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
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/huge_memory.h>
67 * By default transparent hugepage support is disabled in order that avoid
68 * to risk increase the memory footprint of applications without a guaranteed
69 * benefit. When transparent hugepage support is enabled, is for all mappings,
70 * and khugepaged scans all mappings.
71 * Defrag is invoked by khugepaged hugepage allocations and by page faults
72 * for all hugepage allocations.
74 unsigned long transparent_hugepage_flags __read_mostly
=
75 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
76 (1<<TRANSPARENT_HUGEPAGE_FLAG
)|
78 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
79 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
)|
81 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
)|
82 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
)|
83 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
85 /* default scan 8*512 pte (or vmas) every 30 second */
86 static unsigned int khugepaged_pages_to_scan __read_mostly
= HPAGE_PMD_NR
*8;
87 static unsigned int khugepaged_pages_collapsed
;
88 static unsigned int khugepaged_full_scans
;
89 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly
= 10000;
90 /* during fragmentation poll the hugepage allocator once every minute */
91 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly
= 60000;
92 static struct task_struct
*khugepaged_thread __read_mostly
;
93 static DEFINE_MUTEX(khugepaged_mutex
);
94 static DEFINE_SPINLOCK(khugepaged_mm_lock
);
95 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait
);
97 * default collapse hugepages if there is at least one pte mapped like
98 * it would have happened if the vma was large enough during page
101 static unsigned int khugepaged_max_ptes_none __read_mostly
= HPAGE_PMD_NR
-1;
103 static int khugepaged(void *none
);
104 static int khugepaged_slab_init(void);
105 static void khugepaged_slab_exit(void);
107 #define MM_SLOTS_HASH_BITS 10
108 static __read_mostly
DEFINE_HASHTABLE(mm_slots_hash
, MM_SLOTS_HASH_BITS
);
110 static struct kmem_cache
*mm_slot_cache __read_mostly
;
113 * struct mm_slot - hash lookup from mm to mm_slot
114 * @hash: hash collision list
115 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
116 * @mm: the mm that this information is valid for
119 struct hlist_node hash
;
120 struct list_head mm_node
;
121 struct mm_struct
*mm
;
125 * struct khugepaged_scan - cursor for scanning
126 * @mm_head: the head of the mm list to scan
127 * @mm_slot: the current mm_slot we are scanning
128 * @address: the next address inside that to be scanned
130 * There is only the one khugepaged_scan instance of this cursor structure.
132 struct khugepaged_scan
{
133 struct list_head mm_head
;
134 struct mm_slot
*mm_slot
;
135 unsigned long address
;
137 static struct khugepaged_scan khugepaged_scan
= {
138 .mm_head
= LIST_HEAD_INIT(khugepaged_scan
.mm_head
),
141 static struct shrinker deferred_split_shrinker
;
143 static void set_recommended_min_free_kbytes(void)
147 unsigned long recommended_min
;
149 for_each_populated_zone(zone
)
152 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
153 recommended_min
= pageblock_nr_pages
* nr_zones
* 2;
156 * Make sure that on average at least two pageblocks are almost free
157 * of another type, one for a migratetype to fall back to and a
158 * second to avoid subsequent fallbacks of other types There are 3
159 * MIGRATE_TYPES we care about.
161 recommended_min
+= pageblock_nr_pages
* nr_zones
*
162 MIGRATE_PCPTYPES
* MIGRATE_PCPTYPES
;
164 /* don't ever allow to reserve more than 5% of the lowmem */
165 recommended_min
= min(recommended_min
,
166 (unsigned long) nr_free_buffer_pages() / 20);
167 recommended_min
<<= (PAGE_SHIFT
-10);
169 if (recommended_min
> min_free_kbytes
) {
170 if (user_min_free_kbytes
>= 0)
171 pr_info("raising min_free_kbytes from %d to %lu "
172 "to help transparent hugepage allocations\n",
173 min_free_kbytes
, recommended_min
);
175 min_free_kbytes
= recommended_min
;
177 setup_per_zone_wmarks();
180 static int start_stop_khugepaged(void)
183 if (khugepaged_enabled()) {
184 if (!khugepaged_thread
)
185 khugepaged_thread
= kthread_run(khugepaged
, NULL
,
187 if (IS_ERR(khugepaged_thread
)) {
188 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
189 err
= PTR_ERR(khugepaged_thread
);
190 khugepaged_thread
= NULL
;
194 if (!list_empty(&khugepaged_scan
.mm_head
))
195 wake_up_interruptible(&khugepaged_wait
);
197 set_recommended_min_free_kbytes();
198 } else if (khugepaged_thread
) {
199 kthread_stop(khugepaged_thread
);
200 khugepaged_thread
= NULL
;
206 static atomic_t huge_zero_refcount
;
207 struct page
*huge_zero_page __read_mostly
;
209 struct page
*get_huge_zero_page(void)
211 struct page
*zero_page
;
213 if (likely(atomic_inc_not_zero(&huge_zero_refcount
)))
214 return READ_ONCE(huge_zero_page
);
216 zero_page
= alloc_pages((GFP_TRANSHUGE
| __GFP_ZERO
) & ~__GFP_MOVABLE
,
219 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED
);
222 count_vm_event(THP_ZERO_PAGE_ALLOC
);
224 if (cmpxchg(&huge_zero_page
, NULL
, zero_page
)) {
226 __free_pages(zero_page
, compound_order(zero_page
));
230 /* We take additional reference here. It will be put back by shrinker */
231 atomic_set(&huge_zero_refcount
, 2);
233 return READ_ONCE(huge_zero_page
);
236 static void put_huge_zero_page(void)
239 * Counter should never go to zero here. Only shrinker can put
242 BUG_ON(atomic_dec_and_test(&huge_zero_refcount
));
245 static unsigned long shrink_huge_zero_page_count(struct shrinker
*shrink
,
246 struct shrink_control
*sc
)
248 /* we can free zero page only if last reference remains */
249 return atomic_read(&huge_zero_refcount
) == 1 ? HPAGE_PMD_NR
: 0;
252 static unsigned long shrink_huge_zero_page_scan(struct shrinker
*shrink
,
253 struct shrink_control
*sc
)
255 if (atomic_cmpxchg(&huge_zero_refcount
, 1, 0) == 1) {
256 struct page
*zero_page
= xchg(&huge_zero_page
, NULL
);
257 BUG_ON(zero_page
== NULL
);
258 __free_pages(zero_page
, compound_order(zero_page
));
265 static struct shrinker huge_zero_page_shrinker
= {
266 .count_objects
= shrink_huge_zero_page_count
,
267 .scan_objects
= shrink_huge_zero_page_scan
,
268 .seeks
= DEFAULT_SEEKS
,
273 static ssize_t
triple_flag_store(struct kobject
*kobj
,
274 struct kobj_attribute
*attr
,
275 const char *buf
, size_t count
,
276 enum transparent_hugepage_flag enabled
,
277 enum transparent_hugepage_flag deferred
,
278 enum transparent_hugepage_flag req_madv
)
280 if (!memcmp("defer", buf
,
281 min(sizeof("defer")-1, count
))) {
282 if (enabled
== deferred
)
284 clear_bit(enabled
, &transparent_hugepage_flags
);
285 clear_bit(req_madv
, &transparent_hugepage_flags
);
286 set_bit(deferred
, &transparent_hugepage_flags
);
287 } else if (!memcmp("always", buf
,
288 min(sizeof("always")-1, count
))) {
289 clear_bit(deferred
, &transparent_hugepage_flags
);
290 clear_bit(req_madv
, &transparent_hugepage_flags
);
291 set_bit(enabled
, &transparent_hugepage_flags
);
292 } else if (!memcmp("madvise", buf
,
293 min(sizeof("madvise")-1, count
))) {
294 clear_bit(enabled
, &transparent_hugepage_flags
);
295 clear_bit(deferred
, &transparent_hugepage_flags
);
296 set_bit(req_madv
, &transparent_hugepage_flags
);
297 } else if (!memcmp("never", buf
,
298 min(sizeof("never")-1, count
))) {
299 clear_bit(enabled
, &transparent_hugepage_flags
);
300 clear_bit(req_madv
, &transparent_hugepage_flags
);
301 clear_bit(deferred
, &transparent_hugepage_flags
);
308 static ssize_t
enabled_show(struct kobject
*kobj
,
309 struct kobj_attribute
*attr
, char *buf
)
311 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
))
312 return sprintf(buf
, "[always] madvise never\n");
313 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
314 return sprintf(buf
, "always [madvise] never\n");
316 return sprintf(buf
, "always madvise [never]\n");
319 static ssize_t
enabled_store(struct kobject
*kobj
,
320 struct kobj_attribute
*attr
,
321 const char *buf
, size_t count
)
325 ret
= triple_flag_store(kobj
, attr
, buf
, count
,
326 TRANSPARENT_HUGEPAGE_FLAG
,
327 TRANSPARENT_HUGEPAGE_FLAG
,
328 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
333 mutex_lock(&khugepaged_mutex
);
334 err
= start_stop_khugepaged();
335 mutex_unlock(&khugepaged_mutex
);
343 static struct kobj_attribute enabled_attr
=
344 __ATTR(enabled
, 0644, enabled_show
, enabled_store
);
346 static ssize_t
single_flag_show(struct kobject
*kobj
,
347 struct kobj_attribute
*attr
, char *buf
,
348 enum transparent_hugepage_flag flag
)
350 return sprintf(buf
, "%d\n",
351 !!test_bit(flag
, &transparent_hugepage_flags
));
354 static ssize_t
single_flag_store(struct kobject
*kobj
,
355 struct kobj_attribute
*attr
,
356 const char *buf
, size_t count
,
357 enum transparent_hugepage_flag flag
)
362 ret
= kstrtoul(buf
, 10, &value
);
369 set_bit(flag
, &transparent_hugepage_flags
);
371 clear_bit(flag
, &transparent_hugepage_flags
);
377 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
378 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
379 * memory just to allocate one more hugepage.
