2 * Copyright (C) 2009 Red Hat, Inc.
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/sched.h>
12 #include <linux/highmem.h>
13 #include <linux/hugetlb.h>
14 #include <linux/mmu_notifier.h>
15 #include <linux/rmap.h>
16 #include <linux/swap.h>
17 #include <linux/shrinker.h>
18 #include <linux/mm_inline.h>
19 #include <linux/swapops.h>
20 #include <linux/dax.h>
21 #include <linux/kthread.h>
22 #include <linux/khugepaged.h>
23 #include <linux/freezer.h>
24 #include <linux/pfn_t.h>
25 #include <linux/mman.h>
26 #include <linux/memremap.h>
27 #include <linux/pagemap.h>
28 #include <linux/debugfs.h>
29 #include <linux/migrate.h>
30 #include <linux/hashtable.h>
31 #include <linux/userfaultfd_k.h>
32 #include <linux/page_idle.h>
35 #include <asm/pgalloc.h>
45 SCAN_NO_REFERENCED_PAGE,
59 SCAN_ALLOC_HUGE_PAGE_FAIL,
60 SCAN_CGROUP_CHARGE_FAIL
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/huge_memory.h>
67 * By default transparent hugepage support is disabled in order that avoid
68 * to risk increase the memory footprint of applications without a guaranteed
69 * benefit. When transparent hugepage support is enabled, is for all mappings,
70 * and khugepaged scans all mappings.
71 * Defrag is invoked by khugepaged hugepage allocations and by page faults
72 * for all hugepage allocations.
74 unsigned long transparent_hugepage_flags __read_mostly =
75 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
76 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
78 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
79 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
81 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
82 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
83 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
85 /* default scan 8*512 pte (or vmas) every 30 second */
86 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
87 static unsigned int khugepaged_pages_collapsed;
88 static unsigned int khugepaged_full_scans;
89 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
90 /* during fragmentation poll the hugepage allocator once every minute */
91 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
92 static struct task_struct *khugepaged_thread __read_mostly;
93 static DEFINE_MUTEX(khugepaged_mutex);
94 static DEFINE_SPINLOCK(khugepaged_mm_lock);
95 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
97 * default collapse hugepages if there is at least one pte mapped like
98 * it would have happened if the vma was large enough during page
101 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
103 static int khugepaged(void *none);
104 static int khugepaged_slab_init(void);
105 static void khugepaged_slab_exit(void);
107 #define MM_SLOTS_HASH_BITS 10
108 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
110 static struct kmem_cache *mm_slot_cache __read_mostly;
113 * struct mm_slot - hash lookup from mm to mm_slot
114 * @hash: hash collision list
115 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
116 * @mm: the mm that this information is valid for
119 struct hlist_node hash;
120 struct list_head mm_node;
121 struct mm_struct *mm;
125 * struct khugepaged_scan - cursor for scanning
126 * @mm_head: the head of the mm list to scan
127 * @mm_slot: the current mm_slot we are scanning
128 * @address: the next address inside that to be scanned
130 * There is only the one khugepaged_scan instance of this cursor structure.
132 struct khugepaged_scan {
133 struct list_head mm_head;
134 struct mm_slot *mm_slot;
135 unsigned long address;
137 static struct khugepaged_scan khugepaged_scan = {
138 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
141 static struct shrinker deferred_split_shrinker;
143 static void set_recommended_min_free_kbytes(void)
147 unsigned long recommended_min;
149 for_each_populated_zone(zone)
152 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
153 recommended_min = pageblock_nr_pages * nr_zones * 2;
156 * Make sure that on average at least two pageblocks are almost free
157 * of another type, one for a migratetype to fall back to and a
158 * second to avoid subsequent fallbacks of other types There are 3
159 * MIGRATE_TYPES we care about.
161 recommended_min += pageblock_nr_pages * nr_zones *
162 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
164 /* don't ever allow to reserve more than 5% of the lowmem */
165 recommended_min = min(recommended_min,
166 (unsigned long) nr_free_buffer_pages() / 20);
167 recommended_min <<= (PAGE_SHIFT-10);
169 if (recommended_min > min_free_kbytes) {
170 if (user_min_free_kbytes >= 0)
171 pr_info("raising min_free_kbytes from %d to %lu "
172 "to help transparent hugepage allocations\n",
173 min_free_kbytes, recommended_min);
175 min_free_kbytes = recommended_min;
177 setup_per_zone_wmarks();
180 static int start_stop_khugepaged(void)
183 if (khugepaged_enabled()) {
184 if (!khugepaged_thread)
185 khugepaged_thread = kthread_run(khugepaged, NULL,
187 if (IS_ERR(khugepaged_thread)) {
188 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
189 err = PTR_ERR(khugepaged_thread);
190 khugepaged_thread = NULL;
194 if (!list_empty(&khugepaged_scan.mm_head))
195 wake_up_interruptible(&khugepaged_wait);
197 set_recommended_min_free_kbytes();
198 } else if (khugepaged_thread) {
199 kthread_stop(khugepaged_thread);
200 khugepaged_thread = NULL;
206 static atomic_t huge_zero_refcount;
207 struct page *huge_zero_page __read_mostly;
209 struct page *get_huge_zero_page(void)
211 struct page *zero_page;
213 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
214 return READ_ONCE(huge_zero_page);
216 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
219 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
222 count_vm_event(THP_ZERO_PAGE_ALLOC);
224 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
226 __free_pages(zero_page, compound_order(zero_page));
230 /* We take additional reference here. It will be put back by shrinker */
231 atomic_set(&huge_zero_refcount, 2);
233 return READ_ONCE(huge_zero_page);
236 static void put_huge_zero_page(void)
239 * Counter should never go to zero here. Only shrinker can put
242 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
245 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
246 struct shrink_control *sc)
248 /* we can free zero page only if last reference remains */
249 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
252 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
253 struct shrink_control *sc)
255 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
256 struct page *zero_page = xchg(&huge_zero_page, NULL);
257 BUG_ON(zero_page == NULL);
258 __free_pages(zero_page, compound_order(zero_page));
265 static struct shrinker huge_zero_page_shrinker = {
266 .count_objects = shrink_huge_zero_page_count,
267 .scan_objects = shrink_huge_zero_page_scan,
268 .seeks = DEFAULT_SEEKS,
273 static ssize_t double_flag_show(struct kobject *kobj,
274 struct kobj_attribute *attr, char *buf,
275 enum transparent_hugepage_flag enabled,
276 enum transparent_hugepage_flag req_madv)
278 if (test_bit(enabled, &transparent_hugepage_flags)) {
279 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
280 return sprintf(buf, "[always] madvise never\n");
281 } else if (test_bit(req_madv, &transparent_hugepage_flags))
282 return sprintf(buf, "always [madvise] never\n");
284 return sprintf(buf, "always madvise [never]\n");
286 static ssize_t double_flag_store(struct kobject *kobj,
287 struct kobj_attribute *attr,
288 const char *buf, size_t count,
289 enum transparent_hugepage_flag enabled,
290 enum transparent_hugepage_flag req_madv)
292 if (!memcmp("always", buf,
293 min(sizeof("always")-1, count))) {
294 set_bit(enabled, &transparent_hugepage_flags);
295 clear_bit(req_madv, &transparent_hugepage_flags);
296 } else if (!memcmp("madvise", buf,
297 min(sizeof("madvise")-1, count))) {
298 clear_bit(enabled, &transparent_hugepage_flags);
299 set_bit(req_madv, &transparent_hugepage_flags);
300 } else if (!memcmp("never", buf,
301 min(sizeof("never")-1, count))) {
302 clear_bit(enabled, &transparent_hugepage_flags);
303 clear_bit(req_madv, &transparent_hugepage_flags);
310 static ssize_t enabled_show(struct kobject *kobj,
311 struct kobj_attribute *attr, char *buf)
313 return double_flag_show(kobj, attr, buf,
314 TRANSPARENT_HUGEPAGE_FLAG,
315 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
317 static ssize_t enabled_store(struct kobject *kobj,
318 struct kobj_attribute *attr,
319 const char *buf, size_t count)
323 ret = double_flag_store(kobj, attr, buf, count,
324 TRANSPARENT_HUGEPAGE_FLAG,
325 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
330 mutex_lock(&khugepaged_mutex);
331 err = start_stop_khugepaged();
332 mutex_unlock(&khugepaged_mutex);
340 static struct kobj_attribute enabled_attr =
341 __ATTR(enabled, 0644, enabled_show, enabled_store);
343 static ssize_t single_flag_show(struct kobject *kobj,
344 struct kobj_attribute *attr, char *buf,
345 enum transparent_hugepage_flag flag)
347 return sprintf(buf, "%d\n",
348 !!test_bit(flag, &transparent_hugepage_flags));
351 static ssize_t single_flag_store(struct kobject *kobj,
352 struct kobj_attribute *attr,
353 const char *buf, size_t count,
354 enum transparent_hugepage_flag flag)
359 ret = kstrtoul(buf, 10, &value);
366 set_bit(flag, &transparent_hugepage_flags);
368 clear_bit(flag, &transparent_hugepage_flags);
374 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
375 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
376 * memory just to allocate one more hugepage.
