Loading...
1// SPDX-License-Identifier: GPL-2.0
2#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
3
4#include <linux/mm.h>
5#include <linux/sched.h>
6#include <linux/sched/mm.h>
7#include <linux/sched/coredump.h>
8#include <linux/mmu_notifier.h>
9#include <linux/rmap.h>
10#include <linux/swap.h>
11#include <linux/mm_inline.h>
12#include <linux/kthread.h>
13#include <linux/khugepaged.h>
14#include <linux/freezer.h>
15#include <linux/mman.h>
16#include <linux/hashtable.h>
17#include <linux/userfaultfd_k.h>
18#include <linux/page_idle.h>
19#include <linux/page_table_check.h>
20#include <linux/swapops.h>
21#include <linux/shmem_fs.h>
22
23#include <asm/tlb.h>
24#include <asm/pgalloc.h>
25#include "internal.h"
26#include "mm_slot.h"
27
28enum scan_result {
29 SCAN_FAIL,
30 SCAN_SUCCEED,
31 SCAN_PMD_NULL,
32 SCAN_PMD_NONE,
33 SCAN_PMD_MAPPED,
34 SCAN_EXCEED_NONE_PTE,
35 SCAN_EXCEED_SWAP_PTE,
36 SCAN_EXCEED_SHARED_PTE,
37 SCAN_PTE_NON_PRESENT,
38 SCAN_PTE_UFFD_WP,
39 SCAN_PTE_MAPPED_HUGEPAGE,
40 SCAN_PAGE_RO,
41 SCAN_LACK_REFERENCED_PAGE,
42 SCAN_PAGE_NULL,
43 SCAN_SCAN_ABORT,
44 SCAN_PAGE_COUNT,
45 SCAN_PAGE_LRU,
46 SCAN_PAGE_LOCK,
47 SCAN_PAGE_ANON,
48 SCAN_PAGE_COMPOUND,
49 SCAN_ANY_PROCESS,
50 SCAN_VMA_NULL,
51 SCAN_VMA_CHECK,
52 SCAN_ADDRESS_RANGE,
53 SCAN_DEL_PAGE_LRU,
54 SCAN_ALLOC_HUGE_PAGE_FAIL,
55 SCAN_CGROUP_CHARGE_FAIL,
56 SCAN_TRUNCATED,
57 SCAN_PAGE_HAS_PRIVATE,
58};
59
60#define CREATE_TRACE_POINTS
61#include <trace/events/huge_memory.h>
62
63static struct task_struct *khugepaged_thread __read_mostly;
64static DEFINE_MUTEX(khugepaged_mutex);
65
66/* default scan 8*512 pte (or vmas) every 30 second */
67static unsigned int khugepaged_pages_to_scan __read_mostly;
68static unsigned int khugepaged_pages_collapsed;
69static unsigned int khugepaged_full_scans;
70static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
71/* during fragmentation poll the hugepage allocator once every minute */
72static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
73static unsigned long khugepaged_sleep_expire;
74static DEFINE_SPINLOCK(khugepaged_mm_lock);
75static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
76/*
77 * default collapse hugepages if there is at least one pte mapped like
78 * it would have happened if the vma was large enough during page
79 * fault.
80 *
81 * Note that these are only respected if collapse was initiated by khugepaged.
82 */
83static unsigned int khugepaged_max_ptes_none __read_mostly;
84static unsigned int khugepaged_max_ptes_swap __read_mostly;
85static unsigned int khugepaged_max_ptes_shared __read_mostly;
86
87#define MM_SLOTS_HASH_BITS 10
88static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
89
90static struct kmem_cache *mm_slot_cache __read_mostly;
91
92#define MAX_PTE_MAPPED_THP 8
93
94struct collapse_control {
95 bool is_khugepaged;
96
97 /* Num pages scanned per node */
98 u32 node_load[MAX_NUMNODES];
99
100 /* nodemask for allocation fallback */
101 nodemask_t alloc_nmask;
102};
103
104/**
105 * struct khugepaged_mm_slot - khugepaged information per mm that is being scanned
106 * @slot: hash lookup from mm to mm_slot
107 * @nr_pte_mapped_thp: number of pte mapped THP
108 * @pte_mapped_thp: address array corresponding pte mapped THP
109 */
110struct khugepaged_mm_slot {
111 struct mm_slot slot;
112
113 /* pte-mapped THP in this mm */
114 int nr_pte_mapped_thp;
115 unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
116};
117
118/**
119 * struct khugepaged_scan - cursor for scanning
120 * @mm_head: the head of the mm list to scan
121 * @mm_slot: the current mm_slot we are scanning
122 * @address: the next address inside that to be scanned
123 *
124 * There is only the one khugepaged_scan instance of this cursor structure.
125 */
126struct khugepaged_scan {
127 struct list_head mm_head;
128 struct khugepaged_mm_slot *mm_slot;
129 unsigned long address;
130};
131
132static struct khugepaged_scan khugepaged_scan = {
133 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
134};
135
136#ifdef CONFIG_SYSFS
137static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
138 struct kobj_attribute *attr,
139 char *buf)
140{
141 return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
142}
143
144static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
145 struct kobj_attribute *attr,
146 const char *buf, size_t count)
147{
148 unsigned int msecs;
149 int err;
150
151 err = kstrtouint(buf, 10, &msecs);
152 if (err)
153 return -EINVAL;
154
155 khugepaged_scan_sleep_millisecs = msecs;
156 khugepaged_sleep_expire = 0;
157 wake_up_interruptible(&khugepaged_wait);
158
159 return count;
160}
161static struct kobj_attribute scan_sleep_millisecs_attr =
162 __ATTR_RW(scan_sleep_millisecs);
163
164static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
165 struct kobj_attribute *attr,
166 char *buf)
167{
168 return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
169}
170
171static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
172 struct kobj_attribute *attr,
173 const char *buf, size_t count)
174{
175 unsigned int msecs;
176 int err;
177
178 err = kstrtouint(buf, 10, &msecs);
179 if (err)
180 return -EINVAL;
181
182 khugepaged_alloc_sleep_millisecs = msecs;
183 khugepaged_sleep_expire = 0;
184 wake_up_interruptible(&khugepaged_wait);
185
186 return count;
187}
188static struct kobj_attribute alloc_sleep_millisecs_attr =
189 __ATTR_RW(alloc_sleep_millisecs);
190
191static ssize_t pages_to_scan_show(struct kobject *kobj,
192 struct kobj_attribute *attr,
193 char *buf)
194{
195 return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
196}
197static ssize_t pages_to_scan_store(struct kobject *kobj,
198 struct kobj_attribute *attr,
199 const char *buf, size_t count)
200{
201 unsigned int pages;
202 int err;
203
204 err = kstrtouint(buf, 10, &pages);
205 if (err || !pages)
206 return -EINVAL;
207
208 khugepaged_pages_to_scan = pages;
209
210 return count;
211}
212static struct kobj_attribute pages_to_scan_attr =
213 __ATTR_RW(pages_to_scan);
214
215static ssize_t pages_collapsed_show(struct kobject *kobj,
216 struct kobj_attribute *attr,
217 char *buf)
218{
219 return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
220}
221static struct kobj_attribute pages_collapsed_attr =
222 __ATTR_RO(pages_collapsed);
223
224static ssize_t full_scans_show(struct kobject *kobj,
225 struct kobj_attribute *attr,
226 char *buf)
227{
228 return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
229}
230static struct kobj_attribute full_scans_attr =
231 __ATTR_RO(full_scans);
232
233static ssize_t defrag_show(struct kobject *kobj,
234 struct kobj_attribute *attr, char *buf)
235{
236 return single_hugepage_flag_show(kobj, attr, buf,
237 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
238}
239static ssize_t defrag_store(struct kobject *kobj,
240 struct kobj_attribute *attr,
241 const char *buf, size_t count)
242{
243 return single_hugepage_flag_store(kobj, attr, buf, count,
244 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
245}
246static struct kobj_attribute khugepaged_defrag_attr =
247 __ATTR_RW(defrag);
248
249/*
250 * max_ptes_none controls if khugepaged should collapse hugepages over
251 * any unmapped ptes in turn potentially increasing the memory
252 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
253 * reduce the available free memory in the system as it
254 * runs. Increasing max_ptes_none will instead potentially reduce the
255 * free memory in the system during the khugepaged scan.
256 */
257static ssize_t max_ptes_none_show(struct kobject *kobj,
258 struct kobj_attribute *attr,
259 char *buf)
260{
261 return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
262}
263static ssize_t max_ptes_none_store(struct kobject *kobj,
264 struct kobj_attribute *attr,
265 const char *buf, size_t count)
266{
267 int err;
268 unsigned long max_ptes_none;
269
270 err = kstrtoul(buf, 10, &max_ptes_none);
271 if (err || max_ptes_none > HPAGE_PMD_NR - 1)
272 return -EINVAL;
273
274 khugepaged_max_ptes_none = max_ptes_none;
275
276 return count;
277}
278static struct kobj_attribute khugepaged_max_ptes_none_attr =
279 __ATTR_RW(max_ptes_none);
280
281static ssize_t max_ptes_swap_show(struct kobject *kobj,
282 struct kobj_attribute *attr,
283 char *buf)
284{
285 return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
286}
287
288static ssize_t max_ptes_swap_store(struct kobject *kobj,
289 struct kobj_attribute *attr,
290 const char *buf, size_t count)
291{
292 int err;
293 unsigned long max_ptes_swap;
294
295 err = kstrtoul(buf, 10, &max_ptes_swap);
296 if (err || max_ptes_swap > HPAGE_PMD_NR - 1)
297 return -EINVAL;
298
299 khugepaged_max_ptes_swap = max_ptes_swap;
300
301 return count;
302}
303
304static struct kobj_attribute khugepaged_max_ptes_swap_attr =
305 __ATTR_RW(max_ptes_swap);
306
307static ssize_t max_ptes_shared_show(struct kobject *kobj,
308 struct kobj_attribute *attr,
309 char *buf)
310{
311 return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
312}
313
314static ssize_t max_ptes_shared_store(struct kobject *kobj,
315 struct kobj_attribute *attr,
316 const char *buf, size_t count)
317{
318 int err;
319 unsigned long max_ptes_shared;
320
321 err = kstrtoul(buf, 10, &max_ptes_shared);
322 if (err || max_ptes_shared > HPAGE_PMD_NR - 1)
323 return -EINVAL;
324
325 khugepaged_max_ptes_shared = max_ptes_shared;
326
327 return count;
328}
329
330static struct kobj_attribute khugepaged_max_ptes_shared_attr =
331 __ATTR_RW(max_ptes_shared);
332
333static struct attribute *khugepaged_attr[] = {
334 &khugepaged_defrag_attr.attr,
335 &khugepaged_max_ptes_none_attr.attr,
336 &khugepaged_max_ptes_swap_attr.attr,
337 &khugepaged_max_ptes_shared_attr.attr,
338 &pages_to_scan_attr.attr,
339 &pages_collapsed_attr.attr,
340 &full_scans_attr.attr,
341 &scan_sleep_millisecs_attr.attr,
342 &alloc_sleep_millisecs_attr.attr,
343 NULL,
344};
345
346struct attribute_group khugepaged_attr_group = {
347 .attrs = khugepaged_attr,
348 .name = "khugepaged",
349};
350#endif /* CONFIG_SYSFS */
351
352int hugepage_madvise(struct vm_area_struct *vma,
353 unsigned long *vm_flags, int advice)
354{
355 switch (advice) {
356 case MADV_HUGEPAGE:
357#ifdef CONFIG_S390
358 /*
359 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
360 * can't handle this properly after s390_enable_sie, so we simply
361 * ignore the madvise to prevent qemu from causing a SIGSEGV.
362 */
363 if (mm_has_pgste(vma->vm_mm))
364 return 0;
365#endif
366 *vm_flags &= ~VM_NOHUGEPAGE;
367 *vm_flags |= VM_HUGEPAGE;
368 /*
369 * If the vma become good for khugepaged to scan,
370 * register it here without waiting a page fault that
371 * may not happen any time soon.
372 */
373 khugepaged_enter_vma(vma, *vm_flags);
374 break;
375 case MADV_NOHUGEPAGE:
376 *vm_flags &= ~VM_HUGEPAGE;
377 *vm_flags |= VM_NOHUGEPAGE;
378 /*
379 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
380 * this vma even if we leave the mm registered in khugepaged if
381 * it got registered before VM_NOHUGEPAGE was set.
382 */
383 break;
384 }
385
386 return 0;
387}
388
389int __init khugepaged_init(void)
390{
391 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
392 sizeof(struct khugepaged_mm_slot),
393 __alignof__(struct khugepaged_mm_slot),
394 0, NULL);
395 if (!mm_slot_cache)
396 return -ENOMEM;
397
398 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
399 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
400 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
401 khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
402
403 return 0;
404}
405
406void __init khugepaged_destroy(void)
407{
408 kmem_cache_destroy(mm_slot_cache);
409}
410
411static inline int hpage_collapse_test_exit(struct mm_struct *mm)
412{
413 return atomic_read(&mm->mm_users) == 0;
414}
415
416void __khugepaged_enter(struct mm_struct *mm)
417{
418 struct khugepaged_mm_slot *mm_slot;
419 struct mm_slot *slot;
420 int wakeup;
421
422 mm_slot = mm_slot_alloc(mm_slot_cache);
423 if (!mm_slot)
424 return;
425
426 slot = &mm_slot->slot;
427
428 /* __khugepaged_exit() must not run from under us */
429 VM_BUG_ON_MM(hpage_collapse_test_exit(mm), mm);
430 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
431 mm_slot_free(mm_slot_cache, mm_slot);
432 return;
433 }
434
435 spin_lock(&khugepaged_mm_lock);
436 mm_slot_insert(mm_slots_hash, mm, slot);
437 /*
438 * Insert just behind the scanning cursor, to let the area settle
439 * down a little.
440 */
441 wakeup = list_empty(&khugepaged_scan.mm_head);
442 list_add_tail(&slot->mm_node, &khugepaged_scan.mm_head);
443 spin_unlock(&khugepaged_mm_lock);
444
445 mmgrab(mm);
446 if (wakeup)
447 wake_up_interruptible(&khugepaged_wait);
448}
449
450void khugepaged_enter_vma(struct vm_area_struct *vma,
451 unsigned long vm_flags)
452{
453 if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) &&
454 hugepage_flags_enabled()) {
455 if (hugepage_vma_check(vma, vm_flags, false, false, true))
456 __khugepaged_enter(vma->vm_mm);
457 }
458}
459
460void __khugepaged_exit(struct mm_struct *mm)
461{
462 struct khugepaged_mm_slot *mm_slot;
463 struct mm_slot *slot;
464 int free = 0;
465
466 spin_lock(&khugepaged_mm_lock);
467 slot = mm_slot_lookup(mm_slots_hash, mm);
468 mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
469 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
470 hash_del(&slot->hash);
471 list_del(&slot->mm_node);
472 free = 1;
473 }
474 spin_unlock(&khugepaged_mm_lock);
475
476 if (free) {
477 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
478 mm_slot_free(mm_slot_cache, mm_slot);
479 mmdrop(mm);
480 } else if (mm_slot) {
481 /*
482 * This is required to serialize against
483 * hpage_collapse_test_exit() (which is guaranteed to run
484 * under mmap sem read mode). Stop here (after we return all
485 * pagetables will be destroyed) until khugepaged has finished
486 * working on the pagetables under the mmap_lock.
487 */
488 mmap_write_lock(mm);
489 mmap_write_unlock(mm);
490 }
491}
492
493static void release_pte_page(struct page *page)
494{
495 mod_node_page_state(page_pgdat(page),
496 NR_ISOLATED_ANON + page_is_file_lru(page),
497 -compound_nr(page));
498 unlock_page(page);
499 putback_lru_page(page);
500}
501
502static void release_pte_pages(pte_t *pte, pte_t *_pte,
503 struct list_head *compound_pagelist)
504{
505 struct page *page, *tmp;
506
507 while (--_pte >= pte) {
508 pte_t pteval = *_pte;
509
510 page = pte_page(pteval);
511 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
512 !PageCompound(page))
513 release_pte_page(page);
514 }
515
516 list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
517 list_del(&page->lru);
518 release_pte_page(page);
519 }
520}
521
522static bool is_refcount_suitable(struct page *page)
523{
524 int expected_refcount;
525
526 expected_refcount = total_mapcount(page);
527 if (PageSwapCache(page))
528 expected_refcount += compound_nr(page);
529
530 return page_count(page) == expected_refcount;
531}
532
533static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
534 unsigned long address,
535 pte_t *pte,
536 struct collapse_control *cc,
537 struct list_head *compound_pagelist)
538{
539 struct page *page = NULL;
540 pte_t *_pte;
541 int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0;
542 bool writable = false;
543
544 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
545 _pte++, address += PAGE_SIZE) {
546 pte_t pteval = *_pte;
547 if (pte_none(pteval) || (pte_present(pteval) &&
548 is_zero_pfn(pte_pfn(pteval)))) {
549 ++none_or_zero;
550 if (!userfaultfd_armed(vma) &&
551 (!cc->is_khugepaged ||
552 none_or_zero <= khugepaged_max_ptes_none)) {
553 continue;
554 } else {
555 result = SCAN_EXCEED_NONE_PTE;
556 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
557 goto out;
558 }
559 }
560 if (!pte_present(pteval)) {
561 result = SCAN_PTE_NON_PRESENT;
562 goto out;
563 }
564 page = vm_normal_page(vma, address, pteval);
565 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
566 result = SCAN_PAGE_NULL;
567 goto out;
568 }
569
570 VM_BUG_ON_PAGE(!PageAnon(page), page);
571
572 if (page_mapcount(page) > 1) {
573 ++shared;
574 if (cc->is_khugepaged &&
575 shared > khugepaged_max_ptes_shared) {
576 result = SCAN_EXCEED_SHARED_PTE;
577 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
578 goto out;
579 }
580 }
581
582 if (PageCompound(page)) {
583 struct page *p;
584 page = compound_head(page);
585
586 /*
587 * Check if we have dealt with the compound page
588 * already
589 */
590 list_for_each_entry(p, compound_pagelist, lru) {
591 if (page == p)
592 goto next;
593 }
594 }
595
596 /*
597 * We can do it before isolate_lru_page because the
598 * page can't be freed from under us. NOTE: PG_lock
599 * is needed to serialize against split_huge_page
600 * when invoked from the VM.
601 */
602 if (!trylock_page(page)) {
603 result = SCAN_PAGE_LOCK;
604 goto out;
605 }
606
607 /*
608 * Check if the page has any GUP (or other external) pins.
