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/mm_inline.h>
16 #include <linux/kthread.h>
17 #include <linux/khugepaged.h>
18 #include <linux/freezer.h>
19 #include <linux/mman.h>
20 #include <linux/pagemap.h>
22 #include <asm/pgalloc.h>
26 * By default transparent hugepage support is enabled for all mappings
27 * and khugepaged scans all mappings. Defrag is only invoked by
28 * khugepaged hugepage allocations and by page faults inside
29 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
32 unsigned long transparent_hugepage_flags __read_mostly
=
33 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
34 (1<<TRANSPARENT_HUGEPAGE_FLAG
)|
36 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
37 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
)|
39 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
)|
40 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
42 /* default scan 8*512 pte (or vmas) every 30 second */
43 static unsigned int khugepaged_pages_to_scan __read_mostly
= HPAGE_PMD_NR
*8;
44 static unsigned int khugepaged_pages_collapsed
;
45 static unsigned int khugepaged_full_scans
;
46 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly
= 10000;
47 /* during fragmentation poll the hugepage allocator once every minute */
48 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly
= 60000;
49 static struct task_struct
*khugepaged_thread __read_mostly
;
50 static unsigned long huge_zero_pfn __read_mostly
;
51 static DEFINE_MUTEX(khugepaged_mutex
);
52 static DEFINE_SPINLOCK(khugepaged_mm_lock
);
53 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait
);
55 * default collapse hugepages if there is at least one pte mapped like
56 * it would have happened if the vma was large enough during page
59 static unsigned int khugepaged_max_ptes_none __read_mostly
= HPAGE_PMD_NR
-1;
61 static int khugepaged(void *none
);
62 static int mm_slots_hash_init(void);
63 static int khugepaged_slab_init(void);
64 static void khugepaged_slab_free(void);
66 #define MM_SLOTS_HASH_HEADS 1024
67 static struct hlist_head
*mm_slots_hash __read_mostly
;
68 static struct kmem_cache
*mm_slot_cache __read_mostly
;
71 * struct mm_slot - hash lookup from mm to mm_slot
72 * @hash: hash collision list
73 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
74 * @mm: the mm that this information is valid for
77 struct hlist_node hash
;
78 struct list_head mm_node
;
83 * struct khugepaged_scan - cursor for scanning
84 * @mm_head: the head of the mm list to scan
85 * @mm_slot: the current mm_slot we are scanning
86 * @address: the next address inside that to be scanned
88 * There is only the one khugepaged_scan instance of this cursor structure.
90 struct khugepaged_scan
{
91 struct list_head mm_head
;
92 struct mm_slot
*mm_slot
;
93 unsigned long address
;
95 static struct khugepaged_scan khugepaged_scan
= {
96 .mm_head
= LIST_HEAD_INIT(khugepaged_scan
.mm_head
),
100 static int set_recommended_min_free_kbytes(void)
104 unsigned long recommended_min
;
105 extern int min_free_kbytes
;
107 if (!khugepaged_enabled())
110 for_each_populated_zone(zone
)
113 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
114 recommended_min
= pageblock_nr_pages
* nr_zones
* 2;
117 * Make sure that on average at least two pageblocks are almost free
118 * of another type, one for a migratetype to fall back to and a
119 * second to avoid subsequent fallbacks of other types There are 3
120 * MIGRATE_TYPES we care about.
122 recommended_min
+= pageblock_nr_pages
* nr_zones
*
123 MIGRATE_PCPTYPES
* MIGRATE_PCPTYPES
;
125 /* don't ever allow to reserve more than 5% of the lowmem */
126 recommended_min
= min(recommended_min
,
127 (unsigned long) nr_free_buffer_pages() / 20);
128 recommended_min
<<= (PAGE_SHIFT
-10);
130 if (recommended_min
> min_free_kbytes
)
131 min_free_kbytes
= recommended_min
;
132 setup_per_zone_wmarks();
135 late_initcall(set_recommended_min_free_kbytes
);
137 static int start_khugepaged(void)
140 if (khugepaged_enabled()) {
141 if (!khugepaged_thread
)
142 khugepaged_thread
= kthread_run(khugepaged
, NULL
,
144 if (unlikely(IS_ERR(khugepaged_thread
))) {
146 "khugepaged: kthread_run(khugepaged) failed\n");
147 err
= PTR_ERR(khugepaged_thread
);
148 khugepaged_thread
= NULL
;
151 if (!list_empty(&khugepaged_scan
.mm_head
))
152 wake_up_interruptible(&khugepaged_wait
);
154 set_recommended_min_free_kbytes();
155 } else if (khugepaged_thread
) {
156 kthread_stop(khugepaged_thread
);
157 khugepaged_thread
= NULL
;
163 static int __init
init_huge_zero_page(void)
167 hpage
= alloc_pages((GFP_TRANSHUGE
| __GFP_ZERO
) & ~__GFP_MOVABLE
,
172 huge_zero_pfn
= page_to_pfn(hpage
);
176 static inline bool is_huge_zero_pfn(unsigned long pfn
)
178 return pfn
== huge_zero_pfn
;
181 static inline bool is_huge_zero_pmd(pmd_t pmd
)
183 return is_huge_zero_pfn(pmd_pfn(pmd
));
188 static ssize_t
double_flag_show(struct kobject
*kobj
,
189 struct kobj_attribute
*attr
, char *buf
,
190 enum transparent_hugepage_flag enabled
,
191 enum transparent_hugepage_flag req_madv
)
193 if (test_bit(enabled
, &transparent_hugepage_flags
)) {
194 VM_BUG_ON(test_bit(req_madv
, &transparent_hugepage_flags
));
195 return sprintf(buf
, "[always] madvise never\n");
196 } else if (test_bit(req_madv
, &transparent_hugepage_flags
))
197 return sprintf(buf
, "always [madvise] never\n");
199 return sprintf(buf
, "always madvise [never]\n");
201 static ssize_t
double_flag_store(struct kobject
*kobj
,
202 struct kobj_attribute
*attr
,
203 const char *buf
, size_t count
,
204 enum transparent_hugepage_flag enabled
,
205 enum transparent_hugepage_flag req_madv
)
207 if (!memcmp("always", buf
,
208 min(sizeof("always")-1, count
))) {
209 set_bit(enabled
, &transparent_hugepage_flags
);
210 clear_bit(req_madv
, &transparent_hugepage_flags
);
211 } else if (!memcmp("madvise", buf
,
212 min(sizeof("madvise")-1, count
))) {
213 clear_bit(enabled
, &transparent_hugepage_flags
);
214 set_bit(req_madv
, &transparent_hugepage_flags
);
215 } else if (!memcmp("never", buf
,
216 min(sizeof("never")-1, count
))) {
217 clear_bit(enabled
, &transparent_hugepage_flags
);
218 clear_bit(req_madv
, &transparent_hugepage_flags
);
225 static ssize_t
enabled_show(struct kobject
*kobj
,
226 struct kobj_attribute
*attr
, char *buf
)
228 return double_flag_show(kobj
, attr
, buf
,
229 TRANSPARENT_HUGEPAGE_FLAG
,
230 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
232 static ssize_t
enabled_store(struct kobject
*kobj
,
233 struct kobj_attribute
*attr
,
234 const char *buf
, size_t count
)
238 ret
= double_flag_store(kobj
, attr
, buf
, count
,
239 TRANSPARENT_HUGEPAGE_FLAG
,
240 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
245 mutex_lock(&khugepaged_mutex
);
246 err
= start_khugepaged();
247 mutex_unlock(&khugepaged_mutex
);
255 static struct kobj_attribute enabled_attr
=
256 __ATTR(enabled
, 0644, enabled_show
, enabled_store
);
258 static ssize_t
single_flag_show(struct kobject
*kobj
,
259 struct kobj_attribute
*attr
, char *buf
,
260 enum transparent_hugepage_flag flag
)
262 return sprintf(buf
, "%d\n",
263 !!test_bit(flag
, &transparent_hugepage_flags
));
266 static ssize_t
single_flag_store(struct kobject
*kobj
,
267 struct kobj_attribute
*attr
,
268 const char *buf
, size_t count
,
269 enum transparent_hugepage_flag flag
)
274 ret
= kstrtoul(buf
, 10, &value
);
281 set_bit(flag
, &transparent_hugepage_flags
);
283 clear_bit(flag
, &transparent_hugepage_flags
);
289 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
290 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
291 * memory just to allocate one more hugepage.
