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>
21 #include <linux/migrate.h>
23 #include <asm/pgalloc.h>
27 * By default transparent hugepage support is enabled for all mappings
28 * and khugepaged scans all mappings. Defrag is only invoked by
29 * khugepaged hugepage allocations and by page faults inside
30 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
33 unsigned long transparent_hugepage_flags __read_mostly
=
34 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
35 (1<<TRANSPARENT_HUGEPAGE_FLAG
)|
37 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
38 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
)|
40 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
)|
41 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
43 /* default scan 8*512 pte (or vmas) every 30 second */
44 static unsigned int khugepaged_pages_to_scan __read_mostly
= HPAGE_PMD_NR
*8;
45 static unsigned int khugepaged_pages_collapsed
;
46 static unsigned int khugepaged_full_scans
;
47 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly
= 10000;
48 /* during fragmentation poll the hugepage allocator once every minute */
49 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly
= 60000;
50 static struct task_struct
*khugepaged_thread __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
;
165 static ssize_t
double_flag_show(struct kobject
*kobj
,
166 struct kobj_attribute
*attr
, char *buf
,
167 enum transparent_hugepage_flag enabled
,
168 enum transparent_hugepage_flag req_madv
)
170 if (test_bit(enabled
, &transparent_hugepage_flags
)) {
171 VM_BUG_ON(test_bit(req_madv
, &transparent_hugepage_flags
));
172 return sprintf(buf
, "[always] madvise never\n");
173 } else if (test_bit(req_madv
, &transparent_hugepage_flags
))
174 return sprintf(buf
, "always [madvise] never\n");
176 return sprintf(buf
, "always madvise [never]\n");
178 static ssize_t
double_flag_store(struct kobject
*kobj
,
179 struct kobj_attribute
*attr
,
180 const char *buf
, size_t count
,
181 enum transparent_hugepage_flag enabled
,
182 enum transparent_hugepage_flag req_madv
)
184 if (!memcmp("always", buf
,
185 min(sizeof("always")-1, count
))) {
186 set_bit(enabled
, &transparent_hugepage_flags
);
187 clear_bit(req_madv
, &transparent_hugepage_flags
);
188 } else if (!memcmp("madvise", buf
,
189 min(sizeof("madvise")-1, count
))) {
190 clear_bit(enabled
, &transparent_hugepage_flags
);
191 set_bit(req_madv
, &transparent_hugepage_flags
);
192 } else if (!memcmp("never", buf
,
193 min(sizeof("never")-1, count
))) {
194 clear_bit(enabled
, &transparent_hugepage_flags
);
195 clear_bit(req_madv
, &transparent_hugepage_flags
);
202 static ssize_t
enabled_show(struct kobject
*kobj
,
203 struct kobj_attribute
*attr
, char *buf
)
205 return double_flag_show(kobj
, attr
, buf
,
206 TRANSPARENT_HUGEPAGE_FLAG
,
207 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
209 static ssize_t
enabled_store(struct kobject
*kobj
,
210 struct kobj_attribute
*attr
,
211 const char *buf
, size_t count
)
215 ret
= double_flag_store(kobj
, attr
, buf
, count
,
216 TRANSPARENT_HUGEPAGE_FLAG
,
217 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
);
222 mutex_lock(&khugepaged_mutex
);
223 err
= start_khugepaged();
224 mutex_unlock(&khugepaged_mutex
);
232 static struct kobj_attribute enabled_attr
=
233 __ATTR(enabled
, 0644, enabled_show
, enabled_store
);
235 static ssize_t
single_flag_show(struct kobject
*kobj
,
236 struct kobj_attribute
*attr
, char *buf
,
237 enum transparent_hugepage_flag flag
)
239 return sprintf(buf
, "%d\n",
240 !!test_bit(flag
, &transparent_hugepage_flags
));
243 static ssize_t
single_flag_store(struct kobject
*kobj
,
244 struct kobj_attribute
*attr
,
245 const char *buf
, size_t count
,
246 enum transparent_hugepage_flag flag
)
251 ret
= kstrtoul(buf
, 10, &value
);
258 set_bit(flag
, &transparent_hugepage_flags
);
260 clear_bit(flag
, &transparent_hugepage_flags
);
266 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
267 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
268 * memory just to allocate one more hugepage.
270 static ssize_t
defrag_show(struct kobject
*kobj
,
271 struct kobj_attribute
*attr
, char *buf
)
273 return double_flag_show(kobj
, attr
, buf
,
274 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
,
275 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
);
277 static ssize_t
defrag_store(struct kobject
*kobj
,
278 struct kobj_attribute
*attr
,
279 const char *buf
, size_t count
)
281 return double_flag_store(kobj
, attr
, buf
, count
,
282 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG
,
283 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
);
285 static struct kobj_attribute defrag_attr
=
286 __ATTR(defrag
, 0644, defrag_show
, defrag_store
);
288 #ifdef CONFIG_DEBUG_VM
289 static ssize_t
debug_cow_show(struct kobject
*kobj
,
290 struct kobj_attribute
*attr
, char *buf
)
292 return single_flag_show(kobj
, attr
, buf
,
293 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
295 static ssize_t
debug_cow_store(struct kobject
*kobj
,
296 struct kobj_attribute
*attr
,
297 const char *buf
, size_t count
)
299 return single_flag_store(kobj
, attr
, buf
, count
,
300 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
302 static struct kobj_attribute debug_cow_attr
=
303 __ATTR(debug_cow
, 0644, debug_cow_show
, debug_cow_store
);
304 #endif /* CONFIG_DEBUG_VM */
306 static struct attribute
*hugepage_attr
[] = {
309 #ifdef CONFIG_DEBUG_VM
310 &debug_cow_attr
.attr
,
315 static struct attribute_group hugepage_attr_group
= {
316 .attrs
= hugepage_attr
,
319 static ssize_t
scan_sleep_millisecs_show(struct kobject
*kobj
,
320 struct kobj_attribute
*attr
,
323 return sprintf(buf
, "%u\n", khugepaged_scan_sleep_millisecs
);
326 static ssize_t
scan_sleep_millisecs_store(struct kobject
*kobj
,
327 struct kobj_attribute
*attr
,
328 const char *buf
, size_t count
)
333 err
= strict_strtoul(buf
, 10, &msecs
);
334 if (err
|| msecs
> UINT_MAX
)
337 khugepaged_scan_sleep_millisecs
= msecs
;
338 wake_up_interruptible(&khugepaged_wait
);
342 static struct kobj_attribute scan_sleep_millisecs_attr
=
343 __ATTR(scan_sleep_millisecs
, 0644, scan_sleep_millisecs_show
,
344 scan_sleep_millisecs_store
);
346 static ssize_t
alloc_sleep_millisecs_show(struct kobject
*kobj
,
347 struct kobj_attribute
*attr
,
350 return sprintf(buf
, "%u\n", khugepaged_alloc_sleep_millisecs
);
353 static ssize_t
alloc_sleep_millisecs_store(struct kobject
*kobj
,
354 struct kobj_attribute
*attr
,
355 const char *buf
, size_t count
)
360 err
= strict_strtoul(buf
, 10, &msecs
);
361 if (err
|| msecs
> UINT_MAX
)
364 khugepaged_alloc_sleep_millisecs
= msecs
;
365 wake_up_interruptible(&khugepaged_wait
);
369 static struct kobj_attribute alloc_sleep_millisecs_attr
=
370 __ATTR(alloc_sleep_millisecs
, 0644, alloc_sleep_millisecs_show
,
371 alloc_sleep_millisecs_store
);
373 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
374 struct kobj_attribute
*attr
,
377 return sprintf(buf
, "%u\n", khugepaged_pages_to_scan
);
379 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
380 struct kobj_attribute
*attr
,
381 const char *buf
, size_t count
)
386 err
= strict_strtoul(buf
, 10, &pages
);
387 if (err
|| !pages
|| pages
> UINT_MAX
)
390 khugepaged_pages_to_scan
= pages
;
394 static struct kobj_attribute pages_to_scan_attr
=
395 __ATTR(pages_to_scan
, 0644, pages_to_scan_show
,
396 pages_to_scan_store
);
398 static ssize_t
pages_collapsed_show(struct kobject
*kobj
,
399 struct kobj_attribute
*attr
,
402 return sprintf(buf
, "%u\n", khugepaged_pages_collapsed
);
404 static struct kobj_attribute pages_collapsed_attr
=
405 __ATTR_RO(pages_collapsed
);
407 static ssize_t
full_scans_show(struct kobject
*kobj
,
408 struct kobj_attribute
*attr
,
411 return sprintf(buf
, "%u\n", khugepaged_full_scans
);
413 static struct kobj_attribute full_scans_attr
=
414 __ATTR_RO(full_scans
);
416 static ssize_t
khugepaged_defrag_show(struct kobject
*kobj
,
417 struct kobj_attribute
*attr
, char *buf
)
419 return single_flag_show(kobj
, attr
, buf
,
420 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
422 static ssize_t
khugepaged_defrag_store(struct kobject
*kobj
,
423 struct kobj_attribute
*attr
,
424 const char *buf
, size_t count
)
426 return single_flag_store(kobj
, attr
, buf
, count
,
427 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
429 static struct kobj_attribute khugepaged_defrag_attr
=
430 __ATTR(defrag
, 0644, khugepaged_defrag_show
,
431 khugepaged_defrag_store
);
434 * max_ptes_none controls if khugepaged should collapse hugepages over
435 * any unmapped ptes in turn potentially increasing the memory
436 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
437 * reduce the available free memory in the system as it
438 * runs. Increasing max_ptes_none will instead potentially reduce the
439 * free memory in the system during the khugepaged scan.
