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 (!transhuge_vma_enabled(vma, vm_flags))
448 if (vma->vm_file && !IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) -
449 vma->vm_pgoff, HPAGE_PMD_NR))
452 /* Enabled via shmem mount options or sysfs settings. */
453 if (shmem_file(vma->vm_file))
454 return shmem_huge_enabled(vma);
456 /* THP settings require madvise. */
457 if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
460 /* Only regular file is valid */
461 if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
462 (vm_flags & VM_EXEC)) {
463 struct inode *inode = vma->vm_file->f_inode;
465 return !inode_is_open_for_write(inode) &&
466 S_ISREG(inode->i_mode);
469 if (!vma->anon_vma || vma->vm_ops)
471 if (vma_is_temporary_stack(vma))
473 return !(vm_flags & VM_NO_KHUGEPAGED);
476 int __khugepaged_enter(struct mm_struct *mm)
478 struct mm_slot *mm_slot;
481 mm_slot = alloc_mm_slot();
485 /* __khugepaged_exit() must not run from under us */
486 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
487 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
488 free_mm_slot(mm_slot);
492 spin_lock(&khugepaged_mm_lock);
493 insert_to_mm_slots_hash(mm, mm_slot);
495 * Insert just behind the scanning cursor, to let the area settle
498 wakeup = list_empty(&khugepaged_scan.mm_head);
499 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
500 spin_unlock(&khugepaged_mm_lock);
504 wake_up_interruptible(&khugepaged_wait);
509 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
510 unsigned long vm_flags)
512 unsigned long hstart, hend;
515 * khugepaged only supports read-only files for non-shmem files.
516 * khugepaged does not yet work on special mappings. And
517 * file-private shmem THP is not supported.
519 if (!hugepage_vma_check(vma, vm_flags))
522 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
523 hend = vma->vm_end & HPAGE_PMD_MASK;
525 return khugepaged_enter(vma, vm_flags);
529 void __khugepaged_exit(struct mm_struct *mm)
531 struct mm_slot *mm_slot;
534 spin_lock(&khugepaged_mm_lock);
535 mm_slot = get_mm_slot(mm);
536 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
537 hash_del(&mm_slot->hash);
538 list_del(&mm_slot->mm_node);
541 spin_unlock(&khugepaged_mm_lock);
544 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
545 free_mm_slot(mm_slot);
547 } else if (mm_slot) {
549 * This is required to serialize against
550 * khugepaged_test_exit() (which is guaranteed to run
551 * under mmap sem read mode). Stop here (after we
552 * return all pagetables will be destroyed) until
553 * khugepaged has finished working on the pagetables
554 * under the mmap_lock.
557 mmap_write_unlock(mm);
561 static void release_pte_page(struct page *page)
563 mod_node_page_state(page_pgdat(page),
564 NR_ISOLATED_ANON + page_is_file_lru(page),
567 putback_lru_page(page);
570 static void release_pte_pages(pte_t *pte, pte_t *_pte,
571 struct list_head *compound_pagelist)
573 struct page *page, *tmp;
575 while (--_pte >= pte) {
576 pte_t pteval = *_pte;
578 page = pte_page(pteval);
579 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
581 release_pte_page(page);
584 list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
585 list_del(&page->lru);
586 release_pte_page(page);
590 static bool is_refcount_suitable(struct page *page)
592 int expected_refcount;
594 expected_refcount = total_mapcount(page);
595 if (PageSwapCache(page))
596 expected_refcount += compound_nr(page);
598 return page_count(page) == expected_refcount;
601 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
602 unsigned long address,
604 struct list_head *compound_pagelist)
606 struct page *page = NULL;
608 int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
609 bool writable = false;
611 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
612 _pte++, address += PAGE_SIZE) {
613 pte_t pteval = *_pte;
614 if (pte_none(pteval) || (pte_present(pteval) &&
615 is_zero_pfn(pte_pfn(pteval)))) {
616 if (!userfaultfd_armed(vma) &&
617 ++none_or_zero <= khugepaged_max_ptes_none) {
620 result = SCAN_EXCEED_NONE_PTE;
624 if (!pte_present(pteval)) {
625 result = SCAN_PTE_NON_PRESENT;
628 page = vm_normal_page(vma, address, pteval);
629 if (unlikely(!page)) {
630 result = SCAN_PAGE_NULL;
634 VM_BUG_ON_PAGE(!PageAnon(page), page);
636 if (page_mapcount(page) > 1 &&
637 ++shared > khugepaged_max_ptes_shared) {
638 result = SCAN_EXCEED_SHARED_PTE;
642 if (PageCompound(page)) {
644 page = compound_head(page);
647 * Check if we have dealt with the compound page
650 list_for_each_entry(p, compound_pagelist, lru) {
657 * We can do it before isolate_lru_page because the
658 * page can't be freed from under us. NOTE: PG_lock
659 * is needed to serialize against split_huge_page
660 * when invoked from the VM.
662 if (!trylock_page(page)) {
663 result = SCAN_PAGE_LOCK;
668 * Check if the page has any GUP (or other external) pins.
670 * The page table that maps the page has been already unlinked
671 * from the page table tree and this process cannot get
672 * an additional pin on the page.
674 * New pins can come later if the page is shared across fork,
675 * but not from this process. The other process cannot write to
676 * the page, only trigger CoW.
678 if (!is_refcount_suitable(page)) {
680 result = SCAN_PAGE_COUNT;
683 if (!pte_write(pteval) && PageSwapCache(page) &&
684 !reuse_swap_page(page, NULL)) {
686 * Page is in the swap cache and cannot be re-used.
687 * It cannot be collapsed into a THP.
