b5783d81eda90fc9a808478ac6dcea4f74b4424e
[linux-3.10.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
24 #include <asm/tlb.h>
25 #include <asm/pgalloc.h>
26 #include "internal.h"
27
28 /*
29  * By default transparent hugepage support is enabled for all mappings
30  * and khugepaged scans all mappings. Defrag is only invoked by
31  * khugepaged hugepage allocations and by page faults inside
32  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
33  * allocations.
34  */
35 unsigned long transparent_hugepage_flags __read_mostly =
36 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
37         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
38 #endif
39 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
40         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
41 #endif
42         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
43         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
44         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
45
46 /* default scan 8*512 pte (or vmas) every 30 second */
47 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
48 static unsigned int khugepaged_pages_collapsed;
49 static unsigned int khugepaged_full_scans;
50 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
51 /* during fragmentation poll the hugepage allocator once every minute */
52 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
53 static struct task_struct *khugepaged_thread __read_mostly;
54 static DEFINE_MUTEX(khugepaged_mutex);
55 static DEFINE_SPINLOCK(khugepaged_mm_lock);
56 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
57 /*
58  * default collapse hugepages if there is at least one pte mapped like
59  * it would have happened if the vma was large enough during page
60  * fault.
61  */
62 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
63
64 static int khugepaged(void *none);
65 static int mm_slots_hash_init(void);
66 static int khugepaged_slab_init(void);
67 static void khugepaged_slab_free(void);
68
69 #define MM_SLOTS_HASH_HEADS 1024
70 static struct hlist_head *mm_slots_hash __read_mostly;
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         extern int min_free_kbytes;
109
110         if (!khugepaged_enabled())
111                 return 0;
112
113         for_each_populated_zone(zone)
114                 nr_zones++;
115
116         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
117         recommended_min = pageblock_nr_pages * nr_zones * 2;
118
119         /*
120          * Make sure that on average at least two pageblocks are almost free
121          * of another type, one for a migratetype to fall back to and a
122          * second to avoid subsequent fallbacks of other types There are 3
123          * MIGRATE_TYPES we care about.
124          */
125         recommended_min += pageblock_nr_pages * nr_zones *
126                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
127
128         /* don't ever allow to reserve more than 5% of the lowmem */
129         recommended_min = min(recommended_min,
130                               (unsigned long) nr_free_buffer_pages() / 20);
131         recommended_min <<= (PAGE_SHIFT-10);
132
133         if (recommended_min > min_free_kbytes)
134                 min_free_kbytes = recommended_min;
135         setup_per_zone_wmarks();
136         return 0;
137 }
138 late_initcall(set_recommended_min_free_kbytes);
139
140 static int start_khugepaged(void)
141 {
142         int err = 0;
143         if (khugepaged_enabled()) {
144                 if (!khugepaged_thread)
145                         khugepaged_thread = kthread_run(khugepaged, NULL,
146                                                         "khugepaged");
147                 if (unlikely(IS_ERR(khugepaged_thread))) {
148                         printk(KERN_ERR
149                                "khugepaged: kthread_run(khugepaged) failed\n");
150                         err = PTR_ERR(khugepaged_thread);
151                         khugepaged_thread = NULL;
152                 }
153
154                 if (!list_empty(&khugepaged_scan.mm_head))
155                         wake_up_interruptible(&khugepaged_wait);
156
157                 set_recommended_min_free_kbytes();
158         } else if (khugepaged_thread) {
159                 kthread_stop(khugepaged_thread);
160                 khugepaged_thread = NULL;
161         }
162
163         return err;
164 }
165
166 static atomic_t huge_zero_refcount;
167 static unsigned long huge_zero_pfn __read_mostly;
168
169 static inline bool is_huge_zero_pfn(unsigned long pfn)
170 {
171         unsigned long zero_pfn = ACCESS_ONCE(huge_zero_pfn);
172         return zero_pfn && pfn == zero_pfn;
173 }
174
175 static inline bool is_huge_zero_pmd(pmd_t pmd)
176 {
177         return is_huge_zero_pfn(pmd_pfn(pmd));
178 }
179
180 static unsigned long get_huge_zero_page(void)
181 {
182         struct page *zero_page;
183 retry:
184         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
185                 return ACCESS_ONCE(huge_zero_pfn);
186
187         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
188                         HPAGE_PMD_ORDER);
189         if (!zero_page) {
190                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
191                 return 0;
192         }
193         count_vm_event(THP_ZERO_PAGE_ALLOC);
194         preempt_disable();
195         if (cmpxchg(&huge_zero_pfn, 0, page_to_pfn(zero_page))) {
196                 preempt_enable();
197                 __free_page(zero_page);
198                 goto retry;
199         }
200
201         /* We take additional reference here. It will be put back by shrinker */
202         atomic_set(&huge_zero_refcount, 2);
203         preempt_enable();
204         return ACCESS_ONCE(huge_zero_pfn);
205 }
206
207 static void put_huge_zero_page(void)
208 {
209         /*
210          * Counter should never go to zero here. Only shrinker can put
211          * last reference.
212          */
213         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
214 }
215
216 static int shrink_huge_zero_page(struct shrinker *shrink,
217                 struct shrink_control *sc)
218 {
219         if (!sc->nr_to_scan)
220                 /* we can free zero page only if last reference remains */
221                 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
222
223         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
224                 unsigned long zero_pfn = xchg(&huge_zero_pfn, 0);
225                 BUG_ON(zero_pfn == 0);
226                 __free_page(__pfn_to_page(zero_pfn));
227         }
228
229         return 0;
230 }
231
232 static struct shrinker huge_zero_page_shrinker = {
233         .shrink = shrink_huge_zero_page,
234         .seeks = DEFAULT_SEEKS,
235 };
236
237 #ifdef CONFIG_SYSFS
238
239 static ssize_t double_flag_show(struct kobject *kobj,
240                                 struct kobj_attribute *attr, char *buf,
241                                 enum transparent_hugepage_flag enabled,
242                                 enum transparent_hugepage_flag req_madv)
243 {
244         if (test_bit(enabled, &transparent_hugepage_flags)) {
245                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
246                 return sprintf(buf, "[always] madvise never\n");
247         } else if (test_bit(req_madv, &transparent_hugepage_flags))
248                 return sprintf(buf, "always [madvise] never\n");
249         else
250                 return sprintf(buf, "always madvise [never]\n");
251 }
252 static ssize_t double_flag_store(struct kobject *kobj,
253                                  struct kobj_attribute *attr,
254                                  const char *buf, size_t count,
255                                  enum transparent_hugepage_flag enabled,
256                                  enum transparent_hugepage_flag req_madv)
257 {
258         if (!memcmp("always", buf,
259                     min(sizeof("always")-1, count))) {
260                 set_bit(enabled, &transparent_hugepage_flags);
261                 clear_bit(req_madv, &transparent_hugepage_flags);
262         } else if (!memcmp("madvise", buf,
263                            min(sizeof("madvise")-1, count))) {
264                 clear_bit(enabled, &transparent_hugepage_flags);
265                 set_bit(req_madv, &transparent_hugepage_flags);
266         } else if (!memcmp("never", buf,
267                            min(sizeof("never")-1, count))) {
268                 clear_bit(enabled, &transparent_hugepage_flags);
269                 clear_bit(req_madv, &transparent_hugepage_flags);
270         } else
271                 return -EINVAL;
272
273         return count;
274 }
275
276 static ssize_t enabled_show(struct kobject *kobj,
277                             struct kobj_attribute *attr, char *buf)
278 {
279         return double_flag_show(kobj, attr, buf,
280                                 TRANSPARENT_HUGEPAGE_FLAG,
281                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
282 }
283 static ssize_t enabled_store(struct kobject *kobj,
284                              struct kobj_attribute *attr,
285                              const char *buf, size_t count)
286 {
287         ssize_t ret;
288
289         ret = double_flag_store(kobj, attr, buf, count,
290                                 TRANSPARENT_HUGEPAGE_FLAG,
291                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
292
293         if (ret > 0) {
294                 int err;
295
296                 mutex_lock(&khugepaged_mutex);
297                 err = start_khugepaged();
298                 mutex_unlock(&khugepaged_mutex);
299
300                 if (err)
301                         ret = err;
302         }
303
304         return ret;
305 }
306 static struct kobj_attribute enabled_attr =
307         __ATTR(enabled, 0644, enabled_show, enabled_store);
308
309 static ssize_t single_flag_show(struct kobject *kobj,
310                                 struct kobj_attribute *attr, char *buf,
311                                 enum transparent_hugepage_flag flag)
312 {
313         return sprintf(buf, "%d\n",
314                        !!test_bit(flag, &transparent_hugepage_flags));
315 }
316
317 static ssize_t single_flag_store(struct kobject *kobj,
318                                  struct kobj_attribute *attr,
319                                  const char *buf, size_t count,
320                                  enum transparent_hugepage_flag flag)
321 {
322         unsigned long value;
323         int ret;
324
325         ret = kstrtoul(buf, 10, &value);
326         if (ret < 0)
327                 return ret;
328         if (value > 1)
329                 return -EINVAL;
330
331         if (value)
332                 set_bit(flag, &transparent_hugepage_flags);
333         else
334                 clear_bit(flag, &transparent_hugepage_flags);
335
336         return count;
337 }
338
339 /*
340  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
341  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
342  * memory just to allocate one more hugepage.