441 static ssize_t
khugepaged_max_ptes_none_show(struct kobject
*kobj
,
442 struct kobj_attribute
*attr
,
445 return sprintf(buf
, "%u\n", khugepaged_max_ptes_none
);
447 static ssize_t
khugepaged_max_ptes_none_store(struct kobject
*kobj
,
448 struct kobj_attribute
*attr
,
449 const char *buf
, size_t count
)
452 unsigned long max_ptes_none
;
454 err
= strict_strtoul(buf
, 10, &max_ptes_none
);
455 if (err
|| max_ptes_none
> HPAGE_PMD_NR
-1)
458 khugepaged_max_ptes_none
= max_ptes_none
;
462 static struct kobj_attribute khugepaged_max_ptes_none_attr
=
463 __ATTR(max_ptes_none
, 0644, khugepaged_max_ptes_none_show
,
464 khugepaged_max_ptes_none_store
);
466 static struct attribute
*khugepaged_attr
[] = {
467 &khugepaged_defrag_attr
.attr
,
468 &khugepaged_max_ptes_none_attr
.attr
,
469 &pages_to_scan_attr
.attr
,
470 &pages_collapsed_attr
.attr
,
471 &full_scans_attr
.attr
,
472 &scan_sleep_millisecs_attr
.attr
,
473 &alloc_sleep_millisecs_attr
.attr
,
477 static struct attribute_group khugepaged_attr_group
= {
478 .attrs
= khugepaged_attr
,
479 .name
= "khugepaged",
482 static int __init
hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
486 *hugepage_kobj
= kobject_create_and_add("transparent_hugepage", mm_kobj
);
487 if (unlikely(!*hugepage_kobj
)) {
488 printk(KERN_ERR
"hugepage: failed kobject create\n");
492 err
= sysfs_create_group(*hugepage_kobj
, &hugepage_attr_group
);
494 printk(KERN_ERR
"hugepage: failed register hugeage group\n");
498 err
= sysfs_create_group(*hugepage_kobj
, &khugepaged_attr_group
);
500 printk(KERN_ERR
"hugepage: failed register hugeage group\n");
501 goto remove_hp_group
;
507 sysfs_remove_group(*hugepage_kobj
, &hugepage_attr_group
);
509 kobject_put(*hugepage_kobj
);
513 static void __init
hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
515 sysfs_remove_group(hugepage_kobj
, &khugepaged_attr_group
);
516 sysfs_remove_group(hugepage_kobj
, &hugepage_attr_group
);
517 kobject_put(hugepage_kobj
);
520 static inline int hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
525 static inline void hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
528 #endif /* CONFIG_SYSFS */
530 static int __init
hugepage_init(void)
533 struct kobject
*hugepage_kobj
;
535 if (!has_transparent_hugepage()) {
536 transparent_hugepage_flags
= 0;
540 err
= hugepage_init_sysfs(&hugepage_kobj
);
544 err
= khugepaged_slab_init();
548 err
= mm_slots_hash_init();
550 khugepaged_slab_free();
555 * By default disable transparent hugepages on smaller systems,
556 * where the extra memory used could hurt more than TLB overhead
557 * is likely to save. The admin can still enable it through /sys.
559 if (totalram_pages
< (512 << (20 - PAGE_SHIFT
)))
560 transparent_hugepage_flags
= 0;
566 hugepage_exit_sysfs(hugepage_kobj
);
569 module_init(hugepage_init
)
571 static int __init
setup_transparent_hugepage(char *str
)
576 if (!strcmp(str
, "always")) {
577 set_bit(TRANSPARENT_HUGEPAGE_FLAG
,
578 &transparent_hugepage_flags
);
579 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
580 &transparent_hugepage_flags
);
582 } else if (!strcmp(str
, "madvise")) {
583 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
584 &transparent_hugepage_flags
);
585 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
586 &transparent_hugepage_flags
);
588 } else if (!strcmp(str
, "never")) {
589 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
590 &transparent_hugepage_flags
);
591 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
592 &transparent_hugepage_flags
);
598 "transparent_hugepage= cannot parse, ignored\n");
601 __setup("transparent_hugepage=", setup_transparent_hugepage
);
603 static inline pmd_t
maybe_pmd_mkwrite(pmd_t pmd
, struct vm_area_struct
*vma
)
605 if (likely(vma
->vm_flags
& VM_WRITE
))
606 pmd
= pmd_mkwrite(pmd
);
610 static int __do_huge_pmd_anonymous_page(struct mm_struct
*mm
,
611 struct vm_area_struct
*vma
,
612 unsigned long haddr
, pmd_t
*pmd
,
617 VM_BUG_ON(!PageCompound(page
));
618 pgtable
= pte_alloc_one(mm
, haddr
);
619 if (unlikely(!pgtable
))
622 clear_huge_page(page
, haddr
, HPAGE_PMD_NR
);
623 __SetPageUptodate(page
);
625 spin_lock(&mm
->page_table_lock
);
626 if (unlikely(!pmd_none(*pmd
))) {
627 spin_unlock(&mm
->page_table_lock
);
628 mem_cgroup_uncharge_page(page
);
630 pte_free(mm
, pgtable
);
633 entry
= mk_pmd(page
, vma
->vm_page_prot
);
634 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
635 entry
= pmd_mkhuge(entry
);
637 * The spinlocking to take the lru_lock inside
638 * page_add_new_anon_rmap() acts as a full memory
639 * barrier to be sure clear_huge_page writes become
640 * visible after the set_pmd_at() write.
642 page_add_new_anon_rmap(page
, vma
, haddr
);
643 set_pmd_at(mm
, haddr
, pmd
, entry
);
644 pgtable_trans_huge_deposit(mm
, pgtable
);
645 add_mm_counter(mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
647 spin_unlock(&mm
->page_table_lock
);
653 static inline gfp_t
alloc_hugepage_gfpmask(int defrag
, gfp_t extra_gfp
)
655 return (GFP_TRANSHUGE
& ~(defrag
? 0 : __GFP_WAIT
)) | extra_gfp
;
658 static inline struct page
*alloc_hugepage_vma(int defrag
,
659 struct vm_area_struct
*vma
,
660 unsigned long haddr
, int nd
,
663 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag
, extra_gfp
),
664 HPAGE_PMD_ORDER
, vma
, haddr
, nd
);
668 static inline struct page
*alloc_hugepage(int defrag
)
670 return alloc_pages(alloc_hugepage_gfpmask(defrag
, 0),
675 int do_huge_pmd_anonymous_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
676 unsigned long address
, pmd_t
*pmd
,
680 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
683 if (haddr
>= vma
->vm_start
&& haddr
+ HPAGE_PMD_SIZE
<= vma
->vm_end
) {
684 if (unlikely(anon_vma_prepare(vma
)))
686 if (unlikely(khugepaged_enter(vma
)))
688 page
= alloc_hugepage_vma(transparent_hugepage_defrag(vma
),
689 vma
, haddr
, numa_node_id(), 0);
690 if (unlikely(!page
)) {
691 count_vm_event(THP_FAULT_FALLBACK
);
694 count_vm_event(THP_FAULT_ALLOC
);
695 if (unlikely(mem_cgroup_newpage_charge(page
, mm
, GFP_KERNEL
))) {
699 if (unlikely(__do_huge_pmd_anonymous_page(mm
, vma
, haddr
, pmd
,
701 mem_cgroup_uncharge_page(page
);
710 * Use __pte_alloc instead of pte_alloc_map, because we can't
711 * run pte_offset_map on the pmd, if an huge pmd could
712 * materialize from under us from a different thread.
