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/dax.h>
20 #include <linux/kthread.h>
21 #include <linux/khugepaged.h>
22 #include <linux/freezer.h>
23 #include <linux/mman.h>
24 #include <linux/pagemap.h>
25 #include <linux/migrate.h>
26 #include <linux/hashtable.h>
27 #include <linux/userfaultfd_k.h>
30 #include <asm/pgalloc.h>
34 * By default transparent hugepage support is disabled in order that avoid
35 * to risk increase the memory footprint of applications without a guaranteed
36 * benefit. When transparent hugepage support is enabled, is for all mappings,
37 * and khugepaged scans all mappings.
38 * Defrag is invoked by khugepaged hugepage allocations and by page faults
39 * for all hugepage allocations.
41 unsigned long transparent_hugepage_flags __read_mostly =
42 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
43 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
45 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
46 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
48 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
49 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
50 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
52 /* default scan 8*512 pte (or vmas) every 30 second */
53 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
54 static unsigned int khugepaged_pages_collapsed;
55 static unsigned int khugepaged_full_scans;
56 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
57 /* during fragmentation poll the hugepage allocator once every minute */
58 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
59 static struct task_struct *khugepaged_thread __read_mostly;
60 static DEFINE_MUTEX(khugepaged_mutex);
61 static DEFINE_SPINLOCK(khugepaged_mm_lock);
62 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
64 * default collapse hugepages if there is at least one pte mapped like
65 * it would have happened if the vma was large enough during page
68 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
70 static int khugepaged(void *none);
71 static int khugepaged_slab_init(void);
72 static void khugepaged_slab_exit(void);
74 #define MM_SLOTS_HASH_BITS 10
75 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
77 static struct kmem_cache *mm_slot_cache __read_mostly;
80 * struct mm_slot - hash lookup from mm to mm_slot
81 * @hash: hash collision list
82 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
83 * @mm: the mm that this information is valid for
86 struct hlist_node hash;
87 struct list_head mm_node;
92 * struct khugepaged_scan - cursor for scanning
93 * @mm_head: the head of the mm list to scan
94 * @mm_slot: the current mm_slot we are scanning
95 * @address: the next address inside that to be scanned
97 * There is only the one khugepaged_scan instance of this cursor structure.
99 struct khugepaged_scan {
100 struct list_head mm_head;
101 struct mm_slot *mm_slot;
102 unsigned long address;
104 static struct khugepaged_scan khugepaged_scan = {
105 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
109 static void set_recommended_min_free_kbytes(void)
113 unsigned long recommended_min;
115 for_each_populated_zone(zone)
118 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
119 recommended_min = pageblock_nr_pages * nr_zones * 2;
122 * Make sure that on average at least two pageblocks are almost free
123 * of another type, one for a migratetype to fall back to and a
124 * second to avoid subsequent fallbacks of other types There are 3
125 * MIGRATE_TYPES we care about.
127 recommended_min += pageblock_nr_pages * nr_zones *
128 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
130 /* don't ever allow to reserve more than 5% of the lowmem */
131 recommended_min = min(recommended_min,
132 (unsigned long) nr_free_buffer_pages() / 20);
133 recommended_min <<= (PAGE_SHIFT-10);
135 if (recommended_min > min_free_kbytes) {
136 if (user_min_free_kbytes >= 0)
137 pr_info("raising min_free_kbytes from %d to %lu "
138 "to help transparent hugepage allocations\n",
139 min_free_kbytes, recommended_min);
141 min_free_kbytes = recommended_min;
143 setup_per_zone_wmarks();
146 static int start_stop_khugepaged(void)
149 if (khugepaged_enabled()) {
150 if (!khugepaged_thread)
151 khugepaged_thread = kthread_run(khugepaged, NULL,
153 if (unlikely(IS_ERR(khugepaged_thread))) {
154 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
155 err = PTR_ERR(khugepaged_thread);
156 khugepaged_thread = NULL;
160 if (!list_empty(&khugepaged_scan.mm_head))
161 wake_up_interruptible(&khugepaged_wait);
163 set_recommended_min_free_kbytes();
164 } else if (khugepaged_thread) {
165 kthread_stop(khugepaged_thread);
166 khugepaged_thread = NULL;
172 static atomic_t huge_zero_refcount;
173 struct page *huge_zero_page __read_mostly;
175 struct page *get_huge_zero_page(void)
177 struct page *zero_page;
179 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
180 return READ_ONCE(huge_zero_page);
182 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
185 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
188 count_vm_event(THP_ZERO_PAGE_ALLOC);
190 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
192 __free_pages(zero_page, compound_order(zero_page));
196 /* We take additional reference here. It will be put back by shrinker */
197 atomic_set(&huge_zero_refcount, 2);
199 return READ_ONCE(huge_zero_page);
202 static void put_huge_zero_page(void)
205 * Counter should never go to zero here. Only shrinker can put
208 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
211 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
212 struct shrink_control *sc)
214 /* we can free zero page only if last reference remains */
215 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
218 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
219 struct shrink_control *sc)
221 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
222 struct page *zero_page = xchg(&huge_zero_page, NULL);
223 BUG_ON(zero_page == NULL);
224 __free_pages(zero_page, compound_order(zero_page));
231 static struct shrinker huge_zero_page_shrinker = {
232 .count_objects = shrink_huge_zero_page_count,
233 .scan_objects = shrink_huge_zero_page_scan,
234 .seeks = DEFAULT_SEEKS,
239 static ssize_t double_flag_show(struct kobject *kobj,
240 struct kobj_attribute *attr, char *buf,
241 enum transparent_hugepage_flag enabled,
242 enum transparent_hugepage_flag req_madv)
244 if (test_bit(enabled, &transparent_hugepage_flags)) {
245 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
246 return sprintf(buf, "[always] madvise never\n");
247 } else if (test_bit(req_madv, &transparent_hugepage_flags))
248 return sprintf(buf, "always [madvise] never\n");
250 return sprintf(buf, "always madvise [never]\n");
252 static ssize_t double_flag_store(struct kobject *kobj,
253 struct kobj_attribute *attr,
254 const char *buf, size_t count,
255 enum transparent_hugepage_flag enabled,
256 enum transparent_hugepage_flag req_madv)
258 if (!memcmp("always", buf,
259 min(sizeof("always")-1, count))) {
260 set_bit(enabled, &transparent_hugepage_flags);
261 clear_bit(req_madv, &transparent_hugepage_flags);
262 } else if (!memcmp("madvise", buf,
263 min(sizeof("madvise")-1, count))) {
264 clear_bit(enabled, &transparent_hugepage_flags);
265 set_bit(req_madv, &transparent_hugepage_flags);
266 } else if (!memcmp("never", buf,
267 min(sizeof("never")-1, count))) {
268 clear_bit(enabled, &transparent_hugepage_flags);
269 clear_bit(req_madv, &transparent_hugepage_flags);
276 static ssize_t enabled_show(struct kobject *kobj,
277 struct kobj_attribute *attr, char *buf)
279 return double_flag_show(kobj, attr, buf,
280 TRANSPARENT_HUGEPAGE_FLAG,
281 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
283 static ssize_t enabled_store(struct kobject *kobj,
284 struct kobj_attribute *attr,
285 const char *buf, size_t count)
289 ret = double_flag_store(kobj, attr, buf, count,
290 TRANSPARENT_HUGEPAGE_FLAG,
291 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
296 mutex_lock(&khugepaged_mutex);
297 err = start_stop_khugepaged();
298 mutex_unlock(&khugepaged_mutex);
306 static struct kobj_attribute enabled_attr =
307 __ATTR(enabled, 0644, enabled_show, enabled_store);
309 static ssize_t single_flag_show(struct kobject *kobj,
310 struct kobj_attribute *attr, char *buf,
311 enum transparent_hugepage_flag flag)
313 return sprintf(buf, "%d\n",
314 !!test_bit(flag, &transparent_hugepage_flags));
317 static ssize_t single_flag_store(struct kobject *kobj,
318 struct kobj_attribute *attr,
319 const char *buf, size_t count,
320 enum transparent_hugepage_flag flag)
325 ret = kstrtoul(buf, 10, &value);
332 set_bit(flag, &transparent_hugepage_flags);
334 clear_bit(flag, &transparent_hugepage_flags);
340 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
341 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
342 * memory just to allocate one more hugepage.
