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.
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/sched.h>
12 #include <linux/highmem.h>
13 #include <linux/hugetlb.h>
14 #include <linux/mmu_notifier.h>
15 #include <linux/rmap.h>
16 #include <linux/swap.h>
17 #include <linux/shrinker.h>
18 #include <linux/mm_inline.h>
19 #include <linux/swapops.h>
20 #include <linux/dax.h>
21 #include <linux/kthread.h>
22 #include <linux/khugepaged.h>
23 #include <linux/freezer.h>
24 #include <linux/pfn_t.h>
25 #include <linux/mman.h>
26 #include <linux/memremap.h>
27 #include <linux/pagemap.h>
28 #include <linux/debugfs.h>
29 #include <linux/migrate.h>
30 #include <linux/hashtable.h>
31 #include <linux/userfaultfd_k.h>
32 #include <linux/page_idle.h>
35 #include <asm/pgalloc.h>
45 SCAN_NO_REFERENCED_PAGE
,
59 SCAN_ALLOC_HUGE_PAGE_FAIL
,
60 SCAN_CGROUP_CHARGE_FAIL
,
64 #define CREATE_TRACE_POINTS
65 #include <trace/events/huge_memory.h>
68 * By default transparent hugepage support is disabled in order that avoid
69 * to risk increase the memory footprint of applications without a guaranteed
70 * benefit. When transparent hugepage support is enabled, is for all mappings,
71 * and khugepaged scans all mappings.
72 * Defrag is invoked by khugepaged hugepage allocations and by page faults
73 * for all hugepage allocations.
75 unsigned long transparent_hugepage_flags __read_mostly
=
76 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
77 (1<<TRANSPARENT_HUGEPAGE_FLAG
)|
79 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
80 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
)|
82 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
)|
83 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
)|
84 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
86 /* default scan 8*512 pte (or vmas) every 30 second */
87 static unsigned int khugepaged_pages_to_scan __read_mostly
;
88 static unsigned int khugepaged_pages_collapsed
;
89 static unsigned int khugepaged_full_scans
;
90 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly
= 10000;
91 /* during fragmentation poll the hugepage allocator once every minute */
92 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly
= 60000;
93 static unsigned long khugepaged_sleep_expire
;
94 static struct task_struct
*khugepaged_thread __read_mostly
;
95 static DEFINE_MUTEX(khugepaged_mutex
);
96 static DEFINE_SPINLOCK(khugepaged_mm_lock
);
97 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait
);
99 * default collapse hugepages if there is at least one pte mapped like
100 * it would have happened if the vma was large enough during page
103 static unsigned int khugepaged_max_ptes_none __read_mostly
;
104 static unsigned int khugepaged_max_ptes_swap __read_mostly
;
106 static int khugepaged(void *none
);
107 static int khugepaged_slab_init(void);
108 static void khugepaged_slab_exit(void);
110 #define MM_SLOTS_HASH_BITS 10
111 static __read_mostly
DEFINE_HASHTABLE(mm_slots_hash
, MM_SLOTS_HASH_BITS
);
113 static struct kmem_cache
*mm_slot_cache __read_mostly
;
116 * struct mm_slot - hash lookup from mm to mm_slot
117 * @hash: hash collision list
118 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
119 * @mm: the mm that this information is valid for
122 struct hlist_node hash
;
123 struct list_head mm_node
;
124 struct mm_struct
*mm
;
128 * struct khugepaged_scan - cursor for scanning
129 * @mm_head: the head of the mm list to scan
130 * @mm_slot: the current mm_slot we are scanning
131 * @address: the next address inside that to be scanned
133 * There is only the one khugepaged_scan instance of this cursor structure.
135 struct khugepaged_scan
{
136 struct list_head mm_head
;
137 struct mm_slot
*mm_slot
;
138 unsigned long address
;
140 static struct khugepaged_scan khugepaged_scan
= {
141 .mm_head
= LIST_HEAD_INIT(khugepaged_scan
.mm_head
),
144 static struct shrinker deferred_split_shrinker
;
146 static void set_recommended_min_free_kbytes(void)
150 unsigned long recommended_min
;
152 for_each_populated_zone(zone
)
155 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
156 recommended_min
= pageblock_nr_pages
* nr_zones
* 2;
159 * Make sure that on average at least two pageblocks are almost free
160 * of another type, one for a migratetype to fall back to and a
161 * second to avoid subsequent fallbacks of other types There are 3
162 * MIGRATE_TYPES we care about.
164 recommended_min
+= pageblock_nr_pages
* nr_zones
*
165 MIGRATE_PCPTYPES
* MIGRATE_PCPTYPES
;
167 /* don't ever allow to reserve more than 5% of the lowmem */
168 recommended_min
= min(recommended_min
,
169 (unsigned long) nr_free_buffer_pages() / 20);
170 recommended_min
<<= (PAGE_SHIFT
-10);
172 if (recommended_min
> min_free_kbytes
) {
173 if (user_min_free_kbytes
>= 0)
174 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
175 min_free_kbytes
, recommended_min
);
177 min_free_kbytes
= recommended_min
;
179 setup_per_zone_wmarks();
182 static int start_stop_khugepaged(void)
185 if (khugepaged_enabled()) {
186 if (!khugepaged_thread
)
187 khugepaged_thread
= kthread_run(khugepaged
, NULL
,
189 if (IS_ERR(khugepaged_thread
)) {
190 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
191 err
= PTR_ERR(khugepaged_thread
);
192 khugepaged_thread
= NULL
;
196 if (!list_empty(&khugepaged_scan
.mm_head
))
197 wake_up_interruptible(&khugepaged_wait
);
199 set_recommended_min_free_kbytes();
200 } else if (khugepaged_thread
) {
201 kthread_stop(khugepaged_thread
);
202 khugepaged_thread
= NULL
;
208 static atomic_t huge_zero_refcount
;
209 struct page
*huge_zero_page __read_mostly
;
211 struct page
*get_huge_zero_page(void)
213 struct page
*zero_page
;
215 if (likely(atomic_inc_not_zero(&huge_zero_refcount
)))
216 return READ_ONCE(huge_zero_page
);
218 zero_page
= alloc_pages((GFP_TRANSHUGE
| __GFP_ZERO
) & ~__GFP_MOVABLE
,
221 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED
);
224 count_vm_event(THP_ZERO_PAGE_ALLOC
);
226 if (cmpxchg(&huge_zero_page
, NULL
, zero_page
)) {
228 __free_pages(zero_page
, compound_order(zero_page
));
232 /* We take additional reference here. It will be put back by shrinker */
233 atomic_set(&huge_zero_refcount
, 2);
235 return READ_ONCE(huge_zero_page
);
238 void put_huge_zero_page(void)
241 * Counter should never go to zero here. Only shrinker can put
244 BUG_ON(atomic_dec_and_test(&huge_zero_refcount
));
247 static unsigned long shrink_huge_zero_page_count(struct shrinker
*shrink
,
248 struct shrink_control
*sc
)
250 /* we can free zero page only if last reference remains */
251 return atomic_read(&huge_zero_refcount
) == 1 ? HPAGE_PMD_NR
: 0;
254 static unsigned long shrink_huge_zero_page_scan(struct shrinker
*shrink
,
255 struct shrink_control
*sc
)
257 if (atomic_cmpxchg(&huge_zero_refcount
, 1, 0) == 1) {
258 struct page
*zero_page
= xchg(&huge_zero_page
, NULL
);
259 BUG_ON(zero_page
== NULL
);
260 __free_pages(zero_page
, compound_order(zero_page
));
267 static struct shrinker huge_zero_page_shrinker
= {
268 .count_objects
= shrink_huge_zero_page_count
,
269 .scan_objects
= shrink_huge_zero_page_scan
,
270 .seeks
= DEFAULT_SEEKS
,
275 static ssize_t
triple_flag_store(struct kobject
*kobj
,
276 struct kobj_attribute
*attr
,
277 const char *buf
, size_t count
,
278 enum transparent_hugepage_flag enabled
,
279 enum transparent_hugepage_flag deferred
,
280 enum transparent_hugepage_flag req_madv
)
282 if (!memcmp("defer", buf
,
283 min(sizeof("defer")-1, count
))) {
284 if (enabled
== deferred
)
286 clear_bit(enabled
, &transparent_hugepage_flags
);
287 clear_bit(req_madv
, &transparent_hugepage_flags
);
288 set_bit(deferred
, &transparent_hugepage_flags
);
289 } else if (!memcmp("always", buf
,
290 min(sizeof("always")-1, count
))) {
291 clear_bit(deferred
, &transparent_hugepage_flags
);
292 clear_bit(req_madv
, &transparent_hugepage_flags
);
293 set_bit(enabled
, &transparent_hugepage_flags
);
294 } else if (!memcmp("madvise", buf
,
295 min(sizeof("madvise")-1, count
))) {
296 clear_bit(enabled
, &transparent_hugepage_flags
);
297 clear_bit(deferred
, &transparent_hugepage_flags
);
298 set_bit(req_madv
, &transparent_hugepage_flags
);
299 } else if (!memcmp("never", buf
,
300 min(sizeof("never")-1, count
))) {
301 clear_bit(enabled
, &transparent_hugepage_flags
);
302 clear_bit(req_madv
, &transparent_hugepage_flags
);
303 clear_bit(deferred
, &transparent_hugepage_flags
);
310 static ssize_t
enabled_show(struct kobject
*kobj
,
311 struct kobj_attribute
*attr
, char *buf
)
313 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
))
314 return sprintf(buf
, "[always] madvise never\n");
315 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
316 return sprintf(buf
, "always [madvise] never\n");
318 return sprintf(buf
, "always madvise [never]\n");
321 static ssize_t
enabled_store(struct kobject
*kobj
,
322 struct kobj_attribute
*attr
,
323 const char *buf
, size_t count
)
327 ret
= triple_flag_store(kobj
, attr
, buf
, count
,
328 TRANSPARENT_HUGEPAGE_FLAG
,
329 TRANSPARENT_HUGEPAGE_FLAG
,
330 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
335 mutex_lock(&khugepaged_mutex
);
336 err
= start_stop_khugepaged();
337 mutex_unlock(&khugepaged_mutex
);
345 static struct kobj_attribute enabled_attr
=
346 __ATTR(enabled
, 0644, enabled_show
, enabled_store
);
348 static ssize_t
single_flag_show(struct kobject
*kobj
,
349 struct kobj_attribute
*attr
, char *buf
,
350 enum transparent_hugepage_flag flag
)
352 return sprintf(buf
, "%d\n",
353 !!test_bit(flag
, &transparent_hugepage_flags
));
356 static ssize_t
single_flag_store(struct kobject
*kobj
,
357 struct kobj_attribute
*attr
,
358 const char *buf
, size_t count
,
359 enum transparent_hugepage_flag flag
)
364 ret
= kstrtoul(buf
, 10, &value
);
371 set_bit(flag
, &transparent_hugepage_flags
);
373 clear_bit(flag
, &transparent_hugepage_flags
);
379 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
380 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
381 * memory just to allocate one more hugepage.
