projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
mm, thp: change pmd_trans_huge_lock() to return taken lock
[linux-2.6-block.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
AA		 712 pgtable_t pgtable;
713
714 VM_BUG_ON(!PageCompound(page));
715 pgtable = pte_alloc_one(mm, haddr);
edad9d2c 716 if (unlikely(!pgtable))
71e3aac0 717 return VM_FAULT_OOM;
71e3aac0
AA		 718
719 clear_huge_page(page, haddr, HPAGE_PMD_NR);
52f37629
MK		 720 /*
721 * The memory barrier inside __SetPageUptodate makes sure that
722 * clear_huge_page writes become visible before the set_pmd_at()
723 * write.
724 */
71e3aac0
AA		 725 __SetPageUptodate(page);
726
727 spin_lock(&mm->page_table_lock);
728 if (unlikely(!pmd_none(*pmd))) {
729 spin_unlock(&mm->page_table_lock);
b9bbfbe3 730 mem_cgroup_uncharge_page(page);
71e3aac0
AA		 731 put_page(page);
732 pte_free(mm, pgtable);
733 } else {
734 pmd_t entry;
3122359a
KS		 735 entry = mk_huge_pmd(page, vma->vm_page_prot);
736 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
71e3aac0 737 page_add_new_anon_rmap(page, vma, haddr);
6b0b50b0 738 pgtable_trans_huge_deposit(mm, pmd, pgtable);
71e3aac0 739 set_pmd_at(mm, haddr, pmd, entry);
71e3aac0 740 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
e1f56c89 741 atomic_long_inc(&mm->nr_ptes);
71e3aac0
AA		 742 spin_unlock(&mm->page_table_lock);
743 }
744
aa2e878e 745 return 0;
71e3aac0
AA		 746}
747
cc5d462f 748static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
0bbbc0b3 749{
cc5d462f 750 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
0bbbc0b3
AA		 751}
752
753static inline struct page *alloc_hugepage_vma(int defrag,
754 struct vm_area_struct *vma,
cc5d462f
AK		 755 unsigned long haddr, int nd,
756 gfp_t extra_gfp)
0bbbc0b3 757{
cc5d462f 758 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
5c4b4be3 759 HPAGE_PMD_ORDER, vma, haddr, nd);
0bbbc0b3
AA		 760}
761
3ea41e62 762static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
97ae1749 763 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
5918d10a 764 struct page *zero_page)
fc9fe822
KS		 765{
766 pmd_t entry;
3ea41e62
KS		 767 if (!pmd_none(*pmd))
768 return false;
5918d10a 769 entry = mk_pmd(zero_page, vma->vm_page_prot);
fc9fe822
KS		 770 entry = pmd_wrprotect(entry);
771 entry = pmd_mkhuge(entry);
6b0b50b0 772 pgtable_trans_huge_deposit(mm, pmd, pgtable);
fc9fe822 773 set_pmd_at(mm, haddr, pmd, entry);
e1f56c89 774 atomic_long_inc(&mm->nr_ptes);
3ea41e62 775 return true;
fc9fe822
KS		 776}
777
71e3aac0
AA		 778int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
779 unsigned long address, pmd_t *pmd,
780 unsigned int flags)
781{
782 struct page *page;
783 unsigned long haddr = address & HPAGE_PMD_MASK;
71e3aac0 784
128ec037 785 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
c0292554 786 return VM_FAULT_FALLBACK;
128ec037
KS		 787 if (unlikely(anon_vma_prepare(vma)))
788 return VM_FAULT_OOM;
789 if (unlikely(khugepaged_enter(vma)))
790 return VM_FAULT_OOM;
791 if (!(flags & FAULT_FLAG_WRITE) &&
792 transparent_hugepage_use_zero_page()) {
793 pgtable_t pgtable;
794 struct page *zero_page;
795 bool set;
796 pgtable = pte_alloc_one(mm, haddr);
797 if (unlikely(!pgtable))
ba76149f 798 return VM_FAULT_OOM;
128ec037
KS		 799 zero_page = get_huge_zero_page();
800 if (unlikely(!zero_page)) {
801 pte_free(mm, pgtable);
81ab4201 802 count_vm_event(THP_FAULT_FALLBACK);
c0292554 803 return VM_FAULT_FALLBACK;
b9bbfbe3 804 }
128ec037
KS		 805 spin_lock(&mm->page_table_lock);
806 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
807 zero_page);
808 spin_unlock(&mm->page_table_lock);
809 if (!set) {
810 pte_free(mm, pgtable);
811 put_huge_zero_page();
edad9d2c 812 }
edad9d2c 813 return 0;
71e3aac0 814 }
128ec037
KS		 815 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
816 vma, haddr, numa_node_id(), 0);
817 if (unlikely(!page)) {
818 count_vm_event(THP_FAULT_FALLBACK);
c0292554 819 return VM_FAULT_FALLBACK;
128ec037 820 }
128ec037
KS		 821 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
822 put_page(page);
17766dde 823 count_vm_event(THP_FAULT_FALLBACK);
c0292554 824 return VM_FAULT_FALLBACK;
128ec037
KS		 825 }
826 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
827 mem_cgroup_uncharge_page(page);
828 put_page(page);
17766dde 829 count_vm_event(THP_FAULT_FALLBACK);
c0292554 830 return VM_FAULT_FALLBACK;
128ec037
KS		 831 }
832
17766dde 833 count_vm_event(THP_FAULT_ALLOC);
128ec037 834 return 0;
71e3aac0
AA		 835}
836
837int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
838 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
839 struct vm_area_struct *vma)
840{
841 struct page *src_page;
842 pmd_t pmd;
843 pgtable_t pgtable;
844 int ret;
845
846 ret = -ENOMEM;
847 pgtable = pte_alloc_one(dst_mm, addr);
848 if (unlikely(!pgtable))
849 goto out;
850
851 spin_lock(&dst_mm->page_table_lock);
852 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
853
854 ret = -EAGAIN;
855 pmd = *src_pmd;
856 if (unlikely(!pmd_trans_huge(pmd))) {
857 pte_free(dst_mm, pgtable);
858 goto out_unlock;
859 }
fc9fe822
KS		 860 /*
861 * mm->page_table_lock is enough to be sure that huge zero pmd is not
862 * under splitting since we don't split the page itself, only pmd to
863 * a page table.
864 */
865 if (is_huge_zero_pmd(pmd)) {
5918d10a 866 struct page *zero_page;
3ea41e62 867 bool set;
97ae1749
KS		 868 /*
869 * get_huge_zero_page() will never allocate a new page here,
870 * since we already have a zero page to copy. It just takes a
871 * reference.
872 */
5918d10a 873 zero_page = get_huge_zero_page();
3ea41e62 874 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
5918d10a 875 zero_page);
3ea41e62 876 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
fc9fe822
KS		 877 ret = 0;
878 goto out_unlock;
879 }
71e3aac0
AA		 880 if (unlikely(pmd_trans_splitting(pmd))) {
881 /* split huge page running from under us */
882 spin_unlock(&src_mm->page_table_lock);
883 spin_unlock(&dst_mm->page_table_lock);
884 pte_free(dst_mm, pgtable);
885
886 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
887 goto out;
888 }
889 src_page = pmd_page(pmd);
890 VM_BUG_ON(!PageHead(src_page));
891 get_page(src_page);
892 page_dup_rmap(src_page);
893 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
894
895 pmdp_set_wrprotect(src_mm, addr, src_pmd);
896 pmd = pmd_mkold(pmd_wrprotect(pmd));
6b0b50b0 897 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
71e3aac0 898 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
e1f56c89 899 atomic_long_inc(&dst_mm->nr_ptes);
71e3aac0
AA		 900
901 ret = 0;
902out_unlock:
903 spin_unlock(&src_mm->page_table_lock);
904 spin_unlock(&dst_mm->page_table_lock);
905out:
906 return ret;
907}
908
a1dd450b
WD		 909void huge_pmd_set_accessed(struct mm_struct *mm,
910 struct vm_area_struct *vma,
911 unsigned long address,
912 pmd_t *pmd, pmd_t orig_pmd,
913 int dirty)
914{
915 pmd_t entry;
916 unsigned long haddr;
917
918 spin_lock(&mm->page_table_lock);
919 if (unlikely(!pmd_same(*pmd, orig_pmd)))
920 goto unlock;
921
922 entry = pmd_mkyoung(orig_pmd);
923 haddr = address & HPAGE_PMD_MASK;
924 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
925 update_mmu_cache_pmd(vma, address, pmd);
926
927unlock:
928 spin_unlock(&mm->page_table_lock);
929}
930
93b4796d
KS		 931static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
932 struct vm_area_struct *vma, unsigned long address,
3ea41e62 933 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
93b4796d
KS		 934{
935 pgtable_t pgtable;
936 pmd_t _pmd;
937 struct page *page;
938 int i, ret = 0;
939 unsigned long mmun_start; /* For mmu_notifiers */
940 unsigned long mmun_end; /* For mmu_notifiers */
941
942 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
943 if (!page) {
944 ret |= VM_FAULT_OOM;
945 goto out;
946 }
947
948 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
949 put_page(page);
950 ret |= VM_FAULT_OOM;
951 goto out;
952 }
953
954 clear_user_highpage(page, address);
955 __SetPageUptodate(page);
956
957 mmun_start = haddr;
958 mmun_end = haddr + HPAGE_PMD_SIZE;
959 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
960
961 spin_lock(&mm->page_table_lock);
3ea41e62
KS		 962 if (unlikely(!pmd_same(*pmd, orig_pmd)))
963 goto out_free_page;
964
93b4796d
KS		 965 pmdp_clear_flush(vma, haddr, pmd);
966 /* leave pmd empty until pte is filled */
967
6b0b50b0 968 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
93b4796d
KS		 969 pmd_populate(mm, &_pmd, pgtable);
970
971 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
972 pte_t *pte, entry;
973 if (haddr == (address & PAGE_MASK)) {
974 entry = mk_pte(page, vma->vm_page_prot);
975 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
976 page_add_new_anon_rmap(page, vma, haddr);
977 } else {
978 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
