hpsa: fix bad -ENOMEM return value in hpsa_big_passthru_ioctl
[linux/fpc-iii.git] / mm / huge_memory.c
blob389973fd6bb782ac322b52069ec8e25b1e869d47
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 unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
215 struct shrink_control *sc)
217 /* we can free zero page only if last reference remains */
218 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
221 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
222 struct shrink_control *sc)
224 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
225 struct page *zero_page = xchg(&huge_zero_page, NULL);
226 BUG_ON(zero_page == NULL);
227 __free_page(zero_page);
228 return HPAGE_PMD_NR;
231 return 0;
234 static struct shrinker huge_zero_page_shrinker = {
235 .count_objects = shrink_huge_zero_page_count,
236 .scan_objects = shrink_huge_zero_page_scan,
237 .seeks = DEFAULT_SEEKS,
240 #ifdef CONFIG_SYSFS
242 static ssize_t double_flag_show(struct kobject *kobj,
243 struct kobj_attribute *attr, char *buf,
244 enum transparent_hugepage_flag enabled,
245 enum transparent_hugepage_flag req_madv)
247 if (test_bit(enabled, &transparent_hugepage_flags)) {
248 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
249 return sprintf(buf, "[always] madvise never\n");
250 } else if (test_bit(req_madv, &transparent_hugepage_flags))
251 return sprintf(buf, "always [madvise] never\n");
252 else
253 return sprintf(buf, "always madvise [never]\n");
255 static ssize_t double_flag_store(struct kobject *kobj,
256 struct kobj_attribute *attr,
257 const char *buf, size_t count,
258 enum transparent_hugepage_flag enabled,
259 enum transparent_hugepage_flag req_madv)
261 if (!memcmp("always", buf,
262 min(sizeof("always")-1, count))) {
263 set_bit(enabled, &transparent_hugepage_flags);
264 clear_bit(req_madv, &transparent_hugepage_flags);
265 } else if (!memcmp("madvise", buf,
266 min(sizeof("madvise")-1, count))) {
267 clear_bit(enabled, &transparent_hugepage_flags);
268 set_bit(req_madv, &transparent_hugepage_flags);
269 } else if (!memcmp("never", buf,
270 min(sizeof("never")-1, count))) {
271 clear_bit(enabled, &transparent_hugepage_flags);
272 clear_bit(req_madv, &transparent_hugepage_flags);
273 } else
274 return -EINVAL;
276 return count;
279 static ssize_t enabled_show(struct kobject *kobj,
280 struct kobj_attribute *attr, char *buf)
282 return double_flag_show(kobj, attr, buf,
283 TRANSPARENT_HUGEPAGE_FLAG,
284 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
286 static ssize_t enabled_store(struct kobject *kobj,
287 struct kobj_attribute *attr,
288 const char *buf, size_t count)
290 ssize_t ret;
292 ret = double_flag_store(kobj, attr, buf, count,
293 TRANSPARENT_HUGEPAGE_FLAG,
294 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
296 if (ret > 0) {
297 int err;
299 mutex_lock(&khugepaged_mutex);
300 err = start_khugepaged();
301 mutex_unlock(&khugepaged_mutex);
303 if (err)
304 ret = err;
307 return ret;
309 static struct kobj_attribute enabled_attr =
310 __ATTR(enabled, 0644, enabled_show, enabled_store);
312 static ssize_t single_flag_show(struct kobject *kobj,
313 struct kobj_attribute *attr, char *buf,
314 enum transparent_hugepage_flag flag)
316 return sprintf(buf, "%d\n",
317 !!test_bit(flag, &transparent_hugepage_flags));
320 static ssize_t single_flag_store(struct kobject *kobj,
321 struct kobj_attribute *attr,
322 const char *buf, size_t count,
323 enum transparent_hugepage_flag flag)
325 unsigned long value;
326 int ret;
328 ret = kstrtoul(buf, 10, &value);
329 if (ret < 0)
330 return ret;
331 if (value > 1)
332 return -EINVAL;
334 if (value)
335 set_bit(flag, &transparent_hugepage_flags);
336 else
337 clear_bit(flag, &transparent_hugepage_flags);
339 return count;
343 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
344 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
345 * memory just to allocate one more hugepage.
347 static ssize_t defrag_show(struct kobject *kobj,
348 struct kobj_attribute *attr, char *buf)
350 return double_flag_show(kobj, attr, buf,
351 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
352 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
354 static ssize_t defrag_store(struct kobject *kobj,
355 struct kobj_attribute *attr,
356 const char *buf, size_t count)
358 return double_flag_store(kobj, attr, buf, count,
359 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
360 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
362 static struct kobj_attribute defrag_attr =
363 __ATTR(defrag, 0644, defrag_show, defrag_store);
365 static ssize_t use_zero_page_show(struct kobject *kobj,
366 struct kobj_attribute *attr, char *buf)
368 return single_flag_show(kobj, attr, buf,
369 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
371 static ssize_t use_zero_page_store(struct kobject *kobj,
372 struct kobj_attribute *attr, const char *buf, size_t count)
374 return single_flag_store(kobj, attr, buf, count,
375 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
377 static struct kobj_attribute use_zero_page_attr =
378 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
379 #ifdef CONFIG_DEBUG_VM
380 static ssize_t debug_cow_show(struct kobject *kobj,
381 struct kobj_attribute *attr, char *buf)
383 return single_flag_show(kobj, attr, buf,
384 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
386 static ssize_t debug_cow_store(struct kobject *kobj,
387 struct kobj_attribute *attr,
388 const char *buf, size_t count)
390 return single_flag_store(kobj, attr, buf, count,
391 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
393 static struct kobj_attribute debug_cow_attr =
394 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
395 #endif /* CONFIG_DEBUG_VM */
397 static struct attribute *hugepage_attr[] = {
398 &enabled_attr.attr,
399 &defrag_attr.attr,
400 &use_zero_page_attr.attr,
401 #ifdef CONFIG_DEBUG_VM
402 &debug_cow_attr.attr,
403 #endif
404 NULL,
407 static struct attribute_group hugepage_attr_group = {
408 .attrs = hugepage_attr,
411 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
412 struct kobj_attribute *attr,
413 char *buf)
415 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
418 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
419 struct kobj_attribute *attr,
420 const char *buf, size_t count)
422 unsigned long msecs;
423 int err;
425 err = kstrtoul(buf, 10, &msecs);
426 if (err || msecs > UINT_MAX)
427 return -EINVAL;
429 khugepaged_scan_sleep_millisecs = msecs;
430 wake_up_interruptible(&khugepaged_wait);
432 return count;
434 static struct kobj_attribute scan_sleep_millisecs_attr =
435 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
436 scan_sleep_millisecs_store);
438 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
439 struct kobj_attribute *attr,
440 char *buf)
442 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
445 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
446 struct kobj_attribute *attr,
447 const char *buf, size_t count)
449 unsigned long msecs;
450 int err;
452 err = kstrtoul(buf, 10, &msecs);
453 if (err || msecs > UINT_MAX)
454 return -EINVAL;
456 khugepaged_alloc_sleep_millisecs = msecs;
457 wake_up_interruptible(&khugepaged_wait);
459 return count;
461 static struct kobj_attribute alloc_sleep_millisecs_attr =
462 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
463 alloc_sleep_millisecs_store);
465 static ssize_t pages_to_scan_show(struct kobject *kobj,
466 struct kobj_attribute *attr,
467 char *buf)
469 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
471 static ssize_t pages_to_scan_store(struct kobject *kobj,
472 struct kobj_attribute *attr,
473 const char *buf, size_t count)
475 int err;
476 unsigned long pages;
478 err = kstrtoul(buf, 10, &pages);
479 if (err || !pages || pages > UINT_MAX)
480 return -EINVAL;
482 khugepaged_pages_to_scan = pages;
484 return count;
486 static struct kobj_attribute pages_to_scan_attr =
487 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
488 pages_to_scan_store);
490 static ssize_t pages_collapsed_show(struct kobject *kobj,
491 struct kobj_attribute *attr,
492 char *buf)
494 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
496 static struct kobj_attribute pages_collapsed_attr =
497 __ATTR_RO(pages_collapsed);
499 static ssize_t full_scans_show(struct kobject *kobj,
500 struct kobj_attribute *attr,
501 char *buf)
503 return sprintf(buf, "%u\n", khugepaged_full_scans);
505 static struct kobj_attribute full_scans_attr =
506 __ATTR_RO(full_scans);
508 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
509 struct kobj_attribute *attr, char *buf)
511 return single_flag_show(kobj, attr, buf,
512 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
514 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
515 struct kobj_attribute *attr,
516 const char *buf, size_t count)
518 return single_flag_store(kobj, attr, buf, count,
519 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
521 static struct kobj_attribute khugepaged_defrag_attr =
522 __ATTR(defrag, 0644, khugepaged_defrag_show,
523 khugepaged_defrag_store);
526 * max_ptes_none controls if khugepaged should collapse hugepages over
527 * any unmapped ptes in turn potentially increasing the memory
528 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
529 * reduce the available free memory in the system as it
530 * runs. Increasing max_ptes_none will instead potentially reduce the
531 * free memory in the system during the khugepaged scan.