381 static ssize_t
defrag_show(struct kobject
*kobj
,
382 struct kobj_attribute
*attr
, char *buf
)
384 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
))
385 return sprintf(buf
, "[always] defer madvise never\n");
386 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
))
387 return sprintf(buf
, "always [defer] madvise never\n");
388 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
389 return sprintf(buf
, "always defer [madvise] never\n");
391 return sprintf(buf
, "always defer madvise [never]\n");
394 static ssize_t
defrag_store(struct kobject
*kobj
,
395 struct kobj_attribute
*attr
,
396 const char *buf
, size_t count
)
398 return triple_flag_store(kobj
, attr
, buf
, count
,
399 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
,
400 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
,
401 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
);
403 static struct kobj_attribute defrag_attr
=
404 __ATTR(defrag
, 0644, defrag_show
, defrag_store
);
406 static ssize_t
use_zero_page_show(struct kobject
*kobj
,
407 struct kobj_attribute
*attr
, char *buf
)
409 return single_flag_show(kobj
, attr
, buf
,
410 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
412 static ssize_t
use_zero_page_store(struct kobject
*kobj
,
413 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
415 return single_flag_store(kobj
, attr
, buf
, count
,
416 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
418 static struct kobj_attribute use_zero_page_attr
=
419 __ATTR(use_zero_page
, 0644, use_zero_page_show
, use_zero_page_store
);
420 #ifdef CONFIG_DEBUG_VM
421 static ssize_t
debug_cow_show(struct kobject
*kobj
,
422 struct kobj_attribute
*attr
, char *buf
)
424 return single_flag_show(kobj
, attr
, buf
,
425 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
427 static ssize_t
debug_cow_store(struct kobject
*kobj
,
428 struct kobj_attribute
*attr
,
429 const char *buf
, size_t count
)
431 return single_flag_store(kobj
, attr
, buf
, count
,
432 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
434 static struct kobj_attribute debug_cow_attr
=
435 __ATTR(debug_cow
, 0644, debug_cow_show
, debug_cow_store
);
436 #endif /* CONFIG_DEBUG_VM */
438 static struct attribute
*hugepage_attr
[] = {
441 &use_zero_page_attr
.attr
,
442 #ifdef CONFIG_DEBUG_VM
443 &debug_cow_attr
.attr
,
448 static struct attribute_group hugepage_attr_group
= {
449 .attrs
= hugepage_attr
,
452 static ssize_t
scan_sleep_millisecs_show(struct kobject
*kobj
,
453 struct kobj_attribute
*attr
,
456 return sprintf(buf
, "%u\n", khugepaged_scan_sleep_millisecs
);
459 static ssize_t
scan_sleep_millisecs_store(struct kobject
*kobj
,
460 struct kobj_attribute
*attr
,
461 const char *buf
, size_t count
)
466 err
= kstrtoul(buf
, 10, &msecs
);
467 if (err
|| msecs
> UINT_MAX
)
470 khugepaged_scan_sleep_millisecs
= msecs
;
471 wake_up_interruptible(&khugepaged_wait
);
475 static struct kobj_attribute scan_sleep_millisecs_attr
=
476 __ATTR(scan_sleep_millisecs
, 0644, scan_sleep_millisecs_show
,
477 scan_sleep_millisecs_store
);
479 static ssize_t
alloc_sleep_millisecs_show(struct kobject
*kobj
,
480 struct kobj_attribute
*attr
,
483 return sprintf(buf
, "%u\n", khugepaged_alloc_sleep_millisecs
);
486 static ssize_t
alloc_sleep_millisecs_store(struct kobject
*kobj
,
487 struct kobj_attribute
*attr
,
488 const char *buf
, size_t count
)
493 err
= kstrtoul(buf
, 10, &msecs
);
494 if (err
|| msecs
> UINT_MAX
)
497 khugepaged_alloc_sleep_millisecs
= msecs
;
498 wake_up_interruptible(&khugepaged_wait
);
502 static struct kobj_attribute alloc_sleep_millisecs_attr
=
503 __ATTR(alloc_sleep_millisecs
, 0644, alloc_sleep_millisecs_show
,
504 alloc_sleep_millisecs_store
);
506 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
507 struct kobj_attribute
*attr
,
510 return sprintf(buf
, "%u\n", khugepaged_pages_to_scan
);
512 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
513 struct kobj_attribute
*attr
,
514 const char *buf
, size_t count
)
519 err
= kstrtoul(buf
, 10, &pages
);
520 if (err
|| !pages
|| pages
> UINT_MAX
)
523 khugepaged_pages_to_scan
= pages
;
527 static struct kobj_attribute pages_to_scan_attr
=
528 __ATTR(pages_to_scan
, 0644, pages_to_scan_show
,
529 pages_to_scan_store
);
531 static ssize_t
pages_collapsed_show(struct kobject
*kobj
,
532 struct kobj_attribute
*attr
,
535 return sprintf(buf
, "%u\n", khugepaged_pages_collapsed
);
537 static struct kobj_attribute pages_collapsed_attr
=
538 __ATTR_RO(pages_collapsed
);
540 static ssize_t
full_scans_show(struct kobject
*kobj
,
541 struct kobj_attribute
*attr
,
544 return sprintf(buf
, "%u\n", khugepaged_full_scans
);
546 static struct kobj_attribute full_scans_attr
=
547 __ATTR_RO(full_scans
);
549 static ssize_t
khugepaged_defrag_show(struct kobject
*kobj
,
550 struct kobj_attribute
*attr
, char *buf
)
552 return single_flag_show(kobj
, attr
, buf
,
553 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
555 static ssize_t
khugepaged_defrag_store(struct kobject
*kobj
,
556 struct kobj_attribute
*attr
,
557 const char *buf
, size_t count
)
559 return single_flag_store(kobj
, attr
, buf
, count
,
560 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
562 static struct kobj_attribute khugepaged_defrag_attr
=
563 __ATTR(defrag
, 0644, khugepaged_defrag_show
,
564 khugepaged_defrag_store
);
567 * max_ptes_none controls if khugepaged should collapse hugepages over
568 * any unmapped ptes in turn potentially increasing the memory
569 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
570 * reduce the available free memory in the system as it
571 * runs. Increasing max_ptes_none will instead potentially reduce the
572 * free memory in the system during the khugepaged scan.
574 static ssize_t
khugepaged_max_ptes_none_show(struct kobject
*kobj
,
575 struct kobj_attribute
*attr
,
578 return sprintf(buf
, "%u\n", khugepaged_max_ptes_none
);
580 static ssize_t
khugepaged_max_ptes_none_store(struct kobject
*kobj
,
581 struct kobj_attribute
*attr
,
582 const char *buf
, size_t count
)
585 unsigned long max_ptes_none
;
587 err
= kstrtoul(buf
, 10, &max_ptes_none
);
588 if (err
|| max_ptes_none
> HPAGE_PMD_NR
-1)
591 khugepaged_max_ptes_none
= max_ptes_none
;
595 static struct kobj_attribute khugepaged_max_ptes_none_attr
=
596 __ATTR(max_ptes_none
, 0644, khugepaged_max_ptes_none_show
,
597 khugepaged_max_ptes_none_store
);
599 static struct attribute
*khugepaged_attr
[] = {
600 &khugepaged_defrag_attr
.attr
,
601 &khugepaged_max_ptes_none_attr
.attr
,
602 &pages_to_scan_attr
.attr
,
603 &pages_collapsed_attr
.attr
,
604 &full_scans_attr
.attr
,
605 &scan_sleep_millisecs_attr
.attr
,
606 &alloc_sleep_millisecs_attr
.attr
,
610 static struct attribute_group khugepaged_attr_group
= {
611 .attrs
= khugepaged_attr
,
612 .name
= "khugepaged",
615 static int __init
hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
619 *hugepage_kobj
= kobject_create_and_add("transparent_hugepage", mm_kobj
);
620 if (unlikely(!*hugepage_kobj
)) {
621 pr_err("failed to create transparent hugepage kobject\n");
625 err
= sysfs_create_group(*hugepage_kobj
, &hugepage_attr_group
);
627 pr_err("failed to register transparent hugepage group\n");
631 err
= sysfs_create_group(*hugepage_kobj
, &khugepaged_attr_group
);
633 pr_err("failed to register transparent hugepage group\n");
634 goto remove_hp_group
;
640 sysfs_remove_group(*hugepage_kobj
, &hugepage_attr_group
);
642 kobject_put(*hugepage_kobj
);
646 static void __init
hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
648 sysfs_remove_group(hugepage_kobj
, &khugepaged_attr_group
);
649 sysfs_remove_group(hugepage_kobj
, &hugepage_attr_group
);
650 kobject_put(hugepage_kobj
);
653 static inline int hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
658 static inline void hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
661 #endif /* CONFIG_SYSFS */
663 static int __init
hugepage_init(void)
666 struct kobject
*hugepage_kobj
;
668 if (!has_transparent_hugepage()) {
669 transparent_hugepage_flags
= 0;
673 err
= hugepage_init_sysfs(&hugepage_kobj
);
677 err
= khugepaged_slab_init();
681 err
= register_shrinker(&huge_zero_page_shrinker
);
683 goto err_hzp_shrinker
;
684 err
= register_shrinker(&deferred_split_shrinker
);
686 goto err_split_shrinker
;
689 * By default disable transparent hugepages on smaller systems,
690 * where the extra memory used could hurt more than TLB overhead
691 * is likely to save. The admin can still enable it through /sys.
693 if (totalram_pages
< (512 << (20 - PAGE_SHIFT
))) {
694 transparent_hugepage_flags
= 0;
698 err
= start_stop_khugepaged();
704 unregister_shrinker(&deferred_split_shrinker
);
706 unregister_shrinker(&huge_zero_page_shrinker
);
708 khugepaged_slab_exit();
710 hugepage_exit_sysfs(hugepage_kobj
);
714 subsys_initcall(hugepage_init
);
716 static int __init
setup_transparent_hugepage(char *str
)
721 if (!strcmp(str
, "always")) {
722 set_bit(TRANSPARENT_HUGEPAGE_FLAG
,
723 &transparent_hugepage_flags
);
724 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
725 &transparent_hugepage_flags
);
727 } else if (!strcmp(str
, "madvise")) {
728 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
729 &transparent_hugepage_flags
);
730 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
731 &transparent_hugepage_flags
);
733 } else if (!strcmp(str
, "never")) {
734 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
735 &transparent_hugepage_flags
);
736 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
737 &transparent_hugepage_flags
);
742 pr_warn("transparent_hugepage= cannot parse, ignored\n");
745 __setup("transparent_hugepage=", setup_transparent_hugepage
);
747 pmd_t
maybe_pmd_mkwrite(pmd_t pmd
, struct vm_area_struct
*vma
)
749 if (likely(vma
->vm_flags
& VM_WRITE
))
750 pmd
= pmd_mkwrite(pmd
);
754 static inline pmd_t
mk_huge_pmd(struct page
*page
, pgprot_t prot
)
757 entry
= mk_pmd(page
, prot
);
758 entry
= pmd_mkhuge(entry
);
762 static inline struct list_head
*page_deferred_list(struct page
*page
)
765 * ->lru in the tail pages is occupied by compound_head.
766 * Let's use ->mapping + ->index in the second tail page as list_head.
768 return (struct list_head
*)&page
[2].mapping
;
771 void prep_transhuge_page(struct page
*page
)
774 * we use page->mapping and page->indexlru in second tail page
775 * as list_head: assuming THP order >= 2
777 BUILD_BUG_ON(HPAGE_PMD_ORDER
< 2);
779 INIT_LIST_HEAD(page_deferred_list(page
));
780 set_compound_page_dtor(page
, TRANSHUGE_PAGE_DTOR
);
783 static int __do_huge_pmd_anonymous_page(struct mm_struct
*mm
,
784 struct vm_area_struct
*vma
,
785 unsigned long address
, pmd_t
*pmd
,
786 struct page
*page
, gfp_t gfp
,
789 struct mem_cgroup
*memcg
;
792 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
794 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
796 if (mem_cgroup_try_charge(page
, mm
, gfp
, &memcg
, true)) {
798 count_vm_event(THP_FAULT_FALLBACK
);
799 return VM_FAULT_FALLBACK
;
802 pgtable
= pte_alloc_one(mm
, haddr
);
803 if (unlikely(!pgtable
)) {
804 mem_cgroup_cancel_charge(page
, memcg
, true);
809 clear_huge_page(page
, haddr
, HPAGE_PMD_NR
);
811 * The memory barrier inside __SetPageUptodate makes sure that
812 * clear_huge_page writes become visible before the set_pmd_at()
815 __SetPageUptodate(page
);
817 ptl
= pmd_lock(mm
, pmd
);
818 if (unlikely(!pmd_none(*pmd
))) {
820 mem_cgroup_cancel_charge(page
, memcg
, true);
822 pte_free(mm
, pgtable
);
826 /* Deliver the page fault to userland */
827 if (userfaultfd_missing(vma
)) {
831 mem_cgroup_cancel_charge(page
, memcg
, true);
833 pte_free(mm
, pgtable
);
834 ret
= handle_userfault(vma
, address
, flags
,
836 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
840 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
841 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
842 page_add_new_anon_rmap(page
, vma
, haddr
, true);
843 mem_cgroup_commit_charge(page
, memcg
, false, true);
844 lru_cache_add_active_or_unevictable(page
, vma
);
845 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
846 set_pmd_at(mm
, haddr
, pmd
, entry
);
847 add_mm_counter(mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
848 atomic_long_inc(&mm
->nr_ptes
);
850 count_vm_event(THP_FAULT_ALLOC
);
857 * If THP is set to always then directly reclaim/compact as necessary
858 * If set to defer then do no reclaim and defer to khugepaged
859 * If set to madvise and the VMA is flagged then directly reclaim/compact
861 static inline gfp_t
alloc_hugepage_direct_gfpmask(struct vm_area_struct
*vma
)
863 gfp_t reclaim_flags
= 0;
865 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
) &&
866 (vma
->vm_flags
& VM_HUGEPAGE
))
867 reclaim_flags
= __GFP_DIRECT_RECLAIM
;
868 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
))
869 reclaim_flags
= __GFP_KSWAPD_RECLAIM
;
870 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
))
871 reclaim_flags
= __GFP_DIRECT_RECLAIM
;
873 return GFP_TRANSHUGE
| reclaim_flags
;
876 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
877 static inline gfp_t
alloc_hugepage_khugepaged_gfpmask(void)
879 return GFP_TRANSHUGE
| (khugepaged_defrag() ? __GFP_DIRECT_RECLAIM
: 0);
882 /* Caller must hold page table lock. */
883 static bool set_huge_zero_page(pgtable_t pgtable
, struct mm_struct
*mm
,
884 struct vm_area_struct
*vma
, unsigned long haddr
, pmd_t
*pmd
,
885 struct page
*zero_page
)
890 entry
= mk_pmd(zero_page
, vma
->vm_page_prot
);
891 entry
= pmd_mkhuge(entry
);
893 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
894 set_pmd_at(mm
, haddr
, pmd
, entry
);
895 atomic_long_inc(&mm
->nr_ptes
);
899 int do_huge_pmd_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
900 unsigned long address
, pmd_t
*pmd
,
905 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
907 if (haddr
< vma
->vm_start
|| haddr
+ HPAGE_PMD_SIZE
> vma
->vm_end
)
908 return VM_FAULT_FALLBACK
;
909 if (unlikely(anon_vma_prepare(vma
)))
911 if (unlikely(khugepaged_enter(vma
, vma
->vm_flags
)))
913 if (!(flags
& FAULT_FLAG_WRITE
) && !mm_forbids_zeropage(mm
) &&
914 transparent_hugepage_use_zero_page()) {
917 struct page
*zero_page
;
920 pgtable
= pte_alloc_one(mm
, haddr
);
921 if (unlikely(!pgtable
))
923 zero_page
= get_huge_zero_page();
924 if (unlikely(!zero_page
)) {
925 pte_free(mm
, pgtable
);
926 count_vm_event(THP_FAULT_FALLBACK
);
927 return VM_FAULT_FALLBACK
;
929 ptl
= pmd_lock(mm
, pmd
);
932 if (pmd_none(*pmd
)) {
933 if (userfaultfd_missing(vma
)) {
935 ret
= handle_userfault(vma
, address
, flags
,
937 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
939 set_huge_zero_page(pgtable
, mm
, vma
,
948 pte_free(mm
, pgtable
);
949 put_huge_zero_page();
953 gfp
= alloc_hugepage_direct_gfpmask(vma
);
954 page
= alloc_hugepage_vma(gfp
, vma
, haddr
, HPAGE_PMD_ORDER
);
955 if (unlikely(!page
)) {
956 count_vm_event(THP_FAULT_FALLBACK
);
957 return VM_FAULT_FALLBACK
;
959 prep_transhuge_page(page
);
960 return __do_huge_pmd_anonymous_page(mm
, vma
, address
, pmd
, page
, gfp
,
964 static void insert_pfn_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
965 pmd_t
*pmd
, pfn_t pfn
, pgprot_t prot
, bool write
)
967 struct mm_struct
*mm
= vma
->vm_mm
;
971 ptl
= pmd_lock(mm
, pmd
);
972 entry
= pmd_mkhuge(pfn_t_pmd(pfn
, prot
));
973 if (pfn_t_devmap(pfn
))
974 entry
= pmd_mkdevmap(entry
);
976 entry
= pmd_mkyoung(pmd_mkdirty(entry
));
977 entry
= maybe_pmd_mkwrite(entry
, vma
);
979 set_pmd_at(mm
, addr
, pmd
, entry
);
980 update_mmu_cache_pmd(vma
, addr
, pmd
);
984 int vmf_insert_pfn_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
985 pmd_t
*pmd
, pfn_t pfn
, bool write
)
987 pgprot_t pgprot
= vma
->vm_page_prot
;
989 * If we had pmd_special, we could avoid all these restrictions,
990 * but we need to be consistent with PTEs and architectures that
991 * can't support a 'special' bit.
