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/mm_inline.h>
16 #include <linux/kthread.h>
17 #include <linux/khugepaged.h>
18 #include <linux/freezer.h>
19 #include <linux/mman.h>
21 #include <asm/pgalloc.h>
25 * By default transparent hugepage support is enabled for all mappings
26 * and khugepaged scans all mappings. Defrag is only invoked by
27 * khugepaged hugepage allocations and by page faults inside
28 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
31 unsigned long transparent_hugepage_flags __read_mostly =
32 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
33 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
35 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
36 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
38 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
39 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
41 /* default scan 8*512 pte (or vmas) every 30 second */
42 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
43 static unsigned int khugepaged_pages_collapsed;
44 static unsigned int khugepaged_full_scans;
45 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
46 /* during fragmentation poll the hugepage allocator once every minute */
47 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
48 static struct task_struct *khugepaged_thread __read_mostly;
49 static DEFINE_MUTEX(khugepaged_mutex);
50 static DEFINE_SPINLOCK(khugepaged_mm_lock);
51 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
53 * default collapse hugepages if there is at least one pte mapped like
54 * it would have happened if the vma was large enough during page
57 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
59 static int khugepaged(void *none);
60 static int mm_slots_hash_init(void);
61 static int khugepaged_slab_init(void);
62 static void khugepaged_slab_free(void);
64 #define MM_SLOTS_HASH_HEADS 1024
65 static struct hlist_head *mm_slots_hash __read_mostly;
66 static struct kmem_cache *mm_slot_cache __read_mostly;
69 * struct mm_slot - hash lookup from mm to mm_slot
70 * @hash: hash collision list
71 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
72 * @mm: the mm that this information is valid for
75 struct hlist_node hash;
76 struct list_head mm_node;
81 * struct khugepaged_scan - cursor for scanning
82 * @mm_head: the head of the mm list to scan
83 * @mm_slot: the current mm_slot we are scanning
84 * @address: the next address inside that to be scanned
86 * There is only the one khugepaged_scan instance of this cursor structure.
88 struct khugepaged_scan {
89 struct list_head mm_head;
90 struct mm_slot *mm_slot;
91 unsigned long address;
93 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
97 static int set_recommended_min_free_kbytes(void)
101 unsigned long recommended_min;
102 extern int min_free_kbytes;
104 if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG,
105 &transparent_hugepage_flags) &&
106 !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
107 &transparent_hugepage_flags))
110 for_each_populated_zone(zone)
113 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
114 recommended_min = pageblock_nr_pages * nr_zones * 2;
117 * Make sure that on average at least two pageblocks are almost free
118 * of another type, one for a migratetype to fall back to and a
119 * second to avoid subsequent fallbacks of other types There are 3
120 * MIGRATE_TYPES we care about.
122 recommended_min += pageblock_nr_pages * nr_zones *
123 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
125 /* don't ever allow to reserve more than 5% of the lowmem */
126 recommended_min = min(recommended_min,
127 (unsigned long) nr_free_buffer_pages() / 20);
128 recommended_min <<= (PAGE_SHIFT-10);
130 if (recommended_min > min_free_kbytes)
131 min_free_kbytes = recommended_min;
132 setup_per_zone_wmarks();
135 late_initcall(set_recommended_min_free_kbytes);
137 static int start_khugepaged(void)
140 if (khugepaged_enabled()) {
142 if (unlikely(!mm_slot_cache || !mm_slots_hash)) {
146 mutex_lock(&khugepaged_mutex);
147 if (!khugepaged_thread)
148 khugepaged_thread = kthread_run(khugepaged, NULL,
150 if (unlikely(IS_ERR(khugepaged_thread))) {
152 "khugepaged: kthread_run(khugepaged) failed\n");
153 err = PTR_ERR(khugepaged_thread);
154 khugepaged_thread = NULL;
156 wakeup = !list_empty(&khugepaged_scan.mm_head);
157 mutex_unlock(&khugepaged_mutex);
159 wake_up_interruptible(&khugepaged_wait);
161 set_recommended_min_free_kbytes();
164 wake_up_interruptible(&khugepaged_wait);
171 static ssize_t double_flag_show(struct kobject *kobj,
172 struct kobj_attribute *attr, char *buf,
173 enum transparent_hugepage_flag enabled,
174 enum transparent_hugepage_flag req_madv)
176 if (test_bit(enabled, &transparent_hugepage_flags)) {
177 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
178 return sprintf(buf, "[always] madvise never\n");
179 } else if (test_bit(req_madv, &transparent_hugepage_flags))
180 return sprintf(buf, "always [madvise] never\n");
182 return sprintf(buf, "always madvise [never]\n");
184 static ssize_t double_flag_store(struct kobject *kobj,
185 struct kobj_attribute *attr,
186 const char *buf, size_t count,
187 enum transparent_hugepage_flag enabled,
188 enum transparent_hugepage_flag req_madv)
190 if (!memcmp("always", buf,
191 min(sizeof("always")-1, count))) {
192 set_bit(enabled, &transparent_hugepage_flags);
193 clear_bit(req_madv, &transparent_hugepage_flags);
194 } else if (!memcmp("madvise", buf,
195 min(sizeof("madvise")-1, count))) {
196 clear_bit(enabled, &transparent_hugepage_flags);
197 set_bit(req_madv, &transparent_hugepage_flags);
198 } else if (!memcmp("never", buf,
199 min(sizeof("never")-1, count))) {
200 clear_bit(enabled, &transparent_hugepage_flags);
201 clear_bit(req_madv, &transparent_hugepage_flags);
208 static ssize_t enabled_show(struct kobject *kobj,
209 struct kobj_attribute *attr, char *buf)
211 return double_flag_show(kobj, attr, buf,
212 TRANSPARENT_HUGEPAGE_FLAG,
213 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
215 static ssize_t enabled_store(struct kobject *kobj,
216 struct kobj_attribute *attr,
217 const char *buf, size_t count)
221 ret = double_flag_store(kobj, attr, buf, count,
222 TRANSPARENT_HUGEPAGE_FLAG,
223 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
226 int err = start_khugepaged();
232 (test_bit(TRANSPARENT_HUGEPAGE_FLAG,
233 &transparent_hugepage_flags) ||
234 test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
235 &transparent_hugepage_flags)))
236 set_recommended_min_free_kbytes();
240 static struct kobj_attribute enabled_attr =
241 __ATTR(enabled, 0644, enabled_show, enabled_store);
243 static ssize_t single_flag_show(struct kobject *kobj,
244 struct kobj_attribute *attr, char *buf,
245 enum transparent_hugepage_flag flag)
247 return sprintf(buf, "%d\n",
248 !!test_bit(flag, &transparent_hugepage_flags));
251 static ssize_t single_flag_store(struct kobject *kobj,
252 struct kobj_attribute *attr,
253 const char *buf, size_t count,
254 enum transparent_hugepage_flag flag)
259 ret = kstrtoul(buf, 10, &value);
266 set_bit(flag, &transparent_hugepage_flags);
268 clear_bit(flag, &transparent_hugepage_flags);
274 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
275 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
276 * memory just to allocate one more hugepage.
