1 // SPDX-License-Identifier: GPL-2.0
2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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>
24 #include <asm/pgalloc.h>
36 SCAN_EXCEED_SHARED_PTE,
39 SCAN_PTE_MAPPED_HUGEPAGE,
41 SCAN_LACK_REFERENCED_PAGE,
54 SCAN_ALLOC_HUGE_PAGE_FAIL,
55 SCAN_CGROUP_CHARGE_FAIL,
57 SCAN_PAGE_HAS_PRIVATE,
60 #define CREATE_TRACE_POINTS
61 #include <trace/events/huge_memory.h>
63 static struct task_struct *khugepaged_thread __read_mostly;
64 static DEFINE_MUTEX(khugepaged_mutex);
66 /* default scan 8*512 pte (or vmas) every 30 second */
67 static unsigned int khugepaged_pages_to_scan __read_mostly;
68 static unsigned int khugepaged_pages_collapsed;
69 static unsigned int khugepaged_full_scans;
70 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
71 /* during fragmentation poll the hugepage allocator once every minute */
72 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
73 static unsigned long khugepaged_sleep_expire;
74 static DEFINE_SPINLOCK(khugepaged_mm_lock);
75 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
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
81 * Note that these are only respected if collapse was initiated by khugepaged.
83 static unsigned int khugepaged_max_ptes_none __read_mostly;
84 static unsigned int khugepaged_max_ptes_swap __read_mostly;
85 static unsigned int khugepaged_max_ptes_shared __read_mostly;
87 #define MM_SLOTS_HASH_BITS 10
88 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
90 static struct kmem_cache *mm_slot_cache __read_mostly;
92 #define MAX_PTE_MAPPED_THP 8
94 struct collapse_control {
97 /* Num pages scanned per node */
98 u32 node_load[MAX_NUMNODES];
100 /* nodemask for allocation fallback */
101 nodemask_t alloc_nmask;
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
110 struct khugepaged_mm_slot {
113 /* pte-mapped THP in this mm */
114 int nr_pte_mapped_thp;
115 unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
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
124 * There is only the one khugepaged_scan instance of this cursor structure.
126 struct khugepaged_scan {
127 struct list_head mm_head;
128 struct khugepaged_mm_slot *mm_slot;
129 unsigned long address;
132 static struct khugepaged_scan khugepaged_scan = {
133 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
137 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
138 struct kobj_attribute *attr,
141 return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
144 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
145 struct kobj_attribute *attr,
146 const char *buf, size_t count)
151 err = kstrtouint(buf, 10, &msecs);
155 khugepaged_scan_sleep_millisecs = msecs;
156 khugepaged_sleep_expire = 0;
157 wake_up_interruptible(&khugepaged_wait);
161 static struct kobj_attribute scan_sleep_millisecs_attr =
162 __ATTR_RW(scan_sleep_millisecs);
164 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
165 struct kobj_attribute *attr,
168 return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
171 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
172 struct kobj_attribute *attr,
173 const char *buf, size_t count)
178 err = kstrtouint(buf, 10, &msecs);
182 khugepaged_alloc_sleep_millisecs = msecs;
183 khugepaged_sleep_expire = 0;
184 wake_up_interruptible(&khugepaged_wait);
188 static struct kobj_attribute alloc_sleep_millisecs_attr =
189 __ATTR_RW(alloc_sleep_millisecs);
191 static ssize_t pages_to_scan_show(struct kobject *kobj,
192 struct kobj_attribute *attr,
195 return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
197 static ssize_t pages_to_scan_store(struct kobject *kobj,
198 struct kobj_attribute *attr,
199 const char *buf, size_t count)
204 err = kstrtouint(buf, 10, &pages);
208 khugepaged_pages_to_scan = pages;
212 static struct kobj_attribute pages_to_scan_attr =
213 __ATTR_RW(pages_to_scan);
215 static ssize_t pages_collapsed_show(struct kobject *kobj,
216 struct kobj_attribute *attr,
219 return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
221 static struct kobj_attribute pages_collapsed_attr =
222 __ATTR_RO(pages_collapsed);
224 static ssize_t full_scans_show(struct kobject *kobj,
225 struct kobj_attribute *attr,
228 return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
230 static struct kobj_attribute full_scans_attr =
231 __ATTR_RO(full_scans);
233 static ssize_t defrag_show(struct kobject *kobj,
234 struct kobj_attribute *attr, char *buf)
236 return single_hugepage_flag_show(kobj, attr, buf,
237 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
239 static ssize_t defrag_store(struct kobject *kobj,
240 struct kobj_attribute *attr,
241 const char *buf, size_t count)
243 return single_hugepage_flag_store(kobj, attr, buf, count,
244 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
246 static struct kobj_attribute khugepaged_defrag_attr =
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.
257 static ssize_t max_ptes_none_show(struct kobject *kobj,
258 struct kobj_attribute *attr,
261 return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
263 static ssize_t max_ptes_none_store(struct kobject *kobj,
264 struct kobj_attribute *attr,
265 const char *buf, size_t count)
268 unsigned long max_ptes_none;
270 err = kstrtoul(buf, 10, &max_ptes_none);
271 if (err || max_ptes_none > HPAGE_PMD_NR - 1)
274 khugepaged_max_ptes_none = max_ptes_none;
278 static struct kobj_attribute khugepaged_max_ptes_none_attr =
279 __ATTR_RW(max_ptes_none);
281 static ssize_t max_ptes_swap_show(struct kobject *kobj,
282 struct kobj_attribute *attr,
285 return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
288 static ssize_t max_ptes_swap_store(struct kobject *kobj,
289 struct kobj_attribute *attr,
290 const char *buf, size_t count)
293 unsigned long max_ptes_swap;
295 err = kstrtoul(buf, 10, &max_ptes_swap);
296 if (err || max_ptes_swap > HPAGE_PMD_NR - 1)
299 khugepaged_max_ptes_swap = max_ptes_swap;
304 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
305 __ATTR_RW(max_ptes_swap);
307 static ssize_t max_ptes_shared_show(struct kobject *kobj,
308 struct kobj_attribute *attr,
311 return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
314 static ssize_t max_ptes_shared_store(struct kobject *kobj,
315 struct kobj_attribute *attr,
316 const char *buf, size_t count)
319 unsigned long max_ptes_shared;
321 err = kstrtoul(buf, 10, &max_ptes_shared);
322 if (err || max_ptes_shared > HPAGE_PMD_NR - 1)
325 khugepaged_max_ptes_shared = max_ptes_shared;
330 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
331 __ATTR_RW(max_ptes_shared);
333 static 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,
346 struct attribute_group khugepaged_attr_group = {
347 .attrs = khugepaged_attr,
348 .name = "khugepaged",
350 #endif /* CONFIG_SYSFS */
352 int hugepage_madvise(struct vm_area_struct *vma,
353 unsigned long *vm_flags, int advice)
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.
363 if (mm_has_pgste(vma->vm_mm))
366 *vm_flags &= ~VM_NOHUGEPAGE;
367 *vm_flags |= VM_HUGEPAGE;
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.
373 khugepaged_enter_vma(vma, *vm_flags);
375 case MADV_NOHUGEPAGE:
376 *vm_flags &= ~VM_HUGEPAGE;
377 *vm_flags |= VM_NOHUGEPAGE;
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.
