projects / deliverable / linux.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
mm/memcg: iteration skip memcgs not yet fully initialized
[deliverable/linux.git] / mm / huge_memory.c
CommitLineData
71e3aac0
AA		 1/*
2 * Copyright (C) 2009 Red Hat, Inc.
3 *
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
6 */
7
8#include <linux/mm.h>
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>
97ae1749 15#include <linux/shrinker.h>
ba76149f
AA		 16#include <linux/mm_inline.h>
17#include <linux/kthread.h>
18#include <linux/khugepaged.h>
878aee7d 19#include <linux/freezer.h>
a664b2d8 20#include <linux/mman.h>
325adeb5 21#include <linux/pagemap.h>
4daae3b4 22#include <linux/migrate.h>
43b5fbbd 23#include <linux/hashtable.h>
97ae1749 24
71e3aac0
AA		 25#include <asm/tlb.h>
26#include <asm/pgalloc.h>
27#include "internal.h"
28
ba76149f 29/*
8bfa3f9a
JW		 30 * By default transparent hugepage support is disabled in order that avoid
31 * to risk increase the memory footprint of applications without a guaranteed
32 * benefit. When transparent hugepage support is enabled, is for all mappings,
33 * and khugepaged scans all mappings.
34 * Defrag is invoked by khugepaged hugepage allocations and by page faults
35 * for all hugepage allocations.
ba76149f 36 */
71e3aac0 37unsigned long transparent_hugepage_flags __read_mostly =
13ece886 38#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
ba76149f 39 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
13ece886
AA		 40#endif
41#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
42 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
43#endif
d39d33c3 44 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
79da5407
KS		 45 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
46 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
ba76149f
AA		 47
48/* default scan 8*512 pte (or vmas) every 30 second */
49static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
50static unsigned int khugepaged_pages_collapsed;
51static unsigned int khugepaged_full_scans;
52static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
53/* during fragmentation poll the hugepage allocator once every minute */
54static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
55static struct task_struct *khugepaged_thread __read_mostly;
56static DEFINE_MUTEX(khugepaged_mutex);
57static DEFINE_SPINLOCK(khugepaged_mm_lock);
58static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
59/*
60 * default collapse hugepages if there is at least one pte mapped like
61 * it would have happened if the vma was large enough during page
62 * fault.
63 */
64static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
65
66static int khugepaged(void *none);
ba76149f 67static int khugepaged_slab_init(void);
ba76149f 68
43b5fbbd
SL		 69#define MM_SLOTS_HASH_BITS 10
70static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
71
ba76149f
AA		 72static struct kmem_cache *mm_slot_cache __read_mostly;
73
74/**
75 * struct mm_slot - hash lookup from mm to mm_slot
76 * @hash: hash collision list
77 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
78 * @mm: the mm that this information is valid for
79 */
80struct mm_slot {
81 struct hlist_node hash;
82 struct list_head mm_node;
83 struct mm_struct *mm;
84};
85
86/**
87 * struct khugepaged_scan - cursor for scanning
88 * @mm_head: the head of the mm list to scan
89 * @mm_slot: the current mm_slot we are scanning
90 * @address: the next address inside that to be scanned
91 *
92 * There is only the one khugepaged_scan instance of this cursor structure.
93 */
94struct khugepaged_scan {
95 struct list_head mm_head;
96 struct mm_slot *mm_slot;
97 unsigned long address;
2f1da642
HS		 98};
99static struct khugepaged_scan khugepaged_scan = {
ba76149f
AA		 100 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
101};
102
f000565a
AA		 103
104static int set_recommended_min_free_kbytes(void)
105{
106 struct zone *zone;
107 int nr_zones = 0;
108 unsigned long recommended_min;
f000565a 109
17c230af 110 if (!khugepaged_enabled())
f000565a
AA		 111 return 0;
112
113 for_each_populated_zone(zone)
114 nr_zones++;
115
116 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
117 recommended_min = pageblock_nr_pages * nr_zones * 2;
118
119 /*
120 * Make sure that on average at least two pageblocks are almost free
121 * of another type, one for a migratetype to fall back to and a
122 * second to avoid subsequent fallbacks of other types There are 3
123 * MIGRATE_TYPES we care about.
124 */
125 recommended_min += pageblock_nr_pages * nr_zones *
126 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
127
128 /* don't ever allow to reserve more than 5% of the lowmem */
129 recommended_min = min(recommended_min,
130 (unsigned long) nr_free_buffer_pages() / 20);
131 recommended_min <<= (PAGE_SHIFT-10);
132
42aa83cb
HP		 133 if (recommended_min > min_free_kbytes) {
134 if (user_min_free_kbytes >= 0)
135 pr_info("raising min_free_kbytes from %d to %lu "
136 "to help transparent hugepage allocations\n",
137 min_free_kbytes, recommended_min);
138
f000565a 139 min_free_kbytes = recommended_min;
42aa83cb 140 }
f000565a
AA		 141 setup_per_zone_wmarks();
142 return 0;
143}
144late_initcall(set_recommended_min_free_kbytes);
145
ba76149f
AA		 146static int start_khugepaged(void)
147{
148 int err = 0;
149 if (khugepaged_enabled()) {
ba76149f
AA		 150 if (!khugepaged_thread)
151 khugepaged_thread = kthread_run(khugepaged, NULL,
152 "khugepaged");
153 if (unlikely(IS_ERR(khugepaged_thread))) {
154 printk(KERN_ERR
155 "khugepaged: kthread_run(khugepaged) failed\n");
156 err = PTR_ERR(khugepaged_thread);
157 khugepaged_thread = NULL;
158 }
911891af
XG		 159
160 if (!list_empty(&khugepaged_scan.mm_head))
ba76149f 161 wake_up_interruptible(&khugepaged_wait);
f000565a
AA		 162
163 set_recommended_min_free_kbytes();
911891af 164 } else if (khugepaged_thread) {
911891af
XG		 165 kthread_stop(khugepaged_thread);
166 khugepaged_thread = NULL;
167 }
637e3a27 168
ba76149f
AA		 169 return err;
170}
71e3aac0 171
97ae1749 172static atomic_t huge_zero_refcount;
5918d10a 173static struct page *huge_zero_page __read_mostly;
97ae1749 174
5918d10a 175static inline bool is_huge_zero_page(struct page *page)
4a6c1297 176{
5918d10a 177 return ACCESS_ONCE(huge_zero_page) == page;
97ae1749 178}
4a6c1297 179
97ae1749
KS		 180static inline bool is_huge_zero_pmd(pmd_t pmd)
181{
5918d10a 182 return is_huge_zero_page(pmd_page(pmd));
97ae1749
KS		 183}
184
5918d10a 185static struct page *get_huge_zero_page(void)
97ae1749
KS		 186{
187 struct page *zero_page;
188retry:
189 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
5918d10a 190 return ACCESS_ONCE(huge_zero_page);
97ae1749
KS		 191
192 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
4a6c1297 193 HPAGE_PMD_ORDER);
d8a8e1f0
KS		 194 if (!zero_page) {
195 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
5918d10a 196 return NULL;
d8a8e1f0
KS		 197 }
198 count_vm_event(THP_ZERO_PAGE_ALLOC);
97ae1749 199 preempt_disable();
5918d10a 200 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
97ae1749
KS		 201 preempt_enable();
202 __free_page(zero_page);
203 goto retry;
204 }
205
206 /* We take additional reference here. It will be put back by shrinker */
207 atomic_set(&huge_zero_refcount, 2);
208 preempt_enable();
5918d10a 209 return ACCESS_ONCE(huge_zero_page);
4a6c1297
KS		 210}
211
97ae1749 212static void put_huge_zero_page(void)
4a6c1297 213{
97ae1749
KS		 214 /*
215 * Counter should never go to zero here. Only shrinker can put
216 * last reference.
217 */
218 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
4a6c1297
KS		 219}
220
48896466
GC		 221static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
222 struct shrink_control *sc)
4a6c1297 223{
48896466
GC		 224 /* we can free zero page only if last reference remains */
225 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
226}
97ae1749 227
48896466
GC		 228static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
229 struct shrink_control *sc)
230{
97ae1749 231 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
5918d10a
KS		 232 struct page *zero_page = xchg(&huge_zero_page, NULL);
233 BUG_ON(zero_page == NULL);
234 __free_page(zero_page);
48896466 235 return HPAGE_PMD_NR;
97ae1749
KS		 236 }
237
238 return 0;
4a6c1297
KS		 239}
240
97ae1749 241static struct shrinker huge_zero_page_shrinker = {
48896466
GC		 242 .count_objects = shrink_huge_zero_page_count,
243 .scan_objects = shrink_huge_zero_page_scan,
97ae1749
KS		 244 .seeks = DEFAULT_SEEKS,
245};
246
71e3aac0 247#ifdef CONFIG_SYSFS
ba76149f 248
71e3aac0
AA		 249static ssize_t double_flag_show(struct kobject *kobj,
250 struct kobj_attribute *attr, char *buf,
251 enum transparent_hugepage_flag enabled,
252 enum transparent_hugepage_flag req_madv)
253{
254 if (test_bit(enabled, &transparent_hugepage_flags)) {
255 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
256 return sprintf(buf, "[always] madvise never\n");
257 } else if (test_bit(req_madv, &transparent_hugepage_flags))
258 return sprintf(buf, "always [madvise] never\n");
259 else
260 return sprintf(buf, "always madvise [never]\n");
261}
262static ssize_t double_flag_store(struct kobject *kobj,
263 struct kobj_attribute *attr,
264 const char *buf, size_t count,
265 enum transparent_hugepage_flag enabled,
266 enum transparent_hugepage_flag req_madv)
267{
268 if (!memcmp("always", buf,
269 min(sizeof("always")-1, count))) {
270 set_bit(enabled, &transparent_hugepage_flags);
271 clear_bit(req_madv, &transparent_hugepage_flags);
272 } else if (!memcmp("madvise", buf,
273 min(sizeof("madvise")-1, count))) {
274 clear_bit(enabled, &transparent_hugepage_flags);
275 set_bit(req_madv, &transparent_hugepage_flags);
276 } else if (!memcmp("never", buf,
277 min(sizeof("never")-1, count))) {
278 clear_bit(enabled, &transparent_hugepage_flags);
279 clear_bit(req_madv, &transparent_hugepage_flags);
280 } else
281 return -EINVAL;
282
283 return count;
284}
285
286static ssize_t enabled_show(struct kobject *kobj,
287 struct kobj_attribute *attr, char *buf)
288{
289 return double_flag_show(kobj, attr, buf,
290 TRANSPARENT_HUGEPAGE_FLAG,
291 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
292}
293static ssize_t enabled_store(struct kobject *kobj,
294 struct kobj_attribute *attr,
295 const char *buf, size_t count)
296{
ba76149f
AA		 297 ssize_t ret;
298
299 ret = double_flag_store(kobj, attr, buf, count,
300 TRANSPARENT_HUGEPAGE_FLAG,
301 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
302
303 if (ret > 0) {
911891af
XG		 304 int err;
305
306 mutex_lock(&khugepaged_mutex);
307 err = start_khugepaged();
308 mutex_unlock(&khugepaged_mutex);
309
ba76149f
AA		 310 if (err)
311 ret = err;
312 }
313
314 return ret;
71e3aac0
AA		 315}
316static struct kobj_attribute enabled_attr =
317 __ATTR(enabled, 0644, enabled_show, enabled_store);
318
319static ssize_t single_flag_show(struct kobject *kobj,
320 struct kobj_attribute *attr, char *buf,
321 enum transparent_hugepage_flag flag)
322{
e27e6151
BH		 323 return sprintf(buf, "%d\n",
324 !!test_bit(flag, &transparent_hugepage_flags));
71e3aac0 325}
e27e6151 326
71e3aac0
AA		 327static ssize_t single_flag_store(struct kobject *kobj,
328 struct kobj_attribute *attr,
329 const char *buf, size_t count,
330 enum transparent_hugepage_flag flag)
331{
e27e6151
BH		 332 unsigned long value;
333 int ret;
334
335 ret = kstrtoul(buf, 10, &value);
336 if (ret < 0)
337 return ret;
338 if (value > 1)
339 return -EINVAL;
340
341 if (value)
71e3aac0 342 set_bit(flag, &transparent_hugepage_flags);
e27e6151 343 else
71e3aac0 344 clear_bit(flag, &transparent_hugepage_flags);
71e3aac0
AA		 345
346 return count;
347}
348
349/*
350 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
351 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
352 * memory just to allocate one more hugepage.