293 static ssize_t
defrag_show(struct kobject
*kobj
,
294 struct kobj_attribute
*attr
, char *buf
)
296 return double_flag_show(kobj
, attr
, buf
,
297 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
,
298 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
);
300 static ssize_t
defrag_store(struct kobject
*kobj
,
301 struct kobj_attribute
*attr
,
302 const char *buf
, size_t count
)
304 return double_flag_store(kobj
, attr
, buf
, count
,
305 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
,
306 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
);
308 static struct kobj_attribute defrag_attr
=
309 __ATTR(defrag
, 0644, defrag_show
, defrag_store
);
311 #ifdef CONFIG_DEBUG_VM
312 static ssize_t
debug_cow_show(struct kobject
*kobj
,
313 struct kobj_attribute
*attr
, char *buf
)
315 return single_flag_show(kobj
, attr
, buf
,
316 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
318 static ssize_t
debug_cow_store(struct kobject
*kobj
,
319 struct kobj_attribute
*attr
,
320 const char *buf
, size_t count
)
322 return single_flag_store(kobj
, attr
, buf
, count
,
323 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
325 static struct kobj_attribute debug_cow_attr
=
326 __ATTR(debug_cow
, 0644, debug_cow_show
, debug_cow_store
);
327 #endif /* CONFIG_DEBUG_VM */
329 static struct attribute
*hugepage_attr
[] = {
332 #ifdef CONFIG_DEBUG_VM
333 &debug_cow_attr
.attr
,
338 static struct attribute_group hugepage_attr_group
= {
339 .attrs
= hugepage_attr
,
342 static ssize_t
scan_sleep_millisecs_show(struct kobject
*kobj
,
343 struct kobj_attribute
*attr
,
346 return sprintf(buf
, "%u\n", khugepaged_scan_sleep_millisecs
);
349 static ssize_t
scan_sleep_millisecs_store(struct kobject
*kobj
,
350 struct kobj_attribute
*attr
,
351 const char *buf
, size_t count
)
356 err
= strict_strtoul(buf
, 10, &msecs
);
357 if (err
|| msecs
> UINT_MAX
)
360 khugepaged_scan_sleep_millisecs
= msecs
;
361 wake_up_interruptible(&khugepaged_wait
);
365 static struct kobj_attribute scan_sleep_millisecs_attr
=
366 __ATTR(scan_sleep_millisecs
, 0644, scan_sleep_millisecs_show
,
367 scan_sleep_millisecs_store
);
369 static ssize_t
alloc_sleep_millisecs_show(struct kobject
*kobj
,
370 struct kobj_attribute
*attr
,
373 return sprintf(buf
, "%u\n", khugepaged_alloc_sleep_millisecs
);
376 static ssize_t
alloc_sleep_millisecs_store(struct kobject
*kobj
,
377 struct kobj_attribute
*attr
,
378 const char *buf
, size_t count
)
383 err
= strict_strtoul(buf
, 10, &msecs
);
384 if (err
|| msecs
> UINT_MAX
)
387 khugepaged_alloc_sleep_millisecs
= msecs
;
388 wake_up_interruptible(&khugepaged_wait
);
392 static struct kobj_attribute alloc_sleep_millisecs_attr
=
393 __ATTR(alloc_sleep_millisecs
, 0644, alloc_sleep_millisecs_show
,
394 alloc_sleep_millisecs_store
);
396 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
397 struct kobj_attribute
*attr
,
400 return sprintf(buf
, "%u\n", khugepaged_pages_to_scan
);
402 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
403 struct kobj_attribute
*attr
,
404 const char *buf
, size_t count
)
409 err
= strict_strtoul(buf
, 10, &pages
);
410 if (err
|| !pages
|| pages
> UINT_MAX
)
413 khugepaged_pages_to_scan
= pages
;
417 static struct kobj_attribute pages_to_scan_attr
=
418 __ATTR(pages_to_scan
, 0644, pages_to_scan_show
,
419 pages_to_scan_store
);
421 static ssize_t
pages_collapsed_show(struct kobject
*kobj
,
422 struct kobj_attribute
*attr
,
425 return sprintf(buf
, "%u\n", khugepaged_pages_collapsed
);
427 static struct kobj_attribute pages_collapsed_attr
=
428 __ATTR_RO(pages_collapsed
);
430 static ssize_t
full_scans_show(struct kobject
*kobj
,
431 struct kobj_attribute
*attr
,
434 return sprintf(buf
, "%u\n", khugepaged_full_scans
);
436 static struct kobj_attribute full_scans_attr
=
437 __ATTR_RO(full_scans
);
439 static ssize_t
khugepaged_defrag_show(struct kobject
*kobj
,
440 struct kobj_attribute
*attr
, char *buf
)
442 return single_flag_show(kobj
, attr
, buf
,
443 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
445 static ssize_t
khugepaged_defrag_store(struct kobject
*kobj
,
446 struct kobj_attribute
*attr
,
447 const char *buf
, size_t count
)
449 return single_flag_store(kobj
, attr
, buf
, count
,
450 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
452 static struct kobj_attribute khugepaged_defrag_attr
=
453 __ATTR(defrag
, 0644, khugepaged_defrag_show
,
454 khugepaged_defrag_store
);
457 * max_ptes_none controls if khugepaged should collapse hugepages over
458 * any unmapped ptes in turn potentially increasing the memory
459 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
460 * reduce the available free memory in the system as it
461 * runs. Increasing max_ptes_none will instead potentially reduce the
462 * free memory in the system during the khugepaged scan.
464 static ssize_t
khugepaged_max_ptes_none_show(struct kobject
*kobj
,
465 struct kobj_attribute
*attr
,
468 return sprintf(buf
, "%u\n", khugepaged_max_ptes_none
);
470 static ssize_t
khugepaged_max_ptes_none_store(struct kobject
*kobj
,
471 struct kobj_attribute
*attr
,
472 const char *buf
, size_t count
)
475 unsigned long max_ptes_none
;
477 err
= strict_strtoul(buf
, 10, &max_ptes_none
);
478 if (err
|| max_ptes_none
> HPAGE_PMD_NR
-1)
481 khugepaged_max_ptes_none
= max_ptes_none
;
485 static struct kobj_attribute khugepaged_max_ptes_none_attr
=
486 __ATTR(max_ptes_none
, 0644, khugepaged_max_ptes_none_show
,
487 khugepaged_max_ptes_none_store
);
489 static struct attribute
*khugepaged_attr
[] = {
490 &khugepaged_defrag_attr
.attr
,
491 &khugepaged_max_ptes_none_attr
.attr
,
492 &pages_to_scan_attr
.attr
,
493 &pages_collapsed_attr
.attr
,
494 &full_scans_attr
.attr
,
495 &scan_sleep_millisecs_attr
.attr
,
496 &alloc_sleep_millisecs_attr
.attr
,
500 static struct attribute_group khugepaged_attr_group
= {
501 .attrs
= khugepaged_attr
,
502 .name
= "khugepaged",
505 static int __init
hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
509 *hugepage_kobj
= kobject_create_and_add("transparent_hugepage", mm_kobj
);
510 if (unlikely(!*hugepage_kobj
)) {
511 printk(KERN_ERR
"hugepage: failed kobject create\n");
515 err
= sysfs_create_group(*hugepage_kobj
, &hugepage_attr_group
);
517 printk(KERN_ERR
"hugepage: failed register hugeage group\n");
521 err
= sysfs_create_group(*hugepage_kobj
, &khugepaged_attr_group
);
523 printk(KERN_ERR
"hugepage: failed register hugeage group\n");
524 goto remove_hp_group
;
530 sysfs_remove_group(*hugepage_kobj
, &hugepage_attr_group
);
532 kobject_put(*hugepage_kobj
);
536 static void __init
hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
538 sysfs_remove_group(hugepage_kobj
, &khugepaged_attr_group
);
539 sysfs_remove_group(hugepage_kobj
, &hugepage_attr_group
);
540 kobject_put(hugepage_kobj
);
543 static inline int hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
548 static inline void hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
551 #endif /* CONFIG_SYSFS */
553 static int __init
hugepage_init(void)
556 struct kobject
*hugepage_kobj
;
558 if (!has_transparent_hugepage()) {
559 transparent_hugepage_flags
= 0;
563 err
= hugepage_init_sysfs(&hugepage_kobj
);
567 err
= init_huge_zero_page();
571 err
= khugepaged_slab_init();
575 err
= mm_slots_hash_init();
577 khugepaged_slab_free();
582 * By default disable transparent hugepages on smaller systems,
583 * where the extra memory used could hurt more than TLB overhead
584 * is likely to save. The admin can still enable it through /sys.
586 if (totalram_pages
< (512 << (20 - PAGE_SHIFT
)))
587 transparent_hugepage_flags
= 0;
594 __free_page(pfn_to_page(huge_zero_pfn
));
595 hugepage_exit_sysfs(hugepage_kobj
);
598 module_init(hugepage_init
)
600 static int __init
setup_transparent_hugepage(char *str
)
605 if (!strcmp(str
, "always")) {
606 set_bit(TRANSPARENT_HUGEPAGE_FLAG
,
607 &transparent_hugepage_flags
);
608 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
609 &transparent_hugepage_flags
);
611 } else if (!strcmp(str
, "madvise")) {
612 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
613 &transparent_hugepage_flags
);
614 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
615 &transparent_hugepage_flags
);
617 } else if (!strcmp(str
, "never")) {
618 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
619 &transparent_hugepage_flags
);
620 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
621 &transparent_hugepage_flags
);
627 "transparent_hugepage= cannot parse, ignored\n");
630 __setup("transparent_hugepage=", setup_transparent_hugepage
);
632 static inline pmd_t
maybe_pmd_mkwrite(pmd_t pmd
, struct vm_area_struct
*vma
)
634 if (likely(vma
->vm_flags
& VM_WRITE
))
635 pmd
= pmd_mkwrite(pmd
);
639 static inline pmd_t
mk_huge_pmd(struct page
*page
, struct vm_area_struct
*vma
)
642 entry
= mk_pmd(page
, vma
->vm_page_prot
);
643 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
644 entry
= pmd_mkhuge(entry
);
648 static int __do_huge_pmd_anonymous_page(struct mm_struct
*mm
,
649 struct vm_area_struct
*vma
,
650 unsigned long haddr
, pmd_t
*pmd
,
655 VM_BUG_ON(!PageCompound(page
));
656 pgtable
= pte_alloc_one(mm
, haddr
);
657 if (unlikely(!pgtable
))
660 clear_huge_page(page
, haddr
, HPAGE_PMD_NR
);
661 __SetPageUptodate(page
);
663 spin_lock(&mm
->page_table_lock
);
664 if (unlikely(!pmd_none(*pmd
))) {
665 spin_unlock(&mm
->page_table_lock
);
666 mem_cgroup_uncharge_page(page
);
668 pte_free(mm
, pgtable
);
671 entry
= mk_huge_pmd(page
, vma
);
673 * The spinlocking to take the lru_lock inside
674 * page_add_new_anon_rmap() acts as a full memory
675 * barrier to be sure clear_huge_page writes become
676 * visible after the set_pmd_at() write.
