2 * Copyright (C) 2009 Red Hat, Inc.
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
9 #include <linux/sched.h>
10 #include <linux/highmem.h>
11 #include <linux/hugetlb.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/rmap.h>
14 #include <linux/swap.h>
15 #include <linux/mm_inline.h>
16 #include <linux/kthread.h>
17 #include <linux/khugepaged.h>
18 #include <linux/freezer.h>
19 #include <linux/mman.h>
21 #include <asm/pgalloc.h>
25 * By default transparent hugepage support is enabled for all mappings
26 * and khugepaged scans all mappings. Defrag is only invoked by
27 * khugepaged hugepage allocations and by page faults inside
28 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
31 unsigned long transparent_hugepage_flags __read_mostly =
32 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
33 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
35 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
36 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
38 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
39 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
41 /* default scan 8*512 pte (or vmas) every 30 second */
42 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
43 static unsigned int khugepaged_pages_collapsed;
44 static unsigned int khugepaged_full_scans;
45 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
46 /* during fragmentation poll the hugepage allocator once every minute */
47 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
48 static struct task_struct *khugepaged_thread __read_mostly;
49 static DEFINE_MUTEX(khugepaged_mutex);
50 static DEFINE_SPINLOCK(khugepaged_mm_lock);
51 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
53 * default collapse hugepages if there is at least one pte mapped like
54 * it would have happened if the vma was large enough during page
57 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
59 static int khugepaged(void *none);
60 static int mm_slots_hash_init(void);
61 static int khugepaged_slab_init(void);
62 static void khugepaged_slab_free(void);
64 #define MM_SLOTS_HASH_HEADS 1024
65 static struct hlist_head *mm_slots_hash __read_mostly;
66 static struct kmem_cache *mm_slot_cache __read_mostly;
69 extern const struct file_operations fb_fops;
71 #define is_fb_vma(vma) \
72 (vma->vm_file && vma->vm_file->f_op == &fb_fops)
74 #define is_fb_vma(vma) 0
77 static void split_fb_pmd(struct vm_area_struct *vma, pmd_t *pmd);
80 * struct mm_slot - hash lookup from mm to mm_slot
81 * @hash: hash collision list
82 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
83 * @mm: the mm that this information is valid for
86 struct hlist_node hash;
87 struct list_head mm_node;
92 * struct khugepaged_scan - cursor for scanning
93 * @mm_head: the head of the mm list to scan
94 * @mm_slot: the current mm_slot we are scanning
95 * @address: the next address inside that to be scanned
97 * There is only the one khugepaged_scan instance of this cursor structure.
99 struct khugepaged_scan {
100 struct list_head mm_head;
101 struct mm_slot *mm_slot;
102 unsigned long address;
104 static struct khugepaged_scan khugepaged_scan = {
105 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
109 static int set_recommended_min_free_kbytes(void)
113 unsigned long recommended_min;
114 extern int min_free_kbytes;
116 if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG,
117 &transparent_hugepage_flags) &&
118 !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
119 &transparent_hugepage_flags))
122 for_each_populated_zone(zone)
125 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
126 recommended_min = pageblock_nr_pages * nr_zones * 2;
129 * Make sure that on average at least two pageblocks are almost free
130 * of another type, one for a migratetype to fall back to and a
131 * second to avoid subsequent fallbacks of other types There are 3
132 * MIGRATE_TYPES we care about.
134 recommended_min += pageblock_nr_pages * nr_zones *
135 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
137 /* don't ever allow to reserve more than 5% of the lowmem */
138 recommended_min = min(recommended_min,
139 (unsigned long) nr_free_buffer_pages() / 20);
140 recommended_min <<= (PAGE_SHIFT-10);
142 if (recommended_min > min_free_kbytes)
143 min_free_kbytes = recommended_min;
144 setup_per_zone_wmarks();
147 late_initcall(set_recommended_min_free_kbytes);
149 static int start_khugepaged(void)
152 if (khugepaged_enabled()) {
154 if (unlikely(!mm_slot_cache || !mm_slots_hash)) {
158 mutex_lock(&khugepaged_mutex);
159 if (!khugepaged_thread)
160 khugepaged_thread = kthread_run(khugepaged, NULL,
162 if (unlikely(IS_ERR(khugepaged_thread))) {
164 "khugepaged: kthread_run(khugepaged) failed\n");
165 err = PTR_ERR(khugepaged_thread);
166 khugepaged_thread = NULL;
168 wakeup = !list_empty(&khugepaged_scan.mm_head);
169 mutex_unlock(&khugepaged_mutex);
171 wake_up_interruptible(&khugepaged_wait);
173 set_recommended_min_free_kbytes();
176 wake_up_interruptible(&khugepaged_wait);
183 static ssize_t double_flag_show(struct kobject *kobj,
184 struct kobj_attribute *attr, char *buf,
185 enum transparent_hugepage_flag enabled,
186 enum transparent_hugepage_flag req_madv)
188 if (test_bit(enabled, &transparent_hugepage_flags)) {
189 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
190 return sprintf(buf, "[always] madvise never\n");
191 } else if (test_bit(req_madv, &transparent_hugepage_flags))
192 return sprintf(buf, "always [madvise] never\n");
194 return sprintf(buf, "always madvise [never]\n");
196 static ssize_t double_flag_store(struct kobject *kobj,
197 struct kobj_attribute *attr,
198 const char *buf, size_t count,
199 enum transparent_hugepage_flag enabled,
200 enum transparent_hugepage_flag req_madv)
202 if (!memcmp("always", buf,
203 min(sizeof("always")-1, count))) {
204 set_bit(enabled, &transparent_hugepage_flags);
205 clear_bit(req_madv, &transparent_hugepage_flags);
206 } else if (!memcmp("madvise", buf,
207 min(sizeof("madvise")-1, count))) {
208 clear_bit(enabled, &transparent_hugepage_flags);
209 set_bit(req_madv, &transparent_hugepage_flags);
210 } else if (!memcmp("never", buf,
211 min(sizeof("never")-1, count))) {
212 clear_bit(enabled, &transparent_hugepage_flags);
213 clear_bit(req_madv, &transparent_hugepage_flags);
220 static ssize_t enabled_show(struct kobject *kobj,
221 struct kobj_attribute *attr, char *buf)
223 return double_flag_show(kobj, attr, buf,
224 TRANSPARENT_HUGEPAGE_FLAG,
225 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
227 static ssize_t enabled_store(struct kobject *kobj,
228 struct kobj_attribute *attr,
229 const char *buf, size_t count)
233 ret = double_flag_store(kobj, attr, buf, count,
234 TRANSPARENT_HUGEPAGE_FLAG,
235 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
238 int err = start_khugepaged();
244 (test_bit(TRANSPARENT_HUGEPAGE_FLAG,
245 &transparent_hugepage_flags) ||
246 test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
247 &transparent_hugepage_flags)))
248 set_recommended_min_free_kbytes();
252 static struct kobj_attribute enabled_attr =
253 __ATTR(enabled, 0644, enabled_show, enabled_store);
255 static ssize_t single_flag_show(struct kobject *kobj,
256 struct kobj_attribute *attr, char *buf,
257 enum transparent_hugepage_flag flag)
259 return sprintf(buf, "%d\n",
260 !!test_bit(flag, &transparent_hugepage_flags));
263 static ssize_t single_flag_store(struct kobject *kobj,
264 struct kobj_attribute *attr,
265 const char *buf, size_t count,
266 enum transparent_hugepage_flag flag)
271 ret = kstrtoul(buf, 10, &value);
278 set_bit(flag, &transparent_hugepage_flags);
280 clear_bit(flag, &transparent_hugepage_flags);
286 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
287 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
288 * memory just to allocate one more hugepage.
