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 disabled in order that avoid
31 * to risk increase the memory footprint of applications without a guaranteed
32 * benefit. When transparent hugepage support is enabled, is for all mappings,
33 * and khugepaged scans all mappings.
34 * Defrag is invoked by khugepaged hugepage allocations and by page faults
35 * for all hugepage allocations.
37 unsigned long transparent_hugepage_flags __read_mostly
=
38 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
39 (1<<TRANSPARENT_HUGEPAGE_FLAG
)|
41 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
42 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
)|
44 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
)|
45 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
)|
46 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
48 /* default scan 8*512 pte (or vmas) every 30 second */
49 static unsigned int khugepaged_pages_to_scan __read_mostly
= HPAGE_PMD_NR
*8;
50 static unsigned int khugepaged_pages_collapsed
;
51 static unsigned int khugepaged_full_scans
;
52 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly
= 10000;
53 /* during fragmentation poll the hugepage allocator once every minute */
54 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly
= 60000;
55 static struct task_struct
*khugepaged_thread __read_mostly
;
56 static DEFINE_MUTEX(khugepaged_mutex
);
57 static DEFINE_SPINLOCK(khugepaged_mm_lock
);
58 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait
);
60 * default collapse hugepages if there is at least one pte mapped like
61 * it would have happened if the vma was large enough during page
64 static unsigned int khugepaged_max_ptes_none __read_mostly
= HPAGE_PMD_NR
-1;
66 static int khugepaged(void *none
);
67 static int khugepaged_slab_init(void);
69 #define MM_SLOTS_HASH_BITS 10
70 static __read_mostly
DEFINE_HASHTABLE(mm_slots_hash
, MM_SLOTS_HASH_BITS
);
72 static struct kmem_cache
*mm_slot_cache __read_mostly
;
75 * struct mm_slot - hash lookup from mm to mm_slot
76 * @hash: hash collision list
77 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
78 * @mm: the mm that this information is valid for
81 struct hlist_node hash
;
82 struct list_head mm_node
;
87 * struct khugepaged_scan - cursor for scanning
88 * @mm_head: the head of the mm list to scan
89 * @mm_slot: the current mm_slot we are scanning
90 * @address: the next address inside that to be scanned
92 * There is only the one khugepaged_scan instance of this cursor structure.
94 struct khugepaged_scan
{
95 struct list_head mm_head
;
96 struct mm_slot
*mm_slot
;
97 unsigned long address
;
99 static struct khugepaged_scan khugepaged_scan
= {
100 .mm_head
= LIST_HEAD_INIT(khugepaged_scan
.mm_head
),
104 static int set_recommended_min_free_kbytes(void)
108 unsigned long recommended_min
;
110 if (!khugepaged_enabled())
113 for_each_populated_zone(zone
)
116 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
117 recommended_min
= pageblock_nr_pages
* nr_zones
* 2;
120 * Make sure that on average at least two pageblocks are almost free
121 * of another type, one for a migratetype to fall back to and a
122 * second to avoid subsequent fallbacks of other types There are 3
123 * MIGRATE_TYPES we care about.
125 recommended_min
+= pageblock_nr_pages
* nr_zones
*
126 MIGRATE_PCPTYPES
* MIGRATE_PCPTYPES
;
128 /* don't ever allow to reserve more than 5% of the lowmem */
129 recommended_min
= min(recommended_min
,
130 (unsigned long) nr_free_buffer_pages() / 20);
131 recommended_min
<<= (PAGE_SHIFT
-10);
133 if (recommended_min
> min_free_kbytes
)
134 min_free_kbytes
= recommended_min
;
135 setup_per_zone_wmarks();
138 late_initcall(set_recommended_min_free_kbytes
);
140 static int start_khugepaged(void)
143 if (khugepaged_enabled()) {
144 if (!khugepaged_thread
)
145 khugepaged_thread
= kthread_run(khugepaged
, NULL
,
147 if (unlikely(IS_ERR(khugepaged_thread
))) {
149 "khugepaged: kthread_run(khugepaged) failed\n");
150 err
= PTR_ERR(khugepaged_thread
);
151 khugepaged_thread
= NULL
;
154 if (!list_empty(&khugepaged_scan
.mm_head
))
155 wake_up_interruptible(&khugepaged_wait
);
157 set_recommended_min_free_kbytes();
158 } else if (khugepaged_thread
) {
159 kthread_stop(khugepaged_thread
);
160 khugepaged_thread
= NULL
;
166 static atomic_t huge_zero_refcount
;
167 static struct page
*huge_zero_page __read_mostly
;
169 static inline bool is_huge_zero_page(struct page
*page
)
171 return ACCESS_ONCE(huge_zero_page
) == page
;
174 static inline bool is_huge_zero_pmd(pmd_t pmd
)
176 return is_huge_zero_page(pmd_page(pmd
));
179 static struct page
*get_huge_zero_page(void)
181 struct page
*zero_page
;
183 if (likely(atomic_inc_not_zero(&huge_zero_refcount
)))
184 return ACCESS_ONCE(huge_zero_page
);
186 zero_page
= alloc_pages((GFP_TRANSHUGE
| __GFP_ZERO
) & ~__GFP_MOVABLE
,
189 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED
);
192 count_vm_event(THP_ZERO_PAGE_ALLOC
);
194 if (cmpxchg(&huge_zero_page
, NULL
, zero_page
)) {
196 __free_page(zero_page
);
200 /* We take additional reference here. It will be put back by shrinker */
201 atomic_set(&huge_zero_refcount
, 2);
203 return ACCESS_ONCE(huge_zero_page
);
206 static void put_huge_zero_page(void)
209 * Counter should never go to zero here. Only shrinker can put
212 BUG_ON(atomic_dec_and_test(&huge_zero_refcount
));
215 static unsigned long shrink_huge_zero_page_count(struct shrinker
*shrink
,
216 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;
222 static unsigned long shrink_huge_zero_page_scan(struct shrinker
*shrink
,
223 struct shrink_control
*sc
)
225 if (atomic_cmpxchg(&huge_zero_refcount
, 1, 0) == 1) {
226 struct page
*zero_page
= xchg(&huge_zero_page
, NULL
);
227 BUG_ON(zero_page
== NULL
);
228 __free_page(zero_page
);
235 static struct shrinker huge_zero_page_shrinker
= {
236 .count_objects
= shrink_huge_zero_page_count
,
237 .scan_objects
= shrink_huge_zero_page_scan
,
238 .seeks
= DEFAULT_SEEKS
,
243 static ssize_t
double_flag_show(struct kobject
*kobj
,
244 struct kobj_attribute
*attr
, char *buf
,
245 enum transparent_hugepage_flag enabled
,
246 enum transparent_hugepage_flag req_madv
)
248 if (test_bit(enabled
, &transparent_hugepage_flags
)) {
249 VM_BUG_ON(test_bit(req_madv
, &transparent_hugepage_flags
));
250 return sprintf(buf
, "[always] madvise never\n");
251 } else if (test_bit(req_madv
, &transparent_hugepage_flags
))
252 return sprintf(buf
, "always [madvise] never\n");
254 return sprintf(buf
, "always madvise [never]\n");
256 static ssize_t
double_flag_store(struct kobject
*kobj
,
257 struct kobj_attribute
*attr
,
258 const char *buf
, size_t count
,
259 enum transparent_hugepage_flag enabled
,
260 enum transparent_hugepage_flag req_madv
)
262 if (!memcmp("always", buf
,
263 min(sizeof("always")-1, count
))) {
264 set_bit(enabled
, &transparent_hugepage_flags
);
265 clear_bit(req_madv
, &transparent_hugepage_flags
);
266 } else if (!memcmp("madvise", buf
,
267 min(sizeof("madvise")-1, count
))) {
268 clear_bit(enabled
, &transparent_hugepage_flags
);
269 set_bit(req_madv
, &transparent_hugepage_flags
);
270 } else if (!memcmp("never", buf
,
271 min(sizeof("never")-1, count
))) {
272 clear_bit(enabled
, &transparent_hugepage_flags
);
273 clear_bit(req_madv
, &transparent_hugepage_flags
);
280 static ssize_t
enabled_show(struct kobject
*kobj
,
281 struct kobj_attribute
*attr
, char *buf
)
283 return double_flag_show(kobj
, attr
, buf
,
284 TRANSPARENT_HUGEPAGE_FLAG
,
285 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
287 static ssize_t
enabled_store(struct kobject
*kobj
,
288 struct kobj_attribute
*attr
,
289 const char *buf
, size_t count
)
293 ret
= double_flag_store(kobj
, attr
, buf
, count
,
294 TRANSPARENT_HUGEPAGE_FLAG
,
295 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
300 mutex_lock(&khugepaged_mutex
);
301 err
= start_khugepaged();
302 mutex_unlock(&khugepaged_mutex
);
310 static struct kobj_attribute enabled_attr
=
311 __ATTR(enabled
, 0644, enabled_show
, enabled_store
);
313 static ssize_t
single_flag_show(struct kobject
*kobj
,
314 struct kobj_attribute
*attr
, char *buf
,
315 enum transparent_hugepage_flag flag
)
317 return sprintf(buf
, "%d\n",
318 !!test_bit(flag
, &transparent_hugepage_flags
));
321 static ssize_t
single_flag_store(struct kobject
*kobj
,
322 struct kobj_attribute
*attr
,
323 const char *buf
, size_t count
,
324 enum transparent_hugepage_flag flag
)
329 ret
= kstrtoul(buf
, 10, &value
);
336 set_bit(flag
, &transparent_hugepage_flags
);
338 clear_bit(flag
, &transparent_hugepage_flags
);
344 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
345 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
346 * memory just to allocate one more hugepage.