690 result = SCAN_SWAP_CACHE_PAGE;
695 * Isolate the page to avoid collapsing an hugepage
696 * currently in use by the VM.
698 if (isolate_lru_page(page)) {
700 result = SCAN_DEL_PAGE_LRU;
703 mod_node_page_state(page_pgdat(page),
704 NR_ISOLATED_ANON + page_is_file_lru(page),
706 VM_BUG_ON_PAGE(!PageLocked(page), page);
707 VM_BUG_ON_PAGE(PageLRU(page), page);
709 if (PageCompound(page))
710 list_add_tail(&page->lru, compound_pagelist);
712 /* There should be enough young pte to collapse the page */
713 if (pte_young(pteval) ||
714 page_is_young(page) || PageReferenced(page) ||
715 mmu_notifier_test_young(vma->vm_mm, address))
718 if (pte_write(pteval))
722 if (unlikely(!writable)) {
723 result = SCAN_PAGE_RO;
724 } else if (unlikely(!referenced)) {
725 result = SCAN_LACK_REFERENCED_PAGE;
727 result = SCAN_SUCCEED;
728 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
729 referenced, writable, result);
733 release_pte_pages(pte, _pte, compound_pagelist);
734 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
735 referenced, writable, result);
739 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
740 struct vm_area_struct *vma,
741 unsigned long address,
743 struct list_head *compound_pagelist)
745 struct page *src_page, *tmp;
747 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
748 _pte++, page++, address += PAGE_SIZE) {
749 pte_t pteval = *_pte;
751 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
752 clear_user_highpage(page, address);
753 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
754 if (is_zero_pfn(pte_pfn(pteval))) {
756 * ptl mostly unnecessary.
760 * paravirt calls inside pte_clear here are
763 pte_clear(vma->vm_mm, address, _pte);
767 src_page = pte_page(pteval);
768 copy_user_highpage(page, src_page, address, vma);
769 if (!PageCompound(src_page))
770 release_pte_page(src_page);
772 * ptl mostly unnecessary, but preempt has to
773 * be disabled to update the per-cpu stats
774 * inside page_remove_rmap().
778 * paravirt calls inside pte_clear here are
781 pte_clear(vma->vm_mm, address, _pte);
782 page_remove_rmap(src_page, false);
784 free_page_and_swap_cache(src_page);
788 list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
789 list_del(&src_page->lru);
790 release_pte_page(src_page);
794 static void khugepaged_alloc_sleep(void)
798 add_wait_queue(&khugepaged_wait, &wait);
799 freezable_schedule_timeout_interruptible(
800 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
801 remove_wait_queue(&khugepaged_wait, &wait);
804 static int khugepaged_node_load[MAX_NUMNODES];
806 static bool khugepaged_scan_abort(int nid)
811 * If node_reclaim_mode is disabled, then no extra effort is made to
812 * allocate memory locally.
814 if (!node_reclaim_enabled())
817 /* If there is a count for this node already, it must be acceptable */
818 if (khugepaged_node_load[nid])
821 for (i = 0; i < MAX_NUMNODES; i++) {
822 if (!khugepaged_node_load[i])
824 if (node_distance(nid, i) > node_reclaim_distance)
830 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
831 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
833 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
837 static int khugepaged_find_target_node(void)
839 static int last_khugepaged_target_node = NUMA_NO_NODE;
840 int nid, target_node = 0, max_value = 0;
842 /* find first node with max normal pages hit */
843 for (nid = 0; nid < MAX_NUMNODES; nid++)
844 if (khugepaged_node_load[nid] > max_value) {
845 max_value = khugepaged_node_load[nid];
849 /* do some balance if several nodes have the same hit record */
850 if (target_node <= last_khugepaged_target_node)
851 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
853 if (max_value == khugepaged_node_load[nid]) {
858 last_khugepaged_target_node = target_node;
862 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
864 if (IS_ERR(*hpage)) {
870 khugepaged_alloc_sleep();
880 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
882 VM_BUG_ON_PAGE(*hpage, *hpage);
884 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
885 if (unlikely(!*hpage)) {
886 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
887 *hpage = ERR_PTR(-ENOMEM);
891 prep_transhuge_page(*hpage);
892 count_vm_event(THP_COLLAPSE_ALLOC);
896 static int khugepaged_find_target_node(void)
901 static inline struct page *alloc_khugepaged_hugepage(void)
905 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
908 prep_transhuge_page(page);
912 static struct page *khugepaged_alloc_hugepage(bool *wait)
917 hpage = alloc_khugepaged_hugepage();
919 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
924 khugepaged_alloc_sleep();
926 count_vm_event(THP_COLLAPSE_ALLOC);
927 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
932 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
935 * If the hpage allocated earlier was briefly exposed in page cache
936 * before collapse_file() failed, it is possible that racing lookups
937 * have not yet completed, and would then be unpleasantly surprised by
938 * finding the hpage reused for the same mapping at a different offset.
939 * Just release the previous allocation if there is any danger of that.
941 if (*hpage && page_count(*hpage) > 1) {
947 *hpage = khugepaged_alloc_hugepage(wait);
949 if (unlikely(!*hpage))
956 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
965 * If mmap_lock temporarily dropped, revalidate vma
966 * before taking mmap_lock.
967 * Return 0 if succeeds, otherwise return none-zero
971 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
972 struct vm_area_struct **vmap)
974 struct vm_area_struct *vma;
975 unsigned long hstart, hend;
977 if (unlikely(khugepaged_test_exit(mm)))
978 return SCAN_ANY_PROCESS;
980 *vmap = vma = find_vma(mm, address);
982 return SCAN_VMA_NULL;
984 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
985 hend = vma->vm_end & HPAGE_PMD_MASK;
986 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
987 return SCAN_ADDRESS_RANGE;
988 if (!hugepage_vma_check(vma, vma->vm_flags))
989 return SCAN_VMA_CHECK;
990 /* Anon VMA expected */
991 if (!vma->anon_vma || vma->vm_ops)
992 return SCAN_VMA_CHECK;
997 * Bring missing pages in from swap, to complete THP collapse.