344 static ssize_t defrag_show(struct kobject *kobj,
345 struct kobj_attribute *attr, char *buf)
347 return double_flag_show(kobj, attr, buf,
348 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
349 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
351 static ssize_t defrag_store(struct kobject *kobj,
352 struct kobj_attribute *attr,
353 const char *buf, size_t count)
355 return double_flag_store(kobj, attr, buf, count,
356 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
357 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
359 static struct kobj_attribute defrag_attr =
360 __ATTR(defrag, 0644, defrag_show, defrag_store);
362 static ssize_t use_zero_page_show(struct kobject *kobj,
363 struct kobj_attribute *attr, char *buf)
365 return single_flag_show(kobj, attr, buf,
366 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
368 static ssize_t use_zero_page_store(struct kobject *kobj,
369 struct kobj_attribute *attr, const char *buf, size_t count)
371 return single_flag_store(kobj, attr, buf, count,
372 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
374 static struct kobj_attribute use_zero_page_attr =
375 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
376 #ifdef CONFIG_DEBUG_VM
377 static ssize_t debug_cow_show(struct kobject *kobj,
378 struct kobj_attribute *attr, char *buf)
380 return single_flag_show(kobj, attr, buf,
381 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
383 static ssize_t debug_cow_store(struct kobject *kobj,
384 struct kobj_attribute *attr,
385 const char *buf, size_t count)
387 return single_flag_store(kobj, attr, buf, count,
388 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
390 static struct kobj_attribute debug_cow_attr =
391 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
392 #endif /* CONFIG_DEBUG_VM */
394 static struct attribute *hugepage_attr[] = {
397 &use_zero_page_attr.attr,
398 #ifdef CONFIG_DEBUG_VM
399 &debug_cow_attr.attr,
404 static struct attribute_group hugepage_attr_group = {
405 .attrs = hugepage_attr,
408 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
409 struct kobj_attribute *attr,
412 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
415 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
416 struct kobj_attribute *attr,
417 const char *buf, size_t count)
422 err = kstrtoul(buf, 10, &msecs);
423 if (err || msecs > UINT_MAX)
426 khugepaged_scan_sleep_millisecs = msecs;
427 wake_up_interruptible(&khugepaged_wait);
431 static struct kobj_attribute scan_sleep_millisecs_attr =
432 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
433 scan_sleep_millisecs_store);
435 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
436 struct kobj_attribute *attr,
439 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
442 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
443 struct kobj_attribute *attr,
444 const char *buf, size_t count)
449 err = kstrtoul(buf, 10, &msecs);
450 if (err || msecs > UINT_MAX)
453 khugepaged_alloc_sleep_millisecs = msecs;
454 wake_up_interruptible(&khugepaged_wait);
458 static struct kobj_attribute alloc_sleep_millisecs_attr =
459 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
460 alloc_sleep_millisecs_store);
462 static ssize_t pages_to_scan_show(struct kobject *kobj,
463 struct kobj_attribute *attr,
466 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
468 static ssize_t pages_to_scan_store(struct kobject *kobj,
469 struct kobj_attribute *attr,
470 const char *buf, size_t count)
475 err = kstrtoul(buf, 10, &pages);
476 if (err || !pages || pages > UINT_MAX)
479 khugepaged_pages_to_scan = pages;
483 static struct kobj_attribute pages_to_scan_attr =
484 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
485 pages_to_scan_store);
487 static ssize_t pages_collapsed_show(struct kobject *kobj,
488 struct kobj_attribute *attr,
491 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
493 static struct kobj_attribute pages_collapsed_attr =
494 __ATTR_RO(pages_collapsed);
496 static ssize_t full_scans_show(struct kobject *kobj,
497 struct kobj_attribute *attr,
500 return sprintf(buf, "%u\n", khugepaged_full_scans);
502 static struct kobj_attribute full_scans_attr =
503 __ATTR_RO(full_scans);
505 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
506 struct kobj_attribute *attr, char *buf)
508 return single_flag_show(kobj, attr, buf,
509 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
511 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
512 struct kobj_attribute *attr,
513 const char *buf, size_t count)
515 return single_flag_store(kobj, attr, buf, count,
516 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
518 static struct kobj_attribute khugepaged_defrag_attr =
519 __ATTR(defrag, 0644, khugepaged_defrag_show,
520 khugepaged_defrag_store);
523 * max_ptes_none controls if khugepaged should collapse hugepages over
524 * any unmapped ptes in turn potentially increasing the memory
525 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
526 * reduce the available free memory in the system as it
527 * runs. Increasing max_ptes_none will instead potentially reduce the
528 * free memory in the system during the khugepaged scan.
530 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
531 struct kobj_attribute *attr,
534 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
536 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
537 struct kobj_attribute *attr,
538 const char *buf, size_t count)
541 unsigned long max_ptes_none;
543 err = kstrtoul(buf, 10, &max_ptes_none);
544 if (err || max_ptes_none > HPAGE_PMD_NR-1)
547 khugepaged_max_ptes_none = max_ptes_none;
551 static struct kobj_attribute khugepaged_max_ptes_none_attr =
552 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
553 khugepaged_max_ptes_none_store);
555 static struct attribute *khugepaged_attr[] = {
556 &khugepaged_defrag_attr.attr,
557 &khugepaged_max_ptes_none_attr.attr,
558 &pages_to_scan_attr.attr,
559 &pages_collapsed_attr.attr,
560 &full_scans_attr.attr,
561 &scan_sleep_millisecs_attr.attr,
562 &alloc_sleep_millisecs_attr.attr,
566 static struct attribute_group khugepaged_attr_group = {
567 .attrs = khugepaged_attr,
568 .name = "khugepaged",
571 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
575 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
576 if (unlikely(!*hugepage_kobj)) {
577 pr_err("failed to create transparent hugepage kobject\n");
581 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
583 pr_err("failed to register transparent hugepage group\n");
587 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
589 pr_err("failed to register transparent hugepage group\n");
590 goto remove_hp_group;
596 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
598 kobject_put(*hugepage_kobj);
602 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
604 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
605 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
606 kobject_put(hugepage_kobj);
609 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
614 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
617 #endif /* CONFIG_SYSFS */
619 static int __init hugepage_init(void)
622 struct kobject *hugepage_kobj;
624 if (!has_transparent_hugepage()) {
625 transparent_hugepage_flags = 0;
629 err = hugepage_init_sysfs(&hugepage_kobj);
633 err = khugepaged_slab_init();
637 err = register_shrinker(&huge_zero_page_shrinker);
639 goto err_hzp_shrinker;
642 * By default disable transparent hugepages on smaller systems,
643 * where the extra memory used could hurt more than TLB overhead
644 * is likely to save. The admin can still enable it through /sys.
646 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
647 transparent_hugepage_flags = 0;
651 err = start_stop_khugepaged();
657 unregister_shrinker(&huge_zero_page_shrinker);
659 khugepaged_slab_exit();
661 hugepage_exit_sysfs(hugepage_kobj);
665 subsys_initcall(hugepage_init);
667 static int __init setup_transparent_hugepage(char *str)
672 if (!strcmp(str, "always")) {
673 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
674 &transparent_hugepage_flags);
675 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
676 &transparent_hugepage_flags);
678 } else if (!strcmp(str, "madvise")) {
679 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
680 &transparent_hugepage_flags);
681 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
682 &transparent_hugepage_flags);
684 } else if (!strcmp(str, "never")) {
685 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
686 &transparent_hugepage_flags);
687 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
688 &transparent_hugepage_flags);
693 pr_warn("transparent_hugepage= cannot parse, ignored\n");
696 __setup("transparent_hugepage=", setup_transparent_hugepage);
698 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
700 if (likely(vma->vm_flags & VM_WRITE))
701 pmd = pmd_mkwrite(pmd);
705 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
708 entry = mk_pmd(page, prot);
709 entry = pmd_mkhuge(entry);
713 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
714 struct vm_area_struct *vma,
715 unsigned long address, pmd_t *pmd,
716 struct page *page, gfp_t gfp,
719 struct mem_cgroup *memcg;
722 unsigned long haddr = address & HPAGE_PMD_MASK;
724 VM_BUG_ON_PAGE(!PageCompound(page), page);
726 if (mem_cgroup_try_charge(page, mm, gfp, &memcg)) {
728 count_vm_event(THP_FAULT_FALLBACK);
729 return VM_FAULT_FALLBACK;
732 pgtable = pte_alloc_one(mm, haddr);
733 if (unlikely(!pgtable)) {
734 mem_cgroup_cancel_charge(page, memcg);
739 clear_huge_page(page, haddr, HPAGE_PMD_NR);
741 * The memory barrier inside __SetPageUptodate makes sure that
742 * clear_huge_page writes become visible before the set_pmd_at()
745 __SetPageUptodate(page);
747 ptl = pmd_lock(mm, pmd);
748 if (unlikely(!pmd_none(*pmd))) {
750 mem_cgroup_cancel_charge(page, memcg);
752 pte_free(mm, pgtable);
756 /* Deliver the page fault to userland */
757 if (userfaultfd_missing(vma)) {
761 mem_cgroup_cancel_charge(page, memcg);
763 pte_free(mm, pgtable);
764 ret = handle_userfault(vma, address, flags,
766 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
770 entry = mk_huge_pmd(page, vma->vm_page_prot);
771 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
772 page_add_new_anon_rmap(page, vma, haddr);
773 mem_cgroup_commit_charge(page, memcg, false);
774 lru_cache_add_active_or_unevictable(page, vma);
775 pgtable_trans_huge_deposit(mm, pmd, pgtable);
776 set_pmd_at(mm, haddr, pmd, entry);
777 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
778 atomic_long_inc(&mm->nr_ptes);
780 count_vm_event(THP_FAULT_ALLOC);
786 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
788 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
791 /* Caller must hold page table lock. */
792 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
793 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
794 struct page *zero_page)
799 entry = mk_pmd(zero_page, vma->vm_page_prot);
800 entry = pmd_mkhuge(entry);
801 pgtable_trans_huge_deposit(mm, pmd, pgtable);
802 set_pmd_at(mm, haddr, pmd, entry);
803 atomic_long_inc(&mm->nr_ptes);
807 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
808 unsigned long address, pmd_t *pmd,
813 unsigned long haddr = address & HPAGE_PMD_MASK;
815 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
816 return VM_FAULT_FALLBACK;
817 if (unlikely(anon_vma_prepare(vma)))
819 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
821 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
822 transparent_hugepage_use_zero_page()) {
825 struct page *zero_page;
828 pgtable = pte_alloc_one(mm, haddr);
829 if (unlikely(!pgtable))
831 zero_page = get_huge_zero_page();
832 if (unlikely(!zero_page)) {
833 pte_free(mm, pgtable);
834 count_vm_event(THP_FAULT_FALLBACK);
835 return VM_FAULT_FALLBACK;
837 ptl = pmd_lock(mm, pmd);
840 if (pmd_none(*pmd)) {
841 if (userfaultfd_missing(vma)) {
843 ret = handle_userfault(vma, address, flags,
845 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
847 set_huge_zero_page(pgtable, mm, vma,
856 pte_free(mm, pgtable);
857 put_huge_zero_page();
861 gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
862 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
863 if (unlikely(!page)) {
864 count_vm_event(THP_FAULT_FALLBACK);
865 return VM_FAULT_FALLBACK;
867 return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
871 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
872 pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write)
874 struct mm_struct *mm = vma->vm_mm;
878 ptl = pmd_lock(mm, pmd);
879 if (pmd_none(*pmd)) {
880 entry = pmd_mkhuge(pfn_pmd(pfn, prot));
882 entry = pmd_mkyoung(pmd_mkdirty(entry));
883 entry = maybe_pmd_mkwrite(entry, vma);
885 set_pmd_at(mm, addr, pmd, entry);
886 update_mmu_cache_pmd(vma, addr, pmd);
891 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
892 pmd_t *pmd, unsigned long pfn, bool write)
894 pgprot_t pgprot = vma->vm_page_prot;
896 * If we had pmd_special, we could avoid all these restrictions,
897 * but we need to be consistent with PTEs and architectures that
898 * can't support a 'special' bit.