290 static ssize_t defrag_show(struct kobject *kobj,
291 struct kobj_attribute *attr, char *buf)
293 return double_flag_show(kobj, attr, buf,
294 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
295 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
297 static ssize_t defrag_store(struct kobject *kobj,
298 struct kobj_attribute *attr,
299 const char *buf, size_t count)
301 return double_flag_store(kobj, attr, buf, count,
302 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
303 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
305 static struct kobj_attribute defrag_attr =
306 __ATTR(defrag, 0644, defrag_show, defrag_store);
308 #ifdef CONFIG_DEBUG_VM
309 static ssize_t debug_cow_show(struct kobject *kobj,
310 struct kobj_attribute *attr, char *buf)
312 return single_flag_show(kobj, attr, buf,
313 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
315 static ssize_t debug_cow_store(struct kobject *kobj,
316 struct kobj_attribute *attr,
317 const char *buf, size_t count)
319 return single_flag_store(kobj, attr, buf, count,
320 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
322 static struct kobj_attribute debug_cow_attr =
323 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
324 #endif /* CONFIG_DEBUG_VM */
326 static struct attribute *hugepage_attr[] = {
329 #ifdef CONFIG_DEBUG_VM
330 &debug_cow_attr.attr,
335 static struct attribute_group hugepage_attr_group = {
336 .attrs = hugepage_attr,
339 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
340 struct kobj_attribute *attr,
343 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
346 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
347 struct kobj_attribute *attr,
348 const char *buf, size_t count)
353 err = strict_strtoul(buf, 10, &msecs);
354 if (err || msecs > UINT_MAX)
357 khugepaged_scan_sleep_millisecs = msecs;
358 wake_up_interruptible(&khugepaged_wait);
362 static struct kobj_attribute scan_sleep_millisecs_attr =
363 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
364 scan_sleep_millisecs_store);
366 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
367 struct kobj_attribute *attr,
370 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
373 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
374 struct kobj_attribute *attr,
375 const char *buf, size_t count)
380 err = strict_strtoul(buf, 10, &msecs);
381 if (err || msecs > UINT_MAX)
384 khugepaged_alloc_sleep_millisecs = msecs;
385 wake_up_interruptible(&khugepaged_wait);
389 static struct kobj_attribute alloc_sleep_millisecs_attr =
390 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
391 alloc_sleep_millisecs_store);
393 static ssize_t pages_to_scan_show(struct kobject *kobj,
394 struct kobj_attribute *attr,
397 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
399 static ssize_t pages_to_scan_store(struct kobject *kobj,
400 struct kobj_attribute *attr,
401 const char *buf, size_t count)
406 err = strict_strtoul(buf, 10, &pages);
407 if (err || !pages || pages > UINT_MAX)
410 khugepaged_pages_to_scan = pages;
414 static struct kobj_attribute pages_to_scan_attr =
415 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
416 pages_to_scan_store);
418 static ssize_t pages_collapsed_show(struct kobject *kobj,
419 struct kobj_attribute *attr,
422 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
424 static struct kobj_attribute pages_collapsed_attr =
425 __ATTR_RO(pages_collapsed);
427 static ssize_t full_scans_show(struct kobject *kobj,
428 struct kobj_attribute *attr,
431 return sprintf(buf, "%u\n", khugepaged_full_scans);
433 static struct kobj_attribute full_scans_attr =
434 __ATTR_RO(full_scans);
436 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
437 struct kobj_attribute *attr, char *buf)
439 return single_flag_show(kobj, attr, buf,
440 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
442 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
443 struct kobj_attribute *attr,
444 const char *buf, size_t count)
446 return single_flag_store(kobj, attr, buf, count,
447 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
449 static struct kobj_attribute khugepaged_defrag_attr =
450 __ATTR(defrag, 0644, khugepaged_defrag_show,
451 khugepaged_defrag_store);
454 * max_ptes_none controls if khugepaged should collapse hugepages over
455 * any unmapped ptes in turn potentially increasing the memory
456 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
457 * reduce the available free memory in the system as it
458 * runs. Increasing max_ptes_none will instead potentially reduce the
459 * free memory in the system during the khugepaged scan.
461 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
462 struct kobj_attribute *attr,
465 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
467 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
468 struct kobj_attribute *attr,
469 const char *buf, size_t count)
472 unsigned long max_ptes_none;
474 err = strict_strtoul(buf, 10, &max_ptes_none);
475 if (err || max_ptes_none > HPAGE_PMD_NR-1)
478 khugepaged_max_ptes_none = max_ptes_none;
482 static struct kobj_attribute khugepaged_max_ptes_none_attr =
483 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
484 khugepaged_max_ptes_none_store);
486 static struct attribute *khugepaged_attr[] = {
487 &khugepaged_defrag_attr.attr,
488 &khugepaged_max_ptes_none_attr.attr,
489 &pages_to_scan_attr.attr,
490 &pages_collapsed_attr.attr,
491 &full_scans_attr.attr,
492 &scan_sleep_millisecs_attr.attr,
493 &alloc_sleep_millisecs_attr.attr,
497 static struct attribute_group khugepaged_attr_group = {
498 .attrs = khugepaged_attr,
499 .name = "khugepaged",
501 #endif /* CONFIG_SYSFS */
503 static int __init hugepage_init(void)
507 static struct kobject *hugepage_kobj;
511 if (!has_transparent_hugepage()) {
512 transparent_hugepage_flags = 0;
518 hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
519 if (unlikely(!hugepage_kobj)) {
520 printk(KERN_ERR "hugepage: failed kobject create\n");
524 err = sysfs_create_group(hugepage_kobj, &hugepage_attr_group);
526 printk(KERN_ERR "hugepage: failed register hugeage group\n");
530 err = sysfs_create_group(hugepage_kobj, &khugepaged_attr_group);
532 printk(KERN_ERR "hugepage: failed register hugeage group\n");
537 err = khugepaged_slab_init();
541 err = mm_slots_hash_init();
543 khugepaged_slab_free();
548 * By default disable transparent hugepages on smaller systems,
549 * where the extra memory used could hurt more than TLB overhead
550 * is likely to save. The admin can still enable it through /sys.
552 if (totalram_pages < (200 << (20 - PAGE_SHIFT)))
553 transparent_hugepage_flags = 0;
557 set_recommended_min_free_kbytes();
562 module_init(hugepage_init)
564 static int __init setup_transparent_hugepage(char *str)
569 if (!strcmp(str, "always")) {
570 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
571 &transparent_hugepage_flags);
572 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
573 &transparent_hugepage_flags);
575 } else if (!strcmp(str, "madvise")) {
576 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
577 &transparent_hugepage_flags);
578 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
579 &transparent_hugepage_flags);
581 } else if (!strcmp(str, "never")) {
582 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
583 &transparent_hugepage_flags);
584 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
585 &transparent_hugepage_flags);
591 "transparent_hugepage= cannot parse, ignored\n");
594 __setup("transparent_hugepage=", setup_transparent_hugepage);
596 static void prepare_pmd_huge_pte(pgtable_t pgtable,
597 struct mm_struct *mm)
599 assert_spin_locked(&mm->page_table_lock);
602 if (!mm->pmd_huge_pte)
603 INIT_LIST_HEAD(&pgtable->lru);
605 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
606 mm->pmd_huge_pte = pgtable;
609 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
611 if (likely(vma->vm_flags & VM_WRITE))
612 pmd = pmd_mkwrite(pmd);
616 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
617 struct vm_area_struct *vma,
618 unsigned long haddr, pmd_t *pmd,
624 VM_BUG_ON(!PageCompound(page));
625 pgtable = pte_alloc_one(mm, haddr);
626 if (unlikely(!pgtable)) {
627 mem_cgroup_uncharge_page(page);
632 clear_huge_page(page, haddr, HPAGE_PMD_NR);
633 __SetPageUptodate(page);
635 spin_lock(&mm->page_table_lock);
636 if (unlikely(!pmd_none(*pmd))) {
637 spin_unlock(&mm->page_table_lock);
638 mem_cgroup_uncharge_page(page);
640 pte_free(mm, pgtable);
643 entry = mk_pmd(page, vma->vm_page_prot);
644 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
645 entry = pmd_mkhuge(entry);
647 * The spinlocking to take the lru_lock inside
648 * page_add_new_anon_rmap() acts as a full memory
649 * barrier to be sure clear_huge_page writes become
650 * visible after the set_pmd_at() write.
