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 if (test_bit(flag, &transparent_hugepage_flags))
248 return sprintf(buf, "[yes] no\n");
250 return sprintf(buf, "yes [no]\n");
252 static ssize_t single_flag_store(struct kobject *kobj,
253 struct kobj_attribute *attr,
254 const char *buf, size_t count,
255 enum transparent_hugepage_flag flag)
257 if (!memcmp("yes", buf,
258 min(sizeof("yes")-1, count))) {
259 set_bit(flag, &transparent_hugepage_flags);
260 } else if (!memcmp("no", buf,
261 min(sizeof("no")-1, count))) {
262 clear_bit(flag, &transparent_hugepage_flags);
270 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
271 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
272 * memory just to allocate one more hugepage.
274 static ssize_t defrag_show(struct kobject *kobj,
275 struct kobj_attribute *attr, char *buf)
277 return double_flag_show(kobj, attr, buf,
278 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
279 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
281 static ssize_t defrag_store(struct kobject *kobj,
282 struct kobj_attribute *attr,
283 const char *buf, size_t count)
285 return double_flag_store(kobj, attr, buf, count,
286 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
287 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
289 static struct kobj_attribute defrag_attr =
290 __ATTR(defrag, 0644, defrag_show, defrag_store);
292 #ifdef CONFIG_DEBUG_VM
293 static ssize_t debug_cow_show(struct kobject *kobj,
294 struct kobj_attribute *attr, char *buf)
296 return single_flag_show(kobj, attr, buf,
297 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
299 static ssize_t debug_cow_store(struct kobject *kobj,
300 struct kobj_attribute *attr,
301 const char *buf, size_t count)
303 return single_flag_store(kobj, attr, buf, count,
304 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
306 static struct kobj_attribute debug_cow_attr =
307 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
308 #endif /* CONFIG_DEBUG_VM */
310 static struct attribute *hugepage_attr[] = {
313 #ifdef CONFIG_DEBUG_VM
314 &debug_cow_attr.attr,
319 static struct attribute_group hugepage_attr_group = {
320 .attrs = hugepage_attr,
323 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
324 struct kobj_attribute *attr,
327 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
330 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
331 struct kobj_attribute *attr,
332 const char *buf, size_t count)
337 err = strict_strtoul(buf, 10, &msecs);
338 if (err || msecs > UINT_MAX)
341 khugepaged_scan_sleep_millisecs = msecs;
342 wake_up_interruptible(&khugepaged_wait);
346 static struct kobj_attribute scan_sleep_millisecs_attr =
347 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
348 scan_sleep_millisecs_store);
350 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
351 struct kobj_attribute *attr,
354 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
357 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
358 struct kobj_attribute *attr,
359 const char *buf, size_t count)
364 err = strict_strtoul(buf, 10, &msecs);
365 if (err || msecs > UINT_MAX)
368 khugepaged_alloc_sleep_millisecs = msecs;
369 wake_up_interruptible(&khugepaged_wait);
373 static struct kobj_attribute alloc_sleep_millisecs_attr =
374 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
375 alloc_sleep_millisecs_store);
377 static ssize_t pages_to_scan_show(struct kobject *kobj,
378 struct kobj_attribute *attr,
381 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
383 static ssize_t pages_to_scan_store(struct kobject *kobj,
384 struct kobj_attribute *attr,
385 const char *buf, size_t count)
390 err = strict_strtoul(buf, 10, &pages);
391 if (err || !pages || pages > UINT_MAX)
394 khugepaged_pages_to_scan = pages;
398 static struct kobj_attribute pages_to_scan_attr =
399 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
400 pages_to_scan_store);
402 static ssize_t pages_collapsed_show(struct kobject *kobj,
403 struct kobj_attribute *attr,
406 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
408 static struct kobj_attribute pages_collapsed_attr =
409 __ATTR_RO(pages_collapsed);
411 static ssize_t full_scans_show(struct kobject *kobj,
412 struct kobj_attribute *attr,
415 return sprintf(buf, "%u\n", khugepaged_full_scans);
417 static struct kobj_attribute full_scans_attr =
418 __ATTR_RO(full_scans);
420 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
421 struct kobj_attribute *attr, char *buf)
423 return single_flag_show(kobj, attr, buf,
424 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
426 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
427 struct kobj_attribute *attr,
428 const char *buf, size_t count)
430 return single_flag_store(kobj, attr, buf, count,
431 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
433 static struct kobj_attribute khugepaged_defrag_attr =
434 __ATTR(defrag, 0644, khugepaged_defrag_show,
435 khugepaged_defrag_store);
438 * max_ptes_none controls if khugepaged should collapse hugepages over
439 * any unmapped ptes in turn potentially increasing the memory
440 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
441 * reduce the available free memory in the system as it
442 * runs. Increasing max_ptes_none will instead potentially reduce the
443 * free memory in the system during the khugepaged scan.
445 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
446 struct kobj_attribute *attr,
449 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
451 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
452 struct kobj_attribute *attr,
453 const char *buf, size_t count)
456 unsigned long max_ptes_none;
458 err = strict_strtoul(buf, 10, &max_ptes_none);
459 if (err || max_ptes_none > HPAGE_PMD_NR-1)
462 khugepaged_max_ptes_none = max_ptes_none;
466 static struct kobj_attribute khugepaged_max_ptes_none_attr =
467 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
468 khugepaged_max_ptes_none_store);
470 static struct attribute *khugepaged_attr[] = {
471 &khugepaged_defrag_attr.attr,
472 &khugepaged_max_ptes_none_attr.attr,
473 &pages_to_scan_attr.attr,
474 &pages_collapsed_attr.attr,
475 &full_scans_attr.attr,
476 &scan_sleep_millisecs_attr.attr,
477 &alloc_sleep_millisecs_attr.attr,
481 static struct attribute_group khugepaged_attr_group = {
482 .attrs = khugepaged_attr,
483 .name = "khugepaged",
485 #endif /* CONFIG_SYSFS */
487 static int __init hugepage_init(void)
491 static struct kobject *hugepage_kobj;
495 if (!has_transparent_hugepage()) {
496 transparent_hugepage_flags = 0;
502 hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
503 if (unlikely(!hugepage_kobj)) {
504 printk(KERN_ERR "hugepage: failed kobject create\n");
508 err = sysfs_create_group(hugepage_kobj, &hugepage_attr_group);
510 printk(KERN_ERR "hugepage: failed register hugeage group\n");
514 err = sysfs_create_group(hugepage_kobj, &khugepaged_attr_group);
516 printk(KERN_ERR "hugepage: failed register hugeage group\n");
521 err = khugepaged_slab_init();
525 err = mm_slots_hash_init();
527 khugepaged_slab_free();
532 * By default disable transparent hugepages on smaller systems,
533 * where the extra memory used could hurt more than TLB overhead
534 * is likely to save. The admin can still enable it through /sys.