378 static ssize_t defrag_show(struct kobject *kobj,
379 struct kobj_attribute *attr, char *buf)
381 return double_flag_show(kobj, attr, buf,
382 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
383 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
385 static ssize_t defrag_store(struct kobject *kobj,
386 struct kobj_attribute *attr,
387 const char *buf, size_t count)
389 return double_flag_store(kobj, attr, buf, count,
390 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
391 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
393 static struct kobj_attribute defrag_attr =
394 __ATTR(defrag, 0644, defrag_show, defrag_store);
396 static ssize_t use_zero_page_show(struct kobject *kobj,
397 struct kobj_attribute *attr, char *buf)
399 return single_flag_show(kobj, attr, buf,
400 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
402 static ssize_t use_zero_page_store(struct kobject *kobj,
403 struct kobj_attribute *attr, const char *buf, size_t count)
405 return single_flag_store(kobj, attr, buf, count,
406 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
408 static struct kobj_attribute use_zero_page_attr =
409 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
410 #ifdef CONFIG_DEBUG_VM
411 static ssize_t debug_cow_show(struct kobject *kobj,
412 struct kobj_attribute *attr, char *buf)
414 return single_flag_show(kobj, attr, buf,
415 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
417 static ssize_t debug_cow_store(struct kobject *kobj,
418 struct kobj_attribute *attr,
419 const char *buf, size_t count)
421 return single_flag_store(kobj, attr, buf, count,
422 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
424 static struct kobj_attribute debug_cow_attr =
425 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
426 #endif /* CONFIG_DEBUG_VM */
428 static struct attribute *hugepage_attr[] = {
431 &use_zero_page_attr.attr,
432 #ifdef CONFIG_DEBUG_VM
433 &debug_cow_attr.attr,
438 static struct attribute_group hugepage_attr_group = {
439 .attrs = hugepage_attr,
442 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
443 struct kobj_attribute *attr,
446 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
449 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
450 struct kobj_attribute *attr,
451 const char *buf, size_t count)
456 err = kstrtoul(buf, 10, &msecs);
457 if (err || msecs > UINT_MAX)
460 khugepaged_scan_sleep_millisecs = msecs;
461 wake_up_interruptible(&khugepaged_wait);
465 static struct kobj_attribute scan_sleep_millisecs_attr =
466 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
467 scan_sleep_millisecs_store);
469 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
470 struct kobj_attribute *attr,
473 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
476 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
477 struct kobj_attribute *attr,
478 const char *buf, size_t count)
483 err = kstrtoul(buf, 10, &msecs);
484 if (err || msecs > UINT_MAX)
487 khugepaged_alloc_sleep_millisecs = msecs;
488 wake_up_interruptible(&khugepaged_wait);
492 static struct kobj_attribute alloc_sleep_millisecs_attr =
493 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
494 alloc_sleep_millisecs_store);
496 static ssize_t pages_to_scan_show(struct kobject *kobj,
497 struct kobj_attribute *attr,
500 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
502 static ssize_t pages_to_scan_store(struct kobject *kobj,
503 struct kobj_attribute *attr,
504 const char *buf, size_t count)
509 err = kstrtoul(buf, 10, &pages);
510 if (err || !pages || pages > UINT_MAX)
513 khugepaged_pages_to_scan = pages;
517 static struct kobj_attribute pages_to_scan_attr =
518 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
519 pages_to_scan_store);
521 static ssize_t pages_collapsed_show(struct kobject *kobj,
522 struct kobj_attribute *attr,
525 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
527 static struct kobj_attribute pages_collapsed_attr =
528 __ATTR_RO(pages_collapsed);
530 static ssize_t full_scans_show(struct kobject *kobj,
531 struct kobj_attribute *attr,
534 return sprintf(buf, "%u\n", khugepaged_full_scans);
536 static struct kobj_attribute full_scans_attr =
537 __ATTR_RO(full_scans);
539 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
540 struct kobj_attribute *attr, char *buf)
542 return single_flag_show(kobj, attr, buf,
543 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
545 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
546 struct kobj_attribute *attr,
547 const char *buf, size_t count)
549 return single_flag_store(kobj, attr, buf, count,
550 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
552 static struct kobj_attribute khugepaged_defrag_attr =
553 __ATTR(defrag, 0644, khugepaged_defrag_show,
554 khugepaged_defrag_store);
557 * max_ptes_none controls if khugepaged should collapse hugepages over
558 * any unmapped ptes in turn potentially increasing the memory
559 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
560 * reduce the available free memory in the system as it
561 * runs. Increasing max_ptes_none will instead potentially reduce the
562 * free memory in the system during the khugepaged scan.
564 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
565 struct kobj_attribute *attr,
568 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
570 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
571 struct kobj_attribute *attr,
572 const char *buf, size_t count)
575 unsigned long max_ptes_none;
577 err = kstrtoul(buf, 10, &max_ptes_none);
578 if (err || max_ptes_none > HPAGE_PMD_NR-1)
581 khugepaged_max_ptes_none = max_ptes_none;
585 static struct kobj_attribute khugepaged_max_ptes_none_attr =
586 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
587 khugepaged_max_ptes_none_store);
589 static struct attribute *khugepaged_attr[] = {
590 &khugepaged_defrag_attr.attr,
591 &khugepaged_max_ptes_none_attr.attr,
592 &pages_to_scan_attr.attr,
593 &pages_collapsed_attr.attr,
594 &full_scans_attr.attr,
595 &scan_sleep_millisecs_attr.attr,
596 &alloc_sleep_millisecs_attr.attr,
600 static struct attribute_group khugepaged_attr_group = {
601 .attrs = khugepaged_attr,
602 .name = "khugepaged",
605 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
609 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
610 if (unlikely(!*hugepage_kobj)) {
611 pr_err("failed to create transparent hugepage kobject\n");
615 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
617 pr_err("failed to register transparent hugepage group\n");
621 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
623 pr_err("failed to register transparent hugepage group\n");
624 goto remove_hp_group;
630 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
632 kobject_put(*hugepage_kobj);
636 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
638 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
639 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
640 kobject_put(hugepage_kobj);
643 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
648 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
651 #endif /* CONFIG_SYSFS */
653 static int __init hugepage_init(void)
656 struct kobject *hugepage_kobj;
658 if (!has_transparent_hugepage()) {
659 transparent_hugepage_flags = 0;
663 err = hugepage_init_sysfs(&hugepage_kobj);
667 err = khugepaged_slab_init();
671 err = register_shrinker(&huge_zero_page_shrinker);
673 goto err_hzp_shrinker;
674 err = register_shrinker(&deferred_split_shrinker);
676 goto err_split_shrinker;
679 * By default disable transparent hugepages on smaller systems,
680 * where the extra memory used could hurt more than TLB overhead
681 * is likely to save. The admin can still enable it through /sys.
683 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
684 transparent_hugepage_flags = 0;
688 err = start_stop_khugepaged();
694 unregister_shrinker(&deferred_split_shrinker);
696 unregister_shrinker(&huge_zero_page_shrinker);
698 khugepaged_slab_exit();
700 hugepage_exit_sysfs(hugepage_kobj);
704 subsys_initcall(hugepage_init);
706 static int __init setup_transparent_hugepage(char *str)
711 if (!strcmp(str, "always")) {
712 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
713 &transparent_hugepage_flags);
714 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
715 &transparent_hugepage_flags);
717 } else if (!strcmp(str, "madvise")) {
718 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
719 &transparent_hugepage_flags);
720 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
721 &transparent_hugepage_flags);
723 } else if (!strcmp(str, "never")) {
724 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
725 &transparent_hugepage_flags);
726 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
727 &transparent_hugepage_flags);
732 pr_warn("transparent_hugepage= cannot parse, ignored\n");
735 __setup("transparent_hugepage=", setup_transparent_hugepage);
737 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
739 if (likely(vma->vm_flags & VM_WRITE))
740 pmd = pmd_mkwrite(pmd);
744 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
747 entry = mk_pmd(page, prot);
748 entry = pmd_mkhuge(entry);
752 static inline struct list_head *page_deferred_list(struct page *page)
755 * ->lru in the tail pages is occupied by compound_head.
756 * Let's use ->mapping + ->index in the second tail page as list_head.
758 return (struct list_head *)&page[2].mapping;
761 void prep_transhuge_page(struct page *page)
764 * we use page->mapping and page->indexlru in second tail page
765 * as list_head: assuming THP order >= 2
767 BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
769 INIT_LIST_HEAD(page_deferred_list(page));
770 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
773 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
774 struct vm_area_struct *vma,
775 unsigned long address, pmd_t *pmd,
776 struct page *page, gfp_t gfp,
779 struct mem_cgroup *memcg;
782 unsigned long haddr = address & HPAGE_PMD_MASK;
784 VM_BUG_ON_PAGE(!PageCompound(page), page);
786 if (mem_cgroup_try_charge(page, mm, gfp, &memcg, true)) {
788 count_vm_event(THP_FAULT_FALLBACK);
789 return VM_FAULT_FALLBACK;
792 pgtable = pte_alloc_one(mm, haddr);
793 if (unlikely(!pgtable)) {
794 mem_cgroup_cancel_charge(page, memcg, true);
799 clear_huge_page(page, haddr, HPAGE_PMD_NR);
801 * The memory barrier inside __SetPageUptodate makes sure that
802 * clear_huge_page writes become visible before the set_pmd_at()
805 __SetPageUptodate(page);
807 ptl = pmd_lock(mm, pmd);
808 if (unlikely(!pmd_none(*pmd))) {
810 mem_cgroup_cancel_charge(page, memcg, true);
812 pte_free(mm, pgtable);
816 /* Deliver the page fault to userland */
817 if (userfaultfd_missing(vma)) {
821 mem_cgroup_cancel_charge(page, memcg, true);
823 pte_free(mm, pgtable);
824 ret = handle_userfault(vma, address, flags,
826 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
830 entry = mk_huge_pmd(page, vma->vm_page_prot);
831 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
832 page_add_new_anon_rmap(page, vma, haddr, true);
833 mem_cgroup_commit_charge(page, memcg, false, true);
834 lru_cache_add_active_or_unevictable(page, vma);
835 pgtable_trans_huge_deposit(mm, pmd, pgtable);
836 set_pmd_at(mm, haddr, pmd, entry);
837 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
838 atomic_long_inc(&mm->nr_ptes);
840 count_vm_event(THP_FAULT_ALLOC);
846 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
848 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp;
851 /* Caller must hold page table lock. */
852 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
853 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
854 struct page *zero_page)
859 entry = mk_pmd(zero_page, vma->vm_page_prot);
860 entry = pmd_mkhuge(entry);
862 pgtable_trans_huge_deposit(mm, pmd, pgtable);
863 set_pmd_at(mm, haddr, pmd, entry);
864 atomic_long_inc(&mm->nr_ptes);
868 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
869 unsigned long address, pmd_t *pmd,
874 unsigned long haddr = address & HPAGE_PMD_MASK;
876 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
877 return VM_FAULT_FALLBACK;
878 if (unlikely(anon_vma_prepare(vma)))
880 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
882 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
883 transparent_hugepage_use_zero_page()) {
886 struct page *zero_page;
889 pgtable = pte_alloc_one(mm, haddr);
890 if (unlikely(!pgtable))
892 zero_page = get_huge_zero_page();
893 if (unlikely(!zero_page)) {
894 pte_free(mm, pgtable);
895 count_vm_event(THP_FAULT_FALLBACK);
896 return VM_FAULT_FALLBACK;
898 ptl = pmd_lock(mm, pmd);
901 if (pmd_none(*pmd)) {
902 if (userfaultfd_missing(vma)) {
904 ret = handle_userfault(vma, address, flags,
906 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
908 set_huge_zero_page(pgtable, mm, vma,
917 pte_free(mm, pgtable);
918 put_huge_zero_page();
922 gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
923 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
924 if (unlikely(!page)) {
925 count_vm_event(THP_FAULT_FALLBACK);
926 return VM_FAULT_FALLBACK;
928 prep_transhuge_page(page);
929 return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
933 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
934 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
936 struct mm_struct *mm = vma->vm_mm;
940 ptl = pmd_lock(mm, pmd);
941 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
942 if (pfn_t_devmap(pfn))
943 entry = pmd_mkdevmap(entry);
945 entry = pmd_mkyoung(pmd_mkdirty(entry));
946 entry = maybe_pmd_mkwrite(entry, vma);
948 set_pmd_at(mm, addr, pmd, entry);
949 update_mmu_cache_pmd(vma, addr, pmd);
953 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
954 pmd_t *pmd, pfn_t pfn, bool write)
956 pgprot_t pgprot = vma->vm_page_prot;
958 * If we had pmd_special, we could avoid all these restrictions,
959 * but we need to be consistent with PTEs and architectures that
960 * can't support a 'special' bit.
