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>
19 #include <asm/pgalloc.h>
23 * By default transparent hugepage support is enabled for all mappings
24 * and khugepaged scans all mappings. Defrag is only invoked by
25 * khugepaged hugepage allocations and by page faults inside
26 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
29 unsigned long transparent_hugepage_flags __read_mostly =
30 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
31 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
33 /* default scan 8*512 pte (or vmas) every 30 second */
34 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
35 static unsigned int khugepaged_pages_collapsed;
36 static unsigned int khugepaged_full_scans;
37 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
38 /* during fragmentation poll the hugepage allocator once every minute */
39 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
40 static struct task_struct *khugepaged_thread __read_mostly;
41 static DEFINE_MUTEX(khugepaged_mutex);
42 static DEFINE_SPINLOCK(khugepaged_mm_lock);
43 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
45 * default collapse hugepages if there is at least one pte mapped like
46 * it would have happened if the vma was large enough during page
49 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
51 static int khugepaged(void *none);
52 static int mm_slots_hash_init(void);
53 static int khugepaged_slab_init(void);
54 static void khugepaged_slab_free(void);
56 #define MM_SLOTS_HASH_HEADS 1024
57 static struct hlist_head *mm_slots_hash __read_mostly;
58 static struct kmem_cache *mm_slot_cache __read_mostly;
61 * struct mm_slot - hash lookup from mm to mm_slot
62 * @hash: hash collision list
63 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
64 * @mm: the mm that this information is valid for
67 struct hlist_node hash;
68 struct list_head mm_node;
73 * struct khugepaged_scan - cursor for scanning
74 * @mm_head: the head of the mm list to scan
75 * @mm_slot: the current mm_slot we are scanning
76 * @address: the next address inside that to be scanned
78 * There is only the one khugepaged_scan instance of this cursor structure.
80 struct khugepaged_scan {
81 struct list_head mm_head;
82 struct mm_slot *mm_slot;
83 unsigned long address;
85 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
88 static int start_khugepaged(void)
91 if (khugepaged_enabled()) {
93 if (unlikely(!mm_slot_cache || !mm_slots_hash)) {
97 mutex_lock(&khugepaged_mutex);
98 if (!khugepaged_thread)
99 khugepaged_thread = kthread_run(khugepaged, NULL,
101 if (unlikely(IS_ERR(khugepaged_thread))) {
103 "khugepaged: kthread_run(khugepaged) failed\n");
104 err = PTR_ERR(khugepaged_thread);
105 khugepaged_thread = NULL;
107 wakeup = !list_empty(&khugepaged_scan.mm_head);
108 mutex_unlock(&khugepaged_mutex);
110 wake_up_interruptible(&khugepaged_wait);
113 wake_up_interruptible(&khugepaged_wait);
120 static ssize_t double_flag_show(struct kobject *kobj,
121 struct kobj_attribute *attr, char *buf,
122 enum transparent_hugepage_flag enabled,
123 enum transparent_hugepage_flag req_madv)
125 if (test_bit(enabled, &transparent_hugepage_flags)) {
126 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
127 return sprintf(buf, "[always] madvise never\n");
128 } else if (test_bit(req_madv, &transparent_hugepage_flags))
129 return sprintf(buf, "always [madvise] never\n");
131 return sprintf(buf, "always madvise [never]\n");
133 static ssize_t double_flag_store(struct kobject *kobj,
134 struct kobj_attribute *attr,
135 const char *buf, size_t count,
136 enum transparent_hugepage_flag enabled,
137 enum transparent_hugepage_flag req_madv)
139 if (!memcmp("always", buf,
140 min(sizeof("always")-1, count))) {
141 set_bit(enabled, &transparent_hugepage_flags);
142 clear_bit(req_madv, &transparent_hugepage_flags);
143 } else if (!memcmp("madvise", buf,
144 min(sizeof("madvise")-1, count))) {
145 clear_bit(enabled, &transparent_hugepage_flags);
146 set_bit(req_madv, &transparent_hugepage_flags);
147 } else if (!memcmp("never", buf,
148 min(sizeof("never")-1, count))) {
149 clear_bit(enabled, &transparent_hugepage_flags);
150 clear_bit(req_madv, &transparent_hugepage_flags);
157 static ssize_t enabled_show(struct kobject *kobj,
158 struct kobj_attribute *attr, char *buf)
160 return double_flag_show(kobj, attr, buf,
161 TRANSPARENT_HUGEPAGE_FLAG,
162 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
164 static ssize_t enabled_store(struct kobject *kobj,
165 struct kobj_attribute *attr,
166 const char *buf, size_t count)
170 ret = double_flag_store(kobj, attr, buf, count,
171 TRANSPARENT_HUGEPAGE_FLAG,
172 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
175 int err = start_khugepaged();
182 static struct kobj_attribute enabled_attr =
183 __ATTR(enabled, 0644, enabled_show, enabled_store);
185 static ssize_t single_flag_show(struct kobject *kobj,
186 struct kobj_attribute *attr, char *buf,
187 enum transparent_hugepage_flag flag)
189 if (test_bit(flag, &transparent_hugepage_flags))
190 return sprintf(buf, "[yes] no\n");
192 return sprintf(buf, "yes [no]\n");
194 static ssize_t single_flag_store(struct kobject *kobj,
195 struct kobj_attribute *attr,
196 const char *buf, size_t count,
197 enum transparent_hugepage_flag flag)
199 if (!memcmp("yes", buf,
200 min(sizeof("yes")-1, count))) {
201 set_bit(flag, &transparent_hugepage_flags);
202 } else if (!memcmp("no", buf,
203 min(sizeof("no")-1, count))) {
204 clear_bit(flag, &transparent_hugepage_flags);
212 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
213 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
214 * memory just to allocate one more hugepage.
