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/swapops.h>
20 #include <linux/shmem_fs.h>
23 #include <asm/pgalloc.h>
32 SCAN_EXCEED_SHARED_PTE,
36 SCAN_LACK_REFERENCED_PAGE,
50 SCAN_ALLOC_HUGE_PAGE_FAIL,
51 SCAN_CGROUP_CHARGE_FAIL,
53 SCAN_PAGE_HAS_PRIVATE,
56 #define CREATE_TRACE_POINTS
57 #include <trace/events/huge_memory.h>
59 static struct task_struct *khugepaged_thread __read_mostly;
60 static DEFINE_MUTEX(khugepaged_mutex);
62 /* default scan 8*512 pte (or vmas) every 30 second */
63 static unsigned int khugepaged_pages_to_scan __read_mostly;
64 static unsigned int khugepaged_pages_collapsed;
65 static unsigned int khugepaged_full_scans;
66 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
67 /* during fragmentation poll the hugepage allocator once every minute */
68 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
69 static unsigned long khugepaged_sleep_expire;
70 static DEFINE_SPINLOCK(khugepaged_mm_lock);
71 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
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
77 static unsigned int khugepaged_max_ptes_none __read_mostly;
78 static unsigned int khugepaged_max_ptes_swap __read_mostly;
79 static unsigned int khugepaged_max_ptes_shared __read_mostly;
81 #define MM_SLOTS_HASH_BITS 10
82 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
84 static struct kmem_cache *mm_slot_cache __read_mostly;
86 #define MAX_PTE_MAPPED_THP 8
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 * @nr_pte_mapped_thp: number of pte mapped THP
94 * @pte_mapped_thp: address array corresponding pte mapped THP
97 struct hlist_node hash;
98 struct list_head mm_node;
101 /* pte-mapped THP in this mm */
102 int nr_pte_mapped_thp;
103 unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
107 * struct khugepaged_scan - cursor for scanning
108 * @mm_head: the head of the mm list to scan
109 * @mm_slot: the current mm_slot we are scanning
110 * @address: the next address inside that to be scanned
112 * There is only the one khugepaged_scan instance of this cursor structure.
114 struct khugepaged_scan {
115 struct list_head mm_head;
116 struct mm_slot *mm_slot;
117 unsigned long address;
120 static struct khugepaged_scan khugepaged_scan = {
121 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
125 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
126 struct kobj_attribute *attr,
129 return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
132 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
133 struct kobj_attribute *attr,
134 const char *buf, size_t count)
139 err = kstrtouint(buf, 10, &msecs);
143 khugepaged_scan_sleep_millisecs = msecs;
144 khugepaged_sleep_expire = 0;
145 wake_up_interruptible(&khugepaged_wait);
149 static struct kobj_attribute scan_sleep_millisecs_attr =
150 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
151 scan_sleep_millisecs_store);
153 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
154 struct kobj_attribute *attr,
157 return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
160 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
161 struct kobj_attribute *attr,
162 const char *buf, size_t count)
167 err = kstrtouint(buf, 10, &msecs);
171 khugepaged_alloc_sleep_millisecs = msecs;
172 khugepaged_sleep_expire = 0;
173 wake_up_interruptible(&khugepaged_wait);
177 static struct kobj_attribute alloc_sleep_millisecs_attr =
178 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
179 alloc_sleep_millisecs_store);
181 static ssize_t pages_to_scan_show(struct kobject *kobj,
182 struct kobj_attribute *attr,
185 return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
187 static ssize_t pages_to_scan_store(struct kobject *kobj,
188 struct kobj_attribute *attr,
189 const char *buf, size_t count)
194 err = kstrtouint(buf, 10, &pages);
198 khugepaged_pages_to_scan = pages;
202 static struct kobj_attribute pages_to_scan_attr =
203 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
204 pages_to_scan_store);
206 static ssize_t pages_collapsed_show(struct kobject *kobj,
207 struct kobj_attribute *attr,
210 return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
212 static struct kobj_attribute pages_collapsed_attr =
213 __ATTR_RO(pages_collapsed);
215 static ssize_t full_scans_show(struct kobject *kobj,
216 struct kobj_attribute *attr,
219 return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
221 static struct kobj_attribute full_scans_attr =
222 __ATTR_RO(full_scans);
224 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
225 struct kobj_attribute *attr, char *buf)
227 return single_hugepage_flag_show(kobj, attr, buf,
228 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
230 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
231 struct kobj_attribute *attr,
232 const char *buf, size_t count)
234 return single_hugepage_flag_store(kobj, attr, buf, count,
235 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
237 static struct kobj_attribute khugepaged_defrag_attr =
238 __ATTR(defrag, 0644, khugepaged_defrag_show,
239 khugepaged_defrag_store);
242 * max_ptes_none controls if khugepaged should collapse hugepages over
243 * any unmapped ptes in turn potentially increasing the memory
244 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
245 * reduce the available free memory in the system as it
246 * runs. Increasing max_ptes_none will instead potentially reduce the
247 * free memory in the system during the khugepaged scan.
249 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
250 struct kobj_attribute *attr,
253 return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
255 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
256 struct kobj_attribute *attr,
257 const char *buf, size_t count)
260 unsigned long max_ptes_none;
262 err = kstrtoul(buf, 10, &max_ptes_none);
263 if (err || max_ptes_none > HPAGE_PMD_NR-1)
266 khugepaged_max_ptes_none = max_ptes_none;
270 static struct kobj_attribute khugepaged_max_ptes_none_attr =
271 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
272 khugepaged_max_ptes_none_store);
274 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
275 struct kobj_attribute *attr,
278 return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
281 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
282 struct kobj_attribute *attr,
283 const char *buf, size_t count)
286 unsigned long max_ptes_swap;
288 err = kstrtoul(buf, 10, &max_ptes_swap);
289 if (err || max_ptes_swap > HPAGE_PMD_NR-1)
292 khugepaged_max_ptes_swap = max_ptes_swap;
297 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
298 __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
299 khugepaged_max_ptes_swap_store);
301 static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
302 struct kobj_attribute *attr,
305 return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
308 static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
309 struct kobj_attribute *attr,
310 const char *buf, size_t count)
313 unsigned long max_ptes_shared;
315 err = kstrtoul(buf, 10, &max_ptes_shared);
316 if (err || max_ptes_shared > HPAGE_PMD_NR-1)
319 khugepaged_max_ptes_shared = max_ptes_shared;
324 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
325 __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
326 khugepaged_max_ptes_shared_store);
328 static struct attribute *khugepaged_attr[] = {
329 &khugepaged_defrag_attr.attr,
330 &khugepaged_max_ptes_none_attr.attr,
331 &khugepaged_max_ptes_swap_attr.attr,
332 &khugepaged_max_ptes_shared_attr.attr,
333 &pages_to_scan_attr.attr,
334 &pages_collapsed_attr.attr,
335 &full_scans_attr.attr,
336 &scan_sleep_millisecs_attr.attr,
337 &alloc_sleep_millisecs_attr.attr,
341 struct attribute_group khugepaged_attr_group = {
342 .attrs = khugepaged_attr,
343 .name = "khugepaged",
345 #endif /* CONFIG_SYSFS */
347 int hugepage_madvise(struct vm_area_struct *vma,
348 unsigned long *vm_flags, int advice)
354 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
355 * can't handle this properly after s390_enable_sie, so we simply
356 * ignore the madvise to prevent qemu from causing a SIGSEGV.
