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 additinal 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_mode)
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);
1222 static int khugepaged_scan_pmd(struct mm_struct *mm,
1223 struct vm_area_struct *vma,
1224 unsigned long address,
1225 struct page **hpage)
1229 int ret = 0, result = 0, referenced = 0;
1230 int none_or_zero = 0, shared = 0;
1231 struct page *page = NULL;
1232 unsigned long _address;
1234 int node = NUMA_NO_NODE, unmapped = 0;
1235 bool writable = false;
1237 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1239 pmd = mm_find_pmd(mm, address);
1241 result = SCAN_PMD_NULL;
1245 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1246 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1247 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1248 _pte++, _address += PAGE_SIZE) {
1249 pte_t pteval = *_pte;
1250 if (is_swap_pte(pteval)) {
1251 if (++unmapped <= khugepaged_max_ptes_swap) {
1253 * Always be strict with uffd-wp
1254 * enabled swap entries. Please see
1255 * comment below for pte_uffd_wp().
1257 if (pte_swp_uffd_wp(pteval)) {
1258 result = SCAN_PTE_UFFD_WP;
1263 result = SCAN_EXCEED_SWAP_PTE;
1267 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1268 if (!userfaultfd_armed(vma) &&
1269 ++none_or_zero <= khugepaged_max_ptes_none) {
1272 result = SCAN_EXCEED_NONE_PTE;
1276 if (!pte_present(pteval)) {
1277 result = SCAN_PTE_NON_PRESENT;
1280 if (pte_uffd_wp(pteval)) {
1282 * Don't collapse the page if any of the small
1283 * PTEs are armed with uffd write protection.
1284 * Here we can also mark the new huge pmd as
1285 * write protected if any of the small ones is
1286 * marked but that could bring unknown
1287 * userfault messages that falls outside of
1288 * the registered range. So, just be simple.
1290 result = SCAN_PTE_UFFD_WP;
1293 if (pte_write(pteval))
1296 page = vm_normal_page(vma, _address, pteval);
1297 if (unlikely(!page)) {
1298 result = SCAN_PAGE_NULL;
1302 if (page_mapcount(page) > 1 &&
1303 ++shared > khugepaged_max_ptes_shared) {
1304 result = SCAN_EXCEED_SHARED_PTE;
1308 page = compound_head(page);
1311 * Record which node the original page is from and save this
1312 * information to khugepaged_node_load[].
1313 * Khupaged will allocate hugepage from the node has the max
1316 node = page_to_nid(page);
1317 if (khugepaged_scan_abort(node)) {
1318 result = SCAN_SCAN_ABORT;
1321 khugepaged_node_load[node]++;
1322 if (!PageLRU(page)) {
1323 result = SCAN_PAGE_LRU;
1326 if (PageLocked(page)) {
1327 result = SCAN_PAGE_LOCK;
1330 if (!PageAnon(page)) {
1331 result = SCAN_PAGE_ANON;
1336 * Check if the page has any GUP (or other external) pins.
1338 * Here the check is racy it may see totmal_mapcount > refcount
1340 * For example, one process with one forked child process.
1341 * The parent has the PMD split due to MADV_DONTNEED, then
1342 * the child is trying unmap the whole PMD, but khugepaged
1343 * may be scanning the parent between the child has
1344 * PageDoubleMap flag cleared and dec the mapcount. So
1345 * khugepaged may see total_mapcount > refcount.
1347 * But such case is ephemeral we could always retry collapse
1348 * later. However it may report false positive if the page
1349 * has excessive GUP pins (i.e. 512). Anyway the same check
1350 * will be done again later the risk seems low.
1352 if (!is_refcount_suitable(page)) {
1353 result = SCAN_PAGE_COUNT;
1356 if (pte_young(pteval) ||
1357 page_is_young(page) || PageReferenced(page) ||
1358 mmu_notifier_test_young(vma->vm_mm, address))
1362 result = SCAN_PAGE_RO;
1363 } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1364 result = SCAN_LACK_REFERENCED_PAGE;
1366 result = SCAN_SUCCEED;
1370 pte_unmap_unlock(pte, ptl);
1372 node = khugepaged_find_target_node();
1373 /* collapse_huge_page will return with the mmap_lock released */
1374 collapse_huge_page(mm, address, hpage, node,
1375 referenced, unmapped);
1378 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1379 none_or_zero, result, unmapped);
1383 static void collect_mm_slot(struct mm_slot *mm_slot)
1385 struct mm_struct *mm = mm_slot->mm;
1387 lockdep_assert_held(&khugepaged_mm_lock);
1389 if (khugepaged_test_exit(mm)) {
1391 hash_del(&mm_slot->hash);
1392 list_del(&mm_slot->mm_node);
1395 * Not strictly needed because the mm exited already.