979 entry = pte_mkspecial(entry);
980 }
981 pte = pte_offset_map(&_pmd, haddr);
982 VM_BUG_ON(!pte_none(*pte));
983 set_pte_at(mm, haddr, pte, entry);
984 pte_unmap(pte);
985 }
986 smp_wmb(); /* make pte visible before pmd */
987 pmd_populate(mm, pmd, pgtable);
988 spin_unlock(&mm->page_table_lock);
97ae1749 989 put_huge_zero_page();
93b4796d
KS		 990 inc_mm_counter(mm, MM_ANONPAGES);
991
992 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
993
994 ret |= VM_FAULT_WRITE;
995out:
996 return ret;
3ea41e62
KS		 997out_free_page:
998 spin_unlock(&mm->page_table_lock);
999 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1000 mem_cgroup_uncharge_page(page);
1001 put_page(page);
1002 goto out;
93b4796d
KS		 1003}
1004
71e3aac0
AA		 1005static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1006 struct vm_area_struct *vma,
1007 unsigned long address,
1008 pmd_t *pmd, pmd_t orig_pmd,
1009 struct page *page,
1010 unsigned long haddr)
1011{
1012 pgtable_t pgtable;
1013 pmd_t _pmd;
1014 int ret = 0, i;
1015 struct page **pages;
2ec74c3e
SG		 1016 unsigned long mmun_start; /* For mmu_notifiers */
1017 unsigned long mmun_end; /* For mmu_notifiers */
71e3aac0
AA		 1018
1019 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1020 GFP_KERNEL);
1021 if (unlikely(!pages)) {
1022 ret |= VM_FAULT_OOM;
1023 goto out;
1024 }
1025
1026 for (i = 0; i < HPAGE_PMD_NR; i++) {
cc5d462f
AK		 1027 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1028 __GFP_OTHER_NODE,
19ee151e 1029 vma, address, page_to_nid(page));
b9bbfbe3
AA		 1030 if (unlikely(!pages[i] ||
1031 mem_cgroup_newpage_charge(pages[i], mm,
1032 GFP_KERNEL))) {
1033 if (pages[i])
71e3aac0 1034 put_page(pages[i]);
b9bbfbe3
AA		 1035 mem_cgroup_uncharge_start();
1036 while (--i >= 0) {
1037 mem_cgroup_uncharge_page(pages[i]);
1038 put_page(pages[i]);
1039 }
1040 mem_cgroup_uncharge_end();
71e3aac0
AA		 1041 kfree(pages);
1042 ret |= VM_FAULT_OOM;
1043 goto out;
1044 }
1045 }
1046
1047 for (i = 0; i < HPAGE_PMD_NR; i++) {
1048 copy_user_highpage(pages[i], page + i,
0089e485 1049 haddr + PAGE_SIZE * i, vma);
71e3aac0
AA		 1050 __SetPageUptodate(pages[i]);
1051 cond_resched();
1052 }
1053
2ec74c3e
SG		 1054 mmun_start = haddr;
1055 mmun_end = haddr + HPAGE_PMD_SIZE;
1056 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1057
71e3aac0
AA		 1058 spin_lock(&mm->page_table_lock);
1059 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1060 goto out_free_pages;
1061 VM_BUG_ON(!PageHead(page));
1062
2ec74c3e 1063 pmdp_clear_flush(vma, haddr, pmd);
71e3aac0
AA		 1064 /* leave pmd empty until pte is filled */
1065
6b0b50b0 1066 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
71e3aac0
AA		 1067 pmd_populate(mm, &_pmd, pgtable);
1068
1069 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1070 pte_t *pte, entry;
1071 entry = mk_pte(pages[i], vma->vm_page_prot);
1072 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1073 page_add_new_anon_rmap(pages[i], vma, haddr);
1074 pte = pte_offset_map(&_pmd, haddr);
1075 VM_BUG_ON(!pte_none(*pte));
1076 set_pte_at(mm, haddr, pte, entry);
1077 pte_unmap(pte);
1078 }
1079 kfree(pages);
1080
71e3aac0
AA		 1081 smp_wmb(); /* make pte visible before pmd */
1082 pmd_populate(mm, pmd, pgtable);
1083 page_remove_rmap(page);
1084 spin_unlock(&mm->page_table_lock);
1085
2ec74c3e
SG		 1086 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1087
71e3aac0
AA		 1088 ret |= VM_FAULT_WRITE;
1089 put_page(page);
1090
1091out:
1092 return ret;
1093
1094out_free_pages:
1095 spin_unlock(&mm->page_table_lock);
2ec74c3e 1096 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
b9bbfbe3
AA		 1097 mem_cgroup_uncharge_start();
1098 for (i = 0; i < HPAGE_PMD_NR; i++) {
1099 mem_cgroup_uncharge_page(pages[i]);
71e3aac0 1100 put_page(pages[i]);
b9bbfbe3
AA		 1101 }
1102 mem_cgroup_uncharge_end();
71e3aac0
AA		 1103 kfree(pages);
1104 goto out;
1105}
1106
1107int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1108 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1109{
1110 int ret = 0;
93b4796d 1111 struct page *page = NULL, *new_page;
71e3aac0 1112 unsigned long haddr;
2ec74c3e
SG		 1113 unsigned long mmun_start; /* For mmu_notifiers */
1114 unsigned long mmun_end; /* For mmu_notifiers */
71e3aac0
AA		 1115
1116 VM_BUG_ON(!vma->anon_vma);
93b4796d
KS		 1117 haddr = address & HPAGE_PMD_MASK;
1118 if (is_huge_zero_pmd(orig_pmd))
1119 goto alloc;
71e3aac0
AA		 1120 spin_lock(&mm->page_table_lock);
1121 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1122 goto out_unlock;
1123
1124 page = pmd_page(orig_pmd);
1125 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
71e3aac0
AA		 1126 if (page_mapcount(page) == 1) {
1127 pmd_t entry;
1128 entry = pmd_mkyoung(orig_pmd);
1129 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1130 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
b113da65 1131 update_mmu_cache_pmd(vma, address, pmd);
71e3aac0
AA		 1132 ret |= VM_FAULT_WRITE;
1133 goto out_unlock;
1134 }
1135 get_page(page);
1136 spin_unlock(&mm->page_table_lock);
93b4796d 1137alloc:
71e3aac0
AA		 1138 if (transparent_hugepage_enabled(vma) &&
1139 !transparent_hugepage_debug_cow())
0bbbc0b3 1140 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
cc5d462f 1141 vma, haddr, numa_node_id(), 0);
71e3aac0
AA		 1142 else
1143 new_page = NULL;
1144
1145 if (unlikely(!new_page)) {
93b4796d
KS		 1146 if (is_huge_zero_pmd(orig_pmd)) {
1147 ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
3ea41e62 1148 address, pmd, orig_pmd, haddr);
93b4796d
KS		 1149 } else {
1150 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1151 pmd, orig_pmd, page, haddr);
1152 if (ret & VM_FAULT_OOM)
1153 split_huge_page(page);
1154 put_page(page);
1155 }
17766dde 1156 count_vm_event(THP_FAULT_FALLBACK);
71e3aac0
AA		 1157 goto out;
1158 }
1159
b9bbfbe3
AA		 1160 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1161 put_page(new_page);
93b4796d
KS		 1162 if (page) {
1163 split_huge_page(page);
1164 put_page(page);
1165 }
17766dde 1166 count_vm_event(THP_FAULT_FALLBACK);
b9bbfbe3
AA		 1167 ret |= VM_FAULT_OOM;
1168 goto out;
1169 }
1170
17766dde
DR		 1171 count_vm_event(THP_FAULT_ALLOC);
1172
93b4796d
KS		 1173 if (is_huge_zero_pmd(orig_pmd))
1174 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1175 else
1176 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
71e3aac0
AA		 1177 __SetPageUptodate(new_page);
1178
2ec74c3e
SG		 1179 mmun_start = haddr;
1180 mmun_end = haddr + HPAGE_PMD_SIZE;
1181 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1182
71e3aac0 1183 spin_lock(&mm->page_table_lock);
93b4796d
KS		 1184 if (page)
1185 put_page(page);
b9bbfbe3 1186 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
6f60b69d 1187 spin_unlock(&mm->page_table_lock);
b9bbfbe3 1188 mem_cgroup_uncharge_page(new_page);
71e3aac0 1189 put_page(new_page);
2ec74c3e 1190 goto out_mn;
b9bbfbe3 1191 } else {
71e3aac0 1192 pmd_t entry;
3122359a
KS		 1193 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1194 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2ec74c3e 1195 pmdp_clear_flush(vma, haddr, pmd);
71e3aac0
AA		 1196 page_add_new_anon_rmap(new_page, vma, haddr);
1197 set_pmd_at(mm, haddr, pmd, entry);
b113da65 1198 update_mmu_cache_pmd(vma, address, pmd);
97ae1749 1199 if (is_huge_zero_pmd(orig_pmd)) {
93b4796d 1200 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
97ae1749
KS		 1201 put_huge_zero_page();
1202 } else {
93b4796d
KS		 1203 VM_BUG_ON(!PageHead(page));
1204 page_remove_rmap(page);
1205 put_page(page);
1206 }
71e3aac0
AA		 1207 ret |= VM_FAULT_WRITE;
1208 }
71e3aac0 1209 spin_unlock(&mm->page_table_lock);
2ec74c3e
SG		 1210out_mn:
1211 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 1212out:
1213 return ret;
2ec74c3e
SG		 1214out_unlock:
1215 spin_unlock(&mm->page_table_lock);
1216 return ret;
71e3aac0
AA		 1217}
1218
b676b293 1219struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
71e3aac0
AA		 1220 unsigned long addr,
1221 pmd_t *pmd,
1222 unsigned int flags)
1223{
b676b293 1224 struct mm_struct *mm = vma->vm_mm;
71e3aac0
AA		 1225 struct page *page = NULL;
1226
1227 assert_spin_locked(&mm->page_table_lock);
1228
1229 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1230 goto out;
1231
85facf25
KS		 1232 /* Avoid dumping huge zero page */
1233 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1234 return ERR_PTR(-EFAULT);
1235
71e3aac0
AA		 1236 page = pmd_page(*pmd);
1237 VM_BUG_ON(!PageHead(page));
1238 if (flags & FOLL_TOUCH) {
1239 pmd_t _pmd;
1240 /*
1241 * We should set the dirty bit only for FOLL_WRITE but
1242 * for now the dirty bit in the pmd is meaningless.
1243 * And if the dirty bit will become meaningful and
1244 * we'll only set it with FOLL_WRITE, an atomic
1245 * set_bit will be required on the pmd to set the
1246 * young bit, instead of the current set_pmd_at.
