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>
21 #include <asm/pgalloc.h>
25 * By default transparent hugepage support is enabled for all mappings
26 * and khugepaged scans all mappings. Defrag is only invoked by
27 * khugepaged hugepage allocations and by page faults inside
28 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
31 unsigned long transparent_hugepage_flags __read_mostly
=
32 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
33 (1<<TRANSPARENT_HUGEPAGE_FLAG
)|
35 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
36 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
)|
38 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
)|
39 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
41 /* default scan 8*512 pte (or vmas) every 30 second */
42 static unsigned int khugepaged_pages_to_scan __read_mostly
= HPAGE_PMD_NR
*8;
43 static unsigned int khugepaged_pages_collapsed
;
44 static unsigned int khugepaged_full_scans
;
45 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly
= 10000;
46 /* during fragmentation poll the hugepage allocator once every minute */
47 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly
= 60000;
48 static struct task_struct
*khugepaged_thread __read_mostly
;
49 static DEFINE_MUTEX(khugepaged_mutex
);
50 static DEFINE_SPINLOCK(khugepaged_mm_lock
);
51 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait
);
53 * default collapse hugepages if there is at least one pte mapped like
54 * it would have happened if the vma was large enough during page
57 static unsigned int khugepaged_max_ptes_none __read_mostly
= HPAGE_PMD_NR
-1;
59 static int khugepaged(void *none
);
60 static int mm_slots_hash_init(void);
61 static int khugepaged_slab_init(void);
62 static void khugepaged_slab_free(void);
64 #define MM_SLOTS_HASH_HEADS 1024
65 static struct hlist_head
*mm_slots_hash __read_mostly
;
66 static struct kmem_cache
*mm_slot_cache __read_mostly
;
69 * struct mm_slot - hash lookup from mm to mm_slot
70 * @hash: hash collision list
71 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
72 * @mm: the mm that this information is valid for
75 struct hlist_node hash
;
76 struct list_head mm_node
;
81 * struct khugepaged_scan - cursor for scanning
82 * @mm_head: the head of the mm list to scan
83 * @mm_slot: the current mm_slot we are scanning
84 * @address: the next address inside that to be scanned
86 * There is only the one khugepaged_scan instance of this cursor structure.
88 struct khugepaged_scan
{
89 struct list_head mm_head
;
90 struct mm_slot
*mm_slot
;
91 unsigned long address
;
93 static struct khugepaged_scan khugepaged_scan
= {
94 .mm_head
= LIST_HEAD_INIT(khugepaged_scan
.mm_head
),
98 static int set_recommended_min_free_kbytes(void)
102 unsigned long recommended_min
;
103 extern int min_free_kbytes
;
105 if (!khugepaged_enabled())
108 for_each_populated_zone(zone
)
111 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
112 recommended_min
= pageblock_nr_pages
* nr_zones
* 2;
115 * Make sure that on average at least two pageblocks are almost free
116 * of another type, one for a migratetype to fall back to and a
117 * second to avoid subsequent fallbacks of other types There are 3
118 * MIGRATE_TYPES we care about.
120 recommended_min
+= pageblock_nr_pages
* nr_zones
*
121 MIGRATE_PCPTYPES
* MIGRATE_PCPTYPES
;
123 /* don't ever allow to reserve more than 5% of the lowmem */
124 recommended_min
= min(recommended_min
,
125 (unsigned long) nr_free_buffer_pages() / 20);
126 recommended_min
<<= (PAGE_SHIFT
-10);
128 if (recommended_min
> min_free_kbytes
)
129 min_free_kbytes
= recommended_min
;
130 setup_per_zone_wmarks();
133 late_initcall(set_recommended_min_free_kbytes
);
135 static int start_khugepaged(void)
138 if (khugepaged_enabled()) {
139 if (!khugepaged_thread
)
140 khugepaged_thread
= kthread_run(khugepaged
, NULL
,
142 if (unlikely(IS_ERR(khugepaged_thread
))) {
144 "khugepaged: kthread_run(khugepaged) failed\n");
145 err
= PTR_ERR(khugepaged_thread
);
146 khugepaged_thread
= NULL
;
149 if (!list_empty(&khugepaged_scan
.mm_head
))
150 wake_up_interruptible(&khugepaged_wait
);
152 set_recommended_min_free_kbytes();
153 } else if (khugepaged_thread
) {
154 kthread_stop(khugepaged_thread
);
155 khugepaged_thread
= NULL
;
163 static ssize_t
double_flag_show(struct kobject
*kobj
,
164 struct kobj_attribute
*attr
, char *buf
,
165 enum transparent_hugepage_flag enabled
,
166 enum transparent_hugepage_flag req_madv
)
168 if (test_bit(enabled
, &transparent_hugepage_flags
)) {
169 VM_BUG_ON(test_bit(req_madv
, &transparent_hugepage_flags
));
170 return sprintf(buf
, "[always] madvise never\n");
171 } else if (test_bit(req_madv
, &transparent_hugepage_flags
))
172 return sprintf(buf
, "always [madvise] never\n");
174 return sprintf(buf
, "always madvise [never]\n");
176 static ssize_t
double_flag_store(struct kobject
*kobj
,
177 struct kobj_attribute
*attr
,
178 const char *buf
, size_t count
,
179 enum transparent_hugepage_flag enabled
,
180 enum transparent_hugepage_flag req_madv
)
182 if (!memcmp("always", buf
,
183 min(sizeof("always")-1, count
))) {
184 set_bit(enabled
, &transparent_hugepage_flags
);
185 clear_bit(req_madv
, &transparent_hugepage_flags
);
186 } else if (!memcmp("madvise", buf
,
187 min(sizeof("madvise")-1, count
))) {
188 clear_bit(enabled
, &transparent_hugepage_flags
);
189 set_bit(req_madv
, &transparent_hugepage_flags
);
190 } else if (!memcmp("never", buf
,
191 min(sizeof("never")-1, count
))) {
192 clear_bit(enabled
, &transparent_hugepage_flags
);
193 clear_bit(req_madv
, &transparent_hugepage_flags
);
200 static ssize_t
enabled_show(struct kobject
*kobj
,
201 struct kobj_attribute
*attr
, char *buf
)
203 return double_flag_show(kobj
, attr
, buf
,
204 TRANSPARENT_HUGEPAGE_FLAG
,
205 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
207 static ssize_t
enabled_store(struct kobject
*kobj
,
208 struct kobj_attribute
*attr
,
209 const char *buf
, size_t count
)
213 ret
= double_flag_store(kobj
, attr
, buf
, count
,
214 TRANSPARENT_HUGEPAGE_FLAG
,
215 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
220 mutex_lock(&khugepaged_mutex
);
221 err
= start_khugepaged();
222 mutex_unlock(&khugepaged_mutex
);
230 static struct kobj_attribute enabled_attr
=
231 __ATTR(enabled
, 0644, enabled_show
, enabled_store
);
233 static ssize_t
single_flag_show(struct kobject
*kobj
,
234 struct kobj_attribute
*attr
, char *buf
,
235 enum transparent_hugepage_flag flag
)
237 return sprintf(buf
, "%d\n",
238 !!test_bit(flag
, &transparent_hugepage_flags
));
241 static ssize_t
single_flag_store(struct kobject
*kobj
,
242 struct kobj_attribute
*attr
,
243 const char *buf
, size_t count
,
244 enum transparent_hugepage_flag flag
)
249 ret
= kstrtoul(buf
, 10, &value
);
256 set_bit(flag
, &transparent_hugepage_flags
);
258 clear_bit(flag
, &transparent_hugepage_flags
);
264 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
265 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
266 * memory just to allocate one more hugepage.
