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 pmd_t *page_check_address_pmd(struct page *page,
927 struct mm_struct *mm,
928 unsigned long address,
929 enum page_check_address_pmd_flag flag)
933 pmd_t *pmd, *ret = NULL;
935 if (address & ~HPAGE_PMD_MASK)
938 pgd = pgd_offset(mm, address);
939 if (!pgd_present(*pgd))
942 pud = pud_offset(pgd, address);
943 if (!pud_present(*pud))
946 pmd = pmd_offset(pud, address);
949 if (pmd_page(*pmd) != page)
951 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
952 pmd_trans_splitting(*pmd));
953 if (pmd_trans_huge(*pmd)) {
954 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
955 !pmd_trans_splitting(*pmd));
962 static int __split_huge_page_splitting(struct page *page,
963 struct vm_area_struct *vma,
964 unsigned long address)
966 struct mm_struct *mm = vma->vm_mm;
970 spin_lock(&mm->page_table_lock);
971 pmd = page_check_address_pmd(page, mm, address,
972 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
975 * We can't temporarily set the pmd to null in order
976 * to split it, the pmd must remain marked huge at all
977 * times or the VM won't take the pmd_trans_huge paths
978 * and it won't wait on the anon_vma->root->lock to
979 * serialize against split_huge_page*.
981 pmdp_splitting_flush_notify(vma, address, pmd);
984 spin_unlock(&mm->page_table_lock);
989 static void __split_huge_page_refcount(struct page *page)
992 unsigned long head_index = page->index;
993 struct zone *zone = page_zone(page);
995 /* prevent PageLRU to go away from under us, and freeze lru stats */
996 spin_lock_irq(&zone->lru_lock);
999 for (i = 1; i < HPAGE_PMD_NR; i++) {
1000 struct page *page_tail = page + i;
1002 /* tail_page->_count cannot change */
1003 atomic_sub(atomic_read(&page_tail->_count), &page->_count);
1004 BUG_ON(page_count(page) <= 0);
1005 atomic_add(page_mapcount(page) + 1, &page_tail->_count);
1006 BUG_ON(atomic_read(&page_tail->_count) <= 0);
1008 /* after clearing PageTail the gup refcount can be released */
1011 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1012 page_tail->flags |= (page->flags &
1013 ((1L << PG_referenced) |
1014 (1L << PG_swapbacked) |
1015 (1L << PG_mlocked) |
1016 (1L << PG_uptodate)));
1017 page_tail->flags |= (1L << PG_dirty);
1020 * 1) clear PageTail before overwriting first_page
1021 * 2) clear PageTail before clearing PageHead for VM_BUG_ON
1026 * __split_huge_page_splitting() already set the
1027 * splitting bit in all pmd that could map this
1028 * hugepage, that will ensure no CPU can alter the
1029 * mapcount on the head page. The mapcount is only
1030 * accounted in the head page and it has to be
1031 * transferred to all tail pages in the below code. So
1032 * for this code to be safe, the split the mapcount
1033 * can't change. But that doesn't mean userland can't
1034 * keep changing and reading the page contents while
1035 * we transfer the mapcount, so the pmd splitting
1036 * status is achieved setting a reserved bit in the
1037 * pmd, not by clearing the present bit.
1039 BUG_ON(page_mapcount(page_tail));
1040 page_tail->_mapcount = page->_mapcount;
1042 BUG_ON(page_tail->mapping);
1043 page_tail->mapping = page->mapping;
1045 page_tail->index = ++head_index;
1047 BUG_ON(!PageAnon(page_tail));
1048 BUG_ON(!PageUptodate(page_tail));
1049 BUG_ON(!PageDirty(page_tail));
1050 BUG_ON(!PageSwapBacked(page_tail));
1052 lru_add_page_tail(zone, page, page_tail);
1055 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1056 __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1058 ClearPageCompound(page);
1059 compound_unlock(page);
1060 spin_unlock_irq(&zone->lru_lock);
1062 for (i = 1; i < HPAGE_PMD_NR; i++) {
1063 struct page *page_tail = page + i;
1064 BUG_ON(page_count(page_tail) <= 0);
1066 * Tail pages may be freed if there wasn't any mapping
1067 * like if add_to_swap() is running on a lru page that
1068 * had its mapping zapped. And freeing these pages
1069 * requires taking the lru_lock so we do the put_page
1070 * of the tail pages after the split is complete.
