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>
22 #include <linux/migrate.h>
23 #include <linux/hashtable.h>
26 #include <asm/pgalloc.h>
30 * By default transparent hugepage support is disabled in order that avoid
31 * to risk increase the memory footprint of applications without a guaranteed
32 * benefit. When transparent hugepage support is enabled, is for all mappings,
33 * and khugepaged scans all mappings.
34 * Defrag is invoked by khugepaged hugepage allocations and by page faults
35 * for all hugepage allocations.
37 unsigned long transparent_hugepage_flags __read_mostly =
38 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
39 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
41 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
42 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
44 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
45 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
46 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
48 /* default scan 8*512 pte (or vmas) every 30 second */
49 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
50 static unsigned int khugepaged_pages_collapsed;
51 static unsigned int khugepaged_full_scans;
52 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
53 /* during fragmentation poll the hugepage allocator once every minute */
54 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
55 static struct task_struct *khugepaged_thread __read_mostly;
56 static DEFINE_MUTEX(khugepaged_mutex);
57 static DEFINE_SPINLOCK(khugepaged_mm_lock);
58 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
60 * default collapse hugepages if there is at least one pte mapped like
61 * it would have happened if the vma was large enough during page
64 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
66 static int khugepaged(void *none);
67 static int khugepaged_slab_init(void);
69 #define MM_SLOTS_HASH_BITS 10
70 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
72 static struct kmem_cache *mm_slot_cache __read_mostly;
75 * struct mm_slot - hash lookup from mm to mm_slot
76 * @hash: hash collision list
77 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
78 * @mm: the mm that this information is valid for
81 struct hlist_node hash;
82 struct list_head mm_node;
87 * struct khugepaged_scan - cursor for scanning
88 * @mm_head: the head of the mm list to scan
89 * @mm_slot: the current mm_slot we are scanning
90 * @address: the next address inside that to be scanned
92 * There is only the one khugepaged_scan instance of this cursor structure.
94 struct khugepaged_scan {
95 struct list_head mm_head;
96 struct mm_slot *mm_slot;
97 unsigned long address;
99 static struct khugepaged_scan khugepaged_scan = {
100 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
104 static int set_recommended_min_free_kbytes(void)
108 unsigned long recommended_min;
110 if (!khugepaged_enabled())
113 for_each_populated_zone(zone)
116 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
117 recommended_min = pageblock_nr_pages * nr_zones * 2;
120 * Make sure that on average at least two pageblocks are almost free
121 * of another type, one for a migratetype to fall back to and a
122 * second to avoid subsequent fallbacks of other types There are 3
123 * MIGRATE_TYPES we care about.
125 recommended_min += pageblock_nr_pages * nr_zones *
126 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
128 /* don't ever allow to reserve more than 5% of the lowmem */
129 recommended_min = min(recommended_min,
130 (unsigned long) nr_free_buffer_pages() / 20);
131 recommended_min <<= (PAGE_SHIFT-10);
133 if (recommended_min > min_free_kbytes)
134 min_free_kbytes = recommended_min;
135 setup_per_zone_wmarks();
138 late_initcall(set_recommended_min_free_kbytes);
140 static int start_khugepaged(void)
143 if (khugepaged_enabled()) {
144 if (!khugepaged_thread)
145 khugepaged_thread = kthread_run(khugepaged, NULL,
147 if (unlikely(IS_ERR(khugepaged_thread))) {
149 "khugepaged: kthread_run(khugepaged) failed\n");
150 err = PTR_ERR(khugepaged_thread);
151 khugepaged_thread = NULL;
154 if (!list_empty(&khugepaged_scan.mm_head))
155 wake_up_interruptible(&khugepaged_wait);
157 set_recommended_min_free_kbytes();
158 } else if (khugepaged_thread) {
159 kthread_stop(khugepaged_thread);
160 khugepaged_thread = NULL;
166 static atomic_t huge_zero_refcount;
167 static struct page *huge_zero_page __read_mostly;
169 static inline bool is_huge_zero_page(struct page *page)
171 return ACCESS_ONCE(huge_zero_page) == page;
174 static inline bool is_huge_zero_pmd(pmd_t pmd)
176 return is_huge_zero_page(pmd_page(pmd));
179 static struct page *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_page);
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_page, NULL, 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_page);
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 unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
216 struct shrink_control *sc)
218 /* we can free zero page only if last reference remains */
219 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
222 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
223 struct shrink_control *sc)
225 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
226 struct page *zero_page = xchg(&huge_zero_page, NULL);
227 BUG_ON(zero_page == NULL);
228 __free_page(zero_page);
235 static struct shrinker huge_zero_page_shrinker = {
236 .count_objects = shrink_huge_zero_page_count,
237 .scan_objects = shrink_huge_zero_page_scan,
238 .seeks = DEFAULT_SEEKS,
243 static ssize_t double_flag_show(struct kobject *kobj,
244 struct kobj_attribute *attr, char *buf,
245 enum transparent_hugepage_flag enabled,
246 enum transparent_hugepage_flag req_madv)
248 if (test_bit(enabled, &transparent_hugepage_flags)) {
249 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
250 return sprintf(buf, "[always] madvise never\n");
251 } else if (test_bit(req_madv, &transparent_hugepage_flags))
252 return sprintf(buf, "always [madvise] never\n");
254 return sprintf(buf, "always madvise [never]\n");
256 static ssize_t double_flag_store(struct kobject *kobj,
257 struct kobj_attribute *attr,
258 const char *buf, size_t count,
259 enum transparent_hugepage_flag enabled,
260 enum transparent_hugepage_flag req_madv)
262 if (!memcmp("always", buf,
263 min(sizeof("always")-1, count))) {
264 set_bit(enabled, &transparent_hugepage_flags);
265 clear_bit(req_madv, &transparent_hugepage_flags);
266 } else if (!memcmp("madvise", buf,
267 min(sizeof("madvise")-1, count))) {
268 clear_bit(enabled, &transparent_hugepage_flags);
269 set_bit(req_madv, &transparent_hugepage_flags);
270 } else if (!memcmp("never", buf,
271 min(sizeof("never")-1, count))) {
272 clear_bit(enabled, &transparent_hugepage_flags);
273 clear_bit(req_madv, &transparent_hugepage_flags);
280 static ssize_t enabled_show(struct kobject *kobj,
281 struct kobj_attribute *attr, char *buf)
283 return double_flag_show(kobj, attr, buf,
284 TRANSPARENT_HUGEPAGE_FLAG,
285 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
287 static ssize_t enabled_store(struct kobject *kobj,
288 struct kobj_attribute *attr,
289 const char *buf, size_t count)
293 ret = double_flag_store(kobj, attr, buf, count,
294 TRANSPARENT_HUGEPAGE_FLAG,
295 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
300 mutex_lock(&khugepaged_mutex);
301 err = start_khugepaged();
302 mutex_unlock(&khugepaged_mutex);
310 static struct kobj_attribute enabled_attr =
311 __ATTR(enabled, 0644, enabled_show, enabled_store);
313 static ssize_t single_flag_show(struct kobject *kobj,
314 struct kobj_attribute *attr, char *buf,
315 enum transparent_hugepage_flag flag)
317 return sprintf(buf, "%d\n",
318 !!test_bit(flag, &transparent_hugepage_flags));
321 static ssize_t single_flag_store(struct kobject *kobj,
322 struct kobj_attribute *attr,
323 const char *buf, size_t count,
324 enum transparent_hugepage_flag flag)
329 ret = kstrtoul(buf, 10, &value);
336 set_bit(flag, &transparent_hugepage_flags);
338 clear_bit(flag, &transparent_hugepage_flags);
344 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
345 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
346 * memory just to allocate one more hugepage.