900 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
901 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
902 (VM_PFNMAP|VM_MIXEDMAP));
903 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
904 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
906 if (addr < vma->vm_start || addr >= vma->vm_end)
907 return VM_FAULT_SIGBUS;
908 if (track_pfn_insert(vma, &pgprot, pfn))
909 return VM_FAULT_SIGBUS;
910 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
911 return VM_FAULT_NOPAGE;
914 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
915 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
916 struct vm_area_struct *vma)
918 spinlock_t *dst_ptl, *src_ptl;
919 struct page *src_page;
925 pgtable = pte_alloc_one(dst_mm, addr);
926 if (unlikely(!pgtable))
929 dst_ptl = pmd_lock(dst_mm, dst_pmd);
930 src_ptl = pmd_lockptr(src_mm, src_pmd);
931 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
935 if (unlikely(!pmd_trans_huge(pmd))) {
936 pte_free(dst_mm, pgtable);
940 * When page table lock is held, the huge zero pmd should not be
941 * under splitting since we don't split the page itself, only pmd to
944 if (is_huge_zero_pmd(pmd)) {
945 struct page *zero_page;
947 * get_huge_zero_page() will never allocate a new page here,
948 * since we already have a zero page to copy. It just takes a
951 zero_page = get_huge_zero_page();
952 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
958 if (unlikely(pmd_trans_splitting(pmd))) {
959 /* split huge page running from under us */
960 spin_unlock(src_ptl);
961 spin_unlock(dst_ptl);
962 pte_free(dst_mm, pgtable);
964 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
967 src_page = pmd_page(pmd);
968 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
970 page_dup_rmap(src_page);
971 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
973 pmdp_set_wrprotect(src_mm, addr, src_pmd);
974 pmd = pmd_mkold(pmd_wrprotect(pmd));
975 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
976 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
977 atomic_long_inc(&dst_mm->nr_ptes);
981 spin_unlock(src_ptl);
982 spin_unlock(dst_ptl);
987 void huge_pmd_set_accessed(struct mm_struct *mm,
988 struct vm_area_struct *vma,
989 unsigned long address,
990 pmd_t *pmd, pmd_t orig_pmd,
997 ptl = pmd_lock(mm, pmd);
998 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1001 entry = pmd_mkyoung(orig_pmd);
1002 haddr = address & HPAGE_PMD_MASK;
1003 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1004 update_mmu_cache_pmd(vma, address, pmd);
1011 * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
1012 * during copy_user_huge_page()'s copy_page_rep(): in the case when
1013 * the source page gets split and a tail freed before copy completes.
1014 * Called under pmd_lock of checked pmd, so safe from splitting itself.
1016 static void get_user_huge_page(struct page *page)
1018 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1019 struct page *endpage = page + HPAGE_PMD_NR;
1021 atomic_add(HPAGE_PMD_NR, &page->_count);
1022 while (++page < endpage)
1023 get_huge_page_tail(page);
1029 static void put_user_huge_page(struct page *page)
1031 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1032 struct page *endpage = page + HPAGE_PMD_NR;
1034 while (page < endpage)
1041 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1042 struct vm_area_struct *vma,
1043 unsigned long address,
1044 pmd_t *pmd, pmd_t orig_pmd,
1046 unsigned long haddr)
1048 struct mem_cgroup *memcg;
1053 struct page **pages;
1054 unsigned long mmun_start; /* For mmu_notifiers */
1055 unsigned long mmun_end; /* For mmu_notifiers */
1057 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1059 if (unlikely(!pages)) {
1060 ret |= VM_FAULT_OOM;
1064 for (i = 0; i < HPAGE_PMD_NR; i++) {
1065 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1067 vma, address, page_to_nid(page));
1068 if (unlikely(!pages[i] ||
1069 mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1074 memcg = (void *)page_private(pages[i]);
1075 set_page_private(pages[i], 0);
1076 mem_cgroup_cancel_charge(pages[i], memcg);
1080 ret |= VM_FAULT_OOM;
1083 set_page_private(pages[i], (unsigned long)memcg);
1086 for (i = 0; i < HPAGE_PMD_NR; i++) {
1087 copy_user_highpage(pages[i], page + i,
1088 haddr + PAGE_SIZE * i, vma);
1089 __SetPageUptodate(pages[i]);
1094 mmun_end = haddr + HPAGE_PMD_SIZE;
1095 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1097 ptl = pmd_lock(mm, pmd);
1098 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1099 goto out_free_pages;
1100 VM_BUG_ON_PAGE(!PageHead(page), page);
1102 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1103 /* leave pmd empty until pte is filled */
1105 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1106 pmd_populate(mm, &_pmd, pgtable);
1108 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1110 entry = mk_pte(pages[i], vma->vm_page_prot);
1111 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1112 memcg = (void *)page_private(pages[i]);
1113 set_page_private(pages[i], 0);
1114 page_add_new_anon_rmap(pages[i], vma, haddr);
1115 mem_cgroup_commit_charge(pages[i], memcg, false);
1116 lru_cache_add_active_or_unevictable(pages[i], vma);
1117 pte = pte_offset_map(&_pmd, haddr);
1118 VM_BUG_ON(!pte_none(*pte));
1119 set_pte_at(mm, haddr, pte, entry);
1124 smp_wmb(); /* make pte visible before pmd */
1125 pmd_populate(mm, pmd, pgtable);
1126 page_remove_rmap(page);
1129 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1131 ret |= VM_FAULT_WRITE;
1139 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1140 for (i = 0; i < HPAGE_PMD_NR; i++) {
1141 memcg = (void *)page_private(pages[i]);
1142 set_page_private(pages[i], 0);
1143 mem_cgroup_cancel_charge(pages[i], memcg);
1150 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1151 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1155 struct page *page = NULL, *new_page;
1156 struct mem_cgroup *memcg;
1157 unsigned long haddr;
1158 unsigned long mmun_start; /* For mmu_notifiers */
1159 unsigned long mmun_end; /* For mmu_notifiers */
1160 gfp_t huge_gfp; /* for allocation and charge */
1162 ptl = pmd_lockptr(mm, pmd);
1163 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1164 haddr = address & HPAGE_PMD_MASK;
1165 if (is_huge_zero_pmd(orig_pmd))
1168 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1171 page = pmd_page(orig_pmd);
1172 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1173 if (page_mapcount(page) == 1) {
1175 entry = pmd_mkyoung(orig_pmd);
1176 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1177 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
1178 update_mmu_cache_pmd(vma, address, pmd);
1179 ret |= VM_FAULT_WRITE;
1182 get_user_huge_page(page);
1185 if (transparent_hugepage_enabled(vma) &&
1186 !transparent_hugepage_debug_cow()) {
1187 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1188 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1192 if (unlikely(!new_page)) {
1194 split_huge_page_pmd(vma, address, pmd);
1195 ret |= VM_FAULT_FALLBACK;
1197 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1198 pmd, orig_pmd, page, haddr);
1199 if (ret & VM_FAULT_OOM) {
1200 split_huge_page(page);
1201 ret |= VM_FAULT_FALLBACK;
1203 put_user_huge_page(page);
1205 count_vm_event(THP_FAULT_FALLBACK);
1209 if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
1212 split_huge_page(page);
1213 put_user_huge_page(page);
1215 split_huge_page_pmd(vma, address, pmd);
1216 ret |= VM_FAULT_FALLBACK;
1217 count_vm_event(THP_FAULT_FALLBACK);
1221 count_vm_event(THP_FAULT_ALLOC);
1224 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1226 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1227 __SetPageUptodate(new_page);
1230 mmun_end = haddr + HPAGE_PMD_SIZE;
1231 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1235 put_user_huge_page(page);
1236 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1238 mem_cgroup_cancel_charge(new_page, memcg);
1243 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1244 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1245 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1246 page_add_new_anon_rmap(new_page, vma, haddr);
1247 mem_cgroup_commit_charge(new_page, memcg, false);
1248 lru_cache_add_active_or_unevictable(new_page, vma);
1249 set_pmd_at(mm, haddr, pmd, entry);
1250 update_mmu_cache_pmd(vma, address, pmd);
1252 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1253 put_huge_zero_page();
1255 VM_BUG_ON_PAGE(!PageHead(page), page);
1256 page_remove_rmap(page);
1259 ret |= VM_FAULT_WRITE;
1263 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1271 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1276 struct mm_struct *mm = vma->vm_mm;
1277 struct page *page = NULL;
1279 assert_spin_locked(pmd_lockptr(mm, pmd));
1281 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1284 /* Avoid dumping huge zero page */
1285 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1286 return ERR_PTR(-EFAULT);
1288 /* Full NUMA hinting faults to serialise migration in fault paths */
1289 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1292 page = pmd_page(*pmd);
1293 VM_BUG_ON_PAGE(!PageHead(page), page);
1294 if (flags & FOLL_TOUCH) {
1297 * We should set the dirty bit only for FOLL_WRITE but
1298 * for now the dirty bit in the pmd is meaningless.
