2 * Copyright (C) 2009 Red Hat, Inc.
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
9 #include <linux/sched.h>
10 #include <linux/highmem.h>
11 #include <linux/hugetlb.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/rmap.h>
14 #include <linux/swap.h>
15 #include <linux/shrinker.h>
16 #include <linux/mm_inline.h>
17 #include <linux/kthread.h>
18 #include <linux/khugepaged.h>
19 #include <linux/freezer.h>
20 #include <linux/mman.h>
21 #include <linux/pagemap.h>
22 #include <linux/migrate.h>
23 #include <linux/hashtable.h>
26 #include <asm/pgalloc.h>
30 * By default transparent hugepage support is enabled for all mappings
31 * and khugepaged scans all mappings. Defrag is only invoked by
32 * khugepaged hugepage allocations and by page faults inside
33 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
36 unsigned long transparent_hugepage_flags __read_mostly
=
37 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
38 (1<<TRANSPARENT_HUGEPAGE_FLAG
)|
40 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
41 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
)|
43 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
)|
44 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
)|
45 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
47 /* default scan 8*512 pte (or vmas) every 30 second */
48 static unsigned int khugepaged_pages_to_scan __read_mostly
= HPAGE_PMD_NR
*8;
49 static unsigned int khugepaged_pages_collapsed
;
50 static unsigned int khugepaged_full_scans
;
51 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly
= 10000;
52 /* during fragmentation poll the hugepage allocator once every minute */
53 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly
= 60000;
54 static struct task_struct
*khugepaged_thread __read_mostly
;
55 static DEFINE_MUTEX(khugepaged_mutex
);
56 static DEFINE_SPINLOCK(khugepaged_mm_lock
);
57 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait
);
59 * default collapse hugepages if there is at least one pte mapped like
60 * it would have happened if the vma was large enough during page
63 static unsigned int khugepaged_max_ptes_none __read_mostly
= HPAGE_PMD_NR
-1;
65 static int khugepaged(void *none
);
66 static int khugepaged_slab_init(void);
68 #define MM_SLOTS_HASH_BITS 10
69 static __read_mostly
DEFINE_HASHTABLE(mm_slots_hash
, MM_SLOTS_HASH_BITS
);
71 static struct kmem_cache
*mm_slot_cache __read_mostly
;
74 * struct mm_slot - hash lookup from mm to mm_slot
75 * @hash: hash collision list
76 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
77 * @mm: the mm that this information is valid for
80 struct hlist_node hash
;
81 struct list_head mm_node
;
86 * struct khugepaged_scan - cursor for scanning
87 * @mm_head: the head of the mm list to scan
88 * @mm_slot: the current mm_slot we are scanning
89 * @address: the next address inside that to be scanned
91 * There is only the one khugepaged_scan instance of this cursor structure.
93 struct khugepaged_scan
{
94 struct list_head mm_head
;
95 struct mm_slot
*mm_slot
;
96 unsigned long address
;
98 static struct khugepaged_scan khugepaged_scan
= {
99 .mm_head
= LIST_HEAD_INIT(khugepaged_scan
.mm_head
),
103 static int set_recommended_min_free_kbytes(void)
107 unsigned long recommended_min
;
109 if (!khugepaged_enabled())
112 for_each_populated_zone(zone
)
115 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
116 recommended_min
= pageblock_nr_pages
* nr_zones
* 2;
119 * Make sure that on average at least two pageblocks are almost free
120 * of another type, one for a migratetype to fall back to and a
121 * second to avoid subsequent fallbacks of other types There are 3
122 * MIGRATE_TYPES we care about.
124 recommended_min
+= pageblock_nr_pages
* nr_zones
*
125 MIGRATE_PCPTYPES
* MIGRATE_PCPTYPES
;
127 /* don't ever allow to reserve more than 5% of the lowmem */
128 recommended_min
= min(recommended_min
,
129 (unsigned long) nr_free_buffer_pages() / 20);
130 recommended_min
<<= (PAGE_SHIFT
-10);
132 if (recommended_min
> min_free_kbytes
)
133 min_free_kbytes
= recommended_min
;
134 setup_per_zone_wmarks();
137 late_initcall(set_recommended_min_free_kbytes
);
139 static int start_khugepaged(void)
142 if (khugepaged_enabled()) {
143 if (!khugepaged_thread
)
144 khugepaged_thread
= kthread_run(khugepaged
, NULL
,
146 if (unlikely(IS_ERR(khugepaged_thread
))) {
148 "khugepaged: kthread_run(khugepaged) failed\n");
149 err
= PTR_ERR(khugepaged_thread
);
150 khugepaged_thread
= NULL
;
153 if (!list_empty(&khugepaged_scan
.mm_head
))
154 wake_up_interruptible(&khugepaged_wait
);
156 set_recommended_min_free_kbytes();
157 } else if (khugepaged_thread
) {
158 kthread_stop(khugepaged_thread
);
159 khugepaged_thread
= NULL
;
165 static atomic_t huge_zero_refcount
;
166 static struct page
*huge_zero_page __read_mostly
;
168 static inline bool is_huge_zero_page(struct page
*page
)
170 return ACCESS_ONCE(huge_zero_page
) == page
;
173 static inline bool is_huge_zero_pmd(pmd_t pmd
)
175 return is_huge_zero_page(pmd_page(pmd
));
178 static struct page
*get_huge_zero_page(void)
180 struct page
*zero_page
;
182 if (likely(atomic_inc_not_zero(&huge_zero_refcount
)))
183 return ACCESS_ONCE(huge_zero_page
);
185 zero_page
= alloc_pages((GFP_TRANSHUGE
| __GFP_ZERO
) & ~__GFP_MOVABLE
,
188 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED
);
191 count_vm_event(THP_ZERO_PAGE_ALLOC
);
193 if (cmpxchg(&huge_zero_page
, NULL
, zero_page
)) {
195 __free_page(zero_page
);
199 /* We take additional reference here. It will be put back by shrinker */
200 atomic_set(&huge_zero_refcount
, 2);
202 return ACCESS_ONCE(huge_zero_page
);
205 static void put_huge_zero_page(void)
208 * Counter should never go to zero here. Only shrinker can put
211 BUG_ON(atomic_dec_and_test(&huge_zero_refcount
));
214 static int shrink_huge_zero_page(struct shrinker
*shrink
,
215 struct shrink_control
*sc
)
218 /* we can free zero page only if last reference remains */
219 return atomic_read(&huge_zero_refcount
) == 1 ? HPAGE_PMD_NR
: 0;
221 if (atomic_cmpxchg(&huge_zero_refcount
, 1, 0) == 1) {
222 struct page
*zero_page
= xchg(&huge_zero_page
, NULL
);
223 BUG_ON(zero_page
== NULL
);
224 __free_page(zero_page
);
230 static struct shrinker huge_zero_page_shrinker
= {
231 .shrink
= shrink_huge_zero_page
,
232 .seeks
= DEFAULT_SEEKS
,
237 static ssize_t
double_flag_show(struct kobject
*kobj
,
238 struct kobj_attribute
*attr
, char *buf
,
239 enum transparent_hugepage_flag enabled
,
240 enum transparent_hugepage_flag req_madv
)
242 if (test_bit(enabled
, &transparent_hugepage_flags
)) {
243 VM_BUG_ON(test_bit(req_madv
, &transparent_hugepage_flags
));
244 return sprintf(buf
, "[always] madvise never\n");
245 } else if (test_bit(req_madv
, &transparent_hugepage_flags
))
246 return sprintf(buf
, "always [madvise] never\n");
248 return sprintf(buf
, "always madvise [never]\n");
250 static ssize_t
double_flag_store(struct kobject
*kobj
,
251 struct kobj_attribute
*attr
,
252 const char *buf
, size_t count
,
253 enum transparent_hugepage_flag enabled
,
254 enum transparent_hugepage_flag req_madv
)
256 if (!memcmp("always", buf
,
257 min(sizeof("always")-1, count
))) {
258 set_bit(enabled
, &transparent_hugepage_flags
);
259 clear_bit(req_madv
, &transparent_hugepage_flags
);
260 } else if (!memcmp("madvise", buf
,
261 min(sizeof("madvise")-1, count
))) {
262 clear_bit(enabled
, &transparent_hugepage_flags
);
263 set_bit(req_madv
, &transparent_hugepage_flags
);
264 } else if (!memcmp("never", buf
,
265 min(sizeof("never")-1, count
))) {
266 clear_bit(enabled
, &transparent_hugepage_flags
);
267 clear_bit(req_madv
, &transparent_hugepage_flags
);
274 static ssize_t
enabled_show(struct kobject
*kobj
,
275 struct kobj_attribute
*attr
, char *buf
)
277 return double_flag_show(kobj
, attr
, buf
,
278 TRANSPARENT_HUGEPAGE_FLAG
,
279 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
281 static ssize_t
enabled_store(struct kobject
*kobj
,
282 struct kobj_attribute
*attr
,
283 const char *buf
, size_t count
)
287 ret
= double_flag_store(kobj
, attr
, buf
, count
,
288 TRANSPARENT_HUGEPAGE_FLAG
,
289 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
294 mutex_lock(&khugepaged_mutex
);
295 err
= start_khugepaged();
296 mutex_unlock(&khugepaged_mutex
);
304 static struct kobj_attribute enabled_attr
=
305 __ATTR(enabled
, 0644, enabled_show
, enabled_store
);
307 static ssize_t
single_flag_show(struct kobject
*kobj
,
308 struct kobj_attribute
*attr
, char *buf
,
309 enum transparent_hugepage_flag flag
)
311 return sprintf(buf
, "%d\n",
312 !!test_bit(flag
, &transparent_hugepage_flags
));
315 static ssize_t
single_flag_store(struct kobject
*kobj
,
316 struct kobj_attribute
*attr
,
317 const char *buf
, size_t count
,
318 enum transparent_hugepage_flag flag
)
323 ret
= kstrtoul(buf
, 10, &value
);
330 set_bit(flag
, &transparent_hugepage_flags
);
332 clear_bit(flag
, &transparent_hugepage_flags
);
338 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
339 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
340 * memory just to allocate one more hugepage.