278 static ssize_t defrag_show(struct kobject *kobj,
279 struct kobj_attribute *attr, char *buf)
281 return double_flag_show(kobj, attr, buf,
282 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
283 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
285 static ssize_t defrag_store(struct kobject *kobj,
286 struct kobj_attribute *attr,
287 const char *buf, size_t count)
289 return double_flag_store(kobj, attr, buf, count,
290 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
291 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
293 static struct kobj_attribute defrag_attr =
294 __ATTR(defrag, 0644, defrag_show, defrag_store);
296 #ifdef CONFIG_DEBUG_VM
297 static ssize_t debug_cow_show(struct kobject *kobj,
298 struct kobj_attribute *attr, char *buf)
300 return single_flag_show(kobj, attr, buf,
301 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
303 static ssize_t debug_cow_store(struct kobject *kobj,
304 struct kobj_attribute *attr,
305 const char *buf, size_t count)
307 return single_flag_store(kobj, attr, buf, count,
308 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
310 static struct kobj_attribute debug_cow_attr =
311 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
312 #endif /* CONFIG_DEBUG_VM */
314 static struct attribute *hugepage_attr[] = {
317 #ifdef CONFIG_DEBUG_VM
318 &debug_cow_attr.attr,
323 static struct attribute_group hugepage_attr_group = {
324 .attrs = hugepage_attr,
327 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
328 struct kobj_attribute *attr,
331 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
334 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
335 struct kobj_attribute *attr,
336 const char *buf, size_t count)
341 err = strict_strtoul(buf, 10, &msecs);
342 if (err || msecs > UINT_MAX)
345 khugepaged_scan_sleep_millisecs = msecs;
346 wake_up_interruptible(&khugepaged_wait);
350 static struct kobj_attribute scan_sleep_millisecs_attr =
351 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
352 scan_sleep_millisecs_store);
354 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
355 struct kobj_attribute *attr,
358 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
361 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
362 struct kobj_attribute *attr,
363 const char *buf, size_t count)
368 err = strict_strtoul(buf, 10, &msecs);
369 if (err || msecs > UINT_MAX)
372 khugepaged_alloc_sleep_millisecs = msecs;
373 wake_up_interruptible(&khugepaged_wait);
377 static struct kobj_attribute alloc_sleep_millisecs_attr =
378 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
379 alloc_sleep_millisecs_store);
381 static ssize_t pages_to_scan_show(struct kobject *kobj,
382 struct kobj_attribute *attr,
385 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
387 static ssize_t pages_to_scan_store(struct kobject *kobj,
388 struct kobj_attribute *attr,
389 const char *buf, size_t count)
394 err = strict_strtoul(buf, 10, &pages);
395 if (err || !pages || pages > UINT_MAX)
398 khugepaged_pages_to_scan = pages;
402 static struct kobj_attribute pages_to_scan_attr =
403 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
404 pages_to_scan_store);
406 static ssize_t pages_collapsed_show(struct kobject *kobj,
407 struct kobj_attribute *attr,
410 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
412 static struct kobj_attribute pages_collapsed_attr =
413 __ATTR_RO(pages_collapsed);
415 static ssize_t full_scans_show(struct kobject *kobj,
416 struct kobj_attribute *attr,
419 return sprintf(buf, "%u\n", khugepaged_full_scans);
421 static struct kobj_attribute full_scans_attr =
422 __ATTR_RO(full_scans);
424 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
425 struct kobj_attribute *attr, char *buf)
427 return single_flag_show(kobj, attr, buf,
428 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
430 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
431 struct kobj_attribute *attr,
432 const char *buf, size_t count)
434 return single_flag_store(kobj, attr, buf, count,
435 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
437 static struct kobj_attribute khugepaged_defrag_attr =
438 __ATTR(defrag, 0644, khugepaged_defrag_show,
439 khugepaged_defrag_store);
442 * max_ptes_none controls if khugepaged should collapse hugepages over
443 * any unmapped ptes in turn potentially increasing the memory
444 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
445 * reduce the available free memory in the system as it
446 * runs. Increasing max_ptes_none will instead potentially reduce the
447 * free memory in the system during the khugepaged scan.
449 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
450 struct kobj_attribute *attr,
453 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
455 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
456 struct kobj_attribute *attr,
457 const char *buf, size_t count)
460 unsigned long max_ptes_none;
462 err = strict_strtoul(buf, 10, &max_ptes_none);
463 if (err || max_ptes_none > HPAGE_PMD_NR-1)
466 khugepaged_max_ptes_none = max_ptes_none;
470 static struct kobj_attribute khugepaged_max_ptes_none_attr =
471 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
472 khugepaged_max_ptes_none_store);
474 static struct attribute *khugepaged_attr[] = {
475 &khugepaged_defrag_attr.attr,
476 &khugepaged_max_ptes_none_attr.attr,
477 &pages_to_scan_attr.attr,
478 &pages_collapsed_attr.attr,
479 &full_scans_attr.attr,
480 &scan_sleep_millisecs_attr.attr,
481 &alloc_sleep_millisecs_attr.attr,
485 static struct attribute_group khugepaged_attr_group = {
486 .attrs = khugepaged_attr,
487 .name = "khugepaged",
489 #endif /* CONFIG_SYSFS */
491 static int __init hugepage_init(void)
495 static struct kobject *hugepage_kobj;
499 if (!has_transparent_hugepage()) {
500 transparent_hugepage_flags = 0;
506 hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
507 if (unlikely(!hugepage_kobj)) {
508 printk(KERN_ERR "hugepage: failed kobject create\n");
512 err = sysfs_create_group(hugepage_kobj, &hugepage_attr_group);
514 printk(KERN_ERR "hugepage: failed register hugeage group\n");
518 err = sysfs_create_group(hugepage_kobj, &khugepaged_attr_group);
520 printk(KERN_ERR "hugepage: failed register hugeage group\n");
525 err = khugepaged_slab_init();
529 err = mm_slots_hash_init();
531 khugepaged_slab_free();
536 * By default disable transparent hugepages on smaller systems,
537 * where the extra memory used could hurt more than TLB overhead
538 * is likely to save. The admin can still enable it through /sys.
540 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
541 transparent_hugepage_flags = 0;
545 set_recommended_min_free_kbytes();
550 module_init(hugepage_init)
552 static int __init setup_transparent_hugepage(char *str)
557 if (!strcmp(str, "always")) {
558 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
559 &transparent_hugepage_flags);
560 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
561 &transparent_hugepage_flags);
563 } else if (!strcmp(str, "madvise")) {
564 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
565 &transparent_hugepage_flags);
566 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
567 &transparent_hugepage_flags);
569 } else if (!strcmp(str, "never")) {
570 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
571 &transparent_hugepage_flags);
572 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
573 &transparent_hugepage_flags);
579 "transparent_hugepage= cannot parse, ignored\n");
582 __setup("transparent_hugepage=", setup_transparent_hugepage);
584 static void prepare_pmd_huge_pte(pgtable_t pgtable,
585 struct mm_struct *mm)
587 assert_spin_locked(&mm->page_table_lock);
590 if (!mm->pmd_huge_pte)
591 INIT_LIST_HEAD(&pgtable->lru);
593 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
594 mm->pmd_huge_pte = pgtable;
597 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
599 if (likely(vma->vm_flags & VM_WRITE))
600 pmd = pmd_mkwrite(pmd);
604 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
605 struct vm_area_struct *vma,
606 unsigned long haddr, pmd_t *pmd,
612 VM_BUG_ON(!PageCompound(page));
613 pgtable = pte_alloc_one(mm, haddr);
614 if (unlikely(!pgtable)) {
615 mem_cgroup_uncharge_page(page);
620 clear_huge_page(page, haddr, HPAGE_PMD_NR);
621 __SetPageUptodate(page);
623 spin_lock(&mm->page_table_lock);
624 if (unlikely(!pmd_none(*pmd))) {
625 spin_unlock(&mm->page_table_lock);
626 mem_cgroup_uncharge_page(page);
628 pte_free(mm, pgtable);
631 entry = mk_pmd(page, vma->vm_page_prot);
632 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
633 entry = pmd_mkhuge(entry);
635 * The spinlocking to take the lru_lock inside
636 * page_add_new_anon_rmap() acts as a full memory
637 * barrier to be sure clear_huge_page writes become
638 * visible after the set_pmd_at() write.