358 if (mm_has_pgste(vma->vm_mm))
361 *vm_flags &= ~VM_NOHUGEPAGE;
362 *vm_flags |= VM_HUGEPAGE;
364 * If the vma become good for khugepaged to scan,
365 * register it here without waiting a page fault that
366 * may not happen any time soon.
368 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
369 khugepaged_enter_vma_merge(vma, *vm_flags))
372 case MADV_NOHUGEPAGE:
373 *vm_flags &= ~VM_HUGEPAGE;
374 *vm_flags |= VM_NOHUGEPAGE;
376 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
377 * this vma even if we leave the mm registered in khugepaged if
378 * it got registered before VM_NOHUGEPAGE was set.
386 int __init khugepaged_init(void)
388 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
389 sizeof(struct mm_slot),
390 __alignof__(struct mm_slot), 0, NULL);
394 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
395 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
396 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
397 khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
402 void __init khugepaged_destroy(void)
404 kmem_cache_destroy(mm_slot_cache);
407 static inline struct mm_slot *alloc_mm_slot(void)
409 if (!mm_slot_cache) /* initialization failed */
411 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
414 static inline void free_mm_slot(struct mm_slot *mm_slot)
416 kmem_cache_free(mm_slot_cache, mm_slot);
419 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
421 struct mm_slot *mm_slot;
423 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
424 if (mm == mm_slot->mm)
430 static void insert_to_mm_slots_hash(struct mm_struct *mm,
431 struct mm_slot *mm_slot)
434 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
437 static inline int khugepaged_test_exit(struct mm_struct *mm)
439 return atomic_read(&mm->mm_users) == 0;
442 static bool hugepage_vma_check(struct vm_area_struct *vma,
443 unsigned long vm_flags)
445 /* Explicitly disabled through madvise. */
446 if ((vm_flags & VM_NOHUGEPAGE) ||
447 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
450 /* Enabled via shmem mount options or sysfs settings. */
451 if (shmem_file(vma->vm_file) && shmem_huge_enabled(vma)) {
452 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
456 /* THP settings require madvise. */
457 if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
460 /* Read-only file mappings need to be aligned for THP to work. */
461 if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
462 (vm_flags & VM_DENYWRITE)) {
463 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
467 if (!vma->anon_vma || vma->vm_ops)
469 if (vma_is_temporary_stack(vma))
471 return !(vm_flags & VM_NO_KHUGEPAGED);
474 int __khugepaged_enter(struct mm_struct *mm)
476 struct mm_slot *mm_slot;
479 mm_slot = alloc_mm_slot();
483 /* __khugepaged_exit() must not run from under us */
484 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
485 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
486 free_mm_slot(mm_slot);
490 spin_lock(&khugepaged_mm_lock);
491 insert_to_mm_slots_hash(mm, mm_slot);
493 * Insert just behind the scanning cursor, to let the area settle
496 wakeup = list_empty(&khugepaged_scan.mm_head);
497 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
498 spin_unlock(&khugepaged_mm_lock);
502 wake_up_interruptible(&khugepaged_wait);
507 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
508 unsigned long vm_flags)
510 unsigned long hstart, hend;
513 * khugepaged only supports read-only files for non-shmem files.
514 * khugepaged does not yet work on special mappings. And
515 * file-private shmem THP is not supported.
517 if (!hugepage_vma_check(vma, vm_flags))
520 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
521 hend = vma->vm_end & HPAGE_PMD_MASK;
523 return khugepaged_enter(vma, vm_flags);
527 void __khugepaged_exit(struct mm_struct *mm)
529 struct mm_slot *mm_slot;
532 spin_lock(&khugepaged_mm_lock);
533 mm_slot = get_mm_slot(mm);
534 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
535 hash_del(&mm_slot->hash);
536 list_del(&mm_slot->mm_node);
539 spin_unlock(&khugepaged_mm_lock);
542 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
543 free_mm_slot(mm_slot);
545 } else if (mm_slot) {
547 * This is required to serialize against
548 * khugepaged_test_exit() (which is guaranteed to run
549 * under mmap sem read mode). Stop here (after we
550 * return all pagetables will be destroyed) until
551 * khugepaged has finished working on the pagetables
552 * under the mmap_lock.
555 mmap_write_unlock(mm);
559 static void release_pte_page(struct page *page)
561 mod_node_page_state(page_pgdat(page),
562 NR_ISOLATED_ANON + page_is_file_lru(page),
565 putback_lru_page(page);
568 static void release_pte_pages(pte_t *pte, pte_t *_pte,
569 struct list_head *compound_pagelist)
571 struct page *page, *tmp;
573 while (--_pte >= pte) {
574 pte_t pteval = *_pte;
576 page = pte_page(pteval);
577 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
579 release_pte_page(page);
582 list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
583 list_del(&page->lru);
584 release_pte_page(page);
588 static bool is_refcount_suitable(struct page *page)
590 int expected_refcount;
592 expected_refcount = total_mapcount(page);
593 if (PageSwapCache(page))
594 expected_refcount += compound_nr(page);
596 return page_count(page) == expected_refcount;
599 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
600 unsigned long address,
602 struct list_head *compound_pagelist)
604 struct page *page = NULL;
606 int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
607 bool writable = false;
609 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
610 _pte++, address += PAGE_SIZE) {
611 pte_t pteval = *_pte;
612 if (pte_none(pteval) || (pte_present(pteval) &&
613 is_zero_pfn(pte_pfn(pteval)))) {
614 if (!userfaultfd_armed(vma) &&
615 ++none_or_zero <= khugepaged_max_ptes_none) {
618 result = SCAN_EXCEED_NONE_PTE;
622 if (!pte_present(pteval)) {
623 result = SCAN_PTE_NON_PRESENT;
626 page = vm_normal_page(vma, address, pteval);
627 if (unlikely(!page)) {
628 result = SCAN_PAGE_NULL;
632 VM_BUG_ON_PAGE(!PageAnon(page), page);
634 if (page_mapcount(page) > 1 &&
635 ++shared > khugepaged_max_ptes_shared) {
636 result = SCAN_EXCEED_SHARED_PTE;
640 if (PageCompound(page)) {
642 page = compound_head(page);
645 * Check if we have dealt with the compound page
648 list_for_each_entry(p, compound_pagelist, lru) {
655 * We can do it before isolate_lru_page because the
656 * page can't be freed from under us. NOTE: PG_lock
657 * is needed to serialize against split_huge_page
658 * when invoked from the VM.