348 static ssize_t
defrag_show(struct kobject
*kobj
,
349 struct kobj_attribute
*attr
, char *buf
)
351 return double_flag_show(kobj
, attr
, buf
,
352 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
,
353 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
);
355 static ssize_t
defrag_store(struct kobject
*kobj
,
356 struct kobj_attribute
*attr
,
357 const char *buf
, size_t count
)
359 return double_flag_store(kobj
, attr
, buf
, count
,
360 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
,
361 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
);
363 static struct kobj_attribute defrag_attr
=
364 __ATTR(defrag
, 0644, defrag_show
, defrag_store
);
366 static ssize_t
use_zero_page_show(struct kobject
*kobj
,
367 struct kobj_attribute
*attr
, char *buf
)
369 return single_flag_show(kobj
, attr
, buf
,
370 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
372 static ssize_t
use_zero_page_store(struct kobject
*kobj
,
373 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
375 return single_flag_store(kobj
, attr
, buf
, count
,
376 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
378 static struct kobj_attribute use_zero_page_attr
=
379 __ATTR(use_zero_page
, 0644, use_zero_page_show
, use_zero_page_store
);
380 #ifdef CONFIG_DEBUG_VM
381 static ssize_t
debug_cow_show(struct kobject
*kobj
,
382 struct kobj_attribute
*attr
, char *buf
)
384 return single_flag_show(kobj
, attr
, buf
,
385 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
387 static ssize_t
debug_cow_store(struct kobject
*kobj
,
388 struct kobj_attribute
*attr
,
389 const char *buf
, size_t count
)
391 return single_flag_store(kobj
, attr
, buf
, count
,
392 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
394 static struct kobj_attribute debug_cow_attr
=
395 __ATTR(debug_cow
, 0644, debug_cow_show
, debug_cow_store
);
396 #endif /* CONFIG_DEBUG_VM */
398 static struct attribute
*hugepage_attr
[] = {
401 &use_zero_page_attr
.attr
,
402 #ifdef CONFIG_DEBUG_VM
403 &debug_cow_attr
.attr
,
408 static struct attribute_group hugepage_attr_group
= {
409 .attrs
= hugepage_attr
,
412 static ssize_t
scan_sleep_millisecs_show(struct kobject
*kobj
,
413 struct kobj_attribute
*attr
,
416 return sprintf(buf
, "%u\n", khugepaged_scan_sleep_millisecs
);
419 static ssize_t
scan_sleep_millisecs_store(struct kobject
*kobj
,
420 struct kobj_attribute
*attr
,
421 const char *buf
, size_t count
)
426 err
= kstrtoul(buf
, 10, &msecs
);
427 if (err
|| msecs
> UINT_MAX
)
430 khugepaged_scan_sleep_millisecs
= msecs
;
431 wake_up_interruptible(&khugepaged_wait
);
435 static struct kobj_attribute scan_sleep_millisecs_attr
=
436 __ATTR(scan_sleep_millisecs
, 0644, scan_sleep_millisecs_show
,
437 scan_sleep_millisecs_store
);
439 static ssize_t
alloc_sleep_millisecs_show(struct kobject
*kobj
,
440 struct kobj_attribute
*attr
,
443 return sprintf(buf
, "%u\n", khugepaged_alloc_sleep_millisecs
);
446 static ssize_t
alloc_sleep_millisecs_store(struct kobject
*kobj
,
447 struct kobj_attribute
*attr
,
448 const char *buf
, size_t count
)
453 err
= kstrtoul(buf
, 10, &msecs
);
454 if (err
|| msecs
> UINT_MAX
)
457 khugepaged_alloc_sleep_millisecs
= msecs
;
458 wake_up_interruptible(&khugepaged_wait
);
462 static struct kobj_attribute alloc_sleep_millisecs_attr
=
463 __ATTR(alloc_sleep_millisecs
, 0644, alloc_sleep_millisecs_show
,
464 alloc_sleep_millisecs_store
);
466 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
467 struct kobj_attribute
*attr
,
470 return sprintf(buf
, "%u\n", khugepaged_pages_to_scan
);
472 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
473 struct kobj_attribute
*attr
,
474 const char *buf
, size_t count
)
479 err
= kstrtoul(buf
, 10, &pages
);
480 if (err
|| !pages
|| pages
> UINT_MAX
)
483 khugepaged_pages_to_scan
= pages
;
487 static struct kobj_attribute pages_to_scan_attr
=
488 __ATTR(pages_to_scan
, 0644, pages_to_scan_show
,
489 pages_to_scan_store
);
491 static ssize_t
pages_collapsed_show(struct kobject
*kobj
,
492 struct kobj_attribute
*attr
,
495 return sprintf(buf
, "%u\n", khugepaged_pages_collapsed
);
497 static struct kobj_attribute pages_collapsed_attr
=
498 __ATTR_RO(pages_collapsed
);
500 static ssize_t
full_scans_show(struct kobject
*kobj
,
501 struct kobj_attribute
*attr
,
504 return sprintf(buf
, "%u\n", khugepaged_full_scans
);
506 static struct kobj_attribute full_scans_attr
=
507 __ATTR_RO(full_scans
);
509 static ssize_t
khugepaged_defrag_show(struct kobject
*kobj
,
510 struct kobj_attribute
*attr
, char *buf
)
512 return single_flag_show(kobj
, attr
, buf
,
513 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
515 static ssize_t
khugepaged_defrag_store(struct kobject
*kobj
,
516 struct kobj_attribute
*attr
,
517 const char *buf
, size_t count
)
519 return single_flag_store(kobj
, attr
, buf
, count
,
520 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
522 static struct kobj_attribute khugepaged_defrag_attr
=
523 __ATTR(defrag
, 0644, khugepaged_defrag_show
,
524 khugepaged_defrag_store
);
527 * max_ptes_none controls if khugepaged should collapse hugepages over
528 * any unmapped ptes in turn potentially increasing the memory
529 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
530 * reduce the available free memory in the system as it
531 * runs. Increasing max_ptes_none will instead potentially reduce the
532 * free memory in the system during the khugepaged scan.
534 static ssize_t
khugepaged_max_ptes_none_show(struct kobject
*kobj
,
535 struct kobj_attribute
*attr
,
538 return sprintf(buf
, "%u\n", khugepaged_max_ptes_none
);
540 static ssize_t
khugepaged_max_ptes_none_store(struct kobject
*kobj
,
541 struct kobj_attribute
*attr
,
542 const char *buf
, size_t count
)
545 unsigned long max_ptes_none
;
547 err
= kstrtoul(buf
, 10, &max_ptes_none
);
548 if (err
|| max_ptes_none
> HPAGE_PMD_NR
-1)
551 khugepaged_max_ptes_none
= max_ptes_none
;
555 static struct kobj_attribute khugepaged_max_ptes_none_attr
=
556 __ATTR(max_ptes_none
, 0644, khugepaged_max_ptes_none_show
,
557 khugepaged_max_ptes_none_store
);
559 static struct attribute
*khugepaged_attr
[] = {
560 &khugepaged_defrag_attr
.attr
,
561 &khugepaged_max_ptes_none_attr
.attr
,
562 &pages_to_scan_attr
.attr
,
563 &pages_collapsed_attr
.attr
,
564 &full_scans_attr
.attr
,
565 &scan_sleep_millisecs_attr
.attr
,
566 &alloc_sleep_millisecs_attr
.attr
,
570 static struct attribute_group khugepaged_attr_group
= {
571 .attrs
= khugepaged_attr
,
572 .name
= "khugepaged",
575 static int __init
hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
579 *hugepage_kobj
= kobject_create_and_add("transparent_hugepage", mm_kobj
);
580 if (unlikely(!*hugepage_kobj
)) {
581 printk(KERN_ERR
"hugepage: failed to create transparent hugepage kobject\n");
585 err
= sysfs_create_group(*hugepage_kobj
, &hugepage_attr_group
);
587 printk(KERN_ERR
"hugepage: failed to register transparent hugepage group\n");
591 err
= sysfs_create_group(*hugepage_kobj
, &khugepaged_attr_group
);
593 printk(KERN_ERR
"hugepage: failed to register transparent hugepage group\n");
594 goto remove_hp_group
;
600 sysfs_remove_group(*hugepage_kobj
, &hugepage_attr_group
);
602 kobject_put(*hugepage_kobj
);
606 static void __init
hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
608 sysfs_remove_group(hugepage_kobj
, &khugepaged_attr_group
);
609 sysfs_remove_group(hugepage_kobj
, &hugepage_attr_group
);
610 kobject_put(hugepage_kobj
);
613 static inline int hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
618 static inline void hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
621 #endif /* CONFIG_SYSFS */
623 static int __init
hugepage_init(void)
626 struct kobject
*hugepage_kobj
;
628 if (!has_transparent_hugepage()) {
629 transparent_hugepage_flags
= 0;
633 err
= hugepage_init_sysfs(&hugepage_kobj
);
637 err
= khugepaged_slab_init();
641 register_shrinker(&huge_zero_page_shrinker
);
644 * By default disable transparent hugepages on smaller systems,
645 * where the extra memory used could hurt more than TLB overhead
646 * is likely to save. The admin can still enable it through /sys.