998 * Only done if khugepaged_scan_pmd believes it is worthwhile.
1000 * Called and returns without pte mapped or spinlocks held,
1001 * but with mmap_lock held to protect against vma changes.
1004 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
1005 struct vm_area_struct *vma,
1006 unsigned long haddr, pmd_t *pmd,
1011 unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
1013 for (address = haddr; address < end; address += PAGE_SIZE) {
1014 struct vm_fault vmf = {
1017 .pgoff = linear_page_index(vma, haddr),
1018 .flags = FAULT_FLAG_ALLOW_RETRY,
1022 vmf.pte = pte_offset_map(pmd, address);
1023 vmf.orig_pte = *vmf.pte;
1024 if (!is_swap_pte(vmf.orig_pte)) {
1029 ret = do_swap_page(&vmf);
1031 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1032 if (ret & VM_FAULT_RETRY) {
1034 if (hugepage_vma_revalidate(mm, haddr, &vma)) {
1035 /* vma is no longer available, don't continue to swapin */
1036 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1039 /* check if the pmd is still valid */
1040 if (mm_find_pmd(mm, haddr) != pmd) {
1041 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1045 if (ret & VM_FAULT_ERROR) {
1046 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1051 /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1055 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1059 static void collapse_huge_page(struct mm_struct *mm,
1060 unsigned long address,
1061 struct page **hpage,
1062 int node, int referenced, int unmapped)
1064 LIST_HEAD(compound_pagelist);
1068 struct page *new_page;
1069 spinlock_t *pmd_ptl, *pte_ptl;
1070 int isolated = 0, result = 0;
1071 struct vm_area_struct *vma;
1072 struct mmu_notifier_range range;
1075 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1077 /* Only allocate from the target node */
1078 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1081 * Before allocating the hugepage, release the mmap_lock read lock.
1082 * The allocation can take potentially a long time if it involves
1083 * sync compaction, and we do not need to hold the mmap_lock during
1084 * that. We will recheck the vma after taking it again in write mode.
1086 mmap_read_unlock(mm);
1087 new_page = khugepaged_alloc_page(hpage, gfp, node);
1089 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1093 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1094 result = SCAN_CGROUP_CHARGE_FAIL;
1097 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1100 result = hugepage_vma_revalidate(mm, address, &vma);
1102 mmap_read_unlock(mm);
1106 pmd = mm_find_pmd(mm, address);
1108 result = SCAN_PMD_NULL;
1109 mmap_read_unlock(mm);
1114 * __collapse_huge_page_swapin always returns with mmap_lock locked.
1115 * If it fails, we release mmap_lock and jump out_nolock.
1116 * Continuing to collapse causes inconsistency.
1118 if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1120 mmap_read_unlock(mm);
1124 mmap_read_unlock(mm);
1126 * Prevent all access to pagetables with the exception of
1127 * gup_fast later handled by the ptep_clear_flush and the VM
1128 * handled by the anon_vma lock + PG_lock.
1130 mmap_write_lock(mm);
1131 result = hugepage_vma_revalidate(mm, address, &vma);
1134 /* check if the pmd is still valid */
1135 if (mm_find_pmd(mm, address) != pmd)
1138 anon_vma_lock_write(vma->anon_vma);
1140 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1141 address, address + HPAGE_PMD_SIZE);
1142 mmu_notifier_invalidate_range_start(&range);
1144 pte = pte_offset_map(pmd, address);
1145 pte_ptl = pte_lockptr(mm, pmd);
1147 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1149 * This removes any huge TLB entry from the CPU so we won't allow
1150 * huge and small TLB entries for the same virtual address to
1151 * avoid the risk of CPU bugs in that area.
1153 * Parallel fast GUP is fine since fast GUP will back off when
1154 * it detects PMD is changed.
1156 _pmd = pmdp_collapse_flush(vma, address, pmd);
1157 spin_unlock(pmd_ptl);
1158 mmu_notifier_invalidate_range_end(&range);
1159 tlb_remove_table_sync_one();
1162 isolated = __collapse_huge_page_isolate(vma, address, pte,
1163 &compound_pagelist);
1164 spin_unlock(pte_ptl);
1166 if (unlikely(!isolated)) {
1169 BUG_ON(!pmd_none(*pmd));
1171 * We can only use set_pmd_at when establishing
1172 * hugepmds and never for establishing regular pmds that
1173 * points to regular pagetables. Use pmd_populate for that
1175 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1176 spin_unlock(pmd_ptl);
1177 anon_vma_unlock_write(vma->anon_vma);
1183 * All pages are isolated and locked so anon_vma rmap
1184 * can't run anymore.
1186 anon_vma_unlock_write(vma->anon_vma);
1188 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1189 &compound_pagelist);
1192 * spin_lock() below is not the equivalent of smp_wmb(), but
1193 * the smp_wmb() inside __SetPageUptodate() can be reused to
1194 * avoid the copy_huge_page writes to become visible after
1195 * the set_pmd_at() write.