1299 * And if the dirty bit will become meaningful and
1300 * we'll only set it with FOLL_WRITE, an atomic
1301 * set_bit will be required on the pmd to set the
1302 * young bit, instead of the current set_pmd_at.
1304 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1305 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1307 update_mmu_cache_pmd(vma, addr, pmd);
1309 if ((flags & FOLL_POPULATE) && (vma->vm_flags & VM_LOCKED)) {
1310 if (page->mapping && trylock_page(page)) {
1313 mlock_vma_page(page);
1317 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1318 VM_BUG_ON_PAGE(!PageCompound(page), page);
1319 if (flags & FOLL_GET)
1320 get_page_foll(page);
1326 /* NUMA hinting page fault entry point for trans huge pmds */
1327 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1328 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1331 struct anon_vma *anon_vma = NULL;
1333 unsigned long haddr = addr & HPAGE_PMD_MASK;
1334 int page_nid = -1, this_nid = numa_node_id();
1335 int target_nid, last_cpupid = -1;
1337 bool migrated = false;
1341 /* A PROT_NONE fault should not end up here */
1342 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1344 ptl = pmd_lock(mm, pmdp);
1345 if (unlikely(!pmd_same(pmd, *pmdp)))
1349 * If there are potential migrations, wait for completion and retry
1350 * without disrupting NUMA hinting information. Do not relock and
1351 * check_same as the page may no longer be mapped.
1353 if (unlikely(pmd_trans_migrating(*pmdp))) {
1354 page = pmd_page(*pmdp);
1356 wait_on_page_locked(page);
1360 page = pmd_page(pmd);
1361 BUG_ON(is_huge_zero_page(page));
1362 page_nid = page_to_nid(page);
1363 last_cpupid = page_cpupid_last(page);
1364 count_vm_numa_event(NUMA_HINT_FAULTS);
1365 if (page_nid == this_nid) {
1366 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1367 flags |= TNF_FAULT_LOCAL;
1370 /* See similar comment in do_numa_page for explanation */
1371 if (!(vma->vm_flags & VM_WRITE))
1372 flags |= TNF_NO_GROUP;
1375 * Acquire the page lock to serialise THP migrations but avoid dropping
1376 * page_table_lock if at all possible
1378 page_locked = trylock_page(page);
1379 target_nid = mpol_misplaced(page, vma, haddr);
1380 if (target_nid == -1) {
1381 /* If the page was locked, there are no parallel migrations */
1386 /* Migration could have started since the pmd_trans_migrating check */
1389 wait_on_page_locked(page);
1395 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1396 * to serialises splits
1400 anon_vma = page_lock_anon_vma_read(page);
1402 /* Confirm the PMD did not change while page_table_lock was released */
1404 if (unlikely(!pmd_same(pmd, *pmdp))) {
1411 /* Bail if we fail to protect against THP splits for any reason */
1412 if (unlikely(!anon_vma)) {
1419 * Migrate the THP to the requested node, returns with page unlocked
1420 * and access rights restored.
1423 migrated = migrate_misplaced_transhuge_page(mm, vma,
1424 pmdp, pmd, addr, page, target_nid);
1426 flags |= TNF_MIGRATED;
1427 page_nid = target_nid;
1429 flags |= TNF_MIGRATE_FAIL;
1433 BUG_ON(!PageLocked(page));
1434 was_writable = pmd_write(pmd);
1435 pmd = pmd_modify(pmd, vma->vm_page_prot);
1436 pmd = pmd_mkyoung(pmd);
1438 pmd = pmd_mkwrite(pmd);
1439 set_pmd_at(mm, haddr, pmdp, pmd);
1440 update_mmu_cache_pmd(vma, addr, pmdp);
1447 page_unlock_anon_vma_read(anon_vma);
1450 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1455 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1456 pmd_t *pmd, unsigned long addr)
1461 if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1)
1464 * For architectures like ppc64 we look at deposited pgtable
1465 * when calling pmdp_huge_get_and_clear. So do the
1466 * pgtable_trans_huge_withdraw after finishing pmdp related
1469 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1471 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1472 if (vma_is_dax(vma)) {
1474 if (is_huge_zero_pmd(orig_pmd))
1475 put_huge_zero_page();
1476 } else if (is_huge_zero_pmd(orig_pmd)) {
1477 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1478 atomic_long_dec(&tlb->mm->nr_ptes);
1480 put_huge_zero_page();
1482 struct page *page = pmd_page(orig_pmd);
1483 page_remove_rmap(page);
1484 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1485 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1486 VM_BUG_ON_PAGE(!PageHead(page), page);
1487 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1488 atomic_long_dec(&tlb->mm->nr_ptes);
1490 tlb_remove_page(tlb, page);
1495 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1496 unsigned long old_addr,
1497 unsigned long new_addr, unsigned long old_end,
1498 pmd_t *old_pmd, pmd_t *new_pmd)
1500 spinlock_t *old_ptl, *new_ptl;
1504 struct mm_struct *mm = vma->vm_mm;
1506 if ((old_addr & ~HPAGE_PMD_MASK) ||
1507 (new_addr & ~HPAGE_PMD_MASK) ||
1508 old_end - old_addr < HPAGE_PMD_SIZE ||
1509 (new_vma->vm_flags & VM_NOHUGEPAGE))
1513 * The destination pmd shouldn't be established, free_pgtables()
1514 * should have release it.
1516 if (WARN_ON(!pmd_none(*new_pmd))) {
1517 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1522 * We don't have to worry about the ordering of src and dst
1523 * ptlocks because exclusive mmap_sem prevents deadlock.
1525 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1527 new_ptl = pmd_lockptr(mm, new_pmd);
1528 if (new_ptl != old_ptl)
1529 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1530 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1531 VM_BUG_ON(!pmd_none(*new_pmd));
1533 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1535 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1536 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1538 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1539 if (new_ptl != old_ptl)
1540 spin_unlock(new_ptl);
1541 spin_unlock(old_ptl);
1549 * - 0 if PMD could not be locked
1550 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1551 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1553 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1554 unsigned long addr, pgprot_t newprot, int prot_numa)
1556 struct mm_struct *mm = vma->vm_mm;
1560 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1562 bool preserve_write = prot_numa && pmd_write(*pmd);
1566 * Avoid trapping faults against the zero page. The read-only
1567 * data is likely to be read-cached on the local CPU and
1568 * local/remote hits to the zero page are not interesting.
1570 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1575 if (!prot_numa || !pmd_protnone(*pmd)) {
1576 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1577 entry = pmd_modify(entry, newprot);
1579 entry = pmd_mkwrite(entry);
1581 set_pmd_at(mm, addr, pmd, entry);
1582 BUG_ON(!preserve_write && pmd_write(entry));
1591 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1592 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1594 * Note that if it returns 1, this routine returns without unlocking page
1595 * table locks. So callers must unlock them.