268 static ssize_t
defrag_show(struct kobject
*kobj
,
269 struct kobj_attribute
*attr
, char *buf
)
271 return double_flag_show(kobj
, attr
, buf
,
272 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
,
273 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
);
275 static ssize_t
defrag_store(struct kobject
*kobj
,
276 struct kobj_attribute
*attr
,
277 const char *buf
, size_t count
)
279 return double_flag_store(kobj
, attr
, buf
, count
,
280 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
,
281 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
);
283 static struct kobj_attribute defrag_attr
=
284 __ATTR(defrag
, 0644, defrag_show
, defrag_store
);
286 #ifdef CONFIG_DEBUG_VM
287 static ssize_t
debug_cow_show(struct kobject
*kobj
,
288 struct kobj_attribute
*attr
, char *buf
)
290 return single_flag_show(kobj
, attr
, buf
,
291 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
293 static ssize_t
debug_cow_store(struct kobject
*kobj
,
294 struct kobj_attribute
*attr
,
295 const char *buf
, size_t count
)
297 return single_flag_store(kobj
, attr
, buf
, count
,
298 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
300 static struct kobj_attribute debug_cow_attr
=
301 __ATTR(debug_cow
, 0644, debug_cow_show
, debug_cow_store
);
302 #endif /* CONFIG_DEBUG_VM */
304 static struct attribute
*hugepage_attr
[] = {
307 #ifdef CONFIG_DEBUG_VM
308 &debug_cow_attr
.attr
,
313 static struct attribute_group hugepage_attr_group
= {
314 .attrs
= hugepage_attr
,
317 static ssize_t
scan_sleep_millisecs_show(struct kobject
*kobj
,
318 struct kobj_attribute
*attr
,
321 return sprintf(buf
, "%u\n", khugepaged_scan_sleep_millisecs
);
324 static ssize_t
scan_sleep_millisecs_store(struct kobject
*kobj
,
325 struct kobj_attribute
*attr
,
326 const char *buf
, size_t count
)
331 err
= strict_strtoul(buf
, 10, &msecs
);
332 if (err
|| msecs
> UINT_MAX
)
335 khugepaged_scan_sleep_millisecs
= msecs
;
336 wake_up_interruptible(&khugepaged_wait
);
340 static struct kobj_attribute scan_sleep_millisecs_attr
=
341 __ATTR(scan_sleep_millisecs
, 0644, scan_sleep_millisecs_show
,
342 scan_sleep_millisecs_store
);
344 static ssize_t
alloc_sleep_millisecs_show(struct kobject
*kobj
,
345 struct kobj_attribute
*attr
,
348 return sprintf(buf
, "%u\n", khugepaged_alloc_sleep_millisecs
);
351 static ssize_t
alloc_sleep_millisecs_store(struct kobject
*kobj
,
352 struct kobj_attribute
*attr
,
353 const char *buf
, size_t count
)
358 err
= strict_strtoul(buf
, 10, &msecs
);
359 if (err
|| msecs
> UINT_MAX
)
362 khugepaged_alloc_sleep_millisecs
= msecs
;
363 wake_up_interruptible(&khugepaged_wait
);
367 static struct kobj_attribute alloc_sleep_millisecs_attr
=
368 __ATTR(alloc_sleep_millisecs
, 0644, alloc_sleep_millisecs_show
,
369 alloc_sleep_millisecs_store
);
371 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
372 struct kobj_attribute
*attr
,
375 return sprintf(buf
, "%u\n", khugepaged_pages_to_scan
);
377 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
378 struct kobj_attribute
*attr
,
379 const char *buf
, size_t count
)
384 err
= strict_strtoul(buf
, 10, &pages
);
385 if (err
|| !pages
|| pages
> UINT_MAX
)
388 khugepaged_pages_to_scan
= pages
;
392 static struct kobj_attribute pages_to_scan_attr
=
393 __ATTR(pages_to_scan
, 0644, pages_to_scan_show
,
394 pages_to_scan_store
);
396 static ssize_t
pages_collapsed_show(struct kobject
*kobj
,
397 struct kobj_attribute
*attr
,
400 return sprintf(buf
, "%u\n", khugepaged_pages_collapsed
);
402 static struct kobj_attribute pages_collapsed_attr
=
403 __ATTR_RO(pages_collapsed
);
405 static ssize_t
full_scans_show(struct kobject
*kobj
,
406 struct kobj_attribute
*attr
,
409 return sprintf(buf
, "%u\n", khugepaged_full_scans
);
411 static struct kobj_attribute full_scans_attr
=
412 __ATTR_RO(full_scans
);
414 static ssize_t
khugepaged_defrag_show(struct kobject
*kobj
,
415 struct kobj_attribute
*attr
, char *buf
)
417 return single_flag_show(kobj
, attr
, buf
,
418 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
420 static ssize_t
khugepaged_defrag_store(struct kobject
*kobj
,
421 struct kobj_attribute
*attr
,
422 const char *buf
, size_t count
)
424 return single_flag_store(kobj
, attr
, buf
, count
,
425 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
427 static struct kobj_attribute khugepaged_defrag_attr
=
428 __ATTR(defrag
, 0644, khugepaged_defrag_show
,
429 khugepaged_defrag_store
);
432 * max_ptes_none controls if khugepaged should collapse hugepages over
433 * any unmapped ptes in turn potentially increasing the memory
434 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
435 * reduce the available free memory in the system as it
436 * runs. Increasing max_ptes_none will instead potentially reduce the
437 * free memory in the system during the khugepaged scan.
439 static ssize_t
khugepaged_max_ptes_none_show(struct kobject
*kobj
,
440 struct kobj_attribute
*attr
,
443 return sprintf(buf
, "%u\n", khugepaged_max_ptes_none
);
445 static ssize_t
khugepaged_max_ptes_none_store(struct kobject
*kobj
,
446 struct kobj_attribute
*attr
,
447 const char *buf
, size_t count
)
450 unsigned long max_ptes_none
;
452 err
= strict_strtoul(buf
, 10, &max_ptes_none
);
453 if (err
|| max_ptes_none
> HPAGE_PMD_NR
-1)
456 khugepaged_max_ptes_none
= max_ptes_none
;
460 static struct kobj_attribute khugepaged_max_ptes_none_attr
=
461 __ATTR(max_ptes_none
, 0644, khugepaged_max_ptes_none_show
,
462 khugepaged_max_ptes_none_store
);
464 static struct attribute
*khugepaged_attr
[] = {
465 &khugepaged_defrag_attr
.attr
,
466 &khugepaged_max_ptes_none_attr
.attr
,
467 &pages_to_scan_attr
.attr
,
468 &pages_collapsed_attr
.attr
,
469 &full_scans_attr
.attr
,
470 &scan_sleep_millisecs_attr
.attr
,
471 &alloc_sleep_millisecs_attr
.attr
,
475 static struct attribute_group khugepaged_attr_group
= {
476 .attrs
= khugepaged_attr
,
477 .name
= "khugepaged",
480 static int __init
hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
484 *hugepage_kobj
= kobject_create_and_add("transparent_hugepage", mm_kobj
);
485 if (unlikely(!*hugepage_kobj
)) {
486 printk(KERN_ERR
"hugepage: failed kobject create\n");
490 err
= sysfs_create_group(*hugepage_kobj
, &hugepage_attr_group
);
492 printk(KERN_ERR
"hugepage: failed register hugeage group\n");
496 err
= sysfs_create_group(*hugepage_kobj
, &khugepaged_attr_group
);
498 printk(KERN_ERR
"hugepage: failed register hugeage group\n");
499 goto remove_hp_group
;
505 sysfs_remove_group(*hugepage_kobj
, &hugepage_attr_group
);
507 kobject_put(*hugepage_kobj
);
511 static void __init
hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
513 sysfs_remove_group(hugepage_kobj
, &khugepaged_attr_group
);
514 sysfs_remove_group(hugepage_kobj
, &hugepage_attr_group
);
515 kobject_put(hugepage_kobj
);
518 static inline int hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
523 static inline void hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
526 #endif /* CONFIG_SYSFS */
528 static int __init
hugepage_init(void)
531 struct kobject
*hugepage_kobj
;
533 if (!has_transparent_hugepage()) {
534 transparent_hugepage_flags
= 0;
538 err
= hugepage_init_sysfs(&hugepage_kobj
);
542 err
= khugepaged_slab_init();
546 err
= mm_slots_hash_init();
548 khugepaged_slab_free();
553 * By default disable transparent hugepages on smaller systems,
554 * where the extra memory used could hurt more than TLB overhead
555 * is likely to save. The admin can still enable it through /sys.
