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 if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
446 (vm_flags & VM_NOHUGEPAGE) ||
447 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
450 if (shmem_file(vma->vm_file) ||
451 (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
453 (vm_flags & VM_DENYWRITE))) {
454 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
457 if (!vma->anon_vma || vma->vm_ops)
459 if (vma_is_temporary_stack(vma))
461 return !(vm_flags & VM_NO_KHUGEPAGED);
464 int __khugepaged_enter(struct mm_struct *mm)
466 struct mm_slot *mm_slot;
469 mm_slot = alloc_mm_slot();
473 /* __khugepaged_exit() must not run from under us */
474 VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
475 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
476 free_mm_slot(mm_slot);
480 spin_lock(&khugepaged_mm_lock);
481 insert_to_mm_slots_hash(mm, mm_slot);
483 * Insert just behind the scanning cursor, to let the area settle
486 wakeup = list_empty(&khugepaged_scan.mm_head);
487 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
488 spin_unlock(&khugepaged_mm_lock);
492 wake_up_interruptible(&khugepaged_wait);
497 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
498 unsigned long vm_flags)
500 unsigned long hstart, hend;
503 * khugepaged only supports read-only files for non-shmem files.
504 * khugepaged does not yet work on special mappings. And
505 * file-private shmem THP is not supported.
507 if (!hugepage_vma_check(vma, vm_flags))
510 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
511 hend = vma->vm_end & HPAGE_PMD_MASK;
513 return khugepaged_enter(vma, vm_flags);
517 void __khugepaged_exit(struct mm_struct *mm)
519 struct mm_slot *mm_slot;
522 spin_lock(&khugepaged_mm_lock);
523 mm_slot = get_mm_slot(mm);
524 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
525 hash_del(&mm_slot->hash);
526 list_del(&mm_slot->mm_node);
529 spin_unlock(&khugepaged_mm_lock);
532 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
533 free_mm_slot(mm_slot);
535 } else if (mm_slot) {
537 * This is required to serialize against
538 * khugepaged_test_exit() (which is guaranteed to run
539 * under mmap sem read mode). Stop here (after we
540 * return all pagetables will be destroyed) until
541 * khugepaged has finished working on the pagetables
542 * under the mmap_lock.
545 mmap_write_unlock(mm);
549 static void release_pte_page(struct page *page)
551 mod_node_page_state(page_pgdat(page),
552 NR_ISOLATED_ANON + page_is_file_lru(page),
555 putback_lru_page(page);
558 static void release_pte_pages(pte_t *pte, pte_t *_pte,
559 struct list_head *compound_pagelist)
561 struct page *page, *tmp;
563 while (--_pte >= pte) {
564 pte_t pteval = *_pte;
566 page = pte_page(pteval);
567 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
569 release_pte_page(page);
572 list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
573 list_del(&page->lru);
574 release_pte_page(page);
578 static bool is_refcount_suitable(struct page *page)
580 int expected_refcount;
582 expected_refcount = total_mapcount(page);
583 if (PageSwapCache(page))
584 expected_refcount += compound_nr(page);
586 return page_count(page) == expected_refcount;
589 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
590 unsigned long address,
592 struct list_head *compound_pagelist)
594 struct page *page = NULL;
596 int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
597 bool writable = false;
599 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
600 _pte++, address += PAGE_SIZE) {
601 pte_t pteval = *_pte;
602 if (pte_none(pteval) || (pte_present(pteval) &&
603 is_zero_pfn(pte_pfn(pteval)))) {
604 if (!userfaultfd_armed(vma) &&
605 ++none_or_zero <= khugepaged_max_ptes_none) {
608 result = SCAN_EXCEED_NONE_PTE;
612 if (!pte_present(pteval)) {
613 result = SCAN_PTE_NON_PRESENT;
616 page = vm_normal_page(vma, address, pteval);
617 if (unlikely(!page)) {
618 result = SCAN_PAGE_NULL;
622 VM_BUG_ON_PAGE(!PageAnon(page), page);
624 if (page_mapcount(page) > 1 &&
625 ++shared > khugepaged_max_ptes_shared) {
626 result = SCAN_EXCEED_SHARED_PTE;
630 if (PageCompound(page)) {
632 page = compound_head(page);
635 * Check if we have dealt with the compound page
638 list_for_each_entry(p, compound_pagelist, lru) {
645 * We can do it before isolate_lru_page because the
646 * page can't be freed from under us. NOTE: PG_lock
647 * is needed to serialize against split_huge_page
648 * when invoked from the VM.
650 if (!trylock_page(page)) {
651 result = SCAN_PAGE_LOCK;
656 * Check if the page has any GUP (or other external) pins.
658 * The page table that maps the page has been already unlinked
659 * from the page table tree and this process cannot get
660 * an additinal pin on the page.
662 * New pins can come later if the page is shared across fork,
663 * but not from this process. The other process cannot write to
664 * the page, only trigger CoW.
666 if (!is_refcount_suitable(page)) {
668 result = SCAN_PAGE_COUNT;
671 if (!pte_write(pteval) && PageSwapCache(page) &&
672 !reuse_swap_page(page, NULL)) {
674 * Page is in the swap cache and cannot be re-used.
675 * It cannot be collapsed into a THP.
678 result = SCAN_SWAP_CACHE_PAGE;
683 * Isolate the page to avoid collapsing an hugepage
684 * currently in use by the VM.