1597 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1600 *ptl = pmd_lock(vma->vm_mm, pmd);
1601 if (likely(pmd_trans_huge(*pmd))) {
1602 if (unlikely(pmd_trans_splitting(*pmd))) {
1604 wait_split_huge_page(vma->anon_vma, pmd);
1607 /* Thp mapped by 'pmd' is stable, so we can
1608 * handle it as it is. */
1617 * This function returns whether a given @page is mapped onto the @address
1618 * in the virtual space of @mm.
1620 * When it's true, this function returns *pmd with holding the page table lock
1621 * and passing it back to the caller via @ptl.
1622 * If it's false, returns NULL without holding the page table lock.
1624 pmd_t *page_check_address_pmd(struct page *page,
1625 struct mm_struct *mm,
1626 unsigned long address,
1627 enum page_check_address_pmd_flag flag,
1634 if (address & ~HPAGE_PMD_MASK)
1637 pgd = pgd_offset(mm, address);
1638 if (!pgd_present(*pgd))
1640 pud = pud_offset(pgd, address);
1641 if (!pud_present(*pud))
1643 pmd = pmd_offset(pud, address);
1645 *ptl = pmd_lock(mm, pmd);
1646 if (!pmd_present(*pmd))
1648 if (pmd_page(*pmd) != page)
1651 * split_vma() may create temporary aliased mappings. There is
1652 * no risk as long as all huge pmd are found and have their
1653 * splitting bit set before __split_huge_page_refcount
1654 * runs. Finding the same huge pmd more than once during the
1655 * same rmap walk is not a problem.
1657 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1658 pmd_trans_splitting(*pmd))
1660 if (pmd_trans_huge(*pmd)) {
1661 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1662 !pmd_trans_splitting(*pmd));
1670 static int __split_huge_page_splitting(struct page *page,
1671 struct vm_area_struct *vma,
1672 unsigned long address)
1674 struct mm_struct *mm = vma->vm_mm;
1678 /* For mmu_notifiers */
1679 const unsigned long mmun_start = address;
1680 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
1682 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1683 pmd = page_check_address_pmd(page, mm, address,
1684 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1687 * We can't temporarily set the pmd to null in order
1688 * to split it, the pmd must remain marked huge at all
1689 * times or the VM won't take the pmd_trans_huge paths
1690 * and it won't wait on the anon_vma->root->rwsem to
1691 * serialize against split_huge_page*.
1693 pmdp_splitting_flush(vma, address, pmd);
1698 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1703 static void __split_huge_page_refcount(struct page *page,
1704 struct list_head *list)
1707 struct zone *zone = page_zone(page);
1708 struct lruvec *lruvec;
1711 /* prevent PageLRU to go away from under us, and freeze lru stats */
1712 spin_lock_irq(&zone->lru_lock);
1713 lruvec = mem_cgroup_page_lruvec(page, zone);
1715 compound_lock(page);
1716 /* complete memcg works before add pages to LRU */
1717 mem_cgroup_split_huge_fixup(page);
1719 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1720 struct page *page_tail = page + i;
1722 /* tail_page->_mapcount cannot change */
1723 BUG_ON(page_mapcount(page_tail) < 0);
1724 tail_count += page_mapcount(page_tail);
1725 /* check for overflow */
1726 BUG_ON(tail_count < 0);
1727 BUG_ON(atomic_read(&page_tail->_count) != 0);
1729 * tail_page->_count is zero and not changing from
1730 * under us. But get_page_unless_zero() may be running
1731 * from under us on the tail_page. If we used
1732 * atomic_set() below instead of atomic_add(), we
1733 * would then run atomic_set() concurrently with
1734 * get_page_unless_zero(), and atomic_set() is
1735 * implemented in C not using locked ops. spin_unlock
1736 * on x86 sometime uses locked ops because of PPro
1737 * errata 66, 92, so unless somebody can guarantee
1738 * atomic_set() here would be safe on all archs (and
1739 * not only on x86), it's safer to use atomic_add().
1741 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1742 &page_tail->_count);
1744 /* after clearing PageTail the gup refcount can be released */
1745 smp_mb__after_atomic();
1747 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1748 page_tail->flags |= (page->flags &
1749 ((1L << PG_referenced) |
1750 (1L << PG_swapbacked) |
1751 (1L << PG_mlocked) |
1752 (1L << PG_uptodate) |
1754 (1L << PG_unevictable)));
1755 page_tail->flags |= (1L << PG_dirty);
1757 /* clear PageTail before overwriting first_page */
1761 * __split_huge_page_splitting() already set the
1762 * splitting bit in all pmd that could map this
1763 * hugepage, that will ensure no CPU can alter the
1764 * mapcount on the head page. The mapcount is only
1765 * accounted in the head page and it has to be
1766 * transferred to all tail pages in the below code. So
1767 * for this code to be safe, the split the mapcount
1768 * can't change. But that doesn't mean userland can't
1769 * keep changing and reading the page contents while
1770 * we transfer the mapcount, so the pmd splitting
1771 * status is achieved setting a reserved bit in the
1772 * pmd, not by clearing the present bit.
1774 page_tail->_mapcount = page->_mapcount;
1776 BUG_ON(page_tail->mapping);
1777 page_tail->mapping = page->mapping;
1779 page_tail->index = page->index + i;
1780 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1782 BUG_ON(!PageAnon(page_tail));
1783 BUG_ON(!PageUptodate(page_tail));
1784 BUG_ON(!PageDirty(page_tail));
1785 BUG_ON(!PageSwapBacked(page_tail));
1787 lru_add_page_tail(page, page_tail, lruvec, list);
1789 atomic_sub(tail_count, &page->_count);
1790 BUG_ON(atomic_read(&page->_count) <= 0);
1792 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1794 ClearPageCompound(page);
1795 compound_unlock(page);
1796 spin_unlock_irq(&zone->lru_lock);
1798 for (i = 1; i < HPAGE_PMD_NR; i++) {
1799 struct page *page_tail = page + i;
1800 BUG_ON(page_count(page_tail) <= 0);
1802 * Tail pages may be freed if there wasn't any mapping
1803 * like if add_to_swap() is running on a lru page that
1804 * had its mapping zapped. And freeing these pages
1805 * requires taking the lru_lock so we do the put_page
1806 * of the tail pages after the split is complete.
1808 put_page(page_tail);
1812 * Only the head page (now become a regular page) is required
1813 * to be pinned by the caller.
1815 BUG_ON(page_count(page) <= 0);
1818 static int __split_huge_page_map(struct page *page,
1819 struct vm_area_struct *vma,
1820 unsigned long address)
1822 struct mm_struct *mm = vma->vm_mm;
1827 unsigned long haddr;
1829 pmd = page_check_address_pmd(page, mm, address,
1830 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1832 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1833 pmd_populate(mm, &_pmd, pgtable);
1834 if (pmd_write(*pmd))
1835 BUG_ON(page_mapcount(page) != 1);
1838 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1840 BUG_ON(PageCompound(page+i));
1842 * Note that NUMA hinting access restrictions are not
1843 * transferred to avoid any possibility of altering
1844 * permissions across VMAs.
1846 entry = mk_pte(page + i, vma->vm_page_prot);
1847 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1848 if (!pmd_write(*pmd))
1849 entry = pte_wrprotect(entry);
1850 if (!pmd_young(*pmd))
1851 entry = pte_mkold(entry);
1852 pte = pte_offset_map(&_pmd, haddr);
1853 BUG_ON(!pte_none(*pte));
1854 set_pte_at(mm, haddr, pte, entry);
1858 smp_wmb(); /* make pte visible before pmd */
1860 * Up to this point the pmd is present and huge and
1861 * userland has the whole access to the hugepage
1862 * during the split (which happens in place). If we
1863 * overwrite the pmd with the not-huge version
1864 * pointing to the pte here (which of course we could
1865 * if all CPUs were bug free), userland could trigger
1866 * a small page size TLB miss on the small sized TLB
1867 * while the hugepage TLB entry is still established
1868 * in the huge TLB. Some CPU doesn't like that. See
1869 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1870 * Erratum 383 on page 93. Intel should be safe but is
1871 * also warns that it's only safe if the permission
1872 * and cache attributes of the two entries loaded in
1873 * the two TLB is identical (which should be the case
1874 * here). But it is generally safer to never allow
1875 * small and huge TLB entries for the same virtual
1876 * address to be loaded simultaneously. So instead of
1877 * doing "pmd_populate(); flush_tlb_range();" we first
1878 * mark the current pmd notpresent (atomically because
1879 * here the pmd_trans_huge and pmd_trans_splitting
1880 * must remain set at all times on the pmd until the
1881 * split is complete for this pmd), then we flush the
1882 * SMP TLB and finally we write the non-huge version
1883 * of the pmd entry with pmd_populate.