962 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
963 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
964 (VM_PFNMAP|VM_MIXEDMAP));
965 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
966 BUG_ON(!pfn_t_devmap(pfn));
968 if (addr < vma->vm_start || addr >= vma->vm_end)
969 return VM_FAULT_SIGBUS;
970 if (track_pfn_insert(vma, &pgprot, pfn))
971 return VM_FAULT_SIGBUS;
972 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
973 return VM_FAULT_NOPAGE;
976 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
982 * We should set the dirty bit only for FOLL_WRITE but for now
983 * the dirty bit in the pmd is meaningless. And if the dirty
984 * bit will become meaningful and we'll only set it with
985 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
986 * set the young bit, instead of the current set_pmd_at.
988 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
989 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
991 update_mmu_cache_pmd(vma, addr, pmd);
994 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
995 pmd_t *pmd, int flags)
997 unsigned long pfn = pmd_pfn(*pmd);
998 struct mm_struct *mm = vma->vm_mm;
999 struct dev_pagemap *pgmap;
1002 assert_spin_locked(pmd_lockptr(mm, pmd));
1004 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1007 if (pmd_present(*pmd) && pmd_devmap(*pmd))
1012 if (flags & FOLL_TOUCH)
1013 touch_pmd(vma, addr, pmd);
1016 * device mapped pages can only be returned if the
1017 * caller will manage the page reference count.
1019 if (!(flags & FOLL_GET))
1020 return ERR_PTR(-EEXIST);
1022 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1023 pgmap = get_dev_pagemap(pfn, NULL);
1025 return ERR_PTR(-EFAULT);
1026 page = pfn_to_page(pfn);
1028 put_dev_pagemap(pgmap);
1033 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1034 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1035 struct vm_area_struct *vma)
1037 spinlock_t *dst_ptl, *src_ptl;
1038 struct page *src_page;
1040 pgtable_t pgtable = NULL;
1043 if (!vma_is_dax(vma)) {
1045 pgtable = pte_alloc_one(dst_mm, addr);
1046 if (unlikely(!pgtable))
1050 dst_ptl = pmd_lock(dst_mm, dst_pmd);
1051 src_ptl = pmd_lockptr(src_mm, src_pmd);
1052 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1056 if (unlikely(!pmd_trans_huge(pmd) && !pmd_devmap(pmd))) {
1057 pte_free(dst_mm, pgtable);
1061 * When page table lock is held, the huge zero pmd should not be
1062 * under splitting since we don't split the page itself, only pmd to
1065 if (is_huge_zero_pmd(pmd)) {
1066 struct page *zero_page;
1068 * get_huge_zero_page() will never allocate a new page here,
1069 * since we already have a zero page to copy. It just takes a
1072 zero_page = get_huge_zero_page();
1073 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1079 if (!vma_is_dax(vma)) {
1080 /* thp accounting separate from pmd_devmap accounting */
1081 src_page = pmd_page(pmd);
1082 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1084 page_dup_rmap(src_page, true);
1085 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1086 atomic_long_inc(&dst_mm->nr_ptes);
1087 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1090 pmdp_set_wrprotect(src_mm, addr, src_pmd);
1091 pmd = pmd_mkold(pmd_wrprotect(pmd));
1092 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1096 spin_unlock(src_ptl);
1097 spin_unlock(dst_ptl);
1102 void huge_pmd_set_accessed(struct mm_struct *mm,
1103 struct vm_area_struct *vma,
1104 unsigned long address,
1105 pmd_t *pmd, pmd_t orig_pmd,
1110 unsigned long haddr;
1112 ptl = pmd_lock(mm, pmd);
1113 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1116 entry = pmd_mkyoung(orig_pmd);
1117 haddr = address & HPAGE_PMD_MASK;
1118 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1119 update_mmu_cache_pmd(vma, address, pmd);
1125 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1126 struct vm_area_struct *vma,
1127 unsigned long address,
1128 pmd_t *pmd, pmd_t orig_pmd,
1130 unsigned long haddr)
1132 struct mem_cgroup *memcg;
1137 struct page **pages;
1138 unsigned long mmun_start; /* For mmu_notifiers */
1139 unsigned long mmun_end; /* For mmu_notifiers */
1141 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1143 if (unlikely(!pages)) {
1144 ret |= VM_FAULT_OOM;
1148 for (i = 0; i < HPAGE_PMD_NR; i++) {
1149 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1151 vma, address, page_to_nid(page));
1152 if (unlikely(!pages[i] ||
1153 mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1158 memcg = (void *)page_private(pages[i]);
1159 set_page_private(pages[i], 0);
1160 mem_cgroup_cancel_charge(pages[i], memcg,
1165 ret |= VM_FAULT_OOM;
1168 set_page_private(pages[i], (unsigned long)memcg);
1171 for (i = 0; i < HPAGE_PMD_NR; i++) {
1172 copy_user_highpage(pages[i], page + i,
1173 haddr + PAGE_SIZE * i, vma);
1174 __SetPageUptodate(pages[i]);
1179 mmun_end = haddr + HPAGE_PMD_SIZE;
1180 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1182 ptl = pmd_lock(mm, pmd);
1183 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1184 goto out_free_pages;
1185 VM_BUG_ON_PAGE(!PageHead(page), page);
1187 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1188 /* leave pmd empty until pte is filled */
1190 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1191 pmd_populate(mm, &_pmd, pgtable);
1193 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1195 entry = mk_pte(pages[i], vma->vm_page_prot);
1196 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1197 memcg = (void *)page_private(pages[i]);
1198 set_page_private(pages[i], 0);
1199 page_add_new_anon_rmap(pages[i], vma, haddr, false);
1200 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1201 lru_cache_add_active_or_unevictable(pages[i], vma);
1202 pte = pte_offset_map(&_pmd, haddr);
1203 VM_BUG_ON(!pte_none(*pte));
1204 set_pte_at(mm, haddr, pte, entry);
1209 smp_wmb(); /* make pte visible before pmd */
1210 pmd_populate(mm, pmd, pgtable);
1211 page_remove_rmap(page, true);
1214 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1216 ret |= VM_FAULT_WRITE;
1224 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1225 for (i = 0; i < HPAGE_PMD_NR; i++) {
1226 memcg = (void *)page_private(pages[i]);
1227 set_page_private(pages[i], 0);
1228 mem_cgroup_cancel_charge(pages[i], memcg, false);
1235 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1236 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1240 struct page *page = NULL, *new_page;
1241 struct mem_cgroup *memcg;
1242 unsigned long haddr;
1243 unsigned long mmun_start; /* For mmu_notifiers */
1244 unsigned long mmun_end; /* For mmu_notifiers */
1245 gfp_t huge_gfp; /* for allocation and charge */
1247 ptl = pmd_lockptr(mm, pmd);
1248 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1249 haddr = address & HPAGE_PMD_MASK;
1250 if (is_huge_zero_pmd(orig_pmd))
1253 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1256 page = pmd_page(orig_pmd);
1257 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1259 * We can only reuse the page if nobody else maps the huge page or it's
1260 * part. We can do it by checking page_mapcount() on each sub-page, but
1262 * The cheaper way is to check page_count() to be equal 1: every
1263 * mapcount takes page reference reference, so this way we can
1264 * guarantee, that the PMD is the only mapping.
1265 * This can give false negative if somebody pinned the page, but that's
1268 if (page_mapcount(page) == 1 && page_count(page) == 1) {
1270 entry = pmd_mkyoung(orig_pmd);
1271 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1272 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
1273 update_mmu_cache_pmd(vma, address, pmd);
1274 ret |= VM_FAULT_WRITE;
1280 if (transparent_hugepage_enabled(vma) &&
1281 !transparent_hugepage_debug_cow()) {
1282 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1283 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1287 if (likely(new_page)) {
1288 prep_transhuge_page(new_page);
1291 split_huge_pmd(vma, pmd, address);
1292 ret |= VM_FAULT_FALLBACK;
1294 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1295 pmd, orig_pmd, page, haddr);
1296 if (ret & VM_FAULT_OOM) {
1297 split_huge_pmd(vma, pmd, address);
1298 ret |= VM_FAULT_FALLBACK;
1302 count_vm_event(THP_FAULT_FALLBACK);
1306 if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg,
1310 split_huge_pmd(vma, pmd, address);
1313 split_huge_pmd(vma, pmd, address);
1314 ret |= VM_FAULT_FALLBACK;
1315 count_vm_event(THP_FAULT_FALLBACK);
1319 count_vm_event(THP_FAULT_ALLOC);
1322 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1324 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1325 __SetPageUptodate(new_page);
1328 mmun_end = haddr + HPAGE_PMD_SIZE;
1329 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1334 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1336 mem_cgroup_cancel_charge(new_page, memcg, true);
1341 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1342 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1343 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1344 page_add_new_anon_rmap(new_page, vma, haddr, true);
1345 mem_cgroup_commit_charge(new_page, memcg, false, true);
1346 lru_cache_add_active_or_unevictable(new_page, vma);
1347 set_pmd_at(mm, haddr, pmd, entry);
1348 update_mmu_cache_pmd(vma, address, pmd);
1350 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1351 put_huge_zero_page();
1353 VM_BUG_ON_PAGE(!PageHead(page), page);
1354 page_remove_rmap(page, true);
1357 ret |= VM_FAULT_WRITE;
1361 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1369 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1374 struct mm_struct *mm = vma->vm_mm;
1375 struct page *page = NULL;
1377 assert_spin_locked(pmd_lockptr(mm, pmd));
1379 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1382 /* Avoid dumping huge zero page */
1383 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1384 return ERR_PTR(-EFAULT);
1386 /* Full NUMA hinting faults to serialise migration in fault paths */
1387 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1390 page = pmd_page(*pmd);
1391 VM_BUG_ON_PAGE(!PageHead(page), page);
1392 if (flags & FOLL_TOUCH)
1393 touch_pmd(vma, addr, pmd);
1394 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1396 * We don't mlock() pte-mapped THPs. This way we can avoid
1397 * leaking mlocked pages into non-VM_LOCKED VMAs.
