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>
20 #include <linux/pagemap.h>
22 #include <asm/pgalloc.h>
26 * By default transparent hugepage support is enabled for all mappings
27 * and khugepaged scans all mappings. Defrag is only invoked by
28 * khugepaged hugepage allocations and by page faults inside
29 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
32 unsigned long transparent_hugepage_flags __read_mostly =
33 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
34 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
36 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
37 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
39 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
40 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
42 /* default scan 8*512 pte (or vmas) every 30 second */
43 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
44 static unsigned int khugepaged_pages_collapsed;
45 static unsigned int khugepaged_full_scans;
46 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
47 /* during fragmentation poll the hugepage allocator once every minute */
48 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
49 static struct task_struct *khugepaged_thread __read_mostly;
50 static unsigned long huge_zero_pfn __read_mostly;
51 static DEFINE_MUTEX(khugepaged_mutex);
52 static DEFINE_SPINLOCK(khugepaged_mm_lock);
53 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
55 * default collapse hugepages if there is at least one pte mapped like
56 * it would have happened if the vma was large enough during page
59 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
61 static int khugepaged(void *none);
62 static int mm_slots_hash_init(void);
63 static int khugepaged_slab_init(void);
64 static void khugepaged_slab_free(void);
66 #define MM_SLOTS_HASH_HEADS 1024
67 static struct hlist_head *mm_slots_hash __read_mostly;
68 static struct kmem_cache *mm_slot_cache __read_mostly;
71 * struct mm_slot - hash lookup from mm to mm_slot
72 * @hash: hash collision list
73 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
74 * @mm: the mm that this information is valid for
77 struct hlist_node hash;
78 struct list_head mm_node;
83 * struct khugepaged_scan - cursor for scanning
84 * @mm_head: the head of the mm list to scan
85 * @mm_slot: the current mm_slot we are scanning
86 * @address: the next address inside that to be scanned
88 * There is only the one khugepaged_scan instance of this cursor structure.
90 struct khugepaged_scan {
91 struct list_head mm_head;
92 struct mm_slot *mm_slot;
93 unsigned long address;
95 static struct khugepaged_scan khugepaged_scan = {
96 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
100 static int set_recommended_min_free_kbytes(void)
104 unsigned long recommended_min;
105 extern int min_free_kbytes;
107 if (!khugepaged_enabled())
110 for_each_populated_zone(zone)
113 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
114 recommended_min = pageblock_nr_pages * nr_zones * 2;
117 * Make sure that on average at least two pageblocks are almost free
118 * of another type, one for a migratetype to fall back to and a
119 * second to avoid subsequent fallbacks of other types There are 3
120 * MIGRATE_TYPES we care about.
122 recommended_min += pageblock_nr_pages * nr_zones *
123 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
125 /* don't ever allow to reserve more than 5% of the lowmem */
126 recommended_min = min(recommended_min,
127 (unsigned long) nr_free_buffer_pages() / 20);
128 recommended_min <<= (PAGE_SHIFT-10);
130 if (recommended_min > min_free_kbytes)
131 min_free_kbytes = recommended_min;
132 setup_per_zone_wmarks();
135 late_initcall(set_recommended_min_free_kbytes);
137 static int start_khugepaged(void)
140 if (khugepaged_enabled()) {
141 if (!khugepaged_thread)
142 khugepaged_thread = kthread_run(khugepaged, NULL,
144 if (unlikely(IS_ERR(khugepaged_thread))) {
146 "khugepaged: kthread_run(khugepaged) failed\n");
147 err = PTR_ERR(khugepaged_thread);
148 khugepaged_thread = NULL;
151 if (!list_empty(&khugepaged_scan.mm_head))
152 wake_up_interruptible(&khugepaged_wait);
154 set_recommended_min_free_kbytes();
155 } else if (khugepaged_thread) {
156 kthread_stop(khugepaged_thread);
157 khugepaged_thread = NULL;
163 static int init_huge_zero_pfn(void)
168 hpage = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
172 pfn = page_to_pfn(hpage);
173 if (cmpxchg(&huge_zero_pfn, 0, pfn))
178 static inline bool is_huge_zero_pfn(unsigned long pfn)
180 return huge_zero_pfn && pfn == huge_zero_pfn;
183 static inline bool is_huge_zero_pmd(pmd_t pmd)
185 return is_huge_zero_pfn(pmd_pfn(pmd));
190 static ssize_t double_flag_show(struct kobject *kobj,
191 struct kobj_attribute *attr, char *buf,
192 enum transparent_hugepage_flag enabled,
193 enum transparent_hugepage_flag req_madv)
195 if (test_bit(enabled, &transparent_hugepage_flags)) {
196 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
197 return sprintf(buf, "[always] madvise never\n");
198 } else if (test_bit(req_madv, &transparent_hugepage_flags))
199 return sprintf(buf, "always [madvise] never\n");
201 return sprintf(buf, "always madvise [never]\n");
203 static ssize_t double_flag_store(struct kobject *kobj,
204 struct kobj_attribute *attr,
205 const char *buf, size_t count,
206 enum transparent_hugepage_flag enabled,
207 enum transparent_hugepage_flag req_madv)
209 if (!memcmp("always", buf,
210 min(sizeof("always")-1, count))) {
211 set_bit(enabled, &transparent_hugepage_flags);
212 clear_bit(req_madv, &transparent_hugepage_flags);
213 } else if (!memcmp("madvise", buf,
214 min(sizeof("madvise")-1, count))) {
215 clear_bit(enabled, &transparent_hugepage_flags);
216 set_bit(req_madv, &transparent_hugepage_flags);
217 } else if (!memcmp("never", buf,
218 min(sizeof("never")-1, count))) {
219 clear_bit(enabled, &transparent_hugepage_flags);
220 clear_bit(req_madv, &transparent_hugepage_flags);
227 static ssize_t enabled_show(struct kobject *kobj,
228 struct kobj_attribute *attr, char *buf)
230 return double_flag_show(kobj, attr, buf,
231 TRANSPARENT_HUGEPAGE_FLAG,
232 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
234 static ssize_t enabled_store(struct kobject *kobj,
235 struct kobj_attribute *attr,
236 const char *buf, size_t count)
240 ret = double_flag_store(kobj, attr, buf, count,
241 TRANSPARENT_HUGEPAGE_FLAG,
242 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
247 mutex_lock(&khugepaged_mutex);
248 err = start_khugepaged();
249 mutex_unlock(&khugepaged_mutex);
257 static struct kobj_attribute enabled_attr =
258 __ATTR(enabled, 0644, enabled_show, enabled_store);
260 static ssize_t single_flag_show(struct kobject *kobj,
261 struct kobj_attribute *attr, char *buf,
262 enum transparent_hugepage_flag flag)
264 return sprintf(buf, "%d\n",
265 !!test_bit(flag, &transparent_hugepage_flags));
268 static ssize_t single_flag_store(struct kobject *kobj,
269 struct kobj_attribute *attr,
270 const char *buf, size_t count,
271 enum transparent_hugepage_flag flag)
276 ret = kstrtoul(buf, 10, &value);
283 set_bit(flag, &transparent_hugepage_flags);
285 clear_bit(flag, &transparent_hugepage_flags);
291 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
292 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
293 * memory just to allocate one more hugepage.