640 page_add_new_anon_rmap(page, vma, haddr);
641 set_pmd_at(mm, haddr, pmd, entry);
642 prepare_pmd_huge_pte(pgtable, mm);
643 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
645 spin_unlock(&mm->page_table_lock);
651 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
653 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
656 static inline struct page *alloc_hugepage_vma(int defrag,
657 struct vm_area_struct *vma,
658 unsigned long haddr, int nd,
661 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
662 HPAGE_PMD_ORDER, vma, haddr, nd);
666 static inline struct page *alloc_hugepage(int defrag)
668 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
673 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
674 unsigned long address, pmd_t *pmd,
678 unsigned long haddr = address & HPAGE_PMD_MASK;
681 if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
682 if (unlikely(anon_vma_prepare(vma)))
684 if (unlikely(khugepaged_enter(vma)))
686 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
687 vma, haddr, numa_node_id(), 0);
688 if (unlikely(!page)) {
689 count_vm_event(THP_FAULT_FALLBACK);
692 count_vm_event(THP_FAULT_ALLOC);
693 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
698 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
702 * Use __pte_alloc instead of pte_alloc_map, because we can't
703 * run pte_offset_map on the pmd, if an huge pmd could
704 * materialize from under us from a different thread.
706 if (unlikely(__pte_alloc(mm, vma, pmd, address)))
708 /* if an huge pmd materialized from under us just retry later */
709 if (unlikely(pmd_trans_huge(*pmd)))
712 * A regular pmd is established and it can't morph into a huge pmd
713 * from under us anymore at this point because we hold the mmap_sem
714 * read mode and khugepaged takes it in write mode. So now it's
715 * safe to run pte_offset_map().
717 pte = pte_offset_map(pmd, address);
718 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
721 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
722 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
723 struct vm_area_struct *vma)
725 struct page *src_page;
731 pgtable = pte_alloc_one(dst_mm, addr);
732 if (unlikely(!pgtable))
735 spin_lock(&dst_mm->page_table_lock);
736 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
740 if (unlikely(!pmd_trans_huge(pmd))) {
741 pte_free(dst_mm, pgtable);
744 if (unlikely(pmd_trans_splitting(pmd))) {
745 /* split huge page running from under us */
746 spin_unlock(&src_mm->page_table_lock);
747 spin_unlock(&dst_mm->page_table_lock);
748 pte_free(dst_mm, pgtable);
750 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
753 src_page = pmd_page(pmd);
754 VM_BUG_ON(!PageHead(src_page));
756 page_dup_rmap(src_page);
757 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
759 pmdp_set_wrprotect(src_mm, addr, src_pmd);
760 pmd = pmd_mkold(pmd_wrprotect(pmd));
761 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
762 prepare_pmd_huge_pte(pgtable, dst_mm);
767 spin_unlock(&src_mm->page_table_lock);
768 spin_unlock(&dst_mm->page_table_lock);
773 /* no "address" argument so destroys page coloring of some arch */
774 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
778 assert_spin_locked(&mm->page_table_lock);
781 pgtable = mm->pmd_huge_pte;
782 if (list_empty(&pgtable->lru))
783 mm->pmd_huge_pte = NULL;
785 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
787 list_del(&pgtable->lru);
792 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
793 struct vm_area_struct *vma,
794 unsigned long address,
795 pmd_t *pmd, pmd_t orig_pmd,
804 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
806 if (unlikely(!pages)) {
811 for (i = 0; i < HPAGE_PMD_NR; i++) {
812 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
814 vma, address, page_to_nid(page));
815 if (unlikely(!pages[i] ||
816 mem_cgroup_newpage_charge(pages[i], mm,
820 mem_cgroup_uncharge_start();
822 mem_cgroup_uncharge_page(pages[i]);
825 mem_cgroup_uncharge_end();
832 for (i = 0; i < HPAGE_PMD_NR; i++) {
833 copy_user_highpage(pages[i], page + i,
834 haddr + PAGE_SHIFT*i, vma);
835 __SetPageUptodate(pages[i]);
839 spin_lock(&mm->page_table_lock);
840 if (unlikely(!pmd_same(*pmd, orig_pmd)))
842 VM_BUG_ON(!PageHead(page));
844 pmdp_clear_flush_notify(vma, haddr, pmd);
845 /* leave pmd empty until pte is filled */
847 pgtable = get_pmd_huge_pte(mm);
848 pmd_populate(mm, &_pmd, pgtable);
850 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
852 entry = mk_pte(pages[i], vma->vm_page_prot);
853 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
854 page_add_new_anon_rmap(pages[i], vma, haddr);
855 pte = pte_offset_map(&_pmd, haddr);
856 VM_BUG_ON(!pte_none(*pte));
857 set_pte_at(mm, haddr, pte, entry);
862 smp_wmb(); /* make pte visible before pmd */
863 pmd_populate(mm, pmd, pgtable);
864 page_remove_rmap(page);
865 spin_unlock(&mm->page_table_lock);
867 ret |= VM_FAULT_WRITE;
874 spin_unlock(&mm->page_table_lock);
875 mem_cgroup_uncharge_start();
876 for (i = 0; i < HPAGE_PMD_NR; i++) {
877 mem_cgroup_uncharge_page(pages[i]);
880 mem_cgroup_uncharge_end();
885 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
886 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
889 struct page *page, *new_page;
892 VM_BUG_ON(!vma->anon_vma);
893 spin_lock(&mm->page_table_lock);
894 if (unlikely(!pmd_same(*pmd, orig_pmd)))
897 page = pmd_page(orig_pmd);
898 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
899 haddr = address & HPAGE_PMD_MASK;
900 if (page_mapcount(page) == 1) {
902 entry = pmd_mkyoung(orig_pmd);
903 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
904 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
905 update_mmu_cache(vma, address, entry);
906 ret |= VM_FAULT_WRITE;
910 spin_unlock(&mm->page_table_lock);
912 if (transparent_hugepage_enabled(vma) &&
913 !transparent_hugepage_debug_cow())
914 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
915 vma, haddr, numa_node_id(), 0);
919 if (unlikely(!new_page)) {
920 count_vm_event(THP_FAULT_FALLBACK);
921 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
922 pmd, orig_pmd, page, haddr);
923 if (ret & VM_FAULT_OOM)
924 split_huge_page(page);
928 count_vm_event(THP_FAULT_ALLOC);
930 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
932 split_huge_page(page);
938 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
939 __SetPageUptodate(new_page);
941 spin_lock(&mm->page_table_lock);
943 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
944 mem_cgroup_uncharge_page(new_page);
948 VM_BUG_ON(!PageHead(page));
949 entry = mk_pmd(new_page, vma->vm_page_prot);
950 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
951 entry = pmd_mkhuge(entry);
952 pmdp_clear_flush_notify(vma, haddr, pmd);
953 page_add_new_anon_rmap(new_page, vma, haddr);
954 set_pmd_at(mm, haddr, pmd, entry);
955 update_mmu_cache(vma, address, entry);
956 page_remove_rmap(page);
958 ret |= VM_FAULT_WRITE;
961 spin_unlock(&mm->page_table_lock);
966 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
971 struct page *page = NULL;
973 assert_spin_locked(&mm->page_table_lock);
975 if (flags & FOLL_WRITE && !pmd_write(*pmd))
978 page = pmd_page(*pmd);
979 VM_BUG_ON(!PageHead(page));
980 if (flags & FOLL_TOUCH) {
983 * We should set the dirty bit only for FOLL_WRITE but
984 * for now the dirty bit in the pmd is meaningless.
