2 * Copyright (C) 2009 Red Hat, Inc.
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
9 #include <linux/sched.h>
10 #include <linux/highmem.h>
11 #include <linux/hugetlb.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/rmap.h>
14 #include <linux/swap.h>
15 #include <linux/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>
26 #include <asm/pgalloc.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
36 unsigned long transparent_hugepage_flags __read_mostly
=
37 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
38 (1<<TRANSPARENT_HUGEPAGE_FLAG
)|
40 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
41 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
)|
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
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
;
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
80 struct hlist_node hash
;
81 struct list_head mm_node
;
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)
107 unsigned long recommended_min
;
109 if (!khugepaged_enabled())
112 for_each_populated_zone(zone
)
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();
137 late_initcall(set_recommended_min_free_kbytes
);
139 static int start_khugepaged(void)
142 if (khugepaged_enabled()) {
143 if (!khugepaged_thread
)
144 khugepaged_thread
= kthread_run(khugepaged
, NULL
,
146 if (unlikely(IS_ERR(khugepaged_thread
))) {
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
;
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
;
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
,
188 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED
);
191 count_vm_event(THP_ZERO_PAGE_ALLOC
);
193 if (cmpxchg(&huge_zero_page
, NULL
, zero_page
)) {
195 __free_page(zero_page
);
199 /* We take additional reference here. It will be put back by shrinker */
200 atomic_set(&huge_zero_refcount
, 2);
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
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
);
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
,
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");
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
);
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
)
292 ret
= double_flag_store(kobj
, attr
, buf
, count
,
293 TRANSPARENT_HUGEPAGE_FLAG
,
294 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
299 mutex_lock(&khugepaged_mutex
);
300 err
= start_khugepaged();
301 mutex_unlock(&khugepaged_mutex
);
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
)
328 ret
= kstrtoul(buf
, 10, &value
);
335 set_bit(flag
, &transparent_hugepage_flags
);
337 clear_bit(flag
, &transparent_hugepage_flags
);
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
[] = {
400 &use_zero_page_attr
.attr
,
401 #ifdef CONFIG_DEBUG_VM
402 &debug_cow_attr
.attr
,
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
,
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
)
425 err
= kstrtoul(buf
, 10, &msecs
);
426 if (err
|| msecs
> UINT_MAX
)
429 khugepaged_scan_sleep_millisecs
= msecs
;
430 wake_up_interruptible(&khugepaged_wait
);
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
,
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
)
452 err
= kstrtoul(buf
, 10, &msecs
);
453 if (err
|| msecs
> UINT_MAX
)
456 khugepaged_alloc_sleep_millisecs
= msecs
;
457 wake_up_interruptible(&khugepaged_wait
);
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
,
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
)
478 err
= kstrtoul(buf
, 10, &pages
);
479 if (err
|| !pages
|| pages
> UINT_MAX
)
482 khugepaged_pages_to_scan
= pages
;
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
,
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
,
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
,
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
)
544 unsigned long max_ptes_none
;
546 err
= kstrtoul(buf
, 10, &max_ptes_none
);
547 if (err
|| max_ptes_none
> HPAGE_PMD_NR
-1)
550 khugepaged_max_ptes_none
= max_ptes_none
;
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
,
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
)
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");
584 err
= sysfs_create_group(*hugepage_kobj
, &hugepage_attr_group
);
586 printk(KERN_ERR
"hugepage: failed to register transparent hugepage group\n");
590 err
= sysfs_create_group(*hugepage_kobj
, &khugepaged_attr_group
);
592 printk(KERN_ERR
"hugepage: failed to register transparent hugepage group\n");
593 goto remove_hp_group
;
599 sysfs_remove_group(*hugepage_kobj
, &hugepage_attr_group
);
601 kobject_put(*hugepage_kobj
);
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
);
612 static inline int hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
617 static inline void hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
620 #endif /* CONFIG_SYSFS */
622 static int __init
hugepage_init(void)
625 struct kobject
*hugepage_kobj
;
627 if (!has_transparent_hugepage()) {
628 transparent_hugepage_flags
= 0;
632 err
= hugepage_init_sysfs(&hugepage_kobj
);
636 err
= khugepaged_slab_init();
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;
654 hugepage_exit_sysfs(hugepage_kobj
);
657 module_init(hugepage_init
)
659 static int __init
setup_transparent_hugepage(char *str
)
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
);
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
);
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
);
686 "transparent_hugepage= cannot parse, ignored\n");
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
);
698 static inline pmd_t
mk_huge_pmd(struct page
*page
, struct vm_area_struct
*vma
)
701 entry
= mk_pmd(page
, vma
->vm_page_prot
);
702 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
703 entry
= pmd_mkhuge(entry
);
707 static int __do_huge_pmd_anonymous_page(struct mm_struct
*mm
,
708 struct vm_area_struct
*vma
,
709 unsigned long haddr
, pmd_t
*pmd
,
714 VM_BUG_ON(!PageCompound(page
));
715 pgtable
= pte_alloc_one(mm
, haddr
);
716 if (unlikely(!pgtable
))
719 clear_huge_page(page
, haddr
, HPAGE_PMD_NR
);
721 * The memory barrier inside __SetPageUptodate makes sure that
722 * clear_huge_page writes become visible before the set_pmd_at()
725 __SetPageUptodate(page
);
727 spin_lock(&mm
->page_table_lock
);
728 if (unlikely(!pmd_none(*pmd
))) {
729 spin_unlock(&mm
->page_table_lock
);
730 mem_cgroup_uncharge_page(page
);
732 pte_free(mm
, pgtable
);
735 entry
= mk_huge_pmd(page
, 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
);
741 spin_unlock(&mm
->page_table_lock
);
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
,
757 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag
, extra_gfp
),
758 HPAGE_PMD_ORDER
, vma
, haddr
, nd
);
762 static inline struct page
*alloc_hugepage(int defrag
)
764 return alloc_pages(alloc_hugepage_gfpmask(defrag
, 0),
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
)
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
);
785 int do_huge_pmd_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
786 unsigned long address
, pmd_t
*pmd
,
790 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
793 if (haddr
>= vma
->vm_start
&& haddr
+ HPAGE_PMD_SIZE
<= vma
->vm_end
) {
794 if (unlikely(anon_vma_prepare(vma
)))
796 if (unlikely(khugepaged_enter(vma
)))
798 if (!(flags
& FAULT_FLAG_WRITE
) &&
799 transparent_hugepage_use_zero_page()) {
801 struct page
*zero_page
;
803 pgtable
= pte_alloc_one(mm
, haddr
);
804 if (unlikely(!pgtable
))
806 zero_page
= get_huge_zero_page();
807 if (unlikely(!zero_page
)) {
808 pte_free(mm
, pgtable
);
809 count_vm_event(THP_FAULT_FALLBACK
);
812 spin_lock(&mm
->page_table_lock
);
813 set
= set_huge_zero_page(pgtable
, mm
, vma
, haddr
, pmd
,
815 spin_unlock(&mm
->page_table_lock
);
817 pte_free(mm
, pgtable
);
818 put_huge_zero_page();
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
);
828 count_vm_event(THP_FAULT_ALLOC
);
829 if (unlikely(mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))) {
833 if (unlikely(__do_huge_pmd_anonymous_page(mm
, vma
, haddr
, pmd
,
835 mem_cgroup_uncharge_page(page
);
844 * Use __pte_alloc instead of pte_alloc_map, because we can't
845 * run pte_offset_map on the pmd, if an huge pmd could
846 * materialize from under us from a different thread.
848 if (unlikely(pmd_none(*pmd
)) &&
849 unlikely(__pte_alloc(mm
, vma
, pmd
, address
)))
851 /* if an huge pmd materialized from under us just retry later */
852 if (unlikely(pmd_trans_huge(*pmd
)))
855 * A regular pmd is established and it can't morph into a huge pmd
856 * from under us anymore at this point because we hold the mmap_sem
857 * read mode and khugepaged takes it in write mode. So now it's
858 * safe to run pte_offset_map().
