projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
kexec: migrate to reboot cpu
[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 712 pgtable_t pgtable;
c4088ebd 713 spinlock_t *ptl;
71e3aac0
AA		 714
715 VM_BUG_ON(!PageCompound(page));
716 pgtable = pte_alloc_one(mm, haddr);
edad9d2c 717 if (unlikely(!pgtable))
71e3aac0 718 return VM_FAULT_OOM;
71e3aac0
AA		 719
720 clear_huge_page(page, haddr, HPAGE_PMD_NR);
52f37629
MK		 721 /*
722 * The memory barrier inside __SetPageUptodate makes sure that
723 * clear_huge_page writes become visible before the set_pmd_at()
724 * write.
725 */
71e3aac0
AA		 726 __SetPageUptodate(page);
727
c4088ebd 728 ptl = pmd_lock(mm, pmd);
71e3aac0 729 if (unlikely(!pmd_none(*pmd))) {
c4088ebd 730 spin_unlock(ptl);
b9bbfbe3 731 mem_cgroup_uncharge_page(page);
71e3aac0
AA		 732 put_page(page);
733 pte_free(mm, pgtable);
734 } else {
735 pmd_t entry;
3122359a
KS		 736 entry = mk_huge_pmd(page, vma->vm_page_prot);
737 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
71e3aac0 738 page_add_new_anon_rmap(page, vma, haddr);
6b0b50b0 739 pgtable_trans_huge_deposit(mm, pmd, pgtable);
71e3aac0 740 set_pmd_at(mm, haddr, pmd, entry);
71e3aac0 741 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
e1f56c89 742 atomic_long_inc(&mm->nr_ptes);
c4088ebd 743 spin_unlock(ptl);
71e3aac0
AA		 744 }
745
aa2e878e 746 return 0;
71e3aac0
AA		 747}
748
cc5d462f 749static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
0bbbc0b3 750{
cc5d462f 751 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
0bbbc0b3
AA		 752}
753
754static inline struct page *alloc_hugepage_vma(int defrag,
755 struct vm_area_struct *vma,
cc5d462f
AK		 756 unsigned long haddr, int nd,
757 gfp_t extra_gfp)
0bbbc0b3 758{
cc5d462f 759 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
5c4b4be3 760 HPAGE_PMD_ORDER, vma, haddr, nd);
0bbbc0b3
AA		 761}
762
c4088ebd 763/* Caller must hold page table lock. */
3ea41e62 764static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
97ae1749 765 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
5918d10a 766 struct page *zero_page)
fc9fe822
KS		 767{
768 pmd_t entry;
3ea41e62
KS		 769 if (!pmd_none(*pmd))
770 return false;
5918d10a 771 entry = mk_pmd(zero_page, vma->vm_page_prot);
fc9fe822
KS		 772 entry = pmd_wrprotect(entry);
773 entry = pmd_mkhuge(entry);
6b0b50b0 774 pgtable_trans_huge_deposit(mm, pmd, pgtable);
fc9fe822 775 set_pmd_at(mm, haddr, pmd, entry);
e1f56c89 776 atomic_long_inc(&mm->nr_ptes);
3ea41e62 777 return true;
fc9fe822
KS		 778}
779
71e3aac0
AA		 780int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
781 unsigned long address, pmd_t *pmd,
782 unsigned int flags)
783{
784 struct page *page;
785 unsigned long haddr = address & HPAGE_PMD_MASK;
71e3aac0 786
128ec037 787 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
c0292554 788 return VM_FAULT_FALLBACK;
128ec037
KS		 789 if (unlikely(anon_vma_prepare(vma)))
790 return VM_FAULT_OOM;
791 if (unlikely(khugepaged_enter(vma)))
792 return VM_FAULT_OOM;
793 if (!(flags & FAULT_FLAG_WRITE) &&
794 transparent_hugepage_use_zero_page()) {
c4088ebd 795 spinlock_t *ptl;
128ec037
KS		 796 pgtable_t pgtable;
797 struct page *zero_page;
798 bool set;
799 pgtable = pte_alloc_one(mm, haddr);
800 if (unlikely(!pgtable))
ba76149f 801 return VM_FAULT_OOM;
128ec037
KS		 802 zero_page = get_huge_zero_page();
803 if (unlikely(!zero_page)) {
804 pte_free(mm, pgtable);
81ab4201 805 count_vm_event(THP_FAULT_FALLBACK);
c0292554 806 return VM_FAULT_FALLBACK;
b9bbfbe3 807 }
c4088ebd 808 ptl = pmd_lock(mm, pmd);
128ec037
KS		 809 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
810 zero_page);
c4088ebd 811 spin_unlock(ptl);
128ec037
KS		 812 if (!set) {
813 pte_free(mm, pgtable);
814 put_huge_zero_page();
edad9d2c 815 }
edad9d2c 816 return 0;
71e3aac0 817 }
128ec037
KS		 818 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
819 vma, haddr, numa_node_id(), 0);
820 if (unlikely(!page)) {
821 count_vm_event(THP_FAULT_FALLBACK);
c0292554 822 return VM_FAULT_FALLBACK;
128ec037 823 }
128ec037
KS		 824 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
825 put_page(page);
17766dde 826 count_vm_event(THP_FAULT_FALLBACK);
c0292554 827 return VM_FAULT_FALLBACK;
128ec037
KS		 828 }
829 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
830 mem_cgroup_uncharge_page(page);
831 put_page(page);
17766dde 832 count_vm_event(THP_FAULT_FALLBACK);
c0292554 833 return VM_FAULT_FALLBACK;
128ec037
KS		 834 }
835
17766dde 836 count_vm_event(THP_FAULT_ALLOC);
128ec037 837 return 0;
71e3aac0
AA		 838}
839
840int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
841 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
842 struct vm_area_struct *vma)
843{
c4088ebd 844 spinlock_t *dst_ptl, *src_ptl;
71e3aac0
AA		 845 struct page *src_page;
846 pmd_t pmd;
847 pgtable_t pgtable;
848 int ret;
849
850 ret = -ENOMEM;
851 pgtable = pte_alloc_one(dst_mm, addr);
852 if (unlikely(!pgtable))
853 goto out;
854
c4088ebd
KS		 855 dst_ptl = pmd_lock(dst_mm, dst_pmd);
856 src_ptl = pmd_lockptr(src_mm, src_pmd);
857 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
71e3aac0
AA		 858
859 ret = -EAGAIN;
860 pmd = *src_pmd;
861 if (unlikely(!pmd_trans_huge(pmd))) {
862 pte_free(dst_mm, pgtable);
863 goto out_unlock;
864 }
fc9fe822 865 /*
c4088ebd 866 * When page table lock is held, the huge zero pmd should not be
fc9fe822
KS		 867 * under splitting since we don't split the page itself, only pmd to
868 * a page table.
869 */
870 if (is_huge_zero_pmd(pmd)) {
5918d10a 871 struct page *zero_page;
3ea41e62 872 bool set;
97ae1749
KS		 873 /*
874 * get_huge_zero_page() will never allocate a new page here,
875 * since we already have a zero page to copy. It just takes a
876 * reference.
877 */
5918d10a 878 zero_page = get_huge_zero_page();
3ea41e62 879 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
5918d10a 880 zero_page);
3ea41e62 881 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
fc9fe822
KS		 882 ret = 0;
883 goto out_unlock;
884 }
71e3aac0
AA		 885 if (unlikely(pmd_trans_splitting(pmd))) {
886 /* split huge page running from under us */
c4088ebd
KS		 887 spin_unlock(src_ptl);
888 spin_unlock(dst_ptl);
71e3aac0
AA		 889 pte_free(dst_mm, pgtable);
890
891 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
892 goto out;
893 }
894 src_page = pmd_page(pmd);
895 VM_BUG_ON(!PageHead(src_page));
896 get_page(src_page);
897 page_dup_rmap(src_page);
898 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
899
900 pmdp_set_wrprotect(src_mm, addr, src_pmd);
901 pmd = pmd_mkold(pmd_wrprotect(pmd));
6b0b50b0 902 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
71e3aac0 903 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
e1f56c89 904 atomic_long_inc(&dst_mm->nr_ptes);
71e3aac0
AA		 905
906 ret = 0;
907out_unlock:
c4088ebd
KS		 908 spin_unlock(src_ptl);
909 spin_unlock(dst_ptl);
71e3aac0
AA		 910out:
911 return ret;
912}
913
a1dd450b
WD		 914void huge_pmd_set_accessed(struct mm_struct *mm,
915 struct vm_area_struct *vma,
916 unsigned long address,
917 pmd_t *pmd, pmd_t orig_pmd,
918 int dirty)
919{
c4088ebd 920 spinlock_t *ptl;
a1dd450b
WD		 921 pmd_t entry;
922 unsigned long haddr;
923
c4088ebd 924 ptl = pmd_lock(mm, pmd);
a1dd450b
WD		 925 if (unlikely(!pmd_same(*pmd, orig_pmd)))
926 goto unlock;
927
928 entry = pmd_mkyoung(orig_pmd);
929 haddr = address & HPAGE_PMD_MASK;
930 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
931 update_mmu_cache_pmd(vma, address, pmd);
932
933unlock:
c4088ebd 934 spin_unlock(ptl);
a1dd450b
WD		 935}
936
93b4796d
KS		 937static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
938 struct vm_area_struct *vma, unsigned long address,
3ea41e62 939 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
93b4796d 940{
c4088ebd 941 spinlock_t *ptl;
93b4796d
KS		 942 pgtable_t pgtable;
943 pmd_t _pmd;
944 struct page *page;
945 int i, ret = 0;
946 unsigned long mmun_start; /* For mmu_notifiers */
947 unsigned long mmun_end; /* For mmu_notifiers */
948
949 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
950 if (!page) {
951 ret |= VM_FAULT_OOM;
952 goto out;
953 }
954
955 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
956 put_page(page);
957 ret |= VM_FAULT_OOM;
958 goto out;
959 }
960
961 clear_user_highpage(page, address);
962 __SetPageUptodate(page);
963
964 mmun_start = haddr;
965 mmun_end = haddr + HPAGE_PMD_SIZE;
966 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
967
c4088ebd 968 ptl = pmd_lock(mm, pmd);
3ea41e62
KS		 969 if (unlikely(!pmd_same(*pmd, orig_pmd)))
970 goto out_free_page;
971
93b4796d
KS		 972 pmdp_clear_flush(vma, haddr, pmd);
973 /* leave pmd empty until pte is filled */
974
6b0b50b0 975 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
93b4796d
KS		 976 pmd_populate(mm, &_pmd, pgtable);
977
978 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
979 pte_t *pte, entry;
980 if (haddr == (address & PAGE_MASK)) {