652 page_add_new_anon_rmap(page, vma, haddr);
653 set_pmd_at(mm, haddr, pmd, entry);
654 prepare_pmd_huge_pte(pgtable, mm);
655 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
657 spin_unlock(&mm->page_table_lock);
663 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
665 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
668 static inline struct page *alloc_hugepage_vma(int defrag,
669 struct vm_area_struct *vma,
670 unsigned long haddr, int nd,
673 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
674 HPAGE_PMD_ORDER, vma, haddr, nd);
678 static inline struct page *alloc_hugepage(int defrag)
680 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
685 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
686 unsigned long address, pmd_t *pmd,
690 unsigned long haddr = address & HPAGE_PMD_MASK;
693 if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
694 if (unlikely(anon_vma_prepare(vma)))
696 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
698 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
699 vma, haddr, numa_node_id(), 0);
700 if (unlikely(!page)) {
701 count_vm_event(THP_FAULT_FALLBACK);
704 count_vm_event(THP_FAULT_ALLOC);
705 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
710 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
714 * Use __pte_alloc instead of pte_alloc_map, because we can't
715 * run pte_offset_map on the pmd, if an huge pmd could
716 * materialize from under us from a different thread.
718 if (unlikely(__pte_alloc(mm, vma, pmd, address)))
720 /* if an huge pmd materialized from under us just retry later */
721 if (unlikely(pmd_trans_huge(*pmd)))
724 * A regular pmd is established and it can't morph into a huge pmd
725 * from under us anymore at this point because we hold the mmap_sem
726 * read mode and khugepaged takes it in write mode. So now it's
727 * safe to run pte_offset_map().
729 pte = pte_offset_map(pmd, address);
730 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
733 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
734 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
735 struct vm_area_struct *vma)
737 struct page *src_page;
743 pgtable = pte_alloc_one(dst_mm, addr);
744 if (unlikely(!pgtable))
747 spin_lock(&dst_mm->page_table_lock);
748 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
752 if (unlikely(!pmd_trans_huge(pmd))) {
753 pte_free(dst_mm, pgtable);
756 if (unlikely(pmd_trans_splitting(pmd))) {
757 /* split huge page running from under us */
758 spin_unlock(&src_mm->page_table_lock);
759 spin_unlock(&dst_mm->page_table_lock);
760 pte_free(dst_mm, pgtable);
762 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
765 src_page = pmd_page(pmd);
766 VM_BUG_ON(!PageHead(src_page));
768 page_dup_rmap(src_page);
769 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
771 pmdp_set_wrprotect(src_mm, addr, src_pmd);
772 pmd = pmd_mkold(pmd_wrprotect(pmd));
773 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
774 prepare_pmd_huge_pte(pgtable, dst_mm);
779 spin_unlock(&src_mm->page_table_lock);
780 spin_unlock(&dst_mm->page_table_lock);
785 /* no "address" argument so destroys page coloring of some arch */
786 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
790 assert_spin_locked(&mm->page_table_lock);
793 pgtable = mm->pmd_huge_pte;
794 if (list_empty(&pgtable->lru))
795 mm->pmd_huge_pte = NULL;
797 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
799 list_del(&pgtable->lru);
804 void huge_pmd_set_accessed(struct mm_struct *mm,
805 struct vm_area_struct *vma,
806 unsigned long address,
807 pmd_t *pmd, pmd_t orig_pmd,
813 spin_lock(&mm->page_table_lock);
814 if (unlikely(!pmd_same(*pmd, orig_pmd)))
817 entry = pmd_mkyoung(orig_pmd);
818 haddr = address & HPAGE_PMD_MASK;
819 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
820 update_mmu_cache_pmd(vma, address, pmd);
823 spin_unlock(&mm->page_table_lock);
826 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
827 struct vm_area_struct *vma,
828 unsigned long address,
829 pmd_t *pmd, pmd_t orig_pmd,
838 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
840 if (unlikely(!pages)) {
845 for (i = 0; i < HPAGE_PMD_NR; i++) {
846 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
848 vma, address, page_to_nid(page));
849 if (unlikely(!pages[i] ||
850 mem_cgroup_newpage_charge(pages[i], mm,
854 mem_cgroup_uncharge_start();
856 mem_cgroup_uncharge_page(pages[i]);
859 mem_cgroup_uncharge_end();
866 for (i = 0; i < HPAGE_PMD_NR; i++) {
867 copy_user_highpage(pages[i], page + i,
868 haddr + PAGE_SIZE * i, vma);
869 __SetPageUptodate(pages[i]);
873 spin_lock(&mm->page_table_lock);
874 if (unlikely(!pmd_same(*pmd, orig_pmd)))
876 VM_BUG_ON(!PageHead(page));
878 pmdp_clear_flush_notify(vma, haddr, pmd);
879 /* leave pmd empty until pte is filled */
881 pgtable = get_pmd_huge_pte(mm);
882 pmd_populate(mm, &_pmd, pgtable);
884 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
886 entry = mk_pte(pages[i], vma->vm_page_prot);
887 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
888 page_add_new_anon_rmap(pages[i], vma, haddr);
889 pte = pte_offset_map(&_pmd, haddr);
890 VM_BUG_ON(!pte_none(*pte));
891 set_pte_at(mm, haddr, pte, entry);
896 smp_wmb(); /* make pte visible before pmd */
897 pmd_populate(mm, pmd, pgtable);
898 page_remove_rmap(page);
899 spin_unlock(&mm->page_table_lock);
901 ret |= VM_FAULT_WRITE;
908 spin_unlock(&mm->page_table_lock);
909 mem_cgroup_uncharge_start();
910 for (i = 0; i < HPAGE_PMD_NR; i++) {
911 mem_cgroup_uncharge_page(pages[i]);
914 mem_cgroup_uncharge_end();
919 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
920 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
923 struct page *page, *new_page;
926 VM_BUG_ON(!vma->anon_vma);
927 spin_lock(&mm->page_table_lock);
928 if (unlikely(!pmd_same(*pmd, orig_pmd)))
931 page = pmd_page(orig_pmd);
932 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
933 haddr = address & HPAGE_PMD_MASK;
934 if (page_mapcount(page) == 1) {
936 entry = pmd_mkyoung(orig_pmd);
937 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
938 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
939 update_mmu_cache(vma, address, pmd);
940 ret |= VM_FAULT_WRITE;
944 spin_unlock(&mm->page_table_lock);
946 if (transparent_hugepage_enabled(vma) &&
947 !transparent_hugepage_debug_cow())
948 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
949 vma, haddr, numa_node_id(), 0);
953 if (unlikely(!new_page)) {
954 count_vm_event(THP_FAULT_FALLBACK);
955 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
956 pmd, orig_pmd, page, haddr);
957 if (ret & VM_FAULT_OOM)
958 split_huge_page(page);
962 count_vm_event(THP_FAULT_ALLOC);
964 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
966 split_huge_page(page);
972 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
973 __SetPageUptodate(new_page);
975 spin_lock(&mm->page_table_lock);
977 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
978 mem_cgroup_uncharge_page(new_page);
982 VM_BUG_ON(!PageHead(page));
983 entry = mk_pmd(new_page, vma->vm_page_prot);
984 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
985 entry = pmd_mkhuge(entry);
986 pmdp_clear_flush_notify(vma, haddr, pmd);
987 page_add_new_anon_rmap(new_page, vma, haddr);
988 set_pmd_at(mm, haddr, pmd, entry);
989 update_mmu_cache(vma, address, pmd);
990 page_remove_rmap(page);
992 ret |= VM_FAULT_WRITE;
995 spin_unlock(&mm->page_table_lock);
1000 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
1005 struct page *page = NULL;
1007 assert_spin_locked(&mm->page_table_lock);
1009 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1012 page = pmd_page(*pmd);
1013 VM_BUG_ON(!PageHead(page));
1014 if (flags & FOLL_TOUCH) {
1017 * We should set the dirty bit only for FOLL_WRITE but
1018 * for now the dirty bit in the pmd is meaningless.
1019 * And if the dirty bit will become meaningful and
1020 * we'll only set it with FOLL_WRITE, an atomic
1021 * set_bit will be required on the pmd to set the
1022 * young bit, instead of the current set_pmd_at.
