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.
9 #include <linux/sched.h>
10 #include <linux/highmem.h>
11 #include <linux/hugetlb.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/rmap.h>
14 #include <linux/swap.h>
15 #include <linux/shrinker.h>
16 #include <linux/mm_inline.h>
17 #include <linux/kthread.h>
18 #include <linux/khugepaged.h>
19 #include <linux/freezer.h>
20 #include <linux/mman.h>
21 #include <linux/pagemap.h>
24 #include <asm/pgalloc.h>
28 * By default transparent hugepage support is enabled for all mappings
29 * and khugepaged scans all mappings. Defrag is only invoked by
30 * khugepaged hugepage allocations and by page faults inside
31 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
34 unsigned long transparent_hugepage_flags __read_mostly =
35 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
36 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
38 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
39 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
41 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
42 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
43 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
45 /* default scan 8*512 pte (or vmas) every 30 second */
46 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
47 static unsigned int khugepaged_pages_collapsed;
48 static unsigned int khugepaged_full_scans;
49 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
50 /* during fragmentation poll the hugepage allocator once every minute */
51 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
52 static struct task_struct *khugepaged_thread __read_mostly;
53 static DEFINE_MUTEX(khugepaged_mutex);
54 static DEFINE_SPINLOCK(khugepaged_mm_lock);
55 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
57 * default collapse hugepages if there is at least one pte mapped like
58 * it would have happened if the vma was large enough during page
61 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
63 static int khugepaged(void *none);
64 static int mm_slots_hash_init(void);
65 static int khugepaged_slab_init(void);
66 static void khugepaged_slab_free(void);
68 #define MM_SLOTS_HASH_HEADS 1024
69 static struct hlist_head *mm_slots_hash __read_mostly;
70 static struct kmem_cache *mm_slot_cache __read_mostly;
73 * struct mm_slot - hash lookup from mm to mm_slot
74 * @hash: hash collision list
75 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
76 * @mm: the mm that this information is valid for
79 struct hlist_node hash;
80 struct list_head mm_node;
85 * struct khugepaged_scan - cursor for scanning
86 * @mm_head: the head of the mm list to scan
87 * @mm_slot: the current mm_slot we are scanning
88 * @address: the next address inside that to be scanned
90 * There is only the one khugepaged_scan instance of this cursor structure.
92 struct khugepaged_scan {
93 struct list_head mm_head;
94 struct mm_slot *mm_slot;
95 unsigned long address;
97 static struct khugepaged_scan khugepaged_scan = {
98 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
102 static int set_recommended_min_free_kbytes(void)
106 unsigned long recommended_min;
107 extern int min_free_kbytes;
109 if (!khugepaged_enabled())
112 for_each_populated_zone(zone)
115 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
116 recommended_min = pageblock_nr_pages * nr_zones * 2;
119 * Make sure that on average at least two pageblocks are almost free
120 * of another type, one for a migratetype to fall back to and a
121 * second to avoid subsequent fallbacks of other types There are 3
122 * MIGRATE_TYPES we care about.
124 recommended_min += pageblock_nr_pages * nr_zones *
125 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
127 /* don't ever allow to reserve more than 5% of the lowmem */
128 recommended_min = min(recommended_min,
129 (unsigned long) nr_free_buffer_pages() / 20);
130 recommended_min <<= (PAGE_SHIFT-10);
132 if (recommended_min > min_free_kbytes)
133 min_free_kbytes = recommended_min;
134 setup_per_zone_wmarks();
137 late_initcall(set_recommended_min_free_kbytes);
139 static int start_khugepaged(void)
142 if (khugepaged_enabled()) {
143 if (!khugepaged_thread)
144 khugepaged_thread = kthread_run(khugepaged, NULL,
146 if (unlikely(IS_ERR(khugepaged_thread))) {
148 "khugepaged: kthread_run(khugepaged) failed\n");
149 err = PTR_ERR(khugepaged_thread);
150 khugepaged_thread = NULL;
153 if (!list_empty(&khugepaged_scan.mm_head))
154 wake_up_interruptible(&khugepaged_wait);
156 set_recommended_min_free_kbytes();
157 } else if (khugepaged_thread) {
158 kthread_stop(khugepaged_thread);
159 khugepaged_thread = NULL;
165 static atomic_t huge_zero_refcount;
166 static unsigned long huge_zero_pfn __read_mostly;
168 static inline bool is_huge_zero_pfn(unsigned long pfn)
170 unsigned long zero_pfn = ACCESS_ONCE(huge_zero_pfn);
171 return zero_pfn && pfn == zero_pfn;
174 static inline bool is_huge_zero_pmd(pmd_t pmd)
176 return is_huge_zero_pfn(pmd_pfn(pmd));
179 static unsigned long get_huge_zero_page(void)
181 struct page *zero_page;
183 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
184 return ACCESS_ONCE(huge_zero_pfn);
186 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
189 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
192 count_vm_event(THP_ZERO_PAGE_ALLOC);
194 if (cmpxchg(&huge_zero_pfn, 0, page_to_pfn(zero_page))) {
196 __free_page(zero_page);
200 /* We take additional reference here. It will be put back by shrinker */
201 atomic_set(&huge_zero_refcount, 2);
203 return ACCESS_ONCE(huge_zero_pfn);
206 static void put_huge_zero_page(void)
209 * Counter should never go to zero here. Only shrinker can put
212 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
215 static int shrink_huge_zero_page(struct shrinker *shrink,
216 struct shrink_control *sc)
219 /* we can free zero page only if last reference remains */
220 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
222 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
223 unsigned long zero_pfn = xchg(&huge_zero_pfn, 0);
224 BUG_ON(zero_pfn == 0);
225 __free_page(__pfn_to_page(zero_pfn));
231 static struct shrinker huge_zero_page_shrinker = {
232 .shrink = shrink_huge_zero_page,
233 .seeks = DEFAULT_SEEKS,
238 static ssize_t double_flag_show(struct kobject *kobj,
239 struct kobj_attribute *attr, char *buf,
240 enum transparent_hugepage_flag enabled,
241 enum transparent_hugepage_flag req_madv)
243 if (test_bit(enabled, &transparent_hugepage_flags)) {
244 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
245 return sprintf(buf, "[always] madvise never\n");
246 } else if (test_bit(req_madv, &transparent_hugepage_flags))
247 return sprintf(buf, "always [madvise] never\n");
249 return sprintf(buf, "always madvise [never]\n");
251 static ssize_t double_flag_store(struct kobject *kobj,
252 struct kobj_attribute *attr,
253 const char *buf, size_t count,
254 enum transparent_hugepage_flag enabled,
255 enum transparent_hugepage_flag req_madv)
257 if (!memcmp("always", buf,
258 min(sizeof("always")-1, count))) {
259 set_bit(enabled, &transparent_hugepage_flags);
260 clear_bit(req_madv, &transparent_hugepage_flags);
261 } else if (!memcmp("madvise", buf,
262 min(sizeof("madvise")-1, count))) {
263 clear_bit(enabled, &transparent_hugepage_flags);
264 set_bit(req_madv, &transparent_hugepage_flags);
265 } else if (!memcmp("never", buf,
266 min(sizeof("never")-1, count))) {
267 clear_bit(enabled, &transparent_hugepage_flags);
268 clear_bit(req_madv, &transparent_hugepage_flags);
275 static ssize_t enabled_show(struct kobject *kobj,
276 struct kobj_attribute *attr, char *buf)
278 return double_flag_show(kobj, attr, buf,
279 TRANSPARENT_HUGEPAGE_FLAG,
280 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
282 static ssize_t enabled_store(struct kobject *kobj,
283 struct kobj_attribute *attr,
284 const char *buf, size_t count)
288 ret = double_flag_store(kobj, attr, buf, count,
289 TRANSPARENT_HUGEPAGE_FLAG,
290 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
295 mutex_lock(&khugepaged_mutex);
296 err = start_khugepaged();
297 mutex_unlock(&khugepaged_mutex);
305 static struct kobj_attribute enabled_attr =
306 __ATTR(enabled, 0644, enabled_show, enabled_store);
308 static ssize_t single_flag_show(struct kobject *kobj,
309 struct kobj_attribute *attr, char *buf,
310 enum transparent_hugepage_flag flag)
312 return sprintf(buf, "%d\n",
313 !!test_bit(flag, &transparent_hugepage_flags));
316 static ssize_t single_flag_store(struct kobject *kobj,
317 struct kobj_attribute *attr,
318 const char *buf, size_t count,
319 enum transparent_hugepage_flag flag)
324 ret = kstrtoul(buf, 10, &value);
331 set_bit(flag, &transparent_hugepage_flags);
333 clear_bit(flag, &transparent_hugepage_flags);
339 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
340 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
341 * memory just to allocate one more hugepage.