1197 __SetPageUptodate(new_page);
1198 pgtable = pmd_pgtable(_pmd);
1200 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1201 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1204 BUG_ON(!pmd_none(*pmd));
1205 page_add_new_anon_rmap(new_page, vma, address, true);
1206 lru_cache_add_inactive_or_unevictable(new_page, vma);
1207 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1208 set_pmd_at(mm, address, pmd, _pmd);
1209 update_mmu_cache_pmd(vma, address, pmd);
1210 spin_unlock(pmd_ptl);
1214 khugepaged_pages_collapsed++;
1215 result = SCAN_SUCCEED;
1217 mmap_write_unlock(mm);
1219 if (!IS_ERR_OR_NULL(*hpage))
1220 mem_cgroup_uncharge(*hpage);
1221 trace_mm_collapse_huge_page(mm, isolated, result);
1225 static int khugepaged_scan_pmd(struct mm_struct *mm,
1226 struct vm_area_struct *vma,
1227 unsigned long address,
1228 struct page **hpage)
1232 int ret = 0, result = 0, referenced = 0;
1233 int none_or_zero = 0, shared = 0;
1234 struct page *page = NULL;
1235 unsigned long _address;
1237 int node = NUMA_NO_NODE, unmapped = 0;
1238 bool writable = false;
1240 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1242 pmd = mm_find_pmd(mm, address);
1244 result = SCAN_PMD_NULL;
1248 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1249 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1250 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1251 _pte++, _address += PAGE_SIZE) {
1252 pte_t pteval = *_pte;
1253 if (is_swap_pte(pteval)) {
1254 if (++unmapped <= khugepaged_max_ptes_swap) {
1256 * Always be strict with uffd-wp
1257 * enabled swap entries. Please see
1258 * comment below for pte_uffd_wp().
1260 if (pte_swp_uffd_wp(pteval)) {
1261 result = SCAN_PTE_UFFD_WP;
1266 result = SCAN_EXCEED_SWAP_PTE;
1270 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1271 if (!userfaultfd_armed(vma) &&
1272 ++none_or_zero <= khugepaged_max_ptes_none) {
1275 result = SCAN_EXCEED_NONE_PTE;
1279 if (pte_uffd_wp(pteval)) {
1281 * Don't collapse the page if any of the small
1282 * PTEs are armed with uffd write protection.
1283 * Here we can also mark the new huge pmd as
1284 * write protected if any of the small ones is
1285 * marked but that could bring unknown
1286 * userfault messages that falls outside of
1287 * the registered range. So, just be simple.
1289 result = SCAN_PTE_UFFD_WP;
1292 if (pte_write(pteval))
1295 page = vm_normal_page(vma, _address, pteval);
1296 if (unlikely(!page)) {
1297 result = SCAN_PAGE_NULL;
1301 if (page_mapcount(page) > 1 &&
1302 ++shared > khugepaged_max_ptes_shared) {
1303 result = SCAN_EXCEED_SHARED_PTE;
1307 page = compound_head(page);
1310 * Record which node the original page is from and save this
1311 * information to khugepaged_node_load[].
1312 * Khupaged will allocate hugepage from the node has the max
1315 node = page_to_nid(page);
1316 if (khugepaged_scan_abort(node)) {
1317 result = SCAN_SCAN_ABORT;
1320 khugepaged_node_load[node]++;
1321 if (!PageLRU(page)) {
1322 result = SCAN_PAGE_LRU;
1325 if (PageLocked(page)) {
1326 result = SCAN_PAGE_LOCK;
1329 if (!PageAnon(page)) {
1330 result = SCAN_PAGE_ANON;
1335 * Check if the page has any GUP (or other external) pins.
1337 * Here the check is racy it may see totmal_mapcount > refcount
1339 * For example, one process with one forked child process.
1340 * The parent has the PMD split due to MADV_DONTNEED, then
1341 * the child is trying unmap the whole PMD, but khugepaged
1342 * may be scanning the parent between the child has
1343 * PageDoubleMap flag cleared and dec the mapcount. So
1344 * khugepaged may see total_mapcount > refcount.
1346 * But such case is ephemeral we could always retry collapse
1347 * later. However it may report false positive if the page
1348 * has excessive GUP pins (i.e. 512). Anyway the same check
1349 * will be done again later the risk seems low.
1351 if (!is_refcount_suitable(page)) {
1352 result = SCAN_PAGE_COUNT;
1355 if (pte_young(pteval) ||
1356 page_is_young(page) || PageReferenced(page) ||
1357 mmu_notifier_test_young(vma->vm_mm, address))
1361 result = SCAN_PAGE_RO;
1362 } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1363 result = SCAN_LACK_REFERENCED_PAGE;
1365 result = SCAN_SUCCEED;
1369 pte_unmap_unlock(pte, ptl);
1371 node = khugepaged_find_target_node();
1372 /* collapse_huge_page will return with the mmap_lock released */
1373 collapse_huge_page(mm, address, hpage, node,
1374 referenced, unmapped);
1377 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1378 none_or_zero, result, unmapped);
1382 static void collect_mm_slot(struct mm_slot *mm_slot)
1384 struct mm_struct *mm = mm_slot->mm;
1386 lockdep_assert_held(&khugepaged_mm_lock);
1388 if (khugepaged_test_exit(mm)) {
1390 hash_del(&mm_slot->hash);
1391 list_del(&mm_slot->mm_node);
1394 * Not strictly needed because the mm exited already.
1396 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1399 /* khugepaged_mm_lock actually not necessary for the below */
1400 free_mm_slot(mm_slot);
1407 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1408 * khugepaged should try to collapse the page table.
