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