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 /* Explicitly disabled through madvise. */
444 if ((vm_flags & VM_NOHUGEPAGE) ||
445 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
448 /* Enabled via shmem mount options or sysfs settings. */
449 if (shmem_file(vma->vm_file) && shmem_huge_enabled(vma)) {
450 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
454 /* THP settings require madvise. */
455 if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
458 /* Read-only file mappings need to be aligned for THP to work. */
459 if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
460 (vm_flags & VM_DENYWRITE)) {
461 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
465 if (!vma->anon_vma || vma->vm_ops)
467 if (vma_is_temporary_stack(vma))
469 return !(vm_flags & VM_NO_KHUGEPAGED);
472 int __khugepaged_enter(struct mm_struct *mm)
474 struct mm_slot *mm_slot;
477 mm_slot = alloc_mm_slot();
481 /* __khugepaged_exit() must not run from under us */
482 VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
483 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
484 free_mm_slot(mm_slot);
488 spin_lock(&khugepaged_mm_lock);
489 insert_to_mm_slots_hash(mm, mm_slot);
491 * Insert just behind the scanning cursor, to let the area settle
494 wakeup = list_empty(&khugepaged_scan.mm_head);
495 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
496 spin_unlock(&khugepaged_mm_lock);
500 wake_up_interruptible(&khugepaged_wait);
505 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
506 unsigned long vm_flags)
508 unsigned long hstart, hend;
511 * khugepaged only supports read-only files for non-shmem files.
512 * khugepaged does not yet work on special mappings. And
513 * file-private shmem THP is not supported.
515 if (!hugepage_vma_check(vma, vm_flags))
518 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
519 hend = vma->vm_end & HPAGE_PMD_MASK;
521 return khugepaged_enter(vma, vm_flags);
525 void __khugepaged_exit(struct mm_struct *mm)
527 struct mm_slot *mm_slot;
530 spin_lock(&khugepaged_mm_lock);
531 mm_slot = get_mm_slot(mm);
532 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
533 hash_del(&mm_slot->hash);
534 list_del(&mm_slot->mm_node);
537 spin_unlock(&khugepaged_mm_lock);
540 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
541 free_mm_slot(mm_slot);
543 } else if (mm_slot) {
545 * This is required to serialize against
546 * khugepaged_test_exit() (which is guaranteed to run
547 * under mmap sem read mode). Stop here (after we
548 * return all pagetables will be destroyed) until
549 * khugepaged has finished working on the pagetables
550 * under the mmap_lock.
553 mmap_write_unlock(mm);
557 static void release_pte_page(struct page *page)
559 mod_node_page_state(page_pgdat(page),
560 NR_ISOLATED_ANON + page_is_file_lru(page),
563 putback_lru_page(page);
566 static void release_pte_pages(pte_t *pte, pte_t *_pte,
567 struct list_head *compound_pagelist)
569 struct page *page, *tmp;
571 while (--_pte >= pte) {
572 pte_t pteval = *_pte;
574 page = pte_page(pteval);
575 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
577 release_pte_page(page);
580 list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
581 list_del(&page->lru);
582 release_pte_page(page);
586 static bool is_refcount_suitable(struct page *page)
588 int expected_refcount;
590 expected_refcount = total_mapcount(page);
591 if (PageSwapCache(page))
592 expected_refcount += compound_nr(page);
594 return page_count(page) == expected_refcount;
597 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
598 unsigned long address,
600 struct list_head *compound_pagelist)
602 struct page *page = NULL;
604 int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
605 bool writable = false;
607 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
608 _pte++, address += PAGE_SIZE) {
609 pte_t pteval = *_pte;
610 if (pte_none(pteval) || (pte_present(pteval) &&
611 is_zero_pfn(pte_pfn(pteval)))) {
612 if (!userfaultfd_armed(vma) &&
613 ++none_or_zero <= khugepaged_max_ptes_none) {
616 result = SCAN_EXCEED_NONE_PTE;
620 if (!pte_present(pteval)) {
621 result = SCAN_PTE_NON_PRESENT;
624 page = vm_normal_page(vma, address, pteval);
625 if (unlikely(!page)) {
626 result = SCAN_PAGE_NULL;
630 VM_BUG_ON_PAGE(!PageAnon(page), page);
632 if (page_mapcount(page) > 1 &&
633 ++shared > khugepaged_max_ptes_shared) {
634 result = SCAN_EXCEED_SHARED_PTE;
638 if (PageCompound(page)) {
640 page = compound_head(page);
643 * Check if we have dealt with the compound page
646 list_for_each_entry(p, compound_pagelist, lru) {
653 * We can do it before isolate_lru_page because the
654 * page can't be freed from under us. NOTE: PG_lock
655 * is needed to serialize against split_huge_page
656 * when invoked from the VM.
658 if (!trylock_page(page)) {
659 result = SCAN_PAGE_LOCK;
664 * Check if the page has any GUP (or other external) pins.
666 * The page table that maps the page has been already unlinked
667 * from the page table tree and this process cannot get
668 * an additinal pin on the page.
670 * New pins can come later if the page is shared across fork,
671 * but not from this process. The other process cannot write to
672 * the page, only trigger CoW.
674 if (!is_refcount_suitable(page)) {
676 result = SCAN_PAGE_COUNT;
679 if (!pte_write(pteval) && PageSwapCache(page) &&
680 !reuse_swap_page(page, NULL)) {
682 * Page is in the swap cache and cannot be re-used.