993 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
994 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
995 (VM_PFNMAP
|VM_MIXEDMAP
));
996 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
997 BUG_ON(!pfn_t_devmap(pfn
));
999 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1000 return VM_FAULT_SIGBUS
;
1001 if (track_pfn_insert(vma
, &pgprot
, pfn
))
1002 return VM_FAULT_SIGBUS
;
1003 insert_pfn_pmd(vma
, addr
, pmd
, pfn
, pgprot
, write
);
1004 return VM_FAULT_NOPAGE
;
1007 static void touch_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
1013 * We should set the dirty bit only for FOLL_WRITE but for now
1014 * the dirty bit in the pmd is meaningless. And if the dirty
1015 * bit will become meaningful and we'll only set it with
1016 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
1017 * set the young bit, instead of the current set_pmd_at.
1019 _pmd
= pmd_mkyoung(pmd_mkdirty(*pmd
));
1020 if (pmdp_set_access_flags(vma
, addr
& HPAGE_PMD_MASK
,
1022 update_mmu_cache_pmd(vma
, addr
, pmd
);
1025 struct page
*follow_devmap_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
1026 pmd_t
*pmd
, int flags
)
1028 unsigned long pfn
= pmd_pfn(*pmd
);
1029 struct mm_struct
*mm
= vma
->vm_mm
;
1030 struct dev_pagemap
*pgmap
;
1033 assert_spin_locked(pmd_lockptr(mm
, pmd
));
1035 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
1038 if (pmd_present(*pmd
) && pmd_devmap(*pmd
))
1043 if (flags
& FOLL_TOUCH
)
1044 touch_pmd(vma
, addr
, pmd
);
1047 * device mapped pages can only be returned if the
1048 * caller will manage the page reference count.
1050 if (!(flags
& FOLL_GET
))
1051 return ERR_PTR(-EEXIST
);
1053 pfn
+= (addr
& ~PMD_MASK
) >> PAGE_SHIFT
;
1054 pgmap
= get_dev_pagemap(pfn
, NULL
);
1056 return ERR_PTR(-EFAULT
);
1057 page
= pfn_to_page(pfn
);
1059 put_dev_pagemap(pgmap
);
1064 int copy_huge_pmd(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1065 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
1066 struct vm_area_struct
*vma
)
1068 spinlock_t
*dst_ptl
, *src_ptl
;
1069 struct page
*src_page
;
1071 pgtable_t pgtable
= NULL
;
1074 if (!vma_is_dax(vma
)) {
1076 pgtable
= pte_alloc_one(dst_mm
, addr
);
1077 if (unlikely(!pgtable
))
1081 dst_ptl
= pmd_lock(dst_mm
, dst_pmd
);
1082 src_ptl
= pmd_lockptr(src_mm
, src_pmd
);
1083 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1087 if (unlikely(!pmd_trans_huge(pmd
) && !pmd_devmap(pmd
))) {
1088 pte_free(dst_mm
, pgtable
);
1092 * When page table lock is held, the huge zero pmd should not be
1093 * under splitting since we don't split the page itself, only pmd to
1096 if (is_huge_zero_pmd(pmd
)) {
1097 struct page
*zero_page
;
1099 * get_huge_zero_page() will never allocate a new page here,
1100 * since we already have a zero page to copy. It just takes a
1103 zero_page
= get_huge_zero_page();
1104 set_huge_zero_page(pgtable
, dst_mm
, vma
, addr
, dst_pmd
,
1110 if (!vma_is_dax(vma
)) {
1111 /* thp accounting separate from pmd_devmap accounting */
1112 src_page
= pmd_page(pmd
);
1113 VM_BUG_ON_PAGE(!PageHead(src_page
), src_page
);
1115 page_dup_rmap(src_page
, true);
1116 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1117 atomic_long_inc(&dst_mm
->nr_ptes
);
1118 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
1121 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
1122 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
1123 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
1127 spin_unlock(src_ptl
);
1128 spin_unlock(dst_ptl
);
1133 void huge_pmd_set_accessed(struct mm_struct
*mm
,
1134 struct vm_area_struct
*vma
,
1135 unsigned long address
,
1136 pmd_t
*pmd
, pmd_t orig_pmd
,
1141 unsigned long haddr
;
1143 ptl
= pmd_lock(mm
, pmd
);
1144 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
1147 entry
= pmd_mkyoung(orig_pmd
);
1148 haddr
= address
& HPAGE_PMD_MASK
;
1149 if (pmdp_set_access_flags(vma
, haddr
, pmd
, entry
, dirty
))
1150 update_mmu_cache_pmd(vma
, address
, pmd
);
1156 static int do_huge_pmd_wp_page_fallback(struct mm_struct
*mm
,
1157 struct vm_area_struct
*vma
,
1158 unsigned long address
,
1159 pmd_t
*pmd
, pmd_t orig_pmd
,
1161 unsigned long haddr
)
1163 struct mem_cgroup
*memcg
;
1168 struct page
**pages
;
1169 unsigned long mmun_start
; /* For mmu_notifiers */
1170 unsigned long mmun_end
; /* For mmu_notifiers */
1172 pages
= kmalloc(sizeof(struct page
*) * HPAGE_PMD_NR
,
1174 if (unlikely(!pages
)) {
1175 ret
|= VM_FAULT_OOM
;
1179 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1180 pages
[i
] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE
|
1182 vma
, address
, page_to_nid(page
));
1183 if (unlikely(!pages
[i
] ||
1184 mem_cgroup_try_charge(pages
[i
], mm
, GFP_KERNEL
,
1189 memcg
= (void *)page_private(pages
[i
]);
1190 set_page_private(pages
[i
], 0);
1191 mem_cgroup_cancel_charge(pages
[i
], memcg
,
1196 ret
|= VM_FAULT_OOM
;
1199 set_page_private(pages
[i
], (unsigned long)memcg
);
1202 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1203 copy_user_highpage(pages
[i
], page
+ i
,
1204 haddr
+ PAGE_SIZE
* i
, vma
);
1205 __SetPageUptodate(pages
[i
]);
1210 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1211 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1213 ptl
= pmd_lock(mm
, pmd
);
1214 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
1215 goto out_free_pages
;
1216 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1218 pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
1219 /* leave pmd empty until pte is filled */
1221 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1222 pmd_populate(mm
, &_pmd
, pgtable
);
1224 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1226 entry
= mk_pte(pages
[i
], vma
->vm_page_prot
);
1227 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1228 memcg
= (void *)page_private(pages
[i
]);
1229 set_page_private(pages
[i
], 0);
1230 page_add_new_anon_rmap(pages
[i
], vma
, haddr
, false);
1231 mem_cgroup_commit_charge(pages
[i
], memcg
, false, false);
1232 lru_cache_add_active_or_unevictable(pages
[i
], vma
);
1233 pte
= pte_offset_map(&_pmd
, haddr
);
1234 VM_BUG_ON(!pte_none(*pte
));
1235 set_pte_at(mm
, haddr
, pte
, entry
);
1240 smp_wmb(); /* make pte visible before pmd */
1241 pmd_populate(mm
, pmd
, pgtable
);
1242 page_remove_rmap(page
, true);
1245 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1247 ret
|= VM_FAULT_WRITE
;
1255 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1256 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1257 memcg
= (void *)page_private(pages
[i
]);
1258 set_page_private(pages
[i
], 0);
1259 mem_cgroup_cancel_charge(pages
[i
], memcg
, false);
1266 int do_huge_pmd_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1267 unsigned long address
, pmd_t
*pmd
, pmd_t orig_pmd
)
1271 struct page
*page
= NULL
, *new_page
;
1272 struct mem_cgroup
*memcg
;
1273 unsigned long haddr
;
1274 unsigned long mmun_start
; /* For mmu_notifiers */
1275 unsigned long mmun_end
; /* For mmu_notifiers */
1276 gfp_t huge_gfp
; /* for allocation and charge */
1278 ptl
= pmd_lockptr(mm
, pmd
);
1279 VM_BUG_ON_VMA(!vma
->anon_vma
, vma
);
1280 haddr
= address
& HPAGE_PMD_MASK
;
1281 if (is_huge_zero_pmd(orig_pmd
))
1284 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
1287 page
= pmd_page(orig_pmd
);
1288 VM_BUG_ON_PAGE(!PageCompound(page
) || !PageHead(page
), page
);
1290 * We can only reuse the page if nobody else maps the huge page or it's
1291 * part. We can do it by checking page_mapcount() on each sub-page, but
1293 * The cheaper way is to check page_count() to be equal 1: every
1294 * mapcount takes page reference reference, so this way we can
1295 * guarantee, that the PMD is the only mapping.