985 * And if the dirty bit will become meaningful and
986 * we'll only set it with FOLL_WRITE, an atomic
987 * set_bit will be required on the pmd to set the
988 * young bit, instead of the current set_pmd_at.
990 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
991 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
993 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
994 VM_BUG_ON(!PageCompound(page));
995 if (flags & FOLL_GET)
1002 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1007 spin_lock(&tlb->mm->page_table_lock);
1008 if (likely(pmd_trans_huge(*pmd))) {
1009 if (unlikely(pmd_trans_splitting(*pmd))) {
1010 spin_unlock(&tlb->mm->page_table_lock);
1011 wait_split_huge_page(vma->anon_vma,
1016 pgtable = get_pmd_huge_pte(tlb->mm);
1017 page = pmd_page(*pmd);
1019 page_remove_rmap(page);
1020 VM_BUG_ON(page_mapcount(page) < 0);
1021 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1022 VM_BUG_ON(!PageHead(page));
1024 spin_unlock(&tlb->mm->page_table_lock);
1025 tlb_remove_page(tlb, page);
1026 pte_free(tlb->mm, pgtable);
1030 spin_unlock(&tlb->mm->page_table_lock);
1035 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1036 unsigned long addr, unsigned long end,
1041 spin_lock(&vma->vm_mm->page_table_lock);
1042 if (likely(pmd_trans_huge(*pmd))) {
1043 ret = !pmd_trans_splitting(*pmd);
1044 spin_unlock(&vma->vm_mm->page_table_lock);
1046 wait_split_huge_page(vma->anon_vma, pmd);
1049 * All logical pages in the range are present
1050 * if backed by a huge page.
1052 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1055 spin_unlock(&vma->vm_mm->page_table_lock);
1060 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1061 unsigned long addr, pgprot_t newprot)
1063 struct mm_struct *mm = vma->vm_mm;
1066 spin_lock(&mm->page_table_lock);
1067 if (likely(pmd_trans_huge(*pmd))) {
1068 if (unlikely(pmd_trans_splitting(*pmd))) {
1069 spin_unlock(&mm->page_table_lock);
1070 wait_split_huge_page(vma->anon_vma, pmd);
1074 entry = pmdp_get_and_clear(mm, addr, pmd);
1075 entry = pmd_modify(entry, newprot);
1076 set_pmd_at(mm, addr, pmd, entry);
1077 spin_unlock(&vma->vm_mm->page_table_lock);
1078 flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1082 spin_unlock(&vma->vm_mm->page_table_lock);
1087 pmd_t *page_check_address_pmd(struct page *page,
1088 struct mm_struct *mm,
1089 unsigned long address,
1090 enum page_check_address_pmd_flag flag)
1094 pmd_t *pmd, *ret = NULL;
1096 if (address & ~HPAGE_PMD_MASK)
1099 pgd = pgd_offset(mm, address);
1100 if (!pgd_present(*pgd))
1103 pud = pud_offset(pgd, address);
1104 if (!pud_present(*pud))
1107 pmd = pmd_offset(pud, address);
1110 if (pmd_page(*pmd) != page)
1113 * split_vma() may create temporary aliased mappings. There is
1114 * no risk as long as all huge pmd are found and have their
1115 * splitting bit set before __split_huge_page_refcount
1116 * runs. Finding the same huge pmd more than once during the
1117 * same rmap walk is not a problem.
1119 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1120 pmd_trans_splitting(*pmd))
1122 if (pmd_trans_huge(*pmd)) {
1123 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1124 !pmd_trans_splitting(*pmd));
1131 static int __split_huge_page_splitting(struct page *page,
1132 struct vm_area_struct *vma,
1133 unsigned long address)
1135 struct mm_struct *mm = vma->vm_mm;
1139 spin_lock(&mm->page_table_lock);
1140 pmd = page_check_address_pmd(page, mm, address,
1141 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1144 * We can't temporarily set the pmd to null in order
1145 * to split it, the pmd must remain marked huge at all
1146 * times or the VM won't take the pmd_trans_huge paths
1147 * and it won't wait on the anon_vma->root->mutex to
1148 * serialize against split_huge_page*.
1150 pmdp_splitting_flush_notify(vma, address, pmd);
1153 spin_unlock(&mm->page_table_lock);
1158 static void __split_huge_page_refcount(struct page *page)
1161 unsigned long head_index = page->index;
1162 struct zone *zone = page_zone(page);
1166 /* prevent PageLRU to go away from under us, and freeze lru stats */
1167 spin_lock_irq(&zone->lru_lock);
1168 compound_lock(page);
1170 for (i = 1; i < HPAGE_PMD_NR; i++) {
1171 struct page *page_tail = page + i;
1173 /* tail_page->_mapcount cannot change */
1174 BUG_ON(page_mapcount(page_tail) < 0);
1175 tail_count += page_mapcount(page_tail);
1176 /* check for overflow */
1177 BUG_ON(tail_count < 0);
1178 BUG_ON(atomic_read(&page_tail->_count) != 0);
1180 * tail_page->_count is zero and not changing from
1181 * under us. But get_page_unless_zero() may be running
1182 * from under us on the tail_page. If we used
1183 * atomic_set() below instead of atomic_add(), we
1184 * would then run atomic_set() concurrently with
1185 * get_page_unless_zero(), and atomic_set() is
1186 * implemented in C not using locked ops. spin_unlock
1187 * on x86 sometime uses locked ops because of PPro
1188 * errata 66, 92, so unless somebody can guarantee
1189 * atomic_set() here would be safe on all archs (and
1190 * not only on x86), it's safer to use atomic_add().
1192 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1193 &page_tail->_count);
1195 /* after clearing PageTail the gup refcount can be released */
1199 * retain hwpoison flag of the poisoned tail page:
1200 * fix for the unsuitable process killed on Guest Machine(KVM)
1201 * by the memory-failure.
1203 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1204 page_tail->flags |= (page->flags &
1205 ((1L << PG_referenced) |
1206 (1L << PG_swapbacked) |
1207 (1L << PG_mlocked) |
1208 (1L << PG_uptodate)));
1209 page_tail->flags |= (1L << PG_dirty);
1211 /* clear PageTail before overwriting first_page */
1215 * __split_huge_page_splitting() already set the
1216 * splitting bit in all pmd that could map this
1217 * hugepage, that will ensure no CPU can alter the
1218 * mapcount on the head page. The mapcount is only
1219 * accounted in the head page and it has to be
1220 * transferred to all tail pages in the below code. So
1221 * for this code to be safe, the split the mapcount
1222 * can't change. But that doesn't mean userland can't
1223 * keep changing and reading the page contents while
1224 * we transfer the mapcount, so the pmd splitting
1225 * status is achieved setting a reserved bit in the
1226 * pmd, not by clearing the present bit.