1072 put_page(page_tail);
1076 * Only the head page (now become a regular page) is required
1077 * to be pinned by the caller.
1079 BUG_ON(page_count(page) <= 0);
1082 static int __split_huge_page_map(struct page *page,
1083 struct vm_area_struct *vma,
1084 unsigned long address)
1086 struct mm_struct *mm = vma->vm_mm;
1090 unsigned long haddr;
1092 spin_lock(&mm->page_table_lock);
1093 pmd = page_check_address_pmd(page, mm, address,
1094 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1096 pgtable = get_pmd_huge_pte(mm);
1097 pmd_populate(mm, &_pmd, pgtable);
1099 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1100 i++, haddr += PAGE_SIZE) {
1102 BUG_ON(PageCompound(page+i));
1103 entry = mk_pte(page + i, vma->vm_page_prot);
1104 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1105 if (!pmd_write(*pmd))
1106 entry = pte_wrprotect(entry);
1108 BUG_ON(page_mapcount(page) != 1);
1109 if (!pmd_young(*pmd))
1110 entry = pte_mkold(entry);
1111 pte = pte_offset_map(&_pmd, haddr);
1112 BUG_ON(!pte_none(*pte));
1113 set_pte_at(mm, haddr, pte, entry);
1118 smp_wmb(); /* make pte visible before pmd */
1120 * Up to this point the pmd is present and huge and
1121 * userland has the whole access to the hugepage
1122 * during the split (which happens in place). If we
1123 * overwrite the pmd with the not-huge version
1124 * pointing to the pte here (which of course we could
1125 * if all CPUs were bug free), userland could trigger
1126 * a small page size TLB miss on the small sized TLB
1127 * while the hugepage TLB entry is still established
1128 * in the huge TLB. Some CPU doesn't like that. See
1129 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1130 * Erratum 383 on page 93. Intel should be safe but is
1131 * also warns that it's only safe if the permission
1132 * and cache attributes of the two entries loaded in
1133 * the two TLB is identical (which should be the case
1134 * here). But it is generally safer to never allow
1135 * small and huge TLB entries for the same virtual
1136 * address to be loaded simultaneously. So instead of
1137 * doing "pmd_populate(); flush_tlb_range();" we first
1138 * mark the current pmd notpresent (atomically because
1139 * here the pmd_trans_huge and pmd_trans_splitting
1140 * must remain set at all times on the pmd until the
1141 * split is complete for this pmd), then we flush the
1142 * SMP TLB and finally we write the non-huge version
1143 * of the pmd entry with pmd_populate.
1145 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1146 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1147 pmd_populate(mm, pmd, pgtable);
1150 spin_unlock(&mm->page_table_lock);
1155 /* must be called with anon_vma->root->lock hold */
1156 static void __split_huge_page(struct page *page,
1157 struct anon_vma *anon_vma)
1159 int mapcount, mapcount2;
1160 struct anon_vma_chain *avc;
1162 BUG_ON(!PageHead(page));
1163 BUG_ON(PageTail(page));
1166 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1167 struct vm_area_struct *vma = avc->vma;
1168 unsigned long addr = vma_address(page, vma);
1169 BUG_ON(is_vma_temporary_stack(vma));
1170 if (addr == -EFAULT)
1172 mapcount += __split_huge_page_splitting(page, vma, addr);
1175 * It is critical that new vmas are added to the tail of the
1176 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1177 * and establishes a child pmd before
1178 * __split_huge_page_splitting() freezes the parent pmd (so if
1179 * we fail to prevent copy_huge_pmd() from running until the
1180 * whole __split_huge_page() is complete), we will still see
1181 * the newly established pmd of the child later during the
1182 * walk, to be able to set it as pmd_trans_splitting too.