648 if (totalram_pages
< (512 << (20 - PAGE_SHIFT
)))
649 transparent_hugepage_flags
= 0;
655 hugepage_exit_sysfs(hugepage_kobj
);
658 module_init(hugepage_init
)
660 static int __init
setup_transparent_hugepage(char *str
)
665 if (!strcmp(str
, "always")) {
666 set_bit(TRANSPARENT_HUGEPAGE_FLAG
,
667 &transparent_hugepage_flags
);
668 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
669 &transparent_hugepage_flags
);
671 } else if (!strcmp(str
, "madvise")) {
672 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
673 &transparent_hugepage_flags
);
674 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
675 &transparent_hugepage_flags
);
677 } else if (!strcmp(str
, "never")) {
678 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
679 &transparent_hugepage_flags
);
680 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
681 &transparent_hugepage_flags
);
687 "transparent_hugepage= cannot parse, ignored\n");
690 __setup("transparent_hugepage=", setup_transparent_hugepage
);
692 pmd_t
maybe_pmd_mkwrite(pmd_t pmd
, struct vm_area_struct
*vma
)
694 if (likely(vma
->vm_flags
& VM_WRITE
))
695 pmd
= pmd_mkwrite(pmd
);
699 static inline pmd_t
mk_huge_pmd(struct page
*page
, pgprot_t prot
)
702 entry
= mk_pmd(page
, prot
);
703 entry
= pmd_mkhuge(entry
);
707 static int __do_huge_pmd_anonymous_page(struct mm_struct
*mm
,
708 struct vm_area_struct
*vma
,
709 unsigned long haddr
, pmd_t
*pmd
,
715 VM_BUG_ON(!PageCompound(page
));
716 pgtable
= pte_alloc_one(mm
, haddr
);
717 if (unlikely(!pgtable
))
720 clear_huge_page(page
, haddr
, HPAGE_PMD_NR
);
722 * The memory barrier inside __SetPageUptodate makes sure that
723 * clear_huge_page writes become visible before the set_pmd_at()
726 __SetPageUptodate(page
);
728 ptl
= pmd_lock(mm
, pmd
);
729 if (unlikely(!pmd_none(*pmd
))) {
731 mem_cgroup_uncharge_page(page
);
733 pte_free(mm
, pgtable
);
736 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
737 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
738 page_add_new_anon_rmap(page
, vma
, haddr
);
739 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
740 set_pmd_at(mm
, haddr
, pmd
, entry
);
741 add_mm_counter(mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
742 atomic_long_inc(&mm
->nr_ptes
);
749 static inline gfp_t
alloc_hugepage_gfpmask(int defrag
, gfp_t extra_gfp
)
751 return (GFP_TRANSHUGE
& ~(defrag
? 0 : __GFP_WAIT
)) | extra_gfp
;
754 static inline struct page
*alloc_hugepage_vma(int defrag
,
755 struct vm_area_struct
*vma
,
756 unsigned long haddr
, int nd
,
759 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag
, extra_gfp
),
760 HPAGE_PMD_ORDER
, vma
, haddr
, nd
);
763 /* Caller must hold page table lock. */
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 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
775 set_pmd_at(mm
, haddr
, pmd
, entry
);
776 atomic_long_inc(&mm
->nr_ptes
);
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
;
787 if (haddr
< vma
->vm_start
|| haddr
+ HPAGE_PMD_SIZE
> vma
->vm_end
)
788 return VM_FAULT_FALLBACK
;
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()) {
797 struct page
*zero_page
;
799 pgtable
= pte_alloc_one(mm
, haddr
);
800 if (unlikely(!pgtable
))
802 zero_page
= get_huge_zero_page();
803 if (unlikely(!zero_page
)) {
804 pte_free(mm
, pgtable
);
805 count_vm_event(THP_FAULT_FALLBACK
);
806 return VM_FAULT_FALLBACK
;
808 ptl
= pmd_lock(mm
, pmd
);
809 set
= set_huge_zero_page(pgtable
, mm
, vma
, haddr
, pmd
,
813 pte_free(mm
, pgtable
);
814 put_huge_zero_page();
818 page
= alloc_hugepage_vma(transparent_hugepage_defrag(vma
),
819 vma
, haddr
, numa_node_id(), 0);
820 if (unlikely(!page
)) {
821 count_vm_event(THP_FAULT_FALLBACK
);
822 return VM_FAULT_FALLBACK
;
824 if (unlikely(mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))) {
826 count_vm_event(THP_FAULT_FALLBACK
);
827 return VM_FAULT_FALLBACK
;
829 if (unlikely(__do_huge_pmd_anonymous_page(mm
, vma
, haddr
, pmd
, page
))) {
830 mem_cgroup_uncharge_page(page
);
832 count_vm_event(THP_FAULT_FALLBACK
);
833 return VM_FAULT_FALLBACK
;
836 count_vm_event(THP_FAULT_ALLOC
);
840 int copy_huge_pmd(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
841 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
842 struct vm_area_struct
*vma
)
844 spinlock_t
*dst_ptl
, *src_ptl
;
845 struct page
*src_page
;
851 pgtable
= pte_alloc_one(dst_mm
, addr
);
852 if (unlikely(!pgtable
))
855 dst_ptl
= pmd_lock(dst_mm
, dst_pmd
);
856 src_ptl
= pmd_lockptr(src_mm
, src_pmd
);
857 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
861 if (unlikely(!pmd_trans_huge(pmd
))) {
862 pte_free(dst_mm
, pgtable
);
866 * When page table lock is held, the huge zero pmd should not be
867 * under splitting since we don't split the page itself, only pmd to
870 if (is_huge_zero_pmd(pmd
)) {
871 struct page
*zero_page
;
874 * get_huge_zero_page() will never allocate a new page here,
875 * since we already have a zero page to copy. It just takes a
878 zero_page
= get_huge_zero_page();
879 set
= set_huge_zero_page(pgtable
, dst_mm
, vma
, addr
, dst_pmd
,
881 BUG_ON(!set
); /* unexpected !pmd_none(dst_pmd) */
885 if (unlikely(pmd_trans_splitting(pmd
))) {
886 /* split huge page running from under us */
887 spin_unlock(src_ptl
);
888 spin_unlock(dst_ptl
);
889 pte_free(dst_mm
, pgtable
);
891 wait_split_huge_page(vma
->anon_vma
, src_pmd
); /* src_vma */
894 src_page
= pmd_page(pmd
);
895 VM_BUG_ON(!PageHead(src_page
));
897 page_dup_rmap(src_page
);
898 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
900 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
901 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
902 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
903 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
904 atomic_long_inc(&dst_mm
->nr_ptes
);
908 spin_unlock(src_ptl
);
909 spin_unlock(dst_ptl
);
914 void huge_pmd_set_accessed(struct mm_struct
*mm
,
915 struct vm_area_struct
*vma
,
916 unsigned long address
,
917 pmd_t
*pmd
, pmd_t orig_pmd
,
924 ptl
= pmd_lock(mm
, pmd
);
925 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
928 entry
= pmd_mkyoung(orig_pmd
);
929 haddr
= address
& HPAGE_PMD_MASK
;
930 if (pmdp_set_access_flags(vma
, haddr
, pmd
, entry
, dirty
))
931 update_mmu_cache_pmd(vma
, address
, pmd
);
937 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct
*mm
,
938 struct vm_area_struct
*vma
, unsigned long address
,
939 pmd_t
*pmd
, pmd_t orig_pmd
, unsigned long haddr
)
946 unsigned long mmun_start
; /* For mmu_notifiers */
947 unsigned long mmun_end
; /* For mmu_notifiers */
949 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
955 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
)) {
961 clear_user_highpage(page
, address
);
962 __SetPageUptodate(page
);
965 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
966 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
968 ptl
= pmd_lock(mm
, pmd
);
969 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
972 pmdp_clear_flush(vma
, haddr
, pmd
);
973 /* leave pmd empty until pte is filled */
975 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
976 pmd_populate(mm
, &_pmd
, pgtable
);
978 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
980 if (haddr
== (address
& PAGE_MASK
)) {
981 entry
= mk_pte(page
, vma
->vm_page_prot
);
982 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
983 page_add_new_anon_rmap(page
, vma
, haddr
);
985 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
986 entry
= pte_mkspecial(entry
);
988 pte
= pte_offset_map(&_pmd
, haddr
);
989 VM_BUG_ON(!pte_none(*pte
));
990 set_pte_at(mm
, haddr
, pte
, entry
);
993 smp_wmb(); /* make pte visible before pmd */
994 pmd_populate(mm
, pmd
, pgtable
);
996 put_huge_zero_page();
997 inc_mm_counter(mm
, MM_ANONPAGES
);
999 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1001 ret
|= VM_FAULT_WRITE
;
1006 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1007 mem_cgroup_uncharge_page(page
);
1012 static int do_huge_pmd_wp_page_fallback(struct mm_struct
*mm
,
1013 struct vm_area_struct
*vma
,
1014 unsigned long address
,
1015 pmd_t
*pmd
, pmd_t orig_pmd
,
1017 unsigned long haddr
)
1023 struct page
**pages
;
1024 unsigned long mmun_start
; /* For mmu_notifiers */
1025 unsigned long mmun_end
; /* For mmu_notifiers */
1027 pages
= kmalloc(sizeof(struct page
*) * HPAGE_PMD_NR
,
1029 if (unlikely(!pages
)) {
1030 ret
|= VM_FAULT_OOM
;
1034 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1035 pages
[i
] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE
|
1037 vma
, address
, page_to_nid(page
));
1038 if (unlikely(!pages
[i
] ||
1039 mem_cgroup_newpage_charge(pages
[i
], mm
,
1043 mem_cgroup_uncharge_start();
1045 mem_cgroup_uncharge_page(pages
[i
]);
1048 mem_cgroup_uncharge_end();
1050 ret
|= VM_FAULT_OOM
;
1055 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1056 copy_user_highpage(pages
[i
], page
+ i
,
1057 haddr
+ PAGE_SIZE
* i
, vma
);
1058 __SetPageUptodate(pages
[i
]);
1063 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1064 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1066 ptl
= pmd_lock(mm
, pmd
);
1067 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
1068 goto out_free_pages
;
1069 VM_BUG_ON(!PageHead(page
));
1071 pmdp_clear_flush(vma
, haddr
, pmd
);
1072 /* leave pmd empty until pte is filled */
1074 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1075 pmd_populate(mm
, &_pmd
, pgtable
);
1077 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1079 entry