1397 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1400 /* khugepaged_mm_lock actually not necessary for the below */
1401 free_mm_slot(mm_slot);
1408 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1409 * khugepaged should try to collapse the page table.
1411 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1414 struct mm_slot *mm_slot;
1416 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1418 spin_lock(&khugepaged_mm_lock);
1419 mm_slot = get_mm_slot(mm);
1420 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1421 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1422 spin_unlock(&khugepaged_mm_lock);
1427 * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1430 * @mm: process address space where collapse happens
1431 * @addr: THP collapse address
1433 * This function checks whether all the PTEs in the PMD are pointing to the
1434 * right THP. If so, retract the page table so the THP can refault in with
1437 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1439 unsigned long haddr = addr & HPAGE_PMD_MASK;
1440 struct vm_area_struct *vma = find_vma(mm, haddr);
1442 pte_t *start_pte, *pte;
1448 if (!vma || !vma->vm_file ||
1449 vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1453 * This vm_flags may not have VM_HUGEPAGE if the page was not
1454 * collapsed by this mm. But we can still collapse if the page is
1455 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1456 * will not fail the vma for missing VM_HUGEPAGE
1458 if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1461 hpage = find_lock_page(vma->vm_file->f_mapping,
1462 linear_page_index(vma, haddr));
1466 if (!PageHead(hpage))
1469 pmd = mm_find_pmd(mm, haddr);
1473 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1475 /* step 1: check all mapped PTEs are to the right huge page */
1476 for (i = 0, addr = haddr, pte = start_pte;
1477 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1480 /* empty pte, skip */
1484 /* page swapped out, abort */
1485 if (!pte_present(*pte))
1488 page = vm_normal_page(vma, addr, *pte);
1491 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1492 * page table, but the new page will not be a subpage of hpage.
1494 if (hpage + i != page)
1499 /* step 2: adjust rmap */
1500 for (i = 0, addr = haddr, pte = start_pte;
1501 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1506 page = vm_normal_page(vma, addr, *pte);
1507 page_remove_rmap(page, false);
1510 pte_unmap_unlock(start_pte, ptl);
1512 /* step 3: set proper refcount and mm_counters. */
1514 page_ref_sub(hpage, count);
1515 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1518 /* step 4: collapse pmd */
1519 ptl = pmd_lock(vma->vm_mm, pmd);
1520 _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1523 pte_free(mm, pmd_pgtable(_pmd));
1531 pte_unmap_unlock(start_pte, ptl);
1535 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1537 struct mm_struct *mm = mm_slot->mm;
1540 if (likely(mm_slot->nr_pte_mapped_thp == 0))
1543 if (!mmap_write_trylock(mm))
1546 if (unlikely(khugepaged_test_exit(mm)))
1549 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1550 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1553 mm_slot->nr_pte_mapped_thp = 0;
1554 mmap_write_unlock(mm);
1557 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1559 struct vm_area_struct *vma;
1560 struct mm_struct *mm;
1564 i_mmap_lock_write(mapping);
1565 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1567 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1568 * got written to. These VMAs are likely not worth investing
1569 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1572 * Not that vma->anon_vma check is racy: it can be set up after
1573 * the check but before we took mmap_lock by the fault path.
1574 * But page lock would prevent establishing any new ptes of the
1575 * page, so we are safe.
1577 * An alternative would be drop the check, but check that page
1578 * table is clear before calling pmdp_collapse_flush() under
1579 * ptl. It has higher chance to recover THP for the VMA, but
1580 * has higher cost too.
1584 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1585 if (addr & ~HPAGE_PMD_MASK)
1587 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1590 pmd = mm_find_pmd(mm, addr);
1594 * We need exclusive mmap_lock to retract page table.
1596 * We use trylock due to lock inversion: we need to acquire
1597 * mmap_lock while holding page lock. Fault path does it in
1598 * reverse order. Trylock is a way to avoid deadlock.