1184 if (mapcount != page_mapcount(page))
1185 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1186 mapcount, page_mapcount(page));
1187 BUG_ON(mapcount != page_mapcount(page));
1189 __split_huge_page_refcount(page);
1192 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1193 struct vm_area_struct *vma = avc->vma;
1194 unsigned long addr = vma_address(page, vma);
1195 BUG_ON(is_vma_temporary_stack(vma));
1196 if (addr == -EFAULT)
1198 mapcount2 += __split_huge_page_map(page, vma, addr);
1200 if (mapcount != mapcount2)
1201 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1202 mapcount, mapcount2, page_mapcount(page));
1203 BUG_ON(mapcount != mapcount2);
1206 int split_huge_page(struct page *page)
1208 struct anon_vma *anon_vma;
1211 BUG_ON(!PageAnon(page));
1212 anon_vma = page_lock_anon_vma(page);
1216 if (!PageCompound(page))
1219 BUG_ON(!PageSwapBacked(page));
1220 __split_huge_page(page, anon_vma);
1222 BUG_ON(PageCompound(page));
1224 page_unlock_anon_vma(anon_vma);
1229 int hugepage_madvise(unsigned long *vm_flags)
1232 * Be somewhat over-protective like KSM for now!
1234 if (*vm_flags & (VM_HUGEPAGE | VM_SHARED | VM_MAYSHARE |
1235 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1236 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1237 VM_MIXEDMAP | VM_SAO))
1240 *vm_flags |= VM_HUGEPAGE;
1245 static int __init khugepaged_slab_init(void)
1247 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1248 sizeof(struct mm_slot),
1249 __alignof__(struct mm_slot), 0, NULL);
1256 static void __init khugepaged_slab_free(void)
1258 kmem_cache_destroy(mm_slot_cache);
1259 mm_slot_cache = NULL;
1262 static inline struct mm_slot *alloc_mm_slot(void)
1264 if (!mm_slot_cache) /* initialization failed */
1266 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1269 static inline void free_mm_slot(struct mm_slot *mm_slot)
1271 kmem_cache_free(mm_slot_cache, mm_slot);
1274 static int __init mm_slots_hash_init(void)
1276 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1284 static void __init mm_slots_hash_free(void)
1286 kfree(mm_slots_hash);
1287 mm_slots_hash = NULL;
1291 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1293 struct mm_slot *mm_slot;
1294 struct hlist_head *bucket;
1295 struct hlist_node *node;
1297 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1298 % MM_SLOTS_HASH_HEADS];
1299 hlist_for_each_entry(mm_slot, node, bucket, hash) {
1300 if (mm == mm_slot->mm)
1306 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1307 struct mm_slot *mm_slot)
1309 struct hlist_head *bucket;
1311 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1312 % MM_SLOTS_HASH_HEADS];
1314 hlist_add_head(&mm_slot->hash, bucket);
1317 static inline int khugepaged_test_exit(struct mm_struct *mm)
1319 return atomic_read(&mm->mm_users) == 0;
1322 int __khugepaged_enter(struct mm_struct *mm)
1324 struct mm_slot *mm_slot;
1327 mm_slot = alloc_mm_slot();
1331 /* __khugepaged_exit() must not run from under us */
1332 VM_BUG_ON(khugepaged_test_exit(mm));
1333 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1334 free_mm_slot(mm_slot);
1338 spin_lock(&khugepaged_mm_lock);
1339 insert_to_mm_slots_hash(mm, mm_slot);
1341 * Insert just behind the scanning cursor, to let the area settle
1344 wakeup = list_empty(&khugepaged_scan.mm_head);
1345 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1346 spin_unlock(&khugepaged_mm_lock);
1348 atomic_inc(&mm->mm_count);
1350 wake_up_interruptible(&khugepaged_wait);
1355 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1357 unsigned long hstart, hend;
1360 * Not yet faulted in so we will register later in the
1361 * page fault if needed.