= mk_pte(pages
[i
], vma
->vm_page_prot
);
1080 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1081 page_add_new_anon_rmap(pages
[i
], vma
, haddr
);
1082 pte
= pte_offset_map(&_pmd
, haddr
);
1083 VM_BUG_ON(!pte_none(*pte
));
1084 set_pte_at(mm
, haddr
, pte
, entry
);
1089 smp_wmb(); /* make pte visible before pmd */
1090 pmd_populate(mm
, pmd
, pgtable
);
1091 page_remove_rmap(page
);
1094 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1096 ret
|= VM_FAULT_WRITE
;
1104 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1105 mem_cgroup_uncharge_start();
1106 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1107 mem_cgroup_uncharge_page(pages
[i
]);
1110 mem_cgroup_uncharge_end();
1115 int do_huge_pmd_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1116 unsigned long address
, pmd_t
*pmd
, pmd_t orig_pmd
)
1120 struct page
*page
= NULL
, *new_page
;
1121 unsigned long haddr
;
1122 unsigned long mmun_start
; /* For mmu_notifiers */
1123 unsigned long mmun_end
; /* For mmu_notifiers */
1125 ptl
= pmd_lockptr(mm
, pmd
);
1126 VM_BUG_ON(!vma
->anon_vma
);
1127 haddr
= address
& HPAGE_PMD_MASK
;
1128 if (is_huge_zero_pmd(orig_pmd
))
1131 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
1134 page
= pmd_page(orig_pmd
);
1135 VM_BUG_ON(!PageCompound(page
) || !PageHead(page
));
1136 if (page_mapcount(page
) == 1) {
1138 entry
= pmd_mkyoung(orig_pmd
);
1139 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1140 if (pmdp_set_access_flags(vma
, haddr
, pmd
, entry
, 1))
1141 update_mmu_cache_pmd(vma
, address
, pmd
);
1142 ret
|= VM_FAULT_WRITE
;
1148 if (transparent_hugepage_enabled(vma
) &&
1149 !transparent_hugepage_debug_cow())
1150 new_page
= alloc_hugepage_vma(transparent_hugepage_defrag(vma
),
1151 vma
, haddr
, numa_node_id(), 0);
1155 if (unlikely(!new_page
)) {
1156 if (is_huge_zero_pmd(orig_pmd
)) {
1157 ret
= do_huge_pmd_wp_zero_page_fallback(mm
, vma
,
1158 address
, pmd
, orig_pmd
, haddr
);
1160 ret
= do_huge_pmd_wp_page_fallback(mm
, vma
, address
,
1161 pmd
, orig_pmd
, page
, haddr
);
1162 if (ret
& VM_FAULT_OOM
)
1163 split_huge_page(page
);
1166 count_vm_event(THP_FAULT_FALLBACK
);
1170 if (unlikely(mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))) {
1173 split_huge_page(page
);
1176 count_vm_event(THP_FAULT_FALLBACK
);
1177 ret
|= VM_FAULT_OOM
;
1181 count_vm_event(THP_FAULT_ALLOC
);
1183 if (is_huge_zero_pmd(orig_pmd
))
1184 clear_huge_page(new_page
, haddr
, HPAGE_PMD_NR
);
1186 copy_user_huge_page(new_page
, page
, haddr
, vma
, HPAGE_PMD_NR
);
1187 __SetPageUptodate(new_page
);
1190 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1191 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1196 if (unlikely(!pmd_same(*pmd
, orig_pmd
))) {
1198 mem_cgroup_uncharge_page(new_page
);
1203 entry
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
1204 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1205 pmdp_clear_flush(vma
, haddr
, pmd
);
1206 page_add_new_anon_rmap(new_page
, vma
, haddr
);
1207 set_pmd_at(mm
, haddr
, pmd
, entry
);
1208 update_mmu_cache_pmd(vma
, address
, pmd
);
1209 if (is_huge_zero_pmd(orig_pmd
)) {
1210 add_mm_counter(mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1211 put_huge_zero_page();
1213 VM_BUG_ON(!PageHead(page
));
1214 page_remove_rmap(page
);
1217 ret
|= VM_FAULT_WRITE
;
1221 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1229 struct page
*follow_trans_huge_pmd(struct vm_area_struct
*vma
,
1234 struct mm_struct
*mm
= vma
->vm_mm
;
1235 struct page
*page
= NULL
;
1237 assert_spin_locked(pmd_lockptr(mm
, pmd
));
1239 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
1242 /* Avoid dumping huge zero page */
1243 if ((flags
& FOLL_DUMP
) && is_huge_zero_pmd(*pmd
))
1244 return ERR_PTR(-EFAULT
);
1246 page
= pmd_page(*pmd
);
1247 VM_BUG_ON(!PageHead(page
));
1248 if (flags
& FOLL_TOUCH
) {
1251 * We should set the dirty bit only for FOLL_WRITE but
1252 * for now the dirty bit in the pmd is meaningless.
1253 * And if the dirty bit will become meaningful and
1254 * we'll only set it with FOLL_WRITE, an atomic
1255 * set_bit will be required on the pmd to set the
1256 * young bit, instead of the current set_pmd_at.
1258 _pmd
= pmd_mkyoung(pmd_mkdirty(*pmd
));
1259 if (pmdp_set_access_flags(vma
, addr
& HPAGE_PMD_MASK
,
1261 update_mmu_cache_pmd(vma
, addr
, pmd
);
1263 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1264 if (page
->mapping
&& trylock_page(page
)) {
1267 mlock_vma_page(page
);
1271 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
1272 VM_BUG_ON(!PageCompound(page
));
1273 if (flags
& FOLL_GET
)
1274 get_page_foll(page
);
1280 /* NUMA hinting page fault entry point for trans huge pmds */
1281 int do_huge_pmd_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1282 unsigned long addr
, pmd_t pmd
, pmd_t
*pmdp
)
1285 struct anon_vma
*anon_vma
= NULL
;
1287 unsigned long haddr
= addr
& HPAGE_PMD_MASK
;
1288 int page_nid
= -1, this_nid
= numa_node_id();
1289 int target_nid
, last_cpupid
= -1;
1291 bool migrated
= false;
1294 ptl
= pmd_lock(mm
, pmdp
);
1295 if (unlikely(!pmd_same(pmd
, *pmdp
)))
1298 page
= pmd_page(pmd
);
1299 BUG_ON(is_huge_zero_page(page
));
1300 page_nid
= page_to_nid(page
);
1301 last_cpupid
= page_cpupid_last(page
);
1302 count_vm_numa_event(NUMA_HINT_FAULTS
);
1303 if (page_nid
== this_nid
) {
1304 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
1305 flags
|= TNF_FAULT_LOCAL
;
1309 * Avoid grouping on DSO/COW pages in specific and RO pages
1310 * in general, RO pages shouldn't hurt as much anyway since
1311 * they can be in shared cache state.
1313 if (!pmd_write(pmd
))
1314 flags
|= TNF_NO_GROUP
;
1317 * Acquire the page lock to serialise THP migrations but avoid dropping
1318 * page_table_lock if at all possible
1320 page_locked
= trylock_page(page
);
1321 target_nid
= mpol_misplaced(page
, vma
, haddr
);
1322 if (target_nid
== -1) {
1323 /* If the page was locked, there are no parallel migrations */
1328 * Otherwise wait for potential migrations and retry. We do
1329 * relock and check_same as the page may no longer be mapped.
1330 * As the fault is being retried, do not account for it.
1333 wait_on_page_locked(page
);
1338 /* Page is misplaced, serialise migrations and parallel THP splits */
1343 anon_vma
= page_lock_anon_vma_read(page
);
1345 /* Confirm the PMD did not change while page_table_lock was released */
1347 if (unlikely(!pmd_same(pmd
, *pmdp
))) {
1355 * Migrate the THP to the requested node, returns with page unlocked
1356 * and pmd_numa cleared.
1359 migrated
= migrate_misplaced_transhuge_page(mm
, vma
,
1360 pmdp
, pmd
, addr
, page
, target_nid
);
1362 flags
|= TNF_MIGRATED
;
1363 page_nid
= target_nid
;
1368 BUG_ON(!PageLocked(page
));
1369 pmd
= pmd_mknonnuma(pmd
);
1370 set_pmd_at(mm
, haddr
, pmdp
, pmd
);
1371 VM_BUG_ON(pmd_numa(*pmdp
));
1372 update_mmu_cache_pmd(vma
, addr
, pmdp
);
1379 page_unlock_anon_vma_read(anon_vma
);
1382 task_numa_fault(last_cpupid
, page_nid
, HPAGE_PMD_NR
, flags
);
1387 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1388 pmd_t
*pmd
, unsigned long addr
)
1393 if (__pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
1398 * For architectures like ppc64 we look at deposited pgtable
1399 * when calling pmdp_get_and_clear. So do the
1400 * pgtable_trans_huge_withdraw after finishing pmdp related
1403 orig_pmd
= pmdp_get_and_clear(tlb
->mm
, addr
, pmd
);
1404 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1405 pgtable
= pgtable_trans_huge_withdraw(tlb
->mm
, pmd
);
1406 if (is_huge_zero_pmd(orig_pmd
)) {
1407 atomic_long_dec(&tlb
->mm
->nr_ptes
);
1409 put_huge_zero_page();
1411 page
= pmd_page(orig_pmd
);
1412 page_remove_rmap(page
);
1413 VM_BUG_ON(page_mapcount(page
) < 0);
1414 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1415 VM_BUG_ON(!PageHead(page
));
1416 atomic_long_dec(&tlb
->mm
->nr_ptes
);
1418 tlb_remove_page(tlb
, page
);
1420 pte_free(tlb
->mm
, pgtable
);
1426 int mincore_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1427 unsigned long addr
, unsigned long end
,
1433 if (__pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
1435 * All logical pages in the range are present
1436 * if backed by a huge page.
1439 memset(vec
, 1, (end
- addr
) >> PAGE_SHIFT
);
1446 int move_huge_pmd(struct vm_area_struct
*vma
, struct vm_area_struct
*new_vma
,
1447 unsigned long old_addr
,
1448 unsigned long new_addr
, unsigned long old_end
,
1449 pmd_t
*old_pmd
, pmd_t
*new_pmd
)
1451 spinlock_t
*old_ptl
, *new_ptl
;
1455 struct mm_struct
*mm
= vma
->vm_mm
;
1457 if ((old_addr
& ~HPAGE_PMD_MASK
) ||
1458 (new_addr
& ~HPAGE_PMD_MASK
) ||
1459 old_end
- old_addr
< HPAGE_PMD_SIZE
||
1460 (new_vma
->vm_flags
& VM_NOHUGEPAGE
))
1464 * The destination pmd shouldn't be established, free_pgtables()
1465 * should have release it.
1467 if (WARN_ON(!pmd_none(*new_pmd
))) {
1468 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1473 * We don't have to worry about the ordering of src and dst
1474 * ptlocks because exclusive mmap_sem prevents deadlock.
1476 ret
= __pmd_trans_huge_lock(old_pmd
, vma
, &old_ptl
);
1478 new_ptl
= pmd_lockptr(mm
, new_pmd
);
1479 if (new_ptl
!= old_ptl
)
1480 spin_lock_nested(new_ptl
, SINGLE_DEPTH_NESTING
);
1481 pmd
= pmdp_get_and_clear(mm
, old_addr
, old_pmd
);
1482 VM_BUG_ON(!pmd_none(*new_pmd
));
1483 set_pmd_at(mm
, new_addr
, new_pmd
, pmd_mksoft_dirty(pmd
));
1484 if (new_ptl
!= old_ptl
) {
1488 * Move preallocated PTE page table if new_pmd is on
1489 * different PMD page table.