533 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
534 struct kobj_attribute *attr,
535 char *buf)
537 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
539 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
540 struct kobj_attribute *attr,
541 const char *buf, size_t count)
543 int err;
544 unsigned long max_ptes_none;
546 err = kstrtoul(buf, 10, &max_ptes_none);
547 if (err || max_ptes_none > HPAGE_PMD_NR-1)
548 return -EINVAL;
550 khugepaged_max_ptes_none = max_ptes_none;
552 return count;
554 static struct kobj_attribute khugepaged_max_ptes_none_attr =
555 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
556 khugepaged_max_ptes_none_store);
558 static struct attribute *khugepaged_attr[] = {
559 &khugepaged_defrag_attr.attr,
560 &khugepaged_max_ptes_none_attr.attr,
561 &pages_to_scan_attr.attr,
562 &pages_collapsed_attr.attr,
563 &full_scans_attr.attr,
564 &scan_sleep_millisecs_attr.attr,
565 &alloc_sleep_millisecs_attr.attr,
566 NULL,
569 static struct attribute_group khugepaged_attr_group = {
570 .attrs = khugepaged_attr,
571 .name = "khugepaged",
574 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
576 int err;
578 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
579 if (unlikely(!*hugepage_kobj)) {
580 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
581 return -ENOMEM;
584 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
585 if (err) {
586 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
587 goto delete_obj;
590 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
591 if (err) {
592 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
593 goto remove_hp_group;
596 return 0;
598 remove_hp_group:
599 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
600 delete_obj:
601 kobject_put(*hugepage_kobj);
602 return err;
605 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
607 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
608 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
609 kobject_put(hugepage_kobj);
611 #else
612 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
614 return 0;
617 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
620 #endif /* CONFIG_SYSFS */
622 static int __init hugepage_init(void)
624 int err;
625 struct kobject *hugepage_kobj;
627 if (!has_transparent_hugepage()) {
628 transparent_hugepage_flags = 0;
629 return -EINVAL;
632 err = hugepage_init_sysfs(&hugepage_kobj);
633 if (err)
634 return err;
636 err = khugepaged_slab_init();
637 if (err)
638 goto out;
640 register_shrinker(&huge_zero_page_shrinker);
643 * By default disable transparent hugepages on smaller systems,
644 * where the extra memory used could hurt more than TLB overhead
645 * is likely to save. The admin can still enable it through /sys.
647 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
648 transparent_hugepage_flags = 0;
650 start_khugepaged();
652 return 0;
653 out:
654 hugepage_exit_sysfs(hugepage_kobj);
655 return err;
657 module_init(hugepage_init)
659 static int __init setup_transparent_hugepage(char *str)
661 int ret = 0;
662 if (!str)
663 goto out;
664 if (!strcmp(str, "always")) {
665 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
666 &transparent_hugepage_flags);
667 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
668 &transparent_hugepage_flags);
669 ret = 1;
670 } else if (!strcmp(str, "madvise")) {
671 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
672 &transparent_hugepage_flags);
673 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
674 &transparent_hugepage_flags);
675 ret = 1;
676 } else if (!strcmp(str, "never")) {
677 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
678 &transparent_hugepage_flags);
679 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
680 &transparent_hugepage_flags);
681 ret = 1;
683 out:
684 if (!ret)
685 printk(KERN_WARNING
686 "transparent_hugepage= cannot parse, ignored\n");
687 return ret;
689 __setup("transparent_hugepage=", setup_transparent_hugepage);
691 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
693 if (likely(vma->vm_flags & VM_WRITE))
694 pmd = pmd_mkwrite(pmd);
695 return pmd;
698 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
700 pmd_t entry;
701 entry = mk_pmd(page, prot);
702 entry = pmd_mkhuge(entry);
703 return entry;
706 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
707 struct vm_area_struct *vma,
708 unsigned long haddr, pmd_t *pmd,
709 struct page *page)
711 pgtable_t pgtable;
713 VM_BUG_ON(!PageCompound(page));
714 pgtable = pte_alloc_one(mm, haddr);
715 if (unlikely(!pgtable))
716 return VM_FAULT_OOM;
718 clear_huge_page(page, haddr, HPAGE_PMD_NR);
720 * The memory barrier inside __SetPageUptodate makes sure that
721 * clear_huge_page writes become visible before the set_pmd_at()
722 * write.
724 __SetPageUptodate(page);
726 spin_lock(&mm->page_table_lock);
727 if (unlikely(!pmd_none(*pmd))) {
728 spin_unlock(&mm->page_table_lock);
729 mem_cgroup_uncharge_page(page);
730 put_page(page);
731 pte_free(mm, pgtable);
732 } else {
733 pmd_t entry;
734 entry = mk_huge_pmd(page, vma->vm_page_prot);
735 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
736 page_add_new_anon_rmap(page, vma, haddr);
737 pgtable_trans_huge_deposit(mm, pmd, pgtable);
738 set_pmd_at(mm, haddr, pmd, entry);
739 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
740 mm->nr_ptes++;
741 spin_unlock(&mm->page_table_lock);
744 return 0;
747 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
749 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
752 static inline struct page *alloc_hugepage_vma(int defrag,
753 struct vm_area_struct *vma,
754 unsigned long haddr, int nd,
755 gfp_t extra_gfp)
757 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
758 HPAGE_PMD_ORDER, vma, haddr, nd);
761 #ifndef CONFIG_NUMA
762 static inline struct page *alloc_hugepage(int defrag)
764 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
765 HPAGE_PMD_ORDER);
767 #endif
769 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
770 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
771 struct page *zero_page)
773 pmd_t entry;
774 if (!pmd_none(*pmd))
775 return false;
776 entry = mk_pmd(zero_page, vma->vm_page_prot);
777 entry = pmd_wrprotect(entry);
778 entry = pmd_mkhuge(entry);
779 pgtable_trans_huge_deposit(mm, pmd, pgtable);
780 set_pmd_at(mm, haddr, pmd, entry);
781 mm->nr_ptes++;
782 return true;
785 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
786 unsigned long address, pmd_t *pmd,
787 unsigned int flags)
789 struct page *page;
790 unsigned long haddr = address & HPAGE_PMD_MASK;
792 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
793 return VM_FAULT_FALLBACK;
794 if (unlikely(anon_vma_prepare(vma)))
795 return VM_FAULT_OOM;
796 if (unlikely(khugepaged_enter(vma)))
797 return VM_FAULT_OOM;
798 if (!(flags & FAULT_FLAG_WRITE) &&
799 transparent_hugepage_use_zero_page()) {
800 pgtable_t pgtable;
801 struct page *zero_page;
802 bool set;
803 pgtable = pte_alloc_one(mm, haddr);
804 if (unlikely(!pgtable))
805 return VM_FAULT_OOM;
806 zero_page = get_huge_zero_page();
807 if (unlikely(!zero_page)) {
808 pte_free(mm, pgtable);
809 count_vm_event(THP_FAULT_FALLBACK);
810 return VM_FAULT_FALLBACK;
812 spin_lock(&mm->page_table_lock);
813 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
814 zero_page);
815 spin_unlock(&mm->page_table_lock);
816 if (!set) {
817 pte_free(mm, pgtable);
818 put_huge_zero_page();
820 return 0;
822 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
823 vma, haddr, numa_node_id(), 0);
824 if (unlikely(!page)) {
825 count_vm_event(THP_FAULT_FALLBACK);
826 return VM_FAULT_FALLBACK;
828 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
829 put_page(page);
830 count_vm_event(THP_FAULT_FALLBACK);
831 return VM_FAULT_FALLBACK;
833 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
834 mem_cgroup_uncharge_page(page);
835 put_page(page);
836 count_vm_event(THP_FAULT_FALLBACK);
837 return VM_FAULT_FALLBACK;
840 count_vm_event(THP_FAULT_ALLOC);
841 return 0;
844 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
845 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
846 struct vm_area_struct *vma)
848 struct page *src_page;
849 pmd_t pmd;
850 pgtable_t pgtable;
851 int ret;
853 ret = -ENOMEM;
854 pgtable = pte_alloc_one(dst_mm, addr);
855 if (unlikely(!pgtable))
856 goto out;
858 spin_lock(&dst_mm->page_table_lock);
859 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
861 ret = -EAGAIN;
862 pmd = *src_pmd;
863 if (unlikely(!pmd_trans_huge(pmd))) {
864 pte_free(dst_mm, pgtable);
865 goto out_unlock;
868 * mm->page_table_lock is enough to be sure that huge zero pmd is not
869 * under splitting since we don't split the page itself, only pmd to
870 * a page table.
872 if (is_huge_zero_pmd(pmd)) {
873 struct page *zero_page;
874 bool set;
876 * get_huge_zero_page() will never allocate a new page here,
877 * since we already have a zero page to copy. It just takes a
878 * reference.