714 if (unlikely(pmd_none(*pmd
)) &&
715 unlikely(__pte_alloc(mm
, vma
, pmd
, address
)))
717 /* if an huge pmd materialized from under us just retry later */
718 if (unlikely(pmd_trans_huge(*pmd
)))
721 * A regular pmd is established and it can't morph into a huge pmd
722 * from under us anymore at this point because we hold the mmap_sem
723 * read mode and khugepaged takes it in write mode. So now it's
724 * safe to run pte_offset_map().
726 pte
= pte_offset_map(pmd
, address
);
727 return handle_pte_fault(mm
, vma
, address
, pte
, pmd
, flags
);
730 int copy_huge_pmd(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
731 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
732 struct vm_area_struct
*vma
)
734 struct page
*src_page
;
740 pgtable
= pte_alloc_one(dst_mm
, addr
);
741 if (unlikely(!pgtable
))
744 spin_lock(&dst_mm
->page_table_lock
);
745 spin_lock_nested(&src_mm
->page_table_lock
, SINGLE_DEPTH_NESTING
);
749 if (unlikely(!pmd_trans_huge(pmd
))) {
750 pte_free(dst_mm
, pgtable
);
753 if (unlikely(pmd_trans_splitting(pmd
))) {
754 /* split huge page running from under us */
755 spin_unlock(&src_mm
->page_table_lock
);
756 spin_unlock(&dst_mm
->page_table_lock
);
757 pte_free(dst_mm
, pgtable
);
759 wait_split_huge_page(vma
->anon_vma
, src_pmd
); /* src_vma */
762 src_page
= pmd_page(pmd
);
763 VM_BUG_ON(!PageHead(src_page
));
765 page_dup_rmap(src_page
);
766 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
768 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
769 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
770 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
771 pgtable_trans_huge_deposit(dst_mm
, pgtable
);
776 spin_unlock(&src_mm
->page_table_lock
);
777 spin_unlock(&dst_mm
->page_table_lock
);
782 static int do_huge_pmd_wp_page_fallback(struct mm_struct
*mm
,
783 struct vm_area_struct
*vma
,
784 unsigned long address
,
785 pmd_t
*pmd
, pmd_t orig_pmd
,
793 unsigned long mmun_start
; /* For mmu_notifiers */
794 unsigned long mmun_end
; /* For mmu_notifiers */
796 pages
= kmalloc(sizeof(struct page
*) * HPAGE_PMD_NR
,
798 if (unlikely(!pages
)) {
803 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
804 pages
[i
] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE
|
806 vma
, address
, page_to_nid(page
));
807 if (unlikely(!pages
[i
] ||
808 mem_cgroup_newpage_charge(pages
[i
], mm
,
812 mem_cgroup_uncharge_start();
814 mem_cgroup_uncharge_page(pages
[i
]);
817 mem_cgroup_uncharge_end();
824 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
825 copy_user_highpage(pages
[i
], page
+ i
,
826 haddr
+ PAGE_SIZE
* i
, vma
);
827 __SetPageUptodate(pages
[i
]);
832 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
833 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
835 spin_lock(&mm
->page_table_lock
);
836 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
838 VM_BUG_ON(!PageHead(page
));
840 pmdp_clear_flush(vma
, haddr
, pmd
);
841 /* leave pmd empty until pte is filled */
843 pgtable
= pgtable_trans_huge_withdraw(mm
);
844 pmd_populate(mm
, &_pmd
, pgtable
);
846 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
848 entry
= mk_pte(pages
[i
], vma
->vm_page_prot
);
849 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
850 page_add_new_anon_rmap(pages
[i
], vma
, haddr
);
851 pte
= pte_offset_map(&_pmd
, haddr
);
852 VM_BUG_ON(!pte_none(*pte
));
853 set_pte_at(mm
, haddr
, pte
, entry
);
858 smp_wmb(); /* make pte visible before pmd */
859 pmd_populate(mm
, pmd
, pgtable
);
860 page_remove_rmap(page
);
861 spin_unlock(&mm
->page_table_lock
);
863 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
865 ret
|= VM_FAULT_WRITE
;
872 spin_unlock(&mm
->page_table_lock
);
873 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
874 mem_cgroup_uncharge_start();
875 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
876 mem_cgroup_uncharge_page(pages
[i
]);
879 mem_cgroup_uncharge_end();
884 int do_huge_pmd_wp_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
885 unsigned long address
, pmd_t
*pmd
, pmd_t orig_pmd
)
888 struct page
*page
, *new_page
;
890 unsigned long mmun_start
; /* For mmu_notifiers */
891 unsigned long mmun_end
; /* For mmu_notifiers */
893 VM_BUG_ON(!vma
->anon_vma
);
894 spin_lock(&mm
->page_table_lock
);
895 if (unlikely(!pmd_same(*pmd
, orig_pmd
)))
898 page
= pmd_page(orig_pmd
);
899 VM_BUG_ON(!PageCompound(page
) || !PageHead(page
));
900 haddr
= address
& HPAGE_PMD_MASK
;
901 if (page_mapcount(page
) == 1) {
903 entry
= pmd_mkyoung(orig_pmd
);
904 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
905 if (pmdp_set_access_flags(vma
, haddr
, pmd
, entry
, 1))
906 update_mmu_cache_pmd(vma
, address
, pmd
);
907 ret
|= VM_FAULT_WRITE
;
911 spin_unlock(&mm
->page_table_lock
);
913 if (transparent_hugepage_enabled(vma
) &&
914 !transparent_hugepage_debug_cow())
915 new_page
= alloc_hugepage_vma(transparent_hugepage_defrag(vma
),
916 vma
, haddr
, numa_node_id(), 0);
920 if (unlikely(!new_page
)) {
921 count_vm_event(THP_FAULT_FALLBACK
);
922 ret
= do_huge_pmd_wp_page_fallback(mm
, vma
, address
,
923 pmd
, orig_pmd
, page
, haddr
);
924 if (ret
& VM_FAULT_OOM
)
925 split_huge_page(page
);
929 count_vm_event(THP_FAULT_ALLOC
);
931 if (unlikely(mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
))) {
933 split_huge_page(page
);
939 copy_user_huge_page(new_page
, page
, haddr
, vma
, HPAGE_PMD_NR
);
940 __SetPageUptodate(new_page
);
943 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
944 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
946 spin_lock(&mm
->page_table_lock
);
948 if (unlikely(!pmd_same(*pmd
, orig_pmd
))) {
949 spin_unlock(&mm
->page_table_lock
);
950 mem_cgroup_uncharge_page(new_page
);
955 VM_BUG_ON(!PageHead(page
));
956 entry
= mk_pmd(new_page
, vma
->vm_page_prot
);
957 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
958 entry
= pmd_mkhuge(entry
);
959 pmdp_clear_flush(vma
, haddr
, pmd
);
960 page_add_new_anon_rmap(new_page
, vma
, haddr
);
961 set_pmd_at(mm
, haddr
, pmd
, entry
);
962 update_mmu_cache_pmd(vma
, address
, pmd
);
963 page_remove_rmap(page
);
965 ret
|= VM_FAULT_WRITE
;
967 spin_unlock(&mm
->page_table_lock
);
969 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
973 spin_unlock(&mm
->page_table_lock
);
977 struct page
*follow_trans_huge_pmd(struct vm_area_struct
*vma
,
982 struct mm_struct
*mm
= vma
->vm_mm
;
983 struct page
*page
= NULL
;
985 assert_spin_locked(&mm
->page_table_lock
);
987 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
990 page
= pmd_page(*pmd
);
991 VM_BUG_ON(!PageHead(page
));
992 if (flags
& FOLL_TOUCH
) {
995 * We should set the dirty bit only for FOLL_WRITE but
996 * for now the dirty bit in the pmd is meaningless.
997 * And if the dirty bit will become meaningful and
998 * we'll only set it with FOLL_WRITE, an atomic
999 * set_bit will be required on the pmd to set the
1000 * young bit, instead of the current set_pmd_at.
1002 _pmd
= pmd_mkyoung(pmd_mkdirty(*pmd
));
1003 set_pmd_at(mm
, addr
& HPAGE_PMD_MASK
, pmd
, _pmd
);
1005 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1006 if (page
->mapping
&& trylock_page(page
)) {
1009 mlock_vma_page(page
);
1013 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
1014 VM_BUG_ON(!PageCompound(page
));
1015 if (flags
& FOLL_GET
)
1016 get_page_foll(page
);
1022 /* NUMA hinting page fault entry point for trans huge pmds */
1023 int do_huge_pmd_numa_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1024 unsigned long addr
, pmd_t pmd
, pmd_t
*pmdp
)
1026 struct page
*page
= NULL
;
1027 unsigned long haddr
= addr
& HPAGE_PMD_MASK
;
1029 int current_nid
= -1;
1031 spin_lock(&mm
->page_table_lock
);
1032 if (unlikely(!pmd_same(pmd
, *pmdp
)))
1035 page
= pmd_page(pmd
);
1037 spin_unlock(&mm
->page_table_lock
);
1038 current_nid
= page_to_nid(page
);
1039 count_vm_numa_event(NUMA_HINT_FAULTS
);
1040 if (current_nid
== numa_node_id())
1041 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
1043 target_nid
= mpol_misplaced(page
, vma
, haddr
);
1044 if (target_nid
== -1)
1048 * Due to lacking code to migrate thp pages, we'll split
1049 * (which preserves the special PROT_NONE) and re-take the
1050 * fault on the normal pages.