660 if (!trylock_page(page)) {
661 result = SCAN_PAGE_LOCK;
666 * Check if the page has any GUP (or other external) pins.
668 * The page table that maps the page has been already unlinked
669 * from the page table tree and this process cannot get
670 * an additional pin on the page.
672 * New pins can come later if the page is shared across fork,
673 * but not from this process. The other process cannot write to
674 * the page, only trigger CoW.
676 if (!is_refcount_suitable(page)) {
678 result = SCAN_PAGE_COUNT;
681 if (!pte_write(pteval) && PageSwapCache(page) &&
682 !reuse_swap_page(page, NULL)) {
684 * Page is in the swap cache and cannot be re-used.
685 * It cannot be collapsed into a THP.
688 result = SCAN_SWAP_CACHE_PAGE;
693 * Isolate the page to avoid collapsing an hugepage
694 * currently in use by the VM.
696 if (isolate_lru_page(page)) {
698 result = SCAN_DEL_PAGE_LRU;
701 mod_node_page_state(page_pgdat(page),
702 NR_ISOLATED_ANON + page_is_file_lru(page),
704 VM_BUG_ON_PAGE(!PageLocked(page), page);
705 VM_BUG_ON_PAGE(PageLRU(page), page);
707 if (PageCompound(page))
708 list_add_tail(&page->lru, compound_pagelist);
710 /* There should be enough young pte to collapse the page */
711 if (pte_young(pteval) ||
712 page_is_young(page) || PageReferenced(page) ||
713 mmu_notifier_test_young(vma->vm_mm, address))
716 if (pte_write(pteval))
720 if (unlikely(!writable)) {
721 result = SCAN_PAGE_RO;
722 } else if (unlikely(!referenced)) {
723 result = SCAN_LACK_REFERENCED_PAGE;
725 result = SCAN_SUCCEED;
726 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
727 referenced, writable, result);
731 release_pte_pages(pte, _pte, compound_pagelist);
732 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
733 referenced, writable, result);
737 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
738 struct vm_area_struct *vma,
739 unsigned long address,
741 struct list_head *compound_pagelist)
743 struct page *src_page, *tmp;
745 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
746 _pte++, page++, address += PAGE_SIZE) {
747 pte_t pteval = *_pte;
749 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
750 clear_user_highpage(page, address);
751 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
752 if (is_zero_pfn(pte_pfn(pteval))) {
754 * ptl mostly unnecessary.
758 * paravirt calls inside pte_clear here are
761 pte_clear(vma->vm_mm, address, _pte);
765 src_page = pte_page(pteval);
766 copy_user_highpage(page, src_page, address, vma);
767 if (!PageCompound(src_page))
768 release_pte_page(src_page);
770 * ptl mostly unnecessary, but preempt has to
771 * be disabled to update the per-cpu stats
772 * inside page_remove_rmap().
776 * paravirt calls inside pte_clear here are
779 pte_clear(vma->vm_mm, address, _pte);
780 page_remove_rmap(src_page, false);
782 free_page_and_swap_cache(src_page);
786 list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
787 list_del(&src_page->lru);
788 release_pte_page(src_page);
792 static void khugepaged_alloc_sleep(void)
796 add_wait_queue(&khugepaged_wait, &wait);
797 freezable_schedule_timeout_interruptible(
798 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
799 remove_wait_queue(&khugepaged_wait, &wait);
802 static int khugepaged_node_load[MAX_NUMNODES];
804 static bool khugepaged_scan_abort(int nid)
809 * If node_reclaim_mode is disabled, then no extra effort is made to
810 * allocate memory locally.
812 if (!node_reclaim_enabled())
815 /* If there is a count for this node already, it must be acceptable */
816 if (khugepaged_node_load[nid])
819 for (i = 0; i < MAX_NUMNODES; i++) {
820 if (!khugepaged_node_load[i])
822 if (node_distance(nid, i) > node_reclaim_distance)
828 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
829 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
831 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
835 static int khugepaged_find_target_node(void)
837 static int last_khugepaged_target_node = NUMA_NO_NODE;
838 int nid, target_node = 0, max_value = 0;
840 /* find first node with max normal pages hit */
841 for (nid = 0; nid < MAX_NUMNODES; nid++)
842 if (khugepaged_node_load[nid] > max_value) {
843 max_value = khugepaged_node_load[nid];
847 /* do some balance if several nodes have the same hit record */
848 if (target_node <= last_khugepaged_target_node)
849 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
851 if (max_value == khugepaged_node_load[nid]) {
856 last_khugepaged_target_node = target_node;
860 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
862 if (IS_ERR(*hpage)) {
868 khugepaged_alloc_sleep();
878 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
880 VM_BUG_ON_PAGE(*hpage, *hpage);
882 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
883 if (unlikely(!*hpage)) {
884 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
885 *hpage = ERR_PTR(-ENOMEM);
889 prep_transhuge_page(*hpage);
890 count_vm_event(THP_COLLAPSE_ALLOC);
894 static int khugepaged_find_target_node(void)
899 static inline struct page *alloc_khugepaged_hugepage(void)
903 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
906 prep_transhuge_page(page);
910 static struct page *khugepaged_alloc_hugepage(bool *wait)
915 hpage = alloc_khugepaged_hugepage();
917 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
922 khugepaged_alloc_sleep();
924 count_vm_event(THP_COLLAPSE_ALLOC);
925 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
930 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
933 * If the hpage allocated earlier was briefly exposed in page cache
934 * before collapse_file() failed, it is possible that racing lookups
935 * have not yet completed, and would then be unpleasantly surprised by
936 * finding the hpage reused for the same mapping at a different offset.
937 * Just release the previous allocation if there is any danger of that.