353 */
354static ssize_t defrag_show(struct kobject *kobj,
355 struct kobj_attribute *attr, char *buf)
356{
357 return double_flag_show(kobj, attr, buf,
358 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
359 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
360}
361static ssize_t defrag_store(struct kobject *kobj,
362 struct kobj_attribute *attr,
363 const char *buf, size_t count)
364{
365 return double_flag_store(kobj, attr, buf, count,
366 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
367 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
368}
369static struct kobj_attribute defrag_attr =
370 __ATTR(defrag, 0644, defrag_show, defrag_store);
371
79da5407
KS		 372static ssize_t use_zero_page_show(struct kobject *kobj,
373 struct kobj_attribute *attr, char *buf)
374{
375 return single_flag_show(kobj, attr, buf,
376 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
377}
378static ssize_t use_zero_page_store(struct kobject *kobj,
379 struct kobj_attribute *attr, const char *buf, size_t count)
380{
381 return single_flag_store(kobj, attr, buf, count,
382 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
383}
384static struct kobj_attribute use_zero_page_attr =
385 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
71e3aac0
AA		 386#ifdef CONFIG_DEBUG_VM
387static ssize_t debug_cow_show(struct kobject *kobj,
388 struct kobj_attribute *attr, char *buf)
389{
390 return single_flag_show(kobj, attr, buf,
391 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
392}
393static ssize_t debug_cow_store(struct kobject *kobj,
394 struct kobj_attribute *attr,
395 const char *buf, size_t count)
396{
397 return single_flag_store(kobj, attr, buf, count,
398 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
399}
400static struct kobj_attribute debug_cow_attr =
401 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
402#endif /* CONFIG_DEBUG_VM */
403
404static struct attribute *hugepage_attr[] = {
405 &enabled_attr.attr,
406 &defrag_attr.attr,
79da5407 407 &use_zero_page_attr.attr,
71e3aac0
AA		 408#ifdef CONFIG_DEBUG_VM
409 &debug_cow_attr.attr,
410#endif
411 NULL,
412};
413
414static struct attribute_group hugepage_attr_group = {
415 .attrs = hugepage_attr,
ba76149f
AA		 416};
417
418static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
419 struct kobj_attribute *attr,
420 char *buf)
421{
422 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
423}
424
425static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
426 struct kobj_attribute *attr,
427 const char *buf, size_t count)
428{
429 unsigned long msecs;
430 int err;
431
3dbb95f7 432 err = kstrtoul(buf, 10, &msecs);
ba76149f
AA		 433 if (err || msecs > UINT_MAX)
434 return -EINVAL;
435
436 khugepaged_scan_sleep_millisecs = msecs;
437 wake_up_interruptible(&khugepaged_wait);
438
439 return count;
440}
441static struct kobj_attribute scan_sleep_millisecs_attr =
442 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
443 scan_sleep_millisecs_store);
444
445static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
446 struct kobj_attribute *attr,
447 char *buf)
448{
449 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
450}
451
452static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
453 struct kobj_attribute *attr,
454 const char *buf, size_t count)
455{
456 unsigned long msecs;
457 int err;
458
3dbb95f7 459 err = kstrtoul(buf, 10, &msecs);
ba76149f
AA		 460 if (err || msecs > UINT_MAX)
461 return -EINVAL;
462
463 khugepaged_alloc_sleep_millisecs = msecs;
464 wake_up_interruptible(&khugepaged_wait);
465
466 return count;
467}
468static struct kobj_attribute alloc_sleep_millisecs_attr =
469 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
470 alloc_sleep_millisecs_store);
471
472static ssize_t pages_to_scan_show(struct kobject *kobj,
473 struct kobj_attribute *attr,
474 char *buf)
475{
476 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
477}
478static ssize_t pages_to_scan_store(struct kobject *kobj,
479 struct kobj_attribute *attr,
480 const char *buf, size_t count)
481{
482 int err;
483 unsigned long pages;
484
3dbb95f7 485 err = kstrtoul(buf, 10, &pages);
ba76149f
AA		 486 if (err || !pages || pages > UINT_MAX)
487 return -EINVAL;
488
489 khugepaged_pages_to_scan = pages;
490
491 return count;
492}
493static struct kobj_attribute pages_to_scan_attr =
494 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
495 pages_to_scan_store);
496
497static ssize_t pages_collapsed_show(struct kobject *kobj,
498 struct kobj_attribute *attr,
499 char *buf)
500{
501 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
502}
503static struct kobj_attribute pages_collapsed_attr =
504 __ATTR_RO(pages_collapsed);
505
506static ssize_t full_scans_show(struct kobject *kobj,
507 struct kobj_attribute *attr,
508 char *buf)
509{
510 return sprintf(buf, "%u\n", khugepaged_full_scans);
511}
512static struct kobj_attribute full_scans_attr =
513 __ATTR_RO(full_scans);
514
515static ssize_t khugepaged_defrag_show(struct kobject *kobj,
516 struct kobj_attribute *attr, char *buf)
517{
518 return single_flag_show(kobj, attr, buf,
519 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
520}
521static ssize_t khugepaged_defrag_store(struct kobject *kobj,
522 struct kobj_attribute *attr,
523 const char *buf, size_t count)
524{
525 return single_flag_store(kobj, attr, buf, count,
526 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
527}
528static struct kobj_attribute khugepaged_defrag_attr =
529 __ATTR(defrag, 0644, khugepaged_defrag_show,
530 khugepaged_defrag_store);
531
532/*
533 * max_ptes_none controls if khugepaged should collapse hugepages over
534 * any unmapped ptes in turn potentially increasing the memory
535 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
536 * reduce the available free memory in the system as it
537 * runs. Increasing max_ptes_none will instead potentially reduce the
538 * free memory in the system during the khugepaged scan.
539 */
540static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
541 struct kobj_attribute *attr,
542 char *buf)
543{
544 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
545}
546static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
547 struct kobj_attribute *attr,
548 const char *buf, size_t count)
549{
550 int err;
551 unsigned long max_ptes_none;
552
3dbb95f7 553 err = kstrtoul(buf, 10, &max_ptes_none);
ba76149f
AA		 554 if (err || max_ptes_none > HPAGE_PMD_NR-1)
555 return -EINVAL;
556
557 khugepaged_max_ptes_none = max_ptes_none;
558
559 return count;
560}
561static struct kobj_attribute khugepaged_max_ptes_none_attr =
562 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
563 khugepaged_max_ptes_none_store);
564
565static struct attribute *khugepaged_attr[] = {
566 &khugepaged_defrag_attr.attr,
567 &khugepaged_max_ptes_none_attr.attr,
568 &pages_to_scan_attr.attr,
569 &pages_collapsed_attr.attr,
570 &full_scans_attr.attr,
571 &scan_sleep_millisecs_attr.attr,
572 &alloc_sleep_millisecs_attr.attr,
573 NULL,
574};
575
576static struct attribute_group khugepaged_attr_group = {
577 .attrs = khugepaged_attr,
578 .name = "khugepaged",
71e3aac0 579};
71e3aac0 580
569e5590 581static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
71e3aac0 582{
71e3aac0
AA		 583 int err;
584
569e5590
SL		 585 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
586 if (unlikely(!*hugepage_kobj)) {
2c79737a 587 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
569e5590 588 return -ENOMEM;
ba76149f
AA		 589 }
590
569e5590 591 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
ba76149f 592 if (err) {
2c79737a 593 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
569e5590 594 goto delete_obj;
ba76149f
AA		 595 }
596
569e5590 597 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
ba76149f 598 if (err) {
2c79737a 599 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
569e5590 600 goto remove_hp_group;
ba76149f 601 }
569e5590
SL		 602
603 return 0;
604
605remove_hp_group:
606 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
607delete_obj:
608 kobject_put(*hugepage_kobj);
609 return err;
610}
611
612static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
613{
614 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
615 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
616 kobject_put(hugepage_kobj);
617}
618#else
619static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
620{
621 return 0;
622}
623
624static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
625{
626}
627#endif /* CONFIG_SYSFS */
628
629static int __init hugepage_init(void)
630{
631 int err;
632 struct kobject *hugepage_kobj;
633
634 if (!has_transparent_hugepage()) {
635 transparent_hugepage_flags = 0;
636 return -EINVAL;
637 }
638
639 err = hugepage_init_sysfs(&hugepage_kobj);
640 if (err)
641 return err;
ba76149f
AA		 642
643 err = khugepaged_slab_init();
644 if (err)
645 goto out;
646
97ae1749
KS		 647 register_shrinker(&huge_zero_page_shrinker);
648
97562cd2
RR		 649 /*
650 * By default disable transparent hugepages on smaller systems,
651 * where the extra memory used could hurt more than TLB overhead
652 * is likely to save. The admin can still enable it through /sys.
653 */
654 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
655 transparent_hugepage_flags = 0;
656
ba76149f
AA		 657 start_khugepaged();
658
569e5590 659 return 0;
ba76149f 660out:
569e5590 661 hugepage_exit_sysfs(hugepage_kobj);
ba76149f 662 return err;
71e3aac0 663}
a64fb3cd 664subsys_initcall(hugepage_init);
71e3aac0
AA		 665
666static int __init setup_transparent_hugepage(char *str)
667{
668 int ret = 0;
669 if (!str)
670 goto out;
671 if (!strcmp(str, "always")) {
672 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
673 &transparent_hugepage_flags);
674 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
675 &transparent_hugepage_flags);
676 ret = 1;
677 } else if (!strcmp(str, "madvise")) {
678 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
679 &transparent_hugepage_flags);
680 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
681 &transparent_hugepage_flags);
682 ret = 1;
683 } else if (!strcmp(str, "never")) {
684 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
685 &transparent_hugepage_flags);
686 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
687 &transparent_hugepage_flags);
688 ret = 1;
689 }
690out:
691 if (!ret)
692 printk(KERN_WARNING
693 "transparent_hugepage= cannot parse, ignored\n");
694 return ret;
695}
696__setup("transparent_hugepage=", setup_transparent_hugepage);
697
b32967ff 698pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
71e3aac0
AA		 699{
700 if (likely(vma->vm_flags & VM_WRITE))
701 pmd = pmd_mkwrite(pmd);
702 return pmd;
703}
704
3122359a 705static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
b3092b3b
BL		 706{
707 pmd_t entry;
3122359a 708 entry = mk_pmd(page, prot);
b3092b3b
BL		 709 entry = pmd_mkhuge(entry);
710 return entry;
711}
712
71e3aac0
AA		 713static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
714 struct vm_area_struct *vma,
715 unsigned long haddr, pmd_t *pmd,
716 struct page *page)
717{
71e3aac0 718 pgtable_t pgtable;
c4088ebd 719 spinlock_t *ptl;
71e3aac0 720
309381fe 721 VM_BUG_ON_PAGE(!PageCompound(page), page);
71e3aac0 722 pgtable = pte_alloc_one(mm, haddr);
edad9d2c 723 if (unlikely(!pgtable))
71e3aac0 724 return VM_FAULT_OOM;
71e3aac0
AA		 725
726 clear_huge_page(page, haddr, HPAGE_PMD_NR);
52f37629
MK		 727 /*
728 * The memory barrier inside __SetPageUptodate makes sure that
729 * clear_huge_page writes become visible before the set_pmd_at()
730 * write.
731 */
71e3aac0
AA		 732 __SetPageUptodate(page);
733
c4088ebd 734 ptl = pmd_lock(mm, pmd);
71e3aac0 735 if (unlikely(!pmd_none(*pmd))) {
c4088ebd 736 spin_unlock(ptl);
b9bbfbe3 737 mem_cgroup_uncharge_page(page);
71e3aac0
AA		 738 put_page(page);
739 pte_free(mm, pgtable);
740 } else {
741 pmd_t entry;
3122359a
KS		 742 entry = mk_huge_pmd(page, vma->vm_page_prot);
743 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
71e3aac0 744 page_add_new_anon_rmap(page, vma, haddr);
6b0b50b0 745 pgtable_trans_huge_deposit(mm, pmd, pgtable);
71e3aac0 746 set_pmd_at(mm, haddr, pmd, entry);
71e3aac0 747 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
e1f56c89 748 atomic_long_inc(&mm->nr_ptes);
c4088ebd 749 spin_unlock(ptl);
71e3aac0
AA		 750 }
751
aa2e878e 752 return 0;
71e3aac0
AA		 753}
754
cc5d462f 755static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
0bbbc0b3 756{
cc5d462f 757 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
0bbbc0b3
AA		 758}
759
760static inline struct page *alloc_hugepage_vma(int defrag,
761 struct vm_area_struct *vma,
cc5d462f
AK		 762 unsigned long haddr, int nd,
763 gfp_t extra_gfp)
0bbbc0b3 764{
cc5d462f 765 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
5c4b4be3 766 HPAGE_PMD_ORDER, vma, haddr, nd);
0bbbc0b3
AA		 767}
768
c4088ebd 769/* Caller must hold page table lock. */
3ea41e62 770static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
97ae1749 771 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
5918d10a 772 struct page *zero_page)
fc9fe822
KS		 773{
774 pmd_t entry;
3ea41e62
KS		 775 if (!pmd_none(*pmd))
776 return false;
5918d10a 777 entry = mk_pmd(zero_page, vma->vm_page_prot);
fc9fe822
KS		 778 entry = pmd_wrprotect(entry);
779 entry = pmd_mkhuge(entry);
6b0b50b0 780 pgtable_trans_huge_deposit(mm, pmd, pgtable);
fc9fe822 781 set_pmd_at(mm, haddr, pmd, entry);
e1f56c89 782 atomic_long_inc(&mm->nr_ptes);
3ea41e62 783 return true;
fc9fe822
KS		 784}
785
71e3aac0
AA		 786int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
787 unsigned long address, pmd_t *pmd,
788 unsigned int flags)
789{
790 struct page *page;
791 unsigned long haddr = address & HPAGE_PMD_MASK;
71e3aac0 792
128ec037 793 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
c0292554 794 return VM_FAULT_FALLBACK;
128ec037
KS		 795 if (unlikely(anon_vma_prepare(vma)))
796 return VM_FAULT_OOM;
797 if (unlikely(khugepaged_enter(vma)))
798 return VM_FAULT_OOM;
799 if (!(flags & FAULT_FLAG_WRITE) &&
800 transparent_hugepage_use_zero_page()) {
c4088ebd 801 spinlock_t *ptl;
128ec037
KS		 802 pgtable_t pgtable;
803 struct page *zero_page;
804 bool set;
805 pgtable = pte_alloc_one(mm, haddr);
806 if (unlikely(!pgtable))
ba76149f 807 return VM_FAULT_OOM;
128ec037
KS		 808 zero_page = get_huge_zero_page();
809 if (unlikely(!zero_page)) {
810 pte_free(mm, pgtable);
81ab4201 811 count_vm_event(THP_FAULT_FALLBACK);
c0292554 812 return VM_FAULT_FALLBACK;
b9bbfbe3 813 }
c4088ebd 814 ptl = pmd_lock(mm, pmd);
128ec037
KS		 815 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
816 zero_page);
c4088ebd 817 spin_unlock(ptl);
128ec037
KS		 818 if (!set) {
819 pte_free(mm, pgtable);
820 put_huge_zero_page();
edad9d2c 821 }
edad9d2c 822 return 0;
71e3aac0 823 }
128ec037
KS		 824 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
825 vma, haddr, numa_node_id(), 0);
826 if (unlikely(!page)) {
827 count_vm_event(THP_FAULT_FALLBACK);
c0292554 828 return VM_FAULT_FALLBACK;
128ec037 829 }
128ec037
KS		 830 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
831 put_page(page);
17766dde 832 count_vm_event(THP_FAULT_FALLBACK);
c0292554 833 return VM_FAULT_FALLBACK;
128ec037
KS		 834 }
835 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
836 mem_cgroup_uncharge_page(page);
837 put_page(page);
17766dde 838 count_vm_event(THP_FAULT_FALLBACK);
c0292554 839 return VM_FAULT_FALLBACK;
128ec037
KS		 840 }
841
17766dde 842 count_vm_event(THP_FAULT_ALLOC);
128ec037 843 return 0;
71e3aac0
AA		 844}
845
846int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
847 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
848 struct vm_area_struct *vma)
849{
c4088ebd 850 spinlock_t *dst_ptl, *src_ptl;
71e3aac0
AA		 851 struct page *src_page;
852 pmd_t pmd;
853 pgtable_t pgtable;
854 int ret;
855
856 ret = -ENOMEM;
857 pgtable = pte_alloc_one(dst_mm, addr);
858 if (unlikely(!pgtable))
859 goto out;
860
c4088ebd
KS		 861 dst_ptl = pmd_lock(dst_mm, dst_pmd);
862 src_ptl = pmd_lockptr(src_mm, src_pmd);
863 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
71e3aac0
AA		 864
865 ret = -EAGAIN;
866 pmd = *src_pmd;
867 if (unlikely(!pmd_trans_huge(pmd))) {
868 pte_free(dst_mm, pgtable);
869 goto out_unlock;
870 }
fc9fe822 871 /*
c4088ebd 872 * When page table lock is held, the huge zero pmd should not be
fc9fe822
KS		 873 * under splitting since we don't split the page itself, only pmd to
874 * a page table.