609 *
610 * The page table that maps the page has been already unlinked
611 * from the page table tree and this process cannot get
612 * an additional pin on the page.
613 *
614 * New pins can come later if the page is shared across fork,
615 * but not from this process. The other process cannot write to
616 * the page, only trigger CoW.
617 */
618 if (!is_refcount_suitable(page)) {
619 unlock_page(page);
620 result = SCAN_PAGE_COUNT;
621 goto out;
622 }
623
624 /*
625 * Isolate the page to avoid collapsing an hugepage
626 * currently in use by the VM.
627 */
628 if (isolate_lru_page(page)) {
629 unlock_page(page);
630 result = SCAN_DEL_PAGE_LRU;
631 goto out;
632 }
633 mod_node_page_state(page_pgdat(page),
634 NR_ISOLATED_ANON + page_is_file_lru(page),
635 compound_nr(page));
636 VM_BUG_ON_PAGE(!PageLocked(page), page);
637 VM_BUG_ON_PAGE(PageLRU(page), page);
638
639 if (PageCompound(page))
640 list_add_tail(&page->lru, compound_pagelist);
641next:
642 /*
643 * If collapse was initiated by khugepaged, check that there is
644 * enough young pte to justify collapsing the page
645 */
646 if (cc->is_khugepaged &&
647 (pte_young(pteval) || page_is_young(page) ||
648 PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm,
649 address)))
650 referenced++;
651
652 if (pte_write(pteval))
653 writable = true;
654 }
655
656 if (unlikely(!writable)) {
657 result = SCAN_PAGE_RO;
658 } else if (unlikely(cc->is_khugepaged && !referenced)) {
659 result = SCAN_LACK_REFERENCED_PAGE;
660 } else {
661 result = SCAN_SUCCEED;
662 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
663 referenced, writable, result);
664 return result;
665 }
666out:
667 release_pte_pages(pte, _pte, compound_pagelist);
668 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
669 referenced, writable, result);
670 return result;
671}
672
673static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
674 struct vm_area_struct *vma,
675 unsigned long address,
676 spinlock_t *ptl,
677 struct list_head *compound_pagelist)
678{
679 struct page *src_page, *tmp;
680 pte_t *_pte;
681 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
682 _pte++, page++, address += PAGE_SIZE) {
683 pte_t pteval = *_pte;
684
685 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
686 clear_user_highpage(page, address);
687 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
688 if (is_zero_pfn(pte_pfn(pteval))) {
689 /*
690 * ptl mostly unnecessary.
691 */
692 spin_lock(ptl);
693 ptep_clear(vma->vm_mm, address, _pte);
694 spin_unlock(ptl);
695 }
696 } else {
697 src_page = pte_page(pteval);
698 copy_user_highpage(page, src_page, address, vma);
699 if (!PageCompound(src_page))
700 release_pte_page(src_page);
701 /*
702 * ptl mostly unnecessary, but preempt has to
703 * be disabled to update the per-cpu stats
704 * inside page_remove_rmap().
705 */
706 spin_lock(ptl);
707 ptep_clear(vma->vm_mm, address, _pte);
708 page_remove_rmap(src_page, vma, false);
709 spin_unlock(ptl);
710 free_page_and_swap_cache(src_page);
711 }
712 }
713
714 list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
715 list_del(&src_page->lru);
716 mod_node_page_state(page_pgdat(src_page),
717 NR_ISOLATED_ANON + page_is_file_lru(src_page),
718 -compound_nr(src_page));
719 unlock_page(src_page);
720 free_swap_cache(src_page);
721 putback_lru_page(src_page);
722 }
723}
724
725static void khugepaged_alloc_sleep(void)
726{
727 DEFINE_WAIT(wait);
728
729 add_wait_queue(&khugepaged_wait, &wait);
730 __set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
731 schedule_timeout(msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
732 remove_wait_queue(&khugepaged_wait, &wait);
733}
734
735struct collapse_control khugepaged_collapse_control = {
736 .is_khugepaged = true,
737};
738
739static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc)
740{
741 int i;
742
743 /*
744 * If node_reclaim_mode is disabled, then no extra effort is made to
745 * allocate memory locally.
746 */
747 if (!node_reclaim_enabled())
748 return false;
749
750 /* If there is a count for this node already, it must be acceptable */
751 if (cc->node_load[nid])
752 return false;
753
754 for (i = 0; i < MAX_NUMNODES; i++) {
755 if (!cc->node_load[i])
756 continue;
757 if (node_distance(nid, i) > node_reclaim_distance)
758 return true;
759 }
760 return false;
761}
762
763#define khugepaged_defrag() \
764 (transparent_hugepage_flags & \
765 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG))
766
767/* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
768static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
769{
770 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
771}
772
773#ifdef CONFIG_NUMA
774static int hpage_collapse_find_target_node(struct collapse_control *cc)
775{
776 int nid, target_node = 0, max_value = 0;
777
778 /* find first node with max normal pages hit */
779 for (nid = 0; nid < MAX_NUMNODES; nid++)
780 if (cc->node_load[nid] > max_value) {
781 max_value = cc->node_load[nid];
782 target_node = nid;
783 }
784
785 for_each_online_node(nid) {
786 if (max_value == cc->node_load[nid])
787 node_set(nid, cc->alloc_nmask);
788 }
789
790 return target_node;
791}
792#else
793static int hpage_collapse_find_target_node(struct collapse_control *cc)
794{
795 return 0;
796}
797#endif
798
799static bool hpage_collapse_alloc_page(struct page **hpage, gfp_t gfp, int node,
800 nodemask_t *nmask)
801{
802 *hpage = __alloc_pages(gfp, HPAGE_PMD_ORDER, node, nmask);
803 if (unlikely(!*hpage)) {
804 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
805 return false;
806 }
807
808 prep_transhuge_page(*hpage);
809 count_vm_event(THP_COLLAPSE_ALLOC);
810 return true;
811}
812
813/*
814 * If mmap_lock temporarily dropped, revalidate vma
815 * before taking mmap_lock.
816 * Returns enum scan_result value.
817 */
818
819static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
820 bool expect_anon,
821 struct vm_area_struct **vmap,
822 struct collapse_control *cc)
823{
824 struct vm_area_struct *vma;
825
826 if (unlikely(hpage_collapse_test_exit(mm)))
827 return SCAN_ANY_PROCESS;
828
829 *vmap = vma = find_vma(mm, address);
830 if (!vma)
831 return SCAN_VMA_NULL;
832
833 if (!transhuge_vma_suitable(vma, address))
834 return SCAN_ADDRESS_RANGE;
835 if (!hugepage_vma_check(vma, vma->vm_flags, false, false,
836 cc->is_khugepaged))
837 return SCAN_VMA_CHECK;
838 /*
839 * Anon VMA expected, the address may be unmapped then
840 * remapped to file after khugepaged reaquired the mmap_lock.
841 *
842 * hugepage_vma_check may return true for qualified file
843 * vmas.
844 */
845 if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap)))
846 return SCAN_PAGE_ANON;
847 return SCAN_SUCCEED;
848}
849
850/*
851 * See pmd_trans_unstable() for how the result may change out from
852 * underneath us, even if we hold mmap_lock in read.
853 */
854static int find_pmd_or_thp_or_none(struct mm_struct *mm,
855 unsigned long address,
856 pmd_t **pmd)
857{
858 pmd_t pmde;
859
860 *pmd = mm_find_pmd(mm, address);
861 if (!*pmd)
862 return SCAN_PMD_NULL;
863
864 pmde = pmdp_get_lockless(*pmd);
865
866#ifdef CONFIG_TRANSPARENT_HUGEPAGE
867 /* See comments in pmd_none_or_trans_huge_or_clear_bad() */
868 barrier();
869#endif
870 if (pmd_none(pmde))
871 return SCAN_PMD_NONE;
872 if (!pmd_present(pmde))
873 return SCAN_PMD_NULL;
874 if (pmd_trans_huge(pmde))
875 return SCAN_PMD_MAPPED;
876 if (pmd_devmap(pmde))
877 return SCAN_PMD_NULL;
878 if (pmd_bad(pmde))
879 return SCAN_PMD_NULL;
880 return SCAN_SUCCEED;
881}
882
883static int check_pmd_still_valid(struct mm_struct *mm,
884 unsigned long address,
885 pmd_t *pmd)
886{
887 pmd_t *new_pmd;
888 int result = find_pmd_or_thp_or_none(mm, address, &new_pmd);
889
890 if (result != SCAN_SUCCEED)
891 return result;
892 if (new_pmd != pmd)
893 return SCAN_FAIL;
894 return SCAN_SUCCEED;
895}
896
897/*
898 * Bring missing pages in from swap, to complete THP collapse.
899 * Only done if hpage_collapse_scan_pmd believes it is worthwhile.
900 *
901 * Called and returns without pte mapped or spinlocks held.
902 * Note that if false is returned, mmap_lock will be released.
903 */
904
905static int __collapse_huge_page_swapin(struct mm_struct *mm,
906 struct vm_area_struct *vma,
907 unsigned long haddr, pmd_t *pmd,
908 int referenced)
909{
910 int swapped_in = 0;
911 vm_fault_t ret = 0;
912 unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
913
914 for (address = haddr; address < end; address += PAGE_SIZE) {
915 struct vm_fault vmf = {
916 .vma = vma,
917 .address = address,
918 .pgoff = linear_page_index(vma, haddr),
919 .flags = FAULT_FLAG_ALLOW_RETRY,
920 .pmd = pmd,
921 };
922
923 vmf.pte = pte_offset_map(pmd, address);
924 vmf.orig_pte = *vmf.pte;
925 if (!is_swap_pte(vmf.orig_pte)) {
926 pte_unmap(vmf.pte);
927 continue;
928 }
929 ret = do_swap_page(&vmf);
930
931 /*
932 * do_swap_page returns VM_FAULT_RETRY with released mmap_lock.
933 * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because
934 * we do not retry here and swap entry will remain in pagetable
935 * resulting in later failure.
936 */
937 if (ret & VM_FAULT_RETRY) {
938 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
939 /* Likely, but not guaranteed, that page lock failed */
940 return SCAN_PAGE_LOCK;
941 }
942 if (ret & VM_FAULT_ERROR) {
943 mmap_read_unlock(mm);
944 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
945 return SCAN_FAIL;
946 }
947 swapped_in++;
948 }
949
950 /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
951 if (swapped_in)
952 lru_add_drain();
953
954 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
955 return SCAN_SUCCEED;
956}
957
958static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm,
959 struct collapse_control *cc)
960{
961 gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() :
962 GFP_TRANSHUGE);
963 int node = hpage_collapse_find_target_node(cc);
964
965 if (!hpage_collapse_alloc_page(hpage, gfp, node, &cc->alloc_nmask))
966 return SCAN_ALLOC_HUGE_PAGE_FAIL;
967 if (unlikely(mem_cgroup_charge(page_folio(*hpage), mm, gfp)))
968 return SCAN_CGROUP_CHARGE_FAIL;
969 count_memcg_page_event(*hpage, THP_COLLAPSE_ALLOC);
970 return SCAN_SUCCEED;
971}
972
973static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
974 int referenced, int unmapped,
975 struct collapse_control *cc)
976{
977 LIST_HEAD(compound_pagelist);
978 pmd_t *pmd, _pmd;
979 pte_t *pte;
980 pgtable_t pgtable;
981 struct page *hpage;
982 spinlock_t *pmd_ptl, *pte_ptl;
983 int result = SCAN_FAIL;
984 struct vm_area_struct *vma;
985 struct mmu_notifier_range range;
986
987 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
988
989 /*
990 * Before allocating the hugepage, release the mmap_lock read lock.
991 * The allocation can take potentially a long time if it involves
992 * sync compaction, and we do not need to hold the mmap_lock during
993 * that. We will recheck the vma after taking it again in write mode.
994 */
995 mmap_read_unlock(mm);
996
997 result = alloc_charge_hpage(&hpage, mm, cc);
998 if (result != SCAN_SUCCEED)
999 goto out_nolock;
1000
1001 mmap_read_lock(mm);
1002 result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1003 if (result != SCAN_SUCCEED) {
1004 mmap_read_unlock(mm);
1005 goto out_nolock;
1006 }
1007
1008 result = find_pmd_or_thp_or_none(mm, address, &pmd);
1009 if (result != SCAN_SUCCEED) {
1010 mmap_read_unlock(mm);
1011 goto out_nolock;
1012 }
1013
1014 if (unmapped) {
1015 /*
1016 * __collapse_huge_page_swapin will return with mmap_lock
1017 * released when it fails. So we jump out_nolock directly in
1018 * that case. Continuing to collapse causes inconsistency.
1019 */
1020 result = __collapse_huge_page_swapin(mm, vma, address, pmd,
1021 referenced);
1022 if (result != SCAN_SUCCEED)
1023 goto out_nolock;
1024 }
1025
1026 mmap_read_unlock(mm);
1027 /*
1028 * Prevent all access to pagetables with the exception of
1029 * gup_fast later handled by the ptep_clear_flush and the VM
1030 * handled by the anon_vma lock + PG_lock.
1031 */
1032 mmap_write_lock(mm);
1033 result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1034 if (result != SCAN_SUCCEED)
1035 goto out_up_write;
1036 /* check if the pmd is still valid */
1037 result = check_pmd_still_valid(mm, address, pmd);
1038 if (result != SCAN_SUCCEED)
1039 goto out_up_write;
1040
1041 anon_vma_lock_write(vma->anon_vma);
1042
1043 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1044 address, address + HPAGE_PMD_SIZE);
1045 mmu_notifier_invalidate_range_start(&range);
1046
1047 pte = pte_offset_map(pmd, address);
1048 pte_ptl = pte_lockptr(mm, pmd);
1049
1050 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1051 /*
1052 * This removes any huge TLB entry from the CPU so we won't allow
1053 * huge and small TLB entries for the same virtual address to
1054 * avoid the risk of CPU bugs in that area.
1055 *
1056 * Parallel fast GUP is fine since fast GUP will back off when
1057 * it detects PMD is changed.
1058 */
1059 _pmd = pmdp_collapse_flush(vma, address, pmd);
1060 spin_unlock(pmd_ptl);
1061 mmu_notifier_invalidate_range_end(&range);
1062 tlb_remove_table_sync_one();
1063
1064 spin_lock(pte_ptl);
1065 result = __collapse_huge_page_isolate(vma, address, pte, cc,
1066 &compound_pagelist);
1067 spin_unlock(pte_ptl);
1068
1069 if (unlikely(result != SCAN_SUCCEED)) {
1070 pte_unmap(pte);
1071 spin_lock(pmd_ptl);
1072 BUG_ON(!pmd_none(*pmd));
1073 /*
1074 * We can only use set_pmd_at when establishing
1075 * hugepmds and never for establishing regular pmds that
1076 * points to regular pagetables. Use pmd_populate for that
1077 */
1078 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1079 spin_unlock(pmd_ptl);
1080 anon_vma_unlock_write(vma->anon_vma);
1081 goto out_up_write;
1082 }
1083
1084 /*
1085 * All pages are isolated and locked so anon_vma rmap
1086 * can't run anymore.
1087 */
1088 anon_vma_unlock_write(vma->anon_vma);
1089
1090 __collapse_huge_page_copy(pte, hpage, vma, address, pte_ptl,
1091 &compound_pagelist);
1092 pte_unmap(pte);
1093 /*
1094 * spin_lock() below is not the equivalent of smp_wmb(), but
1095 * the smp_wmb() inside __SetPageUptodate() can be reused to
1096 * avoid the copy_huge_page writes to become visible after
1097 * the set_pmd_at() write.
1098 */
1099 __SetPageUptodate(hpage);
1100 pgtable = pmd_pgtable(_pmd);
1101
1102 _pmd = mk_huge_pmd(hpage, vma->vm_page_prot);
1103 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1104
1105 spin_lock(pmd_ptl);
1106 BUG_ON(!pmd_none(*pmd));
1107 page_add_new_anon_rmap(hpage, vma, address);
1108 lru_cache_add_inactive_or_unevictable(hpage, vma);
1109 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1110 set_pmd_at(mm, address, pmd, _pmd);
1111 update_mmu_cache_pmd(vma, address, pmd);
1112 spin_unlock(pmd_ptl);
1113
1114 hpage = NULL;
1115
1116 result = SCAN_SUCCEED;
1117out_up_write:
1118 mmap_write_unlock(mm);
1119out_nolock:
1120 if (hpage) {
1121 mem_cgroup_uncharge(page_folio(hpage));
1122 put_page(hpage);
1123 }
1124 trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result);
1125 return result;
1126}
1127
1128static int hpage_collapse_scan_pmd(struct mm_struct *mm,
1129 struct vm_area_struct *vma,
1130 unsigned long address, bool *mmap_locked,
1131 struct collapse_control *cc)
1132{
1133 pmd_t *pmd;
1134 pte_t *pte, *_pte;
1135 int result = SCAN_FAIL, referenced = 0;
1136 int none_or_zero = 0, shared = 0;
1137 struct page *page = NULL;
1138 unsigned long _address;
1139 spinlock_t *ptl;
1140 int node = NUMA_NO_NODE, unmapped = 0;
1141 bool writable = false;
1142
1143 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1144
1145 result = find_pmd_or_thp_or_none(mm, address, &pmd);
1146 if (result != SCAN_SUCCEED)
1147 goto out;
1148
1149 memset(cc->node_load, 0, sizeof(cc->node_load));
1150 nodes_clear(cc->alloc_nmask);
1151 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1152 for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR;
1153 _pte++, _address += PAGE_SIZE) {
1154 pte_t pteval = *_pte;
1155 if (is_swap_pte(pteval)) {
1156 ++unmapped;
1157 if (!cc->is_khugepaged ||
1158 unmapped <= khugepaged_max_ptes_swap) {
1159 /*
1160 * Always be strict with uffd-wp
1161 * enabled swap entries. Please see
1162 * comment below for pte_uffd_wp().
1163 */
1164 if (pte_swp_uffd_wp(pteval)) {
1165 result = SCAN_PTE_UFFD_WP;
1166 goto out_unmap;
1167 }
1168 continue;
1169 } else {
1170 result = SCAN_EXCEED_SWAP_PTE;
1171 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1172 goto out_unmap;
1173 }
1174 }
1175 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1176 ++none_or_zero;
1177 if (!userfaultfd_armed(vma) &&
1178 (!cc->is_khugepaged ||
1179 none_or_zero <= khugepaged_max_ptes_none)) {
1180 continue;
1181 } else {
1182 result = SCAN_EXCEED_NONE_PTE;
1183 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1184 goto out_unmap;
1185 }
1186 }
1187 if (pte_uffd_wp(pteval)) {
1188 /*
1189 * Don't collapse the page if any of the small
1190 * PTEs are armed with uffd write protection.
