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/pagemap.h>
27 #include <linux/debugfs.h>
28 #include <linux/migrate.h>
29 #include <linux/hashtable.h>
30 #include <linux/userfaultfd_k.h>
31 #include <linux/page_idle.h>
34 #include <asm/pgalloc.h>
44 SCAN_NO_REFERENCED_PAGE
,
58 SCAN_ALLOC_HUGE_PAGE_FAIL
,
59 SCAN_CGROUP_CHARGE_FAIL
62 #define CREATE_TRACE_POINTS
63 #include <trace/events/huge_memory.h>
66 * By default transparent hugepage support is disabled in order that avoid
67 * to risk increase the memory footprint of applications without a guaranteed
68 * benefit. When transparent hugepage support is enabled, is for all mappings,
69 * and khugepaged scans all mappings.
70 * Defrag is invoked by khugepaged hugepage allocations and by page faults
71 * for all hugepage allocations.
73 unsigned long transparent_hugepage_flags __read_mostly
=
74 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
75 (1<<TRANSPARENT_HUGEPAGE_FLAG
)|
77 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
78 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
)|
80 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
)|
81 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
)|
82 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
84 /* default scan 8*512 pte (or vmas) every 30 second */
85 static unsigned int khugepaged_pages_to_scan __read_mostly
= HPAGE_PMD_NR
*8;
86 static unsigned int khugepaged_pages_collapsed
;
87 static unsigned int khugepaged_full_scans
;
88 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly
= 10000;
89 /* during fragmentation poll the hugepage allocator once every minute */
90 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly
= 60000;
91 static struct task_struct
*khugepaged_thread __read_mostly
;
92 static DEFINE_MUTEX(khugepaged_mutex
);
93 static DEFINE_SPINLOCK(khugepaged_mm_lock
);
94 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait
);
96 * default collapse hugepages if there is at least one pte mapped like
97 * it would have happened if the vma was large enough during page
100 static unsigned int khugepaged_max_ptes_none __read_mostly
= HPAGE_PMD_NR
-1;
102 static int khugepaged(void *none
);
103 static int khugepaged_slab_init(void);
104 static void khugepaged_slab_exit(void);
106 #define MM_SLOTS_HASH_BITS 10
107 static __read_mostly
DEFINE_HASHTABLE(mm_slots_hash
, MM_SLOTS_HASH_BITS
);
109 static struct kmem_cache
*mm_slot_cache __read_mostly
;
112 * struct mm_slot - hash lookup from mm to mm_slot
113 * @hash: hash collision list
114 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
115 * @mm: the mm that this information is valid for
118 struct hlist_node hash
;
119 struct list_head mm_node
;
120 struct mm_struct
*mm
;
124 * struct khugepaged_scan - cursor for scanning
125 * @mm_head: the head of the mm list to scan
126 * @mm_slot: the current mm_slot we are scanning
127 * @address: the next address inside that to be scanned
129 * There is only the one khugepaged_scan instance of this cursor structure.
131 struct khugepaged_scan
{
132 struct list_head mm_head
;
133 struct mm_slot
*mm_slot
;
134 unsigned long address
;
136 static struct khugepaged_scan khugepaged_scan
= {
137 .mm_head
= LIST_HEAD_INIT(khugepaged_scan
.mm_head
),
140 static DEFINE_SPINLOCK(split_queue_lock
);
141 static LIST_HEAD(split_queue
);
142 static unsigned long split_queue_len
;
143 static struct shrinker deferred_split_shrinker
;
145 static void set_recommended_min_free_kbytes(void)
149 unsigned long recommended_min
;
151 for_each_populated_zone(zone
)
154 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
155 recommended_min
= pageblock_nr_pages
* nr_zones
* 2;
158 * Make sure that on average at least two pageblocks are almost free
159 * of another type, one for a migratetype to fall back to and a
160 * second to avoid subsequent fallbacks of other types There are 3
161 * MIGRATE_TYPES we care about.
163 recommended_min
+= pageblock_nr_pages
* nr_zones
*
164 MIGRATE_PCPTYPES
* MIGRATE_PCPTYPES
;
166 /* don't ever allow to reserve more than 5% of the lowmem */
167 recommended_min
= min(recommended_min
,
168 (unsigned long) nr_free_buffer_pages() / 20);
169 recommended_min
<<= (PAGE_SHIFT
-10);
171 if (recommended_min
> min_free_kbytes
) {
172 if (user_min_free_kbytes
>= 0)
173 pr_info("raising min_free_kbytes from %d to %lu "
174 "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 static 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
double_flag_show(struct kobject
*kobj
,
276 struct kobj_attribute
*attr
, char *buf
,
277 enum transparent_hugepage_flag enabled
,
278 enum transparent_hugepage_flag req_madv
)
280 if (test_bit(enabled
, &transparent_hugepage_flags
)) {
281 VM_BUG_ON(test_bit(req_madv
, &transparent_hugepage_flags
));
282 return sprintf(buf
, "[always] madvise never\n");
283 } else if (test_bit(req_madv
, &transparent_hugepage_flags
))
284 return sprintf(buf
, "always [madvise] never\n");
286 return sprintf(buf
, "always madvise [never]\n");
288 static ssize_t
double_flag_store(struct kobject
*kobj
,
289 struct kobj_attribute
*attr
,
290 const char *buf
, size_t count
,
291 enum transparent_hugepage_flag enabled
,
292 enum transparent_hugepage_flag req_madv
)
294 if (!memcmp("always", buf
,
295 min(sizeof("always")-1, count
))) {
296 set_bit(enabled
, &transparent_hugepage_flags
);
297 clear_bit(req_madv
, &transparent_hugepage_flags
);
298 } else if (!memcmp("madvise", buf
,
299 min(sizeof("madvise")-1, count
))) {
300 clear_bit(enabled
, &transparent_hugepage_flags
);
301 set_bit(req_madv
, &transparent_hugepage_flags
);
302 } else if (!memcmp("never", buf
,
303 min(sizeof("never")-1, count
))) {
304 clear_bit(enabled
, &transparent_hugepage_flags
);
305 clear_bit(req_madv
, &transparent_hugepage_flags
);
312 static ssize_t
enabled_show(struct kobject
*kobj
,
313 struct kobj_attribute
*attr
, char *buf
)
315 return double_flag_show(kobj
, attr
, buf
,
316 TRANSPARENT_HUGEPAGE_FLAG
,
317 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
319 static ssize_t
enabled_store(struct kobject
*kobj
,
320 struct kobj_attribute
*attr
,
321 const char *buf
, size_t count
)
325 ret
= double_flag_store(kobj
, attr
, buf
, count
,
326 TRANSPARENT_HUGEPAGE_FLAG
,
327 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
332 mutex_lock(&khugepaged_mutex
);
333 err
= start_stop_khugepaged();
334 mutex_unlock(&khugepaged_mutex
);
342 static struct kobj_attribute enabled_attr
=
343 __ATTR(enabled
, 0644, enabled_show
, enabled_store
);
345 static ssize_t
single_flag_show(struct kobject
*kobj
,
346 struct kobj_attribute
*attr
, char *buf
,
347 enum transparent_hugepage_flag flag
)
349 return sprintf(buf
, "%d\n",
350 !!test_bit(flag
, &transparent_hugepage_flags
));
353 static ssize_t
single_flag_store(struct kobject
*kobj
,
354 struct kobj_attribute
*attr
,
355 const char *buf
, size_t count
,
356 enum transparent_hugepage_flag flag
)
361 ret
= kstrtoul(buf
, 10, &value
);
368 set_bit(flag
, &transparent_hugepage_flags
);
370 clear_bit(flag
, &transparent_hugepage_flags
);
376 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
377 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
378 * memory just to allocate one more hugepage.
380 static ssize_t
defrag_show(struct kobject
*kobj
,
381 struct kobj_attribute
*attr
, char *buf
)
383 return double_flag_show(kobj
, attr
, buf
,
384 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
,
385 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
);
387 static ssize_t
defrag_store(struct kobject
*kobj
,
388 struct kobj_attribute
*attr
,
389 const char *buf
, size_t count
)
391 return double_flag_store(kobj
, attr
, buf
, count
,
392 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
,
393 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
);
395 static struct kobj_attribute defrag_attr
=
396 __ATTR(defrag
, 0644, defrag_show
, defrag_store
);
398 static ssize_t
use_zero_page_show(struct kobject
*kobj
,
399 struct kobj_attribute
*attr
, char *buf
)
401 return single_flag_show(kobj
, attr
, buf
,
402 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
404 static ssize_t
use_zero_page_store(struct kobject
*kobj
,
405 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
407 return single_flag_store(kobj
, attr
, buf
, count
,
408 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
410 static struct kobj_attribute use_zero_page_attr
=
411 __ATTR(use_zero_page
, 0644, use_zero_page_show
, use_zero_page_store
);
412 #ifdef CONFIG_DEBUG_VM
413 static ssize_t
debug_cow_show(struct kobject
*kobj
,
414 struct kobj_attribute
*attr
, char *buf
)
416 return single_flag_show(kobj
, attr
, buf
,
417 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
419 static ssize_t
debug_cow_store(struct kobject
*kobj
,
420 struct kobj_attribute
*attr
,
421 const char *buf
, size_t count
)
423 return single_flag_store(kobj
, attr
, buf
, count
,
424 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
426 static struct kobj_attribute debug_cow_attr
=
427 __ATTR(debug_cow
, 0644, debug_cow_show
, debug_cow_store
);
428 #endif /* CONFIG_DEBUG_VM */
430 static struct attribute
*hugepage_attr
[] = {
433 &use_zero_page_attr
.attr
,
434 #ifdef CONFIG_DEBUG_VM
435 &debug_cow_attr
.attr
,
440 static struct attribute_group hugepage_attr_group
= {
441 .attrs
= hugepage_attr
,
444 static ssize_t
scan_sleep_millisecs_show(struct kobject
*kobj
,
445 struct kobj_attribute
*attr
,
448 return sprintf(buf
, "%u\n", khugepaged_scan_sleep_millisecs
);
451 static ssize_t
scan_sleep_millisecs_store(struct kobject
*kobj
,
452 struct kobj_attribute
*attr
,
453 const char *buf
, size_t count
)
458 err
= kstrtoul(buf
, 10, &msecs
);
459 if (err
|| msecs
> UINT_MAX
)
462 khugepaged_scan_sleep_millisecs
= msecs
;
463 wake_up_interruptible(&khugepaged_wait
);
467 static struct kobj_attribute scan_sleep_millisecs_attr
=
468 __ATTR(scan_sleep_millisecs
, 0644, scan_sleep_millisecs_show
,
469 scan_sleep_millisecs_store
);
471 static ssize_t
alloc_sleep_millisecs_show(struct kobject
*kobj
,
472 struct kobj_attribute
*attr
,
475 return sprintf(buf
, "%u\n", khugepaged_alloc_sleep_millisecs
);
478 static ssize_t
alloc_sleep_millisecs_store(struct kobject
*kobj
,
479 struct kobj_attribute
*attr
,
480 const char *buf
, size_t count
)
485 err
= kstrtoul(buf
, 10, &msecs
);
486 if (err
|| msecs
> UINT_MAX
)
489 khugepaged_alloc_sleep_millisecs
= msecs
;
490 wake_up_interruptible(&khugepaged_wait
);
494 static struct kobj_attribute alloc_sleep_millisecs_attr
=
495 __ATTR(alloc_sleep_millisecs
, 0644, alloc_sleep_millisecs_show
,
496 alloc_sleep_millisecs_store
);
498 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
499 struct kobj_attribute
*attr
,
502 return sprintf(buf
, "%u\n", khugepaged_pages_to_scan
);
504 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
505 struct kobj_attribute
*attr
,
506 const char *buf
, size_t count
)
511 err
= kstrtoul(buf
, 10, &pages
);
512 if (err
|| !pages
|| pages
> UINT_MAX
)
515 khugepaged_pages_to_scan
= pages
;
519 static struct kobj_attribute pages_to_scan_attr
=
520 __ATTR(pages_to_scan
, 0644, pages_to_scan_show
,
521 pages_to_scan_store
);
523 static ssize_t
pages_collapsed_show(struct kobject
*kobj
,
524 struct kobj_attribute
*attr
,
527 return sprintf(buf
, "%u\n", khugepaged_pages_collapsed
);
529 static struct kobj_attribute pages_collapsed_attr
=
530 __ATTR_RO(pages_collapsed
);
532 static ssize_t
full_scans_show(struct kobject
*kobj
,
533 struct kobj_attribute
*attr
,
536 return sprintf(buf
, "%u\n", khugepaged_full_scans
);
538 static struct kobj_attribute full_scans_attr
=
539 __ATTR_RO(full_scans
);
541 static ssize_t
khugepaged_defrag_show(struct kobject
*kobj
,
542 struct kobj_attribute
*attr
, char *buf
)
544 return single_flag_show(kobj
, attr
, buf
,
545 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
547 static ssize_t
khugepaged_defrag_store(struct kobject
*kobj
,
548 struct kobj_attribute
*attr
,
549 const char *buf
, size_t count
)
551 return single_flag_store(kobj
, attr
, buf
, count
,
552 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
554 static struct kobj_attribute khugepaged_defrag_attr
=
555 __ATTR(defrag
, 0644, khugepaged_defrag_show
,
556 khugepaged_defrag_store
);
559 * max_ptes_none controls if khugepaged should collapse hugepages over
560 * any unmapped ptes in turn potentially increasing the memory
561 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
562 * reduce the available free memory in the system as it
563 * runs. Increasing max_ptes_none will instead potentially reduce the
564 * free memory in the system during the khugepaged scan.