875 */
876 if (is_huge_zero_pmd(pmd)) {
5918d10a 877 struct page *zero_page;
3ea41e62 878 bool set;
97ae1749
KS		 879 /*
880 * get_huge_zero_page() will never allocate a new page here,
881 * since we already have a zero page to copy. It just takes a
882 * reference.
883 */
5918d10a 884 zero_page = get_huge_zero_page();
3ea41e62 885 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
5918d10a 886 zero_page);
3ea41e62 887 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
fc9fe822
KS		 888 ret = 0;
889 goto out_unlock;
890 }
de466bd6 891
71e3aac0
AA		 892 if (unlikely(pmd_trans_splitting(pmd))) {
893 /* split huge page running from under us */
c4088ebd
KS		 894 spin_unlock(src_ptl);
895 spin_unlock(dst_ptl);
71e3aac0
AA		 896 pte_free(dst_mm, pgtable);
897
898 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
899 goto out;
900 }
901 src_page = pmd_page(pmd);
309381fe 902 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
71e3aac0
AA		 903 get_page(src_page);
904 page_dup_rmap(src_page);
905 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
906
907 pmdp_set_wrprotect(src_mm, addr, src_pmd);
908 pmd = pmd_mkold(pmd_wrprotect(pmd));
6b0b50b0 909 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
71e3aac0 910 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
e1f56c89 911 atomic_long_inc(&dst_mm->nr_ptes);
71e3aac0
AA		 912
913 ret = 0;
914out_unlock:
c4088ebd
KS		 915 spin_unlock(src_ptl);
916 spin_unlock(dst_ptl);
71e3aac0
AA		 917out:
918 return ret;
919}
920
a1dd450b
WD		 921void huge_pmd_set_accessed(struct mm_struct *mm,
922 struct vm_area_struct *vma,
923 unsigned long address,
924 pmd_t *pmd, pmd_t orig_pmd,
925 int dirty)
926{
c4088ebd 927 spinlock_t *ptl;
a1dd450b
WD		 928 pmd_t entry;
929 unsigned long haddr;
930
c4088ebd 931 ptl = pmd_lock(mm, pmd);
a1dd450b
WD		 932 if (unlikely(!pmd_same(*pmd, orig_pmd)))
933 goto unlock;
934
935 entry = pmd_mkyoung(orig_pmd);
936 haddr = address & HPAGE_PMD_MASK;
937 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
938 update_mmu_cache_pmd(vma, address, pmd);
939
940unlock:
c4088ebd 941 spin_unlock(ptl);
a1dd450b
WD		 942}
943
93b4796d
KS		 944static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
945 struct vm_area_struct *vma, unsigned long address,
3ea41e62 946 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
93b4796d 947{
c4088ebd 948 spinlock_t *ptl;
93b4796d
KS		 949 pgtable_t pgtable;
950 pmd_t _pmd;
951 struct page *page;
952 int i, ret = 0;
953 unsigned long mmun_start; /* For mmu_notifiers */
954 unsigned long mmun_end; /* For mmu_notifiers */
955
956 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
957 if (!page) {
958 ret |= VM_FAULT_OOM;
959 goto out;
960 }
961
962 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
963 put_page(page);
964 ret |= VM_FAULT_OOM;
965 goto out;
966 }
967
968 clear_user_highpage(page, address);
969 __SetPageUptodate(page);
970
971 mmun_start = haddr;
972 mmun_end = haddr + HPAGE_PMD_SIZE;
973 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
974
c4088ebd 975 ptl = pmd_lock(mm, pmd);
3ea41e62
KS		 976 if (unlikely(!pmd_same(*pmd, orig_pmd)))
977 goto out_free_page;
978
93b4796d
KS		 979 pmdp_clear_flush(vma, haddr, pmd);
980 /* leave pmd empty until pte is filled */
981
6b0b50b0 982 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
93b4796d
KS		 983 pmd_populate(mm, &_pmd, pgtable);
984
985 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
986 pte_t *pte, entry;
987 if (haddr == (address & PAGE_MASK)) {
988 entry = mk_pte(page, vma->vm_page_prot);
989 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
990 page_add_new_anon_rmap(page, vma, haddr);
991 } else {
992 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
993 entry = pte_mkspecial(entry);
994 }
995 pte = pte_offset_map(&_pmd, haddr);
996 VM_BUG_ON(!pte_none(*pte));
997 set_pte_at(mm, haddr, pte, entry);
998 pte_unmap(pte);
999 }
1000 smp_wmb(); /* make pte visible before pmd */
1001 pmd_populate(mm, pmd, pgtable);
c4088ebd 1002 spin_unlock(ptl);
97ae1749 1003 put_huge_zero_page();
93b4796d
KS		 1004 inc_mm_counter(mm, MM_ANONPAGES);
1005
1006 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1007
1008 ret |= VM_FAULT_WRITE;
1009out:
1010 return ret;
3ea41e62 1011out_free_page:
c4088ebd 1012 spin_unlock(ptl);
3ea41e62
KS		 1013 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1014 mem_cgroup_uncharge_page(page);
1015 put_page(page);
1016 goto out;
93b4796d
KS		 1017}
1018
71e3aac0
AA		 1019static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1020 struct vm_area_struct *vma,
1021 unsigned long address,
1022 pmd_t *pmd, pmd_t orig_pmd,
1023 struct page *page,
1024 unsigned long haddr)
1025{
c4088ebd 1026 spinlock_t *ptl;
71e3aac0
AA		 1027 pgtable_t pgtable;
1028 pmd_t _pmd;
1029 int ret = 0, i;
1030 struct page **pages;
2ec74c3e
SG		 1031 unsigned long mmun_start; /* For mmu_notifiers */
1032 unsigned long mmun_end; /* For mmu_notifiers */
71e3aac0
AA		 1033
1034 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1035 GFP_KERNEL);
1036 if (unlikely(!pages)) {
1037 ret |= VM_FAULT_OOM;
1038 goto out;
1039 }
1040
1041 for (i = 0; i < HPAGE_PMD_NR; i++) {
cc5d462f
AK		 1042 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1043 __GFP_OTHER_NODE,
19ee151e 1044 vma, address, page_to_nid(page));
b9bbfbe3
AA		 1045 if (unlikely(!pages[i] ||
1046 mem_cgroup_newpage_charge(pages[i], mm,
1047 GFP_KERNEL))) {
1048 if (pages[i])
71e3aac0 1049 put_page(pages[i]);
b9bbfbe3
AA		 1050 mem_cgroup_uncharge_start();
1051 while (--i >= 0) {
1052 mem_cgroup_uncharge_page(pages[i]);
1053 put_page(pages[i]);
1054 }
1055 mem_cgroup_uncharge_end();
71e3aac0
AA		 1056 kfree(pages);
1057 ret |= VM_FAULT_OOM;
1058 goto out;
1059 }
1060 }
1061
1062 for (i = 0; i < HPAGE_PMD_NR; i++) {
1063 copy_user_highpage(pages[i], page + i,
0089e485 1064 haddr + PAGE_SIZE * i, vma);
71e3aac0
AA		 1065 __SetPageUptodate(pages[i]);
1066 cond_resched();
1067 }
1068
2ec74c3e
SG		 1069 mmun_start = haddr;
1070 mmun_end = haddr + HPAGE_PMD_SIZE;
1071 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1072
c4088ebd 1073 ptl = pmd_lock(mm, pmd);
71e3aac0
AA		 1074 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1075 goto out_free_pages;
309381fe 1076 VM_BUG_ON_PAGE(!PageHead(page), page);
71e3aac0 1077
2ec74c3e 1078 pmdp_clear_flush(vma, haddr, pmd);
71e3aac0
AA		 1079 /* leave pmd empty until pte is filled */
1080
6b0b50b0 1081 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
71e3aac0
AA		 1082 pmd_populate(mm, &_pmd, pgtable);
1083
1084 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1085 pte_t *pte, entry;
1086 entry = mk_pte(pages[i], vma->vm_page_prot);
1087 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1088 page_add_new_anon_rmap(pages[i], vma, haddr);
1089 pte = pte_offset_map(&_pmd, haddr);
1090 VM_BUG_ON(!pte_none(*pte));
1091 set_pte_at(mm, haddr, pte, entry);
1092 pte_unmap(pte);
1093 }
1094 kfree(pages);
1095
71e3aac0
AA		 1096 smp_wmb(); /* make pte visible before pmd */
1097 pmd_populate(mm, pmd, pgtable);
1098 page_remove_rmap(page);
c4088ebd 1099 spin_unlock(ptl);
71e3aac0 1100
2ec74c3e
SG		 1101 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1102
71e3aac0
AA		 1103 ret |= VM_FAULT_WRITE;
1104 put_page(page);
1105
1106out:
1107 return ret;
1108
1109out_free_pages:
c4088ebd 1110 spin_unlock(ptl);
2ec74c3e 1111 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
b9bbfbe3
AA		 1112 mem_cgroup_uncharge_start();
1113 for (i = 0; i < HPAGE_PMD_NR; i++) {
1114 mem_cgroup_uncharge_page(pages[i]);
71e3aac0 1115 put_page(pages[i]);
b9bbfbe3
AA		 1116 }
1117 mem_cgroup_uncharge_end();
71e3aac0
AA		 1118 kfree(pages);
1119 goto out;
1120}
1121
1122int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1123 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1124{
c4088ebd 1125 spinlock_t *ptl;
71e3aac0 1126 int ret = 0;
93b4796d 1127 struct page *page = NULL, *new_page;
71e3aac0 1128 unsigned long haddr;
2ec74c3e
SG		 1129 unsigned long mmun_start; /* For mmu_notifiers */
1130 unsigned long mmun_end; /* For mmu_notifiers */
71e3aac0 1131
c4088ebd 1132 ptl = pmd_lockptr(mm, pmd);
71e3aac0 1133 VM_BUG_ON(!vma->anon_vma);
93b4796d
KS		 1134 haddr = address & HPAGE_PMD_MASK;
1135 if (is_huge_zero_pmd(orig_pmd))
1136 goto alloc;
c4088ebd 1137 spin_lock(ptl);
71e3aac0
AA		 1138 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1139 goto out_unlock;
1140
1141 page = pmd_page(orig_pmd);
309381fe 1142 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
71e3aac0
AA		 1143 if (page_mapcount(page) == 1) {
1144 pmd_t entry;
1145 entry = pmd_mkyoung(orig_pmd);
1146 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1147 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
b113da65 1148 update_mmu_cache_pmd(vma, address, pmd);
71e3aac0
AA		 1149 ret |= VM_FAULT_WRITE;
1150 goto out_unlock;
1151 }
1152 get_page(page);
c4088ebd 1153 spin_unlock(ptl);
93b4796d 1154alloc:
71e3aac0
AA		 1155 if (transparent_hugepage_enabled(vma) &&
1156 !transparent_hugepage_debug_cow())
0bbbc0b3 1157 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
cc5d462f 1158 vma, haddr, numa_node_id(), 0);
71e3aac0
AA		 1159 else
1160 new_page = NULL;
1161
1162 if (unlikely(!new_page)) {
eecc1e42 1163 if (!page) {
93b4796d 1164 ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
3ea41e62 1165 address, pmd, orig_pmd, haddr);
93b4796d
KS		 1166 } else {
1167 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1168 pmd, orig_pmd, page, haddr);
1169 if (ret & VM_FAULT_OOM)
1170 split_huge_page(page);
1171 put_page(page);
1172 }
17766dde 1173 count_vm_event(THP_FAULT_FALLBACK);
71e3aac0
AA		 1174 goto out;
1175 }
1176
b9bbfbe3
AA		 1177 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1178 put_page(new_page);
93b4796d
KS		 1179 if (page) {
1180 split_huge_page(page);
1181 put_page(page);
1182 }
17766dde 1183 count_vm_event(THP_FAULT_FALLBACK);
b9bbfbe3
AA		 1184 ret |= VM_FAULT_OOM;
1185 goto out;
1186 }
1187
17766dde
DR		 1188 count_vm_event(THP_FAULT_ALLOC);
1189
eecc1e42 1190 if (!page)
93b4796d
KS		 1191 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1192 else
1193 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
71e3aac0
AA		 1194 __SetPageUptodate(new_page);
1195
2ec74c3e
SG		 1196 mmun_start = haddr;
1197 mmun_end = haddr + HPAGE_PMD_SIZE;
1198 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1199
c4088ebd 1200 spin_lock(ptl);
93b4796d
KS		 1201 if (page)
1202 put_page(page);
b9bbfbe3 1203 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
c4088ebd 1204 spin_unlock(ptl);
b9bbfbe3 1205 mem_cgroup_uncharge_page(new_page);
71e3aac0 1206 put_page(new_page);
2ec74c3e 1207 goto out_mn;
b9bbfbe3 1208 } else {
71e3aac0 1209 pmd_t entry;
3122359a
KS		 1210 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1211 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2ec74c3e 1212 pmdp_clear_flush(vma, haddr, pmd);
71e3aac0
AA		 1213 page_add_new_anon_rmap(new_page, vma, haddr);
1214 set_pmd_at(mm, haddr, pmd, entry);
b113da65 1215 update_mmu_cache_pmd(vma, address, pmd);
eecc1e42 1216 if (!page) {
93b4796d 1217 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
97ae1749
KS		 1218 put_huge_zero_page();
1219 } else {
309381fe 1220 VM_BUG_ON_PAGE(!PageHead(page), page);
93b4796d
KS		 1221 page_remove_rmap(page);
1222 put_page(page);
1223 }
71e3aac0
AA		 1224 ret |= VM_FAULT_WRITE;
1225 }
c4088ebd 1226 spin_unlock(ptl);
2ec74c3e
SG		 1227out_mn:
1228 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 1229out:
1230 return ret;
2ec74c3e 1231out_unlock:
c4088ebd 1232 spin_unlock(ptl);
2ec74c3e 1233 return ret;
71e3aac0
AA		 1234}
1235
b676b293 1236struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
71e3aac0
AA		 1237 unsigned long addr,
1238 pmd_t *pmd,
1239 unsigned int flags)
1240{
b676b293 1241 struct mm_struct *mm = vma->vm_mm;
71e3aac0
AA		 1242 struct page *page = NULL;
1243
c4088ebd 1244 assert_spin_locked(pmd_lockptr(mm, pmd));
71e3aac0
AA		 1245
1246 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1247 goto out;
1248
85facf25
KS		 1249 /* Avoid dumping huge zero page */
1250 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1251 return ERR_PTR(-EFAULT);
1252
2b4847e7
MG		 1253 /* Full NUMA hinting faults to serialise migration in fault paths */
1254 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1255 goto out;
1256
71e3aac0 1257 page = pmd_page(*pmd);
309381fe 1258 VM_BUG_ON_PAGE(!PageHead(page), page);
71e3aac0
AA		 1259 if (flags & FOLL_TOUCH) {
1260 pmd_t _pmd;
1261 /*
1262 * We should set the dirty bit only for FOLL_WRITE but
1263 * for now the dirty bit in the pmd is meaningless.