342 static ssize_t
defrag_show(struct kobject
*kobj
,
343 struct kobj_attribute
*attr
, char *buf
)
345 return double_flag_show(kobj
, attr
, buf
,
346 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
,
347 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
);
349 static ssize_t
defrag_store(struct kobject
*kobj
,
350 struct kobj_attribute
*attr
,
351 const char *buf
, size_t count
)
353 return double_flag_store(kobj
, attr
, buf
, count
,
354 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
,
355 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
);
357 static struct kobj_attribute defrag_attr
=
358 __ATTR(defrag
, 0644, defrag_show
, defrag_store
);
360 static ssize_t
use_zero_page_show(struct kobject
*kobj
,
361 struct kobj_attribute
*attr
, char *buf
)
363 return single_flag_show(kobj
, attr
, buf
,
364 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
366 static ssize_t
use_zero_page_store(struct kobject
*kobj
,
367 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
369 return single_flag_store(kobj
, attr
, buf
, count
,
370 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
372 static struct kobj_attribute use_zero_page_attr
=
373 __ATTR(use_zero_page
, 0644, use_zero_page_show
, use_zero_page_store
);
374 #ifdef CONFIG_DEBUG_VM
375 static ssize_t
debug_cow_show(struct kobject
*kobj
,
376 struct kobj_attribute
*attr
, char *buf
)
378 return single_flag_show(kobj
, attr
, buf
,
379 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
381 static ssize_t
debug_cow_store(struct kobject
*kobj
,
382 struct kobj_attribute
*attr
,
383 const char *buf
, size_t count
)
385 return single_flag_store(kobj
, attr
, buf
, count
,
386 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
388 static struct kobj_attribute debug_cow_attr
=
389 __ATTR(debug_cow
, 0644, debug_cow_show
, debug_cow_store
);
390 #endif /* CONFIG_DEBUG_VM */
392 static struct attribute
*hugepage_attr
[] = {
395 &use_zero_page_attr
.attr
,
396 #ifdef CONFIG_DEBUG_VM
397 &debug_cow_attr
.attr
,
402 static struct attribute_group hugepage_attr_group
= {
403 .attrs
= hugepage_attr
,
406 static ssize_t
scan_sleep_millisecs_show(struct kobject
*kobj
,
407 struct kobj_attribute
*attr
,
410 return sprintf(buf
, "%u\n", khugepaged_scan_sleep_millisecs
);
413 static ssize_t
scan_sleep_millisecs_store(struct kobject
*kobj
,
414 struct kobj_attribute
*attr
,
415 const char *buf
, size_t count
)
420 err
= strict_strtoul(buf
, 10, &msecs
);
421 if (err
|| msecs
> UINT_MAX
)
424 khugepaged_scan_sleep_millisecs
= msecs
;
425 wake_up_interruptible(&khugepaged_wait
);
429 static struct kobj_attribute scan_sleep_millisecs_attr
=
430 __ATTR(scan_sleep_millisecs
, 0644, scan_sleep_millisecs_show
,
431 scan_sleep_millisecs_store
);
433 static ssize_t
alloc_sleep_millisecs_show(struct kobject
*kobj
,
434 struct kobj_attribute
*attr
,
437 return sprintf(buf
, "%u\n", khugepaged_alloc_sleep_millisecs
);
440 static ssize_t
alloc_sleep_millisecs_store(struct kobject
*kobj
,
441 struct kobj_attribute
*attr
,
442 const char *buf
, size_t count
)
447 err
= strict_strtoul(buf
, 10, &msecs
);
448 if (err
|| msecs
> UINT_MAX
)
451 khugepaged_alloc_sleep_millisecs
= msecs
;
452 wake_up_interruptible(&khugepaged_wait
);
456 static struct kobj_attribute alloc_sleep_millisecs_attr
=
457 __ATTR(alloc_sleep_millisecs
, 0644, alloc_sleep_millisecs_show
,
458 alloc_sleep_millisecs_store
);
460 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
461 struct kobj_attribute
*attr
,
464 return sprintf(buf
, "%u\n", khugepaged_pages_to_scan
);
466 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
467 struct kobj_attribute
*attr
,
468 const char *buf
, size_t count
)
473 err
= strict_strtoul(buf
, 10, &pages
);
474 if (err
|| !pages
|| pages
> UINT_MAX
)
477 khugepaged_pages_to_scan
= pages
;
481 static struct kobj_attribute pages_to_scan_attr
=
482 __ATTR(pages_to_scan
, 0644, pages_to_scan_show
,
483 pages_to_scan_store
);
485 static ssize_t
pages_collapsed_show(struct kobject
*kobj
,
486 struct kobj_attribute
*attr
,
489 return sprintf(buf
, "%u\n", khugepaged_pages_collapsed
);
491 static struct kobj_attribute pages_collapsed_attr
=
492 __ATTR_RO(pages_collapsed
);
494 static ssize_t
full_scans_show(struct kobject
*kobj
,
495 struct kobj_attribute
*attr
,
498 return sprintf(buf
, "%u\n", khugepaged_full_scans
);
500 static struct kobj_attribute full_scans_attr
=
501 __ATTR_RO(full_scans
);
503 static ssize_t
khugepaged_defrag_show(struct kobject
*kobj
,
504 struct kobj_attribute
*attr
, char *buf
)
506 return single_flag_show(kobj
, attr
, buf
,
507 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
509 static ssize_t
khugepaged_defrag_store(struct kobject
*kobj
,
510 struct kobj_attribute
*attr
,
511 const char *buf
, size_t count
)
513 return single_flag_store(kobj
, attr
, buf
, count
,
514 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
516 static struct kobj_attribute khugepaged_defrag_attr
=
517 __ATTR(defrag
, 0644, khugepaged_defrag_show
,
518 khugepaged_defrag_store
);
521 * max_ptes_none controls if khugepaged should collapse hugepages over
522 * any unmapped ptes in turn potentially increasing the memory
523 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
524 * reduce the available free memory in the system as it
525 * runs. Increasing max_ptes_none will instead potentially reduce the
526 * free memory in the system during the khugepaged scan.
528 static ssize_t
khugepaged_max_ptes_none_show(struct kobject
*kobj
,
529 struct kobj_attribute
*attr
,
532 return sprintf(buf
, "%u\n", khugepaged_max_ptes_none
);
534 static ssize_t
khugepaged_max_ptes_none_store(struct kobject
*kobj
,
535 struct kobj_attribute
*attr
,
536 const char *buf
, size_t count
)
539 unsigned long max_ptes_none
;
541 err
= strict_strtoul(buf
, 10, &max_ptes_none
);
542 if (err
|| max_ptes_none
> HPAGE_PMD_NR
-1)
545 khugepaged_max_ptes_none
= max_ptes_none
;
549 static struct kobj_attribute khugepaged_max_ptes_none_attr
=
550 __ATTR(max_ptes_none
, 0644, khugepaged_max_ptes_none_show
,
551 khugepaged_max_ptes_none_store
);
553 static struct attribute
*khugepaged_attr
[] = {
554 &khugepaged_defrag_attr
.attr
,
555 &khugepaged_max_ptes_none_attr
.attr
,
556 &pages_to_scan_attr
.attr
,
557 &pages_collapsed_attr
.attr
,
558 &full_scans_attr
.attr
,
559 &scan_sleep_millisecs_attr
.attr
,
560 &alloc_sleep_millisecs_attr
.attr
,
564 static struct attribute_group khugepaged_attr_group
= {
565 .attrs
= khugepaged_attr
,
566 .name
= "khugepaged",
569 static int __init
hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
573 *hugepage_kobj
= kobject_create_and_add("transparent_hugepage", mm_kobj
);
574 if (unlikely(!*hugepage_kobj
)) {
575 printk(KERN_ERR
"hugepage: failed to create transparent hugepage kobject\n");
579 err
= sysfs_create_group(*hugepage_kobj
, &hugepage_attr_group
);
581 printk(KERN_ERR
"hugepage: failed to register transparent hugepage group\n");
585 err
= sysfs_create_group(*hugepage_kobj
, &khugepaged_attr_group
);
587 printk(KERN_ERR
"hugepage: failed to register transparent hugepage group\n");
588 goto remove_hp_group
;
594 sysfs_remove_group(*hugepage_kobj
, &hugepage_attr_group
);
596 kobject_put(*hugepage_kobj
);
600 static void __init
hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
602 sysfs_remove_group(hugepage_kobj
, &khugepaged_attr_group
);
603 sysfs_remove_group(hugepage_kobj
, &hugepage_attr_group
);
604 kobject_put(hugepage_kobj
);
607 static inline int hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
612 static inline void hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
615 #endif /* CONFIG_SYSFS */
617 static int __init
hugepage_init(void)
620 struct kobject
*hugepage_kobj
;
622 if (!has_transparent_hugepage()) {
623 transparent_hugepage_flags
= 0;
627 err
= hugepage_init_sysfs(&hugepage_kobj
);
631 err
= khugepaged_slab_init();
635 register_shrinker(&huge_zero_page_shrinker
);
638 * By default disable transparent hugepages on smaller systems,
639 * where the extra memory used could hurt more than TLB overhead
640 * is likely to save. The admin can still enable it through /sys.