880 zero_page = get_huge_zero_page();
881 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
882 zero_page);
883 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
884 ret = 0;
885 goto out_unlock;
888 /* mmap_sem prevents this happening but warn if that changes */
889 WARN_ON(pmd_trans_migrating(pmd));
891 if (unlikely(pmd_trans_splitting(pmd))) {
892 /* split huge page running from under us */
893 spin_unlock(&src_mm->page_table_lock);
894 spin_unlock(&dst_mm->page_table_lock);
895 pte_free(dst_mm, pgtable);
897 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
898 goto out;
900 src_page = pmd_page(pmd);
901 VM_BUG_ON(!PageHead(src_page));
902 get_page(src_page);
903 page_dup_rmap(src_page);
904 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
906 pmdp_set_wrprotect(src_mm, addr, src_pmd);
907 pmd = pmd_mkold(pmd_wrprotect(pmd));
908 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
909 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
910 dst_mm->nr_ptes++;
912 ret = 0;
913 out_unlock:
914 spin_unlock(&src_mm->page_table_lock);
915 spin_unlock(&dst_mm->page_table_lock);
916 out:
917 return ret;
920 void huge_pmd_set_accessed(struct mm_struct *mm,
921 struct vm_area_struct *vma,
922 unsigned long address,
923 pmd_t *pmd, pmd_t orig_pmd,
924 int dirty)
926 pmd_t entry;
927 unsigned long haddr;
929 spin_lock(&mm->page_table_lock);
930 if (unlikely(!pmd_same(*pmd, orig_pmd)))
931 goto unlock;
933 entry = pmd_mkyoung(orig_pmd);
934 haddr = address & HPAGE_PMD_MASK;
935 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
936 update_mmu_cache_pmd(vma, address, pmd);
938 unlock:
939 spin_unlock(&mm->page_table_lock);
942 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
943 struct vm_area_struct *vma, unsigned long address,
944 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
946 pgtable_t pgtable;
947 pmd_t _pmd;
948 struct page *page;
949 int i, ret = 0;
950 unsigned long mmun_start; /* For mmu_notifiers */
951 unsigned long mmun_end; /* For mmu_notifiers */
953 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
954 if (!page) {
955 ret |= VM_FAULT_OOM;
956 goto out;
959 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
960 put_page(page);
961 ret |= VM_FAULT_OOM;
962 goto out;
965 clear_user_highpage(page, address);
966 __SetPageUptodate(page);
968 mmun_start = haddr;
969 mmun_end = haddr + HPAGE_PMD_SIZE;
970 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
972 spin_lock(&mm->page_table_lock);
973 if (unlikely(!pmd_same(*pmd, orig_pmd)))
974 goto out_free_page;
976 pmdp_clear_flush(vma, haddr, pmd);
977 /* leave pmd empty until pte is filled */
979 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
980 pmd_populate(mm, &_pmd, pgtable);
982 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
983 pte_t *pte, entry;
984 if (haddr == (address & PAGE_MASK)) {
985 entry = mk_pte(page, vma->vm_page_prot);
986 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
987 page_add_new_anon_rmap(page, vma, haddr);
988 } else {
989 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
990 entry = pte_mkspecial(entry);
992 pte = pte_offset_map(&_pmd, haddr);
993 VM_BUG_ON(!pte_none(*pte));
994 set_pte_at(mm, haddr, pte, entry);
995 pte_unmap(pte);
997 smp_wmb(); /* make pte visible before pmd */
998 pmd_populate(mm, pmd, pgtable);
999 spin_unlock(&mm->page_table_lock);
1000 put_huge_zero_page();
1001 inc_mm_counter(mm, MM_ANONPAGES);
1003 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1005 ret |= VM_FAULT_WRITE;
1006 out:
1007 return ret;
1008 out_free_page:
1009 spin_unlock(&mm->page_table_lock);
1010 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1011 mem_cgroup_uncharge_page(page);
1012 put_page(page);
1013 goto out;
1016 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1017 struct vm_area_struct *vma,
1018 unsigned long address,
1019 pmd_t *pmd, pmd_t orig_pmd,
1020 struct page *page,
1021 unsigned long haddr)
1023 pgtable_t pgtable;
1024 pmd_t _pmd;
1025 int ret = 0, i;
1026 struct page **pages;
1027 unsigned long mmun_start; /* For mmu_notifiers */
1028 unsigned long mmun_end; /* For mmu_notifiers */
1030 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1031 GFP_KERNEL);
1032 if (unlikely(!pages)) {
1033 ret |= VM_FAULT_OOM;
1034 goto out;
1037 for (i = 0; i < HPAGE_PMD_NR; i++) {
1038 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1039 __GFP_OTHER_NODE,
1040 vma, address, page_to_nid(page));
1041 if (unlikely(!pages[i] ||
1042 mem_cgroup_newpage_charge(pages[i], mm,
1043 GFP_KERNEL))) {
1044 if (pages[i])
1045 put_page(pages[i]);
1046 mem_cgroup_uncharge_start();
1047 while (--i >= 0) {
1048 mem_cgroup_uncharge_page(pages[i]);
1049 put_page(pages[i]);
1051 mem_cgroup_uncharge_end();
1052 kfree(pages);
1053 ret |= VM_FAULT_OOM;
1054 goto out;
1058 for (i = 0; i < HPAGE_PMD_NR; i++) {
1059 copy_user_highpage(pages[i], page + i,
1060 haddr + PAGE_SIZE * i, vma);
1061 __SetPageUptodate(pages[i]);
1062 cond_resched();
1065 mmun_start = haddr;
1066 mmun_end = haddr + HPAGE_PMD_SIZE;
1067 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1069 spin_lock(&mm->page_table_lock);
1070 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1071 goto out_free_pages;
1072 VM_BUG_ON(!PageHead(page));
1074 pmdp_clear_flush(vma, haddr, pmd);
1075 /* leave pmd empty until pte is filled */
1077 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1078 pmd_populate(mm, &_pmd, pgtable);
1080 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1081 pte_t *pte, entry;
1082 entry = mk_pte(pages[i], vma->vm_page_prot);
1083 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1084 page_add_new_anon_rmap(pages[i], vma, haddr);
1085 pte = pte_offset_map(&_pmd, haddr);
1086 VM_BUG_ON(!pte_none(*pte));
1087 set_pte_at(mm, haddr, pte, entry);
1088 pte_unmap(pte);
1090 kfree(pages);
1092 smp_wmb(); /* make pte visible before pmd */
1093 pmd_populate(mm, pmd, pgtable);
1094 page_remove_rmap(page);
1095 spin_unlock(&mm->page_table_lock);
1097 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1099 ret |= VM_FAULT_WRITE;
1100 put_page(page);
1102 out:
1103 return ret;
1105 out_free_pages:
1106 spin_unlock(&mm->page_table_lock);
1107 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1108 mem_cgroup_uncharge_start();
1109 for (i = 0; i < HPAGE_PMD_NR; i++) {
1110 mem_cgroup_uncharge_page(pages[i]);
1111 put_page(pages[i]);
1113 mem_cgroup_uncharge_end();
1114 kfree(pages);
1115 goto out;
1118 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1119 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1121 int ret = 0;
1122 struct page *page = NULL, *new_page;
1123 unsigned long haddr;
1124 unsigned long mmun_start; /* For mmu_notifiers */
1125 unsigned long mmun_end; /* For mmu_notifiers */
1127 VM_BUG_ON(!vma->anon_vma);
1128 haddr = address & HPAGE_PMD_MASK;
1129 if (is_huge_zero_pmd(orig_pmd))
1130 goto alloc;
1131 spin_lock(&mm->page_table_lock);
1132 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1133 goto out_unlock;
1135 page = pmd_page(orig_pmd);
1136 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1137 if (page_mapcount(page) == 1) {
1138 pmd_t entry;
1139 entry = pmd_mkyoung(orig_pmd);
1140 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1141 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
1142 update_mmu_cache_pmd(vma, address, pmd);
1143 ret |= VM_FAULT_WRITE;
1144 goto out_unlock;
1146 get_page(page);
1147 spin_unlock(&mm->page_table_lock);
1148 alloc:
1149 if (transparent_hugepage_enabled(vma) &&
1150 !transparent_hugepage_debug_cow())
1151 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1152 vma, haddr, numa_node_id(), 0);
1153 else
1154 new_page = NULL;
1156 if (unlikely(!new_page)) {
1157 if (!page) {
1158 ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1159 address, pmd, orig_pmd, haddr);
1160 } else {
1161 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1162 pmd, orig_pmd, page, haddr);
1163 if (ret & VM_FAULT_OOM) {
1164 split_huge_page(page);
1165 ret |= VM_FAULT_FALLBACK;
1167 put_page(page);
1169 count_vm_event(THP_FAULT_FALLBACK);
1170 goto out;
1173 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1174 put_page(new_page);
1175 if (page) {
1176 split_huge_page(page);
1177 put_page(page);
1178 } else
1179 split_huge_page_pmd(vma, address, pmd);
1180 ret |= VM_FAULT_FALLBACK;
1181 count_vm_event(THP_FAULT_FALLBACK);
1182 goto out;
1185 count_vm_event(THP_FAULT_ALLOC);
1187 if (!page)
1188 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1189 else
1190 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1191 __SetPageUptodate(new_page);
1193 mmun_start = haddr;
1194 mmun_end = haddr + HPAGE_PMD_SIZE;
1195 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1197 spin_lock(&mm->page_table_lock);
1198 if (page)
1199 put_page(page);
1200 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1201 spin_unlock(&mm->page_table_lock);
1202 mem_cgroup_uncharge_page(new_page);
1203 put_page(new_page);
1204 goto out_mn;
1205 } else {
1206 pmd_t entry;
1207 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1208 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1209 pmdp_clear_flush(vma, haddr, pmd);
1210 page_add_new_anon_rmap(new_page, vma, haddr);
1211 set_pmd_at(mm, haddr, pmd, entry);
1212 update_mmu_cache_pmd(vma, address, pmd);
1213 if (!page) {
1214 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1215 put_huge_zero_page();
1216 } else {
1217 VM_BUG_ON(!PageHead(page));
1218 page_remove_rmap(page);
1219 put_page(page);
1221 ret |= VM_FAULT_WRITE;
1223 spin_unlock(&mm->page_table_lock);
1224 out_mn:
1225 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1226 out:
1227 return ret;
1228 out_unlock:
1229 spin_unlock(&mm->page_table_lock);
1230 return ret;
1233 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1234 unsigned long addr,
1235 pmd_t *pmd,
1236 unsigned int flags)
1238 struct mm_struct *mm = vma->vm_mm;
1239 struct page *page = NULL;
1241 assert_spin_locked(&mm->page_table_lock);
1243 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1244 goto out;
1246 /* Avoid dumping huge zero page */
1247 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1248 return ERR_PTR(-EFAULT);
1250 /* Full NUMA hinting faults to serialise migration in fault paths */
1251 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1252 goto out;
1254 page = pmd_page(*pmd);
1255 VM_BUG_ON(!PageHead(page));
1256 if (flags & FOLL_TOUCH) {
1257 pmd_t _pmd;
1259 * We should set the dirty bit only for FOLL_WRITE but
1260 * for now the dirty bit in the pmd is meaningless.