1264 * And if the dirty bit will become meaningful and
1265 * we'll only set it with FOLL_WRITE, an atomic
1266 * set_bit will be required on the pmd to set the
1267 * young bit, instead of the current set_pmd_at.
1268 */
1269 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
8663890a
AK		 1270 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1271 pmd, _pmd, 1))
1272 update_mmu_cache_pmd(vma, addr, pmd);
71e3aac0 1273 }
b676b293
DR		 1274 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1275 if (page->mapping && trylock_page(page)) {
1276 lru_add_drain();
1277 if (page->mapping)
1278 mlock_vma_page(page);
1279 unlock_page(page);
1280 }
1281 }
71e3aac0 1282 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
309381fe 1283 VM_BUG_ON_PAGE(!PageCompound(page), page);
71e3aac0 1284 if (flags & FOLL_GET)
70b50f94 1285 get_page_foll(page);
71e3aac0
AA		 1286
1287out:
1288 return page;
1289}
1290
d10e63f2 1291/* NUMA hinting page fault entry point for trans huge pmds */
4daae3b4
MG		 1292int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1293 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
d10e63f2 1294{
c4088ebd 1295 spinlock_t *ptl;
b8916634 1296 struct anon_vma *anon_vma = NULL;
b32967ff 1297 struct page *page;
d10e63f2 1298 unsigned long haddr = addr & HPAGE_PMD_MASK;
8191acbd 1299 int page_nid = -1, this_nid = numa_node_id();
90572890 1300 int target_nid, last_cpupid = -1;
8191acbd
MG		 1301 bool page_locked;
1302 bool migrated = false;
6688cc05 1303 int flags = 0;
d10e63f2 1304
c4088ebd 1305 ptl = pmd_lock(mm, pmdp);
d10e63f2
MG		 1306 if (unlikely(!pmd_same(pmd, *pmdp)))
1307 goto out_unlock;
1308
de466bd6
MG		 1309 /*
1310 * If there are potential migrations, wait for completion and retry
1311 * without disrupting NUMA hinting information. Do not relock and
1312 * check_same as the page may no longer be mapped.
1313 */
1314 if (unlikely(pmd_trans_migrating(*pmdp))) {
1315 spin_unlock(ptl);
1316 wait_migrate_huge_page(vma->anon_vma, pmdp);
1317 goto out;
1318 }
1319
d10e63f2 1320 page = pmd_page(pmd);
a1a46184 1321 BUG_ON(is_huge_zero_page(page));
8191acbd 1322 page_nid = page_to_nid(page);
90572890 1323 last_cpupid = page_cpupid_last(page);
03c5a6e1 1324 count_vm_numa_event(NUMA_HINT_FAULTS);
04bb2f94 1325 if (page_nid == this_nid) {
03c5a6e1 1326 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
04bb2f94
RR		 1327 flags |= TNF_FAULT_LOCAL;
1328 }
4daae3b4 1329
6688cc05
PZ		 1330 /*
1331 * Avoid grouping on DSO/COW pages in specific and RO pages
1332 * in general, RO pages shouldn't hurt as much anyway since
1333 * they can be in shared cache state.
1334 */
1335 if (!pmd_write(pmd))
1336 flags |= TNF_NO_GROUP;
1337
ff9042b1
MG		 1338 /*
1339 * Acquire the page lock to serialise THP migrations but avoid dropping
1340 * page_table_lock if at all possible
1341 */
b8916634
MG		 1342 page_locked = trylock_page(page);
1343 target_nid = mpol_misplaced(page, vma, haddr);
1344 if (target_nid == -1) {
1345 /* If the page was locked, there are no parallel migrations */
a54a407f 1346 if (page_locked)
b8916634 1347 goto clear_pmdnuma;
2b4847e7 1348 }
4daae3b4 1349
de466bd6 1350 /* Migration could have started since the pmd_trans_migrating check */
2b4847e7 1351 if (!page_locked) {
c4088ebd 1352 spin_unlock(ptl);
b8916634 1353 wait_on_page_locked(page);
a54a407f 1354 page_nid = -1;
b8916634
MG		 1355 goto out;
1356 }
1357
2b4847e7
MG		 1358 /*
1359 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1360 * to serialises splits
1361 */
b8916634 1362 get_page(page);
c4088ebd 1363 spin_unlock(ptl);
b8916634 1364 anon_vma = page_lock_anon_vma_read(page);
4daae3b4 1365
c69307d5 1366 /* Confirm the PMD did not change while page_table_lock was released */
c4088ebd 1367 spin_lock(ptl);
b32967ff
MG		 1368 if (unlikely(!pmd_same(pmd, *pmdp))) {
1369 unlock_page(page);
1370 put_page(page);
a54a407f 1371 page_nid = -1;
4daae3b4 1372 goto out_unlock;
b32967ff 1373 }
ff9042b1 1374
c3a489ca
MG		 1375 /* Bail if we fail to protect against THP splits for any reason */
1376 if (unlikely(!anon_vma)) {
1377 put_page(page);
1378 page_nid = -1;
1379 goto clear_pmdnuma;
1380 }
1381
a54a407f
MG		 1382 /*
1383 * Migrate the THP to the requested node, returns with page unlocked
1384 * and pmd_numa cleared.
1385 */
c4088ebd 1386 spin_unlock(ptl);
b32967ff 1387 migrated = migrate_misplaced_transhuge_page(mm, vma,
340ef390 1388 pmdp, pmd, addr, page, target_nid);
6688cc05
PZ		 1389 if (migrated) {
1390 flags |= TNF_MIGRATED;
8191acbd 1391 page_nid = target_nid;
6688cc05 1392 }
b32967ff 1393
8191acbd 1394 goto out;
b32967ff 1395clear_pmdnuma:
a54a407f 1396 BUG_ON(!PageLocked(page));
d10e63f2
MG		 1397 pmd = pmd_mknonnuma(pmd);
1398 set_pmd_at(mm, haddr, pmdp, pmd);
1399 VM_BUG_ON(pmd_numa(*pmdp));
1400 update_mmu_cache_pmd(vma, addr, pmdp);
a54a407f 1401 unlock_page(page);
d10e63f2 1402out_unlock:
c4088ebd 1403 spin_unlock(ptl);
b8916634
MG		 1404
1405out:
1406 if (anon_vma)
1407 page_unlock_anon_vma_read(anon_vma);
1408
8191acbd 1409 if (page_nid != -1)
6688cc05 1410 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
8191acbd 1411
d10e63f2
MG		 1412 return 0;
1413}
1414
71e3aac0 1415int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
f21760b1 1416 pmd_t *pmd, unsigned long addr)
71e3aac0 1417{
bf929152 1418 spinlock_t *ptl;
71e3aac0
AA		 1419 int ret = 0;
1420
bf929152 1421 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
025c5b24
NH		 1422 struct page *page;
1423 pgtable_t pgtable;
f5c8ad47 1424 pmd_t orig_pmd;
a6bf2bb0
AK		 1425 /*
1426 * For architectures like ppc64 we look at deposited pgtable
1427 * when calling pmdp_get_and_clear. So do the
1428 * pgtable_trans_huge_withdraw after finishing pmdp related
1429 * operations.
1430 */
f5c8ad47 1431 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
025c5b24 1432 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
a6bf2bb0 1433 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
479f0abb 1434 if (is_huge_zero_pmd(orig_pmd)) {
e1f56c89 1435 atomic_long_dec(&tlb->mm->nr_ptes);
bf929152 1436 spin_unlock(ptl);
97ae1749 1437 put_huge_zero_page();
479f0abb
KS		 1438 } else {
1439 page = pmd_page(orig_pmd);
1440 page_remove_rmap(page);
309381fe 1441 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
479f0abb 1442 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
309381fe 1443 VM_BUG_ON_PAGE(!PageHead(page), page);
e1f56c89 1444 atomic_long_dec(&tlb->mm->nr_ptes);
bf929152 1445 spin_unlock(ptl);
479f0abb
KS		 1446 tlb_remove_page(tlb, page);
1447 }
025c5b24
NH		 1448 pte_free(tlb->mm, pgtable);
1449 ret = 1;
1450 }
71e3aac0
AA		 1451 return ret;
1452}
1453
0ca1634d
JW		 1454int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1455 unsigned long addr, unsigned long end,
1456 unsigned char *vec)
1457{
bf929152 1458 spinlock_t *ptl;
0ca1634d
JW		 1459 int ret = 0;
1460
bf929152 1461 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
025c5b24
NH		 1462 /*
1463 * All logical pages in the range are present
1464 * if backed by a huge page.
1465 */
bf929152 1466 spin_unlock(ptl);
025c5b24
NH		 1467 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1468 ret = 1;
1469 }
0ca1634d
JW		 1470
1471 return ret;
1472}
1473
37a1c49a
AA		 1474int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1475 unsigned long old_addr,
1476 unsigned long new_addr, unsigned long old_end,
1477 pmd_t *old_pmd, pmd_t *new_pmd)
1478{
bf929152 1479 spinlock_t *old_ptl, *new_ptl;
37a1c49a
AA		 1480 int ret = 0;
1481 pmd_t pmd;
1482
1483 struct mm_struct *mm = vma->vm_mm;
1484
1485 if ((old_addr & ~HPAGE_PMD_MASK) ||
1486 (new_addr & ~HPAGE_PMD_MASK) ||
1487 old_end - old_addr < HPAGE_PMD_SIZE ||
1488 (new_vma->vm_flags & VM_NOHUGEPAGE))
1489 goto out;
1490
1491 /*
1492 * The destination pmd shouldn't be established, free_pgtables()
1493 * should have release it.
1494 */
1495 if (WARN_ON(!pmd_none(*new_pmd))) {
1496 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1497 goto out;
1498 }
1499
bf929152
KS		 1500 /*
1501 * We don't have to worry about the ordering of src and dst
1502 * ptlocks because exclusive mmap_sem prevents deadlock.
1503 */
1504 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
025c5b24 1505 if (ret == 1) {
bf929152
KS		 1506 new_ptl = pmd_lockptr(mm, new_pmd);
1507 if (new_ptl != old_ptl)
1508 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
025c5b24
NH		 1509 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1510 VM_BUG_ON(!pmd_none(*new_pmd));
0f8975ec 1511 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
3592806c
KS		 1512 if (new_ptl != old_ptl) {
1513 pgtable_t pgtable;
1514
1515 /*
1516 * Move preallocated PTE page table if new_pmd is on
1517 * different PMD page table.