389 int __init khugepaged_init(void)
391 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
392 sizeof(struct khugepaged_mm_slot),
393 __alignof__(struct khugepaged_mm_slot),
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;
406 void __init khugepaged_destroy(void)
408 kmem_cache_destroy(mm_slot_cache);
411 static inline int hpage_collapse_test_exit(struct mm_struct *mm)
413 return atomic_read(&mm->mm_users) == 0;
416 void __khugepaged_enter(struct mm_struct *mm)
418 struct khugepaged_mm_slot *mm_slot;
419 struct mm_slot *slot;
422 mm_slot = mm_slot_alloc(mm_slot_cache);
426 slot = &mm_slot->slot;
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);
435 spin_lock(&khugepaged_mm_lock);
436 mm_slot_insert(mm_slots_hash, mm, slot);
438 * Insert just behind the scanning cursor, to let the area settle
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);
447 wake_up_interruptible(&khugepaged_wait);
450 void khugepaged_enter_vma(struct vm_area_struct *vma,
451 unsigned long vm_flags)
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);
460 void __khugepaged_exit(struct mm_struct *mm)
462 struct khugepaged_mm_slot *mm_slot;
463 struct mm_slot *slot;
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);
474 spin_unlock(&khugepaged_mm_lock);
477 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
478 mm_slot_free(mm_slot_cache, mm_slot);
480 } else if (mm_slot) {
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.
489 mmap_write_unlock(mm);
493 static void release_pte_page(struct page *page)
495 mod_node_page_state(page_pgdat(page),
496 NR_ISOLATED_ANON + page_is_file_lru(page),
499 putback_lru_page(page);
502 static void release_pte_pages(pte_t *pte, pte_t *_pte,
503 struct list_head *compound_pagelist)
505 struct page *page, *tmp;
507 while (--_pte >= pte) {
508 pte_t pteval = *_pte;
510 page = pte_page(pteval);
511 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
513 release_pte_page(page);
516 list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
517 list_del(&page->lru);
518 release_pte_page(page);
522 static bool is_refcount_suitable(struct page *page)
524 int expected_refcount;
526 expected_refcount = total_mapcount(page);
527 if (PageSwapCache(page))
528 expected_refcount += compound_nr(page);
530 return page_count(page) == expected_refcount;
533 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
534 unsigned long address,
536 struct collapse_control *cc,
537 struct list_head *compound_pagelist)
539 struct page *page = NULL;
541 int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0;
542 bool writable = false;
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)))) {
550 if (!userfaultfd_armed(vma) &&
551 (!cc->is_khugepaged ||
552 none_or_zero <= khugepaged_max_ptes_none)) {
555 result = SCAN_EXCEED_NONE_PTE;
556 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
560 if (!pte_present(pteval)) {
561 result = SCAN_PTE_NON_PRESENT;
564 page = vm_normal_page(vma, address, pteval);
565 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
566 result = SCAN_PAGE_NULL;
570 VM_BUG_ON_PAGE(!PageAnon(page), page);
572 if (page_mapcount(page) > 1) {
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);
582 if (PageCompound(page)) {
584 page = compound_head(page);
587 * Check if we have dealt with the compound page
590 list_for_each_entry(p, compound_pagelist, lru) {
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.
602 if (!trylock_page(page)) {
603 result = SCAN_PAGE_LOCK;
608 * Check if the page has any GUP (or other external) pins.
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.
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.
618 if (!is_refcount_suitable(page)) {
620 result = SCAN_PAGE_COUNT;
625 * Isolate the page to avoid collapsing an hugepage
626 * currently in use by the VM.
628 if (isolate_lru_page(page)) {
630 result = SCAN_DEL_PAGE_LRU;
633 mod_node_page_state(page_pgdat(page),
634 NR_ISOLATED_ANON + page_is_file_lru(page),
636 VM_BUG_ON_PAGE(!PageLocked(page), page);
637 VM_BUG_ON_PAGE(PageLRU(page), page);
639 if (PageCompound(page))
640 list_add_tail(&page->lru, compound_pagelist);
643 * If collapse was initiated by khugepaged, check that there is
644 * enough young pte to justify collapsing the page
646 if (cc->is_khugepaged &&
647 (pte_young(pteval) || page_is_young(page) ||
648 PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm,
652 if (pte_write(pteval))
656 if (unlikely(!writable)) {
657 result = SCAN_PAGE_RO;
658 } else if (unlikely(cc->is_khugepaged && !referenced)) {
659 result = SCAN_LACK_REFERENCED_PAGE;
661 result = SCAN_SUCCEED;
662 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
663 referenced, writable, result);
667 release_pte_pages(pte, _pte, compound_pagelist);
668 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
669 referenced, writable, result);
673 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
674 struct vm_area_struct *vma,
675 unsigned long address,
677 struct list_head *compound_pagelist)
679 struct page *src_page, *tmp;
681 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
682 _pte++, page++, address += PAGE_SIZE) {
683 pte_t pteval = *_pte;
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))) {
690 * ptl mostly unnecessary.
693 ptep_clear(vma->vm_mm, address, _pte);
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);
702 * ptl mostly unnecessary, but preempt has to
703 * be disabled to update the per-cpu stats
704 * inside page_remove_rmap().
707 ptep_clear(vma->vm_mm, address, _pte);
708 page_remove_rmap(src_page, vma, false);
710 free_page_and_swap_cache(src_page);
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);
725 static void khugepaged_alloc_sleep(void)
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);
735 struct collapse_control khugepaged_collapse_control = {
736 .is_khugepaged = true,
739 static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc)
744 * If node_reclaim_mode is disabled, then no extra effort is made to
745 * allocate memory locally.
747 if (!node_reclaim_enabled())
750 /* If there is a count for this node already, it must be acceptable */
751 if (cc->node_load[nid])
754 for (i = 0; i < MAX_NUMNODES; i++) {
755 if (!cc->node_load[i])
757 if (node_distance(nid, i) > node_reclaim_distance)
763 #define khugepaged_defrag() \
764 (transparent_hugepage_flags & \
765 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG))
767 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
768 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
770 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
774 static int hpage_collapse_find_target_node(struct collapse_control *cc)
776 int nid, target_node = 0, max_value = 0;
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];
785 for_each_online_node(nid) {
786 if (max_value == cc->node_load[nid])
787 node_set(nid, cc->alloc_nmask);
793 static int hpage_collapse_find_target_node(struct collapse_control *cc)
799 static bool hpage_collapse_alloc_page(struct page **hpage, gfp_t gfp, int node,
802 *hpage = __alloc_pages(gfp, HPAGE_PMD_ORDER, node, nmask);
803 if (unlikely(!*hpage)) {
804 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
808 prep_transhuge_page(*hpage);
809 count_vm_event(THP_COLLAPSE_ALLOC);
814 * If mmap_lock temporarily dropped, revalidate vma
815 * before taking mmap_lock.
816 * Returns enum scan_result value.
819 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
821 struct vm_area_struct **vmap,
822 struct collapse_control *cc)
824 struct vm_area_struct *vma;
826 if (unlikely(hpage_collapse_test_exit(mm)))
827 return SCAN_ANY_PROCESS;
829 *vmap = vma = find_vma(mm, address);
831 return SCAN_VMA_NULL;
833 if (!transhuge_vma_suitable(vma, address))
834 return SCAN_ADDRESS_RANGE;
835 if (!hugepage_vma_check(vma, vma->vm_flags, false, false,
837 return SCAN_VMA_CHECK;
839 * Anon VMA expected, the address may be unmapped then
840 * remapped to file after khugepaged reaquired the mmap_lock.
842 * hugepage_vma_check may return true for qualified file
845 if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap)))
846 return SCAN_PAGE_ANON;
850 static int find_pmd_or_thp_or_none(struct mm_struct *mm,
851 unsigned long address,
856 *pmd = mm_find_pmd(mm, address);
858 return SCAN_PMD_NULL;
860 pmde = pmdp_get_lockless(*pmd);
862 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
863 /* See comments in pmd_none_or_trans_huge_or_clear_bad() */
867 return SCAN_PMD_NONE;
868 if (pmd_trans_huge(pmde))
869 return SCAN_PMD_MAPPED;
871 return SCAN_PMD_NULL;
875 static int check_pmd_still_valid(struct mm_struct *mm,
876 unsigned long address,
880 int result = find_pmd_or_thp_or_none(mm, address, &new_pmd);
882 if (result != SCAN_SUCCEED)
890 * Bring missing pages in from swap, to complete THP collapse.