981 entry = mk_pte(page, vma->vm_page_prot);
982 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
983 page_add_new_anon_rmap(page, vma, haddr);
984 } else {
985 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
986 entry = pte_mkspecial(entry);
987 }
988 pte = pte_offset_map(&_pmd, haddr);
989 VM_BUG_ON(!pte_none(*pte));
990 set_pte_at(mm, haddr, pte, entry);
991 pte_unmap(pte);
992 }
993 smp_wmb(); /* make pte visible before pmd */
994 pmd_populate(mm, pmd, pgtable);
c4088ebd 995 spin_unlock(ptl);
97ae1749 996 put_huge_zero_page();
93b4796d
KS		 997 inc_mm_counter(mm, MM_ANONPAGES);
998
999 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1000
1001 ret |= VM_FAULT_WRITE;
1002out:
1003 return ret;
3ea41e62 1004out_free_page:
c4088ebd 1005 spin_unlock(ptl);
3ea41e62
KS		 1006 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1007 mem_cgroup_uncharge_page(page);
1008 put_page(page);
1009 goto out;
93b4796d
KS		 1010}
1011
71e3aac0
AA		 1012static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1013 struct vm_area_struct *vma,
1014 unsigned long address,
1015 pmd_t *pmd, pmd_t orig_pmd,
1016 struct page *page,
1017 unsigned long haddr)
1018{
c4088ebd 1019 spinlock_t *ptl;
71e3aac0
AA		 1020 pgtable_t pgtable;
1021 pmd_t _pmd;
1022 int ret = 0, i;
1023 struct page **pages;
2ec74c3e
SG		 1024 unsigned long mmun_start; /* For mmu_notifiers */
1025 unsigned long mmun_end; /* For mmu_notifiers */
71e3aac0
AA		 1026
1027 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1028 GFP_KERNEL);
1029 if (unlikely(!pages)) {
1030 ret |= VM_FAULT_OOM;
1031 goto out;
1032 }
1033
1034 for (i = 0; i < HPAGE_PMD_NR; i++) {
cc5d462f
AK		 1035 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1036 __GFP_OTHER_NODE,
19ee151e 1037 vma, address, page_to_nid(page));
b9bbfbe3
AA		 1038 if (unlikely(!pages[i] ||
1039 mem_cgroup_newpage_charge(pages[i], mm,
1040 GFP_KERNEL))) {
1041 if (pages[i])
71e3aac0 1042 put_page(pages[i]);
b9bbfbe3
AA		 1043 mem_cgroup_uncharge_start();
1044 while (--i >= 0) {
1045 mem_cgroup_uncharge_page(pages[i]);
1046 put_page(pages[i]);
1047 }
1048 mem_cgroup_uncharge_end();
71e3aac0
AA		 1049 kfree(pages);
1050 ret |= VM_FAULT_OOM;
1051 goto out;
1052 }
1053 }
1054
1055 for (i = 0; i < HPAGE_PMD_NR; i++) {
1056 copy_user_highpage(pages[i], page + i,
0089e485 1057 haddr + PAGE_SIZE * i, vma);
71e3aac0
AA		 1058 __SetPageUptodate(pages[i]);
1059 cond_resched();
1060 }
1061
2ec74c3e
SG		 1062 mmun_start = haddr;
1063 mmun_end = haddr + HPAGE_PMD_SIZE;
1064 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1065
c4088ebd 1066 ptl = pmd_lock(mm, pmd);
71e3aac0
AA		 1067 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1068 goto out_free_pages;
1069 VM_BUG_ON(!PageHead(page));
1070
2ec74c3e 1071 pmdp_clear_flush(vma, haddr, pmd);
71e3aac0
AA		 1072 /* leave pmd empty until pte is filled */
1073
6b0b50b0 1074 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
71e3aac0
AA		 1075 pmd_populate(mm, &_pmd, pgtable);
1076
1077 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1078 pte_t *pte, entry;
1079 entry = mk_pte(pages[i], vma->vm_page_prot);
1080 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1081 page_add_new_anon_rmap(pages[i], vma, haddr);
1082 pte = pte_offset_map(&_pmd, haddr);
1083 VM_BUG_ON(!pte_none(*pte));
1084 set_pte_at(mm, haddr, pte, entry);
1085 pte_unmap(pte);
1086 }
1087 kfree(pages);
1088
71e3aac0
AA		 1089 smp_wmb(); /* make pte visible before pmd */
1090 pmd_populate(mm, pmd, pgtable);
1091 page_remove_rmap(page);
c4088ebd 1092 spin_unlock(ptl);
71e3aac0 1093
2ec74c3e
SG		 1094 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1095
71e3aac0
AA		 1096 ret |= VM_FAULT_WRITE;
1097 put_page(page);
1098
1099out:
1100 return ret;
1101
1102out_free_pages:
c4088ebd 1103 spin_unlock(ptl);
2ec74c3e 1104 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
b9bbfbe3
AA		 1105 mem_cgroup_uncharge_start();
1106 for (i = 0; i < HPAGE_PMD_NR; i++) {
1107 mem_cgroup_uncharge_page(pages[i]);
71e3aac0 1108 put_page(pages[i]);
b9bbfbe3
AA		 1109 }
1110 mem_cgroup_uncharge_end();
71e3aac0
AA		 1111 kfree(pages);
1112 goto out;
1113}
1114
1115int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1116 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1117{
c4088ebd 1118 spinlock_t *ptl;
71e3aac0 1119 int ret = 0;
93b4796d 1120 struct page *page = NULL, *new_page;
71e3aac0 1121 unsigned long haddr;
2ec74c3e
SG		 1122 unsigned long mmun_start; /* For mmu_notifiers */
1123 unsigned long mmun_end; /* For mmu_notifiers */
71e3aac0 1124
c4088ebd 1125 ptl = pmd_lockptr(mm, pmd);
71e3aac0 1126 VM_BUG_ON(!vma->anon_vma);
93b4796d
KS		 1127 haddr = address & HPAGE_PMD_MASK;
1128 if (is_huge_zero_pmd(orig_pmd))
1129 goto alloc;
c4088ebd 1130 spin_lock(ptl);
71e3aac0
AA		 1131 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1132 goto out_unlock;
1133
1134 page = pmd_page(orig_pmd);
1135 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
71e3aac0
AA		 1136 if (page_mapcount(page) == 1) {
1137 pmd_t entry;
1138 entry = pmd_mkyoung(orig_pmd);
1139 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1140 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
b113da65 1141 update_mmu_cache_pmd(vma, address, pmd);
71e3aac0
AA		 1142 ret |= VM_FAULT_WRITE;
1143 goto out_unlock;
1144 }
1145 get_page(page);
c4088ebd 1146 spin_unlock(ptl);
93b4796d 1147alloc:
71e3aac0
AA		 1148 if (transparent_hugepage_enabled(vma) &&
1149 !transparent_hugepage_debug_cow())
0bbbc0b3 1150 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
cc5d462f 1151 vma, haddr, numa_node_id(), 0);
71e3aac0
AA		 1152 else
1153 new_page = NULL;
1154
1155 if (unlikely(!new_page)) {
93b4796d
KS		 1156 if (is_huge_zero_pmd(orig_pmd)) {
1157 ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
3ea41e62 1158 address, pmd, orig_pmd, haddr);
93b4796d
KS		 1159 } else {
1160 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1161 pmd, orig_pmd, page, haddr);
1162 if (ret & VM_FAULT_OOM)
1163 split_huge_page(page);
1164 put_page(page);
1165 }
17766dde 1166 count_vm_event(THP_FAULT_FALLBACK);
71e3aac0
AA		 1167 goto out;
1168 }
1169
b9bbfbe3
AA		 1170 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1171 put_page(new_page);
93b4796d
KS		 1172 if (page) {
1173 split_huge_page(page);
1174 put_page(page);
1175 }
17766dde 1176 count_vm_event(THP_FAULT_FALLBACK);
b9bbfbe3
AA		 1177 ret |= VM_FAULT_OOM;
1178 goto out;
1179 }
1180
17766dde
DR		 1181 count_vm_event(THP_FAULT_ALLOC);
1182
93b4796d
KS		 1183 if (is_huge_zero_pmd(orig_pmd))
1184 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1185 else
1186 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
71e3aac0
AA		 1187 __SetPageUptodate(new_page);
1188
2ec74c3e
SG		 1189 mmun_start = haddr;
1190 mmun_end = haddr + HPAGE_PMD_SIZE;
1191 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1192
c4088ebd 1193 spin_lock(ptl);
93b4796d
KS		 1194 if (page)
1195 put_page(page);
b9bbfbe3 1196 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
c4088ebd 1197 spin_unlock(ptl);
b9bbfbe3 1198 mem_cgroup_uncharge_page(new_page);
71e3aac0 1199 put_page(new_page);
2ec74c3e 1200 goto out_mn;
b9bbfbe3 1201 } else {
71e3aac0 1202 pmd_t entry;
3122359a
KS		 1203 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1204 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2ec74c3e 1205 pmdp_clear_flush(vma, haddr, pmd);
71e3aac0
AA		 1206 page_add_new_anon_rmap(new_page, vma, haddr);
1207 set_pmd_at(mm, haddr, pmd, entry);
b113da65 1208 update_mmu_cache_pmd(vma, address, pmd);
97ae1749 1209 if (is_huge_zero_pmd(orig_pmd)) {
93b4796d 1210 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
97ae1749
KS		 1211 put_huge_zero_page();
1212 } else {
93b4796d
KS		 1213 VM_BUG_ON(!PageHead(page));
1214 page_remove_rmap(page);
1215 put_page(page);
1216 }
71e3aac0
AA		 1217 ret |= VM_FAULT_WRITE;
1218 }
c4088ebd 1219 spin_unlock(ptl);
2ec74c3e
SG		 1220out_mn:
1221 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 1222out:
1223 return ret;
2ec74c3e 1224out_unlock:
c4088ebd 1225 spin_unlock(ptl);
2ec74c3e 1226 return ret;
71e3aac0
AA		 1227}
1228
b676b293 1229struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
71e3aac0
AA		 1230 unsigned long addr,
1231 pmd_t *pmd,
1232 unsigned int flags)
1233{
b676b293 1234 struct mm_struct *mm = vma->vm_mm;
71e3aac0
AA		 1235 struct page *page = NULL;
1236
c4088ebd 1237 assert_spin_locked(pmd_lockptr(mm, pmd));
71e3aac0
AA		 1238
1239 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1240 goto out;
1241
85facf25
KS		 1242 /* Avoid dumping huge zero page */
1243 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1244 return ERR_PTR(-EFAULT);
1245
71e3aac0
AA		 1246 page = pmd_page(*pmd);
1247 VM_BUG_ON(!PageHead(page));
1248 if (flags & FOLL_TOUCH) {
1249 pmd_t _pmd;
1250 /*
1251 * We should set the dirty bit only for FOLL_WRITE but
1252 * for now the dirty bit in the pmd is meaningless.