642 if (totalram_pages
< (512 << (20 - PAGE_SHIFT
)))
643 transparent_hugepage_flags
= 0;
649 hugepage_exit_sysfs(hugepage_kobj
);
652 module_init(hugepage_init
)
654 static int __init
setup_transparent_hugepage(char *str
)
659 if (!strcmp(str
, "always")) {
660 set_bit(TRANSPARENT_HUGEPAGE_FLAG
,
661 &transparent_hugepage_flags
);
662 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
663 &transparent_hugepage_flags
);
665 } else if (!strcmp(str
, "madvise")) {
666 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
667 &transparent_hugepage_flags
);
668 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
669 &transparent_hugepage_flags
);
671 } else if (!strcmp(str
, "never")) {
672 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
673 &transparent_hugepage_flags
);
674 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
675 &transparent_hugepage_flags
);
681 "transparent_hugepage= cannot parse, ignored\n");
684 __setup("transparent_hugepage=", setup_transparent_hugepage
);
686 pmd_t
maybe_pmd_mkwrite(pmd_t pmd
, struct vm_area_struct
*vma
)
688 if (likely(vma
->vm_flags
& VM_WRITE
))
689 pmd
= pmd_mkwrite(pmd
);
693 static inline pmd_t
mk_huge_pmd(struct page
*page
, struct vm_area_struct
*vma
)
696 entry
= mk_pmd(page
, vma
->vm_page_prot
);
697 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
698 entry
= pmd_mkhuge(entry
);
702 static int __do_huge_pmd_anonymous_page(struct mm_struct
*mm
,
703 struct vm_area_struct
*vma
,
704 unsigned long haddr
, pmd_t
*pmd
,
709 VM_BUG_ON(!PageCompound(page
));
710 pgtable
= pte_alloc_one(mm
, haddr
);
711 if (unlikely(!pgtable
))
714 clear_huge_page(page
, haddr
, HPAGE_PMD_NR
);
716 * The memory barrier inside __SetPageUptodate makes sure that
717 * clear_huge_page writes become visible before the set_pmd_at()
720 __SetPageUptodate(page
);
722 spin_lock(&mm
->page_table_lock
);
723 if (unlikely(!pmd_none(*pmd
))) {
724 spin_unlock(&mm
->page_table_lock
);
725 mem_cgroup_uncharge_page(page
);
727 pte_free(mm
, pgtable
);
730 entry
= mk_huge_pmd(page
, vma
);
731 page_add_new_anon_rmap(page
, vma
, haddr
);
732 set_pmd_at(mm
, haddr
, pmd
, entry
);
733 pgtable_trans_huge_deposit(mm
, pgtable
);
734 add_mm_counter(mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
736 spin_unlock(&mm
->page_table_lock
);
742 static inline gfp_t
alloc_hugepage_gfpmask(int defrag
, gfp_t extra_gfp
)
744 return (GFP_TRANSHUGE
& ~(defrag
? 0 : __GFP_WAIT
)) | extra_gfp
;
747 static inline struct page
*alloc_hugepage_vma(int defrag
,
748 struct vm_area_struct
*vma
,
749 unsigned long haddr
, int nd
,
752 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag
, extra_gfp
),
753 HPAGE_PMD_ORDER
, vma
, haddr
, nd
);
757 static inline struct page
*alloc_hugepage(int defrag
)
759 return alloc_pages(alloc_hugepage_gfpmask(defrag
, 0),
764 static bool set_huge_zero_page(pgtable_t pgtable
, struct mm_struct
*mm
,
765 struct vm_area_struct
*vma
, unsigned long haddr
, pmd_t
*pmd
,
766 struct page
*zero_page
)
771 entry
= mk_pmd(zero_page
, vma
->vm_page_prot
);
772 entry
= pmd_wrprotect(entry
);
773 entry
= pmd_mkhuge(entry
);
774 set_pmd_at(mm
, haddr
, pmd
, entry
);
775 pgtable_trans_huge_deposit(mm
, pgtable
);
780 int do_huge_pmd_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
781 unsigned long address
, pmd_t
*pmd
,
785 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
788 if (haddr
>= vma
->vm_start
&& haddr
+ HPAGE_PMD_SIZE
<= vma
->vm_end
) {
789 if (unlikely(anon_vma_prepare(vma
)))
791 if (unlikely(khugepaged_enter(vma
)))
793 if (!(flags
& FAULT_FLAG_WRITE
) &&
794 transparent_hugepage_use_zero_page()) {
796 struct page
*zero_page
;
798 pgtable
= pte_alloc_one(mm
, haddr
);
799 if (unlikely(!pgtable
))
801 zero_page
= get_huge_zero_page();
802 if (unlikely(!zero_page
)) {
803 pte_free(mm
, pgtable
);
804 count_vm_event(THP_FAULT_FALLBACK
);
807 spin_lock(&mm
->page_table_lock
);
808 set
= set_huge_zero_page(pgtable
, mm
, vma
, haddr
, pmd
,
810 spin_unlock(&mm
->page_table_lock
);
812 pte_free(mm
, pgtable
);
813 put_huge_zero_page();
817 page
= alloc_hugepage_vma(transparent_hugepage_defrag(vma
),
818 vma
, haddr
, numa_node_id(), 0);
819 if (unlikely(!page
)) {
820 count_vm_event(THP_FAULT_FALLBACK
);
823 count_vm_event(THP_FAULT_ALLOC
);
824 if (unlikely(mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))) {
828 if (unlikely(__do_huge_pmd_anonymous_page(mm
, vma
, haddr
, pmd
,
830 mem_cgroup_uncharge_page(page
);
839 * Use __pte_alloc instead of pte_alloc_map, because we can't
840 * run pte_offset_map on the pmd, if an huge pmd could
841 * materialize from under us from a different thread.
843 if (unlikely(pmd_none(*pmd
)) &&
844 unlikely(__pte_alloc(mm
, vma
, pmd
, address
)))
846 /* if an huge pmd materialized from under us just retry later */
847 if (unlikely(pmd_trans_huge(*pmd
)))
850 * A regular pmd is established and it can't morph into a huge pmd
851 * from under us anymore at this point because we hold the mmap_sem
852 * read mode and khugepaged takes it in write mode. So now it's
853 * safe to run pte_offset_map().
855 pte
= pte_offset_map(pmd
, address
);
856 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
859 int copy_huge_pmd(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
860 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
861 struct vm_area_struct
*vma
)
863 struct page
*src_page
;
869 pgtable
= pte_alloc_one(dst_mm
, addr
);
870 if (unlikely(!pgtable
))
873 spin_lock(&dst_mm
->page_table_lock
);
874 spin_lock_nested(&src_mm
->page_table_lock
, SINGLE_DEPTH_NESTING
);
878 if (unlikely(!pmd_trans_huge(pmd
))) {
879 pte_free(dst_mm
, pgtable
);
883 * mm->page_table_lock is enough to be sure that huge zero pmd is not
884 * under splitting since we don't split the page itself, only pmd to
887 if (is_huge_zero_pmd(pmd
)) {
888 struct page
*zero_page
;
891 * get_huge_zero_page() will never allocate a new page here,
892 * since we already have a zero page to copy. It just takes a
895 zero_page
= get_huge_zero_page();
896 set
= set_huge_zero_page(pgtable
, dst_mm
, vma
, addr
, dst_pmd
,
898 BUG_ON(!set
); /* unexpected !pmd_none(dst_pmd) */
902 if (unlikely(pmd_trans_splitting(pmd
))) {
903 /* split huge page running from under us */
904 spin_unlock(&src_mm
->page_table_lock
);
905 spin_unlock(&dst_mm
->page_table_lock
);
906 pte_free(dst_mm
, pgtable
);
908 wait_split_huge_page(vma
->anon_vma
, src_pmd
); /* src_vma */
911 src_page
= pmd_page(pmd
);
912 VM_BUG_ON(!PageHead(src_page
));
914 page_dup_rmap(src_page
);
915 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
917 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
918 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
919 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
920 pgtable_trans_huge_deposit(dst_mm
, pgtable
);
925 spin_unlock(&src_mm
->page_table_lock
);
926 spin_unlock(&dst_mm
->page_table_lock
);
931 void huge_pmd_set_accessed(struct mm_struct
*mm
,
932 struct vm_area_struct
*vma
,
933 unsigned long address
,
934 pmd_t
*pmd
, pmd_t orig_pmd
,
940 spin_lock(&mm
->page_table_lock
);
941 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
944 entry
= pmd_mkyoung(orig_pmd
);
945 haddr
= address
& HPAGE_PMD_MASK
;
946 if (pmdp_set_access_flags(vma
, haddr
, pmd
, entry
, dirty
))
947 update_mmu_cache_pmd(vma
, address
, pmd
);
950 spin_unlock(&mm
->page_table_lock
);
953 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct
*mm
,
954 struct vm_area_struct
*vma
, unsigned long address
,
955 pmd_t
*pmd
, pmd_t orig_pmd
, unsigned long haddr
)
961 unsigned long mmun_start
; /* For mmu_notifiers */
962 unsigned long mmun_end
; /* For mmu_notifiers */
964 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
970 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
)) {
976 clear_user_highpage(page
, address
);
977 __SetPageUptodate(page
);
980 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
981 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
983 spin_lock(&mm
->page_table_lock
);
984 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
987 pmdp_clear_flush(vma
, haddr
, pmd
);
988 /* leave pmd empty until pte is filled */
990 pgtable
= pgtable_trans_huge_withdraw(mm
);
991 pmd_populate(mm
, &_pmd
, pgtable
);
993 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
995 if (haddr
== (address
& PAGE_MASK
)) {
996 entry
= mk_pte(page
, vma
->vm_page_prot
);
997 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
998 page_add_new_anon_rmap(page
, vma
, haddr
);
1000 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
1001 entry
= pte_mkspecial(entry
);
1003 pte
= pte_offset_map(&_pmd
, haddr
);
1004 VM_BUG_ON(!pte_none(*pte
));
1005 set_pte_at(mm
, haddr
, pte
, entry
);
1008 smp_wmb(); /* make pte visible before pmd */
1009 pmd_populate(mm
, pmd
, pgtable
);
1010 spin_unlock(&mm
->page_table_lock
);
1011 put_huge_zero_page();
1012 inc_mm_counter(mm
, MM_ANONPAGES
);
1014 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1016 ret
|= VM_FAULT_WRITE
;
1020 spin_unlock(&mm
->page_table_lock
);
1021 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1022 mem_cgroup_uncharge_page(page
);
1027 static int do_huge_pmd_wp_page_fallback(struct mm_struct
*mm
,
1028 struct vm_area_struct
*vma
,
1029 unsigned long address
,
1030 pmd_t
*pmd
, pmd_t orig_pmd
,
1032 unsigned long haddr
)
1037 struct page
**pages
;
1038 unsigned long mmun_start
; /* For mmu_notifiers */
1039 unsigned long mmun_end
; /* For mmu_notifiers */
1041 pages
= kmalloc(sizeof(struct page
*) * HPAGE_PMD_NR
,
1043 if (unlikely(!pages
)) {