860 pte
= pte_offset_map(pmd
, address
);
861 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
864 int copy_huge_pmd(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
865 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
866 struct vm_area_struct
*vma
)
868 struct page
*src_page
;
874 pgtable
= pte_alloc_one(dst_mm
, addr
);
875 if (unlikely(!pgtable
))
878 spin_lock(&dst_mm
->page_table_lock
);
879 spin_lock_nested(&src_mm
->page_table_lock
, SINGLE_DEPTH_NESTING
);
883 if (unlikely(!pmd_trans_huge(pmd
))) {
884 pte_free(dst_mm
, pgtable
);
888 * mm->page_table_lock is enough to be sure that huge zero pmd is not
889 * under splitting since we don't split the page itself, only pmd to
892 if (is_huge_zero_pmd(pmd
)) {
893 struct page
*zero_page
;
896 * get_huge_zero_page() will never allocate a new page here,
897 * since we already have a zero page to copy. It just takes a
900 zero_page
= get_huge_zero_page();
901 set
= set_huge_zero_page(pgtable
, dst_mm
, vma
, addr
, dst_pmd
,
903 BUG_ON(!set
); /* unexpected !pmd_none(dst_pmd) */
907 if (unlikely(pmd_trans_splitting(pmd
))) {
908 /* split huge page running from under us */
909 spin_unlock(&src_mm
->page_table_lock
);
910 spin_unlock(&dst_mm
->page_table_lock
);
911 pte_free(dst_mm
, pgtable
);
913 wait_split_huge_page(vma
->anon_vma
, src_pmd
); /* src_vma */
916 src_page
= pmd_page(pmd
);
917 VM_BUG_ON(!PageHead(src_page
));
919 page_dup_rmap(src_page
);
920 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
922 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
923 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
924 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
925 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
930 spin_unlock(&src_mm
->page_table_lock
);
931 spin_unlock(&dst_mm
->page_table_lock
);
936 void huge_pmd_set_accessed(struct mm_struct
*mm
,
937 struct vm_area_struct
*vma
,
938 unsigned long address
,
939 pmd_t
*pmd
, pmd_t orig_pmd
,
945 spin_lock(&mm
->page_table_lock
);
946 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
949 entry
= pmd_mkyoung(orig_pmd
);
950 haddr
= address
& HPAGE_PMD_MASK
;
951 if (pmdp_set_access_flags(vma
, haddr
, pmd
, entry
, dirty
))
952 update_mmu_cache_pmd(vma
, address
, pmd
);
955 spin_unlock(&mm
->page_table_lock
);
958 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct
*mm
,
959 struct vm_area_struct
*vma
, unsigned long address
,
960 pmd_t
*pmd
, pmd_t orig_pmd
, unsigned long haddr
)
966 unsigned long mmun_start
; /* For mmu_notifiers */
967 unsigned long mmun_end
; /* For mmu_notifiers */
969 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
975 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
)) {
981 clear_user_highpage(page
, address
);
982 __SetPageUptodate(page
);
985 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
986 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
988 spin_lock(&mm
->page_table_lock
);
989 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
992 pmdp_clear_flush(vma
, haddr
, pmd
);
993 /* leave pmd empty until pte is filled */
995 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
996 pmd_populate(mm
, &_pmd
, pgtable
);
998 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1000 if (haddr
== (address
& PAGE_MASK
)) {
1001 entry
= mk_pte(page
, vma
->vm_page_prot
);
1002 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1003 page_add_new_anon_rmap(page
, vma
, haddr
);
1005 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
1006 entry
= pte_mkspecial(entry
);
1008 pte
= pte_offset_map(&_pmd
, haddr
);
1009 VM_BUG_ON(!pte_none(*pte
));
1010 set_pte_at(mm
, haddr
, pte
, entry
);
1013 smp_wmb(); /* make pte visible before pmd */
1014 pmd_populate(mm
, pmd
, pgtable
);
1015 spin_unlock(&mm
->page_table_lock
);
1016 put_huge_zero_page();
1017 inc_mm_counter(mm
, MM_ANONPAGES
);
1019 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1021 ret
|= VM_FAULT_WRITE
;
1025 spin_unlock(&mm
->page_table_lock
);
1026 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1027 mem_cgroup_uncharge_page(page
);
1032 static int do_huge_pmd_wp_page_fallback(struct mm_struct
*mm
,
1033 struct vm_area_struct
*vma
,
1034 unsigned long address
,
1035 pmd_t
*pmd
, pmd_t orig_pmd
,
1037 unsigned long haddr
)
1042 struct page
**pages
;
1043 unsigned long mmun_start
; /* For mmu_notifiers */
1044 unsigned long mmun_end
; /* For mmu_notifiers */
1046 pages
= kmalloc(sizeof(struct page
*) * HPAGE_PMD_NR
,
1048 if (unlikely(!pages
)) {
1049 ret
|= VM_FAULT_OOM
;
1053 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1054 pages
[i
] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE
|
1056 vma
, address
, page_to_nid(page
));
1057 if (unlikely(!pages
[i
] ||
1058 mem_cgroup_newpage_charge(pages
[i
], mm
,
1062 mem_cgroup_uncharge_start();
1064 mem_cgroup_uncharge_page(pages
[i
]);
1067 mem_cgroup_uncharge_end();
1069 ret
|= VM_FAULT_OOM
;
1074 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1075 copy_user_highpage(pages
[i
], page
+ i
,
1076 haddr
+ PAGE_SIZE
* i
, vma
);
1077 __SetPageUptodate(pages
[i
]);
1082 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1083 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1085 spin_lock(&mm
->page_table_lock
);
1086 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
1087 goto out_free_pages
;
1088 VM_BUG_ON(!PageHead(page
));
1090 pmdp_clear_flush(vma
, haddr
, pmd
);
1091 /* leave pmd empty until pte is filled */
1093 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1094 pmd_populate(mm
, &_pmd
, pgtable
);
1096 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1098 entry
= mk_pte(pages
[i
], vma
->vm_page_prot
);
1099 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1100 page_add_new_anon_rmap(pages
[i
], vma
, haddr
);
1101 pte
= pte_offset_map(&_pmd
, haddr
);
1102 VM_BUG_ON(!pte_none(*pte
));
1103 set_pte_at(mm
, haddr
, pte
, entry
);
1108 smp_wmb(); /* make pte visible before pmd */
1109 pmd_populate(mm
, pmd
, pgtable
);
1110 page_remove_rmap(page
);
1111 spin_unlock(&mm
->page_table_lock
);
1113 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1115 ret
|= VM_FAULT_WRITE
;
1122 spin_unlock(&mm
->page_table_lock
);
1123 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1124 mem_cgroup_uncharge_start();
1125 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1126 mem_cgroup_uncharge_page(pages
[i
]);
1129 mem_cgroup_uncharge_end();
1134 int do_huge_pmd_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1135 unsigned long address
, pmd_t
*pmd
, pmd_t orig_pmd
)
1138 struct page
*page
= NULL
, *new_page
;
1139 unsigned long haddr
;
1140 unsigned long mmun_start
; /* For mmu_notifiers */
1141 unsigned long mmun_end
; /* For mmu_notifiers */
1143 VM_BUG_ON(!vma
->anon_vma
);
1144 haddr
= address
& HPAGE_PMD_MASK
;
1145 if (is_huge_zero_pmd(orig_pmd
))
1147 spin_lock(&mm
->page_table_lock
);