348 static ssize_t defrag_show(struct kobject *kobj,
349 struct kobj_attribute *attr, char *buf)
351 return double_flag_show(kobj, attr, buf,
352 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
353 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
355 static ssize_t defrag_store(struct kobject *kobj,
356 struct kobj_attribute *attr,
357 const char *buf, size_t count)
359 return double_flag_store(kobj, attr, buf, count,
360 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
361 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
363 static struct kobj_attribute defrag_attr =
364 __ATTR(defrag, 0644, defrag_show, defrag_store);
366 static ssize_t use_zero_page_show(struct kobject *kobj,
367 struct kobj_attribute *attr, char *buf)
369 return single_flag_show(kobj, attr, buf,
370 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
372 static ssize_t use_zero_page_store(struct kobject *kobj,
373 struct kobj_attribute *attr, const char *buf, size_t count)
375 return single_flag_store(kobj, attr, buf, count,
376 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
378 static struct kobj_attribute use_zero_page_attr =
379 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
380 #ifdef CONFIG_DEBUG_VM
381 static ssize_t debug_cow_show(struct kobject *kobj,
382 struct kobj_attribute *attr, char *buf)
384 return single_flag_show(kobj, attr, buf,
385 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
387 static ssize_t debug_cow_store(struct kobject *kobj,
388 struct kobj_attribute *attr,
389 const char *buf, size_t count)
391 return single_flag_store(kobj, attr, buf, count,
392 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
394 static struct kobj_attribute debug_cow_attr =
395 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
396 #endif /* CONFIG_DEBUG_VM */
398 static struct attribute *hugepage_attr[] = {
401 &use_zero_page_attr.attr,
402 #ifdef CONFIG_DEBUG_VM
403 &debug_cow_attr.attr,
408 static struct attribute_group hugepage_attr_group = {
409 .attrs = hugepage_attr,
412 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
413 struct kobj_attribute *attr,
416 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
419 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
420 struct kobj_attribute *attr,
421 const char *buf, size_t count)
426 err = kstrtoul(buf, 10, &msecs);
427 if (err || msecs > UINT_MAX)
430 khugepaged_scan_sleep_millisecs = msecs;
431 wake_up_interruptible(&khugepaged_wait);
435 static struct kobj_attribute scan_sleep_millisecs_attr =
436 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
437 scan_sleep_millisecs_store);
439 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
440 struct kobj_attribute *attr,
443 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
446 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
447 struct kobj_attribute *attr,
448 const char *buf, size_t count)
453 err = kstrtoul(buf, 10, &msecs);
454 if (err || msecs > UINT_MAX)
457 khugepaged_alloc_sleep_millisecs = msecs;
458 wake_up_interruptible(&khugepaged_wait);
462 static struct kobj_attribute alloc_sleep_millisecs_attr =
463 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
464 alloc_sleep_millisecs_store);
466 static ssize_t pages_to_scan_show(struct kobject *kobj,
467 struct kobj_attribute *attr,
470 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
472 static ssize_t pages_to_scan_store(struct kobject *kobj,
473 struct kobj_attribute *attr,
474 const char *buf, size_t count)
479 err = kstrtoul(buf, 10, &pages);
480 if (err || !pages || pages > UINT_MAX)
483 khugepaged_pages_to_scan = pages;
487 static struct kobj_attribute pages_to_scan_attr =
488 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
489 pages_to_scan_store);
491 static ssize_t pages_collapsed_show(struct kobject *kobj,
492 struct kobj_attribute *attr,
495 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
497 static struct kobj_attribute pages_collapsed_attr =
498 __ATTR_RO(pages_collapsed);
500 static ssize_t full_scans_show(struct kobject *kobj,
501 struct kobj_attribute *attr,
504 return sprintf(buf, "%u\n", khugepaged_full_scans);
506 static struct kobj_attribute full_scans_attr =
507 __ATTR_RO(full_scans);
509 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
510 struct kobj_attribute *attr, char *buf)
512 return single_flag_show(kobj, attr, buf,
513 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
515 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
516 struct kobj_attribute *attr,
517 const char *buf, size_t count)
519 return single_flag_store(kobj, attr, buf, count,
520 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
522 static struct kobj_attribute khugepaged_defrag_attr =
523 __ATTR(defrag, 0644, khugepaged_defrag_show,
524 khugepaged_defrag_store);
527 * max_ptes_none controls if khugepaged should collapse hugepages over
528 * any unmapped ptes in turn potentially increasing the memory
529 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
530 * reduce the available free memory in the system as it
531 * runs. Increasing max_ptes_none will instead potentially reduce the
532 * free memory in the system during the khugepaged scan.
534 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
535 struct kobj_attribute *attr,
538 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
540 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
541 struct kobj_attribute *attr,
542 const char *buf, size_t count)
545 unsigned long max_ptes_none;
547 err = kstrtoul(buf, 10, &max_ptes_none);
548 if (err || max_ptes_none > HPAGE_PMD_NR-1)
551 khugepaged_max_ptes_none = max_ptes_none;
555 static struct kobj_attribute khugepaged_max_ptes_none_attr =
556 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
557 khugepaged_max_ptes_none_store);
559 static struct attribute *khugepaged_attr[] = {
560 &khugepaged_defrag_attr.attr,
561 &khugepaged_max_ptes_none_attr.attr,
562 &pages_to_scan_attr.attr,
563 &pages_collapsed_attr.attr,
564 &full_scans_attr.attr,
565 &scan_sleep_millisecs_attr.attr,
566 &alloc_sleep_millisecs_attr.attr,
570 static struct attribute_group khugepaged_attr_group = {
571 .attrs = khugepaged_attr,
572 .name = "khugepaged",
575 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
579 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
580 if (unlikely(!*hugepage_kobj)) {
581 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
585 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
587 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
591 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
593 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
594 goto remove_hp_group;
600 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
602 kobject_put(*hugepage_kobj);
606 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
608 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
609 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
610 kobject_put(hugepage_kobj);
613 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
618 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
621 #endif /* CONFIG_SYSFS */
623 static int __init hugepage_init(void)
626 struct kobject *hugepage_kobj;
628 if (!has_transparent_hugepage()) {
629 transparent_hugepage_flags = 0;
633 err = hugepage_init_sysfs(&hugepage_kobj);
637 err = khugepaged_slab_init();
641 register_shrinker(&huge_zero_page_shrinker);
644 * By default disable transparent hugepages on smaller systems,
645 * where the extra memory used could hurt more than TLB overhead
646 * is likely to save. The admin can still enable it through /sys.
648 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
649 transparent_hugepage_flags = 0;
655 hugepage_exit_sysfs(hugepage_kobj);
658 module_init(hugepage_init)
660 static int __init setup_transparent_hugepage(char *str)
665 if (!strcmp(str, "always")) {
666 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
667 &transparent_hugepage_flags);
668 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
669 &transparent_hugepage_flags);
671 } else if (!strcmp(str, "madvise")) {
672 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
673 &transparent_hugepage_flags);
674 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
675 &transparent_hugepage_flags);
677 } else if (!strcmp(str, "never")) {
678 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
679 &transparent_hugepage_flags);
680 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
681 &transparent_hugepage_flags);
687 "transparent_hugepage= cannot parse, ignored\n");
690 __setup("transparent_hugepage=", setup_transparent_hugepage);
692 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
694 if (likely(vma->vm_flags & VM_WRITE))
695 pmd = pmd_mkwrite(pmd);
699 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
702 entry = mk_pmd(page, prot);
703 entry = pmd_mkhuge(entry);
707 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
708 struct vm_area_struct *vma,
709 unsigned long haddr, pmd_t *pmd,
715 VM_BUG_ON(!PageCompound(page));
716 pgtable = pte_alloc_one(mm, haddr);
717 if (unlikely(!pgtable))
720 clear_huge_page(page, haddr, HPAGE_PMD_NR);
722 * The memory barrier inside __SetPageUptodate makes sure that
723 * clear_huge_page writes become visible before the set_pmd_at()
726 __SetPageUptodate(page);
728 ptl = pmd_lock(mm, pmd);
729 if (unlikely(!pmd_none(*pmd))) {
731 mem_cgroup_uncharge_page(page);
733 pte_free(mm, pgtable);
736 entry = mk_huge_pmd(page, vma->vm_page_prot);
737 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
738 page_add_new_anon_rmap(page, vma, haddr);
739 pgtable_trans_huge_deposit(mm, pmd, pgtable);
740 set_pmd_at(mm, haddr, pmd, entry);
741 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
742 atomic_long_inc(&mm->nr_ptes);
749 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
751 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
754 static inline struct page *alloc_hugepage_vma(int defrag,
755 struct vm_area_struct *vma,
756 unsigned long haddr, int nd,
759 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
760 HPAGE_PMD_ORDER, vma, haddr, nd);
763 /* Caller must hold page table lock. */