383 static ssize_t
defrag_show(struct kobject
*kobj
,
384 struct kobj_attribute
*attr
, char *buf
)
386 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
))
387 return sprintf(buf
, "[always] defer madvise never\n");
388 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
))
389 return sprintf(buf
, "always [defer] madvise never\n");
390 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
391 return sprintf(buf
, "always defer [madvise] never\n");
393 return sprintf(buf
, "always defer madvise [never]\n");
396 static ssize_t
defrag_store(struct kobject
*kobj
,
397 struct kobj_attribute
*attr
,
398 const char *buf
, size_t count
)
400 return triple_flag_store(kobj
, attr
, buf
, count
,
401 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
,
402 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
,
403 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
);
405 static struct kobj_attribute defrag_attr
=
406 __ATTR(defrag
, 0644, defrag_show
, defrag_store
);
408 static ssize_t
use_zero_page_show(struct kobject
*kobj
,
409 struct kobj_attribute
*attr
, char *buf
)
411 return single_flag_show(kobj
, attr
, buf
,
412 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
414 static ssize_t
use_zero_page_store(struct kobject
*kobj
,
415 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
417 return single_flag_store(kobj
, attr
, buf
, count
,
418 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
420 static struct kobj_attribute use_zero_page_attr
=
421 __ATTR(use_zero_page
, 0644, use_zero_page_show
, use_zero_page_store
);
422 #ifdef CONFIG_DEBUG_VM
423 static ssize_t
debug_cow_show(struct kobject
*kobj
,
424 struct kobj_attribute
*attr
, char *buf
)
426 return single_flag_show(kobj
, attr
, buf
,
427 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
429 static ssize_t
debug_cow_store(struct kobject
*kobj
,
430 struct kobj_attribute
*attr
,
431 const char *buf
, size_t count
)
433 return single_flag_store(kobj
, attr
, buf
, count
,
434 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
436 static struct kobj_attribute debug_cow_attr
=
437 __ATTR(debug_cow
, 0644, debug_cow_show
, debug_cow_store
);
438 #endif /* CONFIG_DEBUG_VM */
440 static struct attribute
*hugepage_attr
[] = {
443 &use_zero_page_attr
.attr
,
444 #ifdef CONFIG_DEBUG_VM
445 &debug_cow_attr
.attr
,
450 static struct attribute_group hugepage_attr_group
= {
451 .attrs
= hugepage_attr
,
454 static ssize_t
scan_sleep_millisecs_show(struct kobject
*kobj
,
455 struct kobj_attribute
*attr
,
458 return sprintf(buf
, "%u\n", khugepaged_scan_sleep_millisecs
);
461 static ssize_t
scan_sleep_millisecs_store(struct kobject
*kobj
,
462 struct kobj_attribute
*attr
,
463 const char *buf
, size_t count
)
468 err
= kstrtoul(buf
, 10, &msecs
);
469 if (err
|| msecs
> UINT_MAX
)
472 khugepaged_scan_sleep_millisecs
= msecs
;
473 khugepaged_sleep_expire
= 0;
474 wake_up_interruptible(&khugepaged_wait
);
478 static struct kobj_attribute scan_sleep_millisecs_attr
=
479 __ATTR(scan_sleep_millisecs
, 0644, scan_sleep_millisecs_show
,
480 scan_sleep_millisecs_store
);
482 static ssize_t
alloc_sleep_millisecs_show(struct kobject
*kobj
,
483 struct kobj_attribute
*attr
,
486 return sprintf(buf
, "%u\n", khugepaged_alloc_sleep_millisecs
);
489 static ssize_t
alloc_sleep_millisecs_store(struct kobject
*kobj
,
490 struct kobj_attribute
*attr
,
491 const char *buf
, size_t count
)
496 err
= kstrtoul(buf
, 10, &msecs
);
497 if (err
|| msecs
> UINT_MAX
)
500 khugepaged_alloc_sleep_millisecs
= msecs
;
501 khugepaged_sleep_expire
= 0;
502 wake_up_interruptible(&khugepaged_wait
);
506 static struct kobj_attribute alloc_sleep_millisecs_attr
=
507 __ATTR(alloc_sleep_millisecs
, 0644, alloc_sleep_millisecs_show
,
508 alloc_sleep_millisecs_store
);
510 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
511 struct kobj_attribute
*attr
,
514 return sprintf(buf
, "%u\n", khugepaged_pages_to_scan
);
516 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
517 struct kobj_attribute
*attr
,
518 const char *buf
, size_t count
)
523 err
= kstrtoul(buf
, 10, &pages
);
524 if (err
|| !pages
|| pages
> UINT_MAX
)
527 khugepaged_pages_to_scan
= pages
;
531 static struct kobj_attribute pages_to_scan_attr
=
532 __ATTR(pages_to_scan
, 0644, pages_to_scan_show
,
533 pages_to_scan_store
);
535 static ssize_t
pages_collapsed_show(struct kobject
*kobj
,
536 struct kobj_attribute
*attr
,
539 return sprintf(buf
, "%u\n", khugepaged_pages_collapsed
);
541 static struct kobj_attribute pages_collapsed_attr
=
542 __ATTR_RO(pages_collapsed
);
544 static ssize_t
full_scans_show(struct kobject
*kobj
,
545 struct kobj_attribute
*attr
,
548 return sprintf(buf
, "%u\n", khugepaged_full_scans
);
550 static struct kobj_attribute full_scans_attr
=
551 __ATTR_RO(full_scans
);
553 static ssize_t
khugepaged_defrag_show(struct kobject
*kobj
,
554 struct kobj_attribute
*attr
, char *buf
)
556 return single_flag_show(kobj
, attr
, buf
,
557 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
559 static ssize_t
khugepaged_defrag_store(struct kobject
*kobj
,
560 struct kobj_attribute
*attr
,
561 const char *buf
, size_t count
)
563 return single_flag_store(kobj
, attr
, buf
, count
,
564 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
566 static struct kobj_attribute khugepaged_defrag_attr
=
567 __ATTR(defrag
, 0644, khugepaged_defrag_show
,
568 khugepaged_defrag_store
);
571 * max_ptes_none controls if khugepaged should collapse hugepages over
572 * any unmapped ptes in turn potentially increasing the memory
573 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
574 * reduce the available free memory in the system as it
575 * runs. Increasing max_ptes_none will instead potentially reduce the
576 * free memory in the system during the khugepaged scan.
578 static ssize_t
khugepaged_max_ptes_none_show(struct kobject
*kobj
,
579 struct kobj_attribute
*attr
,
582 return sprintf(buf
, "%u\n", khugepaged_max_ptes_none
);
584 static ssize_t
khugepaged_max_ptes_none_store(struct kobject
*kobj
,
585 struct kobj_attribute
*attr
,
586 const char *buf
, size_t count
)
589 unsigned long max_ptes_none
;
591 err
= kstrtoul(buf
, 10, &max_ptes_none
);
592 if (err
|| max_ptes_none
> HPAGE_PMD_NR
-1)
595 khugepaged_max_ptes_none
= max_ptes_none
;
599 static struct kobj_attribute khugepaged_max_ptes_none_attr
=
600 __ATTR(max_ptes_none
, 0644, khugepaged_max_ptes_none_show
,
601 khugepaged_max_ptes_none_store
);
603 static ssize_t
khugepaged_max_ptes_swap_show(struct kobject
*kobj
,
604 struct kobj_attribute
*attr
,
607 return sprintf(buf
, "%u\n", khugepaged_max_ptes_swap
);
610 static ssize_t
khugepaged_max_ptes_swap_store(struct kobject
*kobj
,
611 struct kobj_attribute
*attr
,
612 const char *buf
, size_t count
)
615 unsigned long max_ptes_swap
;
617 err
= kstrtoul(buf
, 10, &max_ptes_swap
);
618 if (err
|| max_ptes_swap
> HPAGE_PMD_NR
-1)
621 khugepaged_max_ptes_swap
= max_ptes_swap
;
626 static struct kobj_attribute khugepaged_max_ptes_swap_attr
=
627 __ATTR(max_ptes_swap
, 0644, khugepaged_max_ptes_swap_show
,
628 khugepaged_max_ptes_swap_store
);
630 static struct attribute
*khugepaged_attr
[] = {
631 &khugepaged_defrag_attr
.attr
,
632 &khugepaged_max_ptes_none_attr
.attr
,
633 &pages_to_scan_attr
.attr
,
634 &pages_collapsed_attr
.attr
,
635 &full_scans_attr
.attr
,
636 &scan_sleep_millisecs_attr
.attr
,
637 &alloc_sleep_millisecs_attr
.attr
,
638 &khugepaged_max_ptes_swap_attr
.attr
,
642 static struct attribute_group khugepaged_attr_group
= {
643 .attrs
= khugepaged_attr
,
644 .name
= "khugepaged",
647 static int __init
hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
651 *hugepage_kobj
= kobject_create_and_add("transparent_hugepage", mm_kobj
);
652 if (unlikely(!*hugepage_kobj
)) {
653 pr_err("failed to create transparent hugepage kobject\n");
657 err
= sysfs_create_group(*hugepage_kobj
, &hugepage_attr_group
);
659 pr_err("failed to register transparent hugepage group\n");
663 err
= sysfs_create_group(*hugepage_kobj
, &khugepaged_attr_group
);
665 pr_err("failed to register transparent hugepage group\n");
666 goto remove_hp_group
;
672 sysfs_remove_group(*hugepage_kobj
, &hugepage_attr_group
);
674 kobject_put(*hugepage_kobj
);
678 static void __init
hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
680 sysfs_remove_group(hugepage_kobj
, &khugepaged_attr_group
);
681 sysfs_remove_group(hugepage_kobj
, &hugepage_attr_group
);
682 kobject_put(hugepage_kobj
);
685 static inline int hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
690 static inline void hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
693 #endif /* CONFIG_SYSFS */
695 static int __init
hugepage_init(void)
698 struct kobject
*hugepage_kobj
;
700 if (!has_transparent_hugepage()) {
701 transparent_hugepage_flags
= 0;
705 khugepaged_pages_to_scan
= HPAGE_PMD_NR
* 8;
706 khugepaged_max_ptes_none
= HPAGE_PMD_NR
- 1;
707 khugepaged_max_ptes_swap
= HPAGE_PMD_NR
/ 8;
709 * hugepages can't be allocated by the buddy allocator
711 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER
>= MAX_ORDER
);
713 * we use page->mapping and page->index in second tail page
714 * as list_head: assuming THP order >= 2
716 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER
< 2);
718 err
= hugepage_init_sysfs(&hugepage_kobj
);
722 err
= khugepaged_slab_init();
726 err
= register_shrinker(&huge_zero_page_shrinker
);
728 goto err_hzp_shrinker
;
729 err
= register_shrinker(&deferred_split_shrinker
);
731 goto err_split_shrinker
;
734 * By default disable transparent hugepages on smaller systems,
735 * where the extra memory used could hurt more than TLB overhead
736 * is likely to save. The admin can still enable it through /sys.
738 if (totalram_pages
< (512 << (20 - PAGE_SHIFT
))) {
739 transparent_hugepage_flags
= 0;
743 err
= start_stop_khugepaged();
749 unregister_shrinker(&deferred_split_shrinker
);
751 unregister_shrinker(&huge_zero_page_shrinker
);
753 khugepaged_slab_exit();
755 hugepage_exit_sysfs(hugepage_kobj
);
759 subsys_initcall(hugepage_init
);
761 static int __init
setup_transparent_hugepage(char *str
)
766 if (!strcmp(str
, "always")) {
767 set_bit(TRANSPARENT_HUGEPAGE_FLAG
,
768 &transparent_hugepage_flags
);
769 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
770 &transparent_hugepage_flags
);
772 } else if (!strcmp(str
, "madvise")) {
773 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
774 &transparent_hugepage_flags
);
775 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
776 &transparent_hugepage_flags
);
778 } else if (!strcmp(str
, "never")) {
779 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
780 &transparent_hugepage_flags
);
781 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
782 &transparent_hugepage_flags
);
787 pr_warn("transparent_hugepage= cannot parse, ignored\n");
790 __setup("transparent_hugepage=", setup_transparent_hugepage
);
792 pmd_t
maybe_pmd_mkwrite(pmd_t pmd
, struct vm_area_struct
*vma
)
794 if (likely(vma
->vm_flags
& VM_WRITE
))
795 pmd
= pmd_mkwrite(pmd
);
799 static inline struct list_head
*page_deferred_list(struct page
*page
)
802 * ->lru in the tail pages is occupied by compound_head.
803 * Let's use ->mapping + ->index in the second tail page as list_head.