678 page_add_new_anon_rmap(page
, vma
, haddr
);
679 set_pmd_at(mm
, haddr
, pmd
, entry
);
680 pgtable_trans_huge_deposit(mm
, pgtable
);
681 add_mm_counter(mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
683 spin_unlock(&mm
->page_table_lock
);
689 static inline gfp_t
alloc_hugepage_gfpmask(int defrag
, gfp_t extra_gfp
)
691 return (GFP_TRANSHUGE
& ~(defrag
? 0 : __GFP_WAIT
)) | extra_gfp
;
694 static inline struct page
*alloc_hugepage_vma(int defrag
,
695 struct vm_area_struct
*vma
,
696 unsigned long haddr
, int nd
,
699 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag
, extra_gfp
),
700 HPAGE_PMD_ORDER
, vma
, haddr
, nd
);
704 static inline struct page
*alloc_hugepage(int defrag
)
706 return alloc_pages(alloc_hugepage_gfpmask(defrag
, 0),
711 static void set_huge_zero_page(pgtable_t pgtable
, struct mm_struct
*mm
,
712 struct vm_area_struct
*vma
, unsigned long haddr
, pmd_t
*pmd
)
715 entry
= pfn_pmd(huge_zero_pfn
, vma
->vm_page_prot
);
716 entry
= pmd_wrprotect(entry
);
717 entry
= pmd_mkhuge(entry
);
718 set_pmd_at(mm
, haddr
, pmd
, entry
);
719 pgtable_trans_huge_deposit(mm
, pgtable
);
723 int do_huge_pmd_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
724 unsigned long address
, pmd_t
*pmd
,
728 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
731 if (haddr
>= vma
->vm_start
&& haddr
+ HPAGE_PMD_SIZE
<= vma
->vm_end
) {
732 if (unlikely(anon_vma_prepare(vma
)))
734 if (unlikely(khugepaged_enter(vma
)))
736 page
= alloc_hugepage_vma(transparent_hugepage_defrag(vma
),
737 vma
, haddr
, numa_node_id(), 0);
738 if (unlikely(!page
)) {
739 count_vm_event(THP_FAULT_FALLBACK
);
742 count_vm_event(THP_FAULT_ALLOC
);
743 if (unlikely(mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))) {
747 if (unlikely(__do_huge_pmd_anonymous_page(mm
, vma
, haddr
, pmd
,
749 mem_cgroup_uncharge_page(page
);
758 * Use __pte_alloc instead of pte_alloc_map, because we can't
759 * run pte_offset_map on the pmd, if an huge pmd could
760 * materialize from under us from a different thread.
762 if (unlikely(__pte_alloc(mm
, vma
, pmd
, address
)))
764 /* if an huge pmd materialized from under us just retry later */
765 if (unlikely(pmd_trans_huge(*pmd
)))
768 * A regular pmd is established and it can't morph into a huge pmd
769 * from under us anymore at this point because we hold the mmap_sem
770 * read mode and khugepaged takes it in write mode. So now it's
771 * safe to run pte_offset_map().
773 pte
= pte_offset_map(pmd
, address
);
774 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
777 int copy_huge_pmd(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
778 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
779 struct vm_area_struct
*vma
)
781 struct page
*src_page
;
787 pgtable
= pte_alloc_one(dst_mm
, addr
);
788 if (unlikely(!pgtable
))
791 spin_lock(&dst_mm
->page_table_lock
);
792 spin_lock_nested(&src_mm
->page_table_lock
, SINGLE_DEPTH_NESTING
);
796 if (unlikely(!pmd_trans_huge(pmd
))) {
797 pte_free(dst_mm
, pgtable
);
801 * mm->page_table_lock is enough to be sure that huge zero pmd is not
802 * under splitting since we don't split the page itself, only pmd to
805 if (is_huge_zero_pmd(pmd
)) {
806 set_huge_zero_page(pgtable
, dst_mm
, vma
, addr
, dst_pmd
);
810 if (unlikely(pmd_trans_splitting(pmd
))) {
811 /* split huge page running from under us */
812 spin_unlock(&src_mm
->page_table_lock
);
813 spin_unlock(&dst_mm
->page_table_lock
);
814 pte_free(dst_mm
, pgtable
);
816 wait_split_huge_page(vma
->anon_vma
, src_pmd
); /* src_vma */
819 src_page
= pmd_page(pmd
);
820 VM_BUG_ON(!PageHead(src_page
));
822 page_dup_rmap(src_page
);
823 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
825 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
826 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
827 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
828 pgtable_trans_huge_deposit(dst_mm
, pgtable
);
833 spin_unlock(&src_mm
->page_table_lock
);
834 spin_unlock(&dst_mm
->page_table_lock
);
839 void huge_pmd_set_accessed(struct mm_struct
*mm
,
840 struct vm_area_struct
*vma
,
841 unsigned long address
,
842 pmd_t
*pmd
, pmd_t orig_pmd
,
848 spin_lock(&mm
->page_table_lock
);
849 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
852 entry
= pmd_mkyoung(orig_pmd
);
853 haddr
= address
& HPAGE_PMD_MASK
;
854 if (pmdp_set_access_flags(vma
, haddr
, pmd
, entry
, dirty
))
855 update_mmu_cache_pmd(vma
, address
, pmd
);
858 spin_unlock(&mm
->page_table_lock
);
861 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct
*mm
,
862 struct vm_area_struct
*vma
, unsigned long address
,
863 pmd_t
*pmd
, unsigned long haddr
)
869 unsigned long mmun_start
; /* For mmu_notifiers */
870 unsigned long mmun_end
; /* For mmu_notifiers */
872 page
= alloc_page_vma(GFP_HIGHUSER_MOVABLE
, vma
, address
);
878 if (mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
)) {
884 clear_user_highpage(page
, address
);
885 __SetPageUptodate(page
);
888 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
889 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
891 spin_lock(&mm
->page_table_lock
);
892 pmdp_clear_flush(vma
, haddr
, pmd
);
893 /* leave pmd empty until pte is filled */
895 pgtable
= pgtable_trans_huge_withdraw(mm
);
896 pmd_populate(mm
, &_pmd
, pgtable
);
898 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
900 if (haddr
== (address
& PAGE_MASK
)) {
901 entry
= mk_pte(page
, vma
->vm_page_prot
);
902 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
903 page_add_new_anon_rmap(page
, vma
, haddr
);
905 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
906 entry
= pte_mkspecial(entry
);
908 pte
= pte_offset_map(&_pmd
, haddr
);
909 VM_BUG_ON(!pte_none(*pte
));
910 set_pte_at(mm
, haddr
, pte
, entry
);
913 smp_wmb(); /* make pte visible before pmd */
914 pmd_populate(mm
, pmd
, pgtable
);
915 spin_unlock(&mm
->page_table_lock
);
916 inc_mm_counter(mm
, MM_ANONPAGES
);
918 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
920 ret
|= VM_FAULT_WRITE
;
925 static int do_huge_pmd_wp_page_fallback(struct mm_struct
*mm
,
926 struct vm_area_struct
*vma
,
927 unsigned long address
,
928 pmd_t
*pmd
, pmd_t orig_pmd
,
936 unsigned long mmun_start
; /* For mmu_notifiers */
937 unsigned long mmun_end
; /* For mmu_notifiers */
939 pages
= kmalloc(sizeof(struct page
*) * HPAGE_PMD_NR
,
941 if (unlikely(!pages
)) {
946 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
947 pages
[i
] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE
|
949 vma
, address
, page_to_nid(page
));
950 if (unlikely(!pages
[i
] ||
951 mem_cgroup_newpage_charge(pages
[i
], mm
,
955 mem_cgroup_uncharge_start();
957 mem_cgroup_uncharge_page(pages
[i
]);
960 mem_cgroup_uncharge_end();
967 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
968 copy_user_highpage(pages
[i
], page
+ i
,
969 haddr
+ PAGE_SIZE
* i
, vma
);
970 __SetPageUptodate(pages
[i
]);
975 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
976 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
978 spin_lock(&mm
->page_table_lock
);
979 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
981 VM_BUG_ON(!PageHead(page
));
983 pmdp_clear_flush(vma
, haddr
, pmd
);
984 /* leave pmd empty until pte is filled */
986 pgtable
= pgtable_trans_huge_withdraw(mm
);
987 pmd_populate(mm