1518 */
1519 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1520 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1521
bf929152 1522 spin_unlock(new_ptl);
3592806c 1523 }
bf929152 1524 spin_unlock(old_ptl);
37a1c49a
AA		 1525 }
1526out:
1527 return ret;
1528}
1529
f123d74a
MG		 1530/*
1531 * Returns
1532 * - 0 if PMD could not be locked
1533 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1534 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1535 */
cd7548ab 1536int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
4b10e7d5 1537 unsigned long addr, pgprot_t newprot, int prot_numa)
cd7548ab
JW		 1538{
1539 struct mm_struct *mm = vma->vm_mm;
bf929152 1540 spinlock_t *ptl;
cd7548ab
JW		 1541 int ret = 0;
1542
bf929152 1543 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
025c5b24 1544 pmd_t entry;
f123d74a 1545 ret = 1;
a4f1de17 1546 if (!prot_numa) {
f123d74a 1547 entry = pmdp_get_and_clear(mm, addr, pmd);
1667918b
MG		 1548 if (pmd_numa(entry))
1549 entry = pmd_mknonnuma(entry);
4b10e7d5 1550 entry = pmd_modify(entry, newprot);
f123d74a 1551 ret = HPAGE_PMD_NR;
a4f1de17
HD		 1552 BUG_ON(pmd_write(entry));
1553 } else {
4b10e7d5
MG		 1554 struct page *page = pmd_page(*pmd);
1555
a1a46184 1556 /*
1bc115d8
MG		 1557 * Do not trap faults against the zero page. The
1558 * read-only data is likely to be read-cached on the
1559 * local CPU cache and it is less useful to know about
1560 * local vs remote hits on the zero page.
a1a46184 1561 */
1bc115d8 1562 if (!is_huge_zero_page(page) &&
4b10e7d5 1563 !pmd_numa(*pmd)) {
5a6dac3e 1564 entry = *pmd;
4b10e7d5 1565 entry = pmd_mknuma(entry);
f123d74a 1566 ret = HPAGE_PMD_NR;
4b10e7d5
MG		 1567 }
1568 }
f123d74a
MG		 1569
1570 /* Set PMD if cleared earlier */
1571 if (ret == HPAGE_PMD_NR)
1572 set_pmd_at(mm, addr, pmd, entry);
1573
bf929152 1574 spin_unlock(ptl);
025c5b24
NH		 1575 }
1576
1577 return ret;
1578}
1579
1580/*
1581 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1582 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1583 *
1584 * Note that if it returns 1, this routine returns without unlocking page
1585 * table locks. So callers must unlock them.
1586 */
bf929152
KS		 1587int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1588 spinlock_t **ptl)
025c5b24 1589{
bf929152 1590 *ptl = pmd_lock(vma->vm_mm, pmd);
cd7548ab
JW		 1591 if (likely(pmd_trans_huge(*pmd))) {
1592 if (unlikely(pmd_trans_splitting(*pmd))) {
bf929152 1593 spin_unlock(*ptl);
cd7548ab 1594 wait_split_huge_page(vma->anon_vma, pmd);
025c5b24 1595 return -1;
cd7548ab 1596 } else {
025c5b24
NH		 1597 /* Thp mapped by 'pmd' is stable, so we can
1598 * handle it as it is. */
1599 return 1;
cd7548ab 1600 }
025c5b24 1601 }
bf929152 1602 spin_unlock(*ptl);
025c5b24 1603 return 0;
cd7548ab
JW		 1604}
1605
117b0791
KS		 1606/*
1607 * This function returns whether a given @page is mapped onto the @address
1608 * in the virtual space of @mm.
1609 *
1610 * When it's true, this function returns *pmd with holding the page table lock
1611 * and passing it back to the caller via @ptl.
1612 * If it's false, returns NULL without holding the page table lock.
1613 */
71e3aac0
AA		 1614pmd_t *page_check_address_pmd(struct page *page,
1615 struct mm_struct *mm,
1616 unsigned long address,
117b0791
KS		 1617 enum page_check_address_pmd_flag flag,
1618 spinlock_t **ptl)
71e3aac0 1619{
117b0791 1620 pmd_t *pmd;
71e3aac0
AA		 1621
1622 if (address & ~HPAGE_PMD_MASK)
117b0791 1623 return NULL;
71e3aac0 1624
6219049a
BL		 1625 pmd = mm_find_pmd(mm, address);
1626 if (!pmd)
117b0791
KS		 1627 return NULL;
1628 *ptl = pmd_lock(mm, pmd);
71e3aac0 1629 if (pmd_none(*pmd))
117b0791 1630 goto unlock;
71e3aac0 1631 if (pmd_page(*pmd) != page)
117b0791 1632 goto unlock;
94fcc585
AA		 1633 /*
1634 * split_vma() may create temporary aliased mappings. There is
1635 * no risk as long as all huge pmd are found and have their
1636 * splitting bit set before __split_huge_page_refcount
1637 * runs. Finding the same huge pmd more than once during the
1638 * same rmap walk is not a problem.
1639 */
1640 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1641 pmd_trans_splitting(*pmd))
117b0791 1642 goto unlock;
71e3aac0
AA		 1643 if (pmd_trans_huge(*pmd)) {
1644 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1645 !pmd_trans_splitting(*pmd));
117b0791 1646 return pmd;
71e3aac0 1647 }
117b0791
KS		 1648unlock:
1649 spin_unlock(*ptl);
1650 return NULL;
71e3aac0
AA		 1651}
1652
1653static int __split_huge_page_splitting(struct page *page,
1654 struct vm_area_struct *vma,
1655 unsigned long address)
1656{
1657 struct mm_struct *mm = vma->vm_mm;
117b0791 1658 spinlock_t *ptl;
71e3aac0
AA		 1659 pmd_t *pmd;
1660 int ret = 0;
2ec74c3e
SG		 1661 /* For mmu_notifiers */
1662 const unsigned long mmun_start = address;
1663 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
71e3aac0 1664
2ec74c3e 1665 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
71e3aac0 1666 pmd = page_check_address_pmd(page, mm, address,
117b0791 1667 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
71e3aac0
AA		 1668 if (pmd) {
1669 /*
1670 * We can't temporarily set the pmd to null in order
1671 * to split it, the pmd must remain marked huge at all
1672 * times or the VM won't take the pmd_trans_huge paths
5a505085 1673 * and it won't wait on the anon_vma->root->rwsem to
71e3aac0
AA		 1674 * serialize against split_huge_page*.
1675 */
2ec74c3e 1676 pmdp_splitting_flush(vma, address, pmd);
71e3aac0 1677 ret = 1;
117b0791 1678 spin_unlock(ptl);
71e3aac0 1679 }
2ec74c3e 1680 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 1681
1682 return ret;
1683}
1684
5bc7b8ac
SL		 1685static void __split_huge_page_refcount(struct page *page,
1686 struct list_head *list)
71e3aac0
AA		 1687{
1688 int i;
71e3aac0 1689 struct zone *zone = page_zone(page);
fa9add64 1690 struct lruvec *lruvec;
70b50f94 1691 int tail_count = 0;
71e3aac0
AA		 1692
1693 /* prevent PageLRU to go away from under us, and freeze lru stats */
1694 spin_lock_irq(&zone->lru_lock);
fa9add64
HD		 1695 lruvec = mem_cgroup_page_lruvec(page, zone);
1696
71e3aac0 1697 compound_lock(page);
e94c8a9c
KH		 1698 /* complete memcg works before add pages to LRU */
1699 mem_cgroup_split_huge_fixup(page);
71e3aac0 1700
45676885 1701 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
71e3aac0
AA		 1702 struct page *page_tail = page + i;
1703
70b50f94
AA		 1704 /* tail_page->_mapcount cannot change */
1705 BUG_ON(page_mapcount(page_tail) < 0);
1706 tail_count += page_mapcount(page_tail);
1707 /* check for overflow */
1708 BUG_ON(tail_count < 0);
1709 BUG_ON(atomic_read(&page_tail->_count) != 0);
1710 /*
1711 * tail_page->_count is zero and not changing from
1712 * under us. But get_page_unless_zero() may be running
1713 * from under us on the tail_page. If we used
1714 * atomic_set() below instead of atomic_add(), we
1715 * would then run atomic_set() concurrently with
1716 * get_page_unless_zero(), and atomic_set() is
1717 * implemented in C not using locked ops. spin_unlock
1718 * on x86 sometime uses locked ops because of PPro
1719 * errata 66, 92, so unless somebody can guarantee
1720 * atomic_set() here would be safe on all archs (and
1721 * not only on x86), it's safer to use atomic_add().
1722 */
1723 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1724 &page_tail->_count);
71e3aac0
AA		 1725
1726 /* after clearing PageTail the gup refcount can be released */
1727 smp_mb();
1728
a6d30ddd
JD		 1729 /*
1730 * retain hwpoison flag of the poisoned tail page:
1731 * fix for the unsuitable process killed on Guest Machine(KVM)
1732 * by the memory-failure.
1733 */
1734 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
71e3aac0
AA		 1735 page_tail->flags |= (page->flags &
1736 ((1L << PG_referenced) |
1737 (1L << PG_swapbacked) |
1738 (1L << PG_mlocked) |
e180cf80
KS		 1739 (1L << PG_uptodate) |
1740 (1L << PG_active) |
1741 (1L << PG_unevictable)));
71e3aac0
AA		 1742 page_tail->flags |= (1L << PG_dirty);
1743
70b50f94 1744 /* clear PageTail before overwriting first_page */
71e3aac0
AA		 1745 smp_wmb();
1746
1747 /*
1748 * __split_huge_page_splitting() already set the
1749 * splitting bit in all pmd that could map this
1750 * hugepage, that will ensure no CPU can alter the
1751 * mapcount on the head page. The mapcount is only
1752 * accounted in the head page and it has to be
1753 * transferred to all tail pages in the below code. So
1754 * for this code to be safe, the split the mapcount
1755 * can't change. But that doesn't mean userland can't
1756 * keep changing and reading the page contents while
1757 * we transfer the mapcount, so the pmd splitting
1758 * status is achieved setting a reserved bit in the
1759 * pmd, not by clearing the present bit.
1760 */
71e3aac0
AA		 1761 page_tail->_mapcount = page->_mapcount;
1762
1763 BUG_ON(page_tail->mapping);
1764 page_tail->mapping = page->mapping;
1765
45676885 1766 page_tail->index = page->index + i;
90572890 1767 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
71e3aac0
AA		 1768
1769 BUG_ON(!PageAnon(page_tail));
1770 BUG_ON(!PageUptodate(page_tail));
1771 BUG_ON(!PageDirty(page_tail));
1772 BUG_ON(!PageSwapBacked(page_tail));
1773
5bc7b8ac 1774 lru_add_page_tail(page, page_tail, lruvec, list);
71e3aac0 1775 }
70b50f94
AA		 1776 atomic_sub(tail_count, &page->_count);
1777 BUG_ON(atomic_read(&page->_count) <= 0);
71e3aac0 1778
fa9add64 1779 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
79134171 1780
71e3aac0
AA		 1781 ClearPageCompound(page);
1782 compound_unlock(page);
1783 spin_unlock_irq(&zone->lru_lock);
1784
1785 for (i = 1; i < HPAGE_PMD_NR; i++) {
1786 struct page *page_tail = page + i;
1787 BUG_ON(page_count(page_tail) <= 0);
1788 /*
1789 * Tail pages may be freed if there wasn't any mapping
1790 * like if add_to_swap() is running on a lru page that
1791 * had its mapping zapped. And freeing these pages
1792 * requires taking the lru_lock so we do the put_page
1793 * of the tail pages after the split is complete.
1794 */
1795 put_page(page_tail);
1796 }
1797
1798 /*
1799 * Only the head page (now become a regular page) is required
1800 * to be pinned by the caller.
1801 */
1802 BUG_ON(page_count(page) <= 0);
1803}
1804
1805static int __split_huge_page_map(struct page *page,
1806 struct vm_area_struct *vma,
1807 unsigned long address)
1808{
1809 struct mm_struct *mm = vma->vm_mm;
117b0791 1810 spinlock_t *ptl;
71e3aac0
AA		 1811 pmd_t *pmd, _pmd;
1812 int ret = 0, i;
1813 pgtable_t pgtable;
1814 unsigned long haddr;
1815
71e3aac0 1816 pmd = page_check_address_pmd(page, mm, address,
117b0791 1817 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
71e3aac0 1818 if (pmd) {
6b0b50b0 1819 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
71e3aac0
AA		 1820 pmd_populate(mm, &_pmd, pgtable);
1821
e3ebcf64
GS		 1822 haddr = address;
1823 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
71e3aac0
AA		 1824 pte_t *pte, entry;
1825 BUG_ON(PageCompound(page+i));
1826 entry = mk_pte(page + i, vma->vm_page_prot);
1827 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1828 if (!pmd_write(*pmd))
1829 entry = pte_wrprotect(entry);
1830 else
1831 BUG_ON(page_mapcount(page) != 1);
1832 if (!pmd_young(*pmd))
1833 entry = pte_mkold(entry);
1ba6e0b5
AA		 1834 if (pmd_numa(*pmd))
1835 entry = pte_mknuma(entry);
71e3aac0
AA		 1836 pte = pte_offset_map(&_pmd, haddr);
1837 BUG_ON(!pte_none(*pte));
1838 set_pte_at(mm, haddr, pte, entry);
1839 pte_unmap(pte);
1840 }
1841
71e3aac0
AA		 1842 smp_wmb(); /* make pte visible before pmd */
1843 /*
1844 * Up to this point the pmd is present and huge and
1845 * userland has the whole access to the hugepage
1846 * during the split (which happens in place). If we
1847 * overwrite the pmd with the not-huge version
1848 * pointing to the pte here (which of course we could
1849 * if all CPUs were bug free), userland could trigger
1850 * a small page size TLB miss on the small sized TLB
1851 * while the hugepage TLB entry is still established
1852 * in the huge TLB. Some CPU doesn't like that. See
1853 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1854 * Erratum 383 on page 93. Intel should be safe but is
1855 * also warns that it's only safe if the permission
1856 * and cache attributes of the two entries loaded in
1857 * the two TLB is identical (which should be the case
1858 * here). But it is generally safer to never allow
1859 * small and huge TLB entries for the same virtual
1860 * address to be loaded simultaneously. So instead of
1861 * doing "pmd_populate(); flush_tlb_range();" we first
1862 * mark the current pmd notpresent (atomically because
1863 * here the pmd_trans_huge and pmd_trans_splitting
1864 * must remain set at all times on the pmd until the
1865 * split is complete for this pmd), then we flush the
1866 * SMP TLB and finally we write the non-huge version
1867 * of the pmd entry with pmd_populate.