295 static ssize_t defrag_show(struct kobject *kobj,
296 struct kobj_attribute *attr, char *buf)
298 return double_flag_show(kobj, attr, buf,
299 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
300 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
302 static ssize_t defrag_store(struct kobject *kobj,
303 struct kobj_attribute *attr,
304 const char *buf, size_t count)
306 return double_flag_store(kobj, attr, buf, count,
307 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
308 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
310 static struct kobj_attribute defrag_attr =
311 __ATTR(defrag, 0644, defrag_show, defrag_store);
313 #ifdef CONFIG_DEBUG_VM
314 static ssize_t debug_cow_show(struct kobject *kobj,
315 struct kobj_attribute *attr, char *buf)
317 return single_flag_show(kobj, attr, buf,
318 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
320 static ssize_t debug_cow_store(struct kobject *kobj,
321 struct kobj_attribute *attr,
322 const char *buf, size_t count)
324 return single_flag_store(kobj, attr, buf, count,
325 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
327 static struct kobj_attribute debug_cow_attr =
328 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
329 #endif /* CONFIG_DEBUG_VM */
331 static struct attribute *hugepage_attr[] = {
334 #ifdef CONFIG_DEBUG_VM
335 &debug_cow_attr.attr,
340 static struct attribute_group hugepage_attr_group = {
341 .attrs = hugepage_attr,
344 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
345 struct kobj_attribute *attr,
348 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
351 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
352 struct kobj_attribute *attr,
353 const char *buf, size_t count)
358 err = strict_strtoul(buf, 10, &msecs);
359 if (err || msecs > UINT_MAX)
362 khugepaged_scan_sleep_millisecs = msecs;
363 wake_up_interruptible(&khugepaged_wait);
367 static struct kobj_attribute scan_sleep_millisecs_attr =
368 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
369 scan_sleep_millisecs_store);
371 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
372 struct kobj_attribute *attr,
375 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
378 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
379 struct kobj_attribute *attr,
380 const char *buf, size_t count)
385 err = strict_strtoul(buf, 10, &msecs);
386 if (err || msecs > UINT_MAX)
389 khugepaged_alloc_sleep_millisecs = msecs;
390 wake_up_interruptible(&khugepaged_wait);
394 static struct kobj_attribute alloc_sleep_millisecs_attr =
395 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
396 alloc_sleep_millisecs_store);
398 static ssize_t pages_to_scan_show(struct kobject *kobj,
399 struct kobj_attribute *attr,
402 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
404 static ssize_t pages_to_scan_store(struct kobject *kobj,
405 struct kobj_attribute *attr,
406 const char *buf, size_t count)
411 err = strict_strtoul(buf, 10, &pages);
412 if (err || !pages || pages > UINT_MAX)
415 khugepaged_pages_to_scan = pages;
419 static struct kobj_attribute pages_to_scan_attr =
420 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
421 pages_to_scan_store);
423 static ssize_t pages_collapsed_show(struct kobject *kobj,
424 struct kobj_attribute *attr,
427 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
429 static struct kobj_attribute pages_collapsed_attr =
430 __ATTR_RO(pages_collapsed);
432 static ssize_t full_scans_show(struct kobject *kobj,
433 struct kobj_attribute *attr,
436 return sprintf(buf, "%u\n", khugepaged_full_scans);
438 static struct kobj_attribute full_scans_attr =
439 __ATTR_RO(full_scans);
441 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
442 struct kobj_attribute *attr, char *buf)
444 return single_flag_show(kobj, attr, buf,
445 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
447 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
448 struct kobj_attribute *attr,
449 const char *buf, size_t count)
451 return single_flag_store(kobj, attr, buf, count,
452 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
454 static struct kobj_attribute khugepaged_defrag_attr =
455 __ATTR(defrag, 0644, khugepaged_defrag_show,
456 khugepaged_defrag_store);
459 * max_ptes_none controls if khugepaged should collapse hugepages over
460 * any unmapped ptes in turn potentially increasing the memory
461 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
462 * reduce the available free memory in the system as it
463 * runs. Increasing max_ptes_none will instead potentially reduce the
464 * free memory in the system during the khugepaged scan.
466 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
467 struct kobj_attribute *attr,
470 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
472 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
473 struct kobj_attribute *attr,
474 const char *buf, size_t count)
477 unsigned long max_ptes_none;
479 err = strict_strtoul(buf, 10, &max_ptes_none);
480 if (err || max_ptes_none > HPAGE_PMD_NR-1)
483 khugepaged_max_ptes_none = max_ptes_none;
487 static struct kobj_attribute khugepaged_max_ptes_none_attr =
488 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
489 khugepaged_max_ptes_none_store);
491 static struct attribute *khugepaged_attr[] = {
492 &khugepaged_defrag_attr.attr,
493 &khugepaged_max_ptes_none_attr.attr,
494 &pages_to_scan_attr.attr,
495 &pages_collapsed_attr.attr,
496 &full_scans_attr.attr,
497 &scan_sleep_millisecs_attr.attr,
498 &alloc_sleep_millisecs_attr.attr,
502 static struct attribute_group khugepaged_attr_group = {
503 .attrs = khugepaged_attr,
504 .name = "khugepaged",
507 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
511 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
512 if (unlikely(!*hugepage_kobj)) {
513 printk(KERN_ERR "hugepage: failed kobject create\n");
517 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
519 printk(KERN_ERR "hugepage: failed register hugeage group\n");
523 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
525 printk(KERN_ERR "hugepage: failed register hugeage group\n");
526 goto remove_hp_group;
532 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
534 kobject_put(*hugepage_kobj);
538 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
540 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
541 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
542 kobject_put(hugepage_kobj);
545 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
550 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
553 #endif /* CONFIG_SYSFS */
555 static int __init hugepage_init(void)
558 struct kobject *hugepage_kobj;
560 if (!has_transparent_hugepage()) {
561 transparent_hugepage_flags = 0;
565 err = hugepage_init_sysfs(&hugepage_kobj);
569 err = khugepaged_slab_init();
573 err = mm_slots_hash_init();
575 khugepaged_slab_free();
580 * By default disable transparent hugepages on smaller systems,
581 * where the extra memory used could hurt more than TLB overhead
582 * is likely to save. The admin can still enable it through /sys.
584 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
585 transparent_hugepage_flags = 0;
591 hugepage_exit_sysfs(hugepage_kobj);
594 module_init(hugepage_init)
596 static int __init setup_transparent_hugepage(char *str)
601 if (!strcmp(str, "always")) {
602 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
603 &transparent_hugepage_flags);
604 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
605 &transparent_hugepage_flags);
607 } else if (!strcmp(str, "madvise")) {
608 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
609 &transparent_hugepage_flags);
610 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
611 &transparent_hugepage_flags);
613 } else if (!strcmp(str, "never")) {
614 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
615 &transparent_hugepage_flags);
616 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
617 &transparent_hugepage_flags);
623 "transparent_hugepage= cannot parse, ignored\n");
626 __setup("transparent_hugepage=", setup_transparent_hugepage);
628 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
630 if (likely(vma->vm_flags & VM_WRITE))
631 pmd = pmd_mkwrite(pmd);
635 static inline pmd_t mk_huge_pmd(struct page *page, struct vm_area_struct *vma)
638 entry = mk_pmd(page, vma->vm_page_prot);
639 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
640 entry = pmd_mkhuge(entry);
644 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
645 struct vm_area_struct *vma,
646 unsigned long haddr, pmd_t *pmd,
651 VM_BUG_ON(!PageCompound(page));
652 pgtable = pte_alloc_one(mm, haddr);
653 if (unlikely(!pgtable))
656 clear_huge_page(page, haddr, HPAGE_PMD_NR);
657 __SetPageUptodate(page);
659 spin_lock(&mm->page_table_lock);
660 if (unlikely(!pmd_none(*pmd))) {
661 spin_unlock(&mm->page_table_lock);
662 mem_cgroup_uncharge_page(page);
664 pte_free(mm, pgtable);
667 entry = mk_huge_pmd(page, vma);
669 * The spinlocking to take the lru_lock inside
670 * page_add_new_anon_rmap() acts as a full memory
671 * barrier to be sure clear_huge_page writes become
672 * visible after the set_pmd_at() write.