1052 split_huge_page(page
);
1058 spin_lock(&mm
->page_table_lock
);
1059 if (unlikely(!pmd_same(pmd
, *pmdp
)))
1062 pmd
= pmd_mknonnuma(pmd
);
1063 set_pmd_at(mm
, haddr
, pmdp
, pmd
);
1064 VM_BUG_ON(pmd_numa(*pmdp
));
1065 update_mmu_cache_pmd(vma
, addr
, pmdp
);
1068 spin_unlock(&mm
->page_table_lock
);
1071 task_numa_fault(numa_node_id(), HPAGE_PMD_NR
);
1076 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1077 pmd_t
*pmd
, unsigned long addr
)
1081 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
1085 pgtable
= pgtable_trans_huge_withdraw(tlb
->mm
);
1086 orig_pmd
= pmdp_get_and_clear(tlb
->mm
, addr
, pmd
);
1087 page
= pmd_page(orig_pmd
);
1088 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1089 page_remove_rmap(page
);
1090 VM_BUG_ON(page_mapcount(page
) < 0);
1091 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1092 VM_BUG_ON(!PageHead(page
));
1094 spin_unlock(&tlb
->mm
->page_table_lock
);
1095 tlb_remove_page(tlb
, page
);
1096 pte_free(tlb
->mm
, pgtable
);
1102 int mincore_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1103 unsigned long addr
, unsigned long end
,
1108 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
1110 * All logical pages in the range are present
1111 * if backed by a huge page.
1113 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1114 memset(vec
, 1, (end
- addr
) >> PAGE_SHIFT
);
1121 int move_huge_pmd(struct vm_area_struct
*vma
, struct vm_area_struct
*new_vma
,
1122 unsigned long old_addr
,
1123 unsigned long new_addr
, unsigned long old_end
,
1124 pmd_t
*old_pmd
, pmd_t
*new_pmd
)
1129 struct mm_struct
*mm
= vma
->vm_mm
;
1131 if ((old_addr
& ~HPAGE_PMD_MASK
) ||
1132 (new_addr
& ~HPAGE_PMD_MASK
) ||
1133 old_end
- old_addr
< HPAGE_PMD_SIZE
||
1134 (new_vma
->vm_flags
& VM_NOHUGEPAGE
))
1138 * The destination pmd shouldn't be established, free_pgtables()
1139 * should have release it.
1141 if (WARN_ON(!pmd_none(*new_pmd
))) {
1142 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1146 ret
= __pmd_trans_huge_lock(old_pmd
, vma
);
1148 pmd
= pmdp_get_and_clear(mm
, old_addr
, old_pmd
);
1149 VM_BUG_ON(!pmd_none(*new_pmd
));
1150 set_pmd_at(mm
, new_addr
, new_pmd
, pmd
);
1151 spin_unlock(&mm
->page_table_lock
);
1157 int change_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1158 unsigned long addr
, pgprot_t newprot
, int prot_numa
)
1160 struct mm_struct
*mm
= vma
->vm_mm
;
1163 if (__pmd_trans_huge_lock(pmd
, vma
) == 1) {
1165 entry
= pmdp_get_and_clear(mm
, addr
, pmd
);
1167 entry
= pmd_modify(entry
, newprot
);
1169 struct page
*page
= pmd_page(*pmd
);
1171 /* only check non-shared pages */
1172 if (page_mapcount(page
) == 1 &&
1174 entry
= pmd_mknuma(entry
);
1177 set_pmd_at(mm
, addr
, pmd
, entry
);
1178 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1186 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1187 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1189 * Note that if it returns 1, this routine returns without unlocking page
1190 * table locks. So callers must unlock them.
1192 int __pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
)
1194 spin_lock(&vma
->vm_mm
->page_table_lock
);
1195 if (likely(pmd_trans_huge(*pmd
))) {
1196 if (unlikely(pmd_trans_splitting(*pmd
))) {
1197 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1198 wait_split_huge_page(vma
->anon_vma
, pmd
);
1201 /* Thp mapped by 'pmd' is stable, so we can
1202 * handle it as it is. */
1206 spin_unlock(&vma
->vm_mm
->page_table_lock
);
1210 pmd_t
*page_check_address_pmd(struct page
*page
,
1211 struct mm_struct
*mm
,
1212 unsigned long address
,
1213 enum page_check_address_pmd_flag flag
)
1217 pmd_t
*pmd
, *ret
= NULL
;
1219 if (address
& ~HPAGE_PMD_MASK
)
1222 pgd
= pgd_offset(mm
, address
);
1223 if (!pgd_present(*pgd
))
1226 pud
= pud_offset(pgd
, address
);
1227 if (!pud_present(*pud
))
1230 pmd
= pmd_offset(pud
, address
);
1233 if (pmd_page(*pmd
) != page
)
1236 * split_vma() may create temporary aliased mappings. There is
1237 * no risk as long as all huge pmd are found and have their
1238 * splitting bit set before __split_huge_page_refcount
1239 * runs. Finding the same huge pmd more than once during the
1240 * same rmap walk is not a problem.
1242 if (flag
== PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG
&&
1243 pmd_trans_splitting(*pmd
))
1245 if (pmd_trans_huge(*pmd
)) {
1246 VM_BUG_ON(flag
== PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG
&&
1247 !pmd_trans_splitting(*pmd
));
1254 static int __split_huge_page_splitting(struct page
*page
,
1255 struct vm_area_struct
*vma
,
1256 unsigned long address
)
1258 struct mm_struct
*mm
= vma
->vm_mm
;
1261 /* For mmu_notifiers */
1262 const unsigned long mmun_start
= address
;
1263 const unsigned long mmun_end
= address
+ HPAGE_PMD_SIZE
;
1265 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1266 spin_lock(&mm
->page_table_lock
);
1267 pmd
= page_check_address_pmd(page
, mm
, address
,
1268 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG
);
1271 * We can't temporarily set the pmd to null in order
1272 * to split it, the pmd must remain marked huge at all
1273 * times or the VM won't take the pmd_trans_huge paths
1274 * and it won't wait on the anon_vma->root->mutex to
1275 * serialize against split_huge_page*.
1277 pmdp_splitting_flush(vma
, address
, pmd
);
1280 spin_unlock(&mm
->page_table_lock
);
1281 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1286 static void __split_huge_page_refcount(struct page
*page
)
1289 struct zone
*zone
= page_zone(page
);
1290 struct lruvec
*lruvec
;
1293 /* prevent PageLRU to go away from under us, and freeze lru stats */
1294 spin_lock_irq(&zone
->lru_lock
);
1295 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1297 compound_lock(page
);
1298 /* complete memcg works before add pages to LRU */
1299 mem_cgroup_split_huge_fixup(page
);
1301 for (i
= HPAGE_PMD_NR
- 1; i
>= 1; i
--) {
1302 struct page
*page_tail
= page
+ i
;
1304 /* tail_page->_mapcount cannot change */
1305 BUG_ON(page_mapcount(page_tail
) < 0);
1306 tail_count
+= page_mapcount(page_tail
);
1307 /* check for overflow */
1308 BUG_ON(tail_count
< 0);
1309 BUG_ON(atomic_read(&page_tail
->_count
) != 0);
1311 * tail_page->_count is zero and not changing from
1312 * under us. But get_page_unless_zero() may be running
1313 * from under us on the tail_page. If we used
1314 * atomic_set() below instead of atomic_add(), we
1315 * would then run atomic_set() concurrently with
1316 * get_page_unless_zero(), and atomic_set() is
1317 * implemented in C not using locked ops. spin_unlock
1318 * on x86 sometime uses locked ops because of PPro
1319 * errata 66, 92, so unless somebody can guarantee
1320 * atomic_set() here would be safe on all archs (and
1321 * not only on x86), it's safer to use atomic_add().
1323 atomic_add(page_mapcount(page
) + page_mapcount(page_tail
) + 1,
1324 &page_tail
->_count
);
1326 /* after clearing PageTail the gup refcount can be released */
1330 * retain hwpoison flag of the poisoned tail page:
1331 * fix for the unsuitable process killed on Guest Machine(KVM)
1332 * by the memory-failure.
