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 static struct khugepaged_scan khugepaged_scan = {
94 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
98 static int set_recommended_min_free_kbytes(void)
102 unsigned long recommended_min;
103 extern int min_free_kbytes;
105 if (!khugepaged_enabled())
108 for_each_populated_zone(zone)
111 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
112 recommended_min = pageblock_nr_pages * nr_zones * 2;
115 * Make sure that on average at least two pageblocks are almost free
116 * of another type, one for a migratetype to fall back to and a
117 * second to avoid subsequent fallbacks of other types There are 3
118 * MIGRATE_TYPES we care about.
120 recommended_min += pageblock_nr_pages * nr_zones *
121 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
123 /* don't ever allow to reserve more than 5% of the lowmem */
124 recommended_min = min(recommended_min,
125 (unsigned long) nr_free_buffer_pages() / 20);
126 recommended_min <<= (PAGE_SHIFT-10);
128 if (recommended_min > min_free_kbytes)
129 min_free_kbytes = recommended_min;
130 setup_per_zone_wmarks();
133 late_initcall(set_recommended_min_free_kbytes);
135 static int start_khugepaged(void)
138 if (khugepaged_enabled()) {
139 if (!khugepaged_thread)
140 khugepaged_thread = kthread_run(khugepaged, NULL,
142 if (unlikely(IS_ERR(khugepaged_thread))) {
144 "khugepaged: kthread_run(khugepaged) failed\n");
145 err = PTR_ERR(khugepaged_thread);
146 khugepaged_thread = NULL;
149 if (!list_empty(&khugepaged_scan.mm_head))
150 wake_up_interruptible(&khugepaged_wait);
152 set_recommended_min_free_kbytes();
153 } else if (khugepaged_thread) {
154 kthread_stop(khugepaged_thread);
155 khugepaged_thread = NULL;
163 static ssize_t double_flag_show(struct kobject *kobj,
164 struct kobj_attribute *attr, char *buf,
165 enum transparent_hugepage_flag enabled,
166 enum transparent_hugepage_flag req_madv)
168 if (test_bit(enabled, &transparent_hugepage_flags)) {
169 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
170 return sprintf(buf, "[always] madvise never\n");
171 } else if (test_bit(req_madv, &transparent_hugepage_flags))
172 return sprintf(buf, "always [madvise] never\n");
174 return sprintf(buf, "always madvise [never]\n");
176 static ssize_t double_flag_store(struct kobject *kobj,
177 struct kobj_attribute *attr,
178 const char *buf, size_t count,
179 enum transparent_hugepage_flag enabled,
180 enum transparent_hugepage_flag req_madv)
182 if (!memcmp("always", buf,
183 min(sizeof("always")-1, count))) {
184 set_bit(enabled, &transparent_hugepage_flags);
185 clear_bit(req_madv, &transparent_hugepage_flags);
186 } else if (!memcmp("madvise", buf,
187 min(sizeof("madvise")-1, count))) {
188 clear_bit(enabled, &transparent_hugepage_flags);
189 set_bit(req_madv, &transparent_hugepage_flags);
190 } else if (!memcmp("never", buf,
191 min(sizeof("never")-1, count))) {
192 clear_bit(enabled, &transparent_hugepage_flags);
193 clear_bit(req_madv, &transparent_hugepage_flags);
200 static ssize_t enabled_show(struct kobject *kobj,
201 struct kobj_attribute *attr, char *buf)
203 return double_flag_show(kobj, attr, buf,
204 TRANSPARENT_HUGEPAGE_FLAG,
205 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
207 static ssize_t enabled_store(struct kobject *kobj,
208 struct kobj_attribute *attr,
209 const char *buf, size_t count)
213 ret = double_flag_store(kobj, attr, buf, count,
214 TRANSPARENT_HUGEPAGE_FLAG,
215 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
220 mutex_lock(&khugepaged_mutex);
221 err = start_khugepaged();
222 mutex_unlock(&khugepaged_mutex);
228 if (ret > 0 && khugepaged_enabled())
229 set_recommended_min_free_kbytes();
233 static struct kobj_attribute enabled_attr =
234 __ATTR(enabled, 0644, enabled_show, enabled_store);
236 static ssize_t single_flag_show(struct kobject *kobj,
237 struct kobj_attribute *attr, char *buf,
238 enum transparent_hugepage_flag flag)
240 return sprintf(buf, "%d\n",
241 !!test_bit(flag, &transparent_hugepage_flags));
244 static ssize_t single_flag_store(struct kobject *kobj,
245 struct kobj_attribute *attr,
246 const char *buf, size_t count,
247 enum transparent_hugepage_flag flag)
252 ret = kstrtoul(buf, 10, &value);
259 set_bit(flag, &transparent_hugepage_flags);
261 clear_bit(flag, &transparent_hugepage_flags);
267 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
268 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
269 * memory just to allocate one more hugepage.
271 static ssize_t defrag_show(struct kobject *kobj,
272 struct kobj_attribute *attr, char *buf)
274 return double_flag_show(kobj, attr, buf,
275 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
276 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
278 static ssize_t defrag_store(struct kobject *kobj,
279 struct kobj_attribute *attr,
280 const char *buf, size_t count)
282 return double_flag_store(kobj, attr, buf, count,
283 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
284 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
286 static struct kobj_attribute defrag_attr =
287 __ATTR(defrag, 0644, defrag_show, defrag_store);
289 #ifdef CONFIG_DEBUG_VM
290 static ssize_t debug_cow_show(struct kobject *kobj,
291 struct kobj_attribute *attr, char *buf)
293 return single_flag_show(kobj, attr, buf,
294 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
296 static ssize_t debug_cow_store(struct kobject *kobj,
297 struct kobj_attribute *attr,
298 const char *buf, size_t count)
300 return single_flag_store(kobj, attr, buf, count,
301 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
303 static struct kobj_attribute debug_cow_attr =
304 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
305 #endif /* CONFIG_DEBUG_VM */
307 static struct attribute *hugepage_attr[] = {
310 #ifdef CONFIG_DEBUG_VM
311 &debug_cow_attr.attr,
316 static struct attribute_group hugepage_attr_group = {
317 .attrs = hugepage_attr,
320 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
321 struct kobj_attribute *attr,
324 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
327 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
328 struct kobj_attribute *attr,
329 const char *buf, size_t count)
334 err = strict_strtoul(buf, 10, &msecs);
335 if (err || msecs > UINT_MAX)
338 khugepaged_scan_sleep_millisecs = msecs;
339 wake_up_interruptible(&khugepaged_wait);
343 static struct kobj_attribute scan_sleep_millisecs_attr =
344 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
345 scan_sleep_millisecs_store);
347 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
348 struct kobj_attribute *attr,
351 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
354 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
355 struct kobj_attribute *attr,
356 const char *buf, size_t count)
361 err = strict_strtoul(buf, 10, &msecs);
362 if (err || msecs > UINT_MAX)
365 khugepaged_alloc_sleep_millisecs = msecs;
366 wake_up_interruptible(&khugepaged_wait);
370 static struct kobj_attribute alloc_sleep_millisecs_attr =
371 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
372 alloc_sleep_millisecs_store);
374 static ssize_t pages_to_scan_show(struct kobject *kobj,
375 struct kobj_attribute *attr,
378 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
380 static ssize_t pages_to_scan_store(struct kobject *kobj,
381 struct kobj_attribute *attr,
382 const char *buf, size_t count)
387 err = strict_strtoul(buf, 10, &pages);
388 if (err || !pages || pages > UINT_MAX)
391 khugepaged_pages_to_scan = pages;
395 static struct kobj_attribute pages_to_scan_attr =
396 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
397 pages_to_scan_store);
399 static ssize_t pages_collapsed_show(struct kobject *kobj,
400 struct kobj_attribute *attr,
403 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
405 static struct kobj_attribute pages_collapsed_attr =
406 __ATTR_RO(pages_collapsed);
408 static ssize_t full_scans_show(struct kobject *kobj,
409 struct kobj_attribute *attr,
412 return sprintf(buf, "%u\n", khugepaged_full_scans);
414 static struct kobj_attribute full_scans_attr =
415 __ATTR_RO(full_scans);
417 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
418 struct kobj_attribute *attr, char *buf)
420 return single_flag_show(kobj, attr, buf,
421 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
423 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
424 struct kobj_attribute *attr,
425 const char *buf, size_t count)
427 return single_flag_store(kobj, attr, buf, count,
428 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
430 static struct kobj_attribute khugepaged_defrag_attr =
431 __ATTR(defrag, 0644, khugepaged_defrag_show,
432 khugepaged_defrag_store);
435 * max_ptes_none controls if khugepaged should collapse hugepages over
436 * any unmapped ptes in turn potentially increasing the memory
437 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
438 * reduce the available free memory in the system as it
439 * runs. Increasing max_ptes_none will instead potentially reduce the
440 * free memory in the system during the khugepaged scan.