216 static ssize_t defrag_show(struct kobject *kobj,
217 struct kobj_attribute *attr, char *buf)
219 return double_flag_show(kobj, attr, buf,
220 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
221 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
223 static ssize_t defrag_store(struct kobject *kobj,
224 struct kobj_attribute *attr,
225 const char *buf, size_t count)
227 return double_flag_store(kobj, attr, buf, count,
228 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
229 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
231 static struct kobj_attribute defrag_attr =
232 __ATTR(defrag, 0644, defrag_show, defrag_store);
234 #ifdef CONFIG_DEBUG_VM
235 static ssize_t debug_cow_show(struct kobject *kobj,
236 struct kobj_attribute *attr, char *buf)
238 return single_flag_show(kobj, attr, buf,
239 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
241 static ssize_t debug_cow_store(struct kobject *kobj,
242 struct kobj_attribute *attr,
243 const char *buf, size_t count)
245 return single_flag_store(kobj, attr, buf, count,
246 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
248 static struct kobj_attribute debug_cow_attr =
249 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
250 #endif /* CONFIG_DEBUG_VM */
252 static struct attribute *hugepage_attr[] = {
255 #ifdef CONFIG_DEBUG_VM
256 &debug_cow_attr.attr,
261 static struct attribute_group hugepage_attr_group = {
262 .attrs = hugepage_attr,
265 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
266 struct kobj_attribute *attr,
269 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
272 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
273 struct kobj_attribute *attr,
274 const char *buf, size_t count)
279 err = strict_strtoul(buf, 10, &msecs);
280 if (err || msecs > UINT_MAX)
283 khugepaged_scan_sleep_millisecs = msecs;
284 wake_up_interruptible(&khugepaged_wait);
288 static struct kobj_attribute scan_sleep_millisecs_attr =
289 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
290 scan_sleep_millisecs_store);
292 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
293 struct kobj_attribute *attr,
296 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
299 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
300 struct kobj_attribute *attr,
301 const char *buf, size_t count)
306 err = strict_strtoul(buf, 10, &msecs);
307 if (err || msecs > UINT_MAX)
310 khugepaged_alloc_sleep_millisecs = msecs;
311 wake_up_interruptible(&khugepaged_wait);
315 static struct kobj_attribute alloc_sleep_millisecs_attr =
316 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
317 alloc_sleep_millisecs_store);
319 static ssize_t pages_to_scan_show(struct kobject *kobj,
320 struct kobj_attribute *attr,
323 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
325 static ssize_t pages_to_scan_store(struct kobject *kobj,
326 struct kobj_attribute *attr,
327 const char *buf, size_t count)
332 err = strict_strtoul(buf, 10, &pages);
333 if (err || !pages || pages > UINT_MAX)
336 khugepaged_pages_to_scan = pages;
340 static struct kobj_attribute pages_to_scan_attr =
341 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
342 pages_to_scan_store);
344 static ssize_t pages_collapsed_show(struct kobject *kobj,
345 struct kobj_attribute *attr,
348 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
350 static struct kobj_attribute pages_collapsed_attr =
351 __ATTR_RO(pages_collapsed);
353 static ssize_t full_scans_show(struct kobject *kobj,
354 struct kobj_attribute *attr,
357 return sprintf(buf, "%u\n", khugepaged_full_scans);
359 static struct kobj_attribute full_scans_attr =
360 __ATTR_RO(full_scans);
362 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
363 struct kobj_attribute *attr, char *buf)
365 return single_flag_show(kobj, attr, buf,
366 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
368 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
369 struct kobj_attribute *attr,
370 const char *buf, size_t count)
372 return single_flag_store(kobj, attr, buf, count,
373 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
375 static struct kobj_attribute khugepaged_defrag_attr =
376 __ATTR(defrag, 0644, khugepaged_defrag_show,
377 khugepaged_defrag_store);
380 * max_ptes_none controls if khugepaged should collapse hugepages over
381 * any unmapped ptes in turn potentially increasing the memory
382 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
383 * reduce the available free memory in the system as it
384 * runs. Increasing max_ptes_none will instead potentially reduce the
385 * free memory in the system during the khugepaged scan.
387 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
388 struct kobj_attribute *attr,
391 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
393 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
394 struct kobj_attribute *attr,
395 const char *buf, size_t count)
398 unsigned long max_ptes_none;
400 err = strict_strtoul(buf, 10, &max_ptes_none);
401 if (err || max_ptes_none > HPAGE_PMD_NR-1)
404 khugepaged_max_ptes_none = max_ptes_none;
408 static struct kobj_attribute khugepaged_max_ptes_none_attr =
409 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
410 khugepaged_max_ptes_none_store);
412 static struct attribute *khugepaged_attr[] = {
413 &khugepaged_defrag_attr.attr,
414 &khugepaged_max_ptes_none_attr.attr,
415 &pages_to_scan_attr.attr,
416 &pages_collapsed_attr.attr,
417 &full_scans_attr.attr,
418 &scan_sleep_millisecs_attr.attr,
419 &alloc_sleep_millisecs_attr.attr,
423 static struct attribute_group khugepaged_attr_group = {
424 .attrs = khugepaged_attr,
425 .name = "khugepaged",
427 #endif /* CONFIG_SYSFS */
429 static int __init hugepage_init(void)
433 static struct kobject *hugepage_kobj;
436 hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
437 if (unlikely(!hugepage_kobj)) {
438 printk(KERN_ERR "hugepage: failed kobject create\n");
442 err = sysfs_create_group(hugepage_kobj, &hugepage_attr_group);
444 printk(KERN_ERR "hugepage: failed register hugeage group\n");
448 err = sysfs_create_group(hugepage_kobj, &khugepaged_attr_group);
450 printk(KERN_ERR "hugepage: failed register hugeage group\n");
455 err = khugepaged_slab_init();
459 err = mm_slots_hash_init();
461 khugepaged_slab_free();
470 module_init(hugepage_init)
472 static int __init setup_transparent_hugepage(char *str)
477 if (!strcmp(str, "always")) {
478 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
479 &transparent_hugepage_flags);
480 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
481 &transparent_hugepage_flags);
483 } else if (!strcmp(str, "madvise")) {
484 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
485 &transparent_hugepage_flags);
486 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
487 &transparent_hugepage_flags);
489 } else if (!strcmp(str, "never")) {
490 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
491 &transparent_hugepage_flags);
492 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
493 &transparent_hugepage_flags);
499 "transparent_hugepage= cannot parse, ignored\n");
502 __setup("transparent_hugepage=", setup_transparent_hugepage);
504 static void prepare_pmd_huge_pte(pgtable_t pgtable,
505 struct mm_struct *mm)
507 assert_spin_locked(&mm->page_table_lock);
510 if (!mm->pmd_huge_pte)
511 INIT_LIST_HEAD(&pgtable->lru);
513 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
514 mm->pmd_huge_pte = pgtable;
517 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
519 if (likely(vma->vm_flags & VM_WRITE))
520 pmd = pmd_mkwrite(pmd);
524 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
525 struct vm_area_struct *vma,
526 unsigned long haddr, pmd_t *pmd,
532 VM_BUG_ON(!PageCompound(page));
533 pgtable = pte_alloc_one(mm, haddr);
534 if (unlikely(!pgtable)) {
535 mem_cgroup_uncharge_page(page);
540 clear_huge_page(page, haddr, HPAGE_PMD_NR);
541 __SetPageUptodate(page);
543 spin_lock(&mm->page_table_lock);
544 if (unlikely(!pmd_none(*pmd))) {
545 spin_unlock(&mm->page_table_lock);
546 mem_cgroup_uncharge_page(page);
548 pte_free(mm, pgtable);
551 entry = mk_pmd(page, vma->vm_page_prot);
552 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
553 entry = pmd_mkhuge(entry);
555 * The spinlocking to take the lru_lock inside
556 * page_add_new_anon_rmap() acts as a full memory
557 * barrier to be sure clear_huge_page writes become
558 * visible after the set_pmd_at() write.
560 page_add_new_anon_rmap(page, vma, haddr);
561 set_pmd_at(mm, haddr, pmd, entry);
562 prepare_pmd_huge_pte(pgtable, mm);
563 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
564 spin_unlock(&mm->page_table_lock);
570 static inline struct page *alloc_hugepage(int defrag)
572 return alloc_pages(GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT),
576 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
577 unsigned long address, pmd_t *pmd,
581 unsigned long haddr = address & HPAGE_PMD_MASK;
584 if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
585 if (unlikely(anon_vma_prepare(vma)))
587 if (unlikely(khugepaged_enter(vma)))
589 page = alloc_hugepage(transparent_hugepage_defrag(vma));
592 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
597 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
601 * Use __pte_alloc instead of pte_alloc_map, because we can't
602 * run pte_offset_map on the pmd, if an huge pmd could
603 * materialize from under us from a different thread.
605 if (unlikely(__pte_alloc(mm, vma, pmd, address)))
607 /* if an huge pmd materialized from under us just retry later */
608 if (unlikely(pmd_trans_huge(*pmd)))
611 * A regular pmd is established and it can't morph into a huge pmd
612 * from under us anymore at this point because we hold the mmap_sem
613 * read mode and khugepaged takes it in write mode. So now it's
614 * safe to run pte_offset_map().
616 pte = pte_offset_map(pmd, address);
617 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
620 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
621 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
622 struct vm_area_struct *vma)
624 struct page *src_page;
630 pgtable = pte_alloc_one(dst_mm, addr);
631 if (unlikely(!pgtable))
634 spin_lock(&dst_mm->page_table_lock);
635 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
639 if (unlikely(!pmd_trans_huge(pmd))) {
640 pte_free(dst_mm, pgtable);
643 if (unlikely(pmd_trans_splitting(pmd))) {
644 /* split huge page running from under us */
645 spin_unlock(&src_mm->page_table_lock);
646 spin_unlock(&dst_mm->page_table_lock);
647 pte_free(dst_mm, pgtable);
649 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
652 src_page = pmd_page(pmd);
653 VM_BUG_ON(!PageHead(src_page));
655 page_dup_rmap(src_page);
656 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
658 pmdp_set_wrprotect(src_mm, addr, src_pmd);
659 pmd = pmd_mkold(pmd_wrprotect(pmd));
660 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
661 prepare_pmd_huge_pte(pgtable, dst_mm);
665 spin_unlock(&src_mm->page_table_lock);
666 spin_unlock(&dst_mm->page_table_lock);
671 /* no "address" argument so destroys page coloring of some arch */
672 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
676 assert_spin_locked(&mm->page_table_lock);
679 pgtable = mm->pmd_huge_pte;
680 if (list_empty(&pgtable->lru))
681 mm->pmd_huge_pte = NULL;
683 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
685 list_del(&pgtable->lru);
690 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
691 struct vm_area_struct *vma,
692 unsigned long address,
693 pmd_t *pmd, pmd_t orig_pmd,
702 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
704 if (unlikely(!pages)) {
709 for (i = 0; i < HPAGE_PMD_NR; i++) {
710 pages[i] = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
712 if (unlikely(!pages[i] ||
713 mem_cgroup_newpage_charge(pages[i], mm,
717 mem_cgroup_uncharge_start();
719 mem_cgroup_uncharge_page(pages[i]);
722 mem_cgroup_uncharge_end();
729 for (i = 0; i < HPAGE_PMD_NR; i++) {
730 copy_user_highpage(pages[i], page + i,
731 haddr + PAGE_SHIFT*i, vma);
732 __SetPageUptodate(pages[i]);
736 spin_lock(&mm->page_table_lock);
737 if (unlikely(!pmd_same(*pmd, orig_pmd)))
739 VM_BUG_ON(!PageHead(page));
741 pmdp_clear_flush_notify(vma, haddr, pmd);
742 /* leave pmd empty until pte is filled */
744 pgtable = get_pmd_huge_pte(mm);
745 pmd_populate(mm, &_pmd, pgtable);
747 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
749 entry = mk_pte(pages[i], vma->vm_page_prot);
750 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
751 page_add_new_anon_rmap(pages[i], vma, haddr);
752 pte = pte_offset_map(&_pmd, haddr);
753 VM_BUG_ON(!pte_none(*pte));
754 set_pte_at(mm, haddr, pte, entry);
760 smp_wmb(); /* make pte visible before pmd */
761 pmd_populate(mm, pmd, pgtable);
762 page_remove_rmap(page);
763 spin_unlock(&mm->page_table_lock);
765 ret |= VM_FAULT_WRITE;
772 spin_unlock(&mm->page_table_lock);
773 mem_cgroup_uncharge_start();
774 for (i = 0; i < HPAGE_PMD_NR; i++) {
775 mem_cgroup_uncharge_page(pages[i]);
778 mem_cgroup_uncharge_end();
783 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
784 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
787 struct page *page, *new_page;
790 VM_BUG_ON(!vma->anon_vma);
791 spin_lock(&mm->page_table_lock);
792 if (unlikely(!pmd_same(*pmd, orig_pmd)))
795 page = pmd_page(orig_pmd);
796 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
797 haddr = address & HPAGE_PMD_MASK;
798 if (page_mapcount(page) == 1) {
800 entry = pmd_mkyoung(orig_pmd);
801 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
802 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
803 update_mmu_cache(vma, address, entry);
804 ret |= VM_FAULT_WRITE;
808 spin_unlock(&mm->page_table_lock);
810 if (transparent_hugepage_enabled(vma) &&
811 !transparent_hugepage_debug_cow())
812 new_page = alloc_hugepage(transparent_hugepage_defrag(vma));
816 if (unlikely(!new_page)) {
817 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
818 pmd, orig_pmd, page, haddr);
823 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
830 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
831 __SetPageUptodate(new_page);
833 spin_lock(&mm->page_table_lock);
835 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
836 mem_cgroup_uncharge_page(new_page);
840 VM_BUG_ON(!PageHead(page));
841 entry = mk_pmd(new_page, vma->vm_page_prot);
842 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
843 entry = pmd_mkhuge(entry);
844 pmdp_clear_flush_notify(vma, haddr, pmd);
845 page_add_new_anon_rmap(new_page, vma, haddr);
846 set_pmd_at(mm, haddr, pmd, entry);
847 update_mmu_cache(vma, address, entry);
848 page_remove_rmap(page);
850 ret |= VM_FAULT_WRITE;
853 spin_unlock(&mm->page_table_lock);
858 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
863 struct page *page = NULL;
865 assert_spin_locked(&mm->page_table_lock);
867 if (flags & FOLL_WRITE && !pmd_write(*pmd))
870 page = pmd_page(*pmd);
871 VM_BUG_ON(!PageHead(page));
872 if (flags & FOLL_TOUCH) {
875 * We should set the dirty bit only for FOLL_WRITE but
876 * for now the dirty bit in the pmd is meaningless.
877 * And if the dirty bit will become meaningful and
878 * we'll only set it with FOLL_WRITE, an atomic
879 * set_bit will be required on the pmd to set the
880 * young bit, instead of the current set_pmd_at.
882 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
883 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
885 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
886 VM_BUG_ON(!PageCompound(page));
887 if (flags & FOLL_GET)
894 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
899 spin_lock(&tlb->mm->page_table_lock);
900 if (likely(pmd_trans_huge(*pmd))) {
901 if (unlikely(pmd_trans_splitting(*pmd))) {
902 spin_unlock(&tlb->mm->page_table_lock);
903 wait_split_huge_page(vma->anon_vma,
908 pgtable = get_pmd_huge_pte(tlb->mm);
909 page = pmd_page(*pmd);
911 page_remove_rmap(page);
912 VM_BUG_ON(page_mapcount(page) < 0);
913 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
914 VM_BUG_ON(!PageHead(page));
915 spin_unlock(&tlb->mm->page_table_lock);
916 tlb_remove_page(tlb, page);
917 pte_free(tlb->mm, pgtable);
921 spin_unlock(&tlb->mm->page_table_lock);
926 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
927 unsigned long addr, unsigned long end,
932 spin_lock(&vma->vm_mm->page_table_lock);
933 if (likely(pmd_trans_huge(*pmd))) {
934 ret = !pmd_trans_splitting(*pmd);
935 spin_unlock(&vma->vm_mm->page_table_lock);
937 wait_split_huge_page(vma->anon_vma, pmd);
940 * All logical pages in the range are present
941 * if backed by a huge page.
943 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
946 spin_unlock(&vma->vm_mm->page_table_lock);
951 pmd_t *page_check_address_pmd(struct page *page,
952 struct mm_struct *mm,
953 unsigned long address,
954 enum page_check_address_pmd_flag flag)
958 pmd_t *pmd, *ret = NULL;
960 if (address & ~HPAGE_PMD_MASK)
963 pgd = pgd_offset(mm, address);
964 if (!pgd_present(*pgd))
967 pud = pud_offset(pgd, address);
968 if (!pud_present(*pud))
971 pmd = pmd_offset(pud, address);
974 if (pmd_page(*pmd) != page)
976 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
977 pmd_trans_splitting(*pmd));
978 if (pmd_trans_huge(*pmd)) {
979 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
980 !pmd_trans_splitting(*pmd));
987 static int __split_huge_page_splitting(struct page *page,
988 struct vm_area_struct *vma,
989 unsigned long address)
991 struct mm_struct *mm = vma->vm_mm;
995 spin_lock(&mm->page_table_lock);
996 pmd = page_check_address_pmd(page, mm, address,
997 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1000 * We can't temporarily set the pmd to null in order
1001 * to split it, the pmd must remain marked huge at all
1002 * times or the VM won't take the pmd_trans_huge paths
1003 * and it won't wait on the anon_vma->root->lock to
1004 * serialize against split_huge_page*.
1006 pmdp_splitting_flush_notify(vma, address, pmd);
1009 spin_unlock(&mm->page_table_lock);
1014 static void __split_huge_page_refcount(struct page *page)
1017 unsigned long head_index = page->index;
1018 struct zone *zone = page_zone(page);
1020 /* prevent PageLRU to go away from under us, and freeze lru stats */
1021 spin_lock_irq(&zone->lru_lock);
1022 compound_lock(page);
1024 for (i = 1; i < HPAGE_PMD_NR; i++) {
1025 struct page *page_tail = page + i;
1027 /* tail_page->_count cannot change */
1028 atomic_sub(atomic_read(&page_tail->_count), &page->_count);
1029 BUG_ON(page_count(page) <= 0);
1030 atomic_add(page_mapcount(page) + 1, &page_tail->_count);
1031 BUG_ON(atomic_read(&page_tail->_count) <= 0);
1033 /* after clearing PageTail the gup refcount can be released */
1036 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1037 page_tail->flags |= (page->flags &
1038 ((1L << PG_referenced) |
1039 (1L << PG_swapbacked) |
1040 (1L << PG_mlocked) |
1041 (1L << PG_uptodate)));
1042 page_tail->flags |= (1L << PG_dirty);
1045 * 1) clear PageTail before overwriting first_page
1046 * 2) clear PageTail before clearing PageHead for VM_BUG_ON
1051 * __split_huge_page_splitting() already set the
1052 * splitting bit in all pmd that could map this
1053 * hugepage, that will ensure no CPU can alter the
1054 * mapcount on the head page. The mapcount is only
1055 * accounted in the head page and it has to be
1056 * transferred to all tail pages in the below code. So
1057 * for this code to be safe, the split the mapcount
1058 * can't change. But that doesn't mean userland can't
1059 * keep changing and reading the page contents while
1060 * we transfer the mapcount, so the pmd splitting
1061 * status is achieved setting a reserved bit in the
1062 * pmd, not by clearing the present bit.
1064 BUG_ON(page_mapcount(page_tail));
1065 page_tail->_mapcount = page->_mapcount;
1067 BUG_ON(page_tail->mapping);
1068 page_tail->mapping = page->mapping;
1070 page_tail->index = ++head_index;
1072 BUG_ON(!PageAnon(page_tail));
1073 BUG_ON(!PageUptodate(page_tail));
1074 BUG_ON(!PageDirty(page_tail));
1075 BUG_ON(!PageSwapBacked(page_tail));
1077 lru_add_page_tail(zone, page, page_tail);
1080 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1081 __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1083 ClearPageCompound(page);
1084 compound_unlock(page);
1085 spin_unlock_irq(&zone->lru_lock);
1087 for (i = 1; i < HPAGE_PMD_NR; i++) {
1088 struct page *page_tail = page + i;
1089 BUG_ON(page_count(page_tail) <= 0);
1091 * Tail pages may be freed if there wasn't any mapping
1092 * like if add_to_swap() is running on a lru page that
1093 * had its mapping zapped. And freeing these pages
1094 * requires taking the lru_lock so we do the put_page
1095 * of the tail pages after the split is complete.
1097 put_page(page_tail);
1101 * Only the head page (now become a regular page) is required
1102 * to be pinned by the caller.
1104 BUG_ON(page_count(page) <= 0);
1107 static int __split_huge_page_map(struct page *page,
1108 struct vm_area_struct *vma,
1109 unsigned long address)
1111 struct mm_struct *mm = vma->vm_mm;
1115 unsigned long haddr;
1117 spin_lock(&mm->page_table_lock);
1118 pmd = page_check_address_pmd(page, mm, address,
1119 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1121 pgtable = get_pmd_huge_pte(mm);
1122 pmd_populate(mm, &_pmd, pgtable);
1124 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1125 i++, haddr += PAGE_SIZE) {
1127 BUG_ON(PageCompound(page+i));
1128 entry = mk_pte(page + i, vma->vm_page_prot);
1129 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1130 if (!pmd_write(*pmd))
1131 entry = pte_wrprotect(entry);
1133 BUG_ON(page_mapcount(page) != 1);
1134 if (!pmd_young(*pmd))
1135 entry = pte_mkold(entry);
1136 pte = pte_offset_map(&_pmd, haddr);
1137 BUG_ON(!pte_none(*pte));
1138 set_pte_at(mm, haddr, pte, entry);
1143 smp_wmb(); /* make pte visible before pmd */
1145 * Up to this point the pmd is present and huge and
1146 * userland has the whole access to the hugepage
1147 * during the split (which happens in place). If we
1148 * overwrite the pmd with the not-huge version
1149 * pointing to the pte here (which of course we could
1150 * if all CPUs were bug free), userland could trigger
1151 * a small page size TLB miss on the small sized TLB
1152 * while the hugepage TLB entry is still established
1153 * in the huge TLB. Some CPU doesn't like that. See
1154 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1155 * Erratum 383 on page 93. Intel should be safe but is
1156 * also warns that it's only safe if the permission
1157 * and cache attributes of the two entries loaded in
1158 * the two TLB is identical (which should be the case
1159 * here). But it is generally safer to never allow
1160 * small and huge TLB entries for the same virtual
1161 * address to be loaded simultaneously. So instead of
1162 * doing "pmd_populate(); flush_tlb_range();" we first
1163 * mark the current pmd notpresent (atomically because
1164 * here the pmd_trans_huge and pmd_trans_splitting
1165 * must remain set at all times on the pmd until the
1166 * split is complete for this pmd), then we flush the
1167 * SMP TLB and finally we write the non-huge version
1168 * of the pmd entry with pmd_populate.
1170 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1171 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1172 pmd_populate(mm, pmd, pgtable);
1175 spin_unlock(&mm->page_table_lock);
1180 /* must be called with anon_vma->root->lock hold */
1181 static void __split_huge_page(struct page *page,
1182 struct anon_vma *anon_vma)
1184 int mapcount, mapcount2;
1185 struct anon_vma_chain *avc;
1187 BUG_ON(!PageHead(page));
1188 BUG_ON(PageTail(page));
1191 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1192 struct vm_area_struct *vma = avc->vma;
1193 unsigned long addr = vma_address(page, vma);
1194 BUG_ON(is_vma_temporary_stack(vma));
1195 if (addr == -EFAULT)
1197 mapcount += __split_huge_page_splitting(page, vma, addr);
1200 * It is critical that new vmas are added to the tail of the
1201 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1202 * and establishes a child pmd before
1203 * __split_huge_page_splitting() freezes the parent pmd (so if
1204 * we fail to prevent copy_huge_pmd() from running until the
1205 * whole __split_huge_page() is complete), we will still see
1206 * the newly established pmd of the child later during the
1207 * walk, to be able to set it as pmd_trans_splitting too.
1209 if (mapcount != page_mapcount(page))
1210 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1211 mapcount, page_mapcount(page));
1212 BUG_ON(mapcount != page_mapcount(page));
1214 __split_huge_page_refcount(page);
1217 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1218 struct vm_area_struct *vma = avc->vma;
1219 unsigned long addr = vma_address(page, vma);
1220 BUG_ON(is_vma_temporary_stack(vma));
1221 if (addr == -EFAULT)
1223 mapcount2 += __split_huge_page_map(page, vma, addr);
1225 if (mapcount != mapcount2)
1226 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1227 mapcount, mapcount2, page_mapcount(page));
1228 BUG_ON(mapcount != mapcount2);
1231 int split_huge_page(struct page *page)
1233 struct anon_vma *anon_vma;
1236 BUG_ON(!PageAnon(page));
1237 anon_vma = page_lock_anon_vma(page);
1241 if (!PageCompound(page))
1244 BUG_ON(!PageSwapBacked(page));
1245 __split_huge_page(page, anon_vma);
1247 BUG_ON(PageCompound(page));
1249 page_unlock_anon_vma(anon_vma);
1254 int hugepage_madvise(unsigned long *vm_flags)
1257 * Be somewhat over-protective like KSM for now!
1259 if (*vm_flags & (VM_HUGEPAGE | VM_SHARED | VM_MAYSHARE |
1260 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1261 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1262 VM_MIXEDMAP | VM_SAO))
1265 *vm_flags |= VM_HUGEPAGE;
1270 static int __init khugepaged_slab_init(void)
1272 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1273 sizeof(struct mm_slot),
1274 __alignof__(struct mm_slot), 0, NULL);
1281 static void __init khugepaged_slab_free(void)
1283 kmem_cache_destroy(mm_slot_cache);
1284 mm_slot_cache = NULL;
1287 static inline struct mm_slot *alloc_mm_slot(void)
1289 if (!mm_slot_cache) /* initialization failed */
1291 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1294 static inline void free_mm_slot(struct mm_slot *mm_slot)
1296 kmem_cache_free(mm_slot_cache, mm_slot);
1299 static int __init mm_slots_hash_init(void)
1301 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1309 static void __init mm_slots_hash_free(void)
1311 kfree(mm_slots_hash);
1312 mm_slots_hash = NULL;
1316 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1318 struct mm_slot *mm_slot;
1319 struct hlist_head *bucket;
1320 struct hlist_node *node;
1322 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1323 % MM_SLOTS_HASH_HEADS];
1324 hlist_for_each_entry(mm_slot, node, bucket, hash) {
1325 if (mm == mm_slot->mm)
1331 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1332 struct mm_slot *mm_slot)
1334 struct hlist_head *bucket;
1336 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1337 % MM_SLOTS_HASH_HEADS];
1339 hlist_add_head(&mm_slot->hash, bucket);
1342 static inline int khugepaged_test_exit(struct mm_struct *mm)
1344 return atomic_read(&mm->mm_users) == 0;
1347 int __khugepaged_enter(struct mm_struct *mm)
1349 struct mm_slot *mm_slot;
1352 mm_slot = alloc_mm_slot();
1356 /* __khugepaged_exit() must not run from under us */
1357 VM_BUG_ON(khugepaged_test_exit(mm));
1358 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1359 free_mm_slot(mm_slot);
1363 spin_lock(&khugepaged_mm_lock);
1364 insert_to_mm_slots_hash(mm, mm_slot);
1366 * Insert just behind the scanning cursor, to let the area settle
1369 wakeup = list_empty(&khugepaged_scan.mm_head);
1370 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1371 spin_unlock(&khugepaged_mm_lock);
1373 atomic_inc(&mm->mm_count);
1375 wake_up_interruptible(&khugepaged_wait);
1380 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1382 unsigned long hstart, hend;
1385 * Not yet faulted in so we will register later in the
1386 * page fault if needed.
1389 if (vma->vm_file || vma->vm_ops)
1390 /* khugepaged not yet working on file or special mappings */
1392 VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
1393 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1394 hend = vma->vm_end & HPAGE_PMD_MASK;
1396 return khugepaged_enter(vma);
1400 void __khugepaged_exit(struct mm_struct *mm)
1402 struct mm_slot *mm_slot;
1405 spin_lock(&khugepaged_mm_lock);
1406 mm_slot = get_mm_slot(mm);
1407 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1408 hlist_del(&mm_slot->hash);
1409 list_del(&mm_slot->mm_node);
1414 spin_unlock(&khugepaged_mm_lock);
1415 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1416 free_mm_slot(mm_slot);
1418 } else if (mm_slot) {
1419 spin_unlock(&khugepaged_mm_lock);
1421 * This is required to serialize against
1422 * khugepaged_test_exit() (which is guaranteed to run
1423 * under mmap sem read mode). Stop here (after we
1424 * return all pagetables will be destroyed) until
1425 * khugepaged has finished working on the pagetables
1426 * under the mmap_sem.
1428 down_write(&mm->mmap_sem);
1429 up_write(&mm->mmap_sem);
1431 spin_unlock(&khugepaged_mm_lock);
1434 static void release_pte_page(struct page *page)
1436 /* 0 stands for page_is_file_cache(page) == false */
1437 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1439 putback_lru_page(page);
1442 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1444 while (--_pte >= pte) {
1445 pte_t pteval = *_pte;
1446 if (!pte_none(pteval))
1447 release_pte_page(pte_page(pteval));
1451 static void release_all_pte_pages(pte_t *pte)
1453 release_pte_pages(pte, pte + HPAGE_PMD_NR);
1456 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1457 unsigned long address,
1462 int referenced = 0, isolated = 0, none = 0;
1463 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1464 _pte++, address += PAGE_SIZE) {
1465 pte_t pteval = *_pte;
1466 if (pte_none(pteval)) {
1467 if (++none <= khugepaged_max_ptes_none)
1470 release_pte_pages(pte, _pte);
1474 if (!pte_present(pteval) || !pte_write(pteval)) {
1475 release_pte_pages(pte, _pte);
1478 page = vm_normal_page(vma, address, pteval);
1479 if (unlikely(!page)) {
1480 release_pte_pages(pte, _pte);
1483 VM_BUG_ON(PageCompound(page));
1484 BUG_ON(!PageAnon(page));
1485 VM_BUG_ON(!PageSwapBacked(page));
1487 /* cannot use mapcount: can't collapse if there's a gup pin */
1488 if (page_count(page) != 1) {
1489 release_pte_pages(pte, _pte);
1493 * We can do it before isolate_lru_page because the
1494 * page can't be freed from under us. NOTE: PG_lock
1495 * is needed to serialize against split_huge_page
1496 * when invoked from the VM.
1498 if (!trylock_page(page)) {
1499 release_pte_pages(pte, _pte);
1503 * Isolate the page to avoid collapsing an hugepage
1504 * currently in use by the VM.
1506 if (isolate_lru_page(page)) {
1508 release_pte_pages(pte, _pte);
1511 /* 0 stands for page_is_file_cache(page) == false */
1512 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1513 VM_BUG_ON(!PageLocked(page));
1514 VM_BUG_ON(PageLRU(page));
1516 /* If there is no mapped pte young don't collapse the page */
1517 if (pte_young(pteval))
1520 if (unlikely(!referenced))
1521 release_all_pte_pages(pte);
1528 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1529 struct vm_area_struct *vma,
1530 unsigned long address,
1534 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1535 pte_t pteval = *_pte;
1536 struct page *src_page;
1538 if (pte_none(pteval)) {
1539 clear_user_highpage(page, address);
1540 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1542 src_page = pte_page(pteval);
1543 copy_user_highpage(page, src_page, address, vma);
1544 VM_BUG_ON(page_mapcount(src_page) != 1);
1545 VM_BUG_ON(page_count(src_page) != 2);
1546 release_pte_page(src_page);
1548 * ptl mostly unnecessary, but preempt has to
1549 * be disabled to update the per-cpu stats
1550 * inside page_remove_rmap().
1554 * paravirt calls inside pte_clear here are
1557 pte_clear(vma->vm_mm, address, _pte);
1558 page_remove_rmap(src_page);
1560 free_page_and_swap_cache(src_page);
1563 address += PAGE_SIZE;
1568 static void collapse_huge_page(struct mm_struct *mm,
1569 unsigned long address,
1570 struct page **hpage)
1572 struct vm_area_struct *vma;
1578 struct page *new_page;
1581 unsigned long hstart, hend;
1583 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1587 * Prevent all access to pagetables with the exception of
1588 * gup_fast later hanlded by the ptep_clear_flush and the VM
1589 * handled by the anon_vma lock + PG_lock.
1591 down_write(&mm->mmap_sem);
1592 if (unlikely(khugepaged_test_exit(mm)))
1595 vma = find_vma(mm, address);
1596 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1597 hend = vma->vm_end & HPAGE_PMD_MASK;
1598 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1601 if (!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always())
1604 /* VM_PFNMAP vmas may have vm_ops null but vm_file set */
1605 if (!vma->anon_vma || vma->vm_ops || vma->vm_file)
1607 VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
1609 pgd = pgd_offset(mm, address);
1610 if (!pgd_present(*pgd))
1613 pud = pud_offset(pgd, address);
1614 if (!pud_present(*pud))
1617 pmd = pmd_offset(pud, address);
1618 /* pmd can't go away or become huge under us */
1619 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1623 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
1626 anon_vma_lock(vma->anon_vma);
1628 pte = pte_offset_map(pmd, address);
1629 ptl = pte_lockptr(mm, pmd);
1631 spin_lock(&mm->page_table_lock); /* probably unnecessary */
1633 * After this gup_fast can't run anymore. This also removes
1634 * any huge TLB entry from the CPU so we won't allow
1635 * huge and small TLB entries for the same virtual address
1636 * to avoid the risk of CPU bugs in that area.
1638 _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1639 spin_unlock(&mm->page_table_lock);
1642 isolated = __collapse_huge_page_isolate(vma, address, pte);
1646 if (unlikely(!isolated)) {
1647 spin_lock(&mm->page_table_lock);
1648 BUG_ON(!pmd_none(*pmd));
1649 set_pmd_at(mm, address, pmd, _pmd);
1650 spin_unlock(&mm->page_table_lock);
1651 anon_vma_unlock(vma->anon_vma);
1652 mem_cgroup_uncharge_page(new_page);
1657 * All pages are isolated and locked so anon_vma rmap
1658 * can't run anymore.
1660 anon_vma_unlock(vma->anon_vma);
1662 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1663 __SetPageUptodate(new_page);
1664 pgtable = pmd_pgtable(_pmd);
1665 VM_BUG_ON(page_count(pgtable) != 1);
1666 VM_BUG_ON(page_mapcount(pgtable) != 0);
1668 _pmd = mk_pmd(new_page, vma->vm_page_prot);
1669 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1670 _pmd = pmd_mkhuge(_pmd);
1673 * spin_lock() below is not the equivalent of smp_wmb(), so
1674 * this is needed to avoid the copy_huge_page writes to become
1675 * visible after the set_pmd_at() write.
1679 spin_lock(&mm->page_table_lock);
1680 BUG_ON(!pmd_none(*pmd));
1681 page_add_new_anon_rmap(new_page, vma, address);
1682 set_pmd_at(mm, address, pmd, _pmd);
1683 update_mmu_cache(vma, address, entry);
1684 prepare_pmd_huge_pte(pgtable, mm);
1686 spin_unlock(&mm->page_table_lock);
1689 khugepaged_pages_collapsed++;
1691 up_write(&mm->mmap_sem);
1694 static int khugepaged_scan_pmd(struct mm_struct *mm,
1695 struct vm_area_struct *vma,
1696 unsigned long address,
1697 struct page **hpage)
1703 int ret = 0, referenced = 0, none = 0;
1705 unsigned long _address;
1708 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1710 pgd = pgd_offset(mm, address);
1711 if (!pgd_present(*pgd))
1714 pud = pud_offset(pgd, address);
1715 if (!pud_present(*pud))
1718 pmd = pmd_offset(pud, address);
1719 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1722 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1723 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1724 _pte++, _address += PAGE_SIZE) {
1725 pte_t pteval = *_pte;
1726 if (pte_none(pteval)) {
1727 if (++none <= khugepaged_max_ptes_none)
1732 if (!pte_present(pteval) || !pte_write(pteval))
1734 page = vm_normal_page(vma, _address, pteval);
1735 if (unlikely(!page))
1737 VM_BUG_ON(PageCompound(page));
1738 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
1740 /* cannot use mapcount: can't collapse if there's a gup pin */
1741 if (page_count(page) != 1)
1743 if (pte_young(pteval))
1749 pte_unmap_unlock(pte, ptl);
1751 up_read(&mm->mmap_sem);
1752 collapse_huge_page(mm, address, hpage);
1758 static void collect_mm_slot(struct mm_slot *mm_slot)
1760 struct mm_struct *mm = mm_slot->mm;
1762 VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
1764 if (khugepaged_test_exit(mm)) {
1766 hlist_del(&mm_slot->hash);
1767 list_del(&mm_slot->mm_node);
1770 * Not strictly needed because the mm exited already.