1024 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1025 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1027 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1028 VM_BUG_ON(!PageCompound(page));
1029 if (flags & FOLL_GET)
1030 get_page_foll(page);
1036 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1041 spin_lock(&tlb->mm->page_table_lock);
1042 if (likely(pmd_trans_huge(*pmd))) {
1043 if (is_fb_vma(vma)) {
1044 split_fb_pmd(vma, pmd);
1048 if (unlikely(pmd_trans_splitting(*pmd))) {
1049 spin_unlock(&tlb->mm->page_table_lock);
1050 wait_split_huge_page(vma->anon_vma,
1055 pgtable = get_pmd_huge_pte(tlb->mm);
1056 page = pmd_page(*pmd);
1058 page_remove_rmap(page);
1059 VM_BUG_ON(page_mapcount(page) < 0);
1060 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1061 VM_BUG_ON(!PageHead(page));
1063 spin_unlock(&tlb->mm->page_table_lock);
1064 tlb_remove_page(tlb, page);
1065 pte_free(tlb->mm, pgtable);
1069 spin_unlock(&tlb->mm->page_table_lock);
1074 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1075 unsigned long addr, unsigned long end,
1080 spin_lock(&vma->vm_mm->page_table_lock);
1081 if (likely(pmd_trans_huge(*pmd))) {
1082 ret = !pmd_trans_splitting(*pmd);
1083 spin_unlock(&vma->vm_mm->page_table_lock);
1085 wait_split_huge_page(vma->anon_vma, pmd);
1088 * All logical pages in the range are present
1089 * if backed by a huge page.
1091 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1094 spin_unlock(&vma->vm_mm->page_table_lock);
1099 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1100 unsigned long old_addr,
1101 unsigned long new_addr, unsigned long old_end,
1102 pmd_t *old_pmd, pmd_t *new_pmd)
1107 struct mm_struct *mm = vma->vm_mm;
1109 if ((old_addr & ~HPAGE_PMD_MASK) ||
1110 (new_addr & ~HPAGE_PMD_MASK) ||
1111 old_end - old_addr < HPAGE_PMD_SIZE ||
1112 (new_vma->vm_flags & VM_NOHUGEPAGE))
1116 * The destination pmd shouldn't be established, free_pgtables()
1117 * should have release it.
1119 if (WARN_ON(!pmd_none(*new_pmd))) {
1120 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1124 spin_lock(&mm->page_table_lock);
1125 if (likely(pmd_trans_huge(*old_pmd))) {
1126 if (pmd_trans_splitting(*old_pmd)) {
1127 spin_unlock(&mm->page_table_lock);
1128 wait_split_huge_page(vma->anon_vma, old_pmd);
1131 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1132 VM_BUG_ON(!pmd_none(*new_pmd));
1133 set_pmd_at(mm, new_addr, new_pmd, pmd);
1134 spin_unlock(&mm->page_table_lock);
1138 spin_unlock(&mm->page_table_lock);
1144 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1145 unsigned long addr, pgprot_t newprot)
1147 struct mm_struct *mm = vma->vm_mm;
1150 spin_lock(&mm->page_table_lock);
1151 if (likely(pmd_trans_huge(*pmd))) {
1152 if (unlikely(pmd_trans_splitting(*pmd))) {
1153 spin_unlock(&mm->page_table_lock);
1154 wait_split_huge_page(vma->anon_vma, pmd);
1158 entry = pmdp_get_and_clear(mm, addr, pmd);
1159 entry = pmd_modify(entry, newprot);
1160 set_pmd_at(mm, addr, pmd, entry);
1161 spin_unlock(&vma->vm_mm->page_table_lock);
1162 flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1166 spin_unlock(&vma->vm_mm->page_table_lock);
1171 pmd_t *page_check_address_pmd(struct page *page,
1172 struct mm_struct *mm,
1173 unsigned long address,
1174 enum page_check_address_pmd_flag flag)
1178 pmd_t *pmd, *ret = NULL;
1180 if (address & ~HPAGE_PMD_MASK)
1183 pgd = pgd_offset(mm, address);
1184 if (!pgd_present(*pgd))
1187 pud = pud_offset(pgd, address);
1188 if (!pud_present(*pud))
1191 pmd = pmd_offset(pud, address);
1194 if (pmd_page(*pmd) != page)
1197 * split_vma() may create temporary aliased mappings. There is
1198 * no risk as long as all huge pmd are found and have their
1199 * splitting bit set before __split_huge_page_refcount
1200 * runs. Finding the same huge pmd more than once during the
1201 * same rmap walk is not a problem.
1203 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1204 pmd_trans_splitting(*pmd))
1206 if (pmd_trans_huge(*pmd)) {
1207 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1208 !pmd_trans_splitting(*pmd));
1215 static int __split_huge_page_splitting(struct page *page,
1216 struct vm_area_struct *vma,
1217 unsigned long address)
1219 struct mm_struct *mm = vma->vm_mm;
1223 spin_lock(&mm->page_table_lock);
1224 pmd = page_check_address_pmd(page, mm, address,
1225 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1228 * We can't temporarily set the pmd to null in order
1229 * to split it, the pmd must remain marked huge at all
1230 * times or the VM won't take the pmd_trans_huge paths
1231 * and it won't wait on the anon_vma->root->mutex to
1232 * serialize against split_huge_page*.
1234 pmdp_splitting_flush_notify(vma, address, pmd);
1237 spin_unlock(&mm->page_table_lock);
1242 static void __split_huge_page_refcount(struct page *page)
1245 unsigned long head_index = page->index;
1246 struct zone *zone = page_zone(page);
1250 /* prevent PageLRU to go away from under us, and freeze lru stats */
1251 spin_lock_irq(&zone->lru_lock);
1252 compound_lock(page);
1254 for (i = 1; i < HPAGE_PMD_NR; i++) {
1255 struct page *page_tail = page + i;
1257 /* tail_page->_mapcount cannot change */
1258 BUG_ON(page_mapcount(page_tail) < 0);
1259 tail_count += page_mapcount(page_tail);
1260 /* check for overflow */
1261 BUG_ON(tail_count < 0);
1262 BUG_ON(atomic_read(&page_tail->_count) != 0);
1264 * tail_page->_count is zero and not changing from
1265 * under us. But get_page_unless_zero() may be running
1266 * from under us on the tail_page. If we used
1267 * atomic_set() below instead of atomic_add(), we
1268 * would then run atomic_set() concurrently with
1269 * get_page_unless_zero(), and atomic_set() is
1270 * implemented in C not using locked ops. spin_unlock
1271 * on x86 sometime uses locked ops because of PPro
1272 * errata 66, 92, so unless somebody can guarantee
1273 * atomic_set() here would be safe on all archs (and
1274 * not only on x86), it's safer to use atomic_add().
1276 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1277 &page_tail->_count);
1279 /* after clearing PageTail the gup refcount can be released */
1283 * retain hwpoison flag of the poisoned tail page:
1284 * fix for the unsuitable process killed on Guest Machine(KVM)
1285 * by the memory-failure.
1287 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1288 page_tail->flags |= (page->flags &
1289 ((1L << PG_referenced) |
1290 (1L << PG_swapbacked) |
1291 (1L << PG_mlocked) |
1292 (1L << PG_uptodate)));
1293 page_tail->flags |= (1L << PG_dirty);
1295 /* clear PageTail before overwriting first_page */
1299 * __split_huge_page_splitting() already set the
1300 * splitting bit in all pmd that could map this
1301 * hugepage, that will ensure no CPU can alter the
1302 * mapcount on the head page. The mapcount is only
1303 * accounted in the head page and it has to be
1304 * transferred to all tail pages in the below code. So
1305 * for this code to be safe, the split the mapcount
1306 * can't change. But that doesn't mean userland can't
1307 * keep changing and reading the page contents while
1308 * we transfer the mapcount, so the pmd splitting
1309 * status is achieved setting a reserved bit in the
1310 * pmd, not by clearing the present bit.
1312 page_tail->_mapcount = page->_mapcount;
1314 BUG_ON(page_tail->mapping);
1315 page_tail->mapping = page->mapping;
1317 page_tail->index = ++head_index;
1319 BUG_ON(!PageAnon(page_tail));
1320 BUG_ON(!PageUptodate(page_tail));
1321 BUG_ON(!PageDirty(page_tail));
1322 BUG_ON(!PageSwapBacked(page_tail));
1324 mem_cgroup_split_huge_fixup(page, page_tail);
1326 lru_add_page_tail(zone, page, page_tail);
1328 atomic_sub(tail_count, &page->_count);
1329 BUG_ON(atomic_read(&page->_count) <= 0);
1331 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1332 __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1335 * A hugepage counts for HPAGE_PMD_NR pages on the LRU statistics,
1336 * so adjust those appropriately if this page is on the LRU.