442 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
443 struct kobj_attribute *attr,
446 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
448 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
449 struct kobj_attribute *attr,
450 const char *buf, size_t count)
453 unsigned long max_ptes_none;
455 err = strict_strtoul(buf, 10, &max_ptes_none);
456 if (err || max_ptes_none > HPAGE_PMD_NR-1)
459 khugepaged_max_ptes_none = max_ptes_none;
463 static struct kobj_attribute khugepaged_max_ptes_none_attr =
464 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
465 khugepaged_max_ptes_none_store);
467 static struct attribute *khugepaged_attr[] = {
468 &khugepaged_defrag_attr.attr,
469 &khugepaged_max_ptes_none_attr.attr,
470 &pages_to_scan_attr.attr,
471 &pages_collapsed_attr.attr,
472 &full_scans_attr.attr,
473 &scan_sleep_millisecs_attr.attr,
474 &alloc_sleep_millisecs_attr.attr,
478 static struct attribute_group khugepaged_attr_group = {
479 .attrs = khugepaged_attr,
480 .name = "khugepaged",
483 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
487 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
488 if (unlikely(!*hugepage_kobj)) {
489 printk(KERN_ERR "hugepage: failed kobject create\n");
493 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
495 printk(KERN_ERR "hugepage: failed register hugeage group\n");
499 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
501 printk(KERN_ERR "hugepage: failed register hugeage group\n");
502 goto remove_hp_group;
508 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
510 kobject_put(*hugepage_kobj);
514 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
516 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
517 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
518 kobject_put(hugepage_kobj);
521 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
526 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
529 #endif /* CONFIG_SYSFS */
531 static int __init hugepage_init(void)
534 struct kobject *hugepage_kobj;
536 if (!has_transparent_hugepage()) {
537 transparent_hugepage_flags = 0;
541 err = hugepage_init_sysfs(&hugepage_kobj);
545 err = khugepaged_slab_init();
549 err = mm_slots_hash_init();
551 khugepaged_slab_free();
556 * By default disable transparent hugepages on smaller systems,
557 * where the extra memory used could hurt more than TLB overhead
558 * is likely to save. The admin can still enable it through /sys.
560 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
561 transparent_hugepage_flags = 0;
565 set_recommended_min_free_kbytes();
569 hugepage_exit_sysfs(hugepage_kobj);
572 module_init(hugepage_init)
574 static int __init setup_transparent_hugepage(char *str)
579 if (!strcmp(str, "always")) {
580 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
581 &transparent_hugepage_flags);
582 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
583 &transparent_hugepage_flags);
585 } else if (!strcmp(str, "madvise")) {
586 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
587 &transparent_hugepage_flags);
588 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
589 &transparent_hugepage_flags);
591 } else if (!strcmp(str, "never")) {
592 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
593 &transparent_hugepage_flags);
594 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
595 &transparent_hugepage_flags);
601 "transparent_hugepage= cannot parse, ignored\n");
604 __setup("transparent_hugepage=", setup_transparent_hugepage);
606 static void prepare_pmd_huge_pte(pgtable_t pgtable,
607 struct mm_struct *mm)
609 assert_spin_locked(&mm->page_table_lock);
612 if (!mm->pmd_huge_pte)
613 INIT_LIST_HEAD(&pgtable->lru);
615 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
616 mm->pmd_huge_pte = pgtable;
619 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
621 if (likely(vma->vm_flags & VM_WRITE))
622 pmd = pmd_mkwrite(pmd);
626 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
627 struct vm_area_struct *vma,
628 unsigned long haddr, pmd_t *pmd,
633 VM_BUG_ON(!PageCompound(page));
634 pgtable = pte_alloc_one(mm, haddr);
635 if (unlikely(!pgtable))
638 clear_huge_page(page, haddr, HPAGE_PMD_NR);
639 __SetPageUptodate(page);
641 spin_lock(&mm->page_table_lock);
642 if (unlikely(!pmd_none(*pmd))) {
643 spin_unlock(&mm->page_table_lock);
644 mem_cgroup_uncharge_page(page);
646 pte_free(mm, pgtable);
649 entry = mk_pmd(page, vma->vm_page_prot);
650 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
651 entry = pmd_mkhuge(entry);
653 * The spinlocking to take the lru_lock inside
654 * page_add_new_anon_rmap() acts as a full memory
655 * barrier to be sure clear_huge_page writes become
656 * visible after the set_pmd_at() write.
658 page_add_new_anon_rmap(page, vma, haddr);
659 set_pmd_at(mm, haddr, pmd, entry);
660 prepare_pmd_huge_pte(pgtable, mm);
661 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
663 spin_unlock(&mm->page_table_lock);
669 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
671 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
674 static inline struct page *alloc_hugepage_vma(int defrag,
675 struct vm_area_struct *vma,
676 unsigned long haddr, int nd,
679 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
680 HPAGE_PMD_ORDER, vma, haddr, nd);
684 static inline struct page *alloc_hugepage(int defrag)
686 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
691 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
692 unsigned long address, pmd_t *pmd,
696 unsigned long haddr = address & HPAGE_PMD_MASK;
699 if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
700 if (unlikely(anon_vma_prepare(vma)))
702 if (unlikely(khugepaged_enter(vma)))
704 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
705 vma, haddr, numa_node_id(), 0);
706 if (unlikely(!page)) {
707 count_vm_event(THP_FAULT_FALLBACK);
710 count_vm_event(THP_FAULT_ALLOC);
711 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
715 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
717 mem_cgroup_uncharge_page(page);
726 * Use __pte_alloc instead of pte_alloc_map, because we can't
727 * run pte_offset_map on the pmd, if an huge pmd could
728 * materialize from under us from a different thread.