1885 pmdp_invalidate(vma, address, pmd);
1886 pmd_populate(mm, pmd, pgtable);
1894 /* must be called with anon_vma->root->rwsem held */
1895 static void __split_huge_page(struct page *page,
1896 struct anon_vma *anon_vma,
1897 struct list_head *list)
1899 int mapcount, mapcount2;
1900 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1901 struct anon_vma_chain *avc;
1903 BUG_ON(!PageHead(page));
1904 BUG_ON(PageTail(page));
1907 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1908 struct vm_area_struct *vma = avc->vma;
1909 unsigned long addr = vma_address(page, vma);
1910 BUG_ON(is_vma_temporary_stack(vma));
1911 mapcount += __split_huge_page_splitting(page, vma, addr);
1914 * It is critical that new vmas are added to the tail of the
1915 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1916 * and establishes a child pmd before
1917 * __split_huge_page_splitting() freezes the parent pmd (so if
1918 * we fail to prevent copy_huge_pmd() from running until the
1919 * whole __split_huge_page() is complete), we will still see
1920 * the newly established pmd of the child later during the
1921 * walk, to be able to set it as pmd_trans_splitting too.
1923 if (mapcount != page_mapcount(page)) {
1924 pr_err("mapcount %d page_mapcount %d\n",
1925 mapcount, page_mapcount(page));
1929 __split_huge_page_refcount(page, list);
1932 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1933 struct vm_area_struct *vma = avc->vma;
1934 unsigned long addr = vma_address(page, vma);
1935 BUG_ON(is_vma_temporary_stack(vma));
1936 mapcount2 += __split_huge_page_map(page, vma, addr);
1938 if (mapcount != mapcount2) {
1939 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1940 mapcount, mapcount2, page_mapcount(page));
1946 * Split a hugepage into normal pages. This doesn't change the position of head
1947 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1948 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1949 * from the hugepage.
1950 * Return 0 if the hugepage is split successfully otherwise return 1.
1952 int split_huge_page_to_list(struct page *page, struct list_head *list)
1954 struct anon_vma *anon_vma;
1957 BUG_ON(is_huge_zero_page(page));
1958 BUG_ON(!PageAnon(page));
1961 * The caller does not necessarily hold an mmap_sem that would prevent
1962 * the anon_vma disappearing so we first we take a reference to it
1963 * and then lock the anon_vma for write. This is similar to
1964 * page_lock_anon_vma_read except the write lock is taken to serialise
1965 * against parallel split or collapse operations.
1967 anon_vma = page_get_anon_vma(page);
1970 anon_vma_lock_write(anon_vma);
1973 if (!PageCompound(page))
1976 BUG_ON(!PageSwapBacked(page));
1977 __split_huge_page(page, anon_vma, list);
1978 count_vm_event(THP_SPLIT);
1980 BUG_ON(PageCompound(page));
1982 anon_vma_unlock_write(anon_vma);
1983 put_anon_vma(anon_vma);
1988 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1990 int hugepage_madvise(struct vm_area_struct *vma,
1991 unsigned long *vm_flags, int advice)
1997 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1998 * can't handle this properly after s390_enable_sie, so we simply
1999 * ignore the madvise to prevent qemu from causing a SIGSEGV.
2001 if (mm_has_pgste(vma->vm_mm))
2005 * Be somewhat over-protective like KSM for now!
2007 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
2009 *vm_flags &= ~VM_NOHUGEPAGE;
2010 *vm_flags |= VM_HUGEPAGE;
2012 * If the vma become good for khugepaged to scan,
2013 * register it here without waiting a page fault that
2014 * may not happen any time soon.
2016 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
2019 case MADV_NOHUGEPAGE:
2021 * Be somewhat over-protective like KSM for now!
2023 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
2025 *vm_flags &= ~VM_HUGEPAGE;
2026 *vm_flags |= VM_NOHUGEPAGE;
2028 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2029 * this vma even if we leave the mm registered in khugepaged if
2030 * it got registered before VM_NOHUGEPAGE was set.
2038 static int __init khugepaged_slab_init(void)
2040 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2041 sizeof(struct mm_slot),
2042 __alignof__(struct mm_slot), 0, NULL);
2049 static void __init khugepaged_slab_exit(void)
2051 kmem_cache_destroy(mm_slot_cache);
2054 static inline struct mm_slot *alloc_mm_slot(void)
2056 if (!mm_slot_cache) /* initialization failed */
2058 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2061 static inline void free_mm_slot(struct mm_slot *mm_slot)
2063 kmem_cache_free(mm_slot_cache, mm_slot);
2066 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2068 struct mm_slot *mm_slot;
2070 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2071 if (mm == mm_slot->mm)
2077 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2078 struct mm_slot *mm_slot)
2081 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2084 static inline int khugepaged_test_exit(struct mm_struct *mm)
2086 return atomic_read(&mm->mm_users) == 0;
2089 int __khugepaged_enter(struct mm_struct *mm)
2091 struct mm_slot *mm_slot;
2094 mm_slot = alloc_mm_slot();
2098 /* __khugepaged_exit() must not run from under us */
2099 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2100 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2101 free_mm_slot(mm_slot);
2105 spin_lock(&khugepaged_mm_lock);
2106 insert_to_mm_slots_hash(mm, mm_slot);
2108 * Insert just behind the scanning cursor, to let the area settle
2111 wakeup = list_empty(&khugepaged_scan.mm_head);
2112 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2113 spin_unlock(&khugepaged_mm_lock);
2115 atomic_inc(&mm->mm_count);
2117 wake_up_interruptible(&khugepaged_wait);
2122 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2123 unsigned long vm_flags)
2125 unsigned long hstart, hend;
2128 * Not yet faulted in so we will register later in the
2129 * page fault if needed.
2133 /* khugepaged not yet working on file or special mappings */
2135 VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
2136 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2137 hend = vma->vm_end & HPAGE_PMD_MASK;
2139 return khugepaged_enter(vma, vm_flags);
2143 void __khugepaged_exit(struct mm_struct *mm)
2145 struct mm_slot *mm_slot;
2148 spin_lock(&khugepaged_mm_lock);
2149 mm_slot = get_mm_slot(mm);
2150 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2151 hash_del(&mm_slot->hash);
2152 list_del(&mm_slot->mm_node);
2155 spin_unlock(&khugepaged_mm_lock);
2158 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2159 free_mm_slot(mm_slot);
2161 } else if (mm_slot) {
2163 * This is required to serialize against
2164 * khugepaged_test_exit() (which is guaranteed to run
2165 * under mmap sem read mode). Stop here (after we
2166 * return all pagetables will be destroyed) until
2167 * khugepaged has finished working on the pagetables
2168 * under the mmap_sem.
2170 down_write(&mm->mmap_sem);
2171 up_write(&mm->mmap_sem);
2175 static void release_pte_page(struct page *page)
2177 /* 0 stands for page_is_file_cache(page) == false */
2178 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2180 putback_lru_page(page);
2183 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2185 while (--_pte >= pte) {
2186 pte_t pteval = *_pte;
2187 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2188 release_pte_page(pte_page(pteval));
2192 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2193 unsigned long address,
2198 int none_or_zero = 0;
2199 bool referenced = false, writable = false;
2200 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2201 _pte++, address += PAGE_SIZE) {
2202 pte_t pteval = *_pte;
2203 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2204 if (!userfaultfd_armed(vma) &&
2205 ++none_or_zero <= khugepaged_max_ptes_none)
2210 if (!pte_present(pteval))
2212 page = vm_normal_page(vma, address, pteval);
2213 if (unlikely(!page))
2216 VM_BUG_ON_PAGE(PageCompound(page), page);
2217 VM_BUG_ON_PAGE(!PageAnon(page), page);
2218 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2221 * We can do it before isolate_lru_page because the
2222 * page can't be freed from under us. NOTE: PG_lock
2223 * is needed to serialize against split_huge_page
2224 * when invoked from the VM.
2226 if (!trylock_page(page))
2230 * cannot use mapcount: can't collapse if there's a gup pin.
2231 * The page must only be referenced by the scanned process
2232 * and page swap cache.
2234 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2238 if (pte_write(pteval)) {
2241 if (PageSwapCache(page) && !reuse_swap_page(page)) {
2246 * Page is not in the swap cache. It can be collapsed
2252 * Isolate the page to avoid collapsing an hugepage
2253 * currently in use by the VM.
2255 if (isolate_lru_page(page)) {
2259 /* 0 stands for page_is_file_cache(page) == false */
2260 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2261 VM_BUG_ON_PAGE(!PageLocked(page), page);
2262 VM_BUG_ON_PAGE(PageLRU(page), page);
2264 /* If there is no mapped pte young don't collapse the page */
2265 if (pte_young(pteval) || PageReferenced(page) ||
2266 mmu_notifier_test_young(vma->vm_mm, address))
2269 if (likely(referenced && writable))
2272 release_pte_pages(pte, _pte);
2276 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2277 struct vm_area_struct *vma,
2278 unsigned long address,
2282 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2283 pte_t pteval = *_pte;
2284 struct page *src_page;
2286 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2287 clear_user_highpage(page, address);
2288 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2289 if (is_zero_pfn(pte_pfn(pteval))) {
2291 * ptl mostly unnecessary.