686 if (isolate_lru_page(page)) {
688 result = SCAN_DEL_PAGE_LRU;
691 mod_node_page_state(page_pgdat(page),
692 NR_ISOLATED_ANON + page_is_file_lru(page),
694 VM_BUG_ON_PAGE(!PageLocked(page), page);
695 VM_BUG_ON_PAGE(PageLRU(page), page);
697 if (PageCompound(page))
698 list_add_tail(&page->lru, compound_pagelist);
700 /* There should be enough young pte to collapse the page */
701 if (pte_young(pteval) ||
702 page_is_young(page) || PageReferenced(page) ||
703 mmu_notifier_test_young(vma->vm_mm, address))
706 if (pte_write(pteval))
709 if (likely(writable)) {
710 if (likely(referenced)) {
711 result = SCAN_SUCCEED;
712 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
713 referenced, writable, result);
717 result = SCAN_PAGE_RO;
721 release_pte_pages(pte, _pte, compound_pagelist);
722 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
723 referenced, writable, result);
727 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
728 struct vm_area_struct *vma,
729 unsigned long address,
731 struct list_head *compound_pagelist)
733 struct page *src_page, *tmp;
735 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
736 _pte++, page++, address += PAGE_SIZE) {
737 pte_t pteval = *_pte;
739 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
740 clear_user_highpage(page, address);
741 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
742 if (is_zero_pfn(pte_pfn(pteval))) {
744 * ptl mostly unnecessary.
748 * paravirt calls inside pte_clear here are
751 pte_clear(vma->vm_mm, address, _pte);
755 src_page = pte_page(pteval);
756 copy_user_highpage(page, src_page, address, vma);
757 if (!PageCompound(src_page))
758 release_pte_page(src_page);
760 * ptl mostly unnecessary, but preempt has to
761 * be disabled to update the per-cpu stats
762 * inside page_remove_rmap().
766 * paravirt calls inside pte_clear here are
769 pte_clear(vma->vm_mm, address, _pte);
770 page_remove_rmap(src_page, false);
772 free_page_and_swap_cache(src_page);
776 list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
777 list_del(&src_page->lru);
778 release_pte_page(src_page);
782 static void khugepaged_alloc_sleep(void)
786 add_wait_queue(&khugepaged_wait, &wait);
787 freezable_schedule_timeout_interruptible(
788 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
789 remove_wait_queue(&khugepaged_wait, &wait);
792 static int khugepaged_node_load[MAX_NUMNODES];
794 static bool khugepaged_scan_abort(int nid)
799 * If node_reclaim_mode is disabled, then no extra effort is made to
800 * allocate memory locally.
802 if (!node_reclaim_mode)
805 /* If there is a count for this node already, it must be acceptable */
806 if (khugepaged_node_load[nid])
809 for (i = 0; i < MAX_NUMNODES; i++) {
810 if (!khugepaged_node_load[i])
812 if (node_distance(nid, i) > node_reclaim_distance)
818 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
819 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
821 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
825 static int khugepaged_find_target_node(void)
827 static int last_khugepaged_target_node = NUMA_NO_NODE;
828 int nid, target_node = 0, max_value = 0;
830 /* find first node with max normal pages hit */
831 for (nid = 0; nid < MAX_NUMNODES; nid++)
832 if (khugepaged_node_load[nid] > max_value) {
833 max_value = khugepaged_node_load[nid];
837 /* do some balance if several nodes have the same hit record */
838 if (target_node <= last_khugepaged_target_node)
839 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
841 if (max_value == khugepaged_node_load[nid]) {
846 last_khugepaged_target_node = target_node;
850 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
852 if (IS_ERR(*hpage)) {
858 khugepaged_alloc_sleep();
868 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
870 VM_BUG_ON_PAGE(*hpage, *hpage);
872 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
873 if (unlikely(!*hpage)) {
874 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
875 *hpage = ERR_PTR(-ENOMEM);
879 prep_transhuge_page(*hpage);
880 count_vm_event(THP_COLLAPSE_ALLOC);
884 static int khugepaged_find_target_node(void)
889 static inline struct page *alloc_khugepaged_hugepage(void)
893 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
896 prep_transhuge_page(page);
900 static struct page *khugepaged_alloc_hugepage(bool *wait)
905 hpage = alloc_khugepaged_hugepage();
907 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
912 khugepaged_alloc_sleep();
914 count_vm_event(THP_COLLAPSE_ALLOC);
915 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
920 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
923 * If the hpage allocated earlier was briefly exposed in page cache
924 * before collapse_file() failed, it is possible that racing lookups
925 * have not yet completed, and would then be unpleasantly surprised by
926 * finding the hpage reused for the same mapping at a different offset.
927 * Just release the previous allocation if there is any danger of that.
929 if (*hpage && page_count(*hpage) > 1) {
935 *hpage = khugepaged_alloc_hugepage(wait);
937 if (unlikely(!*hpage))
944 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
953 * If mmap_lock temporarily dropped, revalidate vma
954 * before taking mmap_lock.
955 * Return 0 if succeeds, otherwise return none-zero
959 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
960 struct vm_area_struct **vmap)
962 struct vm_area_struct *vma;
963 unsigned long hstart, hend;
965 if (unlikely(khugepaged_test_exit(mm)))
966 return SCAN_ANY_PROCESS;
968 *vmap = vma = find_vma(mm, address);
970 return SCAN_VMA_NULL;
972 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
973 hend = vma->vm_end & HPAGE_PMD_MASK;
974 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
975 return SCAN_ADDRESS_RANGE;
976 if (!hugepage_vma_check(vma, vma->vm_flags))
977 return SCAN_VMA_CHECK;
978 /* Anon VMA expected */
979 if (!vma->anon_vma || vma->vm_ops)
980 return SCAN_VMA_CHECK;
985 * Bring missing pages in from swap, to complete THP collapse.