730 if (unlikely(__pte_alloc(mm, vma, pmd, address)))
732 /* if an huge pmd materialized from under us just retry later */
733 if (unlikely(pmd_trans_huge(*pmd)))
736 * A regular pmd is established and it can't morph into a huge pmd
737 * from under us anymore at this point because we hold the mmap_sem
738 * read mode and khugepaged takes it in write mode. So now it's
739 * safe to run pte_offset_map().
741 pte = pte_offset_map(pmd, address);
742 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
745 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
746 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
747 struct vm_area_struct *vma)
749 struct page *src_page;
755 pgtable = pte_alloc_one(dst_mm, addr);
756 if (unlikely(!pgtable))
759 spin_lock(&dst_mm->page_table_lock);
760 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
764 if (unlikely(!pmd_trans_huge(pmd))) {
765 pte_free(dst_mm, pgtable);
768 if (unlikely(pmd_trans_splitting(pmd))) {
769 /* split huge page running from under us */
770 spin_unlock(&src_mm->page_table_lock);
771 spin_unlock(&dst_mm->page_table_lock);
772 pte_free(dst_mm, pgtable);
774 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
777 src_page = pmd_page(pmd);
778 VM_BUG_ON(!PageHead(src_page));
780 page_dup_rmap(src_page);
781 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
783 pmdp_set_wrprotect(src_mm, addr, src_pmd);
784 pmd = pmd_mkold(pmd_wrprotect(pmd));
785 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
786 prepare_pmd_huge_pte(pgtable, dst_mm);
791 spin_unlock(&src_mm->page_table_lock);
792 spin_unlock(&dst_mm->page_table_lock);
797 /* no "address" argument so destroys page coloring of some arch */
798 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
802 assert_spin_locked(&mm->page_table_lock);
805 pgtable = mm->pmd_huge_pte;
806 if (list_empty(&pgtable->lru))
807 mm->pmd_huge_pte = NULL;
809 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
811 list_del(&pgtable->lru);
816 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
817 struct vm_area_struct *vma,
818 unsigned long address,
819 pmd_t *pmd, pmd_t orig_pmd,
828 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
830 if (unlikely(!pages)) {
835 for (i = 0; i < HPAGE_PMD_NR; i++) {
836 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
838 vma, address, page_to_nid(page));
839 if (unlikely(!pages[i] ||
840 mem_cgroup_newpage_charge(pages[i], mm,
844 mem_cgroup_uncharge_start();
846 mem_cgroup_uncharge_page(pages[i]);
849 mem_cgroup_uncharge_end();
856 for (i = 0; i < HPAGE_PMD_NR; i++) {
857 copy_user_highpage(pages[i], page + i,
858 haddr + PAGE_SIZE * i, vma);
859 __SetPageUptodate(pages[i]);
863 spin_lock(&mm->page_table_lock);
864 if (unlikely(!pmd_same(*pmd, orig_pmd)))
866 VM_BUG_ON(!PageHead(page));
868 pmdp_clear_flush_notify(vma, haddr, pmd);
869 /* leave pmd empty until pte is filled */
871 pgtable = get_pmd_huge_pte(mm);
872 pmd_populate(mm, &_pmd, pgtable);
874 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
876 entry = mk_pte(pages[i], vma->vm_page_prot);
877 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
878 page_add_new_anon_rmap(pages[i], vma, haddr);
879 pte = pte_offset_map(&_pmd, haddr);
880 VM_BUG_ON(!pte_none(*pte));
881 set_pte_at(mm, haddr, pte, entry);
886 smp_wmb(); /* make pte visible before pmd */
887 pmd_populate(mm, pmd, pgtable);
888 page_remove_rmap(page);
889 spin_unlock(&mm->page_table_lock);
891 ret |= VM_FAULT_WRITE;
898 spin_unlock(&mm->page_table_lock);
899 mem_cgroup_uncharge_start();
900 for (i = 0; i < HPAGE_PMD_NR; i++) {
901 mem_cgroup_uncharge_page(pages[i]);
904 mem_cgroup_uncharge_end();
909 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
910 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
913 struct page *page, *new_page;
916 VM_BUG_ON(!vma->anon_vma);
917 spin_lock(&mm->page_table_lock);
918 if (unlikely(!pmd_same(*pmd, orig_pmd)))
921 page = pmd_page(orig_pmd);
922 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
923 haddr = address & HPAGE_PMD_MASK;
924 if (page_mapcount(page) == 1) {
926 entry = pmd_mkyoung(orig_pmd);
927 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
928 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
929 update_mmu_cache(vma, address, entry);
930 ret |= VM_FAULT_WRITE;
934 spin_unlock(&mm->page_table_lock);
936 if (transparent_hugepage_enabled(vma) &&
937 !transparent_hugepage_debug_cow())
938 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
939 vma, haddr, numa_node_id(), 0);
943 if (unlikely(!new_page)) {
944 count_vm_event(THP_FAULT_FALLBACK);
945 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
946 pmd, orig_pmd, page, haddr);
947 if (ret & VM_FAULT_OOM)
948 split_huge_page(page);
952 count_vm_event(THP_FAULT_ALLOC);
954 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
956 split_huge_page(page);
962 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
963 __SetPageUptodate(new_page);
965 spin_lock(&mm->page_table_lock);
967 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
968 spin_unlock(&mm->page_table_lock);
969 mem_cgroup_uncharge_page(new_page);
974 VM_BUG_ON(!PageHead(page));
975 entry = mk_pmd(new_page, vma->vm_page_prot);
976 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
977 entry = pmd_mkhuge(entry);
978 pmdp_clear_flush_notify(vma, haddr, pmd);
979 page_add_new_anon_rmap(new_page, vma, haddr);
980 set_pmd_at(mm, haddr, pmd, entry);
981 update_mmu_cache(vma, address, entry);
982 page_remove_rmap(page);
984 ret |= VM_FAULT_WRITE;
987 spin_unlock(&mm->page_table_lock);
992 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
997 struct page *page = NULL;
999 assert_spin_locked(&mm->page_table_lock);
1001 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1004 page = pmd_page(*pmd);
1005 VM_BUG_ON(!PageHead(page));
1006 if (flags & FOLL_TOUCH) {
1009 * We should set the dirty bit only for FOLL_WRITE but
1010 * for now the dirty bit in the pmd is meaningless.
1011 * And if the dirty bit will become meaningful and
1012 * we'll only set it with FOLL_WRITE, an atomic
1013 * set_bit will be required on the pmd to set the
1014 * young bit, instead of the current set_pmd_at.