566 static ssize_t
khugepaged_max_ptes_none_show(struct kobject
*kobj
,
567 struct kobj_attribute
*attr
,
570 return sprintf(buf
, "%u\n", khugepaged_max_ptes_none
);
572 static ssize_t
khugepaged_max_ptes_none_store(struct kobject
*kobj
,
573 struct kobj_attribute
*attr
,
574 const char *buf
, size_t count
)
577 unsigned long max_ptes_none
;
579 err
= kstrtoul(buf
, 10, &max_ptes_none
);
580 if (err
|| max_ptes_none
> HPAGE_PMD_NR
-1)
583 khugepaged_max_ptes_none
= max_ptes_none
;
587 static struct kobj_attribute khugepaged_max_ptes_none_attr
=
588 __ATTR(max_ptes_none
, 0644, khugepaged_max_ptes_none_show
,
589 khugepaged_max_ptes_none_store
);
591 static struct attribute
*khugepaged_attr
[] = {
592 &khugepaged_defrag_attr
.attr
,
593 &khugepaged_max_ptes_none_attr
.attr
,
594 &pages_to_scan_attr
.attr
,
595 &pages_collapsed_attr
.attr
,
596 &full_scans_attr
.attr
,
597 &scan_sleep_millisecs_attr
.attr
,
598 &alloc_sleep_millisecs_attr
.attr
,
602 static struct attribute_group khugepaged_attr_group
= {
603 .attrs
= khugepaged_attr
,
604 .name
= "khugepaged",
607 static int __init
hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
611 *hugepage_kobj
= kobject_create_and_add("transparent_hugepage", mm_kobj
);
612 if (unlikely(!*hugepage_kobj
)) {
613 pr_err("failed to create transparent hugepage kobject\n");
617 err
= sysfs_create_group(*hugepage_kobj
, &hugepage_attr_group
);
619 pr_err("failed to register transparent hugepage group\n");
623 err
= sysfs_create_group(*hugepage_kobj
, &khugepaged_attr_group
);
625 pr_err("failed to register transparent hugepage group\n");
626 goto remove_hp_group
;
632 sysfs_remove_group(*hugepage_kobj
, &hugepage_attr_group
);
634 kobject_put(*hugepage_kobj
);
638 static void __init
hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
640 sysfs_remove_group(hugepage_kobj
, &khugepaged_attr_group
);
641 sysfs_remove_group(hugepage_kobj
, &hugepage_attr_group
);
642 kobject_put(hugepage_kobj
);
645 static inline int hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
650 static inline void hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
653 #endif /* CONFIG_SYSFS */
655 static int __init
hugepage_init(void)
658 struct kobject
*hugepage_kobj
;
660 if (!has_transparent_hugepage()) {
661 transparent_hugepage_flags
= 0;
665 err
= hugepage_init_sysfs(&hugepage_kobj
);
669 err
= khugepaged_slab_init();
673 err
= register_shrinker(&huge_zero_page_shrinker
);
675 goto err_hzp_shrinker
;
676 err
= register_shrinker(&deferred_split_shrinker
);
678 goto err_split_shrinker
;
681 * By default disable transparent hugepages on smaller systems,
682 * where the extra memory used could hurt more than TLB overhead
683 * is likely to save. The admin can still enable it through /sys.
685 if (totalram_pages
< (512 << (20 - PAGE_SHIFT
))) {
686 transparent_hugepage_flags
= 0;
690 err
= start_stop_khugepaged();
696 unregister_shrinker(&deferred_split_shrinker
);
698 unregister_shrinker(&huge_zero_page_shrinker
);
700 khugepaged_slab_exit();
702 hugepage_exit_sysfs(hugepage_kobj
);
706 subsys_initcall(hugepage_init
);
708 static int __init
setup_transparent_hugepage(char *str
)
713 if (!strcmp(str
, "always")) {
714 set_bit(TRANSPARENT_HUGEPAGE_FLAG
,
715 &transparent_hugepage_flags
);
716 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
717 &transparent_hugepage_flags
);
719 } else if (!strcmp(str
, "madvise")) {
720 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
721 &transparent_hugepage_flags
);
722 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
723 &transparent_hugepage_flags
);
725 } else if (!strcmp(str
, "never")) {
726 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
727 &transparent_hugepage_flags
);
728 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
729 &transparent_hugepage_flags
);
734 pr_warn("transparent_hugepage= cannot parse, ignored\n");
737 __setup("transparent_hugepage=", setup_transparent_hugepage
);
739 pmd_t
maybe_pmd_mkwrite(pmd_t pmd
, struct vm_area_struct
*vma
)
741 if (likely(vma
->vm_flags
& VM_WRITE
))
742 pmd
= pmd_mkwrite(pmd
);
746 static inline pmd_t
mk_huge_pmd(struct page
*page
, pgprot_t prot
)
749 entry
= mk_pmd(page
, prot
);
750 entry
= pmd_mkhuge(entry
);
754 static inline struct list_head
*page_deferred_list(struct page
*page
)
757 * ->lru in the tail pages is occupied by compound_head.
758 * Let's use ->mapping + ->index in the second tail page as list_head.
760 return (struct list_head
*)&page
[2].mapping
;
763 void prep_transhuge_page(struct page
*page
)
766 * we use page->mapping and page->indexlru in second tail page
767 * as list_head: assuming THP order >= 2
769 BUILD_BUG_ON(HPAGE_PMD_ORDER
< 2);
771 INIT_LIST_HEAD(page_deferred_list(page
));
772 set_compound_page_dtor(page
, TRANSHUGE_PAGE_DTOR
);
775 static int __do_huge_pmd_anonymous_page(struct mm_struct
*mm
,
776 struct vm_area_struct
*vma
,
777 unsigned long address
, pmd_t
*pmd
,
778 struct page
*page
, gfp_t gfp
,
781 struct mem_cgroup
*memcg
;
784 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
786 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
788 if (mem_cgroup_try_charge(page
, mm
, gfp
, &memcg
, true)) {
790 count_vm_event(THP_FAULT_FALLBACK
);
791 return VM_FAULT_FALLBACK
;
794 pgtable
= pte_alloc_one(mm
, haddr
);
795 if (unlikely(!pgtable
)) {
796 mem_cgroup_cancel_charge(page
, memcg
, true);
801 clear_huge_page(page
, haddr
, HPAGE_PMD_NR
);
803 * The memory barrier inside __SetPageUptodate makes sure that
804 * clear_huge_page writes become visible before the set_pmd_at()
807 __SetPageUptodate(page
);
809 ptl
= pmd_lock(mm
, pmd
);
810 if (unlikely(!pmd_none(*pmd
))) {
812 mem_cgroup_cancel_charge(page
, memcg
, true);
814 pte_free(mm
, pgtable
);
818 /* Deliver the page fault to userland */
819 if (userfaultfd_missing(vma
)) {
823 mem_cgroup_cancel_charge(page
, memcg
, true);
825 pte_free(mm
, pgtable
);
826 ret
= handle_userfault(vma
, address
, flags
,
828 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
832 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
833 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
834 page_add_new_anon_rmap(page
, vma
, haddr
, true);
835 mem_cgroup_commit_charge(page
, memcg
, false, true);
836 lru_cache_add_active_or_unevictable(page
, vma
);
837 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
838 set_pmd_at(mm
, haddr
, pmd
, entry
);
839 add_mm_counter(mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
840 atomic_long_inc(&mm
->nr_ptes
);
842 count_vm_event(THP_FAULT_ALLOC
);
848 static inline gfp_t
alloc_hugepage_gfpmask(int defrag
, gfp_t extra_gfp
)
850 return (GFP_TRANSHUGE
& ~(defrag
? 0 : __GFP_RECLAIM
)) | extra_gfp
;
853 /* Caller must hold page table lock. */
854 static bool set_huge_zero_page(pgtable_t pgtable
, struct mm_struct
*mm
,
855 struct vm_area_struct
*vma
, unsigned long haddr
, pmd_t
*pmd
,
856 struct page
*zero_page
)
861 entry
= mk_pmd(zero_page
, vma
->vm_page_prot
);
862 entry
= pmd_mkhuge(entry
);
863 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
864 set_pmd_at(mm
, haddr
, pmd
, entry
);
865 atomic_long_inc(&mm
->nr_ptes
);
869 int do_huge_pmd_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
870 unsigned long address
, pmd_t
*pmd
,
875 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
877 if (haddr
< vma
->vm_start
|| haddr
+ HPAGE_PMD_SIZE
> vma
->vm_end
)
878 return VM_FAULT_FALLBACK
;
879 if (unlikely(anon_vma_prepare(vma
)))
881 if (unlikely(khugepaged_enter(vma
, vma
->vm_flags
)))
883 if (!(flags
& FAULT_FLAG_WRITE
) && !mm_forbids_zeropage(mm
) &&
884 transparent_hugepage_use_zero_page()) {
887 struct page
*zero_page
;
890 pgtable
= pte_alloc_one(mm
, haddr
);
891 if (unlikely(!pgtable
))
893 zero_page
= get_huge_zero_page();
894 if (unlikely(!zero_page
)) {
895 pte_free(mm
, pgtable
);
896 count_vm_event(THP_FAULT_FALLBACK
);
897 return VM_FAULT_FALLBACK
;
899 ptl
= pmd_lock(mm
, pmd
);
902 if (pmd_none(*pmd
)) {
903 if (userfaultfd_missing(vma
)) {
905 ret
= handle_userfault(vma
, address
, flags
,
907 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
909 set_huge_zero_page(pgtable
, mm
, vma
,
918 pte_free(mm
, pgtable
);
919 put_huge_zero_page();
923 gfp
= alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma
), 0);
924 page
= alloc_hugepage_vma(gfp
, vma
, haddr
, HPAGE_PMD_ORDER
);
925 if (unlikely(!page
)) {
926 count_vm_event(THP_FAULT_FALLBACK
);
927 return VM_FAULT_FALLBACK
;
929 prep_transhuge_page(page
);
930 return __do_huge_pmd_anonymous_page(mm
, vma
, address
, pmd
, page
, gfp
,
934 static void insert_pfn_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
935 pmd_t
*pmd
, pfn_t pfn
, pgprot_t prot
, bool write
)
937 struct mm_struct
*mm
= vma
->vm_mm
;
941 ptl
= pmd_lock(mm
, pmd
);
942 entry
= pmd_mkhuge(pfn_t_pmd(pfn
, prot
));
943 if (pfn_t_devmap(pfn
))
944 entry
= pmd_mkdevmap(entry
);
946 entry
= pmd_mkyoung(pmd_mkdirty(entry
));
947 entry
= maybe_pmd_mkwrite(entry
, vma
);
949 set_pmd_at(mm
, addr
, pmd
, entry
);
950 update_mmu_cache_pmd(vma
, addr
, pmd
);
954 int vmf_insert_pfn_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
955 pmd_t
*pmd
, pfn_t pfn
, bool write
)
957 pgprot_t pgprot
= vma
->vm_page_prot
;
959 * If we had pmd_special, we could avoid all these restrictions,
960 * but we need to be consistent with PTEs and architectures that
961 * can't support a 'special' bit.