1191 * Here we can also mark the new huge pmd as
1192 * write protected if any of the small ones is
1193 * marked but that could bring unknown
1194 * userfault messages that falls outside of
1195 * the registered range. So, just be simple.
1196 */
1197 result = SCAN_PTE_UFFD_WP;
1198 goto out_unmap;
1199 }
1200 if (pte_write(pteval))
1201 writable = true;
1202
1203 page = vm_normal_page(vma, _address, pteval);
1204 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
1205 result = SCAN_PAGE_NULL;
1206 goto out_unmap;
1207 }
1208
1209 if (page_mapcount(page) > 1) {
1210 ++shared;
1211 if (cc->is_khugepaged &&
1212 shared > khugepaged_max_ptes_shared) {
1213 result = SCAN_EXCEED_SHARED_PTE;
1214 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1215 goto out_unmap;
1216 }
1217 }
1218
1219 page = compound_head(page);
1220
1221 /*
1222 * Record which node the original page is from and save this
1223 * information to cc->node_load[].
1224 * Khugepaged will allocate hugepage from the node has the max
1225 * hit record.
1226 */
1227 node = page_to_nid(page);
1228 if (hpage_collapse_scan_abort(node, cc)) {
1229 result = SCAN_SCAN_ABORT;
1230 goto out_unmap;
1231 }
1232 cc->node_load[node]++;
1233 if (!PageLRU(page)) {
1234 result = SCAN_PAGE_LRU;
1235 goto out_unmap;
1236 }
1237 if (PageLocked(page)) {
1238 result = SCAN_PAGE_LOCK;
1239 goto out_unmap;
1240 }
1241 if (!PageAnon(page)) {
1242 result = SCAN_PAGE_ANON;
1243 goto out_unmap;
1244 }
1245
1246 /*
1247 * Check if the page has any GUP (or other external) pins.
1248 *
1249 * Here the check may be racy:
1250 * it may see total_mapcount > refcount in some cases?
1251 * But such case is ephemeral we could always retry collapse
1252 * later. However it may report false positive if the page
1253 * has excessive GUP pins (i.e. 512). Anyway the same check
1254 * will be done again later the risk seems low.
1255 */
1256 if (!is_refcount_suitable(page)) {
1257 result = SCAN_PAGE_COUNT;
1258 goto out_unmap;
1259 }
1260
1261 /*
1262 * If collapse was initiated by khugepaged, check that there is
1263 * enough young pte to justify collapsing the page
1264 */
1265 if (cc->is_khugepaged &&
1266 (pte_young(pteval) || page_is_young(page) ||
1267 PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm,
1268 address)))
1269 referenced++;
1270 }
1271 if (!writable) {
1272 result = SCAN_PAGE_RO;
1273 } else if (cc->is_khugepaged &&
1274 (!referenced ||
1275 (unmapped && referenced < HPAGE_PMD_NR / 2))) {
1276 result = SCAN_LACK_REFERENCED_PAGE;
1277 } else {
1278 result = SCAN_SUCCEED;
1279 }
1280out_unmap:
1281 pte_unmap_unlock(pte, ptl);
1282 if (result == SCAN_SUCCEED) {
1283 result = collapse_huge_page(mm, address, referenced,
1284 unmapped, cc);
1285 /* collapse_huge_page will return with the mmap_lock released */
1286 *mmap_locked = false;
1287 }
1288out:
1289 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1290 none_or_zero, result, unmapped);
1291 return result;
1292}
1293
1294static void collect_mm_slot(struct khugepaged_mm_slot *mm_slot)
1295{
1296 struct mm_slot *slot = &mm_slot->slot;
1297 struct mm_struct *mm = slot->mm;
1298
1299 lockdep_assert_held(&khugepaged_mm_lock);
1300
1301 if (hpage_collapse_test_exit(mm)) {
1302 /* free mm_slot */
1303 hash_del(&slot->hash);
1304 list_del(&slot->mm_node);
1305
1306 /*
1307 * Not strictly needed because the mm exited already.
1308 *
1309 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1310 */
1311
1312 /* khugepaged_mm_lock actually not necessary for the below */
1313 mm_slot_free(mm_slot_cache, mm_slot);
1314 mmdrop(mm);
1315 }
1316}
1317
1318#ifdef CONFIG_SHMEM
1319/*
1320 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1321 * khugepaged should try to collapse the page table.
1322 *
1323 * Note that following race exists:
1324 * (1) khugepaged calls khugepaged_collapse_pte_mapped_thps() for mm_struct A,
1325 * emptying the A's ->pte_mapped_thp[] array.
1326 * (2) MADV_COLLAPSE collapses some file extent with target mm_struct B, and
1327 * retract_page_tables() finds a VMA in mm_struct A mapping the same extent
1328 * (at virtual address X) and adds an entry (for X) into mm_struct A's
1329 * ->pte-mapped_thp[] array.
1330 * (3) khugepaged calls khugepaged_collapse_scan_file() for mm_struct A at X,
1331 * sees a pte-mapped THP (SCAN_PTE_MAPPED_HUGEPAGE) and adds an entry
1332 * (for X) into mm_struct A's ->pte-mapped_thp[] array.
1333 * Thus, it's possible the same address is added multiple times for the same
1334 * mm_struct. Should this happen, we'll simply attempt
1335 * collapse_pte_mapped_thp() multiple times for the same address, under the same
1336 * exclusive mmap_lock, and assuming the first call is successful, subsequent
1337 * attempts will return quickly (without grabbing any additional locks) when
1338 * a huge pmd is found in find_pmd_or_thp_or_none(). Since this is a cheap
1339 * check, and since this is a rare occurrence, the cost of preventing this
1340 * "multiple-add" is thought to be more expensive than just handling it, should
1341 * it occur.
1342 */
1343static bool khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1344 unsigned long addr)
1345{
1346 struct khugepaged_mm_slot *mm_slot;
1347 struct mm_slot *slot;
1348 bool ret = false;
1349
1350 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1351
1352 spin_lock(&khugepaged_mm_lock);
1353 slot = mm_slot_lookup(mm_slots_hash, mm);
1354 mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
1355 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP)) {
1356 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1357 ret = true;
1358 }
1359 spin_unlock(&khugepaged_mm_lock);
1360 return ret;
1361}
1362
1363/* hpage must be locked, and mmap_lock must be held in write */
1364static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
1365 pmd_t *pmdp, struct page *hpage)
1366{
1367 struct vm_fault vmf = {
1368 .vma = vma,
1369 .address = addr,
1370 .flags = 0,
1371 .pmd = pmdp,
1372 };
1373
1374 VM_BUG_ON(!PageTransHuge(hpage));
1375 mmap_assert_write_locked(vma->vm_mm);
1376
1377 if (do_set_pmd(&vmf, hpage))
1378 return SCAN_FAIL;
1379
1380 get_page(hpage);
1381 return SCAN_SUCCEED;
1382}
1383
1384/*
1385 * A note about locking:
1386 * Trying to take the page table spinlocks would be useless here because those
1387 * are only used to synchronize:
1388 *
1389 * - modifying terminal entries (ones that point to a data page, not to another
1390 * page table)
1391 * - installing *new* non-terminal entries
1392 *
1393 * Instead, we need roughly the same kind of protection as free_pgtables() or
1394 * mm_take_all_locks() (but only for a single VMA):
1395 * The mmap lock together with this VMA's rmap locks covers all paths towards
1396 * the page table entries we're messing with here, except for hardware page
1397 * table walks and lockless_pages_from_mm().
1398 */
1399static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
1400 unsigned long addr, pmd_t *pmdp)
1401{
1402 pmd_t pmd;
1403 struct mmu_notifier_range range;
1404
1405 mmap_assert_write_locked(mm);
1406 if (vma->vm_file)
1407 lockdep_assert_held_write(&vma->vm_file->f_mapping->i_mmap_rwsem);
1408 /*
1409 * All anon_vmas attached to the VMA have the same root and are
1410 * therefore locked by the same lock.
1411 */
1412 if (vma->anon_vma)
1413 lockdep_assert_held_write(&vma->anon_vma->root->rwsem);
1414
1415 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, addr,
1416 addr + HPAGE_PMD_SIZE);
1417 mmu_notifier_invalidate_range_start(&range);
1418 pmd = pmdp_collapse_flush(vma, addr, pmdp);
1419 tlb_remove_table_sync_one();
1420 mmu_notifier_invalidate_range_end(&range);
1421 mm_dec_nr_ptes(mm);
1422 page_table_check_pte_clear_range(mm, addr, pmd);
1423 pte_free(mm, pmd_pgtable(pmd));
1424}
1425
1426/**
1427 * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1428 * address haddr.
1429 *
1430 * @mm: process address space where collapse happens
1431 * @addr: THP collapse address
1432 * @install_pmd: If a huge PMD should be installed
1433 *
1434 * This function checks whether all the PTEs in the PMD are pointing to the
1435 * right THP. If so, retract the page table so the THP can refault in with
1436 * as pmd-mapped. Possibly install a huge PMD mapping the THP.
1437 */
1438int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
1439 bool install_pmd)
1440{
1441 unsigned long haddr = addr & HPAGE_PMD_MASK;
1442 struct vm_area_struct *vma = vma_lookup(mm, haddr);
1443 struct page *hpage;
1444 pte_t *start_pte, *pte;
1445 pmd_t *pmd;
1446 spinlock_t *ptl;
1447 int count = 0, result = SCAN_FAIL;
1448 int i;
1449
1450 mmap_assert_write_locked(mm);
1451
1452 /* Fast check before locking page if already PMD-mapped */
1453 result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
1454 if (result == SCAN_PMD_MAPPED)
1455 return result;
1456
1457 if (!vma || !vma->vm_file ||
1458 !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1459 return SCAN_VMA_CHECK;
1460
1461 /*
1462 * If we are here, we've succeeded in replacing all the native pages
1463 * in the page cache with a single hugepage. If a mm were to fault-in
1464 * this memory (mapped by a suitably aligned VMA), we'd get the hugepage
1465 * and map it by a PMD, regardless of sysfs THP settings. As such, let's
1466 * analogously elide sysfs THP settings here.
1467 */
1468 if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false))
1469 return SCAN_VMA_CHECK;
1470
1471 /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */
1472 if (userfaultfd_wp(vma))
1473 return SCAN_PTE_UFFD_WP;
1474
1475 hpage = find_lock_page(vma->vm_file->f_mapping,
1476 linear_page_index(vma, haddr));
1477 if (!hpage)
1478 return SCAN_PAGE_NULL;
1479
1480 if (!PageHead(hpage)) {
1481 result = SCAN_FAIL;
1482 goto drop_hpage;
1483 }
1484
1485 if (compound_order(hpage) != HPAGE_PMD_ORDER) {
1486 result = SCAN_PAGE_COMPOUND;
1487 goto drop_hpage;
1488 }
1489
1490 switch (result) {
1491 case SCAN_SUCCEED:
1492 break;
1493 case SCAN_PMD_NONE:
1494 /*
1495 * In MADV_COLLAPSE path, possible race with khugepaged where
1496 * all pte entries have been removed and pmd cleared. If so,
1497 * skip all the pte checks and just update the pmd mapping.
1498 */
1499 goto maybe_install_pmd;
1500 default:
1501 goto drop_hpage;
1502 }
1503
1504 /*
1505 * We need to lock the mapping so that from here on, only GUP-fast and
1506 * hardware page walks can access the parts of the page tables that
1507 * we're operating on.
1508 * See collapse_and_free_pmd().
1509 */
1510 i_mmap_lock_write(vma->vm_file->f_mapping);
1511
1512 /*
1513 * This spinlock should be unnecessary: Nobody else should be accessing
1514 * the page tables under spinlock protection here, only
1515 * lockless_pages_from_mm() and the hardware page walker can access page
1516 * tables while all the high-level locks are held in write mode.
1517 */
1518 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1519 result = SCAN_FAIL;
1520
1521 /* step 1: check all mapped PTEs are to the right huge page */
1522 for (i = 0, addr = haddr, pte = start_pte;
1523 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1524 struct page *page;
1525
1526 /* empty pte, skip */
1527 if (pte_none(*pte))
1528 continue;
1529
1530 /* page swapped out, abort */
1531 if (!pte_present(*pte)) {
1532 result = SCAN_PTE_NON_PRESENT;
1533 goto abort;
1534 }
1535
1536 page = vm_normal_page(vma, addr, *pte);
1537 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1538 page = NULL;
1539 /*
1540 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1541 * page table, but the new page will not be a subpage of hpage.
1542 */
1543 if (hpage + i != page)
1544 goto abort;
1545 count++;
1546 }
1547
1548 /* step 2: adjust rmap */
1549 for (i = 0, addr = haddr, pte = start_pte;
1550 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1551 struct page *page;
1552
1553 if (pte_none(*pte))
1554 continue;
1555 page = vm_normal_page(vma, addr, *pte);
1556 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1557 goto abort;
1558 page_remove_rmap(page, vma, false);
1559 }
1560
1561 pte_unmap_unlock(start_pte, ptl);
1562
1563 /* step 3: set proper refcount and mm_counters. */
1564 if (count) {
1565 page_ref_sub(hpage, count);
1566 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1567 }
1568
1569 /* step 4: remove pte entries */
1570 /* we make no change to anon, but protect concurrent anon page lookup */
1571 if (vma->anon_vma)
1572 anon_vma_lock_write(vma->anon_vma);
1573
1574 collapse_and_free_pmd(mm, vma, haddr, pmd);
1575
1576 if (vma->anon_vma)
1577 anon_vma_unlock_write(vma->anon_vma);
1578 i_mmap_unlock_write(vma->vm_file->f_mapping);
1579
1580maybe_install_pmd:
1581 /* step 5: install pmd entry */
1582 result = install_pmd
1583 ? set_huge_pmd(vma, haddr, pmd, hpage)
1584 : SCAN_SUCCEED;
1585
1586drop_hpage:
1587 unlock_page(hpage);
1588 put_page(hpage);
1589 return result;
1590
1591abort:
1592 pte_unmap_unlock(start_pte, ptl);
1593 i_mmap_unlock_write(vma->vm_file->f_mapping);
1594 goto drop_hpage;
1595}
1596
1597static void khugepaged_collapse_pte_mapped_thps(struct khugepaged_mm_slot *mm_slot)
1598{
1599 struct mm_slot *slot = &mm_slot->slot;
1600 struct mm_struct *mm = slot->mm;
1601 int i;
1602
1603 if (likely(mm_slot->nr_pte_mapped_thp == 0))
1604 return;
1605
1606 if (!mmap_write_trylock(mm))
1607 return;
1608
1609 if (unlikely(hpage_collapse_test_exit(mm)))
1610 goto out;
1611
1612 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1613 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i], false);
1614
1615out:
1616 mm_slot->nr_pte_mapped_thp = 0;
1617 mmap_write_unlock(mm);
1618}
1619
1620static int retract_page_tables(struct address_space *mapping, pgoff_t pgoff,
1621 struct mm_struct *target_mm,
1622 unsigned long target_addr, struct page *hpage,
1623 struct collapse_control *cc)
1624{
1625 struct vm_area_struct *vma;
1626 int target_result = SCAN_FAIL;
1627
1628 i_mmap_lock_write(mapping);
1629 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1630 int result = SCAN_FAIL;
1631 struct mm_struct *mm = NULL;
1632 unsigned long addr = 0;
1633 pmd_t *pmd;
1634 bool is_target = false;
1635
1636 /*
1637 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1638 * got written to. These VMAs are likely not worth investing
1639 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1640 * later.
1641 *
1642 * Note that vma->anon_vma check is racy: it can be set up after
1643 * the check but before we took mmap_lock by the fault path.
1644 * But page lock would prevent establishing any new ptes of the
1645 * page, so we are safe.
1646 *
1647 * An alternative would be drop the check, but check that page
1648 * table is clear before calling pmdp_collapse_flush() under
1649 * ptl. It has higher chance to recover THP for the VMA, but
1650 * has higher cost too. It would also probably require locking
1651 * the anon_vma.
1652 */
1653 if (READ_ONCE(vma->anon_vma)) {
1654 result = SCAN_PAGE_ANON;
1655 goto next;
1656 }
1657 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1658 if (addr & ~HPAGE_PMD_MASK ||
1659 vma->vm_end < addr + HPAGE_PMD_SIZE) {
1660 result = SCAN_VMA_CHECK;
1661 goto next;
1662 }
1663 mm = vma->vm_mm;
1664 is_target = mm == target_mm && addr == target_addr;
1665 result = find_pmd_or_thp_or_none(mm, addr, &pmd);
1666 if (result != SCAN_SUCCEED)
1667 goto next;
1668 /*
1669 * We need exclusive mmap_lock to retract page table.
1670 *
1671 * We use trylock due to lock inversion: we need to acquire
1672 * mmap_lock while holding page lock. Fault path does it in
1673 * reverse order. Trylock is a way to avoid deadlock.
1674 *
1675 * Also, it's not MADV_COLLAPSE's job to collapse other
1676 * mappings - let khugepaged take care of them later.
1677 */
1678 result = SCAN_PTE_MAPPED_HUGEPAGE;
1679 if ((cc->is_khugepaged || is_target) &&
1680 mmap_write_trylock(mm)) {
1681 /*
1682 * Re-check whether we have an ->anon_vma, because
1683 * collapse_and_free_pmd() requires that either no
1684 * ->anon_vma exists or the anon_vma is locked.
1685 * We already checked ->anon_vma above, but that check
1686 * is racy because ->anon_vma can be populated under the
1687 * mmap lock in read mode.
1688 */
1689 if (vma->anon_vma) {
1690 result = SCAN_PAGE_ANON;
1691 goto unlock_next;
1692 }
1693 /*
1694 * When a vma is registered with uffd-wp, we can't
1695 * recycle the pmd pgtable because there can be pte
1696 * markers installed. Skip it only, so the rest mm/vma
1697 * can still have the same file mapped hugely, however
1698 * it'll always mapped in small page size for uffd-wp
1699 * registered ranges.
1700 */
1701 if (hpage_collapse_test_exit(mm)) {
1702 result = SCAN_ANY_PROCESS;
1703 goto unlock_next;
1704 }
1705 if (userfaultfd_wp(vma)) {
1706 result = SCAN_PTE_UFFD_WP;
1707 goto unlock_next;
1708 }
1709 collapse_and_free_pmd(mm, vma, addr, pmd);
1710 if (!cc->is_khugepaged && is_target)
1711 result = set_huge_pmd(vma, addr, pmd, hpage);
1712 else
1713 result = SCAN_SUCCEED;
1714
1715unlock_next:
1716 mmap_write_unlock(mm);
1717 goto next;
1718 }
1719 /*
1720 * Calling context will handle target mm/addr. Otherwise, let
1721 * khugepaged try again later.