343  */
344 static ssize_t defrag_show(struct kobject *kobj,
345                            struct kobj_attribute *attr, char *buf)
346 {
347         return double_flag_show(kobj, attr, buf,
348                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
349                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
350 }
351 static ssize_t defrag_store(struct kobject *kobj,
352                             struct kobj_attribute *attr,
353                             const char *buf, size_t count)
354 {
355         return double_flag_store(kobj, attr, buf, count,
356                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
357                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
358 }
359 static struct kobj_attribute defrag_attr =
360         __ATTR(defrag, 0644, defrag_show, defrag_store);
361
362 static ssize_t use_zero_page_show(struct kobject *kobj,
363                 struct kobj_attribute *attr, char *buf)
364 {
365         return single_flag_show(kobj, attr, buf,
366                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
367 }
368 static ssize_t use_zero_page_store(struct kobject *kobj,
369                 struct kobj_attribute *attr, const char *buf, size_t count)
370 {
371         return single_flag_store(kobj, attr, buf, count,
372                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
373 }
374 static struct kobj_attribute use_zero_page_attr =
375         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
376 #ifdef CONFIG_DEBUG_VM
377 static ssize_t debug_cow_show(struct kobject *kobj,
378                                 struct kobj_attribute *attr, char *buf)
379 {
380         return single_flag_show(kobj, attr, buf,
381                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
382 }
383 static ssize_t debug_cow_store(struct kobject *kobj,
384                                struct kobj_attribute *attr,
385                                const char *buf, size_t count)
386 {
387         return single_flag_store(kobj, attr, buf, count,
388                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
389 }
390 static struct kobj_attribute debug_cow_attr =
391         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
392 #endif /* CONFIG_DEBUG_VM */
393
394 static struct attribute *hugepage_attr[] = {
395         &enabled_attr.attr,
396         &defrag_attr.attr,
397         &use_zero_page_attr.attr,
398 #ifdef CONFIG_DEBUG_VM
399         &debug_cow_attr.attr,
400 #endif
401         NULL,
402 };
403
404 static struct attribute_group hugepage_attr_group = {
405         .attrs = hugepage_attr,
406 };
407
408 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
409                                          struct kobj_attribute *attr,
410                                          char *buf)
411 {
412         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
413 }
414
415 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
416                                           struct kobj_attribute *attr,
417                                           const char *buf, size_t count)
418 {
419         unsigned long msecs;
420         int err;
421
422         err = strict_strtoul(buf, 10, &msecs);
423         if (err || msecs > UINT_MAX)
424                 return -EINVAL;
425
426         khugepaged_scan_sleep_millisecs = msecs;
427         wake_up_interruptible(&khugepaged_wait);
428
429         return count;
430 }
431 static struct kobj_attribute scan_sleep_millisecs_attr =
432         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
433                scan_sleep_millisecs_store);
434
435 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
436                                           struct kobj_attribute *attr,
437                                           char *buf)
438 {
439         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
440 }
441
442 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
443                                            struct kobj_attribute *attr,
444                                            const char *buf, size_t count)
445 {
446         unsigned long msecs;
447         int err;
448
449         err = strict_strtoul(buf, 10, &msecs);
450         if (err || msecs > UINT_MAX)
451                 return -EINVAL;
452
453         khugepaged_alloc_sleep_millisecs = msecs;
454         wake_up_interruptible(&khugepaged_wait);
455
456         return count;
457 }
458 static struct kobj_attribute alloc_sleep_millisecs_attr =
459         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
460                alloc_sleep_millisecs_store);
461
462 static ssize_t pages_to_scan_show(struct kobject *kobj,
463                                   struct kobj_attribute *attr,
464                                   char *buf)
465 {
466         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
467 }
468 static ssize_t pages_to_scan_store(struct kobject *kobj,
469                                    struct kobj_attribute *attr,
470                                    const char *buf, size_t count)
471 {
472         int err;
473         unsigned long pages;
474
475         err = strict_strtoul(buf, 10, &pages);
476         if (err || !pages || pages > UINT_MAX)
477                 return -EINVAL;
478
479         khugepaged_pages_to_scan = pages;
480
481         return count;
482 }
483 static struct kobj_attribute pages_to_scan_attr =
484         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
485                pages_to_scan_store);
486
487 static ssize_t pages_collapsed_show(struct kobject *kobj,
488                                     struct kobj_attribute *attr,
489                                     char *buf)
490 {
491         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
492 }
493 static struct kobj_attribute pages_collapsed_attr =
494         __ATTR_RO(pages_collapsed);
495
496 static ssize_t full_scans_show(struct kobject *kobj,
497                                struct kobj_attribute *attr,
498                                char *buf)
499 {
500         return sprintf(buf, "%u\n", khugepaged_full_scans);
501 }
502 static struct kobj_attribute full_scans_attr =
503         __ATTR_RO(full_scans);
504
505 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
506                                       struct kobj_attribute *attr, char *buf)
507 {
508         return single_flag_show(kobj, attr, buf,
509                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
510 }
511 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
512                                        struct kobj_attribute *attr,
513                                        const char *buf, size_t count)
514 {
515         return single_flag_store(kobj, attr, buf, count,
516                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
517 }
518 static struct kobj_attribute khugepaged_defrag_attr =
519         __ATTR(defrag, 0644, khugepaged_defrag_show,
520                khugepaged_defrag_store);
521
522 /*
523  * max_ptes_none controls if khugepaged should collapse hugepages over
524  * any unmapped ptes in turn potentially increasing the memory
525  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
526  * reduce the available free memory in the system as it
527  * runs. Increasing max_ptes_none will instead potentially reduce the
528  * free memory in the system during the khugepaged scan.
529  */
530 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
531                                              struct kobj_attribute *attr,
532                                              char *buf)
533 {
534         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
535 }
536 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
537                                               struct kobj_attribute *attr,
538                                               const char *buf, size_t count)
539 {
540         int err;
541         unsigned long max_ptes_none;
542
543         err = strict_strtoul(buf, 10, &max_ptes_none);
544         if (err || max_ptes_none > HPAGE_PMD_NR-1)
545                 return -EINVAL;
546
547         khugepaged_max_ptes_none = max_ptes_none;
548
549         return count;
550 }
551 static struct kobj_attribute khugepaged_max_ptes_none_attr =
552         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
553                khugepaged_max_ptes_none_store);
554
555 static struct attribute *khugepaged_attr[] = {
556         &khugepaged_defrag_attr.attr,
557         &khugepaged_max_ptes_none_attr.attr,
558         &pages_to_scan_attr.attr,
559         &pages_collapsed_attr.attr,
560         &full_scans_attr.attr,
561         &scan_sleep_millisecs_attr.attr,
562         &alloc_sleep_millisecs_attr.attr,
563         NULL,
564 };
565
566 static struct attribute_group khugepaged_attr_group = {
567         .attrs = khugepaged_attr,
568         .name = "khugepaged",
569 };
570
571 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
572 {
573         int err;
574
575         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
576         if (unlikely(!*hugepage_kobj)) {
577                 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
578                 return -ENOMEM;
579         }
580
581         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
582         if (err) {
583                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
584                 goto delete_obj;
585         }
586
587         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
588         if (err) {
589                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
590                 goto remove_hp_group;
591         }
592
593         return 0;
594
595 remove_hp_group:
596         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
597 delete_obj:
598         kobject_put(*hugepage_kobj);
599         return err;
600 }
601
602 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
603 {
604         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
605         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
606         kobject_put(hugepage_kobj);
607 }
608 #else
609 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
610 {
611         return 0;
612 }
613
614 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
615 {
616 }
617 #endif /* CONFIG_SYSFS */
618
619 static int __init hugepage_init(void)
620 {
621         int err;
622         struct kobject *hugepage_kobj;
623
624         if (!has_transparent_hugepage()) {
625                 transparent_hugepage_flags = 0;
626                 return -EINVAL;
627         }
628
629         err = hugepage_init_sysfs(&hugepage_kobj);
630         if (err)
631                 return err;
632
633         err = khugepaged_slab_init();
634         if (err)
635                 goto out;
636
637         err = mm_slots_hash_init();
638         if (err) {
639                 khugepaged_slab_free();
640                 goto out;
641         }
642
643         register_shrinker(&huge_zero_page_shrinker);
644
645         /*
646          * By default disable transparent hugepages on smaller systems,
647          * where the extra memory used could hurt more than TLB overhead
648          * is likely to save.  The admin can still enable it through /sys.