963 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
964 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
965 (VM_PFNMAP
|VM_MIXEDMAP
));
966 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
967 BUG_ON(!pfn_t_devmap(pfn
));
969 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
970 return VM_FAULT_SIGBUS
;
971 if (track_pfn_insert(vma
, &pgprot
, pfn
))
972 return VM_FAULT_SIGBUS
;
973 insert_pfn_pmd(vma
, addr
, pmd
, pfn
, pgprot
, write
);
974 return VM_FAULT_NOPAGE
;
977 int copy_huge_pmd(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
978 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
979 struct vm_area_struct
*vma
)
981 spinlock_t
*dst_ptl
, *src_ptl
;
982 struct page
*src_page
;
988 pgtable
= pte_alloc_one(dst_mm
, addr
);
989 if (unlikely(!pgtable
))
992 dst_ptl
= pmd_lock(dst_mm
, dst_pmd
);
993 src_ptl
= pmd_lockptr(src_mm
, src_pmd
);
994 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
998 if (unlikely(!pmd_trans_huge(pmd
))) {
999 pte_free(dst_mm
, pgtable
);
1003 * When page table lock is held, the huge zero pmd should not be
1004 * under splitting since we don't split the page itself, only pmd to
1007 if (is_huge_zero_pmd(pmd
)) {
1008 struct page
*zero_page
;
1010 * get_huge_zero_page() will never allocate a new page here,
1011 * since we already have a zero page to copy. It just takes a
1014 zero_page
= get_huge_zero_page();
1015 set_huge_zero_page(pgtable
, dst_mm
, vma
, addr
, dst_pmd
,
1021 src_page
= pmd_page(pmd
);
1022 VM_BUG_ON_PAGE(!PageHead(src_page
), src_page
);
1024 page_dup_rmap(src_page
, true);
1025 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1027 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
1028 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
1029 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
1030 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
1031 atomic_long_inc(&dst_mm
->nr_ptes
);
1035 spin_unlock(src_ptl
);
1036 spin_unlock(dst_ptl
);
1041 void huge_pmd_set_accessed(struct mm_struct
*mm
,
1042 struct vm_area_struct
*vma
,
1043 unsigned long address
,
1044 pmd_t
*pmd
, pmd_t orig_pmd
,
1049 unsigned long haddr
;
1051 ptl
= pmd_lock(mm
, pmd
);
1052 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
1055 entry
= pmd_mkyoung(orig_pmd
);
1056 haddr
= address
& HPAGE_PMD_MASK
;
1057 if (pmdp_set_access_flags(vma
, haddr
, pmd
, entry
, dirty
))
1058 update_mmu_cache_pmd(vma
, address
, pmd
);
1064 static int do_huge_pmd_wp_page_fallback(struct mm_struct
*mm
,
1065 struct vm_area_struct
*vma
,
1066 unsigned long address
,
1067 pmd_t
*pmd
, pmd_t orig_pmd
,
1069 unsigned long haddr
)
1071 struct mem_cgroup
*memcg
;
1076 struct page
**pages
;
1077 unsigned long mmun_start
; /* For mmu_notifiers */
1078 unsigned long mmun_end
; /* For mmu_notifiers */
1080 pages
= kmalloc(sizeof(struct page
*) * HPAGE_PMD_NR
,
1082 if (unlikely(!pages
)) {
1083 ret
|= VM_FAULT_OOM
;
1087 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1088 pages
[i
] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE
|
1090 vma
, address
, page_to_nid(page
));
1091 if (unlikely(!pages
[i
] ||
1092 mem_cgroup_try_charge(pages
[i
], mm
, GFP_KERNEL
,
1097 memcg
= (void *)page_private(pages
[i
]);
1098 set_page_private(pages
[i
], 0);
1099 mem_cgroup_cancel_charge(pages
[i
], memcg
,
1104 ret
|= VM_FAULT_OOM
;
1107 set_page_private(pages
[i
], (unsigned long)memcg
);
1110 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1111 copy_user_highpage(pages
[i
], page
+ i
,
1112 haddr
+ PAGE_SIZE
* i
, vma
);
1113 __SetPageUptodate(pages
[i
]);
1118 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1119 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1121 ptl
= pmd_lock(mm
, pmd
);
1122 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
1123 goto out_free_pages
;
1124 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1126 pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
1127 /* leave pmd empty until pte is filled */
1129 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1130 pmd_populate(mm
, &_pmd
, pgtable
);
1132 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1134 entry
= mk_pte(pages
[i
], vma
->vm_page_prot
);
1135 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1136 memcg
= (void *)page_private(pages
[i
]);
1137 set_page_private(pages
[i
], 0);
1138 page_add_new_anon_rmap(pages
[i
], vma
, haddr
, false);
1139 mem_cgroup_commit_charge(pages
[i
], memcg
, false, false);
1140 lru_cache_add_active_or_unevictable(pages
[i
], vma
);
1141 pte
= pte_offset_map(&_pmd
, haddr
);
1142 VM_BUG_ON(!pte_none(*pte
));
1143 set_pte_at(mm
, haddr
, pte
, entry
);
1148 smp_wmb(); /* make pte visible before pmd */
1149 pmd_populate(mm
, pmd
, pgtable
);
1150 page_remove_rmap(page
, true);
1153 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1155 ret
|= VM_FAULT_WRITE
;
1163 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1164 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1165 memcg
= (void *)page_private(pages
[i
]);
1166 set_page_private(pages
[i
], 0);
1167 mem_cgroup_cancel_charge(pages
[i
], memcg
, false);
1174 int do_huge_pmd_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1175 unsigned long address
, pmd_t
*pmd
, pmd_t orig_pmd
)
1179 struct page
*page
= NULL
, *new_page
;
1180 struct mem_cgroup
*memcg
;
1181 unsigned long haddr
;
1182 unsigned long mmun_start
; /* For mmu_notifiers */
1183 unsigned long mmun_end
; /* For mmu_notifiers */
1184 gfp_t huge_gfp
; /* for allocation and charge */
1186 ptl
= pmd_lockptr(mm
, pmd
);
1187 VM_BUG_ON_VMA(!vma
->anon_vma
, vma
);
1188 haddr
= address
& HPAGE_PMD_MASK
;
1189 if (is_huge_zero_pmd(orig_pmd
))
1192 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
1195 page
= pmd_page(orig_pmd
);
1196 VM_BUG_ON_PAGE(!PageCompound(page
) || !PageHead(page
), page
);
1198 * We can only reuse the page if nobody else maps the huge page or it's
1199 * part. We can do it by checking page_mapcount() on each sub-page, but
1201 * The cheaper way is to check page_count() to be equal 1: every
1202 * mapcount takes page reference reference, so this way we can
1203 * guarantee, that the PMD is the only mapping.
1204 * This can give false negative if somebody pinned the page, but that's
1207 if (page_mapcount(page
) == 1 && page_count(page
) == 1) {
1209 entry
= pmd_mkyoung(orig_pmd
);
1210 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1211 if (pmdp_set_access_flags(vma
, haddr
, pmd
, entry
, 1))
1212 update_mmu_cache_pmd(vma
, address
, pmd
);
1213 ret
|= VM_FAULT_WRITE
;
1219 if (transparent_hugepage_enabled(vma
) &&
1220 !transparent_hugepage_debug_cow()) {
1221 huge_gfp
= alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma
), 0);
1222 new_page
= alloc_hugepage_vma(huge_gfp
, vma
, haddr
, HPAGE_PMD_ORDER
);
1226 if (likely(new_page
)) {
1227 prep_transhuge_page(new_page
);
1230 split_huge_pmd(vma
, pmd
, address
);
1231 ret
|= VM_FAULT_FALLBACK
;
1233 ret
= do_huge_pmd_wp_page_fallback(mm
, vma
, address
,
1234 pmd
, orig_pmd
, page
, haddr
);
1235 if (ret
& VM_FAULT_OOM
) {
1236 split_huge_pmd(vma
, pmd
, address
);
1237 ret
|= VM_FAULT_FALLBACK
;
1241 count_vm_event(THP_FAULT_FALLBACK
);
1245 if (unlikely(mem_cgroup_try_charge(new_page
, mm
, huge_gfp
, &memcg
,
1249 split_huge_pmd(vma
, pmd
, address
);
1252 split_huge_pmd(vma
, pmd
, address
);
1253 ret
|= VM_FAULT_FALLBACK
;
1254 count_vm_event(THP_FAULT_FALLBACK
);
1258 count_vm_event(THP_FAULT_ALLOC
);
1261 clear_huge_page(new_page
, haddr
, HPAGE_PMD_NR
);
1263 copy_user_huge_page(new_page
, page
, haddr
, vma
, HPAGE_PMD_NR
);
1264 __SetPageUptodate(new_page
);
1267 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1268 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1273 if (unlikely(!pmd_same(*pmd
, orig_pmd
))) {
1275 mem_cgroup_cancel_charge(new_page
, memcg
, true);
1280 entry
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
1281 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1282 pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
1283 page_add_new_anon_rmap(new_page
, vma
, haddr
, true);
1284 mem_cgroup_commit_charge(new_page
, memcg
, false, true);
1285 lru_cache_add_active_or_unevictable(new_page
, vma
);
1286 set_pmd_at(mm
, haddr
, pmd
, entry
);
1287 update_mmu_cache_pmd(vma
, address
, pmd
);
1289 add_mm_counter(mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1290 put_huge_zero_page();
1292 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1293 page_remove_rmap(page
, true);
1296 ret
|= VM_FAULT_WRITE
;
1300 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1308 struct page
*follow_trans_huge_pmd(struct vm_area_struct
*vma
,
1313 struct mm_struct
*mm
= vma
->vm_mm
;
1314 struct page
*page
= NULL
;
1316 assert_spin_locked(pmd_lockptr(mm
, pmd
));
1318 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
1321 /* Avoid dumping huge zero page */
1322 if ((flags
& FOLL_DUMP
) && is_huge_zero_pmd(*pmd
))
1323 return ERR_PTR(-EFAULT
);
1325 /* Full NUMA hinting faults to serialise migration in fault paths */
1326 if ((flags
& FOLL_NUMA
) && pmd_protnone(*pmd
))
1329 page
= pmd_page(*pmd
);
1330 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1331 if (flags
& FOLL_TOUCH
) {
1334 * We should set the dirty bit only for FOLL_WRITE but
1335 * for now the dirty bit in the pmd is meaningless.
1336 * And if the dirty bit will become meaningful and
1337 * we'll only set it with FOLL_WRITE, an atomic
1338 * set_bit will be required on the pmd to set the
1339 * young bit, instead of the current set_pmd_at.
1341 _pmd
= pmd_mkyoung(pmd_mkdirty(*pmd
));
1342 if (pmdp_set_access_flags(vma
, addr
& HPAGE_PMD_MASK
,
1344 update_mmu_cache_pmd(vma
, addr
, pmd
);
1346 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1348 * We don't mlock() pte-mapped THPs. This way we can avoid
1349 * leaking mlocked pages into non-VM_LOCKED VMAs.
1351 * In most cases the pmd is the only mapping of the page as we
1352 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1353 * writable private mappings in populate_vma_page_range().
1355 * The only scenario when we have the page shared here is if we
1356 * mlocking read-only mapping shared over fork(). We skip
1357 * mlocking such pages.
1359 if (compound_mapcount(page
) == 1 && !PageDoubleMap(page
) &&
1360 page
->mapping
&& trylock_page(page
)) {
1363 mlock_vma_page(page
);
1367 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
1368 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
1369 if (flags
& FOLL_GET
)
1376 /* NUMA hinting page fault entry point for trans huge pmds */
1377 int do_huge_pmd_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1378 unsigned long addr
, pmd_t pmd
, pmd_t
*pmdp
)
1381 struct anon_vma
*anon_vma
= NULL
;
1383 unsigned long haddr
= addr
& HPAGE_PMD_MASK
;
1384 int page_nid
= -1, this_nid
= numa_node_id();
1385 int target_nid
, last_cpupid
= -1;
1387 bool migrated
= false;
1391 /* A PROT_NONE fault should not end up here */
1392 BUG_ON(!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)));
1394 ptl
= pmd_lock(mm
, pmdp
);
1395 if (unlikely(!pmd_same(pmd
, *pmdp
)))
1399 * If there are potential migrations, wait for completion and retry
1400 * without disrupting NUMA hinting information. Do not relock and
1401 * check_same as the page may no longer be mapped.