805 return (struct list_head
*)&page
[2].mapping
;
808 void prep_transhuge_page(struct page
*page
)
811 * we use page->mapping and page->indexlru in second tail page
812 * as list_head: assuming THP order >= 2
815 INIT_LIST_HEAD(page_deferred_list(page
));
816 set_compound_page_dtor(page
, TRANSHUGE_PAGE_DTOR
);
819 static int __do_huge_pmd_anonymous_page(struct fault_env
*fe
, struct page
*page
,
822 struct vm_area_struct
*vma
= fe
->vma
;
823 struct mem_cgroup
*memcg
;
825 unsigned long haddr
= fe
->address
& HPAGE_PMD_MASK
;
827 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
829 if (mem_cgroup_try_charge(page
, vma
->vm_mm
, gfp
, &memcg
, true)) {
831 count_vm_event(THP_FAULT_FALLBACK
);
832 return VM_FAULT_FALLBACK
;
835 pgtable
= pte_alloc_one(vma
->vm_mm
, haddr
);
836 if (unlikely(!pgtable
)) {
837 mem_cgroup_cancel_charge(page
, memcg
, true);
842 clear_huge_page(page
, haddr
, HPAGE_PMD_NR
);
844 * The memory barrier inside __SetPageUptodate makes sure that
845 * clear_huge_page writes become visible before the set_pmd_at()
848 __SetPageUptodate(page
);
850 fe
->ptl
= pmd_lock(vma
->vm_mm
, fe
->pmd
);
851 if (unlikely(!pmd_none(*fe
->pmd
))) {
852 spin_unlock(fe
->ptl
);
853 mem_cgroup_cancel_charge(page
, memcg
, true);
855 pte_free(vma
->vm_mm
, pgtable
);
859 /* Deliver the page fault to userland */
860 if (userfaultfd_missing(vma
)) {
863 spin_unlock(fe
->ptl
);
864 mem_cgroup_cancel_charge(page
, memcg
, true);
866 pte_free(vma
->vm_mm
, pgtable
);
867 ret
= handle_userfault(fe
, VM_UFFD_MISSING
);
868 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
872 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
873 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
874 page_add_new_anon_rmap(page
, vma
, haddr
, true);
875 mem_cgroup_commit_charge(page
, memcg
, false, true);
876 lru_cache_add_active_or_unevictable(page
, vma
);
877 pgtable_trans_huge_deposit(vma
->vm_mm
, fe
->pmd
, pgtable
);
878 set_pmd_at(vma
->vm_mm
, haddr
, fe
->pmd
, entry
);
879 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
880 atomic_long_inc(&vma
->vm_mm
->nr_ptes
);
881 spin_unlock(fe
->ptl
);
882 count_vm_event(THP_FAULT_ALLOC
);
889 * If THP is set to always then directly reclaim/compact as necessary
890 * If set to defer then do no reclaim and defer to khugepaged
891 * If set to madvise and the VMA is flagged then directly reclaim/compact
893 static inline gfp_t
alloc_hugepage_direct_gfpmask(struct vm_area_struct
*vma
)
895 gfp_t reclaim_flags
= 0;
897 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
) &&
898 (vma
->vm_flags
& VM_HUGEPAGE
))
899 reclaim_flags
= __GFP_DIRECT_RECLAIM
;
900 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
))
901 reclaim_flags
= __GFP_KSWAPD_RECLAIM
;
902 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
))
903 reclaim_flags
= __GFP_DIRECT_RECLAIM
;
905 return GFP_TRANSHUGE
| reclaim_flags
;
908 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
909 static inline gfp_t
alloc_hugepage_khugepaged_gfpmask(void)
911 return GFP_TRANSHUGE
| (khugepaged_defrag() ? __GFP_DIRECT_RECLAIM
: 0);
914 /* Caller must hold page table lock. */
915 static bool set_huge_zero_page(pgtable_t pgtable
, struct mm_struct
*mm
,
916 struct vm_area_struct
*vma
, unsigned long haddr
, pmd_t
*pmd
,
917 struct page
*zero_page
)
922 entry
= mk_pmd(zero_page
, vma
->vm_page_prot
);
923 entry
= pmd_mkhuge(entry
);
925 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
926 set_pmd_at(mm
, haddr
, pmd
, entry
);
927 atomic_long_inc(&mm
->nr_ptes
);
931 int do_huge_pmd_anonymous_page(struct fault_env
*fe
)
933 struct vm_area_struct
*vma
= fe
->vma
;
936 unsigned long haddr
= fe
->address
& HPAGE_PMD_MASK
;
938 if (haddr
< vma
->vm_start
|| haddr
+ HPAGE_PMD_SIZE
> vma
->vm_end
)
939 return VM_FAULT_FALLBACK
;
940 if (unlikely(anon_vma_prepare(vma
)))
942 if (unlikely(khugepaged_enter(vma
, vma
->vm_flags
)))
944 if (!(fe
->flags
& FAULT_FLAG_WRITE
) &&
945 !mm_forbids_zeropage(vma
->vm_mm
) &&
946 transparent_hugepage_use_zero_page()) {
948 struct page
*zero_page
;
951 pgtable
= pte_alloc_one(vma
->vm_mm
, haddr
);
952 if (unlikely(!pgtable
))
954 zero_page
= get_huge_zero_page();
955 if (unlikely(!zero_page
)) {
956 pte_free(vma
->vm_mm
, pgtable
);
957 count_vm_event(THP_FAULT_FALLBACK
);
958 return VM_FAULT_FALLBACK
;
960 fe
->ptl
= pmd_lock(vma
->vm_mm
, fe
->pmd
);
963 if (pmd_none(*fe
->pmd
)) {
964 if (userfaultfd_missing(vma
)) {
965 spin_unlock(fe
->ptl
);
966 ret
= handle_userfault(fe
, VM_UFFD_MISSING
);
967 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
969 set_huge_zero_page(pgtable
, vma
->vm_mm
, vma
,
970 haddr
, fe
->pmd
, zero_page
);
971 spin_unlock(fe
->ptl
);
975 spin_unlock(fe
->ptl
);
977 pte_free(vma
->vm_mm
, pgtable
);
978 put_huge_zero_page();
982 gfp
= alloc_hugepage_direct_gfpmask(vma
);
983 page
= alloc_hugepage_vma(gfp
, vma
, haddr
, HPAGE_PMD_ORDER
);
984 if (unlikely(!page
)) {
985 count_vm_event(THP_FAULT_FALLBACK
);
986 return VM_FAULT_FALLBACK
;
988 prep_transhuge_page(page
);
989 return __do_huge_pmd_anonymous_page(fe
, page
, gfp
);
992 static void insert_pfn_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
993 pmd_t
*pmd
, pfn_t pfn
, pgprot_t prot
, bool write
)
995 struct mm_struct
*mm
= vma
->vm_mm
;
999 ptl
= pmd_lock(mm
, pmd
);
1000 entry
= pmd_mkhuge(pfn_t_pmd(pfn
, prot
));
1001 if (pfn_t_devmap(pfn
))
1002 entry
= pmd_mkdevmap(entry
);
1004 entry
= pmd_mkyoung(pmd_mkdirty(entry
));
1005 entry
= maybe_pmd_mkwrite(entry
, vma
);
1007 set_pmd_at(mm
, addr
, pmd
, entry
);
1008 update_mmu_cache_pmd(vma
, addr
, pmd
);
1012 int vmf_insert_pfn_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
1013 pmd_t
*pmd
, pfn_t pfn
, bool write
)
1015 pgprot_t pgprot
= vma
->vm_page_prot
;
1017 * If we had pmd_special, we could avoid all these restrictions,
1018 * but we need to be consistent with PTEs and architectures that
1019 * can't support a 'special' bit.
1021 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
1022 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1023 (VM_PFNMAP
|VM_MIXEDMAP
));
1024 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1025 BUG_ON(!pfn_t_devmap(pfn
));
1027 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1028 return VM_FAULT_SIGBUS
;
1029 if (track_pfn_insert(vma
, &pgprot
, pfn
))
1030 return VM_FAULT_SIGBUS
;
1031 insert_pfn_pmd(vma
, addr
, pmd
, pfn
, pgprot
, write
);
1032 return VM_FAULT_NOPAGE
;
1034 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd
);
1036 static void touch_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
1042 * We should set the dirty bit only for FOLL_WRITE but for now
1043 * the dirty bit in the pmd is meaningless. And if the dirty
1044 * bit will become meaningful and we'll only set it with
1045 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
1046 * set the young bit, instead of the current set_pmd_at.
1048 _pmd
= pmd_mkyoung(pmd_mkdirty(*pmd
));
1049 if (pmdp_set_access_flags(vma
, addr
& HPAGE_PMD_MASK
,
1051 update_mmu_cache_pmd(vma
, addr
, pmd
);
1054 struct page
*follow_devmap_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
1055 pmd_t
*pmd
, int flags
)
1057 unsigned long pfn
= pmd_pfn(*pmd
);
1058 struct mm_struct
*mm
= vma
->vm_mm
;
1059 struct dev_pagemap
*pgmap
;
1062 assert_spin_locked(pmd_lockptr(mm
, pmd
));
1064 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
1067 if (pmd_present(*pmd
) && pmd_devmap(*pmd
))
1072 if (flags
& FOLL_TOUCH
)
1073 touch_pmd(vma
, addr
, pmd
);
1076 * device mapped pages can only be returned if the
1077 * caller will manage the page reference count.
1079 if (!(flags
& FOLL_GET
))
1080 return ERR_PTR(-EEXIST
);
1082 pfn
+= (addr
& ~PMD_MASK
) >> PAGE_SHIFT
;
1083 pgmap
= get_dev_pagemap(pfn
, NULL
);
1085 return ERR_PTR(-EFAULT
);
1086 page
= pfn_to_page(pfn
);
1088 put_dev_pagemap(pgmap
);
1093 int copy_huge_pmd(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1094 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
1095 struct vm_area_struct
*vma
)
1097 spinlock_t
*dst_ptl
, *src_ptl
;
1098 struct page
*src_page
;
1100 pgtable_t pgtable
= NULL
;
1103 if (!vma_is_dax(vma
)) {
1105 pgtable
= pte_alloc_one(dst_mm
, addr
);
1106 if (unlikely(!pgtable
))
1110 dst_ptl
= pmd_lock(dst_mm
, dst_pmd
);
1111 src_ptl
= pmd_lockptr(src_mm
, src_pmd
);
1112 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1116 if (unlikely(!pmd_trans_huge(pmd
) && !pmd_devmap(pmd
))) {
1117 pte_free(dst_mm
, pgtable
);
1121 * When page table lock is held, the huge zero pmd should not be
1122 * under splitting since we don't split the page itself, only pmd to
1125 if (is_huge_zero_pmd(pmd
)) {
1126 struct page
*zero_page
;
1128 * get_huge_zero_page() will never allocate a new page here,
1129 * since we already have a zero page to copy. It just takes a
1132 zero_page
= get_huge_zero_page();
1133 set_huge_zero_page(pgtable
, dst_mm
, vma
, addr
, dst_pmd
,
1139 if (!vma_is_dax(vma
)) {
1140 /* thp accounting separate from pmd_devmap accounting */
1141 src_page
= pmd_page(pmd
);
1142 VM_BUG_ON_PAGE(!PageHead(src_page
), src_page
);
1144 page_dup_rmap(src_page
, true);
1145 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1146 atomic_long_inc(&dst_mm
->nr_ptes
);
1147 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
1150 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
1151 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
1152 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
1156 spin_unlock(src_ptl
);
1157 spin_unlock(dst_ptl
);
1162 void huge_pmd_set_accessed(struct fault_env
*fe
, pmd_t orig_pmd
)
1165 unsigned long haddr
;
1167 fe
->ptl
= pmd_lock(fe
->vma
->vm_mm
, fe
->pmd
);
1168 if (unlikely(!pmd_same(*fe
->pmd
, orig_pmd
)))
1171 entry
= pmd_mkyoung(orig_pmd
);
1172 haddr
= fe
->address
& HPAGE_PMD_MASK
;
1173 if (pmdp_set_access_flags(fe
->vma
, haddr
, fe
->pmd
, entry
,
1174 fe
->flags
& FAULT_FLAG_WRITE
))
1175 update_mmu_cache_pmd(fe
->vma
, fe
->address
, fe
->pmd
);
1178 spin_unlock(fe
->ptl
);
1181 static int do_huge_pmd_wp_page_fallback(struct fault_env
*fe
, pmd_t orig_pmd
,
1184 struct vm_area_struct
*vma
= fe
->vma
;
1185 unsigned long haddr
= fe
->address
& HPAGE_PMD_MASK
;
1186 struct mem_cgroup
*memcg
;
1190 struct page
**pages
;
1191 unsigned long mmun_start
; /* For mmu_notifiers */
1192 unsigned long mmun_end
; /* For mmu_notifiers */
1194 pages
= kmalloc(sizeof(struct page
*) * HPAGE_PMD_NR
,
1196 if (unlikely(!pages
)) {
1197 ret
|= VM_FAULT_OOM
;
1201 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1202 pages
[i
] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE
|
1203 __GFP_OTHER_NODE
, vma
,
1204 fe
->address
, page_to_nid(page
));
1205 if (unlikely(!pages
[i
] ||
1206 mem_cgroup_try_charge(pages
[i
], vma
->vm_mm
,
1207 GFP_KERNEL
, &memcg
, false))) {
1211 memcg
= (void *)page_private(pages
[i
]);
1212 set_page_private(pages
[i
], 0);
1213 mem_cgroup_cancel_charge(pages
[i
], memcg
,
1218 ret
|= VM_FAULT_OOM
;
1221 set_page_private(pages
[i
], (unsigned long)memcg
);
1224 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1225 copy_user_highpage(pages
[i
], page
+ i
,
1226 haddr
+ PAGE_SIZE
* i
, vma
);
1227 __SetPageUptodate(pages
[i
]);
1232 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1233 mmu_notifier_invalidate_range_start(vma
->vm_mm
, mmun_start
, mmun_end
);
1235 fe
->ptl
= pmd_lock(vma
->vm_mm
, fe
->pmd
);
1236 if (unlikely(!pmd_same(*fe
->pmd
, orig_pmd
)))
1237 goto out_free_pages
;
1238 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1240 pmdp_huge_clear_flush_notify(vma
, haddr
, fe
->pmd
);
1241 /* leave pmd empty until pte is filled */
1243 pgtable
= pgtable_trans_huge_withdraw(vma
->vm_mm
, fe
->pmd
);
1244 pmd_populate(vma
->vm_mm
, &_pmd
, pgtable
);
1246 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1248 entry
= mk_pte(pages
[i
], vma
->vm_page_prot
);
1249 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1250 memcg
= (void *)page_private(pages
[i
]);
1251 set_page_private(pages
[i
], 0);
1252 page_add_new_anon_rmap(pages
[i
], fe
->vma
, haddr
, false);
1253 mem_cgroup_commit_charge(pages
[i
], memcg
, false, false);
1254 lru_cache_add_active_or_unevictable(pages
[i
], vma
);
1255 fe
->pte
= pte_offset_map(&_pmd
, haddr
);
1256 VM_BUG_ON(!pte_none(*fe
->pte
));
1257 set_pte_at(vma
->vm_mm
, haddr
, fe
->pte
, entry
);
1262 smp_wmb(); /* make pte visible before pmd */
1263 pmd_populate(vma
->vm_mm
, fe
->pmd
, pgtable
);
1264 page_remove_rmap(page
, true);