1722 */
1723 if (!is_target) {
1724 khugepaged_add_pte_mapped_thp(mm, addr);
1725 continue;
1726 }
1727next:
1728 if (is_target)
1729 target_result = result;
1730 }
1731 i_mmap_unlock_write(mapping);
1732 return target_result;
1733}
1734
1735/**
1736 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1737 *
1738 * @mm: process address space where collapse happens
1739 * @addr: virtual collapse start address
1740 * @file: file that collapse on
1741 * @start: collapse start address
1742 * @cc: collapse context and scratchpad
1743 *
1744 * Basic scheme is simple, details are more complex:
1745 * - allocate and lock a new huge page;
1746 * - scan page cache replacing old pages with the new one
1747 * + swap/gup in pages if necessary;
1748 * + fill in gaps;
1749 * + keep old pages around in case rollback is required;
1750 * - if replacing succeeds:
1751 * + copy data over;
1752 * + free old pages;
1753 * + unlock huge page;
1754 * - if replacing failed;
1755 * + put all pages back and unfreeze them;
1756 * + restore gaps in the page cache;
1757 * + unlock and free huge page;
1758 */
1759static int collapse_file(struct mm_struct *mm, unsigned long addr,
1760 struct file *file, pgoff_t start,
1761 struct collapse_control *cc)
1762{
1763 struct address_space *mapping = file->f_mapping;
1764 struct page *hpage;
1765 pgoff_t index = 0, end = start + HPAGE_PMD_NR;
1766 LIST_HEAD(pagelist);
1767 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1768 int nr_none = 0, result = SCAN_SUCCEED;
1769 bool is_shmem = shmem_file(file);
1770 int nr = 0;
1771
1772 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1773 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1774
1775 result = alloc_charge_hpage(&hpage, mm, cc);
1776 if (result != SCAN_SUCCEED)
1777 goto out;
1778
1779 /*
1780 * Ensure we have slots for all the pages in the range. This is
1781 * almost certainly a no-op because most of the pages must be present
1782 */
1783 do {
1784 xas_lock_irq(&xas);
1785 xas_create_range(&xas);
1786 if (!xas_error(&xas))
1787 break;
1788 xas_unlock_irq(&xas);
1789 if (!xas_nomem(&xas, GFP_KERNEL)) {
1790 result = SCAN_FAIL;
1791 goto out;
1792 }
1793 } while (1);
1794
1795 __SetPageLocked(hpage);
1796 if (is_shmem)
1797 __SetPageSwapBacked(hpage);
1798 hpage->index = start;
1799 hpage->mapping = mapping;
1800
1801 /*
1802 * At this point the hpage is locked and not up-to-date.
1803 * It's safe to insert it into the page cache, because nobody would
1804 * be able to map it or use it in another way until we unlock it.
1805 */
1806
1807 xas_set(&xas, start);
1808 for (index = start; index < end; index++) {
1809 struct page *page = xas_next(&xas);
1810 struct folio *folio;
1811
1812 VM_BUG_ON(index != xas.xa_index);
1813 if (is_shmem) {
1814 if (!page) {
1815 /*
1816 * Stop if extent has been truncated or
1817 * hole-punched, and is now completely
1818 * empty.
1819 */
1820 if (index == start) {
1821 if (!xas_next_entry(&xas, end - 1)) {
1822 result = SCAN_TRUNCATED;
1823 goto xa_locked;
1824 }
1825 xas_set(&xas, index);
1826 }
1827 if (!shmem_charge(mapping->host, 1)) {
1828 result = SCAN_FAIL;
1829 goto xa_locked;
1830 }
1831 xas_store(&xas, hpage);
1832 nr_none++;
1833 continue;
1834 }
1835
1836 if (xa_is_value(page) || !PageUptodate(page)) {
1837 xas_unlock_irq(&xas);
1838 /* swap in or instantiate fallocated page */
1839 if (shmem_get_folio(mapping->host, index,
1840 &folio, SGP_NOALLOC)) {
1841 result = SCAN_FAIL;
1842 goto xa_unlocked;
1843 }
1844 page = folio_file_page(folio, index);
1845 } else if (trylock_page(page)) {
1846 get_page(page);
1847 xas_unlock_irq(&xas);
1848 } else {
1849 result = SCAN_PAGE_LOCK;
1850 goto xa_locked;
1851 }
1852 } else { /* !is_shmem */
1853 if (!page || xa_is_value(page)) {
1854 xas_unlock_irq(&xas);
1855 page_cache_sync_readahead(mapping, &file->f_ra,
1856 file, index,
1857 end - index);
1858 /* drain pagevecs to help isolate_lru_page() */
1859 lru_add_drain();
1860 page = find_lock_page(mapping, index);
1861 if (unlikely(page == NULL)) {
1862 result = SCAN_FAIL;
1863 goto xa_unlocked;
1864 }
1865 } else if (PageDirty(page)) {
1866 /*
1867 * khugepaged only works on read-only fd,
1868 * so this page is dirty because it hasn't
1869 * been flushed since first write. There
1870 * won't be new dirty pages.
1871 *
1872 * Trigger async flush here and hope the
1873 * writeback is done when khugepaged
1874 * revisits this page.
1875 *
1876 * This is a one-off situation. We are not
1877 * forcing writeback in loop.
1878 */
1879 xas_unlock_irq(&xas);
1880 filemap_flush(mapping);
1881 result = SCAN_FAIL;
1882 goto xa_unlocked;
1883 } else if (PageWriteback(page)) {
1884 xas_unlock_irq(&xas);
1885 result = SCAN_FAIL;
1886 goto xa_unlocked;
1887 } else if (trylock_page(page)) {
1888 get_page(page);
1889 xas_unlock_irq(&xas);
1890 } else {
1891 result = SCAN_PAGE_LOCK;
1892 goto xa_locked;
1893 }
1894 }
1895
1896 /*
1897 * The page must be locked, so we can drop the i_pages lock
1898 * without racing with truncate.
1899 */
1900 VM_BUG_ON_PAGE(!PageLocked(page), page);
1901
1902 /* make sure the page is up to date */
1903 if (unlikely(!PageUptodate(page))) {
1904 result = SCAN_FAIL;
1905 goto out_unlock;
1906 }
1907
1908 /*
1909 * If file was truncated then extended, or hole-punched, before
1910 * we locked the first page, then a THP might be there already.
1911 * This will be discovered on the first iteration.
1912 */
1913 if (PageTransCompound(page)) {
1914 struct page *head = compound_head(page);
1915
1916 result = compound_order(head) == HPAGE_PMD_ORDER &&
1917 head->index == start
1918 /* Maybe PMD-mapped */
1919 ? SCAN_PTE_MAPPED_HUGEPAGE
1920 : SCAN_PAGE_COMPOUND;
1921 goto out_unlock;
1922 }
1923
1924 folio = page_folio(page);
1925
1926 if (folio_mapping(folio) != mapping) {
1927 result = SCAN_TRUNCATED;
1928 goto out_unlock;
1929 }
1930
1931 if (!is_shmem && (folio_test_dirty(folio) ||
1932 folio_test_writeback(folio))) {
1933 /*
1934 * khugepaged only works on read-only fd, so this
1935 * page is dirty because it hasn't been flushed
1936 * since first write.
1937 */
1938 result = SCAN_FAIL;
1939 goto out_unlock;
1940 }
1941
1942 if (folio_isolate_lru(folio)) {
1943 result = SCAN_DEL_PAGE_LRU;
1944 goto out_unlock;
1945 }
1946
1947 if (folio_has_private(folio) &&
1948 !filemap_release_folio(folio, GFP_KERNEL)) {
1949 result = SCAN_PAGE_HAS_PRIVATE;
1950 folio_putback_lru(folio);
1951 goto out_unlock;
1952 }
1953
1954 if (folio_mapped(folio))
1955 try_to_unmap(folio,
1956 TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH);
1957
1958 xas_lock_irq(&xas);
1959 xas_set(&xas, index);
1960
1961 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1962
1963 /*
1964 * The page is expected to have page_count() == 3:
1965 * - we hold a pin on it;
1966 * - one reference from page cache;
1967 * - one from isolate_lru_page;
1968 */
1969 if (!page_ref_freeze(page, 3)) {
1970 result = SCAN_PAGE_COUNT;
1971 xas_unlock_irq(&xas);
1972 putback_lru_page(page);
1973 goto out_unlock;
1974 }
1975
1976 /*
1977 * Add the page to the list to be able to undo the collapse if
1978 * something go wrong.
1979 */
1980 list_add_tail(&page->lru, &pagelist);
1981
1982 /* Finally, replace with the new page. */
1983 xas_store(&xas, hpage);
1984 continue;
1985out_unlock:
1986 unlock_page(page);
1987 put_page(page);
1988 goto xa_unlocked;
1989 }
1990 nr = thp_nr_pages(hpage);
1991
1992 if (is_shmem)
1993 __mod_lruvec_page_state(hpage, NR_SHMEM_THPS, nr);
1994 else {
1995 __mod_lruvec_page_state(hpage, NR_FILE_THPS, nr);
1996 filemap_nr_thps_inc(mapping);
1997 /*
1998 * Paired with smp_mb() in do_dentry_open() to ensure
1999 * i_writecount is up to date and the update to nr_thps is
2000 * visible. Ensures the page cache will be truncated if the
2001 * file is opened writable.
2002 */
2003 smp_mb();
2004 if (inode_is_open_for_write(mapping->host)) {
2005 result = SCAN_FAIL;
2006 __mod_lruvec_page_state(hpage, NR_FILE_THPS, -nr);
2007 filemap_nr_thps_dec(mapping);
2008 goto xa_locked;
2009 }
2010 }
2011
2012 if (nr_none) {
2013 __mod_lruvec_page_state(hpage, NR_FILE_PAGES, nr_none);
2014 /* nr_none is always 0 for non-shmem. */
2015 __mod_lruvec_page_state(hpage, NR_SHMEM, nr_none);
2016 }
2017
2018 /* Join all the small entries into a single multi-index entry */
2019 xas_set_order(&xas, start, HPAGE_PMD_ORDER);
2020 xas_store(&xas, hpage);
2021xa_locked:
2022 xas_unlock_irq(&xas);
2023xa_unlocked:
2024
2025 /*
2026 * If collapse is successful, flush must be done now before copying.
2027 * If collapse is unsuccessful, does flush actually need to be done?
2028 * Do it anyway, to clear the state.
2029 */
2030 try_to_unmap_flush();
2031
2032 if (result == SCAN_SUCCEED) {
2033 struct page *page, *tmp;
2034 struct folio *folio;
2035
2036 /*
2037 * Replacing old pages with new one has succeeded, now we
2038 * need to copy the content and free the old pages.
2039 */
2040 index = start;
2041 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
2042 while (index < page->index) {
2043 clear_highpage(hpage + (index % HPAGE_PMD_NR));
2044 index++;
2045 }
2046 copy_highpage(hpage + (page->index % HPAGE_PMD_NR),
2047 page);
2048 list_del(&page->lru);
2049 page->mapping = NULL;
2050 page_ref_unfreeze(page, 1);
2051 ClearPageActive(page);
2052 ClearPageUnevictable(page);
2053 unlock_page(page);
2054 put_page(page);
2055 index++;
2056 }
2057 while (index < end) {
2058 clear_highpage(hpage + (index % HPAGE_PMD_NR));
2059 index++;
2060 }
2061
2062 folio = page_folio(hpage);
2063 folio_mark_uptodate(folio);
2064 folio_ref_add(folio, HPAGE_PMD_NR - 1);
2065
2066 if (is_shmem)
2067 folio_mark_dirty(folio);
2068 folio_add_lru(folio);
2069
2070 /*
2071 * Remove pte page tables, so we can re-fault the page as huge.
2072 */
2073 result = retract_page_tables(mapping, start, mm, addr, hpage,
2074 cc);
2075 unlock_page(hpage);
2076 hpage = NULL;
2077 } else {
2078 struct page *page;
2079
2080 /* Something went wrong: roll back page cache changes */
2081 xas_lock_irq(&xas);
2082 if (nr_none) {
2083 mapping->nrpages -= nr_none;
2084 shmem_uncharge(mapping->host, nr_none);
2085 }
2086
2087 xas_set(&xas, start);
2088 xas_for_each(&xas, page, end - 1) {
2089 page = list_first_entry_or_null(&pagelist,
2090 struct page, lru);
2091 if (!page || xas.xa_index < page->index) {
2092 if (!nr_none)
2093 break;
2094 nr_none--;
2095 /* Put holes back where they were */
2096 xas_store(&xas, NULL);
2097 continue;
2098 }
2099
2100 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
2101
2102 /* Unfreeze the page. */
2103 list_del(&page->lru);
2104 page_ref_unfreeze(page, 2);
2105 xas_store(&xas, page);
2106 xas_pause(&xas);
2107 xas_unlock_irq(&xas);
2108 unlock_page(page);
2109 putback_lru_page(page);
2110 xas_lock_irq(&xas);
2111 }
2112 VM_BUG_ON(nr_none);
2113 xas_unlock_irq(&xas);
2114
2115 hpage->mapping = NULL;
2116 }
2117
2118 if (hpage)
2119 unlock_page(hpage);
2120out:
2121 VM_BUG_ON(!list_empty(&pagelist));
2122 if (hpage) {
2123 mem_cgroup_uncharge(page_folio(hpage));
2124 put_page(hpage);
2125 }
2126
2127 trace_mm_khugepaged_collapse_file(mm, hpage, index, is_shmem, addr, file, nr, result);
2128 return result;
2129}
2130
2131static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2132 struct file *file, pgoff_t start,
2133 struct collapse_control *cc)
2134{
2135 struct page *page = NULL;
2136 struct address_space *mapping = file->f_mapping;
2137 XA_STATE(xas, &mapping->i_pages, start);
2138 int present, swap;
2139 int node = NUMA_NO_NODE;
2140 int result = SCAN_SUCCEED;
2141
2142 present = 0;
2143 swap = 0;
2144 memset(cc->node_load, 0, sizeof(cc->node_load));
2145 nodes_clear(cc->alloc_nmask);
2146 rcu_read_lock();
2147 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2148 if (xas_retry(&xas, page))
2149 continue;
2150
2151 if (xa_is_value(page)) {
2152 ++swap;
2153 if (cc->is_khugepaged &&
2154 swap > khugepaged_max_ptes_swap) {
2155 result = SCAN_EXCEED_SWAP_PTE;
2156 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
2157 break;
2158 }
2159 continue;
2160 }
2161
2162 /*
2163 * TODO: khugepaged should compact smaller compound pages
2164 * into a PMD sized page
2165 */
2166 if (PageTransCompound(page)) {
2167 struct page *head = compound_head(page);
2168
2169 result = compound_order(head) == HPAGE_PMD_ORDER &&
2170 head->index == start
2171 /* Maybe PMD-mapped */
2172 ? SCAN_PTE_MAPPED_HUGEPAGE
2173 : SCAN_PAGE_COMPOUND;
2174 /*
2175 * For SCAN_PTE_MAPPED_HUGEPAGE, further processing
2176 * by the caller won't touch the page cache, and so
2177 * it's safe to skip LRU and refcount checks before
2178 * returning.
2179 */
2180 break;
2181 }
2182
2183 node = page_to_nid(page);
2184 if (hpage_collapse_scan_abort(node, cc)) {
2185 result = SCAN_SCAN_ABORT;
2186 break;
2187 }
2188 cc->node_load[node]++;
2189
2190 if (!PageLRU(page)) {
2191 result = SCAN_PAGE_LRU;
2192 break;
2193 }
2194
2195 if (page_count(page) !=
2196 1 + page_mapcount(page) + page_has_private(page)) {
2197 result = SCAN_PAGE_COUNT;
2198 break;
2199 }
2200
2201 /*
2202 * We probably should check if the page is referenced here, but
2203 * nobody would transfer pte_young() to PageReferenced() for us.
2204 * And rmap walk here is just too costly...
2205 */
2206
2207 present++;
2208
2209 if (need_resched()) {
2210 xas_pause(&xas);
2211 cond_resched_rcu();
2212 }
2213 }
2214 rcu_read_unlock();
2215
2216 if (result == SCAN_SUCCEED) {
2217 if (cc->is_khugepaged &&
2218 present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2219 result = SCAN_EXCEED_NONE_PTE;
2220 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2221 } else {
2222 result = collapse_file(mm, addr, file, start, cc);
2223 }
2224 }
2225
2226 trace_mm_khugepaged_scan_file(mm, page, file, present, swap, result);
2227 return result;
2228}
2229#else
2230static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2231 struct file *file, pgoff_t start,
2232 struct collapse_control *cc)
2233{
2234 BUILD_BUG();
2235}
2236
2237static void khugepaged_collapse_pte_mapped_thps(struct khugepaged_mm_slot *mm_slot)
2238{
2239}
2240
2241static bool khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
2242 unsigned long addr)
2243{
2244 return false;
2245}
2246#endif
2247
2248static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result,
2249 struct collapse_control *cc)
2250 __releases(&khugepaged_mm_lock)
2251 __acquires(&khugepaged_mm_lock)
2252{
2253 struct vma_iterator vmi;
2254 struct khugepaged_mm_slot *mm_slot;
2255 struct mm_slot *slot;
2256 struct mm_struct *mm;
2257 struct vm_area_struct *vma;
2258 int progress = 0;
2259
2260 VM_BUG_ON(!pages);
2261 lockdep_assert_held(&khugepaged_mm_lock);
2262 *result = SCAN_FAIL;
2263
2264 if (khugepaged_scan.mm_slot) {
2265 mm_slot = khugepaged_scan.mm_slot;
2266 slot = &mm_slot->slot;
2267 } else {
2268 slot = list_entry(khugepaged_scan.mm_head.next,
2269 struct mm_slot, mm_node);
2270 mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2271 khugepaged_scan.address = 0;
2272 khugepaged_scan.mm_slot = mm_slot;
2273 }
2274 spin_unlock(&khugepaged_mm_lock);
2275 khugepaged_collapse_pte_mapped_thps(mm_slot);
2276
2277 mm = slot->mm;
2278 /*
2279 * Don't wait for semaphore (to avoid long wait times). Just move to
2280 * the next mm on the list.