1253 * And if the dirty bit will become meaningful and
1254 * we'll only set it with FOLL_WRITE, an atomic
1255 * set_bit will be required on the pmd to set the
1256 * young bit, instead of the current set_pmd_at.
1257 */
1258 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
8663890a
AK		 1259 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1260 pmd, _pmd, 1))
1261 update_mmu_cache_pmd(vma, addr, pmd);
71e3aac0 1262 }
b676b293
DR		 1263 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1264 if (page->mapping && trylock_page(page)) {
1265 lru_add_drain();
1266 if (page->mapping)
1267 mlock_vma_page(page);
1268 unlock_page(page);
1269 }
1270 }
71e3aac0
AA		 1271 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1272 VM_BUG_ON(!PageCompound(page));
1273 if (flags & FOLL_GET)
70b50f94 1274 get_page_foll(page);
71e3aac0
AA		 1275
1276out:
1277 return page;
1278}
1279
d10e63f2 1280/* NUMA hinting page fault entry point for trans huge pmds */
4daae3b4
MG		 1281int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1282 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
d10e63f2 1283{
c4088ebd 1284 spinlock_t *ptl;
b8916634 1285 struct anon_vma *anon_vma = NULL;
b32967ff 1286 struct page *page;
d10e63f2 1287 unsigned long haddr = addr & HPAGE_PMD_MASK;
8191acbd 1288 int page_nid = -1, this_nid = numa_node_id();
90572890 1289 int target_nid, last_cpupid = -1;
8191acbd
MG		 1290 bool page_locked;
1291 bool migrated = false;
6688cc05 1292 int flags = 0;
d10e63f2 1293
c4088ebd 1294 ptl = pmd_lock(mm, pmdp);
d10e63f2
MG		 1295 if (unlikely(!pmd_same(pmd, *pmdp)))
1296 goto out_unlock;
1297
1298 page = pmd_page(pmd);
a1a46184 1299 BUG_ON(is_huge_zero_page(page));
8191acbd 1300 page_nid = page_to_nid(page);
90572890 1301 last_cpupid = page_cpupid_last(page);
03c5a6e1 1302 count_vm_numa_event(NUMA_HINT_FAULTS);
04bb2f94 1303 if (page_nid == this_nid) {
03c5a6e1 1304 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
04bb2f94
RR		 1305 flags |= TNF_FAULT_LOCAL;
1306 }
4daae3b4 1307
6688cc05
PZ		 1308 /*
1309 * Avoid grouping on DSO/COW pages in specific and RO pages
1310 * in general, RO pages shouldn't hurt as much anyway since
1311 * they can be in shared cache state.
1312 */
1313 if (!pmd_write(pmd))
1314 flags |= TNF_NO_GROUP;
1315
ff9042b1
MG		 1316 /*
1317 * Acquire the page lock to serialise THP migrations but avoid dropping
1318 * page_table_lock if at all possible
1319 */
b8916634
MG		 1320 page_locked = trylock_page(page);
1321 target_nid = mpol_misplaced(page, vma, haddr);
1322 if (target_nid == -1) {
1323 /* If the page was locked, there are no parallel migrations */
a54a407f 1324 if (page_locked)
b8916634 1325 goto clear_pmdnuma;
4daae3b4 1326
a54a407f
MG		 1327 /*
1328 * Otherwise wait for potential migrations and retry. We do
1329 * relock and check_same as the page may no longer be mapped.
1330 * As the fault is being retried, do not account for it.
1331 */
c4088ebd 1332 spin_unlock(ptl);
b8916634 1333 wait_on_page_locked(page);
a54a407f 1334 page_nid = -1;
b8916634
MG		 1335 goto out;
1336 }
1337
1338 /* Page is misplaced, serialise migrations and parallel THP splits */
1339 get_page(page);
c4088ebd 1340 spin_unlock(ptl);
a54a407f 1341 if (!page_locked)
b8916634 1342 lock_page(page);
b8916634 1343 anon_vma = page_lock_anon_vma_read(page);
4daae3b4 1344
c69307d5 1345 /* Confirm the PMD did not change while page_table_lock was released */
c4088ebd 1346 spin_lock(ptl);
b32967ff
MG		 1347 if (unlikely(!pmd_same(pmd, *pmdp))) {
1348 unlock_page(page);
1349 put_page(page);
a54a407f 1350 page_nid = -1;
4daae3b4 1351 goto out_unlock;
b32967ff 1352 }
ff9042b1 1353
a54a407f
MG		 1354 /*
1355 * Migrate the THP to the requested node, returns with page unlocked
1356 * and pmd_numa cleared.
1357 */
c4088ebd 1358 spin_unlock(ptl);
b32967ff 1359 migrated = migrate_misplaced_transhuge_page(mm, vma,
340ef390 1360 pmdp, pmd, addr, page, target_nid);
6688cc05
PZ		 1361 if (migrated) {
1362 flags |= TNF_MIGRATED;
8191acbd 1363 page_nid = target_nid;
6688cc05 1364 }
b32967ff 1365
8191acbd 1366 goto out;
b32967ff 1367clear_pmdnuma:
a54a407f 1368 BUG_ON(!PageLocked(page));
d10e63f2
MG		 1369 pmd = pmd_mknonnuma(pmd);
1370 set_pmd_at(mm, haddr, pmdp, pmd);
1371 VM_BUG_ON(pmd_numa(*pmdp));
1372 update_mmu_cache_pmd(vma, addr, pmdp);
a54a407f 1373 unlock_page(page);
d10e63f2 1374out_unlock:
c4088ebd 1375 spin_unlock(ptl);
b8916634
MG		 1376
1377out:
1378 if (anon_vma)
1379 page_unlock_anon_vma_read(anon_vma);
1380
8191acbd 1381 if (page_nid != -1)
6688cc05 1382 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
8191acbd 1383
d10e63f2
MG		 1384 return 0;
1385}
1386
71e3aac0 1387int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
f21760b1 1388 pmd_t *pmd, unsigned long addr)
71e3aac0 1389{
bf929152 1390 spinlock_t *ptl;
71e3aac0
AA		 1391 int ret = 0;
1392
bf929152 1393 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
025c5b24
NH		 1394 struct page *page;
1395 pgtable_t pgtable;
f5c8ad47 1396 pmd_t orig_pmd;
a6bf2bb0
AK		 1397 /*
1398 * For architectures like ppc64 we look at deposited pgtable
1399 * when calling pmdp_get_and_clear. So do the
1400 * pgtable_trans_huge_withdraw after finishing pmdp related
1401 * operations.
1402 */
f5c8ad47 1403 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
025c5b24 1404 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
a6bf2bb0 1405 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
479f0abb 1406 if (is_huge_zero_pmd(orig_pmd)) {
e1f56c89 1407 atomic_long_dec(&tlb->mm->nr_ptes);
bf929152 1408 spin_unlock(ptl);
97ae1749 1409 put_huge_zero_page();
479f0abb
KS		 1410 } else {
1411 page = pmd_page(orig_pmd);
1412 page_remove_rmap(page);
1413 VM_BUG_ON(page_mapcount(page) < 0);
1414 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1415 VM_BUG_ON(!PageHead(page));
e1f56c89 1416 atomic_long_dec(&tlb->mm->nr_ptes);
bf929152 1417 spin_unlock(ptl);
479f0abb
KS		 1418 tlb_remove_page(tlb, page);
1419 }
025c5b24
NH		 1420 pte_free(tlb->mm, pgtable);
1421 ret = 1;
1422 }
71e3aac0
AA		 1423 return ret;
1424}
1425
0ca1634d
JW		 1426int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1427 unsigned long addr, unsigned long end,
1428 unsigned char *vec)
1429{
bf929152 1430 spinlock_t *ptl;
0ca1634d
JW		 1431 int ret = 0;
1432
bf929152 1433 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
025c5b24
NH		 1434 /*
1435 * All logical pages in the range are present
1436 * if backed by a huge page.
1437 */
bf929152 1438 spin_unlock(ptl);
025c5b24
NH		 1439 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1440 ret = 1;
1441 }
0ca1634d
JW		 1442
1443 return ret;
1444}
1445
37a1c49a
AA		 1446int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1447 unsigned long old_addr,
1448 unsigned long new_addr, unsigned long old_end,
1449 pmd_t *old_pmd, pmd_t *new_pmd)
1450{
bf929152 1451 spinlock_t *old_ptl, *new_ptl;
37a1c49a
AA		 1452 int ret = 0;
1453 pmd_t pmd;
1454
1455 struct mm_struct *mm = vma->vm_mm;
1456
1457 if ((old_addr & ~HPAGE_PMD_MASK) ||
1458 (new_addr & ~HPAGE_PMD_MASK) ||
1459 old_end - old_addr < HPAGE_PMD_SIZE ||
1460 (new_vma->vm_flags & VM_NOHUGEPAGE))
1461 goto out;
1462
1463 /*
1464 * The destination pmd shouldn't be established, free_pgtables()
1465 * should have release it.
1466 */
1467 if (WARN_ON(!pmd_none(*new_pmd))) {
1468 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1469 goto out;
1470 }
1471
bf929152
KS		 1472 /*
1473 * We don't have to worry about the ordering of src and dst
1474 * ptlocks because exclusive mmap_sem prevents deadlock.
1475 */
1476 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
025c5b24 1477 if (ret == 1) {
bf929152
KS		 1478 new_ptl = pmd_lockptr(mm, new_pmd);
1479 if (new_ptl != old_ptl)
1480 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
025c5b24
NH		 1481 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1482 VM_BUG_ON(!pmd_none(*new_pmd));
0f8975ec 1483 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
3592806c
KS		 1484 if (new_ptl != old_ptl) {
1485 pgtable_t pgtable;
1486
1487 /*
1488 * Move preallocated PTE page table if new_pmd is on
1489 * different PMD page table.