939 if (*hpage && page_count(*hpage) > 1) {
945 *hpage = khugepaged_alloc_hugepage(wait);
947 if (unlikely(!*hpage))
954 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
963 * If mmap_lock temporarily dropped, revalidate vma
964 * before taking mmap_lock.
965 * Return 0 if succeeds, otherwise return none-zero
969 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
970 struct vm_area_struct **vmap)
972 struct vm_area_struct *vma;
973 unsigned long hstart, hend;
975 if (unlikely(khugepaged_test_exit(mm)))
976 return SCAN_ANY_PROCESS;
978 *vmap = vma = find_vma(mm, address);
980 return SCAN_VMA_NULL;
982 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
983 hend = vma->vm_end & HPAGE_PMD_MASK;
984 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
985 return SCAN_ADDRESS_RANGE;
986 if (!hugepage_vma_check(vma, vma->vm_flags))
987 return SCAN_VMA_CHECK;
988 /* Anon VMA expected */
989 if (!vma->anon_vma || vma->vm_ops)
990 return SCAN_VMA_CHECK;
995 * Bring missing pages in from swap, to complete THP collapse.
996 * Only done if khugepaged_scan_pmd believes it is worthwhile.
998 * Called and returns without pte mapped or spinlocks held,
999 * but with mmap_lock held to protect against vma changes.
1002 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
1003 struct vm_area_struct *vma,
1004 unsigned long haddr, pmd_t *pmd,
1009 unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
1011 for (address = haddr; address < end; address += PAGE_SIZE) {
1012 struct vm_fault vmf = {
1015 .pgoff = linear_page_index(vma, haddr),
1016 .flags = FAULT_FLAG_ALLOW_RETRY,
1020 vmf.pte = pte_offset_map(pmd, address);
1021 vmf.orig_pte = *vmf.pte;
1022 if (!is_swap_pte(vmf.orig_pte)) {
1027 ret = do_swap_page(&vmf);
1029 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1030 if (ret & VM_FAULT_RETRY) {
1032 if (hugepage_vma_revalidate(mm, haddr, &vma)) {
1033 /* vma is no longer available, don't continue to swapin */
1034 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1037 /* check if the pmd is still valid */
1038 if (mm_find_pmd(mm, haddr) != pmd) {
1039 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1043 if (ret & VM_FAULT_ERROR) {
1044 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1049 /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1053 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1057 static void collapse_huge_page(struct mm_struct *mm,
1058 unsigned long address,
1059 struct page **hpage,
1060 int node, int referenced, int unmapped)
1062 LIST_HEAD(compound_pagelist);
1066 struct page *new_page;
1067 spinlock_t *pmd_ptl, *pte_ptl;
1068 int isolated = 0, result = 0;
1069 struct vm_area_struct *vma;
1070 struct mmu_notifier_range range;
1073 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1075 /* Only allocate from the target node */
1076 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1079 * Before allocating the hugepage, release the mmap_lock read lock.
1080 * The allocation can take potentially a long time if it involves
1081 * sync compaction, and we do not need to hold the mmap_lock during
1082 * that. We will recheck the vma after taking it again in write mode.
1084 mmap_read_unlock(mm);
1085 new_page = khugepaged_alloc_page(hpage, gfp, node);
1087 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1091 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1092 result = SCAN_CGROUP_CHARGE_FAIL;
1095 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1098 result = hugepage_vma_revalidate(mm, address, &vma);
1100 mmap_read_unlock(mm);
1104 pmd = mm_find_pmd(mm, address);
1106 result = SCAN_PMD_NULL;
1107 mmap_read_unlock(mm);
1112 * __collapse_huge_page_swapin always returns with mmap_lock locked.
1113 * If it fails, we release mmap_lock and jump out_nolock.
1114 * Continuing to collapse causes inconsistency.
1116 if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1118 mmap_read_unlock(mm);
1122 mmap_read_unlock(mm);
1124 * Prevent all access to pagetables with the exception of
1125 * gup_fast later handled by the ptep_clear_flush and the VM
1126 * handled by the anon_vma lock + PG_lock.
1128 mmap_write_lock(mm);
1129 result = hugepage_vma_revalidate(mm, address, &vma);
1132 /* check if the pmd is still valid */
1133 if (mm_find_pmd(mm, address) != pmd)
1136 anon_vma_lock_write(vma->anon_vma);
1138 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1139 address, address + HPAGE_PMD_SIZE);
1140 mmu_notifier_invalidate_range_start(&range);
1142 pte = pte_offset_map(pmd, address);
1143 pte_ptl = pte_lockptr(mm, pmd);
1145 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1147 * After this gup_fast can't run anymore. This also removes
1148 * any huge TLB entry from the CPU so we won't allow
1149 * huge and small TLB entries for the same virtual address
1150 * to avoid the risk of CPU bugs in that area.
1152 _pmd = pmdp_collapse_flush(vma, address, pmd);
1153 spin_unlock(pmd_ptl);
1154 mmu_notifier_invalidate_range_end(&range);
1157 isolated = __collapse_huge_page_isolate(vma, address, pte,
1158 &compound_pagelist);
1159 spin_unlock(pte_ptl);
1161 if (unlikely(!isolated)) {
1164 BUG_ON(!pmd_none(*pmd));
1166 * We can only use set_pmd_at when establishing
1167 * hugepmds and never for establishing regular pmds that
1168 * points to regular pagetables. Use pmd_populate for that
1170 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1171 spin_unlock(pmd_ptl);
1172 anon_vma_unlock_write(vma->anon_vma);
1178 * All pages are isolated and locked so anon_vma rmap
1179 * can't run anymore.
1181 anon_vma_unlock_write(vma->anon_vma);
1183 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1184 &compound_pagelist);
1187 * spin_lock() below is not the equivalent of smp_wmb(), but
1188 * the smp_wmb() inside __SetPageUptodate() can be reused to
1189 * avoid the copy_huge_page writes to become visible after
1190 * the set_pmd_at() write.