649          */
650         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
651                 transparent_hugepage_flags = 0;
652
653         start_khugepaged();
654
655         return 0;
656 out:
657         hugepage_exit_sysfs(hugepage_kobj);
658         return err;
659 }
660 module_init(hugepage_init)
661
662 static int __init setup_transparent_hugepage(char *str)
663 {
664         int ret = 0;
665         if (!str)
666                 goto out;
667         if (!strcmp(str, "always")) {
668                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
669                         &transparent_hugepage_flags);
670                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
671                           &transparent_hugepage_flags);
672                 ret = 1;
673         } else if (!strcmp(str, "madvise")) {
674                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
675                           &transparent_hugepage_flags);
676                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
677                         &transparent_hugepage_flags);
678                 ret = 1;
679         } else if (!strcmp(str, "never")) {
680                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
681                           &transparent_hugepage_flags);
682                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
683                           &transparent_hugepage_flags);
684                 ret = 1;
685         }
686 out:
687         if (!ret)
688                 printk(KERN_WARNING
689                        "transparent_hugepage= cannot parse, ignored\n");
690         return ret;
691 }
692 __setup("transparent_hugepage=", setup_transparent_hugepage);
693
694 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
695 {
696         if (likely(vma->vm_flags & VM_WRITE))
697                 pmd = pmd_mkwrite(pmd);
698         return pmd;
699 }
700
701 static inline pmd_t mk_huge_pmd(struct page *page, struct vm_area_struct *vma)
702 {
703         pmd_t entry;
704         entry = mk_pmd(page, vma->vm_page_prot);
705         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
706         entry = pmd_mkhuge(entry);
707         return entry;
708 }
709
710 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
711                                         struct vm_area_struct *vma,
712                                         unsigned long haddr, pmd_t *pmd,
713                                         struct page *page)
714 {
715         pgtable_t pgtable;
716
717         VM_BUG_ON(!PageCompound(page));
718         pgtable = pte_alloc_one(mm, haddr);
719         if (unlikely(!pgtable))
720                 return VM_FAULT_OOM;
721
722         clear_huge_page(page, haddr, HPAGE_PMD_NR);
723         __SetPageUptodate(page);
724
725         spin_lock(&mm->page_table_lock);
726         if (unlikely(!pmd_none(*pmd))) {
727                 spin_unlock(&mm->page_table_lock);
728                 mem_cgroup_uncharge_page(page);
729                 put_page(page);
730                 pte_free(mm, pgtable);
731         } else {
732                 pmd_t entry;
733                 entry = mk_huge_pmd(page, vma);
734                 /*
735                  * The spinlocking to take the lru_lock inside
736                  * page_add_new_anon_rmap() acts as a full memory
737                  * barrier to be sure clear_huge_page writes become
738                  * visible after the set_pmd_at() write.
739                  */
740                 page_add_new_anon_rmap(page, vma, haddr);
741                 set_pmd_at(mm, haddr, pmd, entry);
742                 pgtable_trans_huge_deposit(mm, pgtable);
743                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
744                 mm->nr_ptes++;
745                 spin_unlock(&mm->page_table_lock);
746         }
747
748         return 0;
749 }
750
751 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
752 {
753         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
754 }
755
756 static inline struct page *alloc_hugepage_vma(int defrag,
757                                               struct vm_area_struct *vma,
758                                               unsigned long haddr, int nd,
759                                               gfp_t extra_gfp)
760 {
761         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
762                                HPAGE_PMD_ORDER, vma, haddr, nd);
763 }
764
765 #ifndef CONFIG_NUMA
766 static inline struct page *alloc_hugepage(int defrag)
767 {
768         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
769                            HPAGE_PMD_ORDER);
770 }
771 #endif
772
773 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
774                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
775                 unsigned long zero_pfn)
776 {
777         pmd_t entry;
778         if (!pmd_none(*pmd))
779                 return false;
780         entry = pfn_pmd(zero_pfn, vma->vm_page_prot);
781         entry = pmd_wrprotect(entry);
782         entry = pmd_mkhuge(entry);
783         set_pmd_at(mm, haddr, pmd, entry);
784         pgtable_trans_huge_deposit(mm, pgtable);
785         mm->nr_ptes++;
786         return true;
787 }
788
789 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
790                                unsigned long address, pmd_t *pmd,
791                                unsigned int flags)
792 {
793         struct page *page;
794         unsigned long haddr = address & HPAGE_PMD_MASK;
795         pte_t *pte;
796
797         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
798                 if (unlikely(anon_vma_prepare(vma)))
799                         return VM_FAULT_OOM;
800                 if (unlikely(khugepaged_enter(vma)))
801                         return VM_FAULT_OOM;
802                 if (!(flags & FAULT_FLAG_WRITE) &&
803                                 transparent_hugepage_use_zero_page()) {
804                         pgtable_t pgtable;
805                         unsigned long zero_pfn;
806                         bool set;
807                         pgtable = pte_alloc_one(mm, haddr);
808                         if (unlikely(!pgtable))
809                                 return VM_FAULT_OOM;
810                         zero_pfn = get_huge_zero_page();
811                         if (unlikely(!zero_pfn)) {
812                                 pte_free(mm, pgtable);
813                                 count_vm_event(THP_FAULT_FALLBACK);
814                                 goto out;
815                         }
816                         spin_lock(&mm->page_table_lock);
817                         set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
818                                         zero_pfn);
819                         spin_unlock(&mm->page_table_lock);
820                         if (!set) {
821                                 pte_free(mm, pgtable);
822                                 put_huge_zero_page();
823                         }
824                         return 0;
825                 }
826                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
827                                           vma, haddr, numa_node_id(), 0);
828                 if (unlikely(!page)) {
829                         count_vm_event(THP_FAULT_FALLBACK);
830                         goto out;
831                 }
832                 count_vm_event(THP_FAULT_ALLOC);
833                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
834                         put_page(page);
835                         goto out;
836                 }
837                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
838                                                           page))) {
839                         mem_cgroup_uncharge_page(page);
840                         put_page(page);
841                         goto out;
842                 }
843
844                 return 0;
845         }
846 out:
847         /*
848          * Use __pte_alloc instead of pte_alloc_map, because we can't
849          * run pte_offset_map on the pmd, if an huge pmd could
850          * materialize from under us from a different thread.
851          */
852         if (unlikely(pmd_none(*pmd)) &&
853             unlikely(__pte_alloc(mm, vma, pmd, address)))
854                 return VM_FAULT_OOM;
855         /* if an huge pmd materialized from under us just retry later */
856         if (unlikely(pmd_trans_huge(*pmd)))
857                 return 0;
858         /*
859          * A regular pmd is established and it can't morph into a huge pmd
860          * from under us anymore at this point because we hold the mmap_sem
861          * read mode and khugepaged takes it in write mode. So now it's
862          * safe to run pte_offset_map().
863          */
864         pte = pte_offset_map(pmd, address);
865         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
866 }
867
868 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
869                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
870                   struct vm_area_struct *vma)
871 {
872         struct page *src_page;
873         pmd_t pmd;
874         pgtable_t pgtable;
875         int ret;
876
877         ret = -ENOMEM;
878         pgtable = pte_alloc_one(dst_mm, addr);
879         if (unlikely(!pgtable))
880                 goto out;
881
882         spin_lock(&dst_mm->page_table_lock);
883         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
884
885         ret = -EAGAIN;
886         pmd = *src_pmd;
887         if (unlikely(!pmd_trans_huge(pmd))) {
888                 pte_free(dst_mm, pgtable);
889                 goto out_unlock;
890         }
891         /*
892          * mm->page_table_lock is enough to be sure that huge zero pmd is not
893          * under splitting since we don't split the page itself, only pmd to
894          * a page table.
895          */
896         if (is_huge_zero_pmd(pmd)) {
897                 unsigned long zero_pfn;
898                 bool set;
899                 /*
900                  * get_huge_zero_page() will never allocate a new page here,
901                  * since we already have a zero page to copy. It just takes a
902                  * reference.