1399 * In most cases the pmd is the only mapping of the page as we
1400 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1401 * writable private mappings in populate_vma_page_range().
1403 * The only scenario when we have the page shared here is if we
1404 * mlocking read-only mapping shared over fork(). We skip
1405 * mlocking such pages.
1407 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1408 page->mapping && trylock_page(page)) {
1411 mlock_vma_page(page);
1415 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1416 VM_BUG_ON_PAGE(!PageCompound(page), page);
1417 if (flags & FOLL_GET)
1424 /* NUMA hinting page fault entry point for trans huge pmds */
1425 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1426 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1429 struct anon_vma *anon_vma = NULL;
1431 unsigned long haddr = addr & HPAGE_PMD_MASK;
1432 int page_nid = -1, this_nid = numa_node_id();
1433 int target_nid, last_cpupid = -1;
1435 bool migrated = false;
1439 /* A PROT_NONE fault should not end up here */
1440 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1442 ptl = pmd_lock(mm, pmdp);
1443 if (unlikely(!pmd_same(pmd, *pmdp)))
1447 * If there are potential migrations, wait for completion and retry
1448 * without disrupting NUMA hinting information. Do not relock and
1449 * check_same as the page may no longer be mapped.
1451 if (unlikely(pmd_trans_migrating(*pmdp))) {
1452 page = pmd_page(*pmdp);
1454 wait_on_page_locked(page);
1458 page = pmd_page(pmd);
1459 BUG_ON(is_huge_zero_page(page));
1460 page_nid = page_to_nid(page);
1461 last_cpupid = page_cpupid_last(page);
1462 count_vm_numa_event(NUMA_HINT_FAULTS);
1463 if (page_nid == this_nid) {
1464 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1465 flags |= TNF_FAULT_LOCAL;
1468 /* See similar comment in do_numa_page for explanation */
1469 if (!(vma->vm_flags & VM_WRITE))
1470 flags |= TNF_NO_GROUP;
1473 * Acquire the page lock to serialise THP migrations but avoid dropping
1474 * page_table_lock if at all possible
1476 page_locked = trylock_page(page);
1477 target_nid = mpol_misplaced(page, vma, haddr);
1478 if (target_nid == -1) {
1479 /* If the page was locked, there are no parallel migrations */
1484 /* Migration could have started since the pmd_trans_migrating check */
1487 wait_on_page_locked(page);
1493 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1494 * to serialises splits
1498 anon_vma = page_lock_anon_vma_read(page);
1500 /* Confirm the PMD did not change while page_table_lock was released */
1502 if (unlikely(!pmd_same(pmd, *pmdp))) {
1509 /* Bail if we fail to protect against THP splits for any reason */
1510 if (unlikely(!anon_vma)) {
1517 * Migrate the THP to the requested node, returns with page unlocked
1518 * and access rights restored.
1521 migrated = migrate_misplaced_transhuge_page(mm, vma,
1522 pmdp, pmd, addr, page, target_nid);
1524 flags |= TNF_MIGRATED;
1525 page_nid = target_nid;
1527 flags |= TNF_MIGRATE_FAIL;
1531 BUG_ON(!PageLocked(page));
1532 was_writable = pmd_write(pmd);
1533 pmd = pmd_modify(pmd, vma->vm_page_prot);
1534 pmd = pmd_mkyoung(pmd);
1536 pmd = pmd_mkwrite(pmd);
1537 set_pmd_at(mm, haddr, pmdp, pmd);
1538 update_mmu_cache_pmd(vma, addr, pmdp);
1545 page_unlock_anon_vma_read(anon_vma);
1548 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1553 int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1554 pmd_t *pmd, unsigned long addr, unsigned long next)
1560 struct mm_struct *mm = tlb->mm;
1563 ptl = pmd_trans_huge_lock(pmd, vma);
1568 if (is_huge_zero_pmd(orig_pmd)) {
1573 page = pmd_page(orig_pmd);
1575 * If other processes are mapping this page, we couldn't discard
1576 * the page unless they all do MADV_FREE so let's skip the page.
1578 if (page_mapcount(page) != 1)
1581 if (!trylock_page(page))
1585 * If user want to discard part-pages of THP, split it so MADV_FREE
1586 * will deactivate only them.
1588 if (next - addr != HPAGE_PMD_SIZE) {
1591 if (split_huge_page(page)) {
1602 if (PageDirty(page))
1603 ClearPageDirty(page);
1606 if (PageActive(page))
1607 deactivate_page(page);
1609 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1610 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1612 orig_pmd = pmd_mkold(orig_pmd);
1613 orig_pmd = pmd_mkclean(orig_pmd);
1615 set_pmd_at(mm, addr, pmd, orig_pmd);
1616 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1625 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1626 pmd_t *pmd, unsigned long addr)
1631 ptl = __pmd_trans_huge_lock(pmd, vma);
1635 * For architectures like ppc64 we look at deposited pgtable
1636 * when calling pmdp_huge_get_and_clear. So do the
1637 * pgtable_trans_huge_withdraw after finishing pmdp related
1640 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1642 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1643 if (vma_is_dax(vma)) {
1645 if (is_huge_zero_pmd(orig_pmd))
1646 put_huge_zero_page();
1647 } else if (is_huge_zero_pmd(orig_pmd)) {
1648 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1649 atomic_long_dec(&tlb->mm->nr_ptes);
1651 put_huge_zero_page();
1653 struct page *page = pmd_page(orig_pmd);
1654 page_remove_rmap(page, true);
1655 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1656 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1657 VM_BUG_ON_PAGE(!PageHead(page), page);
1658 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1659 atomic_long_dec(&tlb->mm->nr_ptes);
1661 tlb_remove_page(tlb, page);
1666 bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1667 unsigned long old_addr,
1668 unsigned long new_addr, unsigned long old_end,
1669 pmd_t *old_pmd, pmd_t *new_pmd)
1671 spinlock_t *old_ptl, *new_ptl;
1674 struct mm_struct *mm = vma->vm_mm;
1676 if ((old_addr & ~HPAGE_PMD_MASK) ||
1677 (new_addr & ~HPAGE_PMD_MASK) ||
1678 old_end - old_addr < HPAGE_PMD_SIZE ||
1679 (new_vma->vm_flags & VM_NOHUGEPAGE))
1683 * The destination pmd shouldn't be established, free_pgtables()
1684 * should have release it.
1686 if (WARN_ON(!pmd_none(*new_pmd))) {
1687 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1692 * We don't have to worry about the ordering of src and dst
1693 * ptlocks because exclusive mmap_sem prevents deadlock.
1695 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1697 new_ptl = pmd_lockptr(mm, new_pmd);
1698 if (new_ptl != old_ptl)
1699 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1700 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1701 VM_BUG_ON(!pmd_none(*new_pmd));
1703 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1704 vma_is_anonymous(vma)) {
1706 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1707 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1709 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1710 if (new_ptl != old_ptl)
1711 spin_unlock(new_ptl);
1712 spin_unlock(old_ptl);
1720 * - 0 if PMD could not be locked
1721 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1722 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1724 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1725 unsigned long addr, pgprot_t newprot, int prot_numa)
1727 struct mm_struct *mm = vma->vm_mm;
1731 ptl = __pmd_trans_huge_lock(pmd, vma);
1734 bool preserve_write = prot_numa && pmd_write(*pmd);
1738 * Avoid trapping faults against the zero page. The read-only
1739 * data is likely to be read-cached on the local CPU and
1740 * local/remote hits to the zero page are not interesting.
1742 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1747 if (!prot_numa || !pmd_protnone(*pmd)) {
1748 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1749 entry = pmd_modify(entry, newprot);
1751 entry = pmd_mkwrite(entry);
1753 set_pmd_at(mm, addr, pmd, entry);
1754 BUG_ON(!preserve_write && pmd_write(entry));
1763 * Returns true if a given pmd maps a thp, false otherwise.
1765 * Note that if it returns true, this routine returns without unlocking page
1766 * table lock. So callers must unlock it.
1768 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1771 ptl = pmd_lock(vma->vm_mm, pmd);
1772 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1778 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1780 int hugepage_madvise(struct vm_area_struct *vma,
1781 unsigned long *vm_flags, int advice)
1787 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1788 * can't handle this properly after s390_enable_sie, so we simply
1789 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1791 if (mm_has_pgste(vma->vm_mm))
1795 * Be somewhat over-protective like KSM for now!
1797 if (*vm_flags & VM_NO_THP)
1799 *vm_flags &= ~VM_NOHUGEPAGE;
1800 *vm_flags |= VM_HUGEPAGE;
1802 * If the vma become good for khugepaged to scan,
1803 * register it here without waiting a page fault that
1804 * may not happen any time soon.
1806 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1809 case MADV_NOHUGEPAGE:
1811 * Be somewhat over-protective like KSM for now!
1813 if (*vm_flags & VM_NO_THP)
1815 *vm_flags &= ~VM_HUGEPAGE;
1816 *vm_flags |= VM_NOHUGEPAGE;
1818 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1819 * this vma even if we leave the mm registered in khugepaged if
1820 * it got registered before VM_NOHUGEPAGE was set.
1828 static int __init khugepaged_slab_init(void)
1830 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1831 sizeof(struct mm_slot),
1832 __alignof__(struct mm_slot), 0, NULL);
1839 static void __init khugepaged_slab_exit(void)
1841 kmem_cache_destroy(mm_slot_cache);
1844 static inline struct mm_slot *alloc_mm_slot(void)
1846 if (!mm_slot_cache) /* initialization failed */
1848 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1851 static inline void free_mm_slot(struct mm_slot *mm_slot)
1853 kmem_cache_free(mm_slot_cache, mm_slot);
1856 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1858 struct mm_slot *mm_slot;
1860 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1861 if (mm == mm_slot->mm)
1867 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1868 struct mm_slot *mm_slot)
1871 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1874 static inline int khugepaged_test_exit(struct mm_struct *mm)
1876 return atomic_read(&mm->mm_users) == 0;
1879 int __khugepaged_enter(struct mm_struct *mm)
1881 struct mm_slot *mm_slot;
1884 mm_slot = alloc_mm_slot();
1888 /* __khugepaged_exit() must not run from under us */
1889 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1890 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1891 free_mm_slot(mm_slot);
1895 spin_lock(&khugepaged_mm_lock);
1896 insert_to_mm_slots_hash(mm, mm_slot);
1898 * Insert just behind the scanning cursor, to let the area settle
1901 wakeup = list_empty(&khugepaged_scan.mm_head);
1902 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1903 spin_unlock(&khugepaged_mm_lock);
1905 atomic_inc(&mm->mm_count);
1907 wake_up_interruptible(&khugepaged_wait);
1912 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1913 unsigned long vm_flags)
1915 unsigned long hstart, hend;
1918 * Not yet faulted in so we will register later in the
1919 * page fault if needed.