1410 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1413 struct mm_slot *mm_slot;
1415 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1417 spin_lock(&khugepaged_mm_lock);
1418 mm_slot = get_mm_slot(mm);
1419 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1420 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1421 spin_unlock(&khugepaged_mm_lock);
1426 * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1429 * @mm: process address space where collapse happens
1430 * @addr: THP collapse address
1432 * This function checks whether all the PTEs in the PMD are pointing to the
1433 * right THP. If so, retract the page table so the THP can refault in with
1436 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1438 unsigned long haddr = addr & HPAGE_PMD_MASK;
1439 struct vm_area_struct *vma = find_vma(mm, haddr);
1441 pte_t *start_pte, *pte;
1446 struct mmu_notifier_range range;
1448 if (!vma || !vma->vm_file ||
1449 !range_in_vma(vma, haddr, 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);
1474 * We need to lock the mapping so that from here on, only GUP-fast and
1475 * hardware page walks can access the parts of the page tables that
1476 * we're operating on.
1478 i_mmap_lock_write(vma->vm_file->f_mapping);
1481 * This spinlock should be unnecessary: Nobody else should be accessing
1482 * the page tables under spinlock protection here, only
1483 * lockless_pages_from_mm() and the hardware page walker can access page
1484 * tables while all the high-level locks are held in write mode.
1486 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1488 /* step 1: check all mapped PTEs are to the right huge page */
1489 for (i = 0, addr = haddr, pte = start_pte;
1490 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1493 /* empty pte, skip */
1497 /* page swapped out, abort */
1498 if (!pte_present(*pte))
1501 page = vm_normal_page(vma, addr, *pte);
1504 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1505 * page table, but the new page will not be a subpage of hpage.
1507 if (hpage + i != page)
1512 /* step 2: adjust rmap */
1513 for (i = 0, addr = haddr, pte = start_pte;
1514 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1519 page = vm_normal_page(vma, addr, *pte);
1520 page_remove_rmap(page, false);
1523 pte_unmap_unlock(start_pte, ptl);
1525 /* step 3: set proper refcount and mm_counters. */
1527 page_ref_sub(hpage, count);
1528 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1531 /* step 4: collapse pmd */
1532 /* we make no change to anon, but protect concurrent anon page lookup */
1534 anon_vma_lock_write(vma->anon_vma);
1536 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, haddr,
1537 haddr + HPAGE_PMD_SIZE);
1538 mmu_notifier_invalidate_range_start(&range);
1539 _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1541 tlb_remove_table_sync_one();
1542 mmu_notifier_invalidate_range_end(&range);
1543 pte_free(mm, pmd_pgtable(_pmd));
1546 anon_vma_unlock_write(vma->anon_vma);
1547 i_mmap_unlock_write(vma->vm_file->f_mapping);
1555 pte_unmap_unlock(start_pte, ptl);
1556 i_mmap_unlock_write(vma->vm_file->f_mapping);
1560 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1562 struct mm_struct *mm = mm_slot->mm;
1565 if (likely(mm_slot->nr_pte_mapped_thp == 0))
1568 if (!mmap_write_trylock(mm))
1571 if (unlikely(khugepaged_test_exit(mm)))
1574 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1575 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1578 mm_slot->nr_pte_mapped_thp = 0;
1579 mmap_write_unlock(mm);
1582 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1584 struct vm_area_struct *vma;
1585 struct mm_struct *mm;
1589 i_mmap_lock_write(mapping);
1590 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1592 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1593 * got written to. These VMAs are likely not worth investing
1594 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1597 * Not that vma->anon_vma check is racy: it can be set up after
1598 * the check but before we took mmap_lock by the fault path.
1599 * But page lock would prevent establishing any new ptes of the
1600 * page, so we are safe.
1602 * An alternative would be drop the check, but check that page
1603 * table is clear before calling pmdp_collapse_flush() under
1604 * ptl. It has higher chance to recover THP for the VMA, but
1605 * has higher cost too. It would also probably require locking
1610 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1611 if (addr & ~HPAGE_PMD_MASK)
1613 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1616 pmd = mm_find_pmd(mm, addr);
1620 * We need exclusive mmap_lock to retract page table.
1622 * We use trylock due to lock inversion: we need to acquire
1623 * mmap_lock while holding page lock. Fault path does it in
1624 * reverse order. Trylock is a way to avoid deadlock.
1626 if (mmap_write_trylock(mm)) {
1627 if (!khugepaged_test_exit(mm)) {
1628 struct mmu_notifier_range range;
1630 mmu_notifier_range_init(&range,
1631 MMU_NOTIFY_CLEAR, 0,
1633 addr + HPAGE_PMD_SIZE);
1634 mmu_notifier_invalidate_range_start(&range);
1635 /* assume page table is clear */
1636 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1638 tlb_remove_table_sync_one();
1639 pte_free(mm, pmd_pgtable(_pmd));
1640 mmu_notifier_invalidate_range_end(&range);
1642 mmap_write_unlock(mm);
1644 /* Try again later */
1645 khugepaged_add_pte_mapped_thp(mm, addr);
1648 i_mmap_unlock_write(mapping);
1652 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1654 * @mm: process address space where collapse happens
1655 * @file: file that collapse on
1656 * @start: collapse start address
1657 * @hpage: new allocated huge page for collapse
1658 * @node: appointed node the new huge page allocate from
1660 * Basic scheme is simple, details are more complex:
1661 * - allocate and lock a new huge page;
1662 * - scan page cache replacing old pages with the new one
1663 * + swap/gup in pages if necessary;
1665 * + keep old pages around in case rollback is required;
1666 * - if replacing succeeds:
1669 * + unlock huge page;
1670 * - if replacing failed;
1671 * + put all pages back and unfreeze them;
1672 * + restore gaps in the page cache;
1673 * + unlock and free huge page;
1675 static void collapse_file(struct mm_struct *mm,
1676 struct file *file, pgoff_t start,
1677 struct page **hpage, int node)
1679 struct address_space *mapping = file->f_mapping;
1681 struct page *new_page;
1682 pgoff_t index, end = start + HPAGE_PMD_NR;
1683 LIST_HEAD(pagelist);
1684 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1685 int nr_none = 0, result = SCAN_SUCCEED;
1686 bool is_shmem = shmem_file(file);
1689 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1690 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1692 /* Only allocate from the target node */
1693 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1695 new_page = khugepaged_alloc_page(hpage, gfp, node);
1697 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1701 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1702 result = SCAN_CGROUP_CHARGE_FAIL;
1705 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1707 /* This will be less messy when we use multi-index entries */
1710 xas_create_range(&xas);
1711 if (!xas_error(&xas))
1713 xas_unlock_irq(&xas);
1714 if (!xas_nomem(&xas, GFP_KERNEL)) {
1720 __SetPageLocked(new_page);
1722 __SetPageSwapBacked(new_page);
1723 new_page->index = start;
1724 new_page->mapping = mapping;
1727 * At this point the new_page is locked and not up-to-date.