1044 ret
|= VM_FAULT_OOM
;
1048 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1049 pages
[i
] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE
|
1051 vma
, address
, page_to_nid(page
));
1052 if (unlikely(!pages
[i
] ||
1053 mem_cgroup_newpage_charge(pages
[i
], mm
,
1057 mem_cgroup_uncharge_start();
1059 mem_cgroup_uncharge_page(pages
[i
]);
1062 mem_cgroup_uncharge_end();
1064 ret
|= VM_FAULT_OOM
;
1069 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1070 copy_user_highpage(pages
[i
], page
+ i
,
1071 haddr
+ PAGE_SIZE
* i
, vma
);
1072 __SetPageUptodate(pages
[i
]);
1077 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1078 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1080 spin_lock(&mm
->page_table_lock
);
1081 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
1082 goto out_free_pages
;
1083 VM_BUG_ON(!PageHead(page
));
1085 pmdp_clear_flush(vma
, haddr
, pmd
);
1086 /* leave pmd empty until pte is filled */
1088 pgtable
= pgtable_trans_huge_withdraw(mm
);
1089 pmd_populate(mm
, &_pmd
, pgtable
);
1091 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1093 entry
= mk_pte(pages
[i
], vma
->vm_page_prot
);
1094 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1095 page_add_new_anon_rmap(pages
[i
], vma
, haddr
);
1096 pte
= pte_offset_map(&_pmd
, haddr
);
1097 VM_BUG_ON(!pte_none(*pte
));
1098 set_pte_at(mm
, haddr
, pte
, entry
);
1103 smp_wmb(); /* make pte visible before pmd */
1104 pmd_populate(mm
, pmd
, pgtable
);
1105 page_remove_rmap(page
);
1106 spin_unlock(&mm
->page_table_lock
);
1108 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1110 ret
|= VM_FAULT_WRITE
;
1117 spin_unlock(&mm
->page_table_lock
);
1118 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1119 mem_cgroup_uncharge_start();
1120 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1121 mem_cgroup_uncharge_page(pages
[i
]);
1124 mem_cgroup_uncharge_end();
1129 int do_huge_pmd_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1130 unsigned long address
, pmd_t
*pmd
, pmd_t orig_pmd
)
1133 struct page
*page
= NULL
, *new_page
;
1134 unsigned long haddr
;
1135 unsigned long mmun_start
; /* For mmu_notifiers */
1136 unsigned long mmun_end
; /* For mmu_notifiers */
1138 VM_BUG_ON(!vma
->anon_vma
);
1139 haddr
= address
& HPAGE_PMD_MASK
;
1140 if (is_huge_zero_pmd(orig_pmd
))
1142 spin_lock(&mm
->page_table_lock
);
1143 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
1146 page
= pmd_page(orig_pmd
);
1147 VM_BUG_ON(!PageCompound(page
) || !PageHead(page
));
1148 if (page_mapcount(page
) == 1) {
1150 entry
= pmd_mkyoung(orig_pmd
);
1151 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1152 if (pmdp_set_access_flags(vma
, haddr
, pmd
, entry
, 1))
1153 update_mmu_cache_pmd(vma
, address
, pmd
);
1154 ret
|= VM_FAULT_WRITE
;
1158 spin_unlock(&mm
->page_table_lock
);
1160 if (transparent_hugepage_enabled(vma
) &&
1161 !transparent_hugepage_debug_cow())
1162 new_page
= alloc_hugepage_vma(transparent_hugepage_defrag(vma
),
1163 vma
, haddr
, numa_node_id(), 0);
1167 if (unlikely(!new_page
)) {
1168 count_vm_event(THP_FAULT_FALLBACK
);
1170 ret
= do_huge_pmd_wp_zero_page_fallback(mm
, vma
,
1171 address
, pmd
, orig_pmd
, haddr
);
1173 ret
= do_huge_pmd_wp_page_fallback(mm
, vma
, address
,
1174 pmd
, orig_pmd
, page
, haddr
);
1175 if (ret
& VM_FAULT_OOM
)
1176 split_huge_page(page
);
1181 count_vm_event(THP_FAULT_ALLOC
);
1183 if (unlikely(mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))) {
1186 split_huge_page(page
);
1189 ret
|= VM_FAULT_OOM
;
1194 clear_huge_page(new_page
, haddr
, HPAGE_PMD_NR
);
1196 copy_user_huge_page(new_page
, page
, haddr
, vma
, HPAGE_PMD_NR
);
1197 __SetPageUptodate(new_page
);
1200 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1201 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1203 spin_lock(&mm
->page_table_lock
);
1206 if (unlikely(!pmd_same(*pmd
, orig_pmd
))) {
1207 spin_unlock(&mm
->page_table_lock
);
1208 mem_cgroup_uncharge_page(new_page
);
1213 entry
= mk_huge_pmd(new_page
, vma
);
1214 pmdp_clear_flush(vma
, haddr
, pmd
);
1215 page_add_new_anon_rmap(new_page
, vma
, haddr
);
1216 set_pmd_at(mm
, haddr
, pmd
, entry
);
1217 update_mmu_cache_pmd(vma
, address
, pmd
);
1219 add_mm_counter(mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1220 put_huge_zero_page();
1222 VM_BUG_ON(!PageHead(page
));
1223 page_remove_rmap(page
);
1226 ret
|= VM_FAULT_WRITE
;
1228 spin_unlock(&mm
->page_table_lock
);
1230 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1234 spin_unlock(&mm
->page_table_lock
);
1239 * foll_force can write to even unwritable pmd's, but only
1240 * after we've gone through a cow cycle and they are dirty.
1242 static inline bool can_follow_write_pmd(pmd_t pmd
, struct page
*page
,
1245 return pmd_write(pmd
) ||
1246 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) &&
1247 page
&& PageAnon(page
));
1250 struct page
*follow_trans_huge_pmd(struct vm_area_struct
*vma
,
1255 struct mm_struct
*mm
= vma
->vm_mm
;
1256 struct page
*page
= NULL
;
1258 assert_spin_locked(&mm
->page_table_lock
);
1260 /* Avoid dumping huge zero page */
1261 if ((flags
& FOLL_DUMP
) && is_huge_zero_pmd(*pmd
))
1262 return ERR_PTR(-EFAULT
);
1264 page
= pmd_page(*pmd
);
1265 VM_BUG_ON(!PageHead(page
));
1267 if (flags
& FOLL_WRITE
&& !can_follow_write_pmd(*pmd
, page
, flags
))
1270 if (flags
& FOLL_TOUCH
) {
1273 * We should set the dirty bit only for FOLL_WRITE but
1274 * for now the dirty bit in the pmd is meaningless.
1275 * And if the dirty bit will become meaningful and
1276 * we'll only set it with FOLL_WRITE, an atomic
1277 * set_bit will be required on the pmd to set the
1278 * young bit, instead of the current set_pmd_at.
1280 _pmd
= pmd_mkyoung(pmd_mkdirty(*pmd
));
1281 set_pmd_at(mm
, addr
& HPAGE_PMD_MASK
, pmd
, _pmd
);
1283 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1284 if (page
->mapping
&& trylock_page(page
)) {
1287 mlock_vma_page(page
);
1291 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
1292 VM_BUG_ON(!PageCompound(page
));
1293 if (flags
& FOLL_GET
)
1294 get_page_foll(page
);
1300 /* NUMA hinting page fault entry point for trans huge pmds */
1301 int do_huge_pmd_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1302 unsigned long addr
, pmd_t pmd
, pmd_t
*pmdp
)
1304 struct anon_vma
*anon_vma
= NULL
;
1306 unsigned long haddr
= addr
& HPAGE_PMD_MASK
;
1307 int page_nid
= -1, this_nid
= numa_node_id();
1310 bool migrated
= false;
1312 spin_lock(&mm
->page_table_lock
);
1313 if (unlikely(!pmd_same(pmd
, *pmdp
)))
1316 page
= pmd_page(pmd
);
1317 page_nid
= page_to_nid(page
);
1318 count_vm_numa_event(NUMA_HINT_FAULTS
);
1319 if (page_nid
== this_nid
)
1320 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
1323 * Acquire the page lock to serialise THP migrations but avoid dropping
1324 * page_table_lock if at all possible
1326 page_locked
= trylock_page(page
);
1327 target_nid
= mpol_misplaced(page
, vma
, haddr
);
1328 if (target_nid
== -1) {
1329 /* If the page was locked, there are no parallel migrations */
1334 * Otherwise wait for potential migrations and retry. We do
1335 * relock and check_same as the page may no longer be mapped.
1336 * As the fault is being retried, do not account for it.
1338 spin_unlock(&mm
->page_table_lock
);
1339 wait_on_page_locked(page
);
1344 /* Page is misplaced, serialise migrations and parallel THP splits */
1346 spin_unlock(&mm
->page_table_lock
);
1349 anon_vma
= page_lock_anon_vma_read(page
);
1351 /* Confirm the PTE did not while locked */
1352 spin_lock(&mm
->page_table_lock
);
1353 if (unlikely(!pmd_same(pmd
, *pmdp
))) {
1360 /* Bail if we fail to protect against THP splits for any reason */
1361 if (unlikely(!anon_vma
)) {
1368 * The page_table_lock above provides a memory barrier
1369 * with change_protection_range.
1371 if (mm_tlb_flush_pending(mm
))
1372 flush_tlb_range(vma
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
1375 * Migrate the THP to the requested node, returns with page unlocked
1376 * and pmd_numa cleared.
1378 spin_unlock(&mm
->page_table_lock
);
1379 migrated
= migrate_misplaced_transhuge_page(mm
, vma
,
1380 pmdp
, pmd
, addr
, page
, target_nid
);
1382 page_nid
= target_nid
;
1386 BUG_ON(!PageLocked(page
));
1387 pmd
= pmd_mknonnuma(pmd
);
1388 set_pmd_at(mm
, haddr
, pmdp
, pmd
);
1389 VM_BUG_ON(pmd_numa(*pmdp
));
1390 update_mmu_cache_pmd(vma
, addr
, pmdp
);
1393 spin_unlock(&mm
->page_table_lock
);
1397 page_unlock_anon_vma_read(anon_vma
);
1400 task_numa_fault(page_nid
, HPAGE_PMD_NR
, migrated
);
1405 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1406 pmd_t
*pmd
, unsigned long addr
)
1410 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
1414 pgtable
= pgtable_trans_huge_withdraw(tlb
->mm
);
1415 orig_pmd
= pmdp_get_and_clear(tlb
->mm
, addr
, pmd
);
1416 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1417 if (is_huge_zero_pmd(orig_pmd
)) {
1419 spin_unlock(&tlb
->mm
->page_table_lock
);
1420 put_huge_zero_page();
1422 page
= pmd_page(orig_pmd
);
1423 page_remove_rmap(page
);
1424 VM_BUG_ON(page_mapcount(page
) < 0);
1425 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1426 VM_BUG_ON(!PageHead(page
));
1428 spin_unlock(&tlb
->mm
->page_table_lock
);
1429 tlb_remove_page(tlb
, page
);
1431 pte_free(tlb
->mm
, pgtable
);
1437 int mincore_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1438 unsigned long addr
, unsigned long end
,
1443 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
1445 * All logical pages in the range are present
1446 * if backed by a huge page.
