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
95 struct hlist_node hash;
96 struct list_head mm_node;
99 /* pte-mapped THP in this mm */
100 int nr_pte_mapped_thp;
101 unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
105 * struct khugepaged_scan - cursor for scanning
106 * @mm_head: the head of the mm list to scan
107 * @mm_slot: the current mm_slot we are scanning
108 * @address: the next address inside that to be scanned
110 * There is only the one khugepaged_scan instance of this cursor structure.
112 struct khugepaged_scan {
113 struct list_head mm_head;
114 struct mm_slot *mm_slot;
115 unsigned long address;
118 static struct khugepaged_scan khugepaged_scan = {
119 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
123 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
124 struct kobj_attribute *attr,
127 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
130 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
131 struct kobj_attribute *attr,
132 const char *buf, size_t count)
137 err = kstrtoul(buf, 10, &msecs);
138 if (err || msecs > UINT_MAX)
141 khugepaged_scan_sleep_millisecs = msecs;
142 khugepaged_sleep_expire = 0;
143 wake_up_interruptible(&khugepaged_wait);
147 static struct kobj_attribute scan_sleep_millisecs_attr =
148 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
149 scan_sleep_millisecs_store);
151 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
152 struct kobj_attribute *attr,
155 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
158 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
159 struct kobj_attribute *attr,
160 const char *buf, size_t count)
165 err = kstrtoul(buf, 10, &msecs);
166 if (err || msecs > UINT_MAX)
169 khugepaged_alloc_sleep_millisecs = msecs;
170 khugepaged_sleep_expire = 0;
171 wake_up_interruptible(&khugepaged_wait);
175 static struct kobj_attribute alloc_sleep_millisecs_attr =
176 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
177 alloc_sleep_millisecs_store);
179 static ssize_t pages_to_scan_show(struct kobject *kobj,
180 struct kobj_attribute *attr,
183 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
185 static ssize_t pages_to_scan_store(struct kobject *kobj,
186 struct kobj_attribute *attr,
187 const char *buf, size_t count)
192 err = kstrtoul(buf, 10, &pages);
193 if (err || !pages || pages > UINT_MAX)
196 khugepaged_pages_to_scan = pages;
200 static struct kobj_attribute pages_to_scan_attr =
201 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
202 pages_to_scan_store);
204 static ssize_t pages_collapsed_show(struct kobject *kobj,
205 struct kobj_attribute *attr,
208 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
210 static struct kobj_attribute pages_collapsed_attr =
211 __ATTR_RO(pages_collapsed);
213 static ssize_t full_scans_show(struct kobject *kobj,
214 struct kobj_attribute *attr,
217 return sprintf(buf, "%u\n", khugepaged_full_scans);
219 static struct kobj_attribute full_scans_attr =
220 __ATTR_RO(full_scans);
222 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
223 struct kobj_attribute *attr, char *buf)
225 return single_hugepage_flag_show(kobj, attr, buf,
226 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
228 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
229 struct kobj_attribute *attr,
230 const char *buf, size_t count)
232 return single_hugepage_flag_store(kobj, attr, buf, count,
233 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
235 static struct kobj_attribute khugepaged_defrag_attr =
236 __ATTR(defrag, 0644, khugepaged_defrag_show,
237 khugepaged_defrag_store);
240 * max_ptes_none controls if khugepaged should collapse hugepages over
241 * any unmapped ptes in turn potentially increasing the memory
242 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
243 * reduce the available free memory in the system as it
244 * runs. Increasing max_ptes_none will instead potentially reduce the
245 * free memory in the system during the khugepaged scan.
247 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
248 struct kobj_attribute *attr,
251 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
253 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
254 struct kobj_attribute *attr,
255 const char *buf, size_t count)
258 unsigned long max_ptes_none;
260 err = kstrtoul(buf, 10, &max_ptes_none);
261 if (err || max_ptes_none > HPAGE_PMD_NR-1)
264 khugepaged_max_ptes_none = max_ptes_none;
268 static struct kobj_attribute khugepaged_max_ptes_none_attr =
269 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
270 khugepaged_max_ptes_none_store);
272 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
273 struct kobj_attribute *attr,
276 return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
279 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
280 struct kobj_attribute *attr,
281 const char *buf, size_t count)
284 unsigned long max_ptes_swap;
286 err = kstrtoul(buf, 10, &max_ptes_swap);
287 if (err || max_ptes_swap > HPAGE_PMD_NR-1)
290 khugepaged_max_ptes_swap = max_ptes_swap;
295 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
296 __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
297 khugepaged_max_ptes_swap_store);
299 static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
300 struct kobj_attribute *attr,
303 return sprintf(buf, "%u\n", khugepaged_max_ptes_shared);
306 static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
307 struct kobj_attribute *attr,
308 const char *buf, size_t count)
311 unsigned long max_ptes_shared;
313 err = kstrtoul(buf, 10, &max_ptes_shared);
314 if (err || max_ptes_shared > HPAGE_PMD_NR-1)
317 khugepaged_max_ptes_shared = max_ptes_shared;
322 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
323 __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
324 khugepaged_max_ptes_shared_store);
326 static struct attribute *khugepaged_attr[] = {
327 &khugepaged_defrag_attr.attr,
328 &khugepaged_max_ptes_none_attr.attr,
329 &khugepaged_max_ptes_swap_attr.attr,
330 &khugepaged_max_ptes_shared_attr.attr,
331 &pages_to_scan_attr.attr,
332 &pages_collapsed_attr.attr,
333 &full_scans_attr.attr,
334 &scan_sleep_millisecs_attr.attr,
335 &alloc_sleep_millisecs_attr.attr,
339 struct attribute_group khugepaged_attr_group = {
340 .attrs = khugepaged_attr,
341 .name = "khugepaged",
343 #endif /* CONFIG_SYSFS */
345 int hugepage_madvise(struct vm_area_struct *vma,
346 unsigned long *vm_flags, int advice)
352 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
353 * can't handle this properly after s390_enable_sie, so we simply
354 * ignore the madvise to prevent qemu from causing a SIGSEGV.
356 if (mm_has_pgste(vma->vm_mm))
359 *vm_flags &= ~VM_NOHUGEPAGE;
360 *vm_flags |= VM_HUGEPAGE;
362 * If the vma become good for khugepaged to scan,
363 * register it here without waiting a page fault that
364 * may not happen any time soon.