1228 page_tail->_mapcount = page->_mapcount;
1230 BUG_ON(page_tail->mapping);
1231 page_tail->mapping = page->mapping;
1233 page_tail->index = ++head_index;
1235 BUG_ON(!PageAnon(page_tail));
1236 BUG_ON(!PageUptodate(page_tail));
1237 BUG_ON(!PageDirty(page_tail));
1238 BUG_ON(!PageSwapBacked(page_tail));
1240 mem_cgroup_split_huge_fixup(page, page_tail);
1242 lru_add_page_tail(zone, page, page_tail);
1244 atomic_sub(tail_count, &page->_count);
1245 BUG_ON(atomic_read(&page->_count) <= 0);
1247 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1248 __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1251 * A hugepage counts for HPAGE_PMD_NR pages on the LRU statistics,
1252 * so adjust those appropriately if this page is on the LRU.
1254 if (PageLRU(page)) {
1255 zonestat = NR_LRU_BASE + page_lru(page);
1256 __mod_zone_page_state(zone, zonestat, -(HPAGE_PMD_NR-1));
1259 ClearPageCompound(page);
1260 compound_unlock(page);
1261 spin_unlock_irq(&zone->lru_lock);
1263 for (i = 1; i < HPAGE_PMD_NR; i++) {
1264 struct page *page_tail = page + i;
1265 BUG_ON(page_count(page_tail) <= 0);
1267 * Tail pages may be freed if there wasn't any mapping
1268 * like if add_to_swap() is running on a lru page that
1269 * had its mapping zapped. And freeing these pages
1270 * requires taking the lru_lock so we do the put_page
1271 * of the tail pages after the split is complete.
1273 put_page(page_tail);
1277 * Only the head page (now become a regular page) is required
1278 * to be pinned by the caller.
1280 BUG_ON(page_count(page) <= 0);
1283 static int __split_huge_page_map(struct page *page,
1284 struct vm_area_struct *vma,
1285 unsigned long address)
1287 struct mm_struct *mm = vma->vm_mm;
1291 unsigned long haddr;
1293 spin_lock(&mm->page_table_lock);
1294 pmd = page_check_address_pmd(page, mm, address,
1295 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1297 pgtable = get_pmd_huge_pte(mm);
1298 pmd_populate(mm, &_pmd, pgtable);
1300 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1301 i++, haddr += PAGE_SIZE) {
1303 BUG_ON(PageCompound(page+i));
1304 entry = mk_pte(page + i, vma->vm_page_prot);
1305 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1306 if (!pmd_write(*pmd))
1307 entry = pte_wrprotect(entry);
1309 BUG_ON(page_mapcount(page) != 1);
1310 if (!pmd_young(*pmd))
1311 entry = pte_mkold(entry);
1312 pte = pte_offset_map(&_pmd, haddr);
1313 BUG_ON(!pte_none(*pte));
1314 set_pte_at(mm, haddr, pte, entry);
1318 smp_wmb(); /* make pte visible before pmd */
1320 * Up to this point the pmd is present and huge and
1321 * userland has the whole access to the hugepage
1322 * during the split (which happens in place). If we
1323 * overwrite the pmd with the not-huge version
1324 * pointing to the pte here (which of course we could
1325 * if all CPUs were bug free), userland could trigger
1326 * a small page size TLB miss on the small sized TLB
1327 * while the hugepage TLB entry is still established
1328 * in the huge TLB. Some CPU doesn't like that. See
1329 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1330 * Erratum 383 on page 93. Intel should be safe but is
1331 * also warns that it's only safe if the permission
1332 * and cache attributes of the two entries loaded in
1333 * the two TLB is identical (which should be the case
1334 * here). But it is generally safer to never allow
1335 * small and huge TLB entries for the same virtual
1336 * address to be loaded simultaneously. So instead of
1337 * doing "pmd_populate(); flush_tlb_range();" we first
1338 * mark the current pmd notpresent (atomically because
1339 * here the pmd_trans_huge and pmd_trans_splitting
1340 * must remain set at all times on the pmd until the
1341 * split is complete for this pmd), then we flush the
1342 * SMP TLB and finally we write the non-huge version
1343 * of the pmd entry with pmd_populate.
1345 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1346 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1347 pmd_populate(mm, pmd, pgtable);
1350 spin_unlock(&mm->page_table_lock);
1355 /* must be called with anon_vma->root->mutex hold */
1356 static void __split_huge_page(struct page *page,
1357 struct anon_vma *anon_vma)
1359 int mapcount, mapcount2;
1360 struct anon_vma_chain *avc;
1362 BUG_ON(!PageHead(page));
1363 BUG_ON(PageTail(page));
1366 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1367 struct vm_area_struct *vma = avc->vma;
1368 unsigned long addr = vma_address(page, vma);
1369 BUG_ON(is_vma_temporary_stack(vma));
1370 if (addr == -EFAULT)
1372 mapcount += __split_huge_page_splitting(page, vma, addr);
1375 * It is critical that new vmas are added to the tail of the
1376 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1377 * and establishes a child pmd before
1378 * __split_huge_page_splitting() freezes the parent pmd (so if
1379 * we fail to prevent copy_huge_pmd() from running until the
1380 * whole __split_huge_page() is complete), we will still see
1381 * the newly established pmd of the child later during the
1382 * walk, to be able to set it as pmd_trans_splitting too.
1384 if (mapcount != page_mapcount(page))
1385 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1386 mapcount, page_mapcount(page));
1387 BUG_ON(mapcount != page_mapcount(page));
1389 __split_huge_page_refcount(page);
1392 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1393 struct vm_area_struct *vma = avc->vma;
1394 unsigned long addr = vma_address(page, vma);
1395 BUG_ON(is_vma_temporary_stack(vma));
1396 if (addr == -EFAULT)
1398 mapcount2 += __split_huge_page_map(page, vma, addr);
1400 if (mapcount != mapcount2)
1401 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1402 mapcount, mapcount2, page_mapcount(page));
1403 BUG_ON(mapcount != mapcount2);
1406 int split_huge_page(struct page *page)
1408 struct anon_vma *anon_vma;
1411 BUG_ON(!PageAnon(page));
1412 anon_vma = page_lock_anon_vma(page);
1416 if (!PageCompound(page))
1419 BUG_ON(!PageSwapBacked(page));
1420 __split_huge_page(page, anon_vma);
1421 count_vm_event(THP_SPLIT);
1423 BUG_ON(PageCompound(page));
1425 page_unlock_anon_vma(anon_vma);
1430 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1431 VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1433 int hugepage_madvise(struct vm_area_struct *vma,
1434 unsigned long *vm_flags, int advice)
1439 * Be somewhat over-protective like KSM for now!
1441 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1443 *vm_flags &= ~VM_NOHUGEPAGE;
1444 *vm_flags |= VM_HUGEPAGE;
1446 * If the vma become good for khugepaged to scan,
1447 * register it here without waiting a page fault that
1448 * may not happen any time soon.
1450 if (unlikely(khugepaged_enter_vma_merge(vma)))
1453 case MADV_NOHUGEPAGE:
1455 * Be somewhat over-protective like KSM for now!