1728 * It's safe to insert it into the page cache, because nobody would
1729 * be able to map it or use it in another way until we unlock it.
1732 xas_set(&xas, start);
1733 for (index = start; index < end; index++) {
1734 struct page *page = xas_next(&xas);
1736 VM_BUG_ON(index != xas.xa_index);
1740 * Stop if extent has been truncated or
1741 * hole-punched, and is now completely
1744 if (index == start) {
1745 if (!xas_next_entry(&xas, end - 1)) {
1746 result = SCAN_TRUNCATED;
1749 xas_set(&xas, index);
1751 if (!shmem_charge(mapping->host, 1)) {
1755 xas_store(&xas, new_page);
1760 if (xa_is_value(page) || !PageUptodate(page)) {
1761 xas_unlock_irq(&xas);
1762 /* swap in or instantiate fallocated page */
1763 if (shmem_getpage(mapping->host, index, &page,
1768 } else if (trylock_page(page)) {
1770 xas_unlock_irq(&xas);
1772 result = SCAN_PAGE_LOCK;
1775 } else { /* !is_shmem */
1776 if (!page || xa_is_value(page)) {
1777 xas_unlock_irq(&xas);
1778 page_cache_sync_readahead(mapping, &file->f_ra,
1781 /* drain pagevecs to help isolate_lru_page() */
1783 page = find_lock_page(mapping, index);
1784 if (unlikely(page == NULL)) {
1788 } else if (PageDirty(page)) {
1790 * khugepaged only works on read-only fd,
1791 * so this page is dirty because it hasn't
1792 * been flushed since first write. There
1793 * won't be new dirty pages.
1795 * Trigger async flush here and hope the
1796 * writeback is done when khugepaged
1797 * revisits this page.
1799 * This is a one-off situation. We are not
1800 * forcing writeback in loop.
1802 xas_unlock_irq(&xas);
1803 filemap_flush(mapping);
1806 } else if (PageWriteback(page)) {
1807 xas_unlock_irq(&xas);
1810 } else if (trylock_page(page)) {
1812 xas_unlock_irq(&xas);
1814 result = SCAN_PAGE_LOCK;
1820 * The page must be locked, so we can drop the i_pages lock
1821 * without racing with truncate.
1823 VM_BUG_ON_PAGE(!PageLocked(page), page);
1825 /* make sure the page is up to date */
1826 if (unlikely(!PageUptodate(page))) {
1832 * If file was truncated then extended, or hole-punched, before
1833 * we locked the first page, then a THP might be there already.
1835 if (PageTransCompound(page)) {
1836 result = SCAN_PAGE_COMPOUND;
1840 if (page_mapping(page) != mapping) {
1841 result = SCAN_TRUNCATED;
1845 if (!is_shmem && (PageDirty(page) ||
1846 PageWriteback(page))) {
1848 * khugepaged only works on read-only fd, so this
1849 * page is dirty because it hasn't been flushed
1850 * since first write.
1856 if (isolate_lru_page(page)) {
1857 result = SCAN_DEL_PAGE_LRU;
1861 if (page_has_private(page) &&
1862 !try_to_release_page(page, GFP_KERNEL)) {
1863 result = SCAN_PAGE_HAS_PRIVATE;
1864 putback_lru_page(page);
1868 if (page_mapped(page))
1869 unmap_mapping_pages(mapping, index, 1, false);
1872 xas_set(&xas, index);
1874 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1875 VM_BUG_ON_PAGE(page_mapped(page), page);
1878 * The page is expected to have page_count() == 3:
1879 * - we hold a pin on it;
1880 * - one reference from page cache;
1881 * - one from isolate_lru_page;
1883 if (!page_ref_freeze(page, 3)) {
1884 result = SCAN_PAGE_COUNT;
1885 xas_unlock_irq(&xas);
1886 putback_lru_page(page);
1891 * Add the page to the list to be able to undo the collapse if
1892 * something go wrong.
1894 list_add_tail(&page->lru, &pagelist);
1896 /* Finally, replace with the new page. */
1897 xas_store(&xas, new_page);
1904 nr = thp_nr_pages(new_page);
1907 __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr);
1909 __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
1910 filemap_nr_thps_inc(mapping);
1912 * Paired with smp_mb() in do_dentry_open() to ensure
1913 * i_writecount is up to date and the update to nr_thps is
1914 * visible. Ensures the page cache will be truncated if the
1915 * file is opened writable.
1918 if (inode_is_open_for_write(mapping->host)) {
1920 __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);
1921 filemap_nr_thps_dec(mapping);
1927 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1929 __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1933 xas_unlock_irq(&xas);
1936 if (result == SCAN_SUCCEED) {
1937 struct page *page, *tmp;
1940 * Replacing old pages with new one has succeeded, now we
1941 * need to copy the content and free the old pages.