1148 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
1151 page
= pmd_page(orig_pmd
);
1152 VM_BUG_ON(!PageCompound(page
) || !PageHead(page
));
1153 if (page_mapcount(page
) == 1) {
1155 entry
= pmd_mkyoung(orig_pmd
);
1156 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1157 if (pmdp_set_access_flags(vma
, haddr
, pmd
, entry
, 1))
1158 update_mmu_cache_pmd(vma
, address
, pmd
);
1159 ret
|= VM_FAULT_WRITE
;
1163 spin_unlock(&mm
->page_table_lock
);
1165 if (transparent_hugepage_enabled(vma
) &&
1166 !transparent_hugepage_debug_cow())
1167 new_page
= alloc_hugepage_vma(transparent_hugepage_defrag(vma
),
1168 vma
, haddr
, numa_node_id(), 0);
1172 if (unlikely(!new_page
)) {
1173 count_vm_event(THP_FAULT_FALLBACK
);
1174 if (is_huge_zero_pmd(orig_pmd
)) {
1175 ret
= do_huge_pmd_wp_zero_page_fallback(mm
, vma
,
1176 address
, pmd
, orig_pmd
, haddr
);
1178 ret
= do_huge_pmd_wp_page_fallback(mm
, vma
, address
,
1179 pmd
, orig_pmd
, page
, haddr
);
1180 if (ret
& VM_FAULT_OOM
)
1181 split_huge_page(page
);
1186 count_vm_event(THP_FAULT_ALLOC
);
1188 if (unlikely(mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))) {
1191 split_huge_page(page
);
1194 ret
|= VM_FAULT_OOM
;
1198 if (is_huge_zero_pmd(orig_pmd
))
1199 clear_huge_page(new_page
, haddr
, HPAGE_PMD_NR
);
1201 copy_user_huge_page(new_page
, page
, haddr
, vma
, HPAGE_PMD_NR
);
1202 __SetPageUptodate(new_page
);
1205 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1206 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1208 spin_lock(&mm
->page_table_lock
);
1211 if (unlikely(!pmd_same(*pmd
, orig_pmd
))) {
1212 spin_unlock(&mm
->page_table_lock
);
1213 mem_cgroup_uncharge_page(new_page
);
1218 entry
= mk_huge_pmd(new_page
, vma
);
1219 pmdp_clear_flush(vma
, haddr
, pmd
);
1220 page_add_new_anon_rmap(new_page
, vma
, haddr
);
1221 set_pmd_at(mm
, haddr
, pmd
, entry
);
1222 update_mmu_cache_pmd(vma
, address
, pmd
);
1223 if (is_huge_zero_pmd(orig_pmd
)) {
1224 add_mm_counter(mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1225 put_huge_zero_page();
1227 VM_BUG_ON(!PageHead(page
));
1228 page_remove_rmap(page
);
1231 ret
|= VM_FAULT_WRITE
;
1233 spin_unlock(&mm
->page_table_lock
);
1235 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1239 spin_unlock(&mm
->page_table_lock
);
1243 struct page
*follow_trans_huge_pmd(struct vm_area_struct
*vma
,
1248 struct mm_struct
*mm
= vma
->vm_mm
;
1249 struct page
*page
= NULL
;
1251 assert_spin_locked(&mm
->page_table_lock
);
1253 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
1256 /* Avoid dumping huge zero page */
1257 if ((flags
& FOLL_DUMP
) && is_huge_zero_pmd(*pmd
))
1258 return ERR_PTR(-EFAULT
);
1260 page
= pmd_page(*pmd
);
1261 VM_BUG_ON(!PageHead(page
));
1262 if (flags
& FOLL_TOUCH
) {
1265 * We should set the dirty bit only for FOLL_WRITE but
1266 * for now the dirty bit in the pmd is meaningless.
1267 * And if the dirty bit will become meaningful and
1268 * we'll only set it with FOLL_WRITE, an atomic
1269 * set_bit will be required on the pmd to set the
1270 * young bit, instead of the current set_pmd_at.
1272 _pmd
= pmd_mkyoung(pmd_mkdirty(*pmd
));
1273 if (pmdp_set_access_flags(vma
, addr
& HPAGE_PMD_MASK
,
1275 update_mmu_cache_pmd(vma
, addr
, pmd
);
1277 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1278 if (page
->mapping
&& trylock_page(page
)) {
1281 mlock_vma_page(page
);
1285 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
1286 VM_BUG_ON(!PageCompound(page
));
1287 if (flags
& FOLL_GET
)
1288 get_page_foll(page
);
1294 /* NUMA hinting page fault entry point for trans huge pmds */
1295 int do_huge_pmd_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1296 unsigned long addr
, pmd_t pmd
, pmd_t
*pmdp
)
1299 unsigned long haddr
= addr
& HPAGE_PMD_MASK
;
1301 int current_nid
= -1;
1304 spin_lock(&mm
->page_table_lock
);
1305 if (unlikely(!pmd_same(pmd
, *pmdp
)))
1308 page
= pmd_page(pmd
);
1310 current_nid
= page_to_nid(page
);
1311 count_vm_numa_event(NUMA_HINT_FAULTS
);
1312 if (current_nid
== numa_node_id())
1313 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
1315 target_nid
= mpol_misplaced(page
, vma
, haddr
);
1316 if (target_nid
== -1) {
1321 /* Acquire the page lock to serialise THP migrations */
1322 spin_unlock(&mm
->page_table_lock
);
1325 /* Confirm the PTE did not while locked */
1326 spin_lock(&mm
->page_table_lock
);
1327 if (unlikely(!pmd_same(pmd
, *pmdp
))) {
1332 spin_unlock(&mm
->page_table_lock
);
1334 /* Migrate the THP to the requested node */
1335 migrated
= migrate_misplaced_transhuge_page(mm
, vma
,
1336 pmdp
, pmd
, addr
, page
, target_nid
);
1340 task_numa_fault(target_nid
, HPAGE_PMD_NR
, true);
1344 spin_lock(&mm
->page_table_lock
);
1345 if (unlikely(!pmd_same(pmd
, *pmdp
)))
1348 pmd
= pmd_mknonnuma(pmd
);
1349 set_pmd_at(mm
, haddr
, pmdp
, pmd
);
1350 VM_BUG_ON(pmd_numa(*pmdp
));
1351 update_mmu_cache_pmd(vma
, addr
, pmdp
);
1353 spin_unlock(&mm
->page_table_lock
);
1354 if (current_nid
!= -1)
1355 task_numa_fault(current_nid
, HPAGE_PMD_NR
, false);
1359 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1360 pmd_t
*pmd
, unsigned long addr
)
1364 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
1369 * For architectures like ppc64 we look at deposited pgtable
1370 * when calling pmdp_get_and_clear. So do the
1371 * pgtable_trans_huge_withdraw after finishing pmdp related
1374 orig_pmd
= pmdp_get_and_clear(tlb
->mm
, addr
, pmd
);
1375 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1376 pgtable
= pgtable_trans_huge_withdraw(tlb
->mm
, pmd
);
1377 if (is_huge_zero_pmd(orig_pmd
)) {
1379 spin_unlock(&tlb
->mm
->page_table_lock
);
1380 put_huge_zero_page();
1382 page
= pmd_page(orig_pmd
);
1383 page_remove_rmap(page
);
1384 VM_BUG_ON(page_mapcount(page
) < 0);
1385 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1386 VM_BUG_ON(!PageHead(page
));
1388 spin_unlock(&tlb
->mm
->page_table_lock
);
1389 tlb_remove_page(tlb
, page
);
1391 pte_free(tlb
->mm
, pgtable
);
1397 int mincore_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1398 unsigned long addr
, unsigned long end
,
1403 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
1405 * All logical pages in the range are present
1406 * if backed by a huge page.
1408 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1409 memset(vec
, 1, (end
- addr
) >> PAGE_SHIFT
);
1416 int move_huge_pmd(struct vm_area_struct
*vma
, struct vm_area_struct
*new_vma
,
1417 unsigned long old_addr
,
1418 unsigned long new_addr
, unsigned long old_end
,
1419 pmd_t
*old_pmd
, pmd_t
*new_pmd
)
1424 struct mm_struct
*mm
= vma
->vm_mm
;
1426 if ((old_addr
& ~HPAGE_PMD_MASK
) ||
1427 (new_addr
& ~HPAGE_PMD_MASK
) ||
1428 old_end
- old_addr
< HPAGE_PMD_SIZE
||
1429 (new_vma
->vm_flags
& VM_NOHUGEPAGE
))
1433 * The destination pmd shouldn't be established, free_pgtables()
1434 * should have release it.