1490 */
1491 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1492 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1493
bf929152 1494 spin_unlock(new_ptl);
3592806c 1495 }
bf929152 1496 spin_unlock(old_ptl);
37a1c49a
AA		 1497 }
1498out:
1499 return ret;
1500}
1501
f123d74a
MG		 1502/*
1503 * Returns
1504 * - 0 if PMD could not be locked
1505 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1506 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1507 */
cd7548ab 1508int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
4b10e7d5 1509 unsigned long addr, pgprot_t newprot, int prot_numa)
cd7548ab
JW		 1510{
1511 struct mm_struct *mm = vma->vm_mm;
bf929152 1512 spinlock_t *ptl;
cd7548ab
JW		 1513 int ret = 0;
1514
bf929152 1515 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
025c5b24 1516 pmd_t entry;
f123d74a 1517 ret = 1;
a4f1de17 1518 if (!prot_numa) {
f123d74a 1519 entry = pmdp_get_and_clear(mm, addr, pmd);
4b10e7d5 1520 entry = pmd_modify(entry, newprot);
f123d74a 1521 ret = HPAGE_PMD_NR;
a4f1de17
HD		 1522 BUG_ON(pmd_write(entry));
1523 } else {
4b10e7d5
MG		 1524 struct page *page = pmd_page(*pmd);
1525
a1a46184 1526 /*
1bc115d8
MG		 1527 * Do not trap faults against the zero page. The
1528 * read-only data is likely to be read-cached on the
1529 * local CPU cache and it is less useful to know about
1530 * local vs remote hits on the zero page.
a1a46184 1531 */
1bc115d8 1532 if (!is_huge_zero_page(page) &&
4b10e7d5 1533 !pmd_numa(*pmd)) {
f123d74a 1534 entry = pmdp_get_and_clear(mm, addr, pmd);
4b10e7d5 1535 entry = pmd_mknuma(entry);
f123d74a 1536 ret = HPAGE_PMD_NR;
4b10e7d5
MG		 1537 }
1538 }
f123d74a
MG		 1539
1540 /* Set PMD if cleared earlier */
1541 if (ret == HPAGE_PMD_NR)
1542 set_pmd_at(mm, addr, pmd, entry);
1543
bf929152 1544 spin_unlock(ptl);
025c5b24
NH		 1545 }
1546
1547 return ret;
1548}
1549
1550/*
1551 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1552 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1553 *
1554 * Note that if it returns 1, this routine returns without unlocking page
1555 * table locks. So callers must unlock them.
1556 */
bf929152
KS		 1557int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1558 spinlock_t **ptl)
025c5b24 1559{
bf929152 1560 *ptl = pmd_lock(vma->vm_mm, pmd);
cd7548ab
JW		 1561 if (likely(pmd_trans_huge(*pmd))) {
1562 if (unlikely(pmd_trans_splitting(*pmd))) {
bf929152 1563 spin_unlock(*ptl);
cd7548ab 1564 wait_split_huge_page(vma->anon_vma, pmd);
025c5b24 1565 return -1;
cd7548ab 1566 } else {
025c5b24
NH		 1567 /* Thp mapped by 'pmd' is stable, so we can
1568 * handle it as it is. */
1569 return 1;
cd7548ab 1570 }
025c5b24 1571 }
bf929152 1572 spin_unlock(*ptl);
025c5b24 1573 return 0;
cd7548ab
JW		 1574}
1575
117b0791
KS		 1576/*
1577 * This function returns whether a given @page is mapped onto the @address
1578 * in the virtual space of @mm.
1579 *
1580 * When it's true, this function returns *pmd with holding the page table lock
1581 * and passing it back to the caller via @ptl.
1582 * If it's false, returns NULL without holding the page table lock.
1583 */
71e3aac0
AA		 1584pmd_t *page_check_address_pmd(struct page *page,
1585 struct mm_struct *mm,
1586 unsigned long address,
117b0791
KS		 1587 enum page_check_address_pmd_flag flag,
1588 spinlock_t **ptl)
71e3aac0 1589{
117b0791 1590 pmd_t *pmd;
71e3aac0
AA		 1591
1592 if (address & ~HPAGE_PMD_MASK)
117b0791 1593 return NULL;
71e3aac0 1594
6219049a
BL		 1595 pmd = mm_find_pmd(mm, address);
1596 if (!pmd)
117b0791
KS		 1597 return NULL;
1598 *ptl = pmd_lock(mm, pmd);
71e3aac0 1599 if (pmd_none(*pmd))
117b0791 1600 goto unlock;
71e3aac0 1601 if (pmd_page(*pmd) != page)
117b0791 1602 goto unlock;
94fcc585
AA		 1603 /*
1604 * split_vma() may create temporary aliased mappings. There is
1605 * no risk as long as all huge pmd are found and have their
1606 * splitting bit set before __split_huge_page_refcount
1607 * runs. Finding the same huge pmd more than once during the
1608 * same rmap walk is not a problem.
1609 */
1610 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1611 pmd_trans_splitting(*pmd))
117b0791 1612 goto unlock;
71e3aac0
AA		 1613 if (pmd_trans_huge(*pmd)) {
1614 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1615 !pmd_trans_splitting(*pmd));
117b0791 1616 return pmd;
71e3aac0 1617 }
117b0791
KS		 1618unlock:
1619 spin_unlock(*ptl);
1620 return NULL;
71e3aac0
AA		 1621}
1622
1623static int __split_huge_page_splitting(struct page *page,
1624 struct vm_area_struct *vma,
1625 unsigned long address)
1626{
1627 struct mm_struct *mm = vma->vm_mm;
117b0791 1628 spinlock_t *ptl;
71e3aac0
AA		 1629 pmd_t *pmd;
1630 int ret = 0;
2ec74c3e
SG		 1631 /* For mmu_notifiers */
1632 const unsigned long mmun_start = address;
1633 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
71e3aac0 1634
2ec74c3e 1635 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
71e3aac0 1636 pmd = page_check_address_pmd(page, mm, address,
117b0791 1637 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
71e3aac0
AA		 1638 if (pmd) {
1639 /*
1640 * We can't temporarily set the pmd to null in order
1641 * to split it, the pmd must remain marked huge at all
1642 * times or the VM won't take the pmd_trans_huge paths
5a505085 1643 * and it won't wait on the anon_vma->root->rwsem to
71e3aac0
AA		 1644 * serialize against split_huge_page*.
1645 */
2ec74c3e 1646 pmdp_splitting_flush(vma, address, pmd);
71e3aac0 1647 ret = 1;
117b0791 1648 spin_unlock(ptl);
71e3aac0 1649 }
2ec74c3e 1650 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 1651
1652 return ret;
1653}
1654
5bc7b8ac
SL		 1655static void __split_huge_page_refcount(struct page *page,
1656 struct list_head *list)
71e3aac0
AA		 1657{
1658 int i;
71e3aac0 1659 struct zone *zone = page_zone(page);
fa9add64 1660 struct lruvec *lruvec;
70b50f94 1661 int tail_count = 0;
71e3aac0
AA		 1662
1663 /* prevent PageLRU to go away from under us, and freeze lru stats */
1664 spin_lock_irq(&zone->lru_lock);
fa9add64
HD		 1665 lruvec = mem_cgroup_page_lruvec(page, zone);
1666
71e3aac0 1667 compound_lock(page);
e94c8a9c
KH		 1668 /* complete memcg works before add pages to LRU */
1669 mem_cgroup_split_huge_fixup(page);
71e3aac0 1670
45676885 1671 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
71e3aac0
AA		 1672 struct page *page_tail = page + i;
1673
70b50f94
AA		 1674 /* tail_page->_mapcount cannot change */
1675 BUG_ON(page_mapcount(page_tail) < 0);
1676 tail_count += page_mapcount(page_tail);
1677 /* check for overflow */
1678 BUG_ON(tail_count < 0);
1679 BUG_ON(atomic_read(&page_tail->_count) != 0);
1680 /*
1681 * tail_page->_count is zero and not changing from
1682 * under us. But get_page_unless_zero() may be running
1683 * from under us on the tail_page. If we used
1684 * atomic_set() below instead of atomic_add(), we
1685 * would then run atomic_set() concurrently with
1686 * get_page_unless_zero(), and atomic_set() is
1687 * implemented in C not using locked ops. spin_unlock
1688 * on x86 sometime uses locked ops because of PPro
1689 * errata 66, 92, so unless somebody can guarantee
1690 * atomic_set() here would be safe on all archs (and
1691 * not only on x86), it's safer to use atomic_add().
1692 */
1693 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1694 &page_tail->_count);
71e3aac0
AA		 1695
1696 /* after clearing PageTail the gup refcount can be released */
1697 smp_mb();
1698
a6d30ddd
JD		 1699 /*
1700 * retain hwpoison flag of the poisoned tail page:
1701 * fix for the unsuitable process killed on Guest Machine(KVM)
1702 * by the memory-failure.
1703 */
1704 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
71e3aac0
AA		 1705 page_tail->flags |= (page->flags &
1706 ((1L << PG_referenced) |
1707 (1L << PG_swapbacked) |
1708 (1L << PG_mlocked) |
e180cf80
KS		 1709 (1L << PG_uptodate) |
1710 (1L << PG_active) |
1711 (1L << PG_unevictable)));
71e3aac0
AA		 1712 page_tail->flags |= (1L << PG_dirty);
1713
70b50f94 1714 /* clear PageTail before overwriting first_page */
71e3aac0
AA		 1715 smp_wmb();
1716
1717 /*
1718 * __split_huge_page_splitting() already set the
1719 * splitting bit in all pmd that could map this
1720 * hugepage, that will ensure no CPU can alter the
1721 * mapcount on the head page. The mapcount is only
1722 * accounted in the head page and it has to be
1723 * transferred to all tail pages in the below code. So
1724 * for this code to be safe, the split the mapcount
1725 * can't change. But that doesn't mean userland can't
1726 * keep changing and reading the page contents while
1727 * we transfer the mapcount, so the pmd splitting
1728 * status is achieved setting a reserved bit in the
1729 * pmd, not by clearing the present bit.
1730 */
71e3aac0
AA		 1731 page_tail->_mapcount = page->_mapcount;
1732
1733 BUG_ON(page_tail->mapping);
1734 page_tail->mapping = page->mapping;
1735
45676885 1736 page_tail->index = page->index + i;
90572890 1737 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
71e3aac0
AA		 1738
1739 BUG_ON(!PageAnon(page_tail));
1740 BUG_ON(!PageUptodate(page_tail));
1741 BUG_ON(!PageDirty(page_tail));
1742 BUG_ON(!PageSwapBacked(page_tail));
1743
5bc7b8ac 1744 lru_add_page_tail(page, page_tail, lruvec, list);
71e3aac0 1745 }
70b50f94
AA		 1746 atomic_sub(tail_count, &page->_count);
1747 BUG_ON(atomic_read(&page->_count) <= 0);
71e3aac0 1748
fa9add64 1749 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
79134171 1750
71e3aac0
AA		 1751 ClearPageCompound(page);
1752 compound_unlock(page);
1753 spin_unlock_irq(&zone->lru_lock);
1754
1755 for (i = 1; i < HPAGE_PMD_NR; i++) {
1756 struct page *page_tail = page + i;
1757 BUG_ON(page_count(page_tail) <= 0);
1758 /*
1759 * Tail pages may be freed if there wasn't any mapping
1760 * like if add_to_swap() is running on a lru page that
1761 * had its mapping zapped. And freeing these pages
1762 * requires taking the lru_lock so we do the put_page
1763 * of the tail pages after the split is complete.
1764 */
1765 put_page(page_tail);
1766 }
1767
1768 /*
1769 * Only the head page (now become a regular page) is required
1770 * to be pinned by the caller.