1600 if (mmap_write_trylock(mm)) {
1601 if (!khugepaged_test_exit(mm)) {
1602 spinlock_t *ptl = pmd_lock(mm, pmd);
1603 /* assume page table is clear */
1604 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1607 pte_free(mm, pmd_pgtable(_pmd));
1609 mmap_write_unlock(mm);
1611 /* Try again later */
1612 khugepaged_add_pte_mapped_thp(mm, addr);
1615 i_mmap_unlock_write(mapping);
1619 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1621 * @mm: process address space where collapse happens
1622 * @file: file that collapse on
1623 * @start: collapse start address
1624 * @hpage: new allocated huge page for collapse
1625 * @node: appointed node the new huge page allocate from
1627 * Basic scheme is simple, details are more complex:
1628 * - allocate and lock a new huge page;
1629 * - scan page cache replacing old pages with the new one
1630 * + swap/gup in pages if necessary;
1632 * + keep old pages around in case rollback is required;
1633 * - if replacing succeeds:
1636 * + unlock huge page;
1637 * - if replacing failed;
1638 * + put all pages back and unfreeze them;
1639 * + restore gaps in the page cache;
1640 * + unlock and free huge page;
1642 static void collapse_file(struct mm_struct *mm,
1643 struct file *file, pgoff_t start,
1644 struct page **hpage, int node)
1646 struct address_space *mapping = file->f_mapping;
1648 struct page *new_page;
1649 pgoff_t index, end = start + HPAGE_PMD_NR;
1650 LIST_HEAD(pagelist);
1651 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1652 int nr_none = 0, result = SCAN_SUCCEED;
1653 bool is_shmem = shmem_file(file);
1656 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1657 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1659 /* Only allocate from the target node */
1660 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1662 new_page = khugepaged_alloc_page(hpage, gfp, node);
1664 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1668 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1669 result = SCAN_CGROUP_CHARGE_FAIL;
1672 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1674 /* This will be less messy when we use multi-index entries */
1677 xas_create_range(&xas);
1678 if (!xas_error(&xas))
1680 xas_unlock_irq(&xas);
1681 if (!xas_nomem(&xas, GFP_KERNEL)) {
1687 __SetPageLocked(new_page);
1689 __SetPageSwapBacked(new_page);
1690 new_page->index = start;
1691 new_page->mapping = mapping;
1694 * At this point the new_page is locked and not up-to-date.
1695 * It's safe to insert it into the page cache, because nobody would
1696 * be able to map it or use it in another way until we unlock it.
1699 xas_set(&xas, start);
1700 for (index = start; index < end; index++) {
1701 struct page *page = xas_next(&xas);
1703 VM_BUG_ON(index != xas.xa_index);
1707 * Stop if extent has been truncated or
1708 * hole-punched, and is now completely
1711 if (index == start) {
1712 if (!xas_next_entry(&xas, end - 1)) {
1713 result = SCAN_TRUNCATED;
1716 xas_set(&xas, index);
1718 if (!shmem_charge(mapping->host, 1)) {
1722 xas_store(&xas, new_page);
1727 if (xa_is_value(page) || !PageUptodate(page)) {
1728 xas_unlock_irq(&xas);
1729 /* swap in or instantiate fallocated page */
1730 if (shmem_getpage(mapping->host, index, &page,
1735 } else if (trylock_page(page)) {
1737 xas_unlock_irq(&xas);
1739 result = SCAN_PAGE_LOCK;
1742 } else { /* !is_shmem */
1743 if (!page || xa_is_value(page)) {
1744 xas_unlock_irq(&xas);
1745 page_cache_sync_readahead(mapping, &file->f_ra,
1748 /* drain pagevecs to help isolate_lru_page() */
1750 page = find_lock_page(mapping, index);
1751 if (unlikely(page == NULL)) {
1755 } else if (PageDirty(page)) {
1757 * khugepaged only works on read-only fd,
1758 * so this page is dirty because it hasn't
1759 * been flushed since first write. There
1760 * won't be new dirty pages.
1762 * Trigger async flush here and hope the
1763 * writeback is done when khugepaged
1764 * revisits this page.
1766 * This is a one-off situation. We are not
1767 * forcing writeback in loop.
1769 xas_unlock_irq(&xas);
1770 filemap_flush(mapping);
1773 } else if (trylock_page(page)) {
1775 xas_unlock_irq(&xas);
1777 result = SCAN_PAGE_LOCK;
1783 * The page must be locked, so we can drop the i_pages lock
1784 * without racing with truncate.
1786 VM_BUG_ON_PAGE(!PageLocked(page), page);
1788 /* make sure the page is up to date */
1789 if (unlikely(!PageUptodate(page))) {
1795 * If file was truncated then extended, or hole-punched, before
1796 * we locked the first page, then a THP might be there already.