343 static ssize_t defrag_show(struct kobject *kobj,
344 struct kobj_attribute *attr, char *buf)
346 return double_flag_show(kobj, attr, buf,
347 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
348 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
350 static ssize_t defrag_store(struct kobject *kobj,
351 struct kobj_attribute *attr,
352 const char *buf, size_t count)
354 return double_flag_store(kobj, attr, buf, count,
355 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
356 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
358 static struct kobj_attribute defrag_attr =
359 __ATTR(defrag, 0644, defrag_show, defrag_store);
361 static ssize_t use_zero_page_show(struct kobject *kobj,
362 struct kobj_attribute *attr, char *buf)
364 return single_flag_show(kobj, attr, buf,
365 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
367 static ssize_t use_zero_page_store(struct kobject *kobj,
368 struct kobj_attribute *attr, const char *buf, size_t count)
370 return single_flag_store(kobj, attr, buf, count,
371 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
373 static struct kobj_attribute use_zero_page_attr =
374 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
375 #ifdef CONFIG_DEBUG_VM
376 static ssize_t debug_cow_show(struct kobject *kobj,
377 struct kobj_attribute *attr, char *buf)
379 return single_flag_show(kobj, attr, buf,
380 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
382 static ssize_t debug_cow_store(struct kobject *kobj,
383 struct kobj_attribute *attr,
384 const char *buf, size_t count)
386 return single_flag_store(kobj, attr, buf, count,
387 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
389 static struct kobj_attribute debug_cow_attr =
390 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
391 #endif /* CONFIG_DEBUG_VM */
393 static struct attribute *hugepage_attr[] = {
396 &use_zero_page_attr.attr,
397 #ifdef CONFIG_DEBUG_VM
398 &debug_cow_attr.attr,
403 static struct attribute_group hugepage_attr_group = {
404 .attrs = hugepage_attr,
407 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
408 struct kobj_attribute *attr,
411 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
414 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
415 struct kobj_attribute *attr,
416 const char *buf, size_t count)
421 err = strict_strtoul(buf, 10, &msecs);
422 if (err || msecs > UINT_MAX)
425 khugepaged_scan_sleep_millisecs = msecs;
426 wake_up_interruptible(&khugepaged_wait);
430 static struct kobj_attribute scan_sleep_millisecs_attr =
431 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
432 scan_sleep_millisecs_store);
434 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
435 struct kobj_attribute *attr,
438 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
441 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
442 struct kobj_attribute *attr,
443 const char *buf, size_t count)
448 err = strict_strtoul(buf, 10, &msecs);
449 if (err || msecs > UINT_MAX)
452 khugepaged_alloc_sleep_millisecs = msecs;
453 wake_up_interruptible(&khugepaged_wait);
457 static struct kobj_attribute alloc_sleep_millisecs_attr =
458 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
459 alloc_sleep_millisecs_store);
461 static ssize_t pages_to_scan_show(struct kobject *kobj,
462 struct kobj_attribute *attr,
465 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
467 static ssize_t pages_to_scan_store(struct kobject *kobj,
468 struct kobj_attribute *attr,
469 const char *buf, size_t count)
474 err = strict_strtoul(buf, 10, &pages);
475 if (err || !pages || pages > UINT_MAX)
478 khugepaged_pages_to_scan = pages;
482 static struct kobj_attribute pages_to_scan_attr =
483 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
484 pages_to_scan_store);
486 static ssize_t pages_collapsed_show(struct kobject *kobj,
487 struct kobj_attribute *attr,
490 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
492 static struct kobj_attribute pages_collapsed_attr =
493 __ATTR_RO(pages_collapsed);
495 static ssize_t full_scans_show(struct kobject *kobj,
496 struct kobj_attribute *attr,
499 return sprintf(buf, "%u\n", khugepaged_full_scans);
501 static struct kobj_attribute full_scans_attr =
502 __ATTR_RO(full_scans);
504 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
505 struct kobj_attribute *attr, char *buf)
507 return single_flag_show(kobj, attr, buf,
508 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
510 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
511 struct kobj_attribute *attr,
512 const char *buf, size_t count)
514 return single_flag_store(kobj, attr, buf, count,
515 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
517 static struct kobj_attribute khugepaged_defrag_attr =
518 __ATTR(defrag, 0644, khugepaged_defrag_show,
519 khugepaged_defrag_store);
522 * max_ptes_none controls if khugepaged should collapse hugepages over
523 * any unmapped ptes in turn potentially increasing the memory
524 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
525 * reduce the available free memory in the system as it
526 * runs. Increasing max_ptes_none will instead potentially reduce the
527 * free memory in the system during the khugepaged scan.
529 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
530 struct kobj_attribute *attr,
533 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
535 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
536 struct kobj_attribute *attr,
537 const char *buf, size_t count)
540 unsigned long max_ptes_none;
542 err = strict_strtoul(buf, 10, &max_ptes_none);
543 if (err || max_ptes_none > HPAGE_PMD_NR-1)
546 khugepaged_max_ptes_none = max_ptes_none;
550 static struct kobj_attribute khugepaged_max_ptes_none_attr =
551 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
552 khugepaged_max_ptes_none_store);
554 static struct attribute *khugepaged_attr[] = {
555 &khugepaged_defrag_attr.attr,
556 &khugepaged_max_ptes_none_attr.attr,
557 &pages_to_scan_attr.attr,
558 &pages_collapsed_attr.attr,
559 &full_scans_attr.attr,
560 &scan_sleep_millisecs_attr.attr,
561 &alloc_sleep_millisecs_attr.attr,
565 static struct attribute_group khugepaged_attr_group = {
566 .attrs = khugepaged_attr,
567 .name = "khugepaged",
570 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
574 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
575 if (unlikely(!*hugepage_kobj)) {
576 printk(KERN_ERR "hugepage: failed kobject create\n");
580 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
582 printk(KERN_ERR "hugepage: failed register hugeage group\n");
586 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
588 printk(KERN_ERR "hugepage: failed register hugeage group\n");
589 goto remove_hp_group;
595 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
597 kobject_put(*hugepage_kobj);
601 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
603 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
604 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
605 kobject_put(hugepage_kobj);
608 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
613 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
616 #endif /* CONFIG_SYSFS */
618 static int __init hugepage_init(void)
621 struct kobject *hugepage_kobj;
623 if (!has_transparent_hugepage()) {
624 transparent_hugepage_flags = 0;
628 err = hugepage_init_sysfs(&hugepage_kobj);
632 err = khugepaged_slab_init();
636 err = mm_slots_hash_init();
638 khugepaged_slab_free();
642 register_shrinker(&huge_zero_page_shrinker);
645 * By default disable transparent hugepages on smaller systems,
646 * where the extra memory used could hurt more than TLB overhead
647 * is likely to save. The admin can still enable it through /sys.