764 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
765 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
766 struct page *zero_page)
771 entry = mk_pmd(zero_page, vma->vm_page_prot);
772 entry = pmd_wrprotect(entry);
773 entry = pmd_mkhuge(entry);
774 pgtable_trans_huge_deposit(mm, pmd, pgtable);
775 set_pmd_at(mm, haddr, pmd, entry);
776 atomic_long_inc(&mm->nr_ptes);
780 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
781 unsigned long address, pmd_t *pmd,
785 unsigned long haddr = address & HPAGE_PMD_MASK;
787 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
788 return VM_FAULT_FALLBACK;
789 if (unlikely(anon_vma_prepare(vma)))
791 if (unlikely(khugepaged_enter(vma)))
793 if (!(flags & FAULT_FLAG_WRITE) &&
794 transparent_hugepage_use_zero_page()) {
797 struct page *zero_page;
799 pgtable = pte_alloc_one(mm, haddr);
800 if (unlikely(!pgtable))
802 zero_page = get_huge_zero_page();
803 if (unlikely(!zero_page)) {
804 pte_free(mm, pgtable);
805 count_vm_event(THP_FAULT_FALLBACK);
806 return VM_FAULT_FALLBACK;
808 ptl = pmd_lock(mm, pmd);
809 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
813 pte_free(mm, pgtable);
814 put_huge_zero_page();
818 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
819 vma, haddr, numa_node_id(), 0);
820 if (unlikely(!page)) {
821 count_vm_event(THP_FAULT_FALLBACK);
822 return VM_FAULT_FALLBACK;
824 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
826 count_vm_event(THP_FAULT_FALLBACK);
827 return VM_FAULT_FALLBACK;
829 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
830 mem_cgroup_uncharge_page(page);
832 count_vm_event(THP_FAULT_FALLBACK);
833 return VM_FAULT_FALLBACK;
836 count_vm_event(THP_FAULT_ALLOC);
840 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
841 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
842 struct vm_area_struct *vma)
844 spinlock_t *dst_ptl, *src_ptl;
845 struct page *src_page;
851 pgtable = pte_alloc_one(dst_mm, addr);
852 if (unlikely(!pgtable))
855 dst_ptl = pmd_lock(dst_mm, dst_pmd);
856 src_ptl = pmd_lockptr(src_mm, src_pmd);
857 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
861 if (unlikely(!pmd_trans_huge(pmd))) {
862 pte_free(dst_mm, pgtable);
866 * When page table lock is held, the huge zero pmd should not be
867 * under splitting since we don't split the page itself, only pmd to
870 if (is_huge_zero_pmd(pmd)) {
871 struct page *zero_page;
874 * get_huge_zero_page() will never allocate a new page here,
875 * since we already have a zero page to copy. It just takes a
878 zero_page = get_huge_zero_page();
879 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
881 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
885 if (unlikely(pmd_trans_splitting(pmd))) {
886 /* split huge page running from under us */
887 spin_unlock(src_ptl);
888 spin_unlock(dst_ptl);
889 pte_free(dst_mm, pgtable);
891 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
894 src_page = pmd_page(pmd);
895 VM_BUG_ON(!PageHead(src_page));
897 page_dup_rmap(src_page);
898 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
900 pmdp_set_wrprotect(src_mm, addr, src_pmd);
901 pmd = pmd_mkold(pmd_wrprotect(pmd));
902 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
903 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
904 atomic_long_inc(&dst_mm->nr_ptes);
908 spin_unlock(src_ptl);
909 spin_unlock(dst_ptl);
914 void huge_pmd_set_accessed(struct mm_struct *mm,
915 struct vm_area_struct *vma,
916 unsigned long address,
917 pmd_t *pmd, pmd_t orig_pmd,
924 ptl = pmd_lock(mm, pmd);
925 if (unlikely(!pmd_same(*pmd, orig_pmd)))
928 entry = pmd_mkyoung(orig_pmd);
929 haddr = address & HPAGE_PMD_MASK;
930 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
931 update_mmu_cache_pmd(vma, address, pmd);
937 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
938 struct vm_area_struct *vma, unsigned long address,
939 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
946 unsigned long mmun_start; /* For mmu_notifiers */
947 unsigned long mmun_end; /* For mmu_notifiers */
949 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
955 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
961 clear_user_highpage(page, address);
962 __SetPageUptodate(page);
965 mmun_end = haddr + HPAGE_PMD_SIZE;
966 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
968 ptl = pmd_lock(mm, pmd);
969 if (unlikely(!pmd_same(*pmd, orig_pmd)))
972 pmdp_clear_flush(vma, haddr, pmd);
973 /* leave pmd empty until pte is filled */
975 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
976 pmd_populate(mm, &_pmd, pgtable);
978 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
980 if (haddr == (address & PAGE_MASK)) {
981 entry = mk_pte(page, vma->vm_page_prot);
982 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
983 page_add_new_anon_rmap(page, vma, haddr);
985 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
986 entry = pte_mkspecial(entry);
988 pte = pte_offset_map(&_pmd, haddr);
989 VM_BUG_ON(!pte_none(*pte));
990 set_pte_at(mm, haddr, pte, entry);
993 smp_wmb(); /* make pte visible before pmd */
994 pmd_populate(mm, pmd, pgtable);
996 put_huge_zero_page();
997 inc_mm_counter(mm, MM_ANONPAGES);
999 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1001 ret |= VM_FAULT_WRITE;
1006 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1007 mem_cgroup_uncharge_page(page);
1012 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1013 struct vm_area_struct *vma,
1014 unsigned long address,
1015 pmd_t *pmd, pmd_t orig_pmd,
1017 unsigned long haddr)
1023 struct page **pages;
1024 unsigned long mmun_start; /* For mmu_notifiers */
1025 unsigned long mmun_end; /* For mmu_notifiers */
1027 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1029 if (unlikely(!pages)) {
1030 ret |= VM_FAULT_OOM;
1034 for (i = 0; i < HPAGE_PMD_NR; i++) {
1035 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1037 vma, address, page_to_nid(page));
1038 if (unlikely(!pages[i] ||
1039 mem_cgroup_newpage_charge(pages[i], mm,
1043 mem_cgroup_uncharge_start();
1045 mem_cgroup_uncharge_page(pages[i]);
1048 mem_cgroup_uncharge_end();
1050 ret |= VM_FAULT_OOM;
1055 for (i = 0; i < HPAGE_PMD_NR; i++) {
1056 copy_user_highpage(pages[i], page + i,
1057 haddr + PAGE_SIZE * i, vma);
1058 __SetPageUptodate(pages[i]);
1063 mmun_end = haddr + HPAGE_PMD_SIZE;
1064 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1066 ptl = pmd_lock(mm, pmd);
1067 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1068 goto out_free_pages;
1069 VM_BUG_ON(!PageHead(page));
1071 pmdp_clear_flush(vma, haddr, pmd);
1072 /* leave pmd empty until pte is filled */
1074 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1075 pmd_populate(mm, &_pmd, pgtable);
1077 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1079 entry = mk_pte(pages[i], vma->vm_page_prot);
1080 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1081 page_add_new_anon_rmap(pages[i], vma, haddr);
1082 pte = pte_offset_map(&_pmd, haddr);
1083 VM_BUG_ON(!pte_none(*pte));
1084 set_pte_at(mm, haddr, pte, entry);
1089 smp_wmb(); /* make pte visible before pmd */
1090 pmd_populate(mm, pmd, pgtable);
1091 page_remove_rmap(page);
1094 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1096 ret |= VM_FAULT_WRITE;
1104 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1105 mem_cgroup_uncharge_start();
1106 for (i = 0; i < HPAGE_PMD_NR; i++) {
1107 mem_cgroup_uncharge_page(pages[i]);
1110 mem_cgroup_uncharge_end();
1115 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1116 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1120 struct page *page = NULL, *new_page;
1121 unsigned long haddr;
1122 unsigned long mmun_start; /* For mmu_notifiers */
1123 unsigned long mmun_end; /* For mmu_notifiers */
1125 ptl = pmd_lockptr(mm, pmd);
1126 VM_BUG_ON(!vma->anon_vma);
1127 haddr = address & HPAGE_PMD_MASK;
1128 if (is_huge_zero_pmd(orig_pmd))
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;
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 if (is_huge_zero_pmd(orig_pmd)) {
1157 ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1158 address, pmd, orig_pmd, haddr);
1160 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1161 pmd, orig_pmd, page, haddr);
1162 if (ret & VM_FAULT_OOM)
1163 split_huge_page(page);
1166 count_vm_event(THP_FAULT_FALLBACK);
1170 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1173 split_huge_page(page);
1176 count_vm_event(THP_FAULT_FALLBACK);
1177 ret |= VM_FAULT_OOM;
1181 count_vm_event(THP_FAULT_ALLOC);
1183 if (is_huge_zero_pmd(orig_pmd))
1184 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1186 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1187 __SetPageUptodate(new_page);
1190 mmun_end = haddr + HPAGE_PMD_SIZE;
1191 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1196 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1198 mem_cgroup_uncharge_page(new_page);
1203 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1204 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1205 pmdp_clear_flush(vma, haddr, pmd);
1206 page_add_new_anon_rmap(new_page, vma, haddr);
1207 set_pmd_at(mm, haddr, pmd, entry);
1208 update_mmu_cache_pmd(vma, address, pmd);
1209 if (is_huge_zero_pmd(orig_pmd)) {
1210 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1211 put_huge_zero_page();
1213 VM_BUG_ON(!PageHead(page));
1214 page_remove_rmap(page);
1217 ret |= VM_FAULT_WRITE;
1221 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1229 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1234 struct mm_struct *mm = vma->vm_mm;
1235 struct page *page = NULL;
1237 assert_spin_locked(pmd_lockptr(mm, pmd));
1239 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1242 /* Avoid dumping huge zero page */
1243 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1244 return ERR_PTR(-EFAULT);
1246 /* Full NUMA hinting faults to serialise migration in fault paths */
1247 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1250 page = pmd_page(*pmd);
1251 VM_BUG_ON(!PageHead(page));
1252 if (flags & FOLL_TOUCH) {
1255 * We should set the dirty bit only for FOLL_WRITE but
1256 * for now the dirty bit in the pmd is meaningless.