536 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
537 transparent_hugepage_flags = 0;
541 set_recommended_min_free_kbytes();
546 module_init(hugepage_init)
548 static int __init setup_transparent_hugepage(char *str)
553 if (!strcmp(str, "always")) {
554 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
555 &transparent_hugepage_flags);
556 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
557 &transparent_hugepage_flags);
559 } else if (!strcmp(str, "madvise")) {
560 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
561 &transparent_hugepage_flags);
562 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
563 &transparent_hugepage_flags);
565 } else if (!strcmp(str, "never")) {
566 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
567 &transparent_hugepage_flags);
568 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
569 &transparent_hugepage_flags);
575 "transparent_hugepage= cannot parse, ignored\n");
578 __setup("transparent_hugepage=", setup_transparent_hugepage);
580 static void prepare_pmd_huge_pte(pgtable_t pgtable,
581 struct mm_struct *mm)
583 assert_spin_locked(&mm->page_table_lock);
586 if (!mm->pmd_huge_pte)
587 INIT_LIST_HEAD(&pgtable->lru);
589 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
590 mm->pmd_huge_pte = pgtable;
593 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
595 if (likely(vma->vm_flags & VM_WRITE))
596 pmd = pmd_mkwrite(pmd);
600 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
601 struct vm_area_struct *vma,
602 unsigned long haddr, pmd_t *pmd,
608 VM_BUG_ON(!PageCompound(page));
609 pgtable = pte_alloc_one(mm, haddr);
610 if (unlikely(!pgtable)) {
611 mem_cgroup_uncharge_page(page);
616 clear_huge_page(page, haddr, HPAGE_PMD_NR);
617 __SetPageUptodate(page);
619 spin_lock(&mm->page_table_lock);
620 if (unlikely(!pmd_none(*pmd))) {
621 spin_unlock(&mm->page_table_lock);
622 mem_cgroup_uncharge_page(page);
624 pte_free(mm, pgtable);
627 entry = mk_pmd(page, vma->vm_page_prot);
628 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
629 entry = pmd_mkhuge(entry);
631 * The spinlocking to take the lru_lock inside
632 * page_add_new_anon_rmap() acts as a full memory
633 * barrier to be sure clear_huge_page writes become
634 * visible after the set_pmd_at() write.
636 page_add_new_anon_rmap(page, vma, haddr);
637 set_pmd_at(mm, haddr, pmd, entry);
638 prepare_pmd_huge_pte(pgtable, mm);
639 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
640 spin_unlock(&mm->page_table_lock);
646 static inline gfp_t alloc_hugepage_gfpmask(int defrag)
648 return GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT);
651 static inline struct page *alloc_hugepage_vma(int defrag,
652 struct vm_area_struct *vma,
655 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag),
656 HPAGE_PMD_ORDER, vma, haddr);
660 static inline struct page *alloc_hugepage(int defrag)
662 return alloc_pages(alloc_hugepage_gfpmask(defrag),
667 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
668 unsigned long address, pmd_t *pmd,
672 unsigned long haddr = address & HPAGE_PMD_MASK;
675 if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
676 if (unlikely(anon_vma_prepare(vma)))
678 if (unlikely(khugepaged_enter(vma)))
680 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
684 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
689 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
693 * Use __pte_alloc instead of pte_alloc_map, because we can't
694 * run pte_offset_map on the pmd, if an huge pmd could
695 * materialize from under us from a different thread.
697 if (unlikely(__pte_alloc(mm, vma, pmd, address)))
699 /* if an huge pmd materialized from under us just retry later */
700 if (unlikely(pmd_trans_huge(*pmd)))
703 * A regular pmd is established and it can't morph into a huge pmd
704 * from under us anymore at this point because we hold the mmap_sem
705 * read mode and khugepaged takes it in write mode. So now it's
706 * safe to run pte_offset_map().
708 pte = pte_offset_map(pmd, address);
709 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
712 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
713 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
714 struct vm_area_struct *vma)
716 struct page *src_page;
722 pgtable = pte_alloc_one(dst_mm, addr);
723 if (unlikely(!pgtable))
726 spin_lock(&dst_mm->page_table_lock);
727 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
731 if (unlikely(!pmd_trans_huge(pmd))) {
732 pte_free(dst_mm, pgtable);
735 if (unlikely(pmd_trans_splitting(pmd))) {
736 /* split huge page running from under us */
737 spin_unlock(&src_mm->page_table_lock);
738 spin_unlock(&dst_mm->page_table_lock);
739 pte_free(dst_mm, pgtable);
741 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
744 src_page = pmd_page(pmd);
745 VM_BUG_ON(!PageHead(src_page));
747 page_dup_rmap(src_page);
748 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
750 pmdp_set_wrprotect(src_mm, addr, src_pmd);
751 pmd = pmd_mkold(pmd_wrprotect(pmd));
752 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
753 prepare_pmd_huge_pte(pgtable, dst_mm);
757 spin_unlock(&src_mm->page_table_lock);
758 spin_unlock(&dst_mm->page_table_lock);
763 /* no "address" argument so destroys page coloring of some arch */
764 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
768 assert_spin_locked(&mm->page_table_lock);
771 pgtable = mm->pmd_huge_pte;
772 if (list_empty(&pgtable->lru))
773 mm->pmd_huge_pte = NULL;
775 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
777 list_del(&pgtable->lru);
782 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
783 struct vm_area_struct *vma,
784 unsigned long address,
785 pmd_t *pmd, pmd_t orig_pmd,
794 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
796 if (unlikely(!pages)) {
801 for (i = 0; i < HPAGE_PMD_NR; i++) {
802 pages[i] = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
804 if (unlikely(!pages[i] ||
805 mem_cgroup_newpage_charge(pages[i], mm,
809 mem_cgroup_uncharge_start();
811 mem_cgroup_uncharge_page(pages[i]);
814 mem_cgroup_uncharge_end();
821 for (i = 0; i < HPAGE_PMD_NR; i++) {