1448 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1449 memset(vec
, 1, (end
- addr
) >> PAGE_SHIFT
);
1456 int move_huge_pmd(struct vm_area_struct
*vma
, struct vm_area_struct
*new_vma
,
1457 unsigned long old_addr
,
1458 unsigned long new_addr
, unsigned long old_end
,
1459 pmd_t
*old_pmd
, pmd_t
*new_pmd
)
1464 struct mm_struct
*mm
= vma
->vm_mm
;
1466 if ((old_addr
& ~HPAGE_PMD_MASK
) ||
1467 (new_addr
& ~HPAGE_PMD_MASK
) ||
1468 old_end
- old_addr
< HPAGE_PMD_SIZE
||
1469 (new_vma
->vm_flags
& VM_NOHUGEPAGE
))
1473 * The destination pmd shouldn't be established, free_pgtables()
1474 * should have release it.
1476 if (WARN_ON(!pmd_none(*new_pmd
))) {
1477 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1481 ret
= __pmd_trans_huge_lock(old_pmd
, vma
);
1483 pmd
= pmdp_get_and_clear(mm
, old_addr
, old_pmd
);
1484 VM_BUG_ON(!pmd_none(*new_pmd
));
1485 set_pmd_at(mm
, new_addr
, new_pmd
, pmd
);
1486 spin_unlock(&mm
->page_table_lock
);
1492 int change_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1493 unsigned long addr
, pgprot_t newprot
, int prot_numa
)
1495 struct mm_struct
*mm
= vma
->vm_mm
;
1498 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
1500 entry
= pmdp_get_and_clear(mm
, addr
, pmd
);
1502 entry
= pmd_modify(entry
, newprot
);
1503 BUG_ON(pmd_write(entry
));
1505 struct page
*page
= pmd_page(*pmd
);
1507 /* only check non-shared pages */
1508 if (page_mapcount(page
) == 1 &&
1510 entry
= pmd_mknuma(entry
);
1513 set_pmd_at(mm
, addr
, pmd
, entry
);
1514 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1522 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1523 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1525 * Note that if it returns 1, this routine returns without unlocking page
1526 * table locks. So callers must unlock them.
1528 int __pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
)
1530 spin_lock(&vma
->vm_mm
->page_table_lock
);
1531 if (likely(pmd_trans_huge(*pmd
))) {
1532 if (unlikely(pmd_trans_splitting(*pmd
))) {
1533 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1534 wait_split_huge_page(vma
->anon_vma
, pmd
);
1537 /* Thp mapped by 'pmd' is stable, so we can
1538 * handle it as it is. */
1542 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1546 pmd_t
*page_check_address_pmd(struct page
*page
,
1547 struct mm_struct
*mm
,
1548 unsigned long address
,
1549 enum page_check_address_pmd_flag flag
)
1551 pmd_t
*pmd
, *ret
= NULL
;
1553 if (address
& ~HPAGE_PMD_MASK
)
1556 pmd
= mm_find_pmd(mm
, address
);
1561 if (pmd_page(*pmd
) != page
)
1564 * split_vma() may create temporary aliased mappings. There is
1565 * no risk as long as all huge pmd are found and have their
1566 * splitting bit set before __split_huge_page_refcount
1567 * runs. Finding the same huge pmd more than once during the
1568 * same rmap walk is not a problem.
1570 if (flag
== PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG
&&
1571 pmd_trans_splitting(*pmd
))
1573 if (pmd_trans_huge(*pmd
)) {
1574 VM_BUG_ON(flag
== PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG
&&
1575 !pmd_trans_splitting(*pmd
));
1582 static int __split_huge_page_splitting(struct page
*page
,
1583 struct vm_area_struct
*vma
,
1584 unsigned long address
)
1586 struct mm_struct
*mm
= vma
->vm_mm
;
1589 /* For mmu_notifiers */
1590 const unsigned long mmun_start
= address
;
1591 const unsigned long mmun_end
= address
+ HPAGE_PMD_SIZE
;
1593 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1594 spin_lock(&mm
->page_table_lock
);
1595 pmd
= page_check_address_pmd(page
, mm
, address
,
1596 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG
);
1599 * We can't temporarily set the pmd to null in order
1600 * to split it, the pmd must remain marked huge at all
1601 * times or the VM won't take the pmd_trans_huge paths
1602 * and it won't wait on the anon_vma->root->rwsem to
1603 * serialize against split_huge_page*.
1605 pmdp_splitting_flush(vma
, address
, pmd
);
1608 spin_unlock(&mm
->page_table_lock
);
1609 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1614 static void __split_huge_page_refcount(struct page
*page
,
1615 struct list_head
*list
)
1618 struct zone
*zone
= page_zone(page
);
1619 struct lruvec
*lruvec
;
1622 /* prevent PageLRU to go away from under us, and freeze lru stats */
1623 spin_lock_irq(&zone
->lru_lock
);
1624 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1626 compound_lock(page
);
1627 /* complete memcg works before add pages to LRU */
1628 mem_cgroup_split_huge_fixup(page
);
1630 for (i
= HPAGE_PMD_NR
- 1; i
>= 1; i
--) {
1631 struct page
*page_tail
= page
+ i
;
1633 /* tail_page->_mapcount cannot change */
1634 BUG_ON(page_mapcount(page_tail
) < 0);
1635 tail_count
+= page_mapcount(page_tail
);
1636 /* check for overflow */
1637 BUG_ON(tail_count
< 0);
1638 BUG_ON(atomic_read(&page_tail
->_count
) != 0);
1640 * tail_page->_count is zero and not changing from
1641 * under us. But get_page_unless_zero() may be running
1642 * from under us on the tail_page. If we used
1643 * atomic_set() below instead of atomic_add(), we
1644 * would then run atomic_set() concurrently with
1645 * get_page_unless_zero(), and atomic_set() is
1646 * implemented in C not using locked ops. spin_unlock
1647 * on x86 sometime uses locked ops because of PPro
1648 * errata 66, 92, so unless somebody can guarantee
1649 * atomic_set() here would be safe on all archs (and
1650 * not only on x86), it's safer to use atomic_add().
1652 atomic_add(page_mapcount(page
) + page_mapcount(page_tail
) + 1,
1653 &page_tail
->_count
);
1655 /* after clearing PageTail the gup refcount can be released */
1659 * retain hwpoison flag of the poisoned tail page:
1660 * fix for the unsuitable process killed on Guest Machine(KVM)
1661 * by the memory-failure.
1663 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
| __PG_HWPOISON
;
1664 page_tail
->flags
|= (page
->flags
&
1665 ((1L << PG_referenced
) |
1666 (1L << PG_swapbacked
) |
1667 (1L << PG_mlocked
) |
1668 (1L << PG_uptodate
)));
1669 page_tail
->flags
|= (1L << PG_dirty
);
1671 /* clear PageTail before overwriting first_page */
1675 * __split_huge_page_splitting() already set the
1676 * splitting bit in all pmd that could map this
1677 * hugepage, that will ensure no CPU can alter the
1678 * mapcount on the head page. The mapcount is only
1679 * accounted in the head page and it has to be
1680 * transferred to all tail pages in the below code. So
1681 * for this code to be safe, the split the mapcount
1682 * can't change. But that doesn't mean userland can't
1683 * keep changing and reading the page contents while
1684 * we transfer the mapcount, so the pmd splitting
1685 * status is achieved setting a reserved bit in the
1686 * pmd, not by clearing the present bit.
1688 page_tail
->_mapcount
= page
->_mapcount
;
1690 BUG_ON(page_tail
->mapping
);
1691 page_tail
->mapping
= page
->mapping
;
1693 page_tail
->index
= page
->index
+ i
;
1694 page_nid_xchg_last(page_tail
, page_nid_last(page
));
1696 BUG_ON(!PageAnon(page_tail
));
1697 BUG_ON(!PageUptodate(page_tail
));
1698 BUG_ON(!PageDirty(page_tail
));
1699 BUG_ON(!PageSwapBacked(page_tail
));
1701 lru_add_page_tail(page
, page_tail
, lruvec
, list
);
1703 atomic_sub(tail_count
, &page
->_count
);
1704 BUG_ON(atomic_read(&page
->_count
) <= 0);
1706 __mod_zone_page_state(zone
, NR_ANON_TRANSPARENT_HUGEPAGES
, -1);
1707 __mod_zone_page_state(zone
, NR_ANON_PAGES
, HPAGE_PMD_NR
);
1709 ClearPageCompound(page
);
1710 compound_unlock(page
);
1711 spin_unlock_irq(&zone
->lru_lock
);
1713 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
1714 struct page
*page_tail
= page
+ i
;
1715 BUG_ON(page_count(page_tail
) <= 0);
1717 * Tail pages may be freed if there wasn't any mapping
1718 * like if add_to_swap() is running on a lru page that
1719 * had its mapping zapped. And freeing these pages
1720 * requires taking the lru_lock so we do the put_page
1721 * of the tail pages after the split is complete.
1723 put_page(page_tail
);
1727 * Only the head page (now become a regular page) is required
1728 * to be pinned by the caller.
1730 BUG_ON(page_count(page
) <= 0);
1733 static int __split_huge_page_map(struct page
*page
,
1734 struct vm_area_struct
*vma
,
1735 unsigned long address
)
1737 struct mm_struct
*mm
= vma
->vm_mm
;
1741 unsigned long haddr
;
1743 spin_lock(&mm
->page_table_lock
);
1744 pmd
= page_check_address_pmd(page
, mm
, address
,
1745 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG
);
1747 pgtable
= pgtable_trans_huge_withdraw(mm
);
1748 pmd_populate(mm
, &_pmd
, pgtable
);
1749 if (pmd_write(*pmd
))
1750 BUG_ON(page_mapcount(page
) != 1);
1753 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1755 BUG_ON(PageCompound(page
+i
));
1757 * Note that pmd_numa is not transferred deliberately
1758 * to avoid any possibility that pte_numa leaks to
1759 * a PROT_NONE VMA by accident.