1192 __SetPageUptodate(new_page);
1193 pgtable = pmd_pgtable(_pmd);
1195 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1196 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1199 BUG_ON(!pmd_none(*pmd));
1200 page_add_new_anon_rmap(new_page, vma, address, true);
1201 lru_cache_add_inactive_or_unevictable(new_page, vma);
1202 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1203 set_pmd_at(mm, address, pmd, _pmd);
1204 update_mmu_cache_pmd(vma, address, pmd);
1205 spin_unlock(pmd_ptl);
1209 khugepaged_pages_collapsed++;
1210 result = SCAN_SUCCEED;
1212 mmap_write_unlock(mm);
1214 if (!IS_ERR_OR_NULL(*hpage))
1215 mem_cgroup_uncharge(*hpage);
1216 trace_mm_collapse_huge_page(mm, isolated, result);
1220 static int khugepaged_scan_pmd(struct mm_struct *mm,
1221 struct vm_area_struct *vma,
1222 unsigned long address,
1223 struct page **hpage)
1227 int ret = 0, result = 0, referenced = 0;
1228 int none_or_zero = 0, shared = 0;
1229 struct page *page = NULL;
1230 unsigned long _address;
1232 int node = NUMA_NO_NODE, unmapped = 0;
1233 bool writable = false;
1235 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1237 pmd = mm_find_pmd(mm, address);
1239 result = SCAN_PMD_NULL;
1243 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1244 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1245 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1246 _pte++, _address += PAGE_SIZE) {
1247 pte_t pteval = *_pte;
1248 if (is_swap_pte(pteval)) {
1249 if (++unmapped <= khugepaged_max_ptes_swap) {
1251 * Always be strict with uffd-wp
1252 * enabled swap entries. Please see
1253 * comment below for pte_uffd_wp().
1255 if (pte_swp_uffd_wp(pteval)) {
1256 result = SCAN_PTE_UFFD_WP;
1261 result = SCAN_EXCEED_SWAP_PTE;
1265 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1266 if (!userfaultfd_armed(vma) &&
1267 ++none_or_zero <= khugepaged_max_ptes_none) {
1270 result = SCAN_EXCEED_NONE_PTE;
1274 if (pte_uffd_wp(pteval)) {
1276 * Don't collapse the page if any of the small
1277 * PTEs are armed with uffd write protection.
1278 * Here we can also mark the new huge pmd as
1279 * write protected if any of the small ones is
1280 * marked but that could bring unknown
1281 * userfault messages that falls outside of
1282 * the registered range. So, just be simple.
1284 result = SCAN_PTE_UFFD_WP;
1287 if (pte_write(pteval))
1290 page = vm_normal_page(vma, _address, pteval);
1291 if (unlikely(!page)) {
1292 result = SCAN_PAGE_NULL;
1296 if (page_mapcount(page) > 1 &&
1297 ++shared > khugepaged_max_ptes_shared) {
1298 result = SCAN_EXCEED_SHARED_PTE;
1302 page = compound_head(page);
1305 * Record which node the original page is from and save this
1306 * information to khugepaged_node_load[].
1307 * Khupaged will allocate hugepage from the node has the max
1310 node = page_to_nid(page);
1311 if (khugepaged_scan_abort(node)) {
1312 result = SCAN_SCAN_ABORT;
1315 khugepaged_node_load[node]++;
1316 if (!PageLRU(page)) {
1317 result = SCAN_PAGE_LRU;
1320 if (PageLocked(page)) {
1321 result = SCAN_PAGE_LOCK;
1324 if (!PageAnon(page)) {
1325 result = SCAN_PAGE_ANON;
1330 * Check if the page has any GUP (or other external) pins.
1332 * Here the check is racy it may see totmal_mapcount > refcount
1334 * For example, one process with one forked child process.
1335 * The parent has the PMD split due to MADV_DONTNEED, then
1336 * the child is trying unmap the whole PMD, but khugepaged
1337 * may be scanning the parent between the child has
1338 * PageDoubleMap flag cleared and dec the mapcount. So
1339 * khugepaged may see total_mapcount > refcount.
1341 * But such case is ephemeral we could always retry collapse
1342 * later. However it may report false positive if the page
1343 * has excessive GUP pins (i.e. 512). Anyway the same check
1344 * will be done again later the risk seems low.
1346 if (!is_refcount_suitable(page)) {
1347 result = SCAN_PAGE_COUNT;
1350 if (pte_young(pteval) ||
1351 page_is_young(page) || PageReferenced(page) ||
1352 mmu_notifier_test_young(vma->vm_mm, address))
1356 result = SCAN_PAGE_RO;
1357 } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1358 result = SCAN_LACK_REFERENCED_PAGE;
1360 result = SCAN_SUCCEED;
1364 pte_unmap_unlock(pte, ptl);
1366 node = khugepaged_find_target_node();
1367 /* collapse_huge_page will return with the mmap_lock released */
1368 collapse_huge_page(mm, address, hpage, node,
1369 referenced, unmapped);
1372 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1373 none_or_zero, result, unmapped);
1377 static void collect_mm_slot(struct mm_slot *mm_slot)
1379 struct mm_struct *mm = mm_slot->mm;
1381 lockdep_assert_held(&khugepaged_mm_lock);
1383 if (khugepaged_test_exit(mm)) {
1385 hash_del(&mm_slot->hash);
1386 list_del(&mm_slot->mm_node);
1389 * Not strictly needed because the mm exited already.
1391 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1394 /* khugepaged_mm_lock actually not necessary for the below */
1395 free_mm_slot(mm_slot);
1402 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1403 * khugepaged should try to collapse the page table.
1405 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1408 struct mm_slot *mm_slot;
1410 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1412 spin_lock(&khugepaged_mm_lock);
1413 mm_slot = get_mm_slot(mm);
1414 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1415 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1416 spin_unlock(&khugepaged_mm_lock);
1421 * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1424 * @mm: process address space where collapse happens
1425 * @addr: THP collapse address
1427 * This function checks whether all the PTEs in the PMD are pointing to the
1428 * right THP. If so, retract the page table so the THP can refault in with
1431 void 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 = find_vma(mm, haddr);
1436 pte_t *start_pte, *pte;
1442 if (!vma || !vma->vm_file ||
1443 !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1447 * This vm_flags may not have VM_HUGEPAGE if the page was not
1448 * collapsed by this mm. But we can still collapse if the page is
1449 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1450 * will not fail the vma for missing VM_HUGEPAGE
1452 if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1455 hpage = find_lock_page(vma->vm_file->f_mapping,
1456 linear_page_index(vma, haddr));
1460 if (!PageHead(hpage))
1463 pmd = mm_find_pmd(mm, haddr);
1467 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1469 /* step 1: check all mapped PTEs are to the right huge page */
1470 for (i = 0, addr = haddr, pte = start_pte;
1471 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1474 /* empty pte, skip */
1478 /* page swapped out, abort */
1479 if (!pte_present(*pte))
1482 page = vm_normal_page(vma, addr, *pte);
1485 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1486 * page table, but the new page will not be a subpage of hpage.