1016 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1017 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1019 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1020 VM_BUG_ON(!PageCompound(page));
1021 if (flags & FOLL_GET)
1022 get_page_foll(page);
1028 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1029 pmd_t *pmd, unsigned long addr)
1033 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1036 pgtable = get_pmd_huge_pte(tlb->mm);
1037 page = pmd_page(*pmd);
1039 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1040 page_remove_rmap(page);
1041 VM_BUG_ON(page_mapcount(page) < 0);
1042 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1043 VM_BUG_ON(!PageHead(page));
1045 spin_unlock(&tlb->mm->page_table_lock);
1046 tlb_remove_page(tlb, page);
1047 pte_free(tlb->mm, pgtable);
1053 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1054 unsigned long addr, unsigned long end,
1059 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1061 * All logical pages in the range are present
1062 * if backed by a huge page.
1064 spin_unlock(&vma->vm_mm->page_table_lock);
1065 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1072 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1073 unsigned long old_addr,
1074 unsigned long new_addr, unsigned long old_end,
1075 pmd_t *old_pmd, pmd_t *new_pmd)
1080 struct mm_struct *mm = vma->vm_mm;
1082 if ((old_addr & ~HPAGE_PMD_MASK) ||
1083 (new_addr & ~HPAGE_PMD_MASK) ||
1084 old_end - old_addr < HPAGE_PMD_SIZE ||
1085 (new_vma->vm_flags & VM_NOHUGEPAGE))
1089 * The destination pmd shouldn't be established, free_pgtables()
1090 * should have release it.
1092 if (WARN_ON(!pmd_none(*new_pmd))) {
1093 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1097 ret = __pmd_trans_huge_lock(old_pmd, vma);
1099 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1100 VM_BUG_ON(!pmd_none(*new_pmd));
1101 set_pmd_at(mm, new_addr, new_pmd, pmd);
1102 spin_unlock(&mm->page_table_lock);
1108 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1109 unsigned long addr, pgprot_t newprot)
1111 struct mm_struct *mm = vma->vm_mm;
1114 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1116 entry = pmdp_get_and_clear(mm, addr, pmd);
1117 entry = pmd_modify(entry, newprot);
1118 set_pmd_at(mm, addr, pmd, entry);
1119 spin_unlock(&vma->vm_mm->page_table_lock);
1127 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1128 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1130 * Note that if it returns 1, this routine returns without unlocking page
1131 * table locks. So callers must unlock them.
1133 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1135 spin_lock(&vma->vm_mm->page_table_lock);
1136 if (likely(pmd_trans_huge(*pmd))) {
1137 if (unlikely(pmd_trans_splitting(*pmd))) {
1138 spin_unlock(&vma->vm_mm->page_table_lock);
1139 wait_split_huge_page(vma->anon_vma, pmd);
1142 /* Thp mapped by 'pmd' is stable, so we can
1143 * handle it as it is. */
1147 spin_unlock(&vma->vm_mm->page_table_lock);
1151 pmd_t *page_check_address_pmd(struct page *page,
1152 struct mm_struct *mm,
1153 unsigned long address,
1154 enum page_check_address_pmd_flag flag)
1158 pmd_t *pmd, *ret = NULL;
1160 if (address & ~HPAGE_PMD_MASK)
1163 pgd = pgd_offset(mm, address);
1164 if (!pgd_present(*pgd))
1167 pud = pud_offset(pgd, address);
1168 if (!pud_present(*pud))
1171 pmd = pmd_offset(pud, address);
1174 if (pmd_page(*pmd) != page)
1177 * split_vma() may create temporary aliased mappings. There is
1178 * no risk as long as all huge pmd are found and have their
1179 * splitting bit set before __split_huge_page_refcount
1180 * runs. Finding the same huge pmd more than once during the
1181 * same rmap walk is not a problem.
1183 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1184 pmd_trans_splitting(*pmd))
1186 if (pmd_trans_huge(*pmd)) {
1187 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1188 !pmd_trans_splitting(*pmd));
1195 static int __split_huge_page_splitting(struct page *page,
1196 struct vm_area_struct *vma,
1197 unsigned long address)
1199 struct mm_struct *mm = vma->vm_mm;
1203 spin_lock(&mm->page_table_lock);
1204 pmd = page_check_address_pmd(page, mm, address,
1205 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1208 * We can't temporarily set the pmd to null in order
1209 * to split it, the pmd must remain marked huge at all
1210 * times or the VM won't take the pmd_trans_huge paths
1211 * and it won't wait on the anon_vma->root->mutex to
1212 * serialize against split_huge_page*.
1214 pmdp_splitting_flush_notify(vma, address, pmd);
1217 spin_unlock(&mm->page_table_lock);
1222 static void __split_huge_page_refcount(struct page *page)
1225 struct zone *zone = page_zone(page);
1226 struct lruvec *lruvec;
1229 /* prevent PageLRU to go away from under us, and freeze lru stats */
1230 spin_lock_irq(&zone->lru_lock);
1231 lruvec = mem_cgroup_page_lruvec(page, zone);
1233 compound_lock(page);
1234 /* complete memcg works before add pages to LRU */
1235 mem_cgroup_split_huge_fixup(page);
1237 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1238 struct page *page_tail = page + i;
1240 /* tail_page->_mapcount cannot change */
1241 BUG_ON(page_mapcount(page_tail) < 0);
1242 tail_count += page_mapcount(page_tail);
1243 /* check for overflow */
1244 BUG_ON(tail_count < 0);
1245 BUG_ON(atomic_read(&page_tail->_count) != 0);
1247 * tail_page->_count is zero and not changing from
1248 * under us. But get_page_unless_zero() may be running
1249 * from under us on the tail_page. If we used
1250 * atomic_set() below instead of atomic_add(), we
1251 * would then run atomic_set() concurrently with
1252 * get_page_unless_zero(), and atomic_set() is
1253 * implemented in C not using locked ops. spin_unlock
1254 * on x86 sometime uses locked ops because of PPro
1255 * errata 66, 92, so unless somebody can guarantee
1256 * atomic_set() here would be safe on all archs (and
1257 * not only on x86), it's safer to use atomic_add().
1259 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1260 &page_tail->_count);
1262 /* after clearing PageTail the gup refcount can be released */
1266 * retain hwpoison flag of the poisoned tail page:
1267 * fix for the unsuitable process killed on Guest Machine(KVM)
1268 * by the memory-failure.
1270 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1271 page_tail->flags |= (page->flags &
1272 ((1L << PG_referenced) |
1273 (1L << PG_swapbacked) |
1274 (1L << PG_mlocked) |
1275 (1L << PG_uptodate)));
1276 page_tail->flags |= (1L << PG_dirty);
1278 /* clear PageTail before overwriting first_page */
1282 * __split_huge_page_splitting() already set the
1283 * splitting bit in all pmd that could map this
1284 * hugepage, that will ensure no CPU can alter the
1285 * mapcount on the head page. The mapcount is only
1286 * accounted in the head page and it has to be
1287 * transferred to all tail pages in the below code. So
1288 * for this code to be safe, the split the mapcount
1289 * can't change. But that doesn't mean userland can't
1290 * keep changing and reading the page contents while
1291 * we transfer the mapcount, so the pmd splitting
1292 * status is achieved setting a reserved bit in the
1293 * pmd, not by clearing the present bit.