986 * Only done if khugepaged_scan_pmd believes it is worthwhile.
988 * Called and returns without pte mapped or spinlocks held,
989 * but with mmap_lock held to protect against vma changes.
992 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
993 struct vm_area_struct *vma,
994 unsigned long haddr, pmd_t *pmd,
999 unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
1001 for (address = haddr; address < end; address += PAGE_SIZE) {
1002 struct vm_fault vmf = {
1005 .pgoff = linear_page_index(vma, haddr),
1006 .flags = FAULT_FLAG_ALLOW_RETRY,
1010 vmf.pte = pte_offset_map(pmd, address);
1011 vmf.orig_pte = *vmf.pte;
1012 if (!is_swap_pte(vmf.orig_pte)) {
1017 ret = do_swap_page(&vmf);
1019 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1020 if (ret & VM_FAULT_RETRY) {
1022 if (hugepage_vma_revalidate(mm, haddr, &vma)) {
1023 /* vma is no longer available, don't continue to swapin */
1024 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1027 /* check if the pmd is still valid */
1028 if (mm_find_pmd(mm, haddr) != pmd) {
1029 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1033 if (ret & VM_FAULT_ERROR) {
1034 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1039 /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1043 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1047 static void collapse_huge_page(struct mm_struct *mm,
1048 unsigned long address,
1049 struct page **hpage,
1050 int node, int referenced, int unmapped)
1052 LIST_HEAD(compound_pagelist);
1056 struct page *new_page;
1057 spinlock_t *pmd_ptl, *pte_ptl;
1058 int isolated = 0, result = 0;
1059 struct vm_area_struct *vma;
1060 struct mmu_notifier_range range;
1063 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1065 /* Only allocate from the target node */
1066 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1069 * Before allocating the hugepage, release the mmap_lock read lock.
1070 * The allocation can take potentially a long time if it involves
1071 * sync compaction, and we do not need to hold the mmap_lock during
1072 * that. We will recheck the vma after taking it again in write mode.
1074 mmap_read_unlock(mm);
1075 new_page = khugepaged_alloc_page(hpage, gfp, node);
1077 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1081 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1082 result = SCAN_CGROUP_CHARGE_FAIL;
1085 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1088 result = hugepage_vma_revalidate(mm, address, &vma);
1090 mmap_read_unlock(mm);
1094 pmd = mm_find_pmd(mm, address);
1096 result = SCAN_PMD_NULL;
1097 mmap_read_unlock(mm);
1102 * __collapse_huge_page_swapin always returns with mmap_lock locked.
1103 * If it fails, we release mmap_lock and jump out_nolock.
1104 * Continuing to collapse causes inconsistency.
1106 if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1108 mmap_read_unlock(mm);
1112 mmap_read_unlock(mm);
1114 * Prevent all access to pagetables with the exception of
1115 * gup_fast later handled by the ptep_clear_flush and the VM
1116 * handled by the anon_vma lock + PG_lock.
1118 mmap_write_lock(mm);
1119 result = hugepage_vma_revalidate(mm, address, &vma);
1122 /* check if the pmd is still valid */
1123 if (mm_find_pmd(mm, address) != pmd)
1126 anon_vma_lock_write(vma->anon_vma);
1128 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1129 address, address + HPAGE_PMD_SIZE);
1130 mmu_notifier_invalidate_range_start(&range);
1132 pte = pte_offset_map(pmd, address);
1133 pte_ptl = pte_lockptr(mm, pmd);
1135 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1137 * After this gup_fast can't run anymore. This also removes
1138 * any huge TLB entry from the CPU so we won't allow
1139 * huge and small TLB entries for the same virtual address
1140 * to avoid the risk of CPU bugs in that area.
1142 _pmd = pmdp_collapse_flush(vma, address, pmd);
1143 spin_unlock(pmd_ptl);
1144 mmu_notifier_invalidate_range_end(&range);
1147 isolated = __collapse_huge_page_isolate(vma, address, pte,
1148 &compound_pagelist);
1149 spin_unlock(pte_ptl);
1151 if (unlikely(!isolated)) {
1154 BUG_ON(!pmd_none(*pmd));
1156 * We can only use set_pmd_at when establishing
1157 * hugepmds and never for establishing regular pmds that
1158 * points to regular pagetables. Use pmd_populate for that
1160 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1161 spin_unlock(pmd_ptl);
1162 anon_vma_unlock_write(vma->anon_vma);
1168 * All pages are isolated and locked so anon_vma rmap
1169 * can't run anymore.
1171 anon_vma_unlock_write(vma->anon_vma);
1173 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1174 &compound_pagelist);
1176 __SetPageUptodate(new_page);
1177 pgtable = pmd_pgtable(_pmd);
1179 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1180 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1183 * spin_lock() below is not the equivalent of smp_wmb(), so
1184 * this is needed to avoid the copy_huge_page writes to become
1185 * visible after the set_pmd_at() write.