683 * It cannot be collapsed into a THP.
686 result = SCAN_SWAP_CACHE_PAGE;
691 * Isolate the page to avoid collapsing an hugepage
692 * currently in use by the VM.
694 if (isolate_lru_page(page)) {
696 result = SCAN_DEL_PAGE_LRU;
699 mod_node_page_state(page_pgdat(page),
700 NR_ISOLATED_ANON + page_is_file_lru(page),
702 VM_BUG_ON_PAGE(!PageLocked(page), page);
703 VM_BUG_ON_PAGE(PageLRU(page), page);
705 if (PageCompound(page))
706 list_add_tail(&page->lru, compound_pagelist);
708 /* There should be enough young pte to collapse the page */
709 if (pte_young(pteval) ||
710 page_is_young(page) || PageReferenced(page) ||
711 mmu_notifier_test_young(vma->vm_mm, address))
714 if (pte_write(pteval))
717 if (likely(writable)) {
718 if (likely(referenced)) {
719 result = SCAN_SUCCEED;
720 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
721 referenced, writable, result);
725 result = SCAN_PAGE_RO;
729 release_pte_pages(pte, _pte, compound_pagelist);
730 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
731 referenced, writable, result);
735 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
736 struct vm_area_struct *vma,
737 unsigned long address,
739 struct list_head *compound_pagelist)
741 struct page *src_page, *tmp;
743 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
744 _pte++, page++, address += PAGE_SIZE) {
745 pte_t pteval = *_pte;
747 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
748 clear_user_highpage(page, address);
749 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
750 if (is_zero_pfn(pte_pfn(pteval))) {
752 * ptl mostly unnecessary.
756 * paravirt calls inside pte_clear here are
759 pte_clear(vma->vm_mm, address, _pte);
763 src_page = pte_page(pteval);
764 copy_user_highpage(page, src_page, address, vma);
765 if (!PageCompound(src_page))
766 release_pte_page(src_page);
768 * ptl mostly unnecessary, but preempt has to
769 * be disabled to update the per-cpu stats
770 * inside page_remove_rmap().
774 * paravirt calls inside pte_clear here are
777 pte_clear(vma->vm_mm, address, _pte);
778 page_remove_rmap(src_page, false);
780 free_page_and_swap_cache(src_page);
784 list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
785 list_del(&src_page->lru);
786 release_pte_page(src_page);
790 static void khugepaged_alloc_sleep(void)
794 add_wait_queue(&khugepaged_wait, &wait);
795 freezable_schedule_timeout_interruptible(
796 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
797 remove_wait_queue(&khugepaged_wait, &wait);
800 static int khugepaged_node_load[MAX_NUMNODES];
802 static bool khugepaged_scan_abort(int nid)
807 * If node_reclaim_mode is disabled, then no extra effort is made to
808 * allocate memory locally.
810 if (!node_reclaim_mode)
813 /* If there is a count for this node already, it must be acceptable */
814 if (khugepaged_node_load[nid])
817 for (i = 0; i < MAX_NUMNODES; i++) {
818 if (!khugepaged_node_load[i])
820 if (node_distance(nid, i) > node_reclaim_distance)
826 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
827 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
829 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
833 static int khugepaged_find_target_node(void)
835 static int last_khugepaged_target_node = NUMA_NO_NODE;
836 int nid, target_node = 0, max_value = 0;
838 /* find first node with max normal pages hit */
839 for (nid = 0; nid < MAX_NUMNODES; nid++)
840 if (khugepaged_node_load[nid] > max_value) {
841 max_value = khugepaged_node_load[nid];
845 /* do some balance if several nodes have the same hit record */
846 if (target_node <= last_khugepaged_target_node)
847 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
849 if (max_value == khugepaged_node_load[nid]) {
854 last_khugepaged_target_node = target_node;
858 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
860 if (IS_ERR(*hpage)) {
866 khugepaged_alloc_sleep();
876 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
878 VM_BUG_ON_PAGE(*hpage, *hpage);
880 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
881 if (unlikely(!*hpage)) {
882 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
883 *hpage = ERR_PTR(-ENOMEM);
887 prep_transhuge_page(*hpage);
888 count_vm_event(THP_COLLAPSE_ALLOC);
892 static int khugepaged_find_target_node(void)
897 static inline struct page *alloc_khugepaged_hugepage(void)
901 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
904 prep_transhuge_page(page);
908 static struct page *khugepaged_alloc_hugepage(bool *wait)
913 hpage = alloc_khugepaged_hugepage();
915 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
920 khugepaged_alloc_sleep();
922 count_vm_event(THP_COLLAPSE_ALLOC);
923 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
928 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
931 * If the hpage allocated earlier was briefly exposed in page cache
932 * before collapse_file() failed, it is possible that racing lookups
933 * have not yet completed, and would then be unpleasantly surprised by
934 * finding the hpage reused for the same mapping at a different offset.
935 * Just release the previous allocation if there is any danger of that.
937 if (*hpage && page_count(*hpage) > 1) {
943 *hpage = khugepaged_alloc_hugepage(wait);
945 if (unlikely(!*hpage))
952 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
961 * If mmap_lock temporarily dropped, revalidate vma
962 * before taking mmap_lock.
963 * Return 0 if succeeds, otherwise return none-zero
967 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
968 struct vm_area_struct **vmap)
970 struct vm_area_struct *vma;
971 unsigned long hstart, hend;
973 if (unlikely(khugepaged_test_exit(mm)))
974 return SCAN_ANY_PROCESS;
976 *vmap = vma = find_vma(mm, address);
978 return SCAN_VMA_NULL;
980 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
981 hend = vma->vm_end & HPAGE_PMD_MASK;
982 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
983 return SCAN_ADDRESS_RANGE;
984 if (!hugepage_vma_check(vma, vma->vm_flags))
985 return SCAN_VMA_CHECK;
986 /* Anon VMA expected */
987 if (!vma->anon_vma || vma->vm_ops)
988 return SCAN_VMA_CHECK;
993 * Bring missing pages in from swap, to complete THP collapse.
994 * Only done if khugepaged_scan_pmd believes it is worthwhile.
996 * Called and returns without pte mapped or spinlocks held,
997 * but with mmap_lock held to protect against vma changes.
1000 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
1001 struct vm_area_struct *vma,
1002 unsigned long address, pmd_t *pmd,
1007 struct vm_fault vmf = {
1010 .flags = FAULT_FLAG_ALLOW_RETRY,
1012 .pgoff = linear_page_index(vma, address),
1015 vmf.pte = pte_offset_map(pmd, address);
1016 for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
1017 vmf.pte++, vmf.address += PAGE_SIZE) {
1018 vmf.orig_pte = *vmf.pte;
1019 if (!is_swap_pte(vmf.orig_pte))
1022 ret = do_swap_page(&vmf);
1024 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1025 if (ret & VM_FAULT_RETRY) {
1027 if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
1028 /* vma is no longer available, don't continue to swapin */
1029 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1032 /* check if the pmd is still valid */
1033 if (mm_find_pmd(mm, address) != pmd) {
1034 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1038 if (ret & VM_FAULT_ERROR) {
1039 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1042 /* pte is unmapped now, we need to map it */
1043 vmf.pte = pte_offset_map(pmd, vmf.address);
1048 /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1052 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1056 static void collapse_huge_page(struct mm_struct *mm,
1057 unsigned long address,
1058 struct page **hpage,
1059 int node, int referenced, int unmapped)
1061 LIST_HEAD(compound_pagelist);
1065 struct page *new_page;
1066 spinlock_t *pmd_ptl, *pte_ptl;
1067 int isolated = 0, result = 0;
1068 struct vm_area_struct *vma;
1069 struct mmu_notifier_range range;
1072 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1074 /* Only allocate from the target node */
1075 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1078 * Before allocating the hugepage, release the mmap_lock read lock.
1079 * The allocation can take potentially a long time if it involves
1080 * sync compaction, and we do not need to hold the mmap_lock during
1081 * that. We will recheck the vma after taking it again in write mode.
1083 mmap_read_unlock(mm);
1084 new_page = khugepaged_alloc_page(hpage, gfp, node);
1086 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1090 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1091 result = SCAN_CGROUP_CHARGE_FAIL;
1094 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1097 result = hugepage_vma_revalidate(mm, address, &vma);
1099 mmap_read_unlock(mm);
1103 pmd = mm_find_pmd(mm, address);
1105 result = SCAN_PMD_NULL;
1106 mmap_read_unlock(mm);
1111 * __collapse_huge_page_swapin always returns with mmap_lock locked.
1112 * If it fails, we release mmap_lock and jump out_nolock.
1113 * Continuing to collapse causes inconsistency.
1115 if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1117 mmap_read_unlock(mm);
1121 mmap_read_unlock(mm);
1123 * Prevent all access to pagetables with the exception of
1124 * gup_fast later handled by the ptep_clear_flush and the VM
1125 * handled by the anon_vma lock + PG_lock.
1127 mmap_write_lock(mm);
1128 result = hugepage_vma_revalidate(mm, address, &vma);
1131 /* check if the pmd is still valid */
1132 if (mm_find_pmd(mm, address) != pmd)
1135 anon_vma_lock_write(vma->anon_vma);
1137 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1138 address, address + HPAGE_PMD_SIZE);
1139 mmu_notifier_invalidate_range_start(&range);
1141 pte = pte_offset_map(pmd, address);
1142 pte_ptl = pte_lockptr(mm, pmd);
1144 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1146 * After this gup_fast can't run anymore. This also removes
1147 * any huge TLB entry from the CPU so we won't allow
1148 * huge and small TLB entries for the same virtual address
1149 * to avoid the risk of CPU bugs in that area.
1151 _pmd = pmdp_collapse_flush(vma, address, pmd);
1152 spin_unlock(pmd_ptl);
1153 mmu_notifier_invalidate_range_end(&range);
1156 isolated = __collapse_huge_page_isolate(vma, address, pte,
1157 &compound_pagelist);
1158 spin_unlock(pte_ptl);
1160 if (unlikely(!isolated)) {
1163 BUG_ON(!pmd_none(*pmd));
1165 * We can only use set_pmd_at when establishing
1166 * hugepmds and never for establishing regular pmds that
1167 * points to regular pagetables. Use pmd_populate for that
1169 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1170 spin_unlock(pmd_ptl);
1171 anon_vma_unlock_write(vma->anon_vma);
1177 * All pages are isolated and locked so anon_vma rmap
1178 * can't run anymore.
1180 anon_vma_unlock_write(vma->anon_vma);
1182 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1183 &compound_pagelist);
1185 __SetPageUptodate(new_page);
1186 pgtable = pmd_pgtable(_pmd);
1188 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1189 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1192 * spin_lock() below is not the equivalent of smp_wmb(), so
1193 * this is needed to avoid the copy_huge_page writes to become
1194 * visible after the set_pmd_at() write.
1199 BUG_ON(!pmd_none(*pmd));
1200 page_add_new_anon_rmap(new_page, vma, address, true);
1201 lru_cache_add_inactive_or_unevictable(new_page, vma);
1202 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1203 set_pmd_at(mm, address, pmd, _pmd);
1204 update_mmu_cache_pmd(vma, address, pmd);
1205 spin_unlock(pmd_ptl);
1209 khugepaged_pages_collapsed++;
1210 result = SCAN_SUCCEED;
1212 mmap_write_unlock(mm);
1214 if (!IS_ERR_OR_NULL(*hpage))
1215 mem_cgroup_uncharge(*hpage);
1216 trace_mm_collapse_huge_page(mm, isolated, result);
1222 static int khugepaged_scan_pmd(struct mm_struct *mm,
1223 struct vm_area_struct *vma,
1224 unsigned long address,
1225 struct page **hpage)
1229 int ret = 0, result = 0, referenced = 0;
1230 int none_or_zero = 0, shared = 0;
1231 struct page *page = NULL;
1232 unsigned long _address;
1234 int node = NUMA_NO_NODE, unmapped = 0;
1235 bool writable = false;
1237 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1239 pmd = mm_find_pmd(mm, address);
1241 result = SCAN_PMD_NULL;
1245 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1246 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1247 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1248 _pte++, _address += PAGE_SIZE) {
1249 pte_t pteval = *_pte;
1250 if (is_swap_pte(pteval)) {
1251 if (++unmapped <= khugepaged_max_ptes_swap) {
1253 * Always be strict with uffd-wp
1254 * enabled swap entries. Please see
1255 * comment below for pte_uffd_wp().
1257 if (pte_swp_uffd_wp(pteval)) {
1258 result = SCAN_PTE_UFFD_WP;
1263 result = SCAN_EXCEED_SWAP_PTE;
1267 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1268 if (!userfaultfd_armed(vma) &&
1269 ++none_or_zero <= khugepaged_max_ptes_none) {
1272 result = SCAN_EXCEED_NONE_PTE;
1276 if (!pte_present(pteval)) {
1277 result = SCAN_PTE_NON_PRESENT;
1280 if (pte_uffd_wp(pteval)) {
1282 * Don't collapse the page if any of the small
1283 * PTEs are armed with uffd write protection.
1284 * Here we can also mark the new huge pmd as
1285 * write protected if any of the small ones is
1286 * marked but that could bring uknown
1287 * userfault messages that falls outside of
1288 * the registered range. So, just be simple.
1290 result = SCAN_PTE_UFFD_WP;
1293 if (pte_write(pteval))
1296 page = vm_normal_page(vma, _address, pteval);
1297 if (unlikely(!page)) {
1298 result = SCAN_PAGE_NULL;
1302 if (page_mapcount(page) > 1 &&
1303 ++shared > khugepaged_max_ptes_shared) {
1304 result = SCAN_EXCEED_SHARED_PTE;
1308 page = compound_head(page);
1311 * Record which node the original page is from and save this
1312 * information to khugepaged_node_load[].
1313 * Khupaged will allocate hugepage from the node has the max
1316 node = page_to_nid(page);
1317 if (khugepaged_scan_abort(node)) {
1318 result = SCAN_SCAN_ABORT;
1321 khugepaged_node_load[node]++;
1322 if (!PageLRU(page)) {
1323 result = SCAN_PAGE_LRU;
1326 if (PageLocked(page)) {
1327 result = SCAN_PAGE_LOCK;
1330 if (!PageAnon(page)) {
1331 result = SCAN_PAGE_ANON;
1336 * Check if the page has any GUP (or other external) pins.
1338 * Here the check is racy it may see totmal_mapcount > refcount
1340 * For example, one process with one forked child process.
1341 * The parent has the PMD split due to MADV_DONTNEED, then
1342 * the child is trying unmap the whole PMD, but khugepaged
1343 * may be scanning the parent between the child has
1344 * PageDoubleMap flag cleared and dec the mapcount. So
1345 * khugepaged may see total_mapcount > refcount.
1347 * But such case is ephemeral we could always retry collapse
1348 * later. However it may report false positive if the page
1349 * has excessive GUP pins (i.e. 512). Anyway the same check
1350 * will be done again later the risk seems low.
1352 if (!is_refcount_suitable(page)) {
1353 result = SCAN_PAGE_COUNT;
1356 if (pte_young(pteval) ||
1357 page_is_young(page) || PageReferenced(page) ||
1358 mmu_notifier_test_young(vma->vm_mm, address))
1362 result = SCAN_PAGE_RO;
1363 } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1364 result = SCAN_LACK_REFERENCED_PAGE;
1366 result = SCAN_SUCCEED;
1370 pte_unmap_unlock(pte, ptl);
1372 node = khugepaged_find_target_node();
1373 /* collapse_huge_page will return with the mmap_lock released */
1374 collapse_huge_page(mm, address, hpage, node,
1375 referenced, unmapped);
1378 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1379 none_or_zero, result, unmapped);
1383 static void collect_mm_slot(struct mm_slot *mm_slot)
1385 struct mm_struct *mm = mm_slot->mm;
1387 lockdep_assert_held(&khugepaged_mm_lock);
1389 if (khugepaged_test_exit(mm)) {
1391 hash_del(&mm_slot->hash);
1392 list_del(&mm_slot->mm_node);
1395 * Not strictly needed because the mm exited already.
1397 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1400 /* khugepaged_mm_lock actually not necessary for the below */
1401 free_mm_slot(mm_slot);
1408 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1409 * khugepaged should try to collapse the page table.
1411 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1414 struct mm_slot *mm_slot;
1416 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1418 spin_lock(&khugepaged_mm_lock);
1419 mm_slot = get_mm_slot(mm);
1420 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1421 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1422 spin_unlock(&khugepaged_mm_lock);
1427 * Try to collapse a pte-mapped THP for mm at address haddr.
1429 * This function checks whether all the PTEs in the PMD are pointing to the
1430 * right THP. If so, retract the page table so the THP can refault in with
1433 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1435 unsigned long haddr = addr & HPAGE_PMD_MASK;
1436 struct vm_area_struct *vma = find_vma(mm, haddr);
1438 pte_t *start_pte, *pte;
1444 if (!vma || !vma->vm_file ||
1445 vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1449 * This vm_flags may not have VM_HUGEPAGE if the page was not
1450 * collapsed by this mm. But we can still collapse if the page is
1451 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1452 * will not fail the vma for missing VM_HUGEPAGE
1454 if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1457 hpage = find_lock_page(vma->vm_file->f_mapping,
1458 linear_page_index(vma, haddr));
1462 if (!PageHead(hpage))
1465 pmd = mm_find_pmd(mm, haddr);
1469 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1471 /* step 1: check all mapped PTEs are to the right huge page */
1472 for (i = 0, addr = haddr, pte = start_pte;
1473 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1476 /* empty pte, skip */
1480 /* page swapped out, abort */
1481 if (!pte_present(*pte))
1484 page = vm_normal_page(vma, addr, *pte);
1487 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1488 * page table, but the new page will not be a subpage of hpage.
1490 if (hpage + i != page)
1495 /* step 2: adjust rmap */
1496 for (i = 0, addr = haddr, pte = start_pte;
1497 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1502 page = vm_normal_page(vma, addr, *pte);
1503 page_remove_rmap(page, false);
1506 pte_unmap_unlock(start_pte, ptl);
1508 /* step 3: set proper refcount and mm_counters. */
1510 page_ref_sub(hpage, count);
1511 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1514 /* step 4: collapse pmd */
1515 ptl = pmd_lock(vma->vm_mm, pmd);
1516 _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1519 pte_free(mm, pmd_pgtable(_pmd));
1527 pte_unmap_unlock(start_pte, ptl);
1531 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1533 struct mm_struct *mm = mm_slot->mm;
1536 if (likely(mm_slot->nr_pte_mapped_thp == 0))
1539 if (!mmap_write_trylock(mm))
1542 if (unlikely(khugepaged_test_exit(mm)))
1545 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1546 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1549 mm_slot->nr_pte_mapped_thp = 0;
1550 mmap_write_unlock(mm);
1554 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1556 struct vm_area_struct *vma;
1557 struct mm_struct *mm;
1561 i_mmap_lock_write(mapping);
1562 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1564 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1565 * got written to. These VMAs are likely not worth investing
1566 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1569 * Not that vma->anon_vma check is racy: it can be set up after
1570 * the check but before we took mmap_lock by the fault path.
1571 * But page lock would prevent establishing any new ptes of the
1572 * page, so we are safe.
1574 * An alternative would be drop the check, but check that page
1575 * table is clear before calling pmdp_collapse_flush() under
1576 * ptl. It has higher chance to recover THP for the VMA, but
1577 * has higher cost too.
1581 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1582 if (addr & ~HPAGE_PMD_MASK)
1584 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1587 pmd = mm_find_pmd(mm, addr);
1591 * We need exclusive mmap_lock to retract page table.
1593 * We use trylock due to lock inversion: we need to acquire
1594 * mmap_lock while holding page lock. Fault path does it in
1595 * reverse order. Trylock is a way to avoid deadlock.
1597 if (mmap_write_trylock(mm)) {
1598 if (!khugepaged_test_exit(mm)) {
1599 spinlock_t *ptl = pmd_lock(mm, pmd);
1600 /* assume page table is clear */
1601 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1604 pte_free(mm, pmd_pgtable(_pmd));
1606 mmap_write_unlock(mm);
1608 /* Try again later */
1609 khugepaged_add_pte_mapped_thp(mm, addr);
1612 i_mmap_unlock_write(mapping);
1616 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1618 * Basic scheme is simple, details are more complex:
1619 * - allocate and lock a new huge page;
1620 * - scan page cache replacing old pages with the new one
1621 * + swap/gup in pages if necessary;
1623 * + keep old pages around in case rollback is required;
1624 * - if replacing succeeds:
1627 * + unlock huge page;
1628 * - if replacing failed;
1629 * + put all pages back and unfreeze them;
1630 * + restore gaps in the page cache;
1631 * + unlock and free huge page;
1633 static void collapse_file(struct mm_struct *mm,
1634 struct file *file, pgoff_t start,
1635 struct page **hpage, int node)
1637 struct address_space *mapping = file->f_mapping;
1639 struct page *new_page;
1640 pgoff_t index, end = start + HPAGE_PMD_NR;
1641 LIST_HEAD(pagelist);
1642 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1643 int nr_none = 0, result = SCAN_SUCCEED;
1644 bool is_shmem = shmem_file(file);
1646 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1647 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1649 /* Only allocate from the target node */
1650 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1652 new_page = khugepaged_alloc_page(hpage, gfp, node);
1654 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1658 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1659 result = SCAN_CGROUP_CHARGE_FAIL;
1662 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1664 /* This will be less messy when we use multi-index entries */
1667 xas_create_range(&xas);
1668 if (!xas_error(&xas))
1670 xas_unlock_irq(&xas);
1671 if (!xas_nomem(&xas, GFP_KERNEL)) {
1677 __SetPageLocked(new_page);
1679 __SetPageSwapBacked(new_page);
1680 new_page->index = start;
1681 new_page->mapping = mapping;
1684 * At this point the new_page is locked and not up-to-date.
1685 * It's safe to insert it into the page cache, because nobody would
1686 * be able to map it or use it in another way until we unlock it.
1689 xas_set(&xas, start);
1690 for (index = start; index < end; index++) {
1691 struct page *page = xas_next(&xas);
1693 VM_BUG_ON(index != xas.xa_index);
1697 * Stop if extent has been truncated or
1698 * hole-punched, and is now completely
1701 if (index == start) {
1702 if (!xas_next_entry(&xas, end - 1)) {
1703 result = SCAN_TRUNCATED;
1706 xas_set(&xas, index);
1708 if (!shmem_charge(mapping->host, 1)) {
1712 xas_store(&xas, new_page);
1717 if (xa_is_value(page) || !PageUptodate(page)) {
1718 xas_unlock_irq(&xas);
1719 /* swap in or instantiate fallocated page */
1720 if (shmem_getpage(mapping->host, index, &page,
1725 } else if (trylock_page(page)) {
1727 xas_unlock_irq(&xas);
1729 result = SCAN_PAGE_LOCK;
1732 } else { /* !is_shmem */
1733 if (!page || xa_is_value(page)) {
1734 xas_unlock_irq(&xas);
1735 page_cache_sync_readahead(mapping, &file->f_ra,
1738 /* drain pagevecs to help isolate_lru_page() */
1740 page = find_lock_page(mapping, index);
1741 if (unlikely(page == NULL)) {
1745 } else if (PageDirty(page)) {
1747 * khugepaged only works on read-only fd,
1748 * so this page is dirty because it hasn't
1749 * been flushed since first write. There
1750 * won't be new dirty pages.
1752 * Trigger async flush here and hope the
1753 * writeback is done when khugepaged
1754 * revisits this page.
1756 * This is a one-off situation. We are not
1757 * forcing writeback in loop.
1759 xas_unlock_irq(&xas);
1760 filemap_flush(mapping);
1763 } else if (trylock_page(page)) {
1765 xas_unlock_irq(&xas);
1767 result = SCAN_PAGE_LOCK;
1773 * The page must be locked, so we can drop the i_pages lock
1774 * without racing with truncate.
1776 VM_BUG_ON_PAGE(!PageLocked(page), page);
1778 /* make sure the page is up to date */
1779 if (unlikely(!PageUptodate(page))) {
1785 * If file was truncated then extended, or hole-punched, before
1786 * we locked the first page, then a THP might be there already.
1788 if (PageTransCompound(page)) {
1789 result = SCAN_PAGE_COMPOUND;
1793 if (page_mapping(page) != mapping) {
1794 result = SCAN_TRUNCATED;
1798 if (!is_shmem && PageDirty(page)) {
1800 * khugepaged only works on read-only fd, so this
1801 * page is dirty because it hasn't been flushed
1802 * since first write.
1808 if (isolate_lru_page(page)) {
1809 result = SCAN_DEL_PAGE_LRU;
1813 if (page_has_private(page) &&
1814 !try_to_release_page(page, GFP_KERNEL)) {
1815 result = SCAN_PAGE_HAS_PRIVATE;
1816 putback_lru_page(page);
1820 if (page_mapped(page))
1821 unmap_mapping_pages(mapping, index, 1, false);
1824 xas_set(&xas, index);
1826 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1827 VM_BUG_ON_PAGE(page_mapped(page), page);
1830 * The page is expected to have page_count() == 3:
1831 * - we hold a pin on it;
1832 * - one reference from page cache;
1833 * - one from isolate_lru_page;
1835 if (!page_ref_freeze(page, 3)) {
1836 result = SCAN_PAGE_COUNT;
1837 xas_unlock_irq(&xas);
1838 putback_lru_page(page);
1843 * Add the page to the list to be able to undo the collapse if
1844 * something go wrong.
1846 list_add_tail(&page->lru, &pagelist);
1848 /* Finally, replace with the new page. */
1849 xas_store(&xas, new_page);
1858 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1860 __inc_node_page_state(new_page, NR_FILE_THPS);
1861 filemap_nr_thps_inc(mapping);
1865 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1867 __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1871 xas_unlock_irq(&xas);
1874 if (result == SCAN_SUCCEED) {
1875 struct page *page, *tmp;
1878 * Replacing old pages with new one has succeeded, now we
1879 * need to copy the content and free the old pages.
1882 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1883 while (index < page->index) {
1884 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1887 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1889 list_del(&page->lru);
1890 page->mapping = NULL;
1891 page_ref_unfreeze(page, 1);
1892 ClearPageActive(page);
1893 ClearPageUnevictable(page);
1898 while (index < end) {
1899 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1903 SetPageUptodate(new_page);
1904 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1906 set_page_dirty(new_page);
1907 lru_cache_add(new_page);
1910 * Remove pte page tables, so we can re-fault the page as huge.