1334 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
| __PG_HWPOISON
;
1335 page_tail
->flags
|= (page
->flags
&
1336 ((1L << PG_referenced
) |
1337 (1L << PG_swapbacked
) |
1338 (1L << PG_mlocked
) |
1339 (1L << PG_uptodate
)));
1340 page_tail
->flags
|= (1L << PG_dirty
);
1342 /* clear PageTail before overwriting first_page */
1346 * __split_huge_page_splitting() already set the
1347 * splitting bit in all pmd that could map this
1348 * hugepage, that will ensure no CPU can alter the
1349 * mapcount on the head page. The mapcount is only
1350 * accounted in the head page and it has to be
1351 * transferred to all tail pages in the below code. So
1352 * for this code to be safe, the split the mapcount
1353 * can't change. But that doesn't mean userland can't
1354 * keep changing and reading the page contents while
1355 * we transfer the mapcount, so the pmd splitting
1356 * status is achieved setting a reserved bit in the
1357 * pmd, not by clearing the present bit.
1359 page_tail
->_mapcount
= page
->_mapcount
;
1361 BUG_ON(page_tail
->mapping
);
1362 page_tail
->mapping
= page
->mapping
;
1364 page_tail
->index
= page
->index
+ i
;
1366 BUG_ON(!PageAnon(page_tail
));
1367 BUG_ON(!PageUptodate(page_tail
));
1368 BUG_ON(!PageDirty(page_tail
));
1369 BUG_ON(!PageSwapBacked(page_tail
));
1371 lru_add_page_tail(page
, page_tail
, lruvec
);
1373 atomic_sub(tail_count
, &page
->_count
);
1374 BUG_ON(atomic_read(&page
->_count
) <= 0);
1376 __mod_zone_page_state(zone
, NR_ANON_TRANSPARENT_HUGEPAGES
, -1);
1377 __mod_zone_page_state(zone
, NR_ANON_PAGES
, HPAGE_PMD_NR
);
1379 ClearPageCompound(page
);
1380 compound_unlock(page
);
1381 spin_unlock_irq(&zone
->lru_lock
);
1383 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
1384 struct page
*page_tail
= page
+ i
;
1385 BUG_ON(page_count(page_tail
) <= 0);
1387 * Tail pages may be freed if there wasn't any mapping
1388 * like if add_to_swap() is running on a lru page that
1389 * had its mapping zapped. And freeing these pages
1390 * requires taking the lru_lock so we do the put_page
1391 * of the tail pages after the split is complete.
1393 put_page(page_tail
);
1397 * Only the head page (now become a regular page) is required
1398 * to be pinned by the caller.
1400 BUG_ON(page_count(page
) <= 0);
1403 static int __split_huge_page_map(struct page
*page
,
1404 struct vm_area_struct
*vma
,
1405 unsigned long address
)
1407 struct mm_struct
*mm
= vma
->vm_mm
;
1411 unsigned long haddr
;
1413 spin_lock(&mm
->page_table_lock
);
1414 pmd
= page_check_address_pmd(page
, mm
, address
,
1415 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG
);
1417 pgtable
= pgtable_trans_huge_withdraw(mm
);
1418 pmd_populate(mm
, &_pmd
, pgtable
);
1421 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1423 BUG_ON(PageCompound(page
+i
));
1424 entry
= mk_pte(page
+ i
, vma
->vm_page_prot
);
1425 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1426 if (!pmd_write(*pmd
))
1427 entry
= pte_wrprotect(entry
);
1429 BUG_ON(page_mapcount(page
) != 1);
1430 if (!pmd_young(*pmd
))
1431 entry
= pte_mkold(entry
);
1433 entry
= pte_mknuma(entry
);
1434 pte
= pte_offset_map(&_pmd
, haddr
);
1435 BUG_ON(!pte_none(*pte
));
1436 set_pte_at(mm
, haddr
, pte
, entry
);
1440 smp_wmb(); /* make pte visible before pmd */
1442 * Up to this point the pmd is present and huge and
1443 * userland has the whole access to the hugepage
1444 * during the split (which happens in place). If we
1445 * overwrite the pmd with the not-huge version
1446 * pointing to the pte here (which of course we could
1447 * if all CPUs were bug free), userland could trigger
1448 * a small page size TLB miss on the small sized TLB
1449 * while the hugepage TLB entry is still established
1450 * in the huge TLB. Some CPU doesn't like that. See
1451 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1452 * Erratum 383 on page 93. Intel should be safe but is
1453 * also warns that it's only safe if the permission
1454 * and cache attributes of the two entries loaded in
1455 * the two TLB is identical (which should be the case
1456 * here). But it is generally safer to never allow
1457 * small and huge TLB entries for the same virtual
1458 * address to be loaded simultaneously. So instead of
1459 * doing "pmd_populate(); flush_tlb_range();" we first
1460 * mark the current pmd notpresent (atomically because
1461 * here the pmd_trans_huge and pmd_trans_splitting
1462 * must remain set at all times on the pmd until the
1463 * split is complete for this pmd), then we flush the
1464 * SMP TLB and finally we write the non-huge version
1465 * of the pmd entry with pmd_populate.
1467 pmdp_invalidate(vma
, address
, pmd
);
1468 pmd_populate(mm
, pmd
, pgtable
);
1471 spin_unlock(&mm
->page_table_lock
);
1476 /* must be called with anon_vma->root->mutex hold */
1477 static void __split_huge_page(struct page
*page
,
1478 struct anon_vma
*anon_vma
)
1480 int mapcount
, mapcount2
;
1481 pgoff_t pgoff
= page
->index
<< (PAGE_CACHE_SHIFT
- PAGE_SHIFT
);
1482 struct anon_vma_chain
*avc
;
1484 BUG_ON(!PageHead(page
));
1485 BUG_ON(PageTail(page
));
1488 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
1489 struct vm_area_struct
*vma
= avc
->vma
;
1490 unsigned long addr
= vma_address(page
, vma
);
1491 BUG_ON(is_vma_temporary_stack(vma
));
1492 mapcount
+= __split_huge_page_splitting(page
, vma
, addr
);
1495 * It is critical that new vmas are added to the tail of the
1496 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1497 * and establishes a child pmd before
1498 * __split_huge_page_splitting() freezes the parent pmd (so if
1499 * we fail to prevent copy_huge_pmd() from running until the
1500 * whole __split_huge_page() is complete), we will still see
1501 * the newly established pmd of the child later during the
1502 * walk, to be able to set it as pmd_trans_splitting too.
1504 if (mapcount
!= page_mapcount(page
))
1505 printk(KERN_ERR
"mapcount %d page_mapcount %d\n",
1506 mapcount
, page_mapcount(page
));
1507 BUG_ON(mapcount
!= page_mapcount(page
));
1509 __split_huge_page_refcount(page
);
1512 anon_vma_interval_tree_foreach(avc
, &anon_vma
->rb_root
, pgoff
, pgoff
) {
1513 struct vm_area_struct
*vma
= avc
->vma
;
1514 unsigned long addr
= vma_address(page
, vma
);
1515 BUG_ON(is_vma_temporary_stack(vma
));
1516 mapcount2
+= __split_huge_page_map(page
, vma
, addr
);
1518 if (mapcount
!= mapcount2
)
1519 printk(KERN_ERR
"mapcount %d mapcount2 %d page_mapcount %d\n",
1520 mapcount
, mapcount2
, page_mapcount(page
));
1521 BUG_ON(mapcount
!= mapcount2
);
1524 int split_huge_page(struct page
*page
)
1526 struct anon_vma
*anon_vma
;
1529 BUG_ON(!PageAnon(page
));
1530 anon_vma
= page_lock_anon_vma(page
);
1534 if (!PageCompound(page
))
1537 BUG_ON(!PageSwapBacked(page
));
1538 __split_huge_page(page
, anon_vma
);
1539 count_vm_event(THP_SPLIT
);
1541 BUG_ON(PageCompound(page
));
1543 page_unlock_anon_vma(anon_vma
);
1548 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1550 int hugepage_madvise(struct vm_area_struct
*vma
,
1551 unsigned long *vm_flags
, int advice
)
1553 struct mm_struct
*mm
= vma
->vm_mm
;
1558 * Be somewhat over-protective like KSM for now!
1560 if (*vm_flags
& (VM_HUGEPAGE
| VM_NO_THP
))
1562 if (mm
->def_flags
& VM_NOHUGEPAGE
)
1564 *vm_flags
&= ~VM_NOHUGEPAGE
;
1565 *vm_flags
|= VM_HUGEPAGE
;
1567 * If the vma become good for khugepaged to scan,
1568 * register it here without waiting a page fault that
1569 * may not happen any time soon.
1571 if (unlikely(khugepaged_enter_vma_merge(vma
)))
1574 case MADV_NOHUGEPAGE
:
1576 * Be somewhat over-protective like KSM for now!