891 * Only done if hpage_collapse_scan_pmd believes it is worthwhile.
893 * Called and returns without pte mapped or spinlocks held.
894 * Note that if false is returned, mmap_lock will be released.
897 static int __collapse_huge_page_swapin(struct mm_struct *mm,
898 struct vm_area_struct *vma,
899 unsigned long haddr, pmd_t *pmd,
904 unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
906 for (address = haddr; address < end; address += PAGE_SIZE) {
907 struct vm_fault vmf = {
910 .pgoff = linear_page_index(vma, haddr),
911 .flags = FAULT_FLAG_ALLOW_RETRY,
915 vmf.pte = pte_offset_map(pmd, address);
916 vmf.orig_pte = *vmf.pte;
917 if (!is_swap_pte(vmf.orig_pte)) {
921 ret = do_swap_page(&vmf);
924 * do_swap_page returns VM_FAULT_RETRY with released mmap_lock.
925 * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because
926 * we do not retry here and swap entry will remain in pagetable
927 * resulting in later failure.
929 if (ret & VM_FAULT_RETRY) {
930 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
931 /* Likely, but not guaranteed, that page lock failed */
932 return SCAN_PAGE_LOCK;
934 if (ret & VM_FAULT_ERROR) {
935 mmap_read_unlock(mm);
936 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
942 /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
946 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
950 static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm,
951 struct collapse_control *cc)
953 gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() :
955 int node = hpage_collapse_find_target_node(cc);
957 if (!hpage_collapse_alloc_page(hpage, gfp, node, &cc->alloc_nmask))
958 return SCAN_ALLOC_HUGE_PAGE_FAIL;
959 if (unlikely(mem_cgroup_charge(page_folio(*hpage), mm, gfp)))
960 return SCAN_CGROUP_CHARGE_FAIL;
961 count_memcg_page_event(*hpage, THP_COLLAPSE_ALLOC);
965 static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
966 int referenced, int unmapped,
967 struct collapse_control *cc)
969 LIST_HEAD(compound_pagelist);
974 spinlock_t *pmd_ptl, *pte_ptl;
975 int result = SCAN_FAIL;
976 struct vm_area_struct *vma;
977 struct mmu_notifier_range range;
979 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
982 * Before allocating the hugepage, release the mmap_lock read lock.
983 * The allocation can take potentially a long time if it involves
984 * sync compaction, and we do not need to hold the mmap_lock during
985 * that. We will recheck the vma after taking it again in write mode.
987 mmap_read_unlock(mm);
989 result = alloc_charge_hpage(&hpage, mm, cc);
990 if (result != SCAN_SUCCEED)
994 result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
995 if (result != SCAN_SUCCEED) {
996 mmap_read_unlock(mm);
1000 result = find_pmd_or_thp_or_none(mm, address, &pmd);
1001 if (result != SCAN_SUCCEED) {
1002 mmap_read_unlock(mm);
1008 * __collapse_huge_page_swapin will return with mmap_lock
1009 * released when it fails. So we jump out_nolock directly in
1010 * that case. Continuing to collapse causes inconsistency.
1012 result = __collapse_huge_page_swapin(mm, vma, address, pmd,
1014 if (result != SCAN_SUCCEED)
1018 mmap_read_unlock(mm);
1020 * Prevent all access to pagetables with the exception of
1021 * gup_fast later handled by the ptep_clear_flush and the VM
1022 * handled by the anon_vma lock + PG_lock.
1024 mmap_write_lock(mm);
1025 result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1026 if (result != SCAN_SUCCEED)
1028 /* check if the pmd is still valid */
1029 result = check_pmd_still_valid(mm, address, pmd);
1030 if (result != SCAN_SUCCEED)
1033 anon_vma_lock_write(vma->anon_vma);
1035 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1036 address, address + HPAGE_PMD_SIZE);
1037 mmu_notifier_invalidate_range_start(&range);
1039 pte = pte_offset_map(pmd, address);
1040 pte_ptl = pte_lockptr(mm, pmd);
1042 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1044 * This removes any huge TLB entry from the CPU so we won't allow
1045 * huge and small TLB entries for the same virtual address to
1046 * avoid the risk of CPU bugs in that area.
1048 * Parallel fast GUP is fine since fast GUP will back off when
1049 * it detects PMD is changed.
1051 _pmd = pmdp_collapse_flush(vma, address, pmd);
1052 spin_unlock(pmd_ptl);
1053 mmu_notifier_invalidate_range_end(&range);
1054 tlb_remove_table_sync_one();
1057 result = __collapse_huge_page_isolate(vma, address, pte, cc,
1058 &compound_pagelist);
1059 spin_unlock(pte_ptl);
1061 if (unlikely(result != SCAN_SUCCEED)) {
1064 BUG_ON(!pmd_none(*pmd));
1066 * We can only use set_pmd_at when establishing
1067 * hugepmds and never for establishing regular pmds that
1068 * points to regular pagetables. Use pmd_populate for that
1070 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1071 spin_unlock(pmd_ptl);
1072 anon_vma_unlock_write(vma->anon_vma);
1077 * All pages are isolated and locked so anon_vma rmap
1078 * can't run anymore.
1080 anon_vma_unlock_write(vma->anon_vma);
1082 __collapse_huge_page_copy(pte, hpage, vma, address, pte_ptl,
1083 &compound_pagelist);
1086 * spin_lock() below is not the equivalent of smp_wmb(), but
1087 * the smp_wmb() inside __SetPageUptodate() can be reused to
1088 * avoid the copy_huge_page writes to become visible after
1089 * the set_pmd_at() write.
1091 __SetPageUptodate(hpage);
1092 pgtable = pmd_pgtable(_pmd);
1094 _pmd = mk_huge_pmd(hpage, vma->vm_page_prot);
1095 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1098 BUG_ON(!pmd_none(*pmd));
1099 page_add_new_anon_rmap(hpage, vma, address);
1100 lru_cache_add_inactive_or_unevictable(hpage, vma);
1101 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1102 set_pmd_at(mm, address, pmd, _pmd);
1103 update_mmu_cache_pmd(vma, address, pmd);
1104 spin_unlock(pmd_ptl);
1108 result = SCAN_SUCCEED;
1110 mmap_write_unlock(mm);
1113 mem_cgroup_uncharge(page_folio(hpage));
1116 trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result);
1120 static int hpage_collapse_scan_pmd(struct mm_struct *mm,
1121 struct vm_area_struct *vma,
1122 unsigned long address, bool *mmap_locked,
1123 struct collapse_control *cc)
1127 int result = SCAN_FAIL, referenced = 0;
1128 int none_or_zero = 0, shared = 0;
1129 struct page *page = NULL;
1130 unsigned long _address;
1132 int node = NUMA_NO_NODE, unmapped = 0;
1133 bool writable = false;
1135 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1137 result = find_pmd_or_thp_or_none(mm, address, &pmd);
1138 if (result != SCAN_SUCCEED)
1141 memset(cc->node_load, 0, sizeof(cc->node_load));
1142 nodes_clear(cc->alloc_nmask);
1143 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1144 for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR;
1145 _pte++, _address += PAGE_SIZE) {
1146 pte_t pteval = *_pte;
1147 if (is_swap_pte(pteval)) {
1149 if (!cc->is_khugepaged ||
1150 unmapped <= khugepaged_max_ptes_swap) {
1152 * Always be strict with uffd-wp
1153 * enabled swap entries. Please see
1154 * comment below for pte_uffd_wp().
1156 if (pte_swp_uffd_wp(pteval)) {
1157 result = SCAN_PTE_UFFD_WP;
1162 result = SCAN_EXCEED_SWAP_PTE;
1163 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1167 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1169 if (!userfaultfd_armed(vma) &&
1170 (!cc->is_khugepaged ||
1171 none_or_zero <= khugepaged_max_ptes_none)) {
1174 result = SCAN_EXCEED_NONE_PTE;
1175 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1179 if (pte_uffd_wp(pteval)) {
1181 * Don't collapse the page if any of the small
1182 * PTEs are armed with uffd write protection.