1265 spin_unlock(fe
->ptl
);
1267 mmu_notifier_invalidate_range_end(vma
->vm_mm
, mmun_start
, mmun_end
);
1269 ret
|= VM_FAULT_WRITE
;
1276 spin_unlock(fe
->ptl
);
1277 mmu_notifier_invalidate_range_end(vma
->vm_mm
, mmun_start
, mmun_end
);
1278 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1279 memcg
= (void *)page_private(pages
[i
]);
1280 set_page_private(pages
[i
], 0);
1281 mem_cgroup_cancel_charge(pages
[i
], memcg
, false);
1288 int do_huge_pmd_wp_page(struct fault_env
*fe
, pmd_t orig_pmd
)
1290 struct vm_area_struct
*vma
= fe
->vma
;
1291 struct page
*page
= NULL
, *new_page
;
1292 struct mem_cgroup
*memcg
;
1293 unsigned long haddr
= fe
->address
& HPAGE_PMD_MASK
;
1294 unsigned long mmun_start
; /* For mmu_notifiers */
1295 unsigned long mmun_end
; /* For mmu_notifiers */
1296 gfp_t huge_gfp
; /* for allocation and charge */
1299 fe
->ptl
= pmd_lockptr(vma
->vm_mm
, fe
->pmd
);
1300 VM_BUG_ON_VMA(!vma
->anon_vma
, vma
);
1301 if (is_huge_zero_pmd(orig_pmd
))
1304 if (unlikely(!pmd_same(*fe
->pmd
, orig_pmd
)))
1307 page
= pmd_page(orig_pmd
);
1308 VM_BUG_ON_PAGE(!PageCompound(page
) || !PageHead(page
), page
);
1310 * We can only reuse the page if nobody else maps the huge page or it's
1313 if (page_trans_huge_mapcount(page
, NULL
) == 1) {
1315 entry
= pmd_mkyoung(orig_pmd
);
1316 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1317 if (pmdp_set_access_flags(vma
, haddr
, fe
->pmd
, entry
, 1))
1318 update_mmu_cache_pmd(vma
, fe
->address
, fe
->pmd
);
1319 ret
|= VM_FAULT_WRITE
;
1323 spin_unlock(fe
->ptl
);
1325 if (transparent_hugepage_enabled(vma
) &&
1326 !transparent_hugepage_debug_cow()) {
1327 huge_gfp
= alloc_hugepage_direct_gfpmask(vma
);
1328 new_page
= alloc_hugepage_vma(huge_gfp
, vma
, haddr
, HPAGE_PMD_ORDER
);
1332 if (likely(new_page
)) {
1333 prep_transhuge_page(new_page
);
1336 split_huge_pmd(vma
, fe
->pmd
, fe
->address
);
1337 ret
|= VM_FAULT_FALLBACK
;
1339 ret
= do_huge_pmd_wp_page_fallback(fe
, orig_pmd
, page
);
1340 if (ret
& VM_FAULT_OOM
) {
1341 split_huge_pmd(vma
, fe
->pmd
, fe
->address
);
1342 ret
|= VM_FAULT_FALLBACK
;
1346 count_vm_event(THP_FAULT_FALLBACK
);
1350 if (unlikely(mem_cgroup_try_charge(new_page
, vma
->vm_mm
,
1351 huge_gfp
, &memcg
, true))) {
1353 split_huge_pmd(vma
, fe
->pmd
, fe
->address
);
1356 ret
|= VM_FAULT_FALLBACK
;
1357 count_vm_event(THP_FAULT_FALLBACK
);
1361 count_vm_event(THP_FAULT_ALLOC
);
1364 clear_huge_page(new_page
, haddr
, HPAGE_PMD_NR
);
1366 copy_user_huge_page(new_page
, page
, haddr
, vma
, HPAGE_PMD_NR
);
1367 __SetPageUptodate(new_page
);
1370 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1371 mmu_notifier_invalidate_range_start(vma
->vm_mm
, mmun_start
, mmun_end
);
1376 if (unlikely(!pmd_same(*fe
->pmd
, orig_pmd
))) {
1377 spin_unlock(fe
->ptl
);
1378 mem_cgroup_cancel_charge(new_page
, memcg
, true);
1383 entry
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
1384 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1385 pmdp_huge_clear_flush_notify(vma
, haddr
, fe
->pmd
);
1386 page_add_new_anon_rmap(new_page
, vma
, haddr
, true);
1387 mem_cgroup_commit_charge(new_page
, memcg
, false, true);
1388 lru_cache_add_active_or_unevictable(new_page
, vma
);
1389 set_pmd_at(vma
->vm_mm
, haddr
, fe
->pmd
, entry
);
1390 update_mmu_cache_pmd(vma
, fe
->address
, fe
->pmd
);
1392 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1393 put_huge_zero_page();
1395 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1396 page_remove_rmap(page
, true);
1399 ret
|= VM_FAULT_WRITE
;
1401 spin_unlock(fe
->ptl
);
1403 mmu_notifier_invalidate_range_end(vma
->vm_mm
, mmun_start
, mmun_end
);
1407 spin_unlock(fe
->ptl
);
1411 struct page
*follow_trans_huge_pmd(struct vm_area_struct
*vma
,
1416 struct mm_struct
*mm
= vma
->vm_mm
;
1417 struct page
*page
= NULL
;
1419 assert_spin_locked(pmd_lockptr(mm
, pmd
));
1421 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
1424 /* Avoid dumping huge zero page */
1425 if ((flags
& FOLL_DUMP
) && is_huge_zero_pmd(*pmd
))
1426 return ERR_PTR(-EFAULT
);
1428 /* Full NUMA hinting faults to serialise migration in fault paths */
1429 if ((flags
& FOLL_NUMA
) && pmd_protnone(*pmd
))
1432 page
= pmd_page(*pmd
);
1433 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1434 if (flags
& FOLL_TOUCH
)
1435 touch_pmd(vma
, addr
, pmd
);
1436 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1438 * We don't mlock() pte-mapped THPs. This way we can avoid
1439 * leaking mlocked pages into non-VM_LOCKED VMAs.
1441 * In most cases the pmd is the only mapping of the page as we
1442 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1443 * writable private mappings in populate_vma_page_range().
1445 * The only scenario when we have the page shared here is if we
1446 * mlocking read-only mapping shared over fork(). We skip
1447 * mlocking such pages.
1449 if (compound_mapcount(page
) == 1 && !PageDoubleMap(page
) &&
1450 page
->mapping
&& trylock_page(page
)) {
1453 mlock_vma_page(page
);
1457 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
1458 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
1459 if (flags
& FOLL_GET
)
1466 /* NUMA hinting page fault entry point for trans huge pmds */
1467 int do_huge_pmd_numa_page(struct fault_env
*fe
, pmd_t pmd
)
1469 struct vm_area_struct
*vma
= fe
->vma
;
1470 struct anon_vma
*anon_vma
= NULL
;
1472 unsigned long haddr
= fe
->address
& HPAGE_PMD_MASK
;
1473 int page_nid
= -1, this_nid
= numa_node_id();
1474 int target_nid
, last_cpupid
= -1;
1476 bool migrated
= false;
1480 /* A PROT_NONE fault should not end up here */
1481 BUG_ON(!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)));
1483 fe
->ptl
= pmd_lock(vma
->vm_mm
, fe
->pmd
);
1484 if (unlikely(!pmd_same(pmd
, *fe
->pmd
)))
1488 * If there are potential migrations, wait for completion and retry
1489 * without disrupting NUMA hinting information. Do not relock and
1490 * check_same as the page may no longer be mapped.
1492 if (unlikely(pmd_trans_migrating(*fe
->pmd
))) {
1493 page
= pmd_page(*fe
->pmd
);
1494 spin_unlock(fe
->ptl
);
1495 wait_on_page_locked(page
);
1499 page
= pmd_page(pmd
);
1500 BUG_ON(is_huge_zero_page(page
));
1501 page_nid
= page_to_nid(page
);
1502 last_cpupid
= page_cpupid_last(page
);
1503 count_vm_numa_event(NUMA_HINT_FAULTS
);
1504 if (page_nid
== this_nid
) {
1505 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
1506 flags
|= TNF_FAULT_LOCAL
;
1509 /* See similar comment in do_numa_page for explanation */
1510 if (!(vma
->vm_flags
& VM_WRITE
))
1511 flags
|= TNF_NO_GROUP
;
1514 * Acquire the page lock to serialise THP migrations but avoid dropping
1515 * page_table_lock if at all possible
1517 page_locked
= trylock_page(page
);
1518 target_nid
= mpol_misplaced(page
, vma
, haddr
);
1519 if (target_nid
== -1) {
1520 /* If the page was locked, there are no parallel migrations */
1525 /* Migration could have started since the pmd_trans_migrating check */
1527 spin_unlock(fe
->ptl
);
1528 wait_on_page_locked(page
);
1534 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1535 * to serialises splits
1538 spin_unlock(fe
->ptl
);
1539 anon_vma
= page_lock_anon_vma_read(page
);
1541 /* Confirm the PMD did not change while page_table_lock was released */
1543 if (unlikely(!pmd_same(pmd
, *fe
->pmd
))) {
1550 /* Bail if we fail to protect against THP splits for any reason */
1551 if (unlikely(!anon_vma
)) {
1558 * Migrate the THP to the requested node, returns with page unlocked
1559 * and access rights restored.
1561 spin_unlock(fe
->ptl
);
1562 migrated
= migrate_misplaced_transhuge_page(vma
->vm_mm
, vma
,
1563 fe
->pmd
, pmd
, fe
->address
, page
, target_nid
);
1565 flags
|= TNF_MIGRATED
;
1566 page_nid
= target_nid
;
1568 flags
|= TNF_MIGRATE_FAIL
;
1572 BUG_ON(!PageLocked(page
));
1573 was_writable
= pmd_write(pmd
);
1574 pmd
= pmd_modify(pmd
, vma
->vm_page_prot
);
1575 pmd
= pmd_mkyoung(pmd
);
1577 pmd
= pmd_mkwrite(pmd
);
1578 set_pmd_at(vma
->vm_mm
, haddr
, fe
->pmd
, pmd
);
1579 update_mmu_cache_pmd(vma
, fe
->address
, fe
->pmd
);
1582 spin_unlock(fe
->ptl
);
1586 page_unlock_anon_vma_read(anon_vma
);
1589 task_numa_fault(last_cpupid
, page_nid
, HPAGE_PMD_NR
, fe
->flags
);
1594 int madvise_free_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1595 pmd_t
*pmd
, unsigned long addr
, unsigned long next
)
1601 struct mm_struct
*mm
= tlb
->mm
;
1604 ptl
= pmd_trans_huge_lock(pmd
, vma
);
1609 if (is_huge_zero_pmd(orig_pmd
)) {
1614 page
= pmd_page(orig_pmd
);
1616 * If other processes are mapping this page, we couldn't discard
1617 * the page unless they all do MADV_FREE so let's skip the page.
1619 if (page_mapcount(page
) != 1)
1622 if (!trylock_page(page
))
1626 * If user want to discard part-pages of THP, split it so MADV_FREE
1627 * will deactivate only them.
1629 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1632 split_huge_page(page
);
1638 if (PageDirty(page
))
1639 ClearPageDirty(page
);
1642 if (PageActive(page
))
1643 deactivate_page(page
);
1645 if (pmd_young(orig_pmd
) || pmd_dirty(orig_pmd
)) {
1646 orig_pmd
= pmdp_huge_get_and_clear_full(tlb
->mm
, addr
, pmd
,
1648 orig_pmd
= pmd_mkold(orig_pmd
);
1649 orig_pmd
= pmd_mkclean(orig_pmd
);
1651 set_pmd_at(mm
, addr
, pmd
, orig_pmd
);
1652 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1661 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1662 pmd_t
*pmd
, unsigned long addr
)
1667 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1671 * For architectures like ppc64 we look at deposited pgtable
1672 * when calling pmdp_huge_get_and_clear. So do the
1673 * pgtable_trans_huge_withdraw after finishing pmdp related
1676 orig_pmd
= pmdp_huge_get_and_clear_full(tlb
->mm
, addr
, pmd
,
1678 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1679 if (vma_is_dax(vma
)) {
1681 if (is_huge_zero_pmd(orig_pmd
))
1682 tlb_remove_page(tlb
, pmd_page(orig_pmd
));
1683 } else if (is_huge_zero_pmd(orig_pmd
)) {
1684 pte_free(tlb
->mm
, pgtable_trans_huge_withdraw(tlb
->mm
, pmd
));
1685 atomic_long_dec(&tlb
->mm
->nr_ptes
);
1687 tlb_remove_page(tlb
, pmd_page(orig_pmd
));
1689 struct page
*page
= pmd_page(orig_pmd
);
1690 page_remove_rmap(page
, true);
1691 VM_BUG_ON_PAGE(page_mapcount(page
) < 0, page
);
1692 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1693 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1694 pte_free(tlb
->mm
, pgtable_trans_huge_withdraw(tlb
->mm
, pmd
));
1695 atomic_long_dec(&tlb
->mm
->nr_ptes
);
1697 tlb_remove_page_size(tlb
, page
, HPAGE_PMD_SIZE
);
1702 bool move_huge_pmd(struct vm_area_struct
*vma
, unsigned long old_addr
,
1703 unsigned long new_addr
, unsigned long old_end
,
1704 pmd_t
*old_pmd
, pmd_t
*new_pmd
)
1706 spinlock_t
*old_ptl
, *new_ptl
;
1708 struct mm_struct
*mm
= vma
->vm_mm
;
1710 if ((old_addr
& ~HPAGE_PMD_MASK
) ||
1711 (new_addr
& ~HPAGE_PMD_MASK
) ||
1712 old_end
- old_addr
< HPAGE_PMD_SIZE
)
1716 * The destination pmd shouldn't be established, free_pgtables()
1717 * should have release it.
1719 if (WARN_ON(!pmd_none(*new_pmd
))) {
1720 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1725 * We don't have to worry about the ordering of src and dst
1726 * ptlocks because exclusive mmap_sem prevents deadlock.
1728 old_ptl
= __pmd_trans_huge_lock(old_pmd
, vma
);
1730 new_ptl
= pmd_lockptr(mm
, new_pmd
);
1731 if (new_ptl
!= old_ptl
)
1732 spin_lock_nested(new_ptl
, SINGLE_DEPTH_NESTING
);
1733 pmd
= pmdp_huge_get_and_clear(mm
, old_addr
, old_pmd
);
1734 VM_BUG_ON(!pmd_none(*new_pmd
));
1736 if (pmd_move_must_withdraw(new_ptl
, old_ptl
) &&
1737 vma_is_anonymous(vma
)) {
1739 pgtable
= pgtable_trans_huge_withdraw(mm
, old_pmd
);
1740 pgtable_trans_huge_deposit(mm
, new_pmd
, pgtable
);
1742 set_pmd_at(mm
, new_addr
, new_pmd
, pmd_mksoft_dirty(pmd
));
1743 if (new_ptl
!= old_ptl
)
1744 spin_unlock(new_ptl
);
1745 spin_unlock(old_ptl
);
1753 * - 0 if PMD could not be locked
1754 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1755 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1757 int change_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1758 unsigned long addr
, pgprot_t newprot
, int prot_numa
)
1760 struct mm_struct
*mm
= vma
->vm_mm
;
1764 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1767 bool preserve_write
= prot_numa
&& pmd_write(*pmd
);
1771 * Avoid trapping faults against the zero page. The read-only
1772 * data is likely to be read-cached on the local CPU and
1773 * local/remote hits to the zero page are not interesting.