903                  */
904                 zero_pfn = get_huge_zero_page();
905                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
906                                 zero_pfn);
907                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
908                 ret = 0;
909                 goto out_unlock;
910         }
911         if (unlikely(pmd_trans_splitting(pmd))) {
912                 /* split huge page running from under us */
913                 spin_unlock(&src_mm->page_table_lock);
914                 spin_unlock(&dst_mm->page_table_lock);
915                 pte_free(dst_mm, pgtable);
916
917                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
918                 goto out;
919         }
920         src_page = pmd_page(pmd);
921         VM_BUG_ON(!PageHead(src_page));
922         get_page(src_page);
923         page_dup_rmap(src_page);
924         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
925
926         pmdp_set_wrprotect(src_mm, addr, src_pmd);
927         pmd = pmd_mkold(pmd_wrprotect(pmd));
928         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
929         pgtable_trans_huge_deposit(dst_mm, pgtable);
930         dst_mm->nr_ptes++;
931
932         ret = 0;
933 out_unlock:
934         spin_unlock(&src_mm->page_table_lock);
935         spin_unlock(&dst_mm->page_table_lock);
936 out:
937         return ret;
938 }
939
940 void huge_pmd_set_accessed(struct mm_struct *mm,
941                            struct vm_area_struct *vma,
942                            unsigned long address,
943                            pmd_t *pmd, pmd_t orig_pmd,
944                            int dirty)
945 {
946         pmd_t entry;
947         unsigned long haddr;
948
949         spin_lock(&mm->page_table_lock);
950         if (unlikely(!pmd_same(*pmd, orig_pmd)))
951                 goto unlock;
952
953         entry = pmd_mkyoung(orig_pmd);
954         haddr = address & HPAGE_PMD_MASK;
955         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
956                 update_mmu_cache_pmd(vma, address, pmd);
957
958 unlock:
959         spin_unlock(&mm->page_table_lock);
960 }
961
962 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
963                 struct vm_area_struct *vma, unsigned long address,
964                 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
965 {
966         pgtable_t pgtable;
967         pmd_t _pmd;
968         struct page *page;
969         int i, ret = 0;
970         unsigned long mmun_start;       /* For mmu_notifiers */
971         unsigned long mmun_end;         /* For mmu_notifiers */
972
973         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
974         if (!page) {
975                 ret |= VM_FAULT_OOM;
976                 goto out;
977         }
978
979         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
980                 put_page(page);
981                 ret |= VM_FAULT_OOM;
982                 goto out;
983         }
984
985         clear_user_highpage(page, address);
986         __SetPageUptodate(page);
987
988         mmun_start = haddr;
989         mmun_end   = haddr + HPAGE_PMD_SIZE;
990         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
991
992         spin_lock(&mm->page_table_lock);
993         if (unlikely(!pmd_same(*pmd, orig_pmd)))
994                 goto out_free_page;
995
996         pmdp_clear_flush(vma, haddr, pmd);
997         /* leave pmd empty until pte is filled */
998
999         pgtable = pgtable_trans_huge_withdraw(mm);
1000         pmd_populate(mm, &_pmd, pgtable);
1001
1002         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1003                 pte_t *pte, entry;
1004                 if (haddr == (address & PAGE_MASK)) {
1005                         entry = mk_pte(page, vma->vm_page_prot);
1006                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1007                         page_add_new_anon_rmap(page, vma, haddr);
1008                 } else {
1009                         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1010                         entry = pte_mkspecial(entry);
1011                 }
1012                 pte = pte_offset_map(&_pmd, haddr);
1013                 VM_BUG_ON(!pte_none(*pte));
1014                 set_pte_at(mm, haddr, pte, entry);
1015                 pte_unmap(pte);
1016         }
1017         smp_wmb(); /* make pte visible before pmd */
1018         pmd_populate(mm, pmd, pgtable);
1019         spin_unlock(&mm->page_table_lock);
1020         put_huge_zero_page();
1021         inc_mm_counter(mm, MM_ANONPAGES);
1022
1023         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1024
1025         ret |= VM_FAULT_WRITE;
1026 out:
1027         return ret;
1028 out_free_page:
1029         spin_unlock(&mm->page_table_lock);
1030         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1031         mem_cgroup_uncharge_page(page);
1032         put_page(page);
1033         goto out;
1034 }
1035
1036 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1037                                         struct vm_area_struct *vma,
1038                                         unsigned long address,
1039                                         pmd_t *pmd, pmd_t orig_pmd,
1040                                         struct page *page,
1041                                         unsigned long haddr)
1042 {
1043         pgtable_t pgtable;
1044         pmd_t _pmd;
1045         int ret = 0, i;
1046         struct page **pages;
1047         unsigned long mmun_start;       /* For mmu_notifiers */
1048         unsigned long mmun_end;         /* For mmu_notifiers */
1049
1050         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1051                         GFP_KERNEL);
1052         if (unlikely(!pages)) {
1053                 ret |= VM_FAULT_OOM;
1054                 goto out;
1055         }
1056
1057         for (i = 0; i < HPAGE_PMD_NR; i++) {
1058                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1059                                                __GFP_OTHER_NODE,
1060                                                vma, address, page_to_nid(page));
1061                 if (unlikely(!pages[i] ||
1062                              mem_cgroup_newpage_charge(pages[i], mm,
1063                                                        GFP_KERNEL))) {
1064                         if (pages[i])
1065                                 put_page(pages[i]);
1066                         mem_cgroup_uncharge_start();
1067                         while (--i >= 0) {
1068                                 mem_cgroup_uncharge_page(pages[i]);
1069                                 put_page(pages[i]);
1070                         }
1071                         mem_cgroup_uncharge_end();
1072                         kfree(pages);
1073                         ret |= VM_FAULT_OOM;
1074                         goto out;
1075                 }
1076         }
1077
1078         for (i = 0; i < HPAGE_PMD_NR; i++) {
1079                 copy_user_highpage(pages[i], page + i,
1080                                    haddr + PAGE_SIZE * i, vma);
1081                 __SetPageUptodate(pages[i]);
1082                 cond_resched();
1083         }
1084
1085         mmun_start = haddr;
1086         mmun_end   = haddr + HPAGE_PMD_SIZE;
1087         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1088
1089         spin_lock(&mm->page_table_lock);
1090         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1091                 goto out_free_pages;
1092         VM_BUG_ON(!PageHead(page));
1093
1094         pmdp_clear_flush(vma, haddr, pmd);
1095         /* leave pmd empty until pte is filled */
1096
1097         pgtable = pgtable_trans_huge_withdraw(mm);
1098         pmd_populate(mm, &_pmd, pgtable);
1099
1100         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1101                 pte_t *pte, entry;
1102                 entry = mk_pte(pages[i], vma->vm_page_prot);
1103                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1104                 page_add_new_anon_rmap(pages[i], vma, haddr);
1105                 pte = pte_offset_map(&_pmd, haddr);
1106                 VM_BUG_ON(!pte_none(*pte));
1107                 set_pte_at(mm, haddr, pte, entry);
1108                 pte_unmap(pte);
1109         }
1110         kfree(pages);
1111
1112         smp_wmb(); /* make pte visible before pmd */
1113         pmd_populate(mm, pmd, pgtable);
1114         page_remove_rmap(page);
1115         spin_unlock(&mm->page_table_lock);
1116
1117         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1118
1119         ret |= VM_FAULT_WRITE;
1120         put_page(page);
1121
1122 out:
1123         return ret;
1124
1125 out_free_pages:
1126         spin_unlock(&mm->page_table_lock);
1127         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1128         mem_cgroup_uncharge_start();
1129         for (i = 0; i < HPAGE_PMD_NR; i++) {
1130                 mem_cgroup_uncharge_page(pages[i]);
1131                 put_page(pages[i]);
1132         }
1133         mem_cgroup_uncharge_end();
1134         kfree(pages);
1135         goto out;
1136 }
1137
1138 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1139                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1140 {
1141         int ret = 0;
1142         struct page *page = NULL, *new_page;
1143         unsigned long haddr;
1144         unsigned long mmun_start;       /* For mmu_notifiers */
1145         unsigned long mmun_end;         /* For mmu_notifiers */
1146
1147         VM_BUG_ON(!vma->anon_vma);
1148         haddr = address & HPAGE_PMD_MASK;
1149         if (is_huge_zero_pmd(orig_pmd))
1150                 goto alloc;
1151         spin_lock(&mm->page_table_lock);
1152         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1153                 goto out_unlock;
1154
1155         page = pmd_page(orig_pmd);
1156         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1157         if (page_mapcount(page) == 1) {
1158                 pmd_t entry;
1159                 entry = pmd_mkyoung(orig_pmd);
1160                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1161                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1162                         update_mmu_cache_pmd(vma, address, pmd);
1163                 ret |= VM_FAULT_WRITE;
1164                 goto out_unlock;
1165         }
1166         get_page(page);
1167         spin_unlock(&mm->page_table_lock);
1168 alloc:
1169         if (transparent_hugepage_enabled(vma) &&
1170             !transparent_hugepage_debug_cow())
1171                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1172                                               vma, haddr, numa_node_id(), 0);
1173         else
1174                 new_page = NULL;
1175
1176         if (unlikely(!new_page)) {
1177                 count_vm_event(THP_FAULT_FALLBACK);
1178                 if (is_huge_zero_pmd(orig_pmd)) {
1179                         ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1180                                         address, pmd, orig_pmd, haddr);
1181                 } else {
1182                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1183                                         pmd, orig_pmd, page, haddr);
1184                         if (ret & VM_FAULT_OOM)
1185                                 split_huge_page(page);
1186                         put_page(page);
1187                 }
1188                 goto out;
1189         }
1190         count_vm_event(THP_FAULT_ALLOC);
1191
1192         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1193                 put_page(new_page);
1194                 if (page) {
1195                         split_huge_page(page);
1196                         put_page(page);
1197                 }
1198                 ret |= VM_FAULT_OOM;
1199                 goto out;
1200         }
1201
1202         if (is_huge_zero_pmd(orig_pmd))
1203                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1204         else
1205                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1206         __SetPageUptodate(new_page);
1207
1208         mmun_start = haddr;
1209         mmun_end   = haddr + HPAGE_PMD_SIZE;
1210         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1211
1212         spin_lock(&mm->page_table_lock);
1213         if (page)
1214                 put_page(page);
1215         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1216                 spin_unlock(&mm->page_table_lock);
1217                 mem_cgroup_uncharge_page(new_page);
1218                 put_page(new_page);
1219                 goto out_mn;
1220         } else {
1221                 pmd_t entry;
1222                 entry = mk_huge_pmd(new_page, vma);
1223                 pmdp_clear_flush(vma, haddr, pmd);
1224                 page_add_new_anon_rmap(new_page, vma, haddr);
1225                 set_pmd_at(mm, haddr, pmd, entry);
1226                 update_mmu_cache_pmd(vma, address, pmd);
1227                 if (is_huge_zero_pmd(orig_pmd)) {
1228                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1229                         put_huge_zero_page();
1230                 } else {
1231                         VM_BUG_ON(!PageHead(page));
1232                         page_remove_rmap(page);
1233                         put_page(page);
1234                 }
1235                 ret |= VM_FAULT_WRITE;
1236         }
1237         spin_unlock(&mm->page_table_lock);
1238 out_mn:
1239         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1240 out:
1241         return ret;
1242 out_unlock:
1243         spin_unlock(&mm->page_table_lock);
1244         return ret;
1245 }
1246
1247 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1248                                    unsigned long addr,
1249                                    pmd_t *pmd,
1250                                    unsigned int flags)
1251 {
1252         struct mm_struct *mm = vma->vm_mm;
1253         struct page *page = NULL;
1254
1255         assert_spin_locked(&mm->page_table_lock);
1256
1257         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1258                 goto out;
1259
1260         /* Avoid dumping huge zero page */
1261         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1262                 return ERR_PTR(-EFAULT);
1263
1264         page = pmd_page(*pmd);
1265         VM_BUG_ON(!PageHead(page));
1266         if (flags & FOLL_TOUCH) {
1267                 pmd_t _pmd;
1268                 /*
1269                  * We should set the dirty bit only for FOLL_WRITE but
1270                  * for now the dirty bit in the pmd is meaningless.