1364 if (vma->vm_file || vma->vm_ops)
1365 /* khugepaged not yet working on file or special mappings */
1367 VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
1368 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1369 hend = vma->vm_end & HPAGE_PMD_MASK;
1371 return khugepaged_enter(vma);
1375 void __khugepaged_exit(struct mm_struct *mm)
1377 struct mm_slot *mm_slot;
1380 spin_lock(&khugepaged_mm_lock);
1381 mm_slot = get_mm_slot(mm);
1382 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1383 hlist_del(&mm_slot->hash);
1384 list_del(&mm_slot->mm_node);
1389 spin_unlock(&khugepaged_mm_lock);
1390 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1391 free_mm_slot(mm_slot);
1393 } else if (mm_slot) {
1394 spin_unlock(&khugepaged_mm_lock);
1396 * This is required to serialize against
1397 * khugepaged_test_exit() (which is guaranteed to run
1398 * under mmap sem read mode). Stop here (after we
1399 * return all pagetables will be destroyed) until
1400 * khugepaged has finished working on the pagetables
1401 * under the mmap_sem.
1403 down_write(&mm->mmap_sem);
1404 up_write(&mm->mmap_sem);
1406 spin_unlock(&khugepaged_mm_lock);
1409 static void release_pte_page(struct page *page)
1411 /* 0 stands for page_is_file_cache(page) == false */
1412 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1414 putback_lru_page(page);
1417 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1419 while (--_pte >= pte) {
1420 pte_t pteval = *_pte;
1421 if (!pte_none(pteval))
1422 release_pte_page(pte_page(pteval));
1426 static void release_all_pte_pages(pte_t *pte)
1428 release_pte_pages(pte, pte + HPAGE_PMD_NR);
1431 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1432 unsigned long address,
1437 int referenced = 0, isolated = 0, none = 0;
1438 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1439 _pte++, address += PAGE_SIZE) {
1440 pte_t pteval = *_pte;
1441 if (pte_none(pteval)) {
1442 if (++none <= khugepaged_max_ptes_none)
1445 release_pte_pages(pte, _pte);
1449 if (!pte_present(pteval) || !pte_write(pteval)) {
1450 release_pte_pages(pte, _pte);
1453 page = vm_normal_page(vma, address, pteval);
1454 if (unlikely(!page)) {
1455 release_pte_pages(pte, _pte);
1458 VM_BUG_ON(PageCompound(page));
1459 BUG_ON(!PageAnon(page));
1460 VM_BUG_ON(!PageSwapBacked(page));
1462 /* cannot use mapcount: can't collapse if there's a gup pin */
1463 if (page_count(page) != 1) {
1464 release_pte_pages(pte, _pte);
1468 * We can do it before isolate_lru_page because the
1469 * page can't be freed from under us. NOTE: PG_lock
1470 * is needed to serialize against split_huge_page
1471 * when invoked from the VM.
1473 if (!trylock_page(page)) {
1474 release_pte_pages(pte, _pte);
1478 * Isolate the page to avoid collapsing an hugepage
1479 * currently in use by the VM.
1481 if (isolate_lru_page(page)) {
1483 release_pte_pages(pte, _pte);
1486 /* 0 stands for page_is_file_cache(page) == false */
1487 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1488 VM_BUG_ON(!PageLocked(page));
1489 VM_BUG_ON(PageLRU(page));
1491 /* If there is no mapped pte young don't collapse the page */
1492 if (pte_young(pteval))
1495 if (unlikely(!referenced))
1496 release_all_pte_pages(pte);
1503 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1504 struct vm_area_struct *vma,
1505 unsigned long address,
1509 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1510 pte_t pteval = *_pte;
1511 struct page *src_page;
1513 if (pte_none(pteval)) {
1514 clear_user_highpage(page, address);
1515 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1517 src_page = pte_page(pteval);
1518 copy_user_highpage(page, src_page, address, vma);
1519 VM_BUG_ON(page_mapcount(src_page) != 1);
1520 VM_BUG_ON(page_count(src_page) != 2);
1521 release_pte_page(src_page);
1523 * ptl mostly unnecessary, but preempt has to
1524 * be disabled to update the per-cpu stats
1525 * inside page_remove_rmap().