1775 if (prot_numa
&& is_huge_zero_pmd(*pmd
)) {
1780 if (!prot_numa
|| !pmd_protnone(*pmd
)) {
1781 entry
= pmdp_huge_get_and_clear_notify(mm
, addr
, pmd
);
1782 entry
= pmd_modify(entry
, newprot
);
1784 entry
= pmd_mkwrite(entry
);
1786 set_pmd_at(mm
, addr
, pmd
, entry
);
1787 BUG_ON(!preserve_write
&& pmd_write(entry
));
1796 * Returns true if a given pmd maps a thp, false otherwise.
1798 * Note that if it returns true, this routine returns without unlocking page
1799 * table lock. So callers must unlock it.
1801 spinlock_t
*__pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
)
1804 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
1805 if (likely(pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)))
1811 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1813 int hugepage_madvise(struct vm_area_struct
*vma
,
1814 unsigned long *vm_flags
, int advice
)
1820 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1821 * can't handle this properly after s390_enable_sie, so we simply
1822 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1824 if (mm_has_pgste(vma
->vm_mm
))
1828 * Be somewhat over-protective like KSM for now!
1830 if (*vm_flags
& VM_NO_THP
)
1832 *vm_flags
&= ~VM_NOHUGEPAGE
;
1833 *vm_flags
|= VM_HUGEPAGE
;
1835 * If the vma become good for khugepaged to scan,
1836 * register it here without waiting a page fault that
1837 * may not happen any time soon.
1839 if (unlikely(khugepaged_enter_vma_merge(vma
, *vm_flags
)))
1842 case MADV_NOHUGEPAGE
:
1844 * Be somewhat over-protective like KSM for now!
1846 if (*vm_flags
& VM_NO_THP
)
1848 *vm_flags
&= ~VM_HUGEPAGE
;
1849 *vm_flags
|= VM_NOHUGEPAGE
;
1851 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1852 * this vma even if we leave the mm registered in khugepaged if
1853 * it got registered before VM_NOHUGEPAGE was set.
1861 static int __init
khugepaged_slab_init(void)
1863 mm_slot_cache
= kmem_cache_create("khugepaged_mm_slot",
1864 sizeof(struct mm_slot
),
1865 __alignof__(struct mm_slot
), 0, NULL
);
1872 static void __init
khugepaged_slab_exit(void)
1874 kmem_cache_destroy(mm_slot_cache
);
1877 static inline struct mm_slot
*alloc_mm_slot(void)
1879 if (!mm_slot_cache
) /* initialization failed */
1881 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
1884 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
1886 kmem_cache_free(mm_slot_cache
, mm_slot
);
1889 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
1891 struct mm_slot
*mm_slot
;
1893 hash_for_each_possible(mm_slots_hash
, mm_slot
, hash
, (unsigned long)mm
)
1894 if (mm
== mm_slot
->mm
)
1900 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
1901 struct mm_slot
*mm_slot
)
1904 hash_add(mm_slots_hash
, &mm_slot
->hash
, (long)mm
);
1907 static inline int khugepaged_test_exit(struct mm_struct
*mm
)
1909 return atomic_read(&mm
->mm_users
) == 0;
1912 int __khugepaged_enter(struct mm_struct
*mm
)
1914 struct mm_slot
*mm_slot
;
1917 mm_slot
= alloc_mm_slot();
1921 /* __khugepaged_exit() must not run from under us */
1922 VM_BUG_ON_MM(khugepaged_test_exit(mm
), mm
);
1923 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE
, &mm
->flags
))) {
1924 free_mm_slot(mm_slot
);
1928 spin_lock(&khugepaged_mm_lock
);
1929 insert_to_mm_slots_hash(mm
, mm_slot
);
1931 * Insert just behind the scanning cursor, to let the area settle
1934 wakeup
= list_empty(&khugepaged_scan
.mm_head
);
1935 list_add_tail(&mm_slot
->mm_node
, &khugepaged_scan
.mm_head
);
1936 spin_unlock(&khugepaged_mm_lock
);
1938 atomic_inc(&mm
->mm_count
);
1940 wake_up_interruptible(&khugepaged_wait
);
1945 int khugepaged_enter_vma_merge(struct vm_area_struct
*vma
,
1946 unsigned long vm_flags
)
1948 unsigned long hstart
, hend
;
1951 * Not yet faulted in so we will register later in the
1952 * page fault if needed.
1955 if (vma
->vm_ops
|| (vm_flags
& VM_NO_THP
))
1956 /* khugepaged not yet working on file or special mappings */
1958 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
1959 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
1961 return khugepaged_enter(vma
, vm_flags
);
1965 void __khugepaged_exit(struct mm_struct
*mm
)
1967 struct mm_slot
*mm_slot
;
1970 spin_lock(&khugepaged_mm_lock
);
1971 mm_slot
= get_mm_slot(mm
);
1972 if (mm_slot
&& khugepaged_scan
.mm_slot
!= mm_slot
) {
1973 hash_del(&mm_slot
->hash
);
1974 list_del(&mm_slot
->mm_node
);
1977 spin_unlock(&khugepaged_mm_lock
);
1980 clear_bit(MMF_VM_HUGEPAGE
, &mm
->flags
);
1981 free_mm_slot(mm_slot
);
1983 } else if (mm_slot
) {
1985 * This is required to serialize against
1986 * khugepaged_test_exit() (which is guaranteed to run
1987 * under mmap sem read mode). Stop here (after we
1988 * return all pagetables will be destroyed) until
1989 * khugepaged has finished working on the pagetables
1990 * under the mmap_sem.
1992 down_write(&mm
->mmap_sem
);
1993 up_write(&mm
->mmap_sem
);
1997 static void release_pte_page(struct page
*page
)
1999 /* 0 stands for page_is_file_cache(page) == false */
2000 dec_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
2002 putback_lru_page(page
);
2005 static void release_pte_pages(pte_t
*pte
, pte_t
*_pte
)
2007 while (--_pte
>= pte
) {
2008 pte_t pteval
= *_pte
;
2009 if (!pte_none(pteval
) && !is_zero_pfn(pte_pfn(pteval
)))
2010 release_pte_page(pte_page(pteval
));
2014 static int __collapse_huge_page_isolate(struct vm_area_struct
*vma
,
2015 unsigned long address
,
2018 struct page
*page
= NULL
;
2020 int none_or_zero
= 0, result
= 0;
2021 bool referenced
= false, writable
= false;
2023 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
2024 _pte
++, address
+= PAGE_SIZE
) {
2025 pte_t pteval
= *_pte
;
2026 if (pte_none(pteval
) || (pte_present(pteval
) &&
2027 is_zero_pfn(pte_pfn(pteval
)))) {
2028 if (!userfaultfd_armed(vma
) &&
2029 ++none_or_zero
<= khugepaged_max_ptes_none
) {
2032 result
= SCAN_EXCEED_NONE_PTE
;
2036 if (!pte_present(pteval
)) {
2037 result
= SCAN_PTE_NON_PRESENT
;
2040 page
= vm_normal_page(vma
, address
, pteval
);
2041 if (unlikely(!page
)) {
2042 result
= SCAN_PAGE_NULL
;
2046 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2047 VM_BUG_ON_PAGE(!PageAnon(page
), page
);
2048 VM_BUG_ON_PAGE(!PageSwapBacked(page
), page
);
2051 * We can do it before isolate_lru_page because the
2052 * page can't be freed from under us. NOTE: PG_lock
2053 * is needed to serialize against split_huge_page
2054 * when invoked from the VM.
2056 if (!trylock_page(page
)) {
2057 result
= SCAN_PAGE_LOCK
;
2062 * cannot use mapcount: can't collapse if there's a gup pin.
2063 * The page must only be referenced by the scanned process
2064 * and page swap cache.
2066 if (page_count(page
) != 1 + !!PageSwapCache(page
)) {
2068 result
= SCAN_PAGE_COUNT
;
2071 if (pte_write(pteval
)) {
2074 if (PageSwapCache(page
) &&
2075 !reuse_swap_page(page
, NULL
)) {
2077 result
= SCAN_SWAP_CACHE_PAGE
;
2081 * Page is not in the swap cache. It can be collapsed
2087 * Isolate the page to avoid collapsing an hugepage
2088 * currently in use by the VM.
2090 if (isolate_lru_page(page
)) {
2092 result
= SCAN_DEL_PAGE_LRU
;
2095 /* 0 stands for page_is_file_cache(page) == false */
2096 inc_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
2097 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2098 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2100 /* If there is no mapped pte young don't collapse the page */
2101 if (pte_young(pteval
) ||
2102 page_is_young(page
) || PageReferenced(page
) ||
2103 mmu_notifier_test_young(vma
->vm_mm
, address
))
2106 if (likely(writable
)) {
2107 if (likely(referenced
)) {
2108 result
= SCAN_SUCCEED
;
2109 trace_mm_collapse_huge_page_isolate(page
, none_or_zero
,
2110 referenced
, writable
, result
);
2114 result
= SCAN_PAGE_RO
;
2118 release_pte_pages(pte
, _pte
);
2119 trace_mm_collapse_huge_page_isolate(page
, none_or_zero
,
2120 referenced
, writable
, result
);
2124 static void __collapse_huge_page_copy(pte_t
*pte
, struct page
*page
,
2125 struct vm_area_struct
*vma
,
2126 unsigned long address
,
2130 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
; _pte
++) {
2131 pte_t pteval
= *_pte
;
2132 struct page
*src_page
;
2134 if (pte_none(pteval
) || is_zero_pfn(pte_pfn(pteval
))) {
2135 clear_user_highpage(page
, address
);
2136 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, 1);
2137 if (is_zero_pfn(pte_pfn(pteval
))) {
2139 * ptl mostly unnecessary.
2143 * paravirt calls inside pte_clear here are
2146 pte_clear(vma
->vm_mm
, address
, _pte
);
2150 src_page
= pte_page(pteval
);
2151 copy_user_highpage(page
, src_page
, address
, vma
);
2152 VM_BUG_ON_PAGE(page_mapcount(src_page
) != 1, src_page
);
2153 release_pte_page(src_page
);
2155 * ptl mostly unnecessary, but preempt has to
2156 * be disabled to update the per-cpu stats
2157 * inside page_remove_rmap().
2161 * paravirt calls inside pte_clear here are
2164 pte_clear(vma
->vm_mm
, address
, _pte
);
2165 page_remove_rmap(src_page
, false);
2167 free_page_and_swap_cache(src_page
);
2170 address
+= PAGE_SIZE
;
2175 static void khugepaged_alloc_sleep(void)
2179 add_wait_queue(&khugepaged_wait
, &wait
);
2180 freezable_schedule_timeout_interruptible(
2181 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs
));
2182 remove_wait_queue(&khugepaged_wait
, &wait
);
2185 static int khugepaged_node_load
[MAX_NUMNODES
];
2187 static bool khugepaged_scan_abort(int nid
)
2192 * If zone_reclaim_mode is disabled, then no extra effort is made to
2193 * allocate memory locally.
2195 if (!zone_reclaim_mode
)
2198 /* If there is a count for this node already, it must be acceptable */
2199 if (khugepaged_node_load
[nid
])
2202 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
2203 if (!khugepaged_node_load
[i
])
2205 if (node_distance(nid
, i
) > RECLAIM_DISTANCE
)
2212 static int khugepaged_find_target_node(void)
2214 static int last_khugepaged_target_node
= NUMA_NO_NODE
;
2215 int nid
, target_node
= 0, max_value
= 0;
2217 /* find first node with max normal pages hit */
2218 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
2219 if (khugepaged_node_load
[nid
] > max_value
) {
2220 max_value
= khugepaged_node_load
[nid
];
2224 /* do some balance if several nodes have the same hit record */
2225 if (target_node
<= last_khugepaged_target_node
)
2226 for (nid
= last_khugepaged_target_node
+ 1; nid
< MAX_NUMNODES
;
2228 if (max_value
== khugepaged_node_load
[nid
]) {
2233 last_khugepaged_target_node
= target_node
;
2237 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
2239 if (IS_ERR(*hpage
)) {
2245 khugepaged_alloc_sleep();
2246 } else if (*hpage
) {
2254 static struct page
*
2255 khugepaged_alloc_page(struct page
**hpage
, gfp_t gfp
, struct mm_struct
*mm
,
2256 unsigned long address
, int node
)
2258 VM_BUG_ON_PAGE(*hpage
, *hpage
);
2261 * Before allocating the hugepage, release the mmap_sem read lock.
2262 * The allocation can take potentially a long time if it involves
2263 * sync compaction, and we do not need to hold the mmap_sem during
2264 * that. We will recheck the vma after taking it again in write mode.