649 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
650 transparent_hugepage_flags = 0;
656 hugepage_exit_sysfs(hugepage_kobj);
659 module_init(hugepage_init)
661 static int __init setup_transparent_hugepage(char *str)
666 if (!strcmp(str, "always")) {
667 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
668 &transparent_hugepage_flags);
669 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
670 &transparent_hugepage_flags);
672 } else if (!strcmp(str, "madvise")) {
673 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
674 &transparent_hugepage_flags);
675 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
676 &transparent_hugepage_flags);
678 } else if (!strcmp(str, "never")) {
679 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
680 &transparent_hugepage_flags);
681 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
682 &transparent_hugepage_flags);
688 "transparent_hugepage= cannot parse, ignored\n");
691 __setup("transparent_hugepage=", setup_transparent_hugepage);
693 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
695 if (likely(vma->vm_flags & VM_WRITE))
696 pmd = pmd_mkwrite(pmd);
700 static inline pmd_t mk_huge_pmd(struct page *page, struct vm_area_struct *vma)
703 entry = mk_pmd(page, vma->vm_page_prot);
704 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
705 entry = pmd_mkhuge(entry);
709 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
710 struct vm_area_struct *vma,
711 unsigned long haddr, pmd_t *pmd,
716 VM_BUG_ON(!PageCompound(page));
717 pgtable = pte_alloc_one(mm, haddr);
718 if (unlikely(!pgtable))
721 clear_huge_page(page, haddr, HPAGE_PMD_NR);
722 __SetPageUptodate(page);
724 spin_lock(&mm->page_table_lock);
725 if (unlikely(!pmd_none(*pmd))) {
726 spin_unlock(&mm->page_table_lock);
727 mem_cgroup_uncharge_page(page);
729 pte_free(mm, pgtable);
732 entry = mk_huge_pmd(page, vma);
734 * The spinlocking to take the lru_lock inside
735 * page_add_new_anon_rmap() acts as a full memory
736 * barrier to be sure clear_huge_page writes become
737 * visible after the set_pmd_at() write.
739 page_add_new_anon_rmap(page, vma, haddr);
740 set_pmd_at(mm, haddr, pmd, entry);
741 pgtable_trans_huge_deposit(mm, pgtable);
742 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
744 spin_unlock(&mm->page_table_lock);
750 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
752 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
755 static inline struct page *alloc_hugepage_vma(int defrag,
756 struct vm_area_struct *vma,
757 unsigned long haddr, int nd,
760 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
761 HPAGE_PMD_ORDER, vma, haddr, nd);
765 static inline struct page *alloc_hugepage(int defrag)
767 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
772 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
773 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
774 unsigned long zero_pfn)
777 entry = pfn_pmd(zero_pfn, vma->vm_page_prot);
778 entry = pmd_wrprotect(entry);
779 entry = pmd_mkhuge(entry);
780 set_pmd_at(mm, haddr, pmd, entry);
781 pgtable_trans_huge_deposit(mm, pgtable);
785 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
786 unsigned long address, pmd_t *pmd,
790 unsigned long haddr = address & HPAGE_PMD_MASK;
793 if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
794 if (unlikely(anon_vma_prepare(vma)))
796 if (unlikely(khugepaged_enter(vma)))
798 if (!(flags & FAULT_FLAG_WRITE) &&
799 transparent_hugepage_use_zero_page()) {
801 unsigned long zero_pfn;
802 pgtable = pte_alloc_one(mm, haddr);
803 if (unlikely(!pgtable))
805 zero_pfn = get_huge_zero_page();
806 if (unlikely(!zero_pfn)) {
807 pte_free(mm, pgtable);
808 count_vm_event(THP_FAULT_FALLBACK);
811 spin_lock(&mm->page_table_lock);
812 set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
814 spin_unlock(&mm->page_table_lock);
817 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
818 vma, haddr, numa_node_id(), 0);
819 if (unlikely(!page)) {
820 count_vm_event(THP_FAULT_FALLBACK);
823 count_vm_event(THP_FAULT_ALLOC);
824 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
828 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
830 mem_cgroup_uncharge_page(page);
839 * Use __pte_alloc instead of pte_alloc_map, because we can't
840 * run pte_offset_map on the pmd, if an huge pmd could
841 * materialize from under us from a different thread.
843 if (unlikely(__pte_alloc(mm, vma, pmd, address)))
845 /* if an huge pmd materialized from under us just retry later */
846 if (unlikely(pmd_trans_huge(*pmd)))
849 * A regular pmd is established and it can't morph into a huge pmd
850 * from under us anymore at this point because we hold the mmap_sem
851 * read mode and khugepaged takes it in write mode. So now it's
852 * safe to run pte_offset_map().
854 pte = pte_offset_map(pmd, address);
855 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
858 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
859 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
860 struct vm_area_struct *vma)
862 struct page *src_page;
868 pgtable = pte_alloc_one(dst_mm, addr);
869 if (unlikely(!pgtable))
872 spin_lock(&dst_mm->page_table_lock);
873 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
877 if (unlikely(!pmd_trans_huge(pmd))) {
878 pte_free(dst_mm, pgtable);
882 * mm->page_table_lock is enough to be sure that huge zero pmd is not
883 * under splitting since we don't split the page itself, only pmd to
886 if (is_huge_zero_pmd(pmd)) {
887 unsigned long zero_pfn;
889 * get_huge_zero_page() will never allocate a new page here,
890 * since we already have a zero page to copy. It just takes a
893 zero_pfn = get_huge_zero_page();
894 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
899 if (unlikely(pmd_trans_splitting(pmd))) {
900 /* split huge page running from under us */
901 spin_unlock(&src_mm->page_table_lock);
902 spin_unlock(&dst_mm->page_table_lock);
903 pte_free(dst_mm, pgtable);
905 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
908 src_page = pmd_page(pmd);
909 VM_BUG_ON(!PageHead(src_page));
911 page_dup_rmap(src_page);
912 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
914 pmdp_set_wrprotect(src_mm, addr, src_pmd);
915 pmd = pmd_mkold(pmd_wrprotect(pmd));
916 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
917 pgtable_trans_huge_deposit(dst_mm, pgtable);
922 spin_unlock(&src_mm->page_table_lock);
923 spin_unlock(&dst_mm->page_table_lock);
928 void huge_pmd_set_accessed(struct mm_struct *mm,
929 struct vm_area_struct *vma,
930 unsigned long address,
931 pmd_t *pmd, pmd_t orig_pmd,
937 spin_lock(&mm->page_table_lock);
938 if (unlikely(!pmd_same(*pmd, orig_pmd)))
941 entry = pmd_mkyoung(orig_pmd);
942 haddr = address & HPAGE_PMD_MASK;
943 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
944 update_mmu_cache_pmd(vma, address, pmd);
947 spin_unlock(&mm->page_table_lock);
950 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
951 struct vm_area_struct *vma, unsigned long address,
952 pmd_t *pmd, unsigned long haddr)
958 unsigned long mmun_start; /* For mmu_notifiers */
959 unsigned long mmun_end; /* For mmu_notifiers */
961 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
967 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
973 clear_user_highpage(page, address);
974 __SetPageUptodate(page);
977 mmun_end = haddr + HPAGE_PMD_SIZE;
978 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
980 spin_lock(&mm->page_table_lock);
981 pmdp_clear_flush(vma, haddr, pmd);
982 /* leave pmd empty until pte is filled */
984 pgtable = pgtable_trans_huge_withdraw(mm);
985 pmd_populate(mm, &_pmd, pgtable);
987 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
989 if (haddr == (address & PAGE_MASK)) {
990 entry = mk_pte(page, vma->vm_page_prot);
991 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
992 page_add_new_anon_rmap(page, vma, haddr);
994 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
995 entry = pte_mkspecial(entry);
997 pte = pte_offset_map(&_pmd, haddr);
998 VM_BUG_ON(!pte_none(*pte));
999 set_pte_at(mm, haddr, pte, entry);
1002 smp_wmb(); /* make pte visible before pmd */
1003 pmd_populate(mm, pmd, pgtable);
1004 spin_unlock(&mm->page_table_lock);
1005 put_huge_zero_page();
1006 inc_mm_counter(mm, MM_ANONPAGES);
1008 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1010 ret |= VM_FAULT_WRITE;