1261 * And if the dirty bit will become meaningful and
1262 * we'll only set it with FOLL_WRITE, an atomic
1263 * set_bit will be required on the pmd to set the
1264 * young bit, instead of the current set_pmd_at.
1266 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1267 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1268 pmd, _pmd, 1))
1269 update_mmu_cache_pmd(vma, addr, pmd);
1271 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1272 if (page->mapping && trylock_page(page)) {
1273 lru_add_drain();
1274 if (page->mapping)
1275 mlock_vma_page(page);
1276 unlock_page(page);
1279 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1280 VM_BUG_ON(!PageCompound(page));
1281 if (flags & FOLL_GET)
1282 get_page_foll(page);
1284 out:
1285 return page;
1288 /* NUMA hinting page fault entry point for trans huge pmds */
1289 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1290 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1292 struct anon_vma *anon_vma = NULL;
1293 struct page *page;
1294 unsigned long haddr = addr & HPAGE_PMD_MASK;
1295 int page_nid = -1, this_nid = numa_node_id();
1296 int target_nid;
1297 bool page_locked;
1298 bool migrated = false;
1300 spin_lock(&mm->page_table_lock);
1301 if (unlikely(!pmd_same(pmd, *pmdp)))
1302 goto out_unlock;
1305 * If there are potential migrations, wait for completion and retry
1306 * without disrupting NUMA hinting information. Do not relock and
1307 * check_same as the page may no longer be mapped.
1309 if (unlikely(pmd_trans_migrating(*pmdp))) {
1310 spin_unlock(&mm->page_table_lock);
1311 wait_migrate_huge_page(vma->anon_vma, pmdp);
1312 goto out;
1315 page = pmd_page(pmd);
1316 page_nid = page_to_nid(page);
1317 count_vm_numa_event(NUMA_HINT_FAULTS);
1318 if (page_nid == this_nid)
1319 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1322 * Acquire the page lock to serialise THP migrations but avoid dropping
1323 * page_table_lock if at all possible
1325 page_locked = trylock_page(page);
1326 target_nid = mpol_misplaced(page, vma, haddr);
1327 if (target_nid == -1) {
1328 /* If the page was locked, there are no parallel migrations */
1329 if (page_locked)
1330 goto clear_pmdnuma;
1333 /* Migration could have started since the pmd_trans_migrating check */
1334 if (!page_locked) {
1335 spin_unlock(&mm->page_table_lock);
1336 wait_on_page_locked(page);
1337 page_nid = -1;
1338 goto out;
1342 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1343 * to serialises splits
1345 get_page(page);
1346 spin_unlock(&mm->page_table_lock);
1347 anon_vma = page_lock_anon_vma_read(page);
1349 /* Confirm the PTE did not while locked */
1350 spin_lock(&mm->page_table_lock);
1351 if (unlikely(!pmd_same(pmd, *pmdp))) {
1352 unlock_page(page);
1353 put_page(page);
1354 page_nid = -1;
1355 goto out_unlock;
1358 /* Bail if we fail to protect against THP splits for any reason */
1359 if (unlikely(!anon_vma)) {
1360 put_page(page);
1361 page_nid = -1;
1362 goto clear_pmdnuma;
1366 * Migrate the THP to the requested node, returns with page unlocked
1367 * and pmd_numa cleared.
1369 spin_unlock(&mm->page_table_lock);
1370 migrated = migrate_misplaced_transhuge_page(mm, vma,
1371 pmdp, pmd, addr, page, target_nid);
1372 if (migrated)
1373 page_nid = target_nid;
1375 goto out;
1376 clear_pmdnuma:
1377 BUG_ON(!PageLocked(page));
1378 pmd = pmd_mknonnuma(pmd);
1379 set_pmd_at(mm, haddr, pmdp, pmd);
1380 VM_BUG_ON(pmd_numa(*pmdp));
1381 update_mmu_cache_pmd(vma, addr, pmdp);
1382 unlock_page(page);
1383 out_unlock:
1384 spin_unlock(&mm->page_table_lock);
1386 out:
1387 if (anon_vma)
1388 page_unlock_anon_vma_read(anon_vma);
1390 if (page_nid != -1)
1391 task_numa_fault(page_nid, HPAGE_PMD_NR, migrated);
1393 return 0;
1396 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1397 pmd_t *pmd, unsigned long addr)
1399 int ret = 0;
1401 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1402 struct page *page;
1403 pgtable_t pgtable;
1404 pmd_t orig_pmd;
1406 * For architectures like ppc64 we look at deposited pgtable
1407 * when calling pmdp_get_and_clear. So do the
1408 * pgtable_trans_huge_withdraw after finishing pmdp related
1409 * operations.
1411 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1412 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1413 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1414 if (is_huge_zero_pmd(orig_pmd)) {
1415 tlb->mm->nr_ptes--;
1416 spin_unlock(&tlb->mm->page_table_lock);
1417 put_huge_zero_page();
1418 } else {
1419 page = pmd_page(orig_pmd);
1420 page_remove_rmap(page);
1421 VM_BUG_ON(page_mapcount(page) < 0);
1422 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1423 VM_BUG_ON(!PageHead(page));
1424 tlb->mm->nr_ptes--;
1425 spin_unlock(&tlb->mm->page_table_lock);
1426 tlb_remove_page(tlb, page);
1428 pte_free(tlb->mm, pgtable);
1429 ret = 1;
1431 return ret;
1434 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1435 unsigned long addr, unsigned long end,
1436 unsigned char *vec)
1438 int ret = 0;
1440 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1442 * All logical pages in the range are present
1443 * if backed by a huge page.
1445 spin_unlock(&vma->vm_mm->page_table_lock);
1446 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1447 ret = 1;
1450 return ret;
1453 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1454 unsigned long old_addr,
1455 unsigned long new_addr, unsigned long old_end,
1456 pmd_t *old_pmd, pmd_t *new_pmd)
1458 int ret = 0;
1459 pmd_t pmd;
1461 struct mm_struct *mm = vma->vm_mm;
1463 if ((old_addr & ~HPAGE_PMD_MASK) ||
1464 (new_addr & ~HPAGE_PMD_MASK) ||
1465 old_end - old_addr < HPAGE_PMD_SIZE ||
1466 (new_vma->vm_flags & VM_NOHUGEPAGE))
1467 goto out;
1470 * The destination pmd shouldn't be established, free_pgtables()
1471 * should have release it.
1473 if (WARN_ON(!pmd_none(*new_pmd))) {
1474 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1475 goto out;
1478 ret = __pmd_trans_huge_lock(old_pmd, vma);
1479 if (ret == 1) {
1480 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1481 VM_BUG_ON(!pmd_none(*new_pmd));
1482 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1483 spin_unlock(&mm->page_table_lock);
1485 out:
1486 return ret;
1489 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1490 unsigned long addr, pgprot_t newprot, int prot_numa)
1492 struct mm_struct *mm = vma->vm_mm;
1493 int ret = 0;
1495 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1496 pmd_t entry;
1497 if (!prot_numa) {
1498 entry = pmdp_get_and_clear(mm, addr, pmd);
1499 if (pmd_numa(entry))
1500 entry = pmd_mknonnuma(entry);
1501 entry = pmd_modify(entry, newprot);
1502 BUG_ON(pmd_write(entry));
1503 set_pmd_at(mm, addr, pmd, entry);
1504 } else {
1505 struct page *page = pmd_page(*pmd);
1506 entry = *pmd;
1508 /* only check non-shared pages */
1509 if (page_mapcount(page) == 1 &&
1510 !pmd_numa(*pmd)) {
1511 entry = pmd_mknuma(entry);
1512 set_pmd_at(mm, addr, pmd, entry);
1515 spin_unlock(&vma->vm_mm->page_table_lock);
1516 ret = 1;
1519 return ret;
1523 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1524 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1526 * Note that if it returns 1, this routine returns without unlocking page
1527 * table locks. So callers must unlock them.
1529 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1531 spin_lock(&vma->vm_mm->page_table_lock);
1532 if (likely(pmd_trans_huge(*pmd))) {
1533 if (unlikely(pmd_trans_splitting(*pmd))) {
1534 spin_unlock(&vma->vm_mm->page_table_lock);
1535 wait_split_huge_page(vma->anon_vma, pmd);
1536 return -1;
1537 } else {
1538 /* Thp mapped by 'pmd' is stable, so we can
1539 * handle it as it is. */
1540 return 1;
1543 spin_unlock(&vma->vm_mm->page_table_lock);
1544 return 0;
1547 pmd_t *page_check_address_pmd(struct page *page,
1548 struct mm_struct *mm,
1549 unsigned long address,
1550 enum page_check_address_pmd_flag flag)
1552 pmd_t *pmd, *ret = NULL;
1554 if (address & ~HPAGE_PMD_MASK)
1555 goto out;
1557 pmd = mm_find_pmd(mm, address);
1558 if (!pmd)
1559 goto out;
1560 if (pmd_none(*pmd))
1561 goto out;
1562 if (pmd_page(*pmd) != page)
1563 goto out;
1565 * split_vma() may create temporary aliased mappings. There is
1566 * no risk as long as all huge pmd are found and have their
1567 * splitting bit set before __split_huge_page_refcount
1568 * runs. Finding the same huge pmd more than once during the
1569 * same rmap walk is not a problem.