2266 up_read(&mm
->mmap_sem
);
2268 *hpage
= __alloc_pages_node(node
, gfp
, HPAGE_PMD_ORDER
);
2269 if (unlikely(!*hpage
)) {
2270 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
2271 *hpage
= ERR_PTR(-ENOMEM
);
2275 prep_transhuge_page(*hpage
);
2276 count_vm_event(THP_COLLAPSE_ALLOC
);
2280 static int khugepaged_find_target_node(void)
2285 static inline struct page
*alloc_khugepaged_hugepage(void)
2289 page
= alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
2292 prep_transhuge_page(page
);
2296 static struct page
*khugepaged_alloc_hugepage(bool *wait
)
2301 hpage
= alloc_khugepaged_hugepage();
2303 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
2308 khugepaged_alloc_sleep();
2310 count_vm_event(THP_COLLAPSE_ALLOC
);
2311 } while (unlikely(!hpage
) && likely(khugepaged_enabled()));
2316 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
2319 *hpage
= khugepaged_alloc_hugepage(wait
);
2321 if (unlikely(!*hpage
))
2327 static struct page
*
2328 khugepaged_alloc_page(struct page
**hpage
, gfp_t gfp
, struct mm_struct
*mm
,
2329 unsigned long address
, int node
)
2331 up_read(&mm
->mmap_sem
);
2338 static bool hugepage_vma_check(struct vm_area_struct
*vma
)
2340 if ((!(vma
->vm_flags
& VM_HUGEPAGE
) && !khugepaged_always()) ||
2341 (vma
->vm_flags
& VM_NOHUGEPAGE
))
2343 if (!vma
->anon_vma
|| vma
->vm_ops
)
2345 if (is_vma_temporary_stack(vma
))
2347 return !(vma
->vm_flags
& VM_NO_THP
);
2351 * If mmap_sem temporarily dropped, revalidate vma
2352 * before taking mmap_sem.
2353 * Return 0 if succeeds, otherwise return none-zero
2354 * value (scan code).
2357 static int hugepage_vma_revalidate(struct mm_struct
*mm
, unsigned long address
)
2359 struct vm_area_struct
*vma
;
2360 unsigned long hstart
, hend
;
2362 if (unlikely(khugepaged_test_exit(mm
)))
2363 return SCAN_ANY_PROCESS
;
2365 vma
= find_vma(mm
, address
);
2367 return SCAN_VMA_NULL
;
2369 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2370 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2371 if (address
< hstart
|| address
+ HPAGE_PMD_SIZE
> hend
)
2372 return SCAN_ADDRESS_RANGE
;
2373 if (!hugepage_vma_check(vma
))
2374 return SCAN_VMA_CHECK
;
2379 * Bring missing pages in from swap, to complete THP collapse.
2380 * Only done if khugepaged_scan_pmd believes it is worthwhile.
2382 * Called and returns without pte mapped or spinlocks held,
2383 * but with mmap_sem held to protect against vma changes.
2386 static bool __collapse_huge_page_swapin(struct mm_struct
*mm
,
2387 struct vm_area_struct
*vma
,
2388 unsigned long address
, pmd_t
*pmd
)
2391 int swapped_in
= 0, ret
= 0;
2392 struct fault_env fe
= {
2395 .flags
= FAULT_FLAG_ALLOW_RETRY
,
2399 fe
.pte
= pte_offset_map(pmd
, address
);
2400 for (; fe
.address
< address
+ HPAGE_PMD_NR
*PAGE_SIZE
;
2401 fe
.pte
++, fe
.address
+= PAGE_SIZE
) {
2403 if (!is_swap_pte(pteval
))
2406 ret
= do_swap_page(&fe
, pteval
);
2407 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
2408 if (ret
& VM_FAULT_RETRY
) {
2409 down_read(&mm
->mmap_sem
);
2410 /* vma is no longer available, don't continue to swapin */
2411 if (hugepage_vma_revalidate(mm
, address
))
2413 /* check if the pmd is still valid */
2414 if (mm_find_pmd(mm
, address
) != pmd
)
2417 if (ret
& VM_FAULT_ERROR
) {
2418 trace_mm_collapse_huge_page_swapin(mm
, swapped_in
, 0);
2421 /* pte is unmapped now, we need to map it */
2422 fe
.pte
= pte_offset_map(pmd
, fe
.address
);
2426 trace_mm_collapse_huge_page_swapin(mm
, swapped_in
, 1);
2430 static void collapse_huge_page(struct mm_struct
*mm
,
2431 unsigned long address
,
2432 struct page
**hpage
,
2433 struct vm_area_struct
*vma
,
2439 struct page
*new_page
;
2440 spinlock_t
*pmd_ptl
, *pte_ptl
;
2441 int isolated
= 0, result
= 0;
2442 struct mem_cgroup
*memcg
;
2443 unsigned long mmun_start
; /* For mmu_notifiers */
2444 unsigned long mmun_end
; /* For mmu_notifiers */
2447 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2449 /* Only allocate from the target node */
2450 gfp
= alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE
| __GFP_THISNODE
;
2452 /* release the mmap_sem read lock. */
2453 new_page
= khugepaged_alloc_page(hpage
, gfp
, mm
, address
, node
);
2455 result
= SCAN_ALLOC_HUGE_PAGE_FAIL
;
2459 if (unlikely(mem_cgroup_try_charge(new_page
, mm
, gfp
, &memcg
, true))) {
2460 result
= SCAN_CGROUP_CHARGE_FAIL
;
2464 down_read(&mm
->mmap_sem
);
2465 result
= hugepage_vma_revalidate(mm
, address
);
2467 mem_cgroup_cancel_charge(new_page
, memcg
, true);
2468 up_read(&mm
->mmap_sem
);
2472 pmd
= mm_find_pmd(mm
, address
);
2474 result
= SCAN_PMD_NULL
;
2475 mem_cgroup_cancel_charge(new_page
, memcg
, true);
2476 up_read(&mm
->mmap_sem
);
2481 * __collapse_huge_page_swapin always returns with mmap_sem locked.
2482 * If it fails, release mmap_sem and jump directly out.
2483 * Continuing to collapse causes inconsistency.
2485 if (!__collapse_huge_page_swapin(mm
, vma
, address
, pmd
)) {
2486 mem_cgroup_cancel_charge(new_page
, memcg
, true);
2487 up_read(&mm
->mmap_sem
);
2491 up_read(&mm
->mmap_sem
);
2493 * Prevent all access to pagetables with the exception of
2494 * gup_fast later handled by the ptep_clear_flush and the VM
2495 * handled by the anon_vma lock + PG_lock.
2497 down_write(&mm
->mmap_sem
);
2498 result
= hugepage_vma_revalidate(mm
, address
);
2501 /* check if the pmd is still valid */
2502 if (mm_find_pmd(mm
, address
) != pmd
)
2505 anon_vma_lock_write(vma
->anon_vma
);
2507 pte
= pte_offset_map(pmd
, address
);
2508 pte_ptl
= pte_lockptr(mm
, pmd
);
2510 mmun_start
= address
;
2511 mmun_end
= address
+ HPAGE_PMD_SIZE
;
2512 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2513 pmd_ptl
= pmd_lock(mm
, pmd
); /* probably unnecessary */
2515 * After this gup_fast can't run anymore. This also removes
2516 * any huge TLB entry from the CPU so we won't allow
2517 * huge and small TLB entries for the same virtual address
2518 * to avoid the risk of CPU bugs in that area.
2520 _pmd
= pmdp_collapse_flush(vma
, address
, pmd
);
2521 spin_unlock(pmd_ptl
);
2522 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2525 isolated
= __collapse_huge_page_isolate(vma
, address
, pte
);
2526 spin_unlock(pte_ptl
);
2528 if (unlikely(!isolated
)) {
2531 BUG_ON(!pmd_none(*pmd
));
2533 * We can only use set_pmd_at when establishing
2534 * hugepmds and never for establishing regular pmds that
2535 * points to regular pagetables. Use pmd_populate for that
2537 pmd_populate(mm
, pmd
, pmd_pgtable(_pmd
));
2538 spin_unlock(pmd_ptl
);
2539 anon_vma_unlock_write(vma
->anon_vma
);
2545 * All pages are isolated and locked so anon_vma rmap
2546 * can't run anymore.
2548 anon_vma_unlock_write(vma
->anon_vma
);
2550 __collapse_huge_page_copy(pte
, new_page
, vma
, address
, pte_ptl
);
2552 __SetPageUptodate(new_page
);
2553 pgtable
= pmd_pgtable(_pmd
);
2555 _pmd
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
2556 _pmd
= maybe_pmd_mkwrite(pmd_mkdirty(_pmd
), vma
);
2559 * spin_lock() below is not the equivalent of smp_wmb(), so
2560 * this is needed to avoid the copy_huge_page writes to become
2561 * visible after the set_pmd_at() write.
2566 BUG_ON(!pmd_none(*pmd
));
2567 page_add_new_anon_rmap(new_page
, vma
, address
, true);
2568 mem_cgroup_commit_charge(new_page
, memcg
, false, true);
2569 lru_cache_add_active_or_unevictable(new_page
, vma
);
2570 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
2571 set_pmd_at(mm
, address
, pmd
, _pmd
);
2572 update_mmu_cache_pmd(vma
, address
, pmd
);
2573 spin_unlock(pmd_ptl
);
2577 khugepaged_pages_collapsed
++;
2578 result
= SCAN_SUCCEED
;
2580 up_write(&mm
->mmap_sem
);
2582 trace_mm_collapse_huge_page(mm
, isolated
, result
);
2585 mem_cgroup_cancel_charge(new_page
, memcg
, true);
2589 static int khugepaged_scan_pmd(struct mm_struct
*mm
,
2590 struct vm_area_struct
*vma
,
2591 unsigned long address
,
2592 struct page
**hpage
)
2596 int ret
= 0, none_or_zero
= 0, result
= 0;
2597 struct page
*page
= NULL
;
2598 unsigned long _address
;
2600 int node
= NUMA_NO_NODE
, unmapped
= 0;
2601 bool writable
= false, referenced
= false;
2603 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2605 pmd
= mm_find_pmd(mm
, address
);
2607 result
= SCAN_PMD_NULL
;
2611 memset(khugepaged_node_load
, 0, sizeof(khugepaged_node_load
));
2612 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2613 for (_address
= address
, _pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
2614 _pte
++, _address
+= PAGE_SIZE
) {
2615 pte_t pteval
= *_pte
;
2616 if (is_swap_pte(pteval
)) {
2617 if (++unmapped
<= khugepaged_max_ptes_swap
) {
2620 result
= SCAN_EXCEED_SWAP_PTE
;
2624 if (pte_none(pteval
) || is_zero_pfn(pte_pfn(pteval
))) {
2625 if (!userfaultfd_armed(vma
) &&
2626 ++none_or_zero
<= khugepaged_max_ptes_none
) {
2629 result
= SCAN_EXCEED_NONE_PTE
;
2633 if (!pte_present(pteval
)) {
2634 result
= SCAN_PTE_NON_PRESENT
;
2637 if (pte_write(pteval
))
2640 page
= vm_normal_page(vma
, _address
, pteval
);
2641 if (unlikely(!page
)) {
2642 result
= SCAN_PAGE_NULL
;
2646 /* TODO: teach khugepaged to collapse THP mapped with pte */
2647 if (PageCompound(page
)) {
2648 result
= SCAN_PAGE_COMPOUND
;
2653 * Record which node the original page is from and save this
2654 * information to khugepaged_node_load[].
2655 * Khupaged will allocate hugepage from the node has the max
2658 node
= page_to_nid(page
);
2659 if (khugepaged_scan_abort(node
)) {
2660 result
= SCAN_SCAN_ABORT
;
2663 khugepaged_node_load
[node
]++;
2664 if (!PageLRU(page
)) {
2665 result
= SCAN_PAGE_LRU
;
2668 if (PageLocked(page
)) {
2669 result
= SCAN_PAGE_LOCK
;
2672 if (!PageAnon(page
)) {
2673 result
= SCAN_PAGE_ANON
;
2678 * cannot use mapcount: can't collapse if there's a gup pin.
2679 * The page must only be referenced by the scanned process
2680 * and page swap cache.
2682 if (page_count(page
) != 1 + !!PageSwapCache(page
)) {
2683 result
= SCAN_PAGE_COUNT
;
2686 if (pte_young(pteval
) ||
2687 page_is_young(page
) || PageReferenced(page
) ||
2688 mmu_notifier_test_young(vma
->vm_mm
, address
))
2693 result
= SCAN_SUCCEED
;
2696 result
= SCAN_NO_REFERENCED_PAGE
;
2699 result
= SCAN_PAGE_RO
;
2702 pte_unmap_unlock(pte
, ptl
);
2704 node
= khugepaged_find_target_node();
2705 /* collapse_huge_page will return with the mmap_sem released */
2706 collapse_huge_page(mm
, address
, hpage
, vma
, node
);
2709 trace_mm_khugepaged_scan_pmd(mm
, page
, writable
, referenced
,
2710 none_or_zero
, result
, unmapped
);
2714 static void collect_mm_slot(struct mm_slot
*mm_slot
)
2716 struct mm_struct
*mm
= mm_slot
->mm
;
2718 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2720 if (khugepaged_test_exit(mm
)) {
2722 hash_del(&mm_slot
->hash
);
2723 list_del(&mm_slot
->mm_node
);
2726 * Not strictly needed because the mm exited already.
