projects / deliverable / linux.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
mm: fix TLB flush race between migration, and change_protection_range
[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
133 if (recommended_min > min_free_kbytes)
134 min_free_kbytes = recommended_min;
135 setup_per_zone_wmarks();
136 return 0;
137}
138late_initcall(set_recommended_min_free_kbytes);
139
ba76149f
AA		 140static int start_khugepaged(void)
141{
142 int err = 0;
143 if (khugepaged_enabled()) {
ba76149f
AA		 144 if (!khugepaged_thread)
145 khugepaged_thread = kthread_run(khugepaged, NULL,
146 "khugepaged");
147 if (unlikely(IS_ERR(khugepaged_thread))) {
148 printk(KERN_ERR
149 "khugepaged: kthread_run(khugepaged) failed\n");
150 err = PTR_ERR(khugepaged_thread);
151 khugepaged_thread = NULL;
152 }
911891af
XG		 153
154 if (!list_empty(&khugepaged_scan.mm_head))
ba76149f 155 wake_up_interruptible(&khugepaged_wait);
f000565a
AA		 156
157 set_recommended_min_free_kbytes();
911891af 158 } else if (khugepaged_thread) {
911891af
XG		 159 kthread_stop(khugepaged_thread);
160 khugepaged_thread = NULL;
161 }
637e3a27 162
ba76149f
AA		 163 return err;
164}
71e3aac0 165
97ae1749 166static atomic_t huge_zero_refcount;
5918d10a 167static struct page *huge_zero_page __read_mostly;
97ae1749 168
5918d10a 169static inline bool is_huge_zero_page(struct page *page)
4a6c1297 170{
5918d10a 171 return ACCESS_ONCE(huge_zero_page) == page;
97ae1749 172}
4a6c1297 173
97ae1749
KS		 174static inline bool is_huge_zero_pmd(pmd_t pmd)
175{
5918d10a 176 return is_huge_zero_page(pmd_page(pmd));
97ae1749
KS		 177}
178
5918d10a 179static struct page *get_huge_zero_page(void)
97ae1749
KS		 180{
181 struct page *zero_page;
182retry:
183 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
5918d10a 184 return ACCESS_ONCE(huge_zero_page);
97ae1749
KS		 185
186 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
4a6c1297 187 HPAGE_PMD_ORDER);
d8a8e1f0
KS		 188 if (!zero_page) {
189 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
5918d10a 190 return NULL;
d8a8e1f0
KS		 191 }
192 count_vm_event(THP_ZERO_PAGE_ALLOC);
97ae1749 193 preempt_disable();
5918d10a 194 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
97ae1749
KS		 195 preempt_enable();
196 __free_page(zero_page);
197 goto retry;
198 }
199
200 /* We take additional reference here. It will be put back by shrinker */
201 atomic_set(&huge_zero_refcount, 2);
202 preempt_enable();
5918d10a 203 return ACCESS_ONCE(huge_zero_page);
4a6c1297
KS		 204}
205
97ae1749 206static void put_huge_zero_page(void)
4a6c1297 207{
97ae1749
KS		 208 /*
209 * Counter should never go to zero here. Only shrinker can put
210 * last reference.
211 */
212 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
4a6c1297
KS		 213}
214
48896466
GC		 215static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
216 struct shrink_control *sc)
4a6c1297 217{
48896466
GC		 218 /* we can free zero page only if last reference remains */
219 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
220}
97ae1749 221
48896466
GC		 222static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
223 struct shrink_control *sc)
224{
97ae1749 225 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
5918d10a
KS		 226 struct page *zero_page = xchg(&huge_zero_page, NULL);
227 BUG_ON(zero_page == NULL);
228 __free_page(zero_page);
48896466 229 return HPAGE_PMD_NR;
97ae1749
KS		 230 }
231
232 return 0;
4a6c1297
KS		 233}
234
97ae1749 235static struct shrinker huge_zero_page_shrinker = {
48896466
GC		 236 .count_objects = shrink_huge_zero_page_count,
237 .scan_objects = shrink_huge_zero_page_scan,
97ae1749
KS		 238 .seeks = DEFAULT_SEEKS,
239};
240
71e3aac0 241#ifdef CONFIG_SYSFS
ba76149f 242
71e3aac0
AA		 243static ssize_t double_flag_show(struct kobject *kobj,
244 struct kobj_attribute *attr, char *buf,
245 enum transparent_hugepage_flag enabled,
246 enum transparent_hugepage_flag req_madv)
247{
248 if (test_bit(enabled, &transparent_hugepage_flags)) {
249 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
250 return sprintf(buf, "[always] madvise never\n");
251 } else if (test_bit(req_madv, &transparent_hugepage_flags))
252 return sprintf(buf, "always [madvise] never\n");
253 else
254 return sprintf(buf, "always madvise [never]\n");
255}
256static ssize_t double_flag_store(struct kobject *kobj,
257 struct kobj_attribute *attr,
258 const char *buf, size_t count,
259 enum transparent_hugepage_flag enabled,
260 enum transparent_hugepage_flag req_madv)
261{
262 if (!memcmp("always", buf,
263 min(sizeof("always")-1, count))) {
264 set_bit(enabled, &transparent_hugepage_flags);
265 clear_bit(req_madv, &transparent_hugepage_flags);
266 } else if (!memcmp("madvise", buf,
267 min(sizeof("madvise")-1, count))) {
268 clear_bit(enabled, &transparent_hugepage_flags);
269 set_bit(req_madv, &transparent_hugepage_flags);
270 } else if (!memcmp("never", buf,
271 min(sizeof("never")-1, count))) {
272 clear_bit(enabled, &transparent_hugepage_flags);
273 clear_bit(req_madv, &transparent_hugepage_flags);
274 } else
275 return -EINVAL;
276
277 return count;
278}
279
280static ssize_t enabled_show(struct kobject *kobj,
281 struct kobj_attribute *attr, char *buf)
282{
283 return double_flag_show(kobj, attr, buf,
284 TRANSPARENT_HUGEPAGE_FLAG,
285 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
286}
287static ssize_t enabled_store(struct kobject *kobj,
288 struct kobj_attribute *attr,
289 const char *buf, size_t count)
290{
ba76149f
AA		 291 ssize_t ret;
292
293 ret = double_flag_store(kobj, attr, buf, count,
294 TRANSPARENT_HUGEPAGE_FLAG,
295 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
296
297 if (ret > 0) {
911891af
XG		 298 int err;
299
300 mutex_lock(&khugepaged_mutex);
301 err = start_khugepaged();
302 mutex_unlock(&khugepaged_mutex);
303
ba76149f
AA		 304 if (err)
305 ret = err;
306 }
307
308 return ret;
71e3aac0
AA		 309}
310static struct kobj_attribute enabled_attr =
311 __ATTR(enabled, 0644, enabled_show, enabled_store);
312
313static ssize_t single_flag_show(struct kobject *kobj,
314 struct kobj_attribute *attr, char *buf,
315 enum transparent_hugepage_flag flag)
316{
e27e6151
BH		 317 return sprintf(buf, "%d\n",
318 !!test_bit(flag, &transparent_hugepage_flags));
71e3aac0 319}
e27e6151 320
71e3aac0
AA		 321static ssize_t single_flag_store(struct kobject *kobj,
322 struct kobj_attribute *attr,
323 const char *buf, size_t count,
324 enum transparent_hugepage_flag flag)
325{
e27e6151
BH		 326 unsigned long value;
327 int ret;
328
329 ret = kstrtoul(buf, 10, &value);
330 if (ret < 0)
331 return ret;
332 if (value > 1)
333 return -EINVAL;
334
335 if (value)
71e3aac0 336 set_bit(flag, &transparent_hugepage_flags);
e27e6151 337 else
71e3aac0 338 clear_bit(flag, &transparent_hugepage_flags);
71e3aac0
AA		 339
340 return count;
341}
342
343/*
344 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
345 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
346 * memory just to allocate one more hugepage.
347 */
348static ssize_t defrag_show(struct kobject *kobj,
349 struct kobj_attribute *attr, char *buf)
350{
351 return double_flag_show(kobj, attr, buf,
352 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
353 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
354}
355static ssize_t defrag_store(struct kobject *kobj,
356 struct kobj_attribute *attr,
357 const char *buf, size_t count)
358{
359 return double_flag_store(kobj, attr, buf, count,
360 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
361 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
362}
363static struct kobj_attribute defrag_attr =
364 __ATTR(defrag, 0644, defrag_show, defrag_store);
365
79da5407
KS		 366static ssize_t use_zero_page_show(struct kobject *kobj,
367 struct kobj_attribute *attr, char *buf)
368{
369 return single_flag_show(kobj, attr, buf,
370 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
371}
372static ssize_t use_zero_page_store(struct kobject *kobj,
373 struct kobj_attribute *attr, const char *buf, size_t count)
374{
375 return single_flag_store(kobj, attr, buf, count,
376 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
377}
378static struct kobj_attribute use_zero_page_attr =
379 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
71e3aac0
AA		 380#ifdef CONFIG_DEBUG_VM
381static ssize_t debug_cow_show(struct kobject *kobj,
382 struct kobj_attribute *attr, char *buf)
383{
384 return single_flag_show(kobj, attr, buf,
385 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
386}
387static ssize_t debug_cow_store(struct kobject *kobj,
388 struct kobj_attribute *attr,
389 const char *buf, size_t count)
390{
391 return single_flag_store(kobj, attr, buf, count,
392 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
393}
394static struct kobj_attribute debug_cow_attr =
395 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
396#endif /* CONFIG_DEBUG_VM */
397
398static struct attribute *hugepage_attr[] = {
399 &enabled_attr.attr,
400 &defrag_attr.attr,
79da5407 401 &use_zero_page_attr.attr,
71e3aac0
AA		 402#ifdef CONFIG_DEBUG_VM
403 &debug_cow_attr.attr,
404#endif
405 NULL,
406};
407
408static struct attribute_group hugepage_attr_group = {
409 .attrs = hugepage_attr,
ba76149f
AA		 410};
411
412static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
413 struct kobj_attribute *attr,
414 char *buf)
415{
416 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
417}
418
419static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
420 struct kobj_attribute *attr,
421 const char *buf, size_t count)
422{
423 unsigned long msecs;
424 int err;
425
3dbb95f7 426 err = kstrtoul(buf, 10, &msecs);
ba76149f
AA		 427 if (err || msecs > UINT_MAX)
428 return -EINVAL;
429
430 khugepaged_scan_sleep_millisecs = msecs;
431 wake_up_interruptible(&khugepaged_wait);
432
433 return count;
434}
435static struct kobj_attribute scan_sleep_millisecs_attr =
436 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
437 scan_sleep_millisecs_store);
438
439static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
440 struct kobj_attribute *attr,
441 char *buf)
442{
443 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
444}
445
446static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
447 struct kobj_attribute *attr,
448 const char *buf, size_t count)
449{
450 unsigned long msecs;
451 int err;
452
3dbb95f7 453 err = kstrtoul(buf, 10, &msecs);
ba76149f
AA		 454 if (err || msecs > UINT_MAX)
455 return -EINVAL;
456
457 khugepaged_alloc_sleep_millisecs = msecs;
458 wake_up_interruptible(&khugepaged_wait);
459
460 return count;
461}
462static struct kobj_attribute alloc_sleep_millisecs_attr =
463 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
464 alloc_sleep_millisecs_store);
465
466static ssize_t pages_to_scan_show(struct kobject *kobj,
467 struct kobj_attribute *attr,
468 char *buf)
469{
470 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
471}
472static ssize_t pages_to_scan_store(struct kobject *kobj,
473 struct kobj_attribute *attr,
474 const char *buf, size_t count)
475{
476 int err;
477 unsigned long pages;
478
3dbb95f7 479 err = kstrtoul(buf, 10, &pages);
ba76149f
AA		 480 if (err || !pages || pages > UINT_MAX)
481 return -EINVAL;
482
483 khugepaged_pages_to_scan = pages;
484
485 return count;
486}
487static struct kobj_attribute pages_to_scan_attr =
488 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
489 pages_to_scan_store);
490
491static ssize_t pages_collapsed_show(struct kobject *kobj,
492 struct kobj_attribute *attr,
493 char *buf)
494{
495 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
496}
497static struct kobj_attribute pages_collapsed_attr =
498 __ATTR_RO(pages_collapsed);
499
500static ssize_t full_scans_show(struct kobject *kobj,
501 struct kobj_attribute *attr,
502 char *buf)
503{
504 return sprintf(buf, "%u\n", khugepaged_full_scans);
505}
506static struct kobj_attribute full_scans_attr =
507 __ATTR_RO(full_scans);
508
509static ssize_t khugepaged_defrag_show(struct kobject *kobj,
510 struct kobj_attribute *attr, char *buf)
511{
512 return single_flag_show(kobj, attr, buf,
513 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
514}
515static ssize_t khugepaged_defrag_store(struct kobject *kobj,
516 struct kobj_attribute *attr,
517 const char *buf, size_t count)
518{
519 return single_flag_store(kobj, attr, buf, count,
520 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
521}
522static struct kobj_attribute khugepaged_defrag_attr =
523 __ATTR(defrag, 0644, khugepaged_defrag_show,
524 khugepaged_defrag_store);
525
526/*
527 * max_ptes_none controls if khugepaged should collapse hugepages over
528 * any unmapped ptes in turn potentially increasing the memory
529 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
530 * reduce the available free memory in the system as it
531 * runs. Increasing max_ptes_none will instead potentially reduce the
532 * free memory in the system during the khugepaged scan.