1338 if (PageLRU(page)) {
1339 zonestat = NR_LRU_BASE + page_lru(page);
1340 __mod_zone_page_state(zone, zonestat, -(HPAGE_PMD_NR-1));
1343 ClearPageCompound(page);
1344 compound_unlock(page);
1345 spin_unlock_irq(&zone->lru_lock);
1347 for (i = 1; i < HPAGE_PMD_NR; i++) {
1348 struct page *page_tail = page + i;
1349 BUG_ON(page_count(page_tail) <= 0);
1351 * Tail pages may be freed if there wasn't any mapping
1352 * like if add_to_swap() is running on a lru page that
1353 * had its mapping zapped. And freeing these pages
1354 * requires taking the lru_lock so we do the put_page
1355 * of the tail pages after the split is complete.
1357 put_page(page_tail);
1361 * Only the head page (now become a regular page) is required
1362 * to be pinned by the caller.
1364 BUG_ON(page_count(page) <= 0);
1367 static int __split_huge_page_map(struct page *page,
1368 struct vm_area_struct *vma,
1369 unsigned long address)
1371 struct mm_struct *mm = vma->vm_mm;
1375 unsigned long haddr;
1377 spin_lock(&mm->page_table_lock);
1378 pmd = page_check_address_pmd(page, mm, address,
1379 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1381 pgtable = get_pmd_huge_pte(mm);
1382 pmd_populate(mm, &_pmd, pgtable);
1384 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1385 i++, haddr += PAGE_SIZE) {
1387 BUG_ON(PageCompound(page+i));
1388 entry = mk_pte(page + i, vma->vm_page_prot);
1389 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1390 if (!pmd_write(*pmd))
1391 entry = pte_wrprotect(entry);
1393 BUG_ON(page_mapcount(page) != 1);
1394 if (!pmd_young(*pmd))
1395 entry = pte_mkold(entry);
1396 pte = pte_offset_map(&_pmd, haddr);
1397 BUG_ON(!pte_none(*pte));
1398 set_pte_at(mm, haddr, pte, entry);
1402 smp_wmb(); /* make pte visible before pmd */
1404 * Up to this point the pmd is present and huge and
1405 * userland has the whole access to the hugepage
1406 * during the split (which happens in place). If we
1407 * overwrite the pmd with the not-huge version
1408 * pointing to the pte here (which of course we could
1409 * if all CPUs were bug free), userland could trigger
1410 * a small page size TLB miss on the small sized TLB
1411 * while the hugepage TLB entry is still established
1412 * in the huge TLB. Some CPU doesn't like that. See
1413 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1414 * Erratum 383 on page 93. Intel should be safe but is
1415 * also warns that it's only safe if the permission
1416 * and cache attributes of the two entries loaded in
1417 * the two TLB is identical (which should be the case
1418 * here). But it is generally safer to never allow
1419 * small and huge TLB entries for the same virtual
1420 * address to be loaded simultaneously. So instead of
1421 * doing "pmd_populate(); flush_tlb_range();" we first
1422 * mark the current pmd notpresent (atomically because
1423 * here the pmd_trans_huge and pmd_trans_splitting
1424 * must remain set at all times on the pmd until the
1425 * split is complete for this pmd), then we flush the
1426 * SMP TLB and finally we write the non-huge version
1427 * of the pmd entry with pmd_populate.
1429 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1430 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1431 pmd_populate(mm, pmd, pgtable);
1434 spin_unlock(&mm->page_table_lock);
1439 /* must be called with anon_vma->root->mutex hold */
1440 static void __split_huge_page(struct page *page,
1441 struct anon_vma *anon_vma)
1443 int mapcount, mapcount2;
1444 struct anon_vma_chain *avc;
1446 BUG_ON(!PageHead(page));
1447 BUG_ON(PageTail(page));
1450 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1451 struct vm_area_struct *vma = avc->vma;
1452 unsigned long addr = vma_address(page, vma);
1453 BUG_ON(is_vma_temporary_stack(vma));
1454 if (addr == -EFAULT)
1456 mapcount += __split_huge_page_splitting(page, vma, addr);
1459 * It is critical that new vmas are added to the tail of the
1460 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1461 * and establishes a child pmd before
1462 * __split_huge_page_splitting() freezes the parent pmd (so if
1463 * we fail to prevent copy_huge_pmd() from running until the
1464 * whole __split_huge_page() is complete), we will still see
1465 * the newly established pmd of the child later during the
1466 * walk, to be able to set it as pmd_trans_splitting too.
1468 if (mapcount != page_mapcount(page))
1469 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1470 mapcount, page_mapcount(page));
1471 BUG_ON(mapcount != page_mapcount(page));
1473 __split_huge_page_refcount(page);
1476 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1477 struct vm_area_struct *vma = avc->vma;
1478 unsigned long addr = vma_address(page, vma);
1479 BUG_ON(is_vma_temporary_stack(vma));
1480 if (addr == -EFAULT)
1482 mapcount2 += __split_huge_page_map(page, vma, addr);
1484 if (mapcount != mapcount2)
1485 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1486 mapcount, mapcount2, page_mapcount(page));
1487 BUG_ON(mapcount != mapcount2);
1490 int split_huge_page(struct page *page)
1492 struct anon_vma *anon_vma;
1495 BUG_ON(!PageAnon(page));
1496 anon_vma = page_lock_anon_vma(page);
1500 if (!PageCompound(page))
1503 BUG_ON(!PageSwapBacked(page));
1504 __split_huge_page(page, anon_vma);
1505 count_vm_event(THP_SPLIT);
1507 BUG_ON(PageCompound(page));
1509 page_unlock_anon_vma(anon_vma);
1514 /* callers must hold mmap_sem (madvise() does) */
1515 static int collapse_fb_pmd(struct mm_struct *mm, pmd_t *pmd,
1516 unsigned long addr, struct vm_area_struct *vma)
1518 unsigned long _addr;
1525 pte = pte_offset_map(pmd, addr);
1526 page = pte_page(*pte);
1527 pa = __pfn_to_phys(page_to_pfn(page));
1528 _pmd = pmdp_clear_flush_notify(vma, addr, pmd);
1530 if ((addr | pa) & ~HPAGE_PMD_MASK) {
1531 printk(KERN_ERR "collapse_fb: bad alignment: %08lx->%08x\n",
1537 for (_pte = pte, _addr = addr; _pte < pte + HPAGE_PMD_NR; _pte++) {
1538 pte_t pteval = *_pte;
1539 struct page *src_page;
1541 if (!pte_none(pteval)) {
1542 src_page = pte_page(pteval);
1544 pte_clear(vma->vm_mm, _addr, _pte);
1545 if (pte_present(pteval))
1546 page_remove_rmap(src_page);
1553 pgtable = pmd_pgtable(_pmd);
1554 VM_BUG_ON(page_count(pgtable) != 1);
1555 VM_BUG_ON(page_mapcount(pgtable) != 0);
1557 _pmd = mk_pmd(page, vma->vm_page_prot);
1558 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1559 _pmd = pmd_mkhuge(_pmd);
1563 spin_lock(&mm->page_table_lock);
1564 BUG_ON(!pmd_none(*pmd));
1565 set_pmd_at(mm, addr, pmd, _pmd);
1566 update_mmu_cache(vma, addr, pmd);
1567 prepare_pmd_huge_pte(pgtable, mm);
1568 spin_unlock(&mm->page_table_lock);
1573 static int try_collapse_fb(struct vm_area_struct *vma)
1575 struct mm_struct *mm = vma->vm_mm;
1576 unsigned long hstart, hend, addr;
1582 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1583 hend = vma->vm_end & HPAGE_PMD_MASK;
1587 for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) {
1588 pgd = pgd_offset(mm, addr);
1589 if (!pgd_present(*pgd))
1592 pud = pud_offset(pgd, addr);
1593 if (!pud_present(*pud))
1596 pmd = pmd_offset(pud, addr);
1597 if (!pmd_present(*pmd))
1599 if (pmd_trans_huge(*pmd))
1602 ret = collapse_fb_pmd(mm, pmd, addr, vma);
1610 /* undo collapse_fb_pmd(), restore pages so that mm subsys can release them
1611 * page_table_lock() should be held */
1612 static void split_fb_pmd(struct vm_area_struct *vma, pmd_t *pmd)
1614 struct mm_struct *mm = vma->vm_mm;
1615 unsigned long addr, haddr, pfn;
1621 page = pmd_page(*pmd);
1622 pgtable = get_pmd_huge_pte(mm);
1623 pfn = page_to_pfn(page);
1624 addr = pfn << PAGE_SHIFT;
1626 pmd_populate(mm, &_pmd, pgtable);
1628 for (i = 0, haddr = addr; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1630 BUG_ON(PageCompound(page + i));
1631 entry = mk_pte(page + i, vma->vm_page_prot);
1632 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1633 if (!pmd_young(*pmd))
1634 entry = pte_mkold(entry);
1635 atomic_set(&page[i]._mapcount, 0); // hack?