, &_pmd
, pgtable
);
989 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
991 entry
= mk_pte(pages
[i
], vma
->vm_page_prot
);
992 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
993 page_add_new_anon_rmap(pages
[i
], vma
, haddr
);
994 pte
= pte_offset_map(&_pmd
, haddr
);
995 VM_BUG_ON(!pte_none(*pte
));
996 set_pte_at(mm
, haddr
, pte
, entry
);
1001 smp_wmb(); /* make pte visible before pmd */
1002 pmd_populate(mm
, pmd
, pgtable
);
1003 page_remove_rmap(page
);
1004 spin_unlock(&mm
->page_table_lock
);
1006 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1008 ret
|= VM_FAULT_WRITE
;
1015 spin_unlock(&mm
->page_table_lock
);
1016 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1017 mem_cgroup_uncharge_start();
1018 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1019 mem_cgroup_uncharge_page(pages
[i
]);
1022 mem_cgroup_uncharge_end();
1027 int do_huge_pmd_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1028 unsigned long address
, pmd_t
*pmd
, pmd_t orig_pmd
)
1031 struct page
*page
= NULL
, *new_page
;
1032 unsigned long haddr
;
1033 unsigned long mmun_start
; /* For mmu_notifiers */
1034 unsigned long mmun_end
; /* For mmu_notifiers */
1036 VM_BUG_ON(!vma
->anon_vma
);
1037 haddr
= address
& HPAGE_PMD_MASK
;
1038 if (is_huge_zero_pmd(orig_pmd
))
1040 spin_lock(&mm
->page_table_lock
);
1041 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
1044 page
= pmd_page(orig_pmd
);
1045 VM_BUG_ON(!PageCompound(page
) || !PageHead(page
));
1046 if (page_mapcount(page
) == 1) {
1048 entry
= pmd_mkyoung(orig_pmd
);
1049 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1050 if (pmdp_set_access_flags(vma
, haddr
, pmd
, entry
, 1))
1051 update_mmu_cache_pmd(vma
, address
, pmd
);
1052 ret
|= VM_FAULT_WRITE
;
1056 spin_unlock(&mm
->page_table_lock
);
1058 if (transparent_hugepage_enabled(vma
) &&
1059 !transparent_hugepage_debug_cow())
1060 new_page
= alloc_hugepage_vma(transparent_hugepage_defrag(vma
),
1061 vma
, haddr
, numa_node_id(), 0);
1065 if (unlikely(!new_page
)) {
1066 count_vm_event(THP_FAULT_FALLBACK
);
1067 if (is_huge_zero_pmd(orig_pmd
)) {
1068 ret
= do_huge_pmd_wp_zero_page_fallback(mm
, vma
,
1069 address
, pmd
, haddr
);
1071 ret
= do_huge_pmd_wp_page_fallback(mm
, vma
, address
,
1072 pmd
, orig_pmd
, page
, haddr
);
1073 if (ret
& VM_FAULT_OOM
)
1074 split_huge_page(page
);
1079 count_vm_event(THP_FAULT_ALLOC
);
1081 if (unlikely(mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))) {
1084 split_huge_page(page
);
1087 ret
|= VM_FAULT_OOM
;
1091 if (is_huge_zero_pmd(orig_pmd
))
1092 clear_huge_page(new_page
, haddr
, HPAGE_PMD_NR
);
1094 copy_user_huge_page(new_page
, page
, haddr
, vma
, HPAGE_PMD_NR
);
1095 __SetPageUptodate(new_page
);
1098 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1099 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1101 spin_lock(&mm
->page_table_lock
);
1104 if (unlikely(!pmd_same(*pmd
, orig_pmd
))) {
1105 spin_unlock(&mm
->page_table_lock
);
1106 mem_cgroup_uncharge_page(new_page
);
1111 entry
= mk_huge_pmd(new_page
, vma
);
1112 pmdp_clear_flush(vma
, haddr
, pmd
);
1113 page_add_new_anon_rmap(new_page
, vma
, haddr
);
1114 set_pmd_at(mm
, haddr
, pmd
, entry
);
1115 update_mmu_cache_pmd(vma
, address
, pmd
);
1116 if (is_huge_zero_pmd(orig_pmd
))
1117 add_mm_counter(mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1119 VM_BUG_ON(!PageHead(page
));
1120 page_remove_rmap(page
);
1123 ret
|= VM_FAULT_WRITE
;
1125 spin_unlock(&mm
->page_table_lock
);
1127 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1131 spin_unlock(&mm
->page_table_lock
);
1135 struct page
*follow_trans_huge_pmd(struct vm_area_struct
*vma
,
1140 struct mm_struct
*mm
= vma
->vm_mm
;
1141 struct page
*page
= NULL
;
1143 assert_spin_locked(&mm
->page_table_lock
);
1145 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
1148 page
= pmd_page(*pmd
);
1149 VM_BUG_ON(!PageHead(page
));
1150 if (flags
& FOLL_TOUCH
) {
1153 * We should set the dirty bit only for FOLL_WRITE but
1154 * for now the dirty bit in the pmd is meaningless.
1155 * And if the dirty bit will become meaningful and
1156 * we'll only set it with FOLL_WRITE, an atomic
1157 * set_bit will be required on the pmd to set the
1158 * young bit, instead of the current set_pmd_at.
1160 _pmd
= pmd_mkyoung(pmd_mkdirty(*pmd
));
1161 set_pmd_at(mm
, addr
& HPAGE_PMD_MASK
, pmd
, _pmd
);
1163 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1164 if (page
->mapping
&& trylock_page(page
)) {
1167 mlock_vma_page(page
);
1171 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
1172 VM_BUG_ON(!PageCompound(page
));
1173 if (flags
& FOLL_GET
)
1174 get_page_foll(page
);
1180 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1181 pmd_t
*pmd
, unsigned long addr
)
1185 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
1189 pgtable
= pgtable_trans_huge_withdraw(tlb
->mm
);
1190 orig_pmd
= pmdp_get_and_clear(tlb
->mm
, addr
, pmd
);
1191 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1192 if (is_huge_zero_pmd(orig_pmd
)) {
1194 spin_unlock(&tlb
->mm
->page_table_lock
);
1196 page
= pmd_page(orig_pmd
);
1197 page_remove_rmap(page
);
1198 VM_BUG_ON(page_mapcount(page
) < 0);
1199 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1200 VM_BUG_ON(!PageHead(page
));
1202 spin_unlock(&tlb
->mm
->page_table_lock
);
1203 tlb_remove_page(tlb
, page
);
1205 pte_free(tlb
->mm
, pgtable
);
1211 int mincore_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1212 unsigned long addr
, unsigned long end
,
1217 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
1219 * All logical pages in the range are present
1220 * if backed by a huge page.
1222 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1223 memset(vec
, 1, (end
- addr
) >> PAGE_SHIFT
);
1230 int move_huge_pmd(struct vm_area_struct
*vma
, struct vm_area_struct
*new_vma
,
1231 unsigned long old_addr
,
1232 unsigned long new_addr
, unsigned long old_end
,
1233 pmd_t
*old_pmd
, pmd_t
*new_pmd
)
1238 struct mm_struct
*mm
= vma
->vm_mm
;
1240 if ((old_addr
& ~HPAGE_PMD_MASK
) ||
1241 (new_addr
& ~HPAGE_PMD_MASK
) ||
1242 old_end
- old_addr
< HPAGE_PMD_SIZE
||
1243 (new_vma
->vm_flags
& VM_NOHUGEPAGE
))
1247 * The destination pmd shouldn't be established, free_pgtables()
1248 * should have release it.
1250 if (WARN_ON(!pmd_none(*new_pmd
))) {
1251 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1255 ret
= __pmd_trans_huge_lock(old_pmd
, vma
);
1257 pmd
= pmdp_get_and_clear(mm
, old_addr
, old_pmd
);
1258 VM_BUG_ON(!pmd_none(*new_pmd
));
1259 set_pmd_at(mm
, new_addr
, new_pmd
, pmd
);
1260 spin_unlock(&mm
->page_table_lock
);
1266 int change_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1267 unsigned long addr
, pgprot_t newprot
)
1269 struct mm_struct
*mm
= vma
->vm_mm
;
1272 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
1274 entry
= pmdp_get_and_clear(mm
, addr
, pmd
);
1275 entry
= pmd_modify(entry
, newprot
);
1276 BUG_ON(pmd_write(entry
));
1277 set_pmd_at(mm
, addr
, pmd
, entry
);
1278 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1286 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1287 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1289 * Note that if it returns 1, this routine returns without unlocking page
1290 * table locks. So callers must unlock them.