1798 if (PageTransCompound(page)) {
1799 result = SCAN_PAGE_COMPOUND;
1803 if (page_mapping(page) != mapping) {
1804 result = SCAN_TRUNCATED;
1808 if (!is_shmem && PageDirty(page)) {
1810 * khugepaged only works on read-only fd, so this
1811 * page is dirty because it hasn't been flushed
1812 * since first write.
1818 if (isolate_lru_page(page)) {
1819 result = SCAN_DEL_PAGE_LRU;
1823 if (page_has_private(page) &&
1824 !try_to_release_page(page, GFP_KERNEL)) {
1825 result = SCAN_PAGE_HAS_PRIVATE;
1826 putback_lru_page(page);
1830 if (page_mapped(page))
1831 unmap_mapping_pages(mapping, index, 1, false);
1834 xas_set(&xas, index);
1836 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1837 VM_BUG_ON_PAGE(page_mapped(page), page);
1840 * The page is expected to have page_count() == 3:
1841 * - we hold a pin on it;
1842 * - one reference from page cache;
1843 * - one from isolate_lru_page;
1845 if (!page_ref_freeze(page, 3)) {
1846 result = SCAN_PAGE_COUNT;
1847 xas_unlock_irq(&xas);
1848 putback_lru_page(page);
1853 * Add the page to the list to be able to undo the collapse if
1854 * something go wrong.
1856 list_add_tail(&page->lru, &pagelist);
1858 /* Finally, replace with the new page. */
1859 xas_store(&xas, new_page);
1866 nr = thp_nr_pages(new_page);
1869 __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
1871 __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
1872 filemap_nr_thps_inc(mapping);
1876 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1878 __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1882 xas_unlock_irq(&xas);
1885 if (result == SCAN_SUCCEED) {
1886 struct page *page, *tmp;
1889 * Replacing old pages with new one has succeeded, now we
1890 * need to copy the content and free the old pages.
1893 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1894 while (index < page->index) {
1895 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1898 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1900 list_del(&page->lru);
1901 page->mapping = NULL;
1902 page_ref_unfreeze(page, 1);
1903 ClearPageActive(page);
1904 ClearPageUnevictable(page);
1909 while (index < end) {
1910 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1914 SetPageUptodate(new_page);
1915 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1917 set_page_dirty(new_page);
1918 lru_cache_add(new_page);
1921 * Remove pte page tables, so we can re-fault the page as huge.
1923 retract_page_tables(mapping, start);
1926 khugepaged_pages_collapsed++;
1930 /* Something went wrong: roll back page cache changes */
1932 mapping->nrpages -= nr_none;
1935 shmem_uncharge(mapping->host, nr_none);
1937 xas_set(&xas, start);
1938 xas_for_each(&xas, page, end - 1) {
1939 page = list_first_entry_or_null(&pagelist,
1941 if (!page || xas.xa_index < page->index) {
1945 /* Put holes back where they were */
1946 xas_store(&xas, NULL);
1950 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1952 /* Unfreeze the page. */
1953 list_del(&page->lru);
1954 page_ref_unfreeze(page, 2);
1955 xas_store(&xas, page);
1957 xas_unlock_irq(&xas);
1959 putback_lru_page(page);
1963 xas_unlock_irq(&xas);
1965 new_page->mapping = NULL;
1968 unlock_page(new_page);
1970 VM_BUG_ON(!list_empty(&pagelist));
1971 if (!IS_ERR_OR_NULL(*hpage))
1972 mem_cgroup_uncharge(*hpage);
1973 /* TODO: tracepoints */
1976 static void khugepaged_scan_file(struct mm_struct *mm,
1977 struct file *file, pgoff_t start, struct page **hpage)
1979 struct page *page = NULL;
1980 struct address_space *mapping = file->f_mapping;
1981 XA_STATE(xas, &mapping->i_pages, start);
1983 int node = NUMA_NO_NODE;
1984 int result = SCAN_SUCCEED;
1988 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1990 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1991 if (xas_retry(&xas, page))
1994 if (xa_is_value(page)) {
1995 if (++swap > khugepaged_max_ptes_swap) {
1996 result = SCAN_EXCEED_SWAP_PTE;
2002 if (PageTransCompound(page)) {
2003 result = SCAN_PAGE_COMPOUND;
2007 node = page_to_nid(page);
2008 if (khugepaged_scan_abort(node)) {
2009 result = SCAN_SCAN_ABORT;
2012 khugepaged_node_load[node]++;
2014 if (!PageLRU(page)) {
2015 result = SCAN_PAGE_LRU;
2019 if (page_count(page) !=
2020 1 + page_mapcount(page) + page_has_private(page)) {
2021 result = SCAN_PAGE_COUNT;
2026 * We probably should check if the page is referenced here, but
2027 * nobody would transfer pte_young() to PageReferenced() for us.