1436 if (WARN_ON(!pmd_none(*new_pmd
))) {
1437 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1441 ret
= __pmd_trans_huge_lock(old_pmd
, vma
);
1443 pmd
= pmdp_get_and_clear(mm
, old_addr
, old_pmd
);
1444 VM_BUG_ON(!pmd_none(*new_pmd
));
1445 set_pmd_at(mm
, new_addr
, new_pmd
, pmd_mksoft_dirty(pmd
));
1446 spin_unlock(&mm
->page_table_lock
);
1452 int change_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1453 unsigned long addr
, pgprot_t newprot
, int prot_numa
)
1455 struct mm_struct
*mm
= vma
->vm_mm
;
1458 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
1460 entry
= pmdp_get_and_clear(mm
, addr
, pmd
);
1462 entry
= pmd_modify(entry
, newprot
);
1463 BUG_ON(pmd_write(entry
));
1465 struct page
*page
= pmd_page(*pmd
);
1467 /* only check non-shared pages */
1468 if (page_mapcount(page
) == 1 &&
1470 entry
= pmd_mknuma(entry
);
1473 set_pmd_at(mm
, addr
, pmd
, entry
);
1474 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1482 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1483 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1485 * Note that if it returns 1, this routine returns without unlocking page
1486 * table locks. So callers must unlock them.
1488 int __pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
)
1490 spin_lock(&vma
->vm_mm
->page_table_lock
);
1491 if (likely(pmd_trans_huge(*pmd
))) {
1492 if (unlikely(pmd_trans_splitting(*pmd
))) {
1493 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1494 wait_split_huge_page(vma
->anon_vma
, pmd
);
1497 /* Thp mapped by 'pmd' is stable, so we can
1498 * handle it as it is. */
1502 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1506 pmd_t
*page_check_address_pmd(struct page
*page
,
1507 struct mm_struct
*mm
,
1508 unsigned long address
,
1509 enum page_check_address_pmd_flag flag
)
1511 pmd_t
*pmd
, *ret
= NULL
;
1513 if (address
& ~HPAGE_PMD_MASK
)
1516 pmd
= mm_find_pmd(mm
, address
);
1521 if (pmd_page(*pmd
) != page
)
1524 * split_vma() may create temporary aliased mappings. There is
1525 * no risk as long as all huge pmd are found and have their
1526 * splitting bit set before __split_huge_page_refcount
1527 * runs. Finding the same huge pmd more than once during the
1528 * same rmap walk is not a problem.
1530 if (flag
== PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG
&&
1531 pmd_trans_splitting(*pmd
))
1533 if (pmd_trans_huge(*pmd
)) {
1534 VM_BUG_ON(flag
== PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG
&&
1535 !pmd_trans_splitting(*pmd
));
1542 static int __split_huge_page_splitting(struct page
*page
,
1543 struct vm_area_struct
*vma
,
1544 unsigned long address
)
1546 struct mm_struct
*mm
= vma
->vm_mm
;
1549 /* For mmu_notifiers */
1550 const unsigned long mmun_start
= address
;
1551 const unsigned long mmun_end
= address
+ HPAGE_PMD_SIZE
;
1553 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1554 spin_lock(&mm
->page_table_lock
);
1555 pmd
= page_check_address_pmd(page
, mm
, address
,
1556 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG
);
1559 * We can't temporarily set the pmd to null in order
1560 * to split it, the pmd must remain marked huge at all
1561 * times or the VM won't take the pmd_trans_huge paths
1562 * and it won't wait on the anon_vma->root->rwsem to
1563 * serialize against split_huge_page*.
1565 pmdp_splitting_flush(vma
, address
, pmd
);
1568 spin_unlock(&mm
->page_table_lock
);
1569 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1574 static void __split_huge_page_refcount(struct page
*page
,
1575 struct list_head
*list
)
1578 struct zone
*zone
= page_zone(page
);
1579 struct lruvec
*lruvec
;
1582 /* prevent PageLRU to go away from under us, and freeze lru stats */
1583 spin_lock_irq(&zone
->lru_lock
);
1584 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1586 compound_lock(page
);
1587 /* complete memcg works before add pages to LRU */
1588 mem_cgroup_split_huge_fixup(page
);
1590 for (i
= HPAGE_PMD_NR
- 1; i
>= 1; i
--) {
1591 struct page
*page_tail
= page
+ i
;
1593 /* tail_page->_mapcount cannot change */
1594 BUG_ON(page_mapcount(page_tail
) < 0);
1595 tail_count
+= page_mapcount(page_tail
);
1596 /* check for overflow */
1597 BUG_ON(tail_count
< 0);
1598 BUG_ON(atomic_read(&page_tail
->_count
) != 0);
1600 * tail_page->_count is zero and not changing from
1601 * under us. But get_page_unless_zero() may be running
1602 * from under us on the tail_page. If we used
1603 * atomic_set() below instead of atomic_add(), we
1604 * would then run atomic_set() concurrently with
1605 * get_page_unless_zero(), and atomic_set() is
1606 * implemented in C not using locked ops. spin_unlock
1607 * on x86 sometime uses locked ops because of PPro
1608 * errata 66, 92, so unless somebody can guarantee
1609 * atomic_set() here would be safe on all archs (and
1610 * not only on x86), it's safer to use atomic_add().
1612 atomic_add(page_mapcount(page
) + page_mapcount(page_tail
) + 1,
1613 &page_tail
->_count
);
1615 /* after clearing PageTail the gup refcount can be released */
1619 * retain hwpoison flag of the poisoned tail page:
1620 * fix for the unsuitable process killed on Guest Machine(KVM)
1621 * by the memory-failure.
1623 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
| __PG_HWPOISON
;
1624 page_tail
->flags
|= (page
->flags
&
1625 ((1L << PG_referenced
) |
1626 (1L << PG_swapbacked
) |
1627 (1L << PG_mlocked
) |
1628 (1L << PG_uptodate
) |
1630 (1L << PG_unevictable
)));
1631 page_tail
->flags
|= (1L << PG_dirty
);
1633 /* clear PageTail before overwriting first_page */
1637 * __split_huge_page_splitting() already set the
1638 * splitting bit in all pmd that could map this
1639 * hugepage, that will ensure no CPU can alter the
1640 * mapcount on the head page. The mapcount is only
1641 * accounted in the head page and it has to be
1642 * transferred to all tail pages in the below code. So
1643 * for this code to be safe, the split the mapcount
1644 * can't change. But that doesn't mean userland can't
1645 * keep changing and reading the page contents while
1646 * we transfer the mapcount, so the pmd splitting
1647 * status is achieved setting a reserved bit in the
1648 * pmd, not by clearing the present bit.
1650 page_tail
->_mapcount
= page
->_mapcount
;
1652 BUG_ON(page_tail
->mapping
);
1653 page_tail
->mapping
= page
->mapping
;
1655 page_tail
->index
= page
->index
+ i
;
1656 page_nid_xchg_last(page_tail
, page_nid_last(page
));
1658 BUG_ON(!PageAnon(page_tail
));
1659 BUG_ON(!PageUptodate(page_tail
));
1660 BUG_ON(!PageDirty(page_tail
));
1661 BUG_ON(!PageSwapBacked(page_tail
));
1663 lru_add_page_tail(page
, page_tail
, lruvec
, list
);
1665 atomic_sub(tail_count
, &page
->_count
);
1666 BUG_ON(atomic_read(&page
->_count
) <= 0);
1668 __mod_zone_page_state(zone
, NR_ANON_TRANSPARENT_HUGEPAGES
, -1);
1669 __mod_zone_page_state(zone
, NR_ANON_PAGES
, HPAGE_PMD_NR
);
1671 ClearPageCompound(page
);
1672 compound_unlock(page
);
1673 spin_unlock_irq(&zone
->lru_lock
);
1675 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
1676 struct page
*page_tail
= page
+ i
;
1677 BUG_ON(page_count(page_tail
) <= 0);
1679 * Tail pages may be freed if there wasn't any mapping
1680 * like if add_to_swap() is running on a lru page that
1681 * had its mapping zapped. And freeing these pages
1682 * requires taking the lru_lock so we do the put_page
1683 * of the tail pages after the split is complete.