1190 BUG_ON(!pmd_none(*pmd));
1191 page_add_new_anon_rmap(new_page, vma, address, true);
1192 lru_cache_add_inactive_or_unevictable(new_page, vma);
1193 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1194 set_pmd_at(mm, address, pmd, _pmd);
1195 update_mmu_cache_pmd(vma, address, pmd);
1196 spin_unlock(pmd_ptl);
1200 khugepaged_pages_collapsed++;
1201 result = SCAN_SUCCEED;
1203 mmap_write_unlock(mm);
1205 if (!IS_ERR_OR_NULL(*hpage))
1206 mem_cgroup_uncharge(*hpage);
1207 trace_mm_collapse_huge_page(mm, isolated, result);
1213 static int khugepaged_scan_pmd(struct mm_struct *mm,
1214 struct vm_area_struct *vma,
1215 unsigned long address,
1216 struct page **hpage)
1220 int ret = 0, result = 0, referenced = 0;
1221 int none_or_zero = 0, shared = 0;
1222 struct page *page = NULL;
1223 unsigned long _address;
1225 int node = NUMA_NO_NODE, unmapped = 0;
1226 bool writable = false;
1228 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1230 pmd = mm_find_pmd(mm, address);
1232 result = SCAN_PMD_NULL;
1236 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1237 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1238 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1239 _pte++, _address += PAGE_SIZE) {
1240 pte_t pteval = *_pte;
1241 if (is_swap_pte(pteval)) {
1242 if (++unmapped <= khugepaged_max_ptes_swap) {
1244 * Always be strict with uffd-wp
1245 * enabled swap entries. Please see
1246 * comment below for pte_uffd_wp().
1248 if (pte_swp_uffd_wp(pteval)) {
1249 result = SCAN_PTE_UFFD_WP;
1254 result = SCAN_EXCEED_SWAP_PTE;
1258 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1259 if (!userfaultfd_armed(vma) &&
1260 ++none_or_zero <= khugepaged_max_ptes_none) {
1263 result = SCAN_EXCEED_NONE_PTE;
1267 if (!pte_present(pteval)) {
1268 result = SCAN_PTE_NON_PRESENT;
1271 if (pte_uffd_wp(pteval)) {
1273 * Don't collapse the page if any of the small
1274 * PTEs are armed with uffd write protection.
1275 * Here we can also mark the new huge pmd as
1276 * write protected if any of the small ones is
1277 * marked but that could bring unknown
1278 * userfault messages that falls outside of
1279 * the registered range. So, just be simple.
1281 result = SCAN_PTE_UFFD_WP;
1284 if (pte_write(pteval))
1287 page = vm_normal_page(vma, _address, pteval);
1288 if (unlikely(!page)) {
1289 result = SCAN_PAGE_NULL;
1293 if (page_mapcount(page) > 1 &&
1294 ++shared > khugepaged_max_ptes_shared) {
1295 result = SCAN_EXCEED_SHARED_PTE;
1299 page = compound_head(page);
1302 * Record which node the original page is from and save this
1303 * information to khugepaged_node_load[].
1304 * Khupaged will allocate hugepage from the node has the max
1307 node = page_to_nid(page);
1308 if (khugepaged_scan_abort(node)) {
1309 result = SCAN_SCAN_ABORT;
1312 khugepaged_node_load[node]++;
1313 if (!PageLRU(page)) {
1314 result = SCAN_PAGE_LRU;
1317 if (PageLocked(page)) {
1318 result = SCAN_PAGE_LOCK;
1321 if (!PageAnon(page)) {
1322 result = SCAN_PAGE_ANON;
1327 * Check if the page has any GUP (or other external) pins.
1329 * Here the check is racy it may see totmal_mapcount > refcount
1331 * For example, one process with one forked child process.
1332 * The parent has the PMD split due to MADV_DONTNEED, then
1333 * the child is trying unmap the whole PMD, but khugepaged
1334 * may be scanning the parent between the child has
1335 * PageDoubleMap flag cleared and dec the mapcount. So
1336 * khugepaged may see total_mapcount > refcount.
1338 * But such case is ephemeral we could always retry collapse
1339 * later. However it may report false positive if the page
1340 * has excessive GUP pins (i.e. 512). Anyway the same check
1341 * will be done again later the risk seems low.
1343 if (!is_refcount_suitable(page)) {
1344 result = SCAN_PAGE_COUNT;
1347 if (pte_young(pteval) ||
1348 page_is_young(page) || PageReferenced(page) ||
1349 mmu_notifier_test_young(vma->vm_mm, address))
1353 result = SCAN_PAGE_RO;
1354 } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1355 result = SCAN_LACK_REFERENCED_PAGE;
1357 result = SCAN_SUCCEED;
1361 pte_unmap_unlock(pte, ptl);
1363 node = khugepaged_find_target_node();
1364 /* collapse_huge_page will return with the mmap_lock released */
1365 collapse_huge_page(mm, address, hpage, node,
1366 referenced, unmapped);
1369 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1370 none_or_zero, result, unmapped);
1374 static void collect_mm_slot(struct mm_slot *mm_slot)
1376 struct mm_struct *mm = mm_slot->mm;
1378 lockdep_assert_held(&khugepaged_mm_lock);
1380 if (khugepaged_test_exit(mm)) {
1382 hash_del(&mm_slot->hash);
1383 list_del(&mm_slot->mm_node);
1386 * Not strictly needed because the mm exited already.
1388 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1391 /* khugepaged_mm_lock actually not necessary for the below */
1392 free_mm_slot(mm_slot);
1399 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1400 * khugepaged should try to collapse the page table.