2028 * And rmap walk here is just too costly...
2033 if (need_resched()) {
2040 if (result == SCAN_SUCCEED) {
2041 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2042 result = SCAN_EXCEED_NONE_PTE;
2044 node = khugepaged_find_target_node();
2045 collapse_file(mm, file, start, hpage, node);
2049 /* TODO: tracepoints */
2052 static void khugepaged_scan_file(struct mm_struct *mm,
2053 struct file *file, pgoff_t start, struct page **hpage)
2058 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2063 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2064 struct page **hpage)
2065 __releases(&khugepaged_mm_lock)
2066 __acquires(&khugepaged_mm_lock)
2068 struct mm_slot *mm_slot;
2069 struct mm_struct *mm;
2070 struct vm_area_struct *vma;
2074 lockdep_assert_held(&khugepaged_mm_lock);
2076 if (khugepaged_scan.mm_slot)
2077 mm_slot = khugepaged_scan.mm_slot;
2079 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2080 struct mm_slot, mm_node);
2081 khugepaged_scan.address = 0;
2082 khugepaged_scan.mm_slot = mm_slot;
2084 spin_unlock(&khugepaged_mm_lock);
2085 khugepaged_collapse_pte_mapped_thps(mm_slot);
2089 * Don't wait for semaphore (to avoid long wait times). Just move to
2090 * the next mm on the list.
2093 if (unlikely(!mmap_read_trylock(mm)))
2094 goto breakouterloop_mmap_lock;
2095 if (likely(!khugepaged_test_exit(mm)))
2096 vma = find_vma(mm, khugepaged_scan.address);
2099 for (; vma; vma = vma->vm_next) {
2100 unsigned long hstart, hend;
2103 if (unlikely(khugepaged_test_exit(mm))) {
2107 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2112 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2113 hend = vma->vm_end & HPAGE_PMD_MASK;
2116 if (khugepaged_scan.address > hend)
2118 if (khugepaged_scan.address < hstart)
2119 khugepaged_scan.address = hstart;
2120 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2121 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2124 while (khugepaged_scan.address < hend) {
2127 if (unlikely(khugepaged_test_exit(mm)))
2128 goto breakouterloop;
2130 VM_BUG_ON(khugepaged_scan.address < hstart ||
2131 khugepaged_scan.address + HPAGE_PMD_SIZE >
2133 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2134 struct file *file = get_file(vma->vm_file);
2135 pgoff_t pgoff = linear_page_index(vma,
2136 khugepaged_scan.address);
2138 mmap_read_unlock(mm);
2140 khugepaged_scan_file(mm, file, pgoff, hpage);
2143 ret = khugepaged_scan_pmd(mm, vma,
2144 khugepaged_scan.address,
2147 /* move to next address */
2148 khugepaged_scan.address += HPAGE_PMD_SIZE;
2149 progress += HPAGE_PMD_NR;
2151 /* we released mmap_lock so break loop */
2152 goto breakouterloop_mmap_lock;
2153 if (progress >= pages)
2154 goto breakouterloop;
2158 mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2159 breakouterloop_mmap_lock:
2161 spin_lock(&khugepaged_mm_lock);
2162 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2164 * Release the current mm_slot if this mm is about to die, or
2165 * if we scanned all vmas of this mm.
2167 if (khugepaged_test_exit(mm) || !vma) {
2169 * Make sure that if mm_users is reaching zero while
2170 * khugepaged runs here, khugepaged_exit will find
2171 * mm_slot not pointing to the exiting mm.