1247 */
1248 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
8663890a
AK		 1249 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1250 pmd, _pmd, 1))
1251 update_mmu_cache_pmd(vma, addr, pmd);
71e3aac0 1252 }
b676b293
DR		 1253 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1254 if (page->mapping && trylock_page(page)) {
1255 lru_add_drain();
1256 if (page->mapping)
1257 mlock_vma_page(page);
1258 unlock_page(page);
1259 }
1260 }
71e3aac0
AA		 1261 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1262 VM_BUG_ON(!PageCompound(page));
1263 if (flags & FOLL_GET)
70b50f94 1264 get_page_foll(page);
71e3aac0
AA		 1265
1266out:
1267 return page;
1268}
1269
d10e63f2 1270/* NUMA hinting page fault entry point for trans huge pmds */
4daae3b4
MG		 1271int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1272 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
d10e63f2 1273{
b8916634 1274 struct anon_vma *anon_vma = NULL;
b32967ff 1275 struct page *page;
d10e63f2 1276 unsigned long haddr = addr & HPAGE_PMD_MASK;
8191acbd 1277 int page_nid = -1, this_nid = numa_node_id();
90572890 1278 int target_nid, last_cpupid = -1;
8191acbd
MG		 1279 bool page_locked;
1280 bool migrated = false;
6688cc05 1281 int flags = 0;
d10e63f2
MG		 1282
1283 spin_lock(&mm->page_table_lock);
1284 if (unlikely(!pmd_same(pmd, *pmdp)))
1285 goto out_unlock;
1286
1287 page = pmd_page(pmd);
a1a46184 1288 BUG_ON(is_huge_zero_page(page));
8191acbd 1289 page_nid = page_to_nid(page);
90572890 1290 last_cpupid = page_cpupid_last(page);
03c5a6e1 1291 count_vm_numa_event(NUMA_HINT_FAULTS);
04bb2f94 1292 if (page_nid == this_nid) {
03c5a6e1 1293 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
04bb2f94
RR		 1294 flags |= TNF_FAULT_LOCAL;
1295 }
4daae3b4 1296
6688cc05
PZ		 1297 /*
1298 * Avoid grouping on DSO/COW pages in specific and RO pages
1299 * in general, RO pages shouldn't hurt as much anyway since
1300 * they can be in shared cache state.
1301 */
1302 if (!pmd_write(pmd))
1303 flags |= TNF_NO_GROUP;
1304
ff9042b1
MG		 1305 /*
1306 * Acquire the page lock to serialise THP migrations but avoid dropping
1307 * page_table_lock if at all possible
1308 */
b8916634
MG		 1309 page_locked = trylock_page(page);
1310 target_nid = mpol_misplaced(page, vma, haddr);
1311 if (target_nid == -1) {
1312 /* If the page was locked, there are no parallel migrations */
a54a407f 1313 if (page_locked)
b8916634 1314 goto clear_pmdnuma;
4daae3b4 1315
a54a407f
MG		 1316 /*
1317 * Otherwise wait for potential migrations and retry. We do
1318 * relock and check_same as the page may no longer be mapped.
1319 * As the fault is being retried, do not account for it.
1320 */
b8916634
MG		 1321 spin_unlock(&mm->page_table_lock);
1322 wait_on_page_locked(page);
a54a407f 1323 page_nid = -1;
b8916634
MG		 1324 goto out;
1325 }
1326
1327 /* Page is misplaced, serialise migrations and parallel THP splits */
1328 get_page(page);
b32967ff 1329 spin_unlock(&mm->page_table_lock);
a54a407f 1330 if (!page_locked)
b8916634 1331 lock_page(page);
b8916634 1332 anon_vma = page_lock_anon_vma_read(page);
4daae3b4 1333
c69307d5 1334 /* Confirm the PMD did not change while page_table_lock was released */
4daae3b4 1335 spin_lock(&mm->page_table_lock);
b32967ff
MG		 1336 if (unlikely(!pmd_same(pmd, *pmdp))) {
1337 unlock_page(page);
1338 put_page(page);
a54a407f 1339 page_nid = -1;
4daae3b4 1340 goto out_unlock;
b32967ff 1341 }
ff9042b1 1342
a54a407f
MG		 1343 /*
1344 * Migrate the THP to the requested node, returns with page unlocked
1345 * and pmd_numa cleared.
1346 */
ff9042b1 1347 spin_unlock(&mm->page_table_lock);
b32967ff 1348 migrated = migrate_misplaced_transhuge_page(mm, vma,
340ef390 1349 pmdp, pmd, addr, page, target_nid);
6688cc05
PZ		 1350 if (migrated) {
1351 flags |= TNF_MIGRATED;
8191acbd 1352 page_nid = target_nid;
6688cc05 1353 }
b32967ff 1354
8191acbd 1355 goto out;
b32967ff 1356clear_pmdnuma:
a54a407f 1357 BUG_ON(!PageLocked(page));
d10e63f2
MG		 1358 pmd = pmd_mknonnuma(pmd);
1359 set_pmd_at(mm, haddr, pmdp, pmd);
1360 VM_BUG_ON(pmd_numa(*pmdp));
1361 update_mmu_cache_pmd(vma, addr, pmdp);
a54a407f 1362 unlock_page(page);
d10e63f2
MG		 1363out_unlock:
1364 spin_unlock(&mm->page_table_lock);
b8916634
MG		 1365
1366out:
1367 if (anon_vma)
1368 page_unlock_anon_vma_read(anon_vma);
1369
8191acbd 1370 if (page_nid != -1)
6688cc05 1371 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
8191acbd 1372
d10e63f2
MG		 1373 return 0;
1374}
1375
71e3aac0 1376int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
f21760b1 1377 pmd_t *pmd, unsigned long addr)
71e3aac0 1378{
bf929152 1379 spinlock_t *ptl;
71e3aac0
AA		 1380 int ret = 0;
1381
bf929152 1382 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
025c5b24
NH		 1383 struct page *page;
1384 pgtable_t pgtable;
f5c8ad47 1385 pmd_t orig_pmd;
a6bf2bb0
AK		 1386 /*
1387 * For architectures like ppc64 we look at deposited pgtable
1388 * when calling pmdp_get_and_clear. So do the
1389 * pgtable_trans_huge_withdraw after finishing pmdp related
1390 * operations.
1391 */
f5c8ad47 1392 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
025c5b24 1393 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
a6bf2bb0 1394 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
479f0abb 1395 if (is_huge_zero_pmd(orig_pmd)) {
e1f56c89 1396 atomic_long_dec(&tlb->mm->nr_ptes);
bf929152 1397 spin_unlock(ptl);
97ae1749 1398 put_huge_zero_page();
479f0abb
KS		 1399 } else {
1400 page = pmd_page(orig_pmd);
1401 page_remove_rmap(page);
1402 VM_BUG_ON(page_mapcount(page) < 0);
1403 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1404 VM_BUG_ON(!PageHead(page));
e1f56c89 1405 atomic_long_dec(&tlb->mm->nr_ptes);
bf929152 1406 spin_unlock(ptl);
479f0abb
KS		 1407 tlb_remove_page(tlb, page);
1408 }
025c5b24
NH		 1409 pte_free(tlb->mm, pgtable);
1410 ret = 1;
1411 }
71e3aac0
AA		 1412 return ret;
1413}
1414
0ca1634d
JW		 1415int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1416 unsigned long addr, unsigned long end,
1417 unsigned char *vec)
1418{
bf929152 1419 spinlock_t *ptl;
0ca1634d
JW		 1420 int ret = 0;
1421
bf929152 1422 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
025c5b24
NH		 1423 /*
1424 * All logical pages in the range are present
1425 * if backed by a huge page.
1426 */
bf929152 1427 spin_unlock(ptl);
025c5b24
NH		 1428 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1429 ret = 1;
1430 }
0ca1634d
JW		 1431
1432 return ret;
1433}
1434
37a1c49a
AA		 1435int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1436 unsigned long old_addr,
1437 unsigned long new_addr, unsigned long old_end,
1438 pmd_t *old_pmd, pmd_t *new_pmd)
1439{
bf929152 1440 spinlock_t *old_ptl, *new_ptl;
37a1c49a
AA		 1441 int ret = 0;
1442 pmd_t pmd;
1443
1444 struct mm_struct *mm = vma->vm_mm;
1445
1446 if ((old_addr & ~HPAGE_PMD_MASK) ||
1447 (new_addr & ~HPAGE_PMD_MASK) ||
1448 old_end - old_addr < HPAGE_PMD_SIZE ||
1449 (new_vma->vm_flags & VM_NOHUGEPAGE))
1450 goto out;
1451
1452 /*
1453 * The destination pmd shouldn't be established, free_pgtables()
1454 * should have release it.
1455 */
1456 if (WARN_ON(!pmd_none(*new_pmd))) {
1457 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1458 goto out;
1459 }
1460
bf929152
KS		 1461 /*
1462 * We don't have to worry about the ordering of src and dst
1463 * ptlocks because exclusive mmap_sem prevents deadlock.
1464 */
1465 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
025c5b24 1466 if (ret == 1) {
bf929152
KS		 1467 new_ptl = pmd_lockptr(mm, new_pmd);
1468 if (new_ptl != old_ptl)
1469 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
025c5b24
NH		 1470 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1471 VM_BUG_ON(!pmd_none(*new_pmd));
0f8975ec 1472 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
bf929152
KS		 1473 if (new_ptl != old_ptl)
1474 spin_unlock(new_ptl);
1475 spin_unlock(old_ptl);
37a1c49a
AA		 1476 }
1477out:
1478 return ret;
1479}
1480
f123d74a
MG		 1481/*
1482 * Returns
1483 * - 0 if PMD could not be locked
1484 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1485 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1486 */
cd7548ab 1487int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
4b10e7d5 1488 unsigned long addr, pgprot_t newprot, int prot_numa)
cd7548ab
JW		 1489{
1490 struct mm_struct *mm = vma->vm_mm;
bf929152 1491 spinlock_t *ptl;
cd7548ab
JW		 1492 int ret = 0;
1493
bf929152 1494 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
025c5b24 1495 pmd_t entry;
f123d74a 1496 ret = 1;
a4f1de17 1497 if (!prot_numa) {
f123d74a 1498 entry = pmdp_get_and_clear(mm, addr, pmd);
4b10e7d5 1499 entry = pmd_modify(entry, newprot);
f123d74a 1500 ret = HPAGE_PMD_NR;
a4f1de17
HD		 1501 BUG_ON(pmd_write(entry));
1502 } else {
4b10e7d5
MG		 1503 struct page *page = pmd_page(*pmd);
1504
a1a46184 1505 /*
1bc115d8
MG		 1506 * Do not trap faults against the zero page. The
1507 * read-only data is likely to be read-cached on the
1508 * local CPU cache and it is less useful to know about
1509 * local vs remote hits on the zero page.
a1a46184 1510 */
1bc115d8 1511 if (!is_huge_zero_page(page) &&
4b10e7d5 1512 !pmd_numa(*pmd)) {
f123d74a 1513 entry = pmdp_get_and_clear(mm, addr, pmd);
4b10e7d5 1514 entry = pmd_mknuma(entry);
f123d74a 1515 ret = HPAGE_PMD_NR;
4b10e7d5
MG		 1516 }
1517 }
f123d74a
MG		 1518
1519 /* Set PMD if cleared earlier */
1520 if (ret == HPAGE_PMD_NR)
1521 set_pmd_at(mm, addr, pmd, entry);
1522
bf929152 1523 spin_unlock(ptl);
025c5b24
NH		 1524 }
1525
1526 return ret;
1527}
1528
1529/*
1530 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1531 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1532 *
1533 * Note that if it returns 1, this routine returns without unlocking page
1534 * table locks. So callers must unlock them.
1535 */
bf929152
KS		 1536int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1537 spinlock_t **ptl)
025c5b24 1538{
bf929152 1539 *ptl = pmd_lock(vma->vm_mm, pmd);
cd7548ab
JW		 1540 if (likely(pmd_trans_huge(*pmd))) {
1541 if (unlikely(pmd_trans_splitting(*pmd))) {
bf929152 1542 spin_unlock(*ptl);
cd7548ab 1543 wait_split_huge_page(vma->anon_vma, pmd);
025c5b24 1544 return -1;
cd7548ab 1545 } else {
025c5b24
NH		 1546 /* Thp mapped by 'pmd' is stable, so we can
1547 * handle it as it is. */
1548 return 1;
cd7548ab 1549 }
025c5b24 1550 }
bf929152 1551 spin_unlock(*ptl);
025c5b24 1552 return 0;
cd7548ab
JW		 1553}
1554
71e3aac0
AA		 1555pmd_t *page_check_address_pmd(struct page *page,
1556 struct mm_struct *mm,
1557 unsigned long address,
1558 enum page_check_address_pmd_flag flag)
1559{
71e3aac0
AA		 1560 pmd_t *pmd, *ret = NULL;
1561
1562 if (address & ~HPAGE_PMD_MASK)
1563 goto out;
1564
6219049a
BL		 1565 pmd = mm_find_pmd(mm, address);
1566 if (!pmd)
71e3aac0 1567 goto out;
71e3aac0
AA		 1568 if (pmd_none(*pmd))
1569 goto out;
1570 if (pmd_page(*pmd) != page)
1571 goto out;
94fcc585
AA		 1572 /*
1573 * split_vma() may create temporary aliased mappings. There is
1574 * no risk as long as all huge pmd are found and have their
1575 * splitting bit set before __split_huge_page_refcount
1576 * runs. Finding the same huge pmd more than once during the
1577 * same rmap walk is not a problem.
1578 */
1579 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1580 pmd_trans_splitting(*pmd))
1581 goto out;
71e3aac0
AA		 1582 if (pmd_trans_huge(*pmd)) {
1583 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1584 !pmd_trans_splitting(*pmd));
1585 ret = pmd;
1586 }
1587out:
1588 return ret;
1589}
1590
1591static int __split_huge_page_splitting(struct page *page,
1592 struct vm_area_struct *vma,
1593 unsigned long address)
1594{
1595 struct mm_struct *mm = vma->vm_mm;
1596 pmd_t *pmd;
1597 int ret = 0;
2ec74c3e
SG		 1598 /* For mmu_notifiers */
1599 const unsigned long mmun_start = address;
1600 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
71e3aac0 1601
2ec74c3e 1602 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
71e3aac0
AA		 1603 spin_lock(&mm->page_table_lock);
1604 pmd = page_check_address_pmd(page, mm, address,
1605 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1606 if (pmd) {
1607 /*
1608 * We can't temporarily set the pmd to null in order
1609 * to split it, the pmd must remain marked huge at all
1610 * times or the VM won't take the pmd_trans_huge paths
5a505085 1611 * and it won't wait on the anon_vma->root->rwsem to
71e3aac0
AA		 1612 * serialize against split_huge_page*.
1613 */
2ec74c3e 1614 pmdp_splitting_flush(vma, address, pmd);
71e3aac0
AA		 1615 ret = 1;
1616 }
1617 spin_unlock(&mm->page_table_lock);
2ec74c3e 1618 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 1619
1620 return ret;
1621}
1622
5bc7b8ac
SL		 1623static void __split_huge_page_refcount(struct page *page,
1624 struct list_head *list)
71e3aac0
AA		 1625{
1626 int i;
71e3aac0 1627 struct zone *zone = page_zone(page);
fa9add64 1628 struct lruvec *lruvec;
70b50f94 1629 int tail_count = 0;
71e3aac0
AA		 1630
1631 /* prevent PageLRU to go away from under us, and freeze lru stats */
1632 spin_lock_irq(&zone->lru_lock);
fa9add64
HD		 1633 lruvec = mem_cgroup_page_lruvec(page, zone);
1634
71e3aac0 1635 compound_lock(page);
e94c8a9c
KH		 1636 /* complete memcg works before add pages to LRU */
1637 mem_cgroup_split_huge_fixup(page);
71e3aac0 1638
45676885 1639 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
71e3aac0
AA		 1640 struct page *page_tail = page + i;
1641
70b50f94
AA		 1642 /* tail_page->_mapcount cannot change */
1643 BUG_ON(page_mapcount(page_tail) < 0);
1644 tail_count += page_mapcount(page_tail);
1645 /* check for overflow */
1646 BUG_ON(tail_count < 0);
1647 BUG_ON(atomic_read(&page_tail->_count) != 0);
1648 /*
1649 * tail_page->_count is zero and not changing from
1650 * under us. But get_page_unless_zero() may be running
1651 * from under us on the tail_page. If we used
1652 * atomic_set() below instead of atomic_add(), we
1653 * would then run atomic_set() concurrently with
1654 * get_page_unless_zero(), and atomic_set() is
1655 * implemented in C not using locked ops. spin_unlock
1656 * on x86 sometime uses locked ops because of PPro
1657 * errata 66, 92, so unless somebody can guarantee
1658 * atomic_set() here would be safe on all archs (and
1659 * not only on x86), it's safer to use atomic_add().
1660 */
1661 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1662 &page_tail->_count);
71e3aac0
AA		 1663
1664 /* after clearing PageTail the gup refcount can be released */
1665 smp_mb();
1666
a6d30ddd
JD		 1667 /*
1668 * retain hwpoison flag of the poisoned tail page:
1669 * fix for the unsuitable process killed on Guest Machine(KVM)
1670 * by the memory-failure.
1671 */
1672 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
71e3aac0
AA		 1673 page_tail->flags |= (page->flags &
1674 ((1L << PG_referenced) |
1675 (1L << PG_swapbacked) |
1676 (1L << PG_mlocked) |
e180cf80
KS		 1677 (1L << PG_uptodate) |
1678 (1L << PG_active) |
1679 (1L << PG_unevictable)));
71e3aac0
AA		 1680 page_tail->flags |= (1L << PG_dirty);
1681
70b50f94 1682 /* clear PageTail before overwriting first_page */
71e3aac0
AA		 1683 smp_wmb();
1684
1685 /*
1686 * __split_huge_page_splitting() already set the
1687 * splitting bit in all pmd that could map this
1688 * hugepage, that will ensure no CPU can alter the
1689 * mapcount on the head page. The mapcount is only
1690 * accounted in the head page and it has to be
1691 * transferred to all tail pages in the below code. So
1692 * for this code to be safe, the split the mapcount
1693 * can't change. But that doesn't mean userland can't
1694 * keep changing and reading the page contents while
1695 * we transfer the mapcount, so the pmd splitting
1696 * status is achieved setting a reserved bit in the
1697 * pmd, not by clearing the present bit.
1698 */
71e3aac0
AA		 1699 page_tail->_mapcount = page->_mapcount;
1700
1701 BUG_ON(page_tail->mapping);
1702 page_tail->mapping = page->mapping;
1703
45676885 1704 page_tail->index = page->index + i;
90572890 1705 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
71e3aac0
AA		 1706
1707 BUG_ON(!PageAnon(page_tail));
1708 BUG_ON(!PageUptodate(page_tail));
1709 BUG_ON(!PageDirty(page_tail));
1710 BUG_ON(!PageSwapBacked(page_tail));
1711
5bc7b8ac 1712 lru_add_page_tail(page, page_tail, lruvec, list);
71e3aac0 1713 }
70b50f94
AA		 1714 atomic_sub(tail_count, &page->_count);
1715 BUG_ON(atomic_read(&page->_count) <= 0);
71e3aac0 1716
fa9add64 1717 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
79134171 1718
71e3aac0
AA		 1719 ClearPageCompound(page);
1720 compound_unlock(page);
1721 spin_unlock_irq(&zone->lru_lock);
1722
1723 for (i = 1; i < HPAGE_PMD_NR; i++) {
1724 struct page *page_tail = page + i;
1725 BUG_ON(page_count(page_tail) <= 0);
1726 /*
1727 * Tail pages may be freed if there wasn't any mapping
1728 * like if add_to_swap() is running on a lru page that
1729 * had its mapping zapped. And freeing these pages
1730 * requires taking the lru_lock so we do the put_page
1731 * of the tail pages after the split is complete.
1732 */
1733 put_page(page_tail);
1734 }
1735
1736 /*
1737 * Only the head page (now become a regular page) is required
1738 * to be pinned by the caller.
1739 */
1740 BUG_ON(page_count(page) <= 0);
1741}
1742
1743static int __split_huge_page_map(struct page *page,
1744 struct vm_area_struct *vma,
1745 unsigned long address)
1746{
1747 struct mm_struct *mm = vma->vm_mm;
1748 pmd_t *pmd, _pmd;
1749 int ret = 0, i;
1750 pgtable_t pgtable;
1751 unsigned long haddr;
1752
1753 spin_lock(&mm->page_table_lock);
1754 pmd = page_check_address_pmd(page, mm, address,
1755 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1756 if (pmd) {
6b0b50b0 1757 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
71e3aac0
AA		 1758 pmd_populate(mm, &_pmd, pgtable);
1759
e3ebcf64
GS		 1760 haddr = address;
1761 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
71e3aac0
AA		 1762 pte_t *pte, entry;
1763 BUG_ON(PageCompound(page+i));
1764 entry = mk_pte(page + i, vma->vm_page_prot);
1765 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1766 if (!pmd_write(*pmd))
1767 entry = pte_wrprotect(entry);
1768 else
1769 BUG_ON(page_mapcount(page) != 1);
1770 if (!pmd_young(*pmd))
1771 entry = pte_mkold(entry);
1ba6e0b5
AA		 1772 if (pmd_numa(*pmd))
1773 entry = pte_mknuma(entry);
71e3aac0
AA		 1774 pte = pte_offset_map(&_pmd, haddr);
1775 BUG_ON(!pte_none(*pte));
1776 set_pte_at(mm, haddr, pte, entry);
1777 pte_unmap(pte);
1778 }
1779
71e3aac0
AA		 1780 smp_wmb(); /* make pte visible before pmd */
1781 /*
1782 * Up to this point the pmd is present and huge and
1783 * userland has the whole access to the hugepage
1784 * during the split (which happens in place). If we
1785 * overwrite the pmd with the not-huge version
1786 * pointing to the pte here (which of course we could
1787 * if all CPUs were bug free), userland could trigger
1788 * a small page size TLB miss on the small sized TLB
1789 * while the hugepage TLB entry is still established
1790 * in the huge TLB. Some CPU doesn't like that. See
1791 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1792 * Erratum 383 on page 93. Intel should be safe but is
1793 * also warns that it's only safe if the permission
1794 * and cache attributes of the two entries loaded in
1795 * the two TLB is identical (which should be the case
1796 * here). But it is generally safer to never allow
1797 * small and huge TLB entries for the same virtual
1798 * address to be loaded simultaneously. So instead of
1799 * doing "pmd_populate(); flush_tlb_range();" we first
1800 * mark the current pmd notpresent (atomically because
1801 * here the pmd_trans_huge and pmd_trans_splitting
1802 * must remain set at all times on the pmd until the
1803 * split is complete for this pmd), then we flush the
1804 * SMP TLB and finally we write the non-huge version
1805 * of the pmd entry with pmd_populate.
1806 */
46dcde73 1807 pmdp_invalidate(vma, address, pmd);
71e3aac0
AA		 1808 pmd_populate(mm, pmd, pgtable);
1809 ret = 1;
1810 }
1811 spin_unlock(&mm->page_table_lock);
1812
1813 return ret;
1814}
1815
5a505085 1816/* must be called with anon_vma->root->rwsem held */
71e3aac0 1817static void __split_huge_page(struct page *page,
5bc7b8ac
SL		 1818 struct anon_vma *anon_vma,
1819 struct list_head *list)
71e3aac0
AA		 1820{
1821 int mapcount, mapcount2;
bf181b9f 1822 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
71e3aac0
AA		 1823 struct anon_vma_chain *avc;
1824
1825 BUG_ON(!PageHead(page));
1826 BUG_ON(PageTail(page));
1827
1828 mapcount = 0;
bf181b9f 1829 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
71e3aac0
AA		 1830 struct vm_area_struct *vma = avc->vma;
1831 unsigned long addr = vma_address(page, vma);
1832 BUG_ON(is_vma_temporary_stack(vma));
71e3aac0
AA		 1833 mapcount += __split_huge_page_splitting(page, vma, addr);
1834 }
05759d38
AA		 1835 /*
1836 * It is critical that new vmas are added to the tail of the
1837 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1838 * and establishes a child pmd before
1839 * __split_huge_page_splitting() freezes the parent pmd (so if
1840 * we fail to prevent copy_huge_pmd() from running until the
1841 * whole __split_huge_page() is complete), we will still see
1842 * the newly established pmd of the child later during the
1843 * walk, to be able to set it as pmd_trans_splitting too.
1844 */
1845 if (mapcount != page_mapcount(page))
1846 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1847 mapcount, page_mapcount(page));
71e3aac0
AA		 1848 BUG_ON(mapcount != page_mapcount(page));
1849
5bc7b8ac 1850 __split_huge_page_refcount(page, list);
71e3aac0
AA		 1851
1852 mapcount2 = 0;
bf181b9f 1853 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
71e3aac0
AA		 1854 struct vm_area_struct *vma = avc->vma;
1855 unsigned long addr = vma_address(page, vma);
1856 BUG_ON(is_vma_temporary_stack(vma));
71e3aac0
AA		 1857 mapcount2 += __split_huge_page_map(page, vma, addr);
1858 }
05759d38
AA		 1859 if (mapcount != mapcount2)
1860 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1861 mapcount, mapcount2, page_mapcount(page));
71e3aac0
AA		 1862 BUG_ON(mapcount != mapcount2);
1863}
1864
5bc7b8ac
SL		 1865/*
1866 * Split a hugepage into normal pages. This doesn't change the position of head
1867 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1868 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1869 * from the hugepage.
1870 * Return 0 if the hugepage is split successfully otherwise return 1.
1871 */
1872int split_huge_page_to_list(struct page *page, struct list_head *list)
71e3aac0
AA		 1873{
1874 struct anon_vma *anon_vma;
1875 int ret = 1;
1876
5918d10a 1877 BUG_ON(is_huge_zero_page(page));
71e3aac0 1878 BUG_ON(!PageAnon(page));
062f1af2
MG		 1879
1880 /*
1881 * The caller does not necessarily hold an mmap_sem that would prevent
1882 * the anon_vma disappearing so we first we take a reference to it
1883 * and then lock the anon_vma for write. This is similar to
1884 * page_lock_anon_vma_read except the write lock is taken to serialise
1885 * against parallel split or collapse operations.
1886 */
1887 anon_vma = page_get_anon_vma(page);
71e3aac0
AA		 1888 if (!anon_vma)
1889 goto out;
062f1af2
MG		 1890 anon_vma_lock_write(anon_vma);
1891
71e3aac0
AA		 1892 ret = 0;
1893 if (!PageCompound(page))
1894 goto out_unlock;
1895
1896 BUG_ON(!PageSwapBacked(page));
5bc7b8ac 1897 __split_huge_page(page, anon_vma, list);
81ab4201 1898 count_vm_event(THP_SPLIT);
71e3aac0
AA		 1899
1900 BUG_ON(PageCompound(page));
1901out_unlock:
08b52706 1902 anon_vma_unlock_write(anon_vma);
062f1af2 1903 put_anon_vma(anon_vma);
71e3aac0
AA		 1904out:
1905 return ret;
1906}
1907
4b6e1e37 1908#define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
78f11a25 1909
60ab3244
AA		 1910int hugepage_madvise(struct vm_area_struct *vma,
1911 unsigned long *vm_flags, int advice)
0af4e98b 1912{
8e72033f
GS		 1913 struct mm_struct *mm = vma->vm_mm;
1914
a664b2d8
AA		 1915 switch (advice) {
1916 case MADV_HUGEPAGE:
1917 /*
1918 * Be somewhat over-protective like KSM for now!
1919 */
78f11a25 1920 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
a664b2d8 1921 return -EINVAL;
8e72033f
GS		 1922 if (mm->def_flags & VM_NOHUGEPAGE)
1923 return -EINVAL;
a664b2d8
AA		 1924 *vm_flags &= ~VM_NOHUGEPAGE;
1925 *vm_flags |= VM_HUGEPAGE;
60ab3244
AA		 1926 /*
1927 * If the vma become good for khugepaged to scan,
1928 * register it here without waiting a page fault that
1929 * may not happen any time soon.
1930 */
1931 if (unlikely(khugepaged_enter_vma_merge(vma)))
1932 return -ENOMEM;
a664b2d8
AA		 1933 break;
1934 case MADV_NOHUGEPAGE:
1935 /*
1936 * Be somewhat over-protective like KSM for now!
1937 */
78f11a25 1938 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
a664b2d8
AA		 1939 return -EINVAL;
1940 *vm_flags &= ~VM_HUGEPAGE;
1941 *vm_flags |= VM_NOHUGEPAGE;
60ab3244
AA		 1942 /*
1943 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1944 * this vma even if we leave the mm registered in khugepaged if
1945 * it got registered before VM_NOHUGEPAGE was set.
1946 */
a664b2d8
AA		 1947 break;
1948 }
0af4e98b
AA		 1949
1950 return 0;
1951}
1952
ba76149f
AA		 1953static int __init khugepaged_slab_init(void)
1954{
1955 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1956 sizeof(struct mm_slot),
1957 __alignof__(struct mm_slot), 0, NULL);
1958 if (!mm_slot_cache)
1959 return -ENOMEM;
1960
1961 return 0;
1962}
1963
ba76149f
AA		 1964static inline struct mm_slot *alloc_mm_slot(void)
1965{
1966 if (!mm_slot_cache) /* initialization failed */
1967 return NULL;
1968 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1969}
1970
1971static inline void free_mm_slot(struct mm_slot *mm_slot)
1972{
1973 kmem_cache_free(mm_slot_cache, mm_slot);
1974}
1975
ba76149f
AA		 1976static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1977{
1978 struct mm_slot *mm_slot;
ba76149f 1979
b67bfe0d 1980 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
ba76149f
AA		 1981 if (mm == mm_slot->mm)
1982 return mm_slot;
43b5fbbd 1983
ba76149f
AA		 1984 return NULL;
1985}
1986
1987static void insert_to_mm_slots_hash(struct mm_struct *mm,
1988 struct mm_slot *mm_slot)
1989{
ba76149f 1990 mm_slot->mm = mm;
43b5fbbd 1991 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
ba76149f
AA		 1992}
1993
1994static inline int khugepaged_test_exit(struct mm_struct *mm)
1995{
1996 return atomic_read(&mm->mm_users) == 0;
1997}
1998
1999int __khugepaged_enter(struct mm_struct *mm)
2000{
2001 struct mm_slot *mm_slot;
2002 int wakeup;
2003
2004 mm_slot = alloc_mm_slot();
2005 if (!mm_slot)
2006 return -ENOMEM;
2007
2008 /* __khugepaged_exit() must not run from under us */
2009 VM_BUG_ON(khugepaged_test_exit(mm));
2010 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2011 free_mm_slot(mm_slot);
2012 return 0;
2013 }
2014
2015 spin_lock(&khugepaged_mm_lock);
2016 insert_to_mm_slots_hash(mm, mm_slot);
2017 /*
2018 * Insert just behind the scanning cursor, to let the area settle
2019 * down a little.
2020 */
2021 wakeup = list_empty(&khugepaged_scan.mm_head);
2022 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2023 spin_unlock(&khugepaged_mm_lock);
2024
2025 atomic_inc(&mm->mm_count);
2026 if (wakeup)
2027 wake_up_interruptible(&khugepaged_wait);
2028
2029 return 0;
2030}
2031
2032int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2033{
2034 unsigned long hstart, hend;
2035 if (!vma->anon_vma)
2036 /*
2037 * Not yet faulted in so we will register later in the
2038 * page fault if needed.
2039 */
2040 return 0;
78f11a25 2041 if (vma->vm_ops)
ba76149f
AA		 2042 /* khugepaged not yet working on file or special mappings */
2043 return 0;
b3b9c293 2044 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
ba76149f
AA		 2045 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2046 hend = vma->vm_end & HPAGE_PMD_MASK;
2047 if (hstart < hend)
2048 return khugepaged_enter(vma);
2049 return 0;
2050}
2051
2052void __khugepaged_exit(struct mm_struct *mm)
2053{
2054 struct mm_slot *mm_slot;
2055 int free = 0;
2056
2057 spin_lock(&khugepaged_mm_lock);
2058 mm_slot = get_mm_slot(mm);
2059 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
43b5fbbd 2060 hash_del(&mm_slot->hash);
ba76149f
AA		 2061 list_del(&mm_slot->mm_node);
2062 free = 1;
2063 }
d788e80a 2064 spin_unlock(&khugepaged_mm_lock);
ba76149f
AA		 2065
2066 if (free) {
ba76149f
AA		 2067 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2068 free_mm_slot(mm_slot);
2069 mmdrop(mm);
2070 } else if (mm_slot) {
ba76149f
AA		 2071 /*
2072 * This is required to serialize against
2073 * khugepaged_test_exit() (which is guaranteed to run
2074 * under mmap sem read mode). Stop here (after we
2075 * return all pagetables will be destroyed) until
2076 * khugepaged has finished working on the pagetables
2077 * under the mmap_sem.
2078 */
2079 down_write(&mm->mmap_sem);
2080 up_write(&mm->mmap_sem);
d788e80a 2081 }
ba76149f
AA		 2082}
2083
2084static void release_pte_page(struct page *page)
2085{
2086 /* 0 stands for page_is_file_cache(page) == false */
2087 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2088 unlock_page(page);
2089 putback_lru_page(page);
2090}
2091
2092static void release_pte_pages(pte_t *pte, pte_t *_pte)
2093{
2094 while (--_pte >= pte) {
2095 pte_t pteval = *_pte;
2096 if (!pte_none(pteval))
2097 release_pte_page(pte_page(pteval));
2098 }
2099}
2100
ba76149f
AA		 2101static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2102 unsigned long address,
2103 pte_t *pte)
2104{
2105 struct page *page;
2106 pte_t *_pte;
344aa35c 2107 int referenced = 0, none = 0;
ba76149f
AA		 2108 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2109 _pte++, address += PAGE_SIZE) {
2110 pte_t pteval = *_pte;
2111 if (pte_none(pteval)) {
2112 if (++none <= khugepaged_max_ptes_none)
2113 continue;
344aa35c 2114 else
ba76149f 2115 goto out;
ba76149f 2116 }
344aa35c 2117 if (!pte_present(pteval) || !pte_write(pteval))
ba76149f 2118 goto out;
ba76149f 2119 page = vm_normal_page(vma, address, pteval);
344aa35c 2120 if (unlikely(!page))
ba76149f 2121 goto out;
344aa35c 2122
ba76149f
AA		 2123 VM_BUG_ON(PageCompound(page));
2124 BUG_ON(!PageAnon(page));
2125 VM_BUG_ON(!PageSwapBacked(page));
2126
2127 /* cannot use mapcount: can't collapse if there's a gup pin */
344aa35c 2128 if (page_count(page) != 1)
ba76149f 2129 goto out;
ba76149f
AA		 2130 /*
2131 * We can do it before isolate_lru_page because the
2132 * page can't be freed from under us. NOTE: PG_lock
2133 * is needed to serialize against split_huge_page
2134 * when invoked from the VM.
2135 */
344aa35c 2136 if (!trylock_page(page))
ba76149f 2137 goto out;
ba76149f
AA		 2138 /*
2139 * Isolate the page to avoid collapsing an hugepage
2140 * currently in use by the VM.
2141 */
2142 if (isolate_lru_page(page)) {
2143 unlock_page(page);
ba76149f
AA		 2144 goto out;
2145 }
2146 /* 0 stands for page_is_file_cache(page) == false */
2147 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2148 VM_BUG_ON(!PageLocked(page));
2149 VM_BUG_ON(PageLRU(page));
2150
2151 /* If there is no mapped pte young don't collapse the page */
8ee53820
AA		 2152 if (pte_young(pteval) || PageReferenced(page) ||
2153 mmu_notifier_test_young(vma->vm_mm, address))
ba76149f
AA		 2154 referenced = 1;
2155 }
344aa35c
BL		 2156 if (likely(referenced))
2157 return 1;
ba76149f 2158out:
344aa35c
BL		 2159 release_pte_pages(pte, _pte);
2160 return 0;
ba76149f
AA		 2161}
2162
2163static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2164 struct vm_area_struct *vma,
2165 unsigned long address,
2166 spinlock_t *ptl)
2167{
2168 pte_t *_pte;
2169 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2170 pte_t pteval = *_pte;
2171 struct page *src_page;
2172
2173 if (pte_none(pteval)) {
2174 clear_user_highpage(page, address);
2175 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2176 } else {
2177 src_page = pte_page(pteval);
2178 copy_user_highpage(page, src_page, address, vma);
2179 VM_BUG_ON(page_mapcount(src_page) != 1);
ba76149f
AA		 2180 release_pte_page(src_page);
2181 /*
2182 * ptl mostly unnecessary, but preempt has to
2183 * be disabled to update the per-cpu stats
2184 * inside page_remove_rmap().
2185 */
2186 spin_lock(ptl);
2187 /*
2188 * paravirt calls inside pte_clear here are
2189 * superfluous.
2190 */
2191 pte_clear(vma->vm_mm, address, _pte);
2192 page_remove_rmap(src_page);
2193 spin_unlock(ptl);
2194 free_page_and_swap_cache(src_page);
2195 }
2196
2197 address += PAGE_SIZE;
2198 page++;
2199 }
2200}
2201
26234f36 2202static void khugepaged_alloc_sleep(void)
ba76149f 2203{
26234f36
XG		 2204 wait_event_freezable_timeout(khugepaged_wait, false,
2205 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2206}
ba76149f 2207
9f1b868a
BL		 2208static int khugepaged_node_load[MAX_NUMNODES];
2209
26234f36 2210#ifdef CONFIG_NUMA
9f1b868a
BL		 2211static int khugepaged_find_target_node(void)
2212{
2213 static int last_khugepaged_target_node = NUMA_NO_NODE;
2214 int nid, target_node = 0, max_value = 0;
2215
2216 /* find first node with max normal pages hit */
2217 for (nid = 0; nid < MAX_NUMNODES; nid++)
2218 if (khugepaged_node_load[nid] > max_value) {
2219 max_value = khugepaged_node_load[nid];
2220 target_node = nid;
2221 }
2222
2223 /* do some balance if several nodes have the same hit record */
2224 if (target_node <= last_khugepaged_target_node)
2225 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2226 nid++)
2227 if (max_value == khugepaged_node_load[nid]) {
2228 target_node = nid;
2229 break;
2230 }
2231
2232 last_khugepaged_target_node = target_node;
2233 return target_node;
2234}
2235
26234f36
XG		 2236static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2237{
2238 if (IS_ERR(*hpage)) {
2239 if (!*wait)
2240 return false;
2241
2242 *wait = false;
e3b4126c 2243 *hpage = NULL;
26234f36
XG		 2244 khugepaged_alloc_sleep();
2245 } else if (*hpage) {
2246 put_page(*hpage);
2247 *hpage = NULL;
2248 }
2249
2250 return true;
2251}
2252
2253static struct page
2254*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2255 struct vm_area_struct *vma, unsigned long address,
2256 int node)
2257{
0bbbc0b3 2258 VM_BUG_ON(*hpage);
ce83d217
AA		 2259 /*
2260 * Allocate the page while the vma is still valid and under
2261 * the mmap_sem read mode so there is no memory allocation
2262 * later when we take the mmap_sem in write mode. This is more
2263 * friendly behavior (OTOH it may actually hide bugs) to
2264 * filesystems in userland with daemons allocating memory in
2265 * the userland I/O paths. Allocating memory with the
2266 * mmap_sem in read mode is good idea also to allow greater
2267 * scalability.
2268 */
9f1b868a
BL		 2269 *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2270 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
692e0b35
AA		 2271 /*
2272 * After allocating the hugepage, release the mmap_sem read lock in
2273 * preparation for taking it in write mode.
2274 */
2275 up_read(&mm->mmap_sem);
26234f36 2276 if (unlikely(!*hpage)) {
81ab4201 2277 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
ce83d217 2278 *hpage = ERR_PTR(-ENOMEM);
26234f36 2279 return NULL;
ce83d217 2280 }
26234f36 2281
65b3c07b 2282 count_vm_event(THP_COLLAPSE_ALLOC);
26234f36
XG		 2283 return *hpage;
2284}
2285#else
9f1b868a
BL		 2286static int khugepaged_find_target_node(void)
2287{
2288 return 0;
2289}
2290
10dc4155
BL		 2291static inline struct page *alloc_hugepage(int defrag)
2292{
2293 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2294 HPAGE_PMD_ORDER);
2295}
2296
26234f36
XG		 2297static struct page *khugepaged_alloc_hugepage(bool *wait)
2298{
2299 struct page *hpage;
2300
2301 do {
2302 hpage = alloc_hugepage(khugepaged_defrag());
2303 if (!hpage) {
2304 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2305 if (!*wait)
2306 return NULL;
2307
2308 *wait = false;
2309 khugepaged_alloc_sleep();
2310 } else
2311 count_vm_event(THP_COLLAPSE_ALLOC);
2312 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2313
2314 return hpage;
2315}
2316
2317static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2318{
2319 if (!*hpage)
2320 *hpage = khugepaged_alloc_hugepage(wait);
2321
2322 if (unlikely(!*hpage))
2323 return false;
2324
2325 return true;
2326}
2327
2328static struct page
2329*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2330 struct vm_area_struct *vma, unsigned long address,
2331 int node)
2332{
2333 up_read(&mm->mmap_sem);
2334 VM_BUG_ON(!*hpage);
2335 return *hpage;
2336}
692e0b35
AA		 2337#endif
2338
fa475e51
BL		 2339static bool hugepage_vma_check(struct vm_area_struct *vma)
2340{
2341 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2342 (vma->vm_flags & VM_NOHUGEPAGE))
2343 return false;
2344
2345 if (!vma->anon_vma || vma->vm_ops)
2346 return false;
2347 if (is_vma_temporary_stack(vma))
2348 return false;
2349 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2350 return true;
2351}
2352
26234f36
XG		 2353static void collapse_huge_page(struct mm_struct *mm,
2354 unsigned long address,
2355 struct page **hpage,
2356 struct vm_area_struct *vma,
2357 int node)
2358{
26234f36
XG		 2359 pmd_t *pmd, _pmd;
2360 pte_t *pte;
2361 pgtable_t pgtable;
2362 struct page *new_page;
2363 spinlock_t *ptl;
2364 int isolated;
2365 unsigned long hstart, hend;
2ec74c3e
SG		 2366 unsigned long mmun_start; /* For mmu_notifiers */
2367 unsigned long mmun_end; /* For mmu_notifiers */
26234f36
XG		 2368
2369 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2370
2371 /* release the mmap_sem read lock. */
2372 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2373 if (!new_page)
2374 return;
2375
420256ef 2376 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
ce83d217 2377 return;
ba76149f
AA		 2378
2379 /*
2380 * Prevent all access to pagetables with the exception of
2381 * gup_fast later hanlded by the ptep_clear_flush and the VM
2382 * handled by the anon_vma lock + PG_lock.
2383 */
2384 down_write(&mm->mmap_sem);
2385 if (unlikely(khugepaged_test_exit(mm)))
2386 goto out;
2387
2388 vma = find_vma(mm, address);
a8f531eb
L		 2389 if (!vma)
2390 goto out;
ba76149f
AA		 2391 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2392 hend = vma->vm_end & HPAGE_PMD_MASK;
2393 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2394 goto out;
fa475e51 2395 if (!hugepage_vma_check(vma))
a7d6e4ec 2396 goto out;
6219049a
BL		 2397 pmd = mm_find_pmd(mm, address);
2398 if (!pmd)
ba76149f 2399 goto out;
6219049a 2400 if (pmd_trans_huge(*pmd))
ba76149f
AA		 2401 goto out;
2402
4fc3f1d6 2403 anon_vma_lock_write(vma->anon_vma);
ba76149f
AA		 2404
2405 pte = pte_offset_map(pmd, address);
2406 ptl = pte_lockptr(mm, pmd);
2407
2ec74c3e
SG		 2408 mmun_start = address;
2409 mmun_end = address + HPAGE_PMD_SIZE;
2410 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
ba76149f
AA		 2411 spin_lock(&mm->page_table_lock); /* probably unnecessary */
2412 /*
2413 * After this gup_fast can't run anymore. This also removes
2414 * any huge TLB entry from the CPU so we won't allow
2415 * huge and small TLB entries for the same virtual address
2416 * to avoid the risk of CPU bugs in that area.
2417 */
2ec74c3e 2418 _pmd = pmdp_clear_flush(vma, address, pmd);
ba76149f 2419 spin_unlock(&mm->page_table_lock);
2ec74c3e 2420 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
ba76149f
AA		 2421
2422 spin_lock(ptl);
2423 isolated = __collapse_huge_page_isolate(vma, address, pte);
2424 spin_unlock(ptl);
ba76149f
AA		 2425
2426 if (unlikely(!isolated)) {
453c7192 2427 pte_unmap(pte);
ba76149f
AA		 2428 spin_lock(&mm->page_table_lock);
2429 BUG_ON(!pmd_none(*pmd));
7c342512
AK		 2430 /*
2431 * We can only use set_pmd_at when establishing
2432 * hugepmds and never for establishing regular pmds that
2433 * points to regular pagetables. Use pmd_populate for that
2434 */
2435 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
ba76149f 2436 spin_unlock(&mm->page_table_lock);
08b52706 2437 anon_vma_unlock_write(vma->anon_vma);
ce83d217 2438 goto out;
ba76149f
AA		 2439 }
2440
2441 /*
2442 * All pages are isolated and locked so anon_vma rmap
2443 * can't run anymore.
2444 */
08b52706 2445 anon_vma_unlock_write(vma->anon_vma);
ba76149f
AA		 2446
2447 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
453c7192 2448 pte_unmap(pte);
ba76149f
AA		 2449 __SetPageUptodate(new_page);
2450 pgtable = pmd_pgtable(_pmd);
ba76149f 2451
3122359a
KS		 2452 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2453 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
ba76149f
AA		 2454
2455 /*
2456 * spin_lock() below is not the equivalent of smp_wmb(), so
2457 * this is needed to avoid the copy_huge_page writes to become
2458 * visible after the set_pmd_at() write.
2459 */
2460 smp_wmb();
2461
2462 spin_lock(&mm->page_table_lock);
2463 BUG_ON(!pmd_none(*pmd));
2464 page_add_new_anon_rmap(new_page, vma, address);
fce144b4 2465 pgtable_trans_huge_deposit(mm, pmd, pgtable);
ba76149f 2466 set_pmd_at(mm, address, pmd, _pmd);
b113da65 2467 update_mmu_cache_pmd(vma, address, pmd);
ba76149f
AA		 2468 spin_unlock(&mm->page_table_lock);
2469
2470 *hpage = NULL;
420256ef 2471
ba76149f 2472 khugepaged_pages_collapsed++;
ce83d217 2473out_up_write:
ba76149f 2474 up_write(&mm->mmap_sem);
0bbbc0b3
AA		 2475 return;
2476
ce83d217 2477out:
678ff896 2478 mem_cgroup_uncharge_page(new_page);
ce83d217 2479 goto out_up_write;
ba76149f
AA		 2480}
2481
2482static int khugepaged_scan_pmd(struct mm_struct *mm,
2483 struct vm_area_struct *vma,
2484 unsigned long address,
2485 struct page **hpage)
2486{
ba76149f
AA		 2487 pmd_t *pmd;
2488 pte_t *pte, *_pte;
2489 int ret = 0, referenced = 0, none = 0;
2490 struct page *page;
2491 unsigned long _address;
2492 spinlock_t *ptl;
00ef2d2f 2493 int node = NUMA_NO_NODE;
ba76149f
AA		 2494
2495 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2496
6219049a
BL		 2497 pmd = mm_find_pmd(mm, address);
2498 if (!pmd)
ba76149f 2499 goto out;
6219049a 2500 if (pmd_trans_huge(*pmd))
ba76149f
AA		 2501 goto out;
2502
9f1b868a 2503 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
ba76149f
AA		 2504 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2505 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2506 _pte++, _address += PAGE_SIZE) {
2507 pte_t pteval = *_pte;
2508 if (pte_none(pteval)) {
2509 if (++none <= khugepaged_max_ptes_none)
2510 continue;
2511 else
2512 goto out_unmap;
2513 }
2514 if (!pte_present(pteval) || !pte_write(pteval))
2515 goto out_unmap;
2516 page = vm_normal_page(vma, _address, pteval);
2517 if (unlikely(!page))
2518 goto out_unmap;
5c4b4be3 2519 /*
9f1b868a
BL		 2520 * Record which node the original page is from and save this
2521 * information to khugepaged_node_load[].
2522 * Khupaged will allocate hugepage from the node has the max
2523 * hit record.
5c4b4be3 2524 */
9f1b868a
BL		 2525 node = page_to_nid(page);
2526 khugepaged_node_load[node]++;
ba76149f
AA		 2527 VM_BUG_ON(PageCompound(page));
2528 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2529 goto out_unmap;
2530 /* cannot use mapcount: can't collapse if there's a gup pin */
2531 if (page_count(page) != 1)
2532 goto out_unmap;
8ee53820
AA		 2533 if (pte_young(pteval) || PageReferenced(page) ||
2534 mmu_notifier_test_young(vma->vm_mm, address))
ba76149f
AA		 2535 referenced = 1;
2536 }
2537 if (referenced)
2538 ret = 1;
2539out_unmap:
2540 pte_unmap_unlock(pte, ptl);
9f1b868a
BL		 2541 if (ret) {
2542 node = khugepaged_find_target_node();
ce83d217 2543 /* collapse_huge_page will return with the mmap_sem released */
5c4b4be3 2544 collapse_huge_page(mm, address, hpage, vma, node);
9f1b868a 2545 }
ba76149f
AA		 2546out:
2547 return ret;
2548}
2549
2550static void collect_mm_slot(struct mm_slot *mm_slot)
2551{
2552 struct mm_struct *mm = mm_slot->mm;
2553
b9980cdc 2554 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
ba76149f
AA		 2555
2556 if (khugepaged_test_exit(mm)) {
2557 /* free mm_slot */
43b5fbbd 2558 hash_del(&mm_slot->hash);
ba76149f
AA		 2559 list_del(&mm_slot->mm_node);
2560
2561 /*
2562 * Not strictly needed because the mm exited already.
2563 *
2564 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2565 */
2566
2567 /* khugepaged_mm_lock actually not necessary for the below */
2568 free_mm_slot(mm_slot);
2569 mmdrop(mm);
2570 }
2571}
2572
2573static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2574 struct page **hpage)
2f1da642
HS		 2575 __releases(&khugepaged_mm_lock)
2576 __acquires(&khugepaged_mm_lock)
ba76149f
AA		 2577{
2578 struct mm_slot *mm_slot;
2579 struct mm_struct *mm;
2580 struct vm_area_struct *vma;
2581 int progress = 0;
2582
2583 VM_BUG_ON(!pages);
b9980cdc 2584 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
ba76149f
AA		 2585
2586 if (khugepaged_scan.mm_slot)
2587 mm_slot = khugepaged_scan.mm_slot;
2588 else {
2589 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2590 struct mm_slot, mm_node);
2591 khugepaged_scan.address = 0;
2592 khugepaged_scan.mm_slot = mm_slot;
2593 }
2594 spin_unlock(&khugepaged_mm_lock);
2595
2596 mm = mm_slot->mm;
2597 down_read(&mm->mmap_sem);
2598 if (unlikely(khugepaged_test_exit(mm)))
2599 vma = NULL;
2600 else
2601 vma = find_vma(mm, khugepaged_scan.address);
2602
2603 progress++;
2604 for (; vma; vma = vma->vm_next) {
2605 unsigned long hstart, hend;
2606
2607 cond_resched();
2608 if (unlikely(khugepaged_test_exit(mm))) {
2609 progress++;
2610 break;
2611 }
fa475e51
BL		 2612 if (!hugepage_vma_check(vma)) {
2613skip:
ba76149f
AA		 2614 progress++;
2615 continue;
2616 }
ba76149f
AA		 2617 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2618 hend = vma->vm_end & HPAGE_PMD_MASK;
a7d6e4ec
AA		 2619 if (hstart >= hend)
2620 goto skip;
2621 if (khugepaged_scan.address > hend)
2622 goto skip;
ba76149f
AA		 2623 if (khugepaged_scan.address < hstart)
2624 khugepaged_scan.address = hstart;
a7d6e4ec 2625 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
ba76149f
AA		 2626
2627 while (khugepaged_scan.address < hend) {
2628 int ret;
2629 cond_resched();
2630 if (unlikely(khugepaged_test_exit(mm)))
2631 goto breakouterloop;
2632
2633 VM_BUG_ON(khugepaged_scan.address < hstart ||
2634 khugepaged_scan.address + HPAGE_PMD_SIZE >
2635 hend);
2636 ret = khugepaged_scan_pmd(mm, vma,
2637 khugepaged_scan.address,
2638 hpage);
2639 /* move to next address */
2640 khugepaged_scan.address += HPAGE_PMD_SIZE;
2641 progress += HPAGE_PMD_NR;
2642 if (ret)
2643 /* we released mmap_sem so break loop */
2644 goto breakouterloop_mmap_sem;
2645 if (progress >= pages)
2646 goto breakouterloop;
2647 }
2648 }
2649breakouterloop:
2650 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2651breakouterloop_mmap_sem:
2652
2653 spin_lock(&khugepaged_mm_lock);
a7d6e4ec 2654 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
ba76149f
AA		 2655 /*
2656 * Release the current mm_slot if this mm is about to die, or
2657 * if we scanned all vmas of this mm.
2658 */
2659 if (khugepaged_test_exit(mm) || !vma) {
2660 /*
2661 * Make sure that if mm_users is reaching zero while
2662 * khugepaged runs here, khugepaged_exit will find
2663 * mm_slot not pointing to the exiting mm.
2664 */
2665 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2666 khugepaged_scan.mm_slot = list_entry(
2667 mm_slot->mm_node.next,
2668 struct mm_slot, mm_node);
2669 khugepaged_scan.address = 0;
2670 } else {
2671 khugepaged_scan.mm_slot = NULL;
2672 khugepaged_full_scans++;
2673 }
2674
2675 collect_mm_slot(mm_slot);
2676 }
2677
2678 return progress;
2679}
2680
2681static int khugepaged_has_work(void)
2682{
2683 return !list_empty(&khugepaged_scan.mm_head) &&
2684 khugepaged_enabled();
2685}
2686
2687static int khugepaged_wait_event(void)
2688{
2689 return !list_empty(&khugepaged_scan.mm_head) ||
2017c0bf 2690 kthread_should_stop();
ba76149f
AA		 2691}
2692
d516904b 2693static void khugepaged_do_scan(void)
ba76149f 2694{
d516904b 2695 struct page *hpage = NULL;
ba76149f
AA		 2696 unsigned int progress = 0, pass_through_head = 0;
2697 unsigned int pages = khugepaged_pages_to_scan;
d516904b 2698 bool wait = true;
ba76149f
AA		 2699
2700 barrier(); /* write khugepaged_pages_to_scan to local stack */
2701
2702 while (progress < pages) {
26234f36 2703 if (!khugepaged_prealloc_page(&hpage, &wait))
d516904b 2704 break;
26234f36 2705
420256ef 2706 cond_resched();
ba76149f 2707
878aee7d
AA		 2708 if (unlikely(kthread_should_stop() || freezing(current)))
2709 break;
2710
ba76149f
AA		 2711 spin_lock(&khugepaged_mm_lock);
2712 if (!khugepaged_scan.mm_slot)
2713 pass_through_head++;
2714 if (khugepaged_has_work() &&
2715 pass_through_head < 2)
2716 progress += khugepaged_scan_mm_slot(pages - progress,
d516904b 2717 &hpage);
ba76149f
AA		 2718 else
2719 progress = pages;
2720 spin_unlock(&khugepaged_mm_lock);
2721 }
ba76149f 2722
d516904b
XG		 2723 if (!IS_ERR_OR_NULL(hpage))
2724 put_page(hpage);
0bbbc0b3
AA		 2725}
2726
2017c0bf
XG		 2727static void khugepaged_wait_work(void)
2728{
2729 try_to_freeze();
2730
2731 if (khugepaged_has_work()) {
2732 if (!khugepaged_scan_sleep_millisecs)
2733 return;
2734
2735 wait_event_freezable_timeout(khugepaged_wait,
2736 kthread_should_stop(),
2737 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2738 return;
2739 }
2740
2741 if (khugepaged_enabled())
2742 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2743}
2744
ba76149f
AA		 2745static int khugepaged(void *none)
2746{
2747 struct mm_slot *mm_slot;
2748
878aee7d 2749 set_freezable();
ba76149f
AA		 2750 set_user_nice(current, 19);
2751
b7231789
XG		 2752 while (!kthread_should_stop()) {
2753 khugepaged_do_scan();
2754 khugepaged_wait_work();
2755 }
ba76149f
AA		 2756
2757 spin_lock(&khugepaged_mm_lock);
2758 mm_slot = khugepaged_scan.mm_slot;
2759 khugepaged_scan.mm_slot = NULL;
2760 if (mm_slot)
2761 collect_mm_slot(mm_slot);
2762 spin_unlock(&khugepaged_mm_lock);
ba76149f
AA		 2763 return 0;
2764}
2765
c5a647d0
KS		 2766static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2767 unsigned long haddr, pmd_t *pmd)
2768{
2769 struct mm_struct *mm = vma->vm_mm;
2770 pgtable_t pgtable;
2771 pmd_t _pmd;
2772 int i;
2773
2774 pmdp_clear_flush(vma, haddr, pmd);
2775 /* leave pmd empty until pte is filled */
2776
6b0b50b0 2777 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
c5a647d0
KS		 2778 pmd_populate(mm, &_pmd, pgtable);
2779
2780 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2781 pte_t *pte, entry;
2782 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2783 entry = pte_mkspecial(entry);
2784 pte = pte_offset_map(&_pmd, haddr);
2785 VM_BUG_ON(!pte_none(*pte));
2786 set_pte_at(mm, haddr, pte, entry);
2787 pte_unmap(pte);
2788 }
2789 smp_wmb(); /* make pte visible before pmd */
2790 pmd_populate(mm, pmd, pgtable);
97ae1749 2791 put_huge_zero_page();
c5a647d0
KS		 2792}
2793
e180377f
KS		 2794void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2795 pmd_t *pmd)
71e3aac0
AA		 2796{
2797 struct page *page;
e180377f 2798 struct mm_struct *mm = vma->vm_mm;
c5a647d0
KS		 2799 unsigned long haddr = address & HPAGE_PMD_MASK;
2800 unsigned long mmun_start; /* For mmu_notifiers */
2801 unsigned long mmun_end; /* For mmu_notifiers */
e180377f
KS		 2802
2803 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
71e3aac0 2804
c5a647d0
KS		 2805 mmun_start = haddr;
2806 mmun_end = haddr + HPAGE_PMD_SIZE;
750e8165 2807again:
c5a647d0 2808 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
71e3aac0
AA		 2809 spin_lock(&mm->page_table_lock);
2810 if (unlikely(!pmd_trans_huge(*pmd))) {
2811 spin_unlock(&mm->page_table_lock);
c5a647d0
KS		 2812 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2813 return;
2814 }
2815 if (is_huge_zero_pmd(*pmd)) {
2816 __split_huge_zero_page_pmd(vma, haddr, pmd);
2817 spin_unlock(&mm->page_table_lock);
2818 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 2819 return;
2820 }
2821 page = pmd_page(*pmd);
2822 VM_BUG_ON(!page_count(page));
2823 get_page(page);
2824 spin_unlock(&mm->page_table_lock);
c5a647d0 2825 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 2826
2827 split_huge_page(page);
2828
2829 put_page(page);
750e8165
HD		 2830
2831 /*
2832 * We don't always have down_write of mmap_sem here: a racing
2833 * do_huge_pmd_wp_page() might have copied-on-write to another
2834 * huge page before our split_huge_page() got the anon_vma lock.
2835 */
2836 if (unlikely(pmd_trans_huge(*pmd)))
2837 goto again;
71e3aac0 2838}
94fcc585 2839
e180377f
KS		 2840void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2841 pmd_t *pmd)
2842{
2843 struct vm_area_struct *vma;
2844
2845 vma = find_vma(mm, address);
2846 BUG_ON(vma == NULL);
2847 split_huge_page_pmd(vma, address, pmd);
2848}
2849
94fcc585
AA		 2850static void split_huge_page_address(struct mm_struct *mm,
2851 unsigned long address)
2852{
94fcc585
AA		 2853 pmd_t *pmd;
2854
2855 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2856
6219049a
BL		 2857 pmd = mm_find_pmd(mm, address);
2858 if (!pmd)
94fcc585
AA		 2859 return;
2860 /*
2861 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2862 * materialize from under us.
2863 */
e180377f 2864 split_huge_page_pmd_mm(mm, address, pmd);
94fcc585
AA		 2865}
2866
2867void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2868 unsigned long start,
2869 unsigned long end,
2870 long adjust_next)
2871{
2872 /*
2873 * If the new start address isn't hpage aligned and it could
2874 * previously contain an hugepage: check if we need to split
2875 * an huge pmd.
2876 */
2877 if (start & ~HPAGE_PMD_MASK &&
2878 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2879 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2880 split_huge_page_address(vma->vm_mm, start);
2881
2882 /*
2883 * If the new end address isn't hpage aligned and it could
2884 * previously contain an hugepage: check if we need to split
2885 * an huge pmd.
2886 */
2887 if (end & ~HPAGE_PMD_MASK &&
2888 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2889 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2890 split_huge_page_address(vma->vm_mm, end);
2891
2892 /*
2893 * If we're also updating the vma->vm_next->vm_start, if the new
2894 * vm_next->vm_start isn't page aligned and it could previously
2895 * contain an hugepage: check if we need to split an huge pmd.
2896 */
2897 if (adjust_next > 0) {
2898 struct vm_area_struct *next = vma->vm_next;
2899 unsigned long nstart = next->vm_start;
2900 nstart += adjust_next << PAGE_SHIFT;
2901 if (nstart & ~HPAGE_PMD_MASK &&
2902 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2903 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2904 split_huge_page_address(next->vm_mm, nstart);
2905 }
2906}
Linux 6.x block layer and io_uring tree(s)