533 */
534static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
535 struct kobj_attribute *attr,
536 char *buf)
537{
538 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
539}
540static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
541 struct kobj_attribute *attr,
542 const char *buf, size_t count)
543{
544 int err;
545 unsigned long max_ptes_none;
546
3dbb95f7 547 err = kstrtoul(buf, 10, &max_ptes_none);
ba76149f
AA		 548 if (err || max_ptes_none > HPAGE_PMD_NR-1)
549 return -EINVAL;
550
551 khugepaged_max_ptes_none = max_ptes_none;
552
553 return count;
554}
555static struct kobj_attribute khugepaged_max_ptes_none_attr =
556 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
557 khugepaged_max_ptes_none_store);
558
559static struct attribute *khugepaged_attr[] = {
560 &khugepaged_defrag_attr.attr,
561 &khugepaged_max_ptes_none_attr.attr,
562 &pages_to_scan_attr.attr,
563 &pages_collapsed_attr.attr,
564 &full_scans_attr.attr,
565 &scan_sleep_millisecs_attr.attr,
566 &alloc_sleep_millisecs_attr.attr,
567 NULL,
568};
569
570static struct attribute_group khugepaged_attr_group = {
571 .attrs = khugepaged_attr,
572 .name = "khugepaged",
71e3aac0 573};
71e3aac0 574
569e5590 575static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
71e3aac0 576{
71e3aac0
AA		 577 int err;
578
569e5590
SL		 579 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
580 if (unlikely(!*hugepage_kobj)) {
2c79737a 581 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
569e5590 582 return -ENOMEM;
ba76149f
AA		 583 }
584
569e5590 585 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
ba76149f 586 if (err) {
2c79737a 587 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
569e5590 588 goto delete_obj;
ba76149f
AA		 589 }
590
569e5590 591 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
ba76149f 592 if (err) {
2c79737a 593 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
569e5590 594 goto remove_hp_group;
ba76149f 595 }
569e5590
SL		 596
597 return 0;
598
599remove_hp_group:
600 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
601delete_obj:
602 kobject_put(*hugepage_kobj);
603 return err;
604}
605
606static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
607{
608 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
609 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
610 kobject_put(hugepage_kobj);
611}
612#else
613static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
614{
615 return 0;
616}
617
618static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
619{
620}
621#endif /* CONFIG_SYSFS */
622
623static int __init hugepage_init(void)
624{
625 int err;
626 struct kobject *hugepage_kobj;
627
628 if (!has_transparent_hugepage()) {
629 transparent_hugepage_flags = 0;
630 return -EINVAL;
631 }
632
633 err = hugepage_init_sysfs(&hugepage_kobj);
634 if (err)
635 return err;
ba76149f
AA		 636
637 err = khugepaged_slab_init();
638 if (err)
639 goto out;
640
97ae1749
KS		 641 register_shrinker(&huge_zero_page_shrinker);
642
97562cd2
RR		 643 /*
644 * By default disable transparent hugepages on smaller systems,
645 * where the extra memory used could hurt more than TLB overhead
646 * is likely to save. The admin can still enable it through /sys.
647 */
648 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
649 transparent_hugepage_flags = 0;
650
ba76149f
AA		 651 start_khugepaged();
652
569e5590 653 return 0;
ba76149f 654out:
569e5590 655 hugepage_exit_sysfs(hugepage_kobj);
ba76149f 656 return err;
71e3aac0
AA		 657}
658module_init(hugepage_init)
659
660static int __init setup_transparent_hugepage(char *str)
661{
662 int ret = 0;
663 if (!str)
664 goto out;
665 if (!strcmp(str, "always")) {
666 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
667 &transparent_hugepage_flags);
668 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
669 &transparent_hugepage_flags);
670 ret = 1;
671 } else if (!strcmp(str, "madvise")) {
672 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
673 &transparent_hugepage_flags);
674 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
675 &transparent_hugepage_flags);
676 ret = 1;
677 } else if (!strcmp(str, "never")) {
678 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
679 &transparent_hugepage_flags);
680 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
681 &transparent_hugepage_flags);
682 ret = 1;
683 }
684out:
685 if (!ret)
686 printk(KERN_WARNING
687 "transparent_hugepage= cannot parse, ignored\n");
688 return ret;
689}
690__setup("transparent_hugepage=", setup_transparent_hugepage);
691
b32967ff 692pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
71e3aac0
AA		 693{
694 if (likely(vma->vm_flags & VM_WRITE))
695 pmd = pmd_mkwrite(pmd);
696 return pmd;
697}
698
3122359a 699static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
b3092b3b
BL		 700{
701 pmd_t entry;
3122359a 702 entry = mk_pmd(page, prot);
b3092b3b
BL		 703 entry = pmd_mkhuge(entry);
704 return entry;
705}
706
71e3aac0
AA		 707static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
708 struct vm_area_struct *vma,
709 unsigned long haddr, pmd_t *pmd,
710 struct page *page)
711{
71e3aac0 712 pgtable_t pgtable;
c4088ebd 713 spinlock_t *ptl;
71e3aac0
AA		 714
715 VM_BUG_ON(!PageCompound(page));
716 pgtable = pte_alloc_one(mm, haddr);
edad9d2c 717 if (unlikely(!pgtable))
71e3aac0 718 return VM_FAULT_OOM;
71e3aac0
AA		 719
720 clear_huge_page(page, haddr, HPAGE_PMD_NR);
52f37629
MK		 721 /*
722 * The memory barrier inside __SetPageUptodate makes sure that
723 * clear_huge_page writes become visible before the set_pmd_at()
724 * write.
725 */
71e3aac0
AA		 726 __SetPageUptodate(page);
727
c4088ebd 728 ptl = pmd_lock(mm, pmd);
71e3aac0 729 if (unlikely(!pmd_none(*pmd))) {
c4088ebd 730 spin_unlock(ptl);
b9bbfbe3 731 mem_cgroup_uncharge_page(page);
71e3aac0
AA		 732 put_page(page);
733 pte_free(mm, pgtable);
734 } else {
735 pmd_t entry;
3122359a
KS		 736 entry = mk_huge_pmd(page, vma->vm_page_prot);
737 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
71e3aac0 738 page_add_new_anon_rmap(page, vma, haddr);
6b0b50b0 739 pgtable_trans_huge_deposit(mm, pmd, pgtable);
71e3aac0 740 set_pmd_at(mm, haddr, pmd, entry);
71e3aac0 741 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
e1f56c89 742 atomic_long_inc(&mm->nr_ptes);
c4088ebd 743 spin_unlock(ptl);
71e3aac0
AA		 744 }
745
aa2e878e 746 return 0;
71e3aac0
AA		 747}
748
cc5d462f 749static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
0bbbc0b3 750{
cc5d462f 751 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
0bbbc0b3
AA		 752}
753
754static inline struct page *alloc_hugepage_vma(int defrag,
755 struct vm_area_struct *vma,
cc5d462f
AK		 756 unsigned long haddr, int nd,
757 gfp_t extra_gfp)
0bbbc0b3 758{
cc5d462f 759 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
5c4b4be3 760 HPAGE_PMD_ORDER, vma, haddr, nd);
0bbbc0b3
AA		 761}
762
c4088ebd 763/* Caller must hold page table lock. */
3ea41e62 764static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
97ae1749 765 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
5918d10a 766 struct page *zero_page)
fc9fe822
KS		 767{
768 pmd_t entry;
3ea41e62
KS		 769 if (!pmd_none(*pmd))
770 return false;
5918d10a 771 entry = mk_pmd(zero_page, vma->vm_page_prot);
fc9fe822
KS		 772 entry = pmd_wrprotect(entry);
773 entry = pmd_mkhuge(entry);
6b0b50b0 774 pgtable_trans_huge_deposit(mm, pmd, pgtable);
fc9fe822 775 set_pmd_at(mm, haddr, pmd, entry);
e1f56c89 776 atomic_long_inc(&mm->nr_ptes);
3ea41e62 777 return true;
fc9fe822
KS		 778}
779
71e3aac0
AA		 780int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
781 unsigned long address, pmd_t *pmd,
782 unsigned int flags)
783{
784 struct page *page;
785 unsigned long haddr = address & HPAGE_PMD_MASK;
71e3aac0 786
128ec037 787 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
c0292554 788 return VM_FAULT_FALLBACK;
128ec037
KS		 789 if (unlikely(anon_vma_prepare(vma)))
790 return VM_FAULT_OOM;
791 if (unlikely(khugepaged_enter(vma)))
792 return VM_FAULT_OOM;
793 if (!(flags & FAULT_FLAG_WRITE) &&
794 transparent_hugepage_use_zero_page()) {
c4088ebd 795 spinlock_t *ptl;
128ec037
KS		 796 pgtable_t pgtable;
797 struct page *zero_page;
798 bool set;
799 pgtable = pte_alloc_one(mm, haddr);
800 if (unlikely(!pgtable))
ba76149f 801 return VM_FAULT_OOM;
128ec037
KS		 802 zero_page = get_huge_zero_page();
803 if (unlikely(!zero_page)) {
804 pte_free(mm, pgtable);
81ab4201 805 count_vm_event(THP_FAULT_FALLBACK);
c0292554 806 return VM_FAULT_FALLBACK;
b9bbfbe3 807 }
c4088ebd 808 ptl = pmd_lock(mm, pmd);
128ec037
KS		 809 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
810 zero_page);
c4088ebd 811 spin_unlock(ptl);
128ec037
KS		 812 if (!set) {
813 pte_free(mm, pgtable);
814 put_huge_zero_page();
edad9d2c 815 }
edad9d2c 816 return 0;
71e3aac0 817 }
128ec037
KS		 818 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
819 vma, haddr, numa_node_id(), 0);
820 if (unlikely(!page)) {
821 count_vm_event(THP_FAULT_FALLBACK);
c0292554 822 return VM_FAULT_FALLBACK;
128ec037 823 }
128ec037
KS		 824 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
825 put_page(page);
17766dde 826 count_vm_event(THP_FAULT_FALLBACK);
c0292554 827 return VM_FAULT_FALLBACK;
128ec037
KS		 828 }
829 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
830 mem_cgroup_uncharge_page(page);
831 put_page(page);
17766dde 832 count_vm_event(THP_FAULT_FALLBACK);
c0292554 833 return VM_FAULT_FALLBACK;
128ec037
KS		 834 }
835
17766dde 836 count_vm_event(THP_FAULT_ALLOC);
128ec037 837 return 0;
71e3aac0
AA		 838}
839
840int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
841 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
842 struct vm_area_struct *vma)
843{
c4088ebd 844 spinlock_t *dst_ptl, *src_ptl;
71e3aac0
AA		 845 struct page *src_page;
846 pmd_t pmd;
847 pgtable_t pgtable;
848 int ret;
849
850 ret = -ENOMEM;
851 pgtable = pte_alloc_one(dst_mm, addr);
852 if (unlikely(!pgtable))
853 goto out;
854
c4088ebd
KS		 855 dst_ptl = pmd_lock(dst_mm, dst_pmd);
856 src_ptl = pmd_lockptr(src_mm, src_pmd);
857 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
71e3aac0
AA		 858
859 ret = -EAGAIN;
860 pmd = *src_pmd;
861 if (unlikely(!pmd_trans_huge(pmd))) {
862 pte_free(dst_mm, pgtable);
863 goto out_unlock;
864 }
fc9fe822 865 /*
c4088ebd 866 * When page table lock is held, the huge zero pmd should not be
fc9fe822
KS		 867 * under splitting since we don't split the page itself, only pmd to
868 * a page table.
869 */
870 if (is_huge_zero_pmd(pmd)) {
5918d10a 871 struct page *zero_page;
3ea41e62 872 bool set;
97ae1749
KS		 873 /*
874 * get_huge_zero_page() will never allocate a new page here,
875 * since we already have a zero page to copy. It just takes a
876 * reference.