1183 * Here we can also mark the new huge pmd as
1184 * write protected if any of the small ones is
1185 * marked but that could bring unknown
1186 * userfault messages that falls outside of
1187 * the registered range. So, just be simple.
1189 result = SCAN_PTE_UFFD_WP;
1192 if (pte_write(pteval))
1195 page = vm_normal_page(vma, _address, pteval);
1196 if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
1197 result = SCAN_PAGE_NULL;
1201 if (page_mapcount(page) > 1) {
1203 if (cc->is_khugepaged &&
1204 shared > khugepaged_max_ptes_shared) {
1205 result = SCAN_EXCEED_SHARED_PTE;
1206 count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1211 page = compound_head(page);
1214 * Record which node the original page is from and save this
1215 * information to cc->node_load[].
1216 * Khugepaged will allocate hugepage from the node has the max
1219 node = page_to_nid(page);
1220 if (hpage_collapse_scan_abort(node, cc)) {
1221 result = SCAN_SCAN_ABORT;
1224 cc->node_load[node]++;
1225 if (!PageLRU(page)) {
1226 result = SCAN_PAGE_LRU;
1229 if (PageLocked(page)) {
1230 result = SCAN_PAGE_LOCK;
1233 if (!PageAnon(page)) {
1234 result = SCAN_PAGE_ANON;
1239 * Check if the page has any GUP (or other external) pins.
1241 * Here the check may be racy:
1242 * it may see total_mapcount > refcount in some cases?
1243 * But such case is ephemeral we could always retry collapse
1244 * later. However it may report false positive if the page
1245 * has excessive GUP pins (i.e. 512). Anyway the same check
1246 * will be done again later the risk seems low.
1248 if (!is_refcount_suitable(page)) {
1249 result = SCAN_PAGE_COUNT;
1254 * If collapse was initiated by khugepaged, check that there is
1255 * enough young pte to justify collapsing the page
1257 if (cc->is_khugepaged &&
1258 (pte_young(pteval) || page_is_young(page) ||
1259 PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm,
1264 result = SCAN_PAGE_RO;
1265 } else if (cc->is_khugepaged &&
1267 (unmapped && referenced < HPAGE_PMD_NR / 2))) {
1268 result = SCAN_LACK_REFERENCED_PAGE;
1270 result = SCAN_SUCCEED;
1273 pte_unmap_unlock(pte, ptl);
1274 if (result == SCAN_SUCCEED) {
1275 result = collapse_huge_page(mm, address, referenced,
1277 /* collapse_huge_page will return with the mmap_lock released */
1278 *mmap_locked = false;
1281 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1282 none_or_zero, result, unmapped);
1286 static void collect_mm_slot(struct khugepaged_mm_slot *mm_slot)
1288 struct mm_slot *slot = &mm_slot->slot;
1289 struct mm_struct *mm = slot->mm;
1291 lockdep_assert_held(&khugepaged_mm_lock);
1293 if (hpage_collapse_test_exit(mm)) {
1295 hash_del(&slot->hash);
1296 list_del(&slot->mm_node);
1299 * Not strictly needed because the mm exited already.
1301 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1304 /* khugepaged_mm_lock actually not necessary for the below */
1305 mm_slot_free(mm_slot_cache, mm_slot);
1312 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1313 * khugepaged should try to collapse the page table.
1315 * Note that following race exists:
1316 * (1) khugepaged calls khugepaged_collapse_pte_mapped_thps() for mm_struct A,
1317 * emptying the A's ->pte_mapped_thp[] array.
1318 * (2) MADV_COLLAPSE collapses some file extent with target mm_struct B, and
1319 * retract_page_tables() finds a VMA in mm_struct A mapping the same extent
1320 * (at virtual address X) and adds an entry (for X) into mm_struct A's
1321 * ->pte-mapped_thp[] array.
1322 * (3) khugepaged calls khugepaged_collapse_scan_file() for mm_struct A at X,
1323 * sees a pte-mapped THP (SCAN_PTE_MAPPED_HUGEPAGE) and adds an entry
1324 * (for X) into mm_struct A's ->pte-mapped_thp[] array.
1325 * Thus, it's possible the same address is added multiple times for the same
1326 * mm_struct. Should this happen, we'll simply attempt
1327 * collapse_pte_mapped_thp() multiple times for the same address, under the same
1328 * exclusive mmap_lock, and assuming the first call is successful, subsequent
1329 * attempts will return quickly (without grabbing any additional locks) when
1330 * a huge pmd is found in find_pmd_or_thp_or_none(). Since this is a cheap
1331 * check, and since this is a rare occurrence, the cost of preventing this
1332 * "multiple-add" is thought to be more expensive than just handling it, should
1335 static bool khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1338 struct khugepaged_mm_slot *mm_slot;
1339 struct mm_slot *slot;
1342 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1344 spin_lock(&khugepaged_mm_lock);
1345 slot = mm_slot_lookup(mm_slots_hash, mm);
1346 mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
1347 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP)) {
1348 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1351 spin_unlock(&khugepaged_mm_lock);
1355 /* hpage must be locked, and mmap_lock must be held in write */
1356 static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
1357 pmd_t *pmdp, struct page *hpage)
1359 struct vm_fault vmf = {
1366 VM_BUG_ON(!PageTransHuge(hpage));
1367 mmap_assert_write_locked(vma->vm_mm);
1369 if (do_set_pmd(&vmf, hpage))
1373 return SCAN_SUCCEED;
1377 * A note about locking:
1378 * Trying to take the page table spinlocks would be useless here because those
1379 * are only used to synchronize:
1381 * - modifying terminal entries (ones that point to a data page, not to another
1383 * - installing *new* non-terminal entries
1385 * Instead, we need roughly the same kind of protection as free_pgtables() or
1386 * mm_take_all_locks() (but only for a single VMA):
1387 * The mmap lock together with this VMA's rmap locks covers all paths towards
1388 * the page table entries we're messing with here, except for hardware page
1389 * table walks and lockless_pages_from_mm().
1391 static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
1392 unsigned long addr, pmd_t *pmdp)
1395 struct mmu_notifier_range range;
1397 mmap_assert_write_locked(mm);
1399 lockdep_assert_held_write(&vma->vm_file->f_mapping->i_mmap_rwsem);
1401 * All anon_vmas attached to the VMA have the same root and are
1402 * therefore locked by the same lock.
1405 lockdep_assert_held_write(&vma->anon_vma->root->rwsem);
1407 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, addr,
1408 addr + HPAGE_PMD_SIZE);
1409 mmu_notifier_invalidate_range_start(&range);
1410 pmd = pmdp_collapse_flush(vma, addr, pmdp);
1411 tlb_remove_table_sync_one();
1412 mmu_notifier_invalidate_range_end(&range);
1414 page_table_check_pte_clear_range(mm, addr, pmd);
1415 pte_free(mm, pmd_pgtable(pmd));
1419 * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1422 * @mm: process address space where collapse happens
1423 * @addr: THP collapse address
1424 * @install_pmd: If a huge PMD should be installed
1426 * This function checks whether all the PTEs in the PMD are pointing to the
1427 * right THP. If so, retract the page table so the THP can refault in with
1428 * as pmd-mapped. Possibly install a huge PMD mapping the THP.
1430 int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
1433 unsigned long haddr = addr & HPAGE_PMD_MASK;
1434 struct vm_area_struct *vma = vma_lookup(mm, haddr);
1436 pte_t *start_pte, *pte;
1439 int count = 0, result = SCAN_FAIL;
1442 mmap_assert_write_locked(mm);
1444 /* Fast check before locking page if already PMD-mapped */
1445 result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
1446 if (result == SCAN_PMD_MAPPED)
1449 if (!vma || !vma->vm_file ||
1450 !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1451 return SCAN_VMA_CHECK;
1454 * If we are here, we've succeeded in replacing all the native pages
1455 * in the page cache with a single hugepage. If a mm were to fault-in
1456 * this memory (mapped by a suitably aligned VMA), we'd get the hugepage
1457 * and map it by a PMD, regardless of sysfs THP settings. As such, let's
1458 * analogously elide sysfs THP settings here.