1403 if (unlikely(pmd_trans_migrating(*pmdp
))) {
1404 page
= pmd_page(*pmdp
);
1406 wait_on_page_locked(page
);
1410 page
= pmd_page(pmd
);
1411 BUG_ON(is_huge_zero_page(page
));
1412 page_nid
= page_to_nid(page
);
1413 last_cpupid
= page_cpupid_last(page
);
1414 count_vm_numa_event(NUMA_HINT_FAULTS
);
1415 if (page_nid
== this_nid
) {
1416 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
1417 flags
|= TNF_FAULT_LOCAL
;
1420 /* See similar comment in do_numa_page for explanation */
1421 if (!(vma
->vm_flags
& VM_WRITE
))
1422 flags
|= TNF_NO_GROUP
;
1425 * Acquire the page lock to serialise THP migrations but avoid dropping
1426 * page_table_lock if at all possible
1428 page_locked
= trylock_page(page
);
1429 target_nid
= mpol_misplaced(page
, vma
, haddr
);
1430 if (target_nid
== -1) {
1431 /* If the page was locked, there are no parallel migrations */
1436 /* Migration could have started since the pmd_trans_migrating check */
1439 wait_on_page_locked(page
);
1445 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1446 * to serialises splits
1450 anon_vma
= page_lock_anon_vma_read(page
);
1452 /* Confirm the PMD did not change while page_table_lock was released */
1454 if (unlikely(!pmd_same(pmd
, *pmdp
))) {
1461 /* Bail if we fail to protect against THP splits for any reason */
1462 if (unlikely(!anon_vma
)) {
1469 * Migrate the THP to the requested node, returns with page unlocked
1470 * and access rights restored.
1473 migrated
= migrate_misplaced_transhuge_page(mm
, vma
,
1474 pmdp
, pmd
, addr
, page
, target_nid
);
1476 flags
|= TNF_MIGRATED
;
1477 page_nid
= target_nid
;
1479 flags
|= TNF_MIGRATE_FAIL
;
1483 BUG_ON(!PageLocked(page
));
1484 was_writable
= pmd_write(pmd
);
1485 pmd
= pmd_modify(pmd
, vma
->vm_page_prot
);
1486 pmd
= pmd_mkyoung(pmd
);
1488 pmd
= pmd_mkwrite(pmd
);
1489 set_pmd_at(mm
, haddr
, pmdp
, pmd
);
1490 update_mmu_cache_pmd(vma
, addr
, pmdp
);
1497 page_unlock_anon_vma_read(anon_vma
);
1500 task_numa_fault(last_cpupid
, page_nid
, HPAGE_PMD_NR
, flags
);
1505 int madvise_free_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1506 pmd_t
*pmd
, unsigned long addr
, unsigned long next
)
1512 struct mm_struct
*mm
= tlb
->mm
;
1515 if (!pmd_trans_huge_lock(pmd
, vma
, &ptl
))
1519 if (is_huge_zero_pmd(orig_pmd
)) {
1524 page
= pmd_page(orig_pmd
);
1526 * If other processes are mapping this page, we couldn't discard
1527 * the page unless they all do MADV_FREE so let's skip the page.
1529 if (page_mapcount(page
) != 1)
1532 if (!trylock_page(page
))
1536 * If user want to discard part-pages of THP, split it so MADV_FREE
1537 * will deactivate only them.
1539 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1542 if (split_huge_page(page
)) {
1553 if (PageDirty(page
))
1554 ClearPageDirty(page
);
1557 if (PageActive(page
))
1558 deactivate_page(page
);
1560 if (pmd_young(orig_pmd
) || pmd_dirty(orig_pmd
)) {
1561 orig_pmd
= pmdp_huge_get_and_clear_full(tlb
->mm
, addr
, pmd
,
1563 orig_pmd
= pmd_mkold(orig_pmd
);
1564 orig_pmd
= pmd_mkclean(orig_pmd
);
1566 set_pmd_at(mm
, addr
, pmd
, orig_pmd
);
1567 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1576 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1577 pmd_t
*pmd
, unsigned long addr
)
1582 if (!__pmd_trans_huge_lock(pmd
, vma
, &ptl
))
1585 * For architectures like ppc64 we look at deposited pgtable
1586 * when calling pmdp_huge_get_and_clear. So do the
1587 * pgtable_trans_huge_withdraw after finishing pmdp related
1590 orig_pmd
= pmdp_huge_get_and_clear_full(tlb
->mm
, addr
, pmd
,
1592 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1593 if (vma_is_dax(vma
)) {
1595 if (is_huge_zero_pmd(orig_pmd
))
1596 put_huge_zero_page();
1597 } else if (is_huge_zero_pmd(orig_pmd
)) {
1598 pte_free(tlb
->mm
, pgtable_trans_huge_withdraw(tlb
->mm
, pmd
));
1599 atomic_long_dec(&tlb
->mm
->nr_ptes
);
1601 put_huge_zero_page();
1603 struct page
*page
= pmd_page(orig_pmd
);
1604 page_remove_rmap(page
, true);
1605 VM_BUG_ON_PAGE(page_mapcount(page
) < 0, page
);
1606 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1607 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1608 pte_free(tlb
->mm
, pgtable_trans_huge_withdraw(tlb
->mm
, pmd
));
1609 atomic_long_dec(&tlb
->mm
->nr_ptes
);
1611 tlb_remove_page(tlb
, page
);
1616 bool move_huge_pmd(struct vm_area_struct
*vma
, struct vm_area_struct
*new_vma
,
1617 unsigned long old_addr
,
1618 unsigned long new_addr
, unsigned long old_end
,
1619 pmd_t
*old_pmd
, pmd_t
*new_pmd
)
1621 spinlock_t
*old_ptl
, *new_ptl
;
1624 struct mm_struct
*mm
= vma
->vm_mm
;
1626 if ((old_addr
& ~HPAGE_PMD_MASK
) ||
1627 (new_addr
& ~HPAGE_PMD_MASK
) ||
1628 old_end
- old_addr
< HPAGE_PMD_SIZE
||
1629 (new_vma
->vm_flags
& VM_NOHUGEPAGE
))
1633 * The destination pmd shouldn't be established, free_pgtables()
1634 * should have release it.
1636 if (WARN_ON(!pmd_none(*new_pmd
))) {
1637 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1642 * We don't have to worry about the ordering of src and dst
1643 * ptlocks because exclusive mmap_sem prevents deadlock.
1645 if (__pmd_trans_huge_lock(old_pmd
, vma
, &old_ptl
)) {
1646 new_ptl
= pmd_lockptr(mm
, new_pmd
);
1647 if (new_ptl
!= old_ptl
)
1648 spin_lock_nested(new_ptl
, SINGLE_DEPTH_NESTING
);
1649 pmd
= pmdp_huge_get_and_clear(mm
, old_addr
, old_pmd
);
1650 VM_BUG_ON(!pmd_none(*new_pmd
));
1652 if (pmd_move_must_withdraw(new_ptl
, old_ptl
)) {
1654 pgtable
= pgtable_trans_huge_withdraw(mm
, old_pmd
);
1655 pgtable_trans_huge_deposit(mm
, new_pmd
, pgtable
);
1657 set_pmd_at(mm
, new_addr
, new_pmd
, pmd_mksoft_dirty(pmd
));
1658 if (new_ptl
!= old_ptl
)
1659 spin_unlock(new_ptl
);
1660 spin_unlock(old_ptl
);
1668 * - 0 if PMD could not be locked
1669 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1670 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1672 int change_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1673 unsigned long addr
, pgprot_t newprot
, int prot_numa
)
1675 struct mm_struct
*mm
= vma
->vm_mm
;
1679 if (__pmd_trans_huge_lock(pmd
, vma
, &ptl
)) {
1681 bool preserve_write
= prot_numa
&& pmd_write(*pmd
);
1685 * Avoid trapping faults against the zero page. The read-only
1686 * data is likely to be read-cached on the local CPU and
1687 * local/remote hits to the zero page are not interesting.
1689 if (prot_numa
&& is_huge_zero_pmd(*pmd
)) {
1694 if (!prot_numa
|| !pmd_protnone(*pmd
)) {
1695 entry
= pmdp_huge_get_and_clear_notify(mm
, addr
, pmd
);
1696 entry
= pmd_modify(entry
, newprot
);
1698 entry
= pmd_mkwrite(entry
);
1700 set_pmd_at(mm
, addr
, pmd
, entry
);
1701 BUG_ON(!preserve_write
&& pmd_write(entry
));
1710 * Returns true if a given pmd maps a thp, false otherwise.
1712 * Note that if it returns true, this routine returns without unlocking page
1713 * table lock. So callers must unlock it.
1715 bool __pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
,
1718 *ptl
= pmd_lock(vma
->vm_mm
, pmd
);
1719 if (likely(pmd_trans_huge(*pmd
)))
1725 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1727 int hugepage_madvise(struct vm_area_struct
*vma
,
1728 unsigned long *vm_flags
, int advice
)
1734 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1735 * can't handle this properly after s390_enable_sie, so we simply
1736 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1738 if (mm_has_pgste(vma
->vm_mm
))
1742 * Be somewhat over-protective like KSM for now!
1744 if (*vm_flags
& VM_NO_THP
)
1746 *vm_flags
&= ~VM_NOHUGEPAGE
;
1747 *vm_flags
|= VM_HUGEPAGE
;
1749 * If the vma become good for khugepaged to scan,
1750 * register it here without waiting a page fault that
1751 * may not happen any time soon.
1753 if (unlikely(khugepaged_enter_vma_merge(vma
, *vm_flags
)))
1756 case MADV_NOHUGEPAGE
:
1758 * Be somewhat over-protective like KSM for now!
1760 if (*vm_flags
& VM_NO_THP
)
1762 *vm_flags
&= ~VM_HUGEPAGE
;
1763 *vm_flags
|= VM_NOHUGEPAGE
;
1765 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1766 * this vma even if we leave the mm registered in khugepaged if
1767 * it got registered before VM_NOHUGEPAGE was set.
1775 static int __init
khugepaged_slab_init(void)
1777 mm_slot_cache
= kmem_cache_create("khugepaged_mm_slot",
1778 sizeof(struct mm_slot
),
1779 __alignof__(struct mm_slot
), 0, NULL
);
1786 static void __init
khugepaged_slab_exit(void)
1788 kmem_cache_destroy(mm_slot_cache
);
1791 static inline struct mm_slot
*alloc_mm_slot(void)
1793 if (!mm_slot_cache
) /* initialization failed */
1795 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
1798 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
1800 kmem_cache_free(mm_slot_cache
, mm_slot
);
1803 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
1805 struct mm_slot
*mm_slot
;
1807 hash_for_each_possible(mm_slots_hash
, mm_slot
, hash
, (unsigned long)mm
)
1808 if (mm
== mm_slot
->mm
)
1814 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
1815 struct mm_slot
*mm_slot
)
1818 hash_add(mm_slots_hash
, &mm_slot
->hash
, (long)mm
);
1821 static inline int khugepaged_test_exit(struct mm_struct
*mm
)
1823 return atomic_read(&mm
->mm_users
) == 0;
1826 int __khugepaged_enter(struct mm_struct
*mm
)
1828 struct mm_slot
*mm_slot
;
1831 mm_slot
= alloc_mm_slot();
1835 /* __khugepaged_exit() must not run from under us */
1836 VM_BUG_ON_MM(khugepaged_test_exit(mm
), mm
);
1837 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE
, &mm
->flags
))) {
1838 free_mm_slot(mm_slot
);
1842 spin_lock(&khugepaged_mm_lock
);
1843 insert_to_mm_slots_hash(mm
, mm_slot
);
1845 * Insert just behind the scanning cursor, to let the area settle
1848 wakeup
= list_empty(&khugepaged_scan
.mm_head
);
1849 list_add_tail(&mm_slot
->mm_node
, &khugepaged_scan
.mm_head
);
1850 spin_unlock(&khugepaged_mm_lock
);
1852 atomic_inc(&mm
->mm_count
);
1854 wake_up_interruptible(&khugepaged_wait
);
1859 int khugepaged_enter_vma_merge(struct vm_area_struct
*vma
,
1860 unsigned long vm_flags
)
1862 unsigned long hstart
, hend
;
1865 * Not yet faulted in so we will register later in the
1866 * page fault if needed.