1295 page_tail->_mapcount = page->_mapcount;
1297 BUG_ON(page_tail->mapping);
1298 page_tail->mapping = page->mapping;
1300 page_tail->index = page->index + i;
1302 BUG_ON(!PageAnon(page_tail));
1303 BUG_ON(!PageUptodate(page_tail));
1304 BUG_ON(!PageDirty(page_tail));
1305 BUG_ON(!PageSwapBacked(page_tail));
1307 lru_add_page_tail(page, page_tail, lruvec);
1309 atomic_sub(tail_count, &page->_count);
1310 BUG_ON(atomic_read(&page->_count) <= 0);
1312 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1313 __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1315 ClearPageCompound(page);
1316 compound_unlock(page);
1317 spin_unlock_irq(&zone->lru_lock);
1319 for (i = 1; i < HPAGE_PMD_NR; i++) {
1320 struct page *page_tail = page + i;
1321 BUG_ON(page_count(page_tail) <= 0);
1323 * Tail pages may be freed if there wasn't any mapping
1324 * like if add_to_swap() is running on a lru page that
1325 * had its mapping zapped. And freeing these pages
1326 * requires taking the lru_lock so we do the put_page
1327 * of the tail pages after the split is complete.
1329 put_page(page_tail);
1333 * Only the head page (now become a regular page) is required
1334 * to be pinned by the caller.
1336 BUG_ON(page_count(page) <= 0);
1339 static int __split_huge_page_map(struct page *page,
1340 struct vm_area_struct *vma,
1341 unsigned long address)
1343 struct mm_struct *mm = vma->vm_mm;
1347 unsigned long haddr;
1349 spin_lock(&mm->page_table_lock);
1350 pmd = page_check_address_pmd(page, mm, address,
1351 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1353 pgtable = get_pmd_huge_pte(mm);
1354 pmd_populate(mm, &_pmd, pgtable);
1356 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1357 i++, haddr += PAGE_SIZE) {
1359 BUG_ON(PageCompound(page+i));
1360 entry = mk_pte(page + i, vma->vm_page_prot);
1361 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1362 if (!pmd_write(*pmd))
1363 entry = pte_wrprotect(entry);
1365 BUG_ON(page_mapcount(page) != 1);
1366 if (!pmd_young(*pmd))
1367 entry = pte_mkold(entry);
1368 pte = pte_offset_map(&_pmd, haddr);
1369 BUG_ON(!pte_none(*pte));
1370 set_pte_at(mm, haddr, pte, entry);
1374 smp_wmb(); /* make pte visible before pmd */
1376 * Up to this point the pmd is present and huge and
1377 * userland has the whole access to the hugepage
1378 * during the split (which happens in place). If we
1379 * overwrite the pmd with the not-huge version
1380 * pointing to the pte here (which of course we could
1381 * if all CPUs were bug free), userland could trigger
1382 * a small page size TLB miss on the small sized TLB
1383 * while the hugepage TLB entry is still established
1384 * in the huge TLB. Some CPU doesn't like that. See
1385 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1386 * Erratum 383 on page 93. Intel should be safe but is
1387 * also warns that it's only safe if the permission
1388 * and cache attributes of the two entries loaded in
1389 * the two TLB is identical (which should be the case
1390 * here). But it is generally safer to never allow
1391 * small and huge TLB entries for the same virtual
1392 * address to be loaded simultaneously. So instead of
1393 * doing "pmd_populate(); flush_tlb_range();" we first
1394 * mark the current pmd notpresent (atomically because
1395 * here the pmd_trans_huge and pmd_trans_splitting
1396 * must remain set at all times on the pmd until the
1397 * split is complete for this pmd), then we flush the
1398 * SMP TLB and finally we write the non-huge version
1399 * of the pmd entry with pmd_populate.
1401 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1402 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1403 pmd_populate(mm, pmd, pgtable);
1406 spin_unlock(&mm->page_table_lock);
1411 /* must be called with anon_vma->root->mutex hold */
1412 static void __split_huge_page(struct page *page,
1413 struct anon_vma *anon_vma)
1415 int mapcount, mapcount2;
1416 struct anon_vma_chain *avc;
1418 BUG_ON(!PageHead(page));
1419 BUG_ON(PageTail(page));
1422 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1423 struct vm_area_struct *vma = avc->vma;
1424 unsigned long addr = vma_address(page, vma);
1425 BUG_ON(is_vma_temporary_stack(vma));
1426 if (addr == -EFAULT)
1428 mapcount += __split_huge_page_splitting(page, vma, addr);
1431 * It is critical that new vmas are added to the tail of the
1432 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1433 * and establishes a child pmd before
1434 * __split_huge_page_splitting() freezes the parent pmd (so if
1435 * we fail to prevent copy_huge_pmd() from running until the
1436 * whole __split_huge_page() is complete), we will still see
1437 * the newly established pmd of the child later during the
1438 * walk, to be able to set it as pmd_trans_splitting too.
1440 if (mapcount != page_mapcount(page))
1441 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1442 mapcount, page_mapcount(page));
1443 BUG_ON(mapcount != page_mapcount(page));
1445 __split_huge_page_refcount(page);
1448 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1449 struct vm_area_struct *vma = avc->vma;
1450 unsigned long addr = vma_address(page, vma);
1451 BUG_ON(is_vma_temporary_stack(vma));
1452 if (addr == -EFAULT)
1454 mapcount2 += __split_huge_page_map(page, vma, addr);
1456 if (mapcount != mapcount2)
1457 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1458 mapcount, mapcount2, page_mapcount(page));
1459 BUG_ON(mapcount != mapcount2);
1462 int split_huge_page(struct page *page)
1464 struct anon_vma *anon_vma;
1467 BUG_ON(!PageAnon(page));
1468 anon_vma = page_lock_anon_vma(page);
1472 if (!PageCompound(page))
1475 BUG_ON(!PageSwapBacked(page));
1476 __split_huge_page(page, anon_vma);
1477 count_vm_event(THP_SPLIT);
1479 BUG_ON(PageCompound(page));
1481 page_unlock_anon_vma(anon_vma);
1486 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1488 int hugepage_madvise(struct vm_area_struct *vma,
1489 unsigned long *vm_flags, int advice)
1494 * Be somewhat over-protective like KSM for now!
1496 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1498 *vm_flags &= ~VM_NOHUGEPAGE;
1499 *vm_flags |= VM_HUGEPAGE;
1501 * If the vma become good for khugepaged to scan,
1502 * register it here without waiting a page fault that
1503 * may not happen any time soon.
1505 if (unlikely(khugepaged_enter_vma_merge(vma)))
1508 case MADV_NOHUGEPAGE:
1510 * Be somewhat over-protective like KSM for now!