1772 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1775 /* khugepaged_mm_lock actually not necessary for the below */
1776 free_mm_slot(mm_slot);
1781 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1782 struct page **hpage)
1784 struct mm_slot *mm_slot;
1785 struct mm_struct *mm;
1786 struct vm_area_struct *vma;
1790 VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
1792 if (khugepaged_scan.mm_slot)
1793 mm_slot = khugepaged_scan.mm_slot;
1795 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1796 struct mm_slot, mm_node);
1797 khugepaged_scan.address = 0;
1798 khugepaged_scan.mm_slot = mm_slot;
1800 spin_unlock(&khugepaged_mm_lock);
1803 down_read(&mm->mmap_sem);
1804 if (unlikely(khugepaged_test_exit(mm)))
1807 vma = find_vma(mm, khugepaged_scan.address);
1810 for (; vma; vma = vma->vm_next) {
1811 unsigned long hstart, hend;
1814 if (unlikely(khugepaged_test_exit(mm))) {
1819 if (!(vma->vm_flags & VM_HUGEPAGE) &&
1820 !khugepaged_always()) {
1825 /* VM_PFNMAP vmas may have vm_ops null but vm_file set */
1826 if (!vma->anon_vma || vma->vm_ops || vma->vm_file) {
1827 khugepaged_scan.address = vma->vm_end;
1831 VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
1833 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1834 hend = vma->vm_end & HPAGE_PMD_MASK;
1835 if (hstart >= hend) {
1839 if (khugepaged_scan.address < hstart)
1840 khugepaged_scan.address = hstart;
1841 if (khugepaged_scan.address > hend) {
1842 khugepaged_scan.address = hend + HPAGE_PMD_SIZE;
1846 BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1848 while (khugepaged_scan.address < hend) {
1851 if (unlikely(khugepaged_test_exit(mm)))
1852 goto breakouterloop;
1854 VM_BUG_ON(khugepaged_scan.address < hstart ||
1855 khugepaged_scan.address + HPAGE_PMD_SIZE >
1857 ret = khugepaged_scan_pmd(mm, vma,
1858 khugepaged_scan.address,
1860 /* move to next address */
1861 khugepaged_scan.address += HPAGE_PMD_SIZE;
1862 progress += HPAGE_PMD_NR;
1864 /* we released mmap_sem so break loop */
1865 goto breakouterloop_mmap_sem;
1866 if (progress >= pages)
1867 goto breakouterloop;
1871 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1872 breakouterloop_mmap_sem:
1874 spin_lock(&khugepaged_mm_lock);
1875 BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1877 * Release the current mm_slot if this mm is about to die, or
1878 * if we scanned all vmas of this mm.
1880 if (khugepaged_test_exit(mm) || !vma) {
1882 * Make sure that if mm_users is reaching zero while
1883 * khugepaged runs here, khugepaged_exit will find
1884 * mm_slot not pointing to the exiting mm.
1886 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1887 khugepaged_scan.mm_slot = list_entry(
1888 mm_slot->mm_node.next,
1889 struct mm_slot, mm_node);
1890 khugepaged_scan.address = 0;
1892 khugepaged_scan.mm_slot = NULL;
1893 khugepaged_full_scans++;
1896 collect_mm_slot(mm_slot);
1902 static int khugepaged_has_work(void)
1904 return !list_empty(&khugepaged_scan.mm_head) &&
1905 khugepaged_enabled();
1908 static int khugepaged_wait_event(void)
1910 return !list_empty(&khugepaged_scan.mm_head) ||
1911 !khugepaged_enabled();
1914 static void khugepaged_do_scan(struct page **hpage)
1916 unsigned int progress = 0, pass_through_head = 0;
1917 unsigned int pages = khugepaged_pages_to_scan;
1919 barrier(); /* write khugepaged_pages_to_scan to local stack */
1921 while (progress < pages) {
1925 *hpage = alloc_hugepage(khugepaged_defrag());
1926 if (unlikely(!*hpage))
1930 spin_lock(&khugepaged_mm_lock);
1931 if (!khugepaged_scan.mm_slot)
1932 pass_through_head++;
1933 if (khugepaged_has_work() &&
1934 pass_through_head < 2)
1935 progress += khugepaged_scan_mm_slot(pages - progress,
1939 spin_unlock(&khugepaged_mm_lock);
1943 static struct page *khugepaged_alloc_hugepage(void)
1948 hpage = alloc_hugepage(khugepaged_defrag());
1951 add_wait_queue(&khugepaged_wait, &wait);
1952 schedule_timeout_interruptible(
1954 khugepaged_alloc_sleep_millisecs));
1955 remove_wait_queue(&khugepaged_wait, &wait);
1957 } while (unlikely(!hpage) &&
1958 likely(khugepaged_enabled()));
1962 static void khugepaged_loop(void)
1966 while (likely(khugepaged_enabled())) {
1967 hpage = khugepaged_alloc_hugepage();
1968 if (unlikely(!hpage))
1971 khugepaged_do_scan(&hpage);
1974 if (khugepaged_has_work()) {
1976 if (!khugepaged_scan_sleep_millisecs)
1978 add_wait_queue(&khugepaged_wait, &wait);
1979 schedule_timeout_interruptible(
1981 khugepaged_scan_sleep_millisecs));
1982 remove_wait_queue(&khugepaged_wait, &wait);
1983 } else if (khugepaged_enabled())
1984 wait_event_interruptible(khugepaged_wait,
1985 khugepaged_wait_event());
1989 static int khugepaged(void *none)
1991 struct mm_slot *mm_slot;
1993 set_user_nice(current, 19);
1995 /* serialize with start_khugepaged() */
1996 mutex_lock(&khugepaged_mutex);
1999 mutex_unlock(&khugepaged_mutex);
2000 BUG_ON(khugepaged_thread != current);
2002 BUG_ON(khugepaged_thread != current);
2004 mutex_lock(&khugepaged_mutex);
2005 if (!khugepaged_enabled())
2009 spin_lock(&khugepaged_mm_lock);
2010 mm_slot = khugepaged_scan.mm_slot;
2011 khugepaged_scan.mm_slot = NULL;
2013 collect_mm_slot(mm_slot);
2014 spin_unlock(&khugepaged_mm_lock);
2016 khugepaged_thread = NULL;
2017 mutex_unlock(&khugepaged_mutex);
2022 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2026 spin_lock(&mm->page_table_lock);
2027 if (unlikely(!pmd_trans_huge(*pmd))) {
2028 spin_unlock(&mm->page_table_lock);
2031 page = pmd_page(*pmd);
2032 VM_BUG_ON(!page_count(page));
2034 spin_unlock(&mm->page_table_lock);
2036 split_huge_page(page);
2039 BUG_ON(pmd_trans_huge(*pmd));