1685 put_page(page_tail
);
1689 * Only the head page (now become a regular page) is required
1690 * to be pinned by the caller.
1692 BUG_ON(page_count(page
) <= 0);
1695 static int __split_huge_page_map(struct page
*page
,
1696 struct vm_area_struct
*vma
,
1697 unsigned long address
)
1699 struct mm_struct
*mm
= vma
->vm_mm
;
1703 unsigned long haddr
;
1705 spin_lock(&mm
->page_table_lock
);
1706 pmd
= page_check_address_pmd(page
, mm
, address
,
1707 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG
);
1709 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1710 pmd_populate(mm
, &_pmd
, pgtable
);
1713 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1715 BUG_ON(PageCompound(page
+i
));
1716 entry
= mk_pte(page
+ i
, vma
->vm_page_prot
);
1717 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1718 if (!pmd_write(*pmd
))
1719 entry
= pte_wrprotect(entry
);
1721 BUG_ON(page_mapcount(page
) != 1);
1722 if (!pmd_young(*pmd
))
1723 entry
= pte_mkold(entry
);
1725 entry
= pte_mknuma(entry
);
1726 pte
= pte_offset_map(&_pmd
, haddr
);
1727 BUG_ON(!pte_none(*pte
));
1728 set_pte_at(mm
, haddr
, pte
, entry
);
1732 smp_wmb(); /* make pte visible before pmd */
1734 * Up to this point the pmd is present and huge and
1735 * userland has the whole access to the hugepage
1736 * during the split (which happens in place). If we
1737 * overwrite the pmd with the not-huge version
1738 * pointing to the pte here (which of course we could
1739 * if all CPUs were bug free), userland could trigger
1740 * a small page size TLB miss on the small sized TLB
1741 * while the hugepage TLB entry is still established
1742 * in the huge TLB. Some CPU doesn't like that. See
1743 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1744 * Erratum 383 on page 93. Intel should be safe but is
1745 * also warns that it's only safe if the permission
1746 * and cache attributes of the two entries loaded in
1747 * the two TLB is identical (which should be the case
1748 * here). But it is generally safer to never allow
1749 * small and huge TLB entries for the same virtual
1750 * address to be loaded simultaneously. So instead of
1751 * doing "pmd_populate(); flush_tlb_range();" we first
1752 * mark the current pmd notpresent (atomically because
1753 * here the pmd_trans_huge and pmd_trans_splitting
1754 * must remain set at all times on the pmd until the
1755 * split is complete for this pmd), then we flush the
1756 * SMP TLB and finally we write the non-huge version
1757 * of the pmd entry with pmd_populate.
1759 pmdp_invalidate(vma
, address
, pmd
);
1760 pmd_populate(mm
, pmd
, pgtable
);
1763 spin_unlock(&mm
->page_table_lock
);
1768 /* must be called with anon_vma->root->rwsem held */
1769 static void __split_huge_page(struct page
*page
,
1770 struct anon_vma
*anon_vma
,
1771 struct list_head
*list
)
1773 int mapcount
, mapcount2
;
1774 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1775 struct anon_vma_chain
*avc
;
1777 BUG_ON(!PageHead(page
));
1778 BUG_ON(PageTail(page
));
1781 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
1782 struct vm_area_struct
*vma
= avc
->vma
;
1783 unsigned long addr
= vma_address(page
, vma
);
1784 BUG_ON(is_vma_temporary_stack(vma
));
1785 mapcount
+= __split_huge_page_splitting(page
, vma
, addr
);
1788 * It is critical that new vmas are added to the tail of the
1789 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1790 * and establishes a child pmd before
1791 * __split_huge_page_splitting() freezes the parent pmd (so if
1792 * we fail to prevent copy_huge_pmd() from running until the
1793 * whole __split_huge_page() is complete), we will still see
1794 * the newly established pmd of the child later during the
1795 * walk, to be able to set it as pmd_trans_splitting too.
1797 if (mapcount
!= page_mapcount(page
))
1798 printk(KERN_ERR
"mapcount %d page_mapcount %d\n",
1799 mapcount
, page_mapcount(page
));
1800 BUG_ON(mapcount
!= page_mapcount(page
));
1802 __split_huge_page_refcount(page
, list
);
1805 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
1806 struct vm_area_struct
*vma
= avc
->vma
;
1807 unsigned long addr
= vma_address(page
, vma
);
1808 BUG_ON(is_vma_temporary_stack(vma
));
1809 mapcount2
+= __split_huge_page_map(page
, vma
, addr
);
1811 if (mapcount
!= mapcount2
)
1812 printk(KERN_ERR
"mapcount %d mapcount2 %d page_mapcount %d\n",
1813 mapcount
, mapcount2
, page_mapcount(page
));
1814 BUG_ON(mapcount
!= mapcount2
);
1818 * Split a hugepage into normal pages. This doesn't change the position of head
1819 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1820 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1821 * from the hugepage.
1822 * Return 0 if the hugepage is split successfully otherwise return 1.
1824 int split_huge_page_to_list(struct page
*page
, struct list_head
*list
)
1826 struct anon_vma
*anon_vma
;
1829 BUG_ON(is_huge_zero_page(page
));
1830 BUG_ON(!PageAnon(page
));
1833 * The caller does not necessarily hold an mmap_sem that would prevent
1834 * the anon_vma disappearing so we first we take a reference to it
1835 * and then lock the anon_vma for write. This is similar to
1836 * page_lock_anon_vma_read except the write lock is taken to serialise
1837 * against parallel split or collapse operations.
1839 anon_vma
= page_get_anon_vma(page
);
1842 anon_vma_lock_write(anon_vma
);
1845 if (!PageCompound(page
))
1848 BUG_ON(!PageSwapBacked(page
));
1849 __split_huge_page(page
, anon_vma
, list
);
1850 count_vm_event(THP_SPLIT
);
1852 BUG_ON(PageCompound(page
));
1854 anon_vma_unlock_write(anon_vma
);
1855 put_anon_vma(anon_vma
);
1860 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1862 int hugepage_madvise(struct vm_area_struct
*vma
,
1863 unsigned long *vm_flags
, int advice
)
1865 struct mm_struct
*mm
= vma
->vm_mm
;
1870 * Be somewhat over-protective like KSM for now!
1872 if (*vm_flags
& (VM_HUGEPAGE
| VM_NO_THP
))
1874 if (mm
->def_flags
& VM_NOHUGEPAGE
)
1876 *vm_flags
&= ~VM_NOHUGEPAGE
;
1877 *vm_flags
|= VM_HUGEPAGE
;
1879 * If the vma become good for khugepaged to scan,
1880 * register it here without waiting a page fault that
1881 * may not happen any time soon.
1883 if (unlikely(khugepaged_enter_vma_merge(vma
)))
1886 case MADV_NOHUGEPAGE
:
1888 * Be somewhat over-protective like KSM for now!
1890 if (*vm_flags
& (VM_NOHUGEPAGE
| VM_NO_THP
))
1892 *vm_flags
&= ~VM_HUGEPAGE
;
1893 *vm_flags
|= VM_NOHUGEPAGE
;
1895 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1896 * this vma even if we leave the mm registered in khugepaged if
1897 * it got registered before VM_NOHUGEPAGE was set.