1460 if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false))
1461 return SCAN_VMA_CHECK;
1463 /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */
1464 if (userfaultfd_wp(vma))
1465 return SCAN_PTE_UFFD_WP;
1467 hpage = find_lock_page(vma->vm_file->f_mapping,
1468 linear_page_index(vma, haddr));
1470 return SCAN_PAGE_NULL;
1472 if (!PageHead(hpage)) {
1477 if (compound_order(hpage) != HPAGE_PMD_ORDER) {
1478 result = SCAN_PAGE_COMPOUND;
1487 * In MADV_COLLAPSE path, possible race with khugepaged where
1488 * all pte entries have been removed and pmd cleared. If so,
1489 * skip all the pte checks and just update the pmd mapping.
1491 goto maybe_install_pmd;
1497 * We need to lock the mapping so that from here on, only GUP-fast and
1498 * hardware page walks can access the parts of the page tables that
1499 * we're operating on.
1500 * See collapse_and_free_pmd().
1502 i_mmap_lock_write(vma->vm_file->f_mapping);
1505 * This spinlock should be unnecessary: Nobody else should be accessing
1506 * the page tables under spinlock protection here, only
1507 * lockless_pages_from_mm() and the hardware page walker can access page
1508 * tables while all the high-level locks are held in write mode.
1510 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1513 /* step 1: check all mapped PTEs are to the right huge page */
1514 for (i = 0, addr = haddr, pte = start_pte;
1515 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1518 /* empty pte, skip */
1522 /* page swapped out, abort */
1523 if (!pte_present(*pte)) {
1524 result = SCAN_PTE_NON_PRESENT;
1528 page = vm_normal_page(vma, addr, *pte);
1529 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1532 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1533 * page table, but the new page will not be a subpage of hpage.
1535 if (hpage + i != page)
1540 /* step 2: adjust rmap */
1541 for (i = 0, addr = haddr, pte = start_pte;
1542 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1547 page = vm_normal_page(vma, addr, *pte);
1548 if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1550 page_remove_rmap(page, vma, false);
1553 pte_unmap_unlock(start_pte, ptl);
1555 /* step 3: set proper refcount and mm_counters. */
1557 page_ref_sub(hpage, count);
1558 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1561 /* step 4: remove pte entries */
1562 /* we make no change to anon, but protect concurrent anon page lookup */
1564 anon_vma_lock_write(vma->anon_vma);
1566 collapse_and_free_pmd(mm, vma, haddr, pmd);
1569 anon_vma_unlock_write(vma->anon_vma);
1570 i_mmap_unlock_write(vma->vm_file->f_mapping);
1573 /* step 5: install pmd entry */
1574 result = install_pmd
1575 ? set_huge_pmd(vma, haddr, pmd, hpage)
1584 pte_unmap_unlock(start_pte, ptl);
1585 i_mmap_unlock_write(vma->vm_file->f_mapping);
1589 static void khugepaged_collapse_pte_mapped_thps(struct khugepaged_mm_slot *mm_slot)
1591 struct mm_slot *slot = &mm_slot->slot;
1592 struct mm_struct *mm = slot->mm;
1595 if (likely(mm_slot->nr_pte_mapped_thp == 0))
1598 if (!mmap_write_trylock(mm))
1601 if (unlikely(hpage_collapse_test_exit(mm)))
1604 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1605 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i], false);
1608 mm_slot->nr_pte_mapped_thp = 0;
1609 mmap_write_unlock(mm);
1612 static int retract_page_tables(struct address_space *mapping, pgoff_t pgoff,
1613 struct mm_struct *target_mm,
1614 unsigned long target_addr, struct page *hpage,
1615 struct collapse_control *cc)
1617 struct vm_area_struct *vma;
1618 int target_result = SCAN_FAIL;
1620 i_mmap_lock_write(mapping);
1621 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1622 int result = SCAN_FAIL;
1623 struct mm_struct *mm = NULL;
1624 unsigned long addr = 0;
1626 bool is_target = false;
1629 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1630 * got written to. These VMAs are likely not worth investing
1631 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1634 * Note that vma->anon_vma check is racy: it can be set up after
1635 * the check but before we took mmap_lock by the fault path.
1636 * But page lock would prevent establishing any new ptes of the
1637 * page, so we are safe.
1639 * An alternative would be drop the check, but check that page
1640 * table is clear before calling pmdp_collapse_flush() under
1641 * ptl. It has higher chance to recover THP for the VMA, but
1642 * has higher cost too. It would also probably require locking
1645 if (vma->anon_vma) {
1646 result = SCAN_PAGE_ANON;
1649 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1650 if (addr & ~HPAGE_PMD_MASK ||
1651 vma->vm_end < addr + HPAGE_PMD_SIZE) {
1652 result = SCAN_VMA_CHECK;
1656 is_target = mm == target_mm && addr == target_addr;
1657 result = find_pmd_or_thp_or_none(mm, addr, &pmd);
1658 if (result != SCAN_SUCCEED)
1661 * We need exclusive mmap_lock to retract page table.
1663 * We use trylock due to lock inversion: we need to acquire
1664 * mmap_lock while holding page lock. Fault path does it in
1665 * reverse order. Trylock is a way to avoid deadlock.
1667 * Also, it's not MADV_COLLAPSE's job to collapse other
1668 * mappings - let khugepaged take care of them later.
1670 result = SCAN_PTE_MAPPED_HUGEPAGE;
1671 if ((cc->is_khugepaged || is_target) &&
1672 mmap_write_trylock(mm)) {
1674 * When a vma is registered with uffd-wp, we can't
1675 * recycle the pmd pgtable because there can be pte
1676 * markers installed. Skip it only, so the rest mm/vma
1677 * can still have the same file mapped hugely, however
1678 * it'll always mapped in small page size for uffd-wp
1679 * registered ranges.
1681 if (hpage_collapse_test_exit(mm)) {
1682 result = SCAN_ANY_PROCESS;
1685 if (userfaultfd_wp(vma)) {
1686 result = SCAN_PTE_UFFD_WP;
1689 collapse_and_free_pmd(mm, vma, addr, pmd);
1690 if (!cc->is_khugepaged && is_target)
1691 result = set_huge_pmd(vma, addr, pmd, hpage);
1693 result = SCAN_SUCCEED;
1696 mmap_write_unlock(mm);
1700 * Calling context will handle target mm/addr. Otherwise, let
1701 * khugepaged try again later.
1704 khugepaged_add_pte_mapped_thp(mm, addr);
1709 target_result = result;
1711 i_mmap_unlock_write(mapping);
1712 return target_result;
1716 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1718 * @mm: process address space where collapse happens
1719 * @addr: virtual collapse start address
1720 * @file: file that collapse on
1721 * @start: collapse start address
1722 * @cc: collapse context and scratchpad
1724 * Basic scheme is simple, details are more complex:
1725 * - allocate and lock a new huge page;
1726 * - scan page cache replacing old pages with the new one
1727 * + swap/gup in pages if necessary;
1729 * + keep old pages around in case rollback is required;
1730 * - if replacing succeeds:
1733 * + unlock huge page;
1734 * - if replacing failed;
1735 * + put all pages back and unfreeze them;
1736 * + restore gaps in the page cache;
1737 * + unlock and free huge page;
1739 static int collapse_file(struct mm_struct *mm, unsigned long addr,
1740 struct file *file, pgoff_t start,
1741 struct collapse_control *cc)
1743 struct address_space *mapping = file->f_mapping;
1745 pgoff_t index = 0, end = start + HPAGE_PMD_NR;
1746 LIST_HEAD(pagelist);
1747 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1748 int nr_none = 0, result = SCAN_SUCCEED;
1749 bool is_shmem = shmem_file(file);
1752 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1753 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1755 result = alloc_charge_hpage(&hpage, mm, cc);
1756 if (result != SCAN_SUCCEED)
1760 * Ensure we have slots for all the pages in the range. This is
1761 * almost certainly a no-op because most of the pages must be present
1765 xas_create_range(&xas);
1766 if (!xas_error(&xas))
1768 xas_unlock_irq(&xas);
1769 if (!xas_nomem(&xas, GFP_KERNEL)) {
1775 __SetPageLocked(hpage);
1777 __SetPageSwapBacked(hpage);
1778 hpage->index = start;
1779 hpage->mapping = mapping;
1782 * At this point the hpage is locked and not up-to-date.
1783 * It's safe to insert it into the page cache, because nobody would
1784 * be able to map it or use it in another way until we unlock it.
1787 xas_set(&xas, start);
1788 for (index = start; index < end; index++) {
1789 struct page *page = xas_next(&xas);
1790 struct folio *folio;
1792 VM_BUG_ON(index != xas.xa_index);
1796 * Stop if extent has been truncated or
1797 * hole-punched, and is now completely
1800 if (index == start) {
1801 if (!xas_next_entry(&xas, end - 1)) {
1802 result = SCAN_TRUNCATED;
1805 xas_set(&xas, index);
1807 if (!shmem_charge(mapping->host, 1)) {
1811 xas_store(&xas, hpage);
1816 if (xa_is_value(page) || !PageUptodate(page)) {
1817 xas_unlock_irq(&xas);
1818 /* swap in or instantiate fallocated page */
1819 if (shmem_get_folio(mapping->host, index,
1820 &folio, SGP_NOALLOC)) {
1824 page = folio_file_page(folio, index);
1825 } else if (trylock_page(page)) {
1827 xas_unlock_irq(&xas);
1829 result = SCAN_PAGE_LOCK;
1832 } else { /* !is_shmem */
1833 if (!page || xa_is_value(page)) {
1834 xas_unlock_irq(&xas);
1835 page_cache_sync_readahead(mapping, &file->f_ra,
1838 /* drain pagevecs to help isolate_lru_page() */
1840 page = find_lock_page(mapping, index);
1841 if (unlikely(page == NULL)) {
1845 } else if (PageDirty(page)) {
1847 * khugepaged only works on read-only fd,
1848 * so this page is dirty because it hasn't
1849 * been flushed since first write. There
1850 * won't be new dirty pages.
1852 * Trigger async flush here and hope the
1853 * writeback is done when khugepaged
1854 * revisits this page.
1856 * This is a one-off situation. We are not
1857 * forcing writeback in loop.
1859 xas_unlock_irq(&xas);
1860 filemap_flush(mapping);
1863 } else if (PageWriteback(page)) {
1864 xas_unlock_irq(&xas);
1867 } else if (trylock_page(page)) {
1869 xas_unlock_irq(&xas);
1871 result = SCAN_PAGE_LOCK;
1877 * The page must be locked, so we can drop the i_pages lock
1878 * without racing with truncate.
1880 VM_BUG_ON_PAGE(!PageLocked(page), page);
1882 /* make sure the page is up to date */
1883 if (unlikely(!PageUptodate(page))) {
1889 * If file was truncated then extended, or hole-punched, before
1890 * we locked the first page, then a THP might be there already.
1891 * This will be discovered on the first iteration.
1893 if (PageTransCompound(page)) {
1894 struct page *head = compound_head(page);
1896 result = compound_order(head) == HPAGE_PMD_ORDER &&
1897 head->index == start
1898 /* Maybe PMD-mapped */
1899 ? SCAN_PTE_MAPPED_HUGEPAGE
1900 : SCAN_PAGE_COMPOUND;
1904 folio = page_folio(page);
1906 if (folio_mapping(folio) != mapping) {
1907 result = SCAN_TRUNCATED;
1911 if (!is_shmem && (folio_test_dirty(folio) ||
1912 folio_test_writeback(folio))) {
1914 * khugepaged only works on read-only fd, so this
1915 * page is dirty because it hasn't been flushed
1916 * since first write.
1922 if (folio_isolate_lru(folio)) {
1923 result = SCAN_DEL_PAGE_LRU;
1927 if (folio_has_private(folio) &&
1928 !filemap_release_folio(folio, GFP_KERNEL)) {
1929 result = SCAN_PAGE_HAS_PRIVATE;
1930 folio_putback_lru(folio);
1934 if (folio_mapped(folio))
1936 TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH);
1939 xas_set(&xas, index);
1941 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1944 * The page is expected to have page_count() == 3:
1945 * - we hold a pin on it;
1946 * - one reference from page cache;
1947 * - one from isolate_lru_page;
1949 if (!page_ref_freeze(page, 3)) {
1950 result = SCAN_PAGE_COUNT;
1951 xas_unlock_irq(&xas);
1952 putback_lru_page(page);
1957 * Add the page to the list to be able to undo the collapse if
1958 * something go wrong.
1960 list_add_tail(&page->lru, &pagelist);
1962 /* Finally, replace with the new page. */
1963 xas_store(&xas, hpage);
1970 nr = thp_nr_pages(hpage);
1973 __mod_lruvec_page_state(hpage, NR_SHMEM_THPS, nr);
1975 __mod_lruvec_page_state(hpage, NR_FILE_THPS, nr);
1976 filemap_nr_thps_inc(mapping);
1978 * Paired with smp_mb() in do_dentry_open() to ensure
1979 * i_writecount is up to date and the update to nr_thps is
1980 * visible. Ensures the page cache will be truncated if the
1981 * file is opened writable.
1984 if (inode_is_open_for_write(mapping->host)) {
1986 __mod_lruvec_page_state(hpage, NR_FILE_THPS, -nr);
1987 filemap_nr_thps_dec(mapping);
1993 __mod_lruvec_page_state(hpage, NR_FILE_PAGES, nr_none);
1994 /* nr_none is always 0 for non-shmem. */
1995 __mod_lruvec_page_state(hpage, NR_SHMEM, nr_none);
1998 /* Join all the small entries into a single multi-index entry */
1999 xas_set_order(&xas, start, HPAGE_PMD_ORDER);
2000 xas_store(&xas, hpage);
2002 xas_unlock_irq(&xas);
2006 * If collapse is successful, flush must be done now before copying.
2007 * If collapse is unsuccessful, does flush actually need to be done?
2008 * Do it anyway, to clear the state.
2010 try_to_unmap_flush();
2012 if (result == SCAN_SUCCEED) {
2013 struct page *page, *tmp;
2014 struct folio *folio;
2017 * Replacing old pages with new one has succeeded, now we
2018 * need to copy the content and free the old pages.
2021 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
2022 while (index < page->index) {
2023 clear_highpage(hpage + (index % HPAGE_PMD_NR));
2026 copy_highpage(hpage + (page->index % HPAGE_PMD_NR),
2028 list_del(&page->lru);
2029 page->mapping = NULL;
2030 page_ref_unfreeze(page, 1);
2031 ClearPageActive(page);
2032 ClearPageUnevictable(page);
2037 while (index < end) {
2038 clear_highpage(hpage + (index % HPAGE_PMD_NR));
2042 folio = page_folio(hpage);
2043 folio_mark_uptodate(folio);
2044 folio_ref_add(folio, HPAGE_PMD_NR - 1);
2047 folio_mark_dirty(folio);
2048 folio_add_lru(folio);
2051 * Remove pte page tables, so we can re-fault the page as huge.
2053 result = retract_page_tables(mapping, start, mm, addr, hpage,
2060 /* Something went wrong: roll back page cache changes */
2063 mapping->nrpages -= nr_none;
2064 shmem_uncharge(mapping->host, nr_none);
2067 xas_set(&xas, start);
2068 xas_for_each(&xas, page, end - 1) {
2069 page = list_first_entry_or_null(&pagelist,
2071 if (!page || xas.xa_index < page->index) {
2075 /* Put holes back where they were */
2076 xas_store(&xas, NULL);
2080 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
2082 /* Unfreeze the page. */
2083 list_del(&page->lru);
2084 page_ref_unfreeze(page, 2);
2085 xas_store(&xas, page);
2087 xas_unlock_irq(&xas);
2089 putback_lru_page(page);
2093 xas_unlock_irq(&xas);
2095 hpage->mapping = NULL;
2101 VM_BUG_ON(!list_empty(&pagelist));
2103 mem_cgroup_uncharge(page_folio(hpage));
2107 trace_mm_khugepaged_collapse_file(mm, hpage, index, is_shmem, addr, file, nr, result);
2111 static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2112 struct file *file, pgoff_t start,
2113 struct collapse_control *cc)
2115 struct page *page = NULL;
2116 struct address_space *mapping = file->f_mapping;
2117 XA_STATE(xas, &mapping->i_pages, start);
2119 int node = NUMA_NO_NODE;
2120 int result = SCAN_SUCCEED;
2124 memset(cc->node_load, 0, sizeof(cc->node_load));
2125 nodes_clear(cc->alloc_nmask);
2127 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2128 if (xas_retry(&xas, page))
2131 if (xa_is_value(page)) {
2133 if (cc->is_khugepaged &&
2134 swap > khugepaged_max_ptes_swap) {
2135 result = SCAN_EXCEED_SWAP_PTE;
2136 count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
2143 * TODO: khugepaged should compact smaller compound pages
2144 * into a PMD sized page
2146 if (PageTransCompound(page)) {
2147 struct page *head = compound_head(page);
2149 result = compound_order(head) == HPAGE_PMD_ORDER &&
2150 head->index == start
2151 /* Maybe PMD-mapped */
2152 ? SCAN_PTE_MAPPED_HUGEPAGE
2153 : SCAN_PAGE_COMPOUND;
2155 * For SCAN_PTE_MAPPED_HUGEPAGE, further processing
2156 * by the caller won't touch the page cache, and so
2157 * it's safe to skip LRU and refcount checks before
2163 node = page_to_nid(page);
2164 if (hpage_collapse_scan_abort(node, cc)) {
2165 result = SCAN_SCAN_ABORT;
2168 cc->node_load[node]++;
2170 if (!PageLRU(page)) {
2171 result = SCAN_PAGE_LRU;
2175 if (page_count(page) !=
2176 1 + page_mapcount(page) + page_has_private(page)) {
2177 result = SCAN_PAGE_COUNT;
2182 * We probably should check if the page is referenced here, but
2183 * nobody would transfer pte_young() to PageReferenced() for us.
2184 * And rmap walk here is just too costly...
2189 if (need_resched()) {
2196 if (result == SCAN_SUCCEED) {
2197 if (cc->is_khugepaged &&
2198 present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2199 result = SCAN_EXCEED_NONE_PTE;
2200 count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2202 result = collapse_file(mm, addr, file, start, cc);
2206 trace_mm_khugepaged_scan_file(mm, page, file, present, swap, result);
2210 static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2211 struct file *file, pgoff_t start,
2212 struct collapse_control *cc)
2217 static void khugepaged_collapse_pte_mapped_thps(struct khugepaged_mm_slot *mm_slot)
2221 static bool khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
2228 static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result,
2229 struct collapse_control *cc)
2230 __releases(&khugepaged_mm_lock)
2231 __acquires(&khugepaged_mm_lock)
2233 struct vma_iterator vmi;
2234 struct khugepaged_mm_slot *mm_slot;
2235 struct mm_slot *slot;
2236 struct mm_struct *mm;
2237 struct vm_area_struct *vma;
2241 lockdep_assert_held(&khugepaged_mm_lock);
2242 *result = SCAN_FAIL;
2244 if (khugepaged_scan.mm_slot) {
2245 mm_slot = khugepaged_scan.mm_slot;
2246 slot = &mm_slot->slot;
2248 slot = list_entry(khugepaged_scan.mm_head.next,
2249 struct mm_slot, mm_node);
2250 mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2251 khugepaged_scan.address = 0;
2252 khugepaged_scan.mm_slot = mm_slot;
2254 spin_unlock(&khugepaged_mm_lock);
2255 khugepaged_collapse_pte_mapped_thps(mm_slot);
2259 * Don't wait for semaphore (to avoid long wait times). Just move to
2260 * the next mm on the list.
2263 if (unlikely(!mmap_read_trylock(mm)))
2264 goto breakouterloop_mmap_lock;
2267 if (unlikely(hpage_collapse_test_exit(mm)))
2268 goto breakouterloop;
2270 vma_iter_init(&vmi, mm, khugepaged_scan.address);
2271 for_each_vma(vmi, vma) {
2272 unsigned long hstart, hend;
2275 if (unlikely(hpage_collapse_test_exit(mm))) {
2279 if (!hugepage_vma_check(vma, vma->vm_flags, false, false, true)) {
2284 hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE);
2285 hend = round_down(vma->vm_end, HPAGE_PMD_SIZE);
2286 if (khugepaged_scan.address > hend)
2288 if (khugepaged_scan.address < hstart)
2289 khugepaged_scan.address = hstart;
2290 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2292 while (khugepaged_scan.address < hend) {
2293 bool mmap_locked = true;
2296 if (unlikely(hpage_collapse_test_exit(mm)))
2297 goto breakouterloop;
2299 VM_BUG_ON(khugepaged_scan.address < hstart ||
2300 khugepaged_scan.address + HPAGE_PMD_SIZE >
2302 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2303 struct file *file = get_file(vma->vm_file);
2304 pgoff_t pgoff = linear_page_index(vma,
2305 khugepaged_scan.address);
2307 mmap_read_unlock(mm);
2308 *result = hpage_collapse_scan_file(mm,
2309 khugepaged_scan.address,
2311 mmap_locked = false;
2314 *result = hpage_collapse_scan_pmd(mm, vma,
2315 khugepaged_scan.address,
2320 case SCAN_PTE_MAPPED_HUGEPAGE: {
2323 *result = find_pmd_or_thp_or_none(mm,
2324 khugepaged_scan.address,
2326 if (*result != SCAN_SUCCEED)
2328 if (!khugepaged_add_pte_mapped_thp(mm,
2329 khugepaged_scan.address))
2333 ++khugepaged_pages_collapsed;
2339 /* move to next address */
2340 khugepaged_scan.address += HPAGE_PMD_SIZE;
2341 progress += HPAGE_PMD_NR;
2344 * We released mmap_lock so break loop. Note
2345 * that we drop mmap_lock before all hugepage
2346 * allocations, so if allocation fails, we are
2347 * guaranteed to break here and report the
2348 * correct result back to caller.
2350 goto breakouterloop_mmap_lock;
2351 if (progress >= pages)
2352 goto breakouterloop;
2356 mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2357 breakouterloop_mmap_lock:
2359 spin_lock(&khugepaged_mm_lock);
2360 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2362 * Release the current mm_slot if this mm is about to die, or
2363 * if we scanned all vmas of this mm.
2365 if (hpage_collapse_test_exit(mm) || !vma) {
2367 * Make sure that if mm_users is reaching zero while
2368 * khugepaged runs here, khugepaged_exit will find
2369 * mm_slot not pointing to the exiting mm.
2371 if (slot->mm_node.next != &khugepaged_scan.mm_head) {
2372 slot = list_entry(slot->mm_node.next,
2373 struct mm_slot, mm_node);
2374 khugepaged_scan.mm_slot =
2375 mm_slot_entry(slot, struct khugepaged_mm_slot, slot);
2376 khugepaged_scan.address = 0;
2378 khugepaged_scan.mm_slot = NULL;
2379 khugepaged_full_scans++;
2382 collect_mm_slot(mm_slot);
2388 static int khugepaged_has_work(void)
2390 return !list_empty(&khugepaged_scan.mm_head) &&
2391 hugepage_flags_enabled();
2394 static int khugepaged_wait_event(void)
2396 return !list_empty(&khugepaged_scan.mm_head) ||
2397 kthread_should_stop();
2400 static void khugepaged_do_scan(struct collapse_control *cc)
2402 unsigned int progress = 0, pass_through_head = 0;
2403 unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2405 int result = SCAN_SUCCEED;
2407 lru_add_drain_all();
2412 if (unlikely(kthread_should_stop() || try_to_freeze()))
2415 spin_lock(&khugepaged_mm_lock);
2416 if (!khugepaged_scan.mm_slot)
2417 pass_through_head++;
2418 if (khugepaged_has_work() &&
2419 pass_through_head < 2)
2420 progress += khugepaged_scan_mm_slot(pages - progress,
2424 spin_unlock(&khugepaged_mm_lock);
2426 if (progress >= pages)
2429 if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) {
2431 * If fail to allocate the first time, try to sleep for
2432 * a while. When hit again, cancel the scan.
2437 khugepaged_alloc_sleep();
2442 static bool khugepaged_should_wakeup(void)
2444 return kthread_should_stop() ||
2445 time_after_eq(jiffies, khugepaged_sleep_expire);
2448 static void khugepaged_wait_work(void)
2450 if (khugepaged_has_work()) {
2451 const unsigned long scan_sleep_jiffies =
2452 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2454 if (!scan_sleep_jiffies)
2457 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2458 wait_event_freezable_timeout(khugepaged_wait,
2459 khugepaged_should_wakeup(),
2460 scan_sleep_jiffies);
2464 if (hugepage_flags_enabled())
2465 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2468 static int khugepaged(void *none)
2470 struct khugepaged_mm_slot *mm_slot;
2473 set_user_nice(current, MAX_NICE);
2475 while (!kthread_should_stop()) {
2476 khugepaged_do_scan(&khugepaged_collapse_control);
2477 khugepaged_wait_work();
2480 spin_lock(&khugepaged_mm_lock);
2481 mm_slot = khugepaged_scan.mm_slot;
2482 khugepaged_scan.mm_slot = NULL;
2484 collect_mm_slot(mm_slot);
2485 spin_unlock(&khugepaged_mm_lock);
2489 static void set_recommended_min_free_kbytes(void)
2493 unsigned long recommended_min;
2495 if (!hugepage_flags_enabled()) {
2496 calculate_min_free_kbytes();
2500 for_each_populated_zone(zone) {
2502 * We don't need to worry about fragmentation of
2503 * ZONE_MOVABLE since it only has movable pages.
2505 if (zone_idx(zone) > gfp_zone(GFP_USER))
2511 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2512 recommended_min = pageblock_nr_pages * nr_zones * 2;
2515 * Make sure that on average at least two pageblocks are almost free
2516 * of another type, one for a migratetype to fall back to and a
2517 * second to avoid subsequent fallbacks of other types There are 3
2518 * MIGRATE_TYPES we care about.
2520 recommended_min += pageblock_nr_pages * nr_zones *
2521 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2523 /* don't ever allow to reserve more than 5% of the lowmem */
2524 recommended_min = min(recommended_min,
2525 (unsigned long) nr_free_buffer_pages() / 20);
2526 recommended_min <<= (PAGE_SHIFT-10);
2528 if (recommended_min > min_free_kbytes) {
2529 if (user_min_free_kbytes >= 0)
2530 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2531 min_free_kbytes, recommended_min);
2533 min_free_kbytes = recommended_min;
2537 setup_per_zone_wmarks();
2540 int start_stop_khugepaged(void)
2544 mutex_lock(&khugepaged_mutex);
2545 if (hugepage_flags_enabled()) {
2546 if (!khugepaged_thread)
2547 khugepaged_thread = kthread_run(khugepaged, NULL,
2549 if (IS_ERR(khugepaged_thread)) {
2550 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2551 err = PTR_ERR(khugepaged_thread);
2552 khugepaged_thread = NULL;
2556 if (!list_empty(&khugepaged_scan.mm_head))
2557 wake_up_interruptible(&khugepaged_wait);
2558 } else if (khugepaged_thread) {
2559 kthread_stop(khugepaged_thread);
2560 khugepaged_thread = NULL;
2562 set_recommended_min_free_kbytes();
2564 mutex_unlock(&khugepaged_mutex);
2568 void khugepaged_min_free_kbytes_update(void)
2570 mutex_lock(&khugepaged_mutex);
2571 if (hugepage_flags_enabled() && khugepaged_thread)
2572 set_recommended_min_free_kbytes();
2573 mutex_unlock(&khugepaged_mutex);
2576 bool current_is_khugepaged(void)
2578 return kthread_func(current) == khugepaged;
2581 static int madvise_collapse_errno(enum scan_result r)
2584 * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide
2585 * actionable feedback to caller, so they may take an appropriate
2586 * fallback measure depending on the nature of the failure.
2589 case SCAN_ALLOC_HUGE_PAGE_FAIL:
2591 case SCAN_CGROUP_CHARGE_FAIL:
2593 /* Resource temporary unavailable - trying again might succeed */
2594 case SCAN_PAGE_LOCK:
2596 case SCAN_DEL_PAGE_LRU:
2599 * Other: Trying again likely not to succeed / error intrinsic to
2600 * specified memory range. khugepaged likely won't be able to collapse
2608 int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev,
2609 unsigned long start, unsigned long end)
2611 struct collapse_control *cc;
2612 struct mm_struct *mm = vma->vm_mm;
2613 unsigned long hstart, hend, addr;
2614 int thps = 0, last_fail = SCAN_FAIL;
2615 bool mmap_locked = true;
2617 BUG_ON(vma->vm_start > start);
2618 BUG_ON(vma->vm_end < end);
2622 if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false))
2625 cc = kmalloc(sizeof(*cc), GFP_KERNEL);
2628 cc->is_khugepaged = false;
2631 lru_add_drain_all();
2633 hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2634 hend = end & HPAGE_PMD_MASK;
2636 for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) {
2637 int result = SCAN_FAIL;
2643 result = hugepage_vma_revalidate(mm, addr, false, &vma,
2645 if (result != SCAN_SUCCEED) {
2650 hend = min(hend, vma->vm_end & HPAGE_PMD_MASK);
2652 mmap_assert_locked(mm);
2653 memset(cc->node_load, 0, sizeof(cc->node_load));
2654 nodes_clear(cc->alloc_nmask);
2655 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2656 struct file *file = get_file(vma->vm_file);
2657 pgoff_t pgoff = linear_page_index(vma, addr);
2659 mmap_read_unlock(mm);
2660 mmap_locked = false;
2661 result = hpage_collapse_scan_file(mm, addr, file, pgoff,
2665 result = hpage_collapse_scan_pmd(mm, vma, addr,
2669 *prev = NULL; /* Tell caller we dropped mmap_lock */
2674 case SCAN_PMD_MAPPED:
2677 case SCAN_PTE_MAPPED_HUGEPAGE:
2678 BUG_ON(mmap_locked);
2680 mmap_write_lock(mm);
2681 result = collapse_pte_mapped_thp(mm, addr, true);
2682 mmap_write_unlock(mm);
2684 /* Whitelisted set of results where continuing OK */
2686 case SCAN_PTE_NON_PRESENT:
2687 case SCAN_PTE_UFFD_WP:
2689 case SCAN_LACK_REFERENCED_PAGE:
2690 case SCAN_PAGE_NULL:
2691 case SCAN_PAGE_COUNT:
2692 case SCAN_PAGE_LOCK:
2693 case SCAN_PAGE_COMPOUND:
2695 case SCAN_DEL_PAGE_LRU:
2700 /* Other error, exit */
2706 /* Caller expects us to hold mmap_lock on return */
2710 mmap_assert_locked(mm);
2714 return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0
2715 : madvise_collapse_errno(last_fail);