1512 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1514 *vm_flags &= ~VM_HUGEPAGE;
1515 *vm_flags |= VM_NOHUGEPAGE;
1517 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1518 * this vma even if we leave the mm registered in khugepaged if
1519 * it got registered before VM_NOHUGEPAGE was set.
1527 static int __init khugepaged_slab_init(void)
1529 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1530 sizeof(struct mm_slot),
1531 __alignof__(struct mm_slot), 0, NULL);
1538 static void __init khugepaged_slab_free(void)
1540 kmem_cache_destroy(mm_slot_cache);
1541 mm_slot_cache = NULL;
1544 static inline struct mm_slot *alloc_mm_slot(void)
1546 if (!mm_slot_cache) /* initialization failed */
1548 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1551 static inline void free_mm_slot(struct mm_slot *mm_slot)
1553 kmem_cache_free(mm_slot_cache, mm_slot);
1556 static int __init mm_slots_hash_init(void)
1558 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1566 static void __init mm_slots_hash_free(void)
1568 kfree(mm_slots_hash);
1569 mm_slots_hash = NULL;
1573 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1575 struct mm_slot *mm_slot;
1576 struct hlist_head *bucket;
1577 struct hlist_node *node;
1579 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1580 % MM_SLOTS_HASH_HEADS];
1581 hlist_for_each_entry(mm_slot, node, bucket, hash) {
1582 if (mm == mm_slot->mm)
1588 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1589 struct mm_slot *mm_slot)
1591 struct hlist_head *bucket;
1593 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1594 % MM_SLOTS_HASH_HEADS];
1596 hlist_add_head(&mm_slot->hash, bucket);
1599 static inline int khugepaged_test_exit(struct mm_struct *mm)
1601 return atomic_read(&mm->mm_users) == 0;
1604 int __khugepaged_enter(struct mm_struct *mm)
1606 struct mm_slot *mm_slot;
1609 mm_slot = alloc_mm_slot();
1613 /* __khugepaged_exit() must not run from under us */
1614 VM_BUG_ON(khugepaged_test_exit(mm));
1615 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1616 free_mm_slot(mm_slot);
1620 spin_lock(&khugepaged_mm_lock);
1621 insert_to_mm_slots_hash(mm, mm_slot);
1623 * Insert just behind the scanning cursor, to let the area settle
1626 wakeup = list_empty(&khugepaged_scan.mm_head);
1627 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1628 spin_unlock(&khugepaged_mm_lock);
1630 atomic_inc(&mm->mm_count);
1632 wake_up_interruptible(&khugepaged_wait);
1637 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1639 unsigned long hstart, hend;
1642 * Not yet faulted in so we will register later in the
1643 * page fault if needed.
1647 /* khugepaged not yet working on file or special mappings */
1649 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1650 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1651 hend = vma->vm_end & HPAGE_PMD_MASK;
1653 return khugepaged_enter(vma);
1657 void __khugepaged_exit(struct mm_struct *mm)
1659 struct mm_slot *mm_slot;
1662 spin_lock(&khugepaged_mm_lock);
1663 mm_slot = get_mm_slot(mm);
1664 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1665 hlist_del(&mm_slot->hash);
1666 list_del(&mm_slot->mm_node);
1669 spin_unlock(&khugepaged_mm_lock);
1672 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1673 free_mm_slot(mm_slot);
1675 } else if (mm_slot) {
1677 * This is required to serialize against
1678 * khugepaged_test_exit() (which is guaranteed to run
1679 * under mmap sem read mode). Stop here (after we
1680 * return all pagetables will be destroyed) until
1681 * khugepaged has finished working on the pagetables
1682 * under the mmap_sem.
1684 down_write(&mm->mmap_sem);
1685 up_write(&mm->mmap_sem);
1689 static void release_pte_page(struct page *page)
1691 /* 0 stands for page_is_file_cache(page) == false */
1692 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1694 putback_lru_page(page);
1697 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1699 while (--_pte >= pte) {
1700 pte_t pteval = *_pte;
1701 if (!pte_none(pteval))
1702 release_pte_page(pte_page(pteval));
1706 static void release_all_pte_pages(pte_t *pte)
1708 release_pte_pages(pte, pte + HPAGE_PMD_NR);
1711 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1712 unsigned long address,
1717 int referenced = 0, isolated = 0, none = 0;
1718 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1719 _pte++, address += PAGE_SIZE) {
1720 pte_t pteval = *_pte;
1721 if (pte_none(pteval)) {
1722 if (++none <= khugepaged_max_ptes_none)
1725 release_pte_pages(pte, _pte);
1729 if (!pte_present(pteval) || !pte_write(pteval)) {
1730 release_pte_pages(pte, _pte);
1733 page = vm_normal_page(vma, address, pteval);
1734 if (unlikely(!page)) {
1735 release_pte_pages(pte, _pte);
1738 VM_BUG_ON(PageCompound(page));
1739 BUG_ON(!PageAnon(page));
1740 VM_BUG_ON(!PageSwapBacked(page));
1742 /* cannot use mapcount: can't collapse if there's a gup pin */
1743 if (page_count(page) != 1) {
1744 release_pte_pages(pte, _pte);
1748 * We can do it before isolate_lru_page because the
1749 * page can't be freed from under us. NOTE: PG_lock
1750 * is needed to serialize against split_huge_page
1751 * when invoked from the VM.
1753 if (!trylock_page(page)) {
1754 release_pte_pages(pte, _pte);
1758 * Isolate the page to avoid collapsing an hugepage
1759 * currently in use by the VM.
1761 if (isolate_lru_page(page)) {
1763 release_pte_pages(pte, _pte);
1766 /* 0 stands for page_is_file_cache(page) == false */
1767 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1768 VM_BUG_ON(!PageLocked(page));
1769 VM_BUG_ON(PageLRU(page));
1771 /* If there is no mapped pte young don't collapse the page */
1772 if (pte_young(pteval) || PageReferenced(page) ||
1773 mmu_notifier_test_young(vma->vm_mm, address))
1776 if (unlikely(!referenced))
1777 release_all_pte_pages(pte);
1784 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1785 struct vm_area_struct *vma,
1786 unsigned long address,
1790 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1791 pte_t pteval = *_pte;
1792 struct page *src_page;
1794 if (pte_none(pteval)) {
1795 clear_user_highpage(page, address);
1796 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1798 src_page = pte_page(pteval);
1799 copy_user_highpage(page, src_page, address, vma);
1800 VM_BUG_ON(page_mapcount(src_page) != 1);
1801 release_pte_page(src_page);
1803 * ptl mostly unnecessary, but preempt has to
1804 * be disabled to update the per-cpu stats
1805 * inside page_remove_rmap().