1636 pte = pte_offset_map(&_pmd, haddr);
1637 BUG_ON(!pte_none(*pte));
1638 set_pte_at(mm, haddr, pte, entry);
1642 set_pmd_at(mm, addr, pmd, pmd_mknotpresent(*pmd));
1643 flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1644 pmd_populate(mm, pmd, pgtable);
1650 static u32 pmd_to_va(struct mm_struct *mm, pmd_t *pmd)
1657 pgd = pgd_offset(mm, 0);
1658 pud = pud_offset(pgd, 0);
1659 pmd0 = pmd_offset(pud, 0);
1661 ret = (pmd - pmd0) << SECTION_SHIFT;
1665 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1666 VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1668 int hugepage_madvise(struct vm_area_struct *vma,
1669 unsigned long *vm_flags, int advice)
1674 return try_collapse_fb(vma);
1677 * Be somewhat over-protective like KSM for now!
1679 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1681 *vm_flags &= ~VM_NOHUGEPAGE;
1682 *vm_flags |= VM_HUGEPAGE;
1684 * If the vma become good for khugepaged to scan,
1685 * register it here without waiting a page fault that
1686 * may not happen any time soon.
1688 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1691 case MADV_NOHUGEPAGE:
1693 * Be somewhat over-protective like KSM for now!
1695 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1697 *vm_flags &= ~VM_HUGEPAGE;
1698 *vm_flags |= VM_NOHUGEPAGE;
1700 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1701 * this vma even if we leave the mm registered in khugepaged if
1702 * it got registered before VM_NOHUGEPAGE was set.
1710 static int __init khugepaged_slab_init(void)
1712 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1713 sizeof(struct mm_slot),
1714 __alignof__(struct mm_slot), 0, NULL);
1721 static void __init khugepaged_slab_free(void)
1723 kmem_cache_destroy(mm_slot_cache);
1724 mm_slot_cache = NULL;
1727 static inline struct mm_slot *alloc_mm_slot(void)
1729 if (!mm_slot_cache) /* initialization failed */
1731 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1734 static inline void free_mm_slot(struct mm_slot *mm_slot)
1736 kmem_cache_free(mm_slot_cache, mm_slot);
1739 static int __init mm_slots_hash_init(void)
1741 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1749 static void __init mm_slots_hash_free(void)
1751 kfree(mm_slots_hash);
1752 mm_slots_hash = NULL;
1756 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1758 struct mm_slot *mm_slot;
1759 struct hlist_head *bucket;
1760 struct hlist_node *node;
1762 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1763 % MM_SLOTS_HASH_HEADS];
1764 hlist_for_each_entry(mm_slot, node, bucket, hash) {
1765 if (mm == mm_slot->mm)
1771 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1772 struct mm_slot *mm_slot)
1774 struct hlist_head *bucket;
1776 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1777 % MM_SLOTS_HASH_HEADS];
1779 hlist_add_head(&mm_slot->hash, bucket);
1782 static inline int khugepaged_test_exit(struct mm_struct *mm)
1784 return atomic_read(&mm->mm_users) == 0;
1787 int __khugepaged_enter(struct mm_struct *mm)
1789 struct mm_slot *mm_slot;
1792 mm_slot = alloc_mm_slot();
1796 /* __khugepaged_exit() must not run from under us */
1797 VM_BUG_ON(khugepaged_test_exit(mm));
1798 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1799 free_mm_slot(mm_slot);
1803 spin_lock(&khugepaged_mm_lock);
1804 insert_to_mm_slots_hash(mm, mm_slot);
1806 * Insert just behind the scanning cursor, to let the area settle
1809 wakeup = list_empty(&khugepaged_scan.mm_head);
1810 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1811 spin_unlock(&khugepaged_mm_lock);
1813 atomic_inc(&mm->mm_count);
1815 wake_up_interruptible(&khugepaged_wait);
1820 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1821 unsigned long vm_flags)
1823 unsigned long hstart, hend;
1826 * Not yet faulted in so we will register later in the
1827 * page fault if needed.
1830 if (vma->vm_ops || (vm_flags & VM_NO_THP))
1831 /* khugepaged not yet working on file or special mappings */
1834 * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1835 * true too, verify it here.
1837 VM_BUG_ON(is_linear_pfn_mapping(vma));
1838 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1839 hend = vma->vm_end & HPAGE_PMD_MASK;
1841 return khugepaged_enter(vma, vm_flags);
1845 void __khugepaged_exit(struct mm_struct *mm)
1847 struct mm_slot *mm_slot;
1850 spin_lock(&khugepaged_mm_lock);
1851 mm_slot = get_mm_slot(mm);
1852 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1853 hlist_del(&mm_slot->hash);
1854 list_del(&mm_slot->mm_node);
1857 spin_unlock(&khugepaged_mm_lock);
1860 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1861 free_mm_slot(mm_slot);
1863 } else if (mm_slot) {
1865 * This is required to serialize against
1866 * khugepaged_test_exit() (which is guaranteed to run
1867 * under mmap sem read mode). Stop here (after we
1868 * return all pagetables will be destroyed) until
1869 * khugepaged has finished working on the pagetables
1870 * under the mmap_sem.
1872 down_write(&mm->mmap_sem);
1873 up_write(&mm->mmap_sem);
1877 static void release_pte_page(struct page *page)
1879 /* 0 stands for page_is_file_cache(page) == false */
1880 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1882 putback_lru_page(page);
1885 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1887 while (--_pte >= pte) {
1888 pte_t pteval = *_pte;
1889 if (!pte_none(pteval))
1890 release_pte_page(pte_page(pteval));
1894 static void release_all_pte_pages(pte_t *pte)
1896 release_pte_pages(pte, pte + HPAGE_PMD_NR);
1899 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1900 unsigned long address,
1905 int referenced = 0, isolated = 0, none = 0;
1906 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1907 _pte++, address += PAGE_SIZE) {
1908 pte_t pteval = *_pte;
1909 if (pte_none(pteval)) {
1910 if (++none <= khugepaged_max_ptes_none)
1913 release_pte_pages(pte, _pte);
1917 if (!pte_present(pteval) || !pte_write(pteval)) {
1918 release_pte_pages(pte, _pte);
1921 page = vm_normal_page(vma, address, pteval);
1922 if (unlikely(!page)) {
1923 release_pte_pages(pte, _pte);
1926 VM_BUG_ON(PageCompound(page));
1927 BUG_ON(!PageAnon(page));
1928 VM_BUG_ON(!PageSwapBacked(page));
1930 /* cannot use mapcount: can't collapse if there's a gup pin */
1931 if (page_count(page) != 1) {
1932 release_pte_pages(pte, _pte);
1936 * We can do it before isolate_lru_page because the
1937 * page can't be freed from under us. NOTE: PG_lock
1938 * is needed to serialize against split_huge_page
1939 * when invoked from the VM.
1941 if (!trylock_page(page)) {
1942 release_pte_pages(pte, _pte);
1946 * Isolate the page to avoid collapsing an hugepage
1947 * currently in use by the VM.