1868 */
46dcde73 1869 pmdp_invalidate(vma, address, pmd);
71e3aac0
AA		 1870 pmd_populate(mm, pmd, pgtable);
1871 ret = 1;
117b0791 1872 spin_unlock(ptl);
71e3aac0 1873 }
71e3aac0
AA		 1874
1875 return ret;
1876}
1877
5a505085 1878/* must be called with anon_vma->root->rwsem held */
71e3aac0 1879static void __split_huge_page(struct page *page,
5bc7b8ac
SL		 1880 struct anon_vma *anon_vma,
1881 struct list_head *list)
71e3aac0
AA		 1882{
1883 int mapcount, mapcount2;
bf181b9f 1884 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
71e3aac0
AA		 1885 struct anon_vma_chain *avc;
1886
1887 BUG_ON(!PageHead(page));
1888 BUG_ON(PageTail(page));
1889
1890 mapcount = 0;
bf181b9f 1891 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
71e3aac0
AA		 1892 struct vm_area_struct *vma = avc->vma;
1893 unsigned long addr = vma_address(page, vma);
1894 BUG_ON(is_vma_temporary_stack(vma));
71e3aac0
AA		 1895 mapcount += __split_huge_page_splitting(page, vma, addr);
1896 }
05759d38
AA		 1897 /*
1898 * It is critical that new vmas are added to the tail of the
1899 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1900 * and establishes a child pmd before
1901 * __split_huge_page_splitting() freezes the parent pmd (so if
1902 * we fail to prevent copy_huge_pmd() from running until the
1903 * whole __split_huge_page() is complete), we will still see
1904 * the newly established pmd of the child later during the
1905 * walk, to be able to set it as pmd_trans_splitting too.
1906 */
1907 if (mapcount != page_mapcount(page))
1908 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1909 mapcount, page_mapcount(page));
71e3aac0
AA		 1910 BUG_ON(mapcount != page_mapcount(page));
1911
5bc7b8ac 1912 __split_huge_page_refcount(page, list);
71e3aac0
AA		 1913
1914 mapcount2 = 0;
bf181b9f 1915 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
71e3aac0
AA		 1916 struct vm_area_struct *vma = avc->vma;
1917 unsigned long addr = vma_address(page, vma);
1918 BUG_ON(is_vma_temporary_stack(vma));
71e3aac0
AA		 1919 mapcount2 += __split_huge_page_map(page, vma, addr);
1920 }
05759d38
AA		 1921 if (mapcount != mapcount2)
1922 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1923 mapcount, mapcount2, page_mapcount(page));
71e3aac0
AA		 1924 BUG_ON(mapcount != mapcount2);
1925}
1926
5bc7b8ac
SL		 1927/*
1928 * Split a hugepage into normal pages. This doesn't change the position of head
1929 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1930 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1931 * from the hugepage.
1932 * Return 0 if the hugepage is split successfully otherwise return 1.
1933 */
1934int split_huge_page_to_list(struct page *page, struct list_head *list)
71e3aac0
AA		 1935{
1936 struct anon_vma *anon_vma;
1937 int ret = 1;
1938
5918d10a 1939 BUG_ON(is_huge_zero_page(page));
71e3aac0 1940 BUG_ON(!PageAnon(page));
062f1af2
MG		 1941
1942 /*
1943 * The caller does not necessarily hold an mmap_sem that would prevent
1944 * the anon_vma disappearing so we first we take a reference to it
1945 * and then lock the anon_vma for write. This is similar to
1946 * page_lock_anon_vma_read except the write lock is taken to serialise
1947 * against parallel split or collapse operations.
1948 */
1949 anon_vma = page_get_anon_vma(page);
71e3aac0
AA		 1950 if (!anon_vma)
1951 goto out;
062f1af2
MG		 1952 anon_vma_lock_write(anon_vma);
1953
71e3aac0
AA		 1954 ret = 0;
1955 if (!PageCompound(page))
1956 goto out_unlock;
1957
1958 BUG_ON(!PageSwapBacked(page));
5bc7b8ac 1959 __split_huge_page(page, anon_vma, list);
81ab4201 1960 count_vm_event(THP_SPLIT);
71e3aac0
AA		 1961
1962 BUG_ON(PageCompound(page));
1963out_unlock:
08b52706 1964 anon_vma_unlock_write(anon_vma);
062f1af2 1965 put_anon_vma(anon_vma);
71e3aac0
AA		 1966out:
1967 return ret;
1968}
1969
4b6e1e37 1970#define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
78f11a25 1971
60ab3244
AA		 1972int hugepage_madvise(struct vm_area_struct *vma,
1973 unsigned long *vm_flags, int advice)
0af4e98b 1974{
8e72033f
GS		 1975 struct mm_struct *mm = vma->vm_mm;
1976
a664b2d8
AA		 1977 switch (advice) {
1978 case MADV_HUGEPAGE:
1979 /*
1980 * Be somewhat over-protective like KSM for now!
1981 */
78f11a25 1982 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
a664b2d8 1983 return -EINVAL;
8e72033f
GS		 1984 if (mm->def_flags & VM_NOHUGEPAGE)
1985 return -EINVAL;
a664b2d8
AA		 1986 *vm_flags &= ~VM_NOHUGEPAGE;
1987 *vm_flags |= VM_HUGEPAGE;
60ab3244
AA		 1988 /*
1989 * If the vma become good for khugepaged to scan,
1990 * register it here without waiting a page fault that
1991 * may not happen any time soon.
1992 */
1993 if (unlikely(khugepaged_enter_vma_merge(vma)))
1994 return -ENOMEM;
a664b2d8
AA		 1995 break;
1996 case MADV_NOHUGEPAGE:
1997 /*
1998 * Be somewhat over-protective like KSM for now!
1999 */
78f11a25 2000 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
a664b2d8
AA		 2001 return -EINVAL;
2002 *vm_flags &= ~VM_HUGEPAGE;
2003 *vm_flags |= VM_NOHUGEPAGE;
60ab3244
AA		 2004 /*
2005 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2006 * this vma even if we leave the mm registered in khugepaged if
2007 * it got registered before VM_NOHUGEPAGE was set.
2008 */
a664b2d8
AA		 2009 break;
2010 }
0af4e98b
AA		 2011
2012 return 0;
2013}
2014
ba76149f
AA		 2015static int __init khugepaged_slab_init(void)
2016{
2017 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2018 sizeof(struct mm_slot),
2019 __alignof__(struct mm_slot), 0, NULL);
2020 if (!mm_slot_cache)
2021 return -ENOMEM;
2022
2023 return 0;
2024}
2025
ba76149f
AA		 2026static inline struct mm_slot *alloc_mm_slot(void)
2027{
2028 if (!mm_slot_cache) /* initialization failed */
2029 return NULL;
2030 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2031}
2032
2033static inline void free_mm_slot(struct mm_slot *mm_slot)
2034{
2035 kmem_cache_free(mm_slot_cache, mm_slot);
2036}
2037
ba76149f
AA		 2038static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2039{
2040 struct mm_slot *mm_slot;
ba76149f 2041
b67bfe0d 2042 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
ba76149f
AA		 2043 if (mm == mm_slot->mm)
2044 return mm_slot;
43b5fbbd 2045
ba76149f
AA		 2046 return NULL;
2047}
2048
2049static void insert_to_mm_slots_hash(struct mm_struct *mm,
2050 struct mm_slot *mm_slot)
2051{
ba76149f 2052 mm_slot->mm = mm;
43b5fbbd 2053 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
ba76149f
AA		 2054}
2055
2056static inline int khugepaged_test_exit(struct mm_struct *mm)
2057{
2058 return atomic_read(&mm->mm_users) == 0;
2059}
2060
2061int __khugepaged_enter(struct mm_struct *mm)
2062{
2063 struct mm_slot *mm_slot;
2064 int wakeup;
2065
2066 mm_slot = alloc_mm_slot();
2067 if (!mm_slot)
2068 return -ENOMEM;
2069
2070 /* __khugepaged_exit() must not run from under us */
2071 VM_BUG_ON(khugepaged_test_exit(mm));
2072 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2073 free_mm_slot(mm_slot);
2074 return 0;
2075 }
2076
2077 spin_lock(&khugepaged_mm_lock);
2078 insert_to_mm_slots_hash(mm, mm_slot);
2079 /*
2080 * Insert just behind the scanning cursor, to let the area settle
2081 * down a little.
2082 */
2083 wakeup = list_empty(&khugepaged_scan.mm_head);
2084 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2085 spin_unlock(&khugepaged_mm_lock);
2086
2087 atomic_inc(&mm->mm_count);
2088 if (wakeup)
2089 wake_up_interruptible(&khugepaged_wait);
2090
2091 return 0;
2092}
2093
2094int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2095{
2096 unsigned long hstart, hend;
2097 if (!vma->anon_vma)
2098 /*
2099 * Not yet faulted in so we will register later in the
2100 * page fault if needed.
2101 */
2102 return 0;
78f11a25 2103 if (vma->vm_ops)
ba76149f
AA		 2104 /* khugepaged not yet working on file or special mappings */
2105 return 0;
b3b9c293 2106 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
ba76149f
AA		 2107 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2108 hend = vma->vm_end & HPAGE_PMD_MASK;
2109 if (hstart < hend)
2110 return khugepaged_enter(vma);
2111 return 0;
2112}
2113
2114void __khugepaged_exit(struct mm_struct *mm)
2115{
2116 struct mm_slot *mm_slot;
2117 int free = 0;
2118
2119 spin_lock(&khugepaged_mm_lock);
2120 mm_slot = get_mm_slot(mm);
2121 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
43b5fbbd 2122 hash_del(&mm_slot->hash);
ba76149f
AA		 2123 list_del(&mm_slot->mm_node);
2124 free = 1;
2125 }
d788e80a 2126 spin_unlock(&khugepaged_mm_lock);
ba76149f
AA		 2127
2128 if (free) {
ba76149f
AA		 2129 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2130 free_mm_slot(mm_slot);
2131 mmdrop(mm);
2132 } else if (mm_slot) {
ba76149f
AA		 2133 /*
2134 * This is required to serialize against
2135 * khugepaged_test_exit() (which is guaranteed to run
2136 * under mmap sem read mode). Stop here (after we
2137 * return all pagetables will be destroyed) until
2138 * khugepaged has finished working on the pagetables
2139 * under the mmap_sem.
2140 */
2141 down_write(&mm->mmap_sem);
2142 up_write(&mm->mmap_sem);
d788e80a 2143 }
ba76149f
AA		 2144}
2145
2146static void release_pte_page(struct page *page)
2147{
2148 /* 0 stands for page_is_file_cache(page) == false */
2149 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2150 unlock_page(page);
2151 putback_lru_page(page);
2152}
2153
2154static void release_pte_pages(pte_t *pte, pte_t *_pte)
2155{
2156 while (--_pte >= pte) {
2157 pte_t pteval = *_pte;
2158 if (!pte_none(pteval))
2159 release_pte_page(pte_page(pteval));
2160 }
2161}
2162
ba76149f
AA		 2163static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2164 unsigned long address,
2165 pte_t *pte)
2166{
2167 struct page *page;
2168 pte_t *_pte;
344aa35c 2169 int referenced = 0, none = 0;
ba76149f
AA		 2170 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2171 _pte++, address += PAGE_SIZE) {
2172 pte_t pteval = *_pte;
2173 if (pte_none(pteval)) {
2174 if (++none <= khugepaged_max_ptes_none)
2175 continue;
344aa35c 2176 else
ba76149f 2177 goto out;
ba76149f 2178 }
344aa35c 2179 if (!pte_present(pteval) || !pte_write(pteval))
ba76149f 2180 goto out;
ba76149f 2181 page = vm_normal_page(vma, address, pteval);
344aa35c 2182 if (unlikely(!page))
ba76149f 2183 goto out;
344aa35c 2184
309381fe
SL		 2185 VM_BUG_ON_PAGE(PageCompound(page), page);
2186 VM_BUG_ON_PAGE(!PageAnon(page), page);
2187 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
ba76149f
AA		 2188
2189 /* cannot use mapcount: can't collapse if there's a gup pin */
344aa35c 2190 if (page_count(page) != 1)
ba76149f 2191 goto out;
ba76149f
AA		 2192 /*
2193 * We can do it before isolate_lru_page because the
2194 * page can't be freed from under us. NOTE: PG_lock
2195 * is needed to serialize against split_huge_page
2196 * when invoked from the VM.
2197 */
344aa35c 2198 if (!trylock_page(page))
ba76149f 2199 goto out;
ba76149f
AA		 2200 /*
2201 * Isolate the page to avoid collapsing an hugepage
2202 * currently in use by the VM.