1578 if (*vm_flags
& (VM_NOHUGEPAGE
| VM_NO_THP
))
1580 *vm_flags
&= ~VM_HUGEPAGE
;
1581 *vm_flags
|= VM_NOHUGEPAGE
;
1583 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1584 * this vma even if we leave the mm registered in khugepaged if
1585 * it got registered before VM_NOHUGEPAGE was set.
1593 static int __init
khugepaged_slab_init(void)
1595 mm_slot_cache
= kmem_cache_create("khugepaged_mm_slot",
1596 sizeof(struct mm_slot
),
1597 __alignof__(struct mm_slot
), 0, NULL
);
1604 static void __init
khugepaged_slab_free(void)
1606 kmem_cache_destroy(mm_slot_cache
);
1607 mm_slot_cache
= NULL
;
1610 static inline struct mm_slot
*alloc_mm_slot(void)
1612 if (!mm_slot_cache
) /* initialization failed */
1614 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
1617 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
1619 kmem_cache_free(mm_slot_cache
, mm_slot
);
1622 static int __init
mm_slots_hash_init(void)
1624 mm_slots_hash
= kzalloc(MM_SLOTS_HASH_HEADS
* sizeof(struct hlist_head
),
1632 static void __init
mm_slots_hash_free(void)
1634 kfree(mm_slots_hash
);
1635 mm_slots_hash
= NULL
;
1639 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
1641 struct mm_slot
*mm_slot
;
1642 struct hlist_head
*bucket
;
1643 struct hlist_node
*node
;
1645 bucket
= &mm_slots_hash
[((unsigned long)mm
/ sizeof(struct mm_struct
))
1646 % MM_SLOTS_HASH_HEADS
];
1647 hlist_for_each_entry(mm_slot
, node
, bucket
, hash
) {
1648 if (mm
== mm_slot
->mm
)
1654 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
1655 struct mm_slot
*mm_slot
)
1657 struct hlist_head
*bucket
;
1659 bucket
= &mm_slots_hash
[((unsigned long)mm
/ sizeof(struct mm_struct
))
1660 % MM_SLOTS_HASH_HEADS
];
1662 hlist_add_head(&mm_slot
->hash
, bucket
);
1665 static inline int khugepaged_test_exit(struct mm_struct
*mm
)
1667 return atomic_read(&mm
->mm_users
) == 0;
1670 int __khugepaged_enter(struct mm_struct
*mm
)
1672 struct mm_slot
*mm_slot
;
1675 mm_slot
= alloc_mm_slot();
1679 /* __khugepaged_exit() must not run from under us */
1680 VM_BUG_ON(khugepaged_test_exit(mm
));
1681 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE
, &mm
->flags
))) {
1682 free_mm_slot(mm_slot
);
1686 spin_lock(&khugepaged_mm_lock
);
1687 insert_to_mm_slots_hash(mm
, mm_slot
);
1689 * Insert just behind the scanning cursor, to let the area settle
1692 wakeup
= list_empty(&khugepaged_scan
.mm_head
);
1693 list_add_tail(&mm_slot
->mm_node
, &khugepaged_scan
.mm_head
);
1694 spin_unlock(&khugepaged_mm_lock
);
1696 atomic_inc(&mm
->mm_count
);
1698 wake_up_interruptible(&khugepaged_wait
);
1703 int khugepaged_enter_vma_merge(struct vm_area_struct
*vma
)
1705 unsigned long hstart
, hend
;
1708 * Not yet faulted in so we will register later in the
1709 * page fault if needed.
1713 /* khugepaged not yet working on file or special mappings */
1715 VM_BUG_ON(vma
->vm_flags
& VM_NO_THP
);
1716 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
1717 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
1719 return khugepaged_enter(vma
);
1723 void __khugepaged_exit(struct mm_struct
*mm
)
1725 struct mm_slot
*mm_slot
;
1728 spin_lock(&khugepaged_mm_lock
);
1729 mm_slot
= get_mm_slot(mm
);
1730 if (mm_slot
&& khugepaged_scan
.mm_slot
!= mm_slot
) {
1731 hlist_del(&mm_slot
->hash
);
1732 list_del(&mm_slot
->mm_node
);
1735 spin_unlock(&khugepaged_mm_lock
);
1738 clear_bit(MMF_VM_HUGEPAGE
, &mm
->flags
);
1739 free_mm_slot(mm_slot
);
1741 } else if (mm_slot
) {
1743 * This is required to serialize against
1744 * khugepaged_test_exit() (which is guaranteed to run
1745 * under mmap sem read mode). Stop here (after we
1746 * return all pagetables will be destroyed) until
1747 * khugepaged has finished working on the pagetables
1748 * under the mmap_sem.
1750 down_write(&mm
->mmap_sem
);
1751 up_write(&mm
->mmap_sem
);
1755 static void release_pte_page(struct page
*page
)
1757 /* 0 stands for page_is_file_cache(page) == false */
1758 dec_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
1760 putback_lru_page(page
);
1763 static void release_pte_pages(pte_t
*pte
, pte_t
*_pte
)
1765 while (--_pte
>= pte
) {
1766 pte_t pteval
= *_pte
;
1767 if (!pte_none(pteval
))
1768 release_pte_page(pte_page(pteval
));
1772 static void release_all_pte_pages(pte_t
*pte
)
1774 release_pte_pages(pte
, pte
+ HPAGE_PMD_NR
);
1777 static int __collapse_huge_page_isolate(struct vm_area_struct
*vma
,
1778 unsigned long address
,
1783 int referenced
= 0, isolated
= 0, none
= 0;
1784 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
1785 _pte
++, address
+= PAGE_SIZE
) {
1786 pte_t pteval
= *_pte
;
1787 if (pte_none(pteval
)) {
1788 if (++none
<= khugepaged_max_ptes_none
)
1791 release_pte_pages(pte
, _pte
);
1795 if (!pte_present(pteval
) || !pte_write(pteval
)) {
1796 release_pte_pages(pte
, _pte
);
1799 page
= vm_normal_page(vma
, address
, pteval
);
1800 if (unlikely(!page
)) {
1801 release_pte_pages(pte
, _pte
);
1804 VM_BUG_ON(PageCompound(page
));
1805 BUG_ON(!PageAnon(page
));
1806 VM_BUG_ON(!PageSwapBacked(page
));
1808 /* cannot use mapcount: can't collapse if there's a gup pin */
1809 if (page_count(page
) != 1) {
1810 release_pte_pages(pte
, _pte
);
1814 * We can do it before isolate_lru_page because the
1815 * page can't be freed from under us. NOTE: PG_lock
1816 * is needed to serialize against split_huge_page
1817 * when invoked from the VM.
1819 if (!trylock_page(page
)) {
1820 release_pte_pages(pte
, _pte
);
1824 * Isolate the page to avoid collapsing an hugepage
1825 * currently in use by the VM.
1827 if (isolate_lru_page(page
)) {
1829 release_pte_pages(pte
, _pte
);
1832 /* 0 stands for page_is_file_cache(page) == false */
1833 inc_zone_page_state(page
, NR_ISOLATED_ANON
+ 0);
1834 VM_BUG_ON(!PageLocked(page
));
1835 VM_BUG_ON(PageLRU(page
));
1837 /* If there is no mapped pte young don't collapse the page */
1838 if (pte_young(pteval
) || PageReferenced(page
) ||
1839 mmu_notifier_test_young(vma
->vm_mm
, address
))
1842 if (unlikely(!referenced
))
1843 release_all_pte_pages(pte
);
1850 static void __collapse_huge_page_copy(pte_t
*pte
, struct page
*page
,
1851 struct vm_area_struct
*vma
,
1852 unsigned long address
,
1856 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
; _pte
++) {
1857 pte_t pteval
= *_pte
;
1858 struct page
*src_page
;
1860 if (pte_none(pteval
)) {
1861 clear_user_highpage(page
, address
);
1862 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, 1);
1864 src_page
= pte_page(pteval
);
1865 copy_user_highpage(page
, src_page
, address
, vma
);
1866 VM_BUG_ON(page_mapcount(src_page
) != 1);
1867 release_pte_page(src_page
);
1869 * ptl mostly unnecessary, but preempt has to
1870 * be disabled to update the per-cpu stats
1871 * inside page_remove_rmap().