1944 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1945 while (index < page->index) {
1946 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1949 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1951 list_del(&page->lru);
1952 page->mapping = NULL;
1953 page_ref_unfreeze(page, 1);
1954 ClearPageActive(page);
1955 ClearPageUnevictable(page);
1960 while (index < end) {
1961 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1965 SetPageUptodate(new_page);
1966 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1968 set_page_dirty(new_page);
1969 lru_cache_add(new_page);
1972 * Remove pte page tables, so we can re-fault the page as huge.
1974 retract_page_tables(mapping, start);
1977 khugepaged_pages_collapsed++;
1981 /* Something went wrong: roll back page cache changes */
1983 mapping->nrpages -= nr_none;
1986 shmem_uncharge(mapping->host, nr_none);
1988 xas_set(&xas, start);
1989 xas_for_each(&xas, page, end - 1) {
1990 page = list_first_entry_or_null(&pagelist,
1992 if (!page || xas.xa_index < page->index) {
1996 /* Put holes back where they were */
1997 xas_store(&xas, NULL);
2001 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
2003 /* Unfreeze the page. */
2004 list_del(&page->lru);
2005 page_ref_unfreeze(page, 2);
2006 xas_store(&xas, page);
2008 xas_unlock_irq(&xas);
2010 putback_lru_page(page);
2014 xas_unlock_irq(&xas);
2016 new_page->mapping = NULL;
2019 unlock_page(new_page);
2021 VM_BUG_ON(!list_empty(&pagelist));
2022 if (!IS_ERR_OR_NULL(*hpage))
2023 mem_cgroup_uncharge(*hpage);
2024 /* TODO: tracepoints */
2027 static void khugepaged_scan_file(struct mm_struct *mm,
2028 struct file *file, pgoff_t start, struct page **hpage)
2030 struct page *page = NULL;
2031 struct address_space *mapping = file->f_mapping;
2032 XA_STATE(xas, &mapping->i_pages, start);
2034 int node = NUMA_NO_NODE;
2035 int result = SCAN_SUCCEED;
2039 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2041 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2042 if (xas_retry(&xas, page))
2045 if (xa_is_value(page)) {
2046 if (++swap > khugepaged_max_ptes_swap) {
2047 result = SCAN_EXCEED_SWAP_PTE;
2053 if (PageTransCompound(page)) {
2054 result = SCAN_PAGE_COMPOUND;
2058 node = page_to_nid(page);
2059 if (khugepaged_scan_abort(node)) {
2060 result = SCAN_SCAN_ABORT;
2063 khugepaged_node_load[node]++;
2065 if (!PageLRU(page)) {
2066 result = SCAN_PAGE_LRU;
2070 if (page_count(page) !=
2071 1 + page_mapcount(page) + page_has_private(page)) {
2072 result = SCAN_PAGE_COUNT;
2077 * We probably should check if the page is referenced here, but
2078 * nobody would transfer pte_young() to PageReferenced() for us.
2079 * And rmap walk here is just too costly...
2084 if (need_resched()) {
2091 if (result == SCAN_SUCCEED) {
2092 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2093 result = SCAN_EXCEED_NONE_PTE;
2095 node = khugepaged_find_target_node();
2096 collapse_file(mm, file, start, hpage, node);
2100 /* TODO: tracepoints */
2103 static void khugepaged_scan_file(struct mm_struct *mm,
2104 struct file *file, pgoff_t start, struct page **hpage)
2109 static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2114 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2115 struct page **hpage)
2116 __releases(&khugepaged_mm_lock)
2117 __acquires(&khugepaged_mm_lock)
2119 struct mm_slot *mm_slot;
2120 struct mm_struct *mm;
2121 struct vm_area_struct *vma;
2125 lockdep_assert_held(&khugepaged_mm_lock);
2127 if (khugepaged_scan.mm_slot)
2128 mm_slot = khugepaged_scan.mm_slot;
2130 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2131 struct mm_slot, mm_node);
2132 khugepaged_scan.address = 0;
2133 khugepaged_scan.mm_slot = mm_slot;
2135 spin_unlock(&khugepaged_mm_lock);
2136 khugepaged_collapse_pte_mapped_thps(mm_slot);
2140 * Don't wait for semaphore (to avoid long wait times). Just move to
2141 * the next mm on the list.
2144 if (unlikely(!mmap_read_trylock(mm)))
2145 goto breakouterloop_mmap_lock;
2146 if (likely(!khugepaged_test_exit(mm)))
2147 vma = find_vma(mm, khugepaged_scan.address);
2150 for (; vma; vma = vma->vm_next) {
2151 unsigned long hstart, hend;
2154 if (unlikely(khugepaged_test_exit(mm))) {
2158 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2163 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2164 hend = vma->vm_end & HPAGE_PMD_MASK;
2167 if (khugepaged_scan.address > hend)
2169 if (khugepaged_scan.address < hstart)
2170 khugepaged_scan.address = hstart;
2171 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2172 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2175 while (khugepaged_scan.address < hend) {
2178 if (unlikely(khugepaged_test_exit(mm)))
2179 goto breakouterloop;
2181 VM_BUG_ON(khugepaged_scan.address < hstart ||
2182 khugepaged_scan.address + HPAGE_PMD_SIZE >
2184 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2185 struct file *file = get_file(vma->vm_file);
2186 pgoff_t pgoff = linear_page_index(vma,
2187 khugepaged_scan.address);
2189 mmap_read_unlock(mm);
2191 khugepaged_scan_file(mm, file, pgoff, hpage);
2194 ret = khugepaged_scan_pmd(mm, vma,
2195 khugepaged_scan.address,
2198 /* move to next address */
2199 khugepaged_scan.address += HPAGE_PMD_SIZE;
2200 progress += HPAGE_PMD_NR;
2202 /* we released mmap_lock so break loop */
2203 goto breakouterloop_mmap_lock;
2204 if (progress >= pages)
2205 goto breakouterloop;
2209 mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2210 breakouterloop_mmap_lock:
2212 spin_lock(&khugepaged_mm_lock);
2213 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2215 * Release the current mm_slot if this mm is about to die, or
2216 * if we scanned all vmas of this mm.