877 */
5918d10a 878 zero_page = get_huge_zero_page();
3ea41e62 879 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
5918d10a 880 zero_page);
3ea41e62 881 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
fc9fe822
KS		 882 ret = 0;
883 goto out_unlock;
884 }
de466bd6
MG		 885
886 /* mmap_sem prevents this happening but warn if that changes */
887 WARN_ON(pmd_trans_migrating(pmd));
888
71e3aac0
AA		 889 if (unlikely(pmd_trans_splitting(pmd))) {
890 /* split huge page running from under us */
c4088ebd
KS		 891 spin_unlock(src_ptl);
892 spin_unlock(dst_ptl);
71e3aac0
AA		 893 pte_free(dst_mm, pgtable);
894
895 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
896 goto out;
897 }
898 src_page = pmd_page(pmd);
899 VM_BUG_ON(!PageHead(src_page));
900 get_page(src_page);
901 page_dup_rmap(src_page);
902 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
903
904 pmdp_set_wrprotect(src_mm, addr, src_pmd);
905 pmd = pmd_mkold(pmd_wrprotect(pmd));
6b0b50b0 906 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
71e3aac0 907 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
e1f56c89 908 atomic_long_inc(&dst_mm->nr_ptes);
71e3aac0
AA		 909
910 ret = 0;
911out_unlock:
c4088ebd
KS		 912 spin_unlock(src_ptl);
913 spin_unlock(dst_ptl);
71e3aac0
AA		 914out:
915 return ret;
916}
917
a1dd450b
WD		 918void huge_pmd_set_accessed(struct mm_struct *mm,
919 struct vm_area_struct *vma,
920 unsigned long address,
921 pmd_t *pmd, pmd_t orig_pmd,
922 int dirty)
923{
c4088ebd 924 spinlock_t *ptl;
a1dd450b
WD		 925 pmd_t entry;
926 unsigned long haddr;
927
c4088ebd 928 ptl = pmd_lock(mm, pmd);
a1dd450b
WD		 929 if (unlikely(!pmd_same(*pmd, orig_pmd)))
930 goto unlock;
931
932 entry = pmd_mkyoung(orig_pmd);
933 haddr = address & HPAGE_PMD_MASK;
934 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
935 update_mmu_cache_pmd(vma, address, pmd);
936
937unlock:
c4088ebd 938 spin_unlock(ptl);
a1dd450b
WD		 939}
940
93b4796d
KS		 941static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
942 struct vm_area_struct *vma, unsigned long address,
3ea41e62 943 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
93b4796d 944{
c4088ebd 945 spinlock_t *ptl;
93b4796d
KS		 946 pgtable_t pgtable;
947 pmd_t _pmd;
948 struct page *page;
949 int i, ret = 0;
950 unsigned long mmun_start; /* For mmu_notifiers */
951 unsigned long mmun_end; /* For mmu_notifiers */
952
953 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
954 if (!page) {
955 ret |= VM_FAULT_OOM;
956 goto out;
957 }
958
959 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
960 put_page(page);
961 ret |= VM_FAULT_OOM;
962 goto out;
963 }
964
965 clear_user_highpage(page, address);
966 __SetPageUptodate(page);
967
968 mmun_start = haddr;
969 mmun_end = haddr + HPAGE_PMD_SIZE;
970 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
971
c4088ebd 972 ptl = pmd_lock(mm, pmd);
3ea41e62
KS		 973 if (unlikely(!pmd_same(*pmd, orig_pmd)))
974 goto out_free_page;
975
93b4796d
KS		 976 pmdp_clear_flush(vma, haddr, pmd);
977 /* leave pmd empty until pte is filled */
978
6b0b50b0 979 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
93b4796d
KS		 980 pmd_populate(mm, &_pmd, pgtable);
981
982 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
983 pte_t *pte, entry;
984 if (haddr == (address & PAGE_MASK)) {
985 entry = mk_pte(page, vma->vm_page_prot);
986 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
987 page_add_new_anon_rmap(page, vma, haddr);
988 } else {
989 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
990 entry = pte_mkspecial(entry);
991 }
992 pte = pte_offset_map(&_pmd, haddr);
993 VM_BUG_ON(!pte_none(*pte));
994 set_pte_at(mm, haddr, pte, entry);
995 pte_unmap(pte);
996 }
997 smp_wmb(); /* make pte visible before pmd */
998 pmd_populate(mm, pmd, pgtable);
c4088ebd 999 spin_unlock(ptl);
97ae1749 1000 put_huge_zero_page();
93b4796d
KS		 1001 inc_mm_counter(mm, MM_ANONPAGES);
1002
1003 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1004
1005 ret |= VM_FAULT_WRITE;
1006out:
1007 return ret;
3ea41e62 1008out_free_page:
c4088ebd 1009 spin_unlock(ptl);
3ea41e62
KS		 1010 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1011 mem_cgroup_uncharge_page(page);
1012 put_page(page);
1013 goto out;
93b4796d
KS		 1014}
1015
71e3aac0
AA		 1016static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1017 struct vm_area_struct *vma,
1018 unsigned long address,
1019 pmd_t *pmd, pmd_t orig_pmd,
1020 struct page *page,
1021 unsigned long haddr)
1022{
c4088ebd 1023 spinlock_t *ptl;
71e3aac0
AA		 1024 pgtable_t pgtable;
1025 pmd_t _pmd;
1026 int ret = 0, i;
1027 struct page **pages;
2ec74c3e
SG		 1028 unsigned long mmun_start; /* For mmu_notifiers */
1029 unsigned long mmun_end; /* For mmu_notifiers */
71e3aac0
AA		 1030
1031 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1032 GFP_KERNEL);
1033 if (unlikely(!pages)) {
1034 ret |= VM_FAULT_OOM;
1035 goto out;
1036 }
1037
1038 for (i = 0; i < HPAGE_PMD_NR; i++) {
cc5d462f
AK		 1039 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1040 __GFP_OTHER_NODE,
19ee151e 1041 vma, address, page_to_nid(page));
b9bbfbe3
AA		 1042 if (unlikely(!pages[i] ||
1043 mem_cgroup_newpage_charge(pages[i], mm,
1044 GFP_KERNEL))) {
1045 if (pages[i])
71e3aac0 1046 put_page(pages[i]);
b9bbfbe3
AA		 1047 mem_cgroup_uncharge_start();
1048 while (--i >= 0) {
1049 mem_cgroup_uncharge_page(pages[i]);
1050 put_page(pages[i]);
1051 }
1052 mem_cgroup_uncharge_end();
71e3aac0
AA		 1053 kfree(pages);
1054 ret |= VM_FAULT_OOM;
1055 goto out;
1056 }
1057 }
1058
1059 for (i = 0; i < HPAGE_PMD_NR; i++) {
1060 copy_user_highpage(pages[i], page + i,
0089e485 1061 haddr + PAGE_SIZE * i, vma);
71e3aac0
AA		 1062 __SetPageUptodate(pages[i]);
1063 cond_resched();
1064 }
1065
2ec74c3e
SG		 1066 mmun_start = haddr;
1067 mmun_end = haddr + HPAGE_PMD_SIZE;
1068 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1069
c4088ebd 1070 ptl = pmd_lock(mm, pmd);
71e3aac0
AA		 1071 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1072 goto out_free_pages;
1073 VM_BUG_ON(!PageHead(page));
1074
2ec74c3e 1075 pmdp_clear_flush(vma, haddr, pmd);
71e3aac0
AA		 1076 /* leave pmd empty until pte is filled */
1077
6b0b50b0 1078 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
71e3aac0
AA		 1079 pmd_populate(mm, &_pmd, pgtable);
1080
1081 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1082 pte_t *pte, entry;
1083 entry = mk_pte(pages[i], vma->vm_page_prot);
1084 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1085 page_add_new_anon_rmap(pages[i], vma, haddr);
1086 pte = pte_offset_map(&_pmd, haddr);
1087 VM_BUG_ON(!pte_none(*pte));
1088 set_pte_at(mm, haddr, pte, entry);
1089 pte_unmap(pte);
1090 }
1091 kfree(pages);
1092
71e3aac0
AA		 1093 smp_wmb(); /* make pte visible before pmd */
1094 pmd_populate(mm, pmd, pgtable);
1095 page_remove_rmap(page);
c4088ebd 1096 spin_unlock(ptl);
71e3aac0 1097
2ec74c3e
SG		 1098 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1099
71e3aac0
AA		 1100 ret |= VM_FAULT_WRITE;
1101 put_page(page);
1102
1103out:
1104 return ret;
1105
1106out_free_pages:
c4088ebd 1107 spin_unlock(ptl);
2ec74c3e 1108 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
b9bbfbe3
AA		 1109 mem_cgroup_uncharge_start();
1110 for (i = 0; i < HPAGE_PMD_NR; i++) {
1111 mem_cgroup_uncharge_page(pages[i]);
71e3aac0 1112 put_page(pages[i]);
b9bbfbe3
AA		 1113 }
1114 mem_cgroup_uncharge_end();
71e3aac0
AA		 1115 kfree(pages);
1116 goto out;
1117}
1118
1119int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1120 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1121{
c4088ebd 1122 spinlock_t *ptl;
71e3aac0 1123 int ret = 0;
93b4796d 1124 struct page *page = NULL, *new_page;
71e3aac0 1125 unsigned long haddr;
2ec74c3e
SG		 1126 unsigned long mmun_start; /* For mmu_notifiers */
1127 unsigned long mmun_end; /* For mmu_notifiers */
71e3aac0 1128
c4088ebd 1129 ptl = pmd_lockptr(mm, pmd);
71e3aac0 1130 VM_BUG_ON(!vma->anon_vma);
93b4796d
KS		 1131 haddr = address & HPAGE_PMD_MASK;
1132 if (is_huge_zero_pmd(orig_pmd))
1133 goto alloc;
c4088ebd 1134 spin_lock(ptl);
71e3aac0
AA		 1135 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1136 goto out_unlock;
1137
1138 page = pmd_page(orig_pmd);
1139 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
71e3aac0
AA		 1140 if (page_mapcount(page) == 1) {
1141 pmd_t entry;
1142 entry = pmd_mkyoung(orig_pmd);
1143 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1144 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
b113da65 1145 update_mmu_cache_pmd(vma, address, pmd);
71e3aac0
AA		 1146 ret |= VM_FAULT_WRITE;
1147 goto out_unlock;
1148 }
1149 get_page(page);
c4088ebd 1150 spin_unlock(ptl);
93b4796d 1151alloc:
71e3aac0
AA		 1152 if (transparent_hugepage_enabled(vma) &&
1153 !transparent_hugepage_debug_cow())
0bbbc0b3 1154 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
cc5d462f 1155 vma, haddr, numa_node_id(), 0);
71e3aac0
AA		 1156 else
1157 new_page = NULL;
1158
1159 if (unlikely(!new_page)) {
93b4796d
KS		 1160 if (is_huge_zero_pmd(orig_pmd)) {
1161 ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
3ea41e62 1162 address, pmd, orig_pmd, haddr);
93b4796d
KS		 1163 } else {
1164 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1165 pmd, orig_pmd, page, haddr);
1166 if (ret & VM_FAULT_OOM)
1167 split_huge_page(page);
1168 put_page(page);
1169 }
17766dde 1170 count_vm_event(THP_FAULT_FALLBACK);
71e3aac0
AA		 1171 goto out;
1172 }
1173
b9bbfbe3
AA		 1174 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1175 put_page(new_page);
93b4796d
KS		 1176 if (page) {
1177 split_huge_page(page);
1178 put_page(page);
1179 }
17766dde 1180 count_vm_event(THP_FAULT_FALLBACK);
b9bbfbe3
AA		 1181 ret |= VM_FAULT_OOM;
1182 goto out;
1183 }
1184
17766dde
DR		 1185 count_vm_event(THP_FAULT_ALLOC);
1186
93b4796d
KS		 1187 if (is_huge_zero_pmd(orig_pmd))
1188 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1189 else
1190 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
71e3aac0
AA		 1191 __SetPageUptodate(new_page);
1192
2ec74c3e
SG		 1193 mmun_start = haddr;
1194 mmun_end = haddr + HPAGE_PMD_SIZE;
1195 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1196
c4088ebd 1197 spin_lock(ptl);
93b4796d
KS		 1198 if (page)
1199 put_page(page);
b9bbfbe3 1200 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
c4088ebd 1201 spin_unlock(ptl);
b9bbfbe3 1202 mem_cgroup_uncharge_page(new_page);
71e3aac0 1203 put_page(new_page);
2ec74c3e 1204 goto out_mn;
b9bbfbe3 1205 } else {
71e3aac0 1206 pmd_t entry;
3122359a
KS		 1207 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1208 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2ec74c3e 1209 pmdp_clear_flush(vma, haddr, pmd);
71e3aac0
AA		 1210 page_add_new_anon_rmap(new_page, vma, haddr);
1211 set_pmd_at(mm, haddr, pmd, entry);
b113da65 1212 update_mmu_cache_pmd(vma, address, pmd);
97ae1749 1213 if (is_huge_zero_pmd(orig_pmd)) {
93b4796d 1214 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
97ae1749
KS		 1215 put_huge_zero_page();
1216 } else {
93b4796d
KS		 1217 VM_BUG_ON(!PageHead(page));
1218 page_remove_rmap(page);
1219 put_page(page);
1220 }
71e3aac0
AA		 1221 ret |= VM_FAULT_WRITE;
1222 }
c4088ebd 1223 spin_unlock(ptl);
2ec74c3e
SG		 1224out_mn:
1225 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 1226out:
1227 return ret;
2ec74c3e 1228out_unlock:
c4088ebd 1229 spin_unlock(ptl);
2ec74c3e 1230 return ret;
71e3aac0
AA		 1231}
1232
b676b293 1233struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
71e3aac0
AA		 1234 unsigned long addr,
1235 pmd_t *pmd,
1236 unsigned int flags)
1237{
b676b293 1238 struct mm_struct *mm = vma->vm_mm;
71e3aac0
AA		 1239 struct page *page = NULL;
1240
c4088ebd 1241 assert_spin_locked(pmd_lockptr(mm, pmd));
71e3aac0
AA		 1242
1243 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1244 goto out;
1245
85facf25
KS		 1246 /* Avoid dumping huge zero page */
1247 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1248 return ERR_PTR(-EFAULT);
1249
2b4847e7
MG		 1250 /* Full NUMA hinting faults to serialise migration in fault paths */
1251 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1252 goto out;
1253
71e3aac0
AA		 1254 page = pmd_page(*pmd);
1255 VM_BUG_ON(!PageHead(page));
1256 if (flags & FOLL_TOUCH) {
1257 pmd_t _pmd;
1258 /*
1259 * We should set the dirty bit only for FOLL_WRITE but
1260 * for now the dirty bit in the pmd is meaningless.