1923 /* khugepaged not yet working on file or special mappings */
1925 VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
1926 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1927 hend = vma->vm_end & HPAGE_PMD_MASK;
1929 return khugepaged_enter(vma, vm_flags);
1933 void __khugepaged_exit(struct mm_struct *mm)
1935 struct mm_slot *mm_slot;
1938 spin_lock(&khugepaged_mm_lock);
1939 mm_slot = get_mm_slot(mm);
1940 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1941 hash_del(&mm_slot->hash);
1942 list_del(&mm_slot->mm_node);
1945 spin_unlock(&khugepaged_mm_lock);
1948 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1949 free_mm_slot(mm_slot);
1951 } else if (mm_slot) {
1953 * This is required to serialize against
1954 * khugepaged_test_exit() (which is guaranteed to run
1955 * under mmap sem read mode). Stop here (after we
1956 * return all pagetables will be destroyed) until
1957 * khugepaged has finished working on the pagetables
1958 * under the mmap_sem.
1960 down_write(&mm->mmap_sem);
1961 up_write(&mm->mmap_sem);
1965 static void release_pte_page(struct page *page)
1967 /* 0 stands for page_is_file_cache(page) == false */
1968 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1970 putback_lru_page(page);
1973 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1975 while (--_pte >= pte) {
1976 pte_t pteval = *_pte;
1977 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
1978 release_pte_page(pte_page(pteval));
1982 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1983 unsigned long address,
1986 struct page *page = NULL;
1988 int none_or_zero = 0, result = 0;
1989 bool referenced = false, writable = false;
1991 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1992 _pte++, address += PAGE_SIZE) {
1993 pte_t pteval = *_pte;
1994 if (pte_none(pteval) || (pte_present(pteval) &&
1995 is_zero_pfn(pte_pfn(pteval)))) {
1996 if (!userfaultfd_armed(vma) &&
1997 ++none_or_zero <= khugepaged_max_ptes_none) {
2000 result = SCAN_EXCEED_NONE_PTE;
2004 if (!pte_present(pteval)) {
2005 result = SCAN_PTE_NON_PRESENT;
2008 page = vm_normal_page(vma, address, pteval);
2009 if (unlikely(!page)) {
2010 result = SCAN_PAGE_NULL;
2014 VM_BUG_ON_PAGE(PageCompound(page), page);
2015 VM_BUG_ON_PAGE(!PageAnon(page), page);
2016 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2019 * We can do it before isolate_lru_page because the
2020 * page can't be freed from under us. NOTE: PG_lock
2021 * is needed to serialize against split_huge_page
2022 * when invoked from the VM.
2024 if (!trylock_page(page)) {
2025 result = SCAN_PAGE_LOCK;
2030 * cannot use mapcount: can't collapse if there's a gup pin.
2031 * The page must only be referenced by the scanned process
2032 * and page swap cache.
2034 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2036 result = SCAN_PAGE_COUNT;
2039 if (pte_write(pteval)) {
2042 if (PageSwapCache(page) && !reuse_swap_page(page)) {
2044 result = SCAN_SWAP_CACHE_PAGE;
2048 * Page is not in the swap cache. It can be collapsed
2054 * Isolate the page to avoid collapsing an hugepage
2055 * currently in use by the VM.
2057 if (isolate_lru_page(page)) {
2059 result = SCAN_DEL_PAGE_LRU;
2062 /* 0 stands for page_is_file_cache(page) == false */
2063 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2064 VM_BUG_ON_PAGE(!PageLocked(page), page);
2065 VM_BUG_ON_PAGE(PageLRU(page), page);
2067 /* If there is no mapped pte young don't collapse the page */
2068 if (pte_young(pteval) ||
2069 page_is_young(page) || PageReferenced(page) ||
2070 mmu_notifier_test_young(vma->vm_mm, address))
2073 if (likely(writable)) {
2074 if (likely(referenced)) {
2075 result = SCAN_SUCCEED;
2076 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2077 referenced, writable, result);
2081 result = SCAN_PAGE_RO;
2085 release_pte_pages(pte, _pte);
2086 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2087 referenced, writable, result);
2091 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2092 struct vm_area_struct *vma,
2093 unsigned long address,
2097 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2098 pte_t pteval = *_pte;
2099 struct page *src_page;
2101 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2102 clear_user_highpage(page, address);
2103 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2104 if (is_zero_pfn(pte_pfn(pteval))) {
2106 * ptl mostly unnecessary.
2110 * paravirt calls inside pte_clear here are
2113 pte_clear(vma->vm_mm, address, _pte);
2117 src_page = pte_page(pteval);
2118 copy_user_highpage(page, src_page, address, vma);
2119 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2120 release_pte_page(src_page);
2122 * ptl mostly unnecessary, but preempt has to
2123 * be disabled to update the per-cpu stats
2124 * inside page_remove_rmap().
2128 * paravirt calls inside pte_clear here are
2131 pte_clear(vma->vm_mm, address, _pte);
2132 page_remove_rmap(src_page, false);
2134 free_page_and_swap_cache(src_page);
2137 address += PAGE_SIZE;
2142 static void khugepaged_alloc_sleep(void)
2146 add_wait_queue(&khugepaged_wait, &wait);
2147 freezable_schedule_timeout_interruptible(
2148 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2149 remove_wait_queue(&khugepaged_wait, &wait);
2152 static int khugepaged_node_load[MAX_NUMNODES];
2154 static bool khugepaged_scan_abort(int nid)
2159 * If zone_reclaim_mode is disabled, then no extra effort is made to
2160 * allocate memory locally.
2162 if (!zone_reclaim_mode)
2165 /* If there is a count for this node already, it must be acceptable */
2166 if (khugepaged_node_load[nid])
2169 for (i = 0; i < MAX_NUMNODES; i++) {
2170 if (!khugepaged_node_load[i])
2172 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2179 static int khugepaged_find_target_node(void)
2181 static int last_khugepaged_target_node = NUMA_NO_NODE;
2182 int nid, target_node = 0, max_value = 0;
2184 /* find first node with max normal pages hit */
2185 for (nid = 0; nid < MAX_NUMNODES; nid++)
2186 if (khugepaged_node_load[nid] > max_value) {
2187 max_value = khugepaged_node_load[nid];
2191 /* do some balance if several nodes have the same hit record */
2192 if (target_node <= last_khugepaged_target_node)
2193 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2195 if (max_value == khugepaged_node_load[nid]) {
2200 last_khugepaged_target_node = target_node;
2204 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2206 if (IS_ERR(*hpage)) {
2212 khugepaged_alloc_sleep();
2213 } else if (*hpage) {
2221 static struct page *
2222 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2223 unsigned long address, int node)
2225 VM_BUG_ON_PAGE(*hpage, *hpage);
2228 * Before allocating the hugepage, release the mmap_sem read lock.
2229 * The allocation can take potentially a long time if it involves
2230 * sync compaction, and we do not need to hold the mmap_sem during
2231 * that. We will recheck the vma after taking it again in write mode.
2233 up_read(&mm->mmap_sem);
2235 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2236 if (unlikely(!*hpage)) {
2237 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2238 *hpage = ERR_PTR(-ENOMEM);
2242 prep_transhuge_page(*hpage);
2243 count_vm_event(THP_COLLAPSE_ALLOC);
2247 static int khugepaged_find_target_node(void)
2252 static inline struct page *alloc_hugepage(int defrag)
2256 page = alloc_pages(alloc_hugepage_gfpmask(defrag, 0), HPAGE_PMD_ORDER);
2258 prep_transhuge_page(page);
2262 static struct page *khugepaged_alloc_hugepage(bool *wait)
2267 hpage = alloc_hugepage(khugepaged_defrag());
2269 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2274 khugepaged_alloc_sleep();
2276 count_vm_event(THP_COLLAPSE_ALLOC);
2277 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2282 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2285 *hpage = khugepaged_alloc_hugepage(wait);
2287 if (unlikely(!*hpage))
2293 static struct page *
2294 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2295 unsigned long address, int node)
2297 up_read(&mm->mmap_sem);
2304 static bool hugepage_vma_check(struct vm_area_struct *vma)
2306 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2307 (vma->vm_flags & VM_NOHUGEPAGE))
2309 if (!vma->anon_vma || vma->vm_ops)
2311 if (is_vma_temporary_stack(vma))
2313 VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2317 static void collapse_huge_page(struct mm_struct *mm,
2318 unsigned long address,
2319 struct page **hpage,
2320 struct vm_area_struct *vma,
2326 struct page *new_page;
2327 spinlock_t *pmd_ptl, *pte_ptl;
2328 int isolated = 0, result = 0;
2329 unsigned long hstart, hend;
2330 struct mem_cgroup *memcg;
2331 unsigned long mmun_start; /* For mmu_notifiers */
2332 unsigned long mmun_end; /* For mmu_notifiers */
2335 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2337 /* Only allocate from the target node */
2338 gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2341 /* release the mmap_sem read lock. */
2342 new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2344 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2348 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2349 result = SCAN_CGROUP_CHARGE_FAIL;
2354 * Prevent all access to pagetables with the exception of
2355 * gup_fast later hanlded by the ptep_clear_flush and the VM
2356 * handled by the anon_vma lock + PG_lock.