1296 * This can give false negative if somebody pinned the page, but that's
1299 if (page_mapcount(page
) == 1 && page_count(page
) == 1) {
1301 entry
= pmd_mkyoung(orig_pmd
);
1302 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1303 if (pmdp_set_access_flags(vma
, haddr
, pmd
, entry
, 1))
1304 update_mmu_cache_pmd(vma
, address
, pmd
);
1305 ret
|= VM_FAULT_WRITE
;
1311 if (transparent_hugepage_enabled(vma
) &&
1312 !transparent_hugepage_debug_cow()) {
1313 huge_gfp
= alloc_hugepage_direct_gfpmask(vma
);
1314 new_page
= alloc_hugepage_vma(huge_gfp
, vma
, haddr
, HPAGE_PMD_ORDER
);
1318 if (likely(new_page
)) {
1319 prep_transhuge_page(new_page
);
1322 split_huge_pmd(vma
, pmd
, address
);
1323 ret
|= VM_FAULT_FALLBACK
;
1325 ret
= do_huge_pmd_wp_page_fallback(mm
, vma
, address
,
1326 pmd
, orig_pmd
, page
, haddr
);
1327 if (ret
& VM_FAULT_OOM
) {
1328 split_huge_pmd(vma
, pmd
, address
);
1329 ret
|= VM_FAULT_FALLBACK
;
1333 count_vm_event(THP_FAULT_FALLBACK
);
1337 if (unlikely(mem_cgroup_try_charge(new_page
, mm
, huge_gfp
, &memcg
,
1341 split_huge_pmd(vma
, pmd
, address
);
1344 split_huge_pmd(vma
, pmd
, address
);
1345 ret
|= VM_FAULT_FALLBACK
;
1346 count_vm_event(THP_FAULT_FALLBACK
);
1350 count_vm_event(THP_FAULT_ALLOC
);
1353 clear_huge_page(new_page
, haddr
, HPAGE_PMD_NR
);
1355 copy_user_huge_page(new_page
, page
, haddr
, vma
, HPAGE_PMD_NR
);
1356 __SetPageUptodate(new_page
);
1359 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1360 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1365 if (unlikely(!pmd_same(*pmd
, orig_pmd
))) {
1367 mem_cgroup_cancel_charge(new_page
, memcg
, true);
1372 entry
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
1373 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1374 pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
1375 page_add_new_anon_rmap(new_page
, vma
, haddr
, true);
1376 mem_cgroup_commit_charge(new_page
, memcg
, false, true);
1377 lru_cache_add_active_or_unevictable(new_page
, vma
);
1378 set_pmd_at(mm
, haddr
, pmd
, entry
);
1379 update_mmu_cache_pmd(vma
, address
, pmd
);
1381 add_mm_counter(mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1382 put_huge_zero_page();
1384 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1385 page_remove_rmap(page
, true);
1388 ret
|= VM_FAULT_WRITE
;
1392 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1400 struct page
*follow_trans_huge_pmd(struct vm_area_struct
*vma
,
1405 struct mm_struct
*mm
= vma
->vm_mm
;
1406 struct page
*page
= NULL
;
1408 assert_spin_locked(pmd_lockptr(mm
, pmd
));
1410 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
1413 /* Avoid dumping huge zero page */
1414 if ((flags
& FOLL_DUMP
) && is_huge_zero_pmd(*pmd
))
1415 return ERR_PTR(-EFAULT
);
1417 /* Full NUMA hinting faults to serialise migration in fault paths */
1418 if ((flags
& FOLL_NUMA
) && pmd_protnone(*pmd
))
1421 page
= pmd_page(*pmd
);
1422 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1423 if (flags
& FOLL_TOUCH
)
1424 touch_pmd(vma
, addr
, pmd
);
1425 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1427 * We don't mlock() pte-mapped THPs. This way we can avoid
1428 * leaking mlocked pages into non-VM_LOCKED VMAs.
1430 * In most cases the pmd is the only mapping of the page as we
1431 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1432 * writable private mappings in populate_vma_page_range().
1434 * The only scenario when we have the page shared here is if we
1435 * mlocking read-only mapping shared over fork(). We skip
1436 * mlocking such pages.
1438 if (compound_mapcount(page
) == 1 && !PageDoubleMap(page
) &&
1439 page
->mapping
&& trylock_page(page
)) {
1442 mlock_vma_page(page
);
1446 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
1447 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
1448 if (flags
& FOLL_GET
)
1455 /* NUMA hinting page fault entry point for trans huge pmds */
1456 int do_huge_pmd_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1457 unsigned long addr
, pmd_t pmd
, pmd_t
*pmdp
)
1460 struct anon_vma
*anon_vma
= NULL
;
1462 unsigned long haddr
= addr
& HPAGE_PMD_MASK
;
1463 int page_nid
= -1, this_nid
= numa_node_id();
1464 int target_nid
, last_cpupid
= -1;
1466 bool migrated
= false;
1470 /* A PROT_NONE fault should not end up here */
1471 BUG_ON(!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)));
1473 ptl
= pmd_lock(mm
, pmdp
);
1474 if (unlikely(!pmd_same(pmd
, *pmdp
)))
1478 * If there are potential migrations, wait for completion and retry
1479 * without disrupting NUMA hinting information. Do not relock and
1480 * check_same as the page may no longer be mapped.
1482 if (unlikely(pmd_trans_migrating(*pmdp
))) {
1483 page
= pmd_page(*pmdp
);
1485 wait_on_page_locked(page
);
1489 page
= pmd_page(pmd
);
1490 BUG_ON(is_huge_zero_page(page
));
1491 page_nid
= page_to_nid(page
);
1492 last_cpupid
= page_cpupid_last(page
);
1493 count_vm_numa_event(NUMA_HINT_FAULTS
);
1494 if (page_nid
== this_nid
) {
1495 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
1496 flags
|= TNF_FAULT_LOCAL
;
1499 /* See similar comment in do_numa_page for explanation */
1500 if (!(vma
->vm_flags
& VM_WRITE
))
1501 flags
|= TNF_NO_GROUP
;
1504 * Acquire the page lock to serialise THP migrations but avoid dropping
1505 * page_table_lock if at all possible
1507 page_locked
= trylock_page(page
);
1508 target_nid
= mpol_misplaced(page
, vma
, haddr
);
1509 if (target_nid
== -1) {
1510 /* If the page was locked, there are no parallel migrations */
1515 /* Migration could have started since the pmd_trans_migrating check */
1518 wait_on_page_locked(page
);
1524 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1525 * to serialises splits
1529 anon_vma
= page_lock_anon_vma_read(page
);
1531 /* Confirm the PMD did not change while page_table_lock was released */
1533 if (unlikely(!pmd_same(pmd
, *pmdp
))) {
1540 /* Bail if we fail to protect against THP splits for any reason */
1541 if (unlikely(!anon_vma
)) {
1548 * Migrate the THP to the requested node, returns with page unlocked
1549 * and access rights restored.
1552 migrated
= migrate_misplaced_transhuge_page(mm
, vma
,
1553 pmdp
, pmd
, addr
, page
, target_nid
);
1555 flags
|= TNF_MIGRATED
;
1556 page_nid
= target_nid
;
1558 flags
|= TNF_MIGRATE_FAIL
;
1562 BUG_ON(!PageLocked(page
));
1563 was_writable
= pmd_write(pmd
);
1564 pmd
= pmd_modify(pmd
, vma
->vm_page_prot
);
1565 pmd
= pmd_mkyoung(pmd
);
1567 pmd
= pmd_mkwrite(pmd
);
1568 set_pmd_at(mm
, haddr
, pmdp
, pmd
);
1569 update_mmu_cache_pmd(vma
, addr
, pmdp
);
1576 page_unlock_anon_vma_read(anon_vma
);
1579 task_numa_fault(last_cpupid
, page_nid
, HPAGE_PMD_NR
, flags
);
1584 int madvise_free_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1585 pmd_t
*pmd
, unsigned long addr
, unsigned long next
)
1591 struct mm_struct
*mm
= tlb
->mm
;
1594 ptl
= pmd_trans_huge_lock(pmd
, vma
);
1599 if (is_huge_zero_pmd(orig_pmd
)) {
1604 page
= pmd_page(orig_pmd
);
1606 * If other processes are mapping this page, we couldn't discard
1607 * the page unless they all do MADV_FREE so let's skip the page.
1609 if (page_mapcount(page
) != 1)
1612 if (!trylock_page(page
))
1616 * If user want to discard part-pages of THP, split it so MADV_FREE
1617 * will deactivate only them.
1619 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1622 if (split_huge_page(page
)) {
1633 if (PageDirty(page
))
1634 ClearPageDirty(page
);
1637 if (PageActive(page
))
1638 deactivate_page(page
);
1640 if (pmd_young(orig_pmd
) || pmd_dirty(orig_pmd
)) {
1641 orig_pmd
= pmdp_huge_get_and_clear_full(tlb
->mm
, addr
, pmd
,
1643 orig_pmd
= pmd_mkold(orig_pmd
);
1644 orig_pmd
= pmd_mkclean(orig_pmd
);
1646 set_pmd_at(mm
, addr
, pmd
, orig_pmd
);
1647 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1656 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1657 pmd_t
*pmd
, unsigned long addr
)
1662 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1666 * For architectures like ppc64 we look at deposited pgtable
1667 * when calling pmdp_huge_get_and_clear. So do the
1668 * pgtable_trans_huge_withdraw after finishing pmdp related
1671 orig_pmd
= pmdp_huge_get_and_clear_full(tlb
->mm
, addr
, pmd
,
1673 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1674 if (vma_is_dax(vma
)) {
1676 if (is_huge_zero_pmd(orig_pmd
))
1677 put_huge_zero_page();
1678 } else if (is_huge_zero_pmd(orig_pmd
)) {
1679 pte_free(tlb
->mm
, pgtable_trans_huge_withdraw(tlb
->mm
, pmd
));
1680 atomic_long_dec(&tlb
->mm
->nr_ptes
);
1682 put_huge_zero_page();
1684 struct page
*page
= pmd_page(orig_pmd
);
1685 page_remove_rmap(page
, true);
1686 VM_BUG_ON_PAGE(page_mapcount(page
) < 0, page
);
1687 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1688 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1689 pte_free(tlb
->mm
, pgtable_trans_huge_withdraw(tlb
->mm
, pmd
));
1690 atomic_long_dec(&tlb
->mm
->nr_ptes
);
1692 tlb_remove_page(tlb
, page
);
1697 bool move_huge_pmd(struct vm_area_struct
*vma
, struct vm_area_struct
*new_vma
,
1698 unsigned long old_addr
,
1699 unsigned long new_addr
, unsigned long old_end
,
1700 pmd_t
*old_pmd
, pmd_t
*new_pmd
)
1702 spinlock_t
*old_ptl
, *new_ptl
;
1705 struct mm_struct
*mm
= vma
->vm_mm
;
1707 if ((old_addr
& ~HPAGE_PMD_MASK
) ||
1708 (new_addr
& ~HPAGE_PMD_MASK
) ||
1709 old_end
- old_addr
< HPAGE_PMD_SIZE
||
1710 (new_vma
->vm_flags
& VM_NOHUGEPAGE
))
1714 * The destination pmd shouldn't be established, free_pgtables()
1715 * should have release it.
1717 if (WARN_ON(!pmd_none(*new_pmd
))) {
1718 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1723 * We don't have to worry about the ordering of src and dst
1724 * ptlocks because exclusive mmap_sem prevents deadlock.
1726 old_ptl
= __pmd_trans_huge_lock(old_pmd
, vma
);
1728 new_ptl
= pmd_lockptr(mm
, new_pmd
);
1729 if (new_ptl
!= old_ptl
)
1730 spin_lock_nested(new_ptl
, SINGLE_DEPTH_NESTING
);
1731 pmd
= pmdp_huge_get_and_clear(mm
, old_addr
, old_pmd
);
1732 VM_BUG_ON(!pmd_none(*new_pmd
));
1734 if (pmd_move_must_withdraw(new_ptl
, old_ptl
) &&
1735 vma_is_anonymous(vma
)) {
1737 pgtable
= pgtable_trans_huge_withdraw(mm
, old_pmd
);
1738 pgtable_trans_huge_deposit(mm
, new_pmd
, pgtable
);
1740 set_pmd_at(mm
, new_addr
, new_pmd
, pmd_mksoft_dirty(pmd
));
1741 if (new_ptl
!= old_ptl
)
1742 spin_unlock(new_ptl
);
1743 spin_unlock(old_ptl
);
1751 * - 0 if PMD could not be locked
1752 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1753 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1755 int change_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1756 unsigned long addr
, pgprot_t newprot
, int prot_numa
)
1758 struct mm_struct
*mm
= vma
->vm_mm
;
1762 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1765 bool preserve_write
= prot_numa
&& pmd_write(*pmd
);
1769 * Avoid trapping faults against the zero page. The read-only
1770 * data is likely to be read-cached on the local CPU and
1771 * local/remote hits to the zero page are not interesting.
1773 if (prot_numa
&& is_huge_zero_pmd(*pmd
)) {
1778 if (!prot_numa
|| !pmd_protnone(*pmd
)) {
1779 entry
= pmdp_huge_get_and_clear_notify(mm
, addr
, pmd
);
1780 entry
= pmd_modify(entry
, newprot
);
1782 entry
= pmd_mkwrite(entry
);
1784 set_pmd_at(mm
, addr
, pmd
, entry
);
1785 BUG_ON(!preserve_write
&& pmd_write(entry
));
1794 * Returns true if a given pmd maps a thp, false otherwise.
1796 * Note that if it returns true, this routine returns without unlocking page
1797 * table lock. So callers must unlock it.
1799 spinlock_t
*__pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
)
1802 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
1803 if (likely(pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)))
1809 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1811 int hugepage_madvise(struct vm_area_struct
*vma
,
1812 unsigned long *vm_flags
, int advice
)
1818 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1819 * can't handle this properly after s390_enable_sie, so we simply
1820 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1822 if (mm_has_pgste(vma
->vm_mm
))
1826 * Be somewhat over-protective like KSM for now!
1828 if (*vm_flags
& VM_NO_THP
)
1830 *vm_flags
&= ~VM_NOHUGEPAGE
;
1831 *vm_flags
|= VM_HUGEPAGE
;
1833 * If the vma become good for khugepaged to scan,
1834 * register it here without waiting a page fault that
1835 * may not happen any time soon.