1261 * And if the dirty bit will become meaningful and
1262 * we'll only set it with FOLL_WRITE, an atomic
1263 * set_bit will be required on the pmd to set the
1264 * young bit, instead of the current set_pmd_at.
1265 */
1266 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
8663890a
AK		 1267 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1268 pmd, _pmd, 1))
1269 update_mmu_cache_pmd(vma, addr, pmd);
71e3aac0 1270 }
b676b293
DR		 1271 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1272 if (page->mapping && trylock_page(page)) {
1273 lru_add_drain();
1274 if (page->mapping)
1275 mlock_vma_page(page);
1276 unlock_page(page);
1277 }
1278 }
71e3aac0
AA		 1279 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1280 VM_BUG_ON(!PageCompound(page));
1281 if (flags & FOLL_GET)
70b50f94 1282 get_page_foll(page);
71e3aac0
AA		 1283
1284out:
1285 return page;
1286}
1287
d10e63f2 1288/* NUMA hinting page fault entry point for trans huge pmds */
4daae3b4
MG		 1289int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1290 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
d10e63f2 1291{
c4088ebd 1292 spinlock_t *ptl;
b8916634 1293 struct anon_vma *anon_vma = NULL;
b32967ff 1294 struct page *page;
d10e63f2 1295 unsigned long haddr = addr & HPAGE_PMD_MASK;
8191acbd 1296 int page_nid = -1, this_nid = numa_node_id();
90572890 1297 int target_nid, last_cpupid = -1;
8191acbd
MG		 1298 bool page_locked;
1299 bool migrated = false;
6688cc05 1300 int flags = 0;
d10e63f2 1301
c4088ebd 1302 ptl = pmd_lock(mm, pmdp);
d10e63f2
MG		 1303 if (unlikely(!pmd_same(pmd, *pmdp)))
1304 goto out_unlock;
1305
de466bd6
MG		 1306 /*
1307 * If there are potential migrations, wait for completion and retry
1308 * without disrupting NUMA hinting information. Do not relock and
1309 * check_same as the page may no longer be mapped.
1310 */
1311 if (unlikely(pmd_trans_migrating(*pmdp))) {
1312 spin_unlock(ptl);
1313 wait_migrate_huge_page(vma->anon_vma, pmdp);
1314 goto out;
1315 }
1316
d10e63f2 1317 page = pmd_page(pmd);
a1a46184 1318 BUG_ON(is_huge_zero_page(page));
8191acbd 1319 page_nid = page_to_nid(page);
90572890 1320 last_cpupid = page_cpupid_last(page);
03c5a6e1 1321 count_vm_numa_event(NUMA_HINT_FAULTS);
04bb2f94 1322 if (page_nid == this_nid) {
03c5a6e1 1323 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
04bb2f94
RR		 1324 flags |= TNF_FAULT_LOCAL;
1325 }
4daae3b4 1326
6688cc05
PZ		 1327 /*
1328 * Avoid grouping on DSO/COW pages in specific and RO pages
1329 * in general, RO pages shouldn't hurt as much anyway since
1330 * they can be in shared cache state.
1331 */
1332 if (!pmd_write(pmd))
1333 flags |= TNF_NO_GROUP;
1334
ff9042b1
MG		 1335 /*
1336 * Acquire the page lock to serialise THP migrations but avoid dropping
1337 * page_table_lock if at all possible
1338 */
b8916634
MG		 1339 page_locked = trylock_page(page);
1340 target_nid = mpol_misplaced(page, vma, haddr);
1341 if (target_nid == -1) {
1342 /* If the page was locked, there are no parallel migrations */
a54a407f 1343 if (page_locked)
b8916634 1344 goto clear_pmdnuma;
2b4847e7 1345 }
4daae3b4 1346
de466bd6 1347 /* Migration could have started since the pmd_trans_migrating check */
2b4847e7 1348 if (!page_locked) {
c4088ebd 1349 spin_unlock(ptl);
b8916634 1350 wait_on_page_locked(page);
a54a407f 1351 page_nid = -1;
b8916634
MG		 1352 goto out;
1353 }
1354
2b4847e7
MG		 1355 /*
1356 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1357 * to serialises splits
1358 */
b8916634 1359 get_page(page);
c4088ebd 1360 spin_unlock(ptl);
b8916634 1361 anon_vma = page_lock_anon_vma_read(page);
4daae3b4 1362
c69307d5 1363 /* Confirm the PMD did not change while page_table_lock was released */
c4088ebd 1364 spin_lock(ptl);
b32967ff
MG		 1365 if (unlikely(!pmd_same(pmd, *pmdp))) {
1366 unlock_page(page);
1367 put_page(page);
a54a407f 1368 page_nid = -1;
4daae3b4 1369 goto out_unlock;
b32967ff 1370 }
ff9042b1 1371
c3a489ca
MG		 1372 /* Bail if we fail to protect against THP splits for any reason */
1373 if (unlikely(!anon_vma)) {
1374 put_page(page);
1375 page_nid = -1;
1376 goto clear_pmdnuma;
1377 }
1378
20841405
RR		 1379 /*
1380 * The page_table_lock above provides a memory barrier
1381 * with change_protection_range.
1382 */
1383 if (mm_tlb_flush_pending(mm))
1384 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1385
a54a407f
MG		 1386 /*
1387 * Migrate the THP to the requested node, returns with page unlocked
1388 * and pmd_numa cleared.
1389 */
c4088ebd 1390 spin_unlock(ptl);
b32967ff 1391 migrated = migrate_misplaced_transhuge_page(mm, vma,
340ef390 1392 pmdp, pmd, addr, page, target_nid);
6688cc05
PZ		 1393 if (migrated) {
1394 flags |= TNF_MIGRATED;
8191acbd 1395 page_nid = target_nid;
6688cc05 1396 }
b32967ff 1397
8191acbd 1398 goto out;
b32967ff 1399clear_pmdnuma:
a54a407f 1400 BUG_ON(!PageLocked(page));
d10e63f2
MG		 1401 pmd = pmd_mknonnuma(pmd);
1402 set_pmd_at(mm, haddr, pmdp, pmd);
1403 VM_BUG_ON(pmd_numa(*pmdp));
1404 update_mmu_cache_pmd(vma, addr, pmdp);
a54a407f 1405 unlock_page(page);
d10e63f2 1406out_unlock:
c4088ebd 1407 spin_unlock(ptl);
b8916634
MG		 1408
1409out:
1410 if (anon_vma)
1411 page_unlock_anon_vma_read(anon_vma);
1412
8191acbd 1413 if (page_nid != -1)
6688cc05 1414 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
8191acbd 1415
d10e63f2
MG		 1416 return 0;
1417}
1418
71e3aac0 1419int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
f21760b1 1420 pmd_t *pmd, unsigned long addr)
71e3aac0 1421{
bf929152 1422 spinlock_t *ptl;
71e3aac0
AA		 1423 int ret = 0;
1424
bf929152 1425 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
025c5b24
NH		 1426 struct page *page;
1427 pgtable_t pgtable;
f5c8ad47 1428 pmd_t orig_pmd;
a6bf2bb0
AK		 1429 /*
1430 * For architectures like ppc64 we look at deposited pgtable
1431 * when calling pmdp_get_and_clear. So do the
1432 * pgtable_trans_huge_withdraw after finishing pmdp related
1433 * operations.
1434 */
f5c8ad47 1435 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
025c5b24 1436 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
a6bf2bb0 1437 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
479f0abb 1438 if (is_huge_zero_pmd(orig_pmd)) {
e1f56c89 1439 atomic_long_dec(&tlb->mm->nr_ptes);
bf929152 1440 spin_unlock(ptl);
97ae1749 1441 put_huge_zero_page();
479f0abb
KS		 1442 } else {
1443 page = pmd_page(orig_pmd);
1444 page_remove_rmap(page);
1445 VM_BUG_ON(page_mapcount(page) < 0);
1446 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1447 VM_BUG_ON(!PageHead(page));
e1f56c89 1448 atomic_long_dec(&tlb->mm->nr_ptes);
bf929152 1449 spin_unlock(ptl);
479f0abb
KS		 1450 tlb_remove_page(tlb, page);
1451 }
025c5b24
NH		 1452 pte_free(tlb->mm, pgtable);
1453 ret = 1;
1454 }
71e3aac0
AA		 1455 return ret;
1456}
1457
0ca1634d
JW		 1458int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1459 unsigned long addr, unsigned long end,
1460 unsigned char *vec)
1461{
bf929152 1462 spinlock_t *ptl;
0ca1634d
JW		 1463 int ret = 0;
1464
bf929152 1465 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
025c5b24
NH		 1466 /*
1467 * All logical pages in the range are present
1468 * if backed by a huge page.
1469 */
bf929152 1470 spin_unlock(ptl);
025c5b24
NH		 1471 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1472 ret = 1;
1473 }
0ca1634d
JW		 1474
1475 return ret;
1476}
1477
37a1c49a
AA		 1478int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1479 unsigned long old_addr,
1480 unsigned long new_addr, unsigned long old_end,
1481 pmd_t *old_pmd, pmd_t *new_pmd)
1482{
bf929152 1483 spinlock_t *old_ptl, *new_ptl;
37a1c49a
AA		 1484 int ret = 0;
1485 pmd_t pmd;
1486
1487 struct mm_struct *mm = vma->vm_mm;
1488
1489 if ((old_addr & ~HPAGE_PMD_MASK) ||
1490 (new_addr & ~HPAGE_PMD_MASK) ||
1491 old_end - old_addr < HPAGE_PMD_SIZE ||
1492 (new_vma->vm_flags & VM_NOHUGEPAGE))
1493 goto out;
1494
1495 /*
1496 * The destination pmd shouldn't be established, free_pgtables()
1497 * should have release it.
1498 */
1499 if (WARN_ON(!pmd_none(*new_pmd))) {
1500 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1501 goto out;
1502 }
1503
bf929152
KS		 1504 /*
1505 * We don't have to worry about the ordering of src and dst
1506 * ptlocks because exclusive mmap_sem prevents deadlock.
1507 */
1508 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
025c5b24 1509 if (ret == 1) {
bf929152
KS		 1510 new_ptl = pmd_lockptr(mm, new_pmd);
1511 if (new_ptl != old_ptl)
1512 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
025c5b24
NH		 1513 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1514 VM_BUG_ON(!pmd_none(*new_pmd));
0f8975ec 1515 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
3592806c
KS		 1516 if (new_ptl != old_ptl) {
1517 pgtable_t pgtable;
1518
1519 /*
1520 * Move preallocated PTE page table if new_pmd is on
1521 * different PMD page table.
1522 */
1523 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1524 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1525
bf929152 1526 spin_unlock(new_ptl);
3592806c 1527 }
bf929152 1528 spin_unlock(old_ptl);
37a1c49a
AA		 1529 }
1530out:
1531 return ret;
1532}
1533
f123d74a
MG		 1534/*
1535 * Returns
1536 * - 0 if PMD could not be locked
1537 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1538 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1539 */
cd7548ab 1540int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
4b10e7d5 1541 unsigned long addr, pgprot_t newprot, int prot_numa)
cd7548ab
JW		 1542{
1543 struct mm_struct *mm = vma->vm_mm;
bf929152 1544 spinlock_t *ptl;
cd7548ab
JW		 1545 int ret = 0;
1546
bf929152 1547 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
025c5b24 1548 pmd_t entry;
f123d74a 1549 ret = 1;
a4f1de17 1550 if (!prot_numa) {
f123d74a 1551 entry = pmdp_get_and_clear(mm, addr, pmd);
1667918b
MG		 1552 if (pmd_numa(entry))
1553 entry = pmd_mknonnuma(entry);
4b10e7d5 1554 entry = pmd_modify(entry, newprot);
f123d74a 1555 ret = HPAGE_PMD_NR;
a4f1de17
HD		 1556 BUG_ON(pmd_write(entry));
1557 } else {
4b10e7d5
MG		 1558 struct page *page = pmd_page(*pmd);
1559
a1a46184 1560 /*
1bc115d8
MG		 1561 * Do not trap faults against the zero page. The
1562 * read-only data is likely to be read-cached on the
1563 * local CPU cache and it is less useful to know about
1564 * local vs remote hits on the zero page.
a1a46184 1565 */
1bc115d8 1566 if (!is_huge_zero_page(page) &&
4b10e7d5 1567 !pmd_numa(*pmd)) {
5a6dac3e 1568 entry = *pmd;
4b10e7d5 1569 entry = pmd_mknuma(entry);
f123d74a 1570 ret = HPAGE_PMD_NR;
4b10e7d5
MG		 1571 }
1572 }
f123d74a
MG		 1573
1574 /* Set PMD if cleared earlier */
1575 if (ret == HPAGE_PMD_NR)
1576 set_pmd_at(mm, addr, pmd, entry);
1577
bf929152 1578 spin_unlock(ptl);
025c5b24
NH		 1579 }
1580
1581 return ret;
1582}
1583
1584/*
1585 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1586 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1587 *
1588 * Note that if it returns 1, this routine returns without unlocking page
1589 * table locks. So callers must unlock them.