1402 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1405 struct mm_slot *mm_slot;
1407 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1409 spin_lock(&khugepaged_mm_lock);
1410 mm_slot = get_mm_slot(mm);
1411 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1412 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1413 spin_unlock(&khugepaged_mm_lock);
1418 * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1421 * @mm: process address space where collapse happens
1422 * @addr: THP collapse address
1424 * This function checks whether all the PTEs in the PMD are pointing to the
1425 * right THP. If so, retract the page table so the THP can refault in with
1428 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1430 unsigned long haddr = addr & HPAGE_PMD_MASK;
1431 struct vm_area_struct *vma = find_vma(mm, haddr);
1433 pte_t *start_pte, *pte;
1439 if (!vma || !vma->vm_file ||
1440 vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1444 * This vm_flags may not have VM_HUGEPAGE if the page was not
1445 * collapsed by this mm. But we can still collapse if the page is
1446 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1447 * will not fail the vma for missing VM_HUGEPAGE
1449 if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1452 hpage = find_lock_page(vma->vm_file->f_mapping,
1453 linear_page_index(vma, haddr));
1457 if (!PageHead(hpage))
1460 pmd = mm_find_pmd(mm, haddr);
1464 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1466 /* step 1: check all mapped PTEs are to the right huge page */
1467 for (i = 0, addr = haddr, pte = start_pte;
1468 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1471 /* empty pte, skip */
1475 /* page swapped out, abort */
1476 if (!pte_present(*pte))
1479 page = vm_normal_page(vma, addr, *pte);
1482 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1483 * page table, but the new page will not be a subpage of hpage.
1485 if (hpage + i != page)
1490 /* step 2: adjust rmap */
1491 for (i = 0, addr = haddr, pte = start_pte;
1492 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1497 page = vm_normal_page(vma, addr, *pte);
1498 page_remove_rmap(page, false);
1501 pte_unmap_unlock(start_pte, ptl);
1503 /* step 3: set proper refcount and mm_counters. */
1505 page_ref_sub(hpage, count);
1506 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1509 /* step 4: collapse pmd */
1510 ptl = pmd_lock(vma->vm_mm, pmd);
1511 _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1514 pte_free(mm, pmd_pgtable(_pmd));
1522 pte_unmap_unlock(start_pte, ptl);
1526 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1528 struct mm_struct *mm = mm_slot->mm;
1531 if (likely(mm_slot->nr_pte_mapped_thp == 0))
1534 if (!mmap_write_trylock(mm))
1537 if (unlikely(khugepaged_test_exit(mm)))
1540 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1541 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1544 mm_slot->nr_pte_mapped_thp = 0;
1545 mmap_write_unlock(mm);
1549 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1551 struct vm_area_struct *vma;
1552 struct mm_struct *mm;
1556 i_mmap_lock_write(mapping);
1557 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1559 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1560 * got written to. These VMAs are likely not worth investing
1561 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1564 * Not that vma->anon_vma check is racy: it can be set up after
1565 * the check but before we took mmap_lock by the fault path.
1566 * But page lock would prevent establishing any new ptes of the
1567 * page, so we are safe.
1569 * An alternative would be drop the check, but check that page
1570 * table is clear before calling pmdp_collapse_flush() under
1571 * ptl. It has higher chance to recover THP for the VMA, but
1572 * has higher cost too.
1576 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1577 if (addr & ~HPAGE_PMD_MASK)
1579 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1582 pmd = mm_find_pmd(mm, addr);
1586 * We need exclusive mmap_lock to retract page table.
1588 * We use trylock due to lock inversion: we need to acquire
1589 * mmap_lock while holding page lock. Fault path does it in
1590 * reverse order. Trylock is a way to avoid deadlock.
1592 if (mmap_write_trylock(mm)) {
1593 if (!khugepaged_test_exit(mm)) {
1594 spinlock_t *ptl = pmd_lock(mm, pmd);
1595 /* assume page table is clear */
1596 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1599 pte_free(mm, pmd_pgtable(_pmd));
1601 mmap_write_unlock(mm);
1603 /* Try again later */
1604 khugepaged_add_pte_mapped_thp(mm, addr);
1607 i_mmap_unlock_write(mapping);
1611 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1613 * @mm: process address space where collapse happens
1614 * @file: file that collapse on
1615 * @start: collapse start address
1616 * @hpage: new allocated huge page for collapse
1617 * @node: appointed node the new huge page allocate from
1619 * Basic scheme is simple, details are more complex:
1620 * - allocate and lock a new huge page;
1621 * - scan page cache replacing old pages with the new one
1622 * + swap/gup in pages if necessary;
1624 * + keep old pages around in case rollback is required;
1625 * - if replacing succeeds:
1628 * + unlock huge page;
1629 * - if replacing failed;
1630 * + put all pages back and unfreeze them;
1631 * + restore gaps in the page cache;
1632 * + unlock and free huge page;
1634 static void collapse_file(struct mm_struct *mm,
1635 struct file *file, pgoff_t start,
1636 struct page **hpage, int node)
1638 struct address_space *mapping = file->f_mapping;
1640 struct page *new_page;
1641 pgoff_t index, end = start + HPAGE_PMD_NR;
1642 LIST_HEAD(pagelist);
1643 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1644 int nr_none = 0, result = SCAN_SUCCEED;
1645 bool is_shmem = shmem_file(file);
1647 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1648 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1650 /* Only allocate from the target node */
1651 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1653 new_page = khugepaged_alloc_page(hpage, gfp, node);
1655 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1659 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1660 result = SCAN_CGROUP_CHARGE_FAIL;
1663 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1665 /* This will be less messy when we use multi-index entries */
1668 xas_create_range(&xas);
1669 if (!xas_error(&xas))
1671 xas_unlock_irq(&xas);
1672 if (!xas_nomem(&xas, GFP_KERNEL)) {
1678 __SetPageLocked(new_page);
1680 __SetPageSwapBacked(new_page);
1681 new_page->index = start;
1682 new_page->mapping = mapping;
1685 * At this point the new_page is locked and not up-to-date.