1905 static int __init
khugepaged_slab_init(void)
1907 mm_slot_cache
= kmem_cache_create("khugepaged_mm_slot",
1908 sizeof(struct mm_slot
),
1909 __alignof__(struct mm_slot
), 0, NULL
);
1916 static inline struct mm_slot
*alloc_mm_slot(void)
1918 if (!mm_slot_cache
) /* initialization failed */
1920 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
1923 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
1925 kmem_cache_free(mm_slot_cache
, mm_slot
);
1928 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
1930 struct mm_slot
*mm_slot
;
1932 hash_for_each_possible(mm_slots_hash
, mm_slot
, hash
, (unsigned long)mm
)
1933 if (mm
== mm_slot
->mm
)
1939 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
1940 struct mm_slot
*mm_slot
)
1943 hash_add(mm_slots_hash
, &mm_slot
->hash
, (long)mm
);
1946 static inline int khugepaged_test_exit(struct mm_struct
*mm
)
1948 return atomic_read(&mm
->mm_users
) == 0;
1951 int __khugepaged_enter(struct mm_struct
*mm
)
1953 struct mm_slot
*mm_slot
;
1956 mm_slot
= alloc_mm_slot();
1960 /* __khugepaged_exit() must not run from under us */
1961 VM_BUG_ON(khugepaged_test_exit(mm
));
1962 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE
, &mm
->flags
))) {
1963 free_mm_slot(mm_slot
);
1967 spin_lock(&khugepaged_mm_lock
);
1968 insert_to_mm_slots_hash(mm
, mm_slot
);
1970 * Insert just behind the scanning cursor, to let the area settle
1973 wakeup
= list_empty(&khugepaged_scan
.mm_head
);
1974 list_add_tail(&mm_slot
->mm_node
, &khugepaged_scan
.mm_head
);
1975 spin_unlock(&khugepaged_mm_lock
);
1977 atomic_inc(&mm
->mm_count
);
1979 wake_up_interruptible(&khugepaged_wait
);
1984 int khugepaged_enter_vma_merge(struct vm_area_struct
*vma
)
1986 unsigned long hstart
, hend
;
1989 * Not yet faulted in so we will register later in the
1990 * page fault if needed.
1994 /* khugepaged not yet working on file or special mappings */
1996 VM_BUG_ON(vma
->vm_flags
& VM_NO_THP
);
1997 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
1998 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2000 return khugepaged_enter(vma
);
2004 void __khugepaged_exit(struct mm_struct
*mm
)
2006 struct mm_slot
*mm_slot
;
2009 spin_lock(&khugepaged_mm_lock
);
2010 mm_slot
= get_mm_slot(mm
);
2011 if (mm_slot
&& khugepaged_scan
.mm_slot
!= mm_slot
) {
2012 hash_del(&mm_slot
->hash
);
2013 list_del(&mm_slot
->mm_node
);
2016 spin_unlock(&khugepaged_mm_lock
);
2019 clear_bit(MMF_VM_HUGEPAGE
, &mm
->flags
);
2020 free_mm_slot(mm_slot
);
2022 } else if (mm_slot
) {
2024 * This is required to serialize against
2025 * khugepaged_test_exit() (which is guaranteed to run
2026 * under mmap sem read mode). Stop here (after we
2027 * return all pagetables will be destroyed) until
2028 * khugepaged has finished working on the pagetables
2029 * under the mmap_sem.
2031 down_write(&mm
->mmap_sem
);
2032 up_write(&mm
->mmap_sem
);
2036 static void release_pte_page(struct page
*page
)
2038 /* 0 stands for page_is_file_cache(page) == false */
2039 dec_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
2041 putback_lru_page(page
);
2044 static void release_pte_pages(pte_t
*pte
, pte_t
*_pte
)
2046 while (--_pte
>= pte
) {
2047 pte_t pteval
= *_pte
;
2048 if (!pte_none(pteval
))
2049 release_pte_page(pte_page(pteval
));
2053 static int __collapse_huge_page_isolate(struct vm_area_struct
*vma
,
2054 unsigned long address
,
2059 int referenced
= 0, none
= 0;
2060 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
2061 _pte
++, address
+= PAGE_SIZE
) {
2062 pte_t pteval
= *_pte
;
2063 if (pte_none(pteval
)) {
2064 if (++none
<= khugepaged_max_ptes_none
)
2069 if (!pte_present(pteval
) || !pte_write(pteval
))
2071 page
= vm_normal_page(vma
, address
, pteval
);
2072 if (unlikely(!page
))
2075 VM_BUG_ON(PageCompound(page
));
2076 BUG_ON(!PageAnon(page
));
2077 VM_BUG_ON(!PageSwapBacked(page
));
2079 /* cannot use mapcount: can't collapse if there's a gup pin */
2080 if (page_count(page
) != 1)
2083 * We can do it before isolate_lru_page because the
2084 * page can't be freed from under us. NOTE: PG_lock
2085 * is needed to serialize against split_huge_page
2086 * when invoked from the VM.
2088 if (!trylock_page(page
))
2091 * Isolate the page to avoid collapsing an hugepage
2092 * currently in use by the VM.
2094 if (isolate_lru_page(page
)) {
2098 /* 0 stands for page_is_file_cache(page) == false */
2099 inc_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
2100 VM_BUG_ON(!PageLocked(page
));
2101 VM_BUG_ON(PageLRU(page
));
2103 /* If there is no mapped pte young don't collapse the page */
2104 if (pte_young(pteval
) || PageReferenced(page
) ||
2105 mmu_notifier_test_young(vma
->vm_mm
, address
))
2108 if (likely(referenced
))
2111 release_pte_pages(pte
, _pte
);
2115 static void __collapse_huge_page_copy(pte_t
*pte
, struct page
*page
,
2116 struct vm_area_struct
*vma
,
2117 unsigned long address
,
2121 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
; _pte
++) {
2122 pte_t pteval
= *_pte
;
2123 struct page
*src_page
;
2125 if (pte_none(pteval
)) {
2126 clear_user_highpage(page
, address
);
2127 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, 1);
2129 src_page
= pte_page(pteval
);
2130 copy_user_highpage(page
, src_page
, address
, vma
);
2131 VM_BUG_ON(page_mapcount(src_page
) != 1);
2132 release_pte_page(src_page
);
2134 * ptl mostly unnecessary, but preempt has to
2135 * be disabled to update the per-cpu stats
2136 * inside page_remove_rmap().
2140 * paravirt calls inside pte_clear here are
2143 pte_clear(vma
->vm_mm
, address
, _pte
);
2144 page_remove_rmap(src_page
);
2146 free_page_and_swap_cache(src_page
);
2149 address
+= PAGE_SIZE
;
2154 static void khugepaged_alloc_sleep(void)
2156 wait_event_freezable_timeout(khugepaged_wait
, false,
2157 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs
));
2161 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
2163 if (IS_ERR(*hpage
)) {
2169 khugepaged_alloc_sleep();
2170 } else if (*hpage
) {
2179 *khugepaged_alloc_page(struct page
**hpage
, struct mm_struct
*mm
,
2180 struct vm_area_struct
*vma
, unsigned long address
,
2185 * Allocate the page while the vma is still valid and under
2186 * the mmap_sem read mode so there is no memory allocation
2187 * later when we take the mmap_sem in write mode. This is more
2188 * friendly behavior (OTOH it may actually hide bugs) to
2189 * filesystems in userland with daemons allocating memory in
2190 * the userland I/O paths. Allocating memory with the
2191 * mmap_sem in read mode is good idea also to allow greater
2194 *hpage
= alloc_hugepage_vma(khugepaged_defrag(), vma
, address
,
2195 node
, __GFP_OTHER_NODE
);
2198 * After allocating the hugepage, release the mmap_sem read lock in
2199 * preparation for taking it in write mode.
2201 up_read(&mm
->mmap_sem
);
2202 if (unlikely(!*hpage
)) {
2203 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
2204 *hpage
= ERR_PTR(-ENOMEM
);
2208 count_vm_event(THP_COLLAPSE_ALLOC
);
2212 static struct page
*khugepaged_alloc_hugepage(bool *wait
)
2217 hpage
= alloc_hugepage(khugepaged_defrag());
2219 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
2224 khugepaged_alloc_sleep();
2226 count_vm_event(THP_COLLAPSE_ALLOC
);
2227 } while (unlikely(!hpage
) && likely(khugepaged_enabled()));
2232 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
2235 *hpage
= khugepaged_alloc_hugepage(wait
);
2237 if (unlikely(!*hpage
))
2244 *khugepaged_alloc_page(struct page
**hpage
, struct mm_struct
*mm
,
2245 struct vm_area_struct
*vma
, unsigned long address
,
2248 up_read(&mm
->mmap_sem
);
2254 static bool hugepage_vma_check(struct vm_area_struct
*vma
)
2256 if ((!(vma
->vm_flags
& VM_HUGEPAGE
) && !khugepaged_always()) ||
2257 (vma
->vm_flags
& VM_NOHUGEPAGE
))
2260 if (!vma
->anon_vma
|| vma
->vm_ops
)
2262 if (is_vma_temporary_stack(vma
))
2264 VM_BUG_ON(vma
->vm_flags
& VM_NO_THP
);
2268 static void collapse_huge_page(struct mm_struct
*mm
,
2269 unsigned long address
,
2270 struct page
**hpage
,
2271 struct vm_area_struct
*vma
,
2277 struct page
*new_page
;
2280 unsigned long hstart
, hend
;
2281 unsigned long mmun_start
; /* For mmu_notifiers */
2282 unsigned long mmun_end
; /* For mmu_notifiers */
2284 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2286 /* release the mmap_sem read lock. */
2287 new_page
= khugepaged_alloc_page(hpage
, mm
, vma
, address
, node
);
2291 if (unlikely(mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
)))
2295 * Prevent all access to pagetables with the exception of
2296 * gup_fast later hanlded by the ptep_clear_flush and the VM
2297 * handled by the anon_vma lock + PG_lock.