1590 */
bf929152
KS		 1591int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1592 spinlock_t **ptl)
025c5b24 1593{
bf929152 1594 *ptl = pmd_lock(vma->vm_mm, pmd);
cd7548ab
JW		 1595 if (likely(pmd_trans_huge(*pmd))) {
1596 if (unlikely(pmd_trans_splitting(*pmd))) {
bf929152 1597 spin_unlock(*ptl);
cd7548ab 1598 wait_split_huge_page(vma->anon_vma, pmd);
025c5b24 1599 return -1;
cd7548ab 1600 } else {
025c5b24
NH		 1601 /* Thp mapped by 'pmd' is stable, so we can
1602 * handle it as it is. */
1603 return 1;
cd7548ab 1604 }
025c5b24 1605 }
bf929152 1606 spin_unlock(*ptl);
025c5b24 1607 return 0;
cd7548ab
JW		 1608}
1609
117b0791
KS		 1610/*
1611 * This function returns whether a given @page is mapped onto the @address
1612 * in the virtual space of @mm.
1613 *
1614 * When it's true, this function returns *pmd with holding the page table lock
1615 * and passing it back to the caller via @ptl.
1616 * If it's false, returns NULL without holding the page table lock.
1617 */
71e3aac0
AA		 1618pmd_t *page_check_address_pmd(struct page *page,
1619 struct mm_struct *mm,
1620 unsigned long address,
117b0791
KS		 1621 enum page_check_address_pmd_flag flag,
1622 spinlock_t **ptl)
71e3aac0 1623{
117b0791 1624 pmd_t *pmd;
71e3aac0
AA		 1625
1626 if (address & ~HPAGE_PMD_MASK)
117b0791 1627 return NULL;
71e3aac0 1628
6219049a
BL		 1629 pmd = mm_find_pmd(mm, address);
1630 if (!pmd)
117b0791
KS		 1631 return NULL;
1632 *ptl = pmd_lock(mm, pmd);
71e3aac0 1633 if (pmd_none(*pmd))
117b0791 1634 goto unlock;
71e3aac0 1635 if (pmd_page(*pmd) != page)
117b0791 1636 goto unlock;
94fcc585
AA		 1637 /*
1638 * split_vma() may create temporary aliased mappings. There is
1639 * no risk as long as all huge pmd are found and have their
1640 * splitting bit set before __split_huge_page_refcount
1641 * runs. Finding the same huge pmd more than once during the
1642 * same rmap walk is not a problem.
1643 */
1644 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1645 pmd_trans_splitting(*pmd))
117b0791 1646 goto unlock;
71e3aac0
AA		 1647 if (pmd_trans_huge(*pmd)) {
1648 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1649 !pmd_trans_splitting(*pmd));
117b0791 1650 return pmd;
71e3aac0 1651 }
117b0791
KS		 1652unlock:
1653 spin_unlock(*ptl);
1654 return NULL;
71e3aac0
AA		 1655}
1656
1657static int __split_huge_page_splitting(struct page *page,
1658 struct vm_area_struct *vma,
1659 unsigned long address)
1660{
1661 struct mm_struct *mm = vma->vm_mm;
117b0791 1662 spinlock_t *ptl;
71e3aac0
AA		 1663 pmd_t *pmd;
1664 int ret = 0;
2ec74c3e
SG		 1665 /* For mmu_notifiers */
1666 const unsigned long mmun_start = address;
1667 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
71e3aac0 1668
2ec74c3e 1669 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
71e3aac0 1670 pmd = page_check_address_pmd(page, mm, address,
117b0791 1671 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
71e3aac0
AA		 1672 if (pmd) {
1673 /*
1674 * We can't temporarily set the pmd to null in order
1675 * to split it, the pmd must remain marked huge at all
1676 * times or the VM won't take the pmd_trans_huge paths
5a505085 1677 * and it won't wait on the anon_vma->root->rwsem to
71e3aac0
AA		 1678 * serialize against split_huge_page*.
1679 */
2ec74c3e 1680 pmdp_splitting_flush(vma, address, pmd);
71e3aac0 1681 ret = 1;
117b0791 1682 spin_unlock(ptl);
71e3aac0 1683 }
2ec74c3e 1684 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 1685
1686 return ret;
1687}
1688
5bc7b8ac
SL		 1689static void __split_huge_page_refcount(struct page *page,
1690 struct list_head *list)
71e3aac0
AA		 1691{
1692 int i;
71e3aac0 1693 struct zone *zone = page_zone(page);
fa9add64 1694 struct lruvec *lruvec;
70b50f94 1695 int tail_count = 0;
71e3aac0
AA		 1696
1697 /* prevent PageLRU to go away from under us, and freeze lru stats */
1698 spin_lock_irq(&zone->lru_lock);
fa9add64
HD		 1699 lruvec = mem_cgroup_page_lruvec(page, zone);
1700
71e3aac0 1701 compound_lock(page);
e94c8a9c
KH		 1702 /* complete memcg works before add pages to LRU */
1703 mem_cgroup_split_huge_fixup(page);
71e3aac0 1704
45676885 1705 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
71e3aac0
AA		 1706 struct page *page_tail = page + i;
1707
70b50f94
AA		 1708 /* tail_page->_mapcount cannot change */
1709 BUG_ON(page_mapcount(page_tail) < 0);
1710 tail_count += page_mapcount(page_tail);
1711 /* check for overflow */
1712 BUG_ON(tail_count < 0);
1713 BUG_ON(atomic_read(&page_tail->_count) != 0);
1714 /*
1715 * tail_page->_count is zero and not changing from
1716 * under us. But get_page_unless_zero() may be running
1717 * from under us on the tail_page. If we used
1718 * atomic_set() below instead of atomic_add(), we
1719 * would then run atomic_set() concurrently with
1720 * get_page_unless_zero(), and atomic_set() is
1721 * implemented in C not using locked ops. spin_unlock
1722 * on x86 sometime uses locked ops because of PPro
1723 * errata 66, 92, so unless somebody can guarantee
1724 * atomic_set() here would be safe on all archs (and
1725 * not only on x86), it's safer to use atomic_add().
1726 */
1727 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1728 &page_tail->_count);
71e3aac0
AA		 1729
1730 /* after clearing PageTail the gup refcount can be released */
1731 smp_mb();
1732
a6d30ddd
JD		 1733 /*
1734 * retain hwpoison flag of the poisoned tail page:
1735 * fix for the unsuitable process killed on Guest Machine(KVM)
1736 * by the memory-failure.
1737 */
1738 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
71e3aac0
AA		 1739 page_tail->flags |= (page->flags &
1740 ((1L << PG_referenced) |
1741 (1L << PG_swapbacked) |
1742 (1L << PG_mlocked) |
e180cf80
KS		 1743 (1L << PG_uptodate) |
1744 (1L << PG_active) |
1745 (1L << PG_unevictable)));
71e3aac0
AA		 1746 page_tail->flags |= (1L << PG_dirty);
1747
70b50f94 1748 /* clear PageTail before overwriting first_page */
71e3aac0
AA		 1749 smp_wmb();
1750
1751 /*
1752 * __split_huge_page_splitting() already set the
1753 * splitting bit in all pmd that could map this
1754 * hugepage, that will ensure no CPU can alter the
1755 * mapcount on the head page. The mapcount is only
1756 * accounted in the head page and it has to be
1757 * transferred to all tail pages in the below code. So
1758 * for this code to be safe, the split the mapcount
1759 * can't change. But that doesn't mean userland can't
1760 * keep changing and reading the page contents while
1761 * we transfer the mapcount, so the pmd splitting
1762 * status is achieved setting a reserved bit in the
1763 * pmd, not by clearing the present bit.
1764 */
71e3aac0
AA		 1765 page_tail->_mapcount = page->_mapcount;
1766
1767 BUG_ON(page_tail->mapping);
1768 page_tail->mapping = page->mapping;
1769
45676885 1770 page_tail->index = page->index + i;
90572890 1771 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
71e3aac0
AA		 1772
1773 BUG_ON(!PageAnon(page_tail));
1774 BUG_ON(!PageUptodate(page_tail));
1775 BUG_ON(!PageDirty(page_tail));
1776 BUG_ON(!PageSwapBacked(page_tail));
1777
5bc7b8ac 1778 lru_add_page_tail(page, page_tail, lruvec, list);
71e3aac0 1779 }
70b50f94
AA		 1780 atomic_sub(tail_count, &page->_count);
1781 BUG_ON(atomic_read(&page->_count) <= 0);
71e3aac0 1782
fa9add64 1783 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
79134171 1784
71e3aac0
AA		 1785 ClearPageCompound(page);
1786 compound_unlock(page);
1787 spin_unlock_irq(&zone->lru_lock);
1788
1789 for (i = 1; i < HPAGE_PMD_NR; i++) {
1790 struct page *page_tail = page + i;
1791 BUG_ON(page_count(page_tail) <= 0);
1792 /*
1793 * Tail pages may be freed if there wasn't any mapping
1794 * like if add_to_swap() is running on a lru page that
1795 * had its mapping zapped. And freeing these pages
1796 * requires taking the lru_lock so we do the put_page
1797 * of the tail pages after the split is complete.
1798 */
1799 put_page(page_tail);
1800 }
1801
1802 /*
1803 * Only the head page (now become a regular page) is required
1804 * to be pinned by the caller.
1805 */
1806 BUG_ON(page_count(page) <= 0);
1807}
1808
1809static int __split_huge_page_map(struct page *page,
1810 struct vm_area_struct *vma,
1811 unsigned long address)
1812{
1813 struct mm_struct *mm = vma->vm_mm;
117b0791 1814 spinlock_t *ptl;
71e3aac0
AA		 1815 pmd_t *pmd, _pmd;
1816 int ret = 0, i;
1817 pgtable_t pgtable;
1818 unsigned long haddr;
1819
71e3aac0 1820 pmd = page_check_address_pmd(page, mm, address,
117b0791 1821 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
71e3aac0 1822 if (pmd) {
6b0b50b0 1823 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
71e3aac0
AA		 1824 pmd_populate(mm, &_pmd, pgtable);
1825
e3ebcf64
GS		 1826 haddr = address;
1827 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
71e3aac0
AA		 1828 pte_t *pte, entry;
1829 BUG_ON(PageCompound(page+i));
1830 entry = mk_pte(page + i, vma->vm_page_prot);
1831 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1832 if (!pmd_write(*pmd))
1833 entry = pte_wrprotect(entry);
1834 else
1835 BUG_ON(page_mapcount(page) != 1);
1836 if (!pmd_young(*pmd))
1837 entry = pte_mkold(entry);
1ba6e0b5
AA		 1838 if (pmd_numa(*pmd))
1839 entry = pte_mknuma(entry);
71e3aac0
AA		 1840 pte = pte_offset_map(&_pmd, haddr);
1841 BUG_ON(!pte_none(*pte));
1842 set_pte_at(mm, haddr, pte, entry);
1843 pte_unmap(pte);
1844 }
1845
71e3aac0
AA		 1846 smp_wmb(); /* make pte visible before pmd */
1847 /*
1848 * Up to this point the pmd is present and huge and
1849 * userland has the whole access to the hugepage
1850 * during the split (which happens in place). If we
1851 * overwrite the pmd with the not-huge version
1852 * pointing to the pte here (which of course we could
1853 * if all CPUs were bug free), userland could trigger
1854 * a small page size TLB miss on the small sized TLB
1855 * while the hugepage TLB entry is still established
1856 * in the huge TLB. Some CPU doesn't like that. See
1857 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1858 * Erratum 383 on page 93. Intel should be safe but is
1859 * also warns that it's only safe if the permission
1860 * and cache attributes of the two entries loaded in
1861 * the two TLB is identical (which should be the case
1862 * here). But it is generally safer to never allow
1863 * small and huge TLB entries for the same virtual
1864 * address to be loaded simultaneously. So instead of
1865 * doing "pmd_populate(); flush_tlb_range();" we first
1866 * mark the current pmd notpresent (atomically because
1867 * here the pmd_trans_huge and pmd_trans_splitting
1868 * must remain set at all times on the pmd until the
1869 * split is complete for this pmd), then we flush the
1870 * SMP TLB and finally we write the non-huge version
1871 * of the pmd entry with pmd_populate.
