Merge tag 'x86_seves_fixes_for_v5.10_rc1' of git://git.kernel.org/pub/scm/linux/kerne...
[platform/kernel/linux-rpi.git] / mm / khugepaged.c
1 // SPDX-License-Identifier: GPL-2.0
2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
3
4 #include <linux/mm.h>
5 #include <linux/sched.h>
6 #include <linux/sched/mm.h>
7 #include <linux/sched/coredump.h>
8 #include <linux/mmu_notifier.h>
9 #include <linux/rmap.h>
10 #include <linux/swap.h>
11 #include <linux/mm_inline.h>
12 #include <linux/kthread.h>
13 #include <linux/khugepaged.h>
14 #include <linux/freezer.h>
15 #include <linux/mman.h>
16 #include <linux/hashtable.h>
17 #include <linux/userfaultfd_k.h>
18 #include <linux/page_idle.h>
19 #include <linux/swapops.h>
20 #include <linux/shmem_fs.h>
21
22 #include <asm/tlb.h>
23 #include <asm/pgalloc.h>
24 #include "internal.h"
25
26 enum scan_result {
27         SCAN_FAIL,
28         SCAN_SUCCEED,
29         SCAN_PMD_NULL,
30         SCAN_EXCEED_NONE_PTE,
31         SCAN_EXCEED_SWAP_PTE,
32         SCAN_EXCEED_SHARED_PTE,
33         SCAN_PTE_NON_PRESENT,
34         SCAN_PTE_UFFD_WP,
35         SCAN_PAGE_RO,
36         SCAN_LACK_REFERENCED_PAGE,
37         SCAN_PAGE_NULL,
38         SCAN_SCAN_ABORT,
39         SCAN_PAGE_COUNT,
40         SCAN_PAGE_LRU,
41         SCAN_PAGE_LOCK,
42         SCAN_PAGE_ANON,
43         SCAN_PAGE_COMPOUND,
44         SCAN_ANY_PROCESS,
45         SCAN_VMA_NULL,
46         SCAN_VMA_CHECK,
47         SCAN_ADDRESS_RANGE,
48         SCAN_SWAP_CACHE_PAGE,
49         SCAN_DEL_PAGE_LRU,
50         SCAN_ALLOC_HUGE_PAGE_FAIL,
51         SCAN_CGROUP_CHARGE_FAIL,
52         SCAN_TRUNCATED,
53         SCAN_PAGE_HAS_PRIVATE,
54 };
55
56 #define CREATE_TRACE_POINTS
57 #include <trace/events/huge_memory.h>
58
59 static struct task_struct *khugepaged_thread __read_mostly;
60 static DEFINE_MUTEX(khugepaged_mutex);
61
62 /* default scan 8*512 pte (or vmas) every 30 second */
63 static unsigned int khugepaged_pages_to_scan __read_mostly;
64 static unsigned int khugepaged_pages_collapsed;
65 static unsigned int khugepaged_full_scans;
66 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
67 /* during fragmentation poll the hugepage allocator once every minute */
68 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
69 static unsigned long khugepaged_sleep_expire;
70 static DEFINE_SPINLOCK(khugepaged_mm_lock);
71 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
72 /*
73  * default collapse hugepages if there is at least one pte mapped like
74  * it would have happened if the vma was large enough during page
75  * fault.
76  */
77 static unsigned int khugepaged_max_ptes_none __read_mostly;
78 static unsigned int khugepaged_max_ptes_swap __read_mostly;
79 static unsigned int khugepaged_max_ptes_shared __read_mostly;
80
81 #define MM_SLOTS_HASH_BITS 10
82 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
83
84 static struct kmem_cache *mm_slot_cache __read_mostly;
85
86 #define MAX_PTE_MAPPED_THP 8
87
88 /**
89  * struct mm_slot - hash lookup from mm to mm_slot
90  * @hash: hash collision list
91  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
92  * @mm: the mm that this information is valid for
93  */
94 struct mm_slot {
95         struct hlist_node hash;
96         struct list_head mm_node;
97         struct mm_struct *mm;
98
99         /* pte-mapped THP in this mm */
100         int nr_pte_mapped_thp;
101         unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
102 };
103
104 /**
105  * struct khugepaged_scan - cursor for scanning
106  * @mm_head: the head of the mm list to scan
107  * @mm_slot: the current mm_slot we are scanning
108  * @address: the next address inside that to be scanned
109  *
110  * There is only the one khugepaged_scan instance of this cursor structure.
111  */
112 struct khugepaged_scan {
113         struct list_head mm_head;
114         struct mm_slot *mm_slot;
115         unsigned long address;
116 };
117
118 static struct khugepaged_scan khugepaged_scan = {
119         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
120 };
121
122 #ifdef CONFIG_SYSFS
123 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
124                                          struct kobj_attribute *attr,
125                                          char *buf)
126 {
127         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
128 }
129
130 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
131                                           struct kobj_attribute *attr,
132                                           const char *buf, size_t count)
133 {
134         unsigned long msecs;
135         int err;
136
137         err = kstrtoul(buf, 10, &msecs);
138         if (err || msecs > UINT_MAX)
139                 return -EINVAL;
140
141         khugepaged_scan_sleep_millisecs = msecs;
142         khugepaged_sleep_expire = 0;
143         wake_up_interruptible(&khugepaged_wait);
144
145         return count;
146 }
147 static struct kobj_attribute scan_sleep_millisecs_attr =
148         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
149                scan_sleep_millisecs_store);
150
151 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
152                                           struct kobj_attribute *attr,
153                                           char *buf)
154 {
155         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
156 }
157
158 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
159                                            struct kobj_attribute *attr,
160                                            const char *buf, size_t count)
161 {
162         unsigned long msecs;
163         int err;
164
165         err = kstrtoul(buf, 10, &msecs);
166         if (err || msecs > UINT_MAX)
167                 return -EINVAL;
168
169         khugepaged_alloc_sleep_millisecs = msecs;
170         khugepaged_sleep_expire = 0;
171         wake_up_interruptible(&khugepaged_wait);
172
173         return count;
174 }
175 static struct kobj_attribute alloc_sleep_millisecs_attr =
176         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
177                alloc_sleep_millisecs_store);
178
179 static ssize_t pages_to_scan_show(struct kobject *kobj,
180                                   struct kobj_attribute *attr,
181                                   char *buf)
182 {
183         return sprintf(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         int err;
190         unsigned long pages;
191
192         err = kstrtoul(buf, 10, &pages);
193         if (err || !pages || pages > UINT_MAX)
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(pages_to_scan, 0644, pages_to_scan_show,
202                pages_to_scan_store);
203
204 static ssize_t pages_collapsed_show(struct kobject *kobj,
205                                     struct kobj_attribute *attr,
206                                     char *buf)
207 {
208         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
209 }
210 static struct kobj_attribute pages_collapsed_attr =
211         __ATTR_RO(pages_collapsed);
212
213 static ssize_t full_scans_show(struct kobject *kobj,
214                                struct kobj_attribute *attr,
215                                char *buf)
216 {
217         return sprintf(buf, "%u\n", khugepaged_full_scans);
218 }
219 static struct kobj_attribute full_scans_attr =
220         __ATTR_RO(full_scans);
221
222 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
223                                       struct kobj_attribute *attr, char *buf)
224 {
225         return single_hugepage_flag_show(kobj, attr, buf,
226                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
227 }
228 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
229                                        struct kobj_attribute *attr,
230                                        const char *buf, size_t count)
231 {
232         return single_hugepage_flag_store(kobj, attr, buf, count,
233                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
234 }
235 static struct kobj_attribute khugepaged_defrag_attr =
236         __ATTR(defrag, 0644, khugepaged_defrag_show,
237                khugepaged_defrag_store);
238
239 /*
240  * max_ptes_none controls if khugepaged should collapse hugepages over
241  * any unmapped ptes in turn potentially increasing the memory
242  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
243  * reduce the available free memory in the system as it
244  * runs. Increasing max_ptes_none will instead potentially reduce the
245  * free memory in the system during the khugepaged scan.