822 copy_user_highpage(pages[i], page + i,
823 haddr + PAGE_SHIFT*i, vma);
824 __SetPageUptodate(pages[i]);
828 spin_lock(&mm->page_table_lock);
829 if (unlikely(!pmd_same(*pmd, orig_pmd)))
831 VM_BUG_ON(!PageHead(page));
833 pmdp_clear_flush_notify(vma, haddr, pmd);
834 /* leave pmd empty until pte is filled */
836 pgtable = get_pmd_huge_pte(mm);
837 pmd_populate(mm, &_pmd, pgtable);
839 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
841 entry = mk_pte(pages[i], vma->vm_page_prot);
842 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
843 page_add_new_anon_rmap(pages[i], vma, haddr);
844 pte = pte_offset_map(&_pmd, haddr);
845 VM_BUG_ON(!pte_none(*pte));
846 set_pte_at(mm, haddr, pte, entry);
852 smp_wmb(); /* make pte visible before pmd */
853 pmd_populate(mm, pmd, pgtable);
854 page_remove_rmap(page);
855 spin_unlock(&mm->page_table_lock);
857 ret |= VM_FAULT_WRITE;
864 spin_unlock(&mm->page_table_lock);
865 mem_cgroup_uncharge_start();
866 for (i = 0; i < HPAGE_PMD_NR; i++) {
867 mem_cgroup_uncharge_page(pages[i]);
870 mem_cgroup_uncharge_end();
875 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
876 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
879 struct page *page, *new_page;
882 VM_BUG_ON(!vma->anon_vma);
883 spin_lock(&mm->page_table_lock);
884 if (unlikely(!pmd_same(*pmd, orig_pmd)))
887 page = pmd_page(orig_pmd);
888 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
889 haddr = address & HPAGE_PMD_MASK;
890 if (page_mapcount(page) == 1) {
892 entry = pmd_mkyoung(orig_pmd);
893 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
894 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
895 update_mmu_cache(vma, address, entry);
896 ret |= VM_FAULT_WRITE;
900 spin_unlock(&mm->page_table_lock);
902 if (transparent_hugepage_enabled(vma) &&
903 !transparent_hugepage_debug_cow())
904 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
909 if (unlikely(!new_page)) {
910 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
911 pmd, orig_pmd, page, haddr);
916 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
923 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
924 __SetPageUptodate(new_page);
926 spin_lock(&mm->page_table_lock);
928 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
929 mem_cgroup_uncharge_page(new_page);
933 VM_BUG_ON(!PageHead(page));
934 entry = mk_pmd(new_page, vma->vm_page_prot);
935 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
936 entry = pmd_mkhuge(entry);
937 pmdp_clear_flush_notify(vma, haddr, pmd);
938 page_add_new_anon_rmap(new_page, vma, haddr);
939 set_pmd_at(mm, haddr, pmd, entry);
940 update_mmu_cache(vma, address, entry);
941 page_remove_rmap(page);
943 ret |= VM_FAULT_WRITE;
946 spin_unlock(&mm->page_table_lock);
951 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
956 struct page *page = NULL;
958 assert_spin_locked(&mm->page_table_lock);
960 if (flags & FOLL_WRITE && !pmd_write(*pmd))
963 page = pmd_page(*pmd);
964 VM_BUG_ON(!PageHead(page));
965 if (flags & FOLL_TOUCH) {
968 * We should set the dirty bit only for FOLL_WRITE but
969 * for now the dirty bit in the pmd is meaningless.
970 * And if the dirty bit will become meaningful and
971 * we'll only set it with FOLL_WRITE, an atomic
972 * set_bit will be required on the pmd to set the
973 * young bit, instead of the current set_pmd_at.
975 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
976 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
978 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
979 VM_BUG_ON(!PageCompound(page));
980 if (flags & FOLL_GET)
987 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
992 spin_lock(&tlb->mm->page_table_lock);
993 if (likely(pmd_trans_huge(*pmd))) {
994 if (unlikely(pmd_trans_splitting(*pmd))) {
995 spin_unlock(&tlb->mm->page_table_lock);
996 wait_split_huge_page(vma->anon_vma,
1001 pgtable = get_pmd_huge_pte(tlb->mm);
1002 page = pmd_page(*pmd);
1004 page_remove_rmap(page);
1005 VM_BUG_ON(page_mapcount(page) < 0);
1006 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1007 VM_BUG_ON(!PageHead(page));
1008 spin_unlock(&tlb->mm->page_table_lock);
1009 tlb_remove_page(tlb, page);
1010 pte_free(tlb->mm, pgtable);
1014 spin_unlock(&tlb->mm->page_table_lock);
1019 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1020 unsigned long addr, unsigned long end,
1025 spin_lock(&vma->vm_mm->page_table_lock);
1026 if (likely(pmd_trans_huge(*pmd))) {
1027 ret = !pmd_trans_splitting(*pmd);
1028 spin_unlock(&vma->vm_mm->page_table_lock);
1030 wait_split_huge_page(vma->anon_vma, pmd);
1033 * All logical pages in the range are present
1034 * if backed by a huge page.
1036 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1039 spin_unlock(&vma->vm_mm->page_table_lock);
1044 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1045 unsigned long addr, pgprot_t newprot)
1047 struct mm_struct *mm = vma->vm_mm;
1050 spin_lock(&mm->page_table_lock);
1051 if (likely(pmd_trans_huge(*pmd))) {
1052 if (unlikely(pmd_trans_splitting(*pmd))) {
1053 spin_unlock(&mm->page_table_lock);
1054 wait_split_huge_page(vma->anon_vma, pmd);
1058 entry = pmdp_get_and_clear(mm, addr, pmd);
1059 entry = pmd_modify(entry, newprot);
1060 set_pmd_at(mm, addr, pmd, entry);
1061 spin_unlock(&vma->vm_mm->page_table_lock);
1062 flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1066 spin_unlock(&vma->vm_mm->page_table_lock);
1071 pmd_t *page_check_address_pmd(struct page *page,
1072 struct mm_struct *mm,
1073 unsigned long address,
1074 enum page_check_address_pmd_flag flag)
1078 pmd_t *pmd, *ret = NULL;
1080 if (address & ~HPAGE_PMD_MASK)
1083 pgd = pgd_offset(mm, address);
1084 if (!pgd_present(*pgd))
1087 pud = pud_offset(pgd, address);
1088 if (!pud_present(*pud))
1091 pmd = pmd_offset(pud, address);
1094 if (pmd_page(*pmd) != page)
1097 * split_vma() may create temporary aliased mappings. There is
1098 * no risk as long as all huge pmd are found and have their
1099 * splitting bit set before __split_huge_page_refcount
1100 * runs. Finding the same huge pmd more than once during the
1101 * same rmap walk is not a problem.