366 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
367 khugepaged_enter_vma_merge(vma, *vm_flags))
370 case MADV_NOHUGEPAGE:
371 *vm_flags &= ~VM_HUGEPAGE;
372 *vm_flags |= VM_NOHUGEPAGE;
374 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
375 * this vma even if we leave the mm registered in khugepaged if
376 * it got registered before VM_NOHUGEPAGE was set.
384 int __init khugepaged_init(void)
386 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
387 sizeof(struct mm_slot),
388 __alignof__(struct mm_slot), 0, NULL);
392 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
393 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
394 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
395 khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
400 void __init khugepaged_destroy(void)
402 kmem_cache_destroy(mm_slot_cache);
405 static inline struct mm_slot *alloc_mm_slot(void)
407 if (!mm_slot_cache) /* initialization failed */
409 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
412 static inline void free_mm_slot(struct mm_slot *mm_slot)
414 kmem_cache_free(mm_slot_cache, mm_slot);
417 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
419 struct mm_slot *mm_slot;
421 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
422 if (mm == mm_slot->mm)
428 static void insert_to_mm_slots_hash(struct mm_struct *mm,
429 struct mm_slot *mm_slot)
432 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
435 static inline int khugepaged_test_exit(struct mm_struct *mm)
437 return atomic_read(&mm->mm_users) == 0;
440 static bool hugepage_vma_check(struct vm_area_struct *vma,
441 unsigned long vm_flags)
443 if (!transhuge_vma_enabled(vma, vm_flags))
446 /* Enabled via shmem mount options or sysfs settings. */
447 if (shmem_file(vma->vm_file) && shmem_huge_enabled(vma)) {
448 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
452 /* THP settings require madvise. */
453 if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
456 /* Read-only file mappings need to be aligned for THP to work. */
457 if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
458 (vm_flags & VM_DENYWRITE)) {
459 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
463 if (!vma->anon_vma || vma->vm_ops)
465 if (vma_is_temporary_stack(vma))
467 return !(vm_flags & VM_NO_KHUGEPAGED);
470 int __khugepaged_enter(struct mm_struct *mm)
472 struct mm_slot *mm_slot;
475 mm_slot = alloc_mm_slot();
479 /* __khugepaged_exit() must not run from under us */
480 VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
481 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
482 free_mm_slot(mm_slot);
486 spin_lock(&khugepaged_mm_lock);
487 insert_to_mm_slots_hash(mm, mm_slot);
489 * Insert just behind the scanning cursor, to let the area settle
492 wakeup = list_empty(&khugepaged_scan.mm_head);
493 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
494 spin_unlock(&khugepaged_mm_lock);
498 wake_up_interruptible(&khugepaged_wait);
503 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
504 unsigned long vm_flags)
506 unsigned long hstart, hend;
509 * khugepaged only supports read-only files for non-shmem files.
510 * khugepaged does not yet work on special mappings. And
511 * file-private shmem THP is not supported.
513 if (!hugepage_vma_check(vma, vm_flags))
516 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
517 hend = vma->vm_end & HPAGE_PMD_MASK;
519 return khugepaged_enter(vma, vm_flags);
523 void __khugepaged_exit(struct mm_struct *mm)
525 struct mm_slot *mm_slot;
528 spin_lock(&khugepaged_mm_lock);
529 mm_slot = get_mm_slot(mm);
530 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
531 hash_del(&mm_slot->hash);
532 list_del(&mm_slot->mm_node);
535 spin_unlock(&khugepaged_mm_lock);
538 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
539 free_mm_slot(mm_slot);
541 } else if (mm_slot) {
543 * This is required to serialize against
544 * khugepaged_test_exit() (which is guaranteed to run
545 * under mmap sem read mode). Stop here (after we
546 * return all pagetables will be destroyed) until
547 * khugepaged has finished working on the pagetables
548 * under the mmap_lock.
551 mmap_write_unlock(mm);
555 static void release_pte_page(struct page *page)
557 mod_node_page_state(page_pgdat(page),
558 NR_ISOLATED_ANON + page_is_file_lru(page),
561 putback_lru_page(page);
564 static void release_pte_pages(pte_t *pte, pte_t *_pte,
565 struct list_head *compound_pagelist)
567 struct page *page, *tmp;
569 while (--_pte >= pte) {
570 pte_t pteval = *_pte;
572 page = pte_page(pteval);
573 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
575 release_pte_page(page);
578 list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
579 list_del(&page->lru);
580 release_pte_page(page);
584 static bool is_refcount_suitable(struct page *page)
586 int expected_refcount;
588 expected_refcount = total_mapcount(page);
589 if (PageSwapCache(page))
590 expected_refcount += compound_nr(page);
592 return page_count(page) == expected_refcount;
595 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
596 unsigned long address,
598 struct list_head *compound_pagelist)
600 struct page *page = NULL;
602 int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
603 bool writable = false;
605 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
606 _pte++, address += PAGE_SIZE) {
607 pte_t pteval = *_pte;
608 if (pte_none(pteval) || (pte_present(pteval) &&
609 is_zero_pfn(pte_pfn(pteval)))) {
610 if (!userfaultfd_armed(vma) &&
611 ++none_or_zero <= khugepaged_max_ptes_none) {
614 result = SCAN_EXCEED_NONE_PTE;
618 if (!pte_present(pteval)) {
619 result = SCAN_PTE_NON_PRESENT;
622 page = vm_normal_page(vma, address, pteval);
623 if (unlikely(!page)) {
624 result = SCAN_PAGE_NULL;
628 VM_BUG_ON_PAGE(!PageAnon(page), page);
630 if (page_mapcount(page) > 1 &&
631 ++shared > khugepaged_max_ptes_shared) {
632 result = SCAN_EXCEED_SHARED_PTE;
636 if (PageCompound(page)) {
638 page = compound_head(page);
641 * Check if we have dealt with the compound page
644 list_for_each_entry(p, compound_pagelist, lru) {
651 * We can do it before isolate_lru_page because the
652 * page can't be freed from under us. NOTE: PG_lock
653 * is needed to serialize against split_huge_page
654 * when invoked from the VM.
656 if (!trylock_page(page)) {
657 result = SCAN_PAGE_LOCK;
662 * Check if the page has any GUP (or other external) pins.
664 * The page table that maps the page has been already unlinked
665 * from the page table tree and this process cannot get
666 * an additinal pin on the page.
668 * New pins can come later if the page is shared across fork,
669 * but not from this process. The other process cannot write to
670 * the page, only trigger CoW.