1491 pgtable
= pgtable_trans_huge_withdraw(mm
, old_pmd
);
1492 pgtable_trans_huge_deposit(mm
, new_pmd
, pgtable
);
1494 spin_unlock(new_ptl
);
1496 spin_unlock(old_ptl
);
1504 * - 0 if PMD could not be locked
1505 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1506 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1508 int change_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1509 unsigned long addr
, pgprot_t newprot
, int prot_numa
)
1511 struct mm_struct
*mm
= vma
->vm_mm
;
1515 if (__pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
1519 entry
= pmdp_get_and_clear(mm
, addr
, pmd
);
1520 entry
= pmd_modify(entry
, newprot
);
1522 BUG_ON(pmd_write(entry
));
1524 struct page
*page
= pmd_page(*pmd
);
1527 * Do not trap faults against the zero page. The
1528 * read-only data is likely to be read-cached on the
1529 * local CPU cache and it is less useful to know about
1530 * local vs remote hits on the zero page.
1532 if (!is_huge_zero_page(page
) &&
1534 entry
= pmdp_get_and_clear(mm
, addr
, pmd
);
1535 entry
= pmd_mknuma(entry
);
1540 /* Set PMD if cleared earlier */
1541 if (ret
== HPAGE_PMD_NR
)
1542 set_pmd_at(mm
, addr
, pmd
, entry
);
1551 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1552 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1554 * Note that if it returns 1, this routine returns without unlocking page
1555 * table locks. So callers must unlock them.
1557 int __pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
,
1560 *ptl
= pmd_lock(vma
->vm_mm
, pmd
);
1561 if (likely(pmd_trans_huge(*pmd
))) {
1562 if (unlikely(pmd_trans_splitting(*pmd
))) {
1564 wait_split_huge_page(vma
->anon_vma
, pmd
);
1567 /* Thp mapped by 'pmd' is stable, so we can
1568 * handle it as it is. */
1577 * This function returns whether a given @page is mapped onto the @address
1578 * in the virtual space of @mm.
1580 * When it's true, this function returns *pmd with holding the page table lock
1581 * and passing it back to the caller via @ptl.
1582 * If it's false, returns NULL without holding the page table lock.
1584 pmd_t
*page_check_address_pmd(struct page
*page
,
1585 struct mm_struct
*mm
,
1586 unsigned long address
,
1587 enum page_check_address_pmd_flag flag
,
1592 if (address
& ~HPAGE_PMD_MASK
)
1595 pmd
= mm_find_pmd(mm
, address
);
1598 *ptl
= pmd_lock(mm
, pmd
);
1601 if (pmd_page(*pmd
) != page
)
1604 * split_vma() may create temporary aliased mappings. There is
1605 * no risk as long as all huge pmd are found and have their
1606 * splitting bit set before __split_huge_page_refcount
1607 * runs. Finding the same huge pmd more than once during the
1608 * same rmap walk is not a problem.
1610 if (flag
== PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG
&&
1611 pmd_trans_splitting(*pmd
))
1613 if (pmd_trans_huge(*pmd
)) {
1614 VM_BUG_ON(flag
== PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG
&&
1615 !pmd_trans_splitting(*pmd
));
1623 static int __split_huge_page_splitting(struct page
*page
,
1624 struct vm_area_struct
*vma
,
1625 unsigned long address
)
1627 struct mm_struct
*mm
= vma
->vm_mm
;
1631 /* For mmu_notifiers */
1632 const unsigned long mmun_start
= address
;
1633 const unsigned long mmun_end
= address
+ HPAGE_PMD_SIZE
;
1635 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1636 pmd
= page_check_address_pmd(page
, mm
, address
,
1637 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG
, &ptl
);
1640 * We can't temporarily set the pmd to null in order
1641 * to split it, the pmd must remain marked huge at all
1642 * times or the VM won't take the pmd_trans_huge paths
1643 * and it won't wait on the anon_vma->root->rwsem to
1644 * serialize against split_huge_page*.
1646 pmdp_splitting_flush(vma
, address
, pmd
);
1650 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1655 static void __split_huge_page_refcount(struct page
*page
,
1656 struct list_head
*list
)
1659 struct zone
*zone
= page_zone(page
);
1660 struct lruvec
*lruvec
;
1663 /* prevent PageLRU to go away from under us, and freeze lru stats */
1664 spin_lock_irq(&zone
->lru_lock
);
1665 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1667 compound_lock(page
);
1668 /* complete memcg works before add pages to LRU */
1669 mem_cgroup_split_huge_fixup(page
);
1671 for (i
= HPAGE_PMD_NR
- 1; i
>= 1; i
--) {
1672 struct page
*page_tail
= page
+ i
;
1674 /* tail_page->_mapcount cannot change */
1675 BUG_ON(page_mapcount(page_tail
) < 0);
1676 tail_count
+= page_mapcount(page_tail
);
1677 /* check for overflow */
1678 BUG_ON(tail_count
< 0);
1679 BUG_ON(atomic_read(&page_tail
->_count
) != 0);
1681 * tail_page->_count is zero and not changing from
1682 * under us. But get_page_unless_zero() may be running
1683 * from under us on the tail_page. If we used
1684 * atomic_set() below instead of atomic_add(), we
1685 * would then run atomic_set() concurrently with
1686 * get_page_unless_zero(), and atomic_set() is
1687 * implemented in C not using locked ops. spin_unlock
1688 * on x86 sometime uses locked ops because of PPro
1689 * errata 66, 92, so unless somebody can guarantee
1690 * atomic_set() here would be safe on all archs (and
1691 * not only on x86), it's safer to use atomic_add().
1693 atomic_add(page_mapcount(page
) + page_mapcount(page_tail
) + 1,
1694 &page_tail
->_count
);
1696 /* after clearing PageTail the gup refcount can be released */
1700 * retain hwpoison flag of the poisoned tail page:
1701 * fix for the unsuitable process killed on Guest Machine(KVM)
1702 * by the memory-failure.
1704 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
| __PG_HWPOISON
;
1705 page_tail
->flags
|= (page
->flags
&
1706 ((1L << PG_referenced
) |
1707 (1L << PG_swapbacked
) |
1708 (1L << PG_mlocked
) |
1709 (1L << PG_uptodate
) |
1711 (1L << PG_unevictable
)));
1712 page_tail
->flags
|= (1L << PG_dirty
);
1714 /* clear PageTail before overwriting first_page */
1718 * __split_huge_page_splitting() already set the
1719 * splitting bit in all pmd that could map this
1720 * hugepage, that will ensure no CPU can alter the
1721 * mapcount on the head page. The mapcount is only
1722 * accounted in the head page and it has to be
1723 * transferred to all tail pages in the below code. So
1724 * for this code to be safe, the split the mapcount
1725 * can't change. But that doesn't mean userland can't
1726 * keep changing and reading the page contents while
1727 * we transfer the mapcount, so the pmd splitting
1728 * status is achieved setting a reserved bit in the
1729 * pmd, not by clearing the present bit.
1731 page_tail
->_mapcount
= page
->_mapcount
;
1733 BUG_ON(page_tail
->mapping
);
1734 page_tail
->mapping
= page
->mapping
;
1736 page_tail
->index
= page
->index
+ i
;
1737 page_cpupid_xchg_last(page_tail
, page_cpupid_last(page
));
1739 BUG_ON(!PageAnon(page_tail
));
1740 BUG_ON(!PageUptodate(page_tail
));
1741 BUG_ON(!PageDirty(page_tail
));
1742 BUG_ON(!PageSwapBacked(page_tail
));
1744 lru_add_page_tail(page
, page_tail
, lruvec
, list
);
1746 atomic_sub(tail_count
, &page
->_count
);
1747 BUG_ON(atomic_read(&page
->_count
) <= 0);
1749 __mod_zone_page_state(zone
, NR_ANON_TRANSPARENT_HUGEPAGES
, -1);
1751 ClearPageCompound(page
);
1752 compound_unlock(page
);
1753 spin_unlock_irq(&zone
->lru_lock
);
1755 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
1756 struct page
*page_tail
= page
+ i
;
1757 BUG_ON(page_count(page_tail
) <= 0);
1759 * Tail pages may be freed if there wasn't any mapping
1760 * like if add_to_swap() is running on a lru page that
1761 * had its mapping zapped. And freeing these pages
1762 * requires taking the lru_lock so we do the put_page
1763 * of the tail pages after the split is complete.
1765 put_page(page_tail
);
1769 * Only the head page (now become a regular page) is required
1770 * to be pinned by the caller.
1772 BUG_ON(page_count(page
) <= 0);
1775 static int __split_huge_page_map(struct page
*page
,
1776 struct vm_area_struct
*vma
,
1777 unsigned long address
)
1779 struct mm_struct
*mm
= vma
->vm_mm
;
1784 unsigned long haddr
;
1786 pmd
= page_check_address_pmd(page
, mm
, address
,
1787 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG
, &ptl
);
1789 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1790 pmd_populate(mm
, &_pmd
, pgtable
);
1793 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1795 BUG_ON(PageCompound(page
+i
));
1796 entry
= mk_pte(page
+ i
, vma
->vm_page_prot
);
1797 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1798 if (!pmd_write(*pmd
))
1799 entry
= pte_wrprotect(entry
);
1801 BUG_ON(page_mapcount(page
) != 1);
1802 if (!pmd_young(*pmd
))
1803 entry
= pte_mkold(entry
);
1805 entry
= pte_mknuma(entry
);
1806 pte
= pte_offset_map(&_pmd
, haddr
);
1807 BUG_ON(!pte_none(*pte
));
1808 set_pte_at(mm
, haddr
, pte
, entry
);
1812 smp_wmb(); /* make pte visible before pmd */
1814 * Up to this point the pmd is present and huge and
1815 * userland has the whole access to the hugepage
1816 * during the split (which happens in place). If we
1817 * overwrite the pmd with the not-huge version
1818 * pointing to the pte here (which of course we could
1819 * if all CPUs were bug free), userland could trigger
1820 * a small page size TLB miss on the small sized TLB
1821 * while the hugepage TLB entry is still established
1822 * in the huge TLB. Some CPU doesn't like that. See
1823 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1824 * Erratum 383 on page 93. Intel should be safe but is
1825 * also warns that it's only safe if the permission
1826 * and cache attributes of the two entries loaded in
1827 * the two TLB is identical (which should be the case
1828 * here). But it is generally safer to never allow
1829 * small and huge TLB entries for the same virtual
1830 * address to be loaded simultaneously. So instead of
1831 * doing "pmd_populate(); flush_tlb_range();" we first
1832 * mark the current pmd notpresent (atomically because
1833 * here the pmd_trans_huge and pmd_trans_splitting
1834 * must remain set at all times on the pmd until the
1835 * split is complete for this pmd), then we flush the
1836 * SMP TLB and finally we write the non-huge version
1837 * of the pmd entry with pmd_populate.