1103 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1104 pmd_trans_splitting(*pmd))
1106 if (pmd_trans_huge(*pmd)) {
1107 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1108 !pmd_trans_splitting(*pmd));
1115 static int __split_huge_page_splitting(struct page *page,
1116 struct vm_area_struct *vma,
1117 unsigned long address)
1119 struct mm_struct *mm = vma->vm_mm;
1123 spin_lock(&mm->page_table_lock);
1124 pmd = page_check_address_pmd(page, mm, address,
1125 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1128 * We can't temporarily set the pmd to null in order
1129 * to split it, the pmd must remain marked huge at all
1130 * times or the VM won't take the pmd_trans_huge paths
1131 * and it won't wait on the anon_vma->root->lock to
1132 * serialize against split_huge_page*.
1134 pmdp_splitting_flush_notify(vma, address, pmd);
1137 spin_unlock(&mm->page_table_lock);
1142 static void __split_huge_page_refcount(struct page *page)
1145 unsigned long head_index = page->index;
1146 struct zone *zone = page_zone(page);
1149 /* prevent PageLRU to go away from under us, and freeze lru stats */
1150 spin_lock_irq(&zone->lru_lock);
1151 compound_lock(page);
1153 for (i = 1; i < HPAGE_PMD_NR; i++) {
1154 struct page *page_tail = page + i;
1156 /* tail_page->_count cannot change */
1157 atomic_sub(atomic_read(&page_tail->_count), &page->_count);
1158 BUG_ON(page_count(page) <= 0);
1159 atomic_add(page_mapcount(page) + 1, &page_tail->_count);
1160 BUG_ON(atomic_read(&page_tail->_count) <= 0);
1162 /* after clearing PageTail the gup refcount can be released */
1165 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1166 page_tail->flags |= (page->flags &
1167 ((1L << PG_referenced) |
1168 (1L << PG_swapbacked) |
1169 (1L << PG_mlocked) |
1170 (1L << PG_uptodate)));
1171 page_tail->flags |= (1L << PG_dirty);
1174 * 1) clear PageTail before overwriting first_page
1175 * 2) clear PageTail before clearing PageHead for VM_BUG_ON
1180 * __split_huge_page_splitting() already set the
1181 * splitting bit in all pmd that could map this
1182 * hugepage, that will ensure no CPU can alter the
1183 * mapcount on the head page. The mapcount is only
1184 * accounted in the head page and it has to be
1185 * transferred to all tail pages in the below code. So
1186 * for this code to be safe, the split the mapcount
1187 * can't change. But that doesn't mean userland can't
1188 * keep changing and reading the page contents while
1189 * we transfer the mapcount, so the pmd splitting
1190 * status is achieved setting a reserved bit in the
1191 * pmd, not by clearing the present bit.
1193 BUG_ON(page_mapcount(page_tail));
1194 page_tail->_mapcount = page->_mapcount;
1196 BUG_ON(page_tail->mapping);
1197 page_tail->mapping = page->mapping;
1199 page_tail->index = ++head_index;
1201 BUG_ON(!PageAnon(page_tail));
1202 BUG_ON(!PageUptodate(page_tail));
1203 BUG_ON(!PageDirty(page_tail));
1204 BUG_ON(!PageSwapBacked(page_tail));
1206 lru_add_page_tail(zone, page, page_tail);
1209 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1210 __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1213 * A hugepage counts for HPAGE_PMD_NR pages on the LRU statistics,
1214 * so adjust those appropriately if this page is on the LRU.
1216 if (PageLRU(page)) {
1217 zonestat = NR_LRU_BASE + page_lru(page);
1218 __mod_zone_page_state(zone, zonestat, -(HPAGE_PMD_NR-1));
1221 ClearPageCompound(page);
1222 compound_unlock(page);
1223 spin_unlock_irq(&zone->lru_lock);
1225 for (i = 1; i < HPAGE_PMD_NR; i++) {
1226 struct page *page_tail = page + i;
1227 BUG_ON(page_count(page_tail) <= 0);
1229 * Tail pages may be freed if there wasn't any mapping
1230 * like if add_to_swap() is running on a lru page that
1231 * had its mapping zapped. And freeing these pages
1232 * requires taking the lru_lock so we do the put_page
1233 * of the tail pages after the split is complete.
1235 put_page(page_tail);
1239 * Only the head page (now become a regular page) is required
1240 * to be pinned by the caller.
1242 BUG_ON(page_count(page) <= 0);
1245 static int __split_huge_page_map(struct page *page,
1246 struct vm_area_struct *vma,
1247 unsigned long address)
1249 struct mm_struct *mm = vma->vm_mm;
1253 unsigned long haddr;
1255 spin_lock(&mm->page_table_lock);
1256 pmd = page_check_address_pmd(page, mm, address,
1257 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1259 pgtable = get_pmd_huge_pte(mm);
1260 pmd_populate(mm, &_pmd, pgtable);
1262 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1263 i++, haddr += PAGE_SIZE) {
1265 BUG_ON(PageCompound(page+i));
1266 entry = mk_pte(page + i, vma->vm_page_prot);
1267 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1268 if (!pmd_write(*pmd))
1269 entry = pte_wrprotect(entry);
1271 BUG_ON(page_mapcount(page) != 1);
1272 if (!pmd_young(*pmd))
1273 entry = pte_mkold(entry);
1274 pte = pte_offset_map(&_pmd, haddr);
1275 BUG_ON(!pte_none(*pte));
1276 set_pte_at(mm, haddr, pte, entry);
1281 smp_wmb(); /* make pte visible before pmd */
1283 * Up to this point the pmd is present and huge and
1284 * userland has the whole access to the hugepage
1285 * during the split (which happens in place). If we
1286 * overwrite the pmd with the not-huge version
1287 * pointing to the pte here (which of course we could
1288 * if all CPUs were bug free), userland could trigger
1289 * a small page size TLB miss on the small sized TLB
1290 * while the hugepage TLB entry is still established
1291 * in the huge TLB. Some CPU doesn't like that. See
1292 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1293 * Erratum 383 on page 93. Intel should be safe but is
1294 * also warns that it's only safe if the permission
1295 * and cache attributes of the two entries loaded in
1296 * the two TLB is identical (which should be the case
1297 * here). But it is generally safer to never allow
1298 * small and huge TLB entries for the same virtual
1299 * address to be loaded simultaneously. So instead of
1300 * doing "pmd_populate(); flush_tlb_range();" we first
1301 * mark the current pmd notpresent (atomically because
1302 * here the pmd_trans_huge and pmd_trans_splitting
1303 * must remain set at all times on the pmd until the
1304 * split is complete for this pmd), then we flush the
1305 * SMP TLB and finally we write the non-huge version
1306 * of the pmd entry with pmd_populate.
1308 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1309 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1310 pmd_populate(mm, pmd, pgtable);
1313 spin_unlock(&mm->page_table_lock);
1318 /* must be called with anon_vma->root->lock hold */
1319 static void __split_huge_page(struct page *page,
1320 struct anon_vma *anon_vma)
1322 int mapcount, mapcount2;
1323 struct anon_vma_chain *avc;
1325 BUG_ON(!PageHead(page));
1326 BUG_ON(PageTail(page));
1329 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1330 struct vm_area_struct *vma = avc->vma;
1331 unsigned long addr = vma_address(page, vma);
1332 BUG_ON(is_vma_temporary_stack(vma));
1333 if (addr == -EFAULT)
1335 mapcount += __split_huge_page_splitting(page, vma, addr);
1338 * It is critical that new vmas are added to the tail of the
1339 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1340 * and establishes a child pmd before
1341 * __split_huge_page_splitting() freezes the parent pmd (so if
1342 * we fail to prevent copy_huge_pmd() from running until the
1343 * whole __split_huge_page() is complete), we will still see
1344 * the newly established pmd of the child later during the
1345 * walk, to be able to set it as pmd_trans_splitting too.
1347 if (mapcount != page_mapcount(page))
1348 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1349 mapcount, page_mapcount(page));
1350 BUG_ON(mapcount != page_mapcount(page));
1352 __split_huge_page_refcount(page);
1355 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1356 struct vm_area_struct *vma = avc->vma;
1357 unsigned long addr = vma_address(page, vma);
1358 BUG_ON(is_vma_temporary_stack(vma));
1359 if (addr == -EFAULT)
1361 mapcount2 += __split_huge_page_map(page, vma, addr);
1363 if (mapcount != mapcount2)
1364 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1365 mapcount, mapcount2, page_mapcount(page));
1366 BUG_ON(mapcount != mapcount2);
1369 int split_huge_page(struct page *page)
1371 struct anon_vma *anon_vma;
1374 BUG_ON(!PageAnon(page));
1375 anon_vma = page_lock_anon_vma(page);
1379 if (!PageCompound(page))
1382 BUG_ON(!PageSwapBacked(page));
1383 __split_huge_page(page, anon_vma);
1385 BUG_ON(PageCompound(page));
1387 page_unlock_anon_vma(anon_vma);
1392 int hugepage_madvise(unsigned long *vm_flags, int advice)
1397 * Be somewhat over-protective like KSM for now!
1399 if (*vm_flags & (VM_HUGEPAGE |
1400 VM_SHARED | VM_MAYSHARE |
1401 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1402 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1403 VM_MIXEDMAP | VM_SAO))
1405 *vm_flags &= ~VM_NOHUGEPAGE;
1406 *vm_flags |= VM_HUGEPAGE;
1408 case MADV_NOHUGEPAGE:
1410 * Be somewhat over-protective like KSM for now!
1412 if (*vm_flags & (VM_NOHUGEPAGE |
1413 VM_SHARED | VM_MAYSHARE |
1414 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1415 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1416 VM_MIXEDMAP | VM_SAO))
1418 *vm_flags &= ~VM_HUGEPAGE;
1419 *vm_flags |= VM_NOHUGEPAGE;
1426 static int __init khugepaged_slab_init(void)
1428 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1429 sizeof(struct mm_slot),
1430 __alignof__(struct mm_slot), 0, NULL);
1437 static void __init khugepaged_slab_free(void)
1439 kmem_cache_destroy(mm_slot_cache);
1440 mm_slot_cache = NULL;
1443 static inline struct mm_slot *alloc_mm_slot(void)
1445 if (!mm_slot_cache) /* initialization failed */
1447 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1450 static inline void free_mm_slot(struct mm_slot *mm_slot)
1452 kmem_cache_free(mm_slot_cache, mm_slot);
1455 static int __init mm_slots_hash_init(void)
1457 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1465 static void __init mm_slots_hash_free(void)
1467 kfree(mm_slots_hash);
1468 mm_slots_hash = NULL;
1472 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1474 struct mm_slot *mm_slot;
1475 struct hlist_head *bucket;
1476 struct hlist_node *node;
1478 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1479 % MM_SLOTS_HASH_HEADS];
1480 hlist_for_each_entry(mm_slot, node, bucket, hash) {
1481 if (mm == mm_slot->mm)
1487 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1488 struct mm_slot *mm_slot)
1490 struct hlist_head *bucket;
1492 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1493 % MM_SLOTS_HASH_HEADS];
1495 hlist_add_head(&mm_slot->hash, bucket);
1498 static inline int khugepaged_test_exit(struct mm_struct *mm)
1500 return atomic_read(&mm->mm_users) == 0;
1503 int __khugepaged_enter(struct mm_struct *mm)
1505 struct mm_slot *mm_slot;
1508 mm_slot = alloc_mm_slot();
1512 /* __khugepaged_exit() must not run from under us */
1513 VM_BUG_ON(khugepaged_test_exit(mm));
1514 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1515 free_mm_slot(mm_slot);
1519 spin_lock(&khugepaged_mm_lock);
1520 insert_to_mm_slots_hash(mm, mm_slot);
1522 * Insert just behind the scanning cursor, to let the area settle
1525 wakeup = list_empty(&khugepaged_scan.mm_head);
1526 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1527 spin_unlock(&khugepaged_mm_lock);
1529 atomic_inc(&mm->mm_count);
1531 wake_up_interruptible(&khugepaged_wait);
1536 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1538 unsigned long hstart, hend;
1541 * Not yet faulted in so we will register later in the
1542 * page fault if needed.
1545 if (vma->vm_file || vma->vm_ops)
1546 /* khugepaged not yet working on file or special mappings */
1548 VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
1549 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1550 hend = vma->vm_end & HPAGE_PMD_MASK;
1552 return khugepaged_enter(vma);
1556 void __khugepaged_exit(struct mm_struct *mm)
1558 struct mm_slot *mm_slot;
1561 spin_lock(&khugepaged_mm_lock);
1562 mm_slot = get_mm_slot(mm);
1563 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1564 hlist_del(&mm_slot->hash);
1565 list_del(&mm_slot->mm_node);
1570 spin_unlock(&khugepaged_mm_lock);
1571 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1572 free_mm_slot(mm_slot);
1574 } else if (mm_slot) {
1575 spin_unlock(&khugepaged_mm_lock);
1577 * This is required to serialize against
1578 * khugepaged_test_exit() (which is guaranteed to run
1579 * under mmap sem read mode). Stop here (after we
1580 * return all pagetables will be destroyed) until
1581 * khugepaged has finished working on the pagetables
1582 * under the mmap_sem.
1584 down_write(&mm->mmap_sem);
1585 up_write(&mm->mmap_sem);
1587 spin_unlock(&khugepaged_mm_lock);
1590 static void release_pte_page(struct page *page)
1592 /* 0 stands for page_is_file_cache(page) == false */
1593 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1595 putback_lru_page(page);
1598 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1600 while (--_pte >= pte) {
1601 pte_t pteval = *_pte;
1602 if (!pte_none(pteval))
1603 release_pte_page(pte_page(pteval));
1607 static void release_all_pte_pages(pte_t *pte)
1609 release_pte_pages(pte, pte + HPAGE_PMD_NR);
1612 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1613 unsigned long address,
1618 int referenced = 0, isolated = 0, none = 0;
1619 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1620 _pte++, address += PAGE_SIZE) {
1621 pte_t pteval = *_pte;
1622 if (pte_none(pteval)) {
1623 if (++none <= khugepaged_max_ptes_none)
1626 release_pte_pages(pte, _pte);
1630 if (!pte_present(pteval) || !pte_write(pteval)) {
1631 release_pte_pages(pte, _pte);
1634 page = vm_normal_page(vma, address, pteval);
1635 if (unlikely(!page)) {
1636 release_pte_pages(pte, _pte);
1639 VM_BUG_ON(PageCompound(page));
1640 BUG_ON(!PageAnon(page));
1641 VM_BUG_ON(!PageSwapBacked(page));
1643 /* cannot use mapcount: can't collapse if there's a gup pin */
1644 if (page_count(page) != 1) {
1645 release_pte_pages(pte, _pte);
1649 * We can do it before isolate_lru_page because the
1650 * page can't be freed from under us. NOTE: PG_lock
1651 * is needed to serialize against split_huge_page
1652 * when invoked from the VM.
1654 if (!trylock_page(page)) {
1655 release_pte_pages(pte, _pte);
1659 * Isolate the page to avoid collapsing an hugepage
1660 * currently in use by the VM.
1662 if (isolate_lru_page(page)) {
1664 release_pte_pages(pte, _pte);
1667 /* 0 stands for page_is_file_cache(page) == false */
1668 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1669 VM_BUG_ON(!PageLocked(page));
1670 VM_BUG_ON(PageLRU(page));
1672 /* If there is no mapped pte young don't collapse the page */
1673 if (pte_young(pteval) || PageReferenced(page) ||
1674 mmu_notifier_test_young(vma->vm_mm, address))
1677 if (unlikely(!referenced))
1678 release_all_pte_pages(pte);
1685 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1686 struct vm_area_struct *vma,
1687 unsigned long address,
1691 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1692 pte_t pteval = *_pte;
1693 struct page *src_page;
1695 if (pte_none(pteval)) {
1696 clear_user_highpage(page, address);
1697 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1699 src_page = pte_page(pteval);
1700 copy_user_highpage(page, src_page, address, vma);
1701 VM_BUG_ON(page_mapcount(src_page) != 1);
1702 VM_BUG_ON(page_count(src_page) != 2);
1703 release_pte_page(src_page);
1705 * ptl mostly unnecessary, but preempt has to
1706 * be disabled to update the per-cpu stats
1707 * inside page_remove_rmap().
1711 * paravirt calls inside pte_clear here are
1714 pte_clear(vma->vm_mm, address, _pte);
1715 page_remove_rmap(src_page);
1717 free_page_and_swap_cache(src_page);
1720 address += PAGE_SIZE;
1725 static void collapse_huge_page(struct mm_struct *mm,
1726 unsigned long address,
1727 struct page **hpage,
1728 struct vm_area_struct *vma)
1735 struct page *new_page;
1738 unsigned long hstart, hend;
1740 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1747 * Allocate the page while the vma is still valid and under
1748 * the mmap_sem read mode so there is no memory allocation
1749 * later when we take the mmap_sem in write mode. This is more
1750 * friendly behavior (OTOH it may actually hide bugs) to
1751 * filesystems in userland with daemons allocating memory in
1752 * the userland I/O paths. Allocating memory with the
1753 * mmap_sem in read mode is good idea also to allow greater
1756 new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address);
1757 if (unlikely(!new_page)) {
1758 up_read(&mm->mmap_sem);
1759 *hpage = ERR_PTR(-ENOMEM);
1763 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1764 up_read(&mm->mmap_sem);
1769 /* after allocating the hugepage upgrade to mmap_sem write mode */
1770 up_read(&mm->mmap_sem);
1773 * Prevent all access to pagetables with the exception of
1774 * gup_fast later hanlded by the ptep_clear_flush and the VM
1775 * handled by the anon_vma lock + PG_lock.
1777 down_write(&mm->mmap_sem);
1778 if (unlikely(khugepaged_test_exit(mm)))
1781 vma = find_vma(mm, address);
1782 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1783 hend = vma->vm_end & HPAGE_PMD_MASK;
1784 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1787 if (!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always())
1790 /* VM_PFNMAP vmas may have vm_ops null but vm_file set */
1791 if (!vma->anon_vma || vma->vm_ops || vma->vm_file)
1793 VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
1795 pgd = pgd_offset(mm, address);
1796 if (!pgd_present(*pgd))
1799 pud = pud_offset(pgd, address);
1800 if (!pud_present(*pud))
1803 pmd = pmd_offset(pud, address);
1804 /* pmd can't go away or become huge under us */
1805 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1808 anon_vma_lock(vma->anon_vma);
1810 pte = pte_offset_map(pmd, address);
1811 ptl = pte_lockptr(mm, pmd);
1813 spin_lock(&mm->page_table_lock); /* probably unnecessary */
1815 * After this gup_fast can't run anymore. This also removes
1816 * any huge TLB entry from the CPU so we won't allow
1817 * huge and small TLB entries for the same virtual address
1818 * to avoid the risk of CPU bugs in that area.
1820 _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1821 spin_unlock(&mm->page_table_lock);
1824 isolated = __collapse_huge_page_isolate(vma, address, pte);
1828 if (unlikely(!isolated)) {
1829 spin_lock(&mm->page_table_lock);
1830 BUG_ON(!pmd_none(*pmd));
1831 set_pmd_at(mm, address, pmd, _pmd);
1832 spin_unlock(&mm->page_table_lock);
1833 anon_vma_unlock(vma->anon_vma);
1834 mem_cgroup_uncharge_page(new_page);
1839 * All pages are isolated and locked so anon_vma rmap
1840 * can't run anymore.
1842 anon_vma_unlock(vma->anon_vma);
1844 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1845 __SetPageUptodate(new_page);
1846 pgtable = pmd_pgtable(_pmd);
1847 VM_BUG_ON(page_count(pgtable) != 1);
1848 VM_BUG_ON(page_mapcount(pgtable) != 0);
1850 _pmd = mk_pmd(new_page, vma->vm_page_prot);
1851 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1852 _pmd = pmd_mkhuge(_pmd);
1855 * spin_lock() below is not the equivalent of smp_wmb(), so
1856 * this is needed to avoid the copy_huge_page writes to become
1857 * visible after the set_pmd_at() write.
1861 spin_lock(&mm->page_table_lock);
1862 BUG_ON(!pmd_none(*pmd));
1863 page_add_new_anon_rmap(new_page, vma, address);
1864 set_pmd_at(mm, address, pmd, _pmd);
1865 update_mmu_cache(vma, address, entry);
1866 prepare_pmd_huge_pte(pgtable, mm);
1868 spin_unlock(&mm->page_table_lock);
1873 khugepaged_pages_collapsed++;
1875 up_write(&mm->mmap_sem);
1885 static int khugepaged_scan_pmd(struct mm_struct *mm,
1886 struct vm_area_struct *vma,
1887 unsigned long address,
1888 struct page **hpage)
1894 int ret = 0, referenced = 0, none = 0;
1896 unsigned long _address;
1899 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1901 pgd = pgd_offset(mm, address);
1902 if (!pgd_present(*pgd))
1905 pud = pud_offset(pgd, address);
1906 if (!pud_present(*pud))
1909 pmd = pmd_offset(pud, address);
1910 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1913 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1914 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1915 _pte++, _address += PAGE_SIZE) {
1916 pte_t pteval = *_pte;
1917 if (pte_none(pteval)) {
1918 if (++none <= khugepaged_max_ptes_none)
1923 if (!pte_present(pteval) || !pte_write(pteval))
1925 page = vm_normal_page(vma, _address, pteval);
1926 if (unlikely(!page))
1928 VM_BUG_ON(PageCompound(page));
1929 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
1931 /* cannot use mapcount: can't collapse if there's a gup pin */
1932 if (page_count(page) != 1)
1934 if (pte_young(pteval) || PageReferenced(page) ||
1935 mmu_notifier_test_young(vma->vm_mm, address))
1941 pte_unmap_unlock(pte, ptl);
1943 /* collapse_huge_page will return with the mmap_sem released */
1944 collapse_huge_page(mm, address, hpage, vma);
1949 static void collect_mm_slot(struct mm_slot *mm_slot)
1951 struct mm_struct *mm = mm_slot->mm;
1953 VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
1955 if (khugepaged_test_exit(mm)) {
1957 hlist_del(&mm_slot->hash);
1958 list_del(&mm_slot->mm_node);
1961 * Not strictly needed because the mm exited already.
1963 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1966 /* khugepaged_mm_lock actually not necessary for the below */
1967 free_mm_slot(mm_slot);
1972 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1973 struct page **hpage)
1975 struct mm_slot *mm_slot;
1976 struct mm_struct *mm;
1977 struct vm_area_struct *vma;
1981 VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
1983 if (khugepaged_scan.mm_slot)
1984 mm_slot = khugepaged_scan.mm_slot;
1986 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1987 struct mm_slot, mm_node);
1988 khugepaged_scan.address = 0;
1989 khugepaged_scan.mm_slot = mm_slot;
1991 spin_unlock(&khugepaged_mm_lock);
1994 down_read(&mm->mmap_sem);
1995 if (unlikely(khugepaged_test_exit(mm)))
1998 vma = find_vma(mm, khugepaged_scan.address);
2001 for (; vma; vma = vma->vm_next) {
2002 unsigned long hstart, hend;
2005 if (unlikely(khugepaged_test_exit(mm))) {
2010 if (!(vma->vm_flags & VM_HUGEPAGE) &&
2011 !khugepaged_always()) {
2016 /* VM_PFNMAP vmas may have vm_ops null but vm_file set */
2017 if (!vma->anon_vma || vma->vm_ops || vma->vm_file) {
2018 khugepaged_scan.address = vma->vm_end;
2022 VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
2024 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2025 hend = vma->vm_end & HPAGE_PMD_MASK;
2026 if (hstart >= hend) {
2030 if (khugepaged_scan.address < hstart)
2031 khugepaged_scan.address = hstart;
2032 if (khugepaged_scan.address > hend) {
2033 khugepaged_scan.address = hend + HPAGE_PMD_SIZE;
2037 BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2039 while (khugepaged_scan.address < hend) {
2042 if (unlikely(khugepaged_test_exit(mm)))
2043 goto breakouterloop;
2045 VM_BUG_ON(khugepaged_scan.address < hstart ||
2046 khugepaged_scan.address + HPAGE_PMD_SIZE >
2048 ret = khugepaged_scan_pmd(mm, vma,
2049 khugepaged_scan.address,
2051 /* move to next address */
2052 khugepaged_scan.address += HPAGE_PMD_SIZE;
2053 progress += HPAGE_PMD_NR;
2055 /* we released mmap_sem so break loop */
2056 goto breakouterloop_mmap_sem;
2057 if (progress >= pages)
2058 goto breakouterloop;
2062 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2063 breakouterloop_mmap_sem:
2065 spin_lock(&khugepaged_mm_lock);
2066 BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2068 * Release the current mm_slot if this mm is about to die, or
2069 * if we scanned all vmas of this mm.