1912 retract_page_tables(mapping, start);
1915 khugepaged_pages_collapsed++;
1919 /* Something went wrong: roll back page cache changes */
1921 mapping->nrpages -= nr_none;
1924 shmem_uncharge(mapping->host, nr_none);
1926 xas_set(&xas, start);
1927 xas_for_each(&xas, page, end - 1) {
1928 page = list_first_entry_or_null(&pagelist,
1930 if (!page || xas.xa_index < page->index) {
1934 /* Put holes back where they were */
1935 xas_store(&xas, NULL);
1939 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1941 /* Unfreeze the page. */
1942 list_del(&page->lru);
1943 page_ref_unfreeze(page, 2);
1944 xas_store(&xas, page);
1946 xas_unlock_irq(&xas);
1948 putback_lru_page(page);
1952 xas_unlock_irq(&xas);
1954 new_page->mapping = NULL;
1957 unlock_page(new_page);
1959 VM_BUG_ON(!list_empty(&pagelist));
1960 if (!IS_ERR_OR_NULL(*hpage))
1961 mem_cgroup_uncharge(*hpage);
1962 /* TODO: tracepoints */
1965 static void khugepaged_scan_file(struct mm_struct *mm,
1966 struct file *file, pgoff_t start, struct page **hpage)
1968 struct page *page = NULL;
1969 struct address_space *mapping = file->f_mapping;
1970 XA_STATE(xas, &mapping->i_pages, start);
1972 int node = NUMA_NO_NODE;
1973 int result = SCAN_SUCCEED;
1977 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1979 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1980 if (xas_retry(&xas, page))
1983 if (xa_is_value(page)) {
1984 if (++swap > khugepaged_max_ptes_swap) {
1985 result = SCAN_EXCEED_SWAP_PTE;
1991 if (PageTransCompound(page)) {
1992 result = SCAN_PAGE_COMPOUND;
1996 node = page_to_nid(page);
1997 if (khugepaged_scan_abort(node)) {
1998 result = SCAN_SCAN_ABORT;
2001 khugepaged_node_load[node]++;
2003 if (!PageLRU(page)) {
2004 result = SCAN_PAGE_LRU;
2008 if (page_count(page) !=
2009 1 + page_mapcount(page) + page_has_private(page)) {
2010 result = SCAN_PAGE_COUNT;
2015 * We probably should check if the page is referenced here, but
2016 * nobody would transfer pte_young() to PageReferenced() for us.
2017 * And rmap walk here is just too costly...
2022 if (need_resched()) {
2029 if (result == SCAN_SUCCEED) {
2030 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2031 result = SCAN_EXCEED_NONE_PTE;
2033 node = khugepaged_find_target_node();
2034 collapse_file(mm, file, start, hpage, node);
2038 /* TODO: tracepoints */
2041 static void khugepaged_scan_file(struct mm_struct *mm,
2042 struct file *file, pgoff_t start, struct page **hpage)
2047 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2053 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2054 struct page **hpage)
2055 __releases(&khugepaged_mm_lock)
2056 __acquires(&khugepaged_mm_lock)
2058 struct mm_slot *mm_slot;
2059 struct mm_struct *mm;
2060 struct vm_area_struct *vma;
2064 lockdep_assert_held(&khugepaged_mm_lock);
2066 if (khugepaged_scan.mm_slot)
2067 mm_slot = khugepaged_scan.mm_slot;
2069 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2070 struct mm_slot, mm_node);
2071 khugepaged_scan.address = 0;
2072 khugepaged_scan.mm_slot = mm_slot;
2074 spin_unlock(&khugepaged_mm_lock);
2075 khugepaged_collapse_pte_mapped_thps(mm_slot);
2079 * Don't wait for semaphore (to avoid long wait times). Just move to
2080 * the next mm on the list.
2083 if (unlikely(!mmap_read_trylock(mm)))
2084 goto breakouterloop_mmap_lock;
2085 if (likely(!khugepaged_test_exit(mm)))
2086 vma = find_vma(mm, khugepaged_scan.address);
2089 for (; vma; vma = vma->vm_next) {
2090 unsigned long hstart, hend;
2093 if (unlikely(khugepaged_test_exit(mm))) {
2097 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2102 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2103 hend = vma->vm_end & HPAGE_PMD_MASK;
2106 if (khugepaged_scan.address > hend)
2108 if (khugepaged_scan.address < hstart)
2109 khugepaged_scan.address = hstart;
2110 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2111 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2114 while (khugepaged_scan.address < hend) {
2117 if (unlikely(khugepaged_test_exit(mm)))
2118 goto breakouterloop;
2120 VM_BUG_ON(khugepaged_scan.address < hstart ||
2121 khugepaged_scan.address + HPAGE_PMD_SIZE >
2123 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2124 struct file *file = get_file(vma->vm_file);
2125 pgoff_t pgoff = linear_page_index(vma,
2126 khugepaged_scan.address);
2128 mmap_read_unlock(mm);
2130 khugepaged_scan_file(mm, file, pgoff, hpage);
2133 ret = khugepaged_scan_pmd(mm, vma,
2134 khugepaged_scan.address,
2137 /* move to next address */
2138 khugepaged_scan.address += HPAGE_PMD_SIZE;
2139 progress += HPAGE_PMD_NR;
2141 /* we released mmap_lock so break loop */
2142 goto breakouterloop_mmap_lock;
2143 if (progress >= pages)
2144 goto breakouterloop;
2148 mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2149 breakouterloop_mmap_lock:
2151 spin_lock(&khugepaged_mm_lock);
2152 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2154 * Release the current mm_slot if this mm is about to die, or
2155 * if we scanned all vmas of this mm.