1271                  * And if the dirty bit will become meaningful and
1272                  * we'll only set it with FOLL_WRITE, an atomic
1273                  * set_bit will be required on the pmd to set the
1274                  * young bit, instead of the current set_pmd_at.
1275                  */
1276                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1277                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1278         }
1279         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1280                 if (page->mapping && trylock_page(page)) {
1281                         lru_add_drain();
1282                         if (page->mapping)
1283                                 mlock_vma_page(page);
1284                         unlock_page(page);
1285                 }
1286         }
1287         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1288         VM_BUG_ON(!PageCompound(page));
1289         if (flags & FOLL_GET)
1290                 get_page_foll(page);
1291
1292 out:
1293         return page;
1294 }
1295
1296 /* NUMA hinting page fault entry point for trans huge pmds */
1297 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1298                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1299 {
1300         struct page *page;
1301         unsigned long haddr = addr & HPAGE_PMD_MASK;
1302         int target_nid;
1303         int current_nid = -1;
1304         bool migrated;
1305         bool page_locked = false;
1306
1307         spin_lock(&mm->page_table_lock);
1308         if (unlikely(!pmd_same(pmd, *pmdp)))
1309                 goto out_unlock;
1310
1311         page = pmd_page(pmd);
1312         get_page(page);
1313         current_nid = page_to_nid(page);
1314         count_vm_numa_event(NUMA_HINT_FAULTS);
1315         if (current_nid == numa_node_id())
1316                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1317
1318         target_nid = mpol_misplaced(page, vma, haddr);
1319         if (target_nid == -1) {
1320                 put_page(page);
1321                 goto clear_pmdnuma;
1322         }
1323
1324         /* Acquire the page lock to serialise THP migrations */
1325         spin_unlock(&mm->page_table_lock);
1326         lock_page(page);
1327         page_locked = true;
1328
1329         /* Confirm the PTE did not while locked */
1330         spin_lock(&mm->page_table_lock);
1331         if (unlikely(!pmd_same(pmd, *pmdp))) {
1332                 unlock_page(page);
1333                 put_page(page);
1334                 goto out_unlock;
1335         }
1336         spin_unlock(&mm->page_table_lock);
1337
1338         /* Migrate the THP to the requested node */
1339         migrated = migrate_misplaced_transhuge_page(mm, vma,
1340                                 pmdp, pmd, addr,
1341                                 page, target_nid);
1342         if (migrated)
1343                 current_nid = target_nid;
1344         else {
1345                 spin_lock(&mm->page_table_lock);
1346                 if (unlikely(!pmd_same(pmd, *pmdp))) {
1347                         unlock_page(page);
1348                         goto out_unlock;
1349                 }
1350                 goto clear_pmdnuma;
1351         }
1352
1353         task_numa_fault(current_nid, HPAGE_PMD_NR, migrated);
1354         return 0;
1355
1356 clear_pmdnuma:
1357         pmd = pmd_mknonnuma(pmd);
1358         set_pmd_at(mm, haddr, pmdp, pmd);
1359         VM_BUG_ON(pmd_numa(*pmdp));
1360         update_mmu_cache_pmd(vma, addr, pmdp);
1361         if (page_locked)
1362                 unlock_page(page);
1363
1364 out_unlock:
1365         spin_unlock(&mm->page_table_lock);
1366         if (current_nid != -1)
1367                 task_numa_fault(current_nid, HPAGE_PMD_NR, migrated);
1368         return 0;
1369 }
1370
1371 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1372                  pmd_t *pmd, unsigned long addr)
1373 {
1374         int ret = 0;
1375
1376         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1377                 struct page *page;
1378                 pgtable_t pgtable;
1379                 pmd_t orig_pmd;
1380                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1381                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1382                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1383                 if (is_huge_zero_pmd(orig_pmd)) {
1384                         tlb->mm->nr_ptes--;
1385                         spin_unlock(&tlb->mm->page_table_lock);
1386                         put_huge_zero_page();
1387                 } else {
1388                         page = pmd_page(orig_pmd);
1389                         page_remove_rmap(page);
1390                         VM_BUG_ON(page_mapcount(page) < 0);
1391                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1392                         VM_BUG_ON(!PageHead(page));
1393                         tlb->mm->nr_ptes--;
1394                         spin_unlock(&tlb->mm->page_table_lock);
1395                         tlb_remove_page(tlb, page);
1396                 }
1397                 pte_free(tlb->mm, pgtable);
1398                 ret = 1;
1399         }
1400         return ret;
1401 }
1402
1403 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1404                 unsigned long addr, unsigned long end,
1405                 unsigned char *vec)
1406 {
1407         int ret = 0;
1408
1409         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1410                 /*
1411                  * All logical pages in the range are present
1412                  * if backed by a huge page.
1413                  */
1414                 spin_unlock(&vma->vm_mm->page_table_lock);
1415                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1416                 ret = 1;
1417         }
1418
1419         return ret;
1420 }
1421
1422 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1423                   unsigned long old_addr,
1424                   unsigned long new_addr, unsigned long old_end,
1425                   pmd_t *old_pmd, pmd_t *new_pmd)
1426 {
1427         int ret = 0;
1428         pmd_t pmd;
1429
1430         struct mm_struct *mm = vma->vm_mm;
1431
1432         if ((old_addr & ~HPAGE_PMD_MASK) ||
1433             (new_addr & ~HPAGE_PMD_MASK) ||
1434             old_end - old_addr < HPAGE_PMD_SIZE ||
1435             (new_vma->vm_flags & VM_NOHUGEPAGE))
1436                 goto out;
1437
1438         /*
1439          * The destination pmd shouldn't be established, free_pgtables()
1440          * should have release it.
1441          */
1442         if (WARN_ON(!pmd_none(*new_pmd))) {
1443                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1444                 goto out;
1445         }
1446
1447         ret = __pmd_trans_huge_lock(old_pmd, vma);
1448         if (ret == 1) {
1449                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1450                 VM_BUG_ON(!pmd_none(*new_pmd));
1451                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1452                 spin_unlock(&mm->page_table_lock);
1453         }
1454 out:
1455         return ret;
1456 }
1457
1458 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1459                 unsigned long addr, pgprot_t newprot, int prot_numa)
1460 {
1461         struct mm_struct *mm = vma->vm_mm;
1462         int ret = 0;
1463
1464         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1465                 pmd_t entry;
1466                 entry = pmdp_get_and_clear(mm, addr, pmd);
1467                 if (!prot_numa) {
1468                         entry = pmd_modify(entry, newprot);
1469                         BUG_ON(pmd_write(entry));
1470                 } else {
1471                         struct page *page = pmd_page(*pmd);
1472
1473                         /* only check non-shared pages */
1474                         if (page_mapcount(page) == 1 &&
1475                             !pmd_numa(*pmd)) {
1476                                 entry = pmd_mknuma(entry);
1477                         }
1478                 }
1479                 set_pmd_at(mm, addr, pmd, entry);
1480                 spin_unlock(&vma->vm_mm->page_table_lock);
1481                 ret = 1;
1482         }
1483
1484         return ret;
1485 }
1486
1487 /*
1488  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1489  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1490  *
1491  * Note that if it returns 1, this routine returns without unlocking page
1492  * table locks. So callers must unlock them.