2071 if (khugepaged_test_exit(mm) || !vma) {
2073 * Make sure that if mm_users is reaching zero while
2074 * khugepaged runs here, khugepaged_exit will find
2075 * mm_slot not pointing to the exiting mm.
2077 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2078 khugepaged_scan.mm_slot = list_entry(
2079 mm_slot->mm_node.next,
2080 struct mm_slot, mm_node);
2081 khugepaged_scan.address = 0;
2083 khugepaged_scan.mm_slot = NULL;
2084 khugepaged_full_scans++;
2087 collect_mm_slot(mm_slot);
2093 static int khugepaged_has_work(void)
2095 return !list_empty(&khugepaged_scan.mm_head) &&
2096 khugepaged_enabled();
2099 static int khugepaged_wait_event(void)
2101 return !list_empty(&khugepaged_scan.mm_head) ||
2102 !khugepaged_enabled();
2105 static void khugepaged_do_scan(struct page **hpage)
2107 unsigned int progress = 0, pass_through_head = 0;
2108 unsigned int pages = khugepaged_pages_to_scan;
2110 barrier(); /* write khugepaged_pages_to_scan to local stack */
2112 while (progress < pages) {
2117 *hpage = alloc_hugepage(khugepaged_defrag());
2118 if (unlikely(!*hpage))
2126 if (unlikely(kthread_should_stop() || freezing(current)))
2129 spin_lock(&khugepaged_mm_lock);
2130 if (!khugepaged_scan.mm_slot)
2131 pass_through_head++;
2132 if (khugepaged_has_work() &&
2133 pass_through_head < 2)
2134 progress += khugepaged_scan_mm_slot(pages - progress,
2138 spin_unlock(&khugepaged_mm_lock);
2142 static void khugepaged_alloc_sleep(void)
2145 add_wait_queue(&khugepaged_wait, &wait);
2146 schedule_timeout_interruptible(
2148 khugepaged_alloc_sleep_millisecs));
2149 remove_wait_queue(&khugepaged_wait, &wait);
2153 static struct page *khugepaged_alloc_hugepage(void)
2158 hpage = alloc_hugepage(khugepaged_defrag());
2160 khugepaged_alloc_sleep();
2161 } while (unlikely(!hpage) &&
2162 likely(khugepaged_enabled()));
2167 static void khugepaged_loop(void)
2174 while (likely(khugepaged_enabled())) {
2176 hpage = khugepaged_alloc_hugepage();
2177 if (unlikely(!hpage))
2180 if (IS_ERR(hpage)) {
2181 khugepaged_alloc_sleep();
2186 khugepaged_do_scan(&hpage);
2192 if (unlikely(kthread_should_stop()))
2194 if (khugepaged_has_work()) {
2196 if (!khugepaged_scan_sleep_millisecs)
2198 add_wait_queue(&khugepaged_wait, &wait);
2199 schedule_timeout_interruptible(
2201 khugepaged_scan_sleep_millisecs));
2202 remove_wait_queue(&khugepaged_wait, &wait);
2203 } else if (khugepaged_enabled())
2204 wait_event_freezable(khugepaged_wait,
2205 khugepaged_wait_event());
2209 static int khugepaged(void *none)
2211 struct mm_slot *mm_slot;
2214 set_user_nice(current, 19);
2216 /* serialize with start_khugepaged() */
2217 mutex_lock(&khugepaged_mutex);
2220 mutex_unlock(&khugepaged_mutex);
2221 BUG_ON(khugepaged_thread != current);
2223 BUG_ON(khugepaged_thread != current);
2225 mutex_lock(&khugepaged_mutex);
2226 if (!khugepaged_enabled())
2228 if (unlikely(kthread_should_stop()))
2232 spin_lock(&khugepaged_mm_lock);
2233 mm_slot = khugepaged_scan.mm_slot;
2234 khugepaged_scan.mm_slot = NULL;
2236 collect_mm_slot(mm_slot);
2237 spin_unlock(&khugepaged_mm_lock);
2239 khugepaged_thread = NULL;
2240 mutex_unlock(&khugepaged_mutex);
2245 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2249 spin_lock(&mm->page_table_lock);
2250 if (unlikely(!pmd_trans_huge(*pmd))) {
2251 spin_unlock(&mm->page_table_lock);
2254 page = pmd_page(*pmd);
2255 VM_BUG_ON(!page_count(page));
2257 spin_unlock(&mm->page_table_lock);
2259 split_huge_page(page);
2262 BUG_ON(pmd_trans_huge(*pmd));
2265 static void split_huge_page_address(struct mm_struct *mm,
2266 unsigned long address)
2272 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2274 pgd = pgd_offset(mm, address);
2275 if (!pgd_present(*pgd))
2278 pud = pud_offset(pgd, address);
2279 if (!pud_present(*pud))
2282 pmd = pmd_offset(pud, address);
2283 if (!pmd_present(*pmd))
2286 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2287 * materialize from under us.
2289 split_huge_page_pmd(mm, pmd);
2292 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2293 unsigned long start,
2298 * If the new start address isn't hpage aligned and it could
2299 * previously contain an hugepage: check if we need to split
2302 if (start & ~HPAGE_PMD_MASK &&
2303 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2304 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2305 split_huge_page_address(vma->vm_mm, start);
2308 * If the new end address isn't hpage aligned and it could
2309 * previously contain an hugepage: check if we need to split
2312 if (end & ~HPAGE_PMD_MASK &&
2313 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2314 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2315 split_huge_page_address(vma->vm_mm, end);
2318 * If we're also updating the vma->vm_next->vm_start, if the new
2319 * vm_next->vm_start isn't page aligned and it could previously
2320 * contain an hugepage: check if we need to split an huge pmd.
2322 if (adjust_next > 0) {
2323 struct vm_area_struct *next = vma->vm_next;
2324 unsigned long nstart = next->vm_start;
2325 nstart += adjust_next << PAGE_SHIFT;
2326 if (nstart & ~HPAGE_PMD_MASK &&
2327 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2328 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2329 split_huge_page_address(next->vm_mm, nstart);