1292 int __pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
)
1294 spin_lock(&vma
->vm_mm
->page_table_lock
);
1295 if (likely(pmd_trans_huge(*pmd
))) {
1296 if (unlikely(pmd_trans_splitting(*pmd
))) {
1297 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1298 wait_split_huge_page(vma
->anon_vma
, pmd
);
1301 /* Thp mapped by 'pmd' is stable, so we can
1302 * handle it as it is. */
1306 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1310 pmd_t
*page_check_address_pmd(struct page
*page
,
1311 struct mm_struct
*mm
,
1312 unsigned long address
,
1313 enum page_check_address_pmd_flag flag
)
1315 pmd_t
*pmd
, *ret
= NULL
;
1317 if (address
& ~HPAGE_PMD_MASK
)
1320 pmd
= mm_find_pmd(mm
, address
);
1325 if (pmd_page(*pmd
) != page
)
1328 * split_vma() may create temporary aliased mappings. There is
1329 * no risk as long as all huge pmd are found and have their
1330 * splitting bit set before __split_huge_page_refcount
1331 * runs. Finding the same huge pmd more than once during the
1332 * same rmap walk is not a problem.
1334 if (flag
== PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG
&&
1335 pmd_trans_splitting(*pmd
))
1337 if (pmd_trans_huge(*pmd
)) {
1338 VM_BUG_ON(flag
== PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG
&&
1339 !pmd_trans_splitting(*pmd
));
1346 static int __split_huge_page_splitting(struct page
*page
,
1347 struct vm_area_struct
*vma
,
1348 unsigned long address
)
1350 struct mm_struct
*mm
= vma
->vm_mm
;
1353 /* For mmu_notifiers */
1354 const unsigned long mmun_start
= address
;
1355 const unsigned long mmun_end
= address
+ HPAGE_PMD_SIZE
;
1357 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1358 spin_lock(&mm
->page_table_lock
);
1359 pmd
= page_check_address_pmd(page
, mm
, address
,
1360 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG
);
1363 * We can't temporarily set the pmd to null in order
1364 * to split it, the pmd must remain marked huge at all
1365 * times or the VM won't take the pmd_trans_huge paths
1366 * and it won't wait on the anon_vma->root->mutex to
1367 * serialize against split_huge_page*.
1369 pmdp_splitting_flush(vma
, address
, pmd
);
1372 spin_unlock(&mm
->page_table_lock
);
1373 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1378 static void __split_huge_page_refcount(struct page
*page
)
1381 struct zone
*zone
= page_zone(page
);
1382 struct lruvec
*lruvec
;
1385 /* prevent PageLRU to go away from under us, and freeze lru stats */
1386 spin_lock_irq(&zone
->lru_lock
);
1387 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1389 compound_lock(page
);
1390 /* complete memcg works before add pages to LRU */
1391 mem_cgroup_split_huge_fixup(page
);
1393 for (i
= HPAGE_PMD_NR
- 1; i
>= 1; i
--) {
1394 struct page
*page_tail
= page
+ i
;
1396 /* tail_page->_mapcount cannot change */
1397 BUG_ON(page_mapcount(page_tail
) < 0);
1398 tail_count
+= page_mapcount(page_tail
);
1399 /* check for overflow */
1400 BUG_ON(tail_count
< 0);
1401 BUG_ON(atomic_read(&page_tail
->_count
) != 0);
1403 * tail_page->_count is zero and not changing from
1404 * under us. But get_page_unless_zero() may be running
1405 * from under us on the tail_page. If we used
1406 * atomic_set() below instead of atomic_add(), we
1407 * would then run atomic_set() concurrently with
1408 * get_page_unless_zero(), and atomic_set() is
1409 * implemented in C not using locked ops. spin_unlock
1410 * on x86 sometime uses locked ops because of PPro
1411 * errata 66, 92, so unless somebody can guarantee
1412 * atomic_set() here would be safe on all archs (and
1413 * not only on x86), it's safer to use atomic_add().
1415 atomic_add(page_mapcount(page
) + page_mapcount(page_tail
) + 1,
1416 &page_tail
->_count
);
1418 /* after clearing PageTail the gup refcount can be released */
1422 * retain hwpoison flag of the poisoned tail page:
1423 * fix for the unsuitable process killed on Guest Machine(KVM)
1424 * by the memory-failure.
1426 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
| __PG_HWPOISON
;
1427 page_tail
->flags
|= (page
->flags
&
1428 ((1L << PG_referenced
) |
1429 (1L << PG_swapbacked
) |
1430 (1L << PG_mlocked
) |
1431 (1L << PG_uptodate
)));
1432 page_tail
->flags
|= (1L << PG_dirty
);
1434 /* clear PageTail before overwriting first_page */
1438 * __split_huge_page_splitting() already set the
1439 * splitting bit in all pmd that could map this
1440 * hugepage, that will ensure no CPU can alter the
1441 * mapcount on the head page. The mapcount is only
1442 * accounted in the head page and it has to be
1443 * transferred to all tail pages in the below code. So
1444 * for this code to be safe, the split the mapcount
1445 * can't change. But that doesn't mean userland can't
1446 * keep changing and reading the page contents while
1447 * we transfer the mapcount, so the pmd splitting
1448 * status is achieved setting a reserved bit in the
1449 * pmd, not by clearing the present bit.
1451 page_tail
->_mapcount
= page
->_mapcount
;
1453 BUG_ON(page_tail
->mapping
);
1454 page_tail
->mapping
= page
->mapping
;
1456 page_tail
->index
= page
->index
+ i
;
1458 BUG_ON(!PageAnon(page_tail
));
1459 BUG_ON(!PageUptodate(page_tail
));
1460 BUG_ON(!PageDirty(page_tail
));
1461 BUG_ON(!PageSwapBacked(page_tail
));
1463 lru_add_page_tail(page
, page_tail
, lruvec
);
1465 atomic_sub(tail_count
, &page
->_count
);
1466 BUG_ON(atomic_read(&page
->_count
) <= 0);
1468 __mod_zone_page_state(zone
, NR_ANON_TRANSPARENT_HUGEPAGES
, -1);
1469 __mod_zone_page_state(zone
, NR_ANON_PAGES
, HPAGE_PMD_NR
);
1471 ClearPageCompound(page
);
1472 compound_unlock(page
);
1473 spin_unlock_irq(&zone
->lru_lock
);
1475 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
1476 struct page
*page_tail
= page
+ i
;
1477 BUG_ON(page_count(page_tail
) <= 0);
1479 * Tail pages may be freed if there wasn't any mapping
1480 * like if add_to_swap() is running on a lru page that
1481 * had its mapping zapped. And freeing these pages
1482 * requires taking the lru_lock so we do the put_page
1483 * of the tail pages after the split is complete.
1485 put_page(page_tail
);
1489 * Only the head page (now become a regular page) is required
1490 * to be pinned by the caller.
1492 BUG_ON(page_count(page
) <= 0);
1495 static int __split_huge_page_map(struct page
*page
,
1496 struct vm_area_struct
*vma
,
1497 unsigned long address
)
1499 struct mm_struct
*mm
= vma
->vm_mm
;
1503 unsigned long haddr
;
1505 spin_lock(&mm
->page_table_lock
);
1506 pmd
= page_check_address_pmd(page
, mm
, address
,
1507 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG
);
1509 pgtable
= pgtable_trans_huge_withdraw(mm
);
1510 pmd_populate(mm
, &_pmd
, pgtable
);
1513 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1515 BUG_ON(PageCompound(page
+i
));
1516 entry
= mk_pte(page
+ i
, vma
->vm_page_prot
);
1517 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1518 if (!pmd_write(*pmd
))
1519 entry
= pte_wrprotect(entry
);
1521 BUG_ON(page_mapcount(page
) != 1);
1522 if (!pmd_young(*pmd
))
1523 entry
= pte_mkold(entry
);
1524 pte
= pte_offset_map(&_pmd
, haddr
);
1525 BUG_ON(!pte_none(*pte
));
1526 set_pte_at(mm
, haddr
, pte
, entry
);
1530 smp_wmb(); /* make pte visible before pmd */
1532 * Up to this point the pmd is present and huge and
1533 * userland has the whole access to the hugepage
1534 * during the split (which happens in place). If we
1535 * overwrite the pmd with the not-huge version
1536 * pointing to the pte here (which of course we could
1537 * if all CPUs were bug free), userland could trigger
1538 * a small page size TLB miss on the small sized TLB
1539 * while the hugepage TLB entry is still established
1540 * in the huge TLB. Some CPU doesn't like that. See
1541 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1542 * Erratum 383 on page 93. Intel should be safe but is
1543 * also warns that it's only safe if the permission
1544 * and cache attributes of the two entries loaded in
1545 * the two TLB is identical (which should be the case
1546 * here). But it is generally safer to never allow
1547 * small and huge TLB entries for the same virtual
1548 * address to be loaded simultaneously. So instead of
1549 * doing "pmd_populate(); flush_tlb_range();" we first
1550 * mark the current pmd notpresent (atomically because
1551 * here the pmd_trans_huge and pmd_trans_splitting
1552 * must remain set at all times on the pmd until the
1553 * split is complete for this pmd), then we flush the
1554 * SMP TLB and finally we write the non-huge version
1555 * of the pmd entry with pmd_populate.