2728 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2731 /* khugepaged_mm_lock actually not necessary for the below */
2732 free_mm_slot(mm_slot
);
2737 static unsigned int khugepaged_scan_mm_slot(unsigned int pages
,
2738 struct page
**hpage
)
2739 __releases(&khugepaged_mm_lock
)
2740 __acquires(&khugepaged_mm_lock
)
2742 struct mm_slot
*mm_slot
;
2743 struct mm_struct
*mm
;
2744 struct vm_area_struct
*vma
;
2748 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2750 if (khugepaged_scan
.mm_slot
)
2751 mm_slot
= khugepaged_scan
.mm_slot
;
2753 mm_slot
= list_entry(khugepaged_scan
.mm_head
.next
,
2754 struct mm_slot
, mm_node
);
2755 khugepaged_scan
.address
= 0;
2756 khugepaged_scan
.mm_slot
= mm_slot
;
2758 spin_unlock(&khugepaged_mm_lock
);
2761 down_read(&mm
->mmap_sem
);
2762 if (unlikely(khugepaged_test_exit(mm
)))
2765 vma
= find_vma(mm
, khugepaged_scan
.address
);
2768 for (; vma
; vma
= vma
->vm_next
) {
2769 unsigned long hstart
, hend
;
2772 if (unlikely(khugepaged_test_exit(mm
))) {
2776 if (!hugepage_vma_check(vma
)) {
2781 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2782 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2785 if (khugepaged_scan
.address
> hend
)
2787 if (khugepaged_scan
.address
< hstart
)
2788 khugepaged_scan
.address
= hstart
;
2789 VM_BUG_ON(khugepaged_scan
.address
& ~HPAGE_PMD_MASK
);
2791 while (khugepaged_scan
.address
< hend
) {
2794 if (unlikely(khugepaged_test_exit(mm
)))
2795 goto breakouterloop
;
2797 VM_BUG_ON(khugepaged_scan
.address
< hstart
||
2798 khugepaged_scan
.address
+ HPAGE_PMD_SIZE
>
2800 ret
= khugepaged_scan_pmd(mm
, vma
,
2801 khugepaged_scan
.address
,
2803 /* move to next address */
2804 khugepaged_scan
.address
+= HPAGE_PMD_SIZE
;
2805 progress
+= HPAGE_PMD_NR
;
2807 /* we released mmap_sem so break loop */
2808 goto breakouterloop_mmap_sem
;
2809 if (progress
>= pages
)
2810 goto breakouterloop
;
2814 up_read(&mm
->mmap_sem
); /* exit_mmap will destroy ptes after this */
2815 breakouterloop_mmap_sem
:
2817 spin_lock(&khugepaged_mm_lock
);
2818 VM_BUG_ON(khugepaged_scan
.mm_slot
!= mm_slot
);
2820 * Release the current mm_slot if this mm is about to die, or
2821 * if we scanned all vmas of this mm.
2823 if (khugepaged_test_exit(mm
) || !vma
) {
2825 * Make sure that if mm_users is reaching zero while
2826 * khugepaged runs here, khugepaged_exit will find
2827 * mm_slot not pointing to the exiting mm.
2829 if (mm_slot
->mm_node
.next
!= &khugepaged_scan
.mm_head
) {
2830 khugepaged_scan
.mm_slot
= list_entry(
2831 mm_slot
->mm_node
.next
,
2832 struct mm_slot
, mm_node
);
2833 khugepaged_scan
.address
= 0;
2835 khugepaged_scan
.mm_slot
= NULL
;
2836 khugepaged_full_scans
++;
2839 collect_mm_slot(mm_slot
);
2845 static int khugepaged_has_work(void)
2847 return !list_empty(&khugepaged_scan
.mm_head
) &&
2848 khugepaged_enabled();
2851 static int khugepaged_wait_event(void)
2853 return !list_empty(&khugepaged_scan
.mm_head
) ||
2854 kthread_should_stop();
2857 static void khugepaged_do_scan(void)
2859 struct page
*hpage
= NULL
;
2860 unsigned int progress
= 0, pass_through_head
= 0;
2861 unsigned int pages
= khugepaged_pages_to_scan
;
2864 barrier(); /* write khugepaged_pages_to_scan to local stack */
2866 while (progress
< pages
) {
2867 if (!khugepaged_prealloc_page(&hpage
, &wait
))
2872 if (unlikely(kthread_should_stop() || try_to_freeze()))
2875 spin_lock(&khugepaged_mm_lock
);
2876 if (!khugepaged_scan
.mm_slot
)
2877 pass_through_head
++;
2878 if (khugepaged_has_work() &&
2879 pass_through_head
< 2)
2880 progress
+= khugepaged_scan_mm_slot(pages
- progress
,
2884 spin_unlock(&khugepaged_mm_lock
);
2887 if (!IS_ERR_OR_NULL(hpage
))
2891 static bool khugepaged_should_wakeup(void)
2893 return kthread_should_stop() ||
2894 time_after_eq(jiffies
, khugepaged_sleep_expire
);
2897 static void khugepaged_wait_work(void)
2899 if (khugepaged_has_work()) {
2900 const unsigned long scan_sleep_jiffies
=
2901 msecs_to_jiffies(khugepaged_scan_sleep_millisecs
);
2903 if (!scan_sleep_jiffies
)
2906 khugepaged_sleep_expire
= jiffies
+ scan_sleep_jiffies
;
2907 wait_event_freezable_timeout(khugepaged_wait
,
2908 khugepaged_should_wakeup(),
2909 scan_sleep_jiffies
);
2913 if (khugepaged_enabled())
2914 wait_event_freezable(khugepaged_wait
, khugepaged_wait_event());
2917 static int khugepaged(void *none
)
2919 struct mm_slot
*mm_slot
;
2922 set_user_nice(current
, MAX_NICE
);
2924 while (!kthread_should_stop()) {
2925 khugepaged_do_scan();
2926 khugepaged_wait_work();
2929 spin_lock(&khugepaged_mm_lock
);
2930 mm_slot
= khugepaged_scan
.mm_slot
;
2931 khugepaged_scan
.mm_slot
= NULL
;
2933 collect_mm_slot(mm_slot
);
2934 spin_unlock(&khugepaged_mm_lock
);
2938 static void __split_huge_zero_page_pmd(struct vm_area_struct
*vma
,
2939 unsigned long haddr
, pmd_t
*pmd
)
2941 struct mm_struct
*mm
= vma
->vm_mm
;
2946 /* leave pmd empty until pte is filled */
2947 pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
2949 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2950 pmd_populate(mm
, &_pmd
, pgtable
);
2952 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
2954 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
2955 entry
= pte_mkspecial(entry
);
2956 pte
= pte_offset_map(&_pmd
, haddr
);
2957 VM_BUG_ON(!pte_none(*pte
));
2958 set_pte_at(mm
, haddr
, pte
, entry
);
2961 smp_wmb(); /* make pte visible before pmd */
2962 pmd_populate(mm
, pmd
, pgtable
);
2963 put_huge_zero_page();
2966 static void __split_huge_pmd_locked(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2967 unsigned long haddr
, bool freeze
)
2969 struct mm_struct
*mm
= vma
->vm_mm
;
2973 bool young
, write
, dirty
;
2977 VM_BUG_ON(haddr
& ~HPAGE_PMD_MASK
);
2978 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
2979 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PMD_SIZE
, vma
);
2980 VM_BUG_ON(!pmd_trans_huge(*pmd
) && !pmd_devmap(*pmd
));
2982 count_vm_event(THP_SPLIT_PMD
);
2984 if (vma_is_dax(vma
)) {
2985 pmd_t _pmd
= pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
2986 if (is_huge_zero_pmd(_pmd
))
2987 put_huge_zero_page();
2989 } else if (is_huge_zero_pmd(*pmd
)) {
2990 return __split_huge_zero_page_pmd(vma
, haddr
, pmd
);
2993 page
= pmd_page(*pmd
);
2994 VM_BUG_ON_PAGE(!page_count(page
), page
);
2995 page_ref_add(page
, HPAGE_PMD_NR
- 1);
2996 write
= pmd_write(*pmd
);
2997 young
= pmd_young(*pmd
);
2998 dirty
= pmd_dirty(*pmd
);
3000 pmdp_huge_split_prepare(vma
, haddr
, pmd
);
3001 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
3002 pmd_populate(mm
, &_pmd
, pgtable
);
3004 for (i
= 0, addr
= haddr
; i
< HPAGE_PMD_NR
; i
++, addr
+= PAGE_SIZE
) {
3007 * Note that NUMA hinting access restrictions are not
3008 * transferred to avoid any possibility of altering
3009 * permissions across VMAs.
3012 swp_entry_t swp_entry
;
3013 swp_entry
= make_migration_entry(page
+ i
, write
);
3014 entry
= swp_entry_to_pte(swp_entry
);
3016 entry
= mk_pte(page
+ i
, vma
->vm_page_prot
);
3017 entry
= maybe_mkwrite(entry
, vma
);
3019 entry
= pte_wrprotect(entry
);
3021 entry
= pte_mkold(entry
);
3024 SetPageDirty(page
+ i
);
3025 pte
= pte_offset_map(&_pmd
, addr
);
3026 BUG_ON(!pte_none(*pte
));
3027 set_pte_at(mm
, addr
, pte
, entry
);
3028 atomic_inc(&page
[i
]._mapcount
);
3033 * Set PG_double_map before dropping compound_mapcount to avoid
3034 * false-negative page_mapped().
3036 if (compound_mapcount(page
) > 1 && !TestSetPageDoubleMap(page
)) {
3037 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3038 atomic_inc(&page
[i
]._mapcount
);
3041 if (atomic_add_negative(-1, compound_mapcount_ptr(page
))) {
3042 /* Last compound_mapcount is gone. */
3043 __dec_zone_page_state(page
, NR_ANON_TRANSPARENT_HUGEPAGES
);
3044 if (TestClearPageDoubleMap(page
)) {
3045 /* No need in mapcount reference anymore */
3046 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3047 atomic_dec(&page
[i
]._mapcount
);
3051 smp_wmb(); /* make pte visible before pmd */
3053 * Up to this point the pmd is present and huge and userland has the
3054 * whole access to the hugepage during the split (which happens in
3055 * place). If we overwrite the pmd with the not-huge version pointing
3056 * to the pte here (which of course we could if all CPUs were bug
3057 * free), userland could trigger a small page size TLB miss on the
3058 * small sized TLB while the hugepage TLB entry is still established in
3059 * the huge TLB. Some CPU doesn't like that.
3060 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
3061 * 383 on page 93. Intel should be safe but is also warns that it's
3062 * only safe if the permission and cache attributes of the two entries
3063 * loaded in the two TLB is identical (which should be the case here).
3064 * But it is generally safer to never allow small and huge TLB entries
3065 * for the same virtual address to be loaded simultaneously. So instead
3066 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
3067 * current pmd notpresent (atomically because here the pmd_trans_huge
3068 * and pmd_trans_splitting must remain set at all times on the pmd
3069 * until the split is complete for this pmd), then we flush the SMP TLB
3070 * and finally we write the non-huge version of the pmd entry with
3073 pmdp_invalidate(vma
, haddr
, pmd
);
3074 pmd_populate(mm
, pmd
, pgtable
);
3077 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
3078 page_remove_rmap(page
+ i
, false);
3084 void __split_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
3085 unsigned long address
, bool freeze
, struct page
*page
)
3088 struct mm_struct
*mm
= vma
->vm_mm
;
3089 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
3091 mmu_notifier_invalidate_range_start(mm
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
3092 ptl
= pmd_lock(mm
, pmd
);
3095 * If caller asks to setup a migration entries, we need a page to check
3096 * pmd against. Otherwise we can end up replacing wrong page.
3098 VM_BUG_ON(freeze
&& !page
);
3099 if (page
&& page
!= pmd_page(*pmd
))
3102 if (pmd_trans_huge(*pmd
)) {
3103 page
= pmd_page(*pmd
);
3104 if (PageMlocked(page
))
3105 clear_page_mlock(page
);
3106 } else if (!pmd_devmap(*pmd
))
3108 __split_huge_pmd_locked(vma
, pmd
, haddr
, freeze
);
3111 mmu_notifier_invalidate_range_end(mm
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
3114 void split_huge_pmd_address(struct vm_area_struct
*vma
, unsigned long address
,
3115 bool freeze
, struct page
*page
)
3121 pgd
= pgd_offset(vma
->vm_mm
, address
);
3122 if (!pgd_present(*pgd
))
3125 pud
= pud_offset(pgd
, address
);
3126 if (!pud_present(*pud
))
3129 pmd
= pmd_offset(pud
, address
);
3131 __split_huge_pmd(vma
, pmd
, address
, freeze
, page
);
3134 void vma_adjust_trans_huge(struct vm_area_struct
*vma
,
3135 unsigned long start
,
3140 * If the new start address isn't hpage aligned and it could
3141 * previously contain an hugepage: check if we need to split
3144 if (start
& ~HPAGE_PMD_MASK
&&
3145 (start
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
3146 (start
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
3147 split_huge_pmd_address(vma
, start
, false, NULL
);
3150 * If the new end address isn't hpage aligned and it could
3151 * previously contain an hugepage: check if we need to split
3154 if (end
& ~HPAGE_PMD_MASK
&&
3155 (end
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
3156 (end
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
3157 split_huge_pmd_address(vma
, end
, false, NULL
);
3160 * If we're also updating the vma->vm_next->vm_start, if the new
3161 * vm_next->vm_start isn't page aligned and it could previously
3162 * contain an hugepage: check if we need to split an huge pmd.