2295 * paravirt calls inside pte_clear here are
2298 pte_clear(vma->vm_mm, address, _pte);
2302 src_page = pte_page(pteval);
2303 copy_user_highpage(page, src_page, address, vma);
2304 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2305 release_pte_page(src_page);
2307 * ptl mostly unnecessary, but preempt has to
2308 * be disabled to update the per-cpu stats
2309 * inside page_remove_rmap().
2313 * paravirt calls inside pte_clear here are
2316 pte_clear(vma->vm_mm, address, _pte);
2317 page_remove_rmap(src_page);
2319 free_page_and_swap_cache(src_page);
2322 address += PAGE_SIZE;
2327 static void khugepaged_alloc_sleep(void)
2331 add_wait_queue(&khugepaged_wait, &wait);
2332 freezable_schedule_timeout_interruptible(
2333 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2334 remove_wait_queue(&khugepaged_wait, &wait);
2337 static int khugepaged_node_load[MAX_NUMNODES];
2339 static bool khugepaged_scan_abort(int nid)
2344 * If zone_reclaim_mode is disabled, then no extra effort is made to
2345 * allocate memory locally.
2347 if (!zone_reclaim_mode)
2350 /* If there is a count for this node already, it must be acceptable */
2351 if (khugepaged_node_load[nid])
2354 for (i = 0; i < MAX_NUMNODES; i++) {
2355 if (!khugepaged_node_load[i])
2357 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2364 static int khugepaged_find_target_node(void)
2366 static int last_khugepaged_target_node = NUMA_NO_NODE;
2367 int nid, target_node = 0, max_value = 0;
2369 /* find first node with max normal pages hit */
2370 for (nid = 0; nid < MAX_NUMNODES; nid++)
2371 if (khugepaged_node_load[nid] > max_value) {
2372 max_value = khugepaged_node_load[nid];
2376 /* do some balance if several nodes have the same hit record */
2377 if (target_node <= last_khugepaged_target_node)
2378 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2380 if (max_value == khugepaged_node_load[nid]) {
2385 last_khugepaged_target_node = target_node;
2389 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2391 if (IS_ERR(*hpage)) {
2397 khugepaged_alloc_sleep();
2398 } else if (*hpage) {
2406 static struct page *
2407 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2408 struct vm_area_struct *vma, unsigned long address,
2411 VM_BUG_ON_PAGE(*hpage, *hpage);
2414 * Before allocating the hugepage, release the mmap_sem read lock.
2415 * The allocation can take potentially a long time if it involves
2416 * sync compaction, and we do not need to hold the mmap_sem during
2417 * that. We will recheck the vma after taking it again in write mode.
2419 up_read(&mm->mmap_sem);
2421 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2422 if (unlikely(!*hpage)) {
2423 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2424 *hpage = ERR_PTR(-ENOMEM);
2428 count_vm_event(THP_COLLAPSE_ALLOC);
2432 static int khugepaged_find_target_node(void)
2437 static inline struct page *alloc_hugepage(int defrag)
2439 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2443 static struct page *khugepaged_alloc_hugepage(bool *wait)
2448 hpage = alloc_hugepage(khugepaged_defrag());
2450 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2455 khugepaged_alloc_sleep();
2457 count_vm_event(THP_COLLAPSE_ALLOC);
2458 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2463 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2466 *hpage = khugepaged_alloc_hugepage(wait);
2468 if (unlikely(!*hpage))
2474 static struct page *
2475 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2476 struct vm_area_struct *vma, unsigned long address,
2479 up_read(&mm->mmap_sem);
2486 static bool hugepage_vma_check(struct vm_area_struct *vma)
2488 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2489 (vma->vm_flags & VM_NOHUGEPAGE))
2492 if (!vma->anon_vma || vma->vm_ops)
2494 if (is_vma_temporary_stack(vma))
2496 VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2500 static void collapse_huge_page(struct mm_struct *mm,
2501 unsigned long address,
2502 struct page **hpage,
2503 struct vm_area_struct *vma,
2509 struct page *new_page;
2510 spinlock_t *pmd_ptl, *pte_ptl;
2512 unsigned long hstart, hend;
2513 struct mem_cgroup *memcg;
2514 unsigned long mmun_start; /* For mmu_notifiers */
2515 unsigned long mmun_end; /* For mmu_notifiers */
2518 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2520 /* Only allocate from the target node */
2521 gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2524 /* release the mmap_sem read lock. */
2525 new_page = khugepaged_alloc_page(hpage, gfp, mm, vma, address, node);
2529 if (unlikely(mem_cgroup_try_charge(new_page, mm,
2534 * Prevent all access to pagetables with the exception of
2535 * gup_fast later hanlded by the ptep_clear_flush and the VM
2536 * handled by the anon_vma lock + PG_lock.
2538 down_write(&mm->mmap_sem);
2539 if (unlikely(khugepaged_test_exit(mm)))
2542 vma = find_vma(mm, address);
2545 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2546 hend = vma->vm_end & HPAGE_PMD_MASK;
2547 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2549 if (!hugepage_vma_check(vma))
2551 pmd = mm_find_pmd(mm, address);
2555 anon_vma_lock_write(vma->anon_vma);
2557 pte = pte_offset_map(pmd, address);
2558 pte_ptl = pte_lockptr(mm, pmd);
2560 mmun_start = address;
2561 mmun_end = address + HPAGE_PMD_SIZE;
2562 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2563 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2565 * After this gup_fast can't run anymore. This also removes
2566 * any huge TLB entry from the CPU so we won't allow
2567 * huge and small TLB entries for the same virtual address
2568 * to avoid the risk of CPU bugs in that area.
2570 _pmd = pmdp_collapse_flush(vma, address, pmd);
2571 spin_unlock(pmd_ptl);
2572 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2575 isolated = __collapse_huge_page_isolate(vma, address, pte);
2576 spin_unlock(pte_ptl);
2578 if (unlikely(!isolated)) {
2581 BUG_ON(!pmd_none(*pmd));
2583 * We can only use set_pmd_at when establishing
2584 * hugepmds and never for establishing regular pmds that
2585 * points to regular pagetables. Use pmd_populate for that
2587 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2588 spin_unlock(pmd_ptl);
2589 anon_vma_unlock_write(vma->anon_vma);
2594 * All pages are isolated and locked so anon_vma rmap
2595 * can't run anymore.
2597 anon_vma_unlock_write(vma->anon_vma);
2599 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2601 __SetPageUptodate(new_page);
2602 pgtable = pmd_pgtable(_pmd);
2604 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2605 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2608 * spin_lock() below is not the equivalent of smp_wmb(), so
2609 * this is needed to avoid the copy_huge_page writes to become
2610 * visible after the set_pmd_at() write.
2615 BUG_ON(!pmd_none(*pmd));
2616 page_add_new_anon_rmap(new_page, vma, address);
2617 mem_cgroup_commit_charge(new_page, memcg, false);
2618 lru_cache_add_active_or_unevictable(new_page, vma);
2619 pgtable_trans_huge_deposit(mm, pmd, pgtable);
2620 set_pmd_at(mm, address, pmd, _pmd);
2621 update_mmu_cache_pmd(vma, address, pmd);
2622 spin_unlock(pmd_ptl);
2626 khugepaged_pages_collapsed++;
2628 up_write(&mm->mmap_sem);
2632 mem_cgroup_cancel_charge(new_page, memcg);
2636 static int khugepaged_scan_pmd(struct mm_struct *mm,
2637 struct vm_area_struct *vma,
2638 unsigned long address,
2639 struct page **hpage)
2643 int ret = 0, none_or_zero = 0;
2645 unsigned long _address;
2647 int node = NUMA_NO_NODE;
2648 bool writable = false, referenced = false;
2650 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2652 pmd = mm_find_pmd(mm, address);
2656 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2657 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2658 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2659 _pte++, _address += PAGE_SIZE) {
2660 pte_t pteval = *_pte;
2661 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2662 if (!userfaultfd_armed(vma) &&
2663 ++none_or_zero <= khugepaged_max_ptes_none)
2668 if (!pte_present(pteval))
2670 if (pte_write(pteval))
2673 page = vm_normal_page(vma, _address, pteval);
2674 if (unlikely(!page))
2677 * Record which node the original page is from and save this
2678 * information to khugepaged_node_load[].
2679 * Khupaged will allocate hugepage from the node has the max
2682 node = page_to_nid(page);
2683 if (khugepaged_scan_abort(node))
2685 khugepaged_node_load[node]++;
2686 VM_BUG_ON_PAGE(PageCompound(page), page);
2687 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2690 * cannot use mapcount: can't collapse if there's a gup pin.
2691 * The page must only be referenced by the scanned process
2692 * and page swap cache.