557 if (totalram_pages
< (512 << (20 - PAGE_SHIFT
)))
558 transparent_hugepage_flags
= 0;
564 hugepage_exit_sysfs(hugepage_kobj
);
567 module_init(hugepage_init
)
569 static int __init
setup_transparent_hugepage(char *str
)
574 if (!strcmp(str
, "always")) {
575 set_bit(TRANSPARENT_HUGEPAGE_FLAG
,
576 &transparent_hugepage_flags
);
577 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
578 &transparent_hugepage_flags
);
580 } else if (!strcmp(str
, "madvise")) {
581 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
582 &transparent_hugepage_flags
);
583 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
584 &transparent_hugepage_flags
);
586 } else if (!strcmp(str
, "never")) {
587 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
588 &transparent_hugepage_flags
);
589 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
590 &transparent_hugepage_flags
);
596 "transparent_hugepage= cannot parse, ignored\n");
599 __setup("transparent_hugepage=", setup_transparent_hugepage
);
601 static inline pmd_t
maybe_pmd_mkwrite(pmd_t pmd
, struct vm_area_struct
*vma
)
603 if (likely(vma
->vm_flags
& VM_WRITE
))
604 pmd
= pmd_mkwrite(pmd
);
608 static int __do_huge_pmd_anonymous_page(struct mm_struct
*mm
,
609 struct vm_area_struct
*vma
,
610 unsigned long haddr
, pmd_t
*pmd
,
615 VM_BUG_ON(!PageCompound(page
));
616 pgtable
= pte_alloc_one(mm
, haddr
);
617 if (unlikely(!pgtable
))
620 clear_huge_page(page
, haddr
, HPAGE_PMD_NR
);
621 __SetPageUptodate(page
);
623 spin_lock(&mm
->page_table_lock
);
624 if (unlikely(!pmd_none(*pmd
))) {
625 spin_unlock(&mm
->page_table_lock
);
626 mem_cgroup_uncharge_page(page
);
628 pte_free(mm
, pgtable
);
631 entry
= mk_pmd(page
, vma
->vm_page_prot
);
632 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
633 entry
= pmd_mkhuge(entry
);
635 * The spinlocking to take the lru_lock inside
636 * page_add_new_anon_rmap() acts as a full memory
637 * barrier to be sure clear_huge_page writes become
638 * visible after the set_pmd_at() write.
640 page_add_new_anon_rmap(page
, vma
, haddr
);
641 set_pmd_at(mm
, haddr
, pmd
, entry
);
642 pgtable_trans_huge_deposit(mm
, pgtable
);
643 add_mm_counter(mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
645 spin_unlock(&mm
->page_table_lock
);
651 static inline gfp_t
alloc_hugepage_gfpmask(int defrag
, gfp_t extra_gfp
)
653 return (GFP_TRANSHUGE
& ~(defrag
? 0 : __GFP_WAIT
)) | extra_gfp
;
656 static inline struct page
*alloc_hugepage_vma(int defrag
,
657 struct vm_area_struct
*vma
,
658 unsigned long haddr
, int nd
,
661 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag
, extra_gfp
),
662 HPAGE_PMD_ORDER
, vma
, haddr
, nd
);
666 static inline struct page
*alloc_hugepage(int defrag
)
668 return alloc_pages(alloc_hugepage_gfpmask(defrag
, 0),
673 int do_huge_pmd_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
674 unsigned long address
, pmd_t
*pmd
,
678 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
681 if (haddr
>= vma
->vm_start
&& haddr
+ HPAGE_PMD_SIZE
<= vma
->vm_end
) {
682 if (unlikely(anon_vma_prepare(vma
)))
684 if (unlikely(khugepaged_enter(vma
)))
686 page
= alloc_hugepage_vma(transparent_hugepage_defrag(vma
),
687 vma
, haddr
, numa_node_id(), 0);
688 if (unlikely(!page
)) {
689 count_vm_event(THP_FAULT_FALLBACK
);
692 count_vm_event(THP_FAULT_ALLOC
);
693 if (unlikely(mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))) {
697 if (unlikely(__do_huge_pmd_anonymous_page(mm
, vma
, haddr
, pmd
,
699 mem_cgroup_uncharge_page(page
);
708 * Use __pte_alloc instead of pte_alloc_map, because we can't
709 * run pte_offset_map on the pmd, if an huge pmd could
710 * materialize from under us from a different thread.
712 if (unlikely(__pte_alloc(mm
, vma
, pmd
, address
)))
714 /* if an huge pmd materialized from under us just retry later */
715 if (unlikely(pmd_trans_huge(*pmd
)))
718 * A regular pmd is established and it can't morph into a huge pmd
719 * from under us anymore at this point because we hold the mmap_sem
720 * read mode and khugepaged takes it in write mode. So now it's
721 * safe to run pte_offset_map().
723 pte
= pte_offset_map(pmd
, address
);
724 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
727 int copy_huge_pmd(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
728 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
729 struct vm_area_struct
*vma
)
731 struct page
*src_page
;
737 pgtable
= pte_alloc_one(dst_mm
, addr
);
738 if (unlikely(!pgtable
))
741 spin_lock(&dst_mm
->page_table_lock
);
742 spin_lock_nested(&src_mm
->page_table_lock
, SINGLE_DEPTH_NESTING
);
746 if (unlikely(!pmd_trans_huge(pmd
))) {
747 pte_free(dst_mm
, pgtable
);
750 if (unlikely(pmd_trans_splitting(pmd
))) {
751 /* split huge page running from under us */
752 spin_unlock(&src_mm
->page_table_lock
);
753 spin_unlock(&dst_mm
->page_table_lock
);
754 pte_free(dst_mm
, pgtable
);
756 wait_split_huge_page(vma
->anon_vma
, src_pmd
); /* src_vma */
759 src_page
= pmd_page(pmd
);
760 VM_BUG_ON(!PageHead(src_page
));
762 page_dup_rmap(src_page
);
763 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
765 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
766 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
767 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
768 pgtable_trans_huge_deposit(dst_mm
, pgtable
);
773 spin_unlock(&src_mm
->page_table_lock
);
774 spin_unlock(&dst_mm
->page_table_lock
);
779 static int do_huge_pmd_wp_page_fallback(struct mm_struct
*mm
,
780 struct vm_area_struct
*vma
,
781 unsigned long address
,
782 pmd_t
*pmd
, pmd_t orig_pmd
,
791 pages
= kmalloc(sizeof(struct page
*) * HPAGE_PMD_NR
,
793 if (unlikely(!pages
)) {
798 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
799 pages
[i
] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE
|
801 vma
, address
, page_to_nid(page
));
802 if (unlikely(!pages
[i
] ||
803 mem_cgroup_newpage_charge(pages
[i
], mm
,
807 mem_cgroup_uncharge_start();
809 mem_cgroup_uncharge_page(pages
[i
]);
812 mem_cgroup_uncharge_end();
819 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
820 copy_user_highpage(pages
[i
], page
+ i
,
821 haddr
+ PAGE_SIZE
* i
, vma
);
822 __SetPageUptodate(pages
[i
]);
826 spin_lock(&mm
->page_table_lock
);
827 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
829 VM_BUG_ON(!PageHead(page
));
831 pmdp_clear_flush_notify(vma
, haddr
, pmd
);
832 /* leave pmd empty until pte is filled */
834 pgtable
= pgtable_trans_huge_withdraw(mm
);
835 pmd_populate(mm
, &_pmd
, pgtable
);
837 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
839 entry
= mk_pte(pages
[i
], vma
->vm_page_prot
);
840 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
841 page_add_new_anon_rmap(pages
[i
], vma
, haddr
);
842 pte
= pte_offset_map(&_pmd
, haddr
);
843 VM_BUG_ON(!pte_none(*pte
));
844 set_pte_at(mm
, haddr
, pte
, entry
);
849 smp_wmb(); /* make pte visible before pmd */
850 pmd_populate(mm
, pmd
, pgtable
);
851 page_remove_rmap(page
);
852 spin_unlock(&mm
->page_table_lock
);
854 ret
|= VM_FAULT_WRITE
;
861 spin_unlock(&mm
->page_table_lock
);
862 mem_cgroup_uncharge_start();
863 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
864 mem_cgroup_uncharge_page(pages
[i
]);
867 mem_cgroup_uncharge_end();
872 int do_huge_pmd_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
873 unsigned long address
, pmd_t
*pmd
, pmd_t orig_pmd
)
876 struct page
*page
, *new_page
;
879 VM_BUG_ON(!vma
->anon_vma
);
880 spin_lock(&mm
->page_table_lock
);
881 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
884 page
= pmd_page(orig_pmd
);
885 VM_BUG_ON(!PageCompound(page
) || !PageHead(page
));
886 haddr
= address
& HPAGE_PMD_MASK
;
887 if (page_mapcount(page
) == 1) {
889 entry
= pmd_mkyoung(orig_pmd
);
890 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
891 if (pmdp_set_access_flags(vma
, haddr
, pmd
, entry
, 1))
892 update_mmu_cache(vma
, address
, entry
);
893 ret
|= VM_FAULT_WRITE
;
897 spin_unlock(&mm
->page_table_lock
);
899 if (transparent_hugepage_enabled(vma
) &&
900 !transparent_hugepage_debug_cow())
901 new_page
= alloc_hugepage_vma(transparent_hugepage_defrag(vma
),
902 vma
, haddr
, numa_node_id(), 0);
906 if (unlikely(!new_page
)) {
907 count_vm_event(THP_FAULT_FALLBACK
);
908 ret
= do_huge_pmd_wp_page_fallback(mm
, vma
, address
,
909 pmd
, orig_pmd
, page
, haddr
);
910 if (ret
& VM_FAULT_OOM
)
911 split_huge_page(page
);
915 count_vm_event(THP_FAULT_ALLOC
);
917 if (unlikely(mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))) {
919 split_huge_page(page
);
925 copy_user_huge_page(new_page
, page
, haddr
, vma
, HPAGE_PMD_NR
);
926 __SetPageUptodate(new_page
);
928 spin_lock(&mm
->page_table_lock
);
930 if (unlikely(!pmd_same(*pmd
, orig_pmd
))) {
931 spin_unlock(&mm
->page_table_lock
);
932 mem_cgroup_uncharge_page(new_page
);
937 VM_BUG_ON(!PageHead(page
));
938 entry
= mk_pmd(new_page
, vma
->vm_page_prot
);
939 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
940 entry
= pmd_mkhuge(entry
);
941 pmdp_clear_flush_notify(vma
, haddr
, pmd
);
942 page_add_new_anon_rmap(new_page
, vma
, haddr
);
943 set_pmd_at(mm
, haddr
, pmd
, entry
);
944 update_mmu_cache(vma
, address
, entry
);
945 page_remove_rmap(page
);
947 ret
|= VM_FAULT_WRITE
;
950 spin_unlock(&mm
->page_table_lock
);
955 struct page
*follow_trans_huge_pmd(struct mm_struct
*mm
,
960 struct page
*page
= NULL
;
962 assert_spin_locked(&mm
->page_table_lock
);
964 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
967 page
= pmd_page(*pmd
);
968 VM_BUG_ON(!PageHead(page
));
969 if (flags
& FOLL_TOUCH
) {
972 * We should set the dirty bit only for FOLL_WRITE but
973 * for now the dirty bit in the pmd is meaningless.