1949 if (isolate_lru_page(page)) {
1951 release_pte_pages(pte, _pte);
1954 /* 0 stands for page_is_file_cache(page) == false */
1955 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1956 VM_BUG_ON(!PageLocked(page));
1957 VM_BUG_ON(PageLRU(page));
1959 /* If there is no mapped pte young don't collapse the page */
1960 if (pte_young(pteval) || PageReferenced(page) ||
1961 mmu_notifier_test_young(vma->vm_mm, address))
1964 if (unlikely(!referenced))
1965 release_all_pte_pages(pte);
1972 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1973 struct vm_area_struct *vma,
1974 unsigned long address,
1978 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1979 pte_t pteval = *_pte;
1980 struct page *src_page;
1982 if (pte_none(pteval)) {
1983 clear_user_highpage(page, address);
1984 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1986 src_page = pte_page(pteval);
1987 copy_user_highpage(page, src_page, address, vma);
1988 VM_BUG_ON(page_mapcount(src_page) != 1);
1989 VM_BUG_ON(page_count(src_page) != 2);
1990 release_pte_page(src_page);
1992 * ptl mostly unnecessary, but preempt has to
1993 * be disabled to update the per-cpu stats
1994 * inside page_remove_rmap().
1998 * paravirt calls inside pte_clear here are
2001 pte_clear(vma->vm_mm, address, _pte);
2002 page_remove_rmap(src_page);
2004 free_page_and_swap_cache(src_page);
2007 address += PAGE_SIZE;
2012 static bool hugepage_vma_check(struct vm_area_struct *vma)
2014 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2015 (vma->vm_flags & VM_NOHUGEPAGE))
2018 if (!vma->anon_vma || vma->vm_ops)
2020 if (is_vma_temporary_stack(vma))
2023 * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
2024 * true too, verify it here.
2026 VM_BUG_ON(is_linear_pfn_mapping(vma));
2027 return !(vma->vm_flags & VM_NO_THP);
2030 static void collapse_huge_page(struct mm_struct *mm,
2031 unsigned long address,
2032 struct page **hpage,
2033 struct vm_area_struct *vma,
2041 struct page *new_page;
2044 unsigned long hstart, hend;
2046 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2048 up_read(&mm->mmap_sem);
2054 * Allocate the page while the vma is still valid and under
2055 * the mmap_sem read mode so there is no memory allocation
2056 * later when we take the mmap_sem in write mode. This is more
2057 * friendly behavior (OTOH it may actually hide bugs) to
2058 * filesystems in userland with daemons allocating memory in
2059 * the userland I/O paths. Allocating memory with the
2060 * mmap_sem in read mode is good idea also to allow greater
2063 new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2064 node, __GFP_OTHER_NODE);
2067 * After allocating the hugepage, release the mmap_sem read lock in
2068 * preparation for taking it in write mode.
2070 up_read(&mm->mmap_sem);
2071 if (unlikely(!new_page)) {
2072 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2073 *hpage = ERR_PTR(-ENOMEM);
2078 count_vm_event(THP_COLLAPSE_ALLOC);
2079 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
2087 * Prevent all access to pagetables with the exception of
2088 * gup_fast later hanlded by the ptep_clear_flush and the VM
2089 * handled by the anon_vma lock + PG_lock.
2091 down_write(&mm->mmap_sem);
2092 if (unlikely(khugepaged_test_exit(mm)))
2095 vma = find_vma(mm, address);
2098 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2099 hend = vma->vm_end & HPAGE_PMD_MASK;
2100 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2102 if (!hugepage_vma_check(vma))
2104 pgd = pgd_offset(mm, address);
2105 if (!pgd_present(*pgd))
2108 pud = pud_offset(pgd, address);
2109 if (!pud_present(*pud))
2112 pmd = pmd_offset(pud, address);
2113 /* pmd can't go away or become huge under us */
2114 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2117 anon_vma_lock(vma->anon_vma);
2119 pte = pte_offset_map(pmd, address);
2120 ptl = pte_lockptr(mm, pmd);
2122 spin_lock(&mm->page_table_lock); /* probably unnecessary */
2124 * After this gup_fast can't run anymore. This also removes
2125 * any huge TLB entry from the CPU so we won't allow
2126 * huge and small TLB entries for the same virtual address
2127 * to avoid the risk of CPU bugs in that area.
2129 _pmd = pmdp_clear_flush_notify(vma, address, pmd);
2130 spin_unlock(&mm->page_table_lock);
2133 isolated = __collapse_huge_page_isolate(vma, address, pte);
2136 if (unlikely(!isolated)) {
2138 spin_lock(&mm->page_table_lock);
2139 BUG_ON(!pmd_none(*pmd));
2141 * We can only use set_pmd_at when establishing
2142 * hugepmds and never for establishing regular pmds that
2143 * points to regular pagetables. Use pmd_populate for that
2145 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2146 spin_unlock(&mm->page_table_lock);
2147 anon_vma_unlock(vma->anon_vma);
2152 * All pages are isolated and locked so anon_vma rmap
2153 * can't run anymore.
2155 anon_vma_unlock(vma->anon_vma);
2157 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2159 __SetPageUptodate(new_page);
2160 pgtable = pmd_pgtable(_pmd);
2161 VM_BUG_ON(page_count(pgtable) != 1);
2162 VM_BUG_ON(page_mapcount(pgtable) != 0);
2164 _pmd = mk_pmd(new_page, vma->vm_page_prot);
2165 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2166 _pmd = pmd_mkhuge(_pmd);
2169 * spin_lock() below is not the equivalent of smp_wmb(), so
2170 * this is needed to avoid the copy_huge_page writes to become
2171 * visible after the set_pmd_at() write.
2175 spin_lock(&mm->page_table_lock);
2176 BUG_ON(!pmd_none(*pmd));
2177 page_add_new_anon_rmap(new_page, vma, address);
2178 set_pmd_at(mm, address, pmd, _pmd);
2179 update_mmu_cache(vma, address, pmd);
2180 prepare_pmd_huge_pte(pgtable, mm);
2181 spin_unlock(&mm->page_table_lock);
2186 khugepaged_pages_collapsed++;
2188 up_write(&mm->mmap_sem);
2192 mem_cgroup_uncharge_page(new_page);
2199 static int khugepaged_scan_pmd(struct mm_struct *mm,
2200 struct vm_area_struct *vma,
2201 unsigned long address,
2202 struct page **hpage)
2208 int ret = 0, referenced = 0, none = 0;
2210 unsigned long _address;
2214 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2216 pgd = pgd_offset(mm, address);
2217 if (!pgd_present(*pgd))
2220 pud = pud_offset(pgd, address);
2221 if (!pud_present(*pud))
2224 pmd = pmd_offset(pud, address);
2225 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2228 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2229 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2230 _pte++, _address += PAGE_SIZE) {
2231 pte_t pteval = *_pte;
2232 if (pte_none(pteval)) {
2233 if (++none <= khugepaged_max_ptes_none)
2238 if (!pte_present(pteval) || !pte_write(pteval))
2240 page = vm_normal_page(vma, _address, pteval);
2241 if (unlikely(!page))
2244 * Chose the node of the first page. This could
2245 * be more sophisticated and look at more pages,
2246 * but isn't for now.
2249 node = page_to_nid(page);
2250 VM_BUG_ON(PageCompound(page));
2251 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2253 /* cannot use mapcount: can't collapse if there's a gup pin */
2254 if (page_count(page) != 1)
2256 if (pte_young(pteval) || PageReferenced(page) ||
2257 mmu_notifier_test_young(vma->vm_mm, address))
2263 pte_unmap_unlock(pte, ptl);
2265 /* collapse_huge_page will return with the mmap_sem released */
2266 collapse_huge_page(mm, address, hpage, vma, node);
2271 static void collect_mm_slot(struct mm_slot *mm_slot)
2273 struct mm_struct *mm = mm_slot->mm;
2275 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2277 if (khugepaged_test_exit(mm)) {
2279 hlist_del(&mm_slot->hash);
2280 list_del(&mm_slot->mm_node);
2283 * Not strictly needed because the mm exited already.