1257 * And if the dirty bit will become meaningful and
1258 * we'll only set it with FOLL_WRITE, an atomic
1259 * set_bit will be required on the pmd to set the
1260 * young bit, instead of the current set_pmd_at.
1262 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1263 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1265 update_mmu_cache_pmd(vma, addr, pmd);
1267 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1268 if (page->mapping && trylock_page(page)) {
1271 mlock_vma_page(page);
1275 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1276 VM_BUG_ON(!PageCompound(page));
1277 if (flags & FOLL_GET)
1278 get_page_foll(page);
1284 /* NUMA hinting page fault entry point for trans huge pmds */
1285 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1286 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1289 struct anon_vma *anon_vma = NULL;
1291 unsigned long haddr = addr & HPAGE_PMD_MASK;
1292 int page_nid = -1, this_nid = numa_node_id();
1293 int target_nid, last_cpupid = -1;
1295 bool migrated = false;
1298 ptl = pmd_lock(mm, pmdp);
1299 if (unlikely(!pmd_same(pmd, *pmdp)))
1302 page = pmd_page(pmd);
1303 BUG_ON(is_huge_zero_page(page));
1304 page_nid = page_to_nid(page);
1305 last_cpupid = page_cpupid_last(page);
1306 count_vm_numa_event(NUMA_HINT_FAULTS);
1307 if (page_nid == this_nid) {
1308 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1309 flags |= TNF_FAULT_LOCAL;
1313 * Avoid grouping on DSO/COW pages in specific and RO pages
1314 * in general, RO pages shouldn't hurt as much anyway since
1315 * they can be in shared cache state.
1317 if (!pmd_write(pmd))
1318 flags |= TNF_NO_GROUP;
1321 * Acquire the page lock to serialise THP migrations but avoid dropping
1322 * page_table_lock if at all possible
1324 page_locked = trylock_page(page);
1325 target_nid = mpol_misplaced(page, vma, haddr);
1326 if (target_nid == -1) {
1327 /* If the page was locked, there are no parallel migrations */
1333 * If there are potential migrations, wait for completion and retry. We
1334 * do not relock and check_same as the page may no longer be mapped.
1335 * Furtermore, even if the page is currently misplaced, there is no
1336 * guarantee it is still misplaced after the migration completes.
1340 wait_on_page_locked(page);
1346 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1347 * to serialises splits
1351 anon_vma = page_lock_anon_vma_read(page);
1353 /* Confirm the PMD did not change while page_table_lock was released */
1355 if (unlikely(!pmd_same(pmd, *pmdp))) {
1362 /* Bail if we fail to protect against THP splits for any reason */
1363 if (unlikely(!anon_vma)) {
1370 * Migrate the THP to the requested node, returns with page unlocked
1371 * and pmd_numa cleared.
1374 migrated = migrate_misplaced_transhuge_page(mm, vma,
1375 pmdp, pmd, addr, page, target_nid);
1377 flags |= TNF_MIGRATED;
1378 page_nid = target_nid;
1383 BUG_ON(!PageLocked(page));
1384 pmd = pmd_mknonnuma(pmd);
1385 set_pmd_at(mm, haddr, pmdp, pmd);
1386 VM_BUG_ON(pmd_numa(*pmdp));
1387 update_mmu_cache_pmd(vma, addr, pmdp);
1394 page_unlock_anon_vma_read(anon_vma);
1397 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1402 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1403 pmd_t *pmd, unsigned long addr)
1408 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1413 * For architectures like ppc64 we look at deposited pgtable
1414 * when calling pmdp_get_and_clear. So do the
1415 * pgtable_trans_huge_withdraw after finishing pmdp related
1418 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1419 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1420 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1421 if (is_huge_zero_pmd(orig_pmd)) {
1422 atomic_long_dec(&tlb->mm->nr_ptes);
1424 put_huge_zero_page();
1426 page = pmd_page(orig_pmd);
1427 page_remove_rmap(page);
1428 VM_BUG_ON(page_mapcount(page) < 0);
1429 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1430 VM_BUG_ON(!PageHead(page));
1431 atomic_long_dec(&tlb->mm->nr_ptes);
1433 tlb_remove_page(tlb, page);
1435 pte_free(tlb->mm, pgtable);
1441 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1442 unsigned long addr, unsigned long end,
1448 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1450 * All logical pages in the range are present
1451 * if backed by a huge page.
1454 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1461 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1462 unsigned long old_addr,
1463 unsigned long new_addr, unsigned long old_end,
1464 pmd_t *old_pmd, pmd_t *new_pmd)
1466 spinlock_t *old_ptl, *new_ptl;
1470 struct mm_struct *mm = vma->vm_mm;
1472 if ((old_addr & ~HPAGE_PMD_MASK) ||
1473 (new_addr & ~HPAGE_PMD_MASK) ||
1474 old_end - old_addr < HPAGE_PMD_SIZE ||
1475 (new_vma->vm_flags & VM_NOHUGEPAGE))
1479 * The destination pmd shouldn't be established, free_pgtables()
1480 * should have release it.
1482 if (WARN_ON(!pmd_none(*new_pmd))) {
1483 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1488 * We don't have to worry about the ordering of src and dst
1489 * ptlocks because exclusive mmap_sem prevents deadlock.
1491 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1493 new_ptl = pmd_lockptr(mm, new_pmd);
1494 if (new_ptl != old_ptl)
1495 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1496 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1497 VM_BUG_ON(!pmd_none(*new_pmd));
1498 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1499 if (new_ptl != old_ptl) {
1503 * Move preallocated PTE page table if new_pmd is on
1504 * different PMD page table.
1506 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1507 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1509 spin_unlock(new_ptl);
1511 spin_unlock(old_ptl);
1519 * - 0 if PMD could not be locked
1520 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1521 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1523 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1524 unsigned long addr, pgprot_t newprot, int prot_numa)
1526 struct mm_struct *mm = vma->vm_mm;
1530 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1534 entry = pmdp_get_and_clear(mm, addr, pmd);
1535 if (pmd_numa(entry))
1536 entry = pmd_mknonnuma(entry);
1537 entry = pmd_modify(entry, newprot);
1539 BUG_ON(pmd_write(entry));
1541 struct page *page = pmd_page(*pmd);
1544 * Do not trap faults against the zero page. The
1545 * read-only data is likely to be read-cached on the
1546 * local CPU cache and it is less useful to know about
1547 * local vs remote hits on the zero page.
1549 if (!is_huge_zero_page(page) &&
1552 entry = pmd_mknuma(entry);
1557 /* Set PMD if cleared earlier */
1558 if (ret == HPAGE_PMD_NR)
1559 set_pmd_at(mm, addr, pmd, entry);
1568 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1569 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1571 * Note that if it returns 1, this routine returns without unlocking page
1572 * table locks. So callers must unlock them.
1574 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1577 *ptl = pmd_lock(vma->vm_mm, pmd);
1578 if (likely(pmd_trans_huge(*pmd))) {
1579 if (unlikely(pmd_trans_splitting(*pmd))) {
1581 wait_split_huge_page(vma->anon_vma, pmd);
1584 /* Thp mapped by 'pmd' is stable, so we can
1585 * handle it as it is. */
1594 * This function returns whether a given @page is mapped onto the @address
1595 * in the virtual space of @mm.