1771 */
1772 BUG_ON(page_count(page) <= 0);
1773}
1774
1775static int __split_huge_page_map(struct page *page,
1776 struct vm_area_struct *vma,
1777 unsigned long address)
1778{
1779 struct mm_struct *mm = vma->vm_mm;
117b0791 1780 spinlock_t *ptl;
71e3aac0
AA		 1781 pmd_t *pmd, _pmd;
1782 int ret = 0, i;
1783 pgtable_t pgtable;
1784 unsigned long haddr;
1785
71e3aac0 1786 pmd = page_check_address_pmd(page, mm, address,
117b0791 1787 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
71e3aac0 1788 if (pmd) {
6b0b50b0 1789 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
71e3aac0
AA		 1790 pmd_populate(mm, &_pmd, pgtable);
1791
e3ebcf64
GS		 1792 haddr = address;
1793 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
71e3aac0
AA		 1794 pte_t *pte, entry;
1795 BUG_ON(PageCompound(page+i));
1796 entry = mk_pte(page + i, vma->vm_page_prot);
1797 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1798 if (!pmd_write(*pmd))
1799 entry = pte_wrprotect(entry);
1800 else
1801 BUG_ON(page_mapcount(page) != 1);
1802 if (!pmd_young(*pmd))
1803 entry = pte_mkold(entry);
1ba6e0b5
AA		 1804 if (pmd_numa(*pmd))
1805 entry = pte_mknuma(entry);
71e3aac0
AA		 1806 pte = pte_offset_map(&_pmd, haddr);
1807 BUG_ON(!pte_none(*pte));
1808 set_pte_at(mm, haddr, pte, entry);
1809 pte_unmap(pte);
1810 }
1811
71e3aac0
AA		 1812 smp_wmb(); /* make pte visible before pmd */
1813 /*
1814 * Up to this point the pmd is present and huge and
1815 * userland has the whole access to the hugepage
1816 * during the split (which happens in place). If we
1817 * overwrite the pmd with the not-huge version
1818 * pointing to the pte here (which of course we could
1819 * if all CPUs were bug free), userland could trigger
1820 * a small page size TLB miss on the small sized TLB
1821 * while the hugepage TLB entry is still established
1822 * in the huge TLB. Some CPU doesn't like that. See
1823 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1824 * Erratum 383 on page 93. Intel should be safe but is
1825 * also warns that it's only safe if the permission
1826 * and cache attributes of the two entries loaded in
1827 * the two TLB is identical (which should be the case
1828 * here). But it is generally safer to never allow
1829 * small and huge TLB entries for the same virtual
1830 * address to be loaded simultaneously. So instead of
1831 * doing "pmd_populate(); flush_tlb_range();" we first
1832 * mark the current pmd notpresent (atomically because
1833 * here the pmd_trans_huge and pmd_trans_splitting
1834 * must remain set at all times on the pmd until the
1835 * split is complete for this pmd), then we flush the
1836 * SMP TLB and finally we write the non-huge version
1837 * of the pmd entry with pmd_populate.
1838 */
46dcde73 1839 pmdp_invalidate(vma, address, pmd);
71e3aac0
AA		 1840 pmd_populate(mm, pmd, pgtable);
1841 ret = 1;
117b0791 1842 spin_unlock(ptl);
71e3aac0 1843 }
71e3aac0
AA		 1844
1845 return ret;
1846}
1847
5a505085 1848/* must be called with anon_vma->root->rwsem held */
71e3aac0 1849static void __split_huge_page(struct page *page,
5bc7b8ac
SL		 1850 struct anon_vma *anon_vma,
1851 struct list_head *list)
71e3aac0
AA		 1852{
1853 int mapcount, mapcount2;
bf181b9f 1854 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
71e3aac0
AA		 1855 struct anon_vma_chain *avc;
1856
1857 BUG_ON(!PageHead(page));
1858 BUG_ON(PageTail(page));
1859
1860 mapcount = 0;
bf181b9f 1861 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
71e3aac0
AA		 1862 struct vm_area_struct *vma = avc->vma;
1863 unsigned long addr = vma_address(page, vma);
1864 BUG_ON(is_vma_temporary_stack(vma));
71e3aac0
AA		 1865 mapcount += __split_huge_page_splitting(page, vma, addr);
1866 }
05759d38
AA		 1867 /*
1868 * It is critical that new vmas are added to the tail of the
1869 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1870 * and establishes a child pmd before
1871 * __split_huge_page_splitting() freezes the parent pmd (so if
1872 * we fail to prevent copy_huge_pmd() from running until the
1873 * whole __split_huge_page() is complete), we will still see
1874 * the newly established pmd of the child later during the
1875 * walk, to be able to set it as pmd_trans_splitting too.
1876 */
1877 if (mapcount != page_mapcount(page))
1878 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1879 mapcount, page_mapcount(page));
71e3aac0
AA		 1880 BUG_ON(mapcount != page_mapcount(page));
1881
5bc7b8ac 1882 __split_huge_page_refcount(page, list);
71e3aac0
AA		 1883
1884 mapcount2 = 0;
bf181b9f 1885 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
71e3aac0
AA		 1886 struct vm_area_struct *vma = avc->vma;
1887 unsigned long addr = vma_address(page, vma);
1888 BUG_ON(is_vma_temporary_stack(vma));
71e3aac0
AA		 1889 mapcount2 += __split_huge_page_map(page, vma, addr);
1890 }
05759d38
AA		 1891 if (mapcount != mapcount2)
1892 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1893 mapcount, mapcount2, page_mapcount(page));
71e3aac0
AA		 1894 BUG_ON(mapcount != mapcount2);
1895}
1896
5bc7b8ac
SL		 1897/*
1898 * Split a hugepage into normal pages. This doesn't change the position of head
1899 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1900 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1901 * from the hugepage.
1902 * Return 0 if the hugepage is split successfully otherwise return 1.
1903 */
1904int split_huge_page_to_list(struct page *page, struct list_head *list)
71e3aac0
AA		 1905{
1906 struct anon_vma *anon_vma;
1907 int ret = 1;
1908
5918d10a 1909 BUG_ON(is_huge_zero_page(page));
71e3aac0 1910 BUG_ON(!PageAnon(page));
062f1af2
MG		 1911
1912 /*
1913 * The caller does not necessarily hold an mmap_sem that would prevent
1914 * the anon_vma disappearing so we first we take a reference to it
1915 * and then lock the anon_vma for write. This is similar to
1916 * page_lock_anon_vma_read except the write lock is taken to serialise
1917 * against parallel split or collapse operations.
1918 */
1919 anon_vma = page_get_anon_vma(page);
71e3aac0
AA		 1920 if (!anon_vma)
1921 goto out;
062f1af2
MG		 1922 anon_vma_lock_write(anon_vma);
1923
71e3aac0
AA		 1924 ret = 0;
1925 if (!PageCompound(page))
1926 goto out_unlock;
1927
1928 BUG_ON(!PageSwapBacked(page));
5bc7b8ac 1929 __split_huge_page(page, anon_vma, list);
81ab4201 1930 count_vm_event(THP_SPLIT);
71e3aac0
AA		 1931
1932 BUG_ON(PageCompound(page));
1933out_unlock:
08b52706 1934 anon_vma_unlock_write(anon_vma);
062f1af2 1935 put_anon_vma(anon_vma);
71e3aac0
AA		 1936out:
1937 return ret;
1938}
1939
4b6e1e37 1940#define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
78f11a25 1941
60ab3244
AA		 1942int hugepage_madvise(struct vm_area_struct *vma,
1943 unsigned long *vm_flags, int advice)
0af4e98b 1944{
8e72033f
GS		 1945 struct mm_struct *mm = vma->vm_mm;
1946
a664b2d8
AA		 1947 switch (advice) {
1948 case MADV_HUGEPAGE:
1949 /*
1950 * Be somewhat over-protective like KSM for now!
1951 */
78f11a25 1952 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
a664b2d8 1953 return -EINVAL;
8e72033f
GS		 1954 if (mm->def_flags & VM_NOHUGEPAGE)
1955 return -EINVAL;
a664b2d8
AA		 1956 *vm_flags &= ~VM_NOHUGEPAGE;
1957 *vm_flags |= VM_HUGEPAGE;
60ab3244
AA		 1958 /*
1959 * If the vma become good for khugepaged to scan,
1960 * register it here without waiting a page fault that
1961 * may not happen any time soon.
1962 */
1963 if (unlikely(khugepaged_enter_vma_merge(vma)))
1964 return -ENOMEM;
a664b2d8
AA		 1965 break;
1966 case MADV_NOHUGEPAGE:
1967 /*
1968 * Be somewhat over-protective like KSM for now!
1969 */
78f11a25 1970 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
a664b2d8
AA		 1971 return -EINVAL;
1972 *vm_flags &= ~VM_HUGEPAGE;
1973 *vm_flags |= VM_NOHUGEPAGE;
60ab3244
AA		 1974 /*
1975 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1976 * this vma even if we leave the mm registered in khugepaged if
1977 * it got registered before VM_NOHUGEPAGE was set.
1978 */
a664b2d8
AA		 1979 break;
1980 }
0af4e98b
AA		 1981
1982 return 0;
1983}
1984
ba76149f
AA		 1985static int __init khugepaged_slab_init(void)
1986{
1987 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1988 sizeof(struct mm_slot),
1989 __alignof__(struct mm_slot), 0, NULL);
1990 if (!mm_slot_cache)
1991 return -ENOMEM;
1992
1993 return 0;
1994}
1995
ba76149f
AA		 1996static inline struct mm_slot *alloc_mm_slot(void)
1997{
1998 if (!mm_slot_cache) /* initialization failed */
1999 return NULL;
2000 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2001}
2002
2003static inline void free_mm_slot(struct mm_slot *mm_slot)
2004{
2005 kmem_cache_free(mm_slot_cache, mm_slot);
2006}
2007
ba76149f
AA		 2008static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2009{
2010 struct mm_slot *mm_slot;
ba76149f 2011
b67bfe0d 2012 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
ba76149f
AA		 2013 if (mm == mm_slot->mm)
2014 return mm_slot;
43b5fbbd 2015
ba76149f
AA		 2016 return NULL;
2017}
2018
2019static void insert_to_mm_slots_hash(struct mm_struct *mm,
2020 struct mm_slot *mm_slot)
2021{
ba76149f 2022 mm_slot->mm = mm;
43b5fbbd 2023 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
ba76149f
AA		 2024}
2025
2026static inline int khugepaged_test_exit(struct mm_struct *mm)
2027{
2028 return atomic_read(&mm->mm_users) == 0;
2029}
2030
2031int __khugepaged_enter(struct mm_struct *mm)
2032{
2033 struct mm_slot *mm_slot;
2034 int wakeup;
2035
2036 mm_slot = alloc_mm_slot();
2037 if (!mm_slot)
2038 return -ENOMEM;
2039
2040 /* __khugepaged_exit() must not run from under us */
2041 VM_BUG_ON(khugepaged_test_exit(mm));
2042 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2043 free_mm_slot(mm_slot);
2044 return 0;
2045 }
2046
2047 spin_lock(&khugepaged_mm_lock);
2048 insert_to_mm_slots_hash(mm, mm_slot);
2049 /*
2050 * Insert just behind the scanning cursor, to let the area settle
2051 * down a little.