2285 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2288 /* khugepaged_mm_lock actually not necessary for the below */
2289 free_mm_slot(mm_slot);
2294 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2295 struct page **hpage)
2296 __releases(&khugepaged_mm_lock)
2297 __acquires(&khugepaged_mm_lock)
2299 struct mm_slot *mm_slot;
2300 struct mm_struct *mm;
2301 struct vm_area_struct *vma;
2305 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2307 if (khugepaged_scan.mm_slot)
2308 mm_slot = khugepaged_scan.mm_slot;
2310 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2311 struct mm_slot, mm_node);
2312 khugepaged_scan.address = 0;
2313 khugepaged_scan.mm_slot = mm_slot;
2315 spin_unlock(&khugepaged_mm_lock);
2318 down_read(&mm->mmap_sem);
2319 if (unlikely(khugepaged_test_exit(mm)))
2322 vma = find_vma(mm, khugepaged_scan.address);
2325 for (; vma; vma = vma->vm_next) {
2326 unsigned long hstart, hend;
2329 if (unlikely(khugepaged_test_exit(mm))) {
2333 if (!hugepage_vma_check(vma)) {
2338 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2339 hend = vma->vm_end & HPAGE_PMD_MASK;
2342 if (khugepaged_scan.address > hend)
2344 if (khugepaged_scan.address < hstart)
2345 khugepaged_scan.address = hstart;
2346 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2348 while (khugepaged_scan.address < hend) {
2351 if (unlikely(khugepaged_test_exit(mm)))
2352 goto breakouterloop;
2354 VM_BUG_ON(khugepaged_scan.address < hstart ||
2355 khugepaged_scan.address + HPAGE_PMD_SIZE >
2357 ret = khugepaged_scan_pmd(mm, vma,
2358 khugepaged_scan.address,
2360 /* move to next address */
2361 khugepaged_scan.address += HPAGE_PMD_SIZE;
2362 progress += HPAGE_PMD_NR;
2364 /* we released mmap_sem so break loop */
2365 goto breakouterloop_mmap_sem;
2366 if (progress >= pages)
2367 goto breakouterloop;
2371 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2372 breakouterloop_mmap_sem:
2374 spin_lock(&khugepaged_mm_lock);
2375 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2377 * Release the current mm_slot if this mm is about to die, or
2378 * if we scanned all vmas of this mm.
2380 if (khugepaged_test_exit(mm) || !vma) {
2382 * Make sure that if mm_users is reaching zero while
2383 * khugepaged runs here, khugepaged_exit will find
2384 * mm_slot not pointing to the exiting mm.
2386 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2387 khugepaged_scan.mm_slot = list_entry(
2388 mm_slot->mm_node.next,
2389 struct mm_slot, mm_node);
2390 khugepaged_scan.address = 0;
2392 khugepaged_scan.mm_slot = NULL;
2393 khugepaged_full_scans++;
2396 collect_mm_slot(mm_slot);
2402 static int khugepaged_has_work(void)
2404 return !list_empty(&khugepaged_scan.mm_head) &&
2405 khugepaged_enabled();
2408 static int khugepaged_wait_event(void)
2410 return !list_empty(&khugepaged_scan.mm_head) ||
2411 !khugepaged_enabled();
2414 static void khugepaged_do_scan(struct page **hpage)
2416 unsigned int progress = 0, pass_through_head = 0;
2417 unsigned int pages = khugepaged_pages_to_scan;
2419 barrier(); /* write khugepaged_pages_to_scan to local stack */
2421 while (progress < pages) {
2426 *hpage = alloc_hugepage(khugepaged_defrag());
2427 if (unlikely(!*hpage)) {
2428 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2431 count_vm_event(THP_COLLAPSE_ALLOC);
2438 if (unlikely(kthread_should_stop() || freezing(current)))
2441 spin_lock(&khugepaged_mm_lock);
2442 if (!khugepaged_scan.mm_slot)
2443 pass_through_head++;
2444 if (khugepaged_has_work() &&
2445 pass_through_head < 2)
2446 progress += khugepaged_scan_mm_slot(pages - progress,
2450 spin_unlock(&khugepaged_mm_lock);
2454 static void khugepaged_alloc_sleep(void)
2456 wait_event_freezable_timeout(khugepaged_wait, false,
2457 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2461 static struct page *khugepaged_alloc_hugepage(void)
2466 hpage = alloc_hugepage(khugepaged_defrag());
2468 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2469 khugepaged_alloc_sleep();
2471 count_vm_event(THP_COLLAPSE_ALLOC);
2472 } while (unlikely(!hpage) &&
2473 likely(khugepaged_enabled()));
2478 static void khugepaged_loop(void)
2485 while (likely(khugepaged_enabled())) {
2487 hpage = khugepaged_alloc_hugepage();
2488 if (unlikely(!hpage))
2491 if (IS_ERR(hpage)) {
2492 khugepaged_alloc_sleep();
2497 khugepaged_do_scan(&hpage);
2503 if (unlikely(kthread_should_stop()))
2505 if (khugepaged_has_work()) {
2506 if (!khugepaged_scan_sleep_millisecs)
2508 wait_event_freezable_timeout(khugepaged_wait, false,
2509 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2510 } else if (khugepaged_enabled())
2511 wait_event_freezable(khugepaged_wait,
2512 khugepaged_wait_event());
2516 static int khugepaged(void *none)
2518 struct mm_slot *mm_slot;
2521 set_user_nice(current, 19);
2523 /* serialize with start_khugepaged() */
2524 mutex_lock(&khugepaged_mutex);
2527 mutex_unlock(&khugepaged_mutex);
2528 VM_BUG_ON(khugepaged_thread != current);
2530 VM_BUG_ON(khugepaged_thread != current);
2532 mutex_lock(&khugepaged_mutex);
2533 if (!khugepaged_enabled())
2535 if (unlikely(kthread_should_stop()))
2539 spin_lock(&khugepaged_mm_lock);
2540 mm_slot = khugepaged_scan.mm_slot;
2541 khugepaged_scan.mm_slot = NULL;
2543 collect_mm_slot(mm_slot);
2544 spin_unlock(&khugepaged_mm_lock);
2546 khugepaged_thread = NULL;
2547 mutex_unlock(&khugepaged_mutex);
2552 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2554 struct vm_area_struct *vma;
2557 spin_lock(&mm->page_table_lock);
2558 if (unlikely(!pmd_trans_huge(*pmd))) {
2559 spin_unlock(&mm->page_table_lock);
2562 vma = find_vma(mm, pmd_to_va(mm, pmd));
2563 if (vma && is_fb_vma(vma)) {
2564 split_fb_pmd(vma, pmd);
2565 spin_unlock(&mm->page_table_lock);
2568 page = pmd_page(*pmd);
2569 VM_BUG_ON(!page_count(page));
2571 spin_unlock(&mm->page_table_lock);
2573 split_huge_page(page);
2576 BUG_ON(pmd_trans_huge(*pmd));
2579 static void split_huge_page_address(struct mm_struct *mm,
2580 unsigned long address)
2586 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2588 pgd = pgd_offset(mm, address);
2589 if (!pgd_present(*pgd))
2592 pud = pud_offset(pgd, address);
2593 if (!pud_present(*pud))
2596 pmd = pmd_offset(pud, address);
2597 if (!pmd_present(*pmd))
2600 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2601 * materialize from under us.
2603 split_huge_page_pmd(mm, pmd);
2606 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2607 unsigned long start,
2612 * If the new start address isn't hpage aligned and it could
2613 * previously contain an hugepage: check if we need to split
2616 if (start & ~HPAGE_PMD_MASK &&
2617 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2618 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2619 split_huge_page_address(vma->vm_mm, start);
2622 * If the new end address isn't hpage aligned and it could
2623 * previously contain an hugepage: check if we need to split
2626 if (end & ~HPAGE_PMD_MASK &&
2627 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2628 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2629 split_huge_page_address(vma->vm_mm, end);
2632 * If we're also updating the vma->vm_next->vm_start, if the new
2633 * vm_next->vm_start isn't page aligned and it could previously
2634 * contain an hugepage: check if we need to split an huge pmd.
2636 if (adjust_next > 0) {
2637 struct vm_area_struct *next = vma->vm_next;
2638 unsigned long nstart = next->vm_start;
2639 nstart += adjust_next << PAGE_SHIFT;
2640 if (nstart & ~HPAGE_PMD_MASK &&
2641 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2642 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2643 split_huge_page_address(next->vm_mm, nstart);