1571 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1572 pmd_trans_splitting(*pmd))
1573 goto out;
1574 if (pmd_trans_huge(*pmd)) {
1575 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1576 !pmd_trans_splitting(*pmd));
1577 ret = pmd;
1579 out:
1580 return ret;
1583 static int __split_huge_page_splitting(struct page *page,
1584 struct vm_area_struct *vma,
1585 unsigned long address)
1587 struct mm_struct *mm = vma->vm_mm;
1588 pmd_t *pmd;
1589 int ret = 0;
1590 /* For mmu_notifiers */
1591 const unsigned long mmun_start = address;
1592 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
1594 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1595 spin_lock(&mm->page_table_lock);
1596 pmd = page_check_address_pmd(page, mm, address,
1597 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1598 if (pmd) {
1600 * We can't temporarily set the pmd to null in order
1601 * to split it, the pmd must remain marked huge at all
1602 * times or the VM won't take the pmd_trans_huge paths
1603 * and it won't wait on the anon_vma->root->rwsem to
1604 * serialize against split_huge_page*.
1606 pmdp_splitting_flush(vma, address, pmd);
1607 ret = 1;
1609 spin_unlock(&mm->page_table_lock);
1610 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1612 return ret;
1615 static void __split_huge_page_refcount(struct page *page,
1616 struct list_head *list)
1618 int i;
1619 struct zone *zone = page_zone(page);
1620 struct lruvec *lruvec;
1621 int tail_count = 0;
1623 /* prevent PageLRU to go away from under us, and freeze lru stats */
1624 spin_lock_irq(&zone->lru_lock);
1625 lruvec = mem_cgroup_page_lruvec(page, zone);
1627 compound_lock(page);
1628 /* complete memcg works before add pages to LRU */
1629 mem_cgroup_split_huge_fixup(page);
1631 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1632 struct page *page_tail = page + i;
1634 /* tail_page->_mapcount cannot change */
1635 BUG_ON(page_mapcount(page_tail) < 0);
1636 tail_count += page_mapcount(page_tail);
1637 /* check for overflow */
1638 BUG_ON(tail_count < 0);
1639 BUG_ON(atomic_read(&page_tail->_count) != 0);
1641 * tail_page->_count is zero and not changing from
1642 * under us. But get_page_unless_zero() may be running
1643 * from under us on the tail_page. If we used
1644 * atomic_set() below instead of atomic_add(), we
1645 * would then run atomic_set() concurrently with
1646 * get_page_unless_zero(), and atomic_set() is
1647 * implemented in C not using locked ops. spin_unlock
1648 * on x86 sometime uses locked ops because of PPro
1649 * errata 66, 92, so unless somebody can guarantee
1650 * atomic_set() here would be safe on all archs (and
1651 * not only on x86), it's safer to use atomic_add().
1653 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1654 &page_tail->_count);
1656 /* after clearing PageTail the gup refcount can be released */
1657 smp_mb();
1660 * retain hwpoison flag of the poisoned tail page:
1661 * fix for the unsuitable process killed on Guest Machine(KVM)
1662 * by the memory-failure.
1664 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1665 page_tail->flags |= (page->flags &
1666 ((1L << PG_referenced) |
1667 (1L << PG_swapbacked) |
1668 (1L << PG_mlocked) |
1669 (1L << PG_uptodate) |
1670 (1L << PG_active) |
1671 (1L << PG_unevictable)));
1672 page_tail->flags |= (1L << PG_dirty);
1674 /* clear PageTail before overwriting first_page */
1675 smp_wmb();
1678 * __split_huge_page_splitting() already set the
1679 * splitting bit in all pmd that could map this
1680 * hugepage, that will ensure no CPU can alter the
1681 * mapcount on the head page. The mapcount is only
1682 * accounted in the head page and it has to be
1683 * transferred to all tail pages in the below code. So
1684 * for this code to be safe, the split the mapcount
1685 * can't change. But that doesn't mean userland can't
1686 * keep changing and reading the page contents while
1687 * we transfer the mapcount, so the pmd splitting
1688 * status is achieved setting a reserved bit in the
1689 * pmd, not by clearing the present bit.
1691 page_tail->_mapcount = page->_mapcount;
1693 BUG_ON(page_tail->mapping);
1694 page_tail->mapping = page->mapping;
1696 page_tail->index = page->index + i;
1697 page_nid_xchg_last(page_tail, page_nid_last(page));
1699 BUG_ON(!PageAnon(page_tail));
1700 BUG_ON(!PageUptodate(page_tail));
1701 BUG_ON(!PageDirty(page_tail));
1702 BUG_ON(!PageSwapBacked(page_tail));
1704 lru_add_page_tail(page, page_tail, lruvec, list);
1706 atomic_sub(tail_count, &page->_count);
1707 BUG_ON(atomic_read(&page->_count) <= 0);
1709 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1711 ClearPageCompound(page);
1712 compound_unlock(page);
1713 spin_unlock_irq(&zone->lru_lock);
1715 for (i = 1; i < HPAGE_PMD_NR; i++) {
1716 struct page *page_tail = page + i;
1717 BUG_ON(page_count(page_tail) <= 0);
1719 * Tail pages may be freed if there wasn't any mapping
1720 * like if add_to_swap() is running on a lru page that
1721 * had its mapping zapped. And freeing these pages
1722 * requires taking the lru_lock so we do the put_page
1723 * of the tail pages after the split is complete.
1725 put_page(page_tail);
1729 * Only the head page (now become a regular page) is required
1730 * to be pinned by the caller.
1732 BUG_ON(page_count(page) <= 0);
1735 static int __split_huge_page_map(struct page *page,
1736 struct vm_area_struct *vma,
1737 unsigned long address)
1739 struct mm_struct *mm = vma->vm_mm;
1740 pmd_t *pmd, _pmd;
1741 int ret = 0, i;
1742 pgtable_t pgtable;
1743 unsigned long haddr;
1745 spin_lock(&mm->page_table_lock);
1746 pmd = page_check_address_pmd(page, mm, address,
1747 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1748 if (pmd) {
1749 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1750 pmd_populate(mm, &_pmd, pgtable);
1752 haddr = address;
1753 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1754 pte_t *pte, entry;
1755 BUG_ON(PageCompound(page+i));
1756 entry = mk_pte(page + i, vma->vm_page_prot);
1757 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1758 if (!pmd_write(*pmd))
1759 entry = pte_wrprotect(entry);
1760 else
1761 BUG_ON(page_mapcount(page) != 1);
1762 if (!pmd_young(*pmd))
1763 entry = pte_mkold(entry);
1764 if (pmd_numa(*pmd))
1765 entry = pte_mknuma(entry);
1766 pte = pte_offset_map(&_pmd, haddr);
1767 BUG_ON(!pte_none(*pte));
1768 set_pte_at(mm, haddr, pte, entry);
1769 pte_unmap(pte);
1772 smp_wmb(); /* make pte visible before pmd */
1774 * Up to this point the pmd is present and huge and
1775 * userland has the whole access to the hugepage
1776 * during the split (which happens in place). If we
1777 * overwrite the pmd with the not-huge version
1778 * pointing to the pte here (which of course we could
1779 * if all CPUs were bug free), userland could trigger
1780 * a small page size TLB miss on the small sized TLB
1781 * while the hugepage TLB entry is still established
1782 * in the huge TLB. Some CPU doesn't like that. See
1783 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1784 * Erratum 383 on page 93. Intel should be safe but is
1785 * also warns that it's only safe if the permission
1786 * and cache attributes of the two entries loaded in
1787 * the two TLB is identical (which should be the case
1788 * here). But it is generally safer to never allow
1789 * small and huge TLB entries for the same virtual
1790 * address to be loaded simultaneously. So instead of
1791 * doing "pmd_populate(); flush_tlb_range();" we first
1792 * mark the current pmd notpresent (atomically because
1793 * here the pmd_trans_huge and pmd_trans_splitting
1794 * must remain set at all times on the pmd until the
1795 * split is complete for this pmd), then we flush the
1796 * SMP TLB and finally we write the non-huge version
1797 * of the pmd entry with pmd_populate.
1799 pmdp_invalidate(vma, address, pmd);
1800 pmd_populate(mm, pmd, pgtable);
1801 ret = 1;
1803 spin_unlock(&mm->page_table_lock);
1805 return ret;
1808 /* must be called with anon_vma->root->rwsem held */
1809 static void __split_huge_page(struct page *page,
1810 struct anon_vma *anon_vma,
1811 struct list_head *list)
1813 int mapcount, mapcount2;
1814 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1815 struct anon_vma_chain *avc;
1817 BUG_ON(!PageHead(page));
1818 BUG_ON(PageTail(page));
1820 mapcount = 0;
1821 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1822 struct vm_area_struct *vma = avc->vma;
1823 unsigned long addr = vma_address(page, vma);
1824 BUG_ON(is_vma_temporary_stack(vma));
1825 mapcount += __split_huge_page_splitting(page, vma, addr);
1828 * It is critical that new vmas are added to the tail of the
1829 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1830 * and establishes a child pmd before
1831 * __split_huge_page_splitting() freezes the parent pmd (so if
1832 * we fail to prevent copy_huge_pmd() from running until the
1833 * whole __split_huge_page() is complete), we will still see
1834 * the newly established pmd of the child later during the
1835 * walk, to be able to set it as pmd_trans_splitting too.
