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