2218 if (khugepaged_test_exit(mm) || !vma) {
2220 * Make sure that if mm_users is reaching zero while
2221 * khugepaged runs here, khugepaged_exit will find
2222 * mm_slot not pointing to the exiting mm.
2224 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2225 khugepaged_scan.mm_slot = list_entry(
2226 mm_slot->mm_node.next,
2227 struct mm_slot, mm_node);
2228 khugepaged_scan.address = 0;
2230 khugepaged_scan.mm_slot = NULL;
2231 khugepaged_full_scans++;
2234 collect_mm_slot(mm_slot);
2240 static int khugepaged_has_work(void)
2242 return !list_empty(&khugepaged_scan.mm_head) &&
2243 khugepaged_enabled();
2246 static int khugepaged_wait_event(void)
2248 return !list_empty(&khugepaged_scan.mm_head) ||
2249 kthread_should_stop();
2252 static void khugepaged_do_scan(void)
2254 struct page *hpage = NULL;
2255 unsigned int progress = 0, pass_through_head = 0;
2256 unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2259 lru_add_drain_all();
2261 while (progress < pages) {
2262 if (!khugepaged_prealloc_page(&hpage, &wait))
2267 if (unlikely(kthread_should_stop() || try_to_freeze()))
2270 spin_lock(&khugepaged_mm_lock);
2271 if (!khugepaged_scan.mm_slot)
2272 pass_through_head++;
2273 if (khugepaged_has_work() &&
2274 pass_through_head < 2)
2275 progress += khugepaged_scan_mm_slot(pages - progress,
2279 spin_unlock(&khugepaged_mm_lock);
2282 if (!IS_ERR_OR_NULL(hpage))
2286 static bool khugepaged_should_wakeup(void)
2288 return kthread_should_stop() ||
2289 time_after_eq(jiffies, khugepaged_sleep_expire);
2292 static void khugepaged_wait_work(void)
2294 if (khugepaged_has_work()) {
2295 const unsigned long scan_sleep_jiffies =
2296 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2298 if (!scan_sleep_jiffies)
2301 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2302 wait_event_freezable_timeout(khugepaged_wait,
2303 khugepaged_should_wakeup(),
2304 scan_sleep_jiffies);
2308 if (khugepaged_enabled())
2309 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2312 static int khugepaged(void *none)
2314 struct mm_slot *mm_slot;
2317 set_user_nice(current, MAX_NICE);
2319 while (!kthread_should_stop()) {
2320 khugepaged_do_scan();
2321 khugepaged_wait_work();
2324 spin_lock(&khugepaged_mm_lock);
2325 mm_slot = khugepaged_scan.mm_slot;
2326 khugepaged_scan.mm_slot = NULL;
2328 collect_mm_slot(mm_slot);
2329 spin_unlock(&khugepaged_mm_lock);
2333 static void set_recommended_min_free_kbytes(void)
2337 unsigned long recommended_min;
2339 for_each_populated_zone(zone) {
2341 * We don't need to worry about fragmentation of
2342 * ZONE_MOVABLE since it only has movable pages.
2344 if (zone_idx(zone) > gfp_zone(GFP_USER))
2350 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2351 recommended_min = pageblock_nr_pages * nr_zones * 2;
2354 * Make sure that on average at least two pageblocks are almost free
2355 * of another type, one for a migratetype to fall back to and a
2356 * second to avoid subsequent fallbacks of other types There are 3
2357 * MIGRATE_TYPES we care about.
2359 recommended_min += pageblock_nr_pages * nr_zones *
2360 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2362 /* don't ever allow to reserve more than 5% of the lowmem */
2363 recommended_min = min(recommended_min,
2364 (unsigned long) nr_free_buffer_pages() / 20);
2365 recommended_min <<= (PAGE_SHIFT-10);
2367 if (recommended_min > min_free_kbytes) {
2368 if (user_min_free_kbytes >= 0)
2369 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2370 min_free_kbytes, recommended_min);
2372 min_free_kbytes = recommended_min;
2374 setup_per_zone_wmarks();
2377 int start_stop_khugepaged(void)
2381 mutex_lock(&khugepaged_mutex);
2382 if (khugepaged_enabled()) {
2383 if (!khugepaged_thread)
2384 khugepaged_thread = kthread_run(khugepaged, NULL,
2386 if (IS_ERR(khugepaged_thread)) {
2387 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2388 err = PTR_ERR(khugepaged_thread);
2389 khugepaged_thread = NULL;
2393 if (!list_empty(&khugepaged_scan.mm_head))
2394 wake_up_interruptible(&khugepaged_wait);
2396 set_recommended_min_free_kbytes();
2397 } else if (khugepaged_thread) {
2398 kthread_stop(khugepaged_thread);
2399 khugepaged_thread = NULL;
2402 mutex_unlock(&khugepaged_mutex);
2406 void khugepaged_min_free_kbytes_update(void)
2408 mutex_lock(&khugepaged_mutex);
2409 if (khugepaged_enabled() && khugepaged_thread)
2410 set_recommended_min_free_kbytes();
2411 mutex_unlock(&khugepaged_mutex);