1837 if (mapcount != page_mapcount(page))
1838 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1839 mapcount, page_mapcount(page));
1840 BUG_ON(mapcount != page_mapcount(page));
1842 __split_huge_page_refcount(page, list);
1844 mapcount2 = 0;
1845 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1846 struct vm_area_struct *vma = avc->vma;
1847 unsigned long addr = vma_address(page, vma);
1848 BUG_ON(is_vma_temporary_stack(vma));
1849 mapcount2 += __split_huge_page_map(page, vma, addr);
1851 if (mapcount != mapcount2)
1852 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1853 mapcount, mapcount2, page_mapcount(page));
1854 BUG_ON(mapcount != mapcount2);
1858 * Split a hugepage into normal pages. This doesn't change the position of head
1859 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1860 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1861 * from the hugepage.
1862 * Return 0 if the hugepage is split successfully otherwise return 1.
1864 int split_huge_page_to_list(struct page *page, struct list_head *list)
1866 struct anon_vma *anon_vma;
1867 int ret = 1;
1869 BUG_ON(is_huge_zero_page(page));
1870 BUG_ON(!PageAnon(page));
1873 * The caller does not necessarily hold an mmap_sem that would prevent
1874 * the anon_vma disappearing so we first we take a reference to it
1875 * and then lock the anon_vma for write. This is similar to
1876 * page_lock_anon_vma_read except the write lock is taken to serialise
1877 * against parallel split or collapse operations.
1879 anon_vma = page_get_anon_vma(page);
1880 if (!anon_vma)
1881 goto out;
1882 anon_vma_lock_write(anon_vma);
1884 ret = 0;
1885 if (!PageCompound(page))
1886 goto out_unlock;
1888 BUG_ON(!PageSwapBacked(page));
1889 __split_huge_page(page, anon_vma, list);
1890 count_vm_event(THP_SPLIT);
1892 BUG_ON(PageCompound(page));
1893 out_unlock:
1894 anon_vma_unlock_write(anon_vma);
1895 put_anon_vma(anon_vma);
1896 out:
1897 return ret;
1900 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1902 int hugepage_madvise(struct vm_area_struct *vma,
1903 unsigned long *vm_flags, int advice)
1905 struct mm_struct *mm = vma->vm_mm;
1907 switch (advice) {
1908 case MADV_HUGEPAGE:
1910 * Be somewhat over-protective like KSM for now!
1912 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1913 return -EINVAL;
1914 if (mm->def_flags & VM_NOHUGEPAGE)
1915 return -EINVAL;
1916 *vm_flags &= ~VM_NOHUGEPAGE;
1917 *vm_flags |= VM_HUGEPAGE;
1919 * If the vma become good for khugepaged to scan,
1920 * register it here without waiting a page fault that
1921 * may not happen any time soon.
1923 if (unlikely(khugepaged_enter_vma_merge(vma)))
1924 return -ENOMEM;
1925 break;
1926 case MADV_NOHUGEPAGE:
1928 * Be somewhat over-protective like KSM for now!
1930 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1931 return -EINVAL;
1932 *vm_flags &= ~VM_HUGEPAGE;
1933 *vm_flags |= VM_NOHUGEPAGE;
1935 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1936 * this vma even if we leave the mm registered in khugepaged if
1937 * it got registered before VM_NOHUGEPAGE was set.
1939 break;
1942 return 0;
1945 static int __init khugepaged_slab_init(void)
1947 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1948 sizeof(struct mm_slot),
1949 __alignof__(struct mm_slot), 0, NULL);
1950 if (!mm_slot_cache)
1951 return -ENOMEM;
1953 return 0;
1956 static inline struct mm_slot *alloc_mm_slot(void)
1958 if (!mm_slot_cache) /* initialization failed */
1959 return NULL;
1960 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1963 static inline void free_mm_slot(struct mm_slot *mm_slot)
1965 kmem_cache_free(mm_slot_cache, mm_slot);
1968 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1970 struct mm_slot *mm_slot;
1972 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1973 if (mm == mm_slot->mm)
1974 return mm_slot;
1976 return NULL;
1979 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1980 struct mm_slot *mm_slot)
1982 mm_slot->mm = mm;
1983 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1986 static inline int khugepaged_test_exit(struct mm_struct *mm)
1988 return atomic_read(&mm->mm_users) == 0;
1991 int __khugepaged_enter(struct mm_struct *mm)
1993 struct mm_slot *mm_slot;
1994 int wakeup;
1996 mm_slot = alloc_mm_slot();
1997 if (!mm_slot)
1998 return -ENOMEM;
2000 /* __khugepaged_exit() must not run from under us */
2001 VM_BUG_ON(khugepaged_test_exit(mm));
2002 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2003 free_mm_slot(mm_slot);
2004 return 0;
2007 spin_lock(&khugepaged_mm_lock);
2008 insert_to_mm_slots_hash(mm, mm_slot);
2010 * Insert just behind the scanning cursor, to let the area settle
2011 * down a little.
2013 wakeup = list_empty(&khugepaged_scan.mm_head);
2014 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2015 spin_unlock(&khugepaged_mm_lock);
2017 atomic_inc(&mm->mm_count);
2018 if (wakeup)
2019 wake_up_interruptible(&khugepaged_wait);
2021 return 0;
2024 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2026 unsigned long hstart, hend;
2027 if (!vma->anon_vma)
2029 * Not yet faulted in so we will register later in the
2030 * page fault if needed.
2032 return 0;
2033 if (vma->vm_ops)
2034 /* khugepaged not yet working on file or special mappings */
2035 return 0;
2036 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2037 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2038 hend = vma->vm_end & HPAGE_PMD_MASK;
2039 if (hstart < hend)
2040 return khugepaged_enter(vma);
2041 return 0;
2044 void __khugepaged_exit(struct mm_struct *mm)
2046 struct mm_slot *mm_slot;
2047 int free = 0;
2049 spin_lock(&khugepaged_mm_lock);
2050 mm_slot = get_mm_slot(mm);
2051 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2052 hash_del(&mm_slot->hash);
2053 list_del(&mm_slot->mm_node);
2054 free = 1;
2056 spin_unlock(&khugepaged_mm_lock);
2058 if (free) {
2059 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2060 free_mm_slot(mm_slot);
2061 mmdrop(mm);
2062 } else if (mm_slot) {
2064 * This is required to serialize against
2065 * khugepaged_test_exit() (which is guaranteed to run
2066 * under mmap sem read mode). Stop here (after we
2067 * return all pagetables will be destroyed) until
2068 * khugepaged has finished working on the pagetables
2069 * under the mmap_sem.
2071 down_write(&mm->mmap_sem);
2072 up_write(&mm->mmap_sem);
2076 static void release_pte_page(struct page *page)
2078 /* 0 stands for page_is_file_cache(page) == false */
2079 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2080 unlock_page(page);
2081 putback_lru_page(page);
2084 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2086 while (--_pte >= pte) {
2087 pte_t pteval = *_pte;
2088 if (!pte_none(pteval))
2089 release_pte_page(pte_page(pteval));
2093 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2094 unsigned long address,
2095 pte_t *pte)
2097 struct page *page;
2098 pte_t *_pte;
2099 int referenced = 0, none = 0;
2100 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2101 _pte++, address += PAGE_SIZE) {
2102 pte_t pteval = *_pte;
2103 if (pte_none(pteval)) {
2104 if (++none <= khugepaged_max_ptes_none)
2105 continue;
2106 else
2107 goto out;
2109 if (!pte_present(pteval) || !pte_write(pteval))
2110 goto out;
2111 page = vm_normal_page(vma, address, pteval);
2112 if (unlikely(!page))
2113 goto out;
2115 VM_BUG_ON(PageCompound(page));
2116 BUG_ON(!PageAnon(page));
2117 VM_BUG_ON(!PageSwapBacked(page));
2119 /* cannot use mapcount: can't collapse if there's a gup pin */
2120 if (page_count(page) != 1)
2121 goto out;
2123 * We can do it before isolate_lru_page because the
2124 * page can't be freed from under us. NOTE: PG_lock
2125 * is needed to serialize against split_huge_page
2126 * when invoked from the VM.
2128 if (!trylock_page(page))
2129 goto out;
2131 * Isolate the page to avoid collapsing an hugepage
2132 * currently in use by the VM.
2134 if (isolate_lru_page(page)) {
2135 unlock_page(page);
2136 goto out;
2138 /* 0 stands for page_is_file_cache(page) == false */
2139 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2140 VM_BUG_ON(!PageLocked(page));
2141 VM_BUG_ON(PageLRU(page));
2143 /* If there is no mapped pte young don't collapse the page */
2144 if (pte_young(pteval) || PageReferenced(page) ||
2145 mmu_notifier_test_young(vma->vm_mm, address))
2146 referenced = 1;
2148 if (likely(referenced))
2149 return 1;
2150 out:
2151 release_pte_pages(pte, _pte);
2152 return 0;
2155 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2156 struct vm_area_struct *vma,
2157 unsigned long address,
2158 spinlock_t *ptl)
2160 pte_t *_pte;
2161 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2162 pte_t pteval = *_pte;
2163 struct page *src_page;
2165 if (pte_none(pteval)) {
2166 clear_user_highpage(page, address);
2167 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2168 } else {
2169 src_page = pte_page(pteval);
2170 copy_user_highpage(page, src_page, address, vma);
2171 VM_BUG_ON(page_mapcount(src_page) != 1);
2172 release_pte_page(src_page);
2174 * ptl mostly unnecessary, but preempt has to
2175 * be disabled to update the per-cpu stats
2176 * inside page_remove_rmap().
2178 spin_lock(ptl);
2180 * paravirt calls inside pte_clear here are
2181 * superfluous.