1686 * It's safe to insert it into the page cache, because nobody would
1687 * be able to map it or use it in another way until we unlock it.
1690 xas_set(&xas, start);
1691 for (index = start; index < end; index++) {
1692 struct page *page = xas_next(&xas);
1694 VM_BUG_ON(index != xas.xa_index);
1698 * Stop if extent has been truncated or
1699 * hole-punched, and is now completely
1702 if (index == start) {
1703 if (!xas_next_entry(&xas, end - 1)) {
1704 result = SCAN_TRUNCATED;
1707 xas_set(&xas, index);
1709 if (!shmem_charge(mapping->host, 1)) {
1713 xas_store(&xas, new_page);
1718 if (xa_is_value(page) || !PageUptodate(page)) {
1719 xas_unlock_irq(&xas);
1720 /* swap in or instantiate fallocated page */
1721 if (shmem_getpage(mapping->host, index, &page,
1726 } else if (trylock_page(page)) {
1728 xas_unlock_irq(&xas);
1730 result = SCAN_PAGE_LOCK;
1733 } else { /* !is_shmem */
1734 if (!page || xa_is_value(page)) {
1735 xas_unlock_irq(&xas);
1736 page_cache_sync_readahead(mapping, &file->f_ra,
1739 /* drain pagevecs to help isolate_lru_page() */
1741 page = find_lock_page(mapping, index);
1742 if (unlikely(page == NULL)) {
1746 } else if (PageDirty(page)) {
1748 * khugepaged only works on read-only fd,
1749 * so this page is dirty because it hasn't
1750 * been flushed since first write. There
1751 * won't be new dirty pages.
1753 * Trigger async flush here and hope the
1754 * writeback is done when khugepaged
1755 * revisits this page.
1757 * This is a one-off situation. We are not
1758 * forcing writeback in loop.
1760 xas_unlock_irq(&xas);
1761 filemap_flush(mapping);
1764 } else if (trylock_page(page)) {
1766 xas_unlock_irq(&xas);
1768 result = SCAN_PAGE_LOCK;
1774 * The page must be locked, so we can drop the i_pages lock
1775 * without racing with truncate.
1777 VM_BUG_ON_PAGE(!PageLocked(page), page);
1779 /* make sure the page is up to date */
1780 if (unlikely(!PageUptodate(page))) {
1786 * If file was truncated then extended, or hole-punched, before
1787 * we locked the first page, then a THP might be there already.
1789 if (PageTransCompound(page)) {
1790 result = SCAN_PAGE_COMPOUND;
1794 if (page_mapping(page) != mapping) {
1795 result = SCAN_TRUNCATED;
1799 if (!is_shmem && PageDirty(page)) {
1801 * khugepaged only works on read-only fd, so this
1802 * page is dirty because it hasn't been flushed
1803 * since first write.
1809 if (isolate_lru_page(page)) {
1810 result = SCAN_DEL_PAGE_LRU;
1814 if (page_has_private(page) &&
1815 !try_to_release_page(page, GFP_KERNEL)) {
1816 result = SCAN_PAGE_HAS_PRIVATE;
1817 putback_lru_page(page);
1821 if (page_mapped(page))
1822 unmap_mapping_pages(mapping, index, 1, false);
1825 xas_set(&xas, index);
1827 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1828 VM_BUG_ON_PAGE(page_mapped(page), page);
1831 * The page is expected to have page_count() == 3:
1832 * - we hold a pin on it;
1833 * - one reference from page cache;
1834 * - one from isolate_lru_page;
1836 if (!page_ref_freeze(page, 3)) {
1837 result = SCAN_PAGE_COUNT;
1838 xas_unlock_irq(&xas);
1839 putback_lru_page(page);
1844 * Add the page to the list to be able to undo the collapse if
1845 * something go wrong.
1847 list_add_tail(&page->lru, &pagelist);
1849 /* Finally, replace with the new page. */
1850 xas_store(&xas, new_page);
1859 __inc_lruvec_page_state(new_page, NR_SHMEM_THPS);
1861 __inc_lruvec_page_state(new_page, NR_FILE_THPS);
1862 filemap_nr_thps_inc(mapping);
1866 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1868 __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1872 xas_unlock_irq(&xas);
1875 if (result == SCAN_SUCCEED) {
1876 struct page *page, *tmp;
1879 * Replacing old pages with new one has succeeded, now we
1880 * need to copy the content and free the old pages.
1883 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1884 while (index < page->index) {
1885 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1888 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1890 list_del(&page->lru);
1891 page->mapping = NULL;
1892 page_ref_unfreeze(page, 1);
1893 ClearPageActive(page);
1894 ClearPageUnevictable(page);
1899 while (index < end) {
1900 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1904 SetPageUptodate(new_page);
1905 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1907 set_page_dirty(new_page);
1908 lru_cache_add(new_page);
1911 * Remove pte page tables, so we can re-fault the page as huge.