1837 if (unlikely(khugepaged_enter_vma_merge(vma
, *vm_flags
)))
1840 case MADV_NOHUGEPAGE
:
1842 * Be somewhat over-protective like KSM for now!
1844 if (*vm_flags
& VM_NO_THP
)
1846 *vm_flags
&= ~VM_HUGEPAGE
;
1847 *vm_flags
|= VM_NOHUGEPAGE
;
1849 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1850 * this vma even if we leave the mm registered in khugepaged if
1851 * it got registered before VM_NOHUGEPAGE was set.
1859 static int __init
khugepaged_slab_init(void)
1861 mm_slot_cache
= kmem_cache_create("khugepaged_mm_slot",
1862 sizeof(struct mm_slot
),
1863 __alignof__(struct mm_slot
), 0, NULL
);
1870 static void __init
khugepaged_slab_exit(void)
1872 kmem_cache_destroy(mm_slot_cache
);
1875 static inline struct mm_slot
*alloc_mm_slot(void)
1877 if (!mm_slot_cache
) /* initialization failed */
1879 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
1882 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
1884 kmem_cache_free(mm_slot_cache
, mm_slot
);
1887 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
1889 struct mm_slot
*mm_slot
;
1891 hash_for_each_possible(mm_slots_hash
, mm_slot
, hash
, (unsigned long)mm
)
1892 if (mm
== mm_slot
->mm
)
1898 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
1899 struct mm_slot
*mm_slot
)
1902 hash_add(mm_slots_hash
, &mm_slot
->hash
, (long)mm
);
1905 static inline int khugepaged_test_exit(struct mm_struct
*mm
)
1907 return atomic_read(&mm
->mm_users
) == 0;
1910 int __khugepaged_enter(struct mm_struct
*mm
)
1912 struct mm_slot
*mm_slot
;
1915 mm_slot
= alloc_mm_slot();
1919 /* __khugepaged_exit() must not run from under us */
1920 VM_BUG_ON_MM(khugepaged_test_exit(mm
), mm
);
1921 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE
, &mm
->flags
))) {
1922 free_mm_slot(mm_slot
);
1926 spin_lock(&khugepaged_mm_lock
);
1927 insert_to_mm_slots_hash(mm
, mm_slot
);
1929 * Insert just behind the scanning cursor, to let the area settle
1932 wakeup
= list_empty(&khugepaged_scan
.mm_head
);
1933 list_add_tail(&mm_slot
->mm_node
, &khugepaged_scan
.mm_head
);
1934 spin_unlock(&khugepaged_mm_lock
);
1936 atomic_inc(&mm
->mm_count
);
1938 wake_up_interruptible(&khugepaged_wait
);
1943 int khugepaged_enter_vma_merge(struct vm_area_struct
*vma
,
1944 unsigned long vm_flags
)
1946 unsigned long hstart
, hend
;
1949 * Not yet faulted in so we will register later in the
1950 * page fault if needed.
1954 /* khugepaged not yet working on file or special mappings */
1956 VM_BUG_ON_VMA(vm_flags
& VM_NO_THP
, vma
);
1957 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
1958 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
1960 return khugepaged_enter(vma
, vm_flags
);
1964 void __khugepaged_exit(struct mm_struct
*mm
)
1966 struct mm_slot
*mm_slot
;
1969 spin_lock(&khugepaged_mm_lock
);
1970 mm_slot
= get_mm_slot(mm
);
1971 if (mm_slot
&& khugepaged_scan
.mm_slot
!= mm_slot
) {
1972 hash_del(&mm_slot
->hash
);
1973 list_del(&mm_slot
->mm_node
);
1976 spin_unlock(&khugepaged_mm_lock
);
1979 clear_bit(MMF_VM_HUGEPAGE
, &mm
->flags
);
1980 free_mm_slot(mm_slot
);
1982 } else if (mm_slot
) {
1984 * This is required to serialize against
1985 * khugepaged_test_exit() (which is guaranteed to run
1986 * under mmap sem read mode). Stop here (after we
1987 * return all pagetables will be destroyed) until
1988 * khugepaged has finished working on the pagetables
1989 * under the mmap_sem.
1991 down_write(&mm
->mmap_sem
);
1992 up_write(&mm
->mmap_sem
);
1996 static void release_pte_page(struct page
*page
)
1998 /* 0 stands for page_is_file_cache(page) == false */
1999 dec_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
2001 putback_lru_page(page
);
2004 static void release_pte_pages(pte_t
*pte
, pte_t
*_pte
)
2006 while (--_pte
>= pte
) {
2007 pte_t pteval
= *_pte
;
2008 if (!pte_none(pteval
) && !is_zero_pfn(pte_pfn(pteval
)))
2009 release_pte_page(pte_page(pteval
));
2013 static int __collapse_huge_page_isolate(struct vm_area_struct
*vma
,
2014 unsigned long address
,
2017 struct page
*page
= NULL
;
2019 int none_or_zero
= 0, result
= 0;
2020 bool referenced
= false, writable
= false;
2022 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
2023 _pte
++, address
+= PAGE_SIZE
) {
2024 pte_t pteval
= *_pte
;
2025 if (pte_none(pteval
) || (pte_present(pteval
) &&
2026 is_zero_pfn(pte_pfn(pteval
)))) {
2027 if (!userfaultfd_armed(vma
) &&
2028 ++none_or_zero
<= khugepaged_max_ptes_none
) {
2031 result
= SCAN_EXCEED_NONE_PTE
;
2035 if (!pte_present(pteval
)) {
2036 result
= SCAN_PTE_NON_PRESENT
;
2039 page
= vm_normal_page(vma
, address
, pteval
);
2040 if (unlikely(!page
)) {
2041 result
= SCAN_PAGE_NULL
;
2045 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2046 VM_BUG_ON_PAGE(!PageAnon(page
), page
);
2047 VM_BUG_ON_PAGE(!PageSwapBacked(page
), page
);
2050 * We can do it before isolate_lru_page because the
2051 * page can't be freed from under us. NOTE: PG_lock
2052 * is needed to serialize against split_huge_page
2053 * when invoked from the VM.
2055 if (!trylock_page(page
)) {
2056 result
= SCAN_PAGE_LOCK
;
2061 * cannot use mapcount: can't collapse if there's a gup pin.
2062 * The page must only be referenced by the scanned process
2063 * and page swap cache.
2065 if (page_count(page
) != 1 + !!PageSwapCache(page
)) {
2067 result
= SCAN_PAGE_COUNT
;
2070 if (pte_write(pteval
)) {
2073 if (PageSwapCache(page
) && !reuse_swap_page(page
)) {
2075 result
= SCAN_SWAP_CACHE_PAGE
;
2079 * Page is not in the swap cache. It can be collapsed
2085 * Isolate the page to avoid collapsing an hugepage
2086 * currently in use by the VM.
2088 if (isolate_lru_page(page
)) {
2090 result
= SCAN_DEL_PAGE_LRU
;
2093 /* 0 stands for page_is_file_cache(page) == false */
2094 inc_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
2095 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2096 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2098 /* If there is no mapped pte young don't collapse the page */
2099 if (pte_young(pteval
) ||
2100 page_is_young(page
) || PageReferenced(page
) ||
2101 mmu_notifier_test_young(vma
->vm_mm
, address
))
2104 if (likely(writable
)) {
2105 if (likely(referenced
)) {
2106 result
= SCAN_SUCCEED
;
2107 trace_mm_collapse_huge_page_isolate(page
, none_or_zero
,
2108 referenced
, writable
, result
);
2112 result
= SCAN_PAGE_RO
;
2116 release_pte_pages(pte
, _pte
);
2117 trace_mm_collapse_huge_page_isolate(page
, none_or_zero
,
2118 referenced
, writable
, result
);
2122 static void __collapse_huge_page_copy(pte_t
*pte
, struct page
*page
,
2123 struct vm_area_struct
*vma
,
2124 unsigned long address
,
2128 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
; _pte
++) {
2129 pte_t pteval
= *_pte
;
2130 struct page
*src_page
;
2132 if (pte_none(pteval
) || is_zero_pfn(pte_pfn(pteval
))) {
2133 clear_user_highpage(page
, address
);
2134 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, 1);
2135 if (is_zero_pfn(pte_pfn(pteval
))) {
2137 * ptl mostly unnecessary.
2141 * paravirt calls inside pte_clear here are
2144 pte_clear(vma
->vm_mm
, address
, _pte
);
2148 src_page
= pte_page(pteval
);
2149 copy_user_highpage(page
, src_page
, address
, vma
);
2150 VM_BUG_ON_PAGE(page_mapcount(src_page
) != 1, src_page
);
2151 release_pte_page(src_page
);
2153 * ptl mostly unnecessary, but preempt has to
2154 * be disabled to update the per-cpu stats
2155 * inside page_remove_rmap().
2159 * paravirt calls inside pte_clear here are
2162 pte_clear(vma
->vm_mm
, address
, _pte
);
2163 page_remove_rmap(src_page
, false);
2165 free_page_and_swap_cache(src_page
);
2168 address
+= PAGE_SIZE
;
2173 static void khugepaged_alloc_sleep(void)
2177 add_wait_queue(&khugepaged_wait
, &wait
);
2178 freezable_schedule_timeout_interruptible(
2179 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs
));
2180 remove_wait_queue(&khugepaged_wait
, &wait
);
2183 static int khugepaged_node_load
[MAX_NUMNODES
];
2185 static bool khugepaged_scan_abort(int nid
)
2190 * If zone_reclaim_mode is disabled, then no extra effort is made to
2191 * allocate memory locally.
2193 if (!zone_reclaim_mode
)
2196 /* If there is a count for this node already, it must be acceptable */
2197 if (khugepaged_node_load
[nid
])
2200 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
2201 if (!khugepaged_node_load
[i
])
2203 if (node_distance(nid
, i
) > RECLAIM_DISTANCE
)
2210 static int khugepaged_find_target_node(void)
2212 static int last_khugepaged_target_node
= NUMA_NO_NODE
;
2213 int nid
, target_node
= 0, max_value
= 0;
2215 /* find first node with max normal pages hit */
2216 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
2217 if (khugepaged_node_load
[nid
] > max_value
) {
2218 max_value
= khugepaged_node_load
[nid
];
2222 /* do some balance if several nodes have the same hit record */
2223 if (target_node
<= last_khugepaged_target_node
)
2224 for (nid
= last_khugepaged_target_node
+ 1; nid
< MAX_NUMNODES
;
2226 if (max_value
== khugepaged_node_load
[nid
]) {
2231 last_khugepaged_target_node
= target_node
;
2235 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
2237 if (IS_ERR(*hpage
)) {
2243 khugepaged_alloc_sleep();
2244 } else if (*hpage
) {
2252 static struct page
*
2253 khugepaged_alloc_page(struct page
**hpage
, gfp_t gfp
, struct mm_struct
*mm
,
2254 unsigned long address
, int node
)
2256 VM_BUG_ON_PAGE(*hpage
, *hpage
);
2259 * Before allocating the hugepage, release the mmap_sem read lock.
2260 * The allocation can take potentially a long time if it involves
2261 * sync compaction, and we do not need to hold the mmap_sem during
2262 * that. We will recheck the vma after taking it again in write mode.
2264 up_read(&mm
->mmap_sem
);
2266 *hpage
= __alloc_pages_node(node
, gfp
, HPAGE_PMD_ORDER
);
2267 if (unlikely(!*hpage
)) {
2268 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
2269 *hpage
= ERR_PTR(-ENOMEM
);
2273 prep_transhuge_page(*hpage
);
2274 count_vm_event(THP_COLLAPSE_ALLOC
);
2278 static int khugepaged_find_target_node(void)
2283 static inline struct page
*alloc_khugepaged_hugepage(void)
2287 page
= alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
2290 prep_transhuge_page(page
);
2294 static struct page
*khugepaged_alloc_hugepage(bool *wait
)
2299 hpage
= alloc_khugepaged_hugepage();
2301 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
2306 khugepaged_alloc_sleep();
2308 count_vm_event(THP_COLLAPSE_ALLOC
);
2309 } while (unlikely(!hpage
) && likely(khugepaged_enabled()));
2314 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
2317 *hpage
= khugepaged_alloc_hugepage(wait
);
2319 if (unlikely(!*hpage
))
2325 static struct page
*
2326 khugepaged_alloc_page(struct page
**hpage
, gfp_t gfp
, struct mm_struct
*mm
,
2327 unsigned long address
, int node
)
2329 up_read(&mm
->mmap_sem
);
2336 static bool hugepage_vma_check(struct vm_area_struct
*vma
)
2338 if ((!(vma
->vm_flags
& VM_HUGEPAGE
) && !khugepaged_always()) ||
2339 (vma
->vm_flags
& VM_NOHUGEPAGE
))
2341 if (!vma
->anon_vma
|| vma
->vm_ops
)
2343 if (is_vma_temporary_stack(vma
))
2345 VM_BUG_ON_VMA(vma
->vm_flags
& VM_NO_THP
, vma
);
2349 static void collapse_huge_page(struct mm_struct
*mm
,
2350 unsigned long address
,
2351 struct page
**hpage
,
2352 struct vm_area_struct
*vma
,
2358 struct page
*new_page
;
2359 spinlock_t
*pmd_ptl
, *pte_ptl
;
2360 int isolated
= 0, result
= 0;
2361 unsigned long hstart
, hend
;
2362 struct mem_cgroup
*memcg
;
2363 unsigned long mmun_start
; /* For mmu_notifiers */
2364 unsigned long mmun_end
; /* For mmu_notifiers */
2367 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2369 /* Only allocate from the target node */
2370 gfp
= alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE
| __GFP_THISNODE
;
2372 /* release the mmap_sem read lock. */
2373 new_page
= khugepaged_alloc_page(hpage
, gfp
, mm
, address
, node
);
2375 result
= SCAN_ALLOC_HUGE_PAGE_FAIL
;
2379 if (unlikely(mem_cgroup_try_charge(new_page
, mm
, gfp
, &memcg
, true))) {
2380 result
= SCAN_CGROUP_CHARGE_FAIL
;
2385 * Prevent all access to pagetables with the exception of
2386 * gup_fast later hanlded by the ptep_clear_flush and the VM
2387 * handled by the anon_vma lock + PG_lock.