1872 */
46dcde73 1873 pmdp_invalidate(vma, address, pmd);
71e3aac0
AA		 1874 pmd_populate(mm, pmd, pgtable);
1875 ret = 1;
117b0791 1876 spin_unlock(ptl);
71e3aac0 1877 }
71e3aac0
AA		 1878
1879 return ret;
1880}
1881
5a505085 1882/* must be called with anon_vma->root->rwsem held */
71e3aac0 1883static void __split_huge_page(struct page *page,
5bc7b8ac
SL		 1884 struct anon_vma *anon_vma,
1885 struct list_head *list)
71e3aac0
AA		 1886{
1887 int mapcount, mapcount2;
bf181b9f 1888 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
71e3aac0
AA		 1889 struct anon_vma_chain *avc;
1890
1891 BUG_ON(!PageHead(page));
1892 BUG_ON(PageTail(page));
1893
1894 mapcount = 0;
bf181b9f 1895 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
71e3aac0
AA		 1896 struct vm_area_struct *vma = avc->vma;
1897 unsigned long addr = vma_address(page, vma);
1898 BUG_ON(is_vma_temporary_stack(vma));
71e3aac0
AA		 1899 mapcount += __split_huge_page_splitting(page, vma, addr);
1900 }
05759d38
AA		 1901 /*
1902 * It is critical that new vmas are added to the tail of the
1903 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1904 * and establishes a child pmd before
1905 * __split_huge_page_splitting() freezes the parent pmd (so if
1906 * we fail to prevent copy_huge_pmd() from running until the
1907 * whole __split_huge_page() is complete), we will still see
1908 * the newly established pmd of the child later during the
1909 * walk, to be able to set it as pmd_trans_splitting too.
1910 */
1911 if (mapcount != page_mapcount(page))
1912 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1913 mapcount, page_mapcount(page));
71e3aac0
AA		 1914 BUG_ON(mapcount != page_mapcount(page));
1915
5bc7b8ac 1916 __split_huge_page_refcount(page, list);
71e3aac0
AA		 1917
1918 mapcount2 = 0;
bf181b9f 1919 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
71e3aac0
AA		 1920 struct vm_area_struct *vma = avc->vma;
1921 unsigned long addr = vma_address(page, vma);
1922 BUG_ON(is_vma_temporary_stack(vma));
71e3aac0
AA		 1923 mapcount2 += __split_huge_page_map(page, vma, addr);
1924 }
05759d38
AA		 1925 if (mapcount != mapcount2)
1926 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1927 mapcount, mapcount2, page_mapcount(page));
71e3aac0
AA		 1928 BUG_ON(mapcount != mapcount2);
1929}
1930
5bc7b8ac
SL		 1931/*
1932 * Split a hugepage into normal pages. This doesn't change the position of head
1933 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1934 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1935 * from the hugepage.
1936 * Return 0 if the hugepage is split successfully otherwise return 1.
1937 */
1938int split_huge_page_to_list(struct page *page, struct list_head *list)
71e3aac0
AA		 1939{
1940 struct anon_vma *anon_vma;
1941 int ret = 1;
1942
5918d10a 1943 BUG_ON(is_huge_zero_page(page));
71e3aac0 1944 BUG_ON(!PageAnon(page));
062f1af2
MG		 1945
1946 /*
1947 * The caller does not necessarily hold an mmap_sem that would prevent
1948 * the anon_vma disappearing so we first we take a reference to it
1949 * and then lock the anon_vma for write. This is similar to
1950 * page_lock_anon_vma_read except the write lock is taken to serialise
1951 * against parallel split or collapse operations.
1952 */
1953 anon_vma = page_get_anon_vma(page);
71e3aac0
AA		 1954 if (!anon_vma)
1955 goto out;
062f1af2
MG		 1956 anon_vma_lock_write(anon_vma);
1957
71e3aac0
AA		 1958 ret = 0;
1959 if (!PageCompound(page))
1960 goto out_unlock;
1961
1962 BUG_ON(!PageSwapBacked(page));
5bc7b8ac 1963 __split_huge_page(page, anon_vma, list);
81ab4201 1964 count_vm_event(THP_SPLIT);
71e3aac0
AA		 1965
1966 BUG_ON(PageCompound(page));
1967out_unlock:
08b52706 1968 anon_vma_unlock_write(anon_vma);
062f1af2 1969 put_anon_vma(anon_vma);
71e3aac0
AA		 1970out:
1971 return ret;
1972}
1973
4b6e1e37 1974#define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
78f11a25 1975
60ab3244
AA		 1976int hugepage_madvise(struct vm_area_struct *vma,
1977 unsigned long *vm_flags, int advice)
0af4e98b 1978{
8e72033f
GS		 1979 struct mm_struct *mm = vma->vm_mm;
1980
a664b2d8
AA		 1981 switch (advice) {
1982 case MADV_HUGEPAGE:
1983 /*
1984 * Be somewhat over-protective like KSM for now!
1985 */
78f11a25 1986 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
a664b2d8 1987 return -EINVAL;
8e72033f
GS		 1988 if (mm->def_flags & VM_NOHUGEPAGE)
1989 return -EINVAL;
a664b2d8
AA		 1990 *vm_flags &= ~VM_NOHUGEPAGE;
1991 *vm_flags |= VM_HUGEPAGE;
60ab3244
AA		 1992 /*
1993 * If the vma become good for khugepaged to scan,
1994 * register it here without waiting a page fault that
1995 * may not happen any time soon.
1996 */
1997 if (unlikely(khugepaged_enter_vma_merge(vma)))
1998 return -ENOMEM;
a664b2d8
AA		 1999 break;
2000 case MADV_NOHUGEPAGE:
2001 /*
2002 * Be somewhat over-protective like KSM for now!
2003 */
78f11a25 2004 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
a664b2d8
AA		 2005 return -EINVAL;
2006 *vm_flags &= ~VM_HUGEPAGE;
2007 *vm_flags |= VM_NOHUGEPAGE;
60ab3244
AA		 2008 /*
2009 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2010 * this vma even if we leave the mm registered in khugepaged if
2011 * it got registered before VM_NOHUGEPAGE was set.
2012 */
a664b2d8
AA		 2013 break;
2014 }
0af4e98b
AA		 2015
2016 return 0;
2017}
2018
ba76149f
AA		 2019static int __init khugepaged_slab_init(void)
2020{
2021 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2022 sizeof(struct mm_slot),
2023 __alignof__(struct mm_slot), 0, NULL);
2024 if (!mm_slot_cache)
2025 return -ENOMEM;
2026
2027 return 0;
2028}
2029
ba76149f
AA		 2030static inline struct mm_slot *alloc_mm_slot(void)
2031{
2032 if (!mm_slot_cache) /* initialization failed */
2033 return NULL;
2034 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2035}
2036
2037static inline void free_mm_slot(struct mm_slot *mm_slot)
2038{
2039 kmem_cache_free(mm_slot_cache, mm_slot);
2040}
2041
ba76149f
AA		 2042static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2043{
2044 struct mm_slot *mm_slot;
ba76149f 2045
b67bfe0d 2046 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
ba76149f
AA		 2047 if (mm == mm_slot->mm)
2048 return mm_slot;
43b5fbbd 2049
ba76149f
AA		 2050 return NULL;
2051}
2052
2053static void insert_to_mm_slots_hash(struct mm_struct *mm,
2054 struct mm_slot *mm_slot)
2055{
ba76149f 2056 mm_slot->mm = mm;
43b5fbbd 2057 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
ba76149f
AA		 2058}
2059
2060static inline int khugepaged_test_exit(struct mm_struct *mm)
2061{
2062 return atomic_read(&mm->mm_users) == 0;
2063}
2064
2065int __khugepaged_enter(struct mm_struct *mm)
2066{
2067 struct mm_slot *mm_slot;
2068 int wakeup;
2069
2070 mm_slot = alloc_mm_slot();
2071 if (!mm_slot)
2072 return -ENOMEM;
2073
2074 /* __khugepaged_exit() must not run from under us */
2075 VM_BUG_ON(khugepaged_test_exit(mm));
2076 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2077 free_mm_slot(mm_slot);
2078 return 0;
2079 }
2080
2081 spin_lock(&khugepaged_mm_lock);
2082 insert_to_mm_slots_hash(mm, mm_slot);
2083 /*
2084 * Insert just behind the scanning cursor, to let the area settle
2085 * down a little.
2086 */
2087 wakeup = list_empty(&khugepaged_scan.mm_head);
2088 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2089 spin_unlock(&khugepaged_mm_lock);
2090
2091 atomic_inc(&mm->mm_count);
2092 if (wakeup)
2093 wake_up_interruptible(&khugepaged_wait);
2094
2095 return 0;
2096}
2097
2098int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2099{
2100 unsigned long hstart, hend;
2101 if (!vma->anon_vma)
2102 /*
2103 * Not yet faulted in so we will register later in the
2104 * page fault if needed.
2105 */
2106 return 0;
78f11a25 2107 if (vma->vm_ops)
ba76149f
AA		 2108 /* khugepaged not yet working on file or special mappings */
2109 return 0;
b3b9c293 2110 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
ba76149f
AA		 2111 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2112 hend = vma->vm_end & HPAGE_PMD_MASK;
2113 if (hstart < hend)
2114 return khugepaged_enter(vma);
2115 return 0;
2116}
2117
2118void __khugepaged_exit(struct mm_struct *mm)
2119{
2120 struct mm_slot *mm_slot;
2121 int free = 0;
2122
2123 spin_lock(&khugepaged_mm_lock);
2124 mm_slot = get_mm_slot(mm);
2125 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
43b5fbbd 2126 hash_del(&mm_slot->hash);
ba76149f
AA		 2127 list_del(&mm_slot->mm_node);
2128 free = 1;
2129 }
d788e80a 2130 spin_unlock(&khugepaged_mm_lock);
ba76149f
AA		 2131
2132 if (free) {
ba76149f
AA		 2133 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2134 free_mm_slot(mm_slot);
2135 mmdrop(mm);
2136 } else if (mm_slot) {
ba76149f
AA		 2137 /*
2138 * This is required to serialize against
2139 * khugepaged_test_exit() (which is guaranteed to run
2140 * under mmap sem read mode). Stop here (after we
2141 * return all pagetables will be destroyed) until
2142 * khugepaged has finished working on the pagetables
2143 * under the mmap_sem.
2144 */
2145 down_write(&mm->mmap_sem);
2146 up_write(&mm->mmap_sem);
d788e80a 2147 }
ba76149f
AA		 2148}
2149
2150static void release_pte_page(struct page *page)
2151{
2152 /* 0 stands for page_is_file_cache(page) == false */
2153 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2154 unlock_page(page);
2155 putback_lru_page(page);
2156}
2157
2158static void release_pte_pages(pte_t *pte, pte_t *_pte)
2159{
2160 while (--_pte >= pte) {
2161 pte_t pteval = *_pte;
2162 if (!pte_none(pteval))
2163 release_pte_page(pte_page(pteval));
2164 }
2165}
2166
ba76149f
AA		 2167static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2168 unsigned long address,
2169 pte_t *pte)
2170{
2171 struct page *page;
2172 pte_t *_pte;
344aa35c 2173 int referenced = 0, none = 0;
ba76149f
AA		 2174 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2175 _pte++, address += PAGE_SIZE) {
2176 pte_t pteval = *_pte;
2177 if (pte_none(pteval)) {
2178 if (++none <= khugepaged_max_ptes_none)
2179 continue;
344aa35c 2180 else
ba76149f 2181 goto out;
ba76149f 2182 }
344aa35c 2183 if (!pte_present(pteval) || !pte_write(pteval))
ba76149f 2184 goto out;
ba76149f 2185 page = vm_normal_page(vma, address, pteval);
344aa35c 2186 if (unlikely(!page))
ba76149f 2187 goto out;
344aa35c 2188
ba76149f
AA		 2189 VM_BUG_ON(PageCompound(page));
2190 BUG_ON(!PageAnon(page));
2191 VM_BUG_ON(!PageSwapBacked(page));
2192
2193 /* cannot use mapcount: can't collapse if there's a gup pin */
344aa35c 2194 if (page_count(page) != 1)
ba76149f 2195 goto out;
ba76149f
AA		 2196 /*
2197 * We can do it before isolate_lru_page because the
2198 * page can't be freed from under us. NOTE: PG_lock
2199 * is needed to serialize against split_huge_page
2200 * when invoked from the VM.
2201 */
344aa35c 2202 if (!trylock_page(page))
ba76149f 2203 goto out;
ba76149f
AA		 2204 /*
2205 * Isolate the page to avoid collapsing an hugepage
2206 * currently in use by the VM.
2207 */
2208 if (isolate_lru_page(page)) {
2209 unlock_page(page);
ba76149f
AA		 2210 goto out;
2211 }
2212 /* 0 stands for page_is_file_cache(page) == false */
2213 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2214 VM_BUG_ON(!PageLocked(page));
2215 VM_BUG_ON(PageLRU(page));
2216
2217 /* If there is no mapped pte young don't collapse the page */
8ee53820
AA		 2218 if (pte_young(pteval) || PageReferenced(page) ||
2219 mmu_notifier_test_young(vma->vm_mm, address))
ba76149f
AA		 2220 referenced = 1;
2221 }
344aa35c
BL		 2222 if (likely(referenced))
2223 return 1;
ba76149f 2224out:
344aa35c
BL		 2225 release_pte_pages(pte, _pte);
2226 return 0;
ba76149f
AA		 2227}
2228
2229static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2230 struct vm_area_struct *vma,
2231 unsigned long address,
2232 spinlock_t *ptl)
2233{
2234 pte_t *_pte;
2235 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2236 pte_t pteval = *_pte;
2237 struct page *src_page;
2238
2239 if (pte_none(pteval)) {
2240 clear_user_highpage(page, address);
2241 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2242 } else {
2243 src_page = pte_page(pteval);
2244 copy_user_highpage(page, src_page, address, vma);
2245 VM_BUG_ON(page_mapcount(src_page) != 1);
ba76149f
AA		 2246 release_pte_page(src_page);
2247 /*
2248 * ptl mostly unnecessary, but preempt has to
2249 * be disabled to update the per-cpu stats
2250 * inside page_remove_rmap().