2694 if (page_count(page) != 1 + !!PageSwapCache(page))
2696 if (pte_young(pteval) || PageReferenced(page) ||
2697 mmu_notifier_test_young(vma->vm_mm, address))
2700 if (referenced && writable)
2703 pte_unmap_unlock(pte, ptl);
2705 node = khugepaged_find_target_node();
2706 /* collapse_huge_page will return with the mmap_sem released */
2707 collapse_huge_page(mm, address, hpage, vma, node);
2713 static void collect_mm_slot(struct mm_slot *mm_slot)
2715 struct mm_struct *mm = mm_slot->mm;
2717 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2719 if (khugepaged_test_exit(mm)) {
2721 hash_del(&mm_slot->hash);
2722 list_del(&mm_slot->mm_node);
2725 * Not strictly needed because the mm exited already.
2727 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2730 /* khugepaged_mm_lock actually not necessary for the below */
2731 free_mm_slot(mm_slot);
2736 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2737 struct page **hpage)
2738 __releases(&khugepaged_mm_lock)
2739 __acquires(&khugepaged_mm_lock)
2741 struct mm_slot *mm_slot;
2742 struct mm_struct *mm;
2743 struct vm_area_struct *vma;
2747 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2749 if (khugepaged_scan.mm_slot)
2750 mm_slot = khugepaged_scan.mm_slot;
2752 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2753 struct mm_slot, mm_node);
2754 khugepaged_scan.address = 0;
2755 khugepaged_scan.mm_slot = mm_slot;
2757 spin_unlock(&khugepaged_mm_lock);
2760 down_read(&mm->mmap_sem);
2761 if (unlikely(khugepaged_test_exit(mm)))
2764 vma = find_vma(mm, khugepaged_scan.address);
2767 for (; vma; vma = vma->vm_next) {
2768 unsigned long hstart, hend;
2771 if (unlikely(khugepaged_test_exit(mm))) {
2775 if (!hugepage_vma_check(vma)) {
2780 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2781 hend = vma->vm_end & HPAGE_PMD_MASK;
2784 if (khugepaged_scan.address > hend)
2786 if (khugepaged_scan.address < hstart)
2787 khugepaged_scan.address = hstart;
2788 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2790 while (khugepaged_scan.address < hend) {
2793 if (unlikely(khugepaged_test_exit(mm)))
2794 goto breakouterloop;
2796 VM_BUG_ON(khugepaged_scan.address < hstart ||
2797 khugepaged_scan.address + HPAGE_PMD_SIZE >
2799 ret = khugepaged_scan_pmd(mm, vma,
2800 khugepaged_scan.address,
2802 /* move to next address */
2803 khugepaged_scan.address += HPAGE_PMD_SIZE;
2804 progress += HPAGE_PMD_NR;
2806 /* we released mmap_sem so break loop */
2807 goto breakouterloop_mmap_sem;
2808 if (progress >= pages)
2809 goto breakouterloop;
2813 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2814 breakouterloop_mmap_sem:
2816 spin_lock(&khugepaged_mm_lock);
2817 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2819 * Release the current mm_slot if this mm is about to die, or
2820 * if we scanned all vmas of this mm.
2822 if (khugepaged_test_exit(mm) || !vma) {
2824 * Make sure that if mm_users is reaching zero while
2825 * khugepaged runs here, khugepaged_exit will find
2826 * mm_slot not pointing to the exiting mm.
2828 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2829 khugepaged_scan.mm_slot = list_entry(
2830 mm_slot->mm_node.next,
2831 struct mm_slot, mm_node);
2832 khugepaged_scan.address = 0;
2834 khugepaged_scan.mm_slot = NULL;
2835 khugepaged_full_scans++;
2838 collect_mm_slot(mm_slot);
2844 static int khugepaged_has_work(void)
2846 return !list_empty(&khugepaged_scan.mm_head) &&
2847 khugepaged_enabled();
2850 static int khugepaged_wait_event(void)
2852 return !list_empty(&khugepaged_scan.mm_head) ||
2853 kthread_should_stop();
2856 static void khugepaged_do_scan(void)
2858 struct page *hpage = NULL;
2859 unsigned int progress = 0, pass_through_head = 0;
2860 unsigned int pages = khugepaged_pages_to_scan;
2863 barrier(); /* write khugepaged_pages_to_scan to local stack */
2865 while (progress < pages) {
2866 if (!khugepaged_prealloc_page(&hpage, &wait))
2871 if (unlikely(kthread_should_stop() || try_to_freeze()))
2874 spin_lock(&khugepaged_mm_lock);
2875 if (!khugepaged_scan.mm_slot)
2876 pass_through_head++;
2877 if (khugepaged_has_work() &&
2878 pass_through_head < 2)
2879 progress += khugepaged_scan_mm_slot(pages - progress,
2883 spin_unlock(&khugepaged_mm_lock);
2886 if (!IS_ERR_OR_NULL(hpage))
2890 static void khugepaged_wait_work(void)
2892 if (khugepaged_has_work()) {
2893 if (!khugepaged_scan_sleep_millisecs)
2896 wait_event_freezable_timeout(khugepaged_wait,
2897 kthread_should_stop(),
2898 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2902 if (khugepaged_enabled())
2903 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2906 static int khugepaged(void *none)
2908 struct mm_slot *mm_slot;
2911 set_user_nice(current, MAX_NICE);
2913 while (!kthread_should_stop()) {
2914 khugepaged_do_scan();
2915 khugepaged_wait_work();
2918 spin_lock(&khugepaged_mm_lock);
2919 mm_slot = khugepaged_scan.mm_slot;
2920 khugepaged_scan.mm_slot = NULL;
2922 collect_mm_slot(mm_slot);
2923 spin_unlock(&khugepaged_mm_lock);
2927 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2928 unsigned long haddr, pmd_t *pmd)
2930 struct mm_struct *mm = vma->vm_mm;
2935 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2936 /* leave pmd empty until pte is filled */
2938 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2939 pmd_populate(mm, &_pmd, pgtable);
2941 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2943 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2944 entry = pte_mkspecial(entry);
2945 pte = pte_offset_map(&_pmd, haddr);
2946 VM_BUG_ON(!pte_none(*pte));
2947 set_pte_at(mm, haddr, pte, entry);
2950 smp_wmb(); /* make pte visible before pmd */
2951 pmd_populate(mm, pmd, pgtable);
2952 put_huge_zero_page();
2955 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2959 struct page *page = NULL;
2960 struct mm_struct *mm = vma->vm_mm;
2961 unsigned long haddr = address & HPAGE_PMD_MASK;
2962 unsigned long mmun_start; /* For mmu_notifiers */
2963 unsigned long mmun_end; /* For mmu_notifiers */
2965 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2968 mmun_end = haddr + HPAGE_PMD_SIZE;
2970 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2971 ptl = pmd_lock(mm, pmd);
2972 if (unlikely(!pmd_trans_huge(*pmd)))
2974 if (vma_is_dax(vma)) {
2975 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2976 if (is_huge_zero_pmd(_pmd))
2977 put_huge_zero_page();
2978 } else if (is_huge_zero_pmd(*pmd)) {
2979 __split_huge_zero_page_pmd(vma, haddr, pmd);
2981 page = pmd_page(*pmd);
2982 VM_BUG_ON_PAGE(!page_count(page), page);
2987 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2992 split_huge_page(page);
2996 * We don't always have down_write of mmap_sem here: a racing
2997 * do_huge_pmd_wp_page() might have copied-on-write to another
2998 * huge page before our split_huge_page() got the anon_vma lock.
3000 if (unlikely(pmd_trans_huge(*pmd)))
3004 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
3007 struct vm_area_struct *vma;
3009 vma = find_vma(mm, address);
3010 BUG_ON(vma == NULL);
3011 split_huge_page_pmd(vma, address, pmd);
3014 static void split_huge_page_address(struct mm_struct *mm,
3015 unsigned long address)
3021 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
3023 pgd = pgd_offset(mm, address);
3024 if (!pgd_present(*pgd))
3027 pud = pud_offset(pgd, address);
3028 if (!pud_present(*pud))
3031 pmd = pmd_offset(pud, address);
3032 if (!pmd_present(*pmd))
3035 * Caller holds the mmap_sem write mode, so a huge pmd cannot
3036 * materialize from under us.
3038 split_huge_page_pmd_mm(mm, address, pmd);
3041 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3042 unsigned long start,
3047 * If the new start address isn't hpage aligned and it could
3048 * previously contain an hugepage: check if we need to split
3051 if (start & ~HPAGE_PMD_MASK &&
3052 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3053 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3054 split_huge_page_address(vma->vm_mm, start);
3057 * If the new end address isn't hpage aligned and it could
3058 * previously contain an hugepage: check if we need to split
3061 if (end & ~HPAGE_PMD_MASK &&
3062 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3063 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3064 split_huge_page_address(vma->vm_mm, end);
3067 * If we're also updating the vma->vm_next->vm_start, if the new
3068 * vm_next->vm_start isn't page aligned and it could previously
3069 * contain an hugepage: check if we need to split an huge pmd.
3071 if (adjust_next > 0) {
3072 struct vm_area_struct *next = vma->vm_next;
3073 unsigned long nstart = next->vm_start;
3074 nstart += adjust_next << PAGE_SHIFT;
3075 if (nstart & ~HPAGE_PMD_MASK &&
3076 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3077 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3078 split_huge_page_address(next->vm_mm, nstart);