1597 * When it's true, this function returns *pmd with holding the page table lock
1598 * and passing it back to the caller via @ptl.
1599 * If it's false, returns NULL without holding the page table lock.
1601 pmd_t *page_check_address_pmd(struct page *page,
1602 struct mm_struct *mm,
1603 unsigned long address,
1604 enum page_check_address_pmd_flag flag,
1609 if (address & ~HPAGE_PMD_MASK)
1612 pmd = mm_find_pmd(mm, address);
1615 *ptl = pmd_lock(mm, pmd);
1618 if (pmd_page(*pmd) != page)
1621 * split_vma() may create temporary aliased mappings. There is
1622 * no risk as long as all huge pmd are found and have their
1623 * splitting bit set before __split_huge_page_refcount
1624 * runs. Finding the same huge pmd more than once during the
1625 * same rmap walk is not a problem.
1627 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1628 pmd_trans_splitting(*pmd))
1630 if (pmd_trans_huge(*pmd)) {
1631 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1632 !pmd_trans_splitting(*pmd));
1640 static int __split_huge_page_splitting(struct page *page,
1641 struct vm_area_struct *vma,
1642 unsigned long address)
1644 struct mm_struct *mm = vma->vm_mm;
1648 /* For mmu_notifiers */
1649 const unsigned long mmun_start = address;
1650 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
1652 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1653 pmd = page_check_address_pmd(page, mm, address,
1654 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1657 * We can't temporarily set the pmd to null in order
1658 * to split it, the pmd must remain marked huge at all
1659 * times or the VM won't take the pmd_trans_huge paths
1660 * and it won't wait on the anon_vma->root->rwsem to
1661 * serialize against split_huge_page*.
1663 pmdp_splitting_flush(vma, address, pmd);
1667 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1672 static void __split_huge_page_refcount(struct page *page,
1673 struct list_head *list)
1676 struct zone *zone = page_zone(page);
1677 struct lruvec *lruvec;
1680 /* prevent PageLRU to go away from under us, and freeze lru stats */
1681 spin_lock_irq(&zone->lru_lock);
1682 lruvec = mem_cgroup_page_lruvec(page, zone);
1684 compound_lock(page);
1685 /* complete memcg works before add pages to LRU */
1686 mem_cgroup_split_huge_fixup(page);
1688 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1689 struct page *page_tail = page + i;
1691 /* tail_page->_mapcount cannot change */
1692 BUG_ON(page_mapcount(page_tail) < 0);
1693 tail_count += page_mapcount(page_tail);
1694 /* check for overflow */
1695 BUG_ON(tail_count < 0);
1696 BUG_ON(atomic_read(&page_tail->_count) != 0);
1698 * tail_page->_count is zero and not changing from
1699 * under us. But get_page_unless_zero() may be running
1700 * from under us on the tail_page. If we used
1701 * atomic_set() below instead of atomic_add(), we
1702 * would then run atomic_set() concurrently with
1703 * get_page_unless_zero(), and atomic_set() is
1704 * implemented in C not using locked ops. spin_unlock
1705 * on x86 sometime uses locked ops because of PPro
1706 * errata 66, 92, so unless somebody can guarantee
1707 * atomic_set() here would be safe on all archs (and
1708 * not only on x86), it's safer to use atomic_add().
1710 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1711 &page_tail->_count);
1713 /* after clearing PageTail the gup refcount can be released */
1717 * retain hwpoison flag of the poisoned tail page:
1718 * fix for the unsuitable process killed on Guest Machine(KVM)
1719 * by the memory-failure.
1721 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1722 page_tail->flags |= (page->flags &
1723 ((1L << PG_referenced) |
1724 (1L << PG_swapbacked) |
1725 (1L << PG_mlocked) |
1726 (1L << PG_uptodate) |
1728 (1L << PG_unevictable)));
1729 page_tail->flags |= (1L << PG_dirty);
1731 /* clear PageTail before overwriting first_page */
1735 * __split_huge_page_splitting() already set the
1736 * splitting bit in all pmd that could map this
1737 * hugepage, that will ensure no CPU can alter the
1738 * mapcount on the head page. The mapcount is only
1739 * accounted in the head page and it has to be
1740 * transferred to all tail pages in the below code. So
1741 * for this code to be safe, the split the mapcount
1742 * can't change. But that doesn't mean userland can't
1743 * keep changing and reading the page contents while
1744 * we transfer the mapcount, so the pmd splitting
1745 * status is achieved setting a reserved bit in the
1746 * pmd, not by clearing the present bit.
1748 page_tail->_mapcount = page->_mapcount;
1750 BUG_ON(page_tail->mapping);
1751 page_tail->mapping = page->mapping;
1753 page_tail->index = page->index + i;
1754 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1756 BUG_ON(!PageAnon(page_tail));
1757 BUG_ON(!PageUptodate(page_tail));
1758 BUG_ON(!PageDirty(page_tail));
1759 BUG_ON(!PageSwapBacked(page_tail));
1761 lru_add_page_tail(page, page_tail, lruvec, list);
1763 atomic_sub(tail_count, &page->_count);
1764 BUG_ON(atomic_read(&page->_count) <= 0);
1766 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1768 ClearPageCompound(page);
1769 compound_unlock(page);
1770 spin_unlock_irq(&zone->lru_lock);
1772 for (i = 1; i < HPAGE_PMD_NR; i++) {
1773 struct page *page_tail = page + i;
1774 BUG_ON(page_count(page_tail) <= 0);
1776 * Tail pages may be freed if there wasn't any mapping
1777 * like if add_to_swap() is running on a lru page that
1778 * had its mapping zapped. And freeing these pages
1779 * requires taking the lru_lock so we do the put_page
1780 * of the tail pages after the split is complete.
1782 put_page(page_tail);
1786 * Only the head page (now become a regular page) is required
1787 * to be pinned by the caller.
1789 BUG_ON(page_count(page) <= 0);
1792 static int __split_huge_page_map(struct page *page,
1793 struct vm_area_struct *vma,
1794 unsigned long address)
1796 struct mm_struct *mm = vma->vm_mm;
1801 unsigned long haddr;
1803 pmd = page_check_address_pmd(page, mm, address,
1804 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1806 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1807 pmd_populate(mm, &_pmd, pgtable);
1810 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1812 BUG_ON(PageCompound(page+i));
1813 entry = mk_pte(page + i, vma->vm_page_prot);
1814 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1815 if (!pmd_write(*pmd))
1816 entry = pte_wrprotect(entry);
1818 BUG_ON(page_mapcount(page) != 1);
1819 if (!pmd_young(*pmd))
1820 entry = pte_mkold(entry);
1822 entry = pte_mknuma(entry);
1823 pte = pte_offset_map(&_pmd, haddr);
1824 BUG_ON(!pte_none(*pte));
1825 set_pte_at(mm, haddr, pte, entry);
1829 smp_wmb(); /* make pte visible before pmd */
1831 * Up to this point the pmd is present and huge and
1832 * userland has the whole access to the hugepage
1833 * during the split (which happens in place). If we
1834 * overwrite the pmd with the not-huge version
1835 * pointing to the pte here (which of course we could
1836 * if all CPUs were bug free), userland could trigger
1837 * a small page size TLB miss on the small sized TLB
1838 * while the hugepage TLB entry is still established
1839 * in the huge TLB. Some CPU doesn't like that. See
1840 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1841 * Erratum 383 on page 93. Intel should be safe but is
1842 * also warns that it's only safe if the permission
1843 * and cache attributes of the two entries loaded in
1844 * the two TLB is identical (which should be the case
1845 * here). But it is generally safer to never allow
1846 * small and huge TLB entries for the same virtual
1847 * address to be loaded simultaneously. So instead of
1848 * doing "pmd_populate(); flush_tlb_range();" we first
1849 * mark the current pmd notpresent (atomically because
1850 * here the pmd_trans_huge and pmd_trans_splitting
1851 * must remain set at all times on the pmd until the
1852 * split is complete for this pmd), then we flush the
1853 * SMP TLB and finally we write the non-huge version
1854 * of the pmd entry with pmd_populate.