1557 pmdp_invalidate(vma
, address
, pmd
);
1558 pmd_populate(mm
, pmd
, pgtable
);
1561 spin_unlock(&mm
->page_table_lock
);
1566 /* must be called with anon_vma->root->mutex hold */
1567 static void __split_huge_page(struct page
*page
,
1568 struct anon_vma
*anon_vma
)
1570 int mapcount
, mapcount2
;
1571 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1572 struct anon_vma_chain
*avc
;
1574 BUG_ON(!PageHead(page
));
1575 BUG_ON(PageTail(page
));
1578 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
1579 struct vm_area_struct
*vma
= avc
->vma
;
1580 unsigned long addr
= vma_address(page
, vma
);
1581 BUG_ON(is_vma_temporary_stack(vma
));
1582 mapcount
+= __split_huge_page_splitting(page
, vma
, addr
);
1585 * It is critical that new vmas are added to the tail of the
1586 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1587 * and establishes a child pmd before
1588 * __split_huge_page_splitting() freezes the parent pmd (so if
1589 * we fail to prevent copy_huge_pmd() from running until the
1590 * whole __split_huge_page() is complete), we will still see
1591 * the newly established pmd of the child later during the
1592 * walk, to be able to set it as pmd_trans_splitting too.
1594 if (mapcount
!= page_mapcount(page
))
1595 printk(KERN_ERR
"mapcount %d page_mapcount %d\n",
1596 mapcount
, page_mapcount(page
));
1597 BUG_ON(mapcount
!= page_mapcount(page
));
1599 __split_huge_page_refcount(page
);
1602 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
1603 struct vm_area_struct
*vma
= avc
->vma
;
1604 unsigned long addr
= vma_address(page
, vma
);
1605 BUG_ON(is_vma_temporary_stack(vma
));
1606 mapcount2
+= __split_huge_page_map(page
, vma
, addr
);
1608 if (mapcount
!= mapcount2
)
1609 printk(KERN_ERR
"mapcount %d mapcount2 %d page_mapcount %d\n",
1610 mapcount
, mapcount2
, page_mapcount(page
));
1611 BUG_ON(mapcount
!= mapcount2
);
1614 int split_huge_page(struct page
*page
)
1616 struct anon_vma
*anon_vma
;
1619 BUG_ON(!PageAnon(page
));
1620 anon_vma
= page_lock_anon_vma(page
);
1624 if (!PageCompound(page
))
1627 BUG_ON(!PageSwapBacked(page
));
1628 __split_huge_page(page
, anon_vma
);
1629 count_vm_event(THP_SPLIT
);
1631 BUG_ON(PageCompound(page
));
1633 page_unlock_anon_vma(anon_vma
);
1638 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1640 int hugepage_madvise(struct vm_area_struct
*vma
,
1641 unsigned long *vm_flags
, int advice
)
1643 struct mm_struct
*mm
= vma
->vm_mm
;
1648 * Be somewhat over-protective like KSM for now!
1650 if (*vm_flags
& (VM_HUGEPAGE
| VM_NO_THP
))
1652 if (mm
->def_flags
& VM_NOHUGEPAGE
)
1654 *vm_flags
&= ~VM_NOHUGEPAGE
;
1655 *vm_flags
|= VM_HUGEPAGE
;
1657 * If the vma become good for khugepaged to scan,
1658 * register it here without waiting a page fault that
1659 * may not happen any time soon.
1661 if (unlikely(khugepaged_enter_vma_merge(vma
)))
1664 case MADV_NOHUGEPAGE
:
1666 * Be somewhat over-protective like KSM for now!
1668 if (*vm_flags
& (VM_NOHUGEPAGE
| VM_NO_THP
))
1670 *vm_flags
&= ~VM_HUGEPAGE
;
1671 *vm_flags
|= VM_NOHUGEPAGE
;
1673 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1674 * this vma even if we leave the mm registered in khugepaged if
1675 * it got registered before VM_NOHUGEPAGE was set.
1683 static int __init
khugepaged_slab_init(void)
1685 mm_slot_cache
= kmem_cache_create("khugepaged_mm_slot",
1686 sizeof(struct mm_slot
),
1687 __alignof__(struct mm_slot
), 0, NULL
);
1694 static void __init
khugepaged_slab_free(void)
1696 kmem_cache_destroy(mm_slot_cache
);
1697 mm_slot_cache
= NULL
;
1700 static inline struct mm_slot
*alloc_mm_slot(void)
1702 if (!mm_slot_cache
) /* initialization failed */
1704 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
1707 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
1709 kmem_cache_free(mm_slot_cache
, mm_slot
);
1712 static int __init
mm_slots_hash_init(void)
1714 mm_slots_hash
= kzalloc(MM_SLOTS_HASH_HEADS
* sizeof(struct hlist_head
),
1722 static void __init
mm_slots_hash_free(void)
1724 kfree(mm_slots_hash
);
1725 mm_slots_hash
= NULL
;
1729 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
1731 struct mm_slot
*mm_slot
;
1732 struct hlist_head
*bucket
;
1733 struct hlist_node
*node
;
1735 bucket
= &mm_slots_hash
[((unsigned long)mm
/ sizeof(struct mm_struct
))
1736 % MM_SLOTS_HASH_HEADS
];
1737 hlist_for_each_entry(mm_slot
, node
, bucket
, hash
) {
1738 if (mm
== mm_slot
->mm
)
1744 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
1745 struct mm_slot
*mm_slot
)
1747 struct hlist_head
*bucket
;
1749 bucket
= &mm_slots_hash
[((unsigned long)mm
/ sizeof(struct mm_struct
))
1750 % MM_SLOTS_HASH_HEADS
];
1752 hlist_add_head(&mm_slot
->hash
, bucket
);
1755 static inline int khugepaged_test_exit(struct mm_struct
*mm
)
1757 return atomic_read(&mm
->mm_users
) == 0;
1760 int __khugepaged_enter(struct mm_struct
*mm
)
1762 struct mm_slot
*mm_slot
;
1765 mm_slot
= alloc_mm_slot();
1769 /* __khugepaged_exit() must not run from under us */
1770 VM_BUG_ON(khugepaged_test_exit(mm
));
1771 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE
, &mm
->flags
))) {
1772 free_mm_slot(mm_slot
);
1776 spin_lock(&khugepaged_mm_lock
);
1777 insert_to_mm_slots_hash(mm
, mm_slot
);
1779 * Insert just behind the scanning cursor, to let the area settle
1782 wakeup
= list_empty(&khugepaged_scan
.mm_head
);
1783 list_add_tail(&mm_slot
->mm_node
, &khugepaged_scan
.mm_head
);
1784 spin_unlock(&khugepaged_mm_lock
);
1786 atomic_inc(&mm
->mm_count
);
1788 wake_up_interruptible(&khugepaged_wait
);
1793 int khugepaged_enter_vma_merge(struct vm_area_struct
*vma
)
1795 unsigned long hstart
, hend
;
1798 * Not yet faulted in so we will register later in the
1799 * page fault if needed.
1803 /* khugepaged not yet working on file or special mappings */
1805 VM_BUG_ON(vma
->vm_flags
& VM_NO_THP
);
1806 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
1807 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
1809 return khugepaged_enter(vma
);
1813 void __khugepaged_exit(struct mm_struct
*mm
)
1815 struct mm_slot
*mm_slot
;
1818 spin_lock(&khugepaged_mm_lock
);
1819 mm_slot
= get_mm_slot(mm
);
1820 if (mm_slot
&& khugepaged_scan
.mm_slot
!= mm_slot
) {
1821 hlist_del(&mm_slot
->hash
);
1822 list_del(&mm_slot
->mm_node
);
1825 spin_unlock(&khugepaged_mm_lock
);
1828 clear_bit(MMF_VM_HUGEPAGE
, &mm
->flags
);
1829 free_mm_slot(mm_slot
);
1831 } else if (mm_slot
) {
1833 * This is required to serialize against
1834 * khugepaged_test_exit() (which is guaranteed to run
1835 * under mmap sem read mode). Stop here (after we
1836 * return all pagetables will be destroyed) until
1837 * khugepaged has finished working on the pagetables
1838 * under the mmap_sem.
1840 down_write(&mm
->mmap_sem
);
1841 up_write(&mm
->mmap_sem
);
1845 static void release_pte_page(struct page
*page
)
1847 /* 0 stands for page_is_file_cache(page) == false */
1848 dec_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
1850 putback_lru_page(page
);
1853 static void release_pte_pages(pte_t
*pte
, pte_t
*_pte
)
1855 while (--_pte
>= pte
) {
1856 pte_t pteval
= *_pte
;
1857 if (!pte_none(pteval
))
1858 release_pte_page(pte_page(pteval
));
1862 static int __collapse_huge_page_isolate(struct vm_area_struct
*vma
,
1863 unsigned long address
,
1868 int referenced
= 0, none
= 0;
1869 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
1870 _pte
++, address
+= PAGE_SIZE
) {
1871 pte_t pteval
= *_pte
;
1872 if (pte_none(pteval
)) {
1873 if (++none
<= khugepaged_max_ptes_none
)
1878 if (!pte_present(pteval
) || !pte_write(pteval
))
1880 page
= vm_normal_page(vma
, address
, pteval
);
1881 if (unlikely(!page
))
1884 VM_BUG_ON(PageCompound(page
));
1885 BUG_ON(!PageAnon(page
));
1886 VM_BUG_ON(!PageSwapBacked(page
));
1888 /* cannot use mapcount: can't collapse if there's a gup pin */
1889 if (page_count(page
) != 1)
1892 * We can do it before isolate_lru_page because the
1893 * page can't be freed from under us. NOTE: PG_lock
1894 * is needed to serialize against split_huge_page
1895 * when invoked from the VM.