1488 if (hpage + i != page)
1493 /* step 2: adjust rmap */
1494 for (i = 0, addr = haddr, pte = start_pte;
1495 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1500 page = vm_normal_page(vma, addr, *pte);
1501 page_remove_rmap(page, false);
1504 pte_unmap_unlock(start_pte, ptl);
1506 /* step 3: set proper refcount and mm_counters. */
1508 page_ref_sub(hpage, count);
1509 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1512 /* step 4: collapse pmd */
1513 ptl = pmd_lock(vma->vm_mm, pmd);
1514 _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1517 pte_free(mm, pmd_pgtable(_pmd));
1525 pte_unmap_unlock(start_pte, ptl);
1529 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1531 struct mm_struct *mm = mm_slot->mm;
1534 if (likely(mm_slot->nr_pte_mapped_thp == 0))
1537 if (!mmap_write_trylock(mm))
1540 if (unlikely(khugepaged_test_exit(mm)))
1543 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1544 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1547 mm_slot->nr_pte_mapped_thp = 0;
1548 mmap_write_unlock(mm);
1551 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1553 struct vm_area_struct *vma;
1554 struct mm_struct *mm;
1558 i_mmap_lock_write(mapping);
1559 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1561 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1562 * got written to. These VMAs are likely not worth investing
1563 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1566 * Not that vma->anon_vma check is racy: it can be set up after
1567 * the check but before we took mmap_lock by the fault path.
1568 * But page lock would prevent establishing any new ptes of the
1569 * page, so we are safe.
1571 * An alternative would be drop the check, but check that page
1572 * table is clear before calling pmdp_collapse_flush() under
1573 * ptl. It has higher chance to recover THP for the VMA, but
1574 * has higher cost too.
1578 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1579 if (addr & ~HPAGE_PMD_MASK)
1581 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1584 pmd = mm_find_pmd(mm, addr);
1588 * We need exclusive mmap_lock to retract page table.
1590 * We use trylock due to lock inversion: we need to acquire
1591 * mmap_lock while holding page lock. Fault path does it in
1592 * reverse order. Trylock is a way to avoid deadlock.
1594 if (mmap_write_trylock(mm)) {
1595 if (!khugepaged_test_exit(mm)) {
1596 spinlock_t *ptl = pmd_lock(mm, pmd);
1597 /* assume page table is clear */
1598 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1601 pte_free(mm, pmd_pgtable(_pmd));
1603 mmap_write_unlock(mm);
1605 /* Try again later */
1606 khugepaged_add_pte_mapped_thp(mm, addr);
1609 i_mmap_unlock_write(mapping);
1613 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1615 * @mm: process address space where collapse happens
1616 * @file: file that collapse on
1617 * @start: collapse start address
1618 * @hpage: new allocated huge page for collapse
1619 * @node: appointed node the new huge page allocate from
1621 * Basic scheme is simple, details are more complex:
1622 * - allocate and lock a new huge page;
1623 * - scan page cache replacing old pages with the new one
1624 * + swap/gup in pages if necessary;
1626 * + keep old pages around in case rollback is required;
1627 * - if replacing succeeds:
1630 * + unlock huge page;
1631 * - if replacing failed;
1632 * + put all pages back and unfreeze them;
1633 * + restore gaps in the page cache;
1634 * + unlock and free huge page;
1636 static void collapse_file(struct mm_struct *mm,
1637 struct file *file, pgoff_t start,
1638 struct page **hpage, int node)
1640 struct address_space *mapping = file->f_mapping;
1642 struct page *new_page;
1643 pgoff_t index, end = start + HPAGE_PMD_NR;
1644 LIST_HEAD(pagelist);
1645 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1646 int nr_none = 0, result = SCAN_SUCCEED;
1647 bool is_shmem = shmem_file(file);
1650 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1651 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1653 /* Only allocate from the target node */
1654 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1656 new_page = khugepaged_alloc_page(hpage, gfp, node);
1658 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1662 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1663 result = SCAN_CGROUP_CHARGE_FAIL;
1666 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1668 /* This will be less messy when we use multi-index entries */
1671 xas_create_range(&xas);
1672 if (!xas_error(&xas))
1674 xas_unlock_irq(&xas);
1675 if (!xas_nomem(&xas, GFP_KERNEL)) {
1681 __SetPageLocked(new_page);
1683 __SetPageSwapBacked(new_page);
1684 new_page->index = start;
1685 new_page->mapping = mapping;
1688 * At this point the new_page is locked and not up-to-date.
1689 * It's safe to insert it into the page cache, because nobody would
1690 * be able to map it or use it in another way until we unlock it.
1693 xas_set(&xas, start);
1694 for (index = start; index < end; index++) {
1695 struct page *page = xas_next(&xas);
1697 VM_BUG_ON(index != xas.xa_index);
1701 * Stop if extent has been truncated or
1702 * hole-punched, and is now completely
1705 if (index == start) {
1706 if (!xas_next_entry(&xas, end - 1)) {
1707 result = SCAN_TRUNCATED;
1710 xas_set(&xas, index);
1712 if (!shmem_charge(mapping->host, 1)) {
1716 xas_store(&xas, new_page);
1721 if (xa_is_value(page) || !PageUptodate(page)) {
1722 xas_unlock_irq(&xas);
1723 /* swap in or instantiate fallocated page */
1724 if (shmem_getpage(mapping->host, index, &page,
1729 } else if (trylock_page(page)) {
1731 xas_unlock_irq(&xas);
1733 result = SCAN_PAGE_LOCK;
1736 } else { /* !is_shmem */
1737 if (!page || xa_is_value(page)) {
1738 xas_unlock_irq(&xas);
1739 page_cache_sync_readahead(mapping, &file->f_ra,
1742 /* drain pagevecs to help isolate_lru_page() */
1744 page = find_lock_page(mapping, index);
1745 if (unlikely(page == NULL)) {
1749 } else if (PageDirty(page)) {
1751 * khugepaged only works on read-only fd,
1752 * so this page is dirty because it hasn't
1753 * been flushed since first write. There
1754 * won't be new dirty pages.
1756 * Trigger async flush here and hope the
1757 * writeback is done when khugepaged
1758 * revisits this page.
1760 * This is a one-off situation. We are not
1761 * forcing writeback in loop.
1763 xas_unlock_irq(&xas);
1764 filemap_flush(mapping);
1767 } else if (trylock_page(page)) {
1769 xas_unlock_irq(&xas);
1771 result = SCAN_PAGE_LOCK;
1777 * The page must be locked, so we can drop the i_pages lock
1778 * without racing with truncate.
1780 VM_BUG_ON_PAGE(!PageLocked(page), page);
1782 /* make sure the page is up to date */
1783 if (unlikely(!PageUptodate(page))) {
1789 * If file was truncated then extended, or hole-punched, before
1790 * we locked the first page, then a THP might be there already.