2281 */
2282 vma = NULL;
2283 if (unlikely(!mmap_read_trylock(mm)))
2284 goto breakouterloop_mmap_lock;
2285
2286 progress++;
2287 if (unlikely(hpage_collapse_test_exit(mm)))
2288 goto breakouterloop;
2289
2290 vma_iter_init(&vmi, mm, khugepaged_scan.address);
2291 for_each_vma(vmi, vma) {
2292 unsigned long hstart, hend;
2293
2294 cond_resched();
2295 if (unlikely(hpage_collapse_test_exit(mm))) {
2296 progress++;
2297 break;
2298 }
2299 if (!hugepage_vma_check(vma, vma->vm_flags, false, false, true)) {
2300skip:
2301 progress++;
2302 continue;
2303 }
2304 hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE);
2305 hend = round_down(vma->vm_end, HPAGE_PMD_SIZE);
2306 if (khugepaged_scan.address > hend)
2307 goto skip;
2308 if (khugepaged_scan.address < hstart)
2309 khugepaged_scan.address = hstart;
2310 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2311
2312 while (khugepaged_scan.address < hend) {
2313 bool mmap_locked = true;
2314
2315 cond_resched();
2316 if (unlikely(hpage_collapse_test_exit(mm)))
2317 goto breakouterloop;
2318
2319 VM_BUG_ON(khugepaged_scan.address < hstart ||
2320 khugepaged_scan.address + HPAGE_PMD_SIZE >
2321 hend);
2322 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2323 struct file *file = get_file(vma->vm_file);
2324 pgoff_t pgoff = linear_page_index(vma,
2325 khugepaged_scan.address);
2326
2327 mmap_read_unlock(mm);
2328 *result = hpage_collapse_scan_file(mm,
2329 khugepaged_scan.address,
2330 file, pgoff, cc);
2331 mmap_locked = false;
2332 fput(file);
2333 } else {
2334 *result = hpage_collapse_scan_pmd(mm, vma,
2335 khugepaged_scan.address,
2336 &mmap_locked,
2337 cc);
2338 }
2339 switch (*result) {
2340 case SCAN_PTE_MAPPED_HUGEPAGE: {
2341 pmd_t *pmd;
2342
2343 *result = find_pmd_or_thp_or_none(mm,
2344 khugepaged_scan.address,
2345 &pmd);
2346 if (*result != SCAN_SUCCEED)
2347 break;
2348 if (!khugepaged_add_pte_mapped_thp(mm,
2349 khugepaged_scan.address))
2350 break;
2351 } fallthrough;
2352 case SCAN_SUCCEED:
2353 ++khugepaged_pages_collapsed;
2354 break;
2355 default:
2356 break;
2357 }
2358
2359 /* move to next address */
2360 khugepaged_scan.address += HPAGE_PMD_SIZE;
2361 progress += HPAGE_PMD_NR;
2362 if (!mmap_locked)
2363 /*
2364 * We released mmap_lock so break loop. Note
2365 * that we drop mmap_lock before all hugepage
2366 * allocations, so if allocation fails, we are
2367 * guaranteed to break here and report the
2368 * correct result back to caller.
2369 */
2370 goto breakouterloop_mmap_lock;
2371 if (progress >= pages)
2372 goto breakouterloop;
2373 }
2374 }
2375breakouterloop:
2376 mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2377breakouterloop_mmap_lock:
2378
2379 spin_lock(&khugepaged_mm_lock);
2380 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2381 /*
2382 * Release the current mm_slot if this mm is about to die, or
2383 * if we scanned all vmas of this mm.
2384 */
2385 if (hpage_collapse_test_exit(mm) || !vma) {
2386 /*
2387 * Make sure that if mm_users is reaching zero while
2388 * khugepaged runs here, khugepaged_exit will find
2389 * mm_slot not pointing to the exiting mm.
2390 */
2391 if (slot->mm_node.next != &khugepaged_scan.mm_head) {
2392 slot = list_entry(slot->mm_node.next,
2393 struct mm_slot, mm_node);
2394 khugepaged_scan.mm_slot =
2395 mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2396 khugepaged_scan.address = 0;
2397 } else {
2398 khugepaged_scan.mm_slot = NULL;
2399 khugepaged_full_scans++;
2400 }
2401
2402 collect_mm_slot(mm_slot);
2403 }
2404
2405 return progress;
2406}
2407
2408static int khugepaged_has_work(void)
2409{
2410 return !list_empty(&khugepaged_scan.mm_head) &&
2411 hugepage_flags_enabled();
2412}
2413
2414static int khugepaged_wait_event(void)
2415{
2416 return !list_empty(&khugepaged_scan.mm_head) ||
2417 kthread_should_stop();
2418}
2419
2420static void khugepaged_do_scan(struct collapse_control *cc)
2421{
2422 unsigned int progress = 0, pass_through_head = 0;
2423 unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2424 bool wait = true;
2425 int result = SCAN_SUCCEED;
2426
2427 lru_add_drain_all();
2428
2429 while (true) {
2430 cond_resched();
2431
2432 if (unlikely(kthread_should_stop() || try_to_freeze()))
2433 break;
2434
2435 spin_lock(&khugepaged_mm_lock);
2436 if (!khugepaged_scan.mm_slot)
2437 pass_through_head++;
2438 if (khugepaged_has_work() &&
2439 pass_through_head < 2)
2440 progress += khugepaged_scan_mm_slot(pages - progress,
2441 &result, cc);
2442 else
2443 progress = pages;
2444 spin_unlock(&khugepaged_mm_lock);
2445
2446 if (progress >= pages)
2447 break;
2448
2449 if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) {
2450 /*
2451 * If fail to allocate the first time, try to sleep for
2452 * a while. When hit again, cancel the scan.
2453 */
2454 if (!wait)
2455 break;
2456 wait = false;
2457 khugepaged_alloc_sleep();
2458 }
2459 }
2460}
2461
2462static bool khugepaged_should_wakeup(void)
2463{
2464 return kthread_should_stop() ||
2465 time_after_eq(jiffies, khugepaged_sleep_expire);
2466}
2467
2468static void khugepaged_wait_work(void)
2469{
2470 if (khugepaged_has_work()) {
2471 const unsigned long scan_sleep_jiffies =
2472 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2473
2474 if (!scan_sleep_jiffies)
2475 return;
2476
2477 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2478 wait_event_freezable_timeout(khugepaged_wait,
2479 khugepaged_should_wakeup(),
2480 scan_sleep_jiffies);
2481 return;
2482 }
2483
2484 if (hugepage_flags_enabled())
2485 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2486}
2487
2488static int khugepaged(void *none)
2489{
2490 struct khugepaged_mm_slot *mm_slot;
2491
2492 set_freezable();
2493 set_user_nice(current, MAX_NICE);
2494
2495 while (!kthread_should_stop()) {
2496 khugepaged_do_scan(&khugepaged_collapse_control);
2497 khugepaged_wait_work();
2498 }
2499
2500 spin_lock(&khugepaged_mm_lock);
2501 mm_slot = khugepaged_scan.mm_slot;
2502 khugepaged_scan.mm_slot = NULL;
2503 if (mm_slot)
2504 collect_mm_slot(mm_slot);
2505 spin_unlock(&khugepaged_mm_lock);
2506 return 0;
2507}
2508
2509static void set_recommended_min_free_kbytes(void)
2510{
2511 struct zone *zone;
2512 int nr_zones = 0;
2513 unsigned long recommended_min;
2514
2515 if (!hugepage_flags_enabled()) {
2516 calculate_min_free_kbytes();
2517 goto update_wmarks;
2518 }
2519
2520 for_each_populated_zone(zone) {
2521 /*
2522 * We don't need to worry about fragmentation of
2523 * ZONE_MOVABLE since it only has movable pages.
2524 */
2525 if (zone_idx(zone) > gfp_zone(GFP_USER))
2526 continue;
2527
2528 nr_zones++;
2529 }
2530
2531 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2532 recommended_min = pageblock_nr_pages * nr_zones * 2;
2533
2534 /*
2535 * Make sure that on average at least two pageblocks are almost free
2536 * of another type, one for a migratetype to fall back to and a
2537 * second to avoid subsequent fallbacks of other types There are 3
2538 * MIGRATE_TYPES we care about.
2539 */
2540 recommended_min += pageblock_nr_pages * nr_zones *
2541 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2542
2543 /* don't ever allow to reserve more than 5% of the lowmem */
2544 recommended_min = min(recommended_min,
2545 (unsigned long) nr_free_buffer_pages() / 20);
2546 recommended_min <<= (PAGE_SHIFT-10);
2547
2548 if (recommended_min > min_free_kbytes) {
2549 if (user_min_free_kbytes >= 0)
2550 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2551 min_free_kbytes, recommended_min);
2552
2553 min_free_kbytes = recommended_min;
2554 }
2555
2556update_wmarks:
2557 setup_per_zone_wmarks();
2558}
2559
2560int start_stop_khugepaged(void)
2561{
2562 int err = 0;
2563
2564 mutex_lock(&khugepaged_mutex);
2565 if (hugepage_flags_enabled()) {
2566 if (!khugepaged_thread)
2567 khugepaged_thread = kthread_run(khugepaged, NULL,
2568 "khugepaged");
2569 if (IS_ERR(khugepaged_thread)) {
2570 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2571 err = PTR_ERR(khugepaged_thread);
2572 khugepaged_thread = NULL;
2573 goto fail;
2574 }
2575
2576 if (!list_empty(&khugepaged_scan.mm_head))
2577 wake_up_interruptible(&khugepaged_wait);
2578 } else if (khugepaged_thread) {
2579 kthread_stop(khugepaged_thread);
2580 khugepaged_thread = NULL;
2581 }
2582 set_recommended_min_free_kbytes();
2583fail:
2584 mutex_unlock(&khugepaged_mutex);
2585 return err;
2586}
2587
2588void khugepaged_min_free_kbytes_update(void)
2589{
2590 mutex_lock(&khugepaged_mutex);
2591 if (hugepage_flags_enabled() && khugepaged_thread)
2592 set_recommended_min_free_kbytes();
2593 mutex_unlock(&khugepaged_mutex);
2594}
2595
2596bool current_is_khugepaged(void)
2597{
2598 return kthread_func(current) == khugepaged;
2599}
2600
2601static int madvise_collapse_errno(enum scan_result r)
2602{
2603 /*
2604 * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide
2605 * actionable feedback to caller, so they may take an appropriate
2606 * fallback measure depending on the nature of the failure.
2607 */
2608 switch (r) {
2609 case SCAN_ALLOC_HUGE_PAGE_FAIL:
2610 return -ENOMEM;
2611 case SCAN_CGROUP_CHARGE_FAIL:
2612 return -EBUSY;
2613 /* Resource temporary unavailable - trying again might succeed */
2614 case SCAN_PAGE_COUNT:
2615 case SCAN_PAGE_LOCK:
2616 case SCAN_PAGE_LRU:
2617 case SCAN_DEL_PAGE_LRU:
2618 return -EAGAIN;
2619 /*
2620 * Other: Trying again likely not to succeed / error intrinsic to
2621 * specified memory range. khugepaged likely won't be able to collapse
2622 * either.
2623 */
2624 default:
2625 return -EINVAL;
2626 }
2627}
2628
2629int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev,
2630 unsigned long start, unsigned long end)
2631{
2632 struct collapse_control *cc;
2633 struct mm_struct *mm = vma->vm_mm;
2634 unsigned long hstart, hend, addr;
2635 int thps = 0, last_fail = SCAN_FAIL;
2636 bool mmap_locked = true;
2637
2638 BUG_ON(vma->vm_start > start);
2639 BUG_ON(vma->vm_end < end);
2640
2641 *prev = vma;
2642
2643 if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false))
2644 return -EINVAL;
2645
2646 cc = kmalloc(sizeof(*cc), GFP_KERNEL);
2647 if (!cc)
2648 return -ENOMEM;
2649 cc->is_khugepaged = false;
2650
2651 mmgrab(mm);
2652 lru_add_drain_all();
2653
2654 hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2655 hend = end & HPAGE_PMD_MASK;
2656
2657 for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) {
2658 int result = SCAN_FAIL;
2659
2660 if (!mmap_locked) {
2661 cond_resched();
2662 mmap_read_lock(mm);
2663 mmap_locked = true;
2664 result = hugepage_vma_revalidate(mm, addr, false, &vma,
2665 cc);
2666 if (result != SCAN_SUCCEED) {
2667 last_fail = result;
2668 goto out_nolock;
2669 }
2670
2671 hend = min(hend, vma->vm_end & HPAGE_PMD_MASK);
2672 }
2673 mmap_assert_locked(mm);
2674 memset(cc->node_load, 0, sizeof(cc->node_load));
2675 nodes_clear(cc->alloc_nmask);
2676 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2677 struct file *file = get_file(vma->vm_file);
2678 pgoff_t pgoff = linear_page_index(vma, addr);
2679
2680 mmap_read_unlock(mm);
2681 mmap_locked = false;
2682 result = hpage_collapse_scan_file(mm, addr, file, pgoff,
2683 cc);
2684 fput(file);
2685 } else {
2686 result = hpage_collapse_scan_pmd(mm, vma, addr,
2687 &mmap_locked, cc);
2688 }
2689 if (!mmap_locked)
2690 *prev = NULL; /* Tell caller we dropped mmap_lock */
2691
2692handle_result:
2693 switch (result) {
2694 case SCAN_SUCCEED:
2695 case SCAN_PMD_MAPPED:
2696 ++thps;
2697 break;
2698 case SCAN_PTE_MAPPED_HUGEPAGE:
2699 BUG_ON(mmap_locked);
2700 BUG_ON(*prev);
2701 mmap_write_lock(mm);
2702 result = collapse_pte_mapped_thp(mm, addr, true);
2703 mmap_write_unlock(mm);
2704 goto handle_result;
2705 /* Whitelisted set of results where continuing OK */
2706 case SCAN_PMD_NULL:
2707 case SCAN_PTE_NON_PRESENT:
2708 case SCAN_PTE_UFFD_WP:
2709 case SCAN_PAGE_RO:
2710 case SCAN_LACK_REFERENCED_PAGE:
2711 case SCAN_PAGE_NULL:
2712 case SCAN_PAGE_COUNT:
2713 case SCAN_PAGE_LOCK:
2714 case SCAN_PAGE_COMPOUND:
2715 case SCAN_PAGE_LRU:
2716 case SCAN_DEL_PAGE_LRU:
2717 last_fail = result;
2718 break;
2719 default:
2720 last_fail = result;
2721 /* Other error, exit */
2722 goto out_maybelock;
2723 }
2724 }
2725
2726out_maybelock:
2727 /* Caller expects us to hold mmap_lock on return */
2728 if (!mmap_locked)
2729 mmap_read_lock(mm);
2730out_nolock:
2731 mmap_assert_locked(mm);
2732 mmdrop(mm);
2733 kfree(cc);
2734
2735 return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0
2736 : madvise_collapse_errno(last_fail);
2737}
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
26enum 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
59static struct task_struct *khugepaged_thread __read_mostly;
60static DEFINE_MUTEX(khugepaged_mutex);
61
62/* default scan 8*512 pte (or vmas) every 30 second */
63static unsigned int khugepaged_pages_to_scan __read_mostly;
64static unsigned int khugepaged_pages_collapsed;
65static unsigned int khugepaged_full_scans;
66static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
67/* during fragmentation poll the hugepage allocator once every minute */
68static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
69static unsigned long khugepaged_sleep_expire;
70static DEFINE_SPINLOCK(khugepaged_mm_lock);
71static 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 */
77static unsigned int khugepaged_max_ptes_none __read_mostly;
78static unsigned int khugepaged_max_ptes_swap __read_mostly;
79static unsigned int khugepaged_max_ptes_shared __read_mostly;
80
81#define MM_SLOTS_HASH_BITS 10
82static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
83
84static 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 */
94struct 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 */
112struct khugepaged_scan {
113 struct list_head mm_head;
114 struct mm_slot *mm_slot;
115 unsigned long address;
116};
117
118static struct khugepaged_scan khugepaged_scan = {
119 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
120};
121
122#ifdef CONFIG_SYSFS
123static 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
130static 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}
147static struct kobj_attribute scan_sleep_millisecs_attr =
148 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
149 scan_sleep_millisecs_store);
150
151static 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
158static 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}
175static struct kobj_attribute alloc_sleep_millisecs_attr =
176 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
177 alloc_sleep_millisecs_store);
178
179static 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}
185static 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}
200static struct kobj_attribute pages_to_scan_attr =
201 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
202 pages_to_scan_store);
203
204static 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}
210static struct kobj_attribute pages_collapsed_attr =
211 __ATTR_RO(pages_collapsed);
212
213static 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}
219static struct kobj_attribute full_scans_attr =
220 __ATTR_RO(full_scans);
221
222static 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}
228static 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}
235static 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 */
247static 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}
253static 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}
268static 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
272static 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
279static 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
295static 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
299static 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
306static 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
322static 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
326static 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
339struct attribute_group khugepaged_attr_group = {
340 .attrs = khugepaged_attr,
341 .name = "khugepaged",
342};
343#endif /* CONFIG_SYSFS */
344
345int 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
384int __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
400void __init khugepaged_destroy(void)
401{
402 kmem_cache_destroy(mm_slot_cache);
403}
404
405static 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
412static inline void free_mm_slot(struct mm_slot *mm_slot)
413{
414 kmem_cache_free(mm_slot_cache, mm_slot);
415}
416
417static 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
428static 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
435static inline int khugepaged_test_exit(struct mm_struct *mm)
436{
437 return atomic_read(&mm->mm_users) == 0 || !mmget_still_valid(mm);
438}
439
440static bool hugepage_vma_check(struct vm_area_struct *vma,
441 unsigned long vm_flags)
442{
443 if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
444 (vm_flags & VM_NOHUGEPAGE) ||
445 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
446 return false;
447
448 if (shmem_file(vma->vm_file) ||
449 (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
450 vma->vm_file &&
451 (vm_flags & VM_DENYWRITE))) {
452 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
453 HPAGE_PMD_NR);
454 }
455 if (!vma->anon_vma || vma->vm_ops)
456 return false;
457 if (vma_is_temporary_stack(vma))
458 return false;
459 return !(vm_flags & VM_NO_KHUGEPAGED);
460}
461
462int __khugepaged_enter(struct mm_struct *mm)
463{
464 struct mm_slot *mm_slot;
465 int wakeup;
466
467 mm_slot = alloc_mm_slot();
468 if (!mm_slot)
469 return -ENOMEM;
470
471 /* __khugepaged_exit() must not run from under us */
472 VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
473 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
474 free_mm_slot(mm_slot);
475 return 0;
476 }
477
478 spin_lock(&khugepaged_mm_lock);
479 insert_to_mm_slots_hash(mm, mm_slot);
480 /*
481 * Insert just behind the scanning cursor, to let the area settle
482 * down a little.