1897 if (!trylock_page(page
))
1900 * Isolate the page to avoid collapsing an hugepage
1901 * currently in use by the VM.
1903 if (isolate_lru_page(page
)) {
1907 /* 0 stands for page_is_file_cache(page) == false */
1908 inc_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
1909 VM_BUG_ON(!PageLocked(page
));
1910 VM_BUG_ON(PageLRU(page
));
1912 /* If there is no mapped pte young don't collapse the page */
1913 if (pte_young(pteval
) || PageReferenced(page
) ||
1914 mmu_notifier_test_young(vma
->vm_mm
, address
))
1917 if (likely(referenced
))
1920 release_pte_pages(pte
, _pte
);
1924 static void __collapse_huge_page_copy(pte_t
*pte
, struct page
*page
,
1925 struct vm_area_struct
*vma
,
1926 unsigned long address
,
1930 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
; _pte
++) {
1931 pte_t pteval
= *_pte
;
1932 struct page
*src_page
;
1934 if (pte_none(pteval
)) {
1935 clear_user_highpage(page
, address
);
1936 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, 1);
1938 src_page
= pte_page(pteval
);
1939 copy_user_highpage(page
, src_page
, address
, vma
);
1940 VM_BUG_ON(page_mapcount(src_page
) != 1);
1941 release_pte_page(src_page
);
1943 * ptl mostly unnecessary, but preempt has to
1944 * be disabled to update the per-cpu stats
1945 * inside page_remove_rmap().
1949 * paravirt calls inside pte_clear here are
1952 pte_clear(vma
->vm_mm
, address
, _pte
);
1953 page_remove_rmap(src_page
);
1955 free_page_and_swap_cache(src_page
);
1958 address
+= PAGE_SIZE
;
1963 static void khugepaged_alloc_sleep(void)
1965 wait_event_freezable_timeout(khugepaged_wait
, false,
1966 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs
));
1970 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
1972 if (IS_ERR(*hpage
)) {
1978 khugepaged_alloc_sleep();
1979 } else if (*hpage
) {
1988 *khugepaged_alloc_page(struct page
**hpage
, struct mm_struct
*mm
,
1989 struct vm_area_struct
*vma
, unsigned long address
,
1994 * Allocate the page while the vma is still valid and under
1995 * the mmap_sem read mode so there is no memory allocation
1996 * later when we take the mmap_sem in write mode. This is more
1997 * friendly behavior (OTOH it may actually hide bugs) to
1998 * filesystems in userland with daemons allocating memory in
1999 * the userland I/O paths. Allocating memory with the
2000 * mmap_sem in read mode is good idea also to allow greater
2003 *hpage
= alloc_hugepage_vma(khugepaged_defrag(), vma
, address
,
2004 node
, __GFP_OTHER_NODE
);
2007 * After allocating the hugepage, release the mmap_sem read lock in
2008 * preparation for taking it in write mode.
2010 up_read(&mm
->mmap_sem
);
2011 if (unlikely(!*hpage
)) {
2012 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
2013 *hpage
= ERR_PTR(-ENOMEM
);
2017 count_vm_event(THP_COLLAPSE_ALLOC
);
2021 static struct page
*khugepaged_alloc_hugepage(bool *wait
)
2026 hpage
= alloc_hugepage(khugepaged_defrag());
2028 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
2033 khugepaged_alloc_sleep();
2035 count_vm_event(THP_COLLAPSE_ALLOC
);
2036 } while (unlikely(!hpage
) && likely(khugepaged_enabled()));
2041 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
2044 *hpage
= khugepaged_alloc_hugepage(wait
);
2046 if (unlikely(!*hpage
))
2053 *khugepaged_alloc_page(struct page
**hpage
, struct mm_struct
*mm
,
2054 struct vm_area_struct
*vma
, unsigned long address
,
2057 up_read(&mm
->mmap_sem
);
2063 static bool hugepage_vma_check(struct vm_area_struct
*vma
)
2065 if ((!(vma
->vm_flags
& VM_HUGEPAGE
) && !khugepaged_always()) ||
2066 (vma
->vm_flags
& VM_NOHUGEPAGE
))
2069 if (!vma
->anon_vma
|| vma
->vm_ops
)
2071 if (is_vma_temporary_stack(vma
))
2073 VM_BUG_ON(vma
->vm_flags
& VM_NO_THP
);
2077 static void collapse_huge_page(struct mm_struct
*mm
,
2078 unsigned long address
,
2079 struct page
**hpage
,
2080 struct vm_area_struct
*vma
,
2086 struct page
*new_page
;
2089 unsigned long hstart
, hend
;
2090 unsigned long mmun_start
; /* For mmu_notifiers */
2091 unsigned long mmun_end
; /* For mmu_notifiers */
2093 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2095 /* release the mmap_sem read lock. */
2096 new_page
= khugepaged_alloc_page(hpage
, mm
, vma
, address
, node
);
2100 if (unlikely(mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
)))
2104 * Prevent all access to pagetables with the exception of
2105 * gup_fast later hanlded by the ptep_clear_flush and the VM
2106 * handled by the anon_vma lock + PG_lock.
2108 down_write(&mm
->mmap_sem
);
2109 if (unlikely(khugepaged_test_exit(mm
)))
2112 vma
= find_vma(mm
, address
);
2113 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2114 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2115 if (address
< hstart
|| address
+ HPAGE_PMD_SIZE
> hend
)
2117 if (!hugepage_vma_check(vma
))
2119 pmd
= mm_find_pmd(mm
, address
);
2122 if (pmd_trans_huge(*pmd
))
2125 anon_vma_lock(vma
->anon_vma
);
2127 pte
= pte_offset_map(pmd
, address
);
2128 ptl
= pte_lockptr(mm
, pmd
);
2130 mmun_start
= address
;
2131 mmun_end
= address
+ HPAGE_PMD_SIZE
;
2132 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2133 spin_lock(&mm
->page_table_lock
); /* probably unnecessary */
2135 * After this gup_fast can't run anymore. This also removes
2136 * any huge TLB entry from the CPU so we won't allow
2137 * huge and small TLB entries for the same virtual address
2138 * to avoid the risk of CPU bugs in that area.
2140 _pmd
= pmdp_clear_flush(vma
, address
, pmd
);
2141 spin_unlock(&mm
->page_table_lock
);
2142 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2145 isolated
= __collapse_huge_page_isolate(vma
, address
, pte
);
2148 if (unlikely(!isolated
)) {
2150 spin_lock(&mm
->page_table_lock
);
2151 BUG_ON(!pmd_none(*pmd
));
2152 set_pmd_at(mm
, address
, pmd
, _pmd
);
2153 spin_unlock(&mm
->page_table_lock
);
2154 anon_vma_unlock(vma
->anon_vma
);
2159 * All pages are isolated and locked so anon_vma rmap
2160 * can't run anymore.
2162 anon_vma_unlock(vma
->anon_vma
);
2164 __collapse_huge_page_copy(pte
, new_page
, vma
, address
, ptl
);
2166 __SetPageUptodate(new_page
);
2167 pgtable
= pmd_pgtable(_pmd
);
2169 _pmd
= mk_huge_pmd(new_page
, vma
);
2172 * spin_lock() below is not the equivalent of smp_wmb(), so
2173 * this is needed to avoid the copy_huge_page writes to become
2174 * visible after the set_pmd_at() write.
2178 spin_lock(&mm
->page_table_lock
);
2179 BUG_ON(!pmd_none(*pmd
));
2180 page_add_new_anon_rmap(new_page
, vma
, address
);
2181 set_pmd_at(mm
, address
, pmd
, _pmd
);
2182 update_mmu_cache_pmd(vma
, address
, pmd
);
2183 pgtable_trans_huge_deposit(mm
, pgtable
);
2184 spin_unlock(&mm
->page_table_lock
);
2188 khugepaged_pages_collapsed
++;
2190 up_write(&mm
->mmap_sem
);
2194 mem_cgroup_uncharge_page(new_page
);
2198 static int khugepaged_scan_pmd(struct mm_struct
*mm
,
2199 struct vm_area_struct
*vma
,
2200 unsigned long address
,
2201 struct page
**hpage
)
2205 int ret
= 0, referenced
= 0, none
= 0;
2207 unsigned long _address
;
2211 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2213 pmd
= mm_find_pmd(mm
, address
);
2216 if (pmd_trans_huge(*pmd
))
2219 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2220 for (_address
= address
, _pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
2221 _pte
++, _address
+= PAGE_SIZE
) {
2222 pte_t pteval
= *_pte
;
2223 if (pte_none(pteval
)) {
2224 if (++none
<= khugepaged_max_ptes_none
)
2229 if (!pte_present(pteval
) || !pte_write(pteval
))
2231 page
= vm_normal_page(vma
, _address
, pteval
);
2232 if (unlikely(!page
))
2235 * Chose the node of the first page. This could
2236 * be more sophisticated and look at more pages,
2237 * but isn't for now.