1870 /* khugepaged not yet working on file or special mappings */
1872 VM_BUG_ON_VMA(vm_flags
& VM_NO_THP
, vma
);
1873 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
1874 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
1876 return khugepaged_enter(vma
, vm_flags
);
1880 void __khugepaged_exit(struct mm_struct
*mm
)
1882 struct mm_slot
*mm_slot
;
1885 spin_lock(&khugepaged_mm_lock
);
1886 mm_slot
= get_mm_slot(mm
);
1887 if (mm_slot
&& khugepaged_scan
.mm_slot
!= mm_slot
) {
1888 hash_del(&mm_slot
->hash
);
1889 list_del(&mm_slot
->mm_node
);
1892 spin_unlock(&khugepaged_mm_lock
);
1895 clear_bit(MMF_VM_HUGEPAGE
, &mm
->flags
);
1896 free_mm_slot(mm_slot
);
1898 } else if (mm_slot
) {
1900 * This is required to serialize against
1901 * khugepaged_test_exit() (which is guaranteed to run
1902 * under mmap sem read mode). Stop here (after we
1903 * return all pagetables will be destroyed) until
1904 * khugepaged has finished working on the pagetables
1905 * under the mmap_sem.
1907 down_write(&mm
->mmap_sem
);
1908 up_write(&mm
->mmap_sem
);
1912 static void release_pte_page(struct page
*page
)
1914 /* 0 stands for page_is_file_cache(page) == false */
1915 dec_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
1917 putback_lru_page(page
);
1920 static void release_pte_pages(pte_t
*pte
, pte_t
*_pte
)
1922 while (--_pte
>= pte
) {
1923 pte_t pteval
= *_pte
;
1924 if (!pte_none(pteval
) && !is_zero_pfn(pte_pfn(pteval
)))
1925 release_pte_page(pte_page(pteval
));
1929 static int __collapse_huge_page_isolate(struct vm_area_struct
*vma
,
1930 unsigned long address
,
1933 struct page
*page
= NULL
;
1935 int none_or_zero
= 0, result
= 0;
1936 bool referenced
= false, writable
= false;
1938 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
1939 _pte
++, address
+= PAGE_SIZE
) {
1940 pte_t pteval
= *_pte
;
1941 if (pte_none(pteval
) || (pte_present(pteval
) &&
1942 is_zero_pfn(pte_pfn(pteval
)))) {
1943 if (!userfaultfd_armed(vma
) &&
1944 ++none_or_zero
<= khugepaged_max_ptes_none
) {
1947 result
= SCAN_EXCEED_NONE_PTE
;
1951 if (!pte_present(pteval
)) {
1952 result
= SCAN_PTE_NON_PRESENT
;
1955 page
= vm_normal_page(vma
, address
, pteval
);
1956 if (unlikely(!page
)) {
1957 result
= SCAN_PAGE_NULL
;
1961 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1962 VM_BUG_ON_PAGE(!PageAnon(page
), page
);
1963 VM_BUG_ON_PAGE(!PageSwapBacked(page
), page
);
1966 * We can do it before isolate_lru_page because the
1967 * page can't be freed from under us. NOTE: PG_lock
1968 * is needed to serialize against split_huge_page
1969 * when invoked from the VM.
1971 if (!trylock_page(page
)) {
1972 result
= SCAN_PAGE_LOCK
;
1977 * cannot use mapcount: can't collapse if there's a gup pin.
1978 * The page must only be referenced by the scanned process
1979 * and page swap cache.
1981 if (page_count(page
) != 1 + !!PageSwapCache(page
)) {
1983 result
= SCAN_PAGE_COUNT
;
1986 if (pte_write(pteval
)) {
1989 if (PageSwapCache(page
) && !reuse_swap_page(page
)) {
1991 result
= SCAN_SWAP_CACHE_PAGE
;
1995 * Page is not in the swap cache. It can be collapsed
2001 * Isolate the page to avoid collapsing an hugepage
2002 * currently in use by the VM.
2004 if (isolate_lru_page(page
)) {
2006 result
= SCAN_DEL_PAGE_LRU
;
2009 /* 0 stands for page_is_file_cache(page) == false */
2010 inc_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
2011 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2012 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2014 /* If there is no mapped pte young don't collapse the page */
2015 if (pte_young(pteval
) ||
2016 page_is_young(page
) || PageReferenced(page
) ||
2017 mmu_notifier_test_young(vma
->vm_mm
, address
))
2020 if (likely(writable
)) {
2021 if (likely(referenced
)) {
2022 result
= SCAN_SUCCEED
;
2023 trace_mm_collapse_huge_page_isolate(page_to_pfn(page
), none_or_zero
,
2024 referenced
, writable
, result
);
2028 result
= SCAN_PAGE_RO
;
2032 release_pte_pages(pte
, _pte
);
2033 trace_mm_collapse_huge_page_isolate(page_to_pfn(page
), none_or_zero
,
2034 referenced
, writable
, result
);
2038 static void __collapse_huge_page_copy(pte_t
*pte
, struct page
*page
,
2039 struct vm_area_struct
*vma
,
2040 unsigned long address
,
2044 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
; _pte
++) {
2045 pte_t pteval
= *_pte
;
2046 struct page
*src_page
;
2048 if (pte_none(pteval
) || is_zero_pfn(pte_pfn(pteval
))) {
2049 clear_user_highpage(page
, address
);
2050 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, 1);
2051 if (is_zero_pfn(pte_pfn(pteval
))) {
2053 * ptl mostly unnecessary.
2057 * paravirt calls inside pte_clear here are
2060 pte_clear(vma
->vm_mm
, address
, _pte
);
2064 src_page
= pte_page(pteval
);
2065 copy_user_highpage(page
, src_page
, address
, vma
);
2066 VM_BUG_ON_PAGE(page_mapcount(src_page
) != 1, src_page
);
2067 release_pte_page(src_page
);
2069 * ptl mostly unnecessary, but preempt has to
2070 * be disabled to update the per-cpu stats
2071 * inside page_remove_rmap().
2075 * paravirt calls inside pte_clear here are
2078 pte_clear(vma
->vm_mm
, address
, _pte
);
2079 page_remove_rmap(src_page
, false);
2081 free_page_and_swap_cache(src_page
);
2084 address
+= PAGE_SIZE
;
2089 static void khugepaged_alloc_sleep(void)
2093 add_wait_queue(&khugepaged_wait
, &wait
);
2094 freezable_schedule_timeout_interruptible(
2095 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs
));
2096 remove_wait_queue(&khugepaged_wait
, &wait
);
2099 static int khugepaged_node_load
[MAX_NUMNODES
];
2101 static bool khugepaged_scan_abort(int nid
)
2106 * If zone_reclaim_mode is disabled, then no extra effort is made to
2107 * allocate memory locally.
2109 if (!zone_reclaim_mode
)
2112 /* If there is a count for this node already, it must be acceptable */
2113 if (khugepaged_node_load
[nid
])
2116 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
2117 if (!khugepaged_node_load
[i
])
2119 if (node_distance(nid
, i
) > RECLAIM_DISTANCE
)
2126 static int khugepaged_find_target_node(void)
2128 static int last_khugepaged_target_node
= NUMA_NO_NODE
;
2129 int nid
, target_node
= 0, max_value
= 0;
2131 /* find first node with max normal pages hit */
2132 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
2133 if (khugepaged_node_load
[nid
] > max_value
) {
2134 max_value
= khugepaged_node_load
[nid
];
2138 /* do some balance if several nodes have the same hit record */
2139 if (target_node
<= last_khugepaged_target_node
)
2140 for (nid
= last_khugepaged_target_node
+ 1; nid
< MAX_NUMNODES
;
2142 if (max_value
== khugepaged_node_load
[nid
]) {
2147 last_khugepaged_target_node
= target_node
;
2151 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
2153 if (IS_ERR(*hpage
)) {
2159 khugepaged_alloc_sleep();
2160 } else if (*hpage
) {
2168 static struct page
*
2169 khugepaged_alloc_page(struct page
**hpage
, gfp_t gfp
, struct mm_struct
*mm
,
2170 unsigned long address
, int node
)
2172 VM_BUG_ON_PAGE(*hpage
, *hpage
);
2175 * Before allocating the hugepage, release the mmap_sem read lock.
2176 * The allocation can take potentially a long time if it involves
2177 * sync compaction, and we do not need to hold the mmap_sem during
2178 * that. We will recheck the vma after taking it again in write mode.
2180 up_read(&mm
->mmap_sem
);
2182 *hpage
= __alloc_pages_node(node
, gfp
, HPAGE_PMD_ORDER
);
2183 if (unlikely(!*hpage
)) {
2184 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
2185 *hpage
= ERR_PTR(-ENOMEM
);
2189 prep_transhuge_page(*hpage
);
2190 count_vm_event(THP_COLLAPSE_ALLOC
);
2194 static int khugepaged_find_target_node(void)
2199 static inline struct page
*alloc_hugepage(int defrag
)
2203 page
= alloc_pages(alloc_hugepage_gfpmask(defrag
, 0), HPAGE_PMD_ORDER
);
2205 prep_transhuge_page(page
);
2209 static struct page
*khugepaged_alloc_hugepage(bool *wait
)
2214 hpage
= alloc_hugepage(khugepaged_defrag());
2216 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
2221 khugepaged_alloc_sleep();
2223 count_vm_event(THP_COLLAPSE_ALLOC
);
2224 } while (unlikely(!hpage
) && likely(khugepaged_enabled()));
2229 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
2232 *hpage
= khugepaged_alloc_hugepage(wait
);
2234 if (unlikely(!*hpage
))
2240 static struct page
*
2241 khugepaged_alloc_page(struct page
**hpage
, gfp_t gfp
, struct mm_struct
*mm
,
2242 unsigned long address
, int node
)
2244 up_read(&mm
->mmap_sem
);
2251 static bool hugepage_vma_check(struct vm_area_struct
*vma
)
2253 if ((!(vma
->vm_flags
& VM_HUGEPAGE
) && !khugepaged_always()) ||
2254 (vma
->vm_flags
& VM_NOHUGEPAGE
))
2256 if (!vma
->anon_vma
|| vma
->vm_ops
)
2258 if (is_vma_temporary_stack(vma
))
2260 VM_BUG_ON_VMA(vma
->vm_flags
& VM_NO_THP
, vma
);
2264 static void collapse_huge_page(struct mm_struct
*mm
,
2265 unsigned long address
,
2266 struct page
**hpage
,
2267 struct vm_area_struct
*vma
,
2273 struct page
*new_page
;
2274 spinlock_t
*pmd_ptl
, *pte_ptl
;
2275 int isolated
, result
= 0;
2276 unsigned long hstart
, hend
;
2277 struct mem_cgroup
*memcg
;
2278 unsigned long mmun_start
; /* For mmu_notifiers */
2279 unsigned long mmun_end
; /* For mmu_notifiers */
2282 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2284 /* Only allocate from the target node */
2285 gfp
= alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE
) |
2288 /* release the mmap_sem read lock. */
2289 new_page
= khugepaged_alloc_page(hpage
, gfp
, mm
, address
, node
);
2291 result
= SCAN_ALLOC_HUGE_PAGE_FAIL
;
2295 if (unlikely(mem_cgroup_try_charge(new_page
, mm
, gfp
, &memcg
, true))) {
2296 result
= SCAN_CGROUP_CHARGE_FAIL
;
2301 * Prevent all access to pagetables with the exception of
2302 * gup_fast later hanlded by the ptep_clear_flush and the VM
2303 * handled by the anon_vma lock + PG_lock.