1457 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1459 *vm_flags &= ~VM_HUGEPAGE;
1460 *vm_flags |= VM_NOHUGEPAGE;
1462 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1463 * this vma even if we leave the mm registered in khugepaged if
1464 * it got registered before VM_NOHUGEPAGE was set.
1472 static int __init khugepaged_slab_init(void)
1474 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1475 sizeof(struct mm_slot),
1476 __alignof__(struct mm_slot), 0, NULL);
1483 static void __init khugepaged_slab_free(void)
1485 kmem_cache_destroy(mm_slot_cache);
1486 mm_slot_cache = NULL;
1489 static inline struct mm_slot *alloc_mm_slot(void)
1491 if (!mm_slot_cache) /* initialization failed */
1493 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1496 static inline void free_mm_slot(struct mm_slot *mm_slot)
1498 kmem_cache_free(mm_slot_cache, mm_slot);
1501 static int __init mm_slots_hash_init(void)
1503 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1511 static void __init mm_slots_hash_free(void)
1513 kfree(mm_slots_hash);
1514 mm_slots_hash = NULL;
1518 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1520 struct mm_slot *mm_slot;
1521 struct hlist_head *bucket;
1522 struct hlist_node *node;
1524 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1525 % MM_SLOTS_HASH_HEADS];
1526 hlist_for_each_entry(mm_slot, node, bucket, hash) {
1527 if (mm == mm_slot->mm)
1533 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1534 struct mm_slot *mm_slot)
1536 struct hlist_head *bucket;
1538 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1539 % MM_SLOTS_HASH_HEADS];
1541 hlist_add_head(&mm_slot->hash, bucket);
1544 static inline int khugepaged_test_exit(struct mm_struct *mm)
1546 return atomic_read(&mm->mm_users) == 0;
1549 int __khugepaged_enter(struct mm_struct *mm)
1551 struct mm_slot *mm_slot;
1554 mm_slot = alloc_mm_slot();
1558 /* __khugepaged_exit() must not run from under us */
1559 VM_BUG_ON(khugepaged_test_exit(mm));
1560 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1561 free_mm_slot(mm_slot);
1565 spin_lock(&khugepaged_mm_lock);
1566 insert_to_mm_slots_hash(mm, mm_slot);
1568 * Insert just behind the scanning cursor, to let the area settle
1571 wakeup = list_empty(&khugepaged_scan.mm_head);
1572 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1573 spin_unlock(&khugepaged_mm_lock);
1575 atomic_inc(&mm->mm_count);
1577 wake_up_interruptible(&khugepaged_wait);
1582 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1584 unsigned long hstart, hend;
1587 * Not yet faulted in so we will register later in the
1588 * page fault if needed.
1592 /* khugepaged not yet working on file or special mappings */
1595 * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1596 * true too, verify it here.
1598 VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1599 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1600 hend = vma->vm_end & HPAGE_PMD_MASK;
1602 return khugepaged_enter(vma);
1606 void __khugepaged_exit(struct mm_struct *mm)
1608 struct mm_slot *mm_slot;
1611 spin_lock(&khugepaged_mm_lock);
1612 mm_slot = get_mm_slot(mm);
1613 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1614 hlist_del(&mm_slot->hash);
1615 list_del(&mm_slot->mm_node);
1620 spin_unlock(&khugepaged_mm_lock);
1621 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1622 free_mm_slot(mm_slot);
1624 } else if (mm_slot) {
1625 spin_unlock(&khugepaged_mm_lock);
1627 * This is required to serialize against
1628 * khugepaged_test_exit() (which is guaranteed to run
1629 * under mmap sem read mode). Stop here (after we
1630 * return all pagetables will be destroyed) until
1631 * khugepaged has finished working on the pagetables
1632 * under the mmap_sem.
1634 down_write(&mm->mmap_sem);
1635 up_write(&mm->mmap_sem);
1637 spin_unlock(&khugepaged_mm_lock);
1640 static void release_pte_page(struct page *page)
1642 /* 0 stands for page_is_file_cache(page) == false */
1643 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1645 putback_lru_page(page);
1648 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1650 while (--_pte >= pte) {
1651 pte_t pteval = *_pte;
1652 if (!pte_none(pteval))
1653 release_pte_page(pte_page(pteval));
1657 static void release_all_pte_pages(pte_t *pte)
1659 release_pte_pages(pte, pte + HPAGE_PMD_NR);
1662 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1663 unsigned long address,
1668 int referenced = 0, isolated = 0, none = 0;
1669 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1670 _pte++, address += PAGE_SIZE) {
1671 pte_t pteval = *_pte;
1672 if (pte_none(pteval)) {
1673 if (++none <= khugepaged_max_ptes_none)
1676 release_pte_pages(pte, _pte);
1680 if (!pte_present(pteval) || !pte_write(pteval)) {
1681 release_pte_pages(pte, _pte);
1684 page = vm_normal_page(vma, address, pteval);
1685 if (unlikely(!page)) {
1686 release_pte_pages(pte, _pte);
1689 VM_BUG_ON(PageCompound(page));
1690 BUG_ON(!PageAnon(page));
1691 VM_BUG_ON(!PageSwapBacked(page));
1693 /* cannot use mapcount: can't collapse if there's a gup pin */
1694 if (page_count(page) != 1) {
1695 release_pte_pages(pte, _pte);
1699 * We can do it before isolate_lru_page because the
1700 * page can't be freed from under us. NOTE: PG_lock
1701 * is needed to serialize against split_huge_page
1702 * when invoked from the VM.
1704 if (!trylock_page(page)) {
1705 release_pte_pages(pte, _pte);
1709 * Isolate the page to avoid collapsing an hugepage
1710 * currently in use by the VM.
1712 if (isolate_lru_page(page)) {
1714 release_pte_pages(pte, _pte);
1717 /* 0 stands for page_is_file_cache(page) == false */
1718 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1719 VM_BUG_ON(!PageLocked(page));
1720 VM_BUG_ON(PageLRU(page));
1722 /* If there is no mapped pte young don't collapse the page */
1723 if (pte_young(pteval) || PageReferenced(page) ||
1724 mmu_notifier_test_young(vma->vm_mm, address))
1727 if (unlikely(!referenced))
1728 release_all_pte_pages(pte);
1735 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1736 struct vm_area_struct *vma,
1737 unsigned long address,
1741 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1742 pte_t pteval = *_pte;
1743 struct page *src_page;
1745 if (pte_none(pteval)) {
1746 clear_user_highpage(page, address);
1747 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1749 src_page = pte_page(pteval);
1750 copy_user_highpage(page, src_page, address, vma);
1751 VM_BUG_ON(page_mapcount(src_page) != 1);
1752 VM_BUG_ON(page_count(src_page) != 2);
1753 release_pte_page(src_page);
1755 * ptl mostly unnecessary, but preempt has to
1756 * be disabled to update the per-cpu stats
1757 * inside page_remove_rmap().