3164 if (adjust_next
> 0) {
3165 struct vm_area_struct
*next
= vma
->vm_next
;
3166 unsigned long nstart
= next
->vm_start
;
3167 nstart
+= adjust_next
<< PAGE_SHIFT
;
3168 if (nstart
& ~HPAGE_PMD_MASK
&&
3169 (nstart
& HPAGE_PMD_MASK
) >= next
->vm_start
&&
3170 (nstart
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= next
->vm_end
)
3171 split_huge_pmd_address(next
, nstart
, false, NULL
);
3175 static void freeze_page(struct page
*page
)
3177 enum ttu_flags ttu_flags
= TTU_MIGRATION
| TTU_IGNORE_MLOCK
|
3178 TTU_IGNORE_ACCESS
| TTU_RMAP_LOCKED
;
3181 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3183 /* We only need TTU_SPLIT_HUGE_PMD once */
3184 ret
= try_to_unmap(page
, ttu_flags
| TTU_SPLIT_HUGE_PMD
);
3185 for (i
= 1; !ret
&& i
< HPAGE_PMD_NR
; i
++) {
3186 /* Cut short if the page is unmapped */
3187 if (page_count(page
) == 1)
3190 ret
= try_to_unmap(page
+ i
, ttu_flags
);
3195 static void unfreeze_page(struct page
*page
)
3199 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3200 remove_migration_ptes(page
+ i
, page
+ i
, true);
3203 static void __split_huge_page_tail(struct page
*head
, int tail
,
3204 struct lruvec
*lruvec
, struct list_head
*list
)
3206 struct page
*page_tail
= head
+ tail
;
3208 VM_BUG_ON_PAGE(atomic_read(&page_tail
->_mapcount
) != -1, page_tail
);
3209 VM_BUG_ON_PAGE(page_ref_count(page_tail
) != 0, page_tail
);
3212 * tail_page->_refcount is zero and not changing from under us. But
3213 * get_page_unless_zero() may be running from under us on the
3214 * tail_page. If we used atomic_set() below instead of atomic_inc(), we
3215 * would then run atomic_set() concurrently with
3216 * get_page_unless_zero(), and atomic_set() is implemented in C not
3217 * using locked ops. spin_unlock on x86 sometime uses locked ops
3218 * because of PPro errata 66, 92, so unless somebody can guarantee
3219 * atomic_set() here would be safe on all archs (and not only on x86),
3220 * it's safer to use atomic_inc().
3222 page_ref_inc(page_tail
);
3224 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
3225 page_tail
->flags
|= (head
->flags
&
3226 ((1L << PG_referenced
) |
3227 (1L << PG_swapbacked
) |
3228 (1L << PG_mlocked
) |
3229 (1L << PG_uptodate
) |
3232 (1L << PG_unevictable
) |
3236 * After clearing PageTail the gup refcount can be released.
3237 * Page flags also must be visible before we make the page non-compound.
3241 clear_compound_head(page_tail
);
3243 if (page_is_young(head
))
3244 set_page_young(page_tail
);
3245 if (page_is_idle(head
))
3246 set_page_idle(page_tail
);
3248 /* ->mapping in first tail page is compound_mapcount */
3249 VM_BUG_ON_PAGE(tail
> 2 && page_tail
->mapping
!= TAIL_MAPPING
,
3251 page_tail
->mapping
= head
->mapping
;
3253 page_tail
->index
= head
->index
+ tail
;
3254 page_cpupid_xchg_last(page_tail
, page_cpupid_last(head
));
3255 lru_add_page_tail(head
, page_tail
, lruvec
, list
);
3258 static void __split_huge_page(struct page
*page
, struct list_head
*list
)
3260 struct page
*head
= compound_head(page
);
3261 struct zone
*zone
= page_zone(head
);
3262 struct lruvec
*lruvec
;
3265 /* prevent PageLRU to go away from under us, and freeze lru stats */
3266 spin_lock_irq(&zone
->lru_lock
);
3267 lruvec
= mem_cgroup_page_lruvec(head
, zone
);
3269 /* complete memcg works before add pages to LRU */
3270 mem_cgroup_split_huge_fixup(head
);
3272 for (i
= HPAGE_PMD_NR
- 1; i
>= 1; i
--)
3273 __split_huge_page_tail(head
, i
, lruvec
, list
);
3275 ClearPageCompound(head
);
3276 spin_unlock_irq(&zone
->lru_lock
);
3278 unfreeze_page(head
);
3280 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
3281 struct page
*subpage
= head
+ i
;
3282 if (subpage
== page
)
3284 unlock_page(subpage
);
3287 * Subpages may be freed if there wasn't any mapping
3288 * like if add_to_swap() is running on a lru page that
3289 * had its mapping zapped. And freeing these pages
3290 * requires taking the lru_lock so we do the put_page
3291 * of the tail pages after the split is complete.
3297 int total_mapcount(struct page
*page
)
3299 int i
, compound
, ret
;
3301 VM_BUG_ON_PAGE(PageTail(page
), page
);
3303 if (likely(!PageCompound(page
)))
3304 return atomic_read(&page
->_mapcount
) + 1;
3306 compound
= compound_mapcount(page
);
3310 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3311 ret
+= atomic_read(&page
[i
]._mapcount
) + 1;
3312 /* File pages has compound_mapcount included in _mapcount */
3313 if (!PageAnon(page
))
3314 return ret
- compound
* HPAGE_PMD_NR
;
3315 if (PageDoubleMap(page
))
3316 ret
-= HPAGE_PMD_NR
;
3321 * This calculates accurately how many mappings a transparent hugepage
3322 * has (unlike page_mapcount() which isn't fully accurate). This full
3323 * accuracy is primarily needed to know if copy-on-write faults can
3324 * reuse the page and change the mapping to read-write instead of
3325 * copying them. At the same time this returns the total_mapcount too.
3327 * The function returns the highest mapcount any one of the subpages
3328 * has. If the return value is one, even if different processes are
3329 * mapping different subpages of the transparent hugepage, they can
3330 * all reuse it, because each process is reusing a different subpage.
3332 * The total_mapcount is instead counting all virtual mappings of the
3333 * subpages. If the total_mapcount is equal to "one", it tells the
3334 * caller all mappings belong to the same "mm" and in turn the
3335 * anon_vma of the transparent hugepage can become the vma->anon_vma
3336 * local one as no other process may be mapping any of the subpages.
3338 * It would be more accurate to replace page_mapcount() with
3339 * page_trans_huge_mapcount(), however we only use
3340 * page_trans_huge_mapcount() in the copy-on-write faults where we
3341 * need full accuracy to avoid breaking page pinning, because
3342 * page_trans_huge_mapcount() is slower than page_mapcount().
3344 int page_trans_huge_mapcount(struct page
*page
, int *total_mapcount
)
3346 int i
, ret
, _total_mapcount
, mapcount
;
3348 /* hugetlbfs shouldn't call it */
3349 VM_BUG_ON_PAGE(PageHuge(page
), page
);
3351 if (likely(!PageTransCompound(page
))) {
3352 mapcount
= atomic_read(&page
->_mapcount
) + 1;
3354 *total_mapcount
= mapcount
;
3358 page
= compound_head(page
);
3360 _total_mapcount
= ret
= 0;
3361 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
3362 mapcount
= atomic_read(&page
[i
]._mapcount
) + 1;
3363 ret
= max(ret
, mapcount
);
3364 _total_mapcount
+= mapcount
;
3366 if (PageDoubleMap(page
)) {
3368 _total_mapcount
-= HPAGE_PMD_NR
;
3370 mapcount
= compound_mapcount(page
);
3372 _total_mapcount
+= mapcount
;
3374 *total_mapcount
= _total_mapcount
;
3379 * This function splits huge page into normal pages. @page can point to any
3380 * subpage of huge page to split. Split doesn't change the position of @page.
3382 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3383 * The huge page must be locked.
3385 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3387 * Both head page and tail pages will inherit mapping, flags, and so on from
3390 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3391 * they are not mapped.
3393 * Returns 0 if the hugepage is split successfully.
3394 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3397 int split_huge_page_to_list(struct page
*page
, struct list_head
*list
)
3399 struct page
*head
= compound_head(page
);
3400 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(head
));
3401 struct anon_vma
*anon_vma
;
3402 int count
, mapcount
, ret
;
3404 unsigned long flags
;
3406 VM_BUG_ON_PAGE(is_huge_zero_page(page
), page
);
3407 VM_BUG_ON_PAGE(!PageAnon(page
), page
);
3408 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
3409 VM_BUG_ON_PAGE(!PageSwapBacked(page
), page
);
3410 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
3413 * The caller does not necessarily hold an mmap_sem that would prevent
3414 * the anon_vma disappearing so we first we take a reference to it
3415 * and then lock the anon_vma for write. This is similar to
3416 * page_lock_anon_vma_read except the write lock is taken to serialise
3417 * against parallel split or collapse operations.
3419 anon_vma
= page_get_anon_vma(head
);
3424 anon_vma_lock_write(anon_vma
);
3427 * Racy check if we can split the page, before freeze_page() will
3430 if (total_mapcount(head
) != page_count(head
) - 1) {
3435 mlocked
= PageMlocked(page
);
3437 VM_BUG_ON_PAGE(compound_mapcount(head
), head
);
3439 /* Make sure the page is not on per-CPU pagevec as it takes pin */
3443 /* Prevent deferred_split_scan() touching ->_refcount */
3444 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
3445 count
= page_count(head
);
3446 mapcount
= total_mapcount(head
);
3447 if (!mapcount
&& count
== 1) {
3448 if (!list_empty(page_deferred_list(head
))) {
3449 pgdata
->split_queue_len
--;
3450 list_del(page_deferred_list(head
));
3452 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
3453 __split_huge_page(page
, list
);
3455 } else if (IS_ENABLED(CONFIG_DEBUG_VM
) && mapcount
) {
3456 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
3457 pr_alert("total_mapcount: %u, page_count(): %u\n",
3460 dump_page(head
, NULL
);
3461 dump_page(page
, "total_mapcount(head) > 0");
3464 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
3465 unfreeze_page(head
);
3470 anon_vma_unlock_write(anon_vma
);
3471 put_anon_vma(anon_vma
);
3473 count_vm_event(!ret
? THP_SPLIT_PAGE
: THP_SPLIT_PAGE_FAILED
);
3477 void free_transhuge_page(struct page
*page
)
3479 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(page
));
3480 unsigned long flags
;
3482 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
3483 if (!list_empty(page_deferred_list(page
))) {
3484 pgdata
->split_queue_len
--;
3485 list_del(page_deferred_list(page
));
3487 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
3488 free_compound_page(page
);
3491 void deferred_split_huge_page(struct page
*page
)
3493 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(page
));
3494 unsigned long flags
;
3496 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3498 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
3499 if (list_empty(page_deferred_list(page
))) {
3500 count_vm_event(THP_DEFERRED_SPLIT_PAGE
);
3501 list_add_tail(page_deferred_list(page
), &pgdata
->split_queue
);
3502 pgdata
->split_queue_len
++;
3504 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
3507 static unsigned long deferred_split_count(struct shrinker
*shrink
,
3508 struct shrink_control
*sc
)
3510 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
3511 return ACCESS_ONCE(pgdata
->split_queue_len
);
3514 static unsigned long deferred_split_scan(struct shrinker
*shrink
,
3515 struct shrink_control
*sc
)
3517 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
3518 unsigned long flags
;
3519 LIST_HEAD(list
), *pos
, *next
;
3523 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
3524 /* Take pin on all head pages to avoid freeing them under us */
3525 list_for_each_safe(pos
, next
, &pgdata
->split_queue
) {
3526 page
= list_entry((void *)pos
, struct page
, mapping
);
3527 page
= compound_head(page
);
3528 if (get_page_unless_zero(page
)) {
3529 list_move(page_deferred_list(page
), &list
);
3531 /* We lost race with put_compound_page() */
3532 list_del_init(page_deferred_list(page
));
3533 pgdata
->split_queue_len
--;
3535 if (!--sc
->nr_to_scan
)
3538 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
3540 list_for_each_safe(pos
, next
, &list
) {
3541 page
= list_entry((void *)pos
, struct page
, mapping
);
3543 /* split_huge_page() removes page from list on success */
3544 if (!split_huge_page(page
))
3550 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
3551 list_splice_tail(&list
, &pgdata
->split_queue
);
3552 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
3555 * Stop shrinker if we didn't split any page, but the queue is empty.
3556 * This can happen if pages were freed under us.
3558 if (!split
&& list_empty(&pgdata
->split_queue
))
3563 static struct shrinker deferred_split_shrinker
= {
3564 .count_objects
= deferred_split_count
,
3565 .scan_objects
= deferred_split_scan
,
3566 .seeks
= DEFAULT_SEEKS
,
3567 .flags
= SHRINKER_NUMA_AWARE
,
3570 #ifdef CONFIG_DEBUG_FS
3571 static int split_huge_pages_set(void *data
, u64 val
)
3575 unsigned long pfn
, max_zone_pfn
;
3576 unsigned long total
= 0, split
= 0;
3581 for_each_populated_zone(zone
) {
3582 max_zone_pfn
= zone_end_pfn(zone
);
3583 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++) {
3584 if (!pfn_valid(pfn
))
3587 page
= pfn_to_page(pfn
);
3588 if (!get_page_unless_zero(page
))
3591 if (zone
!= page_zone(page
))
3594 if (!PageHead(page
) || !PageAnon(page
) ||
3600 if (!split_huge_page(page
))
3608 pr_info("%lu of %lu THP split\n", split
, total
);
3612 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops
, NULL
, split_huge_pages_set
,
3615 static int __init
split_huge_pages_debugfs(void)
3619 ret
= debugfs_create_file("split_huge_pages", 0200, NULL
, NULL
,
3620 &split_huge_pages_fops
);
3622 pr_warn("Failed to create split_huge_pages in debugfs");
3625 late_initcall(split_huge_pages_debugfs
);