2299 down_write(&mm
->mmap_sem
);
2300 if (unlikely(khugepaged_test_exit(mm
)))
2303 vma
= find_vma(mm
, address
);
2306 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2307 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2308 if (address
< hstart
|| address
+ HPAGE_PMD_SIZE
> hend
)
2310 if (!hugepage_vma_check(vma
))
2312 pmd
= mm_find_pmd(mm
, address
);
2315 if (pmd_trans_huge(*pmd
))
2318 anon_vma_lock_write(vma
->anon_vma
);
2320 pte
= pte_offset_map(pmd
, address
);
2321 ptl
= pte_lockptr(mm
, pmd
);
2323 mmun_start
= address
;
2324 mmun_end
= address
+ HPAGE_PMD_SIZE
;
2325 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2326 spin_lock(&mm
->page_table_lock
); /* probably unnecessary */
2328 * After this gup_fast can't run anymore. This also removes
2329 * any huge TLB entry from the CPU so we won't allow
2330 * huge and small TLB entries for the same virtual address
2331 * to avoid the risk of CPU bugs in that area.
2333 _pmd
= pmdp_clear_flush(vma
, address
, pmd
);
2334 spin_unlock(&mm
->page_table_lock
);
2335 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2338 isolated
= __collapse_huge_page_isolate(vma
, address
, pte
);
2341 if (unlikely(!isolated
)) {
2343 spin_lock(&mm
->page_table_lock
);
2344 BUG_ON(!pmd_none(*pmd
));
2346 * We can only use set_pmd_at when establishing
2347 * hugepmds and never for establishing regular pmds that
2348 * points to regular pagetables. Use pmd_populate for that
2350 pmd_populate(mm
, pmd
, pmd_pgtable(_pmd
));
2351 spin_unlock(&mm
->page_table_lock
);
2352 anon_vma_unlock_write(vma
->anon_vma
);
2357 * All pages are isolated and locked so anon_vma rmap
2358 * can't run anymore.
2360 anon_vma_unlock_write(vma
->anon_vma
);
2362 __collapse_huge_page_copy(pte
, new_page
, vma
, address
, ptl
);
2364 __SetPageUptodate(new_page
);
2365 pgtable
= pmd_pgtable(_pmd
);
2367 _pmd
= mk_huge_pmd(new_page
, vma
);
2370 * spin_lock() below is not the equivalent of smp_wmb(), so
2371 * this is needed to avoid the copy_huge_page writes to become
2372 * visible after the set_pmd_at() write.
2376 spin_lock(&mm
->page_table_lock
);
2377 BUG_ON(!pmd_none(*pmd
));
2378 page_add_new_anon_rmap(new_page
, vma
, address
);
2379 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
2380 set_pmd_at(mm
, address
, pmd
, _pmd
);
2381 update_mmu_cache_pmd(vma
, address
, pmd
);
2382 spin_unlock(&mm
->page_table_lock
);
2386 khugepaged_pages_collapsed
++;
2388 up_write(&mm
->mmap_sem
);
2392 mem_cgroup_uncharge_page(new_page
);
2396 static int khugepaged_scan_pmd(struct mm_struct
*mm
,
2397 struct vm_area_struct
*vma
,
2398 unsigned long address
,
2399 struct page
**hpage
)
2403 int ret
= 0, referenced
= 0, none
= 0;
2405 unsigned long _address
;
2407 int node
= NUMA_NO_NODE
;
2409 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2411 pmd
= mm_find_pmd(mm
, address
);
2414 if (pmd_trans_huge(*pmd
))
2417 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2418 for (_address
= address
, _pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
2419 _pte
++, _address
+= PAGE_SIZE
) {
2420 pte_t pteval
= *_pte
;
2421 if (pte_none(pteval
)) {
2422 if (++none
<= khugepaged_max_ptes_none
)
2427 if (!pte_present(pteval
) || !pte_write(pteval
))
2429 page
= vm_normal_page(vma
, _address
, pteval
);
2430 if (unlikely(!page
))
2433 * Chose the node of the first page. This could
2434 * be more sophisticated and look at more pages,
2435 * but isn't for now.
2437 if (node
== NUMA_NO_NODE
)
2438 node
= page_to_nid(page
);
2439 VM_BUG_ON(PageCompound(page
));
2440 if (!PageLRU(page
) || PageLocked(page
) || !PageAnon(page
))
2442 /* cannot use mapcount: can't collapse if there's a gup pin */
2443 if (page_count(page
) != 1)
2445 if (pte_young(pteval
) || PageReferenced(page
) ||
2446 mmu_notifier_test_young(vma
->vm_mm
, address
))
2452 pte_unmap_unlock(pte
, ptl
);
2454 /* collapse_huge_page will return with the mmap_sem released */
2455 collapse_huge_page(mm
, address
, hpage
, vma
, node
);
2460 static void collect_mm_slot(struct mm_slot
*mm_slot
)
2462 struct mm_struct
*mm
= mm_slot
->mm
;
2464 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2466 if (khugepaged_test_exit(mm
)) {
2468 hash_del(&mm_slot
->hash
);
2469 list_del(&mm_slot
->mm_node
);
2472 * Not strictly needed because the mm exited already.
2474 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2477 /* khugepaged_mm_lock actually not necessary for the below */
2478 free_mm_slot(mm_slot
);
2483 static unsigned int khugepaged_scan_mm_slot(unsigned int pages
,
2484 struct page
**hpage
)
2485 __releases(&khugepaged_mm_lock
)
2486 __acquires(&khugepaged_mm_lock
)
2488 struct mm_slot
*mm_slot
;
2489 struct mm_struct
*mm
;
2490 struct vm_area_struct
*vma
;
2494 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2496 if (khugepaged_scan
.mm_slot
)
2497 mm_slot
= khugepaged_scan
.mm_slot
;
2499 mm_slot
= list_entry(khugepaged_scan
.mm_head
.next
,
2500 struct mm_slot
, mm_node
);
2501 khugepaged_scan
.address
= 0;
2502 khugepaged_scan
.mm_slot
= mm_slot
;
2504 spin_unlock(&khugepaged_mm_lock
);
2507 down_read(&mm
->mmap_sem
);
2508 if (unlikely(khugepaged_test_exit(mm
)))
2511 vma
= find_vma(mm
, khugepaged_scan
.address
);
2514 for (; vma
; vma
= vma
->vm_next
) {
2515 unsigned long hstart
, hend
;
2518 if (unlikely(khugepaged_test_exit(mm
))) {
2522 if (!hugepage_vma_check(vma
)) {
2527 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2528 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2531 if (khugepaged_scan
.address
> hend
)
2533 if (khugepaged_scan
.address
< hstart
)
2534 khugepaged_scan
.address
= hstart
;
2535 VM_BUG_ON(khugepaged_scan
.address
& ~HPAGE_PMD_MASK
);
2537 while (khugepaged_scan
.address
< hend
) {
2540 if (unlikely(khugepaged_test_exit(mm
)))
2541 goto breakouterloop
;
2543 VM_BUG_ON(khugepaged_scan
.address
< hstart
||
2544 khugepaged_scan
.address
+ HPAGE_PMD_SIZE
>
2546 ret
= khugepaged_scan_pmd(mm
, vma
,
2547 khugepaged_scan
.address
,
2549 /* move to next address */
2550 khugepaged_scan
.address
+= HPAGE_PMD_SIZE
;
2551 progress
+= HPAGE_PMD_NR
;
2553 /* we released mmap_sem so break loop */
2554 goto breakouterloop_mmap_sem
;
2555 if (progress
>= pages
)
2556 goto breakouterloop
;
2560 up_read(&mm
->mmap_sem
); /* exit_mmap will destroy ptes after this */
2561 breakouterloop_mmap_sem
:
2563 spin_lock(&khugepaged_mm_lock
);
2564 VM_BUG_ON(khugepaged_scan
.mm_slot
!= mm_slot
);
2566 * Release the current mm_slot if this mm is about to die, or
2567 * if we scanned all vmas of this mm.