1839 pmdp_invalidate(vma
, address
, pmd
);
1840 pmd_populate(mm
, pmd
, pgtable
);
1848 /* must be called with anon_vma->root->rwsem held */
1849 static void __split_huge_page(struct page
*page
,
1850 struct anon_vma
*anon_vma
,
1851 struct list_head
*list
)
1853 int mapcount
, mapcount2
;
1854 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1855 struct anon_vma_chain
*avc
;
1857 BUG_ON(!PageHead(page
));
1858 BUG_ON(PageTail(page
));
1861 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
1862 struct vm_area_struct
*vma
= avc
->vma
;
1863 unsigned long addr
= vma_address(page
, vma
);
1864 BUG_ON(is_vma_temporary_stack(vma
));
1865 mapcount
+= __split_huge_page_splitting(page
, vma
, addr
);
1868 * It is critical that new vmas are added to the tail of the
1869 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1870 * and establishes a child pmd before
1871 * __split_huge_page_splitting() freezes the parent pmd (so if
1872 * we fail to prevent copy_huge_pmd() from running until the
1873 * whole __split_huge_page() is complete), we will still see
1874 * the newly established pmd of the child later during the
1875 * walk, to be able to set it as pmd_trans_splitting too.
1877 if (mapcount
!= page_mapcount(page
))
1878 printk(KERN_ERR
"mapcount %d page_mapcount %d\n",
1879 mapcount
, page_mapcount(page
));
1880 BUG_ON(mapcount
!= page_mapcount(page
));
1882 __split_huge_page_refcount(page
, list
);
1885 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
1886 struct vm_area_struct
*vma
= avc
->vma
;
1887 unsigned long addr
= vma_address(page
, vma
);
1888 BUG_ON(is_vma_temporary_stack(vma
));
1889 mapcount2
+= __split_huge_page_map(page
, vma
, addr
);
1891 if (mapcount
!= mapcount2
)
1892 printk(KERN_ERR
"mapcount %d mapcount2 %d page_mapcount %d\n",
1893 mapcount
, mapcount2
, page_mapcount(page
));
1894 BUG_ON(mapcount
!= mapcount2
);
1898 * Split a hugepage into normal pages. This doesn't change the position of head
1899 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1900 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1901 * from the hugepage.
1902 * Return 0 if the hugepage is split successfully otherwise return 1.
1904 int split_huge_page_to_list(struct page
*page
, struct list_head
*list
)
1906 struct anon_vma
*anon_vma
;
1909 BUG_ON(is_huge_zero_page(page
));
1910 BUG_ON(!PageAnon(page
));
1913 * The caller does not necessarily hold an mmap_sem that would prevent
1914 * the anon_vma disappearing so we first we take a reference to it
1915 * and then lock the anon_vma for write. This is similar to
1916 * page_lock_anon_vma_read except the write lock is taken to serialise
1917 * against parallel split or collapse operations.
1919 anon_vma
= page_get_anon_vma(page
);
1922 anon_vma_lock_write(anon_vma
);
1925 if (!PageCompound(page
))
1928 BUG_ON(!PageSwapBacked(page
));
1929 __split_huge_page(page
, anon_vma
, list
);
1930 count_vm_event(THP_SPLIT
);
1932 BUG_ON(PageCompound(page
));
1934 anon_vma_unlock_write(anon_vma
);
1935 put_anon_vma(anon_vma
);
1940 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1942 int hugepage_madvise(struct vm_area_struct
*vma
,
1943 unsigned long *vm_flags
, int advice
)
1945 struct mm_struct
*mm
= vma
->vm_mm
;
1950 * Be somewhat over-protective like KSM for now!
1952 if (*vm_flags
& (VM_HUGEPAGE
| VM_NO_THP
))
1954 if (mm
->def_flags
& VM_NOHUGEPAGE
)
1956 *vm_flags
&= ~VM_NOHUGEPAGE
;
1957 *vm_flags
|= VM_HUGEPAGE
;
1959 * If the vma become good for khugepaged to scan,
1960 * register it here without waiting a page fault that
1961 * may not happen any time soon.
1963 if (unlikely(khugepaged_enter_vma_merge(vma
)))
1966 case MADV_NOHUGEPAGE
:
1968 * Be somewhat over-protective like KSM for now!
1970 if (*vm_flags
& (VM_NOHUGEPAGE
| VM_NO_THP
))
1972 *vm_flags
&= ~VM_HUGEPAGE
;
1973 *vm_flags
|= VM_NOHUGEPAGE
;
1975 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1976 * this vma even if we leave the mm registered in khugepaged if
1977 * it got registered before VM_NOHUGEPAGE was set.
1985 static int __init
khugepaged_slab_init(void)
1987 mm_slot_cache
= kmem_cache_create("khugepaged_mm_slot",
1988 sizeof(struct mm_slot
),
1989 __alignof__(struct mm_slot
), 0, NULL
);
1996 static inline struct mm_slot
*alloc_mm_slot(void)
1998 if (!mm_slot_cache
) /* initialization failed */
2000 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
2003 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
2005 kmem_cache_free(mm_slot_cache
, mm_slot
);
2008 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
2010 struct mm_slot
*mm_slot
;
2012 hash_for_each_possible(mm_slots_hash
, mm_slot
, hash
, (unsigned long)mm
)
2013 if (mm
== mm_slot
->mm
)
2019 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
2020 struct mm_slot
*mm_slot
)
2023 hash_add(mm_slots_hash
, &mm_slot
->hash
, (long)mm
);
2026 static inline int khugepaged_test_exit(struct mm_struct
*mm
)
2028 return atomic_read(&mm
->mm_users
) == 0;
2031 int __khugepaged_enter(struct mm_struct
*mm
)
2033 struct mm_slot
*mm_slot
;
2036 mm_slot
= alloc_mm_slot();
2040 /* __khugepaged_exit() must not run from under us */
2041 VM_BUG_ON(khugepaged_test_exit(mm
));
2042 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE
, &mm
->flags
))) {
2043 free_mm_slot(mm_slot
);
2047 spin_lock(&khugepaged_mm_lock
);
2048 insert_to_mm_slots_hash(mm
, mm_slot
);
2050 * Insert just behind the scanning cursor, to let the area settle
2053 wakeup
= list_empty(&khugepaged_scan
.mm_head
);
2054 list_add_tail(&mm_slot
->mm_node
, &khugepaged_scan
.mm_head
);
2055 spin_unlock(&khugepaged_mm_lock
);
2057 atomic_inc(&mm
->mm_count
);
2059 wake_up_interruptible(&khugepaged_wait
);
2064 int khugepaged_enter_vma_merge(struct vm_area_struct
*vma
)
2066 unsigned long hstart
, hend
;
2069 * Not yet faulted in so we will register later in the
2070 * page fault if needed.
2074 /* khugepaged not yet working on file or special mappings */
2076 VM_BUG_ON(vma
->vm_flags
& VM_NO_THP
);
2077 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2078 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2080 return khugepaged_enter(vma
);
2084 void __khugepaged_exit(struct mm_struct
*mm
)
2086 struct mm_slot
*mm_slot
;
2089 spin_lock(&khugepaged_mm_lock
);
2090 mm_slot
= get_mm_slot(mm
);
2091 if (mm_slot
&& khugepaged_scan
.mm_slot
!= mm_slot
) {
2092 hash_del(&mm_slot
->hash
);
2093 list_del(&mm_slot
->mm_node
);
2096 spin_unlock(&khugepaged_mm_lock
);
2099 clear_bit(MMF_VM_HUGEPAGE
, &mm
->flags
);
2100 free_mm_slot(mm_slot
);
2102 } else if (mm_slot
) {
2104 * This is required to serialize against
2105 * khugepaged_test_exit() (which is guaranteed to run
2106 * under mmap sem read mode). Stop here (after we
2107 * return all pagetables will be destroyed) until
2108 * khugepaged has finished working on the pagetables
2109 * under the mmap_sem.
2111 down_write(&mm
->mmap_sem
);
2112 up_write(&mm
->mmap_sem
);
2116 static void release_pte_page(struct page
*page
)
2118 /* 0 stands for page_is_file_cache(page) == false */
2119 dec_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
2121 putback_lru_page(page
);
2124 static void release_pte_pages(pte_t
*pte
, pte_t
*_pte
)
2126 while (--_pte
>= pte
) {
2127 pte_t pteval
= *_pte
;
2128 if (!pte_none(pteval
))
2129 release_pte_page(pte_page(pteval
));
2133 static int __collapse_huge_page_isolate(struct vm_area_struct
*vma
,
2134 unsigned long address
,
2139 int referenced
= 0, none
= 0;
2140 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
2141 _pte
++, address
+= PAGE_SIZE
) {
2142 pte_t pteval
= *_pte
;
2143 if (pte_none(pteval
)) {
2144 if (++none
<= khugepaged_max_ptes_none
)
2149 if (!pte_present(pteval
) || !pte_write(pteval
))
2151 page
= vm_normal_page(vma
, address
, pteval
);
2152 if (unlikely(!page
))
2155 VM_BUG_ON(PageCompound(page
));
2156 BUG_ON(!PageAnon(page
));
2157 VM_BUG_ON(!PageSwapBacked(page
));
2159 /* cannot use mapcount: can't collapse if there's a gup pin */
2160 if (page_count(page
) != 1)
2163 * We can do it before isolate_lru_page because the
2164 * page can't be freed from under us. NOTE: PG_lock
2165 * is needed to serialize against split_huge_page
2166 * when invoked from the VM.
2168 if (!trylock_page(page
))
2171 * Isolate the page to avoid collapsing an hugepage
2172 * currently in use by the VM.