672 if (!is_refcount_suitable(page)) {
674 result = SCAN_PAGE_COUNT;
677 if (!pte_write(pteval) && PageSwapCache(page) &&
678 !reuse_swap_page(page, NULL)) {
680 * Page is in the swap cache and cannot be re-used.
681 * It cannot be collapsed into a THP.
684 result = SCAN_SWAP_CACHE_PAGE;
689 * Isolate the page to avoid collapsing an hugepage
690 * currently in use by the VM.
692 if (isolate_lru_page(page)) {
694 result = SCAN_DEL_PAGE_LRU;
697 mod_node_page_state(page_pgdat(page),
698 NR_ISOLATED_ANON + page_is_file_lru(page),
700 VM_BUG_ON_PAGE(!PageLocked(page), page);
701 VM_BUG_ON_PAGE(PageLRU(page), page);
703 if (PageCompound(page))
704 list_add_tail(&page->lru, compound_pagelist);
706 /* There should be enough young pte to collapse the page */
707 if (pte_young(pteval) ||
708 page_is_young(page) || PageReferenced(page) ||
709 mmu_notifier_test_young(vma->vm_mm, address))
712 if (pte_write(pteval))
716 if (unlikely(!writable)) {
717 result = SCAN_PAGE_RO;
718 } else if (unlikely(!referenced)) {
719 result = SCAN_LACK_REFERENCED_PAGE;
721 result = SCAN_SUCCEED;
722 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
723 referenced, writable, result);
727 release_pte_pages(pte, _pte, compound_pagelist);
728 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
729 referenced, writable, result);
733 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
734 struct vm_area_struct *vma,
735 unsigned long address,
737 struct list_head *compound_pagelist)
739 struct page *src_page, *tmp;
741 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
742 _pte++, page++, address += PAGE_SIZE) {
743 pte_t pteval = *_pte;
745 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
746 clear_user_highpage(page, address);
747 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
748 if (is_zero_pfn(pte_pfn(pteval))) {
750 * ptl mostly unnecessary.
754 * paravirt calls inside pte_clear here are
757 pte_clear(vma->vm_mm, address, _pte);
761 src_page = pte_page(pteval);
762 copy_user_highpage(page, src_page, address, vma);
763 if (!PageCompound(src_page))
764 release_pte_page(src_page);
766 * ptl mostly unnecessary, but preempt has to
767 * be disabled to update the per-cpu stats
768 * inside page_remove_rmap().
772 * paravirt calls inside pte_clear here are
775 pte_clear(vma->vm_mm, address, _pte);
776 page_remove_rmap(src_page, false);
778 free_page_and_swap_cache(src_page);
782 list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
783 list_del(&src_page->lru);
784 release_pte_page(src_page);
788 static void khugepaged_alloc_sleep(void)
792 add_wait_queue(&khugepaged_wait, &wait);
793 freezable_schedule_timeout_interruptible(
794 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
795 remove_wait_queue(&khugepaged_wait, &wait);
798 static int khugepaged_node_load[MAX_NUMNODES];
800 static bool khugepaged_scan_abort(int nid)
805 * If node_reclaim_mode is disabled, then no extra effort is made to
806 * allocate memory locally.
808 if (!node_reclaim_mode)
811 /* If there is a count for this node already, it must be acceptable */
812 if (khugepaged_node_load[nid])
815 for (i = 0; i < MAX_NUMNODES; i++) {
816 if (!khugepaged_node_load[i])
818 if (node_distance(nid, i) > node_reclaim_distance)
824 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
825 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
827 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
831 static int khugepaged_find_target_node(void)
833 static int last_khugepaged_target_node = NUMA_NO_NODE;
834 int nid, target_node = 0, max_value = 0;
836 /* find first node with max normal pages hit */
837 for (nid = 0; nid < MAX_NUMNODES; nid++)
838 if (khugepaged_node_load[nid] > max_value) {
839 max_value = khugepaged_node_load[nid];
843 /* do some balance if several nodes have the same hit record */
844 if (target_node <= last_khugepaged_target_node)
845 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
847 if (max_value == khugepaged_node_load[nid]) {
852 last_khugepaged_target_node = target_node;
856 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
858 if (IS_ERR(*hpage)) {
864 khugepaged_alloc_sleep();
874 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
876 VM_BUG_ON_PAGE(*hpage, *hpage);
878 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
879 if (unlikely(!*hpage)) {
880 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
881 *hpage = ERR_PTR(-ENOMEM);
885 prep_transhuge_page(*hpage);
886 count_vm_event(THP_COLLAPSE_ALLOC);
890 static int khugepaged_find_target_node(void)
895 static inline struct page *alloc_khugepaged_hugepage(void)
899 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
902 prep_transhuge_page(page);
906 static struct page *khugepaged_alloc_hugepage(bool *wait)
911 hpage = alloc_khugepaged_hugepage();
913 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
918 khugepaged_alloc_sleep();
920 count_vm_event(THP_COLLAPSE_ALLOC);
921 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
926 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
929 * If the hpage allocated earlier was briefly exposed in page cache
930 * before collapse_file() failed, it is possible that racing lookups
931 * have not yet completed, and would then be unpleasantly surprised by
932 * finding the hpage reused for the same mapping at a different offset.
933 * Just release the previous allocation if there is any danger of that.
935 if (*hpage && page_count(*hpage) > 1) {
941 *hpage = khugepaged_alloc_hugepage(wait);
943 if (unlikely(!*hpage))
950 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
959 * If mmap_lock temporarily dropped, revalidate vma
960 * before taking mmap_lock.
961 * Return 0 if succeeds, otherwise return none-zero
965 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
966 struct vm_area_struct **vmap)
968 struct vm_area_struct *vma;
969 unsigned long hstart, hend;
971 if (unlikely(khugepaged_test_exit(mm)))
972 return SCAN_ANY_PROCESS;
974 *vmap = vma = find_vma(mm, address);
976 return SCAN_VMA_NULL;
978 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
979 hend = vma->vm_end & HPAGE_PMD_MASK;
980 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
981 return SCAN_ADDRESS_RANGE;
982 if (!hugepage_vma_check(vma, vma->vm_flags))
983 return SCAN_VMA_CHECK;
984 /* Anon VMA expected */
985 if (!vma->anon_vma || vma->vm_ops)
986 return SCAN_VMA_CHECK;
991 * Bring missing pages in from swap, to complete THP collapse.
992 * Only done if khugepaged_scan_pmd believes it is worthwhile.
994 * Called and returns without pte mapped or spinlocks held,
995 * but with mmap_lock held to protect against vma changes.
998 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
999 struct vm_area_struct *vma,
1000 unsigned long address, pmd_t *pmd,
1005 struct vm_fault vmf = {
1008 .flags = FAULT_FLAG_ALLOW_RETRY,
1010 .pgoff = linear_page_index(vma, address),
1013 vmf.pte = pte_offset_map(pmd, address);
1014 for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
1015 vmf.pte++, vmf.address += PAGE_SIZE) {
1016 vmf.orig_pte = *vmf.pte;
1017 if (!is_swap_pte(vmf.orig_pte))
1020 ret = do_swap_page(&vmf);
1022 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1023 if (ret & VM_FAULT_RETRY) {
1025 if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
1026 /* vma is no longer available, don't continue to swapin */
1027 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1030 /* check if the pmd is still valid */
1031 if (mm_find_pmd(mm, address) != pmd) {
1032 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1036 if (ret & VM_FAULT_ERROR) {
1037 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1040 /* pte is unmapped now, we need to map it */
1041 vmf.pte = pte_offset_map(pmd, vmf.address);
1046 /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1050 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1054 static void collapse_huge_page(struct mm_struct *mm,
1055 unsigned long address,
1056 struct page **hpage,
1057 int node, int referenced, int unmapped)
1059 LIST_HEAD(compound_pagelist);
1063 struct page *new_page;
1064 spinlock_t *pmd_ptl, *pte_ptl;
1065 int isolated = 0, result = 0;
1066 struct vm_area_struct *vma;
1067 struct mmu_notifier_range range;
1070 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1072 /* Only allocate from the target node */
1073 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1076 * Before allocating the hugepage, release the mmap_lock read lock.
1077 * The allocation can take potentially a long time if it involves
1078 * sync compaction, and we do not need to hold the mmap_lock during
1079 * that. We will recheck the vma after taking it again in write mode.
1081 mmap_read_unlock(mm);
1082 new_page = khugepaged_alloc_page(hpage, gfp, node);
1084 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1088 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1089 result = SCAN_CGROUP_CHARGE_FAIL;
1092 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1095 result = hugepage_vma_revalidate(mm, address, &vma);
1097 mmap_read_unlock(mm);
1101 pmd = mm_find_pmd(mm, address);
1103 result = SCAN_PMD_NULL;
1104 mmap_read_unlock(mm);
1109 * __collapse_huge_page_swapin always returns with mmap_lock locked.
1110 * If it fails, we release mmap_lock and jump out_nolock.
1111 * Continuing to collapse causes inconsistency.
1113 if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1115 mmap_read_unlock(mm);
1119 mmap_read_unlock(mm);
1121 * Prevent all access to pagetables with the exception of
1122 * gup_fast later handled by the ptep_clear_flush and the VM
1123 * handled by the anon_vma lock + PG_lock.
1125 mmap_write_lock(mm);
1126 result = hugepage_vma_revalidate(mm, address, &vma);
1129 /* check if the pmd is still valid */
1130 if (mm_find_pmd(mm, address) != pmd)
1133 anon_vma_lock_write(vma->anon_vma);
1135 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1136 address, address + HPAGE_PMD_SIZE);
1137 mmu_notifier_invalidate_range_start(&range);
1139 pte = pte_offset_map(pmd, address);
1140 pte_ptl = pte_lockptr(mm, pmd);
1142 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1144 * After this gup_fast can't run anymore. This also removes
1145 * any huge TLB entry from the CPU so we won't allow
1146 * huge and small TLB entries for the same virtual address
1147 * to avoid the risk of CPU bugs in that area.
1149 _pmd = pmdp_collapse_flush(vma, address, pmd);
1150 spin_unlock(pmd_ptl);
1151 mmu_notifier_invalidate_range_end(&range);
1154 isolated = __collapse_huge_page_isolate(vma, address, pte,
1155 &compound_pagelist);
1156 spin_unlock(pte_ptl);
1158 if (unlikely(!isolated)) {
1161 BUG_ON(!pmd_none(*pmd));
1163 * We can only use set_pmd_at when establishing
1164 * hugepmds and never for establishing regular pmds that
1165 * points to regular pagetables. Use pmd_populate for that
1167 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1168 spin_unlock(pmd_ptl);
1169 anon_vma_unlock_write(vma->anon_vma);
1175 * All pages are isolated and locked so anon_vma rmap
1176 * can't run anymore.
1178 anon_vma_unlock_write(vma->anon_vma);
1180 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1181 &compound_pagelist);
1183 __SetPageUptodate(new_page);
1184 pgtable = pmd_pgtable(_pmd);
1186 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1187 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1190 * spin_lock() below is not the equivalent of smp_wmb(), so
1191 * this is needed to avoid the copy_huge_page writes to become
1192 * visible after the set_pmd_at() write.
1197 BUG_ON(!pmd_none(*pmd));
1198 page_add_new_anon_rmap(new_page, vma, address, true);
1199 lru_cache_add_inactive_or_unevictable(new_page, vma);
1200 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1201 set_pmd_at(mm, address, pmd, _pmd);
1202 update_mmu_cache_pmd(vma, address, pmd);
1203 spin_unlock(pmd_ptl);
1207 khugepaged_pages_collapsed++;
1208 result = SCAN_SUCCEED;
1210 mmap_write_unlock(mm);
1212 if (!IS_ERR_OR_NULL(*hpage))
1213 mem_cgroup_uncharge(*hpage);
1214 trace_mm_collapse_huge_page(mm, isolated, result);
1220 static int khugepaged_scan_pmd(struct mm_struct *mm,
1221 struct vm_area_struct *vma,
1222 unsigned long address,
1223 struct page **hpage)
1227 int ret = 0, result = 0, referenced = 0;
1228 int none_or_zero = 0, shared = 0;
1229 struct page *page = NULL;
1230 unsigned long _address;
1232 int node = NUMA_NO_NODE, unmapped = 0;
1233 bool writable = false;
1235 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1237 pmd = mm_find_pmd(mm, address);
1239 result = SCAN_PMD_NULL;
1243 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1244 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1245 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1246 _pte++, _address += PAGE_SIZE) {
1247 pte_t pteval = *_pte;
1248 if (is_swap_pte(pteval)) {
1249 if (++unmapped <= khugepaged_max_ptes_swap) {
1251 * Always be strict with uffd-wp
1252 * enabled swap entries. Please see
1253 * comment below for pte_uffd_wp().
1255 if (pte_swp_uffd_wp(pteval)) {
1256 result = SCAN_PTE_UFFD_WP;
1261 result = SCAN_EXCEED_SWAP_PTE;
1265 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1266 if (!userfaultfd_armed(vma) &&
1267 ++none_or_zero <= khugepaged_max_ptes_none) {
1270 result = SCAN_EXCEED_NONE_PTE;
1274 if (!pte_present(pteval)) {
1275 result = SCAN_PTE_NON_PRESENT;
1278 if (pte_uffd_wp(pteval)) {
1280 * Don't collapse the page if any of the small
1281 * PTEs are armed with uffd write protection.
1282 * Here we can also mark the new huge pmd as
1283 * write protected if any of the small ones is
1284 * marked but that could bring uknown
1285 * userfault messages that falls outside of
1286 * the registered range. So, just be simple.
1288 result = SCAN_PTE_UFFD_WP;
1291 if (pte_write(pteval))
1294 page = vm_normal_page(vma, _address, pteval);
1295 if (unlikely(!page)) {
1296 result = SCAN_PAGE_NULL;
1300 if (page_mapcount(page) > 1 &&
1301 ++shared > khugepaged_max_ptes_shared) {
1302 result = SCAN_EXCEED_SHARED_PTE;
1306 page = compound_head(page);
1309 * Record which node the original page is from and save this
1310 * information to khugepaged_node_load[].
1311 * Khupaged will allocate hugepage from the node has the max
1314 node = page_to_nid(page);
1315 if (khugepaged_scan_abort(node)) {
1316 result = SCAN_SCAN_ABORT;
1319 khugepaged_node_load[node]++;
1320 if (!PageLRU(page)) {
1321 result = SCAN_PAGE_LRU;
1324 if (PageLocked(page)) {
1325 result = SCAN_PAGE_LOCK;
1328 if (!PageAnon(page)) {
1329 result = SCAN_PAGE_ANON;
1334 * Check if the page has any GUP (or other external) pins.
1336 * Here the check is racy it may see totmal_mapcount > refcount
1338 * For example, one process with one forked child process.
1339 * The parent has the PMD split due to MADV_DONTNEED, then
1340 * the child is trying unmap the whole PMD, but khugepaged
1341 * may be scanning the parent between the child has
1342 * PageDoubleMap flag cleared and dec the mapcount. So
1343 * khugepaged may see total_mapcount > refcount.
1345 * But such case is ephemeral we could always retry collapse
1346 * later. However it may report false positive if the page
1347 * has excessive GUP pins (i.e. 512). Anyway the same check
1348 * will be done again later the risk seems low.
1350 if (!is_refcount_suitable(page)) {
1351 result = SCAN_PAGE_COUNT;
1354 if (pte_young(pteval) ||
1355 page_is_young(page) || PageReferenced(page) ||
1356 mmu_notifier_test_young(vma->vm_mm, address))
1360 result = SCAN_PAGE_RO;
1361 } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1362 result = SCAN_LACK_REFERENCED_PAGE;
1364 result = SCAN_SUCCEED;
1368 pte_unmap_unlock(pte, ptl);
1370 node = khugepaged_find_target_node();
1371 /* collapse_huge_page will return with the mmap_lock released */
1372 collapse_huge_page(mm, address, hpage, node,
1373 referenced, unmapped);
1376 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1377 none_or_zero, result, unmapped);
1381 static void collect_mm_slot(struct mm_slot *mm_slot)
1383 struct mm_struct *mm = mm_slot->mm;
1385 lockdep_assert_held(&khugepaged_mm_lock);
1387 if (khugepaged_test_exit(mm)) {
1389 hash_del(&mm_slot->hash);
1390 list_del(&mm_slot->mm_node);
1393 * Not strictly needed because the mm exited already.
1395 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1398 /* khugepaged_mm_lock actually not necessary for the below */
1399 free_mm_slot(mm_slot);
1406 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1407 * khugepaged should try to collapse the page table.
1409 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1412 struct mm_slot *mm_slot;
1414 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1416 spin_lock(&khugepaged_mm_lock);
1417 mm_slot = get_mm_slot(mm);
1418 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1419 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1420 spin_unlock(&khugepaged_mm_lock);
1425 * Try to collapse a pte-mapped THP for mm at address haddr.
1427 * This function checks whether all the PTEs in the PMD are pointing to the
1428 * right THP. If so, retract the page table so the THP can refault in with
1431 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1433 unsigned long haddr = addr & HPAGE_PMD_MASK;
1434 struct vm_area_struct *vma = find_vma(mm, haddr);
1436 pte_t *start_pte, *pte;
1442 if (!vma || !vma->vm_file ||
1443 vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1447 * This vm_flags may not have VM_HUGEPAGE if the page was not
1448 * collapsed by this mm. But we can still collapse if the page is
1449 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1450 * will not fail the vma for missing VM_HUGEPAGE
1452 if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1455 hpage = find_lock_page(vma->vm_file->f_mapping,
1456 linear_page_index(vma, haddr));
1460 if (!PageHead(hpage))
1463 pmd = mm_find_pmd(mm, haddr);
1467 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1469 /* step 1: check all mapped PTEs are to the right huge page */
1470 for (i = 0, addr = haddr, pte = start_pte;
1471 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1474 /* empty pte, skip */
1478 /* page swapped out, abort */
1479 if (!pte_present(*pte))
1482 page = vm_normal_page(vma, addr, *pte);
1485 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1486 * page table, but the new page will not be a subpage of hpage.
1488 if (hpage + i != page)
1493 /* step 2: adjust rmap */
1494 for (i = 0, addr = haddr, pte = start_pte;
1495 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1500 page = vm_normal_page(vma, addr, *pte);
1501 page_remove_rmap(page, false);
1504 pte_unmap_unlock(start_pte, ptl);
1506 /* step 3: set proper refcount and mm_counters. */
1508 page_ref_sub(hpage, count);
1509 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1512 /* step 4: collapse pmd */
1513 ptl = pmd_lock(vma->vm_mm, pmd);
1514 _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1517 pte_free(mm, pmd_pgtable(_pmd));
1525 pte_unmap_unlock(start_pte, ptl);
1529 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1531 struct mm_struct *mm = mm_slot->mm;
1534 if (likely(mm_slot->nr_pte_mapped_thp == 0))
1537 if (!mmap_write_trylock(mm))
1540 if (unlikely(khugepaged_test_exit(mm)))
1543 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1544 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1547 mm_slot->nr_pte_mapped_thp = 0;
1548 mmap_write_unlock(mm);
1552 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1554 struct vm_area_struct *vma;
1555 struct mm_struct *mm;
1559 i_mmap_lock_write(mapping);
1560 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1562 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1563 * got written to. These VMAs are likely not worth investing
1564 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1567 * Not that vma->anon_vma check is racy: it can be set up after
1568 * the check but before we took mmap_lock by the fault path.
1569 * But page lock would prevent establishing any new ptes of the
1570 * page, so we are safe.
1572 * An alternative would be drop the check, but check that page
1573 * table is clear before calling pmdp_collapse_flush() under
1574 * ptl. It has higher chance to recover THP for the VMA, but
1575 * has higher cost too.
1579 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1580 if (addr & ~HPAGE_PMD_MASK)
1582 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1585 pmd = mm_find_pmd(mm, addr);
1589 * We need exclusive mmap_lock to retract page table.
1591 * We use trylock due to lock inversion: we need to acquire
1592 * mmap_lock while holding page lock. Fault path does it in
1593 * reverse order. Trylock is a way to avoid deadlock.
1595 if (mmap_write_trylock(mm)) {
1596 if (!khugepaged_test_exit(mm)) {
1597 spinlock_t *ptl = pmd_lock(mm, pmd);
1598 /* assume page table is clear */
1599 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1602 pte_free(mm, pmd_pgtable(_pmd));
1604 mmap_write_unlock(mm);
1606 /* Try again later */
1607 khugepaged_add_pte_mapped_thp(mm, addr);
1610 i_mmap_unlock_write(mapping);
1614 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1616 * Basic scheme is simple, details are more complex:
1617 * - allocate and lock a new huge page;
1618 * - scan page cache replacing old pages with the new one
1619 * + swap/gup in pages if necessary;
1621 * + keep old pages around in case rollback is required;
1622 * - if replacing succeeds:
1625 * + unlock huge page;
1626 * - if replacing failed;
1627 * + put all pages back and unfreeze them;
1628 * + restore gaps in the page cache;
1629 * + unlock and free huge page;
1631 static void collapse_file(struct mm_struct *mm,
1632 struct file *file, pgoff_t start,
1633 struct page **hpage, int node)
1635 struct address_space *mapping = file->f_mapping;
1637 struct page *new_page;
1638 pgoff_t index, end = start + HPAGE_PMD_NR;
1639 LIST_HEAD(pagelist);
1640 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1641 int nr_none = 0, result = SCAN_SUCCEED;
1642 bool is_shmem = shmem_file(file);
1644 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1645 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1647 /* Only allocate from the target node */
1648 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1650 new_page = khugepaged_alloc_page(hpage, gfp, node);
1652 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1656 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1657 result = SCAN_CGROUP_CHARGE_FAIL;
1660 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1662 /* This will be less messy when we use multi-index entries */
1665 xas_create_range(&xas);
1666 if (!xas_error(&xas))
1668 xas_unlock_irq(&xas);
1669 if (!xas_nomem(&xas, GFP_KERNEL)) {
1675 __SetPageLocked(new_page);
1677 __SetPageSwapBacked(new_page);
1678 new_page->index = start;
1679 new_page->mapping = mapping;
1682 * At this point the new_page is locked and not up-to-date.
1683 * It's safe to insert it into the page cache, because nobody would
1684 * be able to map it or use it in another way until we unlock it.
1687 xas_set(&xas, start);
1688 for (index = start; index < end; index++) {
1689 struct page *page = xas_next(&xas);
1691 VM_BUG_ON(index != xas.xa_index);
1695 * Stop if extent has been truncated or
1696 * hole-punched, and is now completely
1699 if (index == start) {
1700 if (!xas_next_entry(&xas, end - 1)) {
1701 result = SCAN_TRUNCATED;
1704 xas_set(&xas, index);
1706 if (!shmem_charge(mapping->host, 1)) {
1710 xas_store(&xas, new_page);
1715 if (xa_is_value(page) || !PageUptodate(page)) {
1716 xas_unlock_irq(&xas);
1717 /* swap in or instantiate fallocated page */
1718 if (shmem_getpage(mapping->host, index, &page,
1723 } else if (trylock_page(page)) {
1725 xas_unlock_irq(&xas);
1727 result = SCAN_PAGE_LOCK;
1730 } else { /* !is_shmem */
1731 if (!page || xa_is_value(page)) {
1732 xas_unlock_irq(&xas);
1733 page_cache_sync_readahead(mapping, &file->f_ra,
1736 /* drain pagevecs to help isolate_lru_page() */
1738 page = find_lock_page(mapping, index);
1739 if (unlikely(page == NULL)) {
1743 } else if (PageDirty(page)) {
1745 * khugepaged only works on read-only fd,
1746 * so this page is dirty because it hasn't
1747 * been flushed since first write. There
1748 * won't be new dirty pages.
1750 * Trigger async flush here and hope the
1751 * writeback is done when khugepaged
1752 * revisits this page.
1754 * This is a one-off situation. We are not
1755 * forcing writeback in loop.
1757 xas_unlock_irq(&xas);
1758 filemap_flush(mapping);
1761 } else if (PageWriteback(page)) {
1762 xas_unlock_irq(&xas);
1765 } else if (trylock_page(page)) {
1767 xas_unlock_irq(&xas);
1769 result = SCAN_PAGE_LOCK;
1775 * The page must be locked, so we can drop the i_pages lock
1776 * without racing with truncate.
1778 VM_BUG_ON_PAGE(!PageLocked(page), page);
1780 /* make sure the page is up to date */
1781 if (unlikely(!PageUptodate(page))) {
1787 * If file was truncated then extended, or hole-punched, before
1788 * we locked the first page, then a THP might be there already.
1790 if (PageTransCompound(page)) {
1791 result = SCAN_PAGE_COMPOUND;
1795 if (page_mapping(page) != mapping) {
1796 result = SCAN_TRUNCATED;
1800 if (!is_shmem && (PageDirty(page) ||
1801 PageWriteback(page))) {
1803 * khugepaged only works on read-only fd, so this
1804 * page is dirty because it hasn't been flushed
1805 * since first write.
1811 if (isolate_lru_page(page)) {
1812 result = SCAN_DEL_PAGE_LRU;
1816 if (page_has_private(page) &&
1817 !try_to_release_page(page, GFP_KERNEL)) {
1818 result = SCAN_PAGE_HAS_PRIVATE;
1819 putback_lru_page(page);
1823 if (page_mapped(page))
1824 unmap_mapping_pages(mapping, index, 1, false);
1827 xas_set(&xas, index);
1829 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1830 VM_BUG_ON_PAGE(page_mapped(page), page);
1833 * The page is expected to have page_count() == 3:
1834 * - we hold a pin on it;
1835 * - one reference from page cache;
1836 * - one from isolate_lru_page;
1838 if (!page_ref_freeze(page, 3)) {
1839 result = SCAN_PAGE_COUNT;
1840 xas_unlock_irq(&xas);
1841 putback_lru_page(page);
1846 * Add the page to the list to be able to undo the collapse if
1847 * something go wrong.
1849 list_add_tail(&page->lru, &pagelist);
1851 /* Finally, replace with the new page. */
1852 xas_store(&xas, new_page);
1861 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1863 __inc_node_page_state(new_page, NR_FILE_THPS);
1864 filemap_nr_thps_inc(mapping);
1868 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1870 __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1874 xas_unlock_irq(&xas);
1877 if (result == SCAN_SUCCEED) {
1878 struct page *page, *tmp;
1881 * Replacing old pages with new one has succeeded, now we
1882 * need to copy the content and free the old pages.
1885 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1886 while (index < page->index) {
1887 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1890 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1892 list_del(&page->lru);
1893 page->mapping = NULL;
1894 page_ref_unfreeze(page, 1);
1895 ClearPageActive(page);
1896 ClearPageUnevictable(page);
1901 while (index < end) {
1902 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1906 SetPageUptodate(new_page);
1907 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1909 set_page_dirty(new_page);
1910 lru_cache_add(new_page);
1913 * Remove pte page tables, so we can re-fault the page as huge.
1915 retract_page_tables(mapping, start);
1918 khugepaged_pages_collapsed++;
1922 /* Something went wrong: roll back page cache changes */
1924 mapping->nrpages -= nr_none;
1927 shmem_uncharge(mapping->host, nr_none);
1929 xas_set(&xas, start);
1930 xas_for_each(&xas, page, end - 1) {
1931 page = list_first_entry_or_null(&pagelist,
1933 if (!page || xas.xa_index < page->index) {
1937 /* Put holes back where they were */
1938 xas_store(&xas, NULL);
1942 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1944 /* Unfreeze the page. */
1945 list_del(&page->lru);
1946 page_ref_unfreeze(page, 2);
1947 xas_store(&xas, page);
1949 xas_unlock_irq(&xas);
1951 putback_lru_page(page);
1955 xas_unlock_irq(&xas);
1957 new_page->mapping = NULL;
1960 unlock_page(new_page);
1962 VM_BUG_ON(!list_empty(&pagelist));
1963 if (!IS_ERR_OR_NULL(*hpage))
1964 mem_cgroup_uncharge(*hpage);
1965 /* TODO: tracepoints */
1968 static void khugepaged_scan_file(struct mm_struct *mm,
1969 struct file *file, pgoff_t start, struct page **hpage)
1971 struct page *page = NULL;
1972 struct address_space *mapping = file->f_mapping;
1973 XA_STATE(xas, &mapping->i_pages, start);
1975 int node = NUMA_NO_NODE;
1976 int result = SCAN_SUCCEED;
1980 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1982 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1983 if (xas_retry(&xas, page))
1986 if (xa_is_value(page)) {
1987 if (++swap > khugepaged_max_ptes_swap) {
1988 result = SCAN_EXCEED_SWAP_PTE;
1994 if (PageTransCompound(page)) {
1995 result = SCAN_PAGE_COMPOUND;
1999 node = page_to_nid(page);
2000 if (khugepaged_scan_abort(node)) {
2001 result = SCAN_SCAN_ABORT;
2004 khugepaged_node_load[node]++;
2006 if (!PageLRU(page)) {
2007 result = SCAN_PAGE_LRU;
2011 if (page_count(page) !=
2012 1 + page_mapcount(page) + page_has_private(page)) {
2013 result = SCAN_PAGE_COUNT;
2018 * We probably should check if the page is referenced here, but
2019 * nobody would transfer pte_young() to PageReferenced() for us.
2020 * And rmap walk here is just too costly...
2025 if (need_resched()) {
2032 if (result == SCAN_SUCCEED) {
2033 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2034 result = SCAN_EXCEED_NONE_PTE;
2036 node = khugepaged_find_target_node();
2037 collapse_file(mm, file, start, hpage, node);
2041 /* TODO: tracepoints */
2044 static void khugepaged_scan_file(struct mm_struct *mm,
2045 struct file *file, pgoff_t start, struct page **hpage)
2050 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2056 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2057 struct page **hpage)
2058 __releases(&khugepaged_mm_lock)
2059 __acquires(&khugepaged_mm_lock)
2061 struct mm_slot *mm_slot;
2062 struct mm_struct *mm;
2063 struct vm_area_struct *vma;
2067 lockdep_assert_held(&khugepaged_mm_lock);
2069 if (khugepaged_scan.mm_slot)
2070 mm_slot = khugepaged_scan.mm_slot;
2072 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2073 struct mm_slot, mm_node);
2074 khugepaged_scan.address = 0;
2075 khugepaged_scan.mm_slot = mm_slot;
2077 spin_unlock(&khugepaged_mm_lock);
2078 khugepaged_collapse_pte_mapped_thps(mm_slot);
2082 * Don't wait for semaphore (to avoid long wait times). Just move to
2083 * the next mm on the list.
2086 if (unlikely(!mmap_read_trylock(mm)))
2087 goto breakouterloop_mmap_lock;
2088 if (likely(!khugepaged_test_exit(mm)))
2089 vma = find_vma(mm, khugepaged_scan.address);
2092 for (; vma; vma = vma->vm_next) {
2093 unsigned long hstart, hend;
2096 if (unlikely(khugepaged_test_exit(mm))) {
2100 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2105 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2106 hend = vma->vm_end & HPAGE_PMD_MASK;
2109 if (khugepaged_scan.address > hend)
2111 if (khugepaged_scan.address < hstart)
2112 khugepaged_scan.address = hstart;
2113 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2114 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2117 while (khugepaged_scan.address < hend) {
2120 if (unlikely(khugepaged_test_exit(mm)))
2121 goto breakouterloop;
2123 VM_BUG_ON(khugepaged_scan.address < hstart ||
2124 khugepaged_scan.address + HPAGE_PMD_SIZE >
2126 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2127 struct file *file = get_file(vma->vm_file);
2128 pgoff_t pgoff = linear_page_index(vma,
2129 khugepaged_scan.address);
2131 mmap_read_unlock(mm);
2133 khugepaged_scan_file(mm, file, pgoff, hpage);
2136 ret = khugepaged_scan_pmd(mm, vma,
2137 khugepaged_scan.address,
2140 /* move to next address */
2141 khugepaged_scan.address += HPAGE_PMD_SIZE;
2142 progress += HPAGE_PMD_NR;
2144 /* we released mmap_lock so break loop */
2145 goto breakouterloop_mmap_lock;
2146 if (progress >= pages)
2147 goto breakouterloop;
2151 mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2152 breakouterloop_mmap_lock:
2154 spin_lock(&khugepaged_mm_lock);
2155 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2157 * Release the current mm_slot if this mm is about to die, or
2158 * if we scanned all vmas of this mm.
2160 if (khugepaged_test_exit(mm) || !vma) {
2162 * Make sure that if mm_users is reaching zero while
2163 * khugepaged runs here, khugepaged_exit will find
2164 * mm_slot not pointing to the exiting mm.
2166 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2167 khugepaged_scan.mm_slot = list_entry(
2168 mm_slot->mm_node.next,
2169 struct mm_slot, mm_node);
2170 khugepaged_scan.address = 0;
2172 khugepaged_scan.mm_slot = NULL;
2173 khugepaged_full_scans++;
2176 collect_mm_slot(mm_slot);
2182 static int khugepaged_has_work(void)
2184 return !list_empty(&khugepaged_scan.mm_head) &&
2185 khugepaged_enabled();
2188 static int khugepaged_wait_event(void)
2190 return !list_empty(&khugepaged_scan.mm_head) ||
2191 kthread_should_stop();
2194 static void khugepaged_do_scan(void)
2196 struct page *hpage = NULL;
2197 unsigned int progress = 0, pass_through_head = 0;
2198 unsigned int pages = khugepaged_pages_to_scan;
2201 barrier(); /* write khugepaged_pages_to_scan to local stack */
2203 lru_add_drain_all();
2205 while (progress < pages) {
2206 if (!khugepaged_prealloc_page(&hpage, &wait))
2211 if (unlikely(kthread_should_stop() || try_to_freeze()))
2214 spin_lock(&khugepaged_mm_lock);
2215 if (!khugepaged_scan.mm_slot)
2216 pass_through_head++;
2217 if (khugepaged_has_work() &&
2218 pass_through_head < 2)
2219 progress += khugepaged_scan_mm_slot(pages - progress,
2223 spin_unlock(&khugepaged_mm_lock);
2226 if (!IS_ERR_OR_NULL(hpage))
2230 static bool khugepaged_should_wakeup(void)
2232 return kthread_should_stop() ||
2233 time_after_eq(jiffies, khugepaged_sleep_expire);
2236 static void khugepaged_wait_work(void)
2238 if (khugepaged_has_work()) {
2239 const unsigned long scan_sleep_jiffies =
2240 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2242 if (!scan_sleep_jiffies)
2245 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2246 wait_event_freezable_timeout(khugepaged_wait,
2247 khugepaged_should_wakeup(),
2248 scan_sleep_jiffies);
2252 if (khugepaged_enabled())
2253 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2256 static int khugepaged(void *none)
2258 struct mm_slot *mm_slot;
2261 set_user_nice(current, MAX_NICE);
2263 while (!kthread_should_stop()) {
2264 khugepaged_do_scan();
2265 khugepaged_wait_work();
2268 spin_lock(&khugepaged_mm_lock);
2269 mm_slot = khugepaged_scan.mm_slot;
2270 khugepaged_scan.mm_slot = NULL;
2272 collect_mm_slot(mm_slot);
2273 spin_unlock(&khugepaged_mm_lock);
2277 static void set_recommended_min_free_kbytes(void)
2281 unsigned long recommended_min;
2283 for_each_populated_zone(zone) {
2285 * We don't need to worry about fragmentation of
2286 * ZONE_MOVABLE since it only has movable pages.
2288 if (zone_idx(zone) > gfp_zone(GFP_USER))
2294 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2295 recommended_min = pageblock_nr_pages * nr_zones * 2;
2298 * Make sure that on average at least two pageblocks are almost free
2299 * of another type, one for a migratetype to fall back to and a
2300 * second to avoid subsequent fallbacks of other types There are 3
2301 * MIGRATE_TYPES we care about.
2303 recommended_min += pageblock_nr_pages * nr_zones *
2304 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2306 /* don't ever allow to reserve more than 5% of the lowmem */
2307 recommended_min = min(recommended_min,
2308 (unsigned long) nr_free_buffer_pages() / 20);
2309 recommended_min <<= (PAGE_SHIFT-10);
2311 if (recommended_min > min_free_kbytes) {
2312 if (user_min_free_kbytes >= 0)
2313 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2314 min_free_kbytes, recommended_min);
2316 min_free_kbytes = recommended_min;
2318 setup_per_zone_wmarks();
2321 int start_stop_khugepaged(void)
2325 mutex_lock(&khugepaged_mutex);
2326 if (khugepaged_enabled()) {
2327 if (!khugepaged_thread)
2328 khugepaged_thread = kthread_run(khugepaged, NULL,
2330 if (IS_ERR(khugepaged_thread)) {
2331 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2332 err = PTR_ERR(khugepaged_thread);
2333 khugepaged_thread = NULL;
2337 if (!list_empty(&khugepaged_scan.mm_head))
2338 wake_up_interruptible(&khugepaged_wait);
2340 set_recommended_min_free_kbytes();
2341 } else if (khugepaged_thread) {
2342 kthread_stop(khugepaged_thread);
2343 khugepaged_thread = NULL;
2346 mutex_unlock(&khugepaged_mutex);
2350 void khugepaged_min_free_kbytes_update(void)
2352 mutex_lock(&khugepaged_mutex);
2353 if (khugepaged_enabled() && khugepaged_thread)
2354 set_recommended_min_free_kbytes();
2355 mutex_unlock(&khugepaged_mutex);