1792 if (PageTransCompound(page)) {
1793 result = SCAN_PAGE_COMPOUND;
1797 if (page_mapping(page) != mapping) {
1798 result = SCAN_TRUNCATED;
1802 if (!is_shmem && PageDirty(page)) {
1804 * khugepaged only works on read-only fd, so this
1805 * page is dirty because it hasn't been flushed
1806 * since first write.
1812 if (isolate_lru_page(page)) {
1813 result = SCAN_DEL_PAGE_LRU;
1817 if (page_has_private(page) &&
1818 !try_to_release_page(page, GFP_KERNEL)) {
1819 result = SCAN_PAGE_HAS_PRIVATE;
1820 putback_lru_page(page);
1824 if (page_mapped(page))
1825 unmap_mapping_pages(mapping, index, 1, false);
1828 xas_set(&xas, index);
1830 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1831 VM_BUG_ON_PAGE(page_mapped(page), page);
1834 * The page is expected to have page_count() == 3:
1835 * - we hold a pin on it;
1836 * - one reference from page cache;
1837 * - one from isolate_lru_page;
1839 if (!page_ref_freeze(page, 3)) {
1840 result = SCAN_PAGE_COUNT;
1841 xas_unlock_irq(&xas);
1842 putback_lru_page(page);
1847 * Add the page to the list to be able to undo the collapse if
1848 * something go wrong.
1850 list_add_tail(&page->lru, &pagelist);
1852 /* Finally, replace with the new page. */
1853 xas_store(&xas, new_page);
1860 nr = thp_nr_pages(new_page);
1863 __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
1865 __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
1866 filemap_nr_thps_inc(mapping);
1870 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1872 __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1876 xas_unlock_irq(&xas);
1879 if (result == SCAN_SUCCEED) {
1880 struct page *page, *tmp;
1883 * Replacing old pages with new one has succeeded, now we
1884 * need to copy the content and free the old pages.
1887 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1888 while (index < page->index) {
1889 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1892 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1894 list_del(&page->lru);
1895 page->mapping = NULL;
1896 page_ref_unfreeze(page, 1);
1897 ClearPageActive(page);
1898 ClearPageUnevictable(page);
1903 while (index < end) {
1904 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1908 SetPageUptodate(new_page);
1909 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1911 set_page_dirty(new_page);
1912 lru_cache_add(new_page);
1915 * Remove pte page tables, so we can re-fault the page as huge.
1917 retract_page_tables(mapping, start);
1920 khugepaged_pages_collapsed++;
1924 /* Something went wrong: roll back page cache changes */
1926 mapping->nrpages -= nr_none;
1929 shmem_uncharge(mapping->host, nr_none);
1931 xas_set(&xas, start);
1932 xas_for_each(&xas, page, end - 1) {
1933 page = list_first_entry_or_null(&pagelist,
1935 if (!page || xas.xa_index < page->index) {
1939 /* Put holes back where they were */
1940 xas_store(&xas, NULL);
1944 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1946 /* Unfreeze the page. */
1947 list_del(&page->lru);
1948 page_ref_unfreeze(page, 2);
1949 xas_store(&xas, page);
1951 xas_unlock_irq(&xas);
1953 putback_lru_page(page);
1957 xas_unlock_irq(&xas);
1959 new_page->mapping = NULL;
1962 unlock_page(new_page);
1964 VM_BUG_ON(!list_empty(&pagelist));
1965 if (!IS_ERR_OR_NULL(*hpage))
1966 mem_cgroup_uncharge(*hpage);
1967 /* TODO: tracepoints */
1970 static void khugepaged_scan_file(struct mm_struct *mm,
1971 struct file *file, pgoff_t start, struct page **hpage)
1973 struct page *page = NULL;
1974 struct address_space *mapping = file->f_mapping;
1975 XA_STATE(xas, &mapping->i_pages, start);
1977 int node = NUMA_NO_NODE;
1978 int result = SCAN_SUCCEED;
1982 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1984 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1985 if (xas_retry(&xas, page))
1988 if (xa_is_value(page)) {
1989 if (++swap > khugepaged_max_ptes_swap) {
1990 result = SCAN_EXCEED_SWAP_PTE;
1996 if (PageTransCompound(page)) {
1997 result = SCAN_PAGE_COMPOUND;
2001 node = page_to_nid(page);
2002 if (khugepaged_scan_abort(node)) {
2003 result = SCAN_SCAN_ABORT;
2006 khugepaged_node_load[node]++;
2008 if (!PageLRU(page)) {
2009 result = SCAN_PAGE_LRU;
2013 if (page_count(page) !=
2014 1 + page_mapcount(page) + page_has_private(page)) {
2015 result = SCAN_PAGE_COUNT;
2020 * We probably should check if the page is referenced here, but
2021 * nobody would transfer pte_young() to PageReferenced() for us.
2022 * And rmap walk here is just too costly...
2027 if (need_resched()) {
2034 if (result == SCAN_SUCCEED) {
2035 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2036 result = SCAN_EXCEED_NONE_PTE;
2038 node = khugepaged_find_target_node();
2039 collapse_file(mm, file, start, hpage, node);
2043 /* TODO: tracepoints */
2046 static void khugepaged_scan_file(struct mm_struct *mm,
2047 struct file *file, pgoff_t start, struct page **hpage)
2052 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2057 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2058 struct page **hpage)
2059 __releases(&khugepaged_mm_lock)
2060 __acquires(&khugepaged_mm_lock)
2062 struct mm_slot *mm_slot;
2063 struct mm_struct *mm;
2064 struct vm_area_struct *vma;
2068 lockdep_assert_held(&khugepaged_mm_lock);
2070 if (khugepaged_scan.mm_slot)
2071 mm_slot = khugepaged_scan.mm_slot;
2073 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2074 struct mm_slot, mm_node);
2075 khugepaged_scan.address = 0;
2076 khugepaged_scan.mm_slot = mm_slot;
2078 spin_unlock(&khugepaged_mm_lock);
2079 khugepaged_collapse_pte_mapped_thps(mm_slot);
2083 * Don't wait for semaphore (to avoid long wait times). Just move to
2084 * the next mm on the list.