2389 down_write(&mm
->mmap_sem
);
2390 if (unlikely(khugepaged_test_exit(mm
))) {
2391 result
= SCAN_ANY_PROCESS
;
2395 vma
= find_vma(mm
, address
);
2397 result
= SCAN_VMA_NULL
;
2400 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2401 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2402 if (address
< hstart
|| address
+ HPAGE_PMD_SIZE
> hend
) {
2403 result
= SCAN_ADDRESS_RANGE
;
2406 if (!hugepage_vma_check(vma
)) {
2407 result
= SCAN_VMA_CHECK
;
2410 pmd
= mm_find_pmd(mm
, address
);
2412 result
= SCAN_PMD_NULL
;
2416 anon_vma_lock_write(vma
->anon_vma
);
2418 pte
= pte_offset_map(pmd
, address
);
2419 pte_ptl
= pte_lockptr(mm
, pmd
);
2421 mmun_start
= address
;
2422 mmun_end
= address
+ HPAGE_PMD_SIZE
;
2423 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2424 pmd_ptl
= pmd_lock(mm
, pmd
); /* probably unnecessary */
2426 * After this gup_fast can't run anymore. This also removes
2427 * any huge TLB entry from the CPU so we won't allow
2428 * huge and small TLB entries for the same virtual address
2429 * to avoid the risk of CPU bugs in that area.
2431 _pmd
= pmdp_collapse_flush(vma
, address
, pmd
);
2432 spin_unlock(pmd_ptl
);
2433 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2436 isolated
= __collapse_huge_page_isolate(vma
, address
, pte
);
2437 spin_unlock(pte_ptl
);
2439 if (unlikely(!isolated
)) {
2442 BUG_ON(!pmd_none(*pmd
));
2444 * We can only use set_pmd_at when establishing
2445 * hugepmds and never for establishing regular pmds that
2446 * points to regular pagetables. Use pmd_populate for that
2448 pmd_populate(mm
, pmd
, pmd_pgtable(_pmd
));
2449 spin_unlock(pmd_ptl
);
2450 anon_vma_unlock_write(vma
->anon_vma
);
2456 * All pages are isolated and locked so anon_vma rmap
2457 * can't run anymore.
2459 anon_vma_unlock_write(vma
->anon_vma
);
2461 __collapse_huge_page_copy(pte
, new_page
, vma
, address
, pte_ptl
);
2463 __SetPageUptodate(new_page
);
2464 pgtable
= pmd_pgtable(_pmd
);
2466 _pmd
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
2467 _pmd
= maybe_pmd_mkwrite(pmd_mkdirty(_pmd
), vma
);
2470 * spin_lock() below is not the equivalent of smp_wmb(), so
2471 * this is needed to avoid the copy_huge_page writes to become
2472 * visible after the set_pmd_at() write.
2477 BUG_ON(!pmd_none(*pmd
));
2478 page_add_new_anon_rmap(new_page
, vma
, address
, true);
2479 mem_cgroup_commit_charge(new_page
, memcg
, false, true);
2480 lru_cache_add_active_or_unevictable(new_page
, vma
);
2481 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
2482 set_pmd_at(mm
, address
, pmd
, _pmd
);
2483 update_mmu_cache_pmd(vma
, address
, pmd
);
2484 spin_unlock(pmd_ptl
);
2488 khugepaged_pages_collapsed
++;
2489 result
= SCAN_SUCCEED
;
2491 up_write(&mm
->mmap_sem
);
2492 trace_mm_collapse_huge_page(mm
, isolated
, result
);
2496 trace_mm_collapse_huge_page(mm
, isolated
, result
);
2499 mem_cgroup_cancel_charge(new_page
, memcg
, true);
2503 static int khugepaged_scan_pmd(struct mm_struct
*mm
,
2504 struct vm_area_struct
*vma
,
2505 unsigned long address
,
2506 struct page
**hpage
)
2510 int ret
= 0, none_or_zero
= 0, result
= 0;
2511 struct page
*page
= NULL
;
2512 unsigned long _address
;
2514 int node
= NUMA_NO_NODE
;
2515 bool writable
= false, referenced
= false;
2517 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2519 pmd
= mm_find_pmd(mm
, address
);
2521 result
= SCAN_PMD_NULL
;
2525 memset(khugepaged_node_load
, 0, sizeof(khugepaged_node_load
));
2526 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2527 for (_address
= address
, _pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
2528 _pte
++, _address
+= PAGE_SIZE
) {
2529 pte_t pteval
= *_pte
;
2530 if (pte_none(pteval
) || is_zero_pfn(pte_pfn(pteval
))) {
2531 if (!userfaultfd_armed(vma
) &&
2532 ++none_or_zero
<= khugepaged_max_ptes_none
) {
2535 result
= SCAN_EXCEED_NONE_PTE
;
2539 if (!pte_present(pteval
)) {
2540 result
= SCAN_PTE_NON_PRESENT
;
2543 if (pte_write(pteval
))
2546 page
= vm_normal_page(vma
, _address
, pteval
);
2547 if (unlikely(!page
)) {
2548 result
= SCAN_PAGE_NULL
;
2552 /* TODO: teach khugepaged to collapse THP mapped with pte */
2553 if (PageCompound(page
)) {
2554 result
= SCAN_PAGE_COMPOUND
;
2559 * Record which node the original page is from and save this
2560 * information to khugepaged_node_load[].
2561 * Khupaged will allocate hugepage from the node has the max
2564 node
= page_to_nid(page
);
2565 if (khugepaged_scan_abort(node
)) {
2566 result
= SCAN_SCAN_ABORT
;
2569 khugepaged_node_load
[node
]++;
2570 if (!PageLRU(page
)) {
2571 result
= SCAN_SCAN_ABORT
;
2574 if (PageLocked(page
)) {
2575 result
= SCAN_PAGE_LOCK
;
2578 if (!PageAnon(page
)) {
2579 result
= SCAN_PAGE_ANON
;
2584 * cannot use mapcount: can't collapse if there's a gup pin.
2585 * The page must only be referenced by the scanned process
2586 * and page swap cache.
2588 if (page_count(page
) != 1 + !!PageSwapCache(page
)) {
2589 result
= SCAN_PAGE_COUNT
;
2592 if (pte_young(pteval
) ||
2593 page_is_young(page
) || PageReferenced(page
) ||
2594 mmu_notifier_test_young(vma
->vm_mm
, address
))
2599 result
= SCAN_SUCCEED
;
2602 result
= SCAN_NO_REFERENCED_PAGE
;
2605 result
= SCAN_PAGE_RO
;
2608 pte_unmap_unlock(pte
, ptl
);
2610 node
= khugepaged_find_target_node();
2611 /* collapse_huge_page will return with the mmap_sem released */
2612 collapse_huge_page(mm
, address
, hpage
, vma
, node
);
2615 trace_mm_khugepaged_scan_pmd(mm
, page
, writable
, referenced
,
2616 none_or_zero
, result
);
2620 static void collect_mm_slot(struct mm_slot
*mm_slot
)
2622 struct mm_struct
*mm
= mm_slot
->mm
;
2624 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2626 if (khugepaged_test_exit(mm
)) {
2628 hash_del(&mm_slot
->hash
);
2629 list_del(&mm_slot
->mm_node
);
2632 * Not strictly needed because the mm exited already.
2634 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2637 /* khugepaged_mm_lock actually not necessary for the below */
2638 free_mm_slot(mm_slot
);
2643 static unsigned int khugepaged_scan_mm_slot(unsigned int pages
,
2644 struct page
**hpage
)
2645 __releases(&khugepaged_mm_lock
)
2646 __acquires(&khugepaged_mm_lock
)
2648 struct mm_slot
*mm_slot
;
2649 struct mm_struct
*mm
;
2650 struct vm_area_struct
*vma
;
2654 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2656 if (khugepaged_scan
.mm_slot
)
2657 mm_slot
= khugepaged_scan
.mm_slot
;
2659 mm_slot
= list_entry(khugepaged_scan
.mm_head
.next
,
2660 struct mm_slot
, mm_node
);
2661 khugepaged_scan
.address
= 0;
2662 khugepaged_scan
.mm_slot
= mm_slot
;
2664 spin_unlock(&khugepaged_mm_lock
);
2667 down_read(&mm
->mmap_sem
);
2668 if (unlikely(khugepaged_test_exit(mm
)))
2671 vma
= find_vma(mm
, khugepaged_scan
.address
);
2674 for (; vma
; vma
= vma
->vm_next
) {
2675 unsigned long hstart
, hend
;
2678 if (unlikely(khugepaged_test_exit(mm
))) {
2682 if (!hugepage_vma_check(vma
)) {
2687 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2688 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2691 if (khugepaged_scan
.address
> hend
)
2693 if (khugepaged_scan
.address
< hstart
)
2694 khugepaged_scan
.address
= hstart
;
2695 VM_BUG_ON(khugepaged_scan
.address
& ~HPAGE_PMD_MASK
);
2697 while (khugepaged_scan
.address
< hend
) {
2700 if (unlikely(khugepaged_test_exit(mm
)))
2701 goto breakouterloop
;
2703 VM_BUG_ON(khugepaged_scan
.address
< hstart
||
2704 khugepaged_scan
.address
+ HPAGE_PMD_SIZE
>
2706 ret
= khugepaged_scan_pmd(mm
, vma
,
2707 khugepaged_scan
.address
,
2709 /* move to next address */
2710 khugepaged_scan
.address
+= HPAGE_PMD_SIZE
;
2711 progress
+= HPAGE_PMD_NR
;
2713 /* we released mmap_sem so break loop */
2714 goto breakouterloop_mmap_sem
;
2715 if (progress
>= pages
)
2716 goto breakouterloop
;
2720 up_read(&mm
->mmap_sem
); /* exit_mmap will destroy ptes after this */
2721 breakouterloop_mmap_sem
:
2723 spin_lock(&khugepaged_mm_lock
);
2724 VM_BUG_ON(khugepaged_scan
.mm_slot
!= mm_slot
);
2726 * Release the current mm_slot if this mm is about to die, or
2727 * if we scanned all vmas of this mm.
2729 if (khugepaged_test_exit(mm
) || !vma
) {
2731 * Make sure that if mm_users is reaching zero while
2732 * khugepaged runs here, khugepaged_exit will find
2733 * mm_slot not pointing to the exiting mm.