974 * And if the dirty bit will become meaningful and
975 * we'll only set it with FOLL_WRITE, an atomic
976 * set_bit will be required on the pmd to set the
977 * young bit, instead of the current set_pmd_at.
979 _pmd
= pmd_mkyoung(pmd_mkdirty(*pmd
));
980 set_pmd_at(mm
, addr
& HPAGE_PMD_MASK
, pmd
, _pmd
);
982 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
983 VM_BUG_ON(!PageCompound(page
));
984 if (flags
& FOLL_GET
)
991 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
992 pmd_t
*pmd
, unsigned long addr
)
996 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
999 pgtable
= pgtable_trans_huge_withdraw(tlb
->mm
);
1000 page
= pmd_page(*pmd
);
1002 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1003 page_remove_rmap(page
);
1004 VM_BUG_ON(page_mapcount(page
) < 0);
1005 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1006 VM_BUG_ON(!PageHead(page
));
1008 spin_unlock(&tlb
->mm
->page_table_lock
);
1009 tlb_remove_page(tlb
, page
);
1010 pte_free(tlb
->mm
, pgtable
);
1016 int mincore_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1017 unsigned long addr
, unsigned long end
,
1022 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
1024 * All logical pages in the range are present
1025 * if backed by a huge page.
1027 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1028 memset(vec
, 1, (end
- addr
) >> PAGE_SHIFT
);
1035 int move_huge_pmd(struct vm_area_struct
*vma
, struct vm_area_struct
*new_vma
,
1036 unsigned long old_addr
,
1037 unsigned long new_addr
, unsigned long old_end
,
1038 pmd_t
*old_pmd
, pmd_t
*new_pmd
)
1043 struct mm_struct
*mm
= vma
->vm_mm
;
1045 if ((old_addr
& ~HPAGE_PMD_MASK
) ||
1046 (new_addr
& ~HPAGE_PMD_MASK
) ||
1047 old_end
- old_addr
< HPAGE_PMD_SIZE
||
1048 (new_vma
->vm_flags
& VM_NOHUGEPAGE
))
1052 * The destination pmd shouldn't be established, free_pgtables()
1053 * should have release it.
1055 if (WARN_ON(!pmd_none(*new_pmd
))) {
1056 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1060 ret
= __pmd_trans_huge_lock(old_pmd
, vma
);
1062 pmd
= pmdp_get_and_clear(mm
, old_addr
, old_pmd
);
1063 VM_BUG_ON(!pmd_none(*new_pmd
));
1064 set_pmd_at(mm
, new_addr
, new_pmd
, pmd
);
1065 spin_unlock(&mm
->page_table_lock
);
1071 int change_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1072 unsigned long addr
, pgprot_t newprot
)
1074 struct mm_struct
*mm
= vma
->vm_mm
;
1077 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
1079 entry
= pmdp_get_and_clear(mm
, addr
, pmd
);
1080 entry
= pmd_modify(entry
, newprot
);
1081 set_pmd_at(mm
, addr
, pmd
, entry
);
1082 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1090 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1091 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1093 * Note that if it returns 1, this routine returns without unlocking page
1094 * table locks. So callers must unlock them.
1096 int __pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
)
1098 spin_lock(&vma
->vm_mm
->page_table_lock
);
1099 if (likely(pmd_trans_huge(*pmd
))) {
1100 if (unlikely(pmd_trans_splitting(*pmd
))) {
1101 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1102 wait_split_huge_page(vma
->anon_vma
, pmd
);
1105 /* Thp mapped by 'pmd' is stable, so we can
1106 * handle it as it is. */
1110 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1114 pmd_t
*page_check_address_pmd(struct page
*page
,
1115 struct mm_struct
*mm
,
1116 unsigned long address
,
1117 enum page_check_address_pmd_flag flag
)
1121 pmd_t
*pmd
, *ret
= NULL
;
1123 if (address
& ~HPAGE_PMD_MASK
)
1126 pgd
= pgd_offset(mm
, address
);
1127 if (!pgd_present(*pgd
))
1130 pud
= pud_offset(pgd
, address
);
1131 if (!pud_present(*pud
))
1134 pmd
= pmd_offset(pud
, address
);
1137 if (pmd_page(*pmd
) != page
)
1140 * split_vma() may create temporary aliased mappings. There is
1141 * no risk as long as all huge pmd are found and have their
1142 * splitting bit set before __split_huge_page_refcount
1143 * runs. Finding the same huge pmd more than once during the
1144 * same rmap walk is not a problem.
1146 if (flag
== PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG
&&
1147 pmd_trans_splitting(*pmd
))
1149 if (pmd_trans_huge(*pmd
)) {
1150 VM_BUG_ON(flag
== PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG
&&
1151 !pmd_trans_splitting(*pmd
));
1158 static int __split_huge_page_splitting(struct page
*page
,
1159 struct vm_area_struct
*vma
,
1160 unsigned long address
)
1162 struct mm_struct
*mm
= vma
->vm_mm
;
1166 spin_lock(&mm
->page_table_lock
);
1167 pmd
= page_check_address_pmd(page
, mm
, address
,
1168 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG
);
1171 * We can't temporarily set the pmd to null in order
1172 * to split it, the pmd must remain marked huge at all
1173 * times or the VM won't take the pmd_trans_huge paths
1174 * and it won't wait on the anon_vma->root->mutex to
1175 * serialize against split_huge_page*.
1177 pmdp_splitting_flush_notify(vma
, address
, pmd
);
1180 spin_unlock(&mm
->page_table_lock
);
1185 static void __split_huge_page_refcount(struct page
*page
)
1188 struct zone
*zone
= page_zone(page
);
1189 struct lruvec
*lruvec
;
1192 /* prevent PageLRU to go away from under us, and freeze lru stats */
1193 spin_lock_irq(&zone
->lru_lock
);
1194 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1196 compound_lock(page
);
1197 /* complete memcg works before add pages to LRU */
1198 mem_cgroup_split_huge_fixup(page
);
1200 for (i
= HPAGE_PMD_NR
- 1; i
>= 1; i
--) {
1201 struct page
*page_tail
= page
+ i
;
1203 /* tail_page->_mapcount cannot change */
1204 BUG_ON(page_mapcount(page_tail
) < 0);
1205 tail_count
+= page_mapcount(page_tail
);
1206 /* check for overflow */
1207 BUG_ON(tail_count
< 0);
1208 BUG_ON(atomic_read(&page_tail
->_count
) != 0);
1210 * tail_page->_count is zero and not changing from
1211 * under us. But get_page_unless_zero() may be running
1212 * from under us on the tail_page. If we used
1213 * atomic_set() below instead of atomic_add(), we
1214 * would then run atomic_set() concurrently with
1215 * get_page_unless_zero(), and atomic_set() is
1216 * implemented in C not using locked ops. spin_unlock
1217 * on x86 sometime uses locked ops because of PPro
1218 * errata 66, 92, so unless somebody can guarantee
1219 * atomic_set() here would be safe on all archs (and
1220 * not only on x86), it's safer to use atomic_add().