2087 if (unlikely(!mmap_read_trylock(mm)))
2088 goto breakouterloop_mmap_lock;
2089 if (likely(!khugepaged_test_exit(mm)))
2090 vma = find_vma(mm, khugepaged_scan.address);
2093 for (; vma; vma = vma->vm_next) {
2094 unsigned long hstart, hend;
2097 if (unlikely(khugepaged_test_exit(mm))) {
2101 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2106 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2107 hend = vma->vm_end & HPAGE_PMD_MASK;
2110 if (khugepaged_scan.address > hend)
2112 if (khugepaged_scan.address < hstart)
2113 khugepaged_scan.address = hstart;
2114 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2115 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2118 while (khugepaged_scan.address < hend) {
2121 if (unlikely(khugepaged_test_exit(mm)))
2122 goto breakouterloop;
2124 VM_BUG_ON(khugepaged_scan.address < hstart ||
2125 khugepaged_scan.address + HPAGE_PMD_SIZE >
2127 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2128 struct file *file = get_file(vma->vm_file);
2129 pgoff_t pgoff = linear_page_index(vma,
2130 khugepaged_scan.address);
2132 mmap_read_unlock(mm);
2134 khugepaged_scan_file(mm, file, pgoff, hpage);
2137 ret = khugepaged_scan_pmd(mm, vma,
2138 khugepaged_scan.address,
2141 /* move to next address */
2142 khugepaged_scan.address += HPAGE_PMD_SIZE;
2143 progress += HPAGE_PMD_NR;
2145 /* we released mmap_lock so break loop */
2146 goto breakouterloop_mmap_lock;
2147 if (progress >= pages)
2148 goto breakouterloop;
2152 mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2153 breakouterloop_mmap_lock:
2155 spin_lock(&khugepaged_mm_lock);
2156 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2158 * Release the current mm_slot if this mm is about to die, or
2159 * if we scanned all vmas of this mm.
2161 if (khugepaged_test_exit(mm) || !vma) {
2163 * Make sure that if mm_users is reaching zero while
2164 * khugepaged runs here, khugepaged_exit will find
2165 * mm_slot not pointing to the exiting mm.
2167 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2168 khugepaged_scan.mm_slot = list_entry(
2169 mm_slot->mm_node.next,
2170 struct mm_slot, mm_node);
2171 khugepaged_scan.address = 0;
2173 khugepaged_scan.mm_slot = NULL;
2174 khugepaged_full_scans++;
2177 collect_mm_slot(mm_slot);
2183 static int khugepaged_has_work(void)
2185 return !list_empty(&khugepaged_scan.mm_head) &&
2186 khugepaged_enabled();
2189 static int khugepaged_wait_event(void)
2191 return !list_empty(&khugepaged_scan.mm_head) ||
2192 kthread_should_stop();
2195 static void khugepaged_do_scan(void)
2197 struct page *hpage = NULL;
2198 unsigned int progress = 0, pass_through_head = 0;
2199 unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2202 lru_add_drain_all();
2204 while (progress < pages) {
2205 if (!khugepaged_prealloc_page(&hpage, &wait))
2210 if (unlikely(kthread_should_stop() || try_to_freeze()))
2213 spin_lock(&khugepaged_mm_lock);
2214 if (!khugepaged_scan.mm_slot)
2215 pass_through_head++;
2216 if (khugepaged_has_work() &&
2217 pass_through_head < 2)
2218 progress += khugepaged_scan_mm_slot(pages - progress,
2222 spin_unlock(&khugepaged_mm_lock);
2225 if (!IS_ERR_OR_NULL(hpage))
2229 static bool khugepaged_should_wakeup(void)
2231 return kthread_should_stop() ||
2232 time_after_eq(jiffies, khugepaged_sleep_expire);
2235 static void khugepaged_wait_work(void)
2237 if (khugepaged_has_work()) {
2238 const unsigned long scan_sleep_jiffies =
2239 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2241 if (!scan_sleep_jiffies)
2244 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2245 wait_event_freezable_timeout(khugepaged_wait,
2246 khugepaged_should_wakeup(),
2247 scan_sleep_jiffies);
2251 if (khugepaged_enabled())
2252 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2255 static int khugepaged(void *none)
2257 struct mm_slot *mm_slot;
2260 set_user_nice(current, MAX_NICE);
2262 while (!kthread_should_stop()) {
2263 khugepaged_do_scan();
2264 khugepaged_wait_work();
2267 spin_lock(&khugepaged_mm_lock);
2268 mm_slot = khugepaged_scan.mm_slot;
2269 khugepaged_scan.mm_slot = NULL;
2271 collect_mm_slot(mm_slot);
2272 spin_unlock(&khugepaged_mm_lock);
2276 static void set_recommended_min_free_kbytes(void)
2280 unsigned long recommended_min;
2282 for_each_populated_zone(zone) {
2284 * We don't need to worry about fragmentation of
2285 * ZONE_MOVABLE since it only has movable pages.
2287 if (zone_idx(zone) > gfp_zone(GFP_USER))
2293 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2294 recommended_min = pageblock_nr_pages * nr_zones * 2;
2297 * Make sure that on average at least two pageblocks are almost free
2298 * of another type, one for a migratetype to fall back to and a
2299 * second to avoid subsequent fallbacks of other types There are 3
2300 * MIGRATE_TYPES we care about.
2302 recommended_min += pageblock_nr_pages * nr_zones *
2303 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2305 /* don't ever allow to reserve more than 5% of the lowmem */
2306 recommended_min = min(recommended_min,
2307 (unsigned long) nr_free_buffer_pages() / 20);
2308 recommended_min <<= (PAGE_SHIFT-10);
2310 if (recommended_min > min_free_kbytes) {
2311 if (user_min_free_kbytes >= 0)
2312 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2313 min_free_kbytes, recommended_min);
2315 min_free_kbytes = recommended_min;
2317 setup_per_zone_wmarks();
2320 int start_stop_khugepaged(void)
2324 mutex_lock(&khugepaged_mutex);
2325 if (khugepaged_enabled()) {
2326 if (!khugepaged_thread)
2327 khugepaged_thread = kthread_run(khugepaged, NULL,
2329 if (IS_ERR(khugepaged_thread)) {
2330 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2331 err = PTR_ERR(khugepaged_thread);
2332 khugepaged_thread = NULL;
2336 if (!list_empty(&khugepaged_scan.mm_head))
2337 wake_up_interruptible(&khugepaged_wait);
2339 set_recommended_min_free_kbytes();
2340 } else if (khugepaged_thread) {
2341 kthread_stop(khugepaged_thread);
2342 khugepaged_thread = NULL;
2345 mutex_unlock(&khugepaged_mutex);
2349 void khugepaged_min_free_kbytes_update(void)
2351 mutex_lock(&khugepaged_mutex);
2352 if (khugepaged_enabled() && khugepaged_thread)
2353 set_recommended_min_free_kbytes();
2354 mutex_unlock(&khugepaged_mutex);