2203 */
2204 if (isolate_lru_page(page)) {
2205 unlock_page(page);
ba76149f
AA		 2206 goto out;
2207 }
2208 /* 0 stands for page_is_file_cache(page) == false */
2209 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
309381fe
SL		 2210 VM_BUG_ON_PAGE(!PageLocked(page), page);
2211 VM_BUG_ON_PAGE(PageLRU(page), page);
ba76149f
AA		 2212
2213 /* If there is no mapped pte young don't collapse the page */
8ee53820
AA		 2214 if (pte_young(pteval) || PageReferenced(page) ||
2215 mmu_notifier_test_young(vma->vm_mm, address))
ba76149f
AA		 2216 referenced = 1;
2217 }
344aa35c
BL		 2218 if (likely(referenced))
2219 return 1;
ba76149f 2220out:
344aa35c
BL		 2221 release_pte_pages(pte, _pte);
2222 return 0;
ba76149f
AA		 2223}
2224
2225static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2226 struct vm_area_struct *vma,
2227 unsigned long address,
2228 spinlock_t *ptl)
2229{
2230 pte_t *_pte;
2231 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2232 pte_t pteval = *_pte;
2233 struct page *src_page;
2234
2235 if (pte_none(pteval)) {
2236 clear_user_highpage(page, address);
2237 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2238 } else {
2239 src_page = pte_page(pteval);
2240 copy_user_highpage(page, src_page, address, vma);
309381fe 2241 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
ba76149f
AA		 2242 release_pte_page(src_page);
2243 /*
2244 * ptl mostly unnecessary, but preempt has to
2245 * be disabled to update the per-cpu stats
2246 * inside page_remove_rmap().
2247 */
2248 spin_lock(ptl);
2249 /*
2250 * paravirt calls inside pte_clear here are
2251 * superfluous.
2252 */
2253 pte_clear(vma->vm_mm, address, _pte);
2254 page_remove_rmap(src_page);
2255 spin_unlock(ptl);
2256 free_page_and_swap_cache(src_page);
2257 }
2258
2259 address += PAGE_SIZE;
2260 page++;
2261 }
2262}
2263
26234f36 2264static void khugepaged_alloc_sleep(void)
ba76149f 2265{
26234f36
XG		 2266 wait_event_freezable_timeout(khugepaged_wait, false,
2267 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2268}
ba76149f 2269
9f1b868a
BL		 2270static int khugepaged_node_load[MAX_NUMNODES];
2271
26234f36 2272#ifdef CONFIG_NUMA
9f1b868a
BL		 2273static int khugepaged_find_target_node(void)
2274{
2275 static int last_khugepaged_target_node = NUMA_NO_NODE;
2276 int nid, target_node = 0, max_value = 0;
2277
2278 /* find first node with max normal pages hit */
2279 for (nid = 0; nid < MAX_NUMNODES; nid++)
2280 if (khugepaged_node_load[nid] > max_value) {
2281 max_value = khugepaged_node_load[nid];
2282 target_node = nid;
2283 }
2284
2285 /* do some balance if several nodes have the same hit record */
2286 if (target_node <= last_khugepaged_target_node)
2287 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2288 nid++)
2289 if (max_value == khugepaged_node_load[nid]) {
2290 target_node = nid;
2291 break;
2292 }
2293
2294 last_khugepaged_target_node = target_node;
2295 return target_node;
2296}
2297
26234f36
XG		 2298static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2299{
2300 if (IS_ERR(*hpage)) {
2301 if (!*wait)
2302 return false;
2303
2304 *wait = false;
e3b4126c 2305 *hpage = NULL;
26234f36
XG		 2306 khugepaged_alloc_sleep();
2307 } else if (*hpage) {
2308 put_page(*hpage);
2309 *hpage = NULL;
2310 }
2311
2312 return true;
2313}
2314
2315static struct page
2316*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2317 struct vm_area_struct *vma, unsigned long address,
2318 int node)
2319{
309381fe 2320 VM_BUG_ON_PAGE(*hpage, *hpage);
ce83d217
AA		 2321 /*
2322 * Allocate the page while the vma is still valid and under
2323 * the mmap_sem read mode so there is no memory allocation
2324 * later when we take the mmap_sem in write mode. This is more
2325 * friendly behavior (OTOH it may actually hide bugs) to
2326 * filesystems in userland with daemons allocating memory in
2327 * the userland I/O paths. Allocating memory with the
2328 * mmap_sem in read mode is good idea also to allow greater
2329 * scalability.
2330 */
9f1b868a
BL		 2331 *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2332 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
692e0b35
AA		 2333 /*
2334 * After allocating the hugepage, release the mmap_sem read lock in
2335 * preparation for taking it in write mode.
2336 */
2337 up_read(&mm->mmap_sem);
26234f36 2338 if (unlikely(!*hpage)) {
81ab4201 2339 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
ce83d217 2340 *hpage = ERR_PTR(-ENOMEM);
26234f36 2341 return NULL;
ce83d217 2342 }
26234f36 2343
65b3c07b 2344 count_vm_event(THP_COLLAPSE_ALLOC);
26234f36
XG		 2345 return *hpage;
2346}
2347#else
9f1b868a
BL		 2348static int khugepaged_find_target_node(void)
2349{
2350 return 0;
2351}
2352
10dc4155
BL		 2353static inline struct page *alloc_hugepage(int defrag)
2354{
2355 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2356 HPAGE_PMD_ORDER);
2357}
2358
26234f36
XG		 2359static struct page *khugepaged_alloc_hugepage(bool *wait)
2360{
2361 struct page *hpage;
2362
2363 do {
2364 hpage = alloc_hugepage(khugepaged_defrag());
2365 if (!hpage) {
2366 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2367 if (!*wait)
2368 return NULL;
2369
2370 *wait = false;
2371 khugepaged_alloc_sleep();
2372 } else
2373 count_vm_event(THP_COLLAPSE_ALLOC);
2374 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2375
2376 return hpage;
2377}
2378
2379static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2380{
2381 if (!*hpage)
2382 *hpage = khugepaged_alloc_hugepage(wait);
2383
2384 if (unlikely(!*hpage))
2385 return false;
2386
2387 return true;
2388}
2389
2390static struct page
2391*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2392 struct vm_area_struct *vma, unsigned long address,
2393 int node)
2394{
2395 up_read(&mm->mmap_sem);
2396 VM_BUG_ON(!*hpage);
2397 return *hpage;
2398}
692e0b35
AA		 2399#endif
2400
fa475e51
BL		 2401static bool hugepage_vma_check(struct vm_area_struct *vma)
2402{
2403 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2404 (vma->vm_flags & VM_NOHUGEPAGE))
2405 return false;
2406
2407 if (!vma->anon_vma || vma->vm_ops)
2408 return false;
2409 if (is_vma_temporary_stack(vma))
2410 return false;
2411 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2412 return true;
2413}
2414
26234f36
XG		 2415static void collapse_huge_page(struct mm_struct *mm,
2416 unsigned long address,
2417 struct page **hpage,
2418 struct vm_area_struct *vma,
2419 int node)
2420{
26234f36
XG		 2421 pmd_t *pmd, _pmd;
2422 pte_t *pte;
2423 pgtable_t pgtable;
2424 struct page *new_page;
c4088ebd 2425 spinlock_t *pmd_ptl, *pte_ptl;
26234f36
XG		 2426 int isolated;
2427 unsigned long hstart, hend;
2ec74c3e
SG		 2428 unsigned long mmun_start; /* For mmu_notifiers */
2429 unsigned long mmun_end; /* For mmu_notifiers */
26234f36
XG		 2430
2431 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2432
2433 /* release the mmap_sem read lock. */
2434 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2435 if (!new_page)
2436 return;
2437
420256ef 2438 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
ce83d217 2439 return;
ba76149f
AA		 2440
2441 /*
2442 * Prevent all access to pagetables with the exception of
2443 * gup_fast later hanlded by the ptep_clear_flush and the VM
2444 * handled by the anon_vma lock + PG_lock.
2445 */
2446 down_write(&mm->mmap_sem);
2447 if (unlikely(khugepaged_test_exit(mm)))
2448 goto out;
2449
2450 vma = find_vma(mm, address);
a8f531eb
L		 2451 if (!vma)
2452 goto out;
ba76149f
AA		 2453 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2454 hend = vma->vm_end & HPAGE_PMD_MASK;
2455 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2456 goto out;
fa475e51 2457 if (!hugepage_vma_check(vma))
a7d6e4ec 2458 goto out;
6219049a
BL		 2459 pmd = mm_find_pmd(mm, address);
2460 if (!pmd)
ba76149f 2461 goto out;
6219049a 2462 if (pmd_trans_huge(*pmd))
ba76149f
AA		 2463 goto out;
2464
4fc3f1d6 2465 anon_vma_lock_write(vma->anon_vma);
ba76149f
AA		 2466
2467 pte = pte_offset_map(pmd, address);
c4088ebd 2468 pte_ptl = pte_lockptr(mm, pmd);
ba76149f 2469
2ec74c3e
SG		 2470 mmun_start = address;
2471 mmun_end = address + HPAGE_PMD_SIZE;
2472 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
c4088ebd 2473 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
ba76149f
AA		 2474 /*
2475 * After this gup_fast can't run anymore. This also removes
2476 * any huge TLB entry from the CPU so we won't allow
2477 * huge and small TLB entries for the same virtual address
2478 * to avoid the risk of CPU bugs in that area.
2479 */
2ec74c3e 2480 _pmd = pmdp_clear_flush(vma, address, pmd);
c4088ebd 2481 spin_unlock(pmd_ptl);
2ec74c3e 2482 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
ba76149f 2483
c4088ebd 2484 spin_lock(pte_ptl);
ba76149f 2485 isolated = __collapse_huge_page_isolate(vma, address, pte);
c4088ebd 2486 spin_unlock(pte_ptl);
ba76149f
AA		 2487
2488 if (unlikely(!isolated)) {
453c7192 2489 pte_unmap(pte);
c4088ebd 2490 spin_lock(pmd_ptl);
ba76149f 2491 BUG_ON(!pmd_none(*pmd));
7c342512
AK		 2492 /*
2493 * We can only use set_pmd_at when establishing
2494 * hugepmds and never for establishing regular pmds that
2495 * points to regular pagetables. Use pmd_populate for that
2496 */
2497 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
c4088ebd 2498 spin_unlock(pmd_ptl);
08b52706 2499 anon_vma_unlock_write(vma->anon_vma);
ce83d217 2500 goto out;
ba76149f
AA		 2501 }
2502
2503 /*
2504 * All pages are isolated and locked so anon_vma rmap
2505 * can't run anymore.
2506 */
08b52706 2507 anon_vma_unlock_write(vma->anon_vma);
ba76149f 2508
c4088ebd 2509 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
453c7192 2510 pte_unmap(pte);
ba76149f
AA		 2511 __SetPageUptodate(new_page);
2512 pgtable = pmd_pgtable(_pmd);
ba76149f 2513
3122359a
KS		 2514 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2515 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
ba76149f
AA		 2516
2517 /*
2518 * spin_lock() below is not the equivalent of smp_wmb(), so
2519 * this is needed to avoid the copy_huge_page writes to become
2520 * visible after the set_pmd_at() write.
2521 */
2522 smp_wmb();
2523
c4088ebd 2524 spin_lock(pmd_ptl);
ba76149f
AA		 2525 BUG_ON(!pmd_none(*pmd));
2526 page_add_new_anon_rmap(new_page, vma, address);
fce144b4 2527 pgtable_trans_huge_deposit(mm, pmd, pgtable);
ba76149f 2528 set_pmd_at(mm, address, pmd, _pmd);
b113da65 2529 update_mmu_cache_pmd(vma, address, pmd);
c4088ebd 2530 spin_unlock(pmd_ptl);
ba76149f
AA		 2531
2532 *hpage = NULL;
420256ef 2533
ba76149f 2534 khugepaged_pages_collapsed++;
ce83d217 2535out_up_write:
ba76149f 2536 up_write(&mm->mmap_sem);
0bbbc0b3
AA		 2537 return;
2538
ce83d217 2539out:
678ff896 2540 mem_cgroup_uncharge_page(new_page);
ce83d217 2541 goto out_up_write;
ba76149f
AA		 2542}
2543
2544static int khugepaged_scan_pmd(struct mm_struct *mm,
2545 struct vm_area_struct *vma,
2546 unsigned long address,
2547 struct page **hpage)
2548{
ba76149f
AA		 2549 pmd_t *pmd;
2550 pte_t *pte, *_pte;
2551 int ret = 0, referenced = 0, none = 0;
2552 struct page *page;
2553 unsigned long _address;
2554 spinlock_t *ptl;
00ef2d2f 2555 int node = NUMA_NO_NODE;
ba76149f
AA		 2556
2557 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2558
6219049a
BL		 2559 pmd = mm_find_pmd(mm, address);
2560 if (!pmd)
ba76149f 2561 goto out;
6219049a 2562 if (pmd_trans_huge(*pmd))
ba76149f
AA		 2563 goto out;
2564
9f1b868a 2565 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
ba76149f
AA		 2566 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2567 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2568 _pte++, _address += PAGE_SIZE) {
2569 pte_t pteval = *_pte;
2570 if (pte_none(pteval)) {
2571 if (++none <= khugepaged_max_ptes_none)
2572 continue;
2573 else
2574 goto out_unmap;
2575 }
2576 if (!pte_present(pteval) || !pte_write(pteval))
2577 goto out_unmap;
2578 page = vm_normal_page(vma, _address, pteval);
2579 if (unlikely(!page))
2580 goto out_unmap;
5c4b4be3 2581 /*
9f1b868a
BL		 2582 * Record which node the original page is from and save this
2583 * information to khugepaged_node_load[].
2584 * Khupaged will allocate hugepage from the node has the max
2585 * hit record.