2569 if (khugepaged_test_exit(mm
) || !vma
) {
2571 * Make sure that if mm_users is reaching zero while
2572 * khugepaged runs here, khugepaged_exit will find
2573 * mm_slot not pointing to the exiting mm.
2575 if (mm_slot
->mm_node
.next
!= &khugepaged_scan
.mm_head
) {
2576 khugepaged_scan
.mm_slot
= list_entry(
2577 mm_slot
->mm_node
.next
,
2578 struct mm_slot
, mm_node
);
2579 khugepaged_scan
.address
= 0;
2581 khugepaged_scan
.mm_slot
= NULL
;
2582 khugepaged_full_scans
++;
2585 collect_mm_slot(mm_slot
);
2591 static int khugepaged_has_work(void)
2593 return !list_empty(&khugepaged_scan
.mm_head
) &&
2594 khugepaged_enabled();
2597 static int khugepaged_wait_event(void)
2599 return !list_empty(&khugepaged_scan
.mm_head
) ||
2600 kthread_should_stop();
2603 static void khugepaged_do_scan(void)
2605 struct page
*hpage
= NULL
;
2606 unsigned int progress
= 0, pass_through_head
= 0;
2607 unsigned int pages
= khugepaged_pages_to_scan
;
2610 barrier(); /* write khugepaged_pages_to_scan to local stack */
2612 while (progress
< pages
) {
2613 if (!khugepaged_prealloc_page(&hpage
, &wait
))
2618 if (unlikely(kthread_should_stop() || freezing(current
)))
2621 spin_lock(&khugepaged_mm_lock
);
2622 if (!khugepaged_scan
.mm_slot
)
2623 pass_through_head
++;
2624 if (khugepaged_has_work() &&
2625 pass_through_head
< 2)
2626 progress
+= khugepaged_scan_mm_slot(pages
- progress
,
2630 spin_unlock(&khugepaged_mm_lock
);
2633 if (!IS_ERR_OR_NULL(hpage
))
2637 static void khugepaged_wait_work(void)
2641 if (khugepaged_has_work()) {
2642 if (!khugepaged_scan_sleep_millisecs
)
2645 wait_event_freezable_timeout(khugepaged_wait
,
2646 kthread_should_stop(),
2647 msecs_to_jiffies(khugepaged_scan_sleep_millisecs
));
2651 if (khugepaged_enabled())
2652 wait_event_freezable(khugepaged_wait
, khugepaged_wait_event());
2655 static int khugepaged(void *none
)
2657 struct mm_slot
*mm_slot
;
2660 set_user_nice(current
, 19);
2662 while (!kthread_should_stop()) {
2663 khugepaged_do_scan();
2664 khugepaged_wait_work();
2667 spin_lock(&khugepaged_mm_lock
);
2668 mm_slot
= khugepaged_scan
.mm_slot
;
2669 khugepaged_scan
.mm_slot
= NULL
;
2671 collect_mm_slot(mm_slot
);
2672 spin_unlock(&khugepaged_mm_lock
);
2676 static void __split_huge_zero_page_pmd(struct vm_area_struct
*vma
,
2677 unsigned long haddr
, pmd_t
*pmd
)
2679 struct mm_struct
*mm
= vma
->vm_mm
;
2684 pmdp_clear_flush(vma
, haddr
, pmd
);
2685 /* leave pmd empty until pte is filled */
2687 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2688 pmd_populate(mm
, &_pmd
, pgtable
);
2690 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
2692 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
2693 entry
= pte_mkspecial(entry
);
2694 pte
= pte_offset_map(&_pmd
, haddr
);
2695 VM_BUG_ON(!pte_none(*pte
));
2696 set_pte_at(mm
, haddr
, pte
, entry
);
2699 smp_wmb(); /* make pte visible before pmd */
2700 pmd_populate(mm
, pmd
, pgtable
);
2701 put_huge_zero_page();
2704 void __split_huge_page_pmd(struct vm_area_struct
*vma
, unsigned long address
,
2708 struct mm_struct
*mm
= vma
->vm_mm
;
2709 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
2710 unsigned long mmun_start
; /* For mmu_notifiers */
2711 unsigned long mmun_end
; /* For mmu_notifiers */
2713 BUG_ON(vma
->vm_start
> haddr
|| vma
->vm_end
< haddr
+ HPAGE_PMD_SIZE
);
2716 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
2717 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2718 spin_lock(&mm
->page_table_lock
);
2719 if (unlikely(!pmd_trans_huge(*pmd
))) {
2720 spin_unlock(&mm
->page_table_lock
);
2721 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2724 if (is_huge_zero_pmd(*pmd
)) {
2725 __split_huge_zero_page_pmd(vma
, haddr
, pmd
);
2726 spin_unlock(&mm
->page_table_lock
);
2727 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2730 page
= pmd_page(*pmd
);
2731 VM_BUG_ON(!page_count(page
));
2733 spin_unlock(&mm
->page_table_lock
);
2734 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2736 split_huge_page(page
);
2739 BUG_ON(pmd_trans_huge(*pmd
));
2742 void split_huge_page_pmd_mm(struct mm_struct
*mm
, unsigned long address
,
2745 struct vm_area_struct
*vma
;
2747 vma
= find_vma(mm
, address
);
2748 BUG_ON(vma
== NULL
);
2749 split_huge_page_pmd(vma
, address
, pmd
);
2752 static void split_huge_page_address(struct mm_struct
*mm
,
2753 unsigned long address
)
2757 VM_BUG_ON(!(address
& ~HPAGE_PMD_MASK
));
2759 pmd
= mm_find_pmd(mm
, address
);
2763 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2764 * materialize from under us.
2766 split_huge_page_pmd_mm(mm
, address
, pmd
);
2769 void __vma_adjust_trans_huge(struct vm_area_struct
*vma
,
2770 unsigned long start
,
2775 * If the new start address isn't hpage aligned and it could
2776 * previously contain an hugepage: check if we need to split
2779 if (start
& ~HPAGE_PMD_MASK
&&
2780 (start
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2781 (start
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2782 split_huge_page_address(vma
->vm_mm
, start
);
2785 * If the new end address isn't hpage aligned and it could
2786 * previously contain an hugepage: check if we need to split
2789 if (end
& ~HPAGE_PMD_MASK
&&
2790 (end
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2791 (end
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2792 split_huge_page_address(vma
->vm_mm
, end
);
2795 * If we're also updating the vma->vm_next->vm_start, if the new
2796 * vm_next->vm_start isn't page aligned and it could previously
2797 * contain an hugepage: check if we need to split an huge pmd.
2799 if (adjust_next
> 0) {
2800 struct vm_area_struct
*next
= vma
->vm_next
;
2801 unsigned long nstart
= next
->vm_start
;
2802 nstart
+= adjust_next
<< PAGE_SHIFT
;
2803 if (nstart
& ~HPAGE_PMD_MASK
&&
2804 (nstart
& HPAGE_PMD_MASK
) >= next
->vm_start
&&
2805 (nstart
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= next
->vm_end
)
2806 split_huge_page_address(next
->vm_mm
, nstart
);