1015 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1016 struct vm_area_struct *vma,
1017 unsigned long address,
1018 pmd_t *pmd, pmd_t orig_pmd,
1020 unsigned long haddr)
1025 struct page **pages;
1026 unsigned long mmun_start; /* For mmu_notifiers */
1027 unsigned long mmun_end; /* For mmu_notifiers */
1029 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1031 if (unlikely(!pages)) {
1032 ret |= VM_FAULT_OOM;
1036 for (i = 0; i < HPAGE_PMD_NR; i++) {
1037 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1039 vma, address, page_to_nid(page));
1040 if (unlikely(!pages[i] ||
1041 mem_cgroup_newpage_charge(pages[i], mm,
1045 mem_cgroup_uncharge_start();
1047 mem_cgroup_uncharge_page(pages[i]);
1050 mem_cgroup_uncharge_end();
1052 ret |= VM_FAULT_OOM;
1057 for (i = 0; i < HPAGE_PMD_NR; i++) {
1058 copy_user_highpage(pages[i], page + i,
1059 haddr + PAGE_SIZE * i, vma);
1060 __SetPageUptodate(pages[i]);
1065 mmun_end = haddr + HPAGE_PMD_SIZE;
1066 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1068 spin_lock(&mm->page_table_lock);
1069 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1070 goto out_free_pages;
1071 VM_BUG_ON(!PageHead(page));
1073 pmdp_clear_flush(vma, haddr, pmd);
1074 /* leave pmd empty until pte is filled */
1076 pgtable = pgtable_trans_huge_withdraw(mm);
1077 pmd_populate(mm, &_pmd, pgtable);
1079 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1081 entry = mk_pte(pages[i], vma->vm_page_prot);
1082 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1083 page_add_new_anon_rmap(pages[i], vma, haddr);
1084 pte = pte_offset_map(&_pmd, haddr);
1085 VM_BUG_ON(!pte_none(*pte));
1086 set_pte_at(mm, haddr, pte, entry);
1091 smp_wmb(); /* make pte visible before pmd */
1092 pmd_populate(mm, pmd, pgtable);
1093 page_remove_rmap(page);
1094 spin_unlock(&mm->page_table_lock);
1096 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1098 ret |= VM_FAULT_WRITE;
1105 spin_unlock(&mm->page_table_lock);
1106 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1107 mem_cgroup_uncharge_start();
1108 for (i = 0; i < HPAGE_PMD_NR; i++) {
1109 mem_cgroup_uncharge_page(pages[i]);
1112 mem_cgroup_uncharge_end();
1117 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1118 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1121 struct page *page = NULL, *new_page;
1122 unsigned long haddr;
1123 unsigned long mmun_start; /* For mmu_notifiers */
1124 unsigned long mmun_end; /* For mmu_notifiers */
1126 VM_BUG_ON(!vma->anon_vma);
1127 haddr = address & HPAGE_PMD_MASK;
1128 if (is_huge_zero_pmd(orig_pmd))
1130 spin_lock(&mm->page_table_lock);
1131 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1134 page = pmd_page(orig_pmd);
1135 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1136 if (page_mapcount(page) == 1) {
1138 entry = pmd_mkyoung(orig_pmd);
1139 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1140 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
1141 update_mmu_cache_pmd(vma, address, pmd);
1142 ret |= VM_FAULT_WRITE;
1146 spin_unlock(&mm->page_table_lock);
1148 if (transparent_hugepage_enabled(vma) &&
1149 !transparent_hugepage_debug_cow())
1150 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1151 vma, haddr, numa_node_id(), 0);
1155 if (unlikely(!new_page)) {
1156 count_vm_event(THP_FAULT_FALLBACK);
1157 if (is_huge_zero_pmd(orig_pmd)) {
1158 ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1159 address, pmd, haddr);
1161 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1162 pmd, orig_pmd, page, haddr);
1163 if (ret & VM_FAULT_OOM)
1164 split_huge_page(page);
1169 count_vm_event(THP_FAULT_ALLOC);
1171 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1174 split_huge_page(page);
1177 ret |= VM_FAULT_OOM;
1181 if (is_huge_zero_pmd(orig_pmd))
1182 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1184 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1185 __SetPageUptodate(new_page);
1188 mmun_end = haddr + HPAGE_PMD_SIZE;
1189 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1191 spin_lock(&mm->page_table_lock);
1194 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1195 spin_unlock(&mm->page_table_lock);
1196 mem_cgroup_uncharge_page(new_page);
1201 entry = mk_huge_pmd(new_page, vma);
1202 pmdp_clear_flush(vma, haddr, pmd);
1203 page_add_new_anon_rmap(new_page, vma, haddr);
1204 set_pmd_at(mm, haddr, pmd, entry);
1205 update_mmu_cache_pmd(vma, address, pmd);
1206 if (is_huge_zero_pmd(orig_pmd)) {
1207 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1208 put_huge_zero_page();
1210 VM_BUG_ON(!PageHead(page));
1211 page_remove_rmap(page);
1214 ret |= VM_FAULT_WRITE;
1216 spin_unlock(&mm->page_table_lock);
1218 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1222 spin_unlock(&mm->page_table_lock);
1226 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1231 struct mm_struct *mm = vma->vm_mm;
1232 struct page *page = NULL;
1234 assert_spin_locked(&mm->page_table_lock);
1236 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1239 page = pmd_page(*pmd);
1240 VM_BUG_ON(!PageHead(page));
1241 if (flags & FOLL_TOUCH) {
1244 * We should set the dirty bit only for FOLL_WRITE but
1245 * for now the dirty bit in the pmd is meaningless.
1246 * And if the dirty bit will become meaningful and
1247 * we'll only set it with FOLL_WRITE, an atomic
1248 * set_bit will be required on the pmd to set the
1249 * young bit, instead of the current set_pmd_at.
1251 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1252 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1254 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1255 if (page->mapping && trylock_page(page)) {
1258 mlock_vma_page(page);
1262 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1263 VM_BUG_ON(!PageCompound(page));
1264 if (flags & FOLL_GET)
1265 get_page_foll(page);
1271 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1272 pmd_t *pmd, unsigned long addr)
1276 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1280 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1281 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1282 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1283 if (is_huge_zero_pmd(orig_pmd)) {
1285 spin_unlock(&tlb->mm->page_table_lock);
1286 put_huge_zero_page();
1288 page = pmd_page(orig_pmd);
1289 page_remove_rmap(page);
1290 VM_BUG_ON(page_mapcount(page) < 0);
1291 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1292 VM_BUG_ON(!PageHead(page));
1294 spin_unlock(&tlb->mm->page_table_lock);
1295 tlb_remove_page(tlb, page);
1297 pte_free(tlb->mm, pgtable);
1303 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1304 unsigned long addr, unsigned long end,
1309 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1311 * All logical pages in the range are present
1312 * if backed by a huge page.
1314 spin_unlock(&vma->vm_mm->page_table_lock);
1315 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1322 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1323 unsigned long old_addr,
1324 unsigned long new_addr, unsigned long old_end,
1325 pmd_t *old_pmd, pmd_t *new_pmd)
1330 struct mm_struct *mm = vma->vm_mm;
1332 if ((old_addr & ~HPAGE_PMD_MASK) ||
1333 (new_addr & ~HPAGE_PMD_MASK) ||
1334 old_end - old_addr < HPAGE_PMD_SIZE ||
1335 (new_vma->vm_flags & VM_NOHUGEPAGE))
1339 * The destination pmd shouldn't be established, free_pgtables()
1340 * should have release it.
1342 if (WARN_ON(!pmd_none(*new_pmd))) {
1343 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1347 ret = __pmd_trans_huge_lock(old_pmd, vma);
1349 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1350 VM_BUG_ON(!pmd_none(*new_pmd));
1351 set_pmd_at(mm, new_addr, new_pmd, pmd);
1352 spin_unlock(&mm->page_table_lock);
1358 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1359 unsigned long addr, pgprot_t newprot)
1361 struct mm_struct *mm = vma->vm_mm;
1364 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1366 entry = pmdp_get_and_clear(mm, addr, pmd);
1367 entry = pmd_modify(entry, newprot);
1368 BUG_ON(pmd_write(entry));
1369 set_pmd_at(mm, addr, pmd, entry);
1370 spin_unlock(&vma->vm_mm->page_table_lock);
1378 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1379 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1381 * Note that if it returns 1, this routine returns without unlocking page
1382 * table locks. So callers must unlock them.