1529 * paravirt calls inside pte_clear here are
1532 pte_clear(vma->vm_mm, address, _pte);
1533 page_remove_rmap(src_page);
1535 free_page_and_swap_cache(src_page);
1538 address += PAGE_SIZE;
1543 static void collapse_huge_page(struct mm_struct *mm,
1544 unsigned long address,
1545 struct page **hpage)
1547 struct vm_area_struct *vma;
1553 struct page *new_page;
1556 unsigned long hstart, hend;
1558 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1562 * Prevent all access to pagetables with the exception of
1563 * gup_fast later hanlded by the ptep_clear_flush and the VM
1564 * handled by the anon_vma lock + PG_lock.
1566 down_write(&mm->mmap_sem);
1567 if (unlikely(khugepaged_test_exit(mm)))
1570 vma = find_vma(mm, address);
1571 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1572 hend = vma->vm_end & HPAGE_PMD_MASK;
1573 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1576 if (!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always())
1579 /* VM_PFNMAP vmas may have vm_ops null but vm_file set */
1580 if (!vma->anon_vma || vma->vm_ops || vma->vm_file)
1582 VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
1584 pgd = pgd_offset(mm, address);
1585 if (!pgd_present(*pgd))
1588 pud = pud_offset(pgd, address);
1589 if (!pud_present(*pud))
1592 pmd = pmd_offset(pud, address);
1593 /* pmd can't go away or become huge under us */
1594 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1598 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
1601 anon_vma_lock(vma->anon_vma);
1603 pte = pte_offset_map(pmd, address);
1604 ptl = pte_lockptr(mm, pmd);
1606 spin_lock(&mm->page_table_lock); /* probably unnecessary */
1608 * After this gup_fast can't run anymore. This also removes
1609 * any huge TLB entry from the CPU so we won't allow
1610 * huge and small TLB entries for the same virtual address
1611 * to avoid the risk of CPU bugs in that area.
1613 _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1614 spin_unlock(&mm->page_table_lock);
1617 isolated = __collapse_huge_page_isolate(vma, address, pte);
1621 if (unlikely(!isolated)) {
1622 spin_lock(&mm->page_table_lock);
1623 BUG_ON(!pmd_none(*pmd));
1624 set_pmd_at(mm, address, pmd, _pmd);
1625 spin_unlock(&mm->page_table_lock);
1626 anon_vma_unlock(vma->anon_vma);
1627 mem_cgroup_uncharge_page(new_page);
1632 * All pages are isolated and locked so anon_vma rmap
1633 * can't run anymore.
1635 anon_vma_unlock(vma->anon_vma);
1637 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1638 __SetPageUptodate(new_page);
1639 pgtable = pmd_pgtable(_pmd);
1640 VM_BUG_ON(page_count(pgtable) != 1);
1641 VM_BUG_ON(page_mapcount(pgtable) != 0);
1643 _pmd = mk_pmd(new_page, vma->vm_page_prot);
1644 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1645 _pmd = pmd_mkhuge(_pmd);
1648 * spin_lock() below is not the equivalent of smp_wmb(), so
1649 * this is needed to avoid the copy_huge_page writes to become
1650 * visible after the set_pmd_at() write.