1222 atomic_add(page_mapcount(page
) + page_mapcount(page_tail
) + 1,
1223 &page_tail
->_count
);
1225 /* after clearing PageTail the gup refcount can be released */
1229 * retain hwpoison flag of the poisoned tail page:
1230 * fix for the unsuitable process killed on Guest Machine(KVM)
1231 * by the memory-failure.
1233 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
| __PG_HWPOISON
;
1234 page_tail
->flags
|= (page
->flags
&
1235 ((1L << PG_referenced
) |
1236 (1L << PG_swapbacked
) |
1237 (1L << PG_mlocked
) |
1238 (1L << PG_uptodate
)));
1239 page_tail
->flags
|= (1L << PG_dirty
);
1241 /* clear PageTail before overwriting first_page */
1245 * __split_huge_page_splitting() already set the
1246 * splitting bit in all pmd that could map this
1247 * hugepage, that will ensure no CPU can alter the
1248 * mapcount on the head page. The mapcount is only
1249 * accounted in the head page and it has to be
1250 * transferred to all tail pages in the below code. So
1251 * for this code to be safe, the split the mapcount
1252 * can't change. But that doesn't mean userland can't
1253 * keep changing and reading the page contents while
1254 * we transfer the mapcount, so the pmd splitting
1255 * status is achieved setting a reserved bit in the
1256 * pmd, not by clearing the present bit.
1258 page_tail
->_mapcount
= page
->_mapcount
;
1260 BUG_ON(page_tail
->mapping
);
1261 page_tail
->mapping
= page
->mapping
;
1263 page_tail
->index
= page
->index
+ i
;
1265 BUG_ON(!PageAnon(page_tail
));
1266 BUG_ON(!PageUptodate(page_tail
));
1267 BUG_ON(!PageDirty(page_tail
));
1268 BUG_ON(!PageSwapBacked(page_tail
));
1270 lru_add_page_tail(page
, page_tail
, lruvec
);
1272 atomic_sub(tail_count
, &page
->_count
);
1273 BUG_ON(atomic_read(&page
->_count
) <= 0);
1275 __mod_zone_page_state(zone
, NR_ANON_TRANSPARENT_HUGEPAGES
, -1);
1276 __mod_zone_page_state(zone
, NR_ANON_PAGES
, HPAGE_PMD_NR
);
1278 ClearPageCompound(page
);
1279 compound_unlock(page
);
1280 spin_unlock_irq(&zone
->lru_lock
);
1282 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
1283 struct page
*page_tail
= page
+ i
;
1284 BUG_ON(page_count(page_tail
) <= 0);
1286 * Tail pages may be freed if there wasn't any mapping
1287 * like if add_to_swap() is running on a lru page that
1288 * had its mapping zapped. And freeing these pages
1289 * requires taking the lru_lock so we do the put_page
1290 * of the tail pages after the split is complete.
1292 put_page(page_tail
);
1296 * Only the head page (now become a regular page) is required
1297 * to be pinned by the caller.
1299 BUG_ON(page_count(page
) <= 0);
1302 static int __split_huge_page_map(struct page
*page
,
1303 struct vm_area_struct
*vma
,
1304 unsigned long address
)
1306 struct mm_struct
*mm
= vma
->vm_mm
;
1310 unsigned long haddr
;
1312 spin_lock(&mm
->page_table_lock
);
1313 pmd
= page_check_address_pmd(page
, mm
, address
,
1314 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG
);
1316 pgtable
= pgtable_trans_huge_withdraw(mm
);
1317 pmd_populate(mm
, &_pmd
, pgtable
);
1320 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1322 BUG_ON(PageCompound(page
+i
));
1323 entry
= mk_pte(page
+ i
, vma
->vm_page_prot
);
1324 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1325 if (!pmd_write(*pmd
))
1326 entry
= pte_wrprotect(entry
);
1328 BUG_ON(page_mapcount(page
) != 1);
1329 if (!pmd_young(*pmd
))
1330 entry
= pte_mkold(entry
);
1331 pte
= pte_offset_map(&_pmd
, haddr
);
1332 BUG_ON(!pte_none(*pte
));
1333 set_pte_at(mm
, haddr
, pte
, entry
);
1337 smp_wmb(); /* make pte visible before pmd */
1339 * Up to this point the pmd is present and huge and
1340 * userland has the whole access to the hugepage
1341 * during the split (which happens in place). If we
1342 * overwrite the pmd with the not-huge version
1343 * pointing to the pte here (which of course we could
1344 * if all CPUs were bug free), userland could trigger
1345 * a small page size TLB miss on the small sized TLB
1346 * while the hugepage TLB entry is still established
1347 * in the huge TLB. Some CPU doesn't like that. See
1348 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1349 * Erratum 383 on page 93. Intel should be safe but is
1350 * also warns that it's only safe if the permission
1351 * and cache attributes of the two entries loaded in
1352 * the two TLB is identical (which should be the case
1353 * here). But it is generally safer to never allow
1354 * small and huge TLB entries for the same virtual
1355 * address to be loaded simultaneously. So instead of
1356 * doing "pmd_populate(); flush_tlb_range();" we first
1357 * mark the current pmd notpresent (atomically because
1358 * here the pmd_trans_huge and pmd_trans_splitting
1359 * must remain set at all times on the pmd until the
1360 * split is complete for this pmd), then we flush the
1361 * SMP TLB and finally we write the non-huge version
1362 * of the pmd entry with pmd_populate.
1364 pmdp_invalidate(vma
, address
, pmd
);
1365 pmd_populate(mm
, pmd
, pgtable
);
1368 spin_unlock(&mm
->page_table_lock
);
1373 /* must be called with anon_vma->root->mutex hold */
1374 static void __split_huge_page(struct page
*page
,
1375 struct anon_vma
*anon_vma
)
1377 int mapcount
, mapcount2
;
1378 struct anon_vma_chain
*avc
;
1380 BUG_ON(!PageHead(page
));
1381 BUG_ON(PageTail(page
));
1384 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
1385 struct vm_area_struct
*vma
= avc
->vma
;
1386 unsigned long addr
= vma_address(page
, vma
);
1387 BUG_ON(is_vma_temporary_stack(vma
));
1388 if (addr
== -EFAULT
)
1390 mapcount
+= __split_huge_page_splitting(page
, vma
, addr
);
1393 * It is critical that new vmas are added to the tail of the
1394 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1395 * and establishes a child pmd before
1396 * __split_huge_page_splitting() freezes the parent pmd (so if
1397 * we fail to prevent copy_huge_pmd() from running until the
1398 * whole __split_huge_page() is complete), we will still see
1399 * the newly established pmd of the child later during the
1400 * walk, to be able to set it as pmd_trans_splitting too.
1402 if (mapcount
!= page_mapcount(page
))
1403 printk(KERN_ERR
"mapcount %d page_mapcount %d\n",
1404 mapcount
, page_mapcount(page
));
1405 BUG_ON(mapcount
!= page_mapcount(page
));
1407 __split_huge_page_refcount(page
);
1410 list_for_each_entry(avc
, &anon_vma
->head
, same_anon_vma
) {
1411 struct vm_area_struct
*vma
= avc
->vma
;
1412 unsigned long addr
= vma_address(page
, vma
);
1413 BUG_ON(is_vma_temporary_stack(vma
));
1414 if (addr
== -EFAULT
)
1416 mapcount2
+= __split_huge_page_map(page
, vma
, addr
);
1418 if (mapcount
!= mapcount2
)
1419 printk(KERN_ERR
"mapcount %d mapcount2 %d page_mapcount %d\n",
1420 mapcount
, mapcount2
, page_mapcount(page
));
1421 BUG_ON(mapcount
!= mapcount2
);
1424 int split_huge_page(struct page
*page
)
1426 struct anon_vma
*anon_vma
;
1429 BUG_ON(!PageAnon(page
));
1430 anon_vma
= page_lock_anon_vma(page
);
1434 if (!PageCompound(page
))
1437 BUG_ON(!PageSwapBacked(page
));
1438 __split_huge_page(page
, anon_vma
);
1439 count_vm_event(THP_SPLIT
);
1441 BUG_ON(PageCompound(page
));
1443 page_unlock_anon_vma(anon_vma
);
1448 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1450 int hugepage_madvise(struct vm_area_struct
*vma
,
1451 unsigned long *vm_flags
, int advice
)
1453 struct mm_struct
*mm
= vma
->vm_mm
;
1458 * Be somewhat over-protective like KSM for now!
1460 if (*vm_flags
& (VM_HUGEPAGE
| VM_NO_THP
))
1462 if (mm
->def_flags
& VM_NOHUGEPAGE
)
1464 *vm_flags
&= ~VM_NOHUGEPAGE
;
1465 *vm_flags
|= VM_HUGEPAGE
;
1467 * If the vma become good for khugepaged to scan,
1468 * register it here without waiting a page fault that
1469 * may not happen any time soon.