1913 retract_page_tables(mapping, start);
1916 khugepaged_pages_collapsed++;
1920 /* Something went wrong: roll back page cache changes */
1922 mapping->nrpages -= nr_none;
1925 shmem_uncharge(mapping->host, nr_none);
1927 xas_set(&xas, start);
1928 xas_for_each(&xas, page, end - 1) {
1929 page = list_first_entry_or_null(&pagelist,
1931 if (!page || xas.xa_index < page->index) {
1935 /* Put holes back where they were */
1936 xas_store(&xas, NULL);
1940 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1942 /* Unfreeze the page. */
1943 list_del(&page->lru);
1944 page_ref_unfreeze(page, 2);
1945 xas_store(&xas, page);
1947 xas_unlock_irq(&xas);
1949 putback_lru_page(page);
1953 xas_unlock_irq(&xas);
1955 new_page->mapping = NULL;
1958 unlock_page(new_page);
1960 VM_BUG_ON(!list_empty(&pagelist));
1961 if (!IS_ERR_OR_NULL(*hpage))
1962 mem_cgroup_uncharge(*hpage);
1963 /* TODO: tracepoints */
1966 static void khugepaged_scan_file(struct mm_struct *mm,
1967 struct file *file, pgoff_t start, struct page **hpage)
1969 struct page *page = NULL;
1970 struct address_space *mapping = file->f_mapping;
1971 XA_STATE(xas, &mapping->i_pages, start);
1973 int node = NUMA_NO_NODE;
1974 int result = SCAN_SUCCEED;
1978 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1980 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1981 if (xas_retry(&xas, page))
1984 if (xa_is_value(page)) {
1985 if (++swap > khugepaged_max_ptes_swap) {
1986 result = SCAN_EXCEED_SWAP_PTE;
1992 if (PageTransCompound(page)) {
1993 result = SCAN_PAGE_COMPOUND;
1997 node = page_to_nid(page);
1998 if (khugepaged_scan_abort(node)) {
1999 result = SCAN_SCAN_ABORT;
2002 khugepaged_node_load[node]++;
2004 if (!PageLRU(page)) {
2005 result = SCAN_PAGE_LRU;
2009 if (page_count(page) !=
2010 1 + page_mapcount(page) + page_has_private(page)) {
2011 result = SCAN_PAGE_COUNT;
2016 * We probably should check if the page is referenced here, but
2017 * nobody would transfer pte_young() to PageReferenced() for us.
2018 * And rmap walk here is just too costly...
2023 if (need_resched()) {
2030 if (result == SCAN_SUCCEED) {
2031 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2032 result = SCAN_EXCEED_NONE_PTE;
2034 node = khugepaged_find_target_node();
2035 collapse_file(mm, file, start, hpage, node);
2039 /* TODO: tracepoints */
2042 static void khugepaged_scan_file(struct mm_struct *mm,
2043 struct file *file, pgoff_t start, struct page **hpage)
2048 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2054 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2055 struct page **hpage)
2056 __releases(&khugepaged_mm_lock)
2057 __acquires(&khugepaged_mm_lock)
2059 struct mm_slot *mm_slot;
2060 struct mm_struct *mm;
2061 struct vm_area_struct *vma;
2065 lockdep_assert_held(&khugepaged_mm_lock);
2067 if (khugepaged_scan.mm_slot)
2068 mm_slot = khugepaged_scan.mm_slot;
2070 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2071 struct mm_slot, mm_node);
2072 khugepaged_scan.address = 0;
2073 khugepaged_scan.mm_slot = mm_slot;
2075 spin_unlock(&khugepaged_mm_lock);
2076 khugepaged_collapse_pte_mapped_thps(mm_slot);
2080 * Don't wait for semaphore (to avoid long wait times). Just move to
2081 * the next mm on the list.
2084 if (unlikely(!mmap_read_trylock(mm)))
2085 goto breakouterloop_mmap_lock;
2086 if (likely(!khugepaged_test_exit(mm)))
2087 vma = find_vma(mm, khugepaged_scan.address);
2090 for (; vma; vma = vma->vm_next) {
2091 unsigned long hstart, hend;
2094 if (unlikely(khugepaged_test_exit(mm))) {
2098 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2103 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2104 hend = vma->vm_end & HPAGE_PMD_MASK;
2107 if (khugepaged_scan.address > hend)
2109 if (khugepaged_scan.address < hstart)
2110 khugepaged_scan.address = hstart;
2111 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2112 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2115 while (khugepaged_scan.address < hend) {
2118 if (unlikely(khugepaged_test_exit(mm)))
2119 goto breakouterloop;
2121 VM_BUG_ON(khugepaged_scan.address < hstart ||
2122 khugepaged_scan.address + HPAGE_PMD_SIZE >
2124 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2125 struct file *file = get_file(vma->vm_file);
2126 pgoff_t pgoff = linear_page_index(vma,
2127 khugepaged_scan.address);
2129 mmap_read_unlock(mm);
2131 khugepaged_scan_file(mm, file, pgoff, hpage);
2134 ret = khugepaged_scan_pmd(mm, vma,
2135 khugepaged_scan.address,
2138 /* move to next address */
2139 khugepaged_scan.address += HPAGE_PMD_SIZE;
2140 progress += HPAGE_PMD_NR;
2142 /* we released mmap_lock so break loop */
2143 goto breakouterloop_mmap_lock;
2144 if (progress >= pages)
2145 goto breakouterloop;
2149 mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2150 breakouterloop_mmap_lock:
2152 spin_lock(&khugepaged_mm_lock);
2153 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2155 * Release the current mm_slot if this mm is about to die, or
2156 * if we scanned all vmas of this mm.