674 page_add_new_anon_rmap(page, vma, haddr);
675 set_pmd_at(mm, haddr, pmd, entry);
676 pgtable_trans_huge_deposit(mm, pgtable);
677 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
679 spin_unlock(&mm->page_table_lock);
685 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
687 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
690 static inline struct page *alloc_hugepage_vma(int defrag,
691 struct vm_area_struct *vma,
692 unsigned long haddr, int nd,
695 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
696 HPAGE_PMD_ORDER, vma, haddr, nd);
700 static inline struct page *alloc_hugepage(int defrag)
702 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
707 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
708 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd)
711 entry = pfn_pmd(huge_zero_pfn, vma->vm_page_prot);
712 entry = pmd_wrprotect(entry);
713 entry = pmd_mkhuge(entry);
714 set_pmd_at(mm, haddr, pmd, entry);
715 pgtable_trans_huge_deposit(mm, pgtable);
719 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
720 unsigned long address, pmd_t *pmd,
724 unsigned long haddr = address & HPAGE_PMD_MASK;
727 if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
728 if (unlikely(anon_vma_prepare(vma)))
730 if (unlikely(khugepaged_enter(vma)))
732 if (!(flags & FAULT_FLAG_WRITE)) {
734 if (unlikely(!huge_zero_pfn && init_huge_zero_pfn())) {
735 count_vm_event(THP_FAULT_FALLBACK);
738 pgtable = pte_alloc_one(mm, haddr);
739 if (unlikely(!pgtable))
741 spin_lock(&mm->page_table_lock);
742 set_huge_zero_page(pgtable, mm, vma, haddr, pmd);
743 spin_unlock(&mm->page_table_lock);
746 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
747 vma, haddr, numa_node_id(), 0);
748 if (unlikely(!page)) {
749 count_vm_event(THP_FAULT_FALLBACK);
752 count_vm_event(THP_FAULT_ALLOC);
753 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
757 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
759 mem_cgroup_uncharge_page(page);
768 * Use __pte_alloc instead of pte_alloc_map, because we can't
769 * run pte_offset_map on the pmd, if an huge pmd could
770 * materialize from under us from a different thread.
772 if (unlikely(__pte_alloc(mm, vma, pmd, address)))
774 /* if an huge pmd materialized from under us just retry later */
775 if (unlikely(pmd_trans_huge(*pmd)))
778 * A regular pmd is established and it can't morph into a huge pmd
779 * from under us anymore at this point because we hold the mmap_sem
780 * read mode and khugepaged takes it in write mode. So now it's
781 * safe to run pte_offset_map().
783 pte = pte_offset_map(pmd, address);
784 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
787 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
788 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
789 struct vm_area_struct *vma)
791 struct page *src_page;
797 pgtable = pte_alloc_one(dst_mm, addr);
798 if (unlikely(!pgtable))
801 spin_lock(&dst_mm->page_table_lock);
802 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
806 if (unlikely(!pmd_trans_huge(pmd))) {
807 pte_free(dst_mm, pgtable);
811 * mm->page_table_lock is enough to be sure that huge zero pmd is not
812 * under splitting since we don't split the page itself, only pmd to
815 if (is_huge_zero_pmd(pmd)) {
816 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd);
820 if (unlikely(pmd_trans_splitting(pmd))) {
821 /* split huge page running from under us */
822 spin_unlock(&src_mm->page_table_lock);
823 spin_unlock(&dst_mm->page_table_lock);
824 pte_free(dst_mm, pgtable);
826 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
829 src_page = pmd_page(pmd);
830 VM_BUG_ON(!PageHead(src_page));
832 page_dup_rmap(src_page);
833 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
835 pmdp_set_wrprotect(src_mm, addr, src_pmd);
836 pmd = pmd_mkold(pmd_wrprotect(pmd));
837 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
838 pgtable_trans_huge_deposit(dst_mm, pgtable);
843 spin_unlock(&src_mm->page_table_lock);
844 spin_unlock(&dst_mm->page_table_lock);
849 void huge_pmd_set_accessed(struct mm_struct *mm,
850 struct vm_area_struct *vma,
851 unsigned long address,
852 pmd_t *pmd, pmd_t orig_pmd,
858 spin_lock(&mm->page_table_lock);
859 if (unlikely(!pmd_same(*pmd, orig_pmd)))
862 entry = pmd_mkyoung(orig_pmd);
863 haddr = address & HPAGE_PMD_MASK;
864 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
865 update_mmu_cache_pmd(vma, address, pmd);
868 spin_unlock(&mm->page_table_lock);
871 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
872 struct vm_area_struct *vma, unsigned long address,
873 pmd_t *pmd, unsigned long haddr)
879 unsigned long mmun_start; /* For mmu_notifiers */
880 unsigned long mmun_end; /* For mmu_notifiers */
882 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
888 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
894 clear_user_highpage(page, address);
895 __SetPageUptodate(page);
898 mmun_end = haddr + HPAGE_PMD_SIZE;
899 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
901 spin_lock(&mm->page_table_lock);
902 pmdp_clear_flush(vma, haddr, pmd);
903 /* leave pmd empty until pte is filled */
905 pgtable = pgtable_trans_huge_withdraw(mm);
906 pmd_populate(mm, &_pmd, pgtable);
908 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
910 if (haddr == (address & PAGE_MASK)) {
911 entry = mk_pte(page, vma->vm_page_prot);
912 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
913 page_add_new_anon_rmap(page, vma, haddr);
915 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
916 entry = pte_mkspecial(entry);
918 pte = pte_offset_map(&_pmd, haddr);
919 VM_BUG_ON(!pte_none(*pte));
920 set_pte_at(mm, haddr, pte, entry);
923 smp_wmb(); /* make pte visible before pmd */
924 pmd_populate(mm, pmd, pgtable);
925 spin_unlock(&mm->page_table_lock);
926 inc_mm_counter(mm, MM_ANONPAGES);
928 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
930 ret |= VM_FAULT_WRITE;
935 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
936 struct vm_area_struct *vma,
937 unsigned long address,
938 pmd_t *pmd, pmd_t orig_pmd,
946 unsigned long mmun_start; /* For mmu_notifiers */
947 unsigned long mmun_end; /* For mmu_notifiers */
949 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
951 if (unlikely(!pages)) {
956 for (i = 0; i < HPAGE_PMD_NR; i++) {
957 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
959 vma, address, page_to_nid(page));
960 if (unlikely(!pages[i] ||
961 mem_cgroup_newpage_charge(pages[i], mm,
965 mem_cgroup_uncharge_start();
967 mem_cgroup_uncharge_page(pages[i]);
970 mem_cgroup_uncharge_end();
977 for (i = 0; i < HPAGE_PMD_NR; i++) {
978 copy_user_highpage(pages[i], page + i,
979 haddr + PAGE_SIZE * i, vma);
980 __SetPageUptodate(pages[i]);
985 mmun_end = haddr + HPAGE_PMD_SIZE;
986 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
988 spin_lock(&mm->page_table_lock);
989 if (unlikely(!pmd_same(*pmd, orig_pmd)))
991 VM_BUG_ON(!PageHead(page));
993 pmdp_clear_flush(vma, haddr, pmd);
994 /* leave pmd empty until pte is filled */
996 pgtable = pgtable_trans_huge_withdraw(mm);
997 pmd_populate(mm, &_pmd, pgtable);
999 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1001 entry = mk_pte(pages[i], vma->vm_page_prot);
1002 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1003 page_add_new_anon_rmap(pages[i], vma, haddr);
1004 pte = pte_offset_map(&_pmd, haddr);
1005 VM_BUG_ON(!pte_none(*pte));
1006 set_pte_at(mm, haddr, pte, entry);
1011 smp_wmb(); /* make pte visible before pmd */
1012 pmd_populate(mm, pmd, pgtable);
1013 page_remove_rmap(page);
1014 spin_unlock(&mm->page_table_lock);
1016 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1018 ret |= VM_FAULT_WRITE;
1025 spin_unlock(&mm->page_table_lock);
1026 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1027 mem_cgroup_uncharge_start();
1028 for (i = 0; i < HPAGE_PMD_NR; i++) {
1029 mem_cgroup_uncharge_page(pages[i]);
1032 mem_cgroup_uncharge_end();
1037 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1038 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1041 struct page *page = NULL, *new_page;