1384 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1386 spin_lock(&vma->vm_mm->page_table_lock);
1387 if (likely(pmd_trans_huge(*pmd))) {
1388 if (unlikely(pmd_trans_splitting(*pmd))) {
1389 spin_unlock(&vma->vm_mm->page_table_lock);
1390 wait_split_huge_page(vma->anon_vma, pmd);
1393 /* Thp mapped by 'pmd' is stable, so we can
1394 * handle it as it is. */
1398 spin_unlock(&vma->vm_mm->page_table_lock);
1402 pmd_t *page_check_address_pmd(struct page *page,
1403 struct mm_struct *mm,
1404 unsigned long address,
1405 enum page_check_address_pmd_flag flag)
1407 pmd_t *pmd, *ret = NULL;
1409 if (address & ~HPAGE_PMD_MASK)
1412 pmd = mm_find_pmd(mm, address);
1417 if (pmd_page(*pmd) != page)
1420 * split_vma() may create temporary aliased mappings. There is
1421 * no risk as long as all huge pmd are found and have their
1422 * splitting bit set before __split_huge_page_refcount
1423 * runs. Finding the same huge pmd more than once during the
1424 * same rmap walk is not a problem.
1426 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1427 pmd_trans_splitting(*pmd))
1429 if (pmd_trans_huge(*pmd)) {
1430 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1431 !pmd_trans_splitting(*pmd));
1438 static int __split_huge_page_splitting(struct page *page,
1439 struct vm_area_struct *vma,
1440 unsigned long address)
1442 struct mm_struct *mm = vma->vm_mm;
1445 /* For mmu_notifiers */
1446 const unsigned long mmun_start = address;
1447 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
1449 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1450 spin_lock(&mm->page_table_lock);
1451 pmd = page_check_address_pmd(page, mm, address,
1452 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1455 * We can't temporarily set the pmd to null in order
1456 * to split it, the pmd must remain marked huge at all
1457 * times or the VM won't take the pmd_trans_huge paths
1458 * and it won't wait on the anon_vma->root->mutex to
1459 * serialize against split_huge_page*.
1461 pmdp_splitting_flush(vma, address, pmd);
1464 spin_unlock(&mm->page_table_lock);
1465 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1470 static void __split_huge_page_refcount(struct page *page)
1473 struct zone *zone = page_zone(page);
1474 struct lruvec *lruvec;
1477 /* prevent PageLRU to go away from under us, and freeze lru stats */
1478 spin_lock_irq(&zone->lru_lock);
1479 lruvec = mem_cgroup_page_lruvec(page, zone);
1481 compound_lock(page);
1482 /* complete memcg works before add pages to LRU */
1483 mem_cgroup_split_huge_fixup(page);
1485 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1486 struct page *page_tail = page + i;
1488 /* tail_page->_mapcount cannot change */
1489 BUG_ON(page_mapcount(page_tail) < 0);
1490 tail_count += page_mapcount(page_tail);
1491 /* check for overflow */
1492 BUG_ON(tail_count < 0);
1493 BUG_ON(atomic_read(&page_tail->_count) != 0);
1495 * tail_page->_count is zero and not changing from
1496 * under us. But get_page_unless_zero() may be running
1497 * from under us on the tail_page. If we used
1498 * atomic_set() below instead of atomic_add(), we
1499 * would then run atomic_set() concurrently with
1500 * get_page_unless_zero(), and atomic_set() is
1501 * implemented in C not using locked ops. spin_unlock
1502 * on x86 sometime uses locked ops because of PPro
1503 * errata 66, 92, so unless somebody can guarantee
1504 * atomic_set() here would be safe on all archs (and
1505 * not only on x86), it's safer to use atomic_add().
1507 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1508 &page_tail->_count);
1510 /* after clearing PageTail the gup refcount can be released */
1514 * retain hwpoison flag of the poisoned tail page:
1515 * fix for the unsuitable process killed on Guest Machine(KVM)
1516 * by the memory-failure.
1518 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1519 page_tail->flags |= (page->flags &
1520 ((1L << PG_referenced) |
1521 (1L << PG_swapbacked) |
1522 (1L << PG_mlocked) |
1523 (1L << PG_uptodate)));
1524 page_tail->flags |= (1L << PG_dirty);
1526 /* clear PageTail before overwriting first_page */
1530 * __split_huge_page_splitting() already set the
1531 * splitting bit in all pmd that could map this
1532 * hugepage, that will ensure no CPU can alter the
1533 * mapcount on the head page. The mapcount is only
1534 * accounted in the head page and it has to be
1535 * transferred to all tail pages in the below code. So
1536 * for this code to be safe, the split the mapcount
1537 * can't change. But that doesn't mean userland can't
1538 * keep changing and reading the page contents while
1539 * we transfer the mapcount, so the pmd splitting
1540 * status is achieved setting a reserved bit in the
1541 * pmd, not by clearing the present bit.
1543 page_tail->_mapcount = page->_mapcount;
1545 BUG_ON(page_tail->mapping);
1546 page_tail->mapping = page->mapping;
1548 page_tail->index = page->index + i;
1550 BUG_ON(!PageAnon(page_tail));
1551 BUG_ON(!PageUptodate(page_tail));
1552 BUG_ON(!PageDirty(page_tail));
1553 BUG_ON(!PageSwapBacked(page_tail));
1555 lru_add_page_tail(page, page_tail, lruvec);
1557 atomic_sub(tail_count, &page->_count);
1558 BUG_ON(atomic_read(&page->_count) <= 0);
1560 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1561 __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1563 ClearPageCompound(page);
1564 compound_unlock(page);
1565 spin_unlock_irq(&zone->lru_lock);
1567 for (i = 1; i < HPAGE_PMD_NR; i++) {
1568 struct page *page_tail = page + i;
1569 BUG_ON(page_count(page_tail) <= 0);
1571 * Tail pages may be freed if there wasn't any mapping
1572 * like if add_to_swap() is running on a lru page that
1573 * had its mapping zapped. And freeing these pages
1574 * requires taking the lru_lock so we do the put_page
1575 * of the tail pages after the split is complete.
1577 put_page(page_tail);
1581 * Only the head page (now become a regular page) is required
1582 * to be pinned by the caller.
1584 BUG_ON(page_count(page) <= 0);
1587 static int __split_huge_page_map(struct page *page,
1588 struct vm_area_struct *vma,
1589 unsigned long address)
1591 struct mm_struct *mm = vma->vm_mm;
1595 unsigned long haddr;
1597 spin_lock(&mm->page_table_lock);
1598 pmd = page_check_address_pmd(page, mm, address,
1599 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1601 pgtable = pgtable_trans_huge_withdraw(mm);
1602 pmd_populate(mm, &_pmd, pgtable);
1605 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1607 BUG_ON(PageCompound(page+i));
1608 entry = mk_pte(page + i, vma->vm_page_prot);
1609 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1610 if (!pmd_write(*pmd))
1611 entry = pte_wrprotect(entry);
1613 BUG_ON(page_mapcount(page) != 1);
1614 if (!pmd_young(*pmd))
1615 entry = pte_mkold(entry);
1616 pte = pte_offset_map(&_pmd, haddr);
1617 BUG_ON(!pte_none(*pte));
1618 set_pte_at(mm, haddr, pte, entry);
1622 smp_wmb(); /* make pte visible before pmd */
1624 * Up to this point the pmd is present and huge and
1625 * userland has the whole access to the hugepage
1626 * during the split (which happens in place). If we
1627 * overwrite the pmd with the not-huge version
1628 * pointing to the pte here (which of course we could
1629 * if all CPUs were bug free), userland could trigger
1630 * a small page size TLB miss on the small sized TLB
1631 * while the hugepage TLB entry is still established
1632 * in the huge TLB. Some CPU doesn't like that. See
1633 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1634 * Erratum 383 on page 93. Intel should be safe but is
1635 * also warns that it's only safe if the permission
1636 * and cache attributes of the two entries loaded in
1637 * the two TLB is identical (which should be the case
1638 * here). But it is generally safer to never allow
1639 * small and huge TLB entries for the same virtual
1640 * address to be loaded simultaneously. So instead of
1641 * doing "pmd_populate(); flush_tlb_range();" we first
1642 * mark the current pmd notpresent (atomically because
1643 * here the pmd_trans_huge and pmd_trans_splitting
1644 * must remain set at all times on the pmd until the
1645 * split is complete for this pmd), then we flush the
1646 * SMP TLB and finally we write the non-huge version
1647 * of the pmd entry with pmd_populate.