2735 if (mm_slot
->mm_node
.next
!= &khugepaged_scan
.mm_head
) {
2736 khugepaged_scan
.mm_slot
= list_entry(
2737 mm_slot
->mm_node
.next
,
2738 struct mm_slot
, mm_node
);
2739 khugepaged_scan
.address
= 0;
2741 khugepaged_scan
.mm_slot
= NULL
;
2742 khugepaged_full_scans
++;
2745 collect_mm_slot(mm_slot
);
2751 static int khugepaged_has_work(void)
2753 return !list_empty(&khugepaged_scan
.mm_head
) &&
2754 khugepaged_enabled();
2757 static int khugepaged_wait_event(void)
2759 return !list_empty(&khugepaged_scan
.mm_head
) ||
2760 kthread_should_stop();
2763 static void khugepaged_do_scan(void)
2765 struct page
*hpage
= NULL
;
2766 unsigned int progress
= 0, pass_through_head
= 0;
2767 unsigned int pages
= khugepaged_pages_to_scan
;
2770 barrier(); /* write khugepaged_pages_to_scan to local stack */
2772 while (progress
< pages
) {
2773 if (!khugepaged_prealloc_page(&hpage
, &wait
))
2778 if (unlikely(kthread_should_stop() || try_to_freeze()))
2781 spin_lock(&khugepaged_mm_lock
);
2782 if (!khugepaged_scan
.mm_slot
)
2783 pass_through_head
++;
2784 if (khugepaged_has_work() &&
2785 pass_through_head
< 2)
2786 progress
+= khugepaged_scan_mm_slot(pages
- progress
,
2790 spin_unlock(&khugepaged_mm_lock
);
2793 if (!IS_ERR_OR_NULL(hpage
))
2797 static void khugepaged_wait_work(void)
2799 if (khugepaged_has_work()) {
2800 if (!khugepaged_scan_sleep_millisecs
)
2803 wait_event_freezable_timeout(khugepaged_wait
,
2804 kthread_should_stop(),
2805 msecs_to_jiffies(khugepaged_scan_sleep_millisecs
));
2809 if (khugepaged_enabled())
2810 wait_event_freezable(khugepaged_wait
, khugepaged_wait_event());
2813 static int khugepaged(void *none
)
2815 struct mm_slot
*mm_slot
;
2818 set_user_nice(current
, MAX_NICE
);
2820 while (!kthread_should_stop()) {
2821 khugepaged_do_scan();
2822 khugepaged_wait_work();
2825 spin_lock(&khugepaged_mm_lock
);
2826 mm_slot
= khugepaged_scan
.mm_slot
;
2827 khugepaged_scan
.mm_slot
= NULL
;
2829 collect_mm_slot(mm_slot
);
2830 spin_unlock(&khugepaged_mm_lock
);
2834 static void __split_huge_zero_page_pmd(struct vm_area_struct
*vma
,
2835 unsigned long haddr
, pmd_t
*pmd
)
2837 struct mm_struct
*mm
= vma
->vm_mm
;
2842 /* leave pmd empty until pte is filled */
2843 pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
2845 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2846 pmd_populate(mm
, &_pmd
, pgtable
);
2848 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
2850 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
2851 entry
= pte_mkspecial(entry
);
2852 pte
= pte_offset_map(&_pmd
, haddr
);
2853 VM_BUG_ON(!pte_none(*pte
));
2854 set_pte_at(mm
, haddr
, pte
, entry
);
2857 smp_wmb(); /* make pte visible before pmd */
2858 pmd_populate(mm
, pmd
, pgtable
);
2859 put_huge_zero_page();
2862 static void __split_huge_pmd_locked(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2863 unsigned long haddr
, bool freeze
)
2865 struct mm_struct
*mm
= vma
->vm_mm
;
2869 bool young
, write
, dirty
;
2873 VM_BUG_ON(haddr
& ~HPAGE_PMD_MASK
);
2874 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
2875 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PMD_SIZE
, vma
);
2876 VM_BUG_ON(!pmd_trans_huge(*pmd
) && !pmd_devmap(*pmd
));
2878 count_vm_event(THP_SPLIT_PMD
);
2880 if (vma_is_dax(vma
)) {
2881 pmd_t _pmd
= pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
2882 if (is_huge_zero_pmd(_pmd
))
2883 put_huge_zero_page();
2885 } else if (is_huge_zero_pmd(*pmd
)) {
2886 return __split_huge_zero_page_pmd(vma
, haddr
, pmd
);
2889 page
= pmd_page(*pmd
);
2890 VM_BUG_ON_PAGE(!page_count(page
), page
);
2891 atomic_add(HPAGE_PMD_NR
- 1, &page
->_count
);
2892 write
= pmd_write(*pmd
);
2893 young
= pmd_young(*pmd
);
2894 dirty
= pmd_dirty(*pmd
);
2896 pmdp_huge_split_prepare(vma
, haddr
, pmd
);
2897 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2898 pmd_populate(mm
, &_pmd
, pgtable
);
2900 for (i
= 0, addr
= haddr
; i
< HPAGE_PMD_NR
; i
++, addr
+= PAGE_SIZE
) {
2903 * Note that NUMA hinting access restrictions are not
2904 * transferred to avoid any possibility of altering
2905 * permissions across VMAs.
2908 swp_entry_t swp_entry
;
2909 swp_entry
= make_migration_entry(page
+ i
, write
);
2910 entry
= swp_entry_to_pte(swp_entry
);
2912 entry
= mk_pte(page
+ i
, vma
->vm_page_prot
);
2913 entry
= maybe_mkwrite(entry
, vma
);
2915 entry
= pte_wrprotect(entry
);
2917 entry
= pte_mkold(entry
);
2920 SetPageDirty(page
+ i
);
2921 pte
= pte_offset_map(&_pmd
, addr
);
2922 BUG_ON(!pte_none(*pte
));
2923 set_pte_at(mm
, addr
, pte
, entry
);
2924 atomic_inc(&page
[i
]._mapcount
);
2929 * Set PG_double_map before dropping compound_mapcount to avoid
2930 * false-negative page_mapped().
2932 if (compound_mapcount(page
) > 1 && !TestSetPageDoubleMap(page
)) {
2933 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2934 atomic_inc(&page
[i
]._mapcount
);
2937 if (atomic_add_negative(-1, compound_mapcount_ptr(page
))) {
2938 /* Last compound_mapcount is gone. */
2939 __dec_zone_page_state(page
, NR_ANON_TRANSPARENT_HUGEPAGES
);
2940 if (TestClearPageDoubleMap(page
)) {
2941 /* No need in mapcount reference anymore */
2942 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2943 atomic_dec(&page
[i
]._mapcount
);
2947 smp_wmb(); /* make pte visible before pmd */
2949 * Up to this point the pmd is present and huge and userland has the
2950 * whole access to the hugepage during the split (which happens in
2951 * place). If we overwrite the pmd with the not-huge version pointing
2952 * to the pte here (which of course we could if all CPUs were bug
2953 * free), userland could trigger a small page size TLB miss on the
2954 * small sized TLB while the hugepage TLB entry is still established in
2955 * the huge TLB. Some CPU doesn't like that.
2956 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2957 * 383 on page 93. Intel should be safe but is also warns that it's
2958 * only safe if the permission and cache attributes of the two entries
2959 * loaded in the two TLB is identical (which should be the case here).
2960 * But it is generally safer to never allow small and huge TLB entries
2961 * for the same virtual address to be loaded simultaneously. So instead
2962 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2963 * current pmd notpresent (atomically because here the pmd_trans_huge
2964 * and pmd_trans_splitting must remain set at all times on the pmd
2965 * until the split is complete for this pmd), then we flush the SMP TLB
2966 * and finally we write the non-huge version of the pmd entry with
2969 pmdp_invalidate(vma
, haddr
, pmd
);
2970 pmd_populate(mm
, pmd
, pgtable
);
2973 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
2974 page_remove_rmap(page
+ i
, false);
2980 void __split_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2981 unsigned long address
)
2984 struct mm_struct
*mm
= vma
->vm_mm
;
2985 struct page
*page
= NULL
;
2986 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
2988 mmu_notifier_invalidate_range_start(mm
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
2989 ptl
= pmd_lock(mm
, pmd
);
2990 if (pmd_trans_huge(*pmd
)) {
2991 page
= pmd_page(*pmd
);
2992 if (PageMlocked(page
))
2996 } else if (!pmd_devmap(*pmd
))
2998 __split_huge_pmd_locked(vma
, pmd
, haddr
, false);
3001 mmu_notifier_invalidate_range_end(mm
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
3004 munlock_vma_page(page
);
3010 static void split_huge_pmd_address(struct vm_area_struct
*vma
,
3011 unsigned long address
)
3017 VM_BUG_ON(!(address
& ~HPAGE_PMD_MASK
));
3019 pgd
= pgd_offset(vma
->vm_mm
, address
);
3020 if (!pgd_present(*pgd
))
3023 pud
= pud_offset(pgd
, address
);
3024 if (!pud_present(*pud
))
3027 pmd
= pmd_offset(pud
, address
);
3028 if (!pmd_present(*pmd
) || (!pmd_trans_huge(*pmd
) && !pmd_devmap(*pmd
)))
3031 * Caller holds the mmap_sem write mode, so a huge pmd cannot
3032 * materialize from under us.
3034 split_huge_pmd(vma
, pmd
, address
);
3037 void vma_adjust_trans_huge(struct vm_area_struct
*vma
,
3038 unsigned long start
,
3043 * If the new start address isn't hpage aligned and it could
3044 * previously contain an hugepage: check if we need to split
3047 if (start
& ~HPAGE_PMD_MASK
&&
3048 (start
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
3049 (start
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
3050 split_huge_pmd_address(vma
, start
);
3053 * If the new end address isn't hpage aligned and it could
3054 * previously contain an hugepage: check if we need to split
3057 if (end
& ~HPAGE_PMD_MASK
&&
3058 (end
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
3059 (end
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
3060 split_huge_pmd_address(vma
, end
);
3063 * If we're also updating the vma->vm_next->vm_start, if the new
3064 * vm_next->vm_start isn't page aligned and it could previously
3065 * contain an hugepage: check if we need to split an huge pmd.
3067 if (adjust_next
> 0) {
3068 struct vm_area_struct
*next
= vma
->vm_next
;
3069 unsigned long nstart
= next
->vm_start
;
3070 nstart
+= adjust_next
<< PAGE_SHIFT
;
3071 if (nstart
& ~HPAGE_PMD_MASK
&&
3072 (nstart
& HPAGE_PMD_MASK
) >= next
->vm_start
&&
3073 (nstart
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= next
->vm_end
)
3074 split_huge_pmd_address(next
, nstart
);
3078 static void freeze_page_vma(struct vm_area_struct
*vma
, struct page
*page
,
3079 unsigned long address
)
3081 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
3087 int i
, nr
= HPAGE_PMD_NR
;
3089 /* Skip pages which doesn't belong to the VMA */
3090 if (address
< vma
->vm_start
) {
3091 int off
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
3094 address
= vma
->vm_start
;
3097 pgd
= pgd_offset(vma
->vm_mm
, address
);
3098 if (!pgd_present(*pgd
))
3100 pud
= pud_offset(pgd
, address
);
3101 if (!pud_present(*pud
))
3103 pmd
= pmd_offset(pud
, address
);
3104 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
3105 if (!pmd_present(*pmd
)) {
3109 if (pmd_trans_huge(*pmd
)) {
3110 if (page
== pmd_page(*pmd
))
3111 __split_huge_pmd_locked(vma
, pmd
, haddr
, true);
3117 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, address
, &ptl
);
3118 for (i
= 0; i
< nr
; i
++, address
+= PAGE_SIZE
, page
++, pte
++) {
3119 pte_t entry
, swp_pte
;
3120 swp_entry_t swp_entry
;
3123 * We've just crossed page table boundary: need to map next one.
3124 * It can happen if THP was mremaped to non PMD-aligned address.
3126 if (unlikely(address
== haddr
+ HPAGE_PMD_SIZE
)) {
3127 pte_unmap_unlock(pte
- 1, ptl
);
3128 pmd
= mm_find_pmd(vma
->vm_mm
, address
);
3131 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
,
3135 if (!pte_present(*pte
))
3137 if (page_to_pfn(page
) != pte_pfn(*pte
))
3139 flush_cache_page(vma
, address
, page_to_pfn(page
));
3140 entry
= ptep_clear_flush(vma
, address
, pte
);
3141 if (pte_dirty(entry
))
3143 swp_entry
= make_migration_entry(page
, pte_write(entry
));
3144 swp_pte
= swp_entry_to_pte(swp_entry
);
3145 if (pte_soft_dirty(entry
))
3146 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
3147 set_pte_at(vma
->vm_mm
, address
, pte
, swp_pte
);
3148 page_remove_rmap(page
, false);
3151 pte_unmap_unlock(pte
- 1, ptl
);
3154 static void freeze_page(struct anon_vma
*anon_vma
, struct page
*page
)
3156 struct anon_vma_chain
*avc
;
3157 pgoff_t pgoff
= page_to_pgoff(page
);
3159 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3161 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
,
3162 pgoff
+ HPAGE_PMD_NR
- 1) {
3163 unsigned long address
= __vma_address(page
, avc
->vma
);
3165 mmu_notifier_invalidate_range_start(avc
->vma
->vm_mm
,
3166 address
, address
+ HPAGE_PMD_SIZE
);
3167 freeze_page_vma(avc
->vma
, page
, address
);
3168 mmu_notifier_invalidate_range_end(avc
->vma
->vm_mm
,
3169 address
, address
+ HPAGE_PMD_SIZE
);
3173 static void unfreeze_page_vma(struct vm_area_struct
*vma
, struct page
*page
,
3174 unsigned long address
)
3179 swp_entry_t swp_entry
;
3180 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
3181 int i
, nr
= HPAGE_PMD_NR
;
3183 /* Skip pages which doesn't belong to the VMA */
3184 if (address
< vma
->vm_start
) {
3185 int off
= (vma
->vm_start
- address
) >> PAGE_SHIFT
;
3188 address
= vma
->vm_start
;
3191 pmd
= mm_find_pmd(vma
->vm_mm
, address
);
3195 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, address
, &ptl
);
3196 for (i
= 0; i
< nr
; i
++, address
+= PAGE_SIZE
, page
++, pte
++) {
3198 * We've just crossed page table boundary: need to map next one.