1042 unsigned long haddr;
1043 unsigned long mmun_start; /* For mmu_notifiers */
1044 unsigned long mmun_end; /* For mmu_notifiers */
1046 VM_BUG_ON(!vma->anon_vma);
1047 haddr = address & HPAGE_PMD_MASK;
1048 if (is_huge_zero_pmd(orig_pmd))
1050 spin_lock(&mm->page_table_lock);
1051 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1054 page = pmd_page(orig_pmd);
1055 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1056 if (page_mapcount(page) == 1) {
1058 entry = pmd_mkyoung(orig_pmd);
1059 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1060 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
1061 update_mmu_cache_pmd(vma, address, pmd);
1062 ret |= VM_FAULT_WRITE;
1066 spin_unlock(&mm->page_table_lock);
1068 if (transparent_hugepage_enabled(vma) &&
1069 !transparent_hugepage_debug_cow())
1070 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1071 vma, haddr, numa_node_id(), 0);
1075 if (unlikely(!new_page)) {
1076 count_vm_event(THP_FAULT_FALLBACK);
1077 if (is_huge_zero_pmd(orig_pmd)) {
1078 ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1079 address, pmd, haddr);
1081 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1082 pmd, orig_pmd, page, haddr);
1083 if (ret & VM_FAULT_OOM)
1084 split_huge_page(page);
1089 count_vm_event(THP_FAULT_ALLOC);
1091 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1094 split_huge_page(page);
1097 ret |= VM_FAULT_OOM;
1101 if (is_huge_zero_pmd(orig_pmd))
1102 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1104 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1105 __SetPageUptodate(new_page);
1108 mmun_end = haddr + HPAGE_PMD_SIZE;
1109 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1111 spin_lock(&mm->page_table_lock);
1114 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1115 spin_unlock(&mm->page_table_lock);
1116 mem_cgroup_uncharge_page(new_page);
1121 entry = mk_huge_pmd(new_page, vma);
1122 pmdp_clear_flush(vma, haddr, pmd);
1123 page_add_new_anon_rmap(new_page, vma, haddr);
1124 set_pmd_at(mm, haddr, pmd, entry);
1125 update_mmu_cache_pmd(vma, address, pmd);
1126 if (is_huge_zero_pmd(orig_pmd))
1127 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1129 VM_BUG_ON(!PageHead(page));
1130 page_remove_rmap(page);
1133 ret |= VM_FAULT_WRITE;
1135 spin_unlock(&mm->page_table_lock);
1137 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1141 spin_unlock(&mm->page_table_lock);
1145 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1150 struct mm_struct *mm = vma->vm_mm;
1151 struct page *page = NULL;
1153 assert_spin_locked(&mm->page_table_lock);
1155 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1158 page = pmd_page(*pmd);
1159 VM_BUG_ON(!PageHead(page));
1160 if (flags & FOLL_TOUCH) {
1163 * We should set the dirty bit only for FOLL_WRITE but
1164 * for now the dirty bit in the pmd is meaningless.
1165 * And if the dirty bit will become meaningful and
1166 * we'll only set it with FOLL_WRITE, an atomic
1167 * set_bit will be required on the pmd to set the
1168 * young bit, instead of the current set_pmd_at.
1170 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1171 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1173 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1174 if (page->mapping && trylock_page(page)) {
1177 mlock_vma_page(page);
1181 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1182 VM_BUG_ON(!PageCompound(page));
1183 if (flags & FOLL_GET)
1184 get_page_foll(page);
1190 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1191 pmd_t *pmd, unsigned long addr)
1195 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1199 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1200 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1201 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1202 if (is_huge_zero_pmd(orig_pmd)) {
1204 spin_unlock(&tlb->mm->page_table_lock);
1206 page = pmd_page(orig_pmd);
1207 page_remove_rmap(page);
1208 VM_BUG_ON(page_mapcount(page) < 0);
1209 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1210 VM_BUG_ON(!PageHead(page));
1212 spin_unlock(&tlb->mm->page_table_lock);
1213 tlb_remove_page(tlb, page);
1215 pte_free(tlb->mm, pgtable);
1221 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1222 unsigned long addr, unsigned long end,
1227 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1229 * All logical pages in the range are present
1230 * if backed by a huge page.
1232 spin_unlock(&vma->vm_mm->page_table_lock);
1233 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1240 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1241 unsigned long old_addr,
1242 unsigned long new_addr, unsigned long old_end,
1243 pmd_t *old_pmd, pmd_t *new_pmd)
1248 struct mm_struct *mm = vma->vm_mm;
1250 if ((old_addr & ~HPAGE_PMD_MASK) ||
1251 (new_addr & ~HPAGE_PMD_MASK) ||
1252 old_end - old_addr < HPAGE_PMD_SIZE ||
1253 (new_vma->vm_flags & VM_NOHUGEPAGE))
1257 * The destination pmd shouldn't be established, free_pgtables()
1258 * should have release it.
1260 if (WARN_ON(!pmd_none(*new_pmd))) {
1261 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1265 ret = __pmd_trans_huge_lock(old_pmd, vma);
1267 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1268 VM_BUG_ON(!pmd_none(*new_pmd));
1269 set_pmd_at(mm, new_addr, new_pmd, pmd);
1270 spin_unlock(&mm->page_table_lock);
1276 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1277 unsigned long addr, pgprot_t newprot)
1279 struct mm_struct *mm = vma->vm_mm;
1282 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1284 entry = pmdp_get_and_clear(mm, addr, pmd);
1285 entry = pmd_modify(entry, newprot);
1286 BUG_ON(pmd_write(entry));
1287 set_pmd_at(mm, addr, pmd, entry);
1288 spin_unlock(&vma->vm_mm->page_table_lock);
1296 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1297 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1299 * Note that if it returns 1, this routine returns without unlocking page
1300 * table locks. So callers must unlock them.
1302 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1304 spin_lock(&vma->vm_mm->page_table_lock);
1305 if (likely(pmd_trans_huge(*pmd))) {
1306 if (unlikely(pmd_trans_splitting(*pmd))) {
1307 spin_unlock(&vma->vm_mm->page_table_lock);
1308 wait_split_huge_page(vma->anon_vma, pmd);
1311 /* Thp mapped by 'pmd' is stable, so we can
1312 * handle it as it is. */
1316 spin_unlock(&vma->vm_mm->page_table_lock);
1320 pmd_t *page_check_address_pmd(struct page *page,
1321 struct mm_struct *mm,
1322 unsigned long address,
1323 enum page_check_address_pmd_flag flag)
1325 pmd_t *pmd, *ret = NULL;
1327 if (address & ~HPAGE_PMD_MASK)
1330 pmd = mm_find_pmd(mm, address);
1335 if (pmd_page(*pmd) != page)
1338 * split_vma() may create temporary aliased mappings. There is
1339 * no risk as long as all huge pmd are found and have their
1340 * splitting bit set before __split_huge_page_refcount
1341 * runs. Finding the same huge pmd more than once during the
1342 * same rmap walk is not a problem.
1344 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1345 pmd_trans_splitting(*pmd))
1347 if (pmd_trans_huge(*pmd)) {
1348 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1349 !pmd_trans_splitting(*pmd));
1356 static int __split_huge_page_splitting(struct page *page,
1357 struct vm_area_struct *vma,
1358 unsigned long address)
1360 struct mm_struct *mm = vma->vm_mm;
1363 /* For mmu_notifiers */
1364 const unsigned long mmun_start = address;
1365 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
1367 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1368 spin_lock(&mm->page_table_lock);
1369 pmd = page_check_address_pmd(page, mm, address,
1370 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1373 * We can't temporarily set the pmd to null in order
1374 * to split it, the pmd must remain marked huge at all
1375 * times or the VM won't take the pmd_trans_huge paths
1376 * and it won't wait on the anon_vma->root->mutex to
1377 * serialize against split_huge_page*.