2174 if (isolate_lru_page(page
)) {
2178 /* 0 stands for page_is_file_cache(page) == false */
2179 inc_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
2180 VM_BUG_ON(!PageLocked(page
));
2181 VM_BUG_ON(PageLRU(page
));
2183 /* If there is no mapped pte young don't collapse the page */
2184 if (pte_young(pteval
) || PageReferenced(page
) ||
2185 mmu_notifier_test_young(vma
->vm_mm
, address
))
2188 if (likely(referenced
))
2191 release_pte_pages(pte
, _pte
);
2195 static void __collapse_huge_page_copy(pte_t
*pte
, struct page
*page
,
2196 struct vm_area_struct
*vma
,
2197 unsigned long address
,
2201 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
; _pte
++) {
2202 pte_t pteval
= *_pte
;
2203 struct page
*src_page
;
2205 if (pte_none(pteval
)) {
2206 clear_user_highpage(page
, address
);
2207 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, 1);
2209 src_page
= pte_page(pteval
);
2210 copy_user_highpage(page
, src_page
, address
, vma
);
2211 VM_BUG_ON(page_mapcount(src_page
) != 1);
2212 release_pte_page(src_page
);
2214 * ptl mostly unnecessary, but preempt has to
2215 * be disabled to update the per-cpu stats
2216 * inside page_remove_rmap().
2220 * paravirt calls inside pte_clear here are
2223 pte_clear(vma
->vm_mm
, address
, _pte
);
2224 page_remove_rmap(src_page
);
2226 free_page_and_swap_cache(src_page
);
2229 address
+= PAGE_SIZE
;
2234 static void khugepaged_alloc_sleep(void)
2236 wait_event_freezable_timeout(khugepaged_wait
, false,
2237 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs
));
2240 static int khugepaged_node_load
[MAX_NUMNODES
];
2243 static int khugepaged_find_target_node(void)
2245 static int last_khugepaged_target_node
= NUMA_NO_NODE
;
2246 int nid
, target_node
= 0, max_value
= 0;
2248 /* find first node with max normal pages hit */
2249 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
2250 if (khugepaged_node_load
[nid
] > max_value
) {
2251 max_value
= khugepaged_node_load
[nid
];
2255 /* do some balance if several nodes have the same hit record */
2256 if (target_node
<= last_khugepaged_target_node
)
2257 for (nid
= last_khugepaged_target_node
+ 1; nid
< MAX_NUMNODES
;
2259 if (max_value
== khugepaged_node_load
[nid
]) {
2264 last_khugepaged_target_node
= target_node
;
2268 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
2270 if (IS_ERR(*hpage
)) {
2276 khugepaged_alloc_sleep();
2277 } else if (*hpage
) {
2286 *khugepaged_alloc_page(struct page
**hpage
, struct mm_struct
*mm
,
2287 struct vm_area_struct
*vma
, unsigned long address
,
2292 * Allocate the page while the vma is still valid and under
2293 * the mmap_sem read mode so there is no memory allocation
2294 * later when we take the mmap_sem in write mode. This is more
2295 * friendly behavior (OTOH it may actually hide bugs) to
2296 * filesystems in userland with daemons allocating memory in
2297 * the userland I/O paths. Allocating memory with the
2298 * mmap_sem in read mode is good idea also to allow greater
2301 *hpage
= alloc_pages_exact_node(node
, alloc_hugepage_gfpmask(
2302 khugepaged_defrag(), __GFP_OTHER_NODE
), HPAGE_PMD_ORDER
);
2304 * After allocating the hugepage, release the mmap_sem read lock in
2305 * preparation for taking it in write mode.
2307 up_read(&mm
->mmap_sem
);
2308 if (unlikely(!*hpage
)) {
2309 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
2310 *hpage
= ERR_PTR(-ENOMEM
);
2314 count_vm_event(THP_COLLAPSE_ALLOC
);
2318 static int khugepaged_find_target_node(void)
2323 static inline struct page
*alloc_hugepage(int defrag
)
2325 return alloc_pages(alloc_hugepage_gfpmask(defrag
, 0),
2329 static struct page
*khugepaged_alloc_hugepage(bool *wait
)
2334 hpage
= alloc_hugepage(khugepaged_defrag());
2336 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
2341 khugepaged_alloc_sleep();
2343 count_vm_event(THP_COLLAPSE_ALLOC
);
2344 } while (unlikely(!hpage
) && likely(khugepaged_enabled()));
2349 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
2352 *hpage
= khugepaged_alloc_hugepage(wait
);
2354 if (unlikely(!*hpage
))
2361 *khugepaged_alloc_page(struct page
**hpage
, struct mm_struct
*mm
,
2362 struct vm_area_struct
*vma
, unsigned long address
,
2365 up_read(&mm
->mmap_sem
);
2371 static bool hugepage_vma_check(struct vm_area_struct
*vma
)
2373 if ((!(vma
->vm_flags
& VM_HUGEPAGE
) && !khugepaged_always()) ||
2374 (vma
->vm_flags
& VM_NOHUGEPAGE
))
2377 if (!vma
->anon_vma
|| vma
->vm_ops
)
2379 if (is_vma_temporary_stack(vma
))
2381 VM_BUG_ON(vma
->vm_flags
& VM_NO_THP
);
2385 static void collapse_huge_page(struct mm_struct
*mm
,
2386 unsigned long address
,
2387 struct page
**hpage
,
2388 struct vm_area_struct
*vma
,
2394 struct page
*new_page
;
2395 spinlock_t
*pmd_ptl
, *pte_ptl
;
2397 unsigned long hstart
, hend
;
2398 unsigned long mmun_start
; /* For mmu_notifiers */
2399 unsigned long mmun_end
; /* For mmu_notifiers */
2401 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2403 /* release the mmap_sem read lock. */
2404 new_page
= khugepaged_alloc_page(hpage
, mm
, vma
, address
, node
);
2408 if (unlikely(mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
)))
2412 * Prevent all access to pagetables with the exception of
2413 * gup_fast later hanlded by the ptep_clear_flush and the VM
2414 * handled by the anon_vma lock + PG_lock.
2416 down_write(&mm
->mmap_sem
);
2417 if (unlikely(khugepaged_test_exit(mm
)))
2420 vma
= find_vma(mm
, address
);
2423 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2424 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2425 if (address
< hstart
|| address
+ HPAGE_PMD_SIZE
> hend
)
2427 if (!hugepage_vma_check(vma
))
2429 pmd
= mm_find_pmd(mm
, address
);
2432 if (pmd_trans_huge(*pmd
))
2435 anon_vma_lock_write(vma
->anon_vma
);
2437 pte
= pte_offset_map(pmd
, address
);
2438 pte_ptl
= pte_lockptr(mm
, pmd
);
2440 mmun_start
= address
;
2441 mmun_end
= address
+ HPAGE_PMD_SIZE
;
2442 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2443 pmd_ptl
= pmd_lock(mm
, pmd
); /* probably unnecessary */
2445 * After this gup_fast can't run anymore. This also removes
2446 * any huge TLB entry from the CPU so we won't allow
2447 * huge and small TLB entries for the same virtual address
2448 * to avoid the risk of CPU bugs in that area.
2450 _pmd
= pmdp_clear_flush(vma
, address
, pmd
);
2451 spin_unlock(pmd_ptl
);
2452 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2455 isolated
= __collapse_huge_page_isolate(vma
, address
, pte
);
2456 spin_unlock(pte_ptl
);
2458 if (unlikely(!isolated
)) {
2461 BUG_ON(!pmd_none(*pmd
));
2463 * We can only use set_pmd_at when establishing
2464 * hugepmds and never for establishing regular pmds that
2465 * points to regular pagetables. Use pmd_populate for that
2467 pmd_populate(mm
, pmd
, pmd_pgtable(_pmd
));
2468 spin_unlock(pmd_ptl
);
2469 anon_vma_unlock_write(vma
->anon_vma
);
2474 * All pages are isolated and locked so anon_vma rmap
2475 * can't run anymore.
2477 anon_vma_unlock_write(vma
->anon_vma
);
2479 __collapse_huge_page_copy(pte
, new_page
, vma
, address
, pte_ptl
);
2481 __SetPageUptodate(new_page
);
2482 pgtable
= pmd_pgtable(_pmd
);
2484 _pmd
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
2485 _pmd
= maybe_pmd_mkwrite(pmd_mkdirty(_pmd
), vma
);
2488 * spin_lock() below is not the equivalent of smp_wmb(), so
2489 * this is needed to avoid the copy_huge_page writes to become
2490 * visible after the set_pmd_at() write.
2495 BUG_ON(!pmd_none(*pmd
));
2496 page_add_new_anon_rmap(new_page
, vma
, address
);
2497 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
2498 set_pmd_at(mm
, address
, pmd
, _pmd
);
2499 update_mmu_cache_pmd(vma
, address
, pmd
);
2500 spin_unlock(pmd_ptl
);
2504 khugepaged_pages_collapsed
++;
2506 up_write(&mm
->mmap_sem
);
2510 mem_cgroup_uncharge_page(new_page
);
2514 static int khugepaged_scan_pmd(struct mm_struct
*mm
,
2515 struct vm_area_struct
*vma
,
2516 unsigned long address
,
2517 struct page
**hpage
)
2521 int ret
= 0, referenced
= 0, none
= 0;
2523 unsigned long _address
;
2525 int node
= NUMA_NO_NODE
;
2527 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2529 pmd
= mm_find_pmd(mm
, address
);
2532 if (pmd_trans_huge(*pmd
))
2535 memset(khugepaged_node_load
, 0, sizeof(khugepaged_node_load
));
2536 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2537 for (_address
= address
, _pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
2538 _pte
++, _address
+= PAGE_SIZE
) {
2539 pte_t pteval
= *_pte
;
2540 if (pte_none(pteval
)) {
2541 if (++none
<= khugepaged_max_ptes_none
)
2546 if (!pte_present(pteval
) || !pte_write(pteval
))
2548 page
= vm_normal_page(vma
, _address
, pteval
);
2549 if (unlikely(!page
))
2552 * Record which node the original page is from and save this
2553 * information to khugepaged_node_load[].
2554 * Khupaged will allocate hugepage from the node has the max
2557 node
= page_to_nid(page
);
2558 khugepaged_node_load
[node
]++;
2559 VM_BUG_ON(PageCompound(page
));
2560 if (!PageLRU(page
) || PageLocked(page
) || !PageAnon(page
))
2562 /* cannot use mapcount: can't collapse if there's a gup pin */
2563 if (page_count(page
) != 1)
2565 if (pte_young(pteval
) || PageReferenced(page
) ||
2566 mmu_notifier_test_young(vma
->vm_mm
, address
))
2572 pte_unmap_unlock(pte
, ptl
);
2574 node
= khugepaged_find_target_node();
2575 /* collapse_huge_page will return with the mmap_sem released */
2576 collapse_huge_page(mm
, address
, hpage
, vma
, node
);
2582 static void collect_mm_slot(struct mm_slot
*mm_slot
)
2584 struct mm_struct
*mm
= mm_slot
->mm
;
2586 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2588 if (khugepaged_test_exit(mm
)) {
2590 hash_del(&mm_slot
->hash
);
2591 list_del(&mm_slot
->mm_node
);
2594 * Not strictly needed because the mm exited already.