246  */
247 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
248                                              struct kobj_attribute *attr,
249                                              char *buf)
250 {
251         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
252 }
253 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
254                                               struct kobj_attribute *attr,
255                                               const char *buf, size_t count)
256 {
257         int err;
258         unsigned long max_ptes_none;
259
260         err = kstrtoul(buf, 10, &max_ptes_none);
261         if (err || max_ptes_none > HPAGE_PMD_NR-1)
262                 return -EINVAL;
263
264         khugepaged_max_ptes_none = max_ptes_none;
265
266         return count;
267 }
268 static struct kobj_attribute khugepaged_max_ptes_none_attr =
269         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
270                khugepaged_max_ptes_none_store);
271
272 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
273                                              struct kobj_attribute *attr,
274                                              char *buf)
275 {
276         return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
277 }
278
279 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
280                                               struct kobj_attribute *attr,
281                                               const char *buf, size_t count)
282 {
283         int err;
284         unsigned long max_ptes_swap;
285
286         err  = kstrtoul(buf, 10, &max_ptes_swap);
287         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
288                 return -EINVAL;
289
290         khugepaged_max_ptes_swap = max_ptes_swap;
291
292         return count;
293 }
294
295 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
296         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
297                khugepaged_max_ptes_swap_store);
298
299 static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
300                                              struct kobj_attribute *attr,
301                                              char *buf)
302 {
303         return sprintf(buf, "%u\n", khugepaged_max_ptes_shared);
304 }
305
306 static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
307                                               struct kobj_attribute *attr,
308                                               const char *buf, size_t count)
309 {
310         int err;
311         unsigned long max_ptes_shared;
312
313         err  = kstrtoul(buf, 10, &max_ptes_shared);
314         if (err || max_ptes_shared > HPAGE_PMD_NR-1)
315                 return -EINVAL;
316
317         khugepaged_max_ptes_shared = max_ptes_shared;
318
319         return count;
320 }
321
322 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
323         __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
324                khugepaged_max_ptes_shared_store);
325
326 static struct attribute *khugepaged_attr[] = {
327         &khugepaged_defrag_attr.attr,
328         &khugepaged_max_ptes_none_attr.attr,
329         &khugepaged_max_ptes_swap_attr.attr,
330         &khugepaged_max_ptes_shared_attr.attr,
331         &pages_to_scan_attr.attr,
332         &pages_collapsed_attr.attr,
333         &full_scans_attr.attr,
334         &scan_sleep_millisecs_attr.attr,
335         &alloc_sleep_millisecs_attr.attr,
336         NULL,
337 };
338
339 struct attribute_group khugepaged_attr_group = {
340         .attrs = khugepaged_attr,
341         .name = "khugepaged",
342 };
343 #endif /* CONFIG_SYSFS */
344
345 int hugepage_madvise(struct vm_area_struct *vma,
346                      unsigned long *vm_flags, int advice)
347 {
348         switch (advice) {
349         case MADV_HUGEPAGE:
350 #ifdef CONFIG_S390
351                 /*
352                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
353                  * can't handle this properly after s390_enable_sie, so we simply
354                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
355                  */
356                 if (mm_has_pgste(vma->vm_mm))
357                         return 0;
358 #endif
359                 *vm_flags &= ~VM_NOHUGEPAGE;
360                 *vm_flags |= VM_HUGEPAGE;
361                 /*
362                  * If the vma become good for khugepaged to scan,
363                  * register it here without waiting a page fault that
364                  * may not happen any time soon.
365                  */
366                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
367                                 khugepaged_enter_vma_merge(vma, *vm_flags))
368                         return -ENOMEM;
369                 break;
370         case MADV_NOHUGEPAGE:
371                 *vm_flags &= ~VM_HUGEPAGE;
372                 *vm_flags |= VM_NOHUGEPAGE;
373                 /*
374                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
375                  * this vma even if we leave the mm registered in khugepaged if
376                  * it got registered before VM_NOHUGEPAGE was set.
377                  */
378                 break;
379         }
380
381         return 0;
382 }
383
384 int __init khugepaged_init(void)
385 {
386         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
387                                           sizeof(struct mm_slot),
388                                           __alignof__(struct mm_slot), 0, NULL);
389         if (!mm_slot_cache)
390                 return -ENOMEM;
391
392         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
393         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
394         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
395         khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
396
397         return 0;
398 }
399
400 void __init khugepaged_destroy(void)
401 {
402         kmem_cache_destroy(mm_slot_cache);
403 }
404
405 static inline struct mm_slot *alloc_mm_slot(void)
406 {
407         if (!mm_slot_cache)     /* initialization failed */
408                 return NULL;
409         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
410 }
411
412 static inline void free_mm_slot(struct mm_slot *mm_slot)
413 {
414         kmem_cache_free(mm_slot_cache, mm_slot);
415 }
416
417 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
418 {
419         struct mm_slot *mm_slot;
420
421         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
422                 if (mm == mm_slot->mm)
423                         return mm_slot;
424
425         return NULL;
426 }
427
428 static void insert_to_mm_slots_hash(struct mm_struct *mm,
429                                     struct mm_slot *mm_slot)
430 {
431         mm_slot->mm = mm;
432         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
433 }
434
435 static inline int khugepaged_test_exit(struct mm_struct *mm)
436 {
437         return atomic_read(&mm->mm_users) == 0;
438 }
439
440 static bool hugepage_vma_check(struct vm_area_struct *vma,
441                                unsigned long vm_flags)
442 {
443         if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
444             (vm_flags & VM_NOHUGEPAGE) ||
445             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
446                 return false;
447
448         if (shmem_file(vma->vm_file) ||
449             (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
450              vma->vm_file &&
451              (vm_flags & VM_DENYWRITE))) {
452                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
453                                 HPAGE_PMD_NR);
454         }
455         if (!vma->anon_vma || vma->vm_ops)
456                 return false;
457         if (vma_is_temporary_stack(vma))
458                 return false;
459         return !(vm_flags & VM_NO_KHUGEPAGED);
460 }
461
462 int __khugepaged_enter(struct mm_struct *mm)
463 {
464         struct mm_slot *mm_slot;
465         int wakeup;
466
467         mm_slot = alloc_mm_slot();
468         if (!mm_slot)
469                 return -ENOMEM;
470
471         /* __khugepaged_exit() must not run from under us */
472         VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
473         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
474                 free_mm_slot(mm_slot);
475                 return 0;
476         }
477
478         spin_lock(&khugepaged_mm_lock);
479         insert_to_mm_slots_hash(mm, mm_slot);
480         /*
481          * Insert just behind the scanning cursor, to let the area settle
482          * down a little.
483          */
484         wakeup = list_empty(&khugepaged_scan.mm_head);
485         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
486         spin_unlock(&khugepaged_mm_lock);
487
488         mmgrab(mm);
489         if (wakeup)
490                 wake_up_interruptible(&khugepaged_wait);
491
492         return 0;
493 }
494
495 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
496                                unsigned long vm_flags)
497 {
498         unsigned long hstart, hend;
499
500         /*
501          * khugepaged only supports read-only files for non-shmem files.
502          * khugepaged does not yet work on special mappings. And
503          * file-private shmem THP is not supported.
504          */
505         if (!hugepage_vma_check(vma, vm_flags))
506                 return 0;
507
508         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
509         hend = vma->vm_end & HPAGE_PMD_MASK;
510         if (hstart < hend)
511                 return khugepaged_enter(vma, vm_flags);
512         return 0;
513 }
514
515 void __khugepaged_exit(struct mm_struct *mm)
516 {
517         struct mm_slot *mm_slot;
518         int free = 0;
519
520         spin_lock(&khugepaged_mm_lock);
521         mm_slot = get_mm_slot(mm);
522         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
523                 hash_del(&mm_slot->hash);
524                 list_del(&mm_slot->mm_node);
525                 free = 1;
526         }
527         spin_unlock(&khugepaged_mm_lock);
528
529         if (free) {
530                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
531                 free_mm_slot(mm_slot);
532                 mmdrop(mm);
533         } else if (mm_slot) {
534                 /*
535                  * This is required to serialize against
536                  * khugepaged_test_exit() (which is guaranteed to run
537                  * under mmap sem read mode). Stop here (after we
538                  * return all pagetables will be destroyed) until
539                  * khugepaged has finished working on the pagetables
540                  * under the mmap_lock.
541                  */
542                 mmap_write_lock(mm);
543                 mmap_write_unlock(mm);
544         }
545 }
546
547 static void release_pte_page(struct page *page)
548 {
549         mod_node_page_state(page_pgdat(page),
550                         NR_ISOLATED_ANON + page_is_file_lru(page),
551                         -compound_nr(page));
552         unlock_page(page);
553         putback_lru_page(page);
554 }
555
556 static void release_pte_pages(pte_t *pte, pte_t *_pte,
557                 struct list_head *compound_pagelist)
558 {
559         struct page *page, *tmp;
560
561         while (--_pte >= pte) {
562                 pte_t pteval = *_pte;
563
564                 page = pte_page(pteval);
565                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
566                                 !PageCompound(page))
567                         release_pte_page(page);
568         }
569
570         list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
571                 list_del(&page->lru);
572                 release_pte_page(page);
573         }
574 }
575
576 static bool is_refcount_suitable(struct page *page)
577 {
578         int expected_refcount;
579
580         expected_refcount = total_mapcount(page);
581         if (PageSwapCache(page))
582                 expected_refcount += compound_nr(page);
583
584         return page_count(page) == expected_refcount;
585 }
586
587 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
588                                         unsigned long address,
589                                         pte_t *pte,
590                                         struct list_head *compound_pagelist)
591 {
592         struct page *page = NULL;
593         pte_t *_pte;
594         int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
595         bool writable = false;
596
597         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
598              _pte++, address += PAGE_SIZE) {
599                 pte_t pteval = *_pte;
600                 if (pte_none(pteval) || (pte_present(pteval) &&
601                                 is_zero_pfn(pte_pfn(pteval)))) {
602                         if (!userfaultfd_armed(vma) &&
603                             ++none_or_zero <= khugepaged_max_ptes_none) {
604                                 continue;
605                         } else {
606                                 result = 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)) {
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                         goto out;
626                 }
627
628                 if (PageCompound(page)) {
629                         struct page *p;
630                         page = compound_head(page);
631
632                         /*
633                          * Check if we have dealt with the compound page
634                          * already
635                          */
636                         list_for_each_entry(p, compound_pagelist, lru) {
637                                 if (page == p)
638                                         goto next;
639                         }
640                 }
641
642                 /*
643                  * We can do it before isolate_lru_page because the
644                  * page can't be freed from under us. NOTE: PG_lock
645                  * is needed to serialize against split_huge_page
646                  * when invoked from the VM.