1379 pmdp_splitting_flush(vma, address, pmd);
1382 spin_unlock(&mm->page_table_lock);
1383 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1388 static void __split_huge_page_refcount(struct page *page)
1391 struct zone *zone = page_zone(page);
1392 struct lruvec *lruvec;
1395 /* prevent PageLRU to go away from under us, and freeze lru stats */
1396 spin_lock_irq(&zone->lru_lock);
1397 lruvec = mem_cgroup_page_lruvec(page, zone);
1399 compound_lock(page);
1400 /* complete memcg works before add pages to LRU */
1401 mem_cgroup_split_huge_fixup(page);
1403 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1404 struct page *page_tail = page + i;
1406 /* tail_page->_mapcount cannot change */
1407 BUG_ON(page_mapcount(page_tail) < 0);
1408 tail_count += page_mapcount(page_tail);
1409 /* check for overflow */
1410 BUG_ON(tail_count < 0);
1411 BUG_ON(atomic_read(&page_tail->_count) != 0);
1413 * tail_page->_count is zero and not changing from
1414 * under us. But get_page_unless_zero() may be running
1415 * from under us on the tail_page. If we used
1416 * atomic_set() below instead of atomic_add(), we
1417 * would then run atomic_set() concurrently with
1418 * get_page_unless_zero(), and atomic_set() is
1419 * implemented in C not using locked ops. spin_unlock
1420 * on x86 sometime uses locked ops because of PPro
1421 * errata 66, 92, so unless somebody can guarantee
1422 * atomic_set() here would be safe on all archs (and
1423 * not only on x86), it's safer to use atomic_add().
1425 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1426 &page_tail->_count);
1428 /* after clearing PageTail the gup refcount can be released */
1432 * retain hwpoison flag of the poisoned tail page:
1433 * fix for the unsuitable process killed on Guest Machine(KVM)
1434 * by the memory-failure.
1436 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1437 page_tail->flags |= (page->flags &
1438 ((1L << PG_referenced) |
1439 (1L << PG_swapbacked) |
1440 (1L << PG_mlocked) |
1441 (1L << PG_uptodate)));
1442 page_tail->flags |= (1L << PG_dirty);
1444 /* clear PageTail before overwriting first_page */
1448 * __split_huge_page_splitting() already set the
1449 * splitting bit in all pmd that could map this
1450 * hugepage, that will ensure no CPU can alter the
1451 * mapcount on the head page. The mapcount is only
1452 * accounted in the head page and it has to be
1453 * transferred to all tail pages in the below code. So
1454 * for this code to be safe, the split the mapcount
1455 * can't change. But that doesn't mean userland can't
1456 * keep changing and reading the page contents while
1457 * we transfer the mapcount, so the pmd splitting
1458 * status is achieved setting a reserved bit in the
1459 * pmd, not by clearing the present bit.
1461 page_tail->_mapcount = page->_mapcount;
1463 BUG_ON(page_tail->mapping);
1464 page_tail->mapping = page->mapping;
1466 page_tail->index = page->index + i;
1468 BUG_ON(!PageAnon(page_tail));
1469 BUG_ON(!PageUptodate(page_tail));
1470 BUG_ON(!PageDirty(page_tail));
1471 BUG_ON(!PageSwapBacked(page_tail));
1473 lru_add_page_tail(page, page_tail, lruvec);
1475 atomic_sub(tail_count, &page->_count);
1476 BUG_ON(atomic_read(&page->_count) <= 0);
1478 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1479 __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1481 ClearPageCompound(page);
1482 compound_unlock(page);
1483 spin_unlock_irq(&zone->lru_lock);
1485 for (i = 1; i < HPAGE_PMD_NR; i++) {
1486 struct page *page_tail = page + i;
1487 BUG_ON(page_count(page_tail) <= 0);
1489 * Tail pages may be freed if there wasn't any mapping
1490 * like if add_to_swap() is running on a lru page that
1491 * had its mapping zapped. And freeing these pages
1492 * requires taking the lru_lock so we do the put_page
1493 * of the tail pages after the split is complete.
1495 put_page(page_tail);
1499 * Only the head page (now become a regular page) is required
1500 * to be pinned by the caller.
1502 BUG_ON(page_count(page) <= 0);
1505 static int __split_huge_page_map(struct page *page,
1506 struct vm_area_struct *vma,
1507 unsigned long address)
1509 struct mm_struct *mm = vma->vm_mm;
1513 unsigned long haddr;
1515 spin_lock(&mm->page_table_lock);
1516 pmd = page_check_address_pmd(page, mm, address,
1517 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1519 pgtable = pgtable_trans_huge_withdraw(mm);
1520 pmd_populate(mm, &_pmd, pgtable);
1523 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1525 BUG_ON(PageCompound(page+i));
1526 entry = mk_pte(page + i, vma->vm_page_prot);
1527 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1528 if (!pmd_write(*pmd))
1529 entry = pte_wrprotect(entry);
1531 BUG_ON(page_mapcount(page) != 1);
1532 if (!pmd_young(*pmd))
1533 entry = pte_mkold(entry);
1534 pte = pte_offset_map(&_pmd, haddr);
1535 BUG_ON(!pte_none(*pte));
1536 set_pte_at(mm, haddr, pte, entry);
1540 smp_wmb(); /* make pte visible before pmd */
1542 * Up to this point the pmd is present and huge and
1543 * userland has the whole access to the hugepage
1544 * during the split (which happens in place). If we
1545 * overwrite the pmd with the not-huge version
1546 * pointing to the pte here (which of course we could
1547 * if all CPUs were bug free), userland could trigger
1548 * a small page size TLB miss on the small sized TLB
1549 * while the hugepage TLB entry is still established
1550 * in the huge TLB. Some CPU doesn't like that. See
1551 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1552 * Erratum 383 on page 93. Intel should be safe but is
1553 * also warns that it's only safe if the permission
1554 * and cache attributes of the two entries loaded in
1555 * the two TLB is identical (which should be the case
1556 * here). But it is generally safer to never allow
1557 * small and huge TLB entries for the same virtual
1558 * address to be loaded simultaneously. So instead of
1559 * doing "pmd_populate(); flush_tlb_range();" we first
1560 * mark the current pmd notpresent (atomically because
1561 * here the pmd_trans_huge and pmd_trans_splitting
1562 * must remain set at all times on the pmd until the
1563 * split is complete for this pmd), then we flush the
1564 * SMP TLB and finally we write the non-huge version
1565 * of the pmd entry with pmd_populate.
1567 pmdp_invalidate(vma, address, pmd);
1568 pmd_populate(mm, pmd, pgtable);
1571 spin_unlock(&mm->page_table_lock);
1576 /* must be called with anon_vma->root->mutex hold */
1577 static void __split_huge_page(struct page *page,
1578 struct anon_vma *anon_vma)
1580 int mapcount, mapcount2;
1581 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1582 struct anon_vma_chain *avc;
1584 BUG_ON(!PageHead(page));
1585 BUG_ON(PageTail(page));
1588 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1589 struct vm_area_struct *vma = avc->vma;
1590 unsigned long addr = vma_address(page, vma);
1591 BUG_ON(is_vma_temporary_stack(vma));
1592 mapcount += __split_huge_page_splitting(page, vma, addr);
1595 * It is critical that new vmas are added to the tail of the
1596 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1597 * and establishes a child pmd before
1598 * __split_huge_page_splitting() freezes the parent pmd (so if
1599 * we fail to prevent copy_huge_pmd() from running until the
1600 * whole __split_huge_page() is complete), we will still see
1601 * the newly established pmd of the child later during the
1602 * walk, to be able to set it as pmd_trans_splitting too.
1604 if (mapcount != page_mapcount(page))
1605 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1606 mapcount, page_mapcount(page));
1607 BUG_ON(mapcount != page_mapcount(page));
1609 __split_huge_page_refcount(page);
1612 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1613 struct vm_area_struct *vma = avc->vma;
1614 unsigned long addr = vma_address(page, vma);
1615 BUG_ON(is_vma_temporary_stack(vma));
1616 mapcount2 += __split_huge_page_map(page, vma, addr);
1618 if (mapcount != mapcount2)
1619 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1620 mapcount, mapcount2, page_mapcount(page));
1621 BUG_ON(mapcount != mapcount2);
1624 int split_huge_page(struct page *page)
1626 struct anon_vma *anon_vma;
1629 BUG_ON(is_huge_zero_pfn(page_to_pfn(page)));
1630 BUG_ON(!PageAnon(page));
1631 anon_vma = page_lock_anon_vma(page);
1635 if (!PageCompound(page))
1638 BUG_ON(!PageSwapBacked(page));
1639 __split_huge_page(page, anon_vma);
1640 count_vm_event(THP_SPLIT);
1642 BUG_ON(PageCompound(page));
1644 page_unlock_anon_vma(anon_vma);
1649 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1651 int hugepage_madvise(struct vm_area_struct *vma,
1652 unsigned long *vm_flags, int advice)
1654 struct mm_struct *mm = vma->vm_mm;
1659 * Be somewhat over-protective like KSM for now!
1661 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1663 if (mm->def_flags & VM_NOHUGEPAGE)
1665 *vm_flags &= ~VM_NOHUGEPAGE;
1666 *vm_flags |= VM_HUGEPAGE;
1668 * If the vma become good for khugepaged to scan,
1669 * register it here without waiting a page fault that
1670 * may not happen any time soon.
1672 if (unlikely(khugepaged_enter_vma_merge(vma)))
1675 case MADV_NOHUGEPAGE:
1677 * Be somewhat over-protective like KSM for now!
1679 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1681 *vm_flags &= ~VM_HUGEPAGE;
1682 *vm_flags |= VM_NOHUGEPAGE;
1684 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1685 * this vma even if we leave the mm registered in khugepaged if
1686 * it got registered before VM_NOHUGEPAGE was set.
1694 static int __init khugepaged_slab_init(void)
1696 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1697 sizeof(struct mm_slot),
1698 __alignof__(struct mm_slot), 0, NULL);
1705 static void __init khugepaged_slab_free(void)
1707 kmem_cache_destroy(mm_slot_cache);
1708 mm_slot_cache = NULL;
1711 static inline struct mm_slot *alloc_mm_slot(void)
1713 if (!mm_slot_cache) /* initialization failed */
1715 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1718 static inline void free_mm_slot(struct mm_slot *mm_slot)
1720 kmem_cache_free(mm_slot_cache, mm_slot);
1723 static int __init mm_slots_hash_init(void)
1725 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1733 static void __init mm_slots_hash_free(void)
1735 kfree(mm_slots_hash);
1736 mm_slots_hash = NULL;
1740 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1742 struct mm_slot *mm_slot;
1743 struct hlist_head *bucket;
1744 struct hlist_node *node;
1746 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1747 % MM_SLOTS_HASH_HEADS];
1748 hlist_for_each_entry(mm_slot, node, bucket, hash) {
1749 if (mm == mm_slot->mm)
1755 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1756 struct mm_slot *mm_slot)
1758 struct hlist_head *bucket;
1760 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1761 % MM_SLOTS_HASH_HEADS];
1763 hlist_add_head(&mm_slot->hash, bucket);
1766 static inline int khugepaged_test_exit(struct mm_struct *mm)
1768 return atomic_read(&mm->mm_users) == 0;
1771 int __khugepaged_enter(struct mm_struct *mm)
1773 struct mm_slot *mm_slot;
1776 mm_slot = alloc_mm_slot();
1780 /* __khugepaged_exit() must not run from under us */
1781 VM_BUG_ON(khugepaged_test_exit(mm));
1782 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1783 free_mm_slot(mm_slot);
1787 spin_lock(&khugepaged_mm_lock);
1788 insert_to_mm_slots_hash(mm, mm_slot);
1790 * Insert just behind the scanning cursor, to let the area settle
1793 wakeup = list_empty(&khugepaged_scan.mm_head);
1794 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1795 spin_unlock(&khugepaged_mm_lock);
1797 atomic_inc(&mm->mm_count);
1799 wake_up_interruptible(&khugepaged_wait);
1804 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1806 unsigned long hstart, hend;
1809 * Not yet faulted in so we will register later in the
1810 * page fault if needed.
1814 /* khugepaged not yet working on file or special mappings */
1816 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1817 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1818 hend = vma->vm_end & HPAGE_PMD_MASK;
1820 return khugepaged_enter(vma);
1824 void __khugepaged_exit(struct mm_struct *mm)
1826 struct mm_slot *mm_slot;
1829 spin_lock(&khugepaged_mm_lock);
1830 mm_slot = get_mm_slot(mm);
1831 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1832 hlist_del(&mm_slot->hash);
1833 list_del(&mm_slot->mm_node);
1836 spin_unlock(&khugepaged_mm_lock);
1839 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1840 free_mm_slot(mm_slot);
1842 } else if (mm_slot) {
1844 * This is required to serialize against
1845 * khugepaged_test_exit() (which is guaranteed to run
1846 * under mmap sem read mode). Stop here (after we
1847 * return all pagetables will be destroyed) until
1848 * khugepaged has finished working on the pagetables
1849 * under the mmap_sem.
1851 down_write(&mm->mmap_sem);
1852 up_write(&mm->mmap_sem);
1856 static void release_pte_page(struct page *page)
1858 /* 0 stands for page_is_file_cache(page) == false */
1859 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1861 putback_lru_page(page);
1864 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1866 while (--_pte >= pte) {
1867 pte_t pteval = *_pte;
1868 if (!pte_none(pteval))
1869 release_pte_page(pte_page(pteval));
1873 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1874 unsigned long address,
1879 int referenced = 0, none = 0;
1880 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1881 _pte++, address += PAGE_SIZE) {
1882 pte_t pteval = *_pte;
1883 if (pte_none(pteval)) {
1884 if (++none <= khugepaged_max_ptes_none)
1889 if (!pte_present(pteval) || !pte_write(pteval))
1891 page = vm_normal_page(vma, address, pteval);
1892 if (unlikely(!page))
1895 VM_BUG_ON(PageCompound(page));
1896 BUG_ON(!PageAnon(page));
1897 VM_BUG_ON(!PageSwapBacked(page));
1899 /* cannot use mapcount: can't collapse if there's a gup pin */
1900 if (page_count(page) != 1)
1903 * We can do it before isolate_lru_page because the
1904 * page can't be freed from under us. NOTE: PG_lock
1905 * is needed to serialize against split_huge_page
1906 * when invoked from the VM.
1908 if (!trylock_page(page))
1911 * Isolate the page to avoid collapsing an hugepage
1912 * currently in use by the VM.
1914 if (isolate_lru_page(page)) {
1918 /* 0 stands for page_is_file_cache(page) == false */
1919 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1920 VM_BUG_ON(!PageLocked(page));
1921 VM_BUG_ON(PageLRU(page));
1923 /* If there is no mapped pte young don't collapse the page */
1924 if (pte_young(pteval) || PageReferenced(page) ||
1925 mmu_notifier_test_young(vma->vm_mm, address))
1928 if (likely(referenced))
1931 release_pte_pages(pte, _pte);
1935 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1936 struct vm_area_struct *vma,
1937 unsigned long address,
1941 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1942 pte_t pteval = *_pte;
1943 struct page *src_page;
1945 if (pte_none(pteval)) {
1946 clear_user_highpage(page, address);
1947 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1949 src_page = pte_page(pteval);
1950 copy_user_highpage(page, src_page, address, vma);
1951 VM_BUG_ON(page_mapcount(src_page) != 1);
1952 release_pte_page(src_page);
1954 * ptl mostly unnecessary, but preempt has to
1955 * be disabled to update the per-cpu stats
1956 * inside page_remove_rmap().
1960 * paravirt calls inside pte_clear here are
1963 pte_clear(vma->vm_mm, address, _pte);
1964 page_remove_rmap(src_page);
1966 free_page_and_swap_cache(src_page);
1969 address += PAGE_SIZE;
1974 static void khugepaged_alloc_sleep(void)
1976 wait_event_freezable_timeout(khugepaged_wait, false,
1977 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
1981 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1983 if (IS_ERR(*hpage)) {
1989 khugepaged_alloc_sleep();
1990 } else if (*hpage) {
1999 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2000 struct vm_area_struct *vma, unsigned long address,
2005 * Allocate the page while the vma is still valid and under
2006 * the mmap_sem read mode so there is no memory allocation
2007 * later when we take the mmap_sem in write mode. This is more
2008 * friendly behavior (OTOH it may actually hide bugs) to
2009 * filesystems in userland with daemons allocating memory in
2010 * the userland I/O paths. Allocating memory with the
2011 * mmap_sem in read mode is good idea also to allow greater
2014 *hpage = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2015 node, __GFP_OTHER_NODE);
2018 * After allocating the hugepage, release the mmap_sem read lock in
2019 * preparation for taking it in write mode.
2021 up_read(&mm->mmap_sem);
2022 if (unlikely(!*hpage)) {
2023 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2024 *hpage = ERR_PTR(-ENOMEM);
2028 count_vm_event(THP_COLLAPSE_ALLOC);
2032 static struct page *khugepaged_alloc_hugepage(bool *wait)
2037 hpage = alloc_hugepage(khugepaged_defrag());
2039 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2044 khugepaged_alloc_sleep();
2046 count_vm_event(THP_COLLAPSE_ALLOC);
2047 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2052 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2055 *hpage = khugepaged_alloc_hugepage(wait);
2057 if (unlikely(!*hpage))
2064 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2065 struct vm_area_struct *vma, unsigned long address,
2068 up_read(&mm->mmap_sem);
2074 static bool hugepage_vma_check(struct vm_area_struct *vma)
2076 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2077 (vma->vm_flags & VM_NOHUGEPAGE))
2080 if (!vma->anon_vma || vma->vm_ops)
2082 if (is_vma_temporary_stack(vma))
2084 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2088 static void collapse_huge_page(struct mm_struct *mm,
2089 unsigned long address,
2090 struct page **hpage,
2091 struct vm_area_struct *vma,
2097 struct page *new_page;
2100 unsigned long hstart, hend;
2101 unsigned long mmun_start; /* For mmu_notifiers */
2102 unsigned long mmun_end; /* For mmu_notifiers */
2104 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2106 /* release the mmap_sem read lock. */
2107 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2111 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2115 * Prevent all access to pagetables with the exception of
2116 * gup_fast later hanlded by the ptep_clear_flush and the VM
2117 * handled by the anon_vma lock + PG_lock.
2119 down_write(&mm->mmap_sem);
2120 if (unlikely(khugepaged_test_exit(mm)))
2123 vma = find_vma(mm, address);
2124 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2125 hend = vma->vm_end & HPAGE_PMD_MASK;
2126 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2128 if (!hugepage_vma_check(vma))
2130 pmd = mm_find_pmd(mm, address);
2133 if (pmd_trans_huge(*pmd))
2136 anon_vma_lock(vma->anon_vma);
2138 pte = pte_offset_map(pmd, address);
2139 ptl = pte_lockptr(mm, pmd);
2141 mmun_start = address;
2142 mmun_end = address + HPAGE_PMD_SIZE;
2143 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2144 spin_lock(&mm->page_table_lock); /* probably unnecessary */
2146 * After this gup_fast can't run anymore. This also removes
2147 * any huge TLB entry from the CPU so we won't allow
2148 * huge and small TLB entries for the same virtual address
2149 * to avoid the risk of CPU bugs in that area.
2151 _pmd = pmdp_clear_flush(vma, address, pmd);
2152 spin_unlock(&mm->page_table_lock);
2153 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2156 isolated = __collapse_huge_page_isolate(vma, address, pte);
2159 if (unlikely(!isolated)) {
2161 spin_lock(&mm->page_table_lock);
2162 BUG_ON(!pmd_none(*pmd));
2163 set_pmd_at(mm, address, pmd, _pmd);
2164 spin_unlock(&mm->page_table_lock);
2165 anon_vma_unlock(vma->anon_vma);
2170 * All pages are isolated and locked so anon_vma rmap
2171 * can't run anymore.
2173 anon_vma_unlock(vma->anon_vma);
2175 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2177 __SetPageUptodate(new_page);
2178 pgtable = pmd_pgtable(_pmd);
2180 _pmd = mk_huge_pmd(new_page, vma);
2183 * spin_lock() below is not the equivalent of smp_wmb(), so
2184 * this is needed to avoid the copy_huge_page writes to become
2185 * visible after the set_pmd_at() write.
2189 spin_lock(&mm->page_table_lock);
2190 BUG_ON(!pmd_none(*pmd));
2191 page_add_new_anon_rmap(new_page, vma, address);
2192 set_pmd_at(mm, address, pmd, _pmd);
2193 update_mmu_cache_pmd(vma, address, pmd);
2194 pgtable_trans_huge_deposit(mm, pgtable);
2195 spin_unlock(&mm->page_table_lock);
2199 khugepaged_pages_collapsed++;
2201 up_write(&mm->mmap_sem);
2205 mem_cgroup_uncharge_page(new_page);
2209 static int khugepaged_scan_pmd(struct mm_struct *mm,
2210 struct vm_area_struct *vma,
2211 unsigned long address,
2212 struct page **hpage)
2216 int ret = 0, referenced = 0, none = 0;
2218 unsigned long _address;
2222 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2224 pmd = mm_find_pmd(mm, address);
2227 if (pmd_trans_huge(*pmd))
2230 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2231 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2232 _pte++, _address += PAGE_SIZE) {
2233 pte_t pteval = *_pte;
2234 if (pte_none(pteval)) {
2235 if (++none <= khugepaged_max_ptes_none)
2240 if (!pte_present(pteval) || !pte_write(pteval))
2242 page = vm_normal_page(vma, _address, pteval);
2243 if (unlikely(!page))
2246 * Chose the node of the first page. This could
2247 * be more sophisticated and look at more pages,
2248 * but isn't for now.
2251 node = page_to_nid(page);
2252 VM_BUG_ON(PageCompound(page));
2253 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2255 /* cannot use mapcount: can't collapse if there's a gup pin */
2256 if (page_count(page) != 1)
2258 if (pte_young(pteval) || PageReferenced(page) ||
2259 mmu_notifier_test_young(vma->vm_mm, address))
2265 pte_unmap_unlock(pte, ptl);
2267 /* collapse_huge_page will return with the mmap_sem released */
2268 collapse_huge_page(mm, address, hpage, vma, node);
2273 static void collect_mm_slot(struct mm_slot *mm_slot)
2275 struct mm_struct *mm = mm_slot->mm;
2277 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2279 if (khugepaged_test_exit(mm)) {
2281 hlist_del(&mm_slot->hash);
2282 list_del(&mm_slot->mm_node);
2285 * Not strictly needed because the mm exited already.