1493  */
1494 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1495 {
1496         spin_lock(&vma->vm_mm->page_table_lock);
1497         if (likely(pmd_trans_huge(*pmd))) {
1498                 if (unlikely(pmd_trans_splitting(*pmd))) {
1499                         spin_unlock(&vma->vm_mm->page_table_lock);
1500                         wait_split_huge_page(vma->anon_vma, pmd);
1501                         return -1;
1502                 } else {
1503                         /* Thp mapped by 'pmd' is stable, so we can
1504                          * handle it as it is. */
1505                         return 1;
1506                 }
1507         }
1508         spin_unlock(&vma->vm_mm->page_table_lock);
1509         return 0;
1510 }
1511
1512 pmd_t *page_check_address_pmd(struct page *page,
1513                               struct mm_struct *mm,
1514                               unsigned long address,
1515                               enum page_check_address_pmd_flag flag)
1516 {
1517         pmd_t *pmd, *ret = NULL;
1518
1519         if (address & ~HPAGE_PMD_MASK)
1520                 goto out;
1521
1522         pmd = mm_find_pmd(mm, address);
1523         if (!pmd)
1524                 goto out;
1525         if (pmd_none(*pmd))
1526                 goto out;
1527         if (pmd_page(*pmd) != page)
1528                 goto out;
1529         /*
1530          * split_vma() may create temporary aliased mappings. There is
1531          * no risk as long as all huge pmd are found and have their
1532          * splitting bit set before __split_huge_page_refcount
1533          * runs. Finding the same huge pmd more than once during the
1534          * same rmap walk is not a problem.
1535          */
1536         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1537             pmd_trans_splitting(*pmd))
1538                 goto out;
1539         if (pmd_trans_huge(*pmd)) {
1540                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1541                           !pmd_trans_splitting(*pmd));
1542                 ret = pmd;
1543         }
1544 out:
1545         return ret;
1546 }
1547
1548 static int __split_huge_page_splitting(struct page *page,
1549                                        struct vm_area_struct *vma,
1550                                        unsigned long address)
1551 {
1552         struct mm_struct *mm = vma->vm_mm;
1553         pmd_t *pmd;
1554         int ret = 0;
1555         /* For mmu_notifiers */
1556         const unsigned long mmun_start = address;
1557         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1558
1559         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1560         spin_lock(&mm->page_table_lock);
1561         pmd = page_check_address_pmd(page, mm, address,
1562                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1563         if (pmd) {
1564                 /*
1565                  * We can't temporarily set the pmd to null in order
1566                  * to split it, the pmd must remain marked huge at all
1567                  * times or the VM won't take the pmd_trans_huge paths
1568                  * and it won't wait on the anon_vma->root->rwsem to
1569                  * serialize against split_huge_page*.
1570                  */
1571                 pmdp_splitting_flush(vma, address, pmd);
1572                 ret = 1;
1573         }
1574         spin_unlock(&mm->page_table_lock);
1575         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1576
1577         return ret;
1578 }
1579
1580 static void __split_huge_page_refcount(struct page *page)
1581 {
1582         int i;
1583         struct zone *zone = page_zone(page);
1584         struct lruvec *lruvec;
1585         int tail_count = 0;
1586
1587         /* prevent PageLRU to go away from under us, and freeze lru stats */
1588         spin_lock_irq(&zone->lru_lock);
1589         lruvec = mem_cgroup_page_lruvec(page, zone);
1590
1591         compound_lock(page);
1592         /* complete memcg works before add pages to LRU */
1593         mem_cgroup_split_huge_fixup(page);
1594
1595         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1596                 struct page *page_tail = page + i;
1597
1598                 /* tail_page->_mapcount cannot change */
1599                 BUG_ON(page_mapcount(page_tail) < 0);
1600                 tail_count += page_mapcount(page_tail);
1601                 /* check for overflow */
1602                 BUG_ON(tail_count < 0);
1603                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1604                 /*
1605                  * tail_page->_count is zero and not changing from
1606                  * under us. But get_page_unless_zero() may be running
1607                  * from under us on the tail_page. If we used
1608                  * atomic_set() below instead of atomic_add(), we
1609                  * would then run atomic_set() concurrently with
1610                  * get_page_unless_zero(), and atomic_set() is
1611                  * implemented in C not using locked ops. spin_unlock
1612                  * on x86 sometime uses locked ops because of PPro
1613                  * errata 66, 92, so unless somebody can guarantee
1614                  * atomic_set() here would be safe on all archs (and
1615                  * not only on x86), it's safer to use atomic_add().
1616                  */
1617                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1618                            &page_tail->_count);
1619
1620                 /* after clearing PageTail the gup refcount can be released */
1621                 smp_mb();
1622
1623                 /*
1624                  * retain hwpoison flag of the poisoned tail page:
1625                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1626                  *   by the memory-failure.
1627                  */
1628                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1629                 page_tail->flags |= (page->flags &
1630                                      ((1L << PG_referenced) |
1631                                       (1L << PG_swapbacked) |
1632                                       (1L << PG_mlocked) |
1633                                       (1L << PG_uptodate)));
1634                 page_tail->flags |= (1L << PG_dirty);
1635
1636                 /* clear PageTail before overwriting first_page */
1637                 smp_wmb();
1638
1639                 /*
1640                  * __split_huge_page_splitting() already set the
1641                  * splitting bit in all pmd that could map this
1642                  * hugepage, that will ensure no CPU can alter the
1643                  * mapcount on the head page. The mapcount is only
1644                  * accounted in the head page and it has to be
1645                  * transferred to all tail pages in the below code. So
1646                  * for this code to be safe, the split the mapcount
1647                  * can't change. But that doesn't mean userland can't
1648                  * keep changing and reading the page contents while
1649                  * we transfer the mapcount, so the pmd splitting
1650                  * status is achieved setting a reserved bit in the
1651                  * pmd, not by clearing the present bit.
1652                 */
1653                 page_tail->_mapcount = page->_mapcount;
1654
1655                 BUG_ON(page_tail->mapping);
1656                 page_tail->mapping = page->mapping;
1657
1658                 page_tail->index = page->index + i;
1659                 page_xchg_last_nid(page_tail, page_last_nid(page));
1660
1661                 BUG_ON(!PageAnon(page_tail));
1662                 BUG_ON(!PageUptodate(page_tail));
1663                 BUG_ON(!PageDirty(page_tail));
1664                 BUG_ON(!PageSwapBacked(page_tail));
1665
1666                 lru_add_page_tail(page, page_tail, lruvec);
1667         }
1668         atomic_sub(tail_count, &page->_count);
1669         BUG_ON(atomic_read(&page->_count) <= 0);
1670
1671         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1672         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1673
1674         ClearPageCompound(page);
1675         compound_unlock(page);
1676         spin_unlock_irq(&zone->lru_lock);
1677
1678         for (i = 1; i < HPAGE_PMD_NR; i++) {
1679                 struct page *page_tail = page + i;
1680                 BUG_ON(page_count(page_tail) <= 0);
1681                 /*
1682                  * Tail pages may be freed if there wasn't any mapping
1683                  * like if add_to_swap() is running on a lru page that
1684                  * had its mapping zapped. And freeing these pages
1685                  * requires taking the lru_lock so we do the put_page
1686                  * of the tail pages after the split is complete.
1687                  */
1688                 put_page(page_tail);
1689         }
1690
1691         /*
1692          * Only the head page (now become a regular page) is required
1693          * to be pinned by the caller.
1694          */
1695         BUG_ON(page_count(page) <= 0);
1696 }
1697
1698 static int __split_huge_page_map(struct page *page,
1699                                  struct vm_area_struct *vma,
1700                                  unsigned long address)
1701 {
1702         struct mm_struct *mm = vma->vm_mm;
1703         pmd_t *pmd, _pmd;
1704         int ret = 0, i;
1705         pgtable_t pgtable;
1706         unsigned long haddr;
1707
1708         spin_lock(&mm->page_table_lock);
1709         pmd = page_check_address_pmd(page, mm, address,
1710                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1711         if (pmd) {
1712                 pgtable = pgtable_trans_huge_withdraw(mm);
1713                 pmd_populate(mm, &_pmd, pgtable);
1714
1715                 haddr = address;
1716                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1717                         pte_t *pte, entry;
1718                         BUG_ON(PageCompound(page+i));
1719                         entry = mk_pte(page + i, vma->vm_page_prot);
1720                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1721                         if (!pmd_write(*pmd))
1722                                 entry = pte_wrprotect(entry);
1723                         else
1724                                 BUG_ON(page_mapcount(page) != 1);
1725                         if (!pmd_young(*pmd))
1726                                 entry = pte_mkold(entry);
1727                         if (pmd_numa(*pmd))
1728                                 entry = pte_mknuma(entry);
1729                         pte = pte_offset_map(&_pmd, haddr);
1730                         BUG_ON(!pte_none(*pte));
1731                         set_pte_at(mm, haddr, pte, entry);
1732                         pte_unmap(pte);
1733                 }
1734
1735                 smp_wmb(); /* make pte visible before pmd */
1736                 /*
1737                  * Up to this point the pmd is present and huge and
1738                  * userland has the whole access to the hugepage
1739                  * during the split (which happens in place). If we
1740                  * overwrite the pmd with the not-huge version
1741                  * pointing to the pte here (which of course we could
1742                  * if all CPUs were bug free), userland could trigger
1743                  * a small page size TLB miss on the small sized TLB
1744                  * while the hugepage TLB entry is still established
1745                  * in the huge TLB. Some CPU doesn't like that. See
1746                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1747                  * Erratum 383 on page 93. Intel should be safe but is
1748                  * also warns that it's only safe if the permission
1749                  * and cache attributes of the two entries loaded in
1750                  * the two TLB is identical (which should be the case
1751                  * here). But it is generally safer to never allow
1752                  * small and huge TLB entries for the same virtual
1753                  * address to be loaded simultaneously. So instead of
1754                  * doing "pmd_populate(); flush_tlb_range();" we first
1755                  * mark the current pmd notpresent (atomically because
1756                  * here the pmd_trans_huge and pmd_trans_splitting
1757                  * must remain set at all times on the pmd until the
1758                  * split is complete for this pmd), then we flush the
1759                  * SMP TLB and finally we write the non-huge version
1760                  * of the pmd entry with pmd_populate.