2183 pte_clear(vma->vm_mm, address, _pte);
2184 page_remove_rmap(src_page);
2185 spin_unlock(ptl);
2186 free_page_and_swap_cache(src_page);
2189 address += PAGE_SIZE;
2190 page++;
2194 static void khugepaged_alloc_sleep(void)
2196 wait_event_freezable_timeout(khugepaged_wait, false,
2197 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2200 #ifdef CONFIG_NUMA
2201 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2203 if (IS_ERR(*hpage)) {
2204 if (!*wait)
2205 return false;
2207 *wait = false;
2208 *hpage = NULL;
2209 khugepaged_alloc_sleep();
2210 } else if (*hpage) {
2211 put_page(*hpage);
2212 *hpage = NULL;
2215 return true;
2218 static struct page
2219 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2220 struct vm_area_struct *vma, unsigned long address,
2221 int node)
2223 VM_BUG_ON(*hpage);
2225 * Allocate the page while the vma is still valid and under
2226 * the mmap_sem read mode so there is no memory allocation
2227 * later when we take the mmap_sem in write mode. This is more
2228 * friendly behavior (OTOH it may actually hide bugs) to
2229 * filesystems in userland with daemons allocating memory in
2230 * the userland I/O paths. Allocating memory with the
2231 * mmap_sem in read mode is good idea also to allow greater
2232 * scalability.
2234 *hpage = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2235 node, __GFP_OTHER_NODE);
2238 * After allocating the hugepage, release the mmap_sem read lock in
2239 * preparation for taking it in write mode.
2241 up_read(&mm->mmap_sem);
2242 if (unlikely(!*hpage)) {
2243 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2244 *hpage = ERR_PTR(-ENOMEM);
2245 return NULL;
2248 count_vm_event(THP_COLLAPSE_ALLOC);
2249 return *hpage;
2251 #else
2252 static struct page *khugepaged_alloc_hugepage(bool *wait)
2254 struct page *hpage;
2256 do {
2257 hpage = alloc_hugepage(khugepaged_defrag());
2258 if (!hpage) {
2259 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2260 if (!*wait)
2261 return NULL;
2263 *wait = false;
2264 khugepaged_alloc_sleep();
2265 } else
2266 count_vm_event(THP_COLLAPSE_ALLOC);
2267 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2269 return hpage;
2272 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2274 if (!*hpage)
2275 *hpage = khugepaged_alloc_hugepage(wait);
2277 if (unlikely(!*hpage))
2278 return false;
2280 return true;
2283 static struct page
2284 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2285 struct vm_area_struct *vma, unsigned long address,
2286 int node)
2288 up_read(&mm->mmap_sem);
2289 VM_BUG_ON(!*hpage);
2290 return *hpage;
2292 #endif
2294 static bool hugepage_vma_check(struct vm_area_struct *vma)
2296 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2297 (vma->vm_flags & VM_NOHUGEPAGE))
2298 return false;
2300 if (!vma->anon_vma || vma->vm_ops)
2301 return false;
2302 if (is_vma_temporary_stack(vma))
2303 return false;
2304 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2305 return true;
2308 static void collapse_huge_page(struct mm_struct *mm,
2309 unsigned long address,
2310 struct page **hpage,
2311 struct vm_area_struct *vma,
2312 int node)
2314 pmd_t *pmd, _pmd;
2315 pte_t *pte;
2316 pgtable_t pgtable;
2317 struct page *new_page;
2318 spinlock_t *ptl;
2319 int isolated;
2320 unsigned long hstart, hend;
2321 unsigned long mmun_start; /* For mmu_notifiers */
2322 unsigned long mmun_end; /* For mmu_notifiers */
2324 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2326 /* release the mmap_sem read lock. */
2327 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2328 if (!new_page)
2329 return;
2331 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2332 return;
2335 * Prevent all access to pagetables with the exception of
2336 * gup_fast later hanlded by the ptep_clear_flush and the VM
2337 * handled by the anon_vma lock + PG_lock.
2339 down_write(&mm->mmap_sem);
2340 if (unlikely(khugepaged_test_exit(mm)))
2341 goto out;
2343 vma = find_vma(mm, address);
2344 if (!vma)
2345 goto out;
2346 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2347 hend = vma->vm_end & HPAGE_PMD_MASK;
2348 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2349 goto out;
2350 if (!hugepage_vma_check(vma))
2351 goto out;
2352 pmd = mm_find_pmd(mm, address);
2353 if (!pmd)
2354 goto out;
2355 if (pmd_trans_huge(*pmd))
2356 goto out;
2358 anon_vma_lock_write(vma->anon_vma);
2360 pte = pte_offset_map(pmd, address);
2361 ptl = pte_lockptr(mm, pmd);
2363 mmun_start = address;
2364 mmun_end = address + HPAGE_PMD_SIZE;
2365 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2366 spin_lock(&mm->page_table_lock); /* probably unnecessary */
2368 * After this gup_fast can't run anymore. This also removes
2369 * any huge TLB entry from the CPU so we won't allow
2370 * huge and small TLB entries for the same virtual address
2371 * to avoid the risk of CPU bugs in that area.
2373 _pmd = pmdp_clear_flush(vma, address, pmd);
2374 spin_unlock(&mm->page_table_lock);
2375 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2377 spin_lock(ptl);
2378 isolated = __collapse_huge_page_isolate(vma, address, pte);
2379 spin_unlock(ptl);
2381 if (unlikely(!isolated)) {
2382 pte_unmap(pte);
2383 spin_lock(&mm->page_table_lock);
2384 BUG_ON(!pmd_none(*pmd));
2386 * We can only use set_pmd_at when establishing
2387 * hugepmds and never for establishing regular pmds that
2388 * points to regular pagetables. Use pmd_populate for that
2390 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2391 spin_unlock(&mm->page_table_lock);
2392 anon_vma_unlock_write(vma->anon_vma);
2393 goto out;
2397 * All pages are isolated and locked so anon_vma rmap
2398 * can't run anymore.
2400 anon_vma_unlock_write(vma->anon_vma);
2402 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2403 pte_unmap(pte);
2404 __SetPageUptodate(new_page);
2405 pgtable = pmd_pgtable(_pmd);
2407 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2408 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2411 * spin_lock() below is not the equivalent of smp_wmb(), so
2412 * this is needed to avoid the copy_huge_page writes to become
2413 * visible after the set_pmd_at() write.
2415 smp_wmb();
2417 spin_lock(&mm->page_table_lock);
2418 BUG_ON(!pmd_none(*pmd));
2419 page_add_new_anon_rmap(new_page, vma, address);
2420 pgtable_trans_huge_deposit(mm, pmd, pgtable);
2421 set_pmd_at(mm, address, pmd, _pmd);
2422 update_mmu_cache_pmd(vma, address, pmd);
2423 spin_unlock(&mm->page_table_lock);
2425 *hpage = NULL;
2427 khugepaged_pages_collapsed++;
2428 out_up_write:
2429 up_write(&mm->mmap_sem);
2430 return;
2432 out:
2433 mem_cgroup_uncharge_page(new_page);
2434 goto out_up_write;
2437 static int khugepaged_scan_pmd(struct mm_struct *mm,
2438 struct vm_area_struct *vma,
2439 unsigned long address,
2440 struct page **hpage)
2442 pmd_t *pmd;
2443 pte_t *pte, *_pte;
2444 int ret = 0, referenced = 0, none = 0;
2445 struct page *page;
2446 unsigned long _address;
2447 spinlock_t *ptl;
2448 int node = NUMA_NO_NODE;
2450 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2452 pmd = mm_find_pmd(mm, address);
2453 if (!pmd)
2454 goto out;
2455 if (pmd_trans_huge(*pmd))
2456 goto out;
2458 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2459 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2460 _pte++, _address += PAGE_SIZE) {
2461 pte_t pteval = *_pte;
2462 if (pte_none(pteval)) {
2463 if (++none <= khugepaged_max_ptes_none)
2464 continue;
2465 else
2466 goto out_unmap;
2468 if (!pte_present(pteval) || !pte_write(pteval))
2469 goto out_unmap;
2470 page = vm_normal_page(vma, _address, pteval);
2471 if (unlikely(!page))
2472 goto out_unmap;
2474 * Chose the node of the first page. This could
2475 * be more sophisticated and look at more pages,
2476 * but isn't for now.
2478 if (node == NUMA_NO_NODE)
2479 node = page_to_nid(page);
2480 VM_BUG_ON(PageCompound(page));
2481 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2482 goto out_unmap;
2483 /* cannot use mapcount: can't collapse if there's a gup pin */
2484 if (page_count(page) != 1)
2485 goto out_unmap;
2486 if (pte_young(pteval) || PageReferenced(page) ||
2487 mmu_notifier_test_young(vma->vm_mm, address))
2488 referenced = 1;
2490 if (referenced)
2491 ret = 1;
2492 out_unmap:
2493 pte_unmap_unlock(pte, ptl);
2494 if (ret)
2495 /* collapse_huge_page will return with the mmap_sem released */
2496 collapse_huge_page(mm, address, hpage, vma, node);
2497 out:
2498 return ret;
2501 static void collect_mm_slot(struct mm_slot *mm_slot)
2503 struct mm_struct *mm = mm_slot->mm;
2505 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2507 if (khugepaged_test_exit(mm)) {
2508 /* free mm_slot */
2509 hash_del(&mm_slot->hash);
2510 list_del(&mm_slot->mm_node);
2513 * Not strictly needed because the mm exited already.
