmap_sem);
2015 				ret = 1;
2016 				khugepaged_scan_file(mm, file, pgoff, hpage);
2017 				fput(file);
2018 			} else {
2019 				ret = khugepaged_scan_pmd(mm, vma,
2020 						khugepaged_scan.address,
2021 						hpage);
2022 			}
2023 			/* move to next address */
2024 			khugepaged_scan.address += HPAGE_PMD_SIZE;
2025 			progress += HPAGE_PMD_NR;
2026 			if (ret)
2027 				/* we released mmap_sem so break loop */
2028 				goto breakouterloop_mmap_sem;
2029 			if (progress >= pages)
2030 				goto breakouterloop;
2031 		}
2032 	}
2033 breakouterloop:
2034 	up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2035 breakouterloop_mmap_sem:
2036 
2037 	spin_lock(&khugepaged_mm_lock);
2038 	VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2039 	/*
2040 	 * Release the current mm_slot if this mm is about to die, or
2041 	 * if we scanned all vmas of this mm.
2042 	 */
2043 	if (khugepaged_test_exit(mm) || !vma) {
2044 		/*
2045 		 * Make sure that if mm_users is reaching zero while
2046 		 * khugepaged runs here, khugepaged_exit will find
2047 		 * mm_slot not pointing to the exiting mm.
2048 		 */
2049 		if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2050 			khugepaged_scan.mm_slot = list_entry(
2051 				mm_slot->mm_node.next,
2052 				struct mm_slot, mm_node);
2053 			khugepaged_scan.address = 0;
2054 		} else {
2055 			khugepaged_scan.mm_slot = NULL;
2056 			khugepaged_full_scans++;
2057 		}
2058 
2059 		collect_mm_slot(mm_slot);
2060 	}
2061 
2062 	return progress;
2063 }
2064 
2065 static int khugepaged_has_work(void)
2066 {
2067 	return !list_empty(&khugepaged_scan.mm_head) &&
2068 		khugepaged_enabled();
2069 }
2070 
2071 static int khugepaged_wait_event(void)
2072 {
2073 	return !list_empty(&khugepaged_scan.mm_head) ||
2074 		kthread_should_stop();
2075 }
2076 
2077 static void khugepaged_do_scan(void)
2078 {
2079 	struct page *hpage = NULL;
2080 	unsigned int progress = 0, pass_through_head = 0;
2081 	unsigned int pages = khugepaged_pages_to_scan;
2082 	bool wait = true;
2083 
2084 	barrier(); /* write khugepaged_pages_to_scan to local stack */
2085 
2086 	while (progress < pages) {
2087 		if (!khugepaged_prealloc_page(&hpage, &wait))
2088 			break;
2089 
2090 		cond_resched();
2091 
2092 		if (unlikely(kthread_should_stop() || try_to_freeze()))
2093 			break;
2094 
2095 		spin_lock(&khugepaged_mm_lock);
2096 		if (!khugepaged_scan.mm_slot)
2097 			pass_through_head++;
2098 		if (khugepaged_has_work() &&
2099 		    pass_through_head < 2)
2100 			progress += khugepaged_scan_mm_slot(pages - progress,
2101 							    &hpage);
2102 		else
2103 			progress = pages;
2104 		spin_unlock(&khugepaged_mm_lock);
2105 	}
2106 
2107 	if (!IS_ERR_OR_NULL(hpage))
2108 		put_page(hpage);
2109 }
2110 
2111 static bool khugepaged_should_wakeup(void)
2112 {
2113 	return kthread_should_stop() ||
2114 	       time_after_eq(jiffies, khugepaged_sleep_expire);
2115 }
2116 
2117 static void khugepaged_wait_work(void)
2118 {
2119 	if (khugepaged_has_work()) {
2120 		const unsigned long scan_sleep_jiffies =
2121 			msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2122 
2123 		if (!scan_sleep_jiffies)
2124 			return;
2125 