1875 * paravirt calls inside pte_clear here are
1878 pte_clear(vma
->vm_mm
, address
, _pte
);
1879 page_remove_rmap(src_page
);
1881 free_page_and_swap_cache(src_page
);
1884 address
+= PAGE_SIZE
;
1889 static void khugepaged_alloc_sleep(void)
1891 wait_event_freezable_timeout(khugepaged_wait
, false,
1892 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs
));
1896 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
1898 if (IS_ERR(*hpage
)) {
1904 khugepaged_alloc_sleep();
1905 } else if (*hpage
) {
1914 *khugepaged_alloc_page(struct page
**hpage
, struct mm_struct
*mm
,
1915 struct vm_area_struct
*vma
, unsigned long address
,
1920 * Allocate the page while the vma is still valid and under
1921 * the mmap_sem read mode so there is no memory allocation
1922 * later when we take the mmap_sem in write mode. This is more
1923 * friendly behavior (OTOH it may actually hide bugs) to
1924 * filesystems in userland with daemons allocating memory in
1925 * the userland I/O paths. Allocating memory with the
1926 * mmap_sem in read mode is good idea also to allow greater
1929 *hpage
= alloc_hugepage_vma(khugepaged_defrag(), vma
, address
,
1930 node
, __GFP_OTHER_NODE
);
1933 * After allocating the hugepage, release the mmap_sem read lock in
1934 * preparation for taking it in write mode.
1936 up_read(&mm
->mmap_sem
);
1937 if (unlikely(!*hpage
)) {
1938 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
1939 *hpage
= ERR_PTR(-ENOMEM
);
1943 count_vm_event(THP_COLLAPSE_ALLOC
);
1947 static struct page
*khugepaged_alloc_hugepage(bool *wait
)
1952 hpage
= alloc_hugepage(khugepaged_defrag());
1954 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
1959 khugepaged_alloc_sleep();
1961 count_vm_event(THP_COLLAPSE_ALLOC
);
1962 } while (unlikely(!hpage
) && likely(khugepaged_enabled()));
1967 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
1970 *hpage
= khugepaged_alloc_hugepage(wait
);
1972 if (unlikely(!*hpage
))
1979 *khugepaged_alloc_page(struct page
**hpage
, struct mm_struct
*mm
,
1980 struct vm_area_struct
*vma
, unsigned long address
,
1983 up_read(&mm
->mmap_sem
);
1989 static void collapse_huge_page(struct mm_struct
*mm
,
1990 unsigned long address
,
1991 struct page
**hpage
,
1992 struct vm_area_struct
*vma
,
2000 struct page
*new_page
;
2003 unsigned long hstart
, hend
;
2004 unsigned long mmun_start
; /* For mmu_notifiers */
2005 unsigned long mmun_end
; /* For mmu_notifiers */
2007 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2009 /* release the mmap_sem read lock. */
2010 new_page
= khugepaged_alloc_page(hpage
, mm
, vma
, address
, node
);
2014 if (unlikely(mem_cgroup_newpage_charge(new_page
, mm
, GFP_KERNEL
)))
2018 * Prevent all access to pagetables with the exception of
2019 * gup_fast later hanlded by the ptep_clear_flush and the VM
2020 * handled by the anon_vma lock + PG_lock.
2022 down_write(&mm
->mmap_sem
);
2023 if (unlikely(khugepaged_test_exit(mm
)))
2026 vma
= find_vma(mm
, address
);
2027 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2028 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2029 if (address
< hstart
|| address
+ HPAGE_PMD_SIZE
> hend
)
2032 if ((!(vma
->vm_flags
& VM_HUGEPAGE
) && !khugepaged_always()) ||
2033 (vma
->vm_flags
& VM_NOHUGEPAGE
))
2036 if (!vma
->anon_vma
|| vma
->vm_ops
)
2038 if (is_vma_temporary_stack(vma
))
2040 VM_BUG_ON(vma
->vm_flags
& VM_NO_THP
);
2042 pgd
= pgd_offset(mm
, address
);
2043 if (!pgd_present(*pgd
))
2046 pud
= pud_offset(pgd
, address
);
2047 if (!pud_present(*pud
))
2050 pmd
= pmd_offset(pud
, address
);
2051 /* pmd can't go away or become huge under us */
2052 if (!pmd_present(*pmd
) || pmd_trans_huge(*pmd
))
2055 anon_vma_lock(vma
->anon_vma
);
2057 pte
= pte_offset_map(pmd
, address
);
2058 ptl
= pte_lockptr(mm
, pmd
);
2060 mmun_start
= address
;
2061 mmun_end
= address
+ HPAGE_PMD_SIZE
;
2062 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2063 spin_lock(&mm
->page_table_lock
); /* probably unnecessary */
2065 * After this gup_fast can't run anymore. This also removes
2066 * any huge TLB entry from the CPU so we won't allow
2067 * huge and small TLB entries for the same virtual address
2068 * to avoid the risk of CPU bugs in that area.
2070 _pmd
= pmdp_clear_flush(vma
, address
, pmd
);
2071 spin_unlock(&mm
->page_table_lock
);
2072 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2075 isolated
= __collapse_huge_page_isolate(vma
, address
, pte
);
2078 if (unlikely(!isolated
)) {
2080 spin_lock(&mm
->page_table_lock
);
2081 BUG_ON(!pmd_none(*pmd
));
2082 set_pmd_at(mm
, address
, pmd
, _pmd
);
2083 spin_unlock(&mm
->page_table_lock
);
2084 anon_vma_unlock(vma
->anon_vma
);
2089 * All pages are isolated and locked so anon_vma rmap
2090 * can't run anymore.
2092 anon_vma_unlock(vma
->anon_vma
);
2094 __collapse_huge_page_copy(pte
, new_page
, vma
, address
, ptl
);
2096 __SetPageUptodate(new_page
);
2097 pgtable
= pmd_pgtable(_pmd
);
2099 _pmd
= mk_pmd(new_page
, vma
->vm_page_prot
);
2100 _pmd
= maybe_pmd_mkwrite(pmd_mkdirty(_pmd
), vma
);
2101 _pmd
= pmd_mkhuge(_pmd
);
2104 * spin_lock() below is not the equivalent of smp_wmb(), so
2105 * this is needed to avoid the copy_huge_page writes to become
2106 * visible after the set_pmd_at() write.
2110 spin_lock(&mm
->page_table_lock
);
2111 BUG_ON(!pmd_none(*pmd
));
2112 page_add_new_anon_rmap(new_page
, vma
, address
);
2113 set_pmd_at(mm
, address
, pmd
, _pmd
);
2114 update_mmu_cache_pmd(vma
, address
, pmd
);
2115 pgtable_trans_huge_deposit(mm
, pgtable
);
2116 spin_unlock(&mm
->page_table_lock
);
2120 khugepaged_pages_collapsed
++;
2122 up_write(&mm
->mmap_sem
);
2126 mem_cgroup_uncharge_page(new_page
);
2130 static int khugepaged_scan_pmd(struct mm_struct
*mm
,
2131 struct vm_area_struct
*vma
,
2132 unsigned long address
,
2133 struct page
**hpage
)
2139 int ret
= 0, referenced
= 0, none
= 0;
2141 unsigned long _address
;
2145 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
2147 pgd
= pgd_offset(mm
, address
);
2148 if (!pgd_present(*pgd
))
2151 pud
= pud_offset(pgd
, address
);
2152 if (!pud_present(*pud
))
2155 pmd
= pmd_offset(pud
, address
);
2156 if (!pmd_present(*pmd
) || pmd_trans_huge(*pmd
))
2159 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
2160 for (_address
= address
, _pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
2161 _pte
++, _address
+= PAGE_SIZE
) {
2162 pte_t pteval
= *_pte
;
2163 if (pte_none(pteval
)) {
2164 if (++none
<= khugepaged_max_ptes_none
)
2169 if (!pte_present(pteval
) || !pte_write(pteval
))
2171 page
= vm_normal_page(vma
, _address
, pteval
);
2172 if (unlikely(!page
))
2175 * Chose the node of the first page. This could
2176 * be more sophisticated and look at more pages,
2177 * but isn't for now.
2180 node
= page_to_nid(page
);
2181 VM_BUG_ON(PageCompound(page
));
2182 if (!PageLRU(page
) || PageLocked(page
) || !PageAnon(page
))
2184 /* cannot use mapcount: can't collapse if there's a gup pin */
2185 if (page_count(page
) != 1)
2187 if (pte_young(pteval
) || PageReferenced(page
) ||
2188 mmu_notifier_test_young(vma
->vm_mm
, address
))
2194 pte_unmap_unlock(pte
, ptl
);
2196 /* collapse_huge_page will return with the mmap_sem released */
2197 collapse_huge_page(mm
, address
, hpage
, vma
, node
);
2202 static void collect_mm_slot(struct mm_slot
*mm_slot
)
2204 struct mm_struct
*mm
= mm_slot
->mm
;
2206 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2208 if (khugepaged_test_exit(mm
)) {
2210 hlist_del(&mm_slot
->hash
);
2211 list_del(&mm_slot
->mm_node
);
2214 * Not strictly needed because the mm exited already.