1856 pmdp_invalidate(vma, address, pmd);
1857 pmd_populate(mm, pmd, pgtable);
1865 /* must be called with anon_vma->root->rwsem held */
1866 static void __split_huge_page(struct page *page,
1867 struct anon_vma *anon_vma,
1868 struct list_head *list)
1870 int mapcount, mapcount2;
1871 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1872 struct anon_vma_chain *avc;
1874 BUG_ON(!PageHead(page));
1875 BUG_ON(PageTail(page));
1878 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1879 struct vm_area_struct *vma = avc->vma;
1880 unsigned long addr = vma_address(page, vma);
1881 BUG_ON(is_vma_temporary_stack(vma));
1882 mapcount += __split_huge_page_splitting(page, vma, addr);
1885 * It is critical that new vmas are added to the tail of the
1886 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1887 * and establishes a child pmd before
1888 * __split_huge_page_splitting() freezes the parent pmd (so if
1889 * we fail to prevent copy_huge_pmd() from running until the
1890 * whole __split_huge_page() is complete), we will still see
1891 * the newly established pmd of the child later during the
1892 * walk, to be able to set it as pmd_trans_splitting too.
1894 if (mapcount != page_mapcount(page))
1895 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1896 mapcount, page_mapcount(page));
1897 BUG_ON(mapcount != page_mapcount(page));
1899 __split_huge_page_refcount(page, list);
1902 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1903 struct vm_area_struct *vma = avc->vma;
1904 unsigned long addr = vma_address(page, vma);
1905 BUG_ON(is_vma_temporary_stack(vma));
1906 mapcount2 += __split_huge_page_map(page, vma, addr);
1908 if (mapcount != mapcount2)
1909 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1910 mapcount, mapcount2, page_mapcount(page));
1911 BUG_ON(mapcount != mapcount2);
1915 * Split a hugepage into normal pages. This doesn't change the position of head
1916 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1917 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1918 * from the hugepage.
1919 * Return 0 if the hugepage is split successfully otherwise return 1.
1921 int split_huge_page_to_list(struct page *page, struct list_head *list)
1923 struct anon_vma *anon_vma;
1926 BUG_ON(is_huge_zero_page(page));
1927 BUG_ON(!PageAnon(page));
1930 * The caller does not necessarily hold an mmap_sem that would prevent
1931 * the anon_vma disappearing so we first we take a reference to it
1932 * and then lock the anon_vma for write. This is similar to
1933 * page_lock_anon_vma_read except the write lock is taken to serialise
1934 * against parallel split or collapse operations.
1936 anon_vma = page_get_anon_vma(page);
1939 anon_vma_lock_write(anon_vma);
1942 if (!PageCompound(page))
1945 BUG_ON(!PageSwapBacked(page));
1946 __split_huge_page(page, anon_vma, list);
1947 count_vm_event(THP_SPLIT);
1949 BUG_ON(PageCompound(page));
1951 anon_vma_unlock_write(anon_vma);
1952 put_anon_vma(anon_vma);
1957 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1959 int hugepage_madvise(struct vm_area_struct *vma,
1960 unsigned long *vm_flags, int advice)
1962 struct mm_struct *mm = vma->vm_mm;
1967 * Be somewhat over-protective like KSM for now!
1969 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1971 if (mm->def_flags & VM_NOHUGEPAGE)
1973 *vm_flags &= ~VM_NOHUGEPAGE;
1974 *vm_flags |= VM_HUGEPAGE;
1976 * If the vma become good for khugepaged to scan,
1977 * register it here without waiting a page fault that
1978 * may not happen any time soon.
1980 if (unlikely(khugepaged_enter_vma_merge(vma)))
1983 case MADV_NOHUGEPAGE:
1985 * Be somewhat over-protective like KSM for now!
1987 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1989 *vm_flags &= ~VM_HUGEPAGE;
1990 *vm_flags |= VM_NOHUGEPAGE;
1992 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1993 * this vma even if we leave the mm registered in khugepaged if
1994 * it got registered before VM_NOHUGEPAGE was set.
2002 static int __init khugepaged_slab_init(void)
2004 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2005 sizeof(struct mm_slot),
2006 __alignof__(struct mm_slot), 0, NULL);
2013 static inline struct mm_slot *alloc_mm_slot(void)
2015 if (!mm_slot_cache) /* initialization failed */
2017 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2020 static inline void free_mm_slot(struct mm_slot *mm_slot)
2022 kmem_cache_free(mm_slot_cache, mm_slot);
2025 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2027 struct mm_slot *mm_slot;
2029 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2030 if (mm == mm_slot->mm)
2036 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2037 struct mm_slot *mm_slot)
2040 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2043 static inline int khugepaged_test_exit(struct mm_struct *mm)
2045 return atomic_read(&mm->mm_users) == 0;
2048 int __khugepaged_enter(struct mm_struct *mm)
2050 struct mm_slot *mm_slot;
2053 mm_slot = alloc_mm_slot();
2057 /* __khugepaged_exit() must not run from under us */
2058 VM_BUG_ON(khugepaged_test_exit(mm));
2059 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2060 free_mm_slot(mm_slot);
2064 spin_lock(&khugepaged_mm_lock);
2065 insert_to_mm_slots_hash(mm, mm_slot);
2067 * Insert just behind the scanning cursor, to let the area settle
2070 wakeup = list_empty(&khugepaged_scan.mm_head);
2071 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2072 spin_unlock(&khugepaged_mm_lock);
2074 atomic_inc(&mm->mm_count);
2076 wake_up_interruptible(&khugepaged_wait);
2081 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2083 unsigned long hstart, hend;
2086 * Not yet faulted in so we will register later in the
2087 * page fault if needed.
2091 /* khugepaged not yet working on file or special mappings */
2093 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2094 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2095 hend = vma->vm_end & HPAGE_PMD_MASK;
2097 return khugepaged_enter(vma);
2101 void __khugepaged_exit(struct mm_struct *mm)
2103 struct mm_slot *mm_slot;
2106 spin_lock(&khugepaged_mm_lock);
2107 mm_slot = get_mm_slot(mm);
2108 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2109 hash_del(&mm_slot->hash);
2110 list_del(&mm_slot->mm_node);
2113 spin_unlock(&khugepaged_mm_lock);
2116 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2117 free_mm_slot(mm_slot);
2119 } else if (mm_slot) {
2121 * This is required to serialize against
2122 * khugepaged_test_exit() (which is guaranteed to run
2123 * under mmap sem read mode). Stop here (after we
2124 * return all pagetables will be destroyed) until
2125 * khugepaged has finished working on the pagetables
2126 * under the mmap_sem.
2128 down_write(&mm->mmap_sem);
2129 up_write(&mm->mmap_sem);
2133 static void release_pte_page(struct page *page)
2135 /* 0 stands for page_is_file_cache(page) == false */
2136 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2138 putback_lru_page(page);
2141 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2143 while (--_pte >= pte) {
2144 pte_t pteval = *_pte;
2145 if (!pte_none(pteval))
2146 release_pte_page(pte_page(pteval));
2150 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2151 unsigned long address,
2156 int referenced = 0, none = 0;
2157 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2158 _pte++, address += PAGE_SIZE) {
2159 pte_t pteval = *_pte;
2160 if (pte_none(pteval)) {
2161 if (++none <= khugepaged_max_ptes_none)
2166 if (!pte_present(pteval) || !pte_write(pteval))
2168 page = vm_normal_page(vma, address, pteval);
2169 if (unlikely(!page))
2172 VM_BUG_ON(PageCompound(page));
2173 BUG_ON(!PageAnon(page));
2174 VM_BUG_ON(!PageSwapBacked(page));
2176 /* cannot use mapcount: can't collapse if there's a gup pin */
2177 if (page_count(page) != 1)
2180 * We can do it before isolate_lru_page because the
2181 * page can't be freed from under us. NOTE: PG_lock
2182 * is needed to serialize against split_huge_page
2183 * when invoked from the VM.
2185 if (!trylock_page(page))
2188 * Isolate the page to avoid collapsing an hugepage
2189 * currently in use by the VM.