1761 entry
= mk_pte(page
+ i
, vma
->vm_page_prot
);
1762 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1763 if (!pmd_write(*pmd
))
1764 entry
= pte_wrprotect(entry
);
1765 if (!pmd_young(*pmd
))
1766 entry
= pte_mkold(entry
);
1767 pte
= pte_offset_map(&_pmd
, haddr
);
1768 BUG_ON(!pte_none(*pte
));
1769 set_pte_at(mm
, haddr
, pte
, entry
);
1773 smp_wmb(); /* make pte visible before pmd */
1775 * Up to this point the pmd is present and huge and
1776 * userland has the whole access to the hugepage
1777 * during the split (which happens in place). If we
1778 * overwrite the pmd with the not-huge version
1779 * pointing to the pte here (which of course we could
1780 * if all CPUs were bug free), userland could trigger
1781 * a small page size TLB miss on the small sized TLB
1782 * while the hugepage TLB entry is still established
1783 * in the huge TLB. Some CPU doesn't like that. See
1784 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1785 * Erratum 383 on page 93. Intel should be safe but is
1786 * also warns that it's only safe if the permission
1787 * and cache attributes of the two entries loaded in
1788 * the two TLB is identical (which should be the case
1789 * here). But it is generally safer to never allow
1790 * small and huge TLB entries for the same virtual
1791 * address to be loaded simultaneously. So instead of
1792 * doing "pmd_populate(); flush_tlb_range();" we first
1793 * mark the current pmd notpresent (atomically because
1794 * here the pmd_trans_huge and pmd_trans_splitting
1795 * must remain set at all times on the pmd until the
1796 * split is complete for this pmd), then we flush the
1797 * SMP TLB and finally we write the non-huge version
1798 * of the pmd entry with pmd_populate.
1800 pmdp_invalidate(vma
, address
, pmd
);
1801 pmd_populate(mm
, pmd
, pgtable
);
1804 spin_unlock(&mm
->page_table_lock
);
1809 /* must be called with anon_vma->root->rwsem held */
1810 static void __split_huge_page(struct page
*page
,
1811 struct anon_vma
*anon_vma
,
1812 struct list_head
*list
)
1814 int mapcount
, mapcount2
;
1815 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1816 struct anon_vma_chain
*avc
;
1818 BUG_ON(!PageHead(page
));
1819 BUG_ON(PageTail(page
));
1822 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
1823 struct vm_area_struct
*vma
= avc
->vma
;
1824 unsigned long addr
= vma_address(page
, vma
);
1825 BUG_ON(is_vma_temporary_stack(vma
));
1826 mapcount
+= __split_huge_page_splitting(page
, vma
, addr
);
1829 * It is critical that new vmas are added to the tail of the
1830 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1831 * and establishes a child pmd before
1832 * __split_huge_page_splitting() freezes the parent pmd (so if
1833 * we fail to prevent copy_huge_pmd() from running until the
1834 * whole __split_huge_page() is complete), we will still see
1835 * the newly established pmd of the child later during the
1836 * walk, to be able to set it as pmd_trans_splitting too.
1838 if (mapcount
!= page_mapcount(page
))
1839 printk(KERN_ERR
"mapcount %d page_mapcount %d\n",
1840 mapcount
, page_mapcount(page
));
1841 BUG_ON(mapcount
!= page_mapcount(page
));
1843 __split_huge_page_refcount(page
, list
);
1846 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
1847 struct vm_area_struct
*vma
= avc
->vma
;
1848 unsigned long addr
= vma_address(page
, vma
);
1849 BUG_ON(is_vma_temporary_stack(vma
));
1850 mapcount2
+= __split_huge_page_map(page
, vma
, addr
);
1852 if (mapcount
!= mapcount2
)
1853 printk(KERN_ERR
"mapcount %d mapcount2 %d page_mapcount %d\n",
1854 mapcount
, mapcount2
, page_mapcount(page
));
1855 BUG_ON(mapcount
!= mapcount2
);
1859 * Split a hugepage into normal pages. This doesn't change the position of head
1860 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1861 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1862 * from the hugepage.
1863 * Return 0 if the hugepage is split successfully otherwise return 1.
1865 int split_huge_page_to_list(struct page
*page
, struct list_head
*list
)
1867 struct anon_vma
*anon_vma
;
1870 BUG_ON(is_huge_zero_page(page
));
1871 BUG_ON(!PageAnon(page
));
1874 * The caller does not necessarily hold an mmap_sem that would prevent
1875 * the anon_vma disappearing so we first we take a reference to it
1876 * and then lock the anon_vma for write. This is similar to
1877 * page_lock_anon_vma_read except the write lock is taken to serialise
1878 * against parallel split or collapse operations.
1880 anon_vma
= page_get_anon_vma(page
);
1883 anon_vma_lock_write(anon_vma
);
1886 if (!PageCompound(page
))
1889 BUG_ON(!PageSwapBacked(page
));
1890 __split_huge_page(page
, anon_vma
, list
);
1891 count_vm_event(THP_SPLIT
);
1893 BUG_ON(PageCompound(page
));
1895 anon_vma_unlock_write(anon_vma
);
1896 put_anon_vma(anon_vma
);
1901 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1903 int hugepage_madvise(struct vm_area_struct
*vma
,
1904 unsigned long *vm_flags
, int advice
)
1906 struct mm_struct
*mm
= vma
->vm_mm
;
1911 * Be somewhat over-protective like KSM for now!
1913 if (*vm_flags
& (VM_HUGEPAGE
| VM_NO_THP
))
1915 if (mm
->def_flags
& VM_NOHUGEPAGE
)
1917 *vm_flags
&= ~VM_NOHUGEPAGE
;
1918 *vm_flags
|= VM_HUGEPAGE
;
1920 * If the vma become good for khugepaged to scan,
1921 * register it here without waiting a page fault that
1922 * may not happen any time soon.
1924 if (unlikely(khugepaged_enter_vma_merge(vma
)))
1927 case MADV_NOHUGEPAGE
:
1929 * Be somewhat over-protective like KSM for now!
1931 if (*vm_flags
& (VM_NOHUGEPAGE
| VM_NO_THP
))
1933 *vm_flags
&= ~VM_HUGEPAGE
;
1934 *vm_flags
|= VM_NOHUGEPAGE
;
1936 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1937 * this vma even if we leave the mm registered in khugepaged if
1938 * it got registered before VM_NOHUGEPAGE was set.
1946 static int __init
khugepaged_slab_init(void)
1948 mm_slot_cache
= kmem_cache_create("khugepaged_mm_slot",
1949 sizeof(struct mm_slot
),
1950 __alignof__(struct mm_slot
), 0, NULL
);
1957 static inline struct mm_slot
*alloc_mm_slot(void)
1959 if (!mm_slot_cache
) /* initialization failed */
1961 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
1964 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
1966 kmem_cache_free(mm_slot_cache
, mm_slot
);
1969 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
1971 struct mm_slot
*mm_slot
;
1973 hash_for_each_possible(mm_slots_hash
, mm_slot
, hash
, (unsigned long)mm
)
1974 if (mm
== mm_slot
->mm
)
1980 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
1981 struct mm_slot
*mm_slot
)
1984 hash_add(mm_slots_hash
, &mm_slot
->hash
, (long)mm
);
1987 static inline int khugepaged_test_exit(struct mm_struct
*mm
)
1989 return atomic_read(&mm
->mm_users
) == 0;
1992 int __khugepaged_enter(struct mm_struct
*mm
)
1994 struct mm_slot
*mm_slot
;
1997 mm_slot
= alloc_mm_slot();
2001 /* __khugepaged_exit() must not run from under us */
2002 VM_BUG_ON(khugepaged_test_exit(mm
));
2003 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE
, &mm
->flags
))) {
2004 free_mm_slot(mm_slot
);
2008 spin_lock(&khugepaged_mm_lock
);
2009 insert_to_mm_slots_hash(mm
, mm_slot
);
2011 * Insert just behind the scanning cursor, to let the area settle
2014 wakeup
= list_empty(&khugepaged_scan
.mm_head
);
2015 list_add_tail(&mm_slot
->mm_node
, &khugepaged_scan
.mm_head
);
2016 spin_unlock(&khugepaged_mm_lock
);
2018 atomic_inc(&mm
->mm_count
);
2020 wake_up_interruptible(&khugepaged_wait
);
2025 int khugepaged_enter_vma_merge(struct vm_area_struct
*vma
)
2027 unsigned long hstart
, hend
;
2030 * Not yet faulted in so we will register later in the
2031 * page fault if needed.
2035 /* khugepaged not yet working on file or special mappings */
2037 VM_BUG_ON(vma
->vm_flags
& VM_NO_THP
);
2038 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2039 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2041 return khugepaged_enter(vma
);
2045 void __khugepaged_exit(struct mm_struct
*mm
)
2047 struct mm_slot
*mm_slot
;
2050 spin_lock(&khugepaged_mm_lock
);
2051 mm_slot
= get_mm_slot(mm
);
2052 if (mm_slot
&& khugepaged_scan
.mm_slot
!= mm_slot
) {
2053 hash_del(&mm_slot
->hash
);
2054 list_del(&mm_slot
->mm_node
);
2057 spin_unlock(&khugepaged_mm_lock
);
2060 clear_bit(MMF_VM_HUGEPAGE
, &mm
->flags
);
2061 free_mm_slot(mm_slot
);
2063 } else if (mm_slot
) {
2065 * This is required to serialize against
2066 * khugepaged_test_exit() (which is guaranteed to run
2067 * under mmap sem read mode). Stop here (after we
2068 * return all pagetables will be destroyed) until
2069 * khugepaged has finished working on the pagetables
2070 * under the mmap_sem.
2072 down_write(&mm
->mmap_sem
);
2073 up_write(&mm
->mmap_sem
);
2077 static void release_pte_page(struct page
*page
)
2079 /* 0 stands for page_is_file_cache(page) == false */
2080 dec_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
2082 putback_lru_page(page
);
2085 static void release_pte_pages(pte_t
*pte
, pte_t
*_pte
)
2087 while (--_pte
>= pte
) {
2088 pte_t pteval
= *_pte
;
2089 if (!pte_none(pteval
))
2090 release_pte_page(pte_page(pteval
));
2094 static int __collapse_huge_page_isolate(struct vm_area_struct
*vma
,
2095 unsigned long address
,
2100 int referenced
= 0, none
= 0;
2101 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
2102 _pte
++, address
+= PAGE_SIZE
) {
2103 pte_t pteval
= *_pte
;
2104 if (pte_none(pteval
)) {
2105 if (++none
<= khugepaged_max_ptes_none
)
2110 if (!pte_present(pteval
) || !pte_write(pteval
))
2112 page
= vm_normal_page(vma
, address
, pteval
);
2113 if (unlikely(!page
))
2116 VM_BUG_ON(PageCompound(page
));
2117 BUG_ON(!PageAnon(page
));
2118 VM_BUG_ON(!PageSwapBacked(page
));
2120 /* cannot use mapcount: can't collapse if there's a gup pin */
2121 if (page_count(page
) != 1)
2124 * We can do it before isolate_lru_page because the
2125 * page can't be freed from under us. NOTE: PG_lock
2126 * is needed to serialize against split_huge_page
2127 * when invoked from the VM.