5c4b4be3 2586 */
9f1b868a
BL		 2587 node = page_to_nid(page);
2588 khugepaged_node_load[node]++;
309381fe 2589 VM_BUG_ON_PAGE(PageCompound(page), page);
ba76149f
AA		 2590 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2591 goto out_unmap;
2592 /* cannot use mapcount: can't collapse if there's a gup pin */
2593 if (page_count(page) != 1)
2594 goto out_unmap;
8ee53820
AA		 2595 if (pte_young(pteval) || PageReferenced(page) ||
2596 mmu_notifier_test_young(vma->vm_mm, address))
ba76149f
AA		 2597 referenced = 1;
2598 }
2599 if (referenced)
2600 ret = 1;
2601out_unmap:
2602 pte_unmap_unlock(pte, ptl);
9f1b868a
BL		 2603 if (ret) {
2604 node = khugepaged_find_target_node();
ce83d217 2605 /* collapse_huge_page will return with the mmap_sem released */
5c4b4be3 2606 collapse_huge_page(mm, address, hpage, vma, node);
9f1b868a 2607 }
ba76149f
AA		 2608out:
2609 return ret;
2610}
2611
2612static void collect_mm_slot(struct mm_slot *mm_slot)
2613{
2614 struct mm_struct *mm = mm_slot->mm;
2615
b9980cdc 2616 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
ba76149f
AA		 2617
2618 if (khugepaged_test_exit(mm)) {
2619 /* free mm_slot */
43b5fbbd 2620 hash_del(&mm_slot->hash);
ba76149f
AA		 2621 list_del(&mm_slot->mm_node);
2622
2623 /*
2624 * Not strictly needed because the mm exited already.
2625 *
2626 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2627 */
2628
2629 /* khugepaged_mm_lock actually not necessary for the below */
2630 free_mm_slot(mm_slot);
2631 mmdrop(mm);
2632 }
2633}
2634
2635static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2636 struct page **hpage)
2f1da642
HS		 2637 __releases(&khugepaged_mm_lock)
2638 __acquires(&khugepaged_mm_lock)
ba76149f
AA		 2639{
2640 struct mm_slot *mm_slot;
2641 struct mm_struct *mm;
2642 struct vm_area_struct *vma;
2643 int progress = 0;
2644
2645 VM_BUG_ON(!pages);
b9980cdc 2646 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
ba76149f
AA		 2647
2648 if (khugepaged_scan.mm_slot)
2649 mm_slot = khugepaged_scan.mm_slot;
2650 else {
2651 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2652 struct mm_slot, mm_node);
2653 khugepaged_scan.address = 0;
2654 khugepaged_scan.mm_slot = mm_slot;
2655 }
2656 spin_unlock(&khugepaged_mm_lock);
2657
2658 mm = mm_slot->mm;
2659 down_read(&mm->mmap_sem);
2660 if (unlikely(khugepaged_test_exit(mm)))
2661 vma = NULL;
2662 else
2663 vma = find_vma(mm, khugepaged_scan.address);
2664
2665 progress++;
2666 for (; vma; vma = vma->vm_next) {
2667 unsigned long hstart, hend;
2668
2669 cond_resched();
2670 if (unlikely(khugepaged_test_exit(mm))) {
2671 progress++;
2672 break;
2673 }
fa475e51
BL		 2674 if (!hugepage_vma_check(vma)) {
2675skip:
ba76149f
AA		 2676 progress++;
2677 continue;
2678 }
ba76149f
AA		 2679 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2680 hend = vma->vm_end & HPAGE_PMD_MASK;
a7d6e4ec
AA		 2681 if (hstart >= hend)
2682 goto skip;
2683 if (khugepaged_scan.address > hend)
2684 goto skip;
ba76149f
AA		 2685 if (khugepaged_scan.address < hstart)
2686 khugepaged_scan.address = hstart;
a7d6e4ec 2687 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
ba76149f
AA		 2688
2689 while (khugepaged_scan.address < hend) {
2690 int ret;
2691 cond_resched();
2692 if (unlikely(khugepaged_test_exit(mm)))
2693 goto breakouterloop;
2694
2695 VM_BUG_ON(khugepaged_scan.address < hstart ||
2696 khugepaged_scan.address + HPAGE_PMD_SIZE >
2697 hend);
2698 ret = khugepaged_scan_pmd(mm, vma,
2699 khugepaged_scan.address,
2700 hpage);
2701 /* move to next address */
2702 khugepaged_scan.address += HPAGE_PMD_SIZE;
2703 progress += HPAGE_PMD_NR;
2704 if (ret)
2705 /* we released mmap_sem so break loop */
2706 goto breakouterloop_mmap_sem;
2707 if (progress >= pages)
2708 goto breakouterloop;
2709 }
2710 }
2711breakouterloop:
2712 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2713breakouterloop_mmap_sem:
2714
2715 spin_lock(&khugepaged_mm_lock);
a7d6e4ec 2716 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
ba76149f
AA		 2717 /*
2718 * Release the current mm_slot if this mm is about to die, or
2719 * if we scanned all vmas of this mm.
2720 */
2721 if (khugepaged_test_exit(mm) || !vma) {
2722 /*
2723 * Make sure that if mm_users is reaching zero while
2724 * khugepaged runs here, khugepaged_exit will find
2725 * mm_slot not pointing to the exiting mm.
2726 */
2727 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2728 khugepaged_scan.mm_slot = list_entry(
2729 mm_slot->mm_node.next,
2730 struct mm_slot, mm_node);
2731 khugepaged_scan.address = 0;
2732 } else {
2733 khugepaged_scan.mm_slot = NULL;
2734 khugepaged_full_scans++;
2735 }
2736
2737 collect_mm_slot(mm_slot);
2738 }
2739
2740 return progress;
2741}
2742
2743static int khugepaged_has_work(void)
2744{
2745 return !list_empty(&khugepaged_scan.mm_head) &&
2746 khugepaged_enabled();
2747}
2748
2749static int khugepaged_wait_event(void)
2750{
2751 return !list_empty(&khugepaged_scan.mm_head) ||
2017c0bf 2752 kthread_should_stop();
ba76149f
AA		 2753}
2754
d516904b 2755static void khugepaged_do_scan(void)
ba76149f 2756{
d516904b 2757 struct page *hpage = NULL;
ba76149f
AA		 2758 unsigned int progress = 0, pass_through_head = 0;
2759 unsigned int pages = khugepaged_pages_to_scan;
d516904b 2760 bool wait = true;
ba76149f
AA		 2761
2762 barrier(); /* write khugepaged_pages_to_scan to local stack */
2763
2764 while (progress < pages) {
26234f36 2765 if (!khugepaged_prealloc_page(&hpage, &wait))
d516904b 2766 break;
26234f36 2767
420256ef 2768 cond_resched();
ba76149f 2769
878aee7d
AA		 2770 if (unlikely(kthread_should_stop() || freezing(current)))
2771 break;
2772
ba76149f
AA		 2773 spin_lock(&khugepaged_mm_lock);
2774 if (!khugepaged_scan.mm_slot)
2775 pass_through_head++;
2776 if (khugepaged_has_work() &&
2777 pass_through_head < 2)
2778 progress += khugepaged_scan_mm_slot(pages - progress,
d516904b 2779 &hpage);
ba76149f
AA		 2780 else
2781 progress = pages;
2782 spin_unlock(&khugepaged_mm_lock);
2783 }
ba76149f 2784
d516904b
XG		 2785 if (!IS_ERR_OR_NULL(hpage))
2786 put_page(hpage);
0bbbc0b3
AA		 2787}
2788
2017c0bf
XG		 2789static void khugepaged_wait_work(void)
2790{
2791 try_to_freeze();
2792
2793 if (khugepaged_has_work()) {
2794 if (!khugepaged_scan_sleep_millisecs)
2795 return;
2796
2797 wait_event_freezable_timeout(khugepaged_wait,
2798 kthread_should_stop(),
2799 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2800 return;
2801 }
2802
2803 if (khugepaged_enabled())
2804 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2805}
2806
ba76149f
AA		 2807static int khugepaged(void *none)
2808{
2809 struct mm_slot *mm_slot;
2810
878aee7d 2811 set_freezable();
ba76149f
AA		 2812 set_user_nice(current, 19);
2813
b7231789
XG		 2814 while (!kthread_should_stop()) {
2815 khugepaged_do_scan();
2816 khugepaged_wait_work();
2817 }
ba76149f
AA		 2818
2819 spin_lock(&khugepaged_mm_lock);
2820 mm_slot = khugepaged_scan.mm_slot;
2821 khugepaged_scan.mm_slot = NULL;
2822 if (mm_slot)
2823 collect_mm_slot(mm_slot);
2824 spin_unlock(&khugepaged_mm_lock);
ba76149f
AA		 2825 return 0;
2826}
2827
c5a647d0
KS		 2828static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2829 unsigned long haddr, pmd_t *pmd)
2830{
2831 struct mm_struct *mm = vma->vm_mm;
2832 pgtable_t pgtable;
2833 pmd_t _pmd;
2834 int i;
2835
2836 pmdp_clear_flush(vma, haddr, pmd);
2837 /* leave pmd empty until pte is filled */
2838
6b0b50b0 2839 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
c5a647d0
KS		 2840 pmd_populate(mm, &_pmd, pgtable);
2841
2842 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2843 pte_t *pte, entry;
2844 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2845 entry = pte_mkspecial(entry);
2846 pte = pte_offset_map(&_pmd, haddr);
2847 VM_BUG_ON(!pte_none(*pte));
2848 set_pte_at(mm, haddr, pte, entry);
2849 pte_unmap(pte);
2850 }
2851 smp_wmb(); /* make pte visible before pmd */
2852 pmd_populate(mm, pmd, pgtable);
97ae1749 2853 put_huge_zero_page();
c5a647d0
KS		 2854}
2855
e180377f
KS		 2856void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2857 pmd_t *pmd)
71e3aac0 2858{
c4088ebd 2859 spinlock_t *ptl;
71e3aac0 2860 struct page *page;
e180377f 2861 struct mm_struct *mm = vma->vm_mm;
c5a647d0
KS		 2862 unsigned long haddr = address & HPAGE_PMD_MASK;
2863 unsigned long mmun_start; /* For mmu_notifiers */
2864 unsigned long mmun_end; /* For mmu_notifiers */
e180377f
KS		 2865
2866 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
71e3aac0 2867
c5a647d0
KS		 2868 mmun_start = haddr;
2869 mmun_end = haddr + HPAGE_PMD_SIZE;
750e8165 2870again:
c5a647d0 2871 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
c4088ebd 2872 ptl = pmd_lock(mm, pmd);
71e3aac0 2873 if (unlikely(!pmd_trans_huge(*pmd))) {
c4088ebd 2874 spin_unlock(ptl);
c5a647d0
KS		 2875 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2876 return;
2877 }
2878 if (is_huge_zero_pmd(*pmd)) {
2879 __split_huge_zero_page_pmd(vma, haddr, pmd);
c4088ebd 2880 spin_unlock(ptl);
c5a647d0 2881 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 2882 return;
2883 }
2884 page = pmd_page(*pmd);
309381fe 2885 VM_BUG_ON_PAGE(!page_count(page), page);
71e3aac0 2886 get_page(page);
c4088ebd 2887 spin_unlock(ptl);
c5a647d0 2888 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 2889
2890 split_huge_page(page);
2891
2892 put_page(page);
750e8165
HD		 2893
2894 /*
2895 * We don't always have down_write of mmap_sem here: a racing
2896 * do_huge_pmd_wp_page() might have copied-on-write to another
2897 * huge page before our split_huge_page() got the anon_vma lock.
2898 */
2899 if (unlikely(pmd_trans_huge(*pmd)))
2900 goto again;
71e3aac0 2901}
94fcc585 2902
e180377f
KS		 2903void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2904 pmd_t *pmd)
2905{
2906 struct vm_area_struct *vma;
2907
2908 vma = find_vma(mm, address);
2909 BUG_ON(vma == NULL);
2910 split_huge_page_pmd(vma, address, pmd);
2911}
2912
94fcc585
AA		 2913static void split_huge_page_address(struct mm_struct *mm,
2914 unsigned long address)
2915{
94fcc585
AA		 2916 pmd_t *pmd;
2917
2918 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2919
6219049a
BL		 2920 pmd = mm_find_pmd(mm, address);
2921 if (!pmd)
94fcc585
AA		 2922 return;
2923 /*
2924 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2925 * materialize from under us.
2926 */
e180377f 2927 split_huge_page_pmd_mm(mm, address, pmd);
94fcc585
AA		 2928}
2929
2930void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2931 unsigned long start,
2932 unsigned long end,
2933 long adjust_next)
2934{
2935 /*
2936 * If the new start address isn't hpage aligned and it could
2937 * previously contain an hugepage: check if we need to split
2938 * an huge pmd.
2939 */
2940 if (start & ~HPAGE_PMD_MASK &&
2941 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2942 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2943 split_huge_page_address(vma->vm_mm, start);
2944
2945 /*
2946 * If the new end address isn't hpage aligned and it could
2947 * previously contain an hugepage: check if we need to split
2948 * an huge pmd.
2949 */
2950 if (end & ~HPAGE_PMD_MASK &&
2951 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2952 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2953 split_huge_page_address(vma->vm_mm, end);
2954
2955 /*
2956 * If we're also updating the vma->vm_next->vm_start, if the new
2957 * vm_next->vm_start isn't page aligned and it could previously
2958 * contain an hugepage: check if we need to split an huge pmd.
2959 */
2960 if (adjust_next > 0) {
2961 struct vm_area_struct *next = vma->vm_next;
2962 unsigned long nstart = next->vm_start;
2963 nstart += adjust_next << PAGE_SHIFT;
2964 if (nstart & ~HPAGE_PMD_MASK &&
2965 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2966 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2967 split_huge_page_address(next->vm_mm, nstart);
2968 }
2969}
Linux kernel - Deliverables
This page took 0.395378 seconds and 5 git commands to generate.