647                  */
648                 if (!trylock_page(page)) {
649                         result = SCAN_PAGE_LOCK;
650                         goto out;
651                 }
652
653                 /*
654                  * Check if the page has any GUP (or other external) pins.
655                  *
656                  * The page table that maps the page has been already unlinked
657                  * from the page table tree and this process cannot get
658                  * an additinal pin on the page.
659                  *
660                  * New pins can come later if the page is shared across fork,
661                  * but not from this process. The other process cannot write to
662                  * the page, only trigger CoW.
663                  */
664                 if (!is_refcount_suitable(page)) {
665                         unlock_page(page);
666                         result = SCAN_PAGE_COUNT;
667                         goto out;
668                 }
669                 if (!pte_write(pteval) && PageSwapCache(page) &&
670                                 !reuse_swap_page(page, NULL)) {
671                         /*
672                          * Page is in the swap cache and cannot be re-used.
673                          * It cannot be collapsed into a THP.
674                          */
675                         unlock_page(page);
676                         result = SCAN_SWAP_CACHE_PAGE;
677                         goto out;
678                 }
679
680                 /*
681                  * Isolate the page to avoid collapsing an hugepage
682                  * currently in use by the VM.
683                  */
684                 if (isolate_lru_page(page)) {
685                         unlock_page(page);
686                         result = SCAN_DEL_PAGE_LRU;
687                         goto out;
688                 }
689                 mod_node_page_state(page_pgdat(page),
690                                 NR_ISOLATED_ANON + page_is_file_lru(page),
691                                 compound_nr(page));
692                 VM_BUG_ON_PAGE(!PageLocked(page), page);
693                 VM_BUG_ON_PAGE(PageLRU(page), page);
694
695                 if (PageCompound(page))
696                         list_add_tail(&page->lru, compound_pagelist);
697 next:
698                 /* There should be enough young pte to collapse the page */
699                 if (pte_young(pteval) ||
700                     page_is_young(page) || PageReferenced(page) ||
701                     mmu_notifier_test_young(vma->vm_mm, address))
702                         referenced++;
703
704                 if (pte_write(pteval))
705                         writable = true;
706         }
707         if (likely(writable)) {
708                 if (likely(referenced)) {
709                         result = SCAN_SUCCEED;
710                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
711                                                             referenced, writable, result);
712                         return 1;
713                 }
714         } else {
715                 result = SCAN_PAGE_RO;
716         }
717
718 out:
719         release_pte_pages(pte, _pte, compound_pagelist);
720         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
721                                             referenced, writable, result);
722         return 0;
723 }
724
725 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
726                                       struct vm_area_struct *vma,
727                                       unsigned long address,
728                                       spinlock_t *ptl,
729                                       struct list_head *compound_pagelist)
730 {
731         struct page *src_page, *tmp;
732         pte_t *_pte;
733         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
734                                 _pte++, page++, address += PAGE_SIZE) {
735                 pte_t pteval = *_pte;
736
737                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
738                         clear_user_highpage(page, address);
739                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
740                         if (is_zero_pfn(pte_pfn(pteval))) {
741                                 /*
742                                  * ptl mostly unnecessary.
743                                  */
744                                 spin_lock(ptl);
745                                 /*
746                                  * paravirt calls inside pte_clear here are
747                                  * superfluous.
748                                  */
749                                 pte_clear(vma->vm_mm, address, _pte);
750                                 spin_unlock(ptl);
751                         }
752                 } else {
753                         src_page = pte_page(pteval);
754                         copy_user_highpage(page, src_page, address, vma);
755                         if (!PageCompound(src_page))
756                                 release_pte_page(src_page);
757                         /*
758                          * ptl mostly unnecessary, but preempt has to
759                          * be disabled to update the per-cpu stats
760                          * inside page_remove_rmap().
761                          */
762                         spin_lock(ptl);
763                         /*
764                          * paravirt calls inside pte_clear here are
765                          * superfluous.
766                          */
767                         pte_clear(vma->vm_mm, address, _pte);
768                         page_remove_rmap(src_page, false);
769                         spin_unlock(ptl);
770                         free_page_and_swap_cache(src_page);
771                 }
772         }
773
774         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
775                 list_del(&src_page->lru);
776                 release_pte_page(src_page);
777         }
778 }
779
780 static void khugepaged_alloc_sleep(void)
781 {
782         DEFINE_WAIT(wait);
783
784         add_wait_queue(&khugepaged_wait, &wait);
785         freezable_schedule_timeout_interruptible(
786                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
787         remove_wait_queue(&khugepaged_wait, &wait);
788 }
789
790 static int khugepaged_node_load[MAX_NUMNODES];
791
792 static bool khugepaged_scan_abort(int nid)
793 {
794         int i;
795
796         /*
797          * If node_reclaim_mode is disabled, then no extra effort is made to
798          * allocate memory locally.
799          */
800         if (!node_reclaim_mode)
801                 return false;
802
803         /* If there is a count for this node already, it must be acceptable */
804         if (khugepaged_node_load[nid])
805                 return false;
806
807         for (i = 0; i < MAX_NUMNODES; i++) {
808                 if (!khugepaged_node_load[i])
809                         continue;
810                 if (node_distance(nid, i) > node_reclaim_distance)
811                         return true;
812         }
813         return false;
814 }
815
816 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
817 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
818 {
819         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
820 }
821
822 #ifdef CONFIG_NUMA
823 static int khugepaged_find_target_node(void)
824 {
825         static int last_khugepaged_target_node = NUMA_NO_NODE;
826         int nid, target_node = 0, max_value = 0;
827
828         /* find first node with max normal pages hit */
829         for (nid = 0; nid < MAX_NUMNODES; nid++)
830                 if (khugepaged_node_load[nid] > max_value) {
831                         max_value = khugepaged_node_load[nid];
832                         target_node = nid;
833                 }
834
835         /* do some balance if several nodes have the same hit record */
836         if (target_node <= last_khugepaged_target_node)
837                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
838                                 nid++)
839                         if (max_value == khugepaged_node_load[nid]) {
840                                 target_node = nid;
841                                 break;
842                         }
843
844         last_khugepaged_target_node = target_node;
845         return target_node;
846 }
847
848 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
849 {
850         if (IS_ERR(*hpage)) {
851                 if (!*wait)
852                         return false;
853
854                 *wait = false;
855                 *hpage = NULL;
856                 khugepaged_alloc_sleep();
857         } else if (*hpage) {
858                 put_page(*hpage);
859                 *hpage = NULL;
860         }
861
862         return true;
863 }
864
865 static struct page *
866 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
867 {
868         VM_BUG_ON_PAGE(*hpage, *hpage);
869
870         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
871         if (unlikely(!*hpage)) {
872                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
873                 *hpage = ERR_PTR(-ENOMEM);
874                 return NULL;
875         }
876
877         prep_transhuge_page(*hpage);
878         count_vm_event(THP_COLLAPSE_ALLOC);
879         return *hpage;
880 }
881 #else
882 static int khugepaged_find_target_node(void)
883 {
884         return 0;
885 }
886
887 static inline struct page *alloc_khugepaged_hugepage(void)
888 {
889         struct page *page;
890
891         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
892                            HPAGE_PMD_ORDER);
893         if (page)
894                 prep_transhuge_page(page);
895         return page;
896 }
897
898 static struct page *khugepaged_alloc_hugepage(bool *wait)
899 {
900         struct page *hpage;
901
902         do {
903                 hpage = alloc_khugepaged_hugepage();
904                 if (!hpage) {
905                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
906                         if (!*wait)
907                                 return NULL;
908
909                         *wait = false;
910                         khugepaged_alloc_sleep();
911                 } else
912                         count_vm_event(THP_COLLAPSE_ALLOC);
913         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
914
915         return hpage;
916 }
917
918 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
919 {
920         /*
921          * If the hpage allocated earlier was briefly exposed in page cache
922          * before collapse_file() failed, it is possible that racing lookups
923          * have not yet completed, and would then be unpleasantly surprised by
924          * finding the hpage reused for the same mapping at a different offset.
925          * Just release the previous allocation if there is any danger of that.
926          */
927         if (*hpage && page_count(*hpage) > 1) {
928                 put_page(*hpage);
929                 *hpage = NULL;
930         }
931
932         if (!*hpage)
933                 *hpage = khugepaged_alloc_hugepage(wait);
934
935         if (unlikely(!*hpage))
936                 return false;
937
938         return true;
939 }
940
941 static struct page *
942 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
943 {
944         VM_BUG_ON(!*hpage);
945
946         return  *hpage;
947 }
948 #endif
949
950 /*
951  * If mmap_lock temporarily dropped, revalidate vma
952  * before taking mmap_lock.
953  * Return 0 if succeeds, otherwise return none-zero
954  * value (scan code).