483 */
484 wakeup = list_empty(&khugepaged_scan.mm_head);
485 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
486 spin_unlock(&khugepaged_mm_lock);
487
488 mmgrab(mm);
489 if (wakeup)
490 wake_up_interruptible(&khugepaged_wait);
491
492 return 0;
493}
494
495int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
496 unsigned long vm_flags)
497{
498 unsigned long hstart, hend;
499
500 /*
501 * khugepaged only supports read-only files for non-shmem files.
502 * khugepaged does not yet work on special mappings. And
503 * file-private shmem THP is not supported.
504 */
505 if (!hugepage_vma_check(vma, vm_flags))
506 return 0;
507
508 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
509 hend = vma->vm_end & HPAGE_PMD_MASK;
510 if (hstart < hend)
511 return khugepaged_enter(vma, vm_flags);
512 return 0;
513}
514
515void __khugepaged_exit(struct mm_struct *mm)
516{
517 struct mm_slot *mm_slot;
518 int free = 0;
519
520 spin_lock(&khugepaged_mm_lock);
521 mm_slot = get_mm_slot(mm);
522 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
523 hash_del(&mm_slot->hash);
524 list_del(&mm_slot->mm_node);
525 free = 1;
526 }
527 spin_unlock(&khugepaged_mm_lock);
528
529 if (free) {
530 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
531 free_mm_slot(mm_slot);
532 mmdrop(mm);
533 } else if (mm_slot) {
534 /*
535 * This is required to serialize against
536 * khugepaged_test_exit() (which is guaranteed to run
537 * under mmap sem read mode). Stop here (after we
538 * return all pagetables will be destroyed) until
539 * khugepaged has finished working on the pagetables
540 * under the mmap_lock.
541 */
542 mmap_write_lock(mm);
543 mmap_write_unlock(mm);
544 }
545}
546
547static void release_pte_page(struct page *page)
548{
549 mod_node_page_state(page_pgdat(page),
550 NR_ISOLATED_ANON + page_is_file_lru(page),
551 -compound_nr(page));
552 unlock_page(page);
553 putback_lru_page(page);
554}
555
556static void release_pte_pages(pte_t *pte, pte_t *_pte,
557 struct list_head *compound_pagelist)
558{
559 struct page *page, *tmp;
560
561 while (--_pte >= pte) {
562 pte_t pteval = *_pte;
563
564 page = pte_page(pteval);
565 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
566 !PageCompound(page))
567 release_pte_page(page);
568 }
569
570 list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
571 list_del(&page->lru);
572 release_pte_page(page);
573 }
574}
575
576static bool is_refcount_suitable(struct page *page)
577{
578 int expected_refcount;
579
580 expected_refcount = total_mapcount(page);
581 if (PageSwapCache(page))
582 expected_refcount += compound_nr(page);
583
584 return page_count(page) == expected_refcount;
585}
586
587static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
588 unsigned long address,
589 pte_t *pte,
590 struct list_head *compound_pagelist)
591{
592 struct page *page = NULL;
593 pte_t *_pte;
594 int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
595 bool writable = false;
596
597 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
598 _pte++, address += PAGE_SIZE) {
599 pte_t pteval = *_pte;
600 if (pte_none(pteval) || (pte_present(pteval) &&
601 is_zero_pfn(pte_pfn(pteval)))) {
602 if (!userfaultfd_armed(vma) &&
603 ++none_or_zero <= khugepaged_max_ptes_none) {
604 continue;
605 } else {
606 result = SCAN_EXCEED_NONE_PTE;
607 goto out;
608 }
609 }
610 if (!pte_present(pteval)) {
611 result = SCAN_PTE_NON_PRESENT;
612 goto out;
613 }
614 page = vm_normal_page(vma, address, pteval);
615 if (unlikely(!page)) {
616 result = SCAN_PAGE_NULL;
617 goto out;
618 }
619
620 VM_BUG_ON_PAGE(!PageAnon(page), page);
621
622 if (page_mapcount(page) > 1 &&
623 ++shared > khugepaged_max_ptes_shared) {
624 result = SCAN_EXCEED_SHARED_PTE;
625 goto out;
626 }
627
628 if (PageCompound(page)) {
629 struct page *p;
630 page = compound_head(page);
631
632 /*
633 * Check if we have dealt with the compound page
634 * already
635 */
636 list_for_each_entry(p, compound_pagelist, lru) {
637 if (page == p)
638 goto next;
639 }
640 }
641
642 /*
643 * We can do it before isolate_lru_page because the
644 * page can't be freed from under us. NOTE: PG_lock
645 * is needed to serialize against split_huge_page
646 * when invoked from the VM.
647 */
648 if (!trylock_page(page)) {
649 result = SCAN_PAGE_LOCK;
650 goto out;
651 }
652
653 /*
654 * Check if the page has any GUP (or other external) pins.
655 *
656 * The page table that maps the page has been already unlinked
657 * from the page table tree and this process cannot get
658 * an additinal pin on the page.
659 *
660 * New pins can come later if the page is shared across fork,
661 * but not from this process. The other process cannot write to
662 * the page, only trigger CoW.
663 */
664 if (!is_refcount_suitable(page)) {
665 unlock_page(page);
666 result = SCAN_PAGE_COUNT;
667 goto out;
668 }
669 if (!pte_write(pteval) && PageSwapCache(page) &&
670 !reuse_swap_page(page, NULL)) {
671 /*
672 * Page is in the swap cache and cannot be re-used.
673 * It cannot be collapsed into a THP.
674 */
675 unlock_page(page);
676 result = SCAN_SWAP_CACHE_PAGE;
677 goto out;
678 }
679
680 /*
681 * Isolate the page to avoid collapsing an hugepage
682 * currently in use by the VM.
683 */
684 if (isolate_lru_page(page)) {
685 unlock_page(page);
686 result = SCAN_DEL_PAGE_LRU;
687 goto out;
688 }
689 mod_node_page_state(page_pgdat(page),
690 NR_ISOLATED_ANON + page_is_file_lru(page),
691 compound_nr(page));
692 VM_BUG_ON_PAGE(!PageLocked(page), page);
693 VM_BUG_ON_PAGE(PageLRU(page), page);
694
695 if (PageCompound(page))
696 list_add_tail(&page->lru, compound_pagelist);
697next:
698 /* There should be enough young pte to collapse the page */
699 if (pte_young(pteval) ||
700 page_is_young(page) || PageReferenced(page) ||
701 mmu_notifier_test_young(vma->vm_mm, address))
702 referenced++;
703
704 if (pte_write(pteval))
705 writable = true;
706 }
707 if (likely(writable)) {
708 if (likely(referenced)) {
709 result = SCAN_SUCCEED;
710 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
711 referenced, writable, result);
712 return 1;
713 }
714 } else {
715 result = SCAN_PAGE_RO;
716 }
717
718out:
719 release_pte_pages(pte, _pte, compound_pagelist);
720 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
721 referenced, writable, result);
722 return 0;
723}
724
725static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
726 struct vm_area_struct *vma,
727 unsigned long address,
728 spinlock_t *ptl,
729 struct list_head *compound_pagelist)
730{
731 struct page *src_page, *tmp;
732 pte_t *_pte;
733 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
734 _pte++, page++, address += PAGE_SIZE) {
735 pte_t pteval = *_pte;
736
737 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
738 clear_user_highpage(page, address);
739 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
740 if (is_zero_pfn(pte_pfn(pteval))) {
741 /*
742 * ptl mostly unnecessary.
743 */
744 spin_lock(ptl);
745 /*
746 * paravirt calls inside pte_clear here are
747 * superfluous.
748 */
749 pte_clear(vma->vm_mm, address, _pte);
750 spin_unlock(ptl);
751 }
752 } else {
753 src_page = pte_page(pteval);
754 copy_user_highpage(page, src_page, address, vma);
755 if (!PageCompound(src_page))
756 release_pte_page(src_page);
757 /*
758 * ptl mostly unnecessary, but preempt has to
759 * be disabled to update the per-cpu stats
760 * inside page_remove_rmap().
761 */
762 spin_lock(ptl);
763 /*
764 * paravirt calls inside pte_clear here are
765 * superfluous.
766 */
767 pte_clear(vma->vm_mm, address, _pte);
768 page_remove_rmap(src_page, false);
769 spin_unlock(ptl);
770 free_page_and_swap_cache(src_page);
771 }
772 }
773
774 list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
775 list_del(&src_page->lru);
776 release_pte_page(src_page);
777 }
778}
779
780static void khugepaged_alloc_sleep(void)
781{
782 DEFINE_WAIT(wait);
783
784 add_wait_queue(&khugepaged_wait, &wait);
785 freezable_schedule_timeout_interruptible(
786 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
787 remove_wait_queue(&khugepaged_wait, &wait);
788}
789
790static int khugepaged_node_load[MAX_NUMNODES];
791
792static bool khugepaged_scan_abort(int nid)
793{
794 int i;
795
796 /*
797 * If node_reclaim_mode is disabled, then no extra effort is made to
798 * allocate memory locally.
799 */
800 if (!node_reclaim_mode)
801 return false;
802
803 /* If there is a count for this node already, it must be acceptable */
804 if (khugepaged_node_load[nid])
805 return false;
806
807 for (i = 0; i < MAX_NUMNODES; i++) {
808 if (!khugepaged_node_load[i])
809 continue;
810 if (node_distance(nid, i) > node_reclaim_distance)
811 return true;
812 }
813 return false;
814}
815
816/* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
817static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
818{
819 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
820}
821
822#ifdef CONFIG_NUMA
823static int khugepaged_find_target_node(void)
824{
825 static int last_khugepaged_target_node = NUMA_NO_NODE;
826 int nid, target_node = 0, max_value = 0;
827
828 /* find first node with max normal pages hit */
829 for (nid = 0; nid < MAX_NUMNODES; nid++)
830 if (khugepaged_node_load[nid] > max_value) {
831 max_value = khugepaged_node_load[nid];
832 target_node = nid;
833 }
834
835 /* do some balance if several nodes have the same hit record */
836 if (target_node <= last_khugepaged_target_node)
837 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
838 nid++)
839 if (max_value == khugepaged_node_load[nid]) {
840 target_node = nid;
841 break;
842 }
843
844 last_khugepaged_target_node = target_node;
845 return target_node;
846}
847
848static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
849{
850 if (IS_ERR(*hpage)) {
851 if (!*wait)
852 return false;
853
854 *wait = false;
855 *hpage = NULL;
856 khugepaged_alloc_sleep();
857 } else if (*hpage) {
858 put_page(*hpage);
859 *hpage = NULL;
860 }
861
862 return true;
863}
864
865static struct page *
866khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
867{
868 VM_BUG_ON_PAGE(*hpage, *hpage);
869
870 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
871 if (unlikely(!*hpage)) {
872 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
873 *hpage = ERR_PTR(-ENOMEM);
874 return NULL;
875 }
876
877 prep_transhuge_page(*hpage);
878 count_vm_event(THP_COLLAPSE_ALLOC);
879 return *hpage;
880}
881#else
882static int khugepaged_find_target_node(void)
883{
884 return 0;
885}
886
887static inline struct page *alloc_khugepaged_hugepage(void)
888{
889 struct page *page;
890
891 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
892 HPAGE_PMD_ORDER);
893 if (page)
894 prep_transhuge_page(page);
895 return page;
896}
897
898static struct page *khugepaged_alloc_hugepage(bool *wait)
899{
900 struct page *hpage;
901
902 do {
903 hpage = alloc_khugepaged_hugepage();
904 if (!hpage) {
905 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
906 if (!*wait)
907 return NULL;
908
909 *wait = false;
910 khugepaged_alloc_sleep();
911 } else
912 count_vm_event(THP_COLLAPSE_ALLOC);
913 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
914
915 return hpage;
916}
917
918static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
919{
920 /*
921 * If the hpage allocated earlier was briefly exposed in page cache
922 * before collapse_file() failed, it is possible that racing lookups
923 * have not yet completed, and would then be unpleasantly surprised by
924 * finding the hpage reused for the same mapping at a different offset.
925 * Just release the previous allocation if there is any danger of that.
926 */
927 if (*hpage && page_count(*hpage) > 1) {
928 put_page(*hpage);
929 *hpage = NULL;
930 }
931
932 if (!*hpage)
933 *hpage = khugepaged_alloc_hugepage(wait);
934
935 if (unlikely(!*hpage))
936 return false;
937
938 return true;
939}
940
941static struct page *
942khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
943{
944 VM_BUG_ON(!*hpage);
945
946 return *hpage;
947}
948#endif
949
950/*
951 * If mmap_lock temporarily dropped, revalidate vma
952 * before taking mmap_lock.
953 * Return 0 if succeeds, otherwise return none-zero
954 * value (scan code).
955 */
956
957static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
958 struct vm_area_struct **vmap)
959{
960 struct vm_area_struct *vma;
961 unsigned long hstart, hend;
962
963 if (unlikely(khugepaged_test_exit(mm)))
964 return SCAN_ANY_PROCESS;
965
966 *vmap = vma = find_vma(mm, address);
967 if (!vma)
968 return SCAN_VMA_NULL;
969
970 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
971 hend = vma->vm_end & HPAGE_PMD_MASK;
972 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
973 return SCAN_ADDRESS_RANGE;
974 if (!hugepage_vma_check(vma, vma->vm_flags))
975 return SCAN_VMA_CHECK;
976 /* Anon VMA expected */
977 if (!vma->anon_vma || vma->vm_ops)
978 return SCAN_VMA_CHECK;
979 return 0;
980}
981
982/*
983 * Bring missing pages in from swap, to complete THP collapse.
984 * Only done if khugepaged_scan_pmd believes it is worthwhile.
985 *
986 * Called and returns without pte mapped or spinlocks held,
987 * but with mmap_lock held to protect against vma changes.
988 */
989
990static bool __collapse_huge_page_swapin(struct mm_struct *mm,
991 struct vm_area_struct *vma,
992 unsigned long address, pmd_t *pmd,
993 int referenced)
994{
995 int swapped_in = 0;
996 vm_fault_t ret = 0;
997 struct vm_fault vmf = {
998 .vma = vma,
999 .address = address,
1000 .flags = FAULT_FLAG_ALLOW_RETRY,
1001 .pmd = pmd,
1002 .pgoff = linear_page_index(vma, address),
1003 };
1004
1005 vmf.pte = pte_offset_map(pmd, address);
1006 for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
1007 vmf.pte++, vmf.address += PAGE_SIZE) {
1008 vmf.orig_pte = *vmf.pte;
1009 if (!is_swap_pte(vmf.orig_pte))
1010 continue;
1011 swapped_in++;
1012 ret = do_swap_page(&vmf);
1013
1014 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1015 if (ret & VM_FAULT_RETRY) {
1016 mmap_read_lock(mm);
1017 if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
1018 /* vma is no longer available, don't continue to swapin */
1019 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1020 return false;
1021 }
1022 /* check if the pmd is still valid */
1023 if (mm_find_pmd(mm, address) != pmd) {
1024 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1025 return false;
1026 }
1027 }
1028 if (ret & VM_FAULT_ERROR) {
1029 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1030 return false;
1031 }
1032 /* pte is unmapped now, we need to map it */
1033 vmf.pte = pte_offset_map(pmd, vmf.address);
1034 }
1035 vmf.pte--;
1036 pte_unmap(vmf.pte);
1037
1038 /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1039 if (swapped_in)
1040 lru_add_drain();
1041
1042 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1043 return true;
1044}
1045
1046static void collapse_huge_page(struct mm_struct *mm,
1047 unsigned long address,
1048 struct page **hpage,
1049 int node, int referenced, int unmapped)
1050{
1051 LIST_HEAD(compound_pagelist);
1052 pmd_t *pmd, _pmd;
1053 pte_t *pte;
1054 pgtable_t pgtable;
1055 struct page *new_page;
1056 spinlock_t *pmd_ptl, *pte_ptl;
1057 int isolated = 0, result = 0;
1058 struct vm_area_struct *vma;
1059 struct mmu_notifier_range range;
1060 gfp_t gfp;
1061
1062 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1063
1064 /* Only allocate from the target node */
1065 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1066
1067 /*
1068 * Before allocating the hugepage, release the mmap_lock read lock.
1069 * The allocation can take potentially a long time if it involves
1070 * sync compaction, and we do not need to hold the mmap_lock during
1071 * that. We will recheck the vma after taking it again in write mode.
1072 */
1073 mmap_read_unlock(mm);
1074 new_page = khugepaged_alloc_page(hpage, gfp, node);
1075 if (!new_page) {
1076 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1077 goto out_nolock;
1078 }
1079
1080 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1081 result = SCAN_CGROUP_CHARGE_FAIL;
1082 goto out_nolock;
1083 }
1084 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1085
1086 mmap_read_lock(mm);
1087 result = hugepage_vma_revalidate(mm, address, &vma);
1088 if (result) {
1089 mmap_read_unlock(mm);
1090 goto out_nolock;
1091 }
1092
1093 pmd = mm_find_pmd(mm, address);
1094 if (!pmd) {
1095 result = SCAN_PMD_NULL;
1096 mmap_read_unlock(mm);
1097 goto out_nolock;
1098 }
1099
1100 /*
1101 * __collapse_huge_page_swapin always returns with mmap_lock locked.
1102 * If it fails, we release mmap_lock and jump out_nolock.
1103 * Continuing to collapse causes inconsistency.
1104 */
1105 if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1106 pmd, referenced)) {
1107 mmap_read_unlock(mm);
1108 goto out_nolock;
1109 }
1110
1111 mmap_read_unlock(mm);
1112 /*
1113 * Prevent all access to pagetables with the exception of
1114 * gup_fast later handled by the ptep_clear_flush and the VM
1115 * handled by the anon_vma lock + PG_lock.
1116 */
1117 mmap_write_lock(mm);
1118 result = hugepage_vma_revalidate(mm, address, &vma);
1119 if (result)
1120 goto out;
1121 /* check if the pmd is still valid */
1122 if (mm_find_pmd(mm, address) != pmd)
1123 goto out;
1124
1125 anon_vma_lock_write(vma->anon_vma);
1126
1127 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1128 address, address + HPAGE_PMD_SIZE);
1129 mmu_notifier_invalidate_range_start(&range);
1130
1131 pte = pte_offset_map(pmd, address);
1132 pte_ptl = pte_lockptr(mm, pmd);
1133
1134 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1135 /*
1136 * After this gup_fast can't run anymore. This also removes
1137 * any huge TLB entry from the CPU so we won't allow
1138 * huge and small TLB entries for the same virtual address
1139 * to avoid the risk of CPU bugs in that area.