1761                  */
1762                 pmdp_invalidate(vma, address, pmd);
1763                 pmd_populate(mm, pmd, pgtable);
1764                 ret = 1;
1765         }
1766         spin_unlock(&mm->page_table_lock);
1767
1768         return ret;
1769 }
1770
1771 /* must be called with anon_vma->root->rwsem held */
1772 static void __split_huge_page(struct page *page,
1773                               struct anon_vma *anon_vma)
1774 {
1775         int mapcount, mapcount2;
1776         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1777         struct anon_vma_chain *avc;
1778
1779         BUG_ON(!PageHead(page));
1780         BUG_ON(PageTail(page));
1781
1782         mapcount = 0;
1783         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1784                 struct vm_area_struct *vma = avc->vma;
1785                 unsigned long addr = vma_address(page, vma);
1786                 BUG_ON(is_vma_temporary_stack(vma));
1787                 mapcount += __split_huge_page_splitting(page, vma, addr);
1788         }
1789         /*
1790          * It is critical that new vmas are added to the tail of the
1791          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1792          * and establishes a child pmd before
1793          * __split_huge_page_splitting() freezes the parent pmd (so if
1794          * we fail to prevent copy_huge_pmd() from running until the
1795          * whole __split_huge_page() is complete), we will still see
1796          * the newly established pmd of the child later during the
1797          * walk, to be able to set it as pmd_trans_splitting too.
1798          */
1799         if (mapcount != page_mapcount(page))
1800                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1801                        mapcount, page_mapcount(page));
1802         BUG_ON(mapcount != page_mapcount(page));
1803
1804         __split_huge_page_refcount(page);
1805
1806         mapcount2 = 0;
1807         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1808                 struct vm_area_struct *vma = avc->vma;
1809                 unsigned long addr = vma_address(page, vma);
1810                 BUG_ON(is_vma_temporary_stack(vma));
1811                 mapcount2 += __split_huge_page_map(page, vma, addr);
1812         }
1813         if (mapcount != mapcount2)
1814                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1815                        mapcount, mapcount2, page_mapcount(page));
1816         BUG_ON(mapcount != mapcount2);
1817 }
1818
1819 int split_huge_page(struct page *page)
1820 {
1821         struct anon_vma *anon_vma;
1822         int ret = 1;
1823
1824         BUG_ON(is_huge_zero_pfn(page_to_pfn(page)));
1825         BUG_ON(!PageAnon(page));
1826
1827         /*
1828          * The caller does not necessarily hold an mmap_sem that would prevent
1829          * the anon_vma disappearing so we first we take a reference to it
1830          * and then lock the anon_vma for write. This is similar to
1831          * page_lock_anon_vma_read except the write lock is taken to serialise
1832          * against parallel split or collapse operations.
1833          */
1834         anon_vma = page_get_anon_vma(page);
1835         if (!anon_vma)
1836                 goto out;
1837         anon_vma_lock_write(anon_vma);
1838
1839         ret = 0;
1840         if (!PageCompound(page))
1841                 goto out_unlock;
1842
1843         BUG_ON(!PageSwapBacked(page));
1844         __split_huge_page(page, anon_vma);
1845         count_vm_event(THP_SPLIT);
1846
1847         BUG_ON(PageCompound(page));
1848 out_unlock:
1849         anon_vma_unlock(anon_vma);
1850         put_anon_vma(anon_vma);
1851 out:
1852         return ret;
1853 }
1854
1855 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1856
1857 int hugepage_madvise(struct vm_area_struct *vma,
1858                      unsigned long *vm_flags, int advice)
1859 {
1860         struct mm_struct *mm = vma->vm_mm;
1861
1862         switch (advice) {
1863         case MADV_HUGEPAGE:
1864                 /*
1865                  * Be somewhat over-protective like KSM for now!
1866                  */
1867                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1868                         return -EINVAL;
1869                 if (mm->def_flags & VM_NOHUGEPAGE)
1870                         return -EINVAL;
1871                 *vm_flags &= ~VM_NOHUGEPAGE;
1872                 *vm_flags |= VM_HUGEPAGE;
1873                 /*
1874                  * If the vma become good for khugepaged to scan,
1875                  * register it here without waiting a page fault that
1876                  * may not happen any time soon.
1877                  */
1878                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1879                         return -ENOMEM;
1880                 break;
1881         case MADV_NOHUGEPAGE:
1882                 /*
1883                  * Be somewhat over-protective like KSM for now!
1884                  */
1885                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1886                         return -EINVAL;
1887                 *vm_flags &= ~VM_HUGEPAGE;
1888                 *vm_flags |= VM_NOHUGEPAGE;
1889                 /*
1890                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1891                  * this vma even if we leave the mm registered in khugepaged if
1892                  * it got registered before VM_NOHUGEPAGE was set.
1893                  */
1894                 break;
1895         }
1896
1897         return 0;
1898 }
1899
1900 static int __init khugepaged_slab_init(void)
1901 {
1902         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1903                                           sizeof(struct mm_slot),
1904                                           __alignof__(struct mm_slot), 0, NULL);
1905         if (!mm_slot_cache)
1906                 return -ENOMEM;
1907
1908         return 0;
1909 }
1910
1911 static void __init khugepaged_slab_free(void)
1912 {
1913         kmem_cache_destroy(mm_slot_cache);
1914         mm_slot_cache = NULL;
1915 }
1916
1917 static inline struct mm_slot *alloc_mm_slot(void)
1918 {
1919         if (!mm_slot_cache)     /* initialization failed */
1920                 return NULL;
1921         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1922 }
1923
1924 static inline void free_mm_slot(struct mm_slot *mm_slot)
1925 {
1926         kmem_cache_free(mm_slot_cache, mm_slot);
1927 }
1928
1929 static int __init mm_slots_hash_init(void)
1930 {
1931         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1932                                 GFP_KERNEL);
1933         if (!mm_slots_hash)
1934                 return -ENOMEM;
1935         return 0;
1936 }
1937
1938 #if 0
1939 static void __init mm_slots_hash_free(void)
1940 {
1941         kfree(mm_slots_hash);
1942         mm_slots_hash = NULL;
1943 }
1944 #endif
1945
1946 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1947 {
1948         struct mm_slot *mm_slot;
1949         struct hlist_head *bucket;
1950         struct hlist_node *node;
1951
1952         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1953                                 % MM_SLOTS_HASH_HEADS];
1954         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1955                 if (mm == mm_slot->mm)
1956                         return mm_slot;
1957         }
1958         return NULL;
1959 }
1960
1961 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1962                                     struct mm_slot *mm_slot)
1963 {
1964         struct hlist_head *bucket;
1965
1966         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1967                                 % MM_SLOTS_HASH_HEADS];
1968         mm_slot->mm = mm;
1969         hlist_add_head(&mm_slot->hash, bucket);
1970 }
1971
1972 static inline int khugepaged_test_exit(struct mm_struct *mm)
1973 {
1974         return atomic_read(&mm->mm_users) == 0;
1975 }
1976
1977 int __khugepaged_enter(struct mm_struct *mm)
1978 {
1979         struct mm_slot *mm_slot;
1980         int wakeup;
1981
1982         mm_slot = alloc_mm_slot();
1983         if (!mm_slot)
1984                 return -ENOMEM;
1985
1986         /* __khugepaged_exit() must not run from under us */
1987         VM_BUG_ON(khugepaged_test_exit(mm));
1988         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1989                 free_mm_slot(mm_slot);
1990                 return 0;
1991         }
1992
1993         spin_lock(&khugepaged_mm_lock);
1994         insert_to_mm_slots_hash(mm, mm_slot);
1995         /*
1996          * Insert just behind the scanning cursor, to let the area settle
1997          * down a little.
1998          */
1999         wakeup = list_empty(&khugepaged_scan.mm_head);
2000         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2001         spin_unlock(&khugepaged_mm_lock);
2002
2003         atomic_inc(&mm->mm_count);
2004         if (wakeup)
2005                 wake_up_interruptible(&khugepaged_wait);
2006
2007         return 0;
2008 }
2009
2010 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2011 {
2012         unsigned long hstart, hend;
2013         if (!vma->anon_vma)
2014                 /*
2015                  * Not yet faulted in so we will register later in the
2016                  * page fault if needed.