2157 if (khugepaged_test_exit(mm) || !vma) {
2159 * Make sure that if mm_users is reaching zero while
2160 * khugepaged runs here, khugepaged_exit will find
2161 * mm_slot not pointing to the exiting mm.
2163 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2164 khugepaged_scan.mm_slot = list_entry(
2165 mm_slot->mm_node.next,
2166 struct mm_slot, mm_node);
2167 khugepaged_scan.address = 0;
2169 khugepaged_scan.mm_slot = NULL;
2170 khugepaged_full_scans++;
2173 collect_mm_slot(mm_slot);
2179 static int khugepaged_has_work(void)
2181 return !list_empty(&khugepaged_scan.mm_head) &&
2182 khugepaged_enabled();
2185 static int khugepaged_wait_event(void)
2187 return !list_empty(&khugepaged_scan.mm_head) ||
2188 kthread_should_stop();
2191 static void khugepaged_do_scan(void)
2193 struct page *hpage = NULL;
2194 unsigned int progress = 0, pass_through_head = 0;
2195 unsigned int pages = khugepaged_pages_to_scan;
2198 barrier(); /* write khugepaged_pages_to_scan to local stack */
2200 lru_add_drain_all();
2202 while (progress < pages) {
2203 if (!khugepaged_prealloc_page(&hpage, &wait))
2208 if (unlikely(kthread_should_stop() || try_to_freeze()))
2211 spin_lock(&khugepaged_mm_lock);
2212 if (!khugepaged_scan.mm_slot)
2213 pass_through_head++;
2214 if (khugepaged_has_work() &&
2215 pass_through_head < 2)
2216 progress += khugepaged_scan_mm_slot(pages - progress,
2220 spin_unlock(&khugepaged_mm_lock);
2223 if (!IS_ERR_OR_NULL(hpage))
2227 static bool khugepaged_should_wakeup(void)
2229 return kthread_should_stop() ||
2230 time_after_eq(jiffies, khugepaged_sleep_expire);
2233 static void khugepaged_wait_work(void)
2235 if (khugepaged_has_work()) {
2236 const unsigned long scan_sleep_jiffies =
2237 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2239 if (!scan_sleep_jiffies)
2242 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2243 wait_event_freezable_timeout(khugepaged_wait,
2244 khugepaged_should_wakeup(),
2245 scan_sleep_jiffies);
2249 if (khugepaged_enabled())
2250 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2253 static int khugepaged(void *none)
2255 struct mm_slot *mm_slot;
2258 set_user_nice(current, MAX_NICE);
2260 while (!kthread_should_stop()) {
2261 khugepaged_do_scan();
2262 khugepaged_wait_work();
2265 spin_lock(&khugepaged_mm_lock);
2266 mm_slot = khugepaged_scan.mm_slot;
2267 khugepaged_scan.mm_slot = NULL;
2269 collect_mm_slot(mm_slot);
2270 spin_unlock(&khugepaged_mm_lock);
2274 static void set_recommended_min_free_kbytes(void)
2278 unsigned long recommended_min;
2280 for_each_populated_zone(zone) {
2282 * We don't need to worry about fragmentation of
2283 * ZONE_MOVABLE since it only has movable pages.
2285 if (zone_idx(zone) > gfp_zone(GFP_USER))
2291 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2292 recommended_min = pageblock_nr_pages * nr_zones * 2;
2295 * Make sure that on average at least two pageblocks are almost free
2296 * of another type, one for a migratetype to fall back to and a
2297 * second to avoid subsequent fallbacks of other types There are 3
2298 * MIGRATE_TYPES we care about.
2300 recommended_min += pageblock_nr_pages * nr_zones *
2301 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2303 /* don't ever allow to reserve more than 5% of the lowmem */
2304 recommended_min = min(recommended_min,
2305 (unsigned long) nr_free_buffer_pages() / 20);
2306 recommended_min <<= (PAGE_SHIFT-10);
2308 if (recommended_min > min_free_kbytes) {
2309 if (user_min_free_kbytes >= 0)
2310 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2311 min_free_kbytes, recommended_min);
2313 min_free_kbytes = recommended_min;
2315 setup_per_zone_wmarks();
2318 int start_stop_khugepaged(void)
2322 mutex_lock(&khugepaged_mutex);
2323 if (khugepaged_enabled()) {
2324 if (!khugepaged_thread)
2325 khugepaged_thread = kthread_run(khugepaged, NULL,
2327 if (IS_ERR(khugepaged_thread)) {
2328 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2329 err = PTR_ERR(khugepaged_thread);
2330 khugepaged_thread = NULL;
2334 if (!list_empty(&khugepaged_scan.mm_head))
2335 wake_up_interruptible(&khugepaged_wait);
2337 set_recommended_min_free_kbytes();
2338 } else if (khugepaged_thread) {
2339 kthread_stop(khugepaged_thread);
2340 khugepaged_thread = NULL;
2343 mutex_unlock(&khugepaged_mutex);
2347 void khugepaged_min_free_kbytes_update(void)
2349 mutex_lock(&khugepaged_mutex);
2350 if (khugepaged_enabled() && khugepaged_thread)
2351 set_recommended_min_free_kbytes();
2352 mutex_unlock(&khugepaged_mutex);