2305 down_write(&mm
->mmap_sem
);
2306 if (unlikely(khugepaged_test_exit(mm
))) {
2307 result
= SCAN_ANY_PROCESS
;
2311 vma
= find_vma(mm
, address
);
2313 result
= SCAN_VMA_NULL
;
2316 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2317 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2318 if (address
< hstart
|| address
+ HPAGE_PMD_SIZE
> hend
) {
2319 result
= SCAN_ADDRESS_RANGE
;
2322 if (!hugepage_vma_check(vma
)) {
2323 result
= SCAN_VMA_CHECK
;
2326 pmd
= mm_find_pmd(mm
, address
);
2328 result
= SCAN_PMD_NULL
;
2332 anon_vma_lock_write(vma
->anon_vma
);
2334 pte
= pte_offset_map(pmd
, address
);
2335 pte_ptl
= pte_lockptr(mm
, pmd
);
2337 mmun_start
= address
;
2338 mmun_end
= address
+ HPAGE_PMD_SIZE
;
2339 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2340 pmd_ptl
= pmd_lock(mm
, pmd
); /* probably unnecessary */
2342 * After this gup_fast can't run anymore. This also removes
2343 * any huge TLB entry from the CPU so we won't allow
2344 * huge and small TLB entries for the same virtual address
2345 * to avoid the risk of CPU bugs in that area.
2347 _pmd
= pmdp_collapse_flush(vma
, address
, pmd
);
2348 spin_unlock(pmd_ptl
);
2349 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2352 isolated
= __collapse_huge_page_isolate(vma
, address
, pte
);
2353 spin_unlock(pte_ptl
);
2355 if (unlikely(!isolated
)) {
2358 BUG_ON(!pmd_none(*pmd
));
2360 * We can only use set_pmd_at when establishing
2361 * hugepmds and never for establishing regular pmds that
2362 * points to regular pagetables. Use pmd_populate for that
2364 pmd_populate(mm
, pmd
, pmd_pgtable(_pmd
));
2365 spin_unlock(pmd_ptl
);
2366 anon_vma_unlock_write(vma
->anon_vma
);
2372 * All pages are isolated and locked so anon_vma rmap
2373 * can't run anymore.
2375 anon_vma_unlock_write(vma
->anon_vma
);
2377 __collapse_huge_page_copy(pte
, new_page
, vma
, address
, pte_ptl
);
2379 __SetPageUptodate(new_page
);
2380 pgtable
= pmd_pgtable(_pmd
);
2382 _pmd
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
2383 _pmd
= maybe_pmd_mkwrite(pmd_mkdirty(_pmd
), vma
);
2386 * spin_lock() below is not the equivalent of smp_wmb(), so
2387 * this is needed to avoid the copy_huge_page writes to become
2388 * visible after the set_pmd_at() write.
2393 BUG_ON(!pmd_none(*pmd
));
2394 page_add_new_anon_rmap(new_page
, vma
, address
, true);
2395 mem_cgroup_commit_charge(new_page
, memcg
, false, true);
2396 lru_cache_add_active_or_unevictable(new_page
, vma
);
2397 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
2398 set_pmd_at(mm
, address
, pmd
, _pmd
);
2399 update_mmu_cache_pmd(vma
, address
, pmd
);
2400 spin_unlock(pmd_ptl
);
2404 khugepaged_pages_collapsed
++;
2405 result
= SCAN_SUCCEED
;
2407 up_write(&mm
->mmap_sem
);
2408 trace_mm_collapse_huge_page(mm
, isolated
, result
);
2412 trace_mm_collapse_huge_page(mm
, isolated
, result
);
2415 mem_cgroup_cancel_charge(new_page
, memcg
, true);
2419 static int khugepaged_scan_pmd(struct mm_struct
*mm
,
2420 struct vm_area_struct
*vma
,
2421 unsigned long address
,
2422 struct page
**hpage
)
2426 int ret
= 0, none_or_zero
= 0, result
= 0;
2427 struct page
*page
= NULL
;
2428 unsigned long _address
;
2430 int node
= NUMA_NO_NODE
;
2431 bool writable
= false, referenced
= false;
2433 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2435 pmd
= mm_find_pmd(mm
, address
);
2437 result
= SCAN_PMD_NULL
;
2441 memset(khugepaged_node_load
, 0, sizeof(khugepaged_node_load
));
2442 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2443 for (_address
= address
, _pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
2444 _pte
++, _address
+= PAGE_SIZE
) {
2445 pte_t pteval
= *_pte
;
2446 if (pte_none(pteval
) || is_zero_pfn(pte_pfn(pteval
))) {
2447 if (!userfaultfd_armed(vma
) &&
2448 ++none_or_zero
<= khugepaged_max_ptes_none
) {
2451 result
= SCAN_EXCEED_NONE_PTE
;
2455 if (!pte_present(pteval
)) {
2456 result
= SCAN_PTE_NON_PRESENT
;
2459 if (pte_write(pteval
))
2462 page
= vm_normal_page(vma
, _address
, pteval
);
2463 if (unlikely(!page
)) {
2464 result
= SCAN_PAGE_NULL
;
2468 /* TODO: teach khugepaged to collapse THP mapped with pte */
2469 if (PageCompound(page
)) {
2470 result
= SCAN_PAGE_COMPOUND
;
2475 * Record which node the original page is from and save this
2476 * information to khugepaged_node_load[].
2477 * Khupaged will allocate hugepage from the node has the max
2480 node
= page_to_nid(page
);
2481 if (khugepaged_scan_abort(node
)) {
2482 result
= SCAN_SCAN_ABORT
;
2485 khugepaged_node_load
[node
]++;
2486 if (!PageLRU(page
)) {
2487 result
= SCAN_SCAN_ABORT
;
2490 if (PageLocked(page
)) {
2491 result
= SCAN_PAGE_LOCK
;
2494 if (!PageAnon(page
)) {
2495 result
= SCAN_PAGE_ANON
;
2500 * cannot use mapcount: can't collapse if there's a gup pin.
2501 * The page must only be referenced by the scanned process
2502 * and page swap cache.
2504 if (page_count(page
) != 1 + !!PageSwapCache(page
)) {
2505 result
= SCAN_PAGE_COUNT
;
2508 if (pte_young(pteval
) ||
2509 page_is_young(page
) || PageReferenced(page
) ||
2510 mmu_notifier_test_young(vma
->vm_mm
, address
))
2515 result
= SCAN_SUCCEED
;
2518 result
= SCAN_NO_REFERENCED_PAGE
;
2521 result
= SCAN_PAGE_RO
;
2524 pte_unmap_unlock(pte
, ptl
);
2526 node
= khugepaged_find_target_node();
2527 /* collapse_huge_page will return with the mmap_sem released */
2528 collapse_huge_page(mm
, address
, hpage
, vma
, node
);
2531 trace_mm_khugepaged_scan_pmd(mm
, page_to_pfn(page
), writable
, referenced
,
2532 none_or_zero
, result
);
2536 static void collect_mm_slot(struct mm_slot
*mm_slot
)
2538 struct mm_struct
*mm
= mm_slot
->mm
;
2540 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2542 if (khugepaged_test_exit(mm
)) {
2544 hash_del(&mm_slot
->hash
);
2545 list_del(&mm_slot
->mm_node
);
2548 * Not strictly needed because the mm exited already.
2550 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2553 /* khugepaged_mm_lock actually not necessary for the below */
2554 free_mm_slot(mm_slot
);
2559 static unsigned int khugepaged_scan_mm_slot(unsigned int pages
,
2560 struct page
**hpage
)
2561 __releases(&khugepaged_mm_lock
)
2562 __acquires(&khugepaged_mm_lock
)
2564 struct mm_slot
*mm_slot
;
2565 struct mm_struct
*mm
;
2566 struct vm_area_struct
*vma
;
2570 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2572 if (khugepaged_scan
.mm_slot
)
2573 mm_slot
= khugepaged_scan
.mm_slot
;
2575 mm_slot
= list_entry(khugepaged_scan
.mm_head
.next
,
2576 struct mm_slot
, mm_node
);
2577 khugepaged_scan
.address
= 0;
2578 khugepaged_scan
.mm_slot
= mm_slot
;
2580 spin_unlock(&khugepaged_mm_lock
);
2583 down_read(&mm
->mmap_sem
);
2584 if (unlikely(khugepaged_test_exit(mm
)))
2587 vma
= find_vma(mm
, khugepaged_scan
.address
);
2590 for (; vma
; vma
= vma
->vm_next
) {
2591 unsigned long hstart
, hend
;
2594 if (unlikely(khugepaged_test_exit(mm
))) {
2598 if (!hugepage_vma_check(vma
)) {
2603 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2604 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2607 if (khugepaged_scan
.address
> hend
)
2609 if (khugepaged_scan
.address
< hstart
)
2610 khugepaged_scan
.address
= hstart
;
2611 VM_BUG_ON(khugepaged_scan
.address
& ~HPAGE_PMD_MASK
);
2613 while (khugepaged_scan
.address
< hend
) {
2616 if (unlikely(khugepaged_test_exit(mm
)))
2617 goto breakouterloop
;
2619 VM_BUG_ON(khugepaged_scan
.address
< hstart
||
2620 khugepaged_scan
.address
+ HPAGE_PMD_SIZE
>
2622 ret
= khugepaged_scan_pmd(mm
, vma
,
2623 khugepaged_scan
.address
,
2625 /* move to next address */
2626 khugepaged_scan
.address
+= HPAGE_PMD_SIZE
;
2627 progress
+= HPAGE_PMD_NR
;
2629 /* we released mmap_sem so break loop */
2630 goto breakouterloop_mmap_sem
;
2631 if (progress
>= pages
)
2632 goto breakouterloop
;
2636 up_read(&mm
->mmap_sem
); /* exit_mmap will destroy ptes after this */
2637 breakouterloop_mmap_sem
:
2639 spin_lock(&khugepaged_mm_lock
);
2640 VM_BUG_ON(khugepaged_scan
.mm_slot
!= mm_slot
);
2642 * Release the current mm_slot if this mm is about to die, or
2643 * if we scanned all vmas of this mm.
2645 if (khugepaged_test_exit(mm
) || !vma
) {
2647 * Make sure that if mm_users is reaching zero while
2648 * khugepaged runs here, khugepaged_exit will find
2649 * mm_slot not pointing to the exiting mm.