2017                  */
2018                 return 0;
2019         if (vma->vm_ops)
2020                 /* khugepaged not yet working on file or special mappings */
2021                 return 0;
2022         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2023         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2024         hend = vma->vm_end & HPAGE_PMD_MASK;
2025         if (hstart < hend)
2026                 return khugepaged_enter(vma);
2027         return 0;
2028 }
2029
2030 void __khugepaged_exit(struct mm_struct *mm)
2031 {
2032         struct mm_slot *mm_slot;
2033         int free = 0;
2034
2035         spin_lock(&khugepaged_mm_lock);
2036         mm_slot = get_mm_slot(mm);
2037         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2038                 hlist_del(&mm_slot->hash);
2039                 list_del(&mm_slot->mm_node);
2040                 free = 1;
2041         }
2042         spin_unlock(&khugepaged_mm_lock);
2043
2044         if (free) {
2045                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2046                 free_mm_slot(mm_slot);
2047                 mmdrop(mm);
2048         } else if (mm_slot) {
2049                 /*
2050                  * This is required to serialize against
2051                  * khugepaged_test_exit() (which is guaranteed to run
2052                  * under mmap sem read mode). Stop here (after we
2053                  * return all pagetables will be destroyed) until
2054                  * khugepaged has finished working on the pagetables
2055                  * under the mmap_sem.
2056                  */
2057                 down_write(&mm->mmap_sem);
2058                 up_write(&mm->mmap_sem);
2059         }
2060 }
2061
2062 static void release_pte_page(struct page *page)
2063 {
2064         /* 0 stands for page_is_file_cache(page) == false */
2065         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2066         unlock_page(page);
2067         putback_lru_page(page);
2068 }
2069
2070 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2071 {
2072         while (--_pte >= pte) {
2073                 pte_t pteval = *_pte;
2074                 if (!pte_none(pteval))
2075                         release_pte_page(pte_page(pteval));
2076         }
2077 }
2078
2079 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2080                                         unsigned long address,
2081                                         pte_t *pte)
2082 {
2083         struct page *page;
2084         pte_t *_pte;
2085         int referenced = 0, none = 0;
2086         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2087              _pte++, address += PAGE_SIZE) {
2088                 pte_t pteval = *_pte;
2089                 if (pte_none(pteval)) {
2090                         if (++none <= khugepaged_max_ptes_none)
2091                                 continue;
2092                         else
2093                                 goto out;
2094                 }
2095                 if (!pte_present(pteval) || !pte_write(pteval))
2096                         goto out;
2097                 page = vm_normal_page(vma, address, pteval);
2098                 if (unlikely(!page))
2099                         goto out;
2100
2101                 VM_BUG_ON(PageCompound(page));
2102                 BUG_ON(!PageAnon(page));
2103                 VM_BUG_ON(!PageSwapBacked(page));
2104
2105                 /* cannot use mapcount: can't collapse if there's a gup pin */
2106                 if (page_count(page) != 1)
2107                         goto out;
2108                 /*
2109                  * We can do it before isolate_lru_page because the
2110                  * page can't be freed from under us. NOTE: PG_lock
2111                  * is needed to serialize against split_huge_page
2112                  * when invoked from the VM.
2113                  */
2114                 if (!trylock_page(page))
2115                         goto out;
2116                 /*
2117                  * Isolate the page to avoid collapsing an hugepage
2118                  * currently in use by the VM.
2119                  */
2120                 if (isolate_lru_page(page)) {
2121                         unlock_page(page);
2122                         goto out;
2123                 }
2124                 /* 0 stands for page_is_file_cache(page) == false */
2125                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2126                 VM_BUG_ON(!PageLocked(page));
2127                 VM_BUG_ON(PageLRU(page));
2128
2129                 /* If there is no mapped pte young don't collapse the page */
2130                 if (pte_young(pteval) || PageReferenced(page) ||
2131                     mmu_notifier_test_young(vma->vm_mm, address))
2132                         referenced = 1;
2133         }
2134         if (likely(referenced))
2135                 return 1;
2136 out:
2137         release_pte_pages(pte, _pte);
2138         return 0;
2139 }
2140
2141 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2142                                       struct vm_area_struct *vma,
2143                                       unsigned long address,
2144                                       spinlock_t *ptl)
2145 {
2146         pte_t *_pte;
2147         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2148                 pte_t pteval = *_pte;
2149                 struct page *src_page;
2150
2151                 if (pte_none(pteval)) {
2152                         clear_user_highpage(page, address);
2153                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2154                 } else {
2155                         src_page = pte_page(pteval);
2156                         copy_user_highpage(page, src_page, address, vma);
2157                         VM_BUG_ON(page_mapcount(src_page) != 1);
2158                         release_pte_page(src_page);
2159                         /*
2160                          * ptl mostly unnecessary, but preempt has to
2161                          * be disabled to update the per-cpu stats
2162                          * inside page_remove_rmap().
2163                          */
2164                         spin_lock(ptl);
2165                         /*
2166                          * paravirt calls inside pte_clear here are
2167                          * superfluous.
2168                          */
2169                         pte_clear(vma->vm_mm, address, _pte);
2170                         page_remove_rmap(src_page);
2171                         spin_unlock(ptl);
2172                         free_page_and_swap_cache(src_page);
2173                 }
2174
2175                 address += PAGE_SIZE;
2176                 page++;
2177         }
2178 }
2179
2180 static void khugepaged_alloc_sleep(void)
2181 {
2182         wait_event_freezable_timeout(khugepaged_wait, false,
2183                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2184 }
2185
2186 #ifdef CONFIG_NUMA
2187 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2188 {
2189         if (IS_ERR(*hpage)) {
2190                 if (!*wait)
2191                         return false;
2192
2193                 *wait = false;
2194                 *hpage = NULL;
2195                 khugepaged_alloc_sleep();
2196         } else if (*hpage) {
2197                 put_page(*hpage);
2198                 *hpage = NULL;
2199         }
2200
2201         return true;
2202 }
2203
2204 static struct page
2205 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2206                        struct vm_area_struct *vma, unsigned long address,
2207                        int node)
2208 {
2209         VM_BUG_ON(*hpage);
2210         /*
2211          * Allocate the page while the vma is still valid and under
2212          * the mmap_sem read mode so there is no memory allocation
2213          * later when we take the mmap_sem in write mode. This is more
2214          * friendly behavior (OTOH it may actually hide bugs) to
2215          * filesystems in userland with daemons allocating memory in
2216          * the userland I/O paths.  Allocating memory with the
2217          * mmap_sem in read mode is good idea also to allow greater
2218          * scalability.
2219          */
2220         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2221                                       node, __GFP_OTHER_NODE);
2222
2223         /*
2224          * After allocating the hugepage, release the mmap_sem read lock in
2225          * preparation for taking it in write mode.
2226          */
2227         up_read(&mm->mmap_sem);
2228         if (unlikely(!*hpage)) {
2229                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2230                 *hpage = ERR_PTR(-ENOMEM);
2231                 return NULL;
2232         }
2233
2234         count_vm_event(THP_COLLAPSE_ALLOC);
2235         return *hpage;
2236 }
2237 #else
2238 static struct page *khugepaged_alloc_hugepage(bool *wait)
2239 {
2240         struct page *hpage;
2241
2242         do {
2243                 hpage = alloc_hugepage(khugepaged_defrag());
2244                 if (!hpage) {
2245                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2246                         if (!*wait)
2247                                 return NULL;
2248
2249                         *wait = false;
2250                         khugepaged_alloc_sleep();
2251                 } else
2252                         count_vm_event(THP_COLLAPSE_ALLOC);
2253         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2254
2255         return hpage;
2256 }
2257
2258 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2259 {
2260         if (!*hpage)
2261                 *hpage = khugepaged_alloc_hugepage(wait);
2262
2263         if (unlikely(!*hpage))
2264                 return false;
2265
2266         return true;
2267 }
2268
2269 static struct page
2270 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2271                        struct vm_area_struct *vma, unsigned long address,
2272                        int node)
2273 {
2274         up_read(&mm->mmap_sem);
2275         VM_BUG_ON(!*hpage);
2276         return  *hpage;
2277 }
2278 #endif
2279
2280 static bool hugepage_vma_check(struct vm_area_struct *vma)
2281 {
2282         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2283             (vma->vm_flags & VM_NOHUGEPAGE))
2284                 return false;
2285
2286         if (!vma->anon_vma || vma->vm_ops)
2287                 return false;
2288         if (is_vma_temporary_stack(vma))
2289                 return false;
2290         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2291         return true;
2292 }
2293
2294 static void collapse_huge_page(struct mm_struct *mm,
2295                                    unsigned long address,
2296                                    struct page **hpage,
2297                                    struct vm_area_struct *vma,
2298                                    int node)
2299 {
2300         pmd_t *pmd, _pmd;
2301         pte_t *pte;
2302         pgtable_t pgtable;
2303         struct page *new_page;
2304         spinlock_t *ptl;
2305         int isolated;
2306         unsigned long hstart, hend;
2307         unsigned long mmun_start;       /* For mmu_notifiers */
2308         unsigned long mmun_end;         /* For mmu_notifiers */
2309
2310         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2311
2312         /* release the mmap_sem read lock. