Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next
[linux-2.6.git] / mm / huge_memory.c
blob243e710c6039a287f93ec9f303639c7316dbdf0e
1 /*
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.
6 */
8 #include <linux/mm.h>
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/shrinker.h>
16 #include <linux/mm_inline.h>
17 #include <linux/kthread.h>
18 #include <linux/khugepaged.h>
19 #include <linux/freezer.h>
20 #include <linux/mman.h>
21 #include <linux/pagemap.h>
22 #include <linux/migrate.h>
23 #include <linux/hashtable.h>
25 #include <asm/tlb.h>
26 #include <asm/pgalloc.h>
27 #include "internal.h"
30 * By default transparent hugepage support is enabled for all mappings
31 * and khugepaged scans all mappings. Defrag is only invoked by
32 * khugepaged hugepage allocations and by page faults inside
33 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
34 * allocations.
36 unsigned long transparent_hugepage_flags __read_mostly =
37 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
38 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
39 #endif
40 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
41 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
42 #endif
43 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
44 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
45 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
47 /* default scan 8*512 pte (or vmas) every 30 second */
48 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
49 static unsigned int khugepaged_pages_collapsed;
50 static unsigned int khugepaged_full_scans;
51 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
52 /* during fragmentation poll the hugepage allocator once every minute */
53 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
54 static struct task_struct *khugepaged_thread __read_mostly;
55 static DEFINE_MUTEX(khugepaged_mutex);
56 static DEFINE_SPINLOCK(khugepaged_mm_lock);
57 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
59 * default collapse hugepages if there is at least one pte mapped like
60 * it would have happened if the vma was large enough during page
61 * fault.
63 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
65 static int khugepaged(void *none);
66 static int khugepaged_slab_init(void);
68 #define MM_SLOTS_HASH_BITS 10
69 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
71 static struct kmem_cache *mm_slot_cache __read_mostly;
73 /**
74 * struct mm_slot - hash lookup from mm to mm_slot
75 * @hash: hash collision list
76 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
77 * @mm: the mm that this information is valid for
79 struct mm_slot {
80 struct hlist_node hash;
81 struct list_head mm_node;
82 struct mm_struct *mm;
85 /**
86 * struct khugepaged_scan - cursor for scanning
87 * @mm_head: the head of the mm list to scan
88 * @mm_slot: the current mm_slot we are scanning
89 * @address: the next address inside that to be scanned
91 * There is only the one khugepaged_scan instance of this cursor structure.
93 struct khugepaged_scan {
94 struct list_head mm_head;
95 struct mm_slot *mm_slot;
96 unsigned long address;
98 static struct khugepaged_scan khugepaged_scan = {
99 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
103 static int set_recommended_min_free_kbytes(void)
105 struct zone *zone;
106 int nr_zones = 0;
107 unsigned long recommended_min;
109 if (!khugepaged_enabled())
110 return 0;
112 for_each_populated_zone(zone)
113 nr_zones++;
115 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
116 recommended_min = pageblock_nr_pages * nr_zones * 2;
119 * Make sure that on average at least two pageblocks are almost free
120 * of another type, one for a migratetype to fall back to and a
121 * second to avoid subsequent fallbacks of other types There are 3
122 * MIGRATE_TYPES we care about.
124 recommended_min += pageblock_nr_pages * nr_zones *
125 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
127 /* don't ever allow to reserve more than 5% of the lowmem */
128 recommended_min = min(recommended_min,
129 (unsigned long) nr_free_buffer_pages() / 20);
130 recommended_min <<= (PAGE_SHIFT-10);
132 if (recommended_min > min_free_kbytes)
133 min_free_kbytes = recommended_min;
134 setup_per_zone_wmarks();
135 return 0;
137 late_initcall(set_recommended_min_free_kbytes);
139 static int start_khugepaged(void)
141 int err = 0;
142 if (khugepaged_enabled()) {
143 if (!khugepaged_thread)
144 khugepaged_thread = kthread_run(khugepaged, NULL,
145 "khugepaged");
146 if (unlikely(IS_ERR(khugepaged_thread))) {
147 printk(KERN_ERR
148 "khugepaged: kthread_run(khugepaged) failed\n");
149 err = PTR_ERR(khugepaged_thread);
150 khugepaged_thread = NULL;
153 if (!list_empty(&khugepaged_scan.mm_head))
154 wake_up_interruptible(&khugepaged_wait);
156 set_recommended_min_free_kbytes();
157 } else if (khugepaged_thread) {
158 kthread_stop(khugepaged_thread);
159 khugepaged_thread = NULL;
162 return err;
165 static atomic_t huge_zero_refcount;
166 static struct page *huge_zero_page __read_mostly;
168 static inline bool is_huge_zero_page(struct page *page)
170 return ACCESS_ONCE(huge_zero_page) == page;
173 static inline bool is_huge_zero_pmd(pmd_t pmd)
175 return is_huge_zero_page(pmd_page(pmd));
178 static struct page *get_huge_zero_page(void)
180 struct page *zero_page;
181 retry:
182 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
183 return ACCESS_ONCE(huge_zero_page);
185 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
186 HPAGE_PMD_ORDER);
187 if (!zero_page) {
188 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
189 return NULL;
191 count_vm_event(THP_ZERO_PAGE_ALLOC);
192 preempt_disable();
193 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
194 preempt_enable();
195 __free_page(zero_page);
196 goto retry;
199 /* We take additional reference here. It will be put back by shrinker */
200 atomic_set(&huge_zero_refcount, 2);
201 preempt_enable();
202 return ACCESS_ONCE(huge_zero_page);
205 static void put_huge_zero_page(void)
208 * Counter should never go to zero here. Only shrinker can put
209 * last reference.
211 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
214 static int shrink_huge_zero_page(struct shrinker *shrink,
215 struct shrink_control *sc)
217 if (!sc->nr_to_scan)
218 /* we can free zero page only if last reference remains */
219 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
221 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
222 struct page *zero_page = xchg(&huge_zero_page, NULL);
223 BUG_ON(zero_page == NULL);
224 __free_page(zero_page);
227 return 0;
230 static struct shrinker huge_zero_page_shrinker = {
231 .shrink = shrink_huge_zero_page,
232 .seeks = DEFAULT_SEEKS,
235 #ifdef CONFIG_SYSFS
237 static ssize_t double_flag_show(struct kobject *kobj,
238 struct kobj_attribute *attr, char *buf,
239 enum transparent_hugepage_flag enabled,
240 enum transparent_hugepage_flag req_madv)
242 if (test_bit(enabled, &transparent_hugepage_flags)) {
243 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
244 return sprintf(buf, "[always] madvise never\n");
245 } else if (test_bit(req_madv, &transparent_hugepage_flags))
246 return sprintf(buf, "always [madvise] never\n");
247 else
248 return sprintf(buf, "always madvise [never]\n");
250 static ssize_t double_flag_store(struct kobject *kobj,
251 struct kobj_attribute *attr,
252 const char *buf, size_t count,
253 enum transparent_hugepage_flag enabled,
254 enum transparent_hugepage_flag req_madv)
256 if (!memcmp("always", buf,
257 min(sizeof("always")-1, count))) {
258 set_bit(enabled, &transparent_hugepage_flags);
259 clear_bit(req_madv, &transparent_hugepage_flags);
260 } else if (!memcmp("madvise", buf,
261 min(sizeof("madvise")-1, count))) {
262 clear_bit(enabled, &transparent_hugepage_flags);
263 set_bit(req_madv, &transparent_hugepage_flags);
264 } else if (!memcmp("never", buf,
265 min(sizeof("never")-1, count))) {
266 clear_bit(enabled, &transparent_hugepage_flags);
267 clear_bit(req_madv, &transparent_hugepage_flags);
268 } else
269 return -EINVAL;
271 return count;
274 static ssize_t enabled_show(struct kobject *kobj,
275 struct kobj_attribute *attr, char *buf)
277 return double_flag_show(kobj, attr, buf,
278 TRANSPARENT_HUGEPAGE_FLAG,
279 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
281 static ssize_t enabled_store(struct kobject *kobj,
282 struct kobj_attribute *attr,
283 const char *buf, size_t count)
285 ssize_t ret;
287 ret = double_flag_store(kobj, attr, buf, count,
288 TRANSPARENT_HUGEPAGE_FLAG,
289 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
291 if (ret > 0) {
292 int err;
294 mutex_lock(&khugepaged_mutex);
295 err = start_khugepaged();
296 mutex_unlock(&khugepaged_mutex);
298 if (err)
299 ret = err;
302 return ret;
304 static struct kobj_attribute enabled_attr =
305 __ATTR(enabled, 0644, enabled_show, enabled_store);
307 static ssize_t single_flag_show(struct kobject *kobj,
308 struct kobj_attribute *attr, char *buf,
309 enum transparent_hugepage_flag flag)
311 return sprintf(buf, "%d\n",
312 !!test_bit(flag, &transparent_hugepage_flags));
315 static ssize_t single_flag_store(struct kobject *kobj,
316 struct kobj_attribute *attr,
317 const char *buf, size_t count,
318 enum transparent_hugepage_flag flag)
320 unsigned long value;
321 int ret;
323 ret = kstrtoul(buf, 10, &value);
324 if (ret < 0)
325 return ret;
326 if (value > 1)
327 return -EINVAL;
329 if (value)
330 set_bit(flag, &transparent_hugepage_flags);
331 else
332 clear_bit(flag, &transparent_hugepage_flags);
334 return count;
338 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
339 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
340 * memory just to allocate one more hugepage.
342 static ssize_t defrag_show(struct kobject *kobj,
343 struct kobj_attribute *attr, char *buf)
345 return double_flag_show(kobj, attr, buf,
346 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
347 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
349 static ssize_t defrag_store(struct kobject *kobj,
350 struct kobj_attribute *attr,
351 const char *buf, size_t count)
353 return double_flag_store(kobj, attr, buf, count,
354 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
355 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
357 static struct kobj_attribute defrag_attr =
358 __ATTR(defrag, 0644, defrag_show, defrag_store);
360 static ssize_t use_zero_page_show(struct kobject *kobj,
361 struct kobj_attribute *attr, char *buf)
363 return single_flag_show(kobj, attr, buf,
364 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
366 static ssize_t use_zero_page_store(struct kobject *kobj,
367 struct kobj_attribute *attr, const char *buf, size_t count)
369 return single_flag_store(kobj, attr, buf, count,
370 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
372 static struct kobj_attribute use_zero_page_attr =
373 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
374 #ifdef CONFIG_DEBUG_VM
375 static ssize_t debug_cow_show(struct kobject *kobj,
376 struct kobj_attribute *attr, char *buf)
378 return single_flag_show(kobj, attr, buf,
379 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
381 static ssize_t debug_cow_store(struct kobject *kobj,
382 struct kobj_attribute *attr,
383 const char *buf, size_t count)
385 return single_flag_store(kobj, attr, buf, count,
386 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
388 static struct kobj_attribute debug_cow_attr =
389 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
390 #endif /* CONFIG_DEBUG_VM */
392 static struct attribute *hugepage_attr[] = {
393 &enabled_attr.attr,
394 &defrag_attr.attr,
395 &use_zero_page_attr.attr,
396 #ifdef CONFIG_DEBUG_VM
397 &debug_cow_attr.attr,
398 #endif
399 NULL,
402 static struct attribute_group hugepage_attr_group = {
403 .attrs = hugepage_attr,
406 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
407 struct kobj_attribute *attr,
408 char *buf)
410 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
413 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
414 struct kobj_attribute *attr,
415 const char *buf, size_t count)
417 unsigned long msecs;
418 int err;
420 err = strict_strtoul(buf, 10, &msecs);
421 if (err || msecs > UINT_MAX)
422 return -EINVAL;
424 khugepaged_scan_sleep_millisecs = msecs;
425 wake_up_interruptible(&khugepaged_wait);
427 return count;
429 static struct kobj_attribute scan_sleep_millisecs_attr =
430 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
431 scan_sleep_millisecs_store);
433 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
434 struct kobj_attribute *attr,
435 char *buf)
437 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
440 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
441 struct kobj_attribute *attr,
442 const char *buf, size_t count)
444 unsigned long msecs;
445 int err;
447 err = strict_strtoul(buf, 10, &msecs);
448 if (err || msecs > UINT_MAX)
449 return -EINVAL;
451 khugepaged_alloc_sleep_millisecs = msecs;
452 wake_up_interruptible(&khugepaged_wait);
454 return count;
456 static struct kobj_attribute alloc_sleep_millisecs_attr =
457 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
458 alloc_sleep_millisecs_store);
460 static ssize_t pages_to_scan_show(struct kobject *kobj,
461 struct kobj_attribute *attr,
462 char *buf)
464 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
466 static ssize_t pages_to_scan_store(struct kobject *kobj,
467 struct kobj_attribute *attr,
468 const char *buf, size_t count)
470 int err;
471 unsigned long pages;
473 err = strict_strtoul(buf, 10, &pages);
474 if (err || !pages || pages > UINT_MAX)
475 return -EINVAL;
477 khugepaged_pages_to_scan = pages;
479 return count;
481 static struct kobj_attribute pages_to_scan_attr =
482 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
483 pages_to_scan_store);
485 static ssize_t pages_collapsed_show(struct kobject *kobj,
486 struct kobj_attribute *attr,
487 char *buf)
489 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
491 static struct kobj_attribute pages_collapsed_attr =
492 __ATTR_RO(pages_collapsed);
494 static ssize_t full_scans_show(struct kobject *kobj,
495 struct kobj_attribute *attr,
496 char *buf)
498 return sprintf(buf, "%u\n", khugepaged_full_scans);
500 static struct kobj_attribute full_scans_attr =
501 __ATTR_RO(full_scans);
503 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
504 struct kobj_attribute *attr, char *buf)
506 return single_flag_show(kobj, attr, buf,
507 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
509 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
510 struct kobj_attribute *attr,
511 const char *buf, size_t count)
513 return single_flag_store(kobj, attr, buf, count,
514 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
516 static struct kobj_attribute khugepaged_defrag_attr =
517 __ATTR(defrag, 0644, khugepaged_defrag_show,
518 khugepaged_defrag_store);
521 * max_ptes_none controls if khugepaged should collapse hugepages over
522 * any unmapped ptes in turn potentially increasing the memory
523 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
524 * reduce the available free memory in the system as it
525 * runs. Increasing max_ptes_none will instead potentially reduce the
526 * free memory in the system during the khugepaged scan.
528 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
529 struct kobj_attribute *attr,
530 char *buf)
532 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
534 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
535 struct kobj_attribute *attr,
536 const char *buf, size_t count)
538 int err;
539 unsigned long max_ptes_none;
541 err = strict_strtoul(buf, 10, &max_ptes_none);
542 if (err || max_ptes_none > HPAGE_PMD_NR-1)
543 return -EINVAL;
545 khugepaged_max_ptes_none = max_ptes_none;
547 return count;
549 static struct kobj_attribute khugepaged_max_ptes_none_attr =
550 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
551 khugepaged_max_ptes_none_store);
553 static struct attribute *khugepaged_attr[] = {
554 &khugepaged_defrag_attr.attr,
555 &khugepaged_max_ptes_none_attr.attr,
556 &pages_to_scan_attr.attr,
557 &pages_collapsed_attr.attr,
558 &full_scans_attr.attr,
559 &scan_sleep_millisecs_attr.attr,
560 &alloc_sleep_millisecs_attr.attr,
561 NULL,
564 static struct attribute_group khugepaged_attr_group = {
565 .attrs = khugepaged_attr,
566 .name = "khugepaged",
569 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
571 int err;
573 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
574 if (unlikely(!*hugepage_kobj)) {
575 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
576 return -ENOMEM;
579 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
580 if (err) {
581 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
582 goto delete_obj;
585 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
586 if (err) {
587 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
588 goto remove_hp_group;
591 return 0;
593 remove_hp_group:
594 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
595 delete_obj:
596 kobject_put(*hugepage_kobj);
597 return err;
600 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
602 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
603 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
604 kobject_put(hugepage_kobj);
606 #else
607 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
609 return 0;
612 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
615 #endif /* CONFIG_SYSFS */
617 static int __init hugepage_init(void)
619 int err;
620 struct kobject *hugepage_kobj;
622 if (!has_transparent_hugepage()) {
623 transparent_hugepage_flags = 0;
624 return -EINVAL;
627 err = hugepage_init_sysfs(&hugepage_kobj);
628 if (err)
629 return err;
631 err = khugepaged_slab_init();
632 if (err)
633 goto out;
635 register_shrinker(&huge_zero_page_shrinker);
638 * By default disable transparent hugepages on smaller systems,
639 * where the extra memory used could hurt more than TLB overhead
640 * is likely to save. The admin can still enable it through /sys.
642 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
643 transparent_hugepage_flags = 0;
645 start_khugepaged();
647 return 0;
648 out:
649 hugepage_exit_sysfs(hugepage_kobj);
650 return err;
652 module_init(hugepage_init)
654 static int __init setup_transparent_hugepage(char *str)
656 int ret = 0;
657 if (!str)
658 goto out;
659 if (!strcmp(str, "always")) {
660 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
661 &transparent_hugepage_flags);
662 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
663 &transparent_hugepage_flags);
664 ret = 1;
665 } else if (!strcmp(str, "madvise")) {
666 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
667 &transparent_hugepage_flags);
668 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
669 &transparent_hugepage_flags);
670 ret = 1;
671 } else if (!strcmp(str, "never")) {
672 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
673 &transparent_hugepage_flags);
674 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
675 &transparent_hugepage_flags);
676 ret = 1;
678 out:
679 if (!ret)
680 printk(KERN_WARNING
681 "transparent_hugepage= cannot parse, ignored\n");
682 return ret;
684 __setup("transparent_hugepage=", setup_transparent_hugepage);
686 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
688 if (likely(vma->vm_flags & VM_WRITE))
689 pmd = pmd_mkwrite(pmd);
690 return pmd;
693 static inline pmd_t mk_huge_pmd(struct page *page, struct vm_area_struct *vma)
695 pmd_t entry;
696 entry = mk_pmd(page, vma->vm_page_prot);
697 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
698 entry = pmd_mkhuge(entry);
699 return entry;
702 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
703 struct vm_area_struct *vma,
704 unsigned long haddr, pmd_t *pmd,
705 struct page *page)
707 pgtable_t pgtable;
709 VM_BUG_ON(!PageCompound(page));
710 pgtable = pte_alloc_one(mm, haddr);
711 if (unlikely(!pgtable))
712 return VM_FAULT_OOM;
714 clear_huge_page(page, haddr, HPAGE_PMD_NR);
716 * The memory barrier inside __SetPageUptodate makes sure that
717 * clear_huge_page writes become visible before the set_pmd_at()
718 * write.
720 __SetPageUptodate(page);
722 spin_lock(&mm->page_table_lock);
723 if (unlikely(!pmd_none(*pmd))) {
724 spin_unlock(&mm->page_table_lock);
725 mem_cgroup_uncharge_page(page);
726 put_page(page);
727 pte_free(mm, pgtable);
728 } else {
729 pmd_t entry;
730 entry = mk_huge_pmd(page, vma);
731 page_add_new_anon_rmap(page, vma, haddr);
732 pgtable_trans_huge_deposit(mm, pmd, pgtable);
733 set_pmd_at(mm, haddr, pmd, entry);
734 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
735 mm->nr_ptes++;
736 spin_unlock(&mm->page_table_lock);
739 return 0;
742 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
744 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
747 static inline struct page *alloc_hugepage_vma(int defrag,
748 struct vm_area_struct *vma,
749 unsigned long haddr, int nd,
750 gfp_t extra_gfp)
752 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
753 HPAGE_PMD_ORDER, vma, haddr, nd);
756 #ifndef CONFIG_NUMA
757 static inline struct page *alloc_hugepage(int defrag)
759 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
760 HPAGE_PMD_ORDER);
762 #endif
764 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
765 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
766 struct page *zero_page)
768 pmd_t entry;
769 if (!pmd_none(*pmd))
770 return false;
771 entry = mk_pmd(zero_page, vma->vm_page_prot);
772 entry = pmd_wrprotect(entry);
773 entry = pmd_mkhuge(entry);
774 pgtable_trans_huge_deposit(mm, pmd, pgtable);
775 set_pmd_at(mm, haddr, pmd, entry);
776 mm->nr_ptes++;
777 return true;
780 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
781 unsigned long address, pmd_t *pmd,
782 unsigned int flags)
784 struct page *page;
785 unsigned long haddr = address & HPAGE_PMD_MASK;
786 pte_t *pte;
788 if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
789 if (unlikely(anon_vma_prepare(vma)))
790 return VM_FAULT_OOM;
791 if (unlikely(khugepaged_enter(vma)))
792 return VM_FAULT_OOM;
793 if (!(flags & FAULT_FLAG_WRITE) &&
794 transparent_hugepage_use_zero_page()) {
795 pgtable_t pgtable;
796 struct page *zero_page;
797 bool set;
798 pgtable = pte_alloc_one(mm, haddr);
799 if (unlikely(!pgtable))
800 return VM_FAULT_OOM;
801 zero_page = get_huge_zero_page();
802 if (unlikely(!zero_page)) {
803 pte_free(mm, pgtable);
804 count_vm_event(THP_FAULT_FALLBACK);
805 goto out;
807 spin_lock(&mm->page_table_lock);
808 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
809 zero_page);
810 spin_unlock(&mm->page_table_lock);
811 if (!set) {
812 pte_free(mm, pgtable);
813 put_huge_zero_page();
815 return 0;
817 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
818 vma, haddr, numa_node_id(), 0);
819 if (unlikely(!page)) {
820 count_vm_event(THP_FAULT_FALLBACK);
821 goto out;
823 count_vm_event(THP_FAULT_ALLOC);
824 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
825 put_page(page);
826 goto out;
828 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
829 page))) {
830 mem_cgroup_uncharge_page(page);
831 put_page(page);
832 goto out;
835 return 0;
837 out:
839 * Use __pte_alloc instead of pte_alloc_map, because we can't
840 * run pte_offset_map on the pmd, if an huge pmd could
841 * materialize from under us from a different thread.
843 if (unlikely(pmd_none(*pmd)) &&
844 unlikely(__pte_alloc(mm, vma, pmd, address)))
845 return VM_FAULT_OOM;
846 /* if an huge pmd materialized from under us just retry later */
847 if (unlikely(pmd_trans_huge(*pmd)))
848 return 0;
850 * A regular pmd is established and it can't morph into a huge pmd
851 * from under us anymore at this point because we hold the mmap_sem
852 * read mode and khugepaged takes it in write mode. So now it's
853 * safe to run pte_offset_map().
855 pte = pte_offset_map(pmd, address);
856 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
859 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
860 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
861 struct vm_area_struct *vma)
863 struct page *src_page;
864 pmd_t pmd;
865 pgtable_t pgtable;
866 int ret;
868 ret = -ENOMEM;
869 pgtable = pte_alloc_one(dst_mm, addr);
870 if (unlikely(!pgtable))
871 goto out;
873 spin_lock(&dst_mm->page_table_lock);
874 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
876 ret = -EAGAIN;
877 pmd = *src_pmd;
878 if (unlikely(!pmd_trans_huge(pmd))) {
879 pte_free(dst_mm, pgtable);
880 goto out_unlock;
883 * mm->page_table_lock is enough to be sure that huge zero pmd is not
884 * under splitting since we don't split the page itself, only pmd to
885 * a page table.
887 if (is_huge_zero_pmd(pmd)) {
888 struct page *zero_page;
889 bool set;
891 * get_huge_zero_page() will never allocate a new page here,
892 * since we already have a zero page to copy. It just takes a
893 * reference.
895 zero_page = get_huge_zero_page();
896 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
897 zero_page);
898 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
899 ret = 0;
900 goto out_unlock;
902 if (unlikely(pmd_trans_splitting(pmd))) {
903 /* split huge page running from under us */
904 spin_unlock(&src_mm->page_table_lock);
905 spin_unlock(&dst_mm->page_table_lock);
906 pte_free(dst_mm, pgtable);
908 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
909 goto out;
911 src_page = pmd_page(pmd);
912 VM_BUG_ON(!PageHead(src_page));
913 get_page(src_page);
914 page_dup_rmap(src_page);
915 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
917 pmdp_set_wrprotect(src_mm, addr, src_pmd);
918 pmd = pmd_mkold(pmd_wrprotect(pmd));
919 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
920 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
921 dst_mm->nr_ptes++;
923 ret = 0;
924 out_unlock:
925 spin_unlock(&src_mm->page_table_lock);
926 spin_unlock(&dst_mm->page_table_lock);
927 out:
928 return ret;
931 void huge_pmd_set_accessed(struct mm_struct *mm,
932 struct vm_area_struct *vma,
933 unsigned long address,
934 pmd_t *pmd, pmd_t orig_pmd,
935 int dirty)
937 pmd_t entry;
938 unsigned long haddr;
940 spin_lock(&mm->page_table_lock);
941 if (unlikely(!pmd_same(*pmd, orig_pmd)))
942 goto unlock;
944 entry = pmd_mkyoung(orig_pmd);
945 haddr = address & HPAGE_PMD_MASK;
946 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
947 update_mmu_cache_pmd(vma, address, pmd);
949 unlock:
950 spin_unlock(&mm->page_table_lock);
953 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
954 struct vm_area_struct *vma, unsigned long address,
955 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
957 pgtable_t pgtable;
958 pmd_t _pmd;
959 struct page *page;
960 int i, ret = 0;
961 unsigned long mmun_start; /* For mmu_notifiers */
962 unsigned long mmun_end; /* For mmu_notifiers */
964 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
965 if (!page) {
966 ret |= VM_FAULT_OOM;
967 goto out;
970 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
971 put_page(page);
972 ret |= VM_FAULT_OOM;
973 goto out;
976 clear_user_highpage(page, address);
977 __SetPageUptodate(page);
979 mmun_start = haddr;
980 mmun_end = haddr + HPAGE_PMD_SIZE;
981 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
983 spin_lock(&mm->page_table_lock);
984 if (unlikely(!pmd_same(*pmd, orig_pmd)))
985 goto out_free_page;
987 pmdp_clear_flush(vma, haddr, pmd);
988 /* leave pmd empty until pte is filled */
990 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
991 pmd_populate(mm, &_pmd, pgtable);
993 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
994 pte_t *pte, entry;
995 if (haddr == (address & PAGE_MASK)) {
996 entry = mk_pte(page, vma->vm_page_prot);
997 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
998 page_add_new_anon_rmap(page, vma, haddr);
999 } else {
1000 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1001 entry = pte_mkspecial(entry);
1003 pte = pte_offset_map(&_pmd, haddr);
1004 VM_BUG_ON(!pte_none(*pte));
1005 set_pte_at(mm, haddr, pte, entry);
1006 pte_unmap(pte);
1008 smp_wmb(); /* make pte visible before pmd */
1009 pmd_populate(mm, pmd, pgtable);
1010 spin_unlock(&mm->page_table_lock);
1011 put_huge_zero_page();
1012 inc_mm_counter(mm, MM_ANONPAGES);
1014 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1016 ret |= VM_FAULT_WRITE;
1017 out:
1018 return ret;
1019 out_free_page:
1020 spin_unlock(&mm->page_table_lock);
1021 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1022 mem_cgroup_uncharge_page(page);
1023 put_page(page);
1024 goto out;
1027 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1028 struct vm_area_struct *vma,
1029 unsigned long address,
1030 pmd_t *pmd, pmd_t orig_pmd,
1031 struct page *page,
1032 unsigned long haddr)
1034 pgtable_t pgtable;
1035 pmd_t _pmd;
1036 int ret = 0, i;
1037 struct page **pages;
1038 unsigned long mmun_start; /* For mmu_notifiers */
1039 unsigned long mmun_end; /* For mmu_notifiers */
1041 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1042 GFP_KERNEL);
1043 if (unlikely(!pages)) {
1044 ret |= VM_FAULT_OOM;
1045 goto out;
1048 for (i = 0; i < HPAGE_PMD_NR; i++) {
1049 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1050 __GFP_OTHER_NODE,
1051 vma, address, page_to_nid(page));
1052 if (unlikely(!pages[i] ||
1053 mem_cgroup_newpage_charge(pages[i], mm,
1054 GFP_KERNEL))) {
1055 if (pages[i])
1056 put_page(pages[i]);
1057 mem_cgroup_uncharge_start();
1058 while (--i >= 0) {
1059 mem_cgroup_uncharge_page(pages[i]);
1060 put_page(pages[i]);
1062 mem_cgroup_uncharge_end();
1063 kfree(pages);
1064 ret |= VM_FAULT_OOM;
1065 goto out;
1069 for (i = 0; i < HPAGE_PMD_NR; i++) {
1070 copy_user_highpage(pages[i], page + i,
1071 haddr + PAGE_SIZE * i, vma);
1072 __SetPageUptodate(pages[i]);
1073 cond_resched();
1076 mmun_start = haddr;
1077 mmun_end = haddr + HPAGE_PMD_SIZE;
1078 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1080 spin_lock(&mm->page_table_lock);
1081 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1082 goto out_free_pages;
1083 VM_BUG_ON(!PageHead(page));
1085 pmdp_clear_flush(vma, haddr, pmd);
1086 /* leave pmd empty until pte is filled */
1088 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1089 pmd_populate(mm, &_pmd, pgtable);
1091 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1092 pte_t *pte, entry;
1093 entry = mk_pte(pages[i], vma->vm_page_prot);
1094 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1095 page_add_new_anon_rmap(pages[i], vma, haddr);
1096 pte = pte_offset_map(&_pmd, haddr);
1097 VM_BUG_ON(!pte_none(*pte));
1098 set_pte_at(mm, haddr, pte, entry);
1099 pte_unmap(pte);
1101 kfree(pages);
1103 smp_wmb(); /* make pte visible before pmd */
1104 pmd_populate(mm, pmd, pgtable);
1105 page_remove_rmap(page);
1106 spin_unlock(&mm->page_table_lock);
1108 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1110 ret |= VM_FAULT_WRITE;
1111 put_page(page);
1113 out:
1114 return ret;
1116 out_free_pages:
1117 spin_unlock(&mm->page_table_lock);
1118 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1119 mem_cgroup_uncharge_start();
1120 for (i = 0; i < HPAGE_PMD_NR; i++) {
1121 mem_cgroup_uncharge_page(pages[i]);
1122 put_page(pages[i]);
1124 mem_cgroup_uncharge_end();
1125 kfree(pages);
1126 goto out;
1129 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1130 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1132 int ret = 0;
1133 struct page *page = NULL, *new_page;
1134 unsigned long haddr;
1135 unsigned long mmun_start; /* For mmu_notifiers */
1136 unsigned long mmun_end; /* For mmu_notifiers */
1138 VM_BUG_ON(!vma->anon_vma);
1139 haddr = address & HPAGE_PMD_MASK;
1140 if (is_huge_zero_pmd(orig_pmd))
1141 goto alloc;
1142 spin_lock(&mm->page_table_lock);
1143 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1144 goto out_unlock;
1146 page = pmd_page(orig_pmd);
1147 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1148 if (page_mapcount(page) == 1) {
1149 pmd_t entry;
1150 entry = pmd_mkyoung(orig_pmd);
1151 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1152 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
1153 update_mmu_cache_pmd(vma, address, pmd);
1154 ret |= VM_FAULT_WRITE;
1155 goto out_unlock;
1157 get_page(page);
1158 spin_unlock(&mm->page_table_lock);
1159 alloc:
1160 if (transparent_hugepage_enabled(vma) &&
1161 !transparent_hugepage_debug_cow())
1162 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1163 vma, haddr, numa_node_id(), 0);
1164 else
1165 new_page = NULL;
1167 if (unlikely(!new_page)) {
1168 count_vm_event(THP_FAULT_FALLBACK);
1169 if (is_huge_zero_pmd(orig_pmd)) {
1170 ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1171 address, pmd, orig_pmd, haddr);
1172 } else {
1173 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1174 pmd, orig_pmd, page, haddr);
1175 if (ret & VM_FAULT_OOM)
1176 split_huge_page(page);
1177 put_page(page);
1179 goto out;
1181 count_vm_event(THP_FAULT_ALLOC);
1183 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1184 put_page(new_page);
1185 if (page) {
1186 split_huge_page(page);
1187 put_page(page);
1189 ret |= VM_FAULT_OOM;
1190 goto out;
1193 if (is_huge_zero_pmd(orig_pmd))
1194 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1195 else
1196 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1197 __SetPageUptodate(new_page);
1199 mmun_start = haddr;
1200 mmun_end = haddr + HPAGE_PMD_SIZE;
1201 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1203 spin_lock(&mm->page_table_lock);
1204 if (page)
1205 put_page(page);
1206 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1207 spin_unlock(&mm->page_table_lock);
1208 mem_cgroup_uncharge_page(new_page);
1209 put_page(new_page);
1210 goto out_mn;
1211 } else {
1212 pmd_t entry;
1213 entry = mk_huge_pmd(new_page, vma);
1214 pmdp_clear_flush(vma, haddr, pmd);
1215 page_add_new_anon_rmap(new_page, vma, haddr);
1216 set_pmd_at(mm, haddr, pmd, entry);
1217 update_mmu_cache_pmd(vma, address, pmd);
1218 if (is_huge_zero_pmd(orig_pmd)) {
1219 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1220 put_huge_zero_page();
1221 } else {
1222 VM_BUG_ON(!PageHead(page));
1223 page_remove_rmap(page);
1224 put_page(page);
1226 ret |= VM_FAULT_WRITE;
1228 spin_unlock(&mm->page_table_lock);
1229 out_mn:
1230 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1231 out:
1232 return ret;
1233 out_unlock:
1234 spin_unlock(&mm->page_table_lock);
1235 return ret;
1238 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1239 unsigned long addr,
1240 pmd_t *pmd,
1241 unsigned int flags)
1243 struct mm_struct *mm = vma->vm_mm;
1244 struct page *page = NULL;
1246 assert_spin_locked(&mm->page_table_lock);
1248 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1249 goto out;
1251 /* Avoid dumping huge zero page */
1252 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1253 return ERR_PTR(-EFAULT);
1255 page = pmd_page(*pmd);
1256 VM_BUG_ON(!PageHead(page));
1257 if (flags & FOLL_TOUCH) {
1258 pmd_t _pmd;
1260 * We should set the dirty bit only for FOLL_WRITE but
1261 * for now the dirty bit in the pmd is meaningless.
1262 * And if the dirty bit will become meaningful and
1263 * we'll only set it with FOLL_WRITE, an atomic
1264 * set_bit will be required on the pmd to set the
1265 * young bit, instead of the current set_pmd_at.
1267 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1268 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1269 pmd, _pmd, 1))
1270 update_mmu_cache_pmd(vma, addr, pmd);
1272 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1273 if (page->mapping && trylock_page(page)) {
1274 lru_add_drain();
1275 if (page->mapping)
1276 mlock_vma_page(page);
1277 unlock_page(page);
1280 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1281 VM_BUG_ON(!PageCompound(page));
1282 if (flags & FOLL_GET)
1283 get_page_foll(page);
1285 out:
1286 return page;
1289 /* NUMA hinting page fault entry point for trans huge pmds */
1290 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1291 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1293 struct page *page;
1294 unsigned long haddr = addr & HPAGE_PMD_MASK;
1295 int target_nid;
1296 int current_nid = -1;
1297 bool migrated;
1299 spin_lock(&mm->page_table_lock);
1300 if (unlikely(!pmd_same(pmd, *pmdp)))
1301 goto out_unlock;
1303 page = pmd_page(pmd);
1304 get_page(page);
1305 current_nid = page_to_nid(page);
1306 count_vm_numa_event(NUMA_HINT_FAULTS);
1307 if (current_nid == numa_node_id())
1308 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1310 target_nid = mpol_misplaced(page, vma, haddr);
1311 if (target_nid == -1) {
1312 put_page(page);
1313 goto clear_pmdnuma;
1316 /* Acquire the page lock to serialise THP migrations */
1317 spin_unlock(&mm->page_table_lock);
1318 lock_page(page);
1320 /* Confirm the PTE did not while locked */
1321 spin_lock(&mm->page_table_lock);
1322 if (unlikely(!pmd_same(pmd, *pmdp))) {
1323 unlock_page(page);
1324 put_page(page);
1325 goto out_unlock;
1327 spin_unlock(&mm->page_table_lock);
1329 /* Migrate the THP to the requested node */
1330 migrated = migrate_misplaced_transhuge_page(mm, vma,
1331 pmdp, pmd, addr, page, target_nid);
1332 if (!migrated)
1333 goto check_same;
1335 task_numa_fault(target_nid, HPAGE_PMD_NR, true);
1336 return 0;
1338 check_same:
1339 spin_lock(&mm->page_table_lock);
1340 if (unlikely(!pmd_same(pmd, *pmdp)))
1341 goto out_unlock;
1342 clear_pmdnuma:
1343 pmd = pmd_mknonnuma(pmd);
1344 set_pmd_at(mm, haddr, pmdp, pmd);
1345 VM_BUG_ON(pmd_numa(*pmdp));
1346 update_mmu_cache_pmd(vma, addr, pmdp);
1347 out_unlock:
1348 spin_unlock(&mm->page_table_lock);
1349 if (current_nid != -1)
1350 task_numa_fault(current_nid, HPAGE_PMD_NR, false);
1351 return 0;
1354 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1355 pmd_t *pmd, unsigned long addr)
1357 int ret = 0;
1359 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1360 struct page *page;
1361 pgtable_t pgtable;
1362 pmd_t orig_pmd;
1364 * For architectures like ppc64 we look at deposited pgtable
1365 * when calling pmdp_get_and_clear. So do the
1366 * pgtable_trans_huge_withdraw after finishing pmdp related
1367 * operations.
1369 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1370 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1371 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1372 if (is_huge_zero_pmd(orig_pmd)) {
1373 tlb->mm->nr_ptes--;
1374 spin_unlock(&tlb->mm->page_table_lock);
1375 put_huge_zero_page();
1376 } else {
1377 page = pmd_page(orig_pmd);
1378 page_remove_rmap(page);
1379 VM_BUG_ON(page_mapcount(page) < 0);
1380 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1381 VM_BUG_ON(!PageHead(page));
1382 tlb->mm->nr_ptes--;
1383 spin_unlock(&tlb->mm->page_table_lock);
1384 tlb_remove_page(tlb, page);
1386 pte_free(tlb->mm, pgtable);
1387 ret = 1;
1389 return ret;
1392 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1393 unsigned long addr, unsigned long end,
1394 unsigned char *vec)
1396 int ret = 0;
1398 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1400 * All logical pages in the range are present
1401 * if backed by a huge page.
1403 spin_unlock(&vma->vm_mm->page_table_lock);
1404 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1405 ret = 1;
1408 return ret;
1411 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1412 unsigned long old_addr,
1413 unsigned long new_addr, unsigned long old_end,
1414 pmd_t *old_pmd, pmd_t *new_pmd)
1416 int ret = 0;
1417 pmd_t pmd;
1419 struct mm_struct *mm = vma->vm_mm;
1421 if ((old_addr & ~HPAGE_PMD_MASK) ||
1422 (new_addr & ~HPAGE_PMD_MASK) ||
1423 old_end - old_addr < HPAGE_PMD_SIZE ||
1424 (new_vma->vm_flags & VM_NOHUGEPAGE))
1425 goto out;
1428 * The destination pmd shouldn't be established, free_pgtables()
1429 * should have release it.
1431 if (WARN_ON(!pmd_none(*new_pmd))) {
1432 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1433 goto out;
1436 ret = __pmd_trans_huge_lock(old_pmd, vma);
1437 if (ret == 1) {
1438 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1439 VM_BUG_ON(!pmd_none(*new_pmd));
1440 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1441 spin_unlock(&mm->page_table_lock);
1443 out:
1444 return ret;
1447 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1448 unsigned long addr, pgprot_t newprot, int prot_numa)
1450 struct mm_struct *mm = vma->vm_mm;
1451 int ret = 0;
1453 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1454 pmd_t entry;
1455 entry = pmdp_get_and_clear(mm, addr, pmd);
1456 if (!prot_numa) {
1457 entry = pmd_modify(entry, newprot);
1458 BUG_ON(pmd_write(entry));
1459 } else {
1460 struct page *page = pmd_page(*pmd);
1462 /* only check non-shared pages */
1463 if (page_mapcount(page) == 1 &&
1464 !pmd_numa(*pmd)) {
1465 entry = pmd_mknuma(entry);
1468 set_pmd_at(mm, addr, pmd, entry);
1469 spin_unlock(&vma->vm_mm->page_table_lock);
1470 ret = 1;
1473 return ret;
1477 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1478 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1480 * Note that if it returns 1, this routine returns without unlocking page
1481 * table locks. So callers must unlock them.
1483 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1485 spin_lock(&vma->vm_mm->page_table_lock);
1486 if (likely(pmd_trans_huge(*pmd))) {
1487 if (unlikely(pmd_trans_splitting(*pmd))) {
1488 spin_unlock(&vma->vm_mm->page_table_lock);
1489 wait_split_huge_page(vma->anon_vma, pmd);
1490 return -1;
1491 } else {
1492 /* Thp mapped by 'pmd' is stable, so we can
1493 * handle it as it is. */
1494 return 1;
1497 spin_unlock(&vma->vm_mm->page_table_lock);
1498 return 0;
1501 pmd_t *page_check_address_pmd(struct page *page,
1502 struct mm_struct *mm,
1503 unsigned long address,
1504 enum page_check_address_pmd_flag flag)
1506 pmd_t *pmd, *ret = NULL;
1508 if (address & ~HPAGE_PMD_MASK)
1509 goto out;
1511 pmd = mm_find_pmd(mm, address);
1512 if (!pmd)
1513 goto out;
1514 if (pmd_none(*pmd))
1515 goto out;
1516 if (pmd_page(*pmd) != page)
1517 goto out;
1519 * split_vma() may create temporary aliased mappings. There is
1520 * no risk as long as all huge pmd are found and have their
1521 * splitting bit set before __split_huge_page_refcount
1522 * runs. Finding the same huge pmd more than once during the
1523 * same rmap walk is not a problem.
1525 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1526 pmd_trans_splitting(*pmd))
1527 goto out;
1528 if (pmd_trans_huge(*pmd)) {
1529 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1530 !pmd_trans_splitting(*pmd));
1531 ret = pmd;
1533 out:
1534 return ret;
1537 static int __split_huge_page_splitting(struct page *page,
1538 struct vm_area_struct *vma,
1539 unsigned long address)
1541 struct mm_struct *mm = vma->vm_mm;
1542 pmd_t *pmd;
1543 int ret = 0;
1544 /* For mmu_notifiers */
1545 const unsigned long mmun_start = address;
1546 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
1548 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1549 spin_lock(&mm->page_table_lock);
1550 pmd = page_check_address_pmd(page, mm, address,
1551 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1552 if (pmd) {
1554 * We can't temporarily set the pmd to null in order
1555 * to split it, the pmd must remain marked huge at all
1556 * times or the VM won't take the pmd_trans_huge paths
1557 * and it won't wait on the anon_vma->root->rwsem to
1558 * serialize against split_huge_page*.
1560 pmdp_splitting_flush(vma, address, pmd);
1561 ret = 1;
1563 spin_unlock(&mm->page_table_lock);
1564 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1566 return ret;
1569 static void __split_huge_page_refcount(struct page *page,
1570 struct list_head *list)
1572 int i;
1573 struct zone *zone = page_zone(page);
1574 struct lruvec *lruvec;
1575 int tail_count = 0;
1577 /* prevent PageLRU to go away from under us, and freeze lru stats */
1578 spin_lock_irq(&zone->lru_lock);
1579 lruvec = mem_cgroup_page_lruvec(page, zone);
1581 compound_lock(page);
1582 /* complete memcg works before add pages to LRU */
1583 mem_cgroup_split_huge_fixup(page);
1585 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1586 struct page *page_tail = page + i;
1588 /* tail_page->_mapcount cannot change */
1589 BUG_ON(page_mapcount(page_tail) < 0);
1590 tail_count += page_mapcount(page_tail);
1591 /* check for overflow */
1592 BUG_ON(tail_count < 0);
1593 BUG_ON(atomic_read(&page_tail->_count) != 0);
1595 * tail_page->_count is zero and not changing from
1596 * under us. But get_page_unless_zero() may be running
1597 * from under us on the tail_page. If we used
1598 * atomic_set() below instead of atomic_add(), we
1599 * would then run atomic_set() concurrently with
1600 * get_page_unless_zero(), and atomic_set() is
1601 * implemented in C not using locked ops. spin_unlock
1602 * on x86 sometime uses locked ops because of PPro
1603 * errata 66, 92, so unless somebody can guarantee
1604 * atomic_set() here would be safe on all archs (and
1605 * not only on x86), it's safer to use atomic_add().
1607 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1608 &page_tail->_count);
1610 /* after clearing PageTail the gup refcount can be released */
1611 smp_mb();
1614 * retain hwpoison flag of the poisoned tail page:
1615 * fix for the unsuitable process killed on Guest Machine(KVM)
1616 * by the memory-failure.
1618 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1619 page_tail->flags |= (page->flags &
1620 ((1L << PG_referenced) |
1621 (1L << PG_swapbacked) |
1622 (1L << PG_mlocked) |
1623 (1L << PG_uptodate)));
1624 page_tail->flags |= (1L << PG_dirty);
1626 /* clear PageTail before overwriting first_page */
1627 smp_wmb();
1630 * __split_huge_page_splitting() already set the
1631 * splitting bit in all pmd that could map this
1632 * hugepage, that will ensure no CPU can alter the
1633 * mapcount on the head page. The mapcount is only
1634 * accounted in the head page and it has to be
1635 * transferred to all tail pages in the below code. So
1636 * for this code to be safe, the split the mapcount
1637 * can't change. But that doesn't mean userland can't
1638 * keep changing and reading the page contents while
1639 * we transfer the mapcount, so the pmd splitting
1640 * status is achieved setting a reserved bit in the
1641 * pmd, not by clearing the present bit.
1643 page_tail->_mapcount = page->_mapcount;
1645 BUG_ON(page_tail->mapping);
1646 page_tail->mapping = page->mapping;
1648 page_tail->index = page->index + i;
1649 page_nid_xchg_last(page_tail, page_nid_last(page));
1651 BUG_ON(!PageAnon(page_tail));
1652 BUG_ON(!PageUptodate(page_tail));
1653 BUG_ON(!PageDirty(page_tail));
1654 BUG_ON(!PageSwapBacked(page_tail));
1656 lru_add_page_tail(page, page_tail, lruvec, list);
1658 atomic_sub(tail_count, &page->_count);
1659 BUG_ON(atomic_read(&page->_count) <= 0);
1661 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1662 __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1664 ClearPageCompound(page);
1665 compound_unlock(page);
1666 spin_unlock_irq(&zone->lru_lock);
1668 for (i = 1; i < HPAGE_PMD_NR; i++) {
1669 struct page *page_tail = page + i;
1670 BUG_ON(page_count(page_tail) <= 0);
1672 * Tail pages may be freed if there wasn't any mapping
1673 * like if add_to_swap() is running on a lru page that
1674 * had its mapping zapped. And freeing these pages
1675 * requires taking the lru_lock so we do the put_page
1676 * of the tail pages after the split is complete.
1678 put_page(page_tail);
1682 * Only the head page (now become a regular page) is required
1683 * to be pinned by the caller.
1685 BUG_ON(page_count(page) <= 0);
1688 static int __split_huge_page_map(struct page *page,
1689 struct vm_area_struct *vma,
1690 unsigned long address)
1692 struct mm_struct *mm = vma->vm_mm;
1693 pmd_t *pmd, _pmd;
1694 int ret = 0, i;
1695 pgtable_t pgtable;
1696 unsigned long haddr;
1698 spin_lock(&mm->page_table_lock);
1699 pmd = page_check_address_pmd(page, mm, address,
1700 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1701 if (pmd) {
1702 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1703 pmd_populate(mm, &_pmd, pgtable);
1705 haddr = address;
1706 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1707 pte_t *pte, entry;
1708 BUG_ON(PageCompound(page+i));
1709 entry = mk_pte(page + i, vma->vm_page_prot);
1710 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1711 if (!pmd_write(*pmd))
1712 entry = pte_wrprotect(entry);
1713 else
1714 BUG_ON(page_mapcount(page) != 1);
1715 if (!pmd_young(*pmd))
1716 entry = pte_mkold(entry);
1717 if (pmd_numa(*pmd))
1718 entry = pte_mknuma(entry);
1719 pte = pte_offset_map(&_pmd, haddr);
1720 BUG_ON(!pte_none(*pte));
1721 set_pte_at(mm, haddr, pte, entry);
1722 pte_unmap(pte);
1725 smp_wmb(); /* make pte visible before pmd */
1727 * Up to this point the pmd is present and huge and
1728 * userland has the whole access to the hugepage
1729 * during the split (which happens in place). If we
1730 * overwrite the pmd with the not-huge version
1731 * pointing to the pte here (which of course we could
1732 * if all CPUs were bug free), userland could trigger
1733 * a small page size TLB miss on the small sized TLB
1734 * while the hugepage TLB entry is still established
1735 * in the huge TLB. Some CPU doesn't like that. See
1736 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1737 * Erratum 383 on page 93. Intel should be safe but is
1738 * also warns that it's only safe if the permission
1739 * and cache attributes of the two entries loaded in
1740 * the two TLB is identical (which should be the case
1741 * here). But it is generally safer to never allow
1742 * small and huge TLB entries for the same virtual
1743 * address to be loaded simultaneously. So instead of
1744 * doing "pmd_populate(); flush_tlb_range();" we first
1745 * mark the current pmd notpresent (atomically because
1746 * here the pmd_trans_huge and pmd_trans_splitting
1747 * must remain set at all times on the pmd until the
1748 * split is complete for this pmd), then we flush the
1749 * SMP TLB and finally we write the non-huge version
1750 * of the pmd entry with pmd_populate.
1752 pmdp_invalidate(vma, address, pmd);
1753 pmd_populate(mm, pmd, pgtable);
1754 ret = 1;
1756 spin_unlock(&mm->page_table_lock);
1758 return ret;
1761 /* must be called with anon_vma->root->rwsem held */
1762 static void __split_huge_page(struct page *page,
1763 struct anon_vma *anon_vma,
1764 struct list_head *list)
1766 int mapcount, mapcount2;
1767 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1768 struct anon_vma_chain *avc;
1770 BUG_ON(!PageHead(page));
1771 BUG_ON(PageTail(page));
1773 mapcount = 0;
1774 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1775 struct vm_area_struct *vma = avc->vma;
1776 unsigned long addr = vma_address(page, vma);
1777 BUG_ON(is_vma_temporary_stack(vma));
1778 mapcount += __split_huge_page_splitting(page, vma, addr);
1781 * It is critical that new vmas are added to the tail of the
1782 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1783 * and establishes a child pmd before
1784 * __split_huge_page_splitting() freezes the parent pmd (so if
1785 * we fail to prevent copy_huge_pmd() from running until the
1786 * whole __split_huge_page() is complete), we will still see
1787 * the newly established pmd of the child later during the
1788 * walk, to be able to set it as pmd_trans_splitting too.
1790 if (mapcount != page_mapcount(page))
1791 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1792 mapcount, page_mapcount(page));
1793 BUG_ON(mapcount != page_mapcount(page));
1795 __split_huge_page_refcount(page, list);
1797 mapcount2 = 0;
1798 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1799 struct vm_area_struct *vma = avc->vma;
1800 unsigned long addr = vma_address(page, vma);
1801 BUG_ON(is_vma_temporary_stack(vma));
1802 mapcount2 += __split_huge_page_map(page, vma, addr);
1804 if (mapcount != mapcount2)
1805 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1806 mapcount, mapcount2, page_mapcount(page));
1807 BUG_ON(mapcount != mapcount2);
1811 * Split a hugepage into normal pages. This doesn't change the position of head
1812 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1813 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1814 * from the hugepage.
1815 * Return 0 if the hugepage is split successfully otherwise return 1.
1817 int split_huge_page_to_list(struct page *page, struct list_head *list)
1819 struct anon_vma *anon_vma;
1820 int ret = 1;
1822 BUG_ON(is_huge_zero_page(page));
1823 BUG_ON(!PageAnon(page));
1826 * The caller does not necessarily hold an mmap_sem that would prevent
1827 * the anon_vma disappearing so we first we take a reference to it
1828 * and then lock the anon_vma for write. This is similar to
1829 * page_lock_anon_vma_read except the write lock is taken to serialise
1830 * against parallel split or collapse operations.
1832 anon_vma = page_get_anon_vma(page);
1833 if (!anon_vma)
1834 goto out;
1835 anon_vma_lock_write(anon_vma);
1837 ret = 0;
1838 if (!PageCompound(page))
1839 goto out_unlock;
1841 BUG_ON(!PageSwapBacked(page));
1842 __split_huge_page(page, anon_vma, list);
1843 count_vm_event(THP_SPLIT);
1845 BUG_ON(PageCompound(page));
1846 out_unlock:
1847 anon_vma_unlock_write(anon_vma);
1848 put_anon_vma(anon_vma);
1849 out:
1850 return ret;
1853 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1855 int hugepage_madvise(struct vm_area_struct *vma,
1856 unsigned long *vm_flags, int advice)
1858 struct mm_struct *mm = vma->vm_mm;
1860 switch (advice) {
1861 case MADV_HUGEPAGE:
1863 * Be somewhat over-protective like KSM for now!
1865 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1866 return -EINVAL;
1867 if (mm->def_flags & VM_NOHUGEPAGE)
1868 return -EINVAL;
1869 *vm_flags &= ~VM_NOHUGEPAGE;
1870 *vm_flags |= VM_HUGEPAGE;
1872 * If the vma become good for khugepaged to scan,
1873 * register it here without waiting a page fault that
1874 * may not happen any time soon.
1876 if (unlikely(khugepaged_enter_vma_merge(vma)))
1877 return -ENOMEM;
1878 break;
1879 case MADV_NOHUGEPAGE:
1881 * Be somewhat over-protective like KSM for now!
1883 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1884 return -EINVAL;
1885 *vm_flags &= ~VM_HUGEPAGE;
1886 *vm_flags |= VM_NOHUGEPAGE;
1888 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1889 * this vma even if we leave the mm registered in khugepaged if
1890 * it got registered before VM_NOHUGEPAGE was set.
1892 break;
1895 return 0;
1898 static int __init khugepaged_slab_init(void)
1900 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1901 sizeof(struct mm_slot),
1902 __alignof__(struct mm_slot), 0, NULL);
1903 if (!mm_slot_cache)
1904 return -ENOMEM;
1906 return 0;
1909 static inline struct mm_slot *alloc_mm_slot(void)
1911 if (!mm_slot_cache) /* initialization failed */
1912 return NULL;
1913 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1916 static inline void free_mm_slot(struct mm_slot *mm_slot)
1918 kmem_cache_free(mm_slot_cache, mm_slot);
1921 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1923 struct mm_slot *mm_slot;
1925 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1926 if (mm == mm_slot->mm)
1927 return mm_slot;
1929 return NULL;
1932 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1933 struct mm_slot *mm_slot)
1935 mm_slot->mm = mm;
1936 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1939 static inline int khugepaged_test_exit(struct mm_struct *mm)
1941 return atomic_read(&mm->mm_users) == 0;
1944 int __khugepaged_enter(struct mm_struct *mm)
1946 struct mm_slot *mm_slot;
1947 int wakeup;
1949 mm_slot = alloc_mm_slot();
1950 if (!mm_slot)
1951 return -ENOMEM;
1953 /* __khugepaged_exit() must not run from under us */
1954 VM_BUG_ON(khugepaged_test_exit(mm));
1955 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1956 free_mm_slot(mm_slot);
1957 return 0;
1960 spin_lock(&khugepaged_mm_lock);
1961 insert_to_mm_slots_hash(mm, mm_slot);
1963 * Insert just behind the scanning cursor, to let the area settle
1964 * down a little.
1966 wakeup = list_empty(&khugepaged_scan.mm_head);
1967 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1968 spin_unlock(&khugepaged_mm_lock);
1970 atomic_inc(&mm->mm_count);
1971 if (wakeup)
1972 wake_up_interruptible(&khugepaged_wait);
1974 return 0;
1977 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1979 unsigned long hstart, hend;
1980 if (!vma->anon_vma)
1982 * Not yet faulted in so we will register later in the
1983 * page fault if needed.
1985 return 0;
1986 if (vma->vm_ops)
1987 /* khugepaged not yet working on file or special mappings */
1988 return 0;
1989 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1990 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1991 hend = vma->vm_end & HPAGE_PMD_MASK;
1992 if (hstart < hend)
1993 return khugepaged_enter(vma);
1994 return 0;
1997 void __khugepaged_exit(struct mm_struct *mm)
1999 struct mm_slot *mm_slot;
2000 int free = 0;
2002 spin_lock(&khugepaged_mm_lock);
2003 mm_slot = get_mm_slot(mm);
2004 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2005 hash_del(&mm_slot->hash);
2006 list_del(&mm_slot->mm_node);
2007 free = 1;
2009 spin_unlock(&khugepaged_mm_lock);
2011 if (free) {
2012 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2013 free_mm_slot(mm_slot);
2014 mmdrop(mm);
2015 } else if (mm_slot) {
2017 * This is required to serialize against
2018 * khugepaged_test_exit() (which is guaranteed to run
2019 * under mmap sem read mode). Stop here (after we
2020 * return all pagetables will be destroyed) until
2021 * khugepaged has finished working on the pagetables
2022 * under the mmap_sem.
2024 down_write(&mm->mmap_sem);
2025 up_write(&mm->mmap_sem);
2029 static void release_pte_page(struct page *page)
2031 /* 0 stands for page_is_file_cache(page) == false */
2032 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2033 unlock_page(page);
2034 putback_lru_page(page);
2037 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2039 while (--_pte >= pte) {
2040 pte_t pteval = *_pte;
2041 if (!pte_none(pteval))
2042 release_pte_page(pte_page(pteval));
2046 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2047 unsigned long address,
2048 pte_t *pte)
2050 struct page *page;
2051 pte_t *_pte;
2052 int referenced = 0, none = 0;
2053 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2054 _pte++, address += PAGE_SIZE) {
2055 pte_t pteval = *_pte;
2056 if (pte_none(pteval)) {
2057 if (++none <= khugepaged_max_ptes_none)
2058 continue;
2059 else
2060 goto out;
2062 if (!pte_present(pteval) || !pte_write(pteval))
2063 goto out;
2064 page = vm_normal_page(vma, address, pteval);
2065 if (unlikely(!page))
2066 goto out;
2068 VM_BUG_ON(PageCompound(page));
2069 BUG_ON(!PageAnon(page));
2070 VM_BUG_ON(!PageSwapBacked(page));
2072 /* cannot use mapcount: can't collapse if there's a gup pin */
2073 if (page_count(page) != 1)
2074 goto out;
2076 * We can do it before isolate_lru_page because the
2077 * page can't be freed from under us. NOTE: PG_lock
2078 * is needed to serialize against split_huge_page
2079 * when invoked from the VM.
2081 if (!trylock_page(page))
2082 goto out;
2084 * Isolate the page to avoid collapsing an hugepage
2085 * currently in use by the VM.
2087 if (isolate_lru_page(page)) {
2088 unlock_page(page);
2089 goto out;
2091 /* 0 stands for page_is_file_cache(page) == false */
2092 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2093 VM_BUG_ON(!PageLocked(page));
2094 VM_BUG_ON(PageLRU(page));
2096 /* If there is no mapped pte young don't collapse the page */
2097 if (pte_young(pteval) || PageReferenced(page) ||
2098 mmu_notifier_test_young(vma->vm_mm, address))
2099 referenced = 1;
2101 if (likely(referenced))
2102 return 1;
2103 out:
2104 release_pte_pages(pte, _pte);
2105 return 0;
2108 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2109 struct vm_area_struct *vma,
2110 unsigned long address,
2111 spinlock_t *ptl)
2113 pte_t *_pte;
2114 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2115 pte_t pteval = *_pte;
2116 struct page *src_page;
2118 if (pte_none(pteval)) {
2119 clear_user_highpage(page, address);
2120 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2121 } else {
2122 src_page = pte_page(pteval);
2123 copy_user_highpage(page, src_page, address, vma);
2124 VM_BUG_ON(page_mapcount(src_page) != 1);
2125 release_pte_page(src_page);
2127 * ptl mostly unnecessary, but preempt has to
2128 * be disabled to update the per-cpu stats
2129 * inside page_remove_rmap().
2131 spin_lock(ptl);
2133 * paravirt calls inside pte_clear here are
2134 * superfluous.
2136 pte_clear(vma->vm_mm, address, _pte);
2137 page_remove_rmap(src_page);
2138 spin_unlock(ptl);
2139 free_page_and_swap_cache(src_page);
2142 address += PAGE_SIZE;
2143 page++;
2147 static void khugepaged_alloc_sleep(void)
2149 wait_event_freezable_timeout(khugepaged_wait, false,
2150 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2153 #ifdef CONFIG_NUMA
2154 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2156 if (IS_ERR(*hpage)) {
2157 if (!*wait)
2158 return false;
2160 *wait = false;
2161 *hpage = NULL;
2162 khugepaged_alloc_sleep();
2163 } else if (*hpage) {
2164 put_page(*hpage);
2165 *hpage = NULL;
2168 return true;
2171 static struct page
2172 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2173 struct vm_area_struct *vma, unsigned long address,
2174 int node)
2176 VM_BUG_ON(*hpage);
2178 * Allocate the page while the vma is still valid and under
2179 * the mmap_sem read mode so there is no memory allocation
2180 * later when we take the mmap_sem in write mode. This is more
2181 * friendly behavior (OTOH it may actually hide bugs) to
2182 * filesystems in userland with daemons allocating memory in
2183 * the userland I/O paths. Allocating memory with the
2184 * mmap_sem in read mode is good idea also to allow greater
2185 * scalability.
2187 *hpage = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2188 node, __GFP_OTHER_NODE);
2191 * After allocating the hugepage, release the mmap_sem read lock in
2192 * preparation for taking it in write mode.
2194 up_read(&mm->mmap_sem);
2195 if (unlikely(!*hpage)) {
2196 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2197 *hpage = ERR_PTR(-ENOMEM);
2198 return NULL;
2201 count_vm_event(THP_COLLAPSE_ALLOC);
2202 return *hpage;
2204 #else
2205 static struct page *khugepaged_alloc_hugepage(bool *wait)
2207 struct page *hpage;
2209 do {
2210 hpage = alloc_hugepage(khugepaged_defrag());
2211 if (!hpage) {
2212 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2213 if (!*wait)
2214 return NULL;
2216 *wait = false;
2217 khugepaged_alloc_sleep();
2218 } else
2219 count_vm_event(THP_COLLAPSE_ALLOC);
2220 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2222 return hpage;
2225 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2227 if (!*hpage)
2228 *hpage = khugepaged_alloc_hugepage(wait);
2230 if (unlikely(!*hpage))
2231 return false;
2233 return true;
2236 static struct page
2237 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2238 struct vm_area_struct *vma, unsigned long address,
2239 int node)
2241 up_read(&mm->mmap_sem);
2242 VM_BUG_ON(!*hpage);
2243 return *hpage;
2245 #endif
2247 static bool hugepage_vma_check(struct vm_area_struct *vma)
2249 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2250 (vma->vm_flags & VM_NOHUGEPAGE))
2251 return false;
2253 if (!vma->anon_vma || vma->vm_ops)
2254 return false;
2255 if (is_vma_temporary_stack(vma))
2256 return false;
2257 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2258 return true;
2261 static void collapse_huge_page(struct mm_struct *mm,
2262 unsigned long address,
2263 struct page **hpage,
2264 struct vm_area_struct *vma,
2265 int node)
2267 pmd_t *pmd, _pmd;
2268 pte_t *pte;
2269 pgtable_t pgtable;
2270 struct page *new_page;
2271 spinlock_t *ptl;
2272 int isolated;
2273 unsigned long hstart, hend;
2274 unsigned long mmun_start; /* For mmu_notifiers */
2275 unsigned long mmun_end; /* For mmu_notifiers */
2277 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2279 /* release the mmap_sem read lock. */
2280 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2281 if (!new_page)
2282 return;
2284 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2285 return;
2288 * Prevent all access to pagetables with the exception of
2289 * gup_fast later hanlded by the ptep_clear_flush and the VM
2290 * handled by the anon_vma lock + PG_lock.
2292 down_write(&mm->mmap_sem);
2293 if (unlikely(khugepaged_test_exit(mm)))
2294 goto out;
2296 vma = find_vma(mm, address);
2297 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2298 hend = vma->vm_end & HPAGE_PMD_MASK;
2299 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2300 goto out;
2301 if (!hugepage_vma_check(vma))
2302 goto out;
2303 pmd = mm_find_pmd(mm, address);
2304 if (!pmd)
2305 goto out;
2306 if (pmd_trans_huge(*pmd))
2307 goto out;
2309 anon_vma_lock_write(vma->anon_vma);
2311 pte = pte_offset_map(pmd, address);
2312 ptl = pte_lockptr(mm, pmd);
2314 mmun_start = address;
2315 mmun_end = address + HPAGE_PMD_SIZE;
2316 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2317 spin_lock(&mm->page_table_lock); /* probably unnecessary */
2319 * After this gup_fast can't run anymore. This also removes
2320 * any huge TLB entry from the CPU so we won't allow
2321 * huge and small TLB entries for the same virtual address
2322 * to avoid the risk of CPU bugs in that area.
2324 _pmd = pmdp_clear_flush(vma, address, pmd);
2325 spin_unlock(&mm->page_table_lock);
2326 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2328 spin_lock(ptl);
2329 isolated = __collapse_huge_page_isolate(vma, address, pte);
2330 spin_unlock(ptl);
2332 if (unlikely(!isolated)) {
2333 pte_unmap(pte);
2334 spin_lock(&mm->page_table_lock);
2335 BUG_ON(!pmd_none(*pmd));
2337 * We can only use set_pmd_at when establishing
2338 * hugepmds and never for establishing regular pmds that
2339 * points to regular pagetables. Use pmd_populate for that
2341 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2342 spin_unlock(&mm->page_table_lock);
2343 anon_vma_unlock_write(vma->anon_vma);
2344 goto out;
2348 * All pages are isolated and locked so anon_vma rmap
2349 * can't run anymore.
2351 anon_vma_unlock_write(vma->anon_vma);
2353 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2354 pte_unmap(pte);
2355 __SetPageUptodate(new_page);
2356 pgtable = pmd_pgtable(_pmd);
2358 _pmd = mk_huge_pmd(new_page, vma);
2361 * spin_lock() below is not the equivalent of smp_wmb(), so
2362 * this is needed to avoid the copy_huge_page writes to become
2363 * visible after the set_pmd_at() write.
2365 smp_wmb();
2367 spin_lock(&mm->page_table_lock);
2368 BUG_ON(!pmd_none(*pmd));
2369 page_add_new_anon_rmap(new_page, vma, address);
2370 pgtable_trans_huge_deposit(mm, pmd, pgtable);
2371 set_pmd_at(mm, address, pmd, _pmd);
2372 update_mmu_cache_pmd(vma, address, pmd);
2373 spin_unlock(&mm->page_table_lock);
2375 *hpage = NULL;
2377 khugepaged_pages_collapsed++;
2378 out_up_write:
2379 up_write(&mm->mmap_sem);
2380 return;
2382 out:
2383 mem_cgroup_uncharge_page(new_page);
2384 goto out_up_write;
2387 static int khugepaged_scan_pmd(struct mm_struct *mm,
2388 struct vm_area_struct *vma,
2389 unsigned long address,
2390 struct page **hpage)
2392 pmd_t *pmd;
2393 pte_t *pte, *_pte;
2394 int ret = 0, referenced = 0, none = 0;
2395 struct page *page;
2396 unsigned long _address;
2397 spinlock_t *ptl;
2398 int node = NUMA_NO_NODE;
2400 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2402 pmd = mm_find_pmd(mm, address);
2403 if (!pmd)
2404 goto out;
2405 if (pmd_trans_huge(*pmd))
2406 goto out;
2408 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2409 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2410 _pte++, _address += PAGE_SIZE) {
2411 pte_t pteval = *_pte;
2412 if (pte_none(pteval)) {
2413 if (++none <= khugepaged_max_ptes_none)
2414 continue;
2415 else
2416 goto out_unmap;
2418 if (!pte_present(pteval) || !pte_write(pteval))
2419 goto out_unmap;
2420 page = vm_normal_page(vma, _address, pteval);
2421 if (unlikely(!page))
2422 goto out_unmap;
2424 * Chose the node of the first page. This could
2425 * be more sophisticated and look at more pages,
2426 * but isn't for now.
2428 if (node == NUMA_NO_NODE)
2429 node = page_to_nid(page);
2430 VM_BUG_ON(PageCompound(page));
2431 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2432 goto out_unmap;
2433 /* cannot use mapcount: can't collapse if there's a gup pin */
2434 if (page_count(page) != 1)
2435 goto out_unmap;
2436 if (pte_young(pteval) || PageReferenced(page) ||
2437 mmu_notifier_test_young(vma->vm_mm, address))
2438 referenced = 1;
2440 if (referenced)
2441 ret = 1;
2442 out_unmap:
2443 pte_unmap_unlock(pte, ptl);
2444 if (ret)
2445 /* collapse_huge_page will return with the mmap_sem released */
2446 collapse_huge_page(mm, address, hpage, vma, node);
2447 out:
2448 return ret;
2451 static void collect_mm_slot(struct mm_slot *mm_slot)
2453 struct mm_struct *mm = mm_slot->mm;
2455 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2457 if (khugepaged_test_exit(mm)) {
2458 /* free mm_slot */
2459 hash_del(&mm_slot->hash);
2460 list_del(&mm_slot->mm_node);
2463 * Not strictly needed because the mm exited already.
2465 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2468 /* khugepaged_mm_lock actually not necessary for the below */
2469 free_mm_slot(mm_slot);
2470 mmdrop(mm);
2474 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2475 struct page **hpage)
2476 __releases(&khugepaged_mm_lock)
2477 __acquires(&khugepaged_mm_lock)
2479 struct mm_slot *mm_slot;
2480 struct mm_struct *mm;
2481 struct vm_area_struct *vma;
2482 int progress = 0;
2484 VM_BUG_ON(!pages);
2485 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2487 if (khugepaged_scan.mm_slot)
2488 mm_slot = khugepaged_scan.mm_slot;
2489 else {
2490 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2491 struct mm_slot, mm_node);
2492 khugepaged_scan.address = 0;
2493 khugepaged_scan.mm_slot = mm_slot;
2495 spin_unlock(&khugepaged_mm_lock);
2497 mm = mm_slot->mm;
2498 down_read(&mm->mmap_sem);
2499 if (unlikely(khugepaged_test_exit(mm)))
2500 vma = NULL;
2501 else
2502 vma = find_vma(mm, khugepaged_scan.address);
2504 progress++;
2505 for (; vma; vma = vma->vm_next) {
2506 unsigned long hstart, hend;
2508 cond_resched();
2509 if (unlikely(khugepaged_test_exit(mm))) {
2510 progress++;
2511 break;
2513 if (!hugepage_vma_check(vma)) {
2514 skip:
2515 progress++;
2516 continue;
2518 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2519 hend = vma->vm_end & HPAGE_PMD_MASK;
2520 if (hstart >= hend)
2521 goto skip;
2522 if (khugepaged_scan.address > hend)
2523 goto skip;
2524 if (khugepaged_scan.address < hstart)
2525 khugepaged_scan.address = hstart;
2526 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2528 while (khugepaged_scan.address < hend) {
2529 int ret;
2530 cond_resched();
2531 if (unlikely(khugepaged_test_exit(mm)))
2532 goto breakouterloop;
2534 VM_BUG_ON(khugepaged_scan.address < hstart ||
2535 khugepaged_scan.address + HPAGE_PMD_SIZE >
2536 hend);
2537 ret = khugepaged_scan_pmd(mm, vma,
2538 khugepaged_scan.address,
2539 hpage);
2540 /* move to next address */
2541 khugepaged_scan.address += HPAGE_PMD_SIZE;
2542 progress += HPAGE_PMD_NR;
2543 if (ret)
2544 /* we released mmap_sem so break loop */
2545 goto breakouterloop_mmap_sem;
2546 if (progress >= pages)
2547 goto breakouterloop;
2550 breakouterloop:
2551 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2552 breakouterloop_mmap_sem:
2554 spin_lock(&khugepaged_mm_lock);
2555 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2557 * Release the current mm_slot if this mm is about to die, or
2558 * if we scanned all vmas of this mm.
2560 if (khugepaged_test_exit(mm) || !vma) {
2562 * Make sure that if mm_users is reaching zero while
2563 * khugepaged runs here, khugepaged_exit will find
2564 * mm_slot not pointing to the exiting mm.
2566 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2567 khugepaged_scan.mm_slot = list_entry(
2568 mm_slot->mm_node.next,
2569 struct mm_slot, mm_node);
2570 khugepaged_scan.address = 0;
2571 } else {
2572 khugepaged_scan.mm_slot = NULL;
2573 khugepaged_full_scans++;
2576 collect_mm_slot(mm_slot);
2579 return progress;
2582 static int khugepaged_has_work(void)
2584 return !list_empty(&khugepaged_scan.mm_head) &&
2585 khugepaged_enabled();
2588 static int khugepaged_wait_event(void)
2590 return !list_empty(&khugepaged_scan.mm_head) ||
2591 kthread_should_stop();
2594 static void khugepaged_do_scan(void)
2596 struct page *hpage = NULL;
2597 unsigned int progress = 0, pass_through_head = 0;
2598 unsigned int pages = khugepaged_pages_to_scan;
2599 bool wait = true;
2601 barrier(); /* write khugepaged_pages_to_scan to local stack */
2603 while (progress < pages) {
2604 if (!khugepaged_prealloc_page(&hpage, &wait))
2605 break;
2607 cond_resched();
2609 if (unlikely(kthread_should_stop() || freezing(current)))
2610 break;
2612 spin_lock(&khugepaged_mm_lock);
2613 if (!khugepaged_scan.mm_slot)
2614 pass_through_head++;
2615 if (khugepaged_has_work() &&
2616 pass_through_head < 2)
2617 progress += khugepaged_scan_mm_slot(pages - progress,
2618 &hpage);
2619 else
2620 progress = pages;
2621 spin_unlock(&khugepaged_mm_lock);
2624 if (!IS_ERR_OR_NULL(hpage))
2625 put_page(hpage);
2628 static void khugepaged_wait_work(void)
2630 try_to_freeze();
2632 if (khugepaged_has_work()) {
2633 if (!khugepaged_scan_sleep_millisecs)
2634 return;
2636 wait_event_freezable_timeout(khugepaged_wait,
2637 kthread_should_stop(),
2638 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2639 return;
2642 if (khugepaged_enabled())
2643 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2646 static int khugepaged(void *none)
2648 struct mm_slot *mm_slot;
2650 set_freezable();
2651 set_user_nice(current, 19);
2653 while (!kthread_should_stop()) {
2654 khugepaged_do_scan();
2655 khugepaged_wait_work();
2658 spin_lock(&khugepaged_mm_lock);
2659 mm_slot = khugepaged_scan.mm_slot;
2660 khugepaged_scan.mm_slot = NULL;
2661 if (mm_slot)
2662 collect_mm_slot(mm_slot);
2663 spin_unlock(&khugepaged_mm_lock);
2664 return 0;
2667 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2668 unsigned long haddr, pmd_t *pmd)
2670 struct mm_struct *mm = vma->vm_mm;
2671 pgtable_t pgtable;
2672 pmd_t _pmd;
2673 int i;
2675 pmdp_clear_flush(vma, haddr, pmd);
2676 /* leave pmd empty until pte is filled */
2678 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2679 pmd_populate(mm, &_pmd, pgtable);
2681 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2682 pte_t *pte, entry;
2683 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2684 entry = pte_mkspecial(entry);
2685 pte = pte_offset_map(&_pmd, haddr);
2686 VM_BUG_ON(!pte_none(*pte));
2687 set_pte_at(mm, haddr, pte, entry);
2688 pte_unmap(pte);
2690 smp_wmb(); /* make pte visible before pmd */
2691 pmd_populate(mm, pmd, pgtable);
2692 put_huge_zero_page();
2695 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2696 pmd_t *pmd)
2698 struct page *page;
2699 struct mm_struct *mm = vma->vm_mm;
2700 unsigned long haddr = address & HPAGE_PMD_MASK;
2701 unsigned long mmun_start; /* For mmu_notifiers */
2702 unsigned long mmun_end; /* For mmu_notifiers */
2704 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2706 mmun_start = haddr;
2707 mmun_end = haddr + HPAGE_PMD_SIZE;
2708 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2709 spin_lock(&mm->page_table_lock);
2710 if (unlikely(!pmd_trans_huge(*pmd))) {
2711 spin_unlock(&mm->page_table_lock);
2712 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2713 return;
2715 if (is_huge_zero_pmd(*pmd)) {
2716 __split_huge_zero_page_pmd(vma, haddr, pmd);
2717 spin_unlock(&mm->page_table_lock);
2718 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2719 return;
2721 page = pmd_page(*pmd);
2722 VM_BUG_ON(!page_count(page));
2723 get_page(page);
2724 spin_unlock(&mm->page_table_lock);
2725 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2727 split_huge_page(page);
2729 put_page(page);
2730 BUG_ON(pmd_trans_huge(*pmd));
2733 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2734 pmd_t *pmd)
2736 struct vm_area_struct *vma;
2738 vma = find_vma(mm, address);
2739 BUG_ON(vma == NULL);
2740 split_huge_page_pmd(vma, address, pmd);
2743 static void split_huge_page_address(struct mm_struct *mm,
2744 unsigned long address)
2746 pmd_t *pmd;
2748 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2750 pmd = mm_find_pmd(mm, address);
2751 if (!pmd)
2752 return;
2754 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2755 * materialize from under us.
2757 split_huge_page_pmd_mm(mm, address, pmd);
2760 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2761 unsigned long start,
2762 unsigned long end,
2763 long adjust_next)
2766 * If the new start address isn't hpage aligned and it could
2767 * previously contain an hugepage: check if we need to split
2768 * an huge pmd.
2770 if (start & ~HPAGE_PMD_MASK &&
2771 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2772 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2773 split_huge_page_address(vma->vm_mm, start);
2776 * If the new end address isn't hpage aligned and it could
2777 * previously contain an hugepage: check if we need to split
2778 * an huge pmd.
2780 if (end & ~HPAGE_PMD_MASK &&
2781 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2782 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2783 split_huge_page_address(vma->vm_mm, end);
2786 * If we're also updating the vma->vm_next->vm_start, if the new
2787 * vm_next->vm_start isn't page aligned and it could previously
2788 * contain an hugepage: check if we need to split an huge pmd.
2790 if (adjust_next > 0) {
2791 struct vm_area_struct *next = vma->vm_next;
2792 unsigned long nstart = next->vm_start;
2793 nstart += adjust_next << PAGE_SHIFT;
2794 if (nstart & ~HPAGE_PMD_MASK &&
2795 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2796 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2797 split_huge_page_address(next->vm_mm, nstart);