2358 down_write(&mm->mmap_sem);
2359 if (unlikely(khugepaged_test_exit(mm))) {
2360 result = SCAN_ANY_PROCESS;
2364 vma = find_vma(mm, address);
2366 result = SCAN_VMA_NULL;
2369 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2370 hend = vma->vm_end & HPAGE_PMD_MASK;
2371 if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2372 result = SCAN_ADDRESS_RANGE;
2375 if (!hugepage_vma_check(vma)) {
2376 result = SCAN_VMA_CHECK;
2379 pmd = mm_find_pmd(mm, address);
2381 result = SCAN_PMD_NULL;
2385 anon_vma_lock_write(vma->anon_vma);
2387 pte = pte_offset_map(pmd, address);
2388 pte_ptl = pte_lockptr(mm, pmd);
2390 mmun_start = address;
2391 mmun_end = address + HPAGE_PMD_SIZE;
2392 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2393 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2395 * After this gup_fast can't run anymore. This also removes
2396 * any huge TLB entry from the CPU so we won't allow
2397 * huge and small TLB entries for the same virtual address
2398 * to avoid the risk of CPU bugs in that area.
2400 _pmd = pmdp_collapse_flush(vma, address, pmd);
2401 spin_unlock(pmd_ptl);
2402 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2405 isolated = __collapse_huge_page_isolate(vma, address, pte);
2406 spin_unlock(pte_ptl);
2408 if (unlikely(!isolated)) {
2411 BUG_ON(!pmd_none(*pmd));
2413 * We can only use set_pmd_at when establishing
2414 * hugepmds and never for establishing regular pmds that
2415 * points to regular pagetables. Use pmd_populate for that
2417 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2418 spin_unlock(pmd_ptl);
2419 anon_vma_unlock_write(vma->anon_vma);
2425 * All pages are isolated and locked so anon_vma rmap
2426 * can't run anymore.
2428 anon_vma_unlock_write(vma->anon_vma);
2430 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2432 __SetPageUptodate(new_page);
2433 pgtable = pmd_pgtable(_pmd);
2435 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2436 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2439 * spin_lock() below is not the equivalent of smp_wmb(), so
2440 * this is needed to avoid the copy_huge_page writes to become
2441 * visible after the set_pmd_at() write.
2446 BUG_ON(!pmd_none(*pmd));
2447 page_add_new_anon_rmap(new_page, vma, address, true);
2448 mem_cgroup_commit_charge(new_page, memcg, false, true);
2449 lru_cache_add_active_or_unevictable(new_page, vma);
2450 pgtable_trans_huge_deposit(mm, pmd, pgtable);
2451 set_pmd_at(mm, address, pmd, _pmd);
2452 update_mmu_cache_pmd(vma, address, pmd);
2453 spin_unlock(pmd_ptl);
2457 khugepaged_pages_collapsed++;
2458 result = SCAN_SUCCEED;
2460 up_write(&mm->mmap_sem);
2461 trace_mm_collapse_huge_page(mm, isolated, result);
2465 trace_mm_collapse_huge_page(mm, isolated, result);
2468 mem_cgroup_cancel_charge(new_page, memcg, true);
2472 static int khugepaged_scan_pmd(struct mm_struct *mm,
2473 struct vm_area_struct *vma,
2474 unsigned long address,
2475 struct page **hpage)
2479 int ret = 0, none_or_zero = 0, result = 0;
2480 struct page *page = NULL;
2481 unsigned long _address;
2483 int node = NUMA_NO_NODE;
2484 bool writable = false, referenced = false;
2486 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2488 pmd = mm_find_pmd(mm, address);
2490 result = SCAN_PMD_NULL;
2494 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2495 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2496 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2497 _pte++, _address += PAGE_SIZE) {
2498 pte_t pteval = *_pte;
2499 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2500 if (!userfaultfd_armed(vma) &&
2501 ++none_or_zero <= khugepaged_max_ptes_none) {
2504 result = SCAN_EXCEED_NONE_PTE;
2508 if (!pte_present(pteval)) {
2509 result = SCAN_PTE_NON_PRESENT;
2512 if (pte_write(pteval))
2515 page = vm_normal_page(vma, _address, pteval);
2516 if (unlikely(!page)) {
2517 result = SCAN_PAGE_NULL;
2521 /* TODO: teach khugepaged to collapse THP mapped with pte */
2522 if (PageCompound(page)) {
2523 result = SCAN_PAGE_COMPOUND;
2528 * Record which node the original page is from and save this
2529 * information to khugepaged_node_load[].
2530 * Khupaged will allocate hugepage from the node has the max
2533 node = page_to_nid(page);
2534 if (khugepaged_scan_abort(node)) {
2535 result = SCAN_SCAN_ABORT;
2538 khugepaged_node_load[node]++;
2539 if (!PageLRU(page)) {
2540 result = SCAN_SCAN_ABORT;
2543 if (PageLocked(page)) {
2544 result = SCAN_PAGE_LOCK;
2547 if (!PageAnon(page)) {
2548 result = SCAN_PAGE_ANON;
2553 * cannot use mapcount: can't collapse if there's a gup pin.
2554 * The page must only be referenced by the scanned process
2555 * and page swap cache.
2557 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2558 result = SCAN_PAGE_COUNT;
2561 if (pte_young(pteval) ||
2562 page_is_young(page) || PageReferenced(page) ||
2563 mmu_notifier_test_young(vma->vm_mm, address))
2568 result = SCAN_SUCCEED;
2571 result = SCAN_NO_REFERENCED_PAGE;
2574 result = SCAN_PAGE_RO;
2577 pte_unmap_unlock(pte, ptl);
2579 node = khugepaged_find_target_node();
2580 /* collapse_huge_page will return with the mmap_sem released */
2581 collapse_huge_page(mm, address, hpage, vma, node);
2584 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
2585 none_or_zero, result);
2589 static void collect_mm_slot(struct mm_slot *mm_slot)
2591 struct mm_struct *mm = mm_slot->mm;
2593 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2595 if (khugepaged_test_exit(mm)) {
2597 hash_del(&mm_slot->hash);
2598 list_del(&mm_slot->mm_node);
2601 * Not strictly needed because the mm exited already.
2603 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2606 /* khugepaged_mm_lock actually not necessary for the below */
2607 free_mm_slot(mm_slot);
2612 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2613 struct page **hpage)
2614 __releases(&khugepaged_mm_lock)
2615 __acquires(&khugepaged_mm_lock)
2617 struct mm_slot *mm_slot;
2618 struct mm_struct *mm;
2619 struct vm_area_struct *vma;
2623 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2625 if (khugepaged_scan.mm_slot)
2626 mm_slot = khugepaged_scan.mm_slot;
2628 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2629 struct mm_slot, mm_node);
2630 khugepaged_scan.address = 0;
2631 khugepaged_scan.mm_slot = mm_slot;
2633 spin_unlock(&khugepaged_mm_lock);
2636 down_read(&mm->mmap_sem);
2637 if (unlikely(khugepaged_test_exit(mm)))
2640 vma = find_vma(mm, khugepaged_scan.address);
2643 for (; vma; vma = vma->vm_next) {
2644 unsigned long hstart, hend;
2647 if (unlikely(khugepaged_test_exit(mm))) {
2651 if (!hugepage_vma_check(vma)) {
2656 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2657 hend = vma->vm_end & HPAGE_PMD_MASK;
2660 if (khugepaged_scan.address > hend)
2662 if (khugepaged_scan.address < hstart)
2663 khugepaged_scan.address = hstart;
2664 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2666 while (khugepaged_scan.address < hend) {
2669 if (unlikely(khugepaged_test_exit(mm)))
2670 goto breakouterloop;
2672 VM_BUG_ON(khugepaged_scan.address < hstart ||
2673 khugepaged_scan.address + HPAGE_PMD_SIZE >
2675 ret = khugepaged_scan_pmd(mm, vma,
2676 khugepaged_scan.address,
2678 /* move to next address */
2679 khugepaged_scan.address += HPAGE_PMD_SIZE;
2680 progress += HPAGE_PMD_NR;
2682 /* we released mmap_sem so break loop */
2683 goto breakouterloop_mmap_sem;
2684 if (progress >= pages)
2685 goto breakouterloop;
2689 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2690 breakouterloop_mmap_sem:
2692 spin_lock(&khugepaged_mm_lock);
2693 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2695 * Release the current mm_slot if this mm is about to die, or
2696 * if we scanned all vmas of this mm.
2698 if (khugepaged_test_exit(mm) || !vma) {
2700 * Make sure that if mm_users is reaching zero while
2701 * khugepaged runs here, khugepaged_exit will find
2702 * mm_slot not pointing to the exiting mm.
2704 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2705 khugepaged_scan.mm_slot = list_entry(
2706 mm_slot->mm_node.next,
2707 struct mm_slot, mm_node);
2708 khugepaged_scan.address = 0;
2710 khugepaged_scan.mm_slot = NULL;
2711 khugepaged_full_scans++;
2714 collect_mm_slot(mm_slot);
2720 static int khugepaged_has_work(void)
2722 return !list_empty(&khugepaged_scan.mm_head) &&
2723 khugepaged_enabled();
2726 static int khugepaged_wait_event(void)
2728 return !list_empty(&khugepaged_scan.mm_head) ||
2729 kthread_should_stop();
2732 static void khugepaged_do_scan(void)
2734 struct page *hpage = NULL;
2735 unsigned int progress = 0, pass_through_head = 0;
2736 unsigned int pages = khugepaged_pages_to_scan;
2739 barrier(); /* write khugepaged_pages_to_scan to local stack */
2741 while (progress < pages) {
2742 if (!khugepaged_prealloc_page(&hpage, &wait))
2747 if (unlikely(kthread_should_stop() || try_to_freeze()))
2750 spin_lock(&khugepaged_mm_lock);
2751 if (!khugepaged_scan.mm_slot)
2752 pass_through_head++;
2753 if (khugepaged_has_work() &&
2754 pass_through_head < 2)
2755 progress += khugepaged_scan_mm_slot(pages - progress,
2759 spin_unlock(&khugepaged_mm_lock);
2762 if (!IS_ERR_OR_NULL(hpage))
2766 static void khugepaged_wait_work(void)
2768 if (khugepaged_has_work()) {
2769 if (!khugepaged_scan_sleep_millisecs)
2772 wait_event_freezable_timeout(khugepaged_wait,
2773 kthread_should_stop(),
2774 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2778 if (khugepaged_enabled())
2779 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2782 static int khugepaged(void *none)
2784 struct mm_slot *mm_slot;
2787 set_user_nice(current, MAX_NICE);
2789 while (!kthread_should_stop()) {
2790 khugepaged_do_scan();
2791 khugepaged_wait_work();
2794 spin_lock(&khugepaged_mm_lock);
2795 mm_slot = khugepaged_scan.mm_slot;
2796 khugepaged_scan.mm_slot = NULL;
2798 collect_mm_slot(mm_slot);
2799 spin_unlock(&khugepaged_mm_lock);
2803 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2804 unsigned long haddr, pmd_t *pmd)
2806 struct mm_struct *mm = vma->vm_mm;
2811 /* leave pmd empty until pte is filled */
2812 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2814 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2815 pmd_populate(mm, &_pmd, pgtable);
2817 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2819 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2820 entry = pte_mkspecial(entry);
2821 pte = pte_offset_map(&_pmd, haddr);
2822 VM_BUG_ON(!pte_none(*pte));
2823 set_pte_at(mm, haddr, pte, entry);
2826 smp_wmb(); /* make pte visible before pmd */
2827 pmd_populate(mm, pmd, pgtable);
2828 put_huge_zero_page();
2831 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2832 unsigned long haddr, bool freeze)
2834 struct mm_struct *mm = vma->vm_mm;
2838 bool young, write, dirty;
2842 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2843 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2844 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2845 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
2847 count_vm_event(THP_SPLIT_PMD);
2849 if (vma_is_dax(vma)) {
2850 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2851 if (is_huge_zero_pmd(_pmd))
2852 put_huge_zero_page();
2854 } else if (is_huge_zero_pmd(*pmd)) {
2855 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2858 page = pmd_page(*pmd);
2859 VM_BUG_ON_PAGE(!page_count(page), page);
2860 atomic_add(HPAGE_PMD_NR - 1, &page->_count);
2861 write = pmd_write(*pmd);
2862 young = pmd_young(*pmd);
2863 dirty = pmd_dirty(*pmd);
2865 pmdp_huge_split_prepare(vma, haddr, pmd);
2866 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2867 pmd_populate(mm, &_pmd, pgtable);
2869 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2872 * Note that NUMA hinting access restrictions are not
2873 * transferred to avoid any possibility of altering
2874 * permissions across VMAs.