2191 if (isolate_lru_page(page)) {
2195 /* 0 stands for page_is_file_cache(page) == false */
2196 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2197 VM_BUG_ON(!PageLocked(page));
2198 VM_BUG_ON(PageLRU(page));
2200 /* If there is no mapped pte young don't collapse the page */
2201 if (pte_young(pteval) || PageReferenced(page) ||
2202 mmu_notifier_test_young(vma->vm_mm, address))
2205 if (likely(referenced))
2208 release_pte_pages(pte, _pte);
2212 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2213 struct vm_area_struct *vma,
2214 unsigned long address,
2218 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2219 pte_t pteval = *_pte;
2220 struct page *src_page;
2222 if (pte_none(pteval)) {
2223 clear_user_highpage(page, address);
2224 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2226 src_page = pte_page(pteval);
2227 copy_user_highpage(page, src_page, address, vma);
2228 VM_BUG_ON(page_mapcount(src_page) != 1);
2229 release_pte_page(src_page);
2231 * ptl mostly unnecessary, but preempt has to
2232 * be disabled to update the per-cpu stats
2233 * inside page_remove_rmap().
2237 * paravirt calls inside pte_clear here are
2240 pte_clear(vma->vm_mm, address, _pte);
2241 page_remove_rmap(src_page);
2243 free_page_and_swap_cache(src_page);
2246 address += PAGE_SIZE;
2251 static void khugepaged_alloc_sleep(void)
2253 wait_event_freezable_timeout(khugepaged_wait, false,
2254 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2257 static int khugepaged_node_load[MAX_NUMNODES];
2260 static int khugepaged_find_target_node(void)
2262 static int last_khugepaged_target_node = NUMA_NO_NODE;
2263 int nid, target_node = 0, max_value = 0;
2265 /* find first node with max normal pages hit */
2266 for (nid = 0; nid < MAX_NUMNODES; nid++)
2267 if (khugepaged_node_load[nid] > max_value) {
2268 max_value = khugepaged_node_load[nid];
2272 /* do some balance if several nodes have the same hit record */
2273 if (target_node <= last_khugepaged_target_node)
2274 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2276 if (max_value == khugepaged_node_load[nid]) {
2281 last_khugepaged_target_node = target_node;
2285 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2287 if (IS_ERR(*hpage)) {
2293 khugepaged_alloc_sleep();
2294 } else if (*hpage) {
2303 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2304 struct vm_area_struct *vma, unsigned long address,
2309 * Allocate the page while the vma is still valid and under
2310 * the mmap_sem read mode so there is no memory allocation
2311 * later when we take the mmap_sem in write mode. This is more
2312 * friendly behavior (OTOH it may actually hide bugs) to
2313 * filesystems in userland with daemons allocating memory in
2314 * the userland I/O paths. Allocating memory with the
2315 * mmap_sem in read mode is good idea also to allow greater
2318 *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2319 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
2321 * After allocating the hugepage, release the mmap_sem read lock in
2322 * preparation for taking it in write mode.
2324 up_read(&mm->mmap_sem);
2325 if (unlikely(!*hpage)) {
2326 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2327 *hpage = ERR_PTR(-ENOMEM);
2331 count_vm_event(THP_COLLAPSE_ALLOC);
2335 static int khugepaged_find_target_node(void)
2340 static inline struct page *alloc_hugepage(int defrag)
2342 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2346 static struct page *khugepaged_alloc_hugepage(bool *wait)
2351 hpage = alloc_hugepage(khugepaged_defrag());
2353 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2358 khugepaged_alloc_sleep();
2360 count_vm_event(THP_COLLAPSE_ALLOC);
2361 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2366 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2369 *hpage = khugepaged_alloc_hugepage(wait);
2371 if (unlikely(!*hpage))
2378 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2379 struct vm_area_struct *vma, unsigned long address,
2382 up_read(&mm->mmap_sem);
2388 static bool hugepage_vma_check(struct vm_area_struct *vma)
2390 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2391 (vma->vm_flags & VM_NOHUGEPAGE))
2394 if (!vma->anon_vma || vma->vm_ops)
2396 if (is_vma_temporary_stack(vma))
2398 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2402 static void collapse_huge_page(struct mm_struct *mm,
2403 unsigned long address,
2404 struct page **hpage,
2405 struct vm_area_struct *vma,
2411 struct page *new_page;
2412 spinlock_t *pmd_ptl, *pte_ptl;
2414 unsigned long hstart, hend;
2415 unsigned long mmun_start; /* For mmu_notifiers */
2416 unsigned long mmun_end; /* For mmu_notifiers */
2418 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2420 /* release the mmap_sem read lock. */
2421 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2425 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2429 * Prevent all access to pagetables with the exception of
2430 * gup_fast later hanlded by the ptep_clear_flush and the VM
2431 * handled by the anon_vma lock + PG_lock.
2433 down_write(&mm->mmap_sem);
2434 if (unlikely(khugepaged_test_exit(mm)))
2437 vma = find_vma(mm, address);
2440 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2441 hend = vma->vm_end & HPAGE_PMD_MASK;
2442 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2444 if (!hugepage_vma_check(vma))
2446 pmd = mm_find_pmd(mm, address);
2449 if (pmd_trans_huge(*pmd))
2452 anon_vma_lock_write(vma->anon_vma);
2454 pte = pte_offset_map(pmd, address);
2455 pte_ptl = pte_lockptr(mm, pmd);
2457 mmun_start = address;
2458 mmun_end = address + HPAGE_PMD_SIZE;
2459 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2460 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2462 * After this gup_fast can't run anymore. This also removes
2463 * any huge TLB entry from the CPU so we won't allow
2464 * huge and small TLB entries for the same virtual address
2465 * to avoid the risk of CPU bugs in that area.
2467 _pmd = pmdp_clear_flush(vma, address, pmd);
2468 spin_unlock(pmd_ptl);
2469 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2472 isolated = __collapse_huge_page_isolate(vma, address, pte);
2473 spin_unlock(pte_ptl);
2475 if (unlikely(!isolated)) {
2478 BUG_ON(!pmd_none(*pmd));
2480 * We can only use set_pmd_at when establishing
2481 * hugepmds and never for establishing regular pmds that
2482 * points to regular pagetables. Use pmd_populate for that
2484 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2485 spin_unlock(pmd_ptl);
2486 anon_vma_unlock_write(vma->anon_vma);
2491 * All pages are isolated and locked so anon_vma rmap
2492 * can't run anymore.
2494 anon_vma_unlock_write(vma->anon_vma);
2496 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2498 __SetPageUptodate(new_page);
2499 pgtable = pmd_pgtable(_pmd);
2501 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2502 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2505 * spin_lock() below is not the equivalent of smp_wmb(), so
2506 * this is needed to avoid the copy_huge_page writes to become
2507 * visible after the set_pmd_at() write.
2512 BUG_ON(!pmd_none(*pmd));
2513 page_add_new_anon_rmap(new_page, vma, address);
2514 pgtable_trans_huge_deposit(mm, pmd, pgtable);
2515 set_pmd_at(mm, address, pmd, _pmd);
2516 update_mmu_cache_pmd(vma, address, pmd);
2517 spin_unlock(pmd_ptl);
2521 khugepaged_pages_collapsed++;
2523 up_write(&mm->mmap_sem);
2527 mem_cgroup_uncharge_page(new_page);
2531 static int khugepaged_scan_pmd(struct mm_struct *mm,
2532 struct vm_area_struct *vma,
2533 unsigned long address,
2534 struct page **hpage)
2538 int ret = 0, referenced = 0, none = 0;
2540 unsigned long _address;
2542 int node = NUMA_NO_NODE;
2544 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2546 pmd = mm_find_pmd(mm, address);
2549 if (pmd_trans_huge(*pmd))
2552 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2553 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2554 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2555 _pte++, _address += PAGE_SIZE) {
2556 pte_t pteval = *_pte;
2557 if (pte_none(pteval)) {
2558 if (++none <= khugepaged_max_ptes_none)
2563 if (!pte_present(pteval) || !pte_write(pteval))
2565 page = vm_normal_page(vma, _address, pteval);
2566 if (unlikely(!page))
2569 * Record which node the original page is from and save this
2570 * information to khugepaged_node_load[].
2571 * Khupaged will allocate hugepage from the node has the max
2574 node = page_to_nid(page);
2575 khugepaged_node_load[node]++;
2576 VM_BUG_ON(PageCompound(page));
2577 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2579 /* cannot use mapcount: can't collapse if there's a gup pin */
2580 if (page_count(page) != 1)
2582 if (pte_young(pteval) || PageReferenced(page) ||
2583 mmu_notifier_test_young(vma->vm_mm, address))
2589 pte_unmap_unlock(pte, ptl);
2591 node = khugepaged_find_target_node();
2592 /* collapse_huge_page will return with the mmap_sem released */
2593 collapse_huge_page(mm, address, hpage, vma, node);
2599 static void collect_mm_slot(struct mm_slot *mm_slot)
2601 struct mm_struct *mm = mm_slot->mm;
2603 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2605 if (khugepaged_test_exit(mm)) {
2607 hash_del(&mm_slot->hash);
2608 list_del(&mm_slot->mm_node);
2611 * Not strictly needed because the mm exited already.
2613 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2616 /* khugepaged_mm_lock actually not necessary for the below */
2617 free_mm_slot(mm_slot);
2622 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2623 struct page **hpage)
2624 __releases(&khugepaged_mm_lock)
2625 __acquires(&khugepaged_mm_lock)
2627 struct mm_slot *mm_slot;
2628 struct mm_struct *mm;
2629 struct vm_area_struct *vma;
2633 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2635 if (khugepaged_scan.mm_slot)
2636 mm_slot = khugepaged_scan.mm_slot;
2638 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2639 struct mm_slot, mm_node);
2640 khugepaged_scan.address = 0;
2641 khugepaged_scan.mm_slot = mm_slot;
2643 spin_unlock(&khugepaged_mm_lock);
2646 down_read(&mm->mmap_sem);
2647 if (unlikely(khugepaged_test_exit(mm)))
2650 vma = find_vma(mm, khugepaged_scan.address);
2653 for (; vma; vma = vma->vm_next) {
2654 unsigned long hstart, hend;
2657 if (unlikely(khugepaged_test_exit(mm))) {
2661 if (!hugepage_vma_check(vma)) {
2666 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2667 hend = vma->vm_end & HPAGE_PMD_MASK;
2670 if (khugepaged_scan.address > hend)
2672 if (khugepaged_scan.address < hstart)
2673 khugepaged_scan.address = hstart;
2674 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2676 while (khugepaged_scan.address < hend) {
2679 if (unlikely(khugepaged_test_exit(mm)))
2680 goto breakouterloop;
2682 VM_BUG_ON(khugepaged_scan.address < hstart ||
2683 khugepaged_scan.address + HPAGE_PMD_SIZE >
2685 ret = khugepaged_scan_pmd(mm, vma,
2686 khugepaged_scan.address,
2688 /* move to next address */
2689 khugepaged_scan.address += HPAGE_PMD_SIZE;
2690 progress += HPAGE_PMD_NR;
2692 /* we released mmap_sem so break loop */
2693 goto breakouterloop_mmap_sem;
2694 if (progress >= pages)
2695 goto breakouterloop;
2699 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2700 breakouterloop_mmap_sem:
2702 spin_lock(&khugepaged_mm_lock);
2703 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2705 * Release the current mm_slot if this mm is about to die, or
2706 * if we scanned all vmas of this mm.
2708 if (khugepaged_test_exit(mm) || !vma) {
2710 * Make sure that if mm_users is reaching zero while
2711 * khugepaged runs here, khugepaged_exit will find
2712 * mm_slot not pointing to the exiting mm.
2714 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2715 khugepaged_scan.mm_slot = list_entry(
2716 mm_slot->mm_node.next,
2717 struct mm_slot, mm_node);
2718 khugepaged_scan.address = 0;
2720 khugepaged_scan.mm_slot = NULL;
2721 khugepaged_full_scans++;
2724 collect_mm_slot(mm_slot);
2730 static int khugepaged_has_work(void)
2732 return !list_empty(&khugepaged_scan.mm_head) &&
2733 khugepaged_enabled();
2736 static int khugepaged_wait_event(void)
2738 return !list_empty(&khugepaged_scan.mm_head) ||
2739 kthread_should_stop();
2742 static void khugepaged_do_scan(void)
2744 struct page *hpage = NULL;
2745 unsigned int progress = 0, pass_through_head = 0;
2746 unsigned int pages = khugepaged_pages_to_scan;
2749 barrier(); /* write khugepaged_pages_to_scan to local stack */
2751 while (progress < pages) {
2752 if (!khugepaged_prealloc_page(&hpage, &wait))
2757 if (unlikely(kthread_should_stop() || freezing(current)))
2760 spin_lock(&khugepaged_mm_lock);
2761 if (!khugepaged_scan.mm_slot)
2762 pass_through_head++;
2763 if (khugepaged_has_work() &&
2764 pass_through_head < 2)
2765 progress += khugepaged_scan_mm_slot(pages - progress,
2769 spin_unlock(&khugepaged_mm_lock);
2772 if (!IS_ERR_OR_NULL(hpage))
2776 static void khugepaged_wait_work(void)
2780 if (khugepaged_has_work()) {
2781 if (!khugepaged_scan_sleep_millisecs)
2784 wait_event_freezable_timeout(khugepaged_wait,
2785 kthread_should_stop(),
2786 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2790 if (khugepaged_enabled())
2791 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2794 static int khugepaged(void *none)
2796 struct mm_slot *mm_slot;
2799 set_user_nice(current, 19);
2801 while (!kthread_should_stop()) {
2802 khugepaged_do_scan();
2803 khugepaged_wait_work();
2806 spin_lock(&khugepaged_mm_lock);
2807 mm_slot = khugepaged_scan.mm_slot;
2808 khugepaged_scan.mm_slot = NULL;
2810 collect_mm_slot(mm_slot);
2811 spin_unlock(&khugepaged_mm_lock);
2815 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2816 unsigned long haddr, pmd_t *pmd)
2818 struct mm_struct *mm = vma->vm_mm;
2823 pmdp_clear_flush(vma, haddr, pmd);
2824 /* leave pmd empty until pte is filled */
2826 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2827 pmd_populate(mm, &_pmd, pgtable);
2829 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2831 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2832 entry = pte_mkspecial(entry);
2833 pte = pte_offset_map(&_pmd, haddr);
2834 VM_BUG_ON(!pte_none(*pte));
2835 set_pte_at(mm, haddr, pte, entry);
2838 smp_wmb(); /* make pte visible before pmd */
2839 pmd_populate(mm, pmd, pgtable);
2840 put_huge_zero_page();
2843 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2848 struct mm_struct *mm = vma->vm_mm;
2849 unsigned long haddr = address & HPAGE_PMD_MASK;
2850 unsigned long mmun_start; /* For mmu_notifiers */
2851 unsigned long mmun_end; /* For mmu_notifiers */
2853 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2856 mmun_end = haddr + HPAGE_PMD_SIZE;
2858 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2859 ptl = pmd_lock(mm, pmd);
2860 if (unlikely(!pmd_trans_huge(*pmd))) {
2862 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2865 if (is_huge_zero_pmd(*pmd)) {
2866 __split_huge_zero_page_pmd(vma, haddr, pmd);
2868 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2871 page = pmd_page(*pmd);
2872 VM_BUG_ON(!page_count(page));
2875 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2877 split_huge_page(page);
2882 * We don't always have down_write of mmap_sem here: a racing
2883 * do_huge_pmd_wp_page() might have copied-on-write to another
2884 * huge page before our split_huge_page() got the anon_vma lock.
2886 if (unlikely(pmd_trans_huge(*pmd)))
2890 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2893 struct vm_area_struct *vma;
2895 vma = find_vma(mm, address);
2896 BUG_ON(vma == NULL);
2897 split_huge_page_pmd(vma, address, pmd);
2900 static void split_huge_page_address(struct mm_struct *mm,
2901 unsigned long address)
2905 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2907 pmd = mm_find_pmd(mm, address);
2911 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2912 * materialize from under us.
2914 split_huge_page_pmd_mm(mm, address, pmd);
2917 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2918 unsigned long start,
2923 * If the new start address isn't hpage aligned and it could
2924 * previously contain an hugepage: check if we need to split
2927 if (start & ~HPAGE_PMD_MASK &&
2928 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2929 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2930 split_huge_page_address(vma->vm_mm, start);
2933 * If the new end address isn't hpage aligned and it could
2934 * previously contain an hugepage: check if we need to split
2937 if (end & ~HPAGE_PMD_MASK &&
2938 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2939 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2940 split_huge_page_address(vma->vm_mm, end);
2943 * If we're also updating the vma->vm_next->vm_start, if the new
2944 * vm_next->vm_start isn't page aligned and it could previously
2945 * contain an hugepage: check if we need to split an huge pmd.
2947 if (adjust_next > 0) {
2948 struct vm_area_struct *next = vma->vm_next;
2949 unsigned long nstart = next->vm_start;
2950 nstart += adjust_next << PAGE_SHIFT;
2951 if (nstart & ~HPAGE_PMD_MASK &&
2952 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2953 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2954 split_huge_page_address(next->vm_mm, nstart);