2216 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2219 /* khugepaged_mm_lock actually not necessary for the below */
2220 free_mm_slot(mm_slot
);
2225 static unsigned int khugepaged_scan_mm_slot(unsigned int pages
,
2226 struct page
**hpage
)
2227 __releases(&khugepaged_mm_lock
)
2228 __acquires(&khugepaged_mm_lock
)
2230 struct mm_slot
*mm_slot
;
2231 struct mm_struct
*mm
;
2232 struct vm_area_struct
*vma
;
2236 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
2238 if (khugepaged_scan
.mm_slot
)
2239 mm_slot
= khugepaged_scan
.mm_slot
;
2241 mm_slot
= list_entry(khugepaged_scan
.mm_head
.next
,
2242 struct mm_slot
, mm_node
);
2243 khugepaged_scan
.address
= 0;
2244 khugepaged_scan
.mm_slot
= mm_slot
;
2246 spin_unlock(&khugepaged_mm_lock
);
2249 down_read(&mm
->mmap_sem
);
2250 if (unlikely(khugepaged_test_exit(mm
)))
2253 vma
= find_vma(mm
, khugepaged_scan
.address
);
2256 for (; vma
; vma
= vma
->vm_next
) {
2257 unsigned long hstart
, hend
;
2260 if (unlikely(khugepaged_test_exit(mm
))) {
2265 if ((!(vma
->vm_flags
& VM_HUGEPAGE
) &&
2266 !khugepaged_always()) ||
2267 (vma
->vm_flags
& VM_NOHUGEPAGE
)) {
2272 if (!vma
->anon_vma
|| vma
->vm_ops
)
2274 if (is_vma_temporary_stack(vma
))
2276 VM_BUG_ON(vma
->vm_flags
& VM_NO_THP
);
2278 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2279 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2282 if (khugepaged_scan
.address
> hend
)
2284 if (khugepaged_scan
.address
< hstart
)
2285 khugepaged_scan
.address
= hstart
;
2286 VM_BUG_ON(khugepaged_scan
.address
& ~HPAGE_PMD_MASK
);
2288 while (khugepaged_scan
.address
< hend
) {
2291 if (unlikely(khugepaged_test_exit(mm
)))
2292 goto breakouterloop
;
2294 VM_BUG_ON(khugepaged_scan
.address
< hstart
||
2295 khugepaged_scan
.address
+ HPAGE_PMD_SIZE
>
2297 ret
= khugepaged_scan_pmd(mm
, vma
,
2298 khugepaged_scan
.address
,
2300 /* move to next address */
2301 khugepaged_scan
.address
+= HPAGE_PMD_SIZE
;
2302 progress
+= HPAGE_PMD_NR
;
2304 /* we released mmap_sem so break loop */
2305 goto breakouterloop_mmap_sem
;
2306 if (progress
>= pages
)
2307 goto breakouterloop
;
2311 up_read(&mm
->mmap_sem
); /* exit_mmap will destroy ptes after this */
2312 breakouterloop_mmap_sem
:
2314 spin_lock(&khugepaged_mm_lock
);
2315 VM_BUG_ON(khugepaged_scan
.mm_slot
!= mm_slot
);
2317 * Release the current mm_slot if this mm is about to die, or
2318 * if we scanned all vmas of this mm.
2320 if (khugepaged_test_exit(mm
) || !vma
) {
2322 * Make sure that if mm_users is reaching zero while
2323 * khugepaged runs here, khugepaged_exit will find
2324 * mm_slot not pointing to the exiting mm.
2326 if (mm_slot
->mm_node
.next
!= &khugepaged_scan
.mm_head
) {
2327 khugepaged_scan
.mm_slot
= list_entry(
2328 mm_slot
->mm_node
.next
,
2329 struct mm_slot
, mm_node
);
2330 khugepaged_scan
.address
= 0;
2332 khugepaged_scan
.mm_slot
= NULL
;
2333 khugepaged_full_scans
++;
2336 collect_mm_slot(mm_slot
);
2342 static int khugepaged_has_work(void)
2344 return !list_empty(&khugepaged_scan
.mm_head
) &&
2345 khugepaged_enabled();
2348 static int khugepaged_wait_event(void)
2350 return !list_empty(&khugepaged_scan
.mm_head
) ||
2351 kthread_should_stop();
2354 static void khugepaged_do_scan(void)
2356 struct page
*hpage
= NULL
;
2357 unsigned int progress
= 0, pass_through_head
= 0;
2358 unsigned int pages
= khugepaged_pages_to_scan
;
2361 barrier(); /* write khugepaged_pages_to_scan to local stack */
2363 while (progress
< pages
) {
2364 if (!khugepaged_prealloc_page(&hpage
, &wait
))
2369 if (unlikely(kthread_should_stop() || freezing(current
)))
2372 spin_lock(&khugepaged_mm_lock
);
2373 if (!khugepaged_scan
.mm_slot
)
2374 pass_through_head
++;
2375 if (khugepaged_has_work() &&
2376 pass_through_head
< 2)
2377 progress
+= khugepaged_scan_mm_slot(pages
- progress
,
2381 spin_unlock(&khugepaged_mm_lock
);
2384 if (!IS_ERR_OR_NULL(hpage
))
2388 static void khugepaged_wait_work(void)
2392 if (khugepaged_has_work()) {
2393 if (!khugepaged_scan_sleep_millisecs
)
2396 wait_event_freezable_timeout(khugepaged_wait
,
2397 kthread_should_stop(),
2398 msecs_to_jiffies(khugepaged_scan_sleep_millisecs
));
2402 if (khugepaged_enabled())
2403 wait_event_freezable(khugepaged_wait
, khugepaged_wait_event());
2406 static int khugepaged(void *none
)
2408 struct mm_slot
*mm_slot
;
2411 set_user_nice(current
, 19);
2413 while (!kthread_should_stop()) {
2414 khugepaged_do_scan();
2415 khugepaged_wait_work();
2418 spin_lock(&khugepaged_mm_lock
);
2419 mm_slot
= khugepaged_scan
.mm_slot
;
2420 khugepaged_scan
.mm_slot
= NULL
;
2422 collect_mm_slot(mm_slot
);
2423 spin_unlock(&khugepaged_mm_lock
);
2427 void __split_huge_page_pmd(struct mm_struct
*mm
, pmd_t
*pmd
)
2431 spin_lock(&mm
->page_table_lock
);
2432 if (unlikely(!pmd_trans_huge(*pmd
))) {
2433 spin_unlock(&mm
->page_table_lock
);
2436 page
= pmd_page(*pmd
);
2437 VM_BUG_ON(!page_count(page
));
2439 spin_unlock(&mm
->page_table_lock
);
2441 split_huge_page(page
);
2444 BUG_ON(pmd_trans_huge(*pmd
));
2447 static void split_huge_page_address(struct mm_struct
*mm
,
2448 unsigned long address
)
2454 VM_BUG_ON(!(address
& ~HPAGE_PMD_MASK
));
2456 pgd
= pgd_offset(mm
, address
);
2457 if (!pgd_present(*pgd
))
2460 pud
= pud_offset(pgd
, address
);
2461 if (!pud_present(*pud
))
2464 pmd
= pmd_offset(pud
, address
);
2465 if (!pmd_present(*pmd
))
2468 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2469 * materialize from under us.
2471 split_huge_page_pmd(mm
, pmd
);
2474 void __vma_adjust_trans_huge(struct vm_area_struct
*vma
,
2475 unsigned long start
,
2480 * If the new start address isn't hpage aligned and it could
2481 * previously contain an hugepage: check if we need to split
2484 if (start
& ~HPAGE_PMD_MASK
&&
2485 (start
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2486 (start
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2487 split_huge_page_address(vma
->vm_mm
, start
);
2490 * If the new end address isn't hpage aligned and it could
2491 * previously contain an hugepage: check if we need to split
2494 if (end
& ~HPAGE_PMD_MASK
&&
2495 (end
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2496 (end
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2497 split_huge_page_address(vma
->vm_mm
, end
);
2500 * If we're also updating the vma->vm_next->vm_start, if the new
2501 * vm_next->vm_start isn't page aligned and it could previously
2502 * contain an hugepage: check if we need to split an huge pmd.
2504 if (adjust_next
> 0) {
2505 struct vm_area_struct
*next
= vma
->vm_next
;
2506 unsigned long nstart
= next
->vm_start
;
2507 nstart
+= adjust_next
<< PAGE_SHIFT
;
2508 if (nstart
& ~HPAGE_PMD_MASK
&&
2509 (nstart
& HPAGE_PMD_MASK
) >= next
->vm_start
&&
2510 (nstart
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= next
->vm_end
)
2511 split_huge_page_address(next
->vm_mm
, nstart
);