1471 if (unlikely(khugepaged_enter_vma_merge(vma
)))
1474 case MADV_NOHUGEPAGE
:
1476 * Be somewhat over-protective like KSM for now!
1478 if (*vm_flags
& (VM_NOHUGEPAGE
| VM_NO_THP
))
1480 *vm_flags
&= ~VM_HUGEPAGE
;
1481 *vm_flags
|= VM_NOHUGEPAGE
;
1483 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1484 * this vma even if we leave the mm registered in khugepaged if
1485 * it got registered before VM_NOHUGEPAGE was set.
1493 static int __init
khugepaged_slab_init(void)
1495 mm_slot_cache
= kmem_cache_create("khugepaged_mm_slot",
1496 sizeof(struct mm_slot
),
1497 __alignof__(struct mm_slot
), 0, NULL
);
1504 static void __init
khugepaged_slab_free(void)
1506 kmem_cache_destroy(mm_slot_cache
);
1507 mm_slot_cache
= NULL
;
1510 static inline struct mm_slot
*alloc_mm_slot(void)
1512 if (!mm_slot_cache
) /* initialization failed */
1514 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
1517 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
1519 kmem_cache_free(mm_slot_cache
, mm_slot
);
1522 static int __init
mm_slots_hash_init(void)
1524 mm_slots_hash
= kzalloc(MM_SLOTS_HASH_HEADS
* sizeof(struct hlist_head
),
1532 static void __init
mm_slots_hash_free(void)
1534 kfree(mm_slots_hash
);
1535 mm_slots_hash
= NULL
;
1539 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
1541 struct mm_slot
*mm_slot
;
1542 struct hlist_head
*bucket
;
1543 struct hlist_node
*node
;
1545 bucket
= &mm_slots_hash
[((unsigned long)mm
/ sizeof(struct mm_struct
))
1546 % MM_SLOTS_HASH_HEADS
];
1547 hlist_for_each_entry(mm_slot
, node
, bucket
, hash
) {
1548 if (mm
== mm_slot
->mm
)
1554 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
1555 struct mm_slot
*mm_slot
)
1557 struct hlist_head
*bucket
;
1559 bucket
= &mm_slots_hash
[((unsigned long)mm
/ sizeof(struct mm_struct
))
1560 % MM_SLOTS_HASH_HEADS
];
1562 hlist_add_head(&mm_slot
->hash
, bucket
);
1565 static inline int khugepaged_test_exit(struct mm_struct
*mm
)
1567 return atomic_read(&mm
->mm_users
) == 0;
1570 int __khugepaged_enter(struct mm_struct
*mm
)
1572 struct mm_slot
*mm_slot
;
1575 mm_slot
= alloc_mm_slot();
1579 /* __khugepaged_exit() must not run from under us */
1580 VM_BUG_ON(khugepaged_test_exit(mm
));
1581 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE
, &mm
->flags
))) {
1582 free_mm_slot(mm_slot
);
1586 spin_lock(&khugepaged_mm_lock
);
1587 insert_to_mm_slots_hash(mm
, mm_slot
);
1589 * Insert just behind the scanning cursor, to let the area settle
1592 wakeup
= list_empty(&khugepaged_scan
.mm_head
);
1593 list_add_tail(&mm_slot
->mm_node
, &khugepaged_scan
.mm_head
);
1594 spin_unlock(&khugepaged_mm_lock
);
1596 atomic_inc(&mm
->mm_count
);
1598 wake_up_interruptible(&khugepaged_wait
);
1603 int khugepaged_enter_vma_merge(struct vm_area_struct
*vma
)
1605 unsigned long hstart
, hend
;
1608 * Not yet faulted in so we will register later in the
1609 * page fault if needed.
1613 /* khugepaged not yet working on file or special mappings */
1615 VM_BUG_ON(vma
->vm_flags
& VM_NO_THP
);
1616 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
1617 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
1619 return khugepaged_enter(vma
);
1623 void __khugepaged_exit(struct mm_struct
*mm
)
1625 struct mm_slot
*mm_slot
;
1628 spin_lock(&khugepaged_mm_lock
);
1629 mm_slot
= get_mm_slot(mm
);
1630 if (mm_slot
&& khugepaged_scan
.mm_slot
!= mm_slot
) {
1631 hlist_del(&mm_slot
->hash
);
1632 list_del(&mm_slot
->mm_node
);
1635 spin_unlock(&khugepaged_mm_lock
);
1638 clear_bit(MMF_VM_HUGEPAGE
, &mm
->flags
);
1639 free_mm_slot(mm_slot
);
1641 } else if (mm_slot
) {
1643 * This is required to serialize against
1644 * khugepaged_test_exit() (which is guaranteed to run
1645 * under mmap sem read mode). Stop here (after we
1646 * return all pagetables will be destroyed) until
1647 * khugepaged has finished working on the pagetables
1648 * under the mmap_sem.
1650 down_write(&mm
->mmap_sem
);
1651 up_write(&mm
->mmap_sem
);
1655 static void release_pte_page(struct page
*page
)
1657 /* 0 stands for page_is_file_cache(page) == false */
1658 dec_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
1660 putback_lru_page(page
);
1663 static void release_pte_pages(pte_t
*pte
, pte_t
*_pte
)
1665 while (--_pte
>= pte
) {
1666 pte_t pteval
= *_pte
;
1667 if (!pte_none(pteval
))
1668 release_pte_page(pte_page(pteval
));
1672 static void release_all_pte_pages(pte_t
*pte
)
1674 release_pte_pages(pte
, pte
+ HPAGE_PMD_NR
);
1677 static int __collapse_huge_page_isolate(struct vm_area_struct
*vma
,
1678 unsigned long address
,
1683 int referenced
= 0, isolated
= 0, none
= 0;
1684 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
1685 _pte
++, address
+= PAGE_SIZE
) {
1686 pte_t pteval
= *_pte
;
1687 if (pte_none(pteval
)) {
1688 if (++none
<= khugepaged_max_ptes_none
)
1691 release_pte_pages(pte
, _pte
);
1695 if (!pte_present(pteval
) || !pte_write(pteval
)) {
1696 release_pte_pages(pte
, _pte
);
1699 page
= vm_normal_page(vma
, address
, pteval
);
1700 if (unlikely(!page
)) {
1701 release_pte_pages(pte
, _pte
);
1704 VM_BUG_ON(PageCompound(page
));
1705 BUG_ON(!PageAnon(page
));
1706 VM_BUG_ON(!PageSwapBacked(page
));
1708 /* cannot use mapcount: can't collapse if there's a gup pin */
1709 if (page_count(page
) != 1) {
1710 release_pte_pages(pte
, _pte
);
1714 * We can do it before isolate_lru_page because the
1715 * page can't be freed from under us. NOTE: PG_lock
1716 * is needed to serialize against split_huge_page
1717 * when invoked from the VM.
1719 if (!trylock_page(page
)) {
1720 release_pte_pages(pte
, _pte
);
1724 * Isolate the page to avoid collapsing an hugepage
1725 * currently in use by the VM.
1727 if (isolate_lru_page(page
)) {
1729 release_pte_pages(pte
, _pte
);
1732 /* 0 stands for page_is_file_cache(page) == false */
1733 inc_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
1734 VM_BUG_ON(!PageLocked(page
));
1735 VM_BUG_ON(PageLRU(page
));
1737 /* If there is no mapped pte young don't collapse the page */
1738 if (pte_young(pteval
) || PageReferenced(page
) ||
1739 mmu_notifier_test_young(vma
->vm_mm
, address
))
1742 if (unlikely(!referenced
))
1743 release_all_pte_pages(pte
);
1750 static void __collapse_huge_page_copy(pte_t
*pte
, struct page
*page
,
1751 struct vm_area_struct
*vma
,
1752 unsigned long address
,
1756 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
; _pte
++) {
1757 pte_t pteval
= *_pte
;
1758 struct page
*src_page
;
1760 if (pte_none(pteval
)) {
1761 clear_user_highpage(page
, address
);
1762 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, 1);
1764 src_page
= pte_page(pteval
);
1765 copy_user_highpage(page
, src_page
, address
, vma
);
1766 VM_BUG_ON(page_mapcount(src_page
) != 1);
1767 release_pte_page(src_page
);
1769 * ptl mostly unnecessary, but preempt has to
1770 * be disabled to update the per-cpu stats
1771 * inside page_remove_rmap().