2173 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2174 khugepaged_scan.mm_slot = list_entry(
2175 mm_slot->mm_node.next,
2176 struct mm_slot, mm_node);
2177 khugepaged_scan.address = 0;
2179 khugepaged_scan.mm_slot = NULL;
2180 khugepaged_full_scans++;
2183 collect_mm_slot(mm_slot);
2189 static int khugepaged_has_work(void)
2191 return !list_empty(&khugepaged_scan.mm_head) &&
2192 khugepaged_enabled();
2195 static int khugepaged_wait_event(void)
2197 return !list_empty(&khugepaged_scan.mm_head) ||
2198 kthread_should_stop();
2201 static void khugepaged_do_scan(void)
2203 struct page *hpage = NULL;
2204 unsigned int progress = 0, pass_through_head = 0;
2205 unsigned int pages = khugepaged_pages_to_scan;
2208 barrier(); /* write khugepaged_pages_to_scan to local stack */
2210 lru_add_drain_all();
2212 while (progress < pages) {
2213 if (!khugepaged_prealloc_page(&hpage, &wait))
2218 if (unlikely(kthread_should_stop() || try_to_freeze()))
2221 spin_lock(&khugepaged_mm_lock);
2222 if (!khugepaged_scan.mm_slot)
2223 pass_through_head++;
2224 if (khugepaged_has_work() &&
2225 pass_through_head < 2)
2226 progress += khugepaged_scan_mm_slot(pages - progress,
2230 spin_unlock(&khugepaged_mm_lock);
2233 if (!IS_ERR_OR_NULL(hpage))
2237 static bool khugepaged_should_wakeup(void)
2239 return kthread_should_stop() ||
2240 time_after_eq(jiffies, khugepaged_sleep_expire);
2243 static void khugepaged_wait_work(void)
2245 if (khugepaged_has_work()) {
2246 const unsigned long scan_sleep_jiffies =
2247 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2249 if (!scan_sleep_jiffies)
2252 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2253 wait_event_freezable_timeout(khugepaged_wait,
2254 khugepaged_should_wakeup(),
2255 scan_sleep_jiffies);
2259 if (khugepaged_enabled())
2260 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2263 static int khugepaged(void *none)
2265 struct mm_slot *mm_slot;
2268 set_user_nice(current, MAX_NICE);
2270 while (!kthread_should_stop()) {
2271 khugepaged_do_scan();
2272 khugepaged_wait_work();
2275 spin_lock(&khugepaged_mm_lock);
2276 mm_slot = khugepaged_scan.mm_slot;
2277 khugepaged_scan.mm_slot = NULL;
2279 collect_mm_slot(mm_slot);
2280 spin_unlock(&khugepaged_mm_lock);
2284 static void set_recommended_min_free_kbytes(void)
2288 unsigned long recommended_min;
2290 for_each_populated_zone(zone) {
2292 * We don't need to worry about fragmentation of
2293 * ZONE_MOVABLE since it only has movable pages.
2295 if (zone_idx(zone) > gfp_zone(GFP_USER))
2301 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2302 recommended_min = pageblock_nr_pages * nr_zones * 2;
2305 * Make sure that on average at least two pageblocks are almost free
2306 * of another type, one for a migratetype to fall back to and a
2307 * second to avoid subsequent fallbacks of other types There are 3
2308 * MIGRATE_TYPES we care about.
2310 recommended_min += pageblock_nr_pages * nr_zones *
2311 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2313 /* don't ever allow to reserve more than 5% of the lowmem */
2314 recommended_min = min(recommended_min,
2315 (unsigned long) nr_free_buffer_pages() / 20);
2316 recommended_min <<= (PAGE_SHIFT-10);
2318 if (recommended_min > min_free_kbytes) {
2319 if (user_min_free_kbytes >= 0)
2320 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2321 min_free_kbytes, recommended_min);
2323 min_free_kbytes = recommended_min;
2325 setup_per_zone_wmarks();
2328 int start_stop_khugepaged(void)
2332 mutex_lock(&khugepaged_mutex);
2333 if (khugepaged_enabled()) {
2334 if (!khugepaged_thread)
2335 khugepaged_thread = kthread_run(khugepaged, NULL,
2337 if (IS_ERR(khugepaged_thread)) {
2338 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2339 err = PTR_ERR(khugepaged_thread);
2340 khugepaged_thread = NULL;
2344 if (!list_empty(&khugepaged_scan.mm_head))
2345 wake_up_interruptible(&khugepaged_wait);
2347 set_recommended_min_free_kbytes();
2348 } else if (khugepaged_thread) {
2349 kthread_stop(khugepaged_thread);
2350 khugepaged_thread = NULL;
2353 mutex_unlock(&khugepaged_mutex);
2357 void khugepaged_min_free_kbytes_update(void)
2359 mutex_lock(&khugepaged_mutex);
2360 if (khugepaged_enabled() && khugepaged_thread)
2361 set_recommended_min_free_kbytes();
2362 mutex_unlock(&khugepaged_mutex);