2877 swp_entry_t swp_entry;
2878 swp_entry = make_migration_entry(page + i, write);
2879 entry = swp_entry_to_pte(swp_entry);
2881 entry = mk_pte(page + i, vma->vm_page_prot);
2882 entry = maybe_mkwrite(entry, vma);
2884 entry = pte_wrprotect(entry);
2886 entry = pte_mkold(entry);
2889 SetPageDirty(page + i);
2890 pte = pte_offset_map(&_pmd, addr);
2891 BUG_ON(!pte_none(*pte));
2892 set_pte_at(mm, addr, pte, entry);
2893 atomic_inc(&page[i]._mapcount);
2898 * Set PG_double_map before dropping compound_mapcount to avoid
2899 * false-negative page_mapped().
2901 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2902 for (i = 0; i < HPAGE_PMD_NR; i++)
2903 atomic_inc(&page[i]._mapcount);
2906 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2907 /* Last compound_mapcount is gone. */
2908 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
2909 if (TestClearPageDoubleMap(page)) {
2910 /* No need in mapcount reference anymore */
2911 for (i = 0; i < HPAGE_PMD_NR; i++)
2912 atomic_dec(&page[i]._mapcount);
2916 smp_wmb(); /* make pte visible before pmd */
2918 * Up to this point the pmd is present and huge and userland has the
2919 * whole access to the hugepage during the split (which happens in
2920 * place). If we overwrite the pmd with the not-huge version pointing
2921 * to the pte here (which of course we could if all CPUs were bug
2922 * free), userland could trigger a small page size TLB miss on the
2923 * small sized TLB while the hugepage TLB entry is still established in
2924 * the huge TLB. Some CPU doesn't like that.
2925 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2926 * 383 on page 93. Intel should be safe but is also warns that it's
2927 * only safe if the permission and cache attributes of the two entries
2928 * loaded in the two TLB is identical (which should be the case here).
2929 * But it is generally safer to never allow small and huge TLB entries
2930 * for the same virtual address to be loaded simultaneously. So instead
2931 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2932 * current pmd notpresent (atomically because here the pmd_trans_huge
2933 * and pmd_trans_splitting must remain set at all times on the pmd
2934 * until the split is complete for this pmd), then we flush the SMP TLB
2935 * and finally we write the non-huge version of the pmd entry with
2938 pmdp_invalidate(vma, haddr, pmd);
2939 pmd_populate(mm, pmd, pgtable);
2942 for (i = 0; i < HPAGE_PMD_NR; i++) {
2943 page_remove_rmap(page + i, false);
2949 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2950 unsigned long address)
2953 struct mm_struct *mm = vma->vm_mm;
2954 struct page *page = NULL;
2955 unsigned long haddr = address & HPAGE_PMD_MASK;
2957 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2958 ptl = pmd_lock(mm, pmd);
2959 if (pmd_trans_huge(*pmd)) {
2960 page = pmd_page(*pmd);
2961 if (PageMlocked(page))
2965 } else if (!pmd_devmap(*pmd))
2967 __split_huge_pmd_locked(vma, pmd, haddr, false);
2970 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2973 munlock_vma_page(page);
2979 static void split_huge_pmd_address(struct vm_area_struct *vma,
2980 unsigned long address)
2986 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2988 pgd = pgd_offset(vma->vm_mm, address);
2989 if (!pgd_present(*pgd))
2992 pud = pud_offset(pgd, address);
2993 if (!pud_present(*pud))
2996 pmd = pmd_offset(pud, address);
2997 if (!pmd_present(*pmd) || (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)))
3000 * Caller holds the mmap_sem write mode, so a huge pmd cannot
3001 * materialize from under us.
3003 split_huge_pmd(vma, pmd, address);
3006 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3007 unsigned long start,
3012 * If the new start address isn't hpage aligned and it could
3013 * previously contain an hugepage: check if we need to split
3016 if (start & ~HPAGE_PMD_MASK &&
3017 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3018 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3019 split_huge_pmd_address(vma, start);
3022 * If the new end address isn't hpage aligned and it could
3023 * previously contain an hugepage: check if we need to split
3026 if (end & ~HPAGE_PMD_MASK &&
3027 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3028 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3029 split_huge_pmd_address(vma, end);
3032 * If we're also updating the vma->vm_next->vm_start, if the new
3033 * vm_next->vm_start isn't page aligned and it could previously
3034 * contain an hugepage: check if we need to split an huge pmd.
3036 if (adjust_next > 0) {
3037 struct vm_area_struct *next = vma->vm_next;
3038 unsigned long nstart = next->vm_start;
3039 nstart += adjust_next << PAGE_SHIFT;
3040 if (nstart & ~HPAGE_PMD_MASK &&
3041 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3042 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3043 split_huge_pmd_address(next, nstart);
3047 static void freeze_page_vma(struct vm_area_struct *vma, struct page *page,
3048 unsigned long address)
3050 unsigned long haddr = address & HPAGE_PMD_MASK;
3056 int i, nr = HPAGE_PMD_NR;
3058 /* Skip pages which doesn't belong to the VMA */
3059 if (address < vma->vm_start) {
3060 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3063 address = vma->vm_start;
3066 pgd = pgd_offset(vma->vm_mm, address);
3067 if (!pgd_present(*pgd))
3069 pud = pud_offset(pgd, address);
3070 if (!pud_present(*pud))
3072 pmd = pmd_offset(pud, address);
3073 ptl = pmd_lock(vma->vm_mm, pmd);
3074 if (!pmd_present(*pmd)) {
3078 if (pmd_trans_huge(*pmd)) {
3079 if (page == pmd_page(*pmd))
3080 __split_huge_pmd_locked(vma, pmd, haddr, true);
3086 pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3087 for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) {
3088 pte_t entry, swp_pte;
3089 swp_entry_t swp_entry;
3092 * We've just crossed page table boundary: need to map next one.
3093 * It can happen if THP was mremaped to non PMD-aligned address.
3095 if (unlikely(address == haddr + HPAGE_PMD_SIZE)) {
3096 pte_unmap_unlock(pte - 1, ptl);
3097 pmd = mm_find_pmd(vma->vm_mm, address);
3100 pte = pte_offset_map_lock(vma->vm_mm, pmd,
3104 if (!pte_present(*pte))
3106 if (page_to_pfn(page) != pte_pfn(*pte))
3108 flush_cache_page(vma, address, page_to_pfn(page));
3109 entry = ptep_clear_flush(vma, address, pte);
3110 if (pte_dirty(entry))
3112 swp_entry = make_migration_entry(page, pte_write(entry));
3113 swp_pte = swp_entry_to_pte(swp_entry);
3114 if (pte_soft_dirty(entry))
3115 swp_pte = pte_swp_mksoft_dirty(swp_pte);
3116 set_pte_at(vma->vm_mm, address, pte, swp_pte);
3117 page_remove_rmap(page, false);
3120 pte_unmap_unlock(pte - 1, ptl);
3123 static void freeze_page(struct anon_vma *anon_vma, struct page *page)
3125 struct anon_vma_chain *avc;
3126 pgoff_t pgoff = page_to_pgoff(page);
3128 VM_BUG_ON_PAGE(!PageHead(page), page);
3130 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff,
3131 pgoff + HPAGE_PMD_NR - 1) {
3132 unsigned long address = __vma_address(page, avc->vma);
3134 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3135 address, address + HPAGE_PMD_SIZE);
3136 freeze_page_vma(avc->vma, page, address);
3137 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3138 address, address + HPAGE_PMD_SIZE);
3142 static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page,
3143 unsigned long address)
3148 swp_entry_t swp_entry;
3149 unsigned long haddr = address & HPAGE_PMD_MASK;
3150 int i, nr = HPAGE_PMD_NR;
3152 /* Skip pages which doesn't belong to the VMA */
3153 if (address < vma->vm_start) {
3154 int off = (vma->vm_start - address) >> PAGE_SHIFT;
3157 address = vma->vm_start;
3160 pmd = mm_find_pmd(vma->vm_mm, address);
3164 pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl);
3165 for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) {
3167 * We've just crossed page table boundary: need to map next one.
3168 * It can happen if THP was mremaped to non-PMD aligned address.