955  */
956
957 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
958                 struct vm_area_struct **vmap)
959 {
960         struct vm_area_struct *vma;
961         unsigned long hstart, hend;
962
963         if (unlikely(khugepaged_test_exit(mm)))
964                 return SCAN_ANY_PROCESS;
965
966         *vmap = vma = find_vma(mm, address);
967         if (!vma)
968                 return SCAN_VMA_NULL;
969
970         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
971         hend = vma->vm_end & HPAGE_PMD_MASK;
972         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
973                 return SCAN_ADDRESS_RANGE;
974         if (!hugepage_vma_check(vma, vma->vm_flags))
975                 return SCAN_VMA_CHECK;
976         /* Anon VMA expected */
977         if (!vma->anon_vma || vma->vm_ops)
978                 return SCAN_VMA_CHECK;
979         return 0;
980 }
981
982 /*
983  * Bring missing pages in from swap, to complete THP collapse.
984  * Only done if khugepaged_scan_pmd believes it is worthwhile.
985  *
986  * Called and returns without pte mapped or spinlocks held,
987  * but with mmap_lock held to protect against vma changes.
988  */
989
990 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
991                                         struct vm_area_struct *vma,
992                                         unsigned long address, pmd_t *pmd,
993                                         int referenced)
994 {
995         int swapped_in = 0;
996         vm_fault_t ret = 0;
997         struct vm_fault vmf = {
998                 .vma = vma,
999                 .address = address,
1000                 .flags = FAULT_FLAG_ALLOW_RETRY,
1001                 .pmd = pmd,
1002                 .pgoff = linear_page_index(vma, address),
1003         };
1004
1005         vmf.pte = pte_offset_map(pmd, address);
1006         for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
1007                         vmf.pte++, vmf.address += PAGE_SIZE) {
1008                 vmf.orig_pte = *vmf.pte;
1009                 if (!is_swap_pte(vmf.orig_pte))
1010                         continue;
1011                 swapped_in++;
1012                 ret = do_swap_page(&vmf);
1013
1014                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1015                 if (ret & VM_FAULT_RETRY) {
1016                         mmap_read_lock(mm);
1017                         if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
1018                                 /* vma is no longer available, don't continue to swapin */
1019                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1020                                 return false;
1021                         }
1022                         /* check if the pmd is still valid */
1023                         if (mm_find_pmd(mm, address) != pmd) {
1024                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1025                                 return false;
1026                         }
1027                 }
1028                 if (ret & VM_FAULT_ERROR) {
1029                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1030                         return false;
1031                 }
1032                 /* pte is unmapped now, we need to map it */
1033                 vmf.pte = pte_offset_map(pmd, vmf.address);
1034         }
1035         vmf.pte--;
1036         pte_unmap(vmf.pte);
1037
1038         /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1039         if (swapped_in)
1040                 lru_add_drain();
1041
1042         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1043         return true;
1044 }
1045
1046 static void collapse_huge_page(struct mm_struct *mm,
1047                                    unsigned long address,
1048                                    struct page **hpage,
1049                                    int node, int referenced, int unmapped)
1050 {
1051         LIST_HEAD(compound_pagelist);
1052         pmd_t *pmd, _pmd;
1053         pte_t *pte;
1054         pgtable_t pgtable;
1055         struct page *new_page;
1056         spinlock_t *pmd_ptl, *pte_ptl;
1057         int isolated = 0, result = 0;
1058         struct vm_area_struct *vma;
1059         struct mmu_notifier_range range;
1060         gfp_t gfp;
1061
1062         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1063
1064         /* Only allocate from the target node */
1065         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1066
1067         /*
1068          * Before allocating the hugepage, release the mmap_lock read lock.
1069          * The allocation can take potentially a long time if it involves
1070          * sync compaction, and we do not need to hold the mmap_lock during
1071          * that. We will recheck the vma after taking it again in write mode.
1072          */
1073         mmap_read_unlock(mm);
1074         new_page = khugepaged_alloc_page(hpage, gfp, node);
1075         if (!new_page) {
1076                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1077                 goto out_nolock;
1078         }
1079
1080         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1081                 result = SCAN_CGROUP_CHARGE_FAIL;
1082                 goto out_nolock;
1083         }
1084         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1085
1086         mmap_read_lock(mm);
1087         result = hugepage_vma_revalidate(mm, address, &vma);
1088         if (result) {
1089                 mmap_read_unlock(mm);
1090                 goto out_nolock;
1091         }
1092
1093         pmd = mm_find_pmd(mm, address);
1094         if (!pmd) {
1095                 result = SCAN_PMD_NULL;
1096                 mmap_read_unlock(mm);
1097                 goto out_nolock;
1098         }
1099
1100         /*
1101          * __collapse_huge_page_swapin always returns with mmap_lock locked.
1102          * If it fails, we release mmap_lock and jump out_nolock.
1103          * Continuing to collapse causes inconsistency.
1104          */
1105         if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1106                                                      pmd, referenced)) {
1107                 mmap_read_unlock(mm);
1108                 goto out_nolock;
1109         }
1110
1111         mmap_read_unlock(mm);
1112         /*
1113          * Prevent all access to pagetables with the exception of
1114          * gup_fast later handled by the ptep_clear_flush and the VM
1115          * handled by the anon_vma lock + PG_lock.
1116          */
1117         mmap_write_lock(mm);
1118         result = hugepage_vma_revalidate(mm, address, &vma);
1119         if (result)
1120                 goto out;
1121         /* check if the pmd is still valid */
1122         if (mm_find_pmd(mm, address) != pmd)
1123                 goto out;
1124
1125         anon_vma_lock_write(vma->anon_vma);
1126
1127         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1128                                 address, address + HPAGE_PMD_SIZE);
1129         mmu_notifier_invalidate_range_start(&range);
1130
1131         pte = pte_offset_map(pmd, address);
1132         pte_ptl = pte_lockptr(mm, pmd);
1133
1134         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1135         /*
1136          * After this gup_fast can't run anymore. This also removes
1137          * any huge TLB entry from the CPU so we won't allow
1138          * huge and small TLB entries for the same virtual address
1139          * to avoid the risk of CPU bugs in that area.
1140          */
1141         _pmd = pmdp_collapse_flush(vma, address, pmd);
1142         spin_unlock(pmd_ptl);
1143         mmu_notifier_invalidate_range_end(&range);
1144
1145         spin_lock(pte_ptl);
1146         isolated = __collapse_huge_page_isolate(vma, address, pte,
1147                         &compound_pagelist);
1148         spin_unlock(pte_ptl);
1149
1150         if (unlikely(!isolated)) {
1151                 pte_unmap(pte);
1152                 spin_lock(pmd_ptl);
1153                 BUG_ON(!pmd_none(*pmd));
1154                 /*
1155                  * We can only use set_pmd_at when establishing
1156                  * hugepmds and never for establishing regular pmds that
1157                  * points to regular pagetables. Use pmd_populate for that
1158                  */
1159                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1160                 spin_unlock(pmd_ptl);
1161                 anon_vma_unlock_write(vma->anon_vma);
1162                 result = SCAN_FAIL;
1163                 goto out;
1164         }
1165
1166         /*
1167          * All pages are isolated and locked so anon_vma rmap
1168          * can't run anymore.
1169          */
1170         anon_vma_unlock_write(vma->anon_vma);
1171
1172         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1173                         &compound_pagelist);
1174         pte_unmap(pte);
1175         __SetPageUptodate(new_page);
1176         pgtable = pmd_pgtable(_pmd);
1177
1178         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1179         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1180
1181         /*
1182          * spin_lock() below is not the equivalent of smp_wmb(), so
1183          * this is needed to avoid the copy_huge_page writes to become
1184          * visible after the set_pmd_at() write.
1185          */
1186         smp_wmb();
1187
1188         spin_lock(pmd_ptl);
1189         BUG_ON(!pmd_none(*pmd));
1190         page_add_new_anon_rmap(new_page, vma, address, true);
1191         lru_cache_add_inactive_or_unevictable(new_page, vma);
1192         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1193         set_pmd_at(mm, address, pmd, _pmd);
1194         update_mmu_cache_pmd(vma, address, pmd);
1195         spin_unlock(pmd_ptl);
1196
1197         *hpage = NULL;
1198
1199         khugepaged_pages_collapsed++;
1200         result = SCAN_SUCCEED;
1201 out_up_write:
1202         mmap_write_unlock(mm);
1203 out_nolock:
1204         if (!IS_ERR_OR_NULL(*hpage))
1205                 mem_cgroup_uncharge(*hpage);
1206         trace_mm_collapse_huge_page(mm, isolated, result);
1207         return;
1208 out:
1209         goto out_up_write;
1210 }
1211
1212 static int khugepaged_scan_pmd(struct mm_struct *mm,
1213                                struct vm_area_struct *vma,
1214                                unsigned long address,
1215                                struct page **hpage)
1216 {
1217         pmd_t *pmd;
1218         pte_t *pte, *_pte;
1219         int ret = 0, result = 0, referenced = 0;
1220         int none_or_zero = 0, shared = 0;
1221         struct page *page = NULL;
1222         unsigned long _address;
1223         spinlock_t *ptl;
1224         int node = NUMA_NO_NODE, unmapped = 0;
1225         bool writable = false;
1226
1227         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1228
1229         pmd = mm_find_pmd(mm, address);
1230         if (!pmd) {
1231                 result = SCAN_PMD_NULL;
1232                 goto out;
1233         }
1234
1235         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1236         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1237         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1238              _pte++, _address += PAGE_SIZE) {
1239                 pte_t pteval = *_pte;
1240                 if (is_swap_pte(pteval)) {
1241                         if (++unmapped <= khugepaged_max_ptes_swap) {
1242                                 /*
1243                                  * Always be strict with uffd-wp
1244                                  * enabled swap entries.  Please see
1245                                  * comment below for pte_uffd_wp().