1649 pmdp_invalidate(vma, address, pmd);
1650 pmd_populate(mm, pmd, pgtable);
1653 spin_unlock(&mm->page_table_lock);
1658 /* must be called with anon_vma->root->mutex hold */
1659 static void __split_huge_page(struct page *page,
1660 struct anon_vma *anon_vma)
1662 int mapcount, mapcount2;
1663 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1664 struct anon_vma_chain *avc;
1666 BUG_ON(!PageHead(page));
1667 BUG_ON(PageTail(page));
1670 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1671 struct vm_area_struct *vma = avc->vma;
1672 unsigned long addr = vma_address(page, vma);
1673 BUG_ON(is_vma_temporary_stack(vma));
1674 mapcount += __split_huge_page_splitting(page, vma, addr);
1677 * It is critical that new vmas are added to the tail of the
1678 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1679 * and establishes a child pmd before
1680 * __split_huge_page_splitting() freezes the parent pmd (so if
1681 * we fail to prevent copy_huge_pmd() from running until the
1682 * whole __split_huge_page() is complete), we will still see
1683 * the newly established pmd of the child later during the
1684 * walk, to be able to set it as pmd_trans_splitting too.
1686 if (mapcount != page_mapcount(page))
1687 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1688 mapcount, page_mapcount(page));
1689 BUG_ON(mapcount != page_mapcount(page));
1691 __split_huge_page_refcount(page);
1694 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1695 struct vm_area_struct *vma = avc->vma;
1696 unsigned long addr = vma_address(page, vma);
1697 BUG_ON(is_vma_temporary_stack(vma));
1698 mapcount2 += __split_huge_page_map(page, vma, addr);
1700 if (mapcount != mapcount2)
1701 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1702 mapcount, mapcount2, page_mapcount(page));
1703 BUG_ON(mapcount != mapcount2);
1706 int split_huge_page(struct page *page)
1708 struct anon_vma *anon_vma;
1711 BUG_ON(is_huge_zero_pfn(page_to_pfn(page)));
1712 BUG_ON(!PageAnon(page));
1713 anon_vma = page_lock_anon_vma(page);
1717 if (!PageCompound(page))
1720 BUG_ON(!PageSwapBacked(page));
1721 __split_huge_page(page, anon_vma);
1722 count_vm_event(THP_SPLIT);
1724 BUG_ON(PageCompound(page));
1726 page_unlock_anon_vma(anon_vma);
1731 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1733 int hugepage_madvise(struct vm_area_struct *vma,
1734 unsigned long *vm_flags, int advice)
1736 struct mm_struct *mm = vma->vm_mm;
1741 * Be somewhat over-protective like KSM for now!
1743 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1745 if (mm->def_flags & VM_NOHUGEPAGE)
1747 *vm_flags &= ~VM_NOHUGEPAGE;
1748 *vm_flags |= VM_HUGEPAGE;
1750 * If the vma become good for khugepaged to scan,
1751 * register it here without waiting a page fault that
1752 * may not happen any time soon.
1754 if (unlikely(khugepaged_enter_vma_merge(vma)))
1757 case MADV_NOHUGEPAGE:
1759 * Be somewhat over-protective like KSM for now!
1761 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1763 *vm_flags &= ~VM_HUGEPAGE;
1764 *vm_flags |= VM_NOHUGEPAGE;
1766 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1767 * this vma even if we leave the mm registered in khugepaged if
1768 * it got registered before VM_NOHUGEPAGE was set.
1776 static int __init khugepaged_slab_init(void)
1778 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1779 sizeof(struct mm_slot),
1780 __alignof__(struct mm_slot), 0, NULL);
1787 static void __init khugepaged_slab_free(void)
1789 kmem_cache_destroy(mm_slot_cache);
1790 mm_slot_cache = NULL;
1793 static inline struct mm_slot *alloc_mm_slot(void)
1795 if (!mm_slot_cache) /* initialization failed */
1797 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1800 static inline void free_mm_slot(struct mm_slot *mm_slot)
1802 kmem_cache_free(mm_slot_cache, mm_slot);
1805 static int __init mm_slots_hash_init(void)
1807 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1815 static void __init mm_slots_hash_free(void)
1817 kfree(mm_slots_hash);
1818 mm_slots_hash = NULL;
1822 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1824 struct mm_slot *mm_slot;
1825 struct hlist_head *bucket;
1826 struct hlist_node *node;
1828 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1829 % MM_SLOTS_HASH_HEADS];
1830 hlist_for_each_entry(mm_slot, node, bucket, hash) {
1831 if (mm == mm_slot->mm)
1837 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1838 struct mm_slot *mm_slot)
1840 struct hlist_head *bucket;
1842 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1843 % MM_SLOTS_HASH_HEADS];
1845 hlist_add_head(&mm_slot->hash, bucket);
1848 static inline int khugepaged_test_exit(struct mm_struct *mm)
1850 return atomic_read(&mm->mm_users) == 0;
1853 int __khugepaged_enter(struct mm_struct *mm)
1855 struct mm_slot *mm_slot;
1858 mm_slot = alloc_mm_slot();
1862 /* __khugepaged_exit() must not run from under us */
1863 VM_BUG_ON(khugepaged_test_exit(mm));
1864 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1865 free_mm_slot(mm_slot);
1869 spin_lock(&khugepaged_mm_lock);
1870 insert_to_mm_slots_hash(mm, mm_slot);
1872 * Insert just behind the scanning cursor, to let the area settle
1875 wakeup = list_empty(&khugepaged_scan.mm_head);
1876 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1877 spin_unlock(&khugepaged_mm_lock);
1879 atomic_inc(&mm->mm_count);
1881 wake_up_interruptible(&khugepaged_wait);
1886 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1888 unsigned long hstart, hend;
1891 * Not yet faulted in so we will register later in the
1892 * page fault if needed.
1896 /* khugepaged not yet working on file or special mappings */
1898 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1899 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1900 hend = vma->vm_end & HPAGE_PMD_MASK;
1902 return khugepaged_enter(vma);
1906 void __khugepaged_exit(struct mm_struct *mm)
1908 struct mm_slot *mm_slot;
1911 spin_lock(&khugepaged_mm_lock);
1912 mm_slot = get_mm_slot(mm);
1913 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1914 hlist_del(&mm_slot->hash);
1915 list_del(&mm_slot->mm_node);
1918 spin_unlock(&khugepaged_mm_lock);
1921 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1922 free_mm_slot(mm_slot);
1924 } else if (mm_slot) {
1926 * This is required to serialize against
1927 * khugepaged_test_exit() (which is guaranteed to run
1928 * under mmap sem read mode). Stop here (after we
1929 * return all pagetables will be destroyed) until
1930 * khugepaged has finished working on the pagetables
1931 * under the mmap_sem.
1933 down_write(&mm->mmap_sem);
1934 up_write(&mm->mmap_sem);
1938 static void release_pte_page(struct page *page)
1940 /* 0 stands for page_is_file_cache(page) == false */
1941 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1943 putback_lru_page(page);
1946 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1948 while (--_pte >= pte) {
1949 pte_t pteval = *_pte;
1950 if (!pte_none(pteval))
1951 release_pte_page(pte_page(pteval));
1955 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1956 unsigned long address,
1961 int referenced = 0, none = 0;
1962 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1963 _pte++, address += PAGE_SIZE) {
1964 pte_t pteval = *_pte;
1965 if (pte_none(pteval)) {
1966 if (++none <= khugepaged_max_ptes_none)
1971 if (!pte_present(pteval) || !pte_write(pteval))
1973 page = vm_normal_page(vma, address, pteval);
1974 if (unlikely(!page))
1977 VM_BUG_ON(PageCompound(page));
1978 BUG_ON(!PageAnon(page));
1979 VM_BUG_ON(!PageSwapBacked(page));
1981 /* cannot use mapcount: can't collapse if there's a gup pin */
1982 if (page_count(page) != 1)
1985 * We can do it before isolate_lru_page because the
1986 * page can't be freed from under us. NOTE: PG_lock
1987 * is needed to serialize against split_huge_page
1988 * when invoked from the VM.
1990 if (!trylock_page(page))
1993 * Isolate the page to avoid collapsing an hugepage
1994 * currently in use by the VM.
1996 if (isolate_lru_page(page)) {
2000 /* 0 stands for page_is_file_cache(page) == false */
2001 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2002 VM_BUG_ON(!PageLocked(page));
2003 VM_BUG_ON(PageLRU(page));
2005 /* If there is no mapped pte young don't collapse the page */
2006 if (pte_young(pteval) || PageReferenced(page) ||
2007 mmu_notifier_test_young(vma->vm_mm, address))
2010 if (likely(referenced))
2013 release_pte_pages(pte, _pte);
2017 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2018 struct vm_area_struct *vma,
2019 unsigned long address,
2023 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2024 pte_t pteval = *_pte;
2025 struct page *src_page;
2027 if (pte_none(pteval)) {
2028 clear_user_highpage(page, address);
2029 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2031 src_page = pte_page(pteval);
2032 copy_user_highpage(page, src_page, address, vma);
2033 VM_BUG_ON(page_mapcount(src_page) != 1);
2034 release_pte_page(src_page);
2036 * ptl mostly unnecessary, but preempt has to
2037 * be disabled to update the per-cpu stats
2038 * inside page_remove_rmap().
2042 * paravirt calls inside pte_clear here are
2045 pte_clear(vma->vm_mm, address, _pte);
2046 page_remove_rmap(src_page);
2048 free_page_and_swap_cache(src_page);
2051 address += PAGE_SIZE;
2056 static void khugepaged_alloc_sleep(void)
2058 wait_event_freezable_timeout(khugepaged_wait, false,
2059 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2063 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2065 if (IS_ERR(*hpage)) {
2071 khugepaged_alloc_sleep();
2072 } else if (*hpage) {
2081 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2082 struct vm_area_struct *vma, unsigned long address,
2087 * Allocate the page while the vma is still valid and under
2088 * the mmap_sem read mode so there is no memory allocation
2089 * later when we take the mmap_sem in write mode. This is more
2090 * friendly behavior (OTOH it may actually hide bugs) to
2091 * filesystems in userland with daemons allocating memory in
2092 * the userland I/O paths. Allocating memory with the
2093 * mmap_sem in read mode is good idea also to allow greater
2096 *hpage = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2097 node, __GFP_OTHER_NODE);
2100 * After allocating the hugepage, release the mmap_sem read lock in
2101 * preparation for taking it in write mode.
2103 up_read(&mm->mmap_sem);
2104 if (unlikely(!*hpage)) {
2105 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2106 *hpage = ERR_PTR(-ENOMEM);
2110 count_vm_event(THP_COLLAPSE_ALLOC);
2114 static struct page *khugepaged_alloc_hugepage(bool *wait)
2119 hpage = alloc_hugepage(khugepaged_defrag());
2121 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2126 khugepaged_alloc_sleep();
2128 count_vm_event(THP_COLLAPSE_ALLOC);
2129 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2134 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2137 *hpage = khugepaged_alloc_hugepage(wait);
2139 if (unlikely(!*hpage))
2146 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2147 struct vm_area_struct *vma, unsigned long address,
2150 up_read(&mm->mmap_sem);
2156 static bool hugepage_vma_check(struct vm_area_struct *vma)
2158 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2159 (vma->vm_flags & VM_NOHUGEPAGE))
2162 if (!vma->anon_vma || vma->vm_ops)
2164 if (is_vma_temporary_stack(vma))
2166 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2170 static void collapse_huge_page(struct mm_struct *mm,
2171 unsigned long address,
2172 struct page **hpage,
2173 struct vm_area_struct *vma,
2179 struct page *new_page;
2182 unsigned long hstart, hend;
2183 unsigned long mmun_start; /* For mmu_notifiers */
2184 unsigned long mmun_end; /* For mmu_notifiers */
2186 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2188 /* release the mmap_sem read lock. */
2189 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2193 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2197 * Prevent all access to pagetables with the exception of
2198 * gup_fast later hanlded by the ptep_clear_flush and the VM
2199 * handled by the anon_vma lock + PG_lock.
2201 down_write(&mm->mmap_sem);
2202 if (unlikely(khugepaged_test_exit(mm)))
2205 vma = find_vma(mm, address);
2206 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2207 hend = vma->vm_end & HPAGE_PMD_MASK;
2208 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2210 if (!hugepage_vma_check(vma))
2212 pmd = mm_find_pmd(mm, address);
2215 if (pmd_trans_huge(*pmd))
2218 anon_vma_lock(vma->anon_vma);
2220 pte = pte_offset_map(pmd, address);
2221 ptl = pte_lockptr(mm, pmd);
2223 mmun_start = address;
2224 mmun_end = address + HPAGE_PMD_SIZE;
2225 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2226 spin_lock(&mm->page_table_lock); /* probably unnecessary */
2228 * After this gup_fast can't run anymore. This also removes
2229 * any huge TLB entry from the CPU so we won't allow
2230 * huge and small TLB entries for the same virtual address
2231 * to avoid the risk of CPU bugs in that area.
2233 _pmd = pmdp_clear_flush(vma, address, pmd);
2234 spin_unlock(&mm->page_table_lock);
2235 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2238 isolated = __collapse_huge_page_isolate(vma, address, pte);
2241 if (unlikely(!isolated)) {
2243 spin_lock(&mm->page_table_lock);
2244 BUG_ON(!pmd_none(*pmd));
2245 set_pmd_at(mm, address, pmd, _pmd);
2246 spin_unlock(&mm->page_table_lock);
2247 anon_vma_unlock(vma->anon_vma);
2252 * All pages are isolated and locked so anon_vma rmap
2253 * can't run anymore.
2255 anon_vma_unlock(vma->anon_vma);
2257 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2259 __SetPageUptodate(new_page);
2260 pgtable = pmd_pgtable(_pmd);
2262 _pmd = mk_huge_pmd(new_page, vma);
2265 * spin_lock() below is not the equivalent of smp_wmb(), so
2266 * this is needed to avoid the copy_huge_page writes to become
2267 * visible after the set_pmd_at() write.
2271 spin_lock(&mm->page_table_lock);
2272 BUG_ON(!pmd_none(*pmd));
2273 page_add_new_anon_rmap(new_page, vma, address);
2274 set_pmd_at(mm, address, pmd, _pmd);
2275 update_mmu_cache_pmd(vma, address, pmd);
2276 pgtable_trans_huge_deposit(mm, pgtable);
2277 spin_unlock(&mm->page_table_lock);
2281 khugepaged_pages_collapsed++;
2283 up_write(&mm->mmap_sem);
2287 mem_cgroup_uncharge_page(new_page);
2291 static int khugepaged_scan_pmd(struct mm_struct *mm,
2292 struct vm_area_struct *vma,
2293 unsigned long address,
2294 struct page **hpage)
2298 int ret = 0, referenced = 0, none = 0;
2300 unsigned long _address;
2304 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2306 pmd = mm_find_pmd(mm, address);
2309 if (pmd_trans_huge(*pmd))
2312 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2313 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2314 _pte++, _address += PAGE_SIZE) {
2315 pte_t pteval = *_pte;
2316 if (pte_none(pteval)) {
2317 if (++none <= khugepaged_max_ptes_none)
2322 if (!pte_present(pteval) || !pte_write(pteval))
2324 page = vm_normal_page(vma, _address, pteval);
2325 if (unlikely(!page))
2328 * Chose the node of the first page. This could
2329 * be more sophisticated and look at more pages,
2330 * but isn't for now.
2333 node = page_to_nid(page);
2334 VM_BUG_ON(PageCompound(page));
2335 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2337 /* cannot use mapcount: can't collapse if there's a gup pin */
2338 if (page_count(page) != 1)
2340 if (pte_young(pteval) || PageReferenced(page) ||
2341 mmu_notifier_test_young(vma->vm_mm, address))
2347 pte_unmap_unlock(pte, ptl);
2349 /* collapse_huge_page will return with the mmap_sem released */
2350 collapse_huge_page(mm, address, hpage, vma, node);
2355 static void collect_mm_slot(struct mm_slot *mm_slot)
2357 struct mm_struct *mm = mm_slot->mm;
2359 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2361 if (khugepaged_test_exit(mm)) {
2363 hlist_del(&mm_slot->hash);
2364 list_del(&mm_slot->mm_node);
2367 * Not strictly needed because the mm exited already.
2369 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2372 /* khugepaged_mm_lock actually not necessary for the below */
2373 free_mm_slot(mm_slot);
2378 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2379 struct page **hpage)
2380 __releases(&khugepaged_mm_lock)
2381 __acquires(&khugepaged_mm_lock)
2383 struct mm_slot *mm_slot;
2384 struct mm_struct *mm;
2385 struct vm_area_struct *vma;
2389 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2391 if (khugepaged_scan.mm_slot)
2392 mm_slot = khugepaged_scan.mm_slot;
2394 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2395 struct mm_slot, mm_node);
2396 khugepaged_scan.address = 0;
2397 khugepaged_scan.mm_slot = mm_slot;
2399 spin_unlock(&khugepaged_mm_lock);
2402 down_read(&mm->mmap_sem);
2403 if (unlikely(khugepaged_test_exit(mm)))
2406 vma = find_vma(mm, khugepaged_scan.address);
2409 for (; vma; vma = vma->vm_next) {
2410 unsigned long hstart, hend;
2413 if (unlikely(khugepaged_test_exit(mm))) {
2417 if (!hugepage_vma_check(vma)) {
2422 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2423 hend = vma->vm_end & HPAGE_PMD_MASK;
2426 if (khugepaged_scan.address > hend)
2428 if (khugepaged_scan.address < hstart)
2429 khugepaged_scan.address = hstart;
2430 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2432 while (khugepaged_scan.address < hend) {
2435 if (unlikely(khugepaged_test_exit(mm)))
2436 goto breakouterloop;
2438 VM_BUG_ON(khugepaged_scan.address < hstart ||
2439 khugepaged_scan.address + HPAGE_PMD_SIZE >
2441 ret = khugepaged_scan_pmd(mm, vma,
2442 khugepaged_scan.address,
2444 /* move to next address */
2445 khugepaged_scan.address += HPAGE_PMD_SIZE;
2446 progress += HPAGE_PMD_NR;
2448 /* we released mmap_sem so break loop */
2449 goto breakouterloop_mmap_sem;
2450 if (progress >= pages)
2451 goto breakouterloop;
2455 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2456 breakouterloop_mmap_sem:
2458 spin_lock(&khugepaged_mm_lock);
2459 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2461 * Release the current mm_slot if this mm is about to die, or
2462 * if we scanned all vmas of this mm.
2464 if (khugepaged_test_exit(mm) || !vma) {
2466 * Make sure that if mm_users is reaching zero while
2467 * khugepaged runs here, khugepaged_exit will find
2468 * mm_slot not pointing to the exiting mm.
2470 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2471 khugepaged_scan.mm_slot = list_entry(
2472 mm_slot->mm_node.next,
2473 struct mm_slot, mm_node);
2474 khugepaged_scan.address = 0;
2476 khugepaged_scan.mm_slot = NULL;
2477 khugepaged_full_scans++;
2480 collect_mm_slot(mm_slot);
2486 static int khugepaged_has_work(void)
2488 return !list_empty(&khugepaged_scan.mm_head) &&
2489 khugepaged_enabled();
2492 static int khugepaged_wait_event(void)
2494 return !list_empty(&khugepaged_scan.mm_head) ||
2495 kthread_should_stop();
2498 static void khugepaged_do_scan(void)
2500 struct page *hpage = NULL;
2501 unsigned int progress = 0, pass_through_head = 0;
2502 unsigned int pages = khugepaged_pages_to_scan;
2505 barrier(); /* write khugepaged_pages_to_scan to local stack */
2507 while (progress < pages) {
2508 if (!khugepaged_prealloc_page(&hpage, &wait))
2513 if (unlikely(kthread_should_stop() || freezing(current)))
2516 spin_lock(&khugepaged_mm_lock);
2517 if (!khugepaged_scan.mm_slot)
2518 pass_through_head++;
2519 if (khugepaged_has_work() &&
2520 pass_through_head < 2)
2521 progress += khugepaged_scan_mm_slot(pages - progress,
2525 spin_unlock(&khugepaged_mm_lock);
2528 if (!IS_ERR_OR_NULL(hpage))
2532 static void khugepaged_wait_work(void)
2536 if (khugepaged_has_work()) {
2537 if (!khugepaged_scan_sleep_millisecs)
2540 wait_event_freezable_timeout(khugepaged_wait,
2541 kthread_should_stop(),
2542 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2546 if (khugepaged_enabled())
2547 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2550 static int khugepaged(void *none)
2552 struct mm_slot *mm_slot;
2555 set_user_nice(current, 19);
2557 while (!kthread_should_stop()) {
2558 khugepaged_do_scan();
2559 khugepaged_wait_work();
2562 spin_lock(&khugepaged_mm_lock);
2563 mm_slot = khugepaged_scan.mm_slot;
2564 khugepaged_scan.mm_slot = NULL;
2566 collect_mm_slot(mm_slot);
2567 spin_unlock(&khugepaged_mm_lock);
2571 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2572 unsigned long haddr, pmd_t *pmd)
2574 struct mm_struct *mm = vma->vm_mm;
2579 pmdp_clear_flush(vma, haddr, pmd);
2580 /* leave pmd empty until pte is filled */
2582 pgtable = pgtable_trans_huge_withdraw(mm);
2583 pmd_populate(mm, &_pmd, pgtable);
2585 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2587 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2588 entry = pte_mkspecial(entry);
2589 pte = pte_offset_map(&_pmd, haddr);
2590 VM_BUG_ON(!pte_none(*pte));
2591 set_pte_at(mm, haddr, pte, entry);
2594 smp_wmb(); /* make pte visible before pmd */
2595 pmd_populate(mm, pmd, pgtable);
2596 put_huge_zero_page();
2599 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2603 struct mm_struct *mm = vma->vm_mm;
2604 unsigned long haddr = address & HPAGE_PMD_MASK;
2605 unsigned long mmun_start; /* For mmu_notifiers */
2606 unsigned long mmun_end; /* For mmu_notifiers */
2608 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2611 mmun_end = haddr + HPAGE_PMD_SIZE;
2612 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2613 spin_lock(&mm->page_table_lock);
2614 if (unlikely(!pmd_trans_huge(*pmd))) {
2615 spin_unlock(&mm->page_table_lock);
2616 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2619 if (is_huge_zero_pmd(*pmd)) {
2620 __split_huge_zero_page_pmd(vma, haddr, pmd);
2621 spin_unlock(&mm->page_table_lock);
2622 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2625 page = pmd_page(*pmd);
2626 VM_BUG_ON(!page_count(page));
2628 spin_unlock(&mm->page_table_lock);
2629 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2631 split_huge_page(page);
2634 BUG_ON(pmd_trans_huge(*pmd));
2637 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2640 struct vm_area_struct *vma;
2642 vma = find_vma(mm, address);
2643 BUG_ON(vma == NULL);
2644 split_huge_page_pmd(vma, address, pmd);
2647 static void split_huge_page_address(struct mm_struct *mm,
2648 unsigned long address)
2652 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2654 pmd = mm_find_pmd(mm, address);
2658 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2659 * materialize from under us.
2661 split_huge_page_pmd_mm(mm, address, pmd);
2664 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2665 unsigned long start,
2670 * If the new start address isn't hpage aligned and it could
2671 * previously contain an hugepage: check if we need to split
2674 if (start & ~HPAGE_PMD_MASK &&
2675 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2676 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2677 split_huge_page_address(vma->vm_mm, start);
2680 * If the new end address isn't hpage aligned and it could
2681 * previously contain an hugepage: check if we need to split
2684 if (end & ~HPAGE_PMD_MASK &&
2685 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2686 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2687 split_huge_page_address(vma->vm_mm, end);
2690 * If we're also updating the vma->vm_next->vm_start, if the new
2691 * vm_next->vm_start isn't page aligned and it could previously
2692 * contain an hugepage: check if we need to split an huge pmd.
2694 if (adjust_next > 0) {
2695 struct vm_area_struct *next = vma->vm_next;
2696 unsigned long nstart = next->vm_start;
2697 nstart += adjust_next << PAGE_SHIFT;
2698 if (nstart & ~HPAGE_PMD_MASK &&
2699 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2700 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2701 split_huge_page_address(next->vm_mm, nstart);