2126 		khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2127 		wait_event_freezable_timeout(khugepaged_wait,
2128 					     khugepaged_should_wakeup(),
2129 					     scan_sleep_jiffies);
2130 		return;
2131 	}
2132 
2133 	if (khugepaged_enabled())
2134 		wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2135 }
2136 
2137 static int khugepaged(void *none)
2138 {
2139 	struct mm_slot *mm_slot;
2140 
2141 	set_freezable();
2142 	set_user_nice(current, MAX_NICE);
2143 
2144 	while (!kthread_should_stop()) {
2145 		khugepaged_do_scan();
2146 		khugepaged_wait_work();
2147 	}
2148 
2149 	spin_lock(&khugepaged_mm_lock);
2150 	mm_slot = khugepaged_scan.mm_slot;
2151 	khugepaged_scan.mm_slot = NULL;
2152 	if (mm_slot)
2153 		collect_mm_slot(mm_slot);
2154 	spin_unlock(&khugepaged_mm_lock);
2155 	return 0;
2156 }
2157 
2158 static void set_recommended_min_free_kbytes(void)
2159 {
2160 	struct zone *zone;
2161 	int nr_zones = 0;
2162 	unsigned long recommended_min;
2163 
2164 	for_each_populated_zone(zone) {
2165 		/*
2166 		 * We don't need to worry about fragmentation of
2167 		 * ZONE_MOVABLE since it only has movable pages.
2168 		 */
2169 		if (zone_idx(zone) > gfp_zone(GFP_USER))
2170 			continue;
2171 
2172 		nr_zones++;
2173 	}
2174 
2175 	/* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2176 	recommended_min = pageblock_nr_pages * nr_zones * 2;
2177 
2178 	/*
2179 	 * Make sure that on average at least two pageblocks are almost free
2180 	 * of another type, one for a migratetype to fall back to and a
2181 	 * second to avoid subsequent fallbacks of other types There are 3
2182 	 * MIGRATE_TYPES we care about.
2183 	 */
2184 	recommended_min += pageblock_nr_pages * nr_zones *
2185 			   MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2186 
2187 	/* don't ever allow to reserve more than 5% of the lowmem */
2188 	recommended_min = min(recommended_min,
2189 			      (unsigned long) nr_free_buffer_pages() / 20);
2190 	recommended_min <<= (PAGE_SHIFT-10);
2191 
2192 	if (recommended_min > min_free_kbytes) {
2193 		if (user_min_free_kbytes >= 0)
2194 			pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2195 				min_free_kbytes, recommended_min);
2196 
2197 		min_free_kbytes = recommended_min;
2198 	}
2199 	setup_per_zone_wmarks();
2200 }
2201 
2202 int start_stop_khugepaged(void)
2203 {
2204 	static struct task_struct *khugepaged_thread __read_mostly;
2205 	static DEFINE_MUTEX(khugepaged_mutex);
2206 	int err = 0;
2207 
2208 	mutex_lock(&khugepaged_mutex);
2209 	if (khugepaged_enabled()) {
2210 		if (!khugepaged_thread)
2211 			khugepaged_thread = kthread_run(khugepaged, NULL,
2212 							"khugepaged");
2213 		if (IS_ERR(khugepaged_thread)) {
2214 			pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2215 			err = PTR_ERR(khugepaged_thread);
2216 			khugepaged_thread = NULL;
2217 			goto fail;
2218 		}
2219 
2220 		if (!list_empty(&khugepaged_scan.mm_head))
2221 			wake_up_interruptible(&khugepaged_wait);
2222 
2223 		set_recommended_min_free_kbytes();
2224 	} else if (khugepaged_thread) {
2225 		kthread_stop(khugepaged_thread);
2226 		khugepaged_thread = NULL;
2227 	}
2228 fail:
2229 	mutex_unlock(&khugepaged_mutex);
2230 	return err;
2231 }
2232