2251 */
2252 spin_lock(ptl);
2253 /*
2254 * paravirt calls inside pte_clear here are
2255 * superfluous.
2256 */
2257 pte_clear(vma->vm_mm, address, _pte);
2258 page_remove_rmap(src_page);
2259 spin_unlock(ptl);
2260 free_page_and_swap_cache(src_page);
2261 }
2262
2263 address += PAGE_SIZE;
2264 page++;
2265 }
2266}
2267
26234f36 2268static void khugepaged_alloc_sleep(void)
ba76149f 2269{
26234f36
XG		 2270 wait_event_freezable_timeout(khugepaged_wait, false,
2271 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2272}
ba76149f 2273
9f1b868a
BL		 2274static int khugepaged_node_load[MAX_NUMNODES];
2275
26234f36 2276#ifdef CONFIG_NUMA
9f1b868a
BL		 2277static int khugepaged_find_target_node(void)
2278{
2279 static int last_khugepaged_target_node = NUMA_NO_NODE;
2280 int nid, target_node = 0, max_value = 0;
2281
2282 /* find first node with max normal pages hit */
2283 for (nid = 0; nid < MAX_NUMNODES; nid++)
2284 if (khugepaged_node_load[nid] > max_value) {
2285 max_value = khugepaged_node_load[nid];
2286 target_node = nid;
2287 }
2288
2289 /* do some balance if several nodes have the same hit record */
2290 if (target_node <= last_khugepaged_target_node)
2291 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2292 nid++)
2293 if (max_value == khugepaged_node_load[nid]) {
2294 target_node = nid;
2295 break;
2296 }
2297
2298 last_khugepaged_target_node = target_node;
2299 return target_node;
2300}
2301
26234f36
XG		 2302static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2303{
2304 if (IS_ERR(*hpage)) {
2305 if (!*wait)
2306 return false;
2307
2308 *wait = false;
e3b4126c 2309 *hpage = NULL;
26234f36
XG		 2310 khugepaged_alloc_sleep();
2311 } else if (*hpage) {
2312 put_page(*hpage);
2313 *hpage = NULL;
2314 }
2315
2316 return true;
2317}
2318
2319static struct page
2320*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2321 struct vm_area_struct *vma, unsigned long address,
2322 int node)
2323{
0bbbc0b3 2324 VM_BUG_ON(*hpage);
ce83d217
AA		 2325 /*
2326 * Allocate the page while the vma is still valid and under
2327 * the mmap_sem read mode so there is no memory allocation
2328 * later when we take the mmap_sem in write mode. This is more
2329 * friendly behavior (OTOH it may actually hide bugs) to
2330 * filesystems in userland with daemons allocating memory in
2331 * the userland I/O paths. Allocating memory with the
2332 * mmap_sem in read mode is good idea also to allow greater
2333 * scalability.
2334 */
9f1b868a
BL		 2335 *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2336 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
692e0b35
AA		 2337 /*
2338 * After allocating the hugepage, release the mmap_sem read lock in
2339 * preparation for taking it in write mode.
2340 */
2341 up_read(&mm->mmap_sem);
26234f36 2342 if (unlikely(!*hpage)) {
81ab4201 2343 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
ce83d217 2344 *hpage = ERR_PTR(-ENOMEM);
26234f36 2345 return NULL;
ce83d217 2346 }
26234f36 2347
65b3c07b 2348 count_vm_event(THP_COLLAPSE_ALLOC);
26234f36
XG		 2349 return *hpage;
2350}
2351#else
9f1b868a
BL		 2352static int khugepaged_find_target_node(void)
2353{
2354 return 0;
2355}
2356
10dc4155
BL		 2357static inline struct page *alloc_hugepage(int defrag)
2358{
2359 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2360 HPAGE_PMD_ORDER);
2361}
2362
26234f36
XG		 2363static struct page *khugepaged_alloc_hugepage(bool *wait)
2364{
2365 struct page *hpage;
2366
2367 do {
2368 hpage = alloc_hugepage(khugepaged_defrag());
2369 if (!hpage) {
2370 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2371 if (!*wait)
2372 return NULL;
2373
2374 *wait = false;
2375 khugepaged_alloc_sleep();
2376 } else
2377 count_vm_event(THP_COLLAPSE_ALLOC);
2378 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2379
2380 return hpage;
2381}
2382
2383static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2384{
2385 if (!*hpage)
2386 *hpage = khugepaged_alloc_hugepage(wait);
2387
2388 if (unlikely(!*hpage))
2389 return false;
2390
2391 return true;
2392}
2393
2394static struct page
2395*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2396 struct vm_area_struct *vma, unsigned long address,
2397 int node)
2398{
2399 up_read(&mm->mmap_sem);
2400 VM_BUG_ON(!*hpage);
2401 return *hpage;
2402}
692e0b35
AA		 2403#endif
2404
fa475e51
BL		 2405static bool hugepage_vma_check(struct vm_area_struct *vma)
2406{
2407 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2408 (vma->vm_flags & VM_NOHUGEPAGE))
2409 return false;
2410
2411 if (!vma->anon_vma || vma->vm_ops)
2412 return false;
2413 if (is_vma_temporary_stack(vma))
2414 return false;
2415 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2416 return true;
2417}
2418
26234f36
XG		 2419static void collapse_huge_page(struct mm_struct *mm,
2420 unsigned long address,
2421 struct page **hpage,
2422 struct vm_area_struct *vma,
2423 int node)
2424{
26234f36
XG		 2425 pmd_t *pmd, _pmd;
2426 pte_t *pte;
2427 pgtable_t pgtable;
2428 struct page *new_page;
c4088ebd 2429 spinlock_t *pmd_ptl, *pte_ptl;
26234f36
XG		 2430 int isolated;
2431 unsigned long hstart, hend;
2ec74c3e
SG		 2432 unsigned long mmun_start; /* For mmu_notifiers */
2433 unsigned long mmun_end; /* For mmu_notifiers */
26234f36
XG		 2434
2435 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2436
2437 /* release the mmap_sem read lock. */
2438 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2439 if (!new_page)
2440 return;
2441
420256ef 2442 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
ce83d217 2443 return;
ba76149f
AA		 2444
2445 /*
2446 * Prevent all access to pagetables with the exception of
2447 * gup_fast later hanlded by the ptep_clear_flush and the VM
2448 * handled by the anon_vma lock + PG_lock.
2449 */
2450 down_write(&mm->mmap_sem);
2451 if (unlikely(khugepaged_test_exit(mm)))
2452 goto out;
2453
2454 vma = find_vma(mm, address);
a8f531eb
L		 2455 if (!vma)
2456 goto out;
ba76149f
AA		 2457 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2458 hend = vma->vm_end & HPAGE_PMD_MASK;
2459 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2460 goto out;
fa475e51 2461 if (!hugepage_vma_check(vma))
a7d6e4ec 2462 goto out;
6219049a
BL		 2463 pmd = mm_find_pmd(mm, address);
2464 if (!pmd)
ba76149f 2465 goto out;
6219049a 2466 if (pmd_trans_huge(*pmd))
ba76149f
AA		 2467 goto out;
2468
4fc3f1d6 2469 anon_vma_lock_write(vma->anon_vma);
ba76149f
AA		 2470
2471 pte = pte_offset_map(pmd, address);
c4088ebd 2472 pte_ptl = pte_lockptr(mm, pmd);
ba76149f 2473
2ec74c3e
SG		 2474 mmun_start = address;
2475 mmun_end = address + HPAGE_PMD_SIZE;
2476 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
c4088ebd 2477 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
ba76149f
AA		 2478 /*
2479 * After this gup_fast can't run anymore. This also removes
2480 * any huge TLB entry from the CPU so we won't allow
2481 * huge and small TLB entries for the same virtual address
2482 * to avoid the risk of CPU bugs in that area.
2483 */
2ec74c3e 2484 _pmd = pmdp_clear_flush(vma, address, pmd);
c4088ebd 2485 spin_unlock(pmd_ptl);
2ec74c3e 2486 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
ba76149f 2487
c4088ebd 2488 spin_lock(pte_ptl);
ba76149f 2489 isolated = __collapse_huge_page_isolate(vma, address, pte);
c4088ebd 2490 spin_unlock(pte_ptl);
ba76149f
AA		 2491
2492 if (unlikely(!isolated)) {
453c7192 2493 pte_unmap(pte);
c4088ebd 2494 spin_lock(pmd_ptl);
ba76149f 2495 BUG_ON(!pmd_none(*pmd));
7c342512
AK		 2496 /*
2497 * We can only use set_pmd_at when establishing
2498 * hugepmds and never for establishing regular pmds that
2499 * points to regular pagetables. Use pmd_populate for that
2500 */
2501 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
c4088ebd 2502 spin_unlock(pmd_ptl);
08b52706 2503 anon_vma_unlock_write(vma->anon_vma);
ce83d217 2504 goto out;
ba76149f
AA		 2505 }
2506
2507 /*
2508 * All pages are isolated and locked so anon_vma rmap
2509 * can't run anymore.
2510 */
08b52706 2511 anon_vma_unlock_write(vma->anon_vma);
ba76149f 2512
c4088ebd 2513 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
453c7192 2514 pte_unmap(pte);
ba76149f
AA		 2515 __SetPageUptodate(new_page);
2516 pgtable = pmd_pgtable(_pmd);
ba76149f 2517
3122359a
KS		 2518 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2519 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
ba76149f
AA		 2520
2521 /*
2522 * spin_lock() below is not the equivalent of smp_wmb(), so
2523 * this is needed to avoid the copy_huge_page writes to become
2524 * visible after the set_pmd_at() write.
2525 */
2526 smp_wmb();
2527
c4088ebd 2528 spin_lock(pmd_ptl);
ba76149f
AA		 2529 BUG_ON(!pmd_none(*pmd));
2530 page_add_new_anon_rmap(new_page, vma, address);
fce144b4 2531 pgtable_trans_huge_deposit(mm, pmd, pgtable);
ba76149f 2532 set_pmd_at(mm, address, pmd, _pmd);
b113da65 2533 update_mmu_cache_pmd(vma, address, pmd);
c4088ebd 2534 spin_unlock(pmd_ptl);
ba76149f
AA		 2535
2536 *hpage = NULL;
420256ef 2537
ba76149f 2538 khugepaged_pages_collapsed++;
ce83d217 2539out_up_write:
ba76149f 2540 up_write(&mm->mmap_sem);
0bbbc0b3
AA		 2541 return;
2542
ce83d217 2543out:
678ff896 2544 mem_cgroup_uncharge_page(new_page);
ce83d217 2545 goto out_up_write;
ba76149f
AA		 2546}
2547
2548static int khugepaged_scan_pmd(struct mm_struct *mm,
2549 struct vm_area_struct *vma,
2550 unsigned long address,
2551 struct page **hpage)
2552{
ba76149f
AA		 2553 pmd_t *pmd;
2554 pte_t *pte, *_pte;
2555 int ret = 0, referenced = 0, none = 0;
2556 struct page *page;
2557 unsigned long _address;
2558 spinlock_t *ptl;
00ef2d2f 2559 int node = NUMA_NO_NODE;
ba76149f
AA		 2560
2561 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2562
6219049a
BL		 2563 pmd = mm_find_pmd(mm, address);
2564 if (!pmd)
ba76149f 2565 goto out;
6219049a 2566 if (pmd_trans_huge(*pmd))
ba76149f
AA		 2567 goto out;
2568
9f1b868a 2569 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
ba76149f
AA		 2570 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2571 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2572 _pte++, _address += PAGE_SIZE) {
2573 pte_t pteval = *_pte;
2574 if (pte_none(pteval)) {
2575 if (++none <= khugepaged_max_ptes_none)
2576 continue;
2577 else
2578 goto out_unmap;
2579 }
2580 if (!pte_present(pteval) || !pte_write(pteval))
2581 goto out_unmap;
2582 page = vm_normal_page(vma, _address, pteval);
2583 if (unlikely(!page))
2584 goto out_unmap;
5c4b4be3 2585 /*
9f1b868a
BL		 2586 * Record which node the original page is from and save this
2587 * information to khugepaged_node_load[].
2588 * Khupaged will allocate hugepage from the node has the max
2589 * hit record.