2515 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2518 /* khugepaged_mm_lock actually not necessary for the below */
2519 free_mm_slot(mm_slot);
2520 mmdrop(mm);
2524 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2525 struct page **hpage)
2526 __releases(&khugepaged_mm_lock)
2527 __acquires(&khugepaged_mm_lock)
2529 struct mm_slot *mm_slot;
2530 struct mm_struct *mm;
2531 struct vm_area_struct *vma;
2532 int progress = 0;
2534 VM_BUG_ON(!pages);
2535 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2537 if (khugepaged_scan.mm_slot)
2538 mm_slot = khugepaged_scan.mm_slot;
2539 else {
2540 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2541 struct mm_slot, mm_node);
2542 khugepaged_scan.address = 0;
2543 khugepaged_scan.mm_slot = mm_slot;
2545 spin_unlock(&khugepaged_mm_lock);
2547 mm = mm_slot->mm;
2548 down_read(&mm->mmap_sem);
2549 if (unlikely(khugepaged_test_exit(mm)))
2550 vma = NULL;
2551 else
2552 vma = find_vma(mm, khugepaged_scan.address);
2554 progress++;
2555 for (; vma; vma = vma->vm_next) {
2556 unsigned long hstart, hend;
2558 cond_resched();
2559 if (unlikely(khugepaged_test_exit(mm))) {
2560 progress++;
2561 break;
2563 if (!hugepage_vma_check(vma)) {
2564 skip:
2565 progress++;
2566 continue;
2568 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2569 hend = vma->vm_end & HPAGE_PMD_MASK;
2570 if (hstart >= hend)
2571 goto skip;
2572 if (khugepaged_scan.address > hend)
2573 goto skip;
2574 if (khugepaged_scan.address < hstart)
2575 khugepaged_scan.address = hstart;
2576 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2578 while (khugepaged_scan.address < hend) {
2579 int ret;
2580 cond_resched();
2581 if (unlikely(khugepaged_test_exit(mm)))
2582 goto breakouterloop;
2584 VM_BUG_ON(khugepaged_scan.address < hstart ||
2585 khugepaged_scan.address + HPAGE_PMD_SIZE >
2586 hend);
2587 ret = khugepaged_scan_pmd(mm, vma,
2588 khugepaged_scan.address,
2589 hpage);
2590 /* move to next address */
2591 khugepaged_scan.address += HPAGE_PMD_SIZE;
2592 progress += HPAGE_PMD_NR;
2593 if (ret)
2594 /* we released mmap_sem so break loop */
2595 goto breakouterloop_mmap_sem;
2596 if (progress >= pages)
2597 goto breakouterloop;
2600 breakouterloop:
2601 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2602 breakouterloop_mmap_sem:
2604 spin_lock(&khugepaged_mm_lock);
2605 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2607 * Release the current mm_slot if this mm is about to die, or
2608 * if we scanned all vmas of this mm.
2610 if (khugepaged_test_exit(mm) || !vma) {
2612 * Make sure that if mm_users is reaching zero while
2613 * khugepaged runs here, khugepaged_exit will find
2614 * mm_slot not pointing to the exiting mm.
2616 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2617 khugepaged_scan.mm_slot = list_entry(
2618 mm_slot->mm_node.next,
2619 struct mm_slot, mm_node);
2620 khugepaged_scan.address = 0;
2621 } else {
2622 khugepaged_scan.mm_slot = NULL;
2623 khugepaged_full_scans++;
2626 collect_mm_slot(mm_slot);
2629 return progress;
2632 static int khugepaged_has_work(void)
2634 return !list_empty(&khugepaged_scan.mm_head) &&
2635 khugepaged_enabled();
2638 static int khugepaged_wait_event(void)
2640 return !list_empty(&khugepaged_scan.mm_head) ||
2641 kthread_should_stop();
2644 static void khugepaged_do_scan(void)
2646 struct page *hpage = NULL;
2647 unsigned int progress = 0, pass_through_head = 0;
2648 unsigned int pages = khugepaged_pages_to_scan;
2649 bool wait = true;
2651 barrier(); /* write khugepaged_pages_to_scan to local stack */
2653 while (progress < pages) {
2654 if (!khugepaged_prealloc_page(&hpage, &wait))
2655 break;
2657 cond_resched();
2659 if (unlikely(kthread_should_stop() || freezing(current)))
2660 break;
2662 spin_lock(&khugepaged_mm_lock);
2663 if (!khugepaged_scan.mm_slot)
2664 pass_through_head++;
2665 if (khugepaged_has_work() &&
2666 pass_through_head < 2)
2667 progress += khugepaged_scan_mm_slot(pages - progress,
2668 &hpage);
2669 else
2670 progress = pages;
2671 spin_unlock(&khugepaged_mm_lock);
2674 if (!IS_ERR_OR_NULL(hpage))
2675 put_page(hpage);
2678 static void khugepaged_wait_work(void)
2680 try_to_freeze();
2682 if (khugepaged_has_work()) {
2683 if (!khugepaged_scan_sleep_millisecs)
2684 return;
2686 wait_event_freezable_timeout(khugepaged_wait,
2687 kthread_should_stop(),
2688 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2689 return;
2692 if (khugepaged_enabled())
2693 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2696 static int khugepaged(void *none)
2698 struct mm_slot *mm_slot;
2700 set_freezable();
2701 set_user_nice(current, 19);
2703 while (!kthread_should_stop()) {
2704 khugepaged_do_scan();
2705 khugepaged_wait_work();
2708 spin_lock(&khugepaged_mm_lock);
2709 mm_slot = khugepaged_scan.mm_slot;
2710 khugepaged_scan.mm_slot = NULL;
2711 if (mm_slot)
2712 collect_mm_slot(mm_slot);
2713 spin_unlock(&khugepaged_mm_lock);
2714 return 0;
2717 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2718 unsigned long haddr, pmd_t *pmd)
2720 struct mm_struct *mm = vma->vm_mm;
2721 pgtable_t pgtable;
2722 pmd_t _pmd;
2723 int i;
2725 pmdp_clear_flush(vma, haddr, pmd);
2726 /* leave pmd empty until pte is filled */
2728 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2729 pmd_populate(mm, &_pmd, pgtable);
2731 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2732 pte_t *pte, entry;
2733 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2734 entry = pte_mkspecial(entry);
2735 pte = pte_offset_map(&_pmd, haddr);
2736 VM_BUG_ON(!pte_none(*pte));
2737 set_pte_at(mm, haddr, pte, entry);
2738 pte_unmap(pte);
2740 smp_wmb(); /* make pte visible before pmd */
2741 pmd_populate(mm, pmd, pgtable);
2742 put_huge_zero_page();
2745 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2746 pmd_t *pmd)
2748 struct page *page;
2749 struct mm_struct *mm = vma->vm_mm;
2750 unsigned long haddr = address & HPAGE_PMD_MASK;
2751 unsigned long mmun_start; /* For mmu_notifiers */
2752 unsigned long mmun_end; /* For mmu_notifiers */
2754 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2756 mmun_start = haddr;
2757 mmun_end = haddr + HPAGE_PMD_SIZE;
2758 again:
2759 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2760 spin_lock(&mm->page_table_lock);
2761 if (unlikely(!pmd_trans_huge(*pmd))) {
2762 spin_unlock(&mm->page_table_lock);
2763 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2764 return;
2766 if (is_huge_zero_pmd(*pmd)) {
2767 __split_huge_zero_page_pmd(vma, haddr, pmd);
2768 spin_unlock(&mm->page_table_lock);
2769 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2770 return;
2772 page = pmd_page(*pmd);
2773 VM_BUG_ON(!page_count(page));
2774 get_page(page);
2775 spin_unlock(&mm->page_table_lock);
2776 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2778 split_huge_page(page);
2780 put_page(page);
2783 * We don't always have down_write of mmap_sem here: a racing
2784 * do_huge_pmd_wp_page() might have copied-on-write to another
2785 * huge page before our split_huge_page() got the anon_vma lock.
2787 if (unlikely(pmd_trans_huge(*pmd)))
2788 goto again;
2791 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2792 pmd_t *pmd)
2794 struct vm_area_struct *vma;
2796 vma = find_vma(mm, address);
2797 BUG_ON(vma == NULL);
2798 split_huge_page_pmd(vma, address, pmd);
2801 static void split_huge_page_address(struct mm_struct *mm,
2802 unsigned long address)
2804 pmd_t *pmd;
2806 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2808 pmd = mm_find_pmd(mm, address);
2809 if (!pmd)
2810 return;
2812 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2813 * materialize from under us.
2815 split_huge_page_pmd_mm(mm, address, pmd);
2818 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2819 unsigned long start,
2820 unsigned long end,
2821 long adjust_next)
2824 * If the new start address isn't hpage aligned and it could
2825 * previously contain an hugepage: check if we need to split
2826 * an huge pmd.
2828 if (start & ~HPAGE_PMD_MASK &&
2829 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2830 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2831 split_huge_page_address(vma->vm_mm, start);
2834 * If the new end address isn't hpage aligned and it could
2835 * previously contain an hugepage: check if we need to split
2836 * an huge pmd.
2838 if (end & ~HPAGE_PMD_MASK &&
2839 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2840 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2841 split_huge_page_address(vma->vm_mm, end);
2844 * If we're also updating the vma->vm_next->vm_start, if the new
2845 * vm_next->vm_start isn't page aligned and it could previously
2846 * contain an hugepage: check if we need to split an huge pmd.
2848 if (adjust_next > 0) {
2849 struct vm_area_struct *next = vma->vm_next;
2850 unsigned long nstart = next->vm_start;
2851 nstart += adjust_next << PAGE_SHIFT;
2852 if (nstart & ~HPAGE_PMD_MASK &&
2853 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2854 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2855 split_huge_page_address(next->vm_mm, nstart);