3170 if (unlikely(address == haddr + HPAGE_PMD_SIZE)) {
3171 pte_unmap_unlock(pte - 1, ptl);
3172 pmd = mm_find_pmd(vma->vm_mm, address);
3175 pte = pte_offset_map_lock(vma->vm_mm, pmd,
3179 if (!is_swap_pte(*pte))
3182 swp_entry = pte_to_swp_entry(*pte);
3183 if (!is_migration_entry(swp_entry))
3185 if (migration_entry_to_page(swp_entry) != page)
3189 page_add_anon_rmap(page, vma, address, false);
3191 entry = pte_mkold(mk_pte(page, vma->vm_page_prot));
3192 if (PageDirty(page))
3193 entry = pte_mkdirty(entry);
3194 if (is_write_migration_entry(swp_entry))
3195 entry = maybe_mkwrite(entry, vma);
3197 flush_dcache_page(page);
3198 set_pte_at(vma->vm_mm, address, pte, entry);
3200 /* No need to invalidate - it was non-present before */
3201 update_mmu_cache(vma, address, pte);
3203 pte_unmap_unlock(pte - 1, ptl);
3206 static void unfreeze_page(struct anon_vma *anon_vma, struct page *page)
3208 struct anon_vma_chain *avc;
3209 pgoff_t pgoff = page_to_pgoff(page);
3211 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
3212 pgoff, pgoff + HPAGE_PMD_NR - 1) {
3213 unsigned long address = __vma_address(page, avc->vma);
3215 mmu_notifier_invalidate_range_start(avc->vma->vm_mm,
3216 address, address + HPAGE_PMD_SIZE);
3217 unfreeze_page_vma(avc->vma, page, address);
3218 mmu_notifier_invalidate_range_end(avc->vma->vm_mm,
3219 address, address + HPAGE_PMD_SIZE);
3223 static int __split_huge_page_tail(struct page *head, int tail,
3224 struct lruvec *lruvec, struct list_head *list)
3227 struct page *page_tail = head + tail;
3229 mapcount = atomic_read(&page_tail->_mapcount) + 1;
3230 VM_BUG_ON_PAGE(atomic_read(&page_tail->_count) != 0, page_tail);
3233 * tail_page->_count is zero and not changing from under us. But
3234 * get_page_unless_zero() may be running from under us on the
3235 * tail_page. If we used atomic_set() below instead of atomic_add(), we
3236 * would then run atomic_set() concurrently with
3237 * get_page_unless_zero(), and atomic_set() is implemented in C not
3238 * using locked ops. spin_unlock on x86 sometime uses locked ops
3239 * because of PPro errata 66, 92, so unless somebody can guarantee
3240 * atomic_set() here would be safe on all archs (and not only on x86),
3241 * it's safer to use atomic_add().
3243 atomic_add(mapcount + 1, &page_tail->_count);
3246 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3247 page_tail->flags |= (head->flags &
3248 ((1L << PG_referenced) |
3249 (1L << PG_swapbacked) |
3250 (1L << PG_mlocked) |
3251 (1L << PG_uptodate) |
3254 (1L << PG_unevictable) |
3258 * After clearing PageTail the gup refcount can be released.
3259 * Page flags also must be visible before we make the page non-compound.
3263 clear_compound_head(page_tail);
3265 if (page_is_young(head))
3266 set_page_young(page_tail);
3267 if (page_is_idle(head))
3268 set_page_idle(page_tail);
3270 /* ->mapping in first tail page is compound_mapcount */
3271 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3273 page_tail->mapping = head->mapping;
3275 page_tail->index = head->index + tail;
3276 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3277 lru_add_page_tail(head, page_tail, lruvec, list);
3282 static void __split_huge_page(struct page *page, struct list_head *list)
3284 struct page *head = compound_head(page);
3285 struct zone *zone = page_zone(head);
3286 struct lruvec *lruvec;
3287 int i, tail_mapcount;
3289 /* prevent PageLRU to go away from under us, and freeze lru stats */
3290 spin_lock_irq(&zone->lru_lock);
3291 lruvec = mem_cgroup_page_lruvec(head, zone);
3293 /* complete memcg works before add pages to LRU */
3294 mem_cgroup_split_huge_fixup(head);
3297 for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
3298 tail_mapcount += __split_huge_page_tail(head, i, lruvec, list);
3299 atomic_sub(tail_mapcount, &head->_count);
3301 ClearPageCompound(head);
3302 spin_unlock_irq(&zone->lru_lock);
3304 unfreeze_page(page_anon_vma(head), head);
3306 for (i = 0; i < HPAGE_PMD_NR; i++) {
3307 struct page *subpage = head + i;
3308 if (subpage == page)
3310 unlock_page(subpage);
3313 * Subpages may be freed if there wasn't any mapping
3314 * like if add_to_swap() is running on a lru page that
3315 * had its mapping zapped. And freeing these pages
3316 * requires taking the lru_lock so we do the put_page
3317 * of the tail pages after the split is complete.
3323 int total_mapcount(struct page *page)
3327 VM_BUG_ON_PAGE(PageTail(page), page);
3329 if (likely(!PageCompound(page)))
3330 return atomic_read(&page->_mapcount) + 1;
3332 ret = compound_mapcount(page);
3335 for (i = 0; i < HPAGE_PMD_NR; i++)
3336 ret += atomic_read(&page[i]._mapcount) + 1;
3337 if (PageDoubleMap(page))
3338 ret -= HPAGE_PMD_NR;
3343 * This function splits huge page into normal pages. @page can point to any
3344 * subpage of huge page to split. Split doesn't change the position of @page.
3346 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3347 * The huge page must be locked.
3349 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3351 * Both head page and tail pages will inherit mapping, flags, and so on from
3354 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3355 * they are not mapped.
3357 * Returns 0 if the hugepage is split successfully.
3358 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3361 int split_huge_page_to_list(struct page *page, struct list_head *list)
3363 struct page *head = compound_head(page);
3364 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
3365 struct anon_vma *anon_vma;
3366 int count, mapcount, ret;
3368 unsigned long flags;
3370 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3371 VM_BUG_ON_PAGE(!PageAnon(page), page);
3372 VM_BUG_ON_PAGE(!PageLocked(page), page);
3373 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3374 VM_BUG_ON_PAGE(!PageCompound(page), page);
3377 * The caller does not necessarily hold an mmap_sem that would prevent
3378 * the anon_vma disappearing so we first we take a reference to it
3379 * and then lock the anon_vma for write. This is similar to
3380 * page_lock_anon_vma_read except the write lock is taken to serialise
3381 * against parallel split or collapse operations.
3383 anon_vma = page_get_anon_vma(head);
3388 anon_vma_lock_write(anon_vma);
3391 * Racy check if we can split the page, before freeze_page() will
3394 if (total_mapcount(head) != page_count(head) - 1) {
3399 mlocked = PageMlocked(page);
3400 freeze_page(anon_vma, head);
3401 VM_BUG_ON_PAGE(compound_mapcount(head), head);
3403 /* Make sure the page is not on per-CPU pagevec as it takes pin */
3407 /* Prevent deferred_split_scan() touching ->_count */
3408 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3409 count = page_count(head);
3410 mapcount = total_mapcount(head);
3411 if (!mapcount && count == 1) {
3412 if (!list_empty(page_deferred_list(head))) {
3413 pgdata->split_queue_len--;
3414 list_del(page_deferred_list(head));
3416 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3417 __split_huge_page(page, list);
3419 } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
3420 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3421 pr_alert("total_mapcount: %u, page_count(): %u\n",
3424 dump_page(head, NULL);
3425 dump_page(page, "total_mapcount(head) > 0");
3428 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3429 unfreeze_page(anon_vma, head);
3434 anon_vma_unlock_write(anon_vma);
3435 put_anon_vma(anon_vma);
3437 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3441 void free_transhuge_page(struct page *page)
3443 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3444 unsigned long flags;
3446 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3447 if (!list_empty(page_deferred_list(page))) {
3448 pgdata->split_queue_len--;
3449 list_del(page_deferred_list(page));
3451 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3452 free_compound_page(page);
3455 void deferred_split_huge_page(struct page *page)
3457 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3458 unsigned long flags;
3460 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3462 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3463 if (list_empty(page_deferred_list(page))) {
3464 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
3465 pgdata->split_queue_len++;
3467 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3470 static unsigned long deferred_split_count(struct shrinker *shrink,
3471 struct shrink_control *sc)
3473 struct pglist_data *pgdata = NODE_DATA(sc->nid);
3474 return ACCESS_ONCE(pgdata->split_queue_len);
3477 static unsigned long deferred_split_scan(struct shrinker *shrink,
3478 struct shrink_control *sc)
3480 struct pglist_data *pgdata = NODE_DATA(sc->nid);
3481 unsigned long flags;
3482 LIST_HEAD(list), *pos, *next;
3486 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3487 /* Take pin on all head pages to avoid freeing them under us */
3488 list_for_each_safe(pos, next, &pgdata->split_queue) {
3489 page = list_entry((void *)pos, struct page, mapping);
3490 page = compound_head(page);
3491 if (get_page_unless_zero(page)) {
3492 list_move(page_deferred_list(page), &list);
3494 /* We lost race with put_compound_page() */
3495 list_del_init(page_deferred_list(page));
3496 pgdata->split_queue_len--;
3498 if (!--sc->nr_to_scan)
3501 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3503 list_for_each_safe(pos, next, &list) {
3504 page = list_entry((void *)pos, struct page, mapping);
3506 /* split_huge_page() removes page from list on success */
3507 if (!split_huge_page(page))
3513 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3514 list_splice_tail(&list, &pgdata->split_queue);
3515 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3518 * Stop shrinker if we didn't split any page, but the queue is empty.
3519 * This can happen if pages were freed under us.
3521 if (!split && list_empty(&pgdata->split_queue))
3526 static struct shrinker deferred_split_shrinker = {
3527 .count_objects = deferred_split_count,
3528 .scan_objects = deferred_split_scan,
3529 .seeks = DEFAULT_SEEKS,
3530 .flags = SHRINKER_NUMA_AWARE,
3533 #ifdef CONFIG_DEBUG_FS
3534 static int split_huge_pages_set(void *data, u64 val)
3538 unsigned long pfn, max_zone_pfn;
3539 unsigned long total = 0, split = 0;
3544 for_each_populated_zone(zone) {
3545 max_zone_pfn = zone_end_pfn(zone);
3546 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3547 if (!pfn_valid(pfn))
3550 page = pfn_to_page(pfn);
3551 if (!get_page_unless_zero(page))
3554 if (zone != page_zone(page))
3557 if (!PageHead(page) || !PageAnon(page) ||
3563 if (!split_huge_page(page))
3571 pr_info("%lu of %lu THP split", split, total);
3575 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3578 static int __init split_huge_pages_debugfs(void)
3582 ret = debugfs_create_file("split_huge_pages", 0644, NULL, NULL,
3583 &split_huge_pages_fops);
3585 pr_warn("Failed to create split_huge_pages in debugfs");
3588 late_initcall(split_huge_pages_debugfs);