1775 * paravirt calls inside pte_clear here are
1778 pte_clear(vma
->vm_mm
, address
, _pte
);
1779 page_remove_rmap(src_page
);
1781 free_page_and_swap_cache(src_page
);
1784 address
+= PAGE_SIZE
;
1789 static void khugepaged_alloc_sleep(void)
1791 wait_event_freezable_timeout(khugepaged_wait
, false,
1792 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs
));
1796 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
1798 if (IS_ERR(*hpage
)) {
1803 khugepaged_alloc_sleep();
1804 } else if (*hpage
) {
1813 *khugepaged_alloc_page(struct page
**hpage
, struct mm_struct
*mm
,
1814 struct vm_area_struct
*vma
, unsigned long address
,
1819 * Allocate the page while the vma is still valid and under
1820 * the mmap_sem read mode so there is no memory allocation
1821 * later when we take the mmap_sem in write mode. This is more
1822 * friendly behavior (OTOH it may actually hide bugs) to
1823 * filesystems in userland with daemons allocating memory in
1824 * the userland I/O paths. Allocating memory with the
1825 * mmap_sem in read mode is good idea also to allow greater
1828 *hpage
= alloc_hugepage_vma(khugepaged_defrag(), vma
, address
,
1829 node
, __GFP_OTHER_NODE
);
1832 * After allocating the hugepage, release the mmap_sem read lock in
1833 * preparation for taking it in write mode.
1835 up_read(&mm
->mmap_sem
);
1836 if (unlikely(!*hpage
)) {
1837 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
1838 *hpage
= ERR_PTR(-ENOMEM
);
1842 count_vm_event(THP_COLLAPSE_ALLOC
);
1846 static struct page
*khugepaged_alloc_hugepage(bool *wait
)
1851 hpage
= alloc_hugepage(khugepaged_defrag());
1853 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
1858 khugepaged_alloc_sleep();
1860 count_vm_event(THP_COLLAPSE_ALLOC
);
1861 } while (unlikely(!hpage
) && likely(khugepaged_enabled()));
1866 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
1869 *hpage
= khugepaged_alloc_hugepage(wait
);
1871 if (unlikely(!*hpage
))
1878 *khugepaged_alloc_page(struct page
**hpage
, struct mm_struct
*mm
,
1879 struct vm_area_struct
*vma
, unsigned long address
,
1882 up_read(&mm
->mmap_sem
);
1888 static void collapse_huge_page(struct mm_struct
*mm
,
1889 unsigned long address
,
1890 struct page
**hpage
,
1891 struct vm_area_struct
*vma
,
1899 struct page
*new_page
;
1902 unsigned long hstart
, hend
;
1904 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
1906 /* release the mmap_sem read lock. */
1907 new_page
= khugepaged_alloc_page(hpage
, mm
, vma
, address
, node
);
1911 if (unlikely(mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
)))
1915 * Prevent all access to pagetables with the exception of
1916 * gup_fast later hanlded by the ptep_clear_flush and the VM
1917 * handled by the anon_vma lock + PG_lock.
1919 down_write(&mm
->mmap_sem
);
1920 if (unlikely(khugepaged_test_exit(mm
)))
1923 vma
= find_vma(mm
, address
);
1924 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
1925 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
1926 if (address
< hstart
|| address
+ HPAGE_PMD_SIZE
> hend
)
1929 if ((!(vma
->vm_flags
& VM_HUGEPAGE
) && !khugepaged_always()) ||
1930 (vma
->vm_flags
& VM_NOHUGEPAGE
))
1933 if (!vma
->anon_vma
|| vma
->vm_ops
)
1935 if (is_vma_temporary_stack(vma
))
1937 VM_BUG_ON(vma
->vm_flags
& VM_NO_THP
);
1939 pgd
= pgd_offset(mm
, address
);
1940 if (!pgd_present(*pgd
))
1943 pud
= pud_offset(pgd
, address
);
1944 if (!pud_present(*pud
))
1947 pmd
= pmd_offset(pud
, address
);
1948 /* pmd can't go away or become huge under us */
1949 if (!pmd_present(*pmd
) || pmd_trans_huge(*pmd
))
1952 anon_vma_lock(vma
->anon_vma
);
1954 pte
= pte_offset_map(pmd
, address
);
1955 ptl
= pte_lockptr(mm
, pmd
);
1957 spin_lock(&mm
->page_table_lock
); /* probably unnecessary */
1959 * After this gup_fast can't run anymore. This also removes
1960 * any huge TLB entry from the CPU so we won't allow
1961 * huge and small TLB entries for the same virtual address
1962 * to avoid the risk of CPU bugs in that area.
1964 _pmd
= pmdp_clear_flush_notify(vma
, address
, pmd
);
1965 spin_unlock(&mm
->page_table_lock
);
1968 isolated
= __collapse_huge_page_isolate(vma
, address
, pte
);
1971 if (unlikely(!isolated
)) {
1973 spin_lock(&mm
->page_table_lock
);
1974 BUG_ON(!pmd_none(*pmd
));
1975 set_pmd_at(mm
, address
, pmd
, _pmd
);
1976 spin_unlock(&mm
->page_table_lock
);
1977 anon_vma_unlock(vma
->anon_vma
);
1982 * All pages are isolated and locked so anon_vma rmap
1983 * can't run anymore.
1985 anon_vma_unlock(vma
->anon_vma
);
1987 __collapse_huge_page_copy(pte
, new_page
, vma
, address
, ptl
);
1989 __SetPageUptodate(new_page
);
1990 pgtable
= pmd_pgtable(_pmd
);
1992 _pmd
= mk_pmd(new_page
, vma
->vm_page_prot
);
1993 _pmd
= maybe_pmd_mkwrite(pmd_mkdirty(_pmd
), vma
);
1994 _pmd
= pmd_mkhuge(_pmd
);
1997 * spin_lock() below is not the equivalent of smp_wmb(), so
1998 * this is needed to avoid the copy_huge_page writes to become
1999 * visible after the set_pmd_at() write.
2003 spin_lock(&mm
->page_table_lock
);
2004 BUG_ON(!pmd_none(*pmd
));
2005 page_add_new_anon_rmap(new_page
, vma
, address
);
2006 set_pmd_at(mm
, address
, pmd
, _pmd
);
2007 update_mmu_cache(vma
, address
, _pmd
);
2008 pgtable_trans_huge_deposit(mm
, pgtable
);
2009 spin_unlock(&mm
->page_table_lock
);
2013 khugepaged_pages_collapsed
++;
2015 up_write(&mm
->mmap_sem
);
2019 mem_cgroup_uncharge_page(new_page
);
2023 static int khugepaged_scan_pmd(struct mm_struct
*mm
,
2024 struct vm_area_struct
*vma
,
2025 unsigned long address
,
2026 struct page
**hpage
)
2032 int ret
= 0, referenced
= 0, none
= 0;
2034 unsigned long _address
;
2038 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2040 pgd
= pgd_offset(mm
, address
);
2041 if (!pgd_present(*pgd
))
2044 pud
= pud_offset(pgd
, address
);
2045 if (!pud_present(*pud
))
2048 pmd
= pmd_offset(pud
, address
);
2049 if (!pmd_present(*pmd
) || pmd_trans_huge(*pmd
))
2052 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2053 for (_address
= address
, _pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
2054 _pte
++, _address
+= PAGE_SIZE
) {
2055 pte_t pteval
= *_pte
;
2056 if (pte_none(pteval
)) {
2057 if (++none
<= khugepaged_max_ptes_none
)
2062 if (!pte_present(pteval
) || !pte_write(pteval
))
2064 page
= vm_normal_page(vma
, _address
, pteval
);
2065 if (unlikely(!page
))
2068 * Chose the node of the first page. This could
2069 * be more sophisticated and look at more pages,
2070 * but isn't for now.
2073 node
= page_to_nid(page
);
2074 VM_BUG_ON(PageCompound(page
));
2075 if (!PageLRU(page
) || PageLocked(page
) || !PageAnon(page
))
2077 /* cannot use mapcount: can't collapse if there's a gup pin */
2078 if (page_count(page
) != 1)
2080 if (pte_young(pteval
) || PageReferenced(page
) ||
2081 mmu_notifier_test_young(vma
->vm_mm
, address
))
2087 pte_unmap_unlock(pte
, ptl
);
2089 /* collapse_huge_page will return with the mmap_sem released */
2090 collapse_huge_page(mm
, address
, hpage
, vma
, node
);
2095 static void collect_mm_slot(struct mm_slot
*mm_slot
)
2097 struct mm_struct
*mm
= mm_slot
->mm
;
2099 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2101 if (khugepaged_test_exit(mm
)) {
2103 hlist_del(&mm_slot
->hash
);
2104 list_del(&mm_slot
->mm_node
);
2107 * Not strictly needed because the mm exited already.