5c4b4be3 2590 */
9f1b868a
BL		 2591 node = page_to_nid(page);
2592 khugepaged_node_load[node]++;
ba76149f
AA		 2593 VM_BUG_ON(PageCompound(page));
2594 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2595 goto out_unmap;
2596 /* cannot use mapcount: can't collapse if there's a gup pin */
2597 if (page_count(page) != 1)
2598 goto out_unmap;
8ee53820
AA		 2599 if (pte_young(pteval) || PageReferenced(page) ||
2600 mmu_notifier_test_young(vma->vm_mm, address))
ba76149f
AA		 2601 referenced = 1;
2602 }
2603 if (referenced)
2604 ret = 1;
2605out_unmap:
2606 pte_unmap_unlock(pte, ptl);
9f1b868a
BL		 2607 if (ret) {
2608 node = khugepaged_find_target_node();
ce83d217 2609 /* collapse_huge_page will return with the mmap_sem released */
5c4b4be3 2610 collapse_huge_page(mm, address, hpage, vma, node);
9f1b868a 2611 }
ba76149f
AA		 2612out:
2613 return ret;
2614}
2615
2616static void collect_mm_slot(struct mm_slot *mm_slot)
2617{
2618 struct mm_struct *mm = mm_slot->mm;
2619
b9980cdc 2620 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
ba76149f
AA		 2621
2622 if (khugepaged_test_exit(mm)) {
2623 /* free mm_slot */
43b5fbbd 2624 hash_del(&mm_slot->hash);
ba76149f
AA		 2625 list_del(&mm_slot->mm_node);
2626
2627 /*
2628 * Not strictly needed because the mm exited already.
2629 *
2630 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2631 */
2632
2633 /* khugepaged_mm_lock actually not necessary for the below */
2634 free_mm_slot(mm_slot);
2635 mmdrop(mm);
2636 }
2637}
2638
2639static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2640 struct page **hpage)
2f1da642
HS		 2641 __releases(&khugepaged_mm_lock)
2642 __acquires(&khugepaged_mm_lock)
ba76149f
AA		 2643{
2644 struct mm_slot *mm_slot;
2645 struct mm_struct *mm;
2646 struct vm_area_struct *vma;
2647 int progress = 0;
2648
2649 VM_BUG_ON(!pages);
b9980cdc 2650 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
ba76149f
AA		 2651
2652 if (khugepaged_scan.mm_slot)
2653 mm_slot = khugepaged_scan.mm_slot;
2654 else {
2655 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2656 struct mm_slot, mm_node);
2657 khugepaged_scan.address = 0;
2658 khugepaged_scan.mm_slot = mm_slot;
2659 }
2660 spin_unlock(&khugepaged_mm_lock);
2661
2662 mm = mm_slot->mm;
2663 down_read(&mm->mmap_sem);
2664 if (unlikely(khugepaged_test_exit(mm)))
2665 vma = NULL;
2666 else
2667 vma = find_vma(mm, khugepaged_scan.address);
2668
2669 progress++;
2670 for (; vma; vma = vma->vm_next) {
2671 unsigned long hstart, hend;
2672
2673 cond_resched();
2674 if (unlikely(khugepaged_test_exit(mm))) {
2675 progress++;
2676 break;
2677 }
fa475e51
BL		 2678 if (!hugepage_vma_check(vma)) {
2679skip:
ba76149f
AA		 2680 progress++;
2681 continue;
2682 }
ba76149f
AA		 2683 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2684 hend = vma->vm_end & HPAGE_PMD_MASK;
a7d6e4ec
AA		 2685 if (hstart >= hend)
2686 goto skip;
2687 if (khugepaged_scan.address > hend)
2688 goto skip;
ba76149f
AA		 2689 if (khugepaged_scan.address < hstart)
2690 khugepaged_scan.address = hstart;
a7d6e4ec 2691 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
ba76149f
AA		 2692
2693 while (khugepaged_scan.address < hend) {
2694 int ret;
2695 cond_resched();
2696 if (unlikely(khugepaged_test_exit(mm)))
2697 goto breakouterloop;
2698
2699 VM_BUG_ON(khugepaged_scan.address < hstart ||
2700 khugepaged_scan.address + HPAGE_PMD_SIZE >
2701 hend);
2702 ret = khugepaged_scan_pmd(mm, vma,
2703 khugepaged_scan.address,
2704 hpage);
2705 /* move to next address */
2706 khugepaged_scan.address += HPAGE_PMD_SIZE;
2707 progress += HPAGE_PMD_NR;
2708 if (ret)
2709 /* we released mmap_sem so break loop */
2710 goto breakouterloop_mmap_sem;
2711 if (progress >= pages)
2712 goto breakouterloop;
2713 }
2714 }
2715breakouterloop:
2716 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2717breakouterloop_mmap_sem:
2718
2719 spin_lock(&khugepaged_mm_lock);
a7d6e4ec 2720 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
ba76149f
AA		 2721 /*
2722 * Release the current mm_slot if this mm is about to die, or
2723 * if we scanned all vmas of this mm.
2724 */
2725 if (khugepaged_test_exit(mm) || !vma) {
2726 /*
2727 * Make sure that if mm_users is reaching zero while
2728 * khugepaged runs here, khugepaged_exit will find
2729 * mm_slot not pointing to the exiting mm.
2730 */
2731 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2732 khugepaged_scan.mm_slot = list_entry(
2733 mm_slot->mm_node.next,
2734 struct mm_slot, mm_node);
2735 khugepaged_scan.address = 0;
2736 } else {
2737 khugepaged_scan.mm_slot = NULL;
2738 khugepaged_full_scans++;
2739 }
2740
2741 collect_mm_slot(mm_slot);
2742 }
2743
2744 return progress;
2745}
2746
2747static int khugepaged_has_work(void)
2748{
2749 return !list_empty(&khugepaged_scan.mm_head) &&
2750 khugepaged_enabled();
2751}
2752
2753static int khugepaged_wait_event(void)
2754{
2755 return !list_empty(&khugepaged_scan.mm_head) ||
2017c0bf 2756 kthread_should_stop();
ba76149f
AA		 2757}
2758
d516904b 2759static void khugepaged_do_scan(void)
ba76149f 2760{
d516904b 2761 struct page *hpage = NULL;
ba76149f
AA		 2762 unsigned int progress = 0, pass_through_head = 0;
2763 unsigned int pages = khugepaged_pages_to_scan;
d516904b 2764 bool wait = true;
ba76149f
AA		 2765
2766 barrier(); /* write khugepaged_pages_to_scan to local stack */
2767
2768 while (progress < pages) {
26234f36 2769 if (!khugepaged_prealloc_page(&hpage, &wait))
d516904b 2770 break;
26234f36 2771
420256ef 2772 cond_resched();
ba76149f 2773
878aee7d
AA		 2774 if (unlikely(kthread_should_stop() || freezing(current)))
2775 break;
2776
ba76149f
AA		 2777 spin_lock(&khugepaged_mm_lock);
2778 if (!khugepaged_scan.mm_slot)
2779 pass_through_head++;
2780 if (khugepaged_has_work() &&
2781 pass_through_head < 2)
2782 progress += khugepaged_scan_mm_slot(pages - progress,
d516904b 2783 &hpage);
ba76149f
AA		 2784 else
2785 progress = pages;
2786 spin_unlock(&khugepaged_mm_lock);
2787 }
ba76149f 2788
d516904b
XG		 2789 if (!IS_ERR_OR_NULL(hpage))
2790 put_page(hpage);
0bbbc0b3
AA		 2791}
2792
2017c0bf
XG		 2793static void khugepaged_wait_work(void)
2794{
2795 try_to_freeze();
2796
2797 if (khugepaged_has_work()) {
2798 if (!khugepaged_scan_sleep_millisecs)
2799 return;
2800
2801 wait_event_freezable_timeout(khugepaged_wait,
2802 kthread_should_stop(),
2803 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2804 return;
2805 }
2806
2807 if (khugepaged_enabled())
2808 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2809}
2810
ba76149f
AA		 2811static int khugepaged(void *none)
2812{
2813 struct mm_slot *mm_slot;
2814
878aee7d 2815 set_freezable();
ba76149f
AA		 2816 set_user_nice(current, 19);
2817
b7231789
XG		 2818 while (!kthread_should_stop()) {
2819 khugepaged_do_scan();
2820 khugepaged_wait_work();
2821 }
ba76149f
AA		 2822
2823 spin_lock(&khugepaged_mm_lock);
2824 mm_slot = khugepaged_scan.mm_slot;
2825 khugepaged_scan.mm_slot = NULL;
2826 if (mm_slot)
2827 collect_mm_slot(mm_slot);
2828 spin_unlock(&khugepaged_mm_lock);
ba76149f
AA		 2829 return 0;
2830}
2831
c5a647d0
KS		 2832static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2833 unsigned long haddr, pmd_t *pmd)
2834{
2835 struct mm_struct *mm = vma->vm_mm;
2836 pgtable_t pgtable;
2837 pmd_t _pmd;
2838 int i;
2839
2840 pmdp_clear_flush(vma, haddr, pmd);
2841 /* leave pmd empty until pte is filled */
2842
6b0b50b0 2843 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
c5a647d0
KS		 2844 pmd_populate(mm, &_pmd, pgtable);
2845
2846 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2847 pte_t *pte, entry;
2848 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2849 entry = pte_mkspecial(entry);
2850 pte = pte_offset_map(&_pmd, haddr);
2851 VM_BUG_ON(!pte_none(*pte));
2852 set_pte_at(mm, haddr, pte, entry);
2853 pte_unmap(pte);
2854 }
2855 smp_wmb(); /* make pte visible before pmd */
2856 pmd_populate(mm, pmd, pgtable);
97ae1749 2857 put_huge_zero_page();
c5a647d0
KS		 2858}
2859
e180377f
KS		 2860void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2861 pmd_t *pmd)
71e3aac0 2862{
c4088ebd 2863 spinlock_t *ptl;
71e3aac0 2864 struct page *page;
e180377f 2865 struct mm_struct *mm = vma->vm_mm;
c5a647d0
KS		 2866 unsigned long haddr = address & HPAGE_PMD_MASK;
2867 unsigned long mmun_start; /* For mmu_notifiers */
2868 unsigned long mmun_end; /* For mmu_notifiers */
e180377f
KS		 2869
2870 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
71e3aac0 2871
c5a647d0
KS		 2872 mmun_start = haddr;
2873 mmun_end = haddr + HPAGE_PMD_SIZE;
750e8165 2874again:
c5a647d0 2875 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
c4088ebd 2876 ptl = pmd_lock(mm, pmd);
71e3aac0 2877 if (unlikely(!pmd_trans_huge(*pmd))) {
c4088ebd 2878 spin_unlock(ptl);
c5a647d0
KS		 2879 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2880 return;
2881 }
2882 if (is_huge_zero_pmd(*pmd)) {
2883 __split_huge_zero_page_pmd(vma, haddr, pmd);
c4088ebd 2884 spin_unlock(ptl);
c5a647d0 2885 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 2886 return;
2887 }
2888 page = pmd_page(*pmd);
2889 VM_BUG_ON(!page_count(page));
2890 get_page(page);
c4088ebd 2891 spin_unlock(ptl);
c5a647d0 2892 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 2893
2894 split_huge_page(page);
2895
2896 put_page(page);
750e8165
HD		 2897
2898 /*
2899 * We don't always have down_write of mmap_sem here: a racing
2900 * do_huge_pmd_wp_page() might have copied-on-write to another
2901 * huge page before our split_huge_page() got the anon_vma lock.
2902 */
2903 if (unlikely(pmd_trans_huge(*pmd)))
2904 goto again;
71e3aac0 2905}
94fcc585 2906
e180377f
KS		 2907void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2908 pmd_t *pmd)
2909{
2910 struct vm_area_struct *vma;
2911
2912 vma = find_vma(mm, address);
2913 BUG_ON(vma == NULL);
2914 split_huge_page_pmd(vma, address, pmd);
2915}
2916
94fcc585
AA		 2917static void split_huge_page_address(struct mm_struct *mm,
2918 unsigned long address)
2919{
94fcc585
AA		 2920 pmd_t *pmd;
2921
2922 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2923
6219049a
BL		 2924 pmd = mm_find_pmd(mm, address);
2925 if (!pmd)
94fcc585
AA		 2926 return;
2927 /*
2928 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2929 * materialize from under us.
2930 */
e180377f 2931 split_huge_page_pmd_mm(mm, address, pmd);
94fcc585
AA		 2932}
2933
2934void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2935 unsigned long start,
2936 unsigned long end,
2937 long adjust_next)
2938{
2939 /*
2940 * If the new start address isn't hpage aligned and it could
2941 * previously contain an hugepage: check if we need to split
2942 * an huge pmd.
2943 */
2944 if (start & ~HPAGE_PMD_MASK &&
2945 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2946 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2947 split_huge_page_address(vma->vm_mm, start);
2948
2949 /*
2950 * If the new end address isn't hpage aligned and it could
2951 * previously contain an hugepage: check if we need to split
2952 * an huge pmd.
2953 */
2954 if (end & ~HPAGE_PMD_MASK &&
2955 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2956 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2957 split_huge_page_address(vma->vm_mm, end);
2958
2959 /*
2960 * If we're also updating the vma->vm_next->vm_start, if the new
2961 * vm_next->vm_start isn't page aligned and it could previously
2962 * contain an hugepage: check if we need to split an huge pmd.
2963 */
2964 if (adjust_next > 0) {
2965 struct vm_area_struct *next = vma->vm_next;
2966 unsigned long nstart = next->vm_start;
2967 nstart += adjust_next << PAGE_SHIFT;
2968 if (nstart & ~HPAGE_PMD_MASK &&
2969 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2970 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2971 split_huge_page_address(next->vm_mm, nstart);
2972 }
2973}
Linux kernel - Deliverables
This page took 0.374101 seconds and 5 git commands to generate.