1809 * paravirt calls inside pte_clear here are
1812 pte_clear(vma->vm_mm, address, _pte);
1813 page_remove_rmap(src_page);
1815 free_page_and_swap_cache(src_page);
1818 address += PAGE_SIZE;
1823 static void khugepaged_alloc_sleep(void)
1825 wait_event_freezable_timeout(khugepaged_wait, false,
1826 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
1830 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1832 if (IS_ERR(*hpage)) {
1837 khugepaged_alloc_sleep();
1838 } else if (*hpage) {
1847 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1848 struct vm_area_struct *vma, unsigned long address,
1853 * Allocate the page while the vma is still valid and under
1854 * the mmap_sem read mode so there is no memory allocation
1855 * later when we take the mmap_sem in write mode. This is more
1856 * friendly behavior (OTOH it may actually hide bugs) to
1857 * filesystems in userland with daemons allocating memory in
1858 * the userland I/O paths. Allocating memory with the
1859 * mmap_sem in read mode is good idea also to allow greater
1862 *hpage = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1863 node, __GFP_OTHER_NODE);
1866 * After allocating the hugepage, release the mmap_sem read lock in
1867 * preparation for taking it in write mode.
1869 up_read(&mm->mmap_sem);
1870 if (unlikely(!*hpage)) {
1871 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1872 *hpage = ERR_PTR(-ENOMEM);
1876 count_vm_event(THP_COLLAPSE_ALLOC);
1880 static struct page *khugepaged_alloc_hugepage(bool *wait)
1885 hpage = alloc_hugepage(khugepaged_defrag());
1887 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1892 khugepaged_alloc_sleep();
1894 count_vm_event(THP_COLLAPSE_ALLOC);
1895 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
1900 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1903 *hpage = khugepaged_alloc_hugepage(wait);
1905 if (unlikely(!*hpage))
1912 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1913 struct vm_area_struct *vma, unsigned long address,
1916 up_read(&mm->mmap_sem);
1922 static void collapse_huge_page(struct mm_struct *mm,
1923 unsigned long address,
1924 struct page **hpage,
1925 struct vm_area_struct *vma,
1933 struct page *new_page;
1936 unsigned long hstart, hend;
1938 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1940 /* release the mmap_sem read lock. */
1941 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
1945 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
1949 * Prevent all access to pagetables with the exception of
1950 * gup_fast later hanlded by the ptep_clear_flush and the VM
1951 * handled by the anon_vma lock + PG_lock.
1953 down_write(&mm->mmap_sem);
1954 if (unlikely(khugepaged_test_exit(mm)))
1957 vma = find_vma(mm, address);
1958 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1959 hend = vma->vm_end & HPAGE_PMD_MASK;
1960 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1963 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1964 (vma->vm_flags & VM_NOHUGEPAGE))
1967 if (!vma->anon_vma || vma->vm_ops)
1969 if (is_vma_temporary_stack(vma))
1971 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1973 pgd = pgd_offset(mm, address);
1974 if (!pgd_present(*pgd))
1977 pud = pud_offset(pgd, address);
1978 if (!pud_present(*pud))
1981 pmd = pmd_offset(pud, address);
1982 /* pmd can't go away or become huge under us */
1983 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1986 anon_vma_lock(vma->anon_vma);
1988 pte = pte_offset_map(pmd, address);
1989 ptl = pte_lockptr(mm, pmd);
1991 spin_lock(&mm->page_table_lock); /* probably unnecessary */
1993 * After this gup_fast can't run anymore. This also removes
1994 * any huge TLB entry from the CPU so we won't allow
1995 * huge and small TLB entries for the same virtual address
1996 * to avoid the risk of CPU bugs in that area.
1998 _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1999 spin_unlock(&mm->page_table_lock);
2002 isolated = __collapse_huge_page_isolate(vma, address, pte);
2005 if (unlikely(!isolated)) {
2007 spin_lock(&mm->page_table_lock);
2008 BUG_ON(!pmd_none(*pmd));
2009 set_pmd_at(mm, address, pmd, _pmd);
2010 spin_unlock(&mm->page_table_lock);
2011 anon_vma_unlock(vma->anon_vma);
2016 * All pages are isolated and locked so anon_vma rmap
2017 * can't run anymore.
2019 anon_vma_unlock(vma->anon_vma);
2021 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2023 __SetPageUptodate(new_page);
2024 pgtable = pmd_pgtable(_pmd);
2025 VM_BUG_ON(page_count(pgtable) != 1);
2026 VM_BUG_ON(page_mapcount(pgtable) != 0);
2028 _pmd = mk_pmd(new_page, vma->vm_page_prot);
2029 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2030 _pmd = pmd_mkhuge(_pmd);
2033 * spin_lock() below is not the equivalent of smp_wmb(), so
2034 * this is needed to avoid the copy_huge_page writes to become
2035 * visible after the set_pmd_at() write.
2039 spin_lock(&mm->page_table_lock);
2040 BUG_ON(!pmd_none(*pmd));
2041 page_add_new_anon_rmap(new_page, vma, address);
2042 set_pmd_at(mm, address, pmd, _pmd);
2043 update_mmu_cache(vma, address, _pmd);
2044 prepare_pmd_huge_pte(pgtable, mm);
2045 spin_unlock(&mm->page_table_lock);
2049 khugepaged_pages_collapsed++;
2051 up_write(&mm->mmap_sem);
2055 mem_cgroup_uncharge_page(new_page);
2059 static int khugepaged_scan_pmd(struct mm_struct *mm,
2060 struct vm_area_struct *vma,
2061 unsigned long address,
2062 struct page **hpage)
2068 int ret = 0, referenced = 0, none = 0;
2070 unsigned long _address;
2074 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2076 pgd = pgd_offset(mm, address);
2077 if (!pgd_present(*pgd))
2080 pud = pud_offset(pgd, address);
2081 if (!pud_present(*pud))
2084 pmd = pmd_offset(pud, address);
2085 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2088 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2089 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2090 _pte++, _address += PAGE_SIZE) {
2091 pte_t pteval = *_pte;
2092 if (pte_none(pteval)) {
2093 if (++none <= khugepaged_max_ptes_none)
2098 if (!pte_present(pteval) || !pte_write(pteval))
2100 page = vm_normal_page(vma, _address, pteval);
2101 if (unlikely(!page))
2104 * Chose the node of the first page. This could
2105 * be more sophisticated and look at more pages,
2106 * but isn't for now.
2109 node = page_to_nid(page);
2110 VM_BUG_ON(PageCompound(page));
2111 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2113 /* cannot use mapcount: can't collapse if there's a gup pin */
2114 if (page_count(page) != 1)
2116 if (pte_young(pteval) || PageReferenced(page) ||
2117 mmu_notifier_test_young(vma->vm_mm, address))
2123 pte_unmap_unlock(pte, ptl);
2125 /* collapse_huge_page will return with the mmap_sem released */
2126 collapse_huge_page(mm, address, hpage, vma, node);
2131 static void collect_mm_slot(struct mm_slot *mm_slot)
2133 struct mm_struct *mm = mm_slot->mm;
2135 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2137 if (khugepaged_test_exit(mm)) {
2139 hlist_del(&mm_slot->hash);
2140 list_del(&mm_slot->mm_node);
2143 * Not strictly needed because the mm exited already.
2145 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2148 /* khugepaged_mm_lock actually not necessary for the below */
2149 free_mm_slot(mm_slot);
2154 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2155 struct page **hpage)
2156 __releases(&khugepaged_mm_lock)
2157 __acquires(&khugepaged_mm_lock)
2159 struct mm_slot *mm_slot;
2160 struct mm_struct *mm;
2161 struct vm_area_struct *vma;
2165 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2167 if (khugepaged_scan.mm_slot)
2168 mm_slot = khugepaged_scan.mm_slot;
2170 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2171 struct mm_slot, mm_node);
2172 khugepaged_scan.address = 0;
2173 khugepaged_scan.mm_slot = mm_slot;
2175 spin_unlock(&khugepaged_mm_lock);
2178 down_read(&mm->mmap_sem);
2179 if (unlikely(khugepaged_test_exit(mm)))
2182 vma = find_vma(mm, khugepaged_scan.address);
2185 for (; vma; vma = vma->vm_next) {
2186 unsigned long hstart, hend;
2189 if (unlikely(khugepaged_test_exit(mm))) {
2194 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2195 !khugepaged_always()) ||
2196 (vma->vm_flags & VM_NOHUGEPAGE)) {
2201 if (!vma->anon_vma || vma->vm_ops)
2203 if (is_vma_temporary_stack(vma))
2205 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2207 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2208 hend = vma->vm_end & HPAGE_PMD_MASK;
2211 if (khugepaged_scan.address > hend)
2213 if (khugepaged_scan.address < hstart)
2214 khugepaged_scan.address = hstart;
2215 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2217 while (khugepaged_scan.address < hend) {
2220 if (unlikely(khugepaged_test_exit(mm)))
2221 goto breakouterloop;
2223 VM_BUG_ON(khugepaged_scan.address < hstart ||
2224 khugepaged_scan.address + HPAGE_PMD_SIZE >
2226 ret = khugepaged_scan_pmd(mm, vma,
2227 khugepaged_scan.address,
2229 /* move to next address */
2230 khugepaged_scan.address += HPAGE_PMD_SIZE;
2231 progress += HPAGE_PMD_NR;
2233 /* we released mmap_sem so break loop */
2234 goto breakouterloop_mmap_sem;
2235 if (progress >= pages)
2236 goto breakouterloop;
2240 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2241 breakouterloop_mmap_sem:
2243 spin_lock(&khugepaged_mm_lock);
2244 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2246 * Release the current mm_slot if this mm is about to die, or
2247 * if we scanned all vmas of this mm.
2249 if (khugepaged_test_exit(mm) || !vma) {
2251 * Make sure that if mm_users is reaching zero while
2252 * khugepaged runs here, khugepaged_exit will find
2253 * mm_slot not pointing to the exiting mm.
2255 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2256 khugepaged_scan.mm_slot = list_entry(
2257 mm_slot->mm_node.next,
2258 struct mm_slot, mm_node);
2259 khugepaged_scan.address = 0;
2261 khugepaged_scan.mm_slot = NULL;
2262 khugepaged_full_scans++;
2265 collect_mm_slot(mm_slot);
2271 static int khugepaged_has_work(void)
2273 return !list_empty(&khugepaged_scan.mm_head) &&
2274 khugepaged_enabled();
2277 static int khugepaged_wait_event(void)
2279 return !list_empty(&khugepaged_scan.mm_head) ||
2280 kthread_should_stop();
2283 static void khugepaged_do_scan(void)
2285 struct page *hpage = NULL;
2286 unsigned int progress = 0, pass_through_head = 0;
2287 unsigned int pages = khugepaged_pages_to_scan;
2290 barrier(); /* write khugepaged_pages_to_scan to local stack */
2292 while (progress < pages) {
2293 if (!khugepaged_prealloc_page(&hpage, &wait))
2298 if (unlikely(kthread_should_stop() || freezing(current)))
2301 spin_lock(&khugepaged_mm_lock);
2302 if (!khugepaged_scan.mm_slot)
2303 pass_through_head++;
2304 if (khugepaged_has_work() &&
2305 pass_through_head < 2)
2306 progress += khugepaged_scan_mm_slot(pages - progress,
2310 spin_unlock(&khugepaged_mm_lock);
2313 if (!IS_ERR_OR_NULL(hpage))
2317 static void khugepaged_wait_work(void)
2321 if (khugepaged_has_work()) {
2322 if (!khugepaged_scan_sleep_millisecs)
2325 wait_event_freezable_timeout(khugepaged_wait,
2326 kthread_should_stop(),
2327 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2331 if (khugepaged_enabled())
2332 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2335 static int khugepaged(void *none)
2337 struct mm_slot *mm_slot;
2340 set_user_nice(current, 19);
2342 while (!kthread_should_stop()) {
2343 khugepaged_do_scan();
2344 khugepaged_wait_work();
2347 spin_lock(&khugepaged_mm_lock);
2348 mm_slot = khugepaged_scan.mm_slot;
2349 khugepaged_scan.mm_slot = NULL;
2351 collect_mm_slot(mm_slot);
2352 spin_unlock(&khugepaged_mm_lock);
2356 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2360 spin_lock(&mm->page_table_lock);
2361 if (unlikely(!pmd_trans_huge(*pmd))) {
2362 spin_unlock(&mm->page_table_lock);
2365 page = pmd_page(*pmd);
2366 VM_BUG_ON(!page_count(page));
2368 spin_unlock(&mm->page_table_lock);
2370 split_huge_page(page);
2373 BUG_ON(pmd_trans_huge(*pmd));
2376 static void split_huge_page_address(struct mm_struct *mm,
2377 unsigned long address)
2383 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2385 pgd = pgd_offset(mm, address);
2386 if (!pgd_present(*pgd))
2389 pud = pud_offset(pgd, address);
2390 if (!pud_present(*pud))
2393 pmd = pmd_offset(pud, address);
2394 if (!pmd_present(*pmd))
2397 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2398 * materialize from under us.
2400 split_huge_page_pmd(mm, pmd);
2403 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2404 unsigned long start,
2409 * If the new start address isn't hpage aligned and it could
2410 * previously contain an hugepage: check if we need to split
2413 if (start & ~HPAGE_PMD_MASK &&
2414 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2415 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2416 split_huge_page_address(vma->vm_mm, start);
2419 * If the new end address isn't hpage aligned and it could
2420 * previously contain an hugepage: check if we need to split
2423 if (end & ~HPAGE_PMD_MASK &&
2424 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2425 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2426 split_huge_page_address(vma->vm_mm, end);
2429 * If we're also updating the vma->vm_next->vm_start, if the new
2430 * vm_next->vm_start isn't page aligned and it could previously
2431 * contain an hugepage: check if we need to split an huge pmd.
2433 if (adjust_next > 0) {
2434 struct vm_area_struct *next = vma->vm_next;
2435 unsigned long nstart = next->vm_start;
2436 nstart += adjust_next << PAGE_SHIFT;
2437 if (nstart & ~HPAGE_PMD_MASK &&
2438 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2439 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2440 split_huge_page_address(next->vm_mm, nstart);