1140 */
1141 _pmd = pmdp_collapse_flush(vma, address, pmd);
1142 spin_unlock(pmd_ptl);
1143 mmu_notifier_invalidate_range_end(&range);
1144
1145 spin_lock(pte_ptl);
1146 isolated = __collapse_huge_page_isolate(vma, address, pte,
1147 &compound_pagelist);
1148 spin_unlock(pte_ptl);
1149
1150 if (unlikely(!isolated)) {
1151 pte_unmap(pte);
1152 spin_lock(pmd_ptl);
1153 BUG_ON(!pmd_none(*pmd));
1154 /*
1155 * We can only use set_pmd_at when establishing
1156 * hugepmds and never for establishing regular pmds that
1157 * points to regular pagetables. Use pmd_populate for that
1158 */
1159 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1160 spin_unlock(pmd_ptl);
1161 anon_vma_unlock_write(vma->anon_vma);
1162 result = SCAN_FAIL;
1163 goto out;
1164 }
1165
1166 /*
1167 * All pages are isolated and locked so anon_vma rmap
1168 * can't run anymore.
1169 */
1170 anon_vma_unlock_write(vma->anon_vma);
1171
1172 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1173 &compound_pagelist);
1174 pte_unmap(pte);
1175 __SetPageUptodate(new_page);
1176 pgtable = pmd_pgtable(_pmd);
1177
1178 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1179 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1180
1181 /*
1182 * spin_lock() below is not the equivalent of smp_wmb(), so
1183 * this is needed to avoid the copy_huge_page writes to become
1184 * visible after the set_pmd_at() write.
1185 */
1186 smp_wmb();
1187
1188 spin_lock(pmd_ptl);
1189 BUG_ON(!pmd_none(*pmd));
1190 page_add_new_anon_rmap(new_page, vma, address, true);
1191 lru_cache_add_inactive_or_unevictable(new_page, vma);
1192 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1193 set_pmd_at(mm, address, pmd, _pmd);
1194 update_mmu_cache_pmd(vma, address, pmd);
1195 spin_unlock(pmd_ptl);
1196
1197 *hpage = NULL;
1198
1199 khugepaged_pages_collapsed++;
1200 result = SCAN_SUCCEED;
1201out_up_write:
1202 mmap_write_unlock(mm);
1203out_nolock:
1204 if (!IS_ERR_OR_NULL(*hpage))
1205 mem_cgroup_uncharge(*hpage);
1206 trace_mm_collapse_huge_page(mm, isolated, result);
1207 return;
1208out:
1209 goto out_up_write;
1210}
1211
1212static int khugepaged_scan_pmd(struct mm_struct *mm,
1213 struct vm_area_struct *vma,
1214 unsigned long address,
1215 struct page **hpage)
1216{
1217 pmd_t *pmd;
1218 pte_t *pte, *_pte;
1219 int ret = 0, result = 0, referenced = 0;
1220 int none_or_zero = 0, shared = 0;
1221 struct page *page = NULL;
1222 unsigned long _address;
1223 spinlock_t *ptl;
1224 int node = NUMA_NO_NODE, unmapped = 0;
1225 bool writable = false;
1226
1227 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1228
1229 pmd = mm_find_pmd(mm, address);
1230 if (!pmd) {
1231 result = SCAN_PMD_NULL;
1232 goto out;
1233 }
1234
1235 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1236 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1237 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1238 _pte++, _address += PAGE_SIZE) {
1239 pte_t pteval = *_pte;
1240 if (is_swap_pte(pteval)) {
1241 if (++unmapped <= khugepaged_max_ptes_swap) {
1242 /*
1243 * Always be strict with uffd-wp
1244 * enabled swap entries. Please see
1245 * comment below for pte_uffd_wp().
1246 */
1247 if (pte_swp_uffd_wp(pteval)) {
1248 result = SCAN_PTE_UFFD_WP;
1249 goto out_unmap;
1250 }
1251 continue;
1252 } else {
1253 result = SCAN_EXCEED_SWAP_PTE;
1254 goto out_unmap;
1255 }
1256 }
1257 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1258 if (!userfaultfd_armed(vma) &&
1259 ++none_or_zero <= khugepaged_max_ptes_none) {
1260 continue;
1261 } else {
1262 result = SCAN_EXCEED_NONE_PTE;
1263 goto out_unmap;
1264 }
1265 }
1266 if (!pte_present(pteval)) {
1267 result = SCAN_PTE_NON_PRESENT;
1268 goto out_unmap;
1269 }
1270 if (pte_uffd_wp(pteval)) {
1271 /*
1272 * Don't collapse the page if any of the small
1273 * PTEs are armed with uffd write protection.
1274 * Here we can also mark the new huge pmd as
1275 * write protected if any of the small ones is
1276 * marked but that could bring uknown
1277 * userfault messages that falls outside of
1278 * the registered range. So, just be simple.
1279 */
1280 result = SCAN_PTE_UFFD_WP;
1281 goto out_unmap;
1282 }
1283 if (pte_write(pteval))
1284 writable = true;
1285
1286 page = vm_normal_page(vma, _address, pteval);
1287 if (unlikely(!page)) {
1288 result = SCAN_PAGE_NULL;
1289 goto out_unmap;
1290 }
1291
1292 if (page_mapcount(page) > 1 &&
1293 ++shared > khugepaged_max_ptes_shared) {
1294 result = SCAN_EXCEED_SHARED_PTE;
1295 goto out_unmap;
1296 }
1297
1298 page = compound_head(page);
1299
1300 /*
1301 * Record which node the original page is from and save this
1302 * information to khugepaged_node_load[].
1303 * Khupaged will allocate hugepage from the node has the max
1304 * hit record.
1305 */
1306 node = page_to_nid(page);
1307 if (khugepaged_scan_abort(node)) {
1308 result = SCAN_SCAN_ABORT;
1309 goto out_unmap;
1310 }
1311 khugepaged_node_load[node]++;
1312 if (!PageLRU(page)) {
1313 result = SCAN_PAGE_LRU;
1314 goto out_unmap;
1315 }
1316 if (PageLocked(page)) {
1317 result = SCAN_PAGE_LOCK;
1318 goto out_unmap;
1319 }
1320 if (!PageAnon(page)) {
1321 result = SCAN_PAGE_ANON;
1322 goto out_unmap;
1323 }
1324
1325 /*
1326 * Check if the page has any GUP (or other external) pins.
1327 *
1328 * Here the check is racy it may see totmal_mapcount > refcount
1329 * in some cases.
1330 * For example, one process with one forked child process.
1331 * The parent has the PMD split due to MADV_DONTNEED, then
1332 * the child is trying unmap the whole PMD, but khugepaged
1333 * may be scanning the parent between the child has
1334 * PageDoubleMap flag cleared and dec the mapcount. So
1335 * khugepaged may see total_mapcount > refcount.
1336 *
1337 * But such case is ephemeral we could always retry collapse
1338 * later. However it may report false positive if the page
1339 * has excessive GUP pins (i.e. 512). Anyway the same check
1340 * will be done again later the risk seems low.
1341 */
1342 if (!is_refcount_suitable(page)) {
1343 result = SCAN_PAGE_COUNT;
1344 goto out_unmap;
1345 }
1346 if (pte_young(pteval) ||
1347 page_is_young(page) || PageReferenced(page) ||
1348 mmu_notifier_test_young(vma->vm_mm, address))
1349 referenced++;
1350 }
1351 if (!writable) {
1352 result = SCAN_PAGE_RO;
1353 } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1354 result = SCAN_LACK_REFERENCED_PAGE;
1355 } else {
1356 result = SCAN_SUCCEED;
1357 ret = 1;
1358 }
1359out_unmap:
1360 pte_unmap_unlock(pte, ptl);
1361 if (ret) {
1362 node = khugepaged_find_target_node();
1363 /* collapse_huge_page will return with the mmap_lock released */
1364 collapse_huge_page(mm, address, hpage, node,
1365 referenced, unmapped);
1366 }
1367out:
1368 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1369 none_or_zero, result, unmapped);
1370 return ret;
1371}
1372
1373static void collect_mm_slot(struct mm_slot *mm_slot)
1374{
1375 struct mm_struct *mm = mm_slot->mm;
1376
1377 lockdep_assert_held(&khugepaged_mm_lock);
1378
1379 if (khugepaged_test_exit(mm)) {
1380 /* free mm_slot */
1381 hash_del(&mm_slot->hash);
1382 list_del(&mm_slot->mm_node);
1383
1384 /*
1385 * Not strictly needed because the mm exited already.
1386 *
1387 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1388 */
1389
1390 /* khugepaged_mm_lock actually not necessary for the below */
1391 free_mm_slot(mm_slot);
1392 mmdrop(mm);
1393 }
1394}
1395
1396#ifdef CONFIG_SHMEM
1397/*
1398 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1399 * khugepaged should try to collapse the page table.
1400 */
1401static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1402 unsigned long addr)
1403{
1404 struct mm_slot *mm_slot;
1405
1406 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1407
1408 spin_lock(&khugepaged_mm_lock);
1409 mm_slot = get_mm_slot(mm);
1410 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1411 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1412 spin_unlock(&khugepaged_mm_lock);
1413 return 0;
1414}
1415
1416/**
1417 * Try to collapse a pte-mapped THP for mm at address haddr.
1418 *
1419 * This function checks whether all the PTEs in the PMD are pointing to the
1420 * right THP. If so, retract the page table so the THP can refault in with
1421 * as pmd-mapped.
1422 */
1423void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1424{
1425 unsigned long haddr = addr & HPAGE_PMD_MASK;
1426 struct vm_area_struct *vma = find_vma(mm, haddr);
1427 struct page *hpage;
1428 pte_t *start_pte, *pte;
1429 pmd_t *pmd, _pmd;
1430 spinlock_t *ptl;
1431 int count = 0;
1432 int i;
1433
1434 if (!vma || !vma->vm_file ||
1435 vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1436 return;
1437
1438 /*
1439 * This vm_flags may not have VM_HUGEPAGE if the page was not
1440 * collapsed by this mm. But we can still collapse if the page is
1441 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1442 * will not fail the vma for missing VM_HUGEPAGE
1443 */
1444 if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1445 return;
1446
1447 hpage = find_lock_page(vma->vm_file->f_mapping,
1448 linear_page_index(vma, haddr));
1449 if (!hpage)
1450 return;
1451
1452 if (!PageHead(hpage))
1453 goto drop_hpage;
1454
1455 pmd = mm_find_pmd(mm, haddr);
1456 if (!pmd)
1457 goto drop_hpage;
1458
1459 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1460
1461 /* step 1: check all mapped PTEs are to the right huge page */
1462 for (i = 0, addr = haddr, pte = start_pte;
1463 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1464 struct page *page;
1465
1466 /* empty pte, skip */
1467 if (pte_none(*pte))
1468 continue;
1469
1470 /* page swapped out, abort */
1471 if (!pte_present(*pte))
1472 goto abort;
1473
1474 page = vm_normal_page(vma, addr, *pte);
1475
1476 /*
1477 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1478 * page table, but the new page will not be a subpage of hpage.
1479 */
1480 if (hpage + i != page)
1481 goto abort;
1482 count++;
1483 }
1484
1485 /* step 2: adjust rmap */
1486 for (i = 0, addr = haddr, pte = start_pte;
1487 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1488 struct page *page;
1489
1490 if (pte_none(*pte))
1491 continue;
1492 page = vm_normal_page(vma, addr, *pte);
1493 page_remove_rmap(page, false);
1494 }
1495
1496 pte_unmap_unlock(start_pte, ptl);
1497
1498 /* step 3: set proper refcount and mm_counters. */
1499 if (count) {
1500 page_ref_sub(hpage, count);
1501 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1502 }
1503
1504 /* step 4: collapse pmd */
1505 ptl = pmd_lock(vma->vm_mm, pmd);
1506 _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1507 spin_unlock(ptl);
1508 mm_dec_nr_ptes(mm);
1509 pte_free(mm, pmd_pgtable(_pmd));
1510
1511drop_hpage:
1512 unlock_page(hpage);
1513 put_page(hpage);
1514 return;
1515
1516abort:
1517 pte_unmap_unlock(start_pte, ptl);
1518 goto drop_hpage;
1519}
1520
1521static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1522{
1523 struct mm_struct *mm = mm_slot->mm;
1524 int i;
1525
1526 if (likely(mm_slot->nr_pte_mapped_thp == 0))
1527 return 0;
1528
1529 if (!mmap_write_trylock(mm))
1530 return -EBUSY;
1531
1532 if (unlikely(khugepaged_test_exit(mm)))
1533 goto out;
1534
1535 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1536 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1537
1538out:
1539 mm_slot->nr_pte_mapped_thp = 0;
1540 mmap_write_unlock(mm);
1541 return 0;
1542}
1543
1544static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1545{
1546 struct vm_area_struct *vma;
1547 struct mm_struct *mm;
1548 unsigned long addr;
1549 pmd_t *pmd, _pmd;
1550
1551 i_mmap_lock_write(mapping);
1552 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1553 /*
1554 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1555 * got written to. These VMAs are likely not worth investing
1556 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1557 * later.
1558 *
1559 * Not that vma->anon_vma check is racy: it can be set up after
1560 * the check but before we took mmap_lock by the fault path.
1561 * But page lock would prevent establishing any new ptes of the
1562 * page, so we are safe.
1563 *
1564 * An alternative would be drop the check, but check that page
1565 * table is clear before calling pmdp_collapse_flush() under
1566 * ptl. It has higher chance to recover THP for the VMA, but
1567 * has higher cost too.
1568 */
1569 if (vma->anon_vma)
1570 continue;
1571 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1572 if (addr & ~HPAGE_PMD_MASK)
1573 continue;
1574 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1575 continue;
1576 mm = vma->vm_mm;
1577 pmd = mm_find_pmd(mm, addr);
1578 if (!pmd)
1579 continue;
1580 /*
1581 * We need exclusive mmap_lock to retract page table.
1582 *
1583 * We use trylock due to lock inversion: we need to acquire
1584 * mmap_lock while holding page lock. Fault path does it in
1585 * reverse order. Trylock is a way to avoid deadlock.
1586 */
1587 if (mmap_write_trylock(mm)) {
1588 if (!khugepaged_test_exit(mm)) {
1589 spinlock_t *ptl = pmd_lock(mm, pmd);
1590 /* assume page table is clear */
1591 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1592 spin_unlock(ptl);
1593 mm_dec_nr_ptes(mm);
1594 pte_free(mm, pmd_pgtable(_pmd));
1595 }
1596 mmap_write_unlock(mm);
1597 } else {
1598 /* Try again later */
1599 khugepaged_add_pte_mapped_thp(mm, addr);
1600 }
1601 }
1602 i_mmap_unlock_write(mapping);
1603}
1604
1605/**
1606 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1607 *
1608 * Basic scheme is simple, details are more complex:
1609 * - allocate and lock a new huge page;
1610 * - scan page cache replacing old pages with the new one
1611 * + swap/gup in pages if necessary;
1612 * + fill in gaps;
1613 * + keep old pages around in case rollback is required;
1614 * - if replacing succeeds:
1615 * + copy data over;
1616 * + free old pages;
1617 * + unlock huge page;
1618 * - if replacing failed;
1619 * + put all pages back and unfreeze them;
1620 * + restore gaps in the page cache;
1621 * + unlock and free huge page;
1622 */
1623static void collapse_file(struct mm_struct *mm,
1624 struct file *file, pgoff_t start,
1625 struct page **hpage, int node)
1626{
1627 struct address_space *mapping = file->f_mapping;
1628 gfp_t gfp;
1629 struct page *new_page;
1630 pgoff_t index, end = start + HPAGE_PMD_NR;
1631 LIST_HEAD(pagelist);
1632 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1633 int nr_none = 0, result = SCAN_SUCCEED;
1634 bool is_shmem = shmem_file(file);
1635
1636 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1637 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1638
1639 /* Only allocate from the target node */
1640 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1641
1642 new_page = khugepaged_alloc_page(hpage, gfp, node);
1643 if (!new_page) {
1644 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1645 goto out;
1646 }
1647
1648 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1649 result = SCAN_CGROUP_CHARGE_FAIL;
1650 goto out;
1651 }
1652 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1653
1654 /* This will be less messy when we use multi-index entries */
1655 do {
1656 xas_lock_irq(&xas);
1657 xas_create_range(&xas);
1658 if (!xas_error(&xas))
1659 break;
1660 xas_unlock_irq(&xas);
1661 if (!xas_nomem(&xas, GFP_KERNEL)) {
1662 result = SCAN_FAIL;
1663 goto out;
1664 }
1665 } while (1);
1666
1667 __SetPageLocked(new_page);
1668 if (is_shmem)
1669 __SetPageSwapBacked(new_page);
1670 new_page->index = start;
1671 new_page->mapping = mapping;
1672
1673 /*
1674 * At this point the new_page is locked and not up-to-date.
1675 * It's safe to insert it into the page cache, because nobody would
1676 * be able to map it or use it in another way until we unlock it.
1677 */
1678
1679 xas_set(&xas, start);
1680 for (index = start; index < end; index++) {
1681 struct page *page = xas_next(&xas);
1682
1683 VM_BUG_ON(index != xas.xa_index);
1684 if (is_shmem) {
1685 if (!page) {
1686 /*
1687 * Stop if extent has been truncated or
1688 * hole-punched, and is now completely
1689 * empty.
1690 */
1691 if (index == start) {
1692 if (!xas_next_entry(&xas, end - 1)) {
1693 result = SCAN_TRUNCATED;
1694 goto xa_locked;
1695 }
1696 xas_set(&xas, index);
1697 }
1698 if (!shmem_charge(mapping->host, 1)) {
1699 result = SCAN_FAIL;
1700 goto xa_locked;
1701 }
1702 xas_store(&xas, new_page);
1703 nr_none++;
1704 continue;
1705 }
1706
1707 if (xa_is_value(page) || !PageUptodate(page)) {
1708 xas_unlock_irq(&xas);
1709 /* swap in or instantiate fallocated page */
1710 if (shmem_getpage(mapping->host, index, &page,
1711 SGP_NOHUGE)) {
1712 result = SCAN_FAIL;
1713 goto xa_unlocked;
1714 }
1715 } else if (trylock_page(page)) {
1716 get_page(page);
1717 xas_unlock_irq(&xas);
1718 } else {
1719 result = SCAN_PAGE_LOCK;
1720 goto xa_locked;
1721 }
1722 } else { /* !is_shmem */
1723 if (!page || xa_is_value(page)) {
1724 xas_unlock_irq(&xas);
1725 page_cache_sync_readahead(mapping, &file->f_ra,
1726 file, index,
1727 end - index);
1728 /* drain pagevecs to help isolate_lru_page() */
1729 lru_add_drain();
1730 page = find_lock_page(mapping, index);
1731 if (unlikely(page == NULL)) {
1732 result = SCAN_FAIL;
1733 goto xa_unlocked;
1734 }
1735 } else if (PageDirty(page)) {
1736 /*
1737 * khugepaged only works on read-only fd,
1738 * so this page is dirty because it hasn't
1739 * been flushed since first write. There
1740 * won't be new dirty pages.