1654 spin_lock(&mm->page_table_lock);
1655 BUG_ON(!pmd_none(*pmd));
1656 page_add_new_anon_rmap(new_page, vma, address);
1657 set_pmd_at(mm, address, pmd, _pmd);
1658 update_mmu_cache(vma, address, entry);
1659 prepare_pmd_huge_pte(pgtable, mm);
1661 spin_unlock(&mm->page_table_lock);
1664 khugepaged_pages_collapsed++;
1666 up_write(&mm->mmap_sem);
1669 static int khugepaged_scan_pmd(struct mm_struct *mm,
1670 struct vm_area_struct *vma,
1671 unsigned long address,
1672 struct page **hpage)
1678 int ret = 0, referenced = 0, none = 0;
1680 unsigned long _address;
1683 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1685 pgd = pgd_offset(mm, address);
1686 if (!pgd_present(*pgd))
1689 pud = pud_offset(pgd, address);
1690 if (!pud_present(*pud))
1693 pmd = pmd_offset(pud, address);
1694 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1697 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1698 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1699 _pte++, _address += PAGE_SIZE) {
1700 pte_t pteval = *_pte;
1701 if (pte_none(pteval)) {
1702 if (++none <= khugepaged_max_ptes_none)
1707 if (!pte_present(pteval) || !pte_write(pteval))
1709 page = vm_normal_page(vma, _address, pteval);
1710 if (unlikely(!page))
1712 VM_BUG_ON(PageCompound(page));
1713 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
1715 /* cannot use mapcount: can't collapse if there's a gup pin */
1716 if (page_count(page) != 1)
1718 if (pte_young(pteval))
1724 pte_unmap_unlock(pte, ptl);
1726 up_read(&mm->mmap_sem);
1727 collapse_huge_page(mm, address, hpage);
1733 static void collect_mm_slot(struct mm_slot *mm_slot)
1735 struct mm_struct *mm = mm_slot->mm;
1737 VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
1739 if (khugepaged_test_exit(mm)) {
1741 hlist_del(&mm_slot->hash);
1742 list_del(&mm_slot->mm_node);
1745 * Not strictly needed because the mm exited already.
1747 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1750 /* khugepaged_mm_lock actually not necessary for the below */
1751 free_mm_slot(mm_slot);
1756 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1757 struct page **hpage)
1759 struct mm_slot *mm_slot;
1760 struct mm_struct *mm;
1761 struct vm_area_struct *vma;
1765 VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
1767 if (khugepaged_scan.mm_slot)
1768 mm_slot = khugepaged_scan.mm_slot;
1770 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1771 struct mm_slot, mm_node);
1772 khugepaged_scan.address = 0;
1773 khugepaged_scan.mm_slot = mm_slot;
1775 spin_unlock(&khugepaged_mm_lock);
1778 down_read(&mm->mmap_sem);
1779 if (unlikely(khugepaged_test_exit(mm)))
1782 vma = find_vma(mm, khugepaged_scan.address);
1785 for (; vma; vma = vma->vm_next) {
1786 unsigned long hstart, hend;
1789 if (unlikely(khugepaged_test_exit(mm))) {
1794 if (!(vma->vm_flags & VM_HUGEPAGE) &&
1795 !khugepaged_always()) {
1800 /* VM_PFNMAP vmas may have vm_ops null but vm_file set */
1801 if (!vma->anon_vma || vma->vm_ops || vma->vm_file) {
1802 khugepaged_scan.address = vma->vm_end;
1806 VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
1808 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1809 hend = vma->vm_end & HPAGE_PMD_MASK;
1810 if (hstart >= hend) {
1814 if (khugepaged_scan.address < hstart)
1815 khugepaged_scan.address = hstart;
1816 if (khugepaged_scan.address > hend) {
1817 khugepaged_scan.address = hend + HPAGE_PMD_SIZE;
1821 BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1823 while (khugepaged_scan.address < hend) {
1826 if (unlikely(khugepaged_test_exit(mm)))
1827 goto breakouterloop;
1829 VM_BUG_ON(khugepaged_scan.address < hstart ||
1830 khugepaged_scan.address + HPAGE_PMD_SIZE >
1832 ret = khugepaged_scan_pmd(mm, vma,
1833 khugepaged_scan.address,
1835 /* move to next address */
1836 khugepaged_scan.address += HPAGE_PMD_SIZE;
1837 progress += HPAGE_PMD_NR;
1839 /* we released mmap_sem so break loop */
1840 goto breakouterloop_mmap_sem;
1841 if (progress >= pages)
1842 goto breakouterloop;
1846 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1847 breakouterloop_mmap_sem:
1849 spin_lock(&khugepaged_mm_lock);
1850 BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1852 * Release the current mm_slot if this mm is about to die, or
1853 * if we scanned all vmas of this mm.