1246                                  */
1247                                 if (pte_swp_uffd_wp(pteval)) {
1248                                         result = SCAN_PTE_UFFD_WP;
1249                                         goto out_unmap;
1250                                 }
1251                                 continue;
1252                         } else {
1253                                 result = SCAN_EXCEED_SWAP_PTE;
1254                                 goto out_unmap;
1255                         }
1256                 }
1257                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1258                         if (!userfaultfd_armed(vma) &&
1259                             ++none_or_zero <= khugepaged_max_ptes_none) {
1260                                 continue;
1261                         } else {
1262                                 result = SCAN_EXCEED_NONE_PTE;
1263                                 goto out_unmap;
1264                         }
1265                 }
1266                 if (!pte_present(pteval)) {
1267                         result = SCAN_PTE_NON_PRESENT;
1268                         goto out_unmap;
1269                 }
1270                 if (pte_uffd_wp(pteval)) {
1271                         /*
1272                          * Don't collapse the page if any of the small
1273                          * PTEs are armed with uffd write protection.
1274                          * Here we can also mark the new huge pmd as
1275                          * write protected if any of the small ones is
1276                          * marked but that could bring uknown
1277                          * userfault messages that falls outside of
1278                          * the registered range.  So, just be simple.
1279                          */
1280                         result = SCAN_PTE_UFFD_WP;
1281                         goto out_unmap;
1282                 }
1283                 if (pte_write(pteval))
1284                         writable = true;
1285
1286                 page = vm_normal_page(vma, _address, pteval);
1287                 if (unlikely(!page)) {
1288                         result = SCAN_PAGE_NULL;
1289                         goto out_unmap;
1290                 }
1291
1292                 if (page_mapcount(page) > 1 &&
1293                                 ++shared > khugepaged_max_ptes_shared) {
1294                         result = SCAN_EXCEED_SHARED_PTE;
1295                         goto out_unmap;
1296                 }
1297
1298                 page = compound_head(page);
1299
1300                 /*
1301                  * Record which node the original page is from and save this
1302                  * information to khugepaged_node_load[].
1303                  * Khupaged will allocate hugepage from the node has the max
1304                  * hit record.
1305                  */
1306                 node = page_to_nid(page);
1307                 if (khugepaged_scan_abort(node)) {
1308                         result = SCAN_SCAN_ABORT;
1309                         goto out_unmap;
1310                 }
1311                 khugepaged_node_load[node]++;
1312                 if (!PageLRU(page)) {
1313                         result = SCAN_PAGE_LRU;
1314                         goto out_unmap;
1315                 }
1316                 if (PageLocked(page)) {
1317                         result = SCAN_PAGE_LOCK;
1318                         goto out_unmap;
1319                 }
1320                 if (!PageAnon(page)) {
1321                         result = SCAN_PAGE_ANON;
1322                         goto out_unmap;
1323                 }
1324
1325                 /*
1326                  * Check if the page has any GUP (or other external) pins.
1327                  *
1328                  * Here the check is racy it may see totmal_mapcount > refcount
1329                  * in some cases.
1330                  * For example, one process with one forked child process.
1331                  * The parent has the PMD split due to MADV_DONTNEED, then
1332                  * the child is trying unmap the whole PMD, but khugepaged
1333                  * may be scanning the parent between the child has
1334                  * PageDoubleMap flag cleared and dec the mapcount.  So
1335                  * khugepaged may see total_mapcount > refcount.
1336                  *
1337                  * But such case is ephemeral we could always retry collapse
1338                  * later.  However it may report false positive if the page
1339                  * has excessive GUP pins (i.e. 512).  Anyway the same check
1340                  * will be done again later the risk seems low.
1341                  */
1342                 if (!is_refcount_suitable(page)) {
1343                         result = SCAN_PAGE_COUNT;
1344                         goto out_unmap;
1345                 }
1346                 if (pte_young(pteval) ||
1347                     page_is_young(page) || PageReferenced(page) ||
1348                     mmu_notifier_test_young(vma->vm_mm, address))
1349                         referenced++;
1350         }
1351         if (!writable) {
1352                 result = SCAN_PAGE_RO;
1353         } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1354                 result = SCAN_LACK_REFERENCED_PAGE;
1355         } else {
1356                 result = SCAN_SUCCEED;
1357                 ret = 1;
1358         }
1359 out_unmap:
1360         pte_unmap_unlock(pte, ptl);
1361         if (ret) {
1362                 node = khugepaged_find_target_node();
1363                 /* collapse_huge_page will return with the mmap_lock released */
1364                 collapse_huge_page(mm, address, hpage, node,
1365                                 referenced, unmapped);
1366         }
1367 out:
1368         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1369                                      none_or_zero, result, unmapped);
1370         return ret;
1371 }
1372
1373 static void collect_mm_slot(struct mm_slot *mm_slot)
1374 {
1375         struct mm_struct *mm = mm_slot->mm;
1376
1377         lockdep_assert_held(&khugepaged_mm_lock);
1378
1379         if (khugepaged_test_exit(mm)) {
1380                 /* free mm_slot */
1381                 hash_del(&mm_slot->hash);
1382                 list_del(&mm_slot->mm_node);
1383
1384                 /*
1385                  * Not strictly needed because the mm exited already.
1386                  *
1387                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1388                  */
1389
1390                 /* khugepaged_mm_lock actually not necessary for the below */
1391                 free_mm_slot(mm_slot);
1392                 mmdrop(mm);
1393         }
1394 }
1395
1396 #ifdef CONFIG_SHMEM
1397 /*
1398  * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1399  * khugepaged should try to collapse the page table.
1400  */
1401 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1402                                          unsigned long addr)
1403 {
1404         struct mm_slot *mm_slot;
1405
1406         VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1407
1408         spin_lock(&khugepaged_mm_lock);
1409         mm_slot = get_mm_slot(mm);
1410         if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1411                 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1412         spin_unlock(&khugepaged_mm_lock);
1413         return 0;
1414 }
1415
1416 /**
1417  * Try to collapse a pte-mapped THP for mm at address haddr.
1418  *
1419  * This function checks whether all the PTEs in the PMD are pointing to the
1420  * right THP. If so, retract the page table so the THP can refault in with
1421  * as pmd-mapped.
1422  */
1423 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1424 {
1425         unsigned long haddr = addr & HPAGE_PMD_MASK;
1426         struct vm_area_struct *vma = find_vma(mm, haddr);
1427         struct page *hpage;
1428         pte_t *start_pte, *pte;
1429         pmd_t *pmd, _pmd;
1430         spinlock_t *ptl;
1431         int count = 0;
1432         int i;
1433
1434         if (!vma || !vma->vm_file ||
1435             vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1436                 return;
1437
1438         /*
1439          * This vm_flags may not have VM_HUGEPAGE if the page was not
1440          * collapsed by this mm. But we can still collapse if the page is
1441          * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1442          * will not fail the vma for missing VM_HUGEPAGE
1443          */
1444         if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1445                 return;
1446
1447         hpage = find_lock_page(vma->vm_file->f_mapping,
1448                                linear_page_index(vma, haddr));
1449         if (!hpage)
1450                 return;
1451
1452         if (!PageHead(hpage))
1453                 goto drop_hpage;
1454
1455         pmd = mm_find_pmd(mm, haddr);
1456         if (!pmd)
1457                 goto drop_hpage;
1458
1459         start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1460
1461         /* step 1: check all mapped PTEs are to the right huge page */
1462         for (i = 0, addr = haddr, pte = start_pte;
1463              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1464                 struct page *page;
1465
1466                 /* empty pte, skip */
1467                 if (pte_none(*pte))
1468                         continue;
1469
1470                 /* page swapped out, abort */
1471                 if (!pte_present(*pte))
1472                         goto abort;
1473
1474                 page = vm_normal_page(vma, addr, *pte);
1475
1476                 /*
1477                  * Note that uprobe, debugger, or MAP_PRIVATE may change the
1478                  * page table, but the new page will not be a subpage of hpage.