2287 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2290 /* khugepaged_mm_lock actually not necessary for the below */
2291 free_mm_slot(mm_slot);
2296 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2297 struct page **hpage)
2298 __releases(&khugepaged_mm_lock)
2299 __acquires(&khugepaged_mm_lock)
2301 struct mm_slot *mm_slot;
2302 struct mm_struct *mm;
2303 struct vm_area_struct *vma;
2307 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2309 if (khugepaged_scan.mm_slot)
2310 mm_slot = khugepaged_scan.mm_slot;
2312 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2313 struct mm_slot, mm_node);
2314 khugepaged_scan.address = 0;
2315 khugepaged_scan.mm_slot = mm_slot;
2317 spin_unlock(&khugepaged_mm_lock);
2320 down_read(&mm->mmap_sem);
2321 if (unlikely(khugepaged_test_exit(mm)))
2324 vma = find_vma(mm, khugepaged_scan.address);
2327 for (; vma; vma = vma->vm_next) {
2328 unsigned long hstart, hend;
2331 if (unlikely(khugepaged_test_exit(mm))) {
2335 if (!hugepage_vma_check(vma)) {
2340 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2341 hend = vma->vm_end & HPAGE_PMD_MASK;
2344 if (khugepaged_scan.address > hend)
2346 if (khugepaged_scan.address < hstart)
2347 khugepaged_scan.address = hstart;
2348 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2350 while (khugepaged_scan.address < hend) {
2353 if (unlikely(khugepaged_test_exit(mm)))
2354 goto breakouterloop;
2356 VM_BUG_ON(khugepaged_scan.address < hstart ||
2357 khugepaged_scan.address + HPAGE_PMD_SIZE >
2359 ret = khugepaged_scan_pmd(mm, vma,
2360 khugepaged_scan.address,
2362 /* move to next address */
2363 khugepaged_scan.address += HPAGE_PMD_SIZE;
2364 progress += HPAGE_PMD_NR;
2366 /* we released mmap_sem so break loop */
2367 goto breakouterloop_mmap_sem;
2368 if (progress >= pages)
2369 goto breakouterloop;
2373 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2374 breakouterloop_mmap_sem:
2376 spin_lock(&khugepaged_mm_lock);
2377 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2379 * Release the current mm_slot if this mm is about to die, or
2380 * if we scanned all vmas of this mm.
2382 if (khugepaged_test_exit(mm) || !vma) {
2384 * Make sure that if mm_users is reaching zero while
2385 * khugepaged runs here, khugepaged_exit will find
2386 * mm_slot not pointing to the exiting mm.
2388 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2389 khugepaged_scan.mm_slot = list_entry(
2390 mm_slot->mm_node.next,
2391 struct mm_slot, mm_node);
2392 khugepaged_scan.address = 0;
2394 khugepaged_scan.mm_slot = NULL;
2395 khugepaged_full_scans++;
2398 collect_mm_slot(mm_slot);
2404 static int khugepaged_has_work(void)
2406 return !list_empty(&khugepaged_scan.mm_head) &&
2407 khugepaged_enabled();
2410 static int khugepaged_wait_event(void)
2412 return !list_empty(&khugepaged_scan.mm_head) ||
2413 kthread_should_stop();
2416 static void khugepaged_do_scan(void)
2418 struct page *hpage = NULL;
2419 unsigned int progress = 0, pass_through_head = 0;
2420 unsigned int pages = khugepaged_pages_to_scan;
2423 barrier(); /* write khugepaged_pages_to_scan to local stack */
2425 while (progress < pages) {
2426 if (!khugepaged_prealloc_page(&hpage, &wait))
2431 if (unlikely(kthread_should_stop() || freezing(current)))
2434 spin_lock(&khugepaged_mm_lock);
2435 if (!khugepaged_scan.mm_slot)
2436 pass_through_head++;
2437 if (khugepaged_has_work() &&
2438 pass_through_head < 2)
2439 progress += khugepaged_scan_mm_slot(pages - progress,
2443 spin_unlock(&khugepaged_mm_lock);
2446 if (!IS_ERR_OR_NULL(hpage))
2450 static void khugepaged_wait_work(void)
2454 if (khugepaged_has_work()) {
2455 if (!khugepaged_scan_sleep_millisecs)
2458 wait_event_freezable_timeout(khugepaged_wait,
2459 kthread_should_stop(),
2460 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2464 if (khugepaged_enabled())
2465 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2468 static int khugepaged(void *none)
2470 struct mm_slot *mm_slot;
2473 set_user_nice(current, 19);
2475 while (!kthread_should_stop()) {
2476 khugepaged_do_scan();
2477 khugepaged_wait_work();
2480 spin_lock(&khugepaged_mm_lock);
2481 mm_slot = khugepaged_scan.mm_slot;
2482 khugepaged_scan.mm_slot = NULL;
2484 collect_mm_slot(mm_slot);
2485 spin_unlock(&khugepaged_mm_lock);
2489 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2490 unsigned long haddr, pmd_t *pmd)
2492 struct mm_struct *mm = vma->vm_mm;
2497 pmdp_clear_flush(vma, haddr, pmd);
2498 /* leave pmd empty until pte is filled */
2500 pgtable = pgtable_trans_huge_withdraw(mm);
2501 pmd_populate(mm, &_pmd, pgtable);
2503 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2505 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2506 entry = pte_mkspecial(entry);
2507 pte = pte_offset_map(&_pmd, haddr);
2508 VM_BUG_ON(!pte_none(*pte));
2509 set_pte_at(mm, haddr, pte, entry);
2512 smp_wmb(); /* make pte visible before pmd */
2513 pmd_populate(mm, pmd, pgtable);
2516 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2520 struct mm_struct *mm = vma->vm_mm;
2521 unsigned long haddr = address & HPAGE_PMD_MASK;
2522 unsigned long mmun_start; /* For mmu_notifiers */
2523 unsigned long mmun_end; /* For mmu_notifiers */
2525 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2528 mmun_end = haddr + HPAGE_PMD_SIZE;
2529 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2530 spin_lock(&mm->page_table_lock);
2531 if (unlikely(!pmd_trans_huge(*pmd))) {
2532 spin_unlock(&mm->page_table_lock);
2533 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2536 if (is_huge_zero_pmd(*pmd)) {
2537 __split_huge_zero_page_pmd(vma, haddr, pmd);
2538 spin_unlock(&mm->page_table_lock);
2539 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2542 page = pmd_page(*pmd);
2543 VM_BUG_ON(!page_count(page));
2545 spin_unlock(&mm->page_table_lock);
2546 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2548 split_huge_page(page);
2551 BUG_ON(pmd_trans_huge(*pmd));
2554 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2557 struct vm_area_struct *vma;
2559 vma = find_vma(mm, address);
2560 BUG_ON(vma == NULL);
2561 split_huge_page_pmd(vma, address, pmd);
2564 static void split_huge_page_address(struct mm_struct *mm,
2565 unsigned long address)
2569 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2571 pmd = mm_find_pmd(mm, address);
2575 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2576 * materialize from under us.
2578 split_huge_page_pmd_mm(mm, address, pmd);
2581 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2582 unsigned long start,
2587 * If the new start address isn't hpage aligned and it could
2588 * previously contain an hugepage: check if we need to split
2591 if (start & ~HPAGE_PMD_MASK &&
2592 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2593 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2594 split_huge_page_address(vma->vm_mm, start);
2597 * If the new end address isn't hpage aligned and it could
2598 * previously contain an hugepage: check if we need to split
2601 if (end & ~HPAGE_PMD_MASK &&
2602 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2603 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2604 split_huge_page_address(vma->vm_mm, end);
2607 * If we're also updating the vma->vm_next->vm_start, if the new
2608 * vm_next->vm_start isn't page aligned and it could previously
2609 * contain an hugepage: check if we need to split an huge pmd.
2611 if (adjust_next > 0) {
2612 struct vm_area_struct *next = vma->vm_next;
2613 unsigned long nstart = next->vm_start;
2614 nstart += adjust_next << PAGE_SHIFT;
2615 if (nstart & ~HPAGE_PMD_MASK &&
2616 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2617 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2618 split_huge_page_address(next->vm_mm, nstart);