2651 if (mm_slot
->mm_node
.next
!= &khugepaged_scan
.mm_head
) {
2652 khugepaged_scan
.mm_slot
= list_entry(
2653 mm_slot
->mm_node
.next
,
2654 struct mm_slot
, mm_node
);
2655 khugepaged_scan
.address
= 0;
2657 khugepaged_scan
.mm_slot
= NULL
;
2658 khugepaged_full_scans
++;
2661 collect_mm_slot(mm_slot
);
2667 static int khugepaged_has_work(void)
2669 return !list_empty(&khugepaged_scan
.mm_head
) &&
2670 khugepaged_enabled();
2673 static int khugepaged_wait_event(void)
2675 return !list_empty(&khugepaged_scan
.mm_head
) ||
2676 kthread_should_stop();
2679 static void khugepaged_do_scan(void)
2681 struct page
*hpage
= NULL
;
2682 unsigned int progress
= 0, pass_through_head
= 0;
2683 unsigned int pages
= khugepaged_pages_to_scan
;
2686 barrier(); /* write khugepaged_pages_to_scan to local stack */
2688 while (progress
< pages
) {
2689 if (!khugepaged_prealloc_page(&hpage
, &wait
))
2694 if (unlikely(kthread_should_stop() || try_to_freeze()))
2697 spin_lock(&khugepaged_mm_lock
);
2698 if (!khugepaged_scan
.mm_slot
)
2699 pass_through_head
++;
2700 if (khugepaged_has_work() &&
2701 pass_through_head
< 2)
2702 progress
+= khugepaged_scan_mm_slot(pages
- progress
,
2706 spin_unlock(&khugepaged_mm_lock
);
2709 if (!IS_ERR_OR_NULL(hpage
))
2713 static void khugepaged_wait_work(void)
2715 if (khugepaged_has_work()) {
2716 if (!khugepaged_scan_sleep_millisecs
)
2719 wait_event_freezable_timeout(khugepaged_wait
,
2720 kthread_should_stop(),
2721 msecs_to_jiffies(khugepaged_scan_sleep_millisecs
));
2725 if (khugepaged_enabled())
2726 wait_event_freezable(khugepaged_wait
, khugepaged_wait_event());
2729 static int khugepaged(void *none
)
2731 struct mm_slot
*mm_slot
;
2734 set_user_nice(current
, MAX_NICE
);
2736 while (!kthread_should_stop()) {
2737 khugepaged_do_scan();
2738 khugepaged_wait_work();
2741 spin_lock(&khugepaged_mm_lock
);
2742 mm_slot
= khugepaged_scan
.mm_slot
;
2743 khugepaged_scan
.mm_slot
= NULL
;
2745 collect_mm_slot(mm_slot
);
2746 spin_unlock(&khugepaged_mm_lock
);
2750 static void __split_huge_zero_page_pmd(struct vm_area_struct
*vma
,
2751 unsigned long haddr
, pmd_t
*pmd
)
2753 struct mm_struct
*mm
= vma
->vm_mm
;
2758 /* leave pmd empty until pte is filled */
2759 pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
2761 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2762 pmd_populate(mm
, &_pmd
, pgtable
);
2764 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
2766 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
2767 entry
= pte_mkspecial(entry
);
2768 pte
= pte_offset_map(&_pmd
, haddr
);
2769 VM_BUG_ON(!pte_none(*pte
));
2770 set_pte_at(mm
, haddr
, pte
, entry
);
2773 smp_wmb(); /* make pte visible before pmd */
2774 pmd_populate(mm
, pmd
, pgtable
);
2775 put_huge_zero_page();
2778 static void __split_huge_pmd_locked(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2779 unsigned long haddr
, bool freeze
)
2781 struct mm_struct
*mm
= vma
->vm_mm
;
2785 bool young
, write
, dirty
;
2788 VM_BUG_ON(haddr
& ~HPAGE_PMD_MASK
);
2789 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
2790 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PMD_SIZE
, vma
);
2791 VM_BUG_ON(!pmd_trans_huge(*pmd
));
2793 count_vm_event(THP_SPLIT_PMD
);
2795 if (vma_is_dax(vma
)) {
2796 pmd_t _pmd
= pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
2797 if (is_huge_zero_pmd(_pmd
))
2798 put_huge_zero_page();
2800 } else if (is_huge_zero_pmd(*pmd
)) {
2801 return __split_huge_zero_page_pmd(vma
, haddr
, pmd
);
2804 page
= pmd_page(*pmd
);
2805 VM_BUG_ON_PAGE(!page_count(page
), page
);
2806 atomic_add(HPAGE_PMD_NR
- 1, &page
->_count
);
2807 write
= pmd_write(*pmd
);
2808 young
= pmd_young(*pmd
);
2809 dirty
= pmd_dirty(*pmd
);
2811 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2812 pmd_populate(mm
, &_pmd
, pgtable
);
2814 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
2817 * Note that NUMA hinting access restrictions are not
2818 * transferred to avoid any possibility of altering
2819 * permissions across VMAs.
2822 swp_entry_t swp_entry
;
2823 swp_entry
= make_migration_entry(page
+ i
, write
);
2824 entry
= swp_entry_to_pte(swp_entry
);
2826 entry
= mk_pte(page
+ i
, vma
->vm_page_prot
);
2827 entry
= maybe_mkwrite(entry
, vma
);
2829 entry
= pte_wrprotect(entry
);
2831 entry
= pte_mkold(entry
);
2834 SetPageDirty(page
+ i
);
2835 pte
= pte_offset_map(&_pmd
, haddr
);
2836 BUG_ON(!pte_none(*pte
));
2837 set_pte_at(mm
, haddr
, pte
, entry
);
2838 atomic_inc(&page
[i
]._mapcount
);
2843 * Set PG_double_map before dropping compound_mapcount to avoid
2844 * false-negative page_mapped().
2846 if (compound_mapcount(page
) > 1 && !TestSetPageDoubleMap(page
)) {
2847 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2848 atomic_inc(&page
[i
]._mapcount
);
2851 if (atomic_add_negative(-1, compound_mapcount_ptr(page
))) {
2852 /* Last compound_mapcount is gone. */
2853 __dec_zone_page_state(page
, NR_ANON_TRANSPARENT_HUGEPAGES
);
2854 if (TestClearPageDoubleMap(page
)) {
2855 /* No need in mapcount reference anymore */
2856 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2857 atomic_dec(&page
[i
]._mapcount
);
2861 smp_wmb(); /* make pte visible before pmd */
2863 * Up to this point the pmd is present and huge and userland has the
2864 * whole access to the hugepage during the split (which happens in
2865 * place). If we overwrite the pmd with the not-huge version pointing
2866 * to the pte here (which of course we could if all CPUs were bug
2867 * free), userland could trigger a small page size TLB miss on the
2868 * small sized TLB while the hugepage TLB entry is still established in
2869 * the huge TLB. Some CPU doesn't like that.
2870 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2871 * 383 on page 93. Intel should be safe but is also warns that it's
2872 * only safe if the permission and cache attributes of the two entries
2873 * loaded in the two TLB is identical (which should be the case here).
2874 * But it is generally safer to never allow small and huge TLB entries
2875 * for the same virtual address to be loaded simultaneously. So instead
2876 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2877 * current pmd notpresent (atomically because here the pmd_trans_huge
2878 * and pmd_trans_splitting must remain set at all times on the pmd
2879 * until the split is complete for this pmd), then we flush the SMP TLB
2880 * and finally we write the non-huge version of the pmd entry with
2883 pmdp_invalidate(vma
, haddr
, pmd
);
2884 pmd_populate(mm
, pmd
, pgtable
);
2887 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
2888 page_remove_rmap(page
+ i
, false);
2894 void __split_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2895 unsigned long address
)
2898 struct mm_struct
*mm
= vma
->vm_mm
;
2899 struct page
*page
= NULL
;
2900 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
2902 mmu_notifier_invalidate_range_start(mm
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
2903 ptl
= pmd_lock(mm
, pmd
);
2904 if (unlikely(!pmd_trans_huge(*pmd
)))
2906 page
= pmd_page(*pmd
);
2907 __split_huge_pmd_locked(vma
, pmd
, haddr
, false);
2908 if (PageMlocked(page
))
2914 mmu_notifier_invalidate_range_end(mm
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
2917 munlock_vma_page(page
);
2923 static void split_huge_pmd_address(struct vm_area_struct
*vma
,
2924 unsigned long address
)
2930 VM_BUG_ON(!(address
& ~HPAGE_PMD_MASK
));
2932 pgd
= pgd_offset(vma
->vm_mm
, address
);
2933 if (!pgd_present(*pgd
))
2936 pud
= pud_offset(pgd
, address
);
2937 if (!pud_present(*pud
))
2940 pmd
= pmd_offset(pud
, address
);
2941 if (!pmd_present(*pmd
) || !pmd_trans_huge(*pmd
))
2944 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2945 * materialize from under us.
2947 split_huge_pmd(vma
, pmd
, address
);
2950 void vma_adjust_trans_huge(struct vm_area_struct
*vma
,
2951 unsigned long start
,
2956 * If the new start address isn't hpage aligned and it could
2957 * previously contain an hugepage: check if we need to split
2960 if (start
& ~HPAGE_PMD_MASK
&&
2961 (start
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2962 (start
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2963 split_huge_pmd_address(vma
, start
);
2966 * If the new end address isn't hpage aligned and it could
2967 * previously contain an hugepage: check if we need to split
2970 if (end
& ~HPAGE_PMD_MASK
&&
2971 (end
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2972 (end
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2973 split_huge_pmd_address(vma
, end
);
2976 * If we're also updating the vma->vm_next->vm_start, if the new
2977 * vm_next->vm_start isn't page aligned and it could previously
2978 * contain an hugepage: check if we need to split an huge pmd.
2980 if (adjust_next
> 0) {
2981 struct vm_area_struct
*next
= vma
->vm_next
;
2982 unsigned long nstart
= next
->vm_start
;
2983 nstart
+= adjust_next
<< PAGE_SHIFT
;
2984 if (nstart
& ~HPAGE_PMD_MASK
&&
2985 (nstart
& HPAGE_PMD_MASK
) >= next
->vm_start
&&
2986 (nstart
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= next
->vm_end
)
2987 split_huge_pmd_address(next
, nstart
);
2991 static void freeze_page_vma(struct vm_area_struct
*vma
, struct page
*page
,
2992 unsigned long address
)
2994 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
3000 int i
, nr
= HPAGE_PMD_NR
;
3002 /* Skip pages which doesn't belong to the VMA */
3003 if (address
< vma
->vm_start
) {
3004 int off
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
3007 address
= vma
->vm_start
;
3010 pgd
= pgd_offset(vma
->vm_mm
, address
);
3011 if (!pgd_present(*pgd
))
3013 pud
= pud_offset(pgd
, address
);
3014 if (!pud_present(*pud
))
3016 pmd
= pmd_offset(pud
, address
);
3017 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
3018 if (!pmd_present(*pmd
)) {
3022 if (pmd_trans_huge(*pmd
)) {
3023 if (page
== pmd_page(*pmd
))
3024 __split_huge_pmd_locked(vma
, pmd
, haddr
, true);
3030 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, address
, &ptl
);
3031 for (i
= 0; i
< nr
; i
++, address
+= PAGE_SIZE
, page
++, pte
++) {
3032 pte_t entry
, swp_pte
;
3033 swp_entry_t swp_entry
;
3036 * We've just crossed page table boundary: need to map next one.
3037 * It can happen if THP was mremaped to non PMD-aligned address.
3039 if (unlikely(address
== haddr
+ HPAGE_PMD_SIZE
)) {
3040 pte_unmap_unlock(pte
- 1, ptl
);
3041 pmd
= mm_find_pmd(vma
->vm_mm
, address
);
3044 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
,
3048 if (!pte_present(*pte
))
3050 if (page_to_pfn(page
) != pte_pfn(*pte
))
3052 flush_cache_page(vma
, address
, page_to_pfn(page
));
3053 entry
= ptep_clear_flush(vma
, address
, pte
);
3054 if (pte_dirty(entry
))
3056 swp_entry
= make_migration_entry(page
, pte_write(entry
));
3057 swp_pte
= swp_entry_to_pte(swp_entry
);
3058 if (pte_soft_dirty(entry
))
3059 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
3060 set_pte_at(vma
->vm_mm
, address
, pte
, swp_pte
);
3061 page_remove_rmap(page
, false);
3064 pte_unmap_unlock(pte
- 1, ptl
);
3067 static void freeze_page(struct anon_vma
*anon_vma
, struct page
*page
)
3069 struct anon_vma_chain
*avc
;
3070 pgoff_t pgoff
= page_to_pgoff(page
);
3072 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3074 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
,
3075 pgoff
+ HPAGE_PMD_NR
- 1) {
3076 unsigned long address
= __vma_address(page
, avc
->vma
);
3078 mmu_notifier_invalidate_range_start(avc
->vma
->vm_mm
,
3079 address
, address
+ HPAGE_PMD_SIZE
);
3080 freeze_page_vma(avc
->vma
, page
, address
);
3081 mmu_notifier_invalidate_range_end(avc
->vma
->vm_mm
,
3082 address
, address
+ HPAGE_PMD_SIZE
);
3086 static void unfreeze_page_vma(struct vm_area_struct
*vma
, struct page
*page
,
3087 unsigned long address
)
3092 swp_entry_t swp_entry
;
3093 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
3094 int i
, nr
= HPAGE_PMD_NR
;
3096 /* Skip pages which doesn't belong to the VMA */
3097 if (address
< vma
->vm_start
) {
3098 int off
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
3101 address
= vma
->vm_start
;
3104 pmd
= mm_find_pmd(vma
->vm_mm
, address
);
3108 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, address
, &ptl
);
3109 for (i
= 0; i
< nr
; i
++, address
+= PAGE_SIZE
, page
++, pte
++) {
3111 * We've just crossed page table boundary: need to map next one.
3112 * It can happen if THP was mremaped to non-PMD aligned address.