2109 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2112 /* khugepaged_mm_lock actually not necessary for the below */
2113 free_mm_slot(mm_slot
);
2118 static unsigned int khugepaged_scan_mm_slot(unsigned int pages
,
2119 struct page
**hpage
)
2120 __releases(&khugepaged_mm_lock
)
2121 __acquires(&khugepaged_mm_lock
)
2123 struct mm_slot
*mm_slot
;
2124 struct mm_struct
*mm
;
2125 struct vm_area_struct
*vma
;
2129 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2131 if (khugepaged_scan
.mm_slot
)
2132 mm_slot
= khugepaged_scan
.mm_slot
;
2134 mm_slot
= list_entry(khugepaged_scan
.mm_head
.next
,
2135 struct mm_slot
, mm_node
);
2136 khugepaged_scan
.address
= 0;
2137 khugepaged_scan
.mm_slot
= mm_slot
;
2139 spin_unlock(&khugepaged_mm_lock
);
2142 down_read(&mm
->mmap_sem
);
2143 if (unlikely(khugepaged_test_exit(mm
)))
2146 vma
= find_vma(mm
, khugepaged_scan
.address
);
2149 for (; vma
; vma
= vma
->vm_next
) {
2150 unsigned long hstart
, hend
;
2153 if (unlikely(khugepaged_test_exit(mm
))) {
2158 if ((!(vma
->vm_flags
& VM_HUGEPAGE
) &&
2159 !khugepaged_always()) ||
2160 (vma
->vm_flags
& VM_NOHUGEPAGE
)) {
2165 if (!vma
->anon_vma
|| vma
->vm_ops
)
2167 if (is_vma_temporary_stack(vma
))
2169 VM_BUG_ON(vma
->vm_flags
& VM_NO_THP
);
2171 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2172 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2175 if (khugepaged_scan
.address
> hend
)
2177 if (khugepaged_scan
.address
< hstart
)
2178 khugepaged_scan
.address
= hstart
;
2179 VM_BUG_ON(khugepaged_scan
.address
& ~HPAGE_PMD_MASK
);
2181 while (khugepaged_scan
.address
< hend
) {
2184 if (unlikely(khugepaged_test_exit(mm
)))
2185 goto breakouterloop
;
2187 VM_BUG_ON(khugepaged_scan
.address
< hstart
||
2188 khugepaged_scan
.address
+ HPAGE_PMD_SIZE
>
2190 ret
= khugepaged_scan_pmd(mm
, vma
,
2191 khugepaged_scan
.address
,
2193 /* move to next address */
2194 khugepaged_scan
.address
+= HPAGE_PMD_SIZE
;
2195 progress
+= HPAGE_PMD_NR
;
2197 /* we released mmap_sem so break loop */
2198 goto breakouterloop_mmap_sem
;
2199 if (progress
>= pages
)
2200 goto breakouterloop
;
2204 up_read(&mm
->mmap_sem
); /* exit_mmap will destroy ptes after this */
2205 breakouterloop_mmap_sem
:
2207 spin_lock(&khugepaged_mm_lock
);
2208 VM_BUG_ON(khugepaged_scan
.mm_slot
!= mm_slot
);
2210 * Release the current mm_slot if this mm is about to die, or
2211 * if we scanned all vmas of this mm.
2213 if (khugepaged_test_exit(mm
) || !vma
) {
2215 * Make sure that if mm_users is reaching zero while
2216 * khugepaged runs here, khugepaged_exit will find
2217 * mm_slot not pointing to the exiting mm.
2219 if (mm_slot
->mm_node
.next
!= &khugepaged_scan
.mm_head
) {
2220 khugepaged_scan
.mm_slot
= list_entry(
2221 mm_slot
->mm_node
.next
,
2222 struct mm_slot
, mm_node
);
2223 khugepaged_scan
.address
= 0;
2225 khugepaged_scan
.mm_slot
= NULL
;
2226 khugepaged_full_scans
++;
2229 collect_mm_slot(mm_slot
);
2235 static int khugepaged_has_work(void)
2237 return !list_empty(&khugepaged_scan
.mm_head
) &&
2238 khugepaged_enabled();
2241 static int khugepaged_wait_event(void)
2243 return !list_empty(&khugepaged_scan
.mm_head
) ||
2244 kthread_should_stop();
2247 static void khugepaged_do_scan(void)
2249 struct page
*hpage
= NULL
;
2250 unsigned int progress
= 0, pass_through_head
= 0;
2251 unsigned int pages
= khugepaged_pages_to_scan
;
2254 barrier(); /* write khugepaged_pages_to_scan to local stack */
2256 while (progress
< pages
) {
2257 if (!khugepaged_prealloc_page(&hpage
, &wait
))
2262 if (unlikely(kthread_should_stop() || freezing(current
)))
2265 spin_lock(&khugepaged_mm_lock
);
2266 if (!khugepaged_scan
.mm_slot
)
2267 pass_through_head
++;
2268 if (khugepaged_has_work() &&
2269 pass_through_head
< 2)
2270 progress
+= khugepaged_scan_mm_slot(pages
- progress
,
2274 spin_unlock(&khugepaged_mm_lock
);
2277 if (!IS_ERR_OR_NULL(hpage
))
2281 static void khugepaged_wait_work(void)
2285 if (khugepaged_has_work()) {
2286 if (!khugepaged_scan_sleep_millisecs
)
2289 wait_event_freezable_timeout(khugepaged_wait
,
2290 kthread_should_stop(),
2291 msecs_to_jiffies(khugepaged_scan_sleep_millisecs
));
2295 if (khugepaged_enabled())
2296 wait_event_freezable(khugepaged_wait
, khugepaged_wait_event());
2299 static int khugepaged(void *none
)
2301 struct mm_slot
*mm_slot
;
2304 set_user_nice(current
, 19);
2306 while (!kthread_should_stop()) {
2307 khugepaged_do_scan();
2308 khugepaged_wait_work();
2311 spin_lock(&khugepaged_mm_lock
);
2312 mm_slot
= khugepaged_scan
.mm_slot
;
2313 khugepaged_scan
.mm_slot
= NULL
;
2315 collect_mm_slot(mm_slot
);
2316 spin_unlock(&khugepaged_mm_lock
);
2320 void __split_huge_page_pmd(struct mm_struct
*mm
, pmd_t
*pmd
)
2324 spin_lock(&mm
->page_table_lock
);
2325 if (unlikely(!pmd_trans_huge(*pmd
))) {
2326 spin_unlock(&mm
->page_table_lock
);
2329 page
= pmd_page(*pmd
);
2330 VM_BUG_ON(!page_count(page
));
2332 spin_unlock(&mm
->page_table_lock
);
2334 split_huge_page(page
);
2337 BUG_ON(pmd_trans_huge(*pmd
));
2340 static void split_huge_page_address(struct mm_struct
*mm
,
2341 unsigned long address
)
2347 VM_BUG_ON(!(address
& ~HPAGE_PMD_MASK
));
2349 pgd
= pgd_offset(mm
, address
);
2350 if (!pgd_present(*pgd
))
2353 pud
= pud_offset(pgd
, address
);
2354 if (!pud_present(*pud
))
2357 pmd
= pmd_offset(pud
, address
);
2358 if (!pmd_present(*pmd
))
2361 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2362 * materialize from under us.
2364 split_huge_page_pmd(mm
, pmd
);
2367 void __vma_adjust_trans_huge(struct vm_area_struct
*vma
,
2368 unsigned long start
,
2373 * If the new start address isn't hpage aligned and it could
2374 * previously contain an hugepage: check if we need to split
2377 if (start
& ~HPAGE_PMD_MASK
&&
2378 (start
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2379 (start
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2380 split_huge_page_address(vma
->vm_mm
, start
);
2383 * If the new end address isn't hpage aligned and it could
2384 * previously contain an hugepage: check if we need to split
2387 if (end
& ~HPAGE_PMD_MASK
&&
2388 (end
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2389 (end
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2390 split_huge_page_address(vma
->vm_mm
, end
);
2393 * If we're also updating the vma->vm_next->vm_start, if the new
2394 * vm_next->vm_start isn't page aligned and it could previously
2395 * contain an hugepage: check if we need to split an huge pmd.
2397 if (adjust_next
> 0) {
2398 struct vm_area_struct
*next
= vma
->vm_next
;
2399 unsigned long nstart
= next
->vm_start
;
2400 nstart
+= adjust_next
<< PAGE_SHIFT
;
2401 if (nstart
& ~HPAGE_PMD_MASK
&&
2402 (nstart
& HPAGE_PMD_MASK
) >= next
->vm_start
&&
2403 (nstart
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= next
->vm_end
)
2404 split_huge_page_address(next
->vm_mm
, nstart
);