2158 if (khugepaged_test_exit(mm) || !vma) {
2160 * Make sure that if mm_users is reaching zero while
2161 * khugepaged runs here, khugepaged_exit will find
2162 * mm_slot not pointing to the exiting mm.
2164 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2165 khugepaged_scan.mm_slot = list_entry(
2166 mm_slot->mm_node.next,
2167 struct mm_slot, mm_node);
2168 khugepaged_scan.address = 0;
2170 khugepaged_scan.mm_slot = NULL;
2171 khugepaged_full_scans++;
2174 collect_mm_slot(mm_slot);
2180 static int khugepaged_has_work(void)
2182 return !list_empty(&khugepaged_scan.mm_head) &&
2183 khugepaged_enabled();
2186 static int khugepaged_wait_event(void)
2188 return !list_empty(&khugepaged_scan.mm_head) ||
2189 kthread_should_stop();
2192 static void khugepaged_do_scan(void)
2194 struct page *hpage = NULL;
2195 unsigned int progress = 0, pass_through_head = 0;
2196 unsigned int pages = khugepaged_pages_to_scan;
2199 barrier(); /* write khugepaged_pages_to_scan to local stack */
2201 lru_add_drain_all();
2203 while (progress < pages) {
2204 if (!khugepaged_prealloc_page(&hpage, &wait))
2209 if (unlikely(kthread_should_stop() || try_to_freeze()))
2212 spin_lock(&khugepaged_mm_lock);
2213 if (!khugepaged_scan.mm_slot)
2214 pass_through_head++;
2215 if (khugepaged_has_work() &&
2216 pass_through_head < 2)
2217 progress += khugepaged_scan_mm_slot(pages - progress,
2221 spin_unlock(&khugepaged_mm_lock);
2224 if (!IS_ERR_OR_NULL(hpage))
2228 static bool khugepaged_should_wakeup(void)
2230 return kthread_should_stop() ||
2231 time_after_eq(jiffies, khugepaged_sleep_expire);
2234 static void khugepaged_wait_work(void)
2236 if (khugepaged_has_work()) {
2237 const unsigned long scan_sleep_jiffies =
2238 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2240 if (!scan_sleep_jiffies)
2243 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2244 wait_event_freezable_timeout(khugepaged_wait,
2245 khugepaged_should_wakeup(),
2246 scan_sleep_jiffies);
2250 if (khugepaged_enabled())
2251 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2254 static int khugepaged(void *none)
2256 struct mm_slot *mm_slot;
2259 set_user_nice(current, MAX_NICE);
2261 while (!kthread_should_stop()) {
2262 khugepaged_do_scan();
2263 khugepaged_wait_work();
2266 spin_lock(&khugepaged_mm_lock);
2267 mm_slot = khugepaged_scan.mm_slot;
2268 khugepaged_scan.mm_slot = NULL;
2270 collect_mm_slot(mm_slot);
2271 spin_unlock(&khugepaged_mm_lock);
2275 static void set_recommended_min_free_kbytes(void)
2279 unsigned long recommended_min;
2281 for_each_populated_zone(zone) {
2283 * We don't need to worry about fragmentation of
2284 * ZONE_MOVABLE since it only has movable pages.
2286 if (zone_idx(zone) > gfp_zone(GFP_USER))
2292 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2293 recommended_min = pageblock_nr_pages * nr_zones * 2;
2296 * Make sure that on average at least two pageblocks are almost free
2297 * of another type, one for a migratetype to fall back to and a
2298 * second to avoid subsequent fallbacks of other types There are 3
2299 * MIGRATE_TYPES we care about.
2301 recommended_min += pageblock_nr_pages * nr_zones *
2302 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2304 /* don't ever allow to reserve more than 5% of the lowmem */
2305 recommended_min = min(recommended_min,
2306 (unsigned long) nr_free_buffer_pages() / 20);
2307 recommended_min <<= (PAGE_SHIFT-10);
2309 if (recommended_min > min_free_kbytes) {
2310 if (user_min_free_kbytes >= 0)
2311 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2312 min_free_kbytes, recommended_min);
2314 min_free_kbytes = recommended_min;
2316 setup_per_zone_wmarks();
2319 int start_stop_khugepaged(void)
2323 mutex_lock(&khugepaged_mutex);
2324 if (khugepaged_enabled()) {
2325 if (!khugepaged_thread)
2326 khugepaged_thread = kthread_run(khugepaged, NULL,
2328 if (IS_ERR(khugepaged_thread)) {
2329 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2330 err = PTR_ERR(khugepaged_thread);
2331 khugepaged_thread = NULL;
2335 if (!list_empty(&khugepaged_scan.mm_head))
2336 wake_up_interruptible(&khugepaged_wait);
2338 set_recommended_min_free_kbytes();
2339 } else if (khugepaged_thread) {
2340 kthread_stop(khugepaged_thread);
2341 khugepaged_thread = NULL;
2344 mutex_unlock(&khugepaged_mutex);
2348 void khugepaged_min_free_kbytes_update(void)
2350 mutex_lock(&khugepaged_mutex);
2351 if (khugepaged_enabled() && khugepaged_thread)
2352 set_recommended_min_free_kbytes();
2353 mutex_unlock(&khugepaged_mutex);