2052 */
2053 wakeup = list_empty(&khugepaged_scan.mm_head);
2054 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2055 spin_unlock(&khugepaged_mm_lock);
2056
2057 atomic_inc(&mm->mm_count);
2058 if (wakeup)
2059 wake_up_interruptible(&khugepaged_wait);
2060
2061 return 0;
2062}
2063
2064int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2065{
2066 unsigned long hstart, hend;
2067 if (!vma->anon_vma)
2068 /*
2069 * Not yet faulted in so we will register later in the
2070 * page fault if needed.
2071 */
2072 return 0;
78f11a25 2073 if (vma->vm_ops)
ba76149f
AA		 2074 /* khugepaged not yet working on file or special mappings */
2075 return 0;
b3b9c293 2076 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
ba76149f
AA		 2077 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2078 hend = vma->vm_end & HPAGE_PMD_MASK;
2079 if (hstart < hend)
2080 return khugepaged_enter(vma);
2081 return 0;
2082}
2083
2084void __khugepaged_exit(struct mm_struct *mm)
2085{
2086 struct mm_slot *mm_slot;
2087 int free = 0;
2088
2089 spin_lock(&khugepaged_mm_lock);
2090 mm_slot = get_mm_slot(mm);
2091 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
43b5fbbd 2092 hash_del(&mm_slot->hash);
ba76149f
AA		 2093 list_del(&mm_slot->mm_node);
2094 free = 1;
2095 }
d788e80a 2096 spin_unlock(&khugepaged_mm_lock);
ba76149f
AA		 2097
2098 if (free) {
ba76149f
AA		 2099 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2100 free_mm_slot(mm_slot);
2101 mmdrop(mm);
2102 } else if (mm_slot) {
ba76149f
AA		 2103 /*
2104 * This is required to serialize against
2105 * khugepaged_test_exit() (which is guaranteed to run
2106 * under mmap sem read mode). Stop here (after we
2107 * return all pagetables will be destroyed) until
2108 * khugepaged has finished working on the pagetables
2109 * under the mmap_sem.
2110 */
2111 down_write(&mm->mmap_sem);
2112 up_write(&mm->mmap_sem);
d788e80a 2113 }
ba76149f
AA		 2114}
2115
2116static void release_pte_page(struct page *page)
2117{
2118 /* 0 stands for page_is_file_cache(page) == false */
2119 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2120 unlock_page(page);
2121 putback_lru_page(page);
2122}
2123
2124static void release_pte_pages(pte_t *pte, pte_t *_pte)
2125{
2126 while (--_pte >= pte) {
2127 pte_t pteval = *_pte;
2128 if (!pte_none(pteval))
2129 release_pte_page(pte_page(pteval));
2130 }
2131}
2132
ba76149f
AA		 2133static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2134 unsigned long address,
2135 pte_t *pte)
2136{
2137 struct page *page;
2138 pte_t *_pte;
344aa35c 2139 int referenced = 0, none = 0;
ba76149f
AA		 2140 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2141 _pte++, address += PAGE_SIZE) {
2142 pte_t pteval = *_pte;
2143 if (pte_none(pteval)) {
2144 if (++none <= khugepaged_max_ptes_none)
2145 continue;
344aa35c 2146 else
ba76149f 2147 goto out;
ba76149f 2148 }
344aa35c 2149 if (!pte_present(pteval) || !pte_write(pteval))
ba76149f 2150 goto out;
ba76149f 2151 page = vm_normal_page(vma, address, pteval);
344aa35c 2152 if (unlikely(!page))
ba76149f 2153 goto out;
344aa35c 2154
ba76149f
AA		 2155 VM_BUG_ON(PageCompound(page));
2156 BUG_ON(!PageAnon(page));
2157 VM_BUG_ON(!PageSwapBacked(page));
2158
2159 /* cannot use mapcount: can't collapse if there's a gup pin */
344aa35c 2160 if (page_count(page) != 1)
ba76149f 2161 goto out;
ba76149f
AA		 2162 /*
2163 * We can do it before isolate_lru_page because the
2164 * page can't be freed from under us. NOTE: PG_lock
2165 * is needed to serialize against split_huge_page
2166 * when invoked from the VM.
2167 */
344aa35c 2168 if (!trylock_page(page))
ba76149f 2169 goto out;
ba76149f
AA		 2170 /*
2171 * Isolate the page to avoid collapsing an hugepage
2172 * currently in use by the VM.
2173 */
2174 if (isolate_lru_page(page)) {
2175 unlock_page(page);
ba76149f
AA		 2176 goto out;
2177 }
2178 /* 0 stands for page_is_file_cache(page) == false */
2179 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2180 VM_BUG_ON(!PageLocked(page));
2181 VM_BUG_ON(PageLRU(page));
2182
2183 /* If there is no mapped pte young don't collapse the page */
8ee53820
AA		 2184 if (pte_young(pteval) || PageReferenced(page) ||
2185 mmu_notifier_test_young(vma->vm_mm, address))
ba76149f
AA		 2186 referenced = 1;
2187 }
344aa35c
BL		 2188 if (likely(referenced))
2189 return 1;
ba76149f 2190out:
344aa35c
BL		 2191 release_pte_pages(pte, _pte);
2192 return 0;
ba76149f
AA		 2193}
2194
2195static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2196 struct vm_area_struct *vma,
2197 unsigned long address,
2198 spinlock_t *ptl)
2199{
2200 pte_t *_pte;
2201 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2202 pte_t pteval = *_pte;
2203 struct page *src_page;
2204
2205 if (pte_none(pteval)) {
2206 clear_user_highpage(page, address);
2207 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2208 } else {
2209 src_page = pte_page(pteval);
2210 copy_user_highpage(page, src_page, address, vma);
2211 VM_BUG_ON(page_mapcount(src_page) != 1);
ba76149f
AA		 2212 release_pte_page(src_page);
2213 /*
2214 * ptl mostly unnecessary, but preempt has to
2215 * be disabled to update the per-cpu stats
2216 * inside page_remove_rmap().
2217 */
2218 spin_lock(ptl);
2219 /*
2220 * paravirt calls inside pte_clear here are
2221 * superfluous.
2222 */
2223 pte_clear(vma->vm_mm, address, _pte);
2224 page_remove_rmap(src_page);
2225 spin_unlock(ptl);
2226 free_page_and_swap_cache(src_page);
2227 }
2228
2229 address += PAGE_SIZE;
2230 page++;
2231 }
2232}
2233
26234f36 2234static void khugepaged_alloc_sleep(void)
ba76149f 2235{
26234f36
XG		 2236 wait_event_freezable_timeout(khugepaged_wait, false,
2237 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2238}
ba76149f 2239
9f1b868a
BL		 2240static int khugepaged_node_load[MAX_NUMNODES];
2241
26234f36 2242#ifdef CONFIG_NUMA
9f1b868a
BL		 2243static int khugepaged_find_target_node(void)
2244{
2245 static int last_khugepaged_target_node = NUMA_NO_NODE;
2246 int nid, target_node = 0, max_value = 0;
2247
2248 /* find first node with max normal pages hit */
2249 for (nid = 0; nid < MAX_NUMNODES; nid++)
2250 if (khugepaged_node_load[nid] > max_value) {
2251 max_value = khugepaged_node_load[nid];
2252 target_node = nid;
2253 }
2254
2255 /* do some balance if several nodes have the same hit record */
2256 if (target_node <= last_khugepaged_target_node)
2257 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2258 nid++)
2259 if (max_value == khugepaged_node_load[nid]) {
2260 target_node = nid;
2261 break;
2262 }
2263
2264 last_khugepaged_target_node = target_node;
2265 return target_node;
2266}
2267
26234f36
XG		 2268static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2269{
2270 if (IS_ERR(*hpage)) {
2271 if (!*wait)
2272 return false;
2273
2274 *wait = false;
e3b4126c 2275 *hpage = NULL;
26234f36
XG		 2276 khugepaged_alloc_sleep();
2277 } else if (*hpage) {
2278 put_page(*hpage);
2279 *hpage = NULL;
2280 }
2281
2282 return true;
2283}
2284
2285static struct page
2286*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2287 struct vm_area_struct *vma, unsigned long address,
2288 int node)
2289{
0bbbc0b3 2290 VM_BUG_ON(*hpage);
ce83d217
AA		 2291 /*
2292 * Allocate the page while the vma is still valid and under
2293 * the mmap_sem read mode so there is no memory allocation
2294 * later when we take the mmap_sem in write mode. This is more
2295 * friendly behavior (OTOH it may actually hide bugs) to
2296 * filesystems in userland with daemons allocating memory in
2297 * the userland I/O paths. Allocating memory with the
2298 * mmap_sem in read mode is good idea also to allow greater
2299 * scalability.
2300 */
9f1b868a
BL		 2301 *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2302 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
692e0b35
AA		 2303 /*
2304 * After allocating the hugepage, release the mmap_sem read lock in
2305 * preparation for taking it in write mode.
2306 */
2307 up_read(&mm->mmap_sem);
26234f36 2308 if (unlikely(!*hpage)) {
81ab4201 2309 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
ce83d217 2310 *hpage = ERR_PTR(-ENOMEM);
26234f36 2311 return NULL;
ce83d217 2312 }
26234f36 2313
65b3c07b 2314 count_vm_event(THP_COLLAPSE_ALLOC);
26234f36
XG		 2315 return *hpage;
2316}
2317#else
9f1b868a
BL		 2318static int khugepaged_find_target_node(void)
2319{
2320 return 0;
2321}
2322
10dc4155
BL		 2323static inline struct page *alloc_hugepage(int defrag)
2324{
2325 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2326 HPAGE_PMD_ORDER);
2327}
2328
26234f36
XG		 2329static struct page *khugepaged_alloc_hugepage(bool *wait)
2330{
2331 struct page *hpage;
2332
2333 do {
2334 hpage = alloc_hugepage(khugepaged_defrag());
2335 if (!hpage) {
2336 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2337 if (!*wait)
2338 return NULL;
2339
2340 *wait = false;
2341 khugepaged_alloc_sleep();
2342 } else
2343 count_vm_event(THP_COLLAPSE_ALLOC);
2344 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2345
2346 return hpage;
2347}
2348
2349static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2350{
2351 if (!*hpage)
2352 *hpage = khugepaged_alloc_hugepage(wait);
2353
2354 if (unlikely(!*hpage))
2355 return false;
2356
2357 return true;
2358}
2359
2360static struct page
2361*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2362 struct vm_area_struct *vma, unsigned long address,
2363 int node)
2364{
2365 up_read(&mm->mmap_sem);
2366 VM_BUG_ON(!*hpage);
2367 return *hpage;
2368}
692e0b35
AA		 2369#endif
2370
fa475e51
BL		 2371static bool hugepage_vma_check(struct vm_area_struct *vma)
2372{
2373 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2374 (vma->vm_flags & VM_NOHUGEPAGE))
2375 return false;
2376
2377 if (!vma->anon_vma || vma->vm_ops)
2378 return false;
2379 if (is_vma_temporary_stack(vma))
2380 return false;
2381 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2382 return true;
2383}
2384
26234f36
XG		 2385static void collapse_huge_page(struct mm_struct *mm,
2386 unsigned long address,
2387 struct page **hpage,
2388 struct vm_area_struct *vma,
2389 int node)
2390{
26234f36
XG		 2391 pmd_t *pmd, _pmd;
2392 pte_t *pte;
2393 pgtable_t pgtable;
2394 struct page *new_page;
c4088ebd 2395 spinlock_t *pmd_ptl, *pte_ptl;
26234f36
XG		 2396 int isolated;
2397 unsigned long hstart, hend;
2ec74c3e
SG		 2398 unsigned long mmun_start; /* For mmu_notifiers */
2399 unsigned long mmun_end; /* For mmu_notifiers */
26234f36
XG		 2400
2401 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2402
2403 /* release the mmap_sem read lock. */
2404 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2405 if (!new_page)
2406 return;
2407
420256ef 2408 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
ce83d217 2409 return;
ba76149f
AA		 2410
2411 /*
2412 * Prevent all access to pagetables with the exception of
2413 * gup_fast later hanlded by the ptep_clear_flush and the VM
2414 * handled by the anon_vma lock + PG_lock.
2415 */
2416 down_write(&mm->mmap_sem);
2417 if (unlikely(khugepaged_test_exit(mm)))
2418 goto out;
2419
2420 vma = find_vma(mm, address);
a8f531eb
L		 2421 if (!vma)
2422 goto out;
ba76149f
AA		 2423 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2424 hend = vma->vm_end & HPAGE_PMD_MASK;
2425 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2426 goto out;
fa475e51 2427 if (!hugepage_vma_check(vma))
a7d6e4ec 2428 goto out;
6219049a
BL		 2429 pmd = mm_find_pmd(mm, address);
2430 if (!pmd)
ba76149f 2431 goto out;
6219049a 2432 if (pmd_trans_huge(*pmd))
ba76149f
AA		 2433 goto out;
2434
4fc3f1d6 2435 anon_vma_lock_write(vma->anon_vma);
ba76149f
AA		 2436
2437 pte = pte_offset_map(pmd, address);
c4088ebd 2438 pte_ptl = pte_lockptr(mm, pmd);
ba76149f 2439
2ec74c3e
SG		 2440 mmun_start = address;
2441 mmun_end = address + HPAGE_PMD_SIZE;
2442 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
c4088ebd 2443 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
ba76149f
AA		 2444 /*
2445 * After this gup_fast can't run anymore. This also removes
2446 * any huge TLB entry from the CPU so we won't allow
2447 * huge and small TLB entries for the same virtual address
2448 * to avoid the risk of CPU bugs in that area.
2449 */
2ec74c3e 2450 _pmd = pmdp_clear_flush(vma, address, pmd);
c4088ebd 2451 spin_unlock(pmd_ptl);
2ec74c3e 2452 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
ba76149f 2453
c4088ebd 2454 spin_lock(pte_ptl);
ba76149f 2455 isolated = __collapse_huge_page_isolate(vma, address, pte);
c4088ebd 2456 spin_unlock(pte_ptl);
ba76149f
AA		 2457
2458 if (unlikely(!isolated)) {
453c7192 2459 pte_unmap(pte);
c4088ebd 2460 spin_lock(pmd_ptl);
ba76149f 2461 BUG_ON(!pmd_none(*pmd));
7c342512
AK		 2462 /*
2463 * We can only use set_pmd_at when establishing
2464 * hugepmds and never for establishing regular pmds that
2465 * points to regular pagetables. Use pmd_populate for that
2466 */
2467 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
c4088ebd 2468 spin_unlock(pmd_ptl);
08b52706 2469 anon_vma_unlock_write(vma->anon_vma);
ce83d217 2470 goto out;
ba76149f
AA		 2471 }
2472
2473 /*
2474 * All pages are isolated and locked so anon_vma rmap
2475 * can't run anymore.
2476 */
08b52706 2477 anon_vma_unlock_write(vma->anon_vma);
ba76149f 2478
c4088ebd 2479 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
453c7192 2480 pte_unmap(pte);
ba76149f
AA		 2481 __SetPageUptodate(new_page);
2482 pgtable = pmd_pgtable(_pmd);
ba76149f 2483
3122359a
KS		 2484 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2485 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
ba76149f
AA		 2486
2487 /*
2488 * spin_lock() below is not the equivalent of smp_wmb(), so
2489 * this is needed to avoid the copy_huge_page writes to become
2490 * visible after the set_pmd_at() write.
2491 */
2492 smp_wmb();
2493
c4088ebd 2494 spin_lock(pmd_ptl);
ba76149f
AA		 2495 BUG_ON(!pmd_none(*pmd));
2496 page_add_new_anon_rmap(new_page, vma, address);
fce144b4 2497 pgtable_trans_huge_deposit(mm, pmd, pgtable);
ba76149f 2498 set_pmd_at(mm, address, pmd, _pmd);
b113da65 2499 update_mmu_cache_pmd(vma, address, pmd);
c4088ebd 2500 spin_unlock(pmd_ptl);
ba76149f
AA		 2501
2502 *hpage = NULL;
420256ef 2503
ba76149f 2504 khugepaged_pages_collapsed++;
ce83d217 2505out_up_write:
ba76149f 2506 up_write(&mm->mmap_sem);
0bbbc0b3
AA		 2507 return;
2508
ce83d217 2509out:
678ff896 2510 mem_cgroup_uncharge_page(new_page);
ce83d217 2511 goto out_up_write;
ba76149f
AA		 2512}
2513
2514static int khugepaged_scan_pmd(struct mm_struct *mm,
2515 struct vm_area_struct *vma,
2516 unsigned long address,
2517 struct page **hpage)
2518{
ba76149f
AA		 2519 pmd_t *pmd;
2520 pte_t *pte, *_pte;
2521 int ret = 0, referenced = 0, none = 0;
2522 struct page *page;
2523 unsigned long _address;
2524 spinlock_t *ptl;
00ef2d2f 2525 int node = NUMA_NO_NODE;
ba76149f
AA		 2526
2527 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2528
6219049a
BL		 2529 pmd = mm_find_pmd(mm, address);
2530 if (!pmd)
ba76149f 2531 goto out;
6219049a 2532 if (pmd_trans_huge(*pmd))
ba76149f
AA		 2533 goto out;
2534
9f1b868a 2535 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
ba76149f
AA		 2536 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2537 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2538 _pte++, _address += PAGE_SIZE) {
2539 pte_t pteval = *_pte;
2540 if (pte_none(pteval)) {
2541 if (++none <= khugepaged_max_ptes_none)
2542 continue;
2543 else
2544 goto out_unmap;
2545 }
2546 if (!pte_present(pteval) || !pte_write(pteval))
2547 goto out_unmap;
2548 page = vm_normal_page(vma, _address, pteval);
2549 if (unlikely(!page))
2550 goto out_unmap;
5c4b4be3 2551 /*
9f1b868a
BL		 2552 * Record which node the original page is from and save this
2553 * information to khugepaged_node_load[].
2554 * Khupaged will allocate hugepage from the node has the max
2555 * hit record.
5c4b4be3 2556 */
9f1b868a
BL		 2557 node = page_to_nid(page);
2558 khugepaged_node_load[node]++;
ba76149f
AA		 2559 VM_BUG_ON(PageCompound(page));
2560 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2561 goto out_unmap;
2562 /* cannot use mapcount: can't collapse if there's a gup pin */
2563 if (page_count(page) != 1)
2564 goto out_unmap;
8ee53820
AA		 2565 if (pte_young(pteval) || PageReferenced(page) ||
2566 mmu_notifier_test_young(vma->vm_mm, address))
ba76149f
AA		 2567 referenced = 1;
2568 }
2569 if (referenced)
2570 ret = 1;
2571out_unmap:
2572 pte_unmap_unlock(pte, ptl);
9f1b868a
BL		 2573 if (ret) {
2574 node = khugepaged_find_target_node();
ce83d217 2575 /* collapse_huge_page will return with the mmap_sem released */
5c4b4be3 2576 collapse_huge_page(mm, address, hpage, vma, node);
9f1b868a 2577 }
ba76149f
AA		 2578out:
2579 return ret;
2580}
2581
2582static void collect_mm_slot(struct mm_slot *mm_slot)
2583{
2584 struct mm_struct *mm = mm_slot->mm;
2585
b9980cdc 2586 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
ba76149f
AA		 2587
2588 if (khugepaged_test_exit(mm)) {
2589 /* free mm_slot */
43b5fbbd 2590 hash_del(&mm_slot->hash);
ba76149f
AA		 2591 list_del(&mm_slot->mm_node);
2592
2593 /*
2594 * Not strictly needed because the mm exited already.