2596 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2599 /* khugepaged_mm_lock actually not necessary for the below */
2600 free_mm_slot(mm_slot
);
2605 static unsigned int khugepaged_scan_mm_slot(unsigned int pages
,
2606 struct page
**hpage
)
2607 __releases(&khugepaged_mm_lock
)
2608 __acquires(&khugepaged_mm_lock
)
2610 struct mm_slot
*mm_slot
;
2611 struct mm_struct
*mm
;
2612 struct vm_area_struct
*vma
;
2616 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2618 if (khugepaged_scan
.mm_slot
)
2619 mm_slot
= khugepaged_scan
.mm_slot
;
2621 mm_slot
= list_entry(khugepaged_scan
.mm_head
.next
,
2622 struct mm_slot
, mm_node
);
2623 khugepaged_scan
.address
= 0;
2624 khugepaged_scan
.mm_slot
= mm_slot
;
2626 spin_unlock(&khugepaged_mm_lock
);
2629 down_read(&mm
->mmap_sem
);
2630 if (unlikely(khugepaged_test_exit(mm
)))
2633 vma
= find_vma(mm
, khugepaged_scan
.address
);
2636 for (; vma
; vma
= vma
->vm_next
) {
2637 unsigned long hstart
, hend
;
2640 if (unlikely(khugepaged_test_exit(mm
))) {
2644 if (!hugepage_vma_check(vma
)) {
2649 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2650 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2653 if (khugepaged_scan
.address
> hend
)
2655 if (khugepaged_scan
.address
< hstart
)
2656 khugepaged_scan
.address
= hstart
;
2657 VM_BUG_ON(khugepaged_scan
.address
& ~HPAGE_PMD_MASK
);
2659 while (khugepaged_scan
.address
< hend
) {
2662 if (unlikely(khugepaged_test_exit(mm
)))
2663 goto breakouterloop
;
2665 VM_BUG_ON(khugepaged_scan
.address
< hstart
||
2666 khugepaged_scan
.address
+ HPAGE_PMD_SIZE
>
2668 ret
= khugepaged_scan_pmd(mm
, vma
,
2669 khugepaged_scan
.address
,
2671 /* move to next address */
2672 khugepaged_scan
.address
+= HPAGE_PMD_SIZE
;
2673 progress
+= HPAGE_PMD_NR
;
2675 /* we released mmap_sem so break loop */
2676 goto breakouterloop_mmap_sem
;
2677 if (progress
>= pages
)
2678 goto breakouterloop
;
2682 up_read(&mm
->mmap_sem
); /* exit_mmap will destroy ptes after this */
2683 breakouterloop_mmap_sem
:
2685 spin_lock(&khugepaged_mm_lock
);
2686 VM_BUG_ON(khugepaged_scan
.mm_slot
!= mm_slot
);
2688 * Release the current mm_slot if this mm is about to die, or
2689 * if we scanned all vmas of this mm.
2691 if (khugepaged_test_exit(mm
) || !vma
) {
2693 * Make sure that if mm_users is reaching zero while
2694 * khugepaged runs here, khugepaged_exit will find
2695 * mm_slot not pointing to the exiting mm.
2697 if (mm_slot
->mm_node
.next
!= &khugepaged_scan
.mm_head
) {
2698 khugepaged_scan
.mm_slot
= list_entry(
2699 mm_slot
->mm_node
.next
,
2700 struct mm_slot
, mm_node
);
2701 khugepaged_scan
.address
= 0;
2703 khugepaged_scan
.mm_slot
= NULL
;
2704 khugepaged_full_scans
++;
2707 collect_mm_slot(mm_slot
);
2713 static int khugepaged_has_work(void)
2715 return !list_empty(&khugepaged_scan
.mm_head
) &&
2716 khugepaged_enabled();
2719 static int khugepaged_wait_event(void)
2721 return !list_empty(&khugepaged_scan
.mm_head
) ||
2722 kthread_should_stop();
2725 static void khugepaged_do_scan(void)
2727 struct page
*hpage
= NULL
;
2728 unsigned int progress
= 0, pass_through_head
= 0;
2729 unsigned int pages
= khugepaged_pages_to_scan
;
2732 barrier(); /* write khugepaged_pages_to_scan to local stack */
2734 while (progress
< pages
) {
2735 if (!khugepaged_prealloc_page(&hpage
, &wait
))
2740 if (unlikely(kthread_should_stop() || freezing(current
)))
2743 spin_lock(&khugepaged_mm_lock
);
2744 if (!khugepaged_scan
.mm_slot
)
2745 pass_through_head
++;
2746 if (khugepaged_has_work() &&
2747 pass_through_head
< 2)
2748 progress
+= khugepaged_scan_mm_slot(pages
- progress
,
2752 spin_unlock(&khugepaged_mm_lock
);
2755 if (!IS_ERR_OR_NULL(hpage
))
2759 static void khugepaged_wait_work(void)
2763 if (khugepaged_has_work()) {
2764 if (!khugepaged_scan_sleep_millisecs
)
2767 wait_event_freezable_timeout(khugepaged_wait
,
2768 kthread_should_stop(),
2769 msecs_to_jiffies(khugepaged_scan_sleep_millisecs
));
2773 if (khugepaged_enabled())
2774 wait_event_freezable(khugepaged_wait
, khugepaged_wait_event());
2777 static int khugepaged(void *none
)
2779 struct mm_slot
*mm_slot
;
2782 set_user_nice(current
, 19);
2784 while (!kthread_should_stop()) {
2785 khugepaged_do_scan();
2786 khugepaged_wait_work();
2789 spin_lock(&khugepaged_mm_lock
);
2790 mm_slot
= khugepaged_scan
.mm_slot
;
2791 khugepaged_scan
.mm_slot
= NULL
;
2793 collect_mm_slot(mm_slot
);
2794 spin_unlock(&khugepaged_mm_lock
);
2798 static void __split_huge_zero_page_pmd(struct vm_area_struct
*vma
,
2799 unsigned long haddr
, pmd_t
*pmd
)
2801 struct mm_struct
*mm
= vma
->vm_mm
;
2806 pmdp_clear_flush(vma
, haddr
, pmd
);
2807 /* leave pmd empty until pte is filled */
2809 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2810 pmd_populate(mm
, &_pmd
, pgtable
);
2812 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
2814 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
2815 entry
= pte_mkspecial(entry
);
2816 pte
= pte_offset_map(&_pmd
, haddr
);
2817 VM_BUG_ON(!pte_none(*pte
));
2818 set_pte_at(mm
, haddr
, pte
, entry
);
2821 smp_wmb(); /* make pte visible before pmd */
2822 pmd_populate(mm
, pmd
, pgtable
);
2823 put_huge_zero_page();
2826 void __split_huge_page_pmd(struct vm_area_struct
*vma
, unsigned long address
,
2831 struct mm_struct
*mm
= vma
->vm_mm
;
2832 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
2833 unsigned long mmun_start
; /* For mmu_notifiers */
2834 unsigned long mmun_end
; /* For mmu_notifiers */
2836 BUG_ON(vma
->vm_start
> haddr
|| vma
->vm_end
< haddr
+ HPAGE_PMD_SIZE
);
2839 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
2841 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2842 ptl
= pmd_lock(mm
, pmd
);
2843 if (unlikely(!pmd_trans_huge(*pmd
))) {
2845 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2848 if (is_huge_zero_pmd(*pmd
)) {
2849 __split_huge_zero_page_pmd(vma
, haddr
, pmd
);
2851 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2854 page
= pmd_page(*pmd
);
2855 VM_BUG_ON(!page_count(page
));
2858 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2860 split_huge_page(page
);
2865 * We don't always have down_write of mmap_sem here: a racing
2866 * do_huge_pmd_wp_page() might have copied-on-write to another
2867 * huge page before our split_huge_page() got the anon_vma lock.
2869 if (unlikely(pmd_trans_huge(*pmd
)))
2873 void split_huge_page_pmd_mm(struct mm_struct
*mm
, unsigned long address
,
2876 struct vm_area_struct
*vma
;
2878 vma
= find_vma(mm
, address
);
2879 BUG_ON(vma
== NULL
);
2880 split_huge_page_pmd(vma
, address
, pmd
);
2883 static void split_huge_page_address(struct mm_struct
*mm
,
2884 unsigned long address
)
2888 VM_BUG_ON(!(address
& ~HPAGE_PMD_MASK
));
2890 pmd
= mm_find_pmd(mm
, address
);
2894 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2895 * materialize from under us.
2897 split_huge_page_pmd_mm(mm
, address
, pmd
);
2900 void __vma_adjust_trans_huge(struct vm_area_struct
*vma
,
2901 unsigned long start
,
2906 * If the new start address isn't hpage aligned and it could
2907 * previously contain an hugepage: check if we need to split
2910 if (start
& ~HPAGE_PMD_MASK
&&
2911 (start
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2912 (start
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2913 split_huge_page_address(vma
->vm_mm
, start
);
2916 * If the new end address isn't hpage aligned and it could
2917 * previously contain an hugepage: check if we need to split
2920 if (end
& ~HPAGE_PMD_MASK
&&
2921 (end
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2922 (end
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2923 split_huge_page_address(vma
->vm_mm
, end
);
2926 * If we're also updating the vma->vm_next->vm_start, if the new
2927 * vm_next->vm_start isn't page aligned and it could previously
2928 * contain an hugepage: check if we need to split an huge pmd.
2930 if (adjust_next
> 0) {
2931 struct vm_area_struct
*next
= vma
->vm_next
;
2932 unsigned long nstart
= next
->vm_start
;
2933 nstart
+= adjust_next
<< PAGE_SHIFT
;
2934 if (nstart
& ~HPAGE_PMD_MASK
&&
2935 (nstart
& HPAGE_PMD_MASK
) >= next
->vm_start
&&
2936 (nstart
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= next
->vm_end
)
2937 split_huge_page_address(next
->vm_mm
, nstart
);