2240 node
= page_to_nid(page
);
2241 VM_BUG_ON(PageCompound(page
));
2242 if (!PageLRU(page
) || PageLocked(page
) || !PageAnon(page
))
2244 /* cannot use mapcount: can't collapse if there's a gup pin */
2245 if (page_count(page
) != 1)
2247 if (pte_young(pteval
) || PageReferenced(page
) ||
2248 mmu_notifier_test_young(vma
->vm_mm
, address
))
2254 pte_unmap_unlock(pte
, ptl
);
2256 /* collapse_huge_page will return with the mmap_sem released */
2257 collapse_huge_page(mm
, address
, hpage
, vma
, node
);
2262 static void collect_mm_slot(struct mm_slot
*mm_slot
)
2264 struct mm_struct
*mm
= mm_slot
->mm
;
2266 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2268 if (khugepaged_test_exit(mm
)) {
2270 hlist_del(&mm_slot
->hash
);
2271 list_del(&mm_slot
->mm_node
);
2274 * Not strictly needed because the mm exited already.
2276 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2279 /* khugepaged_mm_lock actually not necessary for the below */
2280 free_mm_slot(mm_slot
);
2285 static unsigned int khugepaged_scan_mm_slot(unsigned int pages
,
2286 struct page
**hpage
)
2287 __releases(&khugepaged_mm_lock
)
2288 __acquires(&khugepaged_mm_lock
)
2290 struct mm_slot
*mm_slot
;
2291 struct mm_struct
*mm
;
2292 struct vm_area_struct
*vma
;
2296 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2298 if (khugepaged_scan
.mm_slot
)
2299 mm_slot
= khugepaged_scan
.mm_slot
;
2301 mm_slot
= list_entry(khugepaged_scan
.mm_head
.next
,
2302 struct mm_slot
, mm_node
);
2303 khugepaged_scan
.address
= 0;
2304 khugepaged_scan
.mm_slot
= mm_slot
;
2306 spin_unlock(&khugepaged_mm_lock
);
2309 down_read(&mm
->mmap_sem
);
2310 if (unlikely(khugepaged_test_exit(mm
)))
2313 vma
= find_vma(mm
, khugepaged_scan
.address
);
2316 for (; vma
; vma
= vma
->vm_next
) {
2317 unsigned long hstart
, hend
;
2320 if (unlikely(khugepaged_test_exit(mm
))) {
2324 if (!hugepage_vma_check(vma
)) {
2329 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2330 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2333 if (khugepaged_scan
.address
> hend
)
2335 if (khugepaged_scan
.address
< hstart
)
2336 khugepaged_scan
.address
= hstart
;
2337 VM_BUG_ON(khugepaged_scan
.address
& ~HPAGE_PMD_MASK
);
2339 while (khugepaged_scan
.address
< hend
) {
2342 if (unlikely(khugepaged_test_exit(mm
)))
2343 goto breakouterloop
;
2345 VM_BUG_ON(khugepaged_scan
.address
< hstart
||
2346 khugepaged_scan
.address
+ HPAGE_PMD_SIZE
>
2348 ret
= khugepaged_scan_pmd(mm
, vma
,
2349 khugepaged_scan
.address
,
2351 /* move to next address */
2352 khugepaged_scan
.address
+= HPAGE_PMD_SIZE
;
2353 progress
+= HPAGE_PMD_NR
;
2355 /* we released mmap_sem so break loop */
2356 goto breakouterloop_mmap_sem
;
2357 if (progress
>= pages
)
2358 goto breakouterloop
;
2362 up_read(&mm
->mmap_sem
); /* exit_mmap will destroy ptes after this */
2363 breakouterloop_mmap_sem
:
2365 spin_lock(&khugepaged_mm_lock
);
2366 VM_BUG_ON(khugepaged_scan
.mm_slot
!= mm_slot
);
2368 * Release the current mm_slot if this mm is about to die, or
2369 * if we scanned all vmas of this mm.
2371 if (khugepaged_test_exit(mm
) || !vma
) {
2373 * Make sure that if mm_users is reaching zero while
2374 * khugepaged runs here, khugepaged_exit will find
2375 * mm_slot not pointing to the exiting mm.
2377 if (mm_slot
->mm_node
.next
!= &khugepaged_scan
.mm_head
) {
2378 khugepaged_scan
.mm_slot
= list_entry(
2379 mm_slot
->mm_node
.next
,
2380 struct mm_slot
, mm_node
);
2381 khugepaged_scan
.address
= 0;
2383 khugepaged_scan
.mm_slot
= NULL
;
2384 khugepaged_full_scans
++;
2387 collect_mm_slot(mm_slot
);
2393 static int khugepaged_has_work(void)
2395 return !list_empty(&khugepaged_scan
.mm_head
) &&
2396 khugepaged_enabled();
2399 static int khugepaged_wait_event(void)
2401 return !list_empty(&khugepaged_scan
.mm_head
) ||
2402 kthread_should_stop();
2405 static void khugepaged_do_scan(void)
2407 struct page
*hpage
= NULL
;
2408 unsigned int progress
= 0, pass_through_head
= 0;
2409 unsigned int pages
= khugepaged_pages_to_scan
;
2412 barrier(); /* write khugepaged_pages_to_scan to local stack */
2414 while (progress
< pages
) {
2415 if (!khugepaged_prealloc_page(&hpage
, &wait
))
2420 if (unlikely(kthread_should_stop() || freezing(current
)))
2423 spin_lock(&khugepaged_mm_lock
);
2424 if (!khugepaged_scan
.mm_slot
)
2425 pass_through_head
++;
2426 if (khugepaged_has_work() &&
2427 pass_through_head
< 2)
2428 progress
+= khugepaged_scan_mm_slot(pages
- progress
,
2432 spin_unlock(&khugepaged_mm_lock
);
2435 if (!IS_ERR_OR_NULL(hpage
))
2439 static void khugepaged_wait_work(void)
2443 if (khugepaged_has_work()) {
2444 if (!khugepaged_scan_sleep_millisecs
)
2447 wait_event_freezable_timeout(khugepaged_wait
,
2448 kthread_should_stop(),
2449 msecs_to_jiffies(khugepaged_scan_sleep_millisecs
));
2453 if (khugepaged_enabled())
2454 wait_event_freezable(khugepaged_wait
, khugepaged_wait_event());
2457 static int khugepaged(void *none
)
2459 struct mm_slot
*mm_slot
;
2462 set_user_nice(current
, 19);
2464 while (!kthread_should_stop()) {
2465 khugepaged_do_scan();
2466 khugepaged_wait_work();
2469 spin_lock(&khugepaged_mm_lock
);
2470 mm_slot
= khugepaged_scan
.mm_slot
;
2471 khugepaged_scan
.mm_slot
= NULL
;
2473 collect_mm_slot(mm_slot
);
2474 spin_unlock(&khugepaged_mm_lock
);
2478 void __split_huge_page_pmd(struct vm_area_struct
*vma
, unsigned long address
,
2482 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
2483 struct mm_struct
*mm
= vma
->vm_mm
;
2485 BUG_ON(vma
->vm_start
> haddr
|| vma
->vm_end
< haddr
+ HPAGE_PMD_SIZE
);
2487 spin_lock(&mm
->page_table_lock
);
2488 if (unlikely(!pmd_trans_huge(*pmd
))) {
2489 spin_unlock(&mm
->page_table_lock
);
2492 page
= pmd_page(*pmd
);
2493 VM_BUG_ON(!page_count(page
));
2495 spin_unlock(&mm
->page_table_lock
);
2497 split_huge_page(page
);
2500 BUG_ON(pmd_trans_huge(*pmd
));
2503 void split_huge_page_pmd_mm(struct mm_struct
*mm
, unsigned long address
,
2506 struct vm_area_struct
*vma
;
2508 vma
= find_vma(mm
, address
);
2509 BUG_ON(vma
== NULL
);
2510 split_huge_page_pmd(vma
, address
, pmd
);
2513 static void split_huge_page_address(struct mm_struct
*mm
,
2514 unsigned long address
)
2518 VM_BUG_ON(!(address
& ~HPAGE_PMD_MASK
));
2520 pmd
= mm_find_pmd(mm
, address
);
2524 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2525 * materialize from under us.
2527 split_huge_page_pmd_mm(mm
, address
, pmd
);
2530 void __vma_adjust_trans_huge(struct vm_area_struct
*vma
,
2531 unsigned long start
,
2536 * If the new start address isn't hpage aligned and it could
2537 * previously contain an hugepage: check if we need to split
2540 if (start
& ~HPAGE_PMD_MASK
&&
2541 (start
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2542 (start
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2543 split_huge_page_address(vma
->vm_mm
, start
);
2546 * If the new end address isn't hpage aligned and it could
2547 * previously contain an hugepage: check if we need to split
2550 if (end
& ~HPAGE_PMD_MASK
&&
2551 (end
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2552 (end
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2553 split_huge_page_address(vma
->vm_mm
, end
);
2556 * If we're also updating the vma->vm_next->vm_start, if the new
2557 * vm_next->vm_start isn't page aligned and it could previously
2558 * contain an hugepage: check if we need to split an huge pmd.
2560 if (adjust_next
> 0) {
2561 struct vm_area_struct
*next
= vma
->vm_next
;
2562 unsigned long nstart
= next
->vm_start
;
2563 nstart
+= adjust_next
<< PAGE_SHIFT
;
2564 if (nstart
& ~HPAGE_PMD_MASK
&&
2565 (nstart
& HPAGE_PMD_MASK
) >= next
->vm_start
&&
2566 (nstart
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= next
->vm_end
)
2567 split_huge_page_address(next
->vm_mm
, nstart
);