1741 *
1742 * Trigger async flush here and hope the
1743 * writeback is done when khugepaged
1744 * revisits this page.
1745 *
1746 * This is a one-off situation. We are not
1747 * forcing writeback in loop.
1748 */
1749 xas_unlock_irq(&xas);
1750 filemap_flush(mapping);
1751 result = SCAN_FAIL;
1752 goto xa_unlocked;
1753 } else if (trylock_page(page)) {
1754 get_page(page);
1755 xas_unlock_irq(&xas);
1756 } else {
1757 result = SCAN_PAGE_LOCK;
1758 goto xa_locked;
1759 }
1760 }
1761
1762 /*
1763 * The page must be locked, so we can drop the i_pages lock
1764 * without racing with truncate.
1765 */
1766 VM_BUG_ON_PAGE(!PageLocked(page), page);
1767
1768 /* make sure the page is up to date */
1769 if (unlikely(!PageUptodate(page))) {
1770 result = SCAN_FAIL;
1771 goto out_unlock;
1772 }
1773
1774 /*
1775 * If file was truncated then extended, or hole-punched, before
1776 * we locked the first page, then a THP might be there already.
1777 */
1778 if (PageTransCompound(page)) {
1779 result = SCAN_PAGE_COMPOUND;
1780 goto out_unlock;
1781 }
1782
1783 if (page_mapping(page) != mapping) {
1784 result = SCAN_TRUNCATED;
1785 goto out_unlock;
1786 }
1787
1788 if (!is_shmem && PageDirty(page)) {
1789 /*
1790 * khugepaged only works on read-only fd, so this
1791 * page is dirty because it hasn't been flushed
1792 * since first write.
1793 */
1794 result = SCAN_FAIL;
1795 goto out_unlock;
1796 }
1797
1798 if (isolate_lru_page(page)) {
1799 result = SCAN_DEL_PAGE_LRU;
1800 goto out_unlock;
1801 }
1802
1803 if (page_has_private(page) &&
1804 !try_to_release_page(page, GFP_KERNEL)) {
1805 result = SCAN_PAGE_HAS_PRIVATE;
1806 putback_lru_page(page);
1807 goto out_unlock;
1808 }
1809
1810 if (page_mapped(page))
1811 unmap_mapping_pages(mapping, index, 1, false);
1812
1813 xas_lock_irq(&xas);
1814 xas_set(&xas, index);
1815
1816 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1817 VM_BUG_ON_PAGE(page_mapped(page), page);
1818
1819 /*
1820 * The page is expected to have page_count() == 3:
1821 * - we hold a pin on it;
1822 * - one reference from page cache;
1823 * - one from isolate_lru_page;
1824 */
1825 if (!page_ref_freeze(page, 3)) {
1826 result = SCAN_PAGE_COUNT;
1827 xas_unlock_irq(&xas);
1828 putback_lru_page(page);
1829 goto out_unlock;
1830 }
1831
1832 /*
1833 * Add the page to the list to be able to undo the collapse if
1834 * something go wrong.
1835 */
1836 list_add_tail(&page->lru, &pagelist);
1837
1838 /* Finally, replace with the new page. */
1839 xas_store(&xas, new_page);
1840 continue;
1841out_unlock:
1842 unlock_page(page);
1843 put_page(page);
1844 goto xa_unlocked;
1845 }
1846
1847 if (is_shmem)
1848 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1849 else {
1850 __inc_node_page_state(new_page, NR_FILE_THPS);
1851 filemap_nr_thps_inc(mapping);
1852 }
1853
1854 if (nr_none) {
1855 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1856 if (is_shmem)
1857 __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1858 }
1859
1860xa_locked:
1861 xas_unlock_irq(&xas);
1862xa_unlocked:
1863
1864 if (result == SCAN_SUCCEED) {
1865 struct page *page, *tmp;
1866
1867 /*
1868 * Replacing old pages with new one has succeeded, now we
1869 * need to copy the content and free the old pages.
1870 */
1871 index = start;
1872 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1873 while (index < page->index) {
1874 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1875 index++;
1876 }
1877 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1878 page);
1879 list_del(&page->lru);
1880 page->mapping = NULL;
1881 page_ref_unfreeze(page, 1);
1882 ClearPageActive(page);
1883 ClearPageUnevictable(page);
1884 unlock_page(page);
1885 put_page(page);
1886 index++;
1887 }
1888 while (index < end) {
1889 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1890 index++;
1891 }
1892
1893 SetPageUptodate(new_page);
1894 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1895 if (is_shmem)
1896 set_page_dirty(new_page);
1897 lru_cache_add(new_page);
1898
1899 /*
1900 * Remove pte page tables, so we can re-fault the page as huge.
1901 */
1902 retract_page_tables(mapping, start);
1903 *hpage = NULL;
1904
1905 khugepaged_pages_collapsed++;
1906 } else {
1907 struct page *page;
1908
1909 /* Something went wrong: roll back page cache changes */
1910 xas_lock_irq(&xas);
1911 mapping->nrpages -= nr_none;
1912
1913 if (is_shmem)
1914 shmem_uncharge(mapping->host, nr_none);
1915
1916 xas_set(&xas, start);
1917 xas_for_each(&xas, page, end - 1) {
1918 page = list_first_entry_or_null(&pagelist,
1919 struct page, lru);
1920 if (!page || xas.xa_index < page->index) {
1921 if (!nr_none)
1922 break;
1923 nr_none--;
1924 /* Put holes back where they were */
1925 xas_store(&xas, NULL);
1926 continue;
1927 }
1928
1929 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1930
1931 /* Unfreeze the page. */
1932 list_del(&page->lru);
1933 page_ref_unfreeze(page, 2);
1934 xas_store(&xas, page);
1935 xas_pause(&xas);
1936 xas_unlock_irq(&xas);
1937 unlock_page(page);
1938 putback_lru_page(page);
1939 xas_lock_irq(&xas);
1940 }
1941 VM_BUG_ON(nr_none);
1942 xas_unlock_irq(&xas);
1943
1944 new_page->mapping = NULL;
1945 }
1946
1947 unlock_page(new_page);
1948out:
1949 VM_BUG_ON(!list_empty(&pagelist));
1950 if (!IS_ERR_OR_NULL(*hpage))
1951 mem_cgroup_uncharge(*hpage);
1952 /* TODO: tracepoints */
1953}
1954
1955static void khugepaged_scan_file(struct mm_struct *mm,
1956 struct file *file, pgoff_t start, struct page **hpage)
1957{
1958 struct page *page = NULL;
1959 struct address_space *mapping = file->f_mapping;
1960 XA_STATE(xas, &mapping->i_pages, start);
1961 int present, swap;
1962 int node = NUMA_NO_NODE;
1963 int result = SCAN_SUCCEED;
1964
1965 present = 0;
1966 swap = 0;
1967 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1968 rcu_read_lock();
1969 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1970 if (xas_retry(&xas, page))
1971 continue;
1972
1973 if (xa_is_value(page)) {
1974 if (++swap > khugepaged_max_ptes_swap) {
1975 result = SCAN_EXCEED_SWAP_PTE;
1976 break;
1977 }
1978 continue;
1979 }
1980
1981 if (PageTransCompound(page)) {
1982 result = SCAN_PAGE_COMPOUND;
1983 break;
1984 }
1985
1986 node = page_to_nid(page);
1987 if (khugepaged_scan_abort(node)) {
1988 result = SCAN_SCAN_ABORT;
1989 break;
1990 }
1991 khugepaged_node_load[node]++;
1992
1993 if (!PageLRU(page)) {
1994 result = SCAN_PAGE_LRU;
1995 break;
1996 }
1997
1998 if (page_count(page) !=
1999 1 + page_mapcount(page) + page_has_private(page)) {
2000 result = SCAN_PAGE_COUNT;
2001 break;
2002 }
2003
2004 /*
2005 * We probably should check if the page is referenced here, but
2006 * nobody would transfer pte_young() to PageReferenced() for us.
2007 * And rmap walk here is just too costly...
2008 */
2009
2010 present++;
2011
2012 if (need_resched()) {
2013 xas_pause(&xas);
2014 cond_resched_rcu();
2015 }
2016 }
2017 rcu_read_unlock();
2018
2019 if (result == SCAN_SUCCEED) {
2020 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2021 result = SCAN_EXCEED_NONE_PTE;
2022 } else {
2023 node = khugepaged_find_target_node();
2024 collapse_file(mm, file, start, hpage, node);
2025 }
2026 }
2027
2028 /* TODO: tracepoints */
2029}
2030#else
2031static void khugepaged_scan_file(struct mm_struct *mm,
2032 struct file *file, pgoff_t start, struct page **hpage)
2033{
2034 BUILD_BUG();
2035}
2036
2037static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2038{
2039 return 0;
2040}
2041#endif
2042
2043static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2044 struct page **hpage)
2045 __releases(&khugepaged_mm_lock)
2046 __acquires(&khugepaged_mm_lock)
2047{
2048 struct mm_slot *mm_slot;
2049 struct mm_struct *mm;
2050 struct vm_area_struct *vma;
2051 int progress = 0;
2052
2053 VM_BUG_ON(!pages);
2054 lockdep_assert_held(&khugepaged_mm_lock);
2055
2056 if (khugepaged_scan.mm_slot)
2057 mm_slot = khugepaged_scan.mm_slot;
2058 else {
2059 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2060 struct mm_slot, mm_node);
2061 khugepaged_scan.address = 0;
2062 khugepaged_scan.mm_slot = mm_slot;
2063 }
2064 spin_unlock(&khugepaged_mm_lock);
2065 khugepaged_collapse_pte_mapped_thps(mm_slot);
2066
2067 mm = mm_slot->mm;
2068 /*
2069 * Don't wait for semaphore (to avoid long wait times). Just move to
2070 * the next mm on the list.
2071 */
2072 vma = NULL;
2073 if (unlikely(!mmap_read_trylock(mm)))
2074 goto breakouterloop_mmap_lock;
2075 if (likely(!khugepaged_test_exit(mm)))
2076 vma = find_vma(mm, khugepaged_scan.address);
2077
2078 progress++;
2079 for (; vma; vma = vma->vm_next) {
2080 unsigned long hstart, hend;
2081
2082 cond_resched();
2083 if (unlikely(khugepaged_test_exit(mm))) {
2084 progress++;
2085 break;
2086 }
2087 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2088skip:
2089 progress++;
2090 continue;
2091 }
2092 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2093 hend = vma->vm_end & HPAGE_PMD_MASK;
2094 if (hstart >= hend)
2095 goto skip;
2096 if (khugepaged_scan.address > hend)
2097 goto skip;
2098 if (khugepaged_scan.address < hstart)
2099 khugepaged_scan.address = hstart;
2100 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2101 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2102 goto skip;
2103
2104 while (khugepaged_scan.address < hend) {
2105 int ret;
2106 cond_resched();
2107 if (unlikely(khugepaged_test_exit(mm)))
2108 goto breakouterloop;
2109
2110 VM_BUG_ON(khugepaged_scan.address < hstart ||
2111 khugepaged_scan.address + HPAGE_PMD_SIZE >
2112 hend);
2113 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2114 struct file *file = get_file(vma->vm_file);
2115 pgoff_t pgoff = linear_page_index(vma,
2116 khugepaged_scan.address);
2117
2118 mmap_read_unlock(mm);
2119 ret = 1;
2120 khugepaged_scan_file(mm, file, pgoff, hpage);
2121 fput(file);
2122 } else {
2123 ret = khugepaged_scan_pmd(mm, vma,
2124 khugepaged_scan.address,
2125 hpage);
2126 }
2127 /* move to next address */
2128 khugepaged_scan.address += HPAGE_PMD_SIZE;
2129 progress += HPAGE_PMD_NR;
2130 if (ret)
2131 /* we released mmap_lock so break loop */
2132 goto breakouterloop_mmap_lock;
2133 if (progress >= pages)
2134 goto breakouterloop;
2135 }
2136 }
2137breakouterloop:
2138 mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2139breakouterloop_mmap_lock:
2140
2141 spin_lock(&khugepaged_mm_lock);
2142 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2143 /*
2144 * Release the current mm_slot if this mm is about to die, or
2145 * if we scanned all vmas of this mm.
2146 */
2147 if (khugepaged_test_exit(mm) || !vma) {
2148 /*
2149 * Make sure that if mm_users is reaching zero while
2150 * khugepaged runs here, khugepaged_exit will find
2151 * mm_slot not pointing to the exiting mm.
2152 */
2153 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2154 khugepaged_scan.mm_slot = list_entry(
2155 mm_slot->mm_node.next,
2156 struct mm_slot, mm_node);
2157 khugepaged_scan.address = 0;
2158 } else {
2159 khugepaged_scan.mm_slot = NULL;
2160 khugepaged_full_scans++;
2161 }
2162
2163 collect_mm_slot(mm_slot);
2164 }
2165
2166 return progress;
2167}
2168
2169static int khugepaged_has_work(void)
2170{
2171 return !list_empty(&khugepaged_scan.mm_head) &&
2172 khugepaged_enabled();
2173}
2174
2175static int khugepaged_wait_event(void)
2176{
2177 return !list_empty(&khugepaged_scan.mm_head) ||
2178 kthread_should_stop();
2179}
2180
2181static void khugepaged_do_scan(void)
2182{
2183 struct page *hpage = NULL;
2184 unsigned int progress = 0, pass_through_head = 0;
2185 unsigned int pages = khugepaged_pages_to_scan;
2186 bool wait = true;
2187
2188 barrier(); /* write khugepaged_pages_to_scan to local stack */
2189
2190 lru_add_drain_all();
2191
2192 while (progress < pages) {
2193 if (!khugepaged_prealloc_page(&hpage, &wait))
2194 break;
2195
2196 cond_resched();
2197
2198 if (unlikely(kthread_should_stop() || try_to_freeze()))
2199 break;
2200
2201 spin_lock(&khugepaged_mm_lock);
2202 if (!khugepaged_scan.mm_slot)
2203 pass_through_head++;
2204 if (khugepaged_has_work() &&
2205 pass_through_head < 2)
2206 progress += khugepaged_scan_mm_slot(pages - progress,
2207 &hpage);
2208 else
2209 progress = pages;
2210 spin_unlock(&khugepaged_mm_lock);
2211 }
2212
2213 if (!IS_ERR_OR_NULL(hpage))
2214 put_page(hpage);
2215}
2216
2217static bool khugepaged_should_wakeup(void)
2218{
2219 return kthread_should_stop() ||
2220 time_after_eq(jiffies, khugepaged_sleep_expire);
2221}
2222
2223static void khugepaged_wait_work(void)
2224{
2225 if (khugepaged_has_work()) {
2226 const unsigned long scan_sleep_jiffies =
2227 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2228
2229 if (!scan_sleep_jiffies)
2230 return;
2231
2232 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2233 wait_event_freezable_timeout(khugepaged_wait,
2234 khugepaged_should_wakeup(),
2235 scan_sleep_jiffies);
2236 return;
2237 }
2238
2239 if (khugepaged_enabled())
2240 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2241}
2242
2243static int khugepaged(void *none)
2244{
2245 struct mm_slot *mm_slot;
2246
2247 set_freezable();
2248 set_user_nice(current, MAX_NICE);
2249
2250 while (!kthread_should_stop()) {
2251 khugepaged_do_scan();
2252 khugepaged_wait_work();
2253 }
2254
2255 spin_lock(&khugepaged_mm_lock);
2256 mm_slot = khugepaged_scan.mm_slot;
2257 khugepaged_scan.mm_slot = NULL;
2258 if (mm_slot)
2259 collect_mm_slot(mm_slot);
2260 spin_unlock(&khugepaged_mm_lock);
2261 return 0;
2262}
2263
2264static void set_recommended_min_free_kbytes(void)
2265{
2266 struct zone *zone;
2267 int nr_zones = 0;
2268 unsigned long recommended_min;
2269
2270 for_each_populated_zone(zone) {
2271 /*
2272 * We don't need to worry about fragmentation of
2273 * ZONE_MOVABLE since it only has movable pages.
2274 */
2275 if (zone_idx(zone) > gfp_zone(GFP_USER))
2276 continue;
2277
2278 nr_zones++;
2279 }
2280
2281 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2282 recommended_min = pageblock_nr_pages * nr_zones * 2;
2283
2284 /*
2285 * Make sure that on average at least two pageblocks are almost free
2286 * of another type, one for a migratetype to fall back to and a
2287 * second to avoid subsequent fallbacks of other types There are 3
2288 * MIGRATE_TYPES we care about.
2289 */
2290 recommended_min += pageblock_nr_pages * nr_zones *
2291 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2292
2293 /* don't ever allow to reserve more than 5% of the lowmem */
2294 recommended_min = min(recommended_min,
2295 (unsigned long) nr_free_buffer_pages() / 20);
2296 recommended_min <<= (PAGE_SHIFT-10);
2297
2298 if (recommended_min > min_free_kbytes) {
2299 if (user_min_free_kbytes >= 0)
2300 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2301 min_free_kbytes, recommended_min);
2302
2303 min_free_kbytes = recommended_min;
2304 }
2305 setup_per_zone_wmarks();
2306}
2307
2308int start_stop_khugepaged(void)
2309{
2310 int err = 0;
2311
2312 mutex_lock(&khugepaged_mutex);
2313 if (khugepaged_enabled()) {
2314 if (!khugepaged_thread)
2315 khugepaged_thread = kthread_run(khugepaged, NULL,
2316 "khugepaged");
2317 if (IS_ERR(khugepaged_thread)) {
2318 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2319 err = PTR_ERR(khugepaged_thread);
2320 khugepaged_thread = NULL;
2321 goto fail;
2322 }
2323
2324 if (!list_empty(&khugepaged_scan.mm_head))
2325 wake_up_interruptible(&khugepaged_wait);
2326
2327 set_recommended_min_free_kbytes();
2328 } else if (khugepaged_thread) {
2329 kthread_stop(khugepaged_thread);
2330 khugepaged_thread = NULL;
2331 }
2332fail:
2333 mutex_unlock(&khugepaged_mutex);
2334 return err;
2335}
2336
2337void khugepaged_min_free_kbytes_update(void)
2338{
2339 mutex_lock(&khugepaged_mutex);
2340 if (khugepaged_enabled() && khugepaged_thread)
2341 set_recommended_min_free_kbytes();
2342 mutex_unlock(&khugepaged_mutex);
2343}