2595 *
2596 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2597 */
2598
2599 /* khugepaged_mm_lock actually not necessary for the below */
2600 free_mm_slot(mm_slot);
2601 mmdrop(mm);
2602 }
2603}
2604
2605static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2606 struct page **hpage)
2f1da642
HS		 2607 __releases(&khugepaged_mm_lock)
2608 __acquires(&khugepaged_mm_lock)
ba76149f
AA		 2609{
2610 struct mm_slot *mm_slot;
2611 struct mm_struct *mm;
2612 struct vm_area_struct *vma;
2613 int progress = 0;
2614
2615 VM_BUG_ON(!pages);
b9980cdc 2616 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
ba76149f
AA		 2617
2618 if (khugepaged_scan.mm_slot)
2619 mm_slot = khugepaged_scan.mm_slot;
2620 else {
2621 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2622 struct mm_slot, mm_node);
2623 khugepaged_scan.address = 0;
2624 khugepaged_scan.mm_slot = mm_slot;
2625 }
2626 spin_unlock(&khugepaged_mm_lock);
2627
2628 mm = mm_slot->mm;
2629 down_read(&mm->mmap_sem);
2630 if (unlikely(khugepaged_test_exit(mm)))
2631 vma = NULL;
2632 else
2633 vma = find_vma(mm, khugepaged_scan.address);
2634
2635 progress++;
2636 for (; vma; vma = vma->vm_next) {
2637 unsigned long hstart, hend;
2638
2639 cond_resched();
2640 if (unlikely(khugepaged_test_exit(mm))) {
2641 progress++;
2642 break;
2643 }
fa475e51
BL		 2644 if (!hugepage_vma_check(vma)) {
2645skip:
ba76149f
AA		 2646 progress++;
2647 continue;
2648 }
ba76149f
AA		 2649 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2650 hend = vma->vm_end & HPAGE_PMD_MASK;
a7d6e4ec
AA		 2651 if (hstart >= hend)
2652 goto skip;
2653 if (khugepaged_scan.address > hend)
2654 goto skip;
ba76149f
AA		 2655 if (khugepaged_scan.address < hstart)
2656 khugepaged_scan.address = hstart;
a7d6e4ec 2657 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
ba76149f
AA		 2658
2659 while (khugepaged_scan.address < hend) {
2660 int ret;
2661 cond_resched();
2662 if (unlikely(khugepaged_test_exit(mm)))
2663 goto breakouterloop;
2664
2665 VM_BUG_ON(khugepaged_scan.address < hstart ||
2666 khugepaged_scan.address + HPAGE_PMD_SIZE >
2667 hend);
2668 ret = khugepaged_scan_pmd(mm, vma,
2669 khugepaged_scan.address,
2670 hpage);
2671 /* move to next address */
2672 khugepaged_scan.address += HPAGE_PMD_SIZE;
2673 progress += HPAGE_PMD_NR;
2674 if (ret)
2675 /* we released mmap_sem so break loop */
2676 goto breakouterloop_mmap_sem;
2677 if (progress >= pages)
2678 goto breakouterloop;
2679 }
2680 }
2681breakouterloop:
2682 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2683breakouterloop_mmap_sem:
2684
2685 spin_lock(&khugepaged_mm_lock);
a7d6e4ec 2686 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
ba76149f
AA		 2687 /*
2688 * Release the current mm_slot if this mm is about to die, or
2689 * if we scanned all vmas of this mm.
2690 */
2691 if (khugepaged_test_exit(mm) || !vma) {
2692 /*
2693 * Make sure that if mm_users is reaching zero while
2694 * khugepaged runs here, khugepaged_exit will find
2695 * mm_slot not pointing to the exiting mm.
2696 */
2697 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2698 khugepaged_scan.mm_slot = list_entry(
2699 mm_slot->mm_node.next,
2700 struct mm_slot, mm_node);
2701 khugepaged_scan.address = 0;
2702 } else {
2703 khugepaged_scan.mm_slot = NULL;
2704 khugepaged_full_scans++;
2705 }
2706
2707 collect_mm_slot(mm_slot);
2708 }
2709
2710 return progress;
2711}
2712
2713static int khugepaged_has_work(void)
2714{
2715 return !list_empty(&khugepaged_scan.mm_head) &&
2716 khugepaged_enabled();
2717}
2718
2719static int khugepaged_wait_event(void)
2720{
2721 return !list_empty(&khugepaged_scan.mm_head) ||
2017c0bf 2722 kthread_should_stop();
ba76149f
AA		 2723}
2724
d516904b 2725static void khugepaged_do_scan(void)
ba76149f 2726{
d516904b 2727 struct page *hpage = NULL;
ba76149f
AA		 2728 unsigned int progress = 0, pass_through_head = 0;
2729 unsigned int pages = khugepaged_pages_to_scan;
d516904b 2730 bool wait = true;
ba76149f
AA		 2731
2732 barrier(); /* write khugepaged_pages_to_scan to local stack */
2733
2734 while (progress < pages) {
26234f36 2735 if (!khugepaged_prealloc_page(&hpage, &wait))
d516904b 2736 break;
26234f36 2737
420256ef 2738 cond_resched();
ba76149f 2739
878aee7d
AA		 2740 if (unlikely(kthread_should_stop() || freezing(current)))
2741 break;
2742
ba76149f
AA		 2743 spin_lock(&khugepaged_mm_lock);
2744 if (!khugepaged_scan.mm_slot)
2745 pass_through_head++;
2746 if (khugepaged_has_work() &&
2747 pass_through_head < 2)
2748 progress += khugepaged_scan_mm_slot(pages - progress,
d516904b 2749 &hpage);
ba76149f
AA		 2750 else
2751 progress = pages;
2752 spin_unlock(&khugepaged_mm_lock);
2753 }
ba76149f 2754
d516904b
XG		 2755 if (!IS_ERR_OR_NULL(hpage))
2756 put_page(hpage);
0bbbc0b3
AA		 2757}
2758
2017c0bf
XG		 2759static void khugepaged_wait_work(void)
2760{
2761 try_to_freeze();
2762
2763 if (khugepaged_has_work()) {
2764 if (!khugepaged_scan_sleep_millisecs)
2765 return;
2766
2767 wait_event_freezable_timeout(khugepaged_wait,
2768 kthread_should_stop(),
2769 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2770 return;
2771 }
2772
2773 if (khugepaged_enabled())
2774 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2775}
2776
ba76149f
AA		 2777static int khugepaged(void *none)
2778{
2779 struct mm_slot *mm_slot;
2780
878aee7d 2781 set_freezable();
ba76149f
AA		 2782 set_user_nice(current, 19);
2783
b7231789
XG		 2784 while (!kthread_should_stop()) {
2785 khugepaged_do_scan();
2786 khugepaged_wait_work();
2787 }
ba76149f
AA		 2788
2789 spin_lock(&khugepaged_mm_lock);
2790 mm_slot = khugepaged_scan.mm_slot;
2791 khugepaged_scan.mm_slot = NULL;
2792 if (mm_slot)
2793 collect_mm_slot(mm_slot);
2794 spin_unlock(&khugepaged_mm_lock);
ba76149f
AA		 2795 return 0;
2796}
2797
c5a647d0
KS		 2798static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2799 unsigned long haddr, pmd_t *pmd)
2800{
2801 struct mm_struct *mm = vma->vm_mm;
2802 pgtable_t pgtable;
2803 pmd_t _pmd;
2804 int i;
2805
2806 pmdp_clear_flush(vma, haddr, pmd);
2807 /* leave pmd empty until pte is filled */
2808
6b0b50b0 2809 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
c5a647d0
KS		 2810 pmd_populate(mm, &_pmd, pgtable);
2811
2812 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2813 pte_t *pte, entry;
2814 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2815 entry = pte_mkspecial(entry);
2816 pte = pte_offset_map(&_pmd, haddr);
2817 VM_BUG_ON(!pte_none(*pte));
2818 set_pte_at(mm, haddr, pte, entry);
2819 pte_unmap(pte);
2820 }
2821 smp_wmb(); /* make pte visible before pmd */
2822 pmd_populate(mm, pmd, pgtable);
97ae1749 2823 put_huge_zero_page();
c5a647d0
KS		 2824}
2825
e180377f
KS		 2826void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2827 pmd_t *pmd)
71e3aac0 2828{
c4088ebd 2829 spinlock_t *ptl;
71e3aac0 2830 struct page *page;
e180377f 2831 struct mm_struct *mm = vma->vm_mm;
c5a647d0
KS		 2832 unsigned long haddr = address & HPAGE_PMD_MASK;
2833 unsigned long mmun_start; /* For mmu_notifiers */
2834 unsigned long mmun_end; /* For mmu_notifiers */
e180377f
KS		 2835
2836 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
71e3aac0 2837
c5a647d0
KS		 2838 mmun_start = haddr;
2839 mmun_end = haddr + HPAGE_PMD_SIZE;
750e8165 2840again:
c5a647d0 2841 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
c4088ebd 2842 ptl = pmd_lock(mm, pmd);
71e3aac0 2843 if (unlikely(!pmd_trans_huge(*pmd))) {
c4088ebd 2844 spin_unlock(ptl);
c5a647d0
KS		 2845 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2846 return;
2847 }
2848 if (is_huge_zero_pmd(*pmd)) {
2849 __split_huge_zero_page_pmd(vma, haddr, pmd);
c4088ebd 2850 spin_unlock(ptl);
c5a647d0 2851 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 2852 return;
2853 }
2854 page = pmd_page(*pmd);
2855 VM_BUG_ON(!page_count(page));
2856 get_page(page);
c4088ebd 2857 spin_unlock(ptl);
c5a647d0 2858 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
71e3aac0
AA		 2859
2860 split_huge_page(page);
2861
2862 put_page(page);
750e8165
HD		 2863
2864 /*
2865 * We don't always have down_write of mmap_sem here: a racing
2866 * do_huge_pmd_wp_page() might have copied-on-write to another
2867 * huge page before our split_huge_page() got the anon_vma lock.
2868 */
2869 if (unlikely(pmd_trans_huge(*pmd)))
2870 goto again;
71e3aac0 2871}
94fcc585 2872
e180377f
KS		 2873void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2874 pmd_t *pmd)
2875{
2876 struct vm_area_struct *vma;
2877
2878 vma = find_vma(mm, address);
2879 BUG_ON(vma == NULL);
2880 split_huge_page_pmd(vma, address, pmd);
2881}
2882
94fcc585
AA		 2883static void split_huge_page_address(struct mm_struct *mm,
2884 unsigned long address)
2885{
94fcc585
AA		 2886 pmd_t *pmd;
2887
2888 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2889
6219049a
BL		 2890 pmd = mm_find_pmd(mm, address);
2891 if (!pmd)
94fcc585
AA		 2892 return;
2893 /*
2894 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2895 * materialize from under us.
2896 */
e180377f 2897 split_huge_page_pmd_mm(mm, address, pmd);
94fcc585
AA		 2898}
2899
2900void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2901 unsigned long start,
2902 unsigned long end,
2903 long adjust_next)
2904{
2905 /*
2906 * If the new start address isn't hpage aligned and it could
2907 * previously contain an hugepage: check if we need to split
2908 * an huge pmd.
2909 */
2910 if (start & ~HPAGE_PMD_MASK &&
2911 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2912 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2913 split_huge_page_address(vma->vm_mm, start);
2914
2915 /*
2916 * If the new end address isn't hpage aligned and it could
2917 * previously contain an hugepage: check if we need to split
2918 * an huge pmd.
2919 */
2920 if (end & ~HPAGE_PMD_MASK &&
2921 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2922 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2923 split_huge_page_address(vma->vm_mm, end);
2924
2925 /*
2926 * If we're also updating the vma->vm_next->vm_start, if the new
2927 * vm_next->vm_start isn't page aligned and it could previously
2928 * contain an hugepage: check if we need to split an huge pmd.
2929 */
2930 if (adjust_next > 0) {
2931 struct vm_area_struct *next = vma->vm_next;
2932 unsigned long nstart = next->vm_start;
2933 nstart += adjust_next << PAGE_SHIFT;
2934 if (nstart & ~HPAGE_PMD_MASK &&
2935 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2936 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2937 split_huge_page_address(next->vm_mm, nstart);
2938 }
2939}
Linux 6.x block layer and io_uring tree(s)