1479                  */
1480                 if (hpage + i != page)
1481                         goto abort;
1482                 count++;
1483         }
1484
1485         /* step 2: adjust rmap */
1486         for (i = 0, addr = haddr, pte = start_pte;
1487              i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1488                 struct page *page;
1489
1490                 if (pte_none(*pte))
1491                         continue;
1492                 page = vm_normal_page(vma, addr, *pte);
1493                 page_remove_rmap(page, false);
1494         }
1495
1496         pte_unmap_unlock(start_pte, ptl);
1497
1498         /* step 3: set proper refcount and mm_counters. */
1499         if (count) {
1500                 page_ref_sub(hpage, count);
1501                 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1502         }
1503
1504         /* step 4: collapse pmd */
1505         ptl = pmd_lock(vma->vm_mm, pmd);
1506         _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1507         spin_unlock(ptl);
1508         mm_dec_nr_ptes(mm);
1509         pte_free(mm, pmd_pgtable(_pmd));
1510
1511 drop_hpage:
1512         unlock_page(hpage);
1513         put_page(hpage);
1514         return;
1515
1516 abort:
1517         pte_unmap_unlock(start_pte, ptl);
1518         goto drop_hpage;
1519 }
1520
1521 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1522 {
1523         struct mm_struct *mm = mm_slot->mm;
1524         int i;
1525
1526         if (likely(mm_slot->nr_pte_mapped_thp == 0))
1527                 return 0;
1528
1529         if (!mmap_write_trylock(mm))
1530                 return -EBUSY;
1531
1532         if (unlikely(khugepaged_test_exit(mm)))
1533                 goto out;
1534
1535         for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1536                 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1537
1538 out:
1539         mm_slot->nr_pte_mapped_thp = 0;
1540         mmap_write_unlock(mm);
1541         return 0;
1542 }
1543
1544 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1545 {
1546         struct vm_area_struct *vma;
1547         struct mm_struct *mm;
1548         unsigned long addr;
1549         pmd_t *pmd, _pmd;
1550
1551         i_mmap_lock_write(mapping);
1552         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1553                 /*
1554                  * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1555                  * got written to. These VMAs are likely not worth investing
1556                  * mmap_write_lock(mm) as PMD-mapping is likely to be split
1557                  * later.
1558                  *
1559                  * Not that vma->anon_vma check is racy: it can be set up after
1560                  * the check but before we took mmap_lock by the fault path.
1561                  * But page lock would prevent establishing any new ptes of the
1562                  * page, so we are safe.
1563                  *
1564                  * An alternative would be drop the check, but check that page
1565                  * table is clear before calling pmdp_collapse_flush() under
1566                  * ptl. It has higher chance to recover THP for the VMA, but
1567                  * has higher cost too.
1568                  */
1569                 if (vma->anon_vma)
1570                         continue;
1571                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1572                 if (addr & ~HPAGE_PMD_MASK)
1573                         continue;
1574                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1575                         continue;
1576                 mm = vma->vm_mm;
1577                 pmd = mm_find_pmd(mm, addr);
1578                 if (!pmd)
1579                         continue;
1580                 /*
1581                  * We need exclusive mmap_lock to retract page table.
1582                  *
1583                  * We use trylock due to lock inversion: we need to acquire
1584                  * mmap_lock while holding page lock. Fault path does it in
1585                  * reverse order. Trylock is a way to avoid deadlock.
1586                  */
1587                 if (mmap_write_trylock(mm)) {
1588                         if (!khugepaged_test_exit(mm)) {
1589                                 spinlock_t *ptl = pmd_lock(mm, pmd);
1590                                 /* assume page table is clear */
1591                                 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1592                                 spin_unlock(ptl);
1593                                 mm_dec_nr_ptes(mm);
1594                                 pte_free(mm, pmd_pgtable(_pmd));
1595                         }
1596                         mmap_write_unlock(mm);
1597                 } else {
1598                         /* Try again later */
1599                         khugepaged_add_pte_mapped_thp(mm, addr);
1600                 }
1601         }
1602         i_mmap_unlock_write(mapping);
1603 }
1604
1605 /**
1606  * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1607  *
1608  * Basic scheme is simple, details are more complex:
1609  *  - allocate and lock a new huge page;
1610  *  - scan page cache replacing old pages with the new one
1611  *    + swap/gup in pages if necessary;
1612  *    + fill in gaps;
1613  *    + keep old pages around in case rollback is required;
1614  *  - if replacing succeeds:
1615  *    + copy data over;
1616  *    + free old pages;
1617  *    + unlock huge page;
1618  *  - if replacing failed;
1619  *    + put all pages back and unfreeze them;
1620  *    + restore gaps in the page cache;
1621  *    + unlock and free huge page;
1622  */
1623 static void collapse_file(struct mm_struct *mm,
1624                 struct file *file, pgoff_t start,
1625                 struct page **hpage, int node)
1626 {
1627         struct address_space *mapping = file->f_mapping;
1628         gfp_t gfp;
1629         struct page *new_page;
1630         pgoff_t index, end = start + HPAGE_PMD_NR;
1631         LIST_HEAD(pagelist);
1632         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1633         int nr_none = 0, result = SCAN_SUCCEED;
1634         bool is_shmem = shmem_file(file);
1635
1636         VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1637         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1638
1639         /* Only allocate from the target node */
1640         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1641
1642         new_page = khugepaged_alloc_page(hpage, gfp, node);
1643         if (!new_page) {
1644                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1645                 goto out;
1646         }
1647
1648         if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1649                 result = SCAN_CGROUP_CHARGE_FAIL;
1650                 goto out;
1651         }
1652         count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1653
1654         /* This will be less messy when we use multi-index entries */
1655         do {
1656                 xas_lock_irq(&xas);
1657                 xas_create_range(&xas);
1658                 if (!xas_error(&xas))
1659                         break;
1660                 xas_unlock_irq(&xas);
1661                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1662                         result = SCAN_FAIL;
1663                         goto out;
1664                 }
1665         } while (1);
1666
1667         __SetPageLocked(new_page);
1668         if (is_shmem)
1669                 __SetPageSwapBacked(new_page);
1670         new_page->index = start;
1671         new_page->mapping = mapping;
1672
1673         /*
1674          * At this point the new_page is locked and not up-to-date.
1675          * It's safe to insert it into the page cache, because nobody would
1676          * be able to map it or use it in another way until we unlock it.
1677          */
1678
1679         xas_set(&xas, start);
1680         for (index = start; index < end; index++) {
1681                 struct page *page = xas_next(&xas);
1682
1683                 VM_BUG_ON(index != xas.xa_index);
1684                 if (is_shmem) {
1685                         if (!page) {
1686                                 /*
1687                                  * Stop if extent has been truncated or
1688                                  * hole-punched, and is now completely
1689                                  * empty.
1690                                  */
1691                                 if (index == start) {
1692                                         if (!xas_next_entry(&xas, end - 1)) {
1693                                                 result = SCAN_TRUNCATED;
1694                                                 goto xa_locked;
1695                                         }
1696                                         xas_set(&xas, index);
1697                                 }
1698                                 if (!shmem_charge(mapping->host, 1)) {
1699                                         result = SCAN_FAIL;
1700                                         goto xa_locked;
1701                                 }
1702                                 xas_store(&xas, new_page);
1703                                 nr_none++;
1704                                 continue;
1705                         }
1706
1707                         if (xa_is_value(page) || !PageUptodate(page)) {
1708                                 xas_unlock_irq(&xas);
1709                                 /* swap in or instantiate fallocated page */
1710                                 if (shmem_getpage(mapping->host, index, &page,
1711                                                   SGP_NOHUGE)) {
1712                                         result = SCAN_FAIL;
1713                                         goto xa_unlocked;
1714                                 }
1715                         } else if (trylock_page(page)) {
1716                                 get_page(page);
1717                                 xas_unlock_irq(&xas);
1718                         } else {
1719                                 result = SCAN_PAGE_LOCK;
1720                                 goto xa_locked;
1721                         }
1722                 } else {        /* !is_shmem */
1723                         if (!page || xa_is_value(page)) {
1724                                 xas_unlock_irq(&xas);
1725                                 page_cache_sync_readahead(mapping, &file->f_ra,
1726                                                           file, index,
1727                                                           end - index);
1728                                 /* drain pagevecs to help isolate_lru_page() */
1729                                 lru_add_drain();
1730                                 page = find_lock_page(mapping, index);
1731                                 if (unlikely(page == NULL)) {
1732                                         result = SCAN_FAIL;
1733                                         goto xa_unlocked;
1734                                 }
1735                         } else if (PageDirty(page)) {
1736                                 /*
1737                                  * khugepaged only works on read-only fd,
1738                                  * so this page is dirty because it hasn't
1739                                  * been flushed since first write. There
1740                                  * won't be new dirty pages.
1741                                  *
1742                                  * Trigger async flush here and hope the
1743                                  * writeback is done when khugepaged
1744                                  * revisits this page.
1745                                  *
1746                                  * This is a one-off situation. We are not
1747                                  * forcing writeback in loop.
1748                                  */
1749                                 xas_unlock_irq(&xas);
1750                                 filemap_flush(mapping);
1751                                 result = SCAN_FAIL;
1752                                 goto xa_unlocked;
1753                         } else if (trylock_page(page)) {
1754                                 get_page(page);
1755                                 xas_unlock_irq(&xas);
1756                         } else {
1757                                 result = SCAN_PAGE_LOCK;
1758                                 goto xa_locked;
1759                         }
1760                 }
1761
1762                 /*
1763                  * The page must be locked, so we can drop the i_pages lock
1764                  * without racing with truncate.
1765                  */
1766                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1767
1768                 /* make sure the page is up to date */
1769                 if (unlikely(!PageUptodate(page))) {
1770                         result = SCAN_FAIL;
1771                         goto out_unlock;
1772                 }
1773
1774                 /*
1775                  * If file was truncated then extended, or hole-punched, before
1776                  * we locked the first page, then a THP might be there already.
1777                  */
1778                 if (PageTransCompound(page)) {
1779                         result = SCAN_PAGE_COMPOUND;
1780                         goto out_unlock;
1781                 }
1782
1783                 if (page_mapping(page) != mapping) {
1784                         result = SCAN_TRUNCATED;
1785                         goto out_unlock;
1786                 }
1787
1788                 if (!is_shmem && PageDirty(page)) {
1789                         /*
1790                          * khugepaged only works on read-only fd, so this
1791                          * page is dirty because it hasn't been flushed
1792                          * since first write.
1793                          */
1794                         result = SCAN_FAIL;
1795                         goto out_unlock;
1796                 }
1797
1798                 if (isolate_lru_page(page)) {
1799                         result = SCAN_DEL_PAGE_LRU;
1800                         goto out_unlock;
1801                 }
1802
1803                 if (page_has_private(page) &&
1804                     !try_to_release_page(page, GFP_KERNEL)) {
1805                         result = SCAN_PAGE_HAS_PRIVATE;
1806                         putback_lru_page(page);
1807                         goto out_unlock;
1808                 }
1809
1810                 if (page_mapped(page))
1811                         unmap_mapping_pages(mapping, index, 1, false);
1812
1813                 xas_lock_irq(&xas);
1814                 xas_set(&xas, index);
1815
1816                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1817                 VM_BUG_ON_PAGE(page_mapped(page), page);
1818
1819                 /*
1820                  * The page is expected to have page_count() == 3:
1821                  *  - we hold a pin on it;
1822                  *  - one reference from page cache;
1823                  *  - one from isolate_lru_page;
1824                  */
1825                 if (!page_ref_freeze(page, 3)) {
1826                         result = SCAN_PAGE_COUNT;
1827                         xas_unlock_irq(&xas);
1828                         putback_lru_page(page);
1829                         goto out_unlock;
1830                 }
1831
1832                 /*
1833                  * Add the page to the list to be able to undo the collapse if
1834                  * something go wrong.
1835                  */
1836                 list_add_tail(&page->lru, &pagelist);
1837
1838                 /* Finally, replace with the new page. */
1839                 xas_store(&xas, new_page);
1840                 continue;
1841 out_unlock:
1842                 unlock_page(page);
1843                 put_page(page);
1844                 goto xa_unlocked;
1845         }
1846
1847         if (is_shmem)
1848                 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1849         else {
1850                 __inc_node_page_state(new_page, NR_FILE_THPS);
1851                 filemap_nr_thps_inc(mapping);
1852         }
1853
1854         if (nr_none) {
1855                 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1856                 if (is_shmem)
1857                         __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1858         }
1859
1860 xa_locked:
1861         xas_unlock_irq(&xas);
1862 xa_unlocked:
1863
1864         if (result == SCAN_SUCCEED) {
1865                 struct page *page, *tmp;
1866
1867                 /*
1868                  * Replacing old pages with new one has succeeded, now we
1869                  * need to copy the content and free the old pages.
1870                  */
1871                 index = start;
1872                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1873                         while (index < page->index) {
1874                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1875                                 index++;
1876                         }
1877                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1878                                         page);
1879                         list_del(&page->lru);
1880                         page->mapping = NULL;
1881                         page_ref_unfreeze(page, 1);
1882                         ClearPageActive(page);
1883                         ClearPageUnevictable(page);
1884                         unlock_page(page);
1885                         put_page(page);
1886                         index++;
1887                 }
1888                 while (index < end) {
1889                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1890                         index++;
1891                 }
1892
1893                 SetPageUptodate(new_page);
1894                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1895                 if (is_shmem)
1896                         set_page_dirty(new_page);
1897                 lru_cache_add(new_page);
1898
1899                 /*
1900                  * Remove pte page tables, so we can re-fault the page as huge.
1901                  */
1902                 retract_page_tables(mapping, start);
1903                 *hpage = NULL;
1904
1905                 khugepaged_pages_collapsed++;
1906         } else {
1907                 struct page *page;
1908
1909                 /* Something went wrong: roll back page cache changes */
1910                 xas_lock_irq(&xas);
1911                 mapping->nrpages -= nr_none;
1912
1913                 if (is_shmem)
1914                         shmem_uncharge(mapping->host, nr_none);
1915
1916                 xas_set(&xas, start);
1917                 xas_for_each(&xas, page, end - 1) {
1918                         page = list_first_entry_or_null(&pagelist,
1919                                         struct page, lru);
1920                         if (!page || xas.xa_index < page->index) {
1921                                 if (!nr_none)
1922                                         break;
1923                                 nr_none--;
1924                                 /* Put holes back where they were */
1925                                 xas_store(&xas, NULL);
1926                                 continue;
1927                         }
1928
1929                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1930
1931                         /* Unfreeze the page. */
1932                         list_del(&page->lru);
1933                         page_ref_unfreeze(page, 2);
1934                         xas_store(&xas, page);
1935                         xas_pause(&xas);
1936                         xas_unlock_irq(&xas);
1937                         unlock_page(page);
1938                         putback_lru_page(page);
1939                         xas_lock_irq(&xas);
1940                 }
1941                 VM_BUG_ON(nr_none);
1942                 xas_unlock_irq(&xas);
1943
1944                 new_page->mapping = NULL;
1945         }
1946
1947         unlock_page(new_page);
1948 out:
1949         VM_BUG_ON(!list_empty(&pagelist));
1950         if (!IS_ERR_OR_NULL(*hpage))
1951                 mem_cgroup_uncharge(*hpage);
1952         /* TODO: tracepoints */
1953 }
1954
1955 static void khugepaged_scan_file(struct mm_struct *mm,
1956                 struct file *file, pgoff_t start, struct page **hpage)
1957 {
1958         struct page *page = NULL;
1959         struct address_space *mapping = file->f_mapping;
1960         XA_STATE(xas, &mapping->i_pages, start);
1961         int present, swap;
1962         int node = NUMA_NO_NODE;
1963         int result = SCAN_SUCCEED;
1964
1965         present = 0;
1966         swap = 0;
1967         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1968         rcu_read_lock();
1969         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1970                 if (xas_retry(&xas, page))
1971                         continue;
1972
1973                 if (xa_is_value(page)) {
1974                         if (++swap > khugepaged_max_ptes_swap) {
1975                                 result = SCAN_EXCEED_SWAP_PTE;
1976                                 break;
1977                         }
1978                         continue;
1979                 }
1980
1981                 if (PageTransCompound(page)) {
1982                         result = SCAN_PAGE_COMPOUND;
1983                         break;
1984                 }
1985
1986                 node = page_to_nid(page);
1987                 if (khugepaged_scan_abort(node)) {
1988                         result = SCAN_SCAN_ABORT;
1989                         break;
1990                 }
1991                 khugepaged_node_load[node]++;
1992
1993                 if (!PageLRU(page)) {
1994                         result = SCAN_PAGE_LRU;
1995                         break;
1996                 }
1997
1998                 if (page_count(page) !=
1999                     1 + page_mapcount(page) + page_has_private(page)) {
2000                         result = SCAN_PAGE_COUNT;
2001                         break;
2002                 }
2003
2004                 /*
2005                  * We probably should check if the page is referenced here, but
2006                  * nobody would transfer pte_young() to PageReferenced() for us.
2007                  * And rmap walk here is just too costly...
2008                  */
2009
2010                 present++;
2011
2012                 if (need_resched()) {
2013                         xas_pause(&xas);
2014                         cond_resched_rcu();
2015                 }
2016         }
2017         rcu_read_unlock();
2018
2019         if (result == SCAN_SUCCEED) {
2020                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2021                         result = SCAN_EXCEED_NONE_PTE;
2022                 } else {
2023                         node = khugepaged_find_target_node();
2024                         collapse_file(mm, file, start, hpage, node);
2025                 }
2026         }
2027
2028         /* TODO: tracepoints */
2029 }
2030 #else
2031 static void khugepaged_scan_file(struct mm_struct *mm,
2032                 struct file *file, pgoff_t start, struct page **hpage)
2033 {
2034         BUILD_BUG();
2035 }
2036
2037 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2038 {
2039         return 0;
2040 }
2041 #endif
2042
2043 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2044                                             struct page **hpage)
2045         __releases(&khugepaged_mm_lock)
2046         __acquires(&khugepaged_mm_lock)
2047 {
2048         struct mm_slot *mm_slot;
2049         struct mm_struct *mm;
2050         struct vm_area_struct *vma;
2051         int progress = 0;
2052
2053         VM_BUG_ON(!pages);
2054         lockdep_assert_held(&khugepaged_mm_lock);
2055
2056         if (khugepaged_scan.mm_slot)
2057                 mm_slot = khugepaged_scan.mm_slot;
2058         else {
2059                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2060                                      struct mm_slot, mm_node);
2061                 khugepaged_scan.address = 0;
2062                 khugepaged_scan.mm_slot = mm_slot;
2063         }
2064         spin_unlock(&khugepaged_mm_lock);
2065         khugepaged_collapse_pte_mapped_thps(mm_slot);
2066
2067         mm = mm_slot->mm;
2068         /*
2069          * Don't wait for semaphore (to avoid long wait times).  Just move to
2070          * the next mm on the list.
2071          */
2072         vma = NULL;
2073         if (unlikely(!mmap_read_trylock(mm)))
2074                 goto breakouterloop_mmap_lock;
2075         if (likely(!khugepaged_test_exit(mm)))
2076                 vma = find_vma(mm, khugepaged_scan.address);
2077
2078         progress++;
2079         for (; vma; vma = vma->vm_next) {
2080                 unsigned long hstart, hend;
2081
2082                 cond_resched();
2083                 if (unlikely(khugepaged_test_exit(mm))) {
2084                         progress++;
2085                         break;
2086                 }
2087                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2088 skip:
2089                         progress++;
2090                         continue;
2091                 }
2092                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2093                 hend = vma->vm_end & HPAGE_PMD_MASK;
2094                 if (hstart >= hend)
2095                         goto skip;
2096                 if (khugepaged_scan.address > hend)
2097                         goto skip;
2098                 if (khugepaged_scan.address < hstart)
2099                         khugepaged_scan.address = hstart;
2100                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2101                 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2102                         goto skip;
2103
2104                 while (khugepaged_scan.address < hend) {
2105                         int ret;
2106                         cond_resched();
2107                         if (unlikely(khugepaged_test_exit(mm)))
2108                                 goto breakouterloop;
2109
2110                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2111                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2112                                   hend);
2113                         if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2114                                 struct file *file = get_file(vma->vm_file);
2115                                 pgoff_t pgoff = linear_page_index(vma,
2116                                                 khugepaged_scan.address);
2117
2118                                 mmap_read_unlock(mm);
2119                                 ret = 1;
2120                                 khugepaged_scan_file(mm, file, pgoff, hpage);
2121                                 fput(file);
2122                         } else {
2123                                 ret = khugepaged_scan_pmd(mm, vma,
2124                                                 khugepaged_scan.address,
2125                                                 hpage);
2126                         }
2127                         /* move to next address */
2128                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2129                         progress += HPAGE_PMD_NR;
2130                         if (ret)
2131                                 /* we released mmap_lock so break loop */
2132                                 goto breakouterloop_mmap_lock;
2133                         if (progress >= pages)
2134                                 goto breakouterloop;
2135                 }
2136         }
2137 breakouterloop:
2138         mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2139 breakouterloop_mmap_lock:
2140
2141         spin_lock(&khugepaged_mm_lock);
2142         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2143         /*
2144          * Release the current mm_slot if this mm is about to die, or
2145          * if we scanned all vmas of this mm.
2146          */
2147         if (khugepaged_test_exit(mm) || !vma) {
2148                 /*
2149                  * Make sure that if mm_users is reaching zero while
2150                  * khugepaged runs here, khugepaged_exit will find
2151                  * mm_slot not pointing to the exiting mm.
2152                  */
2153                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2154                         khugepaged_scan.mm_slot = list_entry(
2155                                 mm_slot->mm_node.next,
2156                                 struct mm_slot, mm_node);
2157                         khugepaged_scan.address = 0;
2158                 } else {
2159                         khugepaged_scan.mm_slot = NULL;
2160                         khugepaged_full_scans++;
2161                 }
2162
2163                 collect_mm_slot(mm_slot);
2164         }
2165
2166         return progress;
2167 }
2168
2169 static int khugepaged_has_work(void)
2170 {
2171         return !list_empty(&khugepaged_scan.mm_head) &&
2172                 khugepaged_enabled();
2173 }
2174
2175 static int khugepaged_wait_event(void)
2176 {
2177         return !list_empty(&khugepaged_scan.mm_head) ||
2178                 kthread_should_stop();
2179 }
2180
2181 static void khugepaged_do_scan(void)
2182 {
2183         struct page *hpage = NULL;
2184         unsigned int progress = 0, pass_through_head = 0;
2185         unsigned int pages = khugepaged_pages_to_scan;
2186         bool wait = true;
2187
2188         barrier(); /* write khugepaged_pages_to_scan to local stack */
2189
2190         lru_add_drain_all();
2191
2192         while (progress < pages) {
2193                 if (!khugepaged_prealloc_page(&hpage, &wait))
2194                         break;
2195
2196                 cond_resched();
2197
2198                 if (unlikely(kthread_should_stop() || try_to_freeze()))
2199                         break;
2200
2201                 spin_lock(&khugepaged_mm_lock);
2202                 if (!khugepaged_scan.mm_slot)
2203                         pass_through_head++;
2204                 if (khugepaged_has_work() &&
2205                     pass_through_head < 2)
2206                         progress += khugepaged_scan_mm_slot(pages - progress,
2207                                                             &hpage);
2208                 else
2209                         progress = pages;
2210                 spin_unlock(&khugepaged_mm_lock);
2211         }
2212
2213         if (!IS_ERR_OR_NULL(hpage))
2214                 put_page(hpage);
2215 }
2216
2217 static bool khugepaged_should_wakeup(void)
2218 {
2219         return kthread_should_stop() ||
2220                time_after_eq(jiffies, khugepaged_sleep_expire);
2221 }
2222
2223 static void khugepaged_wait_work(void)
2224 {
2225         if (khugepaged_has_work()) {
2226                 const unsigned long scan_sleep_jiffies =
2227                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2228
2229                 if (!scan_sleep_jiffies)
2230                         return;
2231
2232                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2233                 wait_event_freezable_timeout(khugepaged_wait,
2234                                              khugepaged_should_wakeup(),
2235                                              scan_sleep_jiffies);
2236                 return;
2237         }
2238
2239         if (khugepaged_enabled())
2240                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2241 }
2242
2243 static int khugepaged(void *none)
2244 {
2245         struct mm_slot *mm_slot;
2246
2247         set_freezable();
2248         set_user_nice(current, MAX_NICE);
2249
2250         while (!kthread_should_stop()) {
2251                 khugepaged_do_scan();
2252                 khugepaged_wait_work();
2253         }
2254
2255         spin_lock(&khugepaged_mm_lock);
2256         mm_slot = khugepaged_scan.mm_slot;
2257         khugepaged_scan.mm_slot = NULL;
2258         if (mm_slot)
2259                 collect_mm_slot(mm_slot);
2260         spin_unlock(&khugepaged_mm_lock);
2261         return 0;
2262 }
2263
2264 static void set_recommended_min_free_kbytes(void)
2265 {
2266         struct zone *zone;
2267         int nr_zones = 0;
2268         unsigned long recommended_min;
2269
2270         for_each_populated_zone(zone) {
2271                 /*
2272                  * We don't need to worry about fragmentation of
2273                  * ZONE_MOVABLE since it only has movable pages.
2274                  */
2275                 if (zone_idx(zone) > gfp_zone(GFP_USER))
2276                         continue;
2277
2278                 nr_zones++;
2279         }
2280
2281         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2282         recommended_min = pageblock_nr_pages * nr_zones * 2;
2283
2284         /*
2285          * Make sure that on average at least two pageblocks are almost free
2286          * of another type, one for a migratetype to fall back to and a
2287          * second to avoid subsequent fallbacks of other types There are 3
2288          * MIGRATE_TYPES we care about.
2289          */
2290         recommended_min += pageblock_nr_pages * nr_zones *
2291                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2292
2293         /* don't ever allow to reserve more than 5% of the lowmem */
2294         recommended_min = min(recommended_min,
2295                               (unsigned long) nr_free_buffer_pages() / 20);
2296         recommended_min <<= (PAGE_SHIFT-10);
2297
2298         if (recommended_min > min_free_kbytes) {
2299                 if (user_min_free_kbytes >= 0)
2300                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2301                                 min_free_kbytes, recommended_min);
2302
2303                 min_free_kbytes = recommended_min;
2304         }
2305         setup_per_zone_wmarks();
2306 }
2307
2308 int start_stop_khugepaged(void)
2309 {
2310         int err = 0;
2311
2312         mutex_lock(&khugepaged_mutex);
2313         if (khugepaged_enabled()) {
2314                 if (!khugepaged_thread)
2315                         khugepaged_thread = kthread_run(khugepaged, NULL,
2316                                                         "khugepaged");
2317                 if (IS_ERR(khugepaged_thread)) {
2318                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2319                         err = PTR_ERR(khugepaged_thread);
2320                         khugepaged_thread = NULL;
2321                         goto fail;
2322                 }
2323
2324                 if (!list_empty(&khugepaged_scan.mm_head))
2325                         wake_up_interruptible(&khugepaged_wait);
2326
2327                 set_recommended_min_free_kbytes();
2328         } else if (khugepaged_thread) {
2329                 kthread_stop(khugepaged_thread);
2330                 khugepaged_thread = NULL;
2331         }
2332 fail:
2333         mutex_unlock(&khugepaged_mutex);
2334         return err;
2335 }
2336
2337 void khugepaged_min_free_kbytes_update(void)
2338 {
2339         mutex_lock(&khugepaged_mutex);
2340         if (khugepaged_enabled() && khugepaged_thread)
2341                 set_recommended_min_free_kbytes();
2342         mutex_unlock(&khugepaged_mutex);
2343 }