1761 * paravirt calls inside pte_clear here are
1764 pte_clear(vma->vm_mm, address, _pte);
1765 page_remove_rmap(src_page);
1767 free_page_and_swap_cache(src_page);
1770 address += PAGE_SIZE;
1775 static void collapse_huge_page(struct mm_struct *mm,
1776 unsigned long address,
1777 struct page **hpage,
1778 struct vm_area_struct *vma,
1786 struct page *new_page;
1789 unsigned long hstart, hend;
1791 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1793 up_read(&mm->mmap_sem);
1799 * Allocate the page while the vma is still valid and under
1800 * the mmap_sem read mode so there is no memory allocation
1801 * later when we take the mmap_sem in write mode. This is more
1802 * friendly behavior (OTOH it may actually hide bugs) to
1803 * filesystems in userland with daemons allocating memory in
1804 * the userland I/O paths. Allocating memory with the
1805 * mmap_sem in read mode is good idea also to allow greater
1808 new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1809 node, __GFP_OTHER_NODE);
1812 * After allocating the hugepage, release the mmap_sem read lock in
1813 * preparation for taking it in write mode.
1815 up_read(&mm->mmap_sem);
1816 if (unlikely(!new_page)) {
1817 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1818 *hpage = ERR_PTR(-ENOMEM);
1823 count_vm_event(THP_COLLAPSE_ALLOC);
1824 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1832 * Prevent all access to pagetables with the exception of
1833 * gup_fast later hanlded by the ptep_clear_flush and the VM
1834 * handled by the anon_vma lock + PG_lock.
1836 down_write(&mm->mmap_sem);
1837 if (unlikely(khugepaged_test_exit(mm)))
1840 vma = find_vma(mm, address);
1841 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1842 hend = vma->vm_end & HPAGE_PMD_MASK;
1843 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1846 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1847 (vma->vm_flags & VM_NOHUGEPAGE))
1850 if (!vma->anon_vma || vma->vm_ops)
1852 if (is_vma_temporary_stack(vma))
1855 * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1856 * true too, verify it here.
1858 VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1860 pgd = pgd_offset(mm, address);
1861 if (!pgd_present(*pgd))
1864 pud = pud_offset(pgd, address);
1865 if (!pud_present(*pud))
1868 pmd = pmd_offset(pud, address);
1869 /* pmd can't go away or become huge under us */
1870 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1873 anon_vma_lock(vma->anon_vma);
1875 pte = pte_offset_map(pmd, address);
1876 ptl = pte_lockptr(mm, pmd);
1878 spin_lock(&mm->page_table_lock); /* probably unnecessary */
1880 * After this gup_fast can't run anymore. This also removes
1881 * any huge TLB entry from the CPU so we won't allow
1882 * huge and small TLB entries for the same virtual address
1883 * to avoid the risk of CPU bugs in that area.
1885 _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1886 spin_unlock(&mm->page_table_lock);
1889 isolated = __collapse_huge_page_isolate(vma, address, pte);
1892 if (unlikely(!isolated)) {
1894 spin_lock(&mm->page_table_lock);
1895 BUG_ON(!pmd_none(*pmd));
1896 set_pmd_at(mm, address, pmd, _pmd);
1897 spin_unlock(&mm->page_table_lock);
1898 anon_vma_unlock(vma->anon_vma);
1903 * All pages are isolated and locked so anon_vma rmap
1904 * can't run anymore.
1906 anon_vma_unlock(vma->anon_vma);
1908 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1910 __SetPageUptodate(new_page);
1911 pgtable = pmd_pgtable(_pmd);
1912 VM_BUG_ON(page_count(pgtable) != 1);
1913 VM_BUG_ON(page_mapcount(pgtable) != 0);
1915 _pmd = mk_pmd(new_page, vma->vm_page_prot);
1916 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1917 _pmd = pmd_mkhuge(_pmd);
1920 * spin_lock() below is not the equivalent of smp_wmb(), so
1921 * this is needed to avoid the copy_huge_page writes to become
1922 * visible after the set_pmd_at() write.
1926 spin_lock(&mm->page_table_lock);
1927 BUG_ON(!pmd_none(*pmd));
1928 page_add_new_anon_rmap(new_page, vma, address);
1929 set_pmd_at(mm, address, pmd, _pmd);
1930 update_mmu_cache(vma, address, entry);
1931 prepare_pmd_huge_pte(pgtable, mm);
1932 spin_unlock(&mm->page_table_lock);
1937 khugepaged_pages_collapsed++;
1939 up_write(&mm->mmap_sem);
1943 mem_cgroup_uncharge_page(new_page);
1950 static int khugepaged_scan_pmd(struct mm_struct *mm,
1951 struct vm_area_struct *vma,
1952 unsigned long address,
1953 struct page **hpage)
1959 int ret = 0, referenced = 0, none = 0;
1961 unsigned long _address;
1965 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1967 pgd = pgd_offset(mm, address);
1968 if (!pgd_present(*pgd))
1971 pud = pud_offset(pgd, address);
1972 if (!pud_present(*pud))
1975 pmd = pmd_offset(pud, address);
1976 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1979 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1980 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1981 _pte++, _address += PAGE_SIZE) {
1982 pte_t pteval = *_pte;
1983 if (pte_none(pteval)) {
1984 if (++none <= khugepaged_max_ptes_none)
1989 if (!pte_present(pteval) || !pte_write(pteval))
1991 page = vm_normal_page(vma, _address, pteval);
1992 if (unlikely(!page))
1995 * Chose the node of the first page. This could
1996 * be more sophisticated and look at more pages,
1997 * but isn't for now.
2000 node = page_to_nid(page);
2001 VM_BUG_ON(PageCompound(page));
2002 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2004 /* cannot use mapcount: can't collapse if there's a gup pin */
2005 if (page_count(page) != 1)
2007 if (pte_young(pteval) || PageReferenced(page) ||
2008 mmu_notifier_test_young(vma->vm_mm, address))
2014 pte_unmap_unlock(pte, ptl);
2016 /* collapse_huge_page will return with the mmap_sem released */
2017 collapse_huge_page(mm, address, hpage, vma, node);
2022 static void collect_mm_slot(struct mm_slot *mm_slot)
2024 struct mm_struct *mm = mm_slot->mm;
2026 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2028 if (khugepaged_test_exit(mm)) {
2030 hlist_del(&mm_slot->hash);
2031 list_del(&mm_slot->mm_node);
2034 * Not strictly needed because the mm exited already.
2036 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2039 /* khugepaged_mm_lock actually not necessary for the below */
2040 free_mm_slot(mm_slot);
2045 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2046 struct page **hpage)
2048 struct mm_slot *mm_slot;
2049 struct mm_struct *mm;
2050 struct vm_area_struct *vma;
2054 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2056 if (khugepaged_scan.mm_slot)
2057 mm_slot = khugepaged_scan.mm_slot;
2059 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2060 struct mm_slot, mm_node);
2061 khugepaged_scan.address = 0;
2062 khugepaged_scan.mm_slot = mm_slot;
2064 spin_unlock(&khugepaged_mm_lock);
2067 down_read(&mm->mmap_sem);
2068 if (unlikely(khugepaged_test_exit(mm)))
2071 vma = find_vma(mm, khugepaged_scan.address);
2074 for (; vma; vma = vma->vm_next) {
2075 unsigned long hstart, hend;
2078 if (unlikely(khugepaged_test_exit(mm))) {
2083 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2084 !khugepaged_always()) ||
2085 (vma->vm_flags & VM_NOHUGEPAGE)) {
2090 if (!vma->anon_vma || vma->vm_ops)
2092 if (is_vma_temporary_stack(vma))
2095 * If is_pfn_mapping() is true is_learn_pfn_mapping()
2096 * must be true too, verify it here.
2098 VM_BUG_ON(is_linear_pfn_mapping(vma) ||
2099 vma->vm_flags & VM_NO_THP);
2101 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2102 hend = vma->vm_end & HPAGE_PMD_MASK;
2105 if (khugepaged_scan.address > hend)
2107 if (khugepaged_scan.address < hstart)
2108 khugepaged_scan.address = hstart;
2109 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2111 while (khugepaged_scan.address < hend) {
2114 if (unlikely(khugepaged_test_exit(mm)))
2115 goto breakouterloop;
2117 VM_BUG_ON(khugepaged_scan.address < hstart ||
2118 khugepaged_scan.address + HPAGE_PMD_SIZE >
2120 ret = khugepaged_scan_pmd(mm, vma,
2121 khugepaged_scan.address,
2123 /* move to next address */
2124 khugepaged_scan.address += HPAGE_PMD_SIZE;
2125 progress += HPAGE_PMD_NR;
2127 /* we released mmap_sem so break loop */
2128 goto breakouterloop_mmap_sem;
2129 if (progress >= pages)
2130 goto breakouterloop;
2134 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2135 breakouterloop_mmap_sem:
2137 spin_lock(&khugepaged_mm_lock);
2138 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2140 * Release the current mm_slot if this mm is about to die, or
2141 * if we scanned all vmas of this mm.
2143 if (khugepaged_test_exit(mm) || !vma) {
2145 * Make sure that if mm_users is reaching zero while
2146 * khugepaged runs here, khugepaged_exit will find
2147 * mm_slot not pointing to the exiting mm.
2149 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2150 khugepaged_scan.mm_slot = list_entry(
2151 mm_slot->mm_node.next,
2152 struct mm_slot, mm_node);
2153 khugepaged_scan.address = 0;
2155 khugepaged_scan.mm_slot = NULL;
2156 khugepaged_full_scans++;
2159 collect_mm_slot(mm_slot);
2165 static int khugepaged_has_work(void)
2167 return !list_empty(&khugepaged_scan.mm_head) &&
2168 khugepaged_enabled();
2171 static int khugepaged_wait_event(void)
2173 return !list_empty(&khugepaged_scan.mm_head) ||
2174 !khugepaged_enabled();
2177 static void khugepaged_do_scan(struct page **hpage)
2179 unsigned int progress = 0, pass_through_head = 0;
2180 unsigned int pages = khugepaged_pages_to_scan;
2182 barrier(); /* write khugepaged_pages_to_scan to local stack */
2184 while (progress < pages) {
2189 *hpage = alloc_hugepage(khugepaged_defrag());
2190 if (unlikely(!*hpage)) {
2191 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2194 count_vm_event(THP_COLLAPSE_ALLOC);
2201 if (unlikely(kthread_should_stop() || freezing(current)))
2204 spin_lock(&khugepaged_mm_lock);
2205 if (!khugepaged_scan.mm_slot)
2206 pass_through_head++;
2207 if (khugepaged_has_work() &&
2208 pass_through_head < 2)
2209 progress += khugepaged_scan_mm_slot(pages - progress,
2213 spin_unlock(&khugepaged_mm_lock);
2217 static void khugepaged_alloc_sleep(void)
2220 add_wait_queue(&khugepaged_wait, &wait);
2221 schedule_timeout_interruptible(
2223 khugepaged_alloc_sleep_millisecs));
2224 remove_wait_queue(&khugepaged_wait, &wait);
2228 static struct page *khugepaged_alloc_hugepage(void)
2233 hpage = alloc_hugepage(khugepaged_defrag());
2235 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2236 khugepaged_alloc_sleep();
2238 count_vm_event(THP_COLLAPSE_ALLOC);
2239 } while (unlikely(!hpage) &&
2240 likely(khugepaged_enabled()));
2245 static void khugepaged_loop(void)
2252 while (likely(khugepaged_enabled())) {
2254 hpage = khugepaged_alloc_hugepage();
2255 if (unlikely(!hpage))
2258 if (IS_ERR(hpage)) {
2259 khugepaged_alloc_sleep();
2264 khugepaged_do_scan(&hpage);
2270 if (unlikely(kthread_should_stop()))
2272 if (khugepaged_has_work()) {
2274 if (!khugepaged_scan_sleep_millisecs)
2276 add_wait_queue(&khugepaged_wait, &wait);
2277 schedule_timeout_interruptible(
2279 khugepaged_scan_sleep_millisecs));
2280 remove_wait_queue(&khugepaged_wait, &wait);
2281 } else if (khugepaged_enabled())
2282 wait_event_freezable(khugepaged_wait,
2283 khugepaged_wait_event());
2287 static int khugepaged(void *none)
2289 struct mm_slot *mm_slot;
2292 set_user_nice(current, 19);
2294 /* serialize with start_khugepaged() */
2295 mutex_lock(&khugepaged_mutex);
2298 mutex_unlock(&khugepaged_mutex);
2299 VM_BUG_ON(khugepaged_thread != current);
2301 VM_BUG_ON(khugepaged_thread != current);
2303 mutex_lock(&khugepaged_mutex);
2304 if (!khugepaged_enabled())
2306 if (unlikely(kthread_should_stop()))
2310 spin_lock(&khugepaged_mm_lock);
2311 mm_slot = khugepaged_scan.mm_slot;
2312 khugepaged_scan.mm_slot = NULL;
2314 collect_mm_slot(mm_slot);
2315 spin_unlock(&khugepaged_mm_lock);
2317 khugepaged_thread = NULL;
2318 mutex_unlock(&khugepaged_mutex);
2323 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2327 spin_lock(&mm->page_table_lock);
2328 if (unlikely(!pmd_trans_huge(*pmd))) {
2329 spin_unlock(&mm->page_table_lock);
2332 page = pmd_page(*pmd);
2333 VM_BUG_ON(!page_count(page));
2335 spin_unlock(&mm->page_table_lock);
2337 split_huge_page(page);
2340 BUG_ON(pmd_trans_huge(*pmd));
2343 static void split_huge_page_address(struct mm_struct *mm,
2344 unsigned long address)
2350 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2352 pgd = pgd_offset(mm, address);
2353 if (!pgd_present(*pgd))
2356 pud = pud_offset(pgd, address);
2357 if (!pud_present(*pud))
2360 pmd = pmd_offset(pud, address);
2361 if (!pmd_present(*pmd))
2364 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2365 * materialize from under us.
2367 split_huge_page_pmd(mm, pmd);
2370 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2371 unsigned long start,
2376 * If the new start address isn't hpage aligned and it could
2377 * previously contain an hugepage: check if we need to split
2380 if (start & ~HPAGE_PMD_MASK &&
2381 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2382 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2383 split_huge_page_address(vma->vm_mm, start);
2386 * If the new end address isn't hpage aligned and it could
2387 * previously contain an hugepage: check if we need to split
2390 if (end & ~HPAGE_PMD_MASK &&
2391 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2392 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2393 split_huge_page_address(vma->vm_mm, end);
2396 * If we're also updating the vma->vm_next->vm_start, if the new
2397 * vm_next->vm_start isn't page aligned and it could previously
2398 * contain an hugepage: check if we need to split an huge pmd.
2400 if (adjust_next > 0) {
2401 struct vm_area_struct *next = vma->vm_next;
2402 unsigned long nstart = next->vm_start;
2403 nstart += adjust_next << PAGE_SHIFT;
2404 if (nstart & ~HPAGE_PMD_MASK &&
2405 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2406 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2407 split_huge_page_address(next->vm_mm, nstart);