1855 if (khugepaged_test_exit(mm) || !vma) {
1857 * Make sure that if mm_users is reaching zero while
1858 * khugepaged runs here, khugepaged_exit will find
1859 * mm_slot not pointing to the exiting mm.
1861 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1862 khugepaged_scan.mm_slot = list_entry(
1863 mm_slot->mm_node.next,
1864 struct mm_slot, mm_node);
1865 khugepaged_scan.address = 0;
1867 khugepaged_scan.mm_slot = NULL;
1868 khugepaged_full_scans++;
1871 collect_mm_slot(mm_slot);
1877 static int khugepaged_has_work(void)
1879 return !list_empty(&khugepaged_scan.mm_head) &&
1880 khugepaged_enabled();
1883 static int khugepaged_wait_event(void)
1885 return !list_empty(&khugepaged_scan.mm_head) ||
1886 !khugepaged_enabled();
1889 static void khugepaged_do_scan(struct page **hpage)
1891 unsigned int progress = 0, pass_through_head = 0;
1892 unsigned int pages = khugepaged_pages_to_scan;
1894 barrier(); /* write khugepaged_pages_to_scan to local stack */
1896 while (progress < pages) {
1900 *hpage = alloc_hugepage(khugepaged_defrag());
1901 if (unlikely(!*hpage))
1905 spin_lock(&khugepaged_mm_lock);
1906 if (!khugepaged_scan.mm_slot)
1907 pass_through_head++;
1908 if (khugepaged_has_work() &&
1909 pass_through_head < 2)
1910 progress += khugepaged_scan_mm_slot(pages - progress,
1914 spin_unlock(&khugepaged_mm_lock);
1918 static struct page *khugepaged_alloc_hugepage(void)
1923 hpage = alloc_hugepage(khugepaged_defrag());
1926 add_wait_queue(&khugepaged_wait, &wait);
1927 schedule_timeout_interruptible(
1929 khugepaged_alloc_sleep_millisecs));
1930 remove_wait_queue(&khugepaged_wait, &wait);
1932 } while (unlikely(!hpage) &&
1933 likely(khugepaged_enabled()));
1937 static void khugepaged_loop(void)
1941 while (likely(khugepaged_enabled())) {
1942 hpage = khugepaged_alloc_hugepage();
1943 if (unlikely(!hpage))
1946 khugepaged_do_scan(&hpage);
1949 if (khugepaged_has_work()) {
1951 if (!khugepaged_scan_sleep_millisecs)
1953 add_wait_queue(&khugepaged_wait, &wait);
1954 schedule_timeout_interruptible(
1956 khugepaged_scan_sleep_millisecs));
1957 remove_wait_queue(&khugepaged_wait, &wait);
1958 } else if (khugepaged_enabled())
1959 wait_event_interruptible(khugepaged_wait,
1960 khugepaged_wait_event());
1964 static int khugepaged(void *none)
1966 struct mm_slot *mm_slot;
1968 set_user_nice(current, 19);
1970 /* serialize with start_khugepaged() */
1971 mutex_lock(&khugepaged_mutex);
1974 mutex_unlock(&khugepaged_mutex);
1975 BUG_ON(khugepaged_thread != current);
1977 BUG_ON(khugepaged_thread != current);
1979 mutex_lock(&khugepaged_mutex);
1980 if (!khugepaged_enabled())
1984 spin_lock(&khugepaged_mm_lock);
1985 mm_slot = khugepaged_scan.mm_slot;
1986 khugepaged_scan.mm_slot = NULL;
1988 collect_mm_slot(mm_slot);
1989 spin_unlock(&khugepaged_mm_lock);
1991 khugepaged_thread = NULL;
1992 mutex_unlock(&khugepaged_mutex);
1997 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2001 spin_lock(&mm->page_table_lock);
2002 if (unlikely(!pmd_trans_huge(*pmd))) {
2003 spin_unlock(&mm->page_table_lock);
2006 page = pmd_page(*pmd);
2007 VM_BUG_ON(!page_count(page));
2009 spin_unlock(&mm->page_table_lock);
2011 split_huge_page(page);
2014 BUG_ON(pmd_trans_huge(*pmd));