3199 * It can happen if THP was mremaped to non-PMD aligned address.
3201 if (unlikely(address
== haddr
+ HPAGE_PMD_SIZE
)) {
3202 pte_unmap_unlock(pte
- 1, ptl
);
3203 pmd
= mm_find_pmd(vma
->vm_mm
, address
);
3206 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
,
3210 if (!is_swap_pte(*pte
))
3213 swp_entry
= pte_to_swp_entry(*pte
);
3214 if (!is_migration_entry(swp_entry
))
3216 if (migration_entry_to_page(swp_entry
) != page
)
3220 page_add_anon_rmap(page
, vma
, address
, false);
3222 entry
= pte_mkold(mk_pte(page
, vma
->vm_page_prot
));
3223 if (PageDirty(page
))
3224 entry
= pte_mkdirty(entry
);
3225 if (is_write_migration_entry(swp_entry
))
3226 entry
= maybe_mkwrite(entry
, vma
);
3228 flush_dcache_page(page
);
3229 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
3231 /* No need to invalidate - it was non-present before */
3232 update_mmu_cache(vma
, address
, pte
);
3234 pte_unmap_unlock(pte
- 1, ptl
);
3237 static void unfreeze_page(struct anon_vma
*anon_vma
, struct page
*page
)
3239 struct anon_vma_chain
*avc
;
3240 pgoff_t pgoff
= page_to_pgoff(page
);
3242 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
,
3243 pgoff
, pgoff
+ HPAGE_PMD_NR
- 1) {
3244 unsigned long address
= __vma_address(page
, avc
->vma
);
3246 mmu_notifier_invalidate_range_start(avc
->vma
->vm_mm
,
3247 address
, address
+ HPAGE_PMD_SIZE
);
3248 unfreeze_page_vma(avc
->vma
, page
, address
);
3249 mmu_notifier_invalidate_range_end(avc
->vma
->vm_mm
,
3250 address
, address
+ HPAGE_PMD_SIZE
);
3254 static void __split_huge_page_tail(struct page
*head
, int tail
,
3255 struct lruvec
*lruvec
, struct list_head
*list
)
3257 struct page
*page_tail
= head
+ tail
;
3259 VM_BUG_ON_PAGE(atomic_read(&page_tail
->_mapcount
) != -1, page_tail
);
3260 VM_BUG_ON_PAGE(atomic_read(&page_tail
->_count
) != 0, page_tail
);
3263 * tail_page->_count is zero and not changing from under us. But
3264 * get_page_unless_zero() may be running from under us on the
3265 * tail_page. If we used atomic_set() below instead of atomic_inc(), we
3266 * would then run atomic_set() concurrently with
3267 * get_page_unless_zero(), and atomic_set() is implemented in C not
3268 * using locked ops. spin_unlock on x86 sometime uses locked ops
3269 * because of PPro errata 66, 92, so unless somebody can guarantee
3270 * atomic_set() here would be safe on all archs (and not only on x86),
3271 * it's safer to use atomic_inc().
3273 atomic_inc(&page_tail
->_count
);
3275 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
3276 page_tail
->flags
|= (head
->flags
&
3277 ((1L << PG_referenced
) |
3278 (1L << PG_swapbacked
) |
3279 (1L << PG_mlocked
) |
3280 (1L << PG_uptodate
) |
3283 (1L << PG_unevictable
) |
3287 * After clearing PageTail the gup refcount can be released.
3288 * Page flags also must be visible before we make the page non-compound.
3292 clear_compound_head(page_tail
);
3294 if (page_is_young(head
))
3295 set_page_young(page_tail
);
3296 if (page_is_idle(head
))
3297 set_page_idle(page_tail
);
3299 /* ->mapping in first tail page is compound_mapcount */
3300 VM_BUG_ON_PAGE(tail
> 2 && page_tail
->mapping
!= TAIL_MAPPING
,
3302 page_tail
->mapping
= head
->mapping
;
3304 page_tail
->index
= head
->index
+ tail
;
3305 page_cpupid_xchg_last(page_tail
, page_cpupid_last(head
));
3306 lru_add_page_tail(head
, page_tail
, lruvec
, list
);
3309 static void __split_huge_page(struct page
*page
, struct list_head
*list
)
3311 struct page
*head
= compound_head(page
);
3312 struct zone
*zone
= page_zone(head
);
3313 struct lruvec
*lruvec
;
3316 /* prevent PageLRU to go away from under us, and freeze lru stats */
3317 spin_lock_irq(&zone
->lru_lock
);
3318 lruvec
= mem_cgroup_page_lruvec(head
, zone
);
3320 /* complete memcg works before add pages to LRU */
3321 mem_cgroup_split_huge_fixup(head
);
3323 for (i
= HPAGE_PMD_NR
- 1; i
>= 1; i
--)
3324 __split_huge_page_tail(head
, i
, lruvec
, list
);
3326 ClearPageCompound(head
);
3327 spin_unlock_irq(&zone
->lru_lock
);
3329 unfreeze_page(page_anon_vma(head
), head
);
3331 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
3332 struct page
*subpage
= head
+ i
;
3333 if (subpage
== page
)
3335 unlock_page(subpage
);
3338 * Subpages may be freed if there wasn't any mapping
3339 * like if add_to_swap() is running on a lru page that
3340 * had its mapping zapped. And freeing these pages
3341 * requires taking the lru_lock so we do the put_page
3342 * of the tail pages after the split is complete.
3348 int total_mapcount(struct page
*page
)
3352 VM_BUG_ON_PAGE(PageTail(page
), page
);
3354 if (likely(!PageCompound(page
)))
3355 return atomic_read(&page
->_mapcount
) + 1;
3357 ret
= compound_mapcount(page
);
3360 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
3361 ret
+= atomic_read(&page
[i
]._mapcount
) + 1;
3362 if (PageDoubleMap(page
))
3363 ret
-= HPAGE_PMD_NR
;
3368 * This function splits huge page into normal pages. @page can point to any
3369 * subpage of huge page to split. Split doesn't change the position of @page.
3371 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3372 * The huge page must be locked.
3374 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3376 * Both head page and tail pages will inherit mapping, flags, and so on from
3379 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3380 * they are not mapped.
3382 * Returns 0 if the hugepage is split successfully.
3383 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3386 int split_huge_page_to_list(struct page
*page
, struct list_head
*list
)
3388 struct page
*head
= compound_head(page
);
3389 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(head
));
3390 struct anon_vma
*anon_vma
;
3391 int count
, mapcount
, ret
;
3393 unsigned long flags
;
3395 VM_BUG_ON_PAGE(is_huge_zero_page(page
), page
);
3396 VM_BUG_ON_PAGE(!PageAnon(page
), page
);
3397 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
3398 VM_BUG_ON_PAGE(!PageSwapBacked(page
), page
);
3399 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
3402 * The caller does not necessarily hold an mmap_sem that would prevent
3403 * the anon_vma disappearing so we first we take a reference to it
3404 * and then lock the anon_vma for write. This is similar to
3405 * page_lock_anon_vma_read except the write lock is taken to serialise
3406 * against parallel split or collapse operations.
3408 anon_vma
= page_get_anon_vma(head
);
3413 anon_vma_lock_write(anon_vma
);
3416 * Racy check if we can split the page, before freeze_page() will
3419 if (total_mapcount(head
) != page_count(head
) - 1) {
3424 mlocked
= PageMlocked(page
);
3425 freeze_page(anon_vma
, head
);
3426 VM_BUG_ON_PAGE(compound_mapcount(head
), head
);
3428 /* Make sure the page is not on per-CPU pagevec as it takes pin */
3432 /* Prevent deferred_split_scan() touching ->_count */
3433 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
3434 count
= page_count(head
);
3435 mapcount
= total_mapcount(head
);
3436 if (!mapcount
&& count
== 1) {
3437 if (!list_empty(page_deferred_list(head
))) {
3438 pgdata
->split_queue_len
--;
3439 list_del(page_deferred_list(head
));
3441 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
3442 __split_huge_page(page
, list
);
3444 } else if (IS_ENABLED(CONFIG_DEBUG_VM
) && mapcount
) {
3445 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
3446 pr_alert("total_mapcount: %u, page_count(): %u\n",
3449 dump_page(head
, NULL
);
3450 dump_page(page
, "total_mapcount(head) > 0");
3453 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
3454 unfreeze_page(anon_vma
, head
);
3459 anon_vma_unlock_write(anon_vma
);
3460 put_anon_vma(anon_vma
);
3462 count_vm_event(!ret
? THP_SPLIT_PAGE
: THP_SPLIT_PAGE_FAILED
);
3466 void free_transhuge_page(struct page
*page
)
3468 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(page
));
3469 unsigned long flags
;
3471 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
3472 if (!list_empty(page_deferred_list(page
))) {
3473 pgdata
->split_queue_len
--;
3474 list_del(page_deferred_list(page
));
3476 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
3477 free_compound_page(page
);
3480 void deferred_split_huge_page(struct page
*page
)
3482 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(page
));
3483 unsigned long flags
;
3485 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3487 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
3488 if (list_empty(page_deferred_list(page
))) {
3489 count_vm_event(THP_DEFERRED_SPLIT_PAGE
);
3490 list_add_tail(page_deferred_list(page
), &pgdata
->split_queue
);
3491 pgdata
->split_queue_len
++;
3493 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
3496 static unsigned long deferred_split_count(struct shrinker
*shrink
,
3497 struct shrink_control
*sc
)
3499 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
3500 return ACCESS_ONCE(pgdata
->split_queue_len
);
3503 static unsigned long deferred_split_scan(struct shrinker
*shrink
,
3504 struct shrink_control
*sc
)
3506 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
3507 unsigned long flags
;
3508 LIST_HEAD(list
), *pos
, *next
;
3512 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
3513 /* Take pin on all head pages to avoid freeing them under us */
3514 list_for_each_safe(pos
, next
, &pgdata
->split_queue
) {
3515 page
= list_entry((void *)pos
, struct page
, mapping
);
3516 page
= compound_head(page
);
3517 if (get_page_unless_zero(page
)) {
3518 list_move(page_deferred_list(page
), &list
);
3520 /* We lost race with put_compound_page() */
3521 list_del_init(page_deferred_list(page
));
3522 pgdata
->split_queue_len
--;
3524 if (!--sc
->nr_to_scan
)
3527 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
3529 list_for_each_safe(pos
, next
, &list
) {
3530 page
= list_entry((void *)pos
, struct page
, mapping
);
3532 /* split_huge_page() removes page from list on success */
3533 if (!split_huge_page(page
))
3539 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
3540 list_splice_tail(&list
, &pgdata
->split_queue
);
3541 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
3544 * Stop shrinker if we didn't split any page, but the queue is empty.
3545 * This can happen if pages were freed under us.
3547 if (!split
&& list_empty(&pgdata
->split_queue
))
3552 static struct shrinker deferred_split_shrinker
= {
3553 .count_objects
= deferred_split_count
,
3554 .scan_objects
= deferred_split_scan
,
3555 .seeks
= DEFAULT_SEEKS
,
3556 .flags
= SHRINKER_NUMA_AWARE
,
3559 #ifdef CONFIG_DEBUG_FS
3560 static int split_huge_pages_set(void *data
, u64 val
)
3564 unsigned long pfn
, max_zone_pfn
;
3565 unsigned long total
= 0, split
= 0;
3570 for_each_populated_zone(zone
) {
3571 max_zone_pfn
= zone_end_pfn(zone
);
3572 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++) {
3573 if (!pfn_valid(pfn
))
3576 page
= pfn_to_page(pfn
);
3577 if (!get_page_unless_zero(page
))
3580 if (zone
!= page_zone(page
))
3583 if (!PageHead(page
) || !PageAnon(page
) ||
3589 if (!split_huge_page(page
))
3597 pr_info("%lu of %lu THP split", split
, total
);
3601 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops
, NULL
, split_huge_pages_set
,
3604 static int __init
split_huge_pages_debugfs(void)
3608 ret
= debugfs_create_file("split_huge_pages", 0644, NULL
, NULL
,
3609 &split_huge_pages_fops
);
3611 pr_warn("Failed to create split_huge_pages in debugfs");
3614 late_initcall(split_huge_pages_debugfs
);