3114 if (unlikely(address
== haddr
+ HPAGE_PMD_SIZE
)) {
3115 pte_unmap_unlock(pte
- 1, ptl
);
3116 pmd
= mm_find_pmd(vma
->vm_mm
, address
);
3119 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
,
3123 if (!is_swap_pte(*pte
))
3126 swp_entry
= pte_to_swp_entry(*pte
);
3127 if (!is_migration_entry(swp_entry
))
3129 if (migration_entry_to_page(swp_entry
) != page
)
3133 page_add_anon_rmap(page
, vma
, address
, false);
3135 entry
= pte_mkold(mk_pte(page
, vma
->vm_page_prot
));
3136 if (PageDirty(page
))
3137 entry
= pte_mkdirty(entry
);
3138 if (is_write_migration_entry(swp_entry
))
3139 entry
= maybe_mkwrite(entry
, vma
);
3141 flush_dcache_page(page
);
3142 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
3144 /* No need to invalidate - it was non-present before */
3145 update_mmu_cache(vma
, address
, pte
);
3147 pte_unmap_unlock(pte
- 1, ptl
);
3150 static void unfreeze_page(struct anon_vma
*anon_vma
, struct page
*page
)
3152 struct anon_vma_chain
*avc
;
3153 pgoff_t pgoff
= page_to_pgoff(page
);
3155 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
,
3156 pgoff
, pgoff
+ HPAGE_PMD_NR
- 1) {
3157 unsigned long address
= __vma_address(page
, avc
->vma
);
3159 mmu_notifier_invalidate_range_start(avc
->vma
->vm_mm
,
3160 address
, address
+ HPAGE_PMD_SIZE
);
3161 unfreeze_page_vma(avc
->vma
, page
, address
);
3162 mmu_notifier_invalidate_range_end(avc
->vma
->vm_mm
,
3163 address
, address
+ HPAGE_PMD_SIZE
);
3167 static int __split_huge_page_tail(struct page
*head
, int tail
,
3168 struct lruvec
*lruvec
, struct list_head
*list
)
3171 struct page
*page_tail
= head
+ tail
;
3173 mapcount
= atomic_read(&page_tail
->_mapcount
) + 1;
3174 VM_BUG_ON_PAGE(atomic_read(&page_tail
->_count
) != 0, page_tail
);
3177 * tail_page->_count is zero and not changing from under us. But
3178 * get_page_unless_zero() may be running from under us on the
3179 * tail_page. If we used atomic_set() below instead of atomic_add(), we
3180 * would then run atomic_set() concurrently with
3181 * get_page_unless_zero(), and atomic_set() is implemented in C not
3182 * using locked ops. spin_unlock on x86 sometime uses locked ops
3183 * because of PPro errata 66, 92, so unless somebody can guarantee
3184 * atomic_set() here would be safe on all archs (and not only on x86),
3185 * it's safer to use atomic_add().
3187 atomic_add(mapcount
+ 1, &page_tail
->_count
);
3190 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
3191 page_tail
->flags
|= (head
->flags
&
3192 ((1L << PG_referenced
) |
3193 (1L << PG_swapbacked
) |
3194 (1L << PG_mlocked
) |
3195 (1L << PG_uptodate
) |
3198 (1L << PG_unevictable
) |
3202 * After clearing PageTail the gup refcount can be released.
3203 * Page flags also must be visible before we make the page non-compound.
3207 clear_compound_head(page_tail
);
3209 if (page_is_young(head
))
3210 set_page_young(page_tail
);
3211 if (page_is_idle(head
))
3212 set_page_idle(page_tail
);
3214 /* ->mapping in first tail page is compound_mapcount */
3215 VM_BUG_ON_PAGE(tail
> 2 && page_tail
->mapping
!= TAIL_MAPPING
,
3217 page_tail
->mapping
= head
->mapping
;
3219 page_tail
->index
= head
->index
+ tail
;
3220 page_cpupid_xchg_last(page_tail
, page_cpupid_last(head
));
3221 lru_add_page_tail(head
, page_tail
, lruvec
, list
);
3226 static void __split_huge_page(struct page
*page
, struct list_head
*list
)
3228 struct page
*head
= compound_head(page
);
3229 struct zone
*zone
= page_zone(head
);
3230 struct lruvec
*lruvec
;
3231 int i
, tail_mapcount
;
3233 /* prevent PageLRU to go away from under us, and freeze lru stats */
3234 spin_lock_irq(&zone
->lru_lock
);
3235 lruvec
= mem_cgroup_page_lruvec(head
, zone
);
3237 /* complete memcg works before add pages to LRU */
3238 mem_cgroup_split_huge_fixup(head
);
3241 for (i
= HPAGE_PMD_NR
- 1; i
>= 1; i
--)
3242 tail_mapcount
+= __split_huge_page_tail(head
, i
, lruvec
, list
);
3243 atomic_sub(tail_mapcount
, &head
->_count
);
3245 ClearPageCompound(head
);
3246 spin_unlock_irq(&zone
->lru_lock
);
3248 unfreeze_page(page_anon_vma(head
), head
);
3250 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
3251 struct page
*subpage
= head
+ i
;
3252 if (subpage
== page
)
3254 unlock_page(subpage
);
3257 * Subpages may be freed if there wasn't any mapping
3258 * like if add_to_swap() is running on a lru page that
3259 * had its mapping zapped. And freeing these pages
3260 * requires taking the lru_lock so we do the put_page
3261 * of the tail pages after the split is complete.
3267 int total_mapcount(struct page
*page
)
3271 VM_BUG_ON_PAGE(PageTail(page
), page
);
3273 if (likely(!PageCompound(page
)))
3274 return atomic_read(&page
->_mapcount
) + 1;
3276 ret
= compound_mapcount(page
);
3279 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3280 ret
+= atomic_read(&page
[i
]._mapcount
) + 1;
3281 if (PageDoubleMap(page
))
3282 ret
-= HPAGE_PMD_NR
;
3287 * This function splits huge page into normal pages. @page can point to any
3288 * subpage of huge page to split. Split doesn't change the position of @page.
3290 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3291 * The huge page must be locked.
3293 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3295 * Both head page and tail pages will inherit mapping, flags, and so on from
3298 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3299 * they are not mapped.
3301 * Returns 0 if the hugepage is split successfully.
3302 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3305 int split_huge_page_to_list(struct page
*page
, struct list_head
*list
)
3307 struct page
*head
= compound_head(page
);
3308 struct anon_vma
*anon_vma
;
3309 int count
, mapcount
, ret
;
3312 VM_BUG_ON_PAGE(is_huge_zero_page(page
), page
);
3313 VM_BUG_ON_PAGE(!PageAnon(page
), page
);
3314 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
3315 VM_BUG_ON_PAGE(!PageSwapBacked(page
), page
);
3316 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
3319 * The caller does not necessarily hold an mmap_sem that would prevent
3320 * the anon_vma disappearing so we first we take a reference to it
3321 * and then lock the anon_vma for write. This is similar to
3322 * page_lock_anon_vma_read except the write lock is taken to serialise
3323 * against parallel split or collapse operations.
3325 anon_vma
= page_get_anon_vma(head
);
3330 anon_vma_lock_write(anon_vma
);
3333 * Racy check if we can split the page, before freeze_page() will
3336 if (total_mapcount(head
) != page_count(head
) - 1) {
3341 mlocked
= PageMlocked(page
);
3342 freeze_page(anon_vma
, head
);
3343 VM_BUG_ON_PAGE(compound_mapcount(head
), head
);
3345 /* Make sure the page is not on per-CPU pagevec as it takes pin */
3349 /* Prevent deferred_split_scan() touching ->_count */
3350 spin_lock(&split_queue_lock
);
3351 count
= page_count(head
);
3352 mapcount
= total_mapcount(head
);
3353 if (!mapcount
&& count
== 1) {
3354 if (!list_empty(page_deferred_list(head
))) {
3356 list_del(page_deferred_list(head
));
3358 spin_unlock(&split_queue_lock
);
3359 __split_huge_page(page
, list
);
3361 } else if (IS_ENABLED(CONFIG_DEBUG_VM
) && mapcount
) {
3362 spin_unlock(&split_queue_lock
);
3363 pr_alert("total_mapcount: %u, page_count(): %u\n",
3366 dump_page(head
, NULL
);
3367 dump_page(page
, "total_mapcount(head) > 0");
3370 spin_unlock(&split_queue_lock
);
3371 unfreeze_page(anon_vma
, head
);
3376 anon_vma_unlock_write(anon_vma
);
3377 put_anon_vma(anon_vma
);
3379 count_vm_event(!ret
? THP_SPLIT_PAGE
: THP_SPLIT_PAGE_FAILED
);
3383 void free_transhuge_page(struct page
*page
)
3385 unsigned long flags
;
3387 spin_lock_irqsave(&split_queue_lock
, flags
);
3388 if (!list_empty(page_deferred_list(page
))) {
3390 list_del(page_deferred_list(page
));
3392 spin_unlock_irqrestore(&split_queue_lock
, flags
);
3393 free_compound_page(page
);
3396 void deferred_split_huge_page(struct page
*page
)
3398 unsigned long flags
;
3400 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3402 spin_lock_irqsave(&split_queue_lock
, flags
);
3403 if (list_empty(page_deferred_list(page
))) {
3404 list_add_tail(page_deferred_list(page
), &split_queue
);
3407 spin_unlock_irqrestore(&split_queue_lock
, flags
);
3410 static unsigned long deferred_split_count(struct shrinker
*shrink
,
3411 struct shrink_control
*sc
)
3414 * Split a page from split_queue will free up at least one page,
3415 * at most HPAGE_PMD_NR - 1. We don't track exact number.
3416 * Let's use HPAGE_PMD_NR / 2 as ballpark.
3418 return ACCESS_ONCE(split_queue_len
) * HPAGE_PMD_NR
/ 2;
3421 static unsigned long deferred_split_scan(struct shrinker
*shrink
,
3422 struct shrink_control
*sc
)
3424 unsigned long flags
;
3425 LIST_HEAD(list
), *pos
, *next
;
3429 spin_lock_irqsave(&split_queue_lock
, flags
);
3430 list_splice_init(&split_queue
, &list
);
3432 /* Take pin on all head pages to avoid freeing them under us */
3433 list_for_each_safe(pos
, next
, &list
) {
3434 page
= list_entry((void *)pos
, struct page
, mapping
);
3435 page
= compound_head(page
);
3436 /* race with put_compound_page() */
3437 if (!get_page_unless_zero(page
)) {
3438 list_del_init(page_deferred_list(page
));
3442 spin_unlock_irqrestore(&split_queue_lock
, flags
);
3444 list_for_each_safe(pos
, next
, &list
) {
3445 page
= list_entry((void *)pos
, struct page
, mapping
);
3447 /* split_huge_page() removes page from list on success */
3448 if (!split_huge_page(page
))
3454 spin_lock_irqsave(&split_queue_lock
, flags
);
3455 list_splice_tail(&list
, &split_queue
);
3456 spin_unlock_irqrestore(&split_queue_lock
, flags
);
3458 return split
* HPAGE_PMD_NR
/ 2;
3461 static struct shrinker deferred_split_shrinker
= {
3462 .count_objects
= deferred_split_count
,
3463 .scan_objects
= deferred_split_scan
,
3464 .seeks
= DEFAULT_SEEKS
,
3467 #ifdef CONFIG_DEBUG_FS
3468 static int split_huge_pages_set(void *data
, u64 val
)
3472 unsigned long pfn
, max_zone_pfn
;
3473 unsigned long total
= 0, split
= 0;
3478 for_each_populated_zone(zone
) {
3479 max_zone_pfn
= zone_end_pfn(zone
);
3480 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++) {
3481 if (!pfn_valid(pfn
))
3484 page
= pfn_to_page(pfn
);
3485 if (!get_page_unless_zero(page
))
3488 if (zone
!= page_zone(page
))
3491 if (!PageHead(page
) || !PageAnon(page
) ||
3497 if (!split_huge_page(page
))
3505 pr_info("%lu of %lu THP split", split
, total
);
3509 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops
, NULL
, split_huge_pages_set
,
3512 static int __init
split_huge_pages_debugfs(void)
3516 ret
= debugfs_create_file("split_huge_pages", 0644, NULL
, NULL
,
3517 &split_huge_pages_fops
);
3519 pr_warn("Failed to create split_huge_pages in debugfs");
3522 late_initcall(split_huge_pages_debugfs
);