2129 if (!trylock_page(page
))
2132 * Isolate the page to avoid collapsing an hugepage
2133 * currently in use by the VM.
2135 if (isolate_lru_page(page
)) {
2139 /* 0 stands for page_is_file_cache(page) == false */
2140 inc_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
2141 VM_BUG_ON(!PageLocked(page
));
2142 VM_BUG_ON(PageLRU(page
));
2144 /* If there is no mapped pte young don't collapse the page */
2145 if (pte_young(pteval
) || PageReferenced(page
) ||
2146 mmu_notifier_test_young(vma
->vm_mm
, address
))
2149 if (likely(referenced
))
2152 release_pte_pages(pte
, _pte
);
2156 static void __collapse_huge_page_copy(pte_t
*pte
, struct page
*page
,
2157 struct vm_area_struct
*vma
,
2158 unsigned long address
,
2162 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
; _pte
++) {
2163 pte_t pteval
= *_pte
;
2164 struct page
*src_page
;
2166 if (pte_none(pteval
)) {
2167 clear_user_highpage(page
, address
);
2168 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, 1);
2170 src_page
= pte_page(pteval
);
2171 copy_user_highpage(page
, src_page
, address
, vma
);
2172 VM_BUG_ON(page_mapcount(src_page
) != 1);
2173 release_pte_page(src_page
);
2175 * ptl mostly unnecessary, but preempt has to
2176 * be disabled to update the per-cpu stats
2177 * inside page_remove_rmap().
2181 * paravirt calls inside pte_clear here are
2184 pte_clear(vma
->vm_mm
, address
, _pte
);
2185 page_remove_rmap(src_page
);
2187 free_page_and_swap_cache(src_page
);
2190 address
+= PAGE_SIZE
;
2195 static void khugepaged_alloc_sleep(void)
2197 wait_event_freezable_timeout(khugepaged_wait
, false,
2198 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs
));
2202 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
2204 if (IS_ERR(*hpage
)) {
2210 khugepaged_alloc_sleep();
2211 } else if (*hpage
) {
2220 *khugepaged_alloc_page(struct page
**hpage
, struct mm_struct
*mm
,
2221 struct vm_area_struct
*vma
, unsigned long address
,
2226 * Allocate the page while the vma is still valid and under
2227 * the mmap_sem read mode so there is no memory allocation
2228 * later when we take the mmap_sem in write mode. This is more
2229 * friendly behavior (OTOH it may actually hide bugs) to
2230 * filesystems in userland with daemons allocating memory in
2231 * the userland I/O paths. Allocating memory with the
2232 * mmap_sem in read mode is good idea also to allow greater
2235 *hpage
= alloc_hugepage_vma(khugepaged_defrag(), vma
, address
,
2236 node
, __GFP_OTHER_NODE
);
2239 * After allocating the hugepage, release the mmap_sem read lock in
2240 * preparation for taking it in write mode.
2242 up_read(&mm
->mmap_sem
);
2243 if (unlikely(!*hpage
)) {
2244 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
2245 *hpage
= ERR_PTR(-ENOMEM
);
2249 count_vm_event(THP_COLLAPSE_ALLOC
);
2253 static struct page
*khugepaged_alloc_hugepage(bool *wait
)
2258 hpage
= alloc_hugepage(khugepaged_defrag());
2260 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
2265 khugepaged_alloc_sleep();
2267 count_vm_event(THP_COLLAPSE_ALLOC
);
2268 } while (unlikely(!hpage
) && likely(khugepaged_enabled()));
2273 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
2276 *hpage
= khugepaged_alloc_hugepage(wait
);
2278 if (unlikely(!*hpage
))
2285 *khugepaged_alloc_page(struct page
**hpage
, struct mm_struct
*mm
,
2286 struct vm_area_struct
*vma
, unsigned long address
,
2289 up_read(&mm
->mmap_sem
);
2295 static bool hugepage_vma_check(struct vm_area_struct
*vma
)
2297 if ((!(vma
->vm_flags
& VM_HUGEPAGE
) && !khugepaged_always()) ||
2298 (vma
->vm_flags
& VM_NOHUGEPAGE
))
2301 if (!vma
->anon_vma
|| vma
->vm_ops
)
2303 if (is_vma_temporary_stack(vma
))
2305 VM_BUG_ON(vma
->vm_flags
& VM_NO_THP
);
2309 static void collapse_huge_page(struct mm_struct
*mm
,
2310 unsigned long address
,
2311 struct page
**hpage
,
2312 struct vm_area_struct
*vma
,
2318 struct page
*new_page
;
2321 unsigned long hstart
, hend
;
2322 unsigned long mmun_start
; /* For mmu_notifiers */
2323 unsigned long mmun_end
; /* For mmu_notifiers */
2325 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2327 /* release the mmap_sem read lock. */
2328 new_page
= khugepaged_alloc_page(hpage
, mm
, vma
, address
, node
);
2332 if (unlikely(mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
)))
2336 * Prevent all access to pagetables with the exception of
2337 * gup_fast later hanlded by the ptep_clear_flush and the VM
2338 * handled by the anon_vma lock + PG_lock.
2340 down_write(&mm
->mmap_sem
);
2341 if (unlikely(khugepaged_test_exit(mm
)))
2344 vma
= find_vma(mm
, address
);
2347 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2348 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2349 if (address
< hstart
|| address
+ HPAGE_PMD_SIZE
> hend
)
2351 if (!hugepage_vma_check(vma
))
2353 pmd
= mm_find_pmd(mm
, address
);
2356 if (pmd_trans_huge(*pmd
))
2359 anon_vma_lock_write(vma
->anon_vma
);
2361 pte
= pte_offset_map(pmd
, address
);
2362 ptl
= pte_lockptr(mm
, pmd
);
2364 mmun_start
= address
;
2365 mmun_end
= address
+ HPAGE_PMD_SIZE
;
2366 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2367 spin_lock(&mm
->page_table_lock
); /* probably unnecessary */
2369 * After this gup_fast can't run anymore. This also removes
2370 * any huge TLB entry from the CPU so we won't allow
2371 * huge and small TLB entries for the same virtual address
2372 * to avoid the risk of CPU bugs in that area.
2374 _pmd
= pmdp_clear_flush(vma
, address
, pmd
);
2375 spin_unlock(&mm
->page_table_lock
);
2376 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2379 isolated
= __collapse_huge_page_isolate(vma
, address
, pte
);
2382 if (unlikely(!isolated
)) {
2384 spin_lock(&mm
->page_table_lock
);
2385 BUG_ON(!pmd_none(*pmd
));
2387 * We can only use set_pmd_at when establishing
2388 * hugepmds and never for establishing regular pmds that
2389 * points to regular pagetables. Use pmd_populate for that
2391 pmd_populate(mm
, pmd
, pmd_pgtable(_pmd
));
2392 spin_unlock(&mm
->page_table_lock
);
2393 anon_vma_unlock_write(vma
->anon_vma
);
2398 * All pages are isolated and locked so anon_vma rmap
2399 * can't run anymore.
2401 anon_vma_unlock_write(vma
->anon_vma
);
2403 __collapse_huge_page_copy(pte
, new_page
, vma
, address
, ptl
);
2405 __SetPageUptodate(new_page
);
2406 pgtable
= pmd_pgtable(_pmd
);
2408 _pmd
= mk_huge_pmd(new_page
, vma
);
2411 * spin_lock() below is not the equivalent of smp_wmb(), so
2412 * this is needed to avoid the copy_huge_page writes to become
2413 * visible after the set_pmd_at() write.
2417 spin_lock(&mm
->page_table_lock
);
2418 BUG_ON(!pmd_none(*pmd
));
2419 page_add_new_anon_rmap(new_page
, vma
, address
);
2420 set_pmd_at(mm
, address
, pmd
, _pmd
);
2421 update_mmu_cache_pmd(vma
, address
, pmd
);
2422 pgtable_trans_huge_deposit(mm
, pgtable
);
2423 spin_unlock(&mm
->page_table_lock
);
2427 khugepaged_pages_collapsed
++;
2429 up_write(&mm
->mmap_sem
);
2433 mem_cgroup_uncharge_page(new_page
);
2437 static int khugepaged_scan_pmd(struct mm_struct
*mm
,
2438 struct vm_area_struct
*vma
,
2439 unsigned long address
,
2440 struct page
**hpage
)
2444 int ret
= 0, referenced
= 0, none
= 0;
2446 unsigned long _address
;
2448 int node
= NUMA_NO_NODE
;
2450 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2452 pmd
= mm_find_pmd(mm
, address
);
2455 if (pmd_trans_huge(*pmd
))
2458 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2459 for (_address
= address
, _pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
2460 _pte
++, _address
+= PAGE_SIZE
) {
2461 pte_t pteval
= *_pte
;
2462 if (pte_none(pteval
)) {
2463 if (++none
<= khugepaged_max_ptes_none
)
2468 if (!pte_present(pteval
) || !pte_write(pteval
))
2470 page
= vm_normal_page(vma
, _address
, pteval
);
2471 if (unlikely(!page
))
2474 * Chose the node of the first page. This could
2475 * be more sophisticated and look at more pages,
2476 * but isn't for now.
2478 if (node
== NUMA_NO_NODE
)
2479 node
= page_to_nid(page
);
2480 VM_BUG_ON(PageCompound(page
));
2481 if (!PageLRU(page
) || PageLocked(page
) || !PageAnon(page
))
2483 /* cannot use mapcount: can't collapse if there's a gup pin */
2484 if (page_count(page
) != 1)
2486 if (pte_young(pteval
) || PageReferenced(page
) ||
2487 mmu_notifier_test_young(vma
->vm_mm
, address
))
2493 pte_unmap_unlock(pte
, ptl
);
2495 /* collapse_huge_page will return with the mmap_sem released */
2496 collapse_huge_page(mm
, address
, hpage
, vma
, node
);
2501 static void collect_mm_slot(struct mm_slot
*mm_slot
)
2503 struct mm_struct
*mm
= mm_slot
->mm
;
2505 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2507 if (khugepaged_test_exit(mm
)) {
2509 hash_del(&mm_slot
->hash
);
2510 list_del(&mm_slot
->mm_node
);
2513 * Not strictly needed because the mm exited already.
2515 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2518 /* khugepaged_mm_lock actually not necessary for the below */
2519 free_mm_slot(mm_slot
);
2524 static unsigned int khugepaged_scan_mm_slot(unsigned int pages
,
2525 struct page
**hpage
)
2526 __releases(&khugepaged_mm_lock
)
2527 __acquires(&khugepaged_mm_lock
)
2529 struct mm_slot
*mm_slot
;
2530 struct mm_struct
*mm
;
2531 struct vm_area_struct
*vma
;
2535 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2537 if (khugepaged_scan
.mm_slot
)
2538 mm_slot
= khugepaged_scan
.mm_slot
;
2540 mm_slot
= list_entry(khugepaged_scan
.mm_head
.next
,
2541 struct mm_slot
, mm_node
);
2542 khugepaged_scan
.address
= 0;
2543 khugepaged_scan
.mm_slot
= mm_slot
;
2545 spin_unlock(&khugepaged_mm_lock
);
2548 down_read(&mm
->mmap_sem
);
2549 if (unlikely(khugepaged_test_exit(mm
)))
2552 vma
= find_vma(mm
, khugepaged_scan
.address
);
2555 for (; vma
; vma
= vma
->vm_next
) {
2556 unsigned long hstart
, hend
;
2559 if (unlikely(khugepaged_test_exit(mm
))) {
2563 if (!hugepage_vma_check(vma
)) {
2568 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2569 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2572 if (khugepaged_scan
.address
> hend
)
2574 if (khugepaged_scan
.address
< hstart
)
2575 khugepaged_scan
.address
= hstart
;
2576 VM_BUG_ON(khugepaged_scan
.address
& ~HPAGE_PMD_MASK
);
2578 while (khugepaged_scan
.address
< hend
) {
2581 if (unlikely(khugepaged_test_exit(mm
)))
2582 goto breakouterloop
;
2584 VM_BUG_ON(khugepaged_scan
.address
< hstart
||
2585 khugepaged_scan
.address
+ HPAGE_PMD_SIZE
>
2587 ret
= khugepaged_scan_pmd(mm
, vma
,
2588 khugepaged_scan
.address
,
2590 /* move to next address */
2591 khugepaged_scan
.address
+= HPAGE_PMD_SIZE
;
2592 progress
+= HPAGE_PMD_NR
;
2594 /* we released mmap_sem so break loop */
2595 goto breakouterloop_mmap_sem
;
2596 if (progress
>= pages
)
2597 goto breakouterloop
;
2601 up_read(&mm
->mmap_sem
); /* exit_mmap will destroy ptes after this */
2602 breakouterloop_mmap_sem
:
2604 spin_lock(&khugepaged_mm_lock
);
2605 VM_BUG_ON(khugepaged_scan
.mm_slot
!= mm_slot
);
2607 * Release the current mm_slot if this mm is about to die, or
2608 * if we scanned all vmas of this mm.
2610 if (khugepaged_test_exit(mm
) || !vma
) {
2612 * Make sure that if mm_users is reaching zero while
2613 * khugepaged runs here, khugepaged_exit will find
2614 * mm_slot not pointing to the exiting mm.
2616 if (mm_slot
->mm_node
.next
!= &khugepaged_scan
.mm_head
) {
2617 khugepaged_scan
.mm_slot
= list_entry(
2618 mm_slot
->mm_node
.next
,
2619 struct mm_slot
, mm_node
);
2620 khugepaged_scan
.address
= 0;
2622 khugepaged_scan
.mm_slot
= NULL
;
2623 khugepaged_full_scans
++;
2626 collect_mm_slot(mm_slot
);
2632 static int khugepaged_has_work(void)
2634 return !list_empty(&khugepaged_scan
.mm_head
) &&
2635 khugepaged_enabled();
2638 static int khugepaged_wait_event(void)
2640 return !list_empty(&khugepaged_scan
.mm_head
) ||
2641 kthread_should_stop();
2644 static void khugepaged_do_scan(void)
2646 struct page
*hpage
= NULL
;
2647 unsigned int progress
= 0, pass_through_head
= 0;
2648 unsigned int pages
= khugepaged_pages_to_scan
;
2651 barrier(); /* write khugepaged_pages_to_scan to local stack */
2653 while (progress
< pages
) {
2654 if (!khugepaged_prealloc_page(&hpage
, &wait
))
2659 if (unlikely(kthread_should_stop() || freezing(current
)))
2662 spin_lock(&khugepaged_mm_lock
);
2663 if (!khugepaged_scan
.mm_slot
)
2664 pass_through_head
++;
2665 if (khugepaged_has_work() &&
2666 pass_through_head
< 2)
2667 progress
+= khugepaged_scan_mm_slot(pages
- progress
,
2671 spin_unlock(&khugepaged_mm_lock
);
2674 if (!IS_ERR_OR_NULL(hpage
))
2678 static void khugepaged_wait_work(void)
2682 if (khugepaged_has_work()) {
2683 if (!khugepaged_scan_sleep_millisecs
)
2686 wait_event_freezable_timeout(khugepaged_wait
,
2687 kthread_should_stop(),
2688 msecs_to_jiffies(khugepaged_scan_sleep_millisecs
));
2692 if (khugepaged_enabled())
2693 wait_event_freezable(khugepaged_wait
, khugepaged_wait_event());
2696 static int khugepaged(void *none
)
2698 struct mm_slot
*mm_slot
;
2701 set_user_nice(current
, 19);
2703 while (!kthread_should_stop()) {
2704 khugepaged_do_scan();
2705 khugepaged_wait_work();
2708 spin_lock(&khugepaged_mm_lock
);
2709 mm_slot
= khugepaged_scan
.mm_slot
;
2710 khugepaged_scan
.mm_slot
= NULL
;
2712 collect_mm_slot(mm_slot
);
2713 spin_unlock(&khugepaged_mm_lock
);
2717 static void __split_huge_zero_page_pmd(struct vm_area_struct
*vma
,
2718 unsigned long haddr
, pmd_t
*pmd
)
2720 struct mm_struct
*mm
= vma
->vm_mm
;
2725 pmdp_clear_flush(vma
, haddr
, pmd
);
2726 /* leave pmd empty until pte is filled */
2728 pgtable
= pgtable_trans_huge_withdraw(mm
);
2729 pmd_populate(mm
, &_pmd
, pgtable
);
2731 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
2733 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
2734 entry
= pte_mkspecial(entry
);
2735 pte
= pte_offset_map(&_pmd
, haddr
);
2736 VM_BUG_ON(!pte_none(*pte
));
2737 set_pte_at(mm
, haddr
, pte
, entry
);
2740 smp_wmb(); /* make pte visible before pmd */
2741 pmd_populate(mm
, pmd
, pgtable
);
2742 put_huge_zero_page();
2745 void __split_huge_page_pmd(struct vm_area_struct
*vma
, unsigned long address
,
2749 struct mm_struct
*mm
= vma
->vm_mm
;
2750 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
2751 unsigned long mmun_start
; /* For mmu_notifiers */
2752 unsigned long mmun_end
; /* For mmu_notifiers */
2754 BUG_ON(vma
->vm_start
> haddr
|| vma
->vm_end
< haddr
+ HPAGE_PMD_SIZE
);
2757 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
2759 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2760 spin_lock(&mm
->page_table_lock
);
2761 if (unlikely(!pmd_trans_huge(*pmd
))) {
2762 spin_unlock(&mm
->page_table_lock
);
2763 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2766 if (is_huge_zero_pmd(*pmd
)) {
2767 __split_huge_zero_page_pmd(vma
, haddr
, pmd
);
2768 spin_unlock(&mm
->page_table_lock
);
2769 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2772 page
= pmd_page(*pmd
);
2773 VM_BUG_ON(!page_count(page
));
2775 spin_unlock(&mm
->page_table_lock
);
2776 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2778 split_huge_page(page
);
2783 * We don't always have down_write of mmap_sem here: a racing
2784 * do_huge_pmd_wp_page() might have copied-on-write to another
2785 * huge page before our split_huge_page() got the anon_vma lock.
2787 if (unlikely(pmd_trans_huge(*pmd
)))
2791 void split_huge_page_pmd_mm(struct mm_struct
*mm
, unsigned long address
,
2794 struct vm_area_struct
*vma
;
2796 vma
= find_vma(mm
, address
);
2797 BUG_ON(vma
== NULL
);
2798 split_huge_page_pmd(vma
, address
, pmd
);
2801 static void split_huge_page_address(struct mm_struct
*mm
,
2802 unsigned long address
)
2806 VM_BUG_ON(!(address
& ~HPAGE_PMD_MASK
));
2808 pmd
= mm_find_pmd(mm
, address
);
2812 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2813 * materialize from under us.
2815 split_huge_page_pmd_mm(mm
, address
, pmd
);
2818 void __vma_adjust_trans_huge(struct vm_area_struct
*vma
,
2819 unsigned long start
,
2824 * If the new start address isn't hpage aligned and it could
2825 * previously contain an hugepage: check if we need to split
2828 if (start
& ~HPAGE_PMD_MASK
&&
2829 (start
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2830 (start
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2831 split_huge_page_address(vma
->vm_mm
, start
);
2834 * If the new end address isn't hpage aligned and it could
2835 * previously contain an hugepage: check if we need to split
2838 if (end
& ~HPAGE_PMD_MASK
&&
2839 (end
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2840 (end
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2841 split_huge_page_address(vma
->vm_mm
, end
);
2844 * If we're also updating the vma->vm_next->vm_start, if the new
2845 * vm_next->vm_start isn't page aligned and it could previously
2846 * contain an hugepage: check if we need to split an huge pmd.
2848 if (adjust_next
> 0) {
2849 struct vm_area_struct
*next
= vma
->vm_next
;
2850 unsigned long nstart
= next
->vm_start
;
2851 nstart
+= adjust_next
<< PAGE_SHIFT
;
2852 if (nstart
& ~HPAGE_PMD_MASK
&&
2853 (nstart
& HPAGE_PMD_MASK
) >= next
->vm_start
&&
2854 (nstart
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= next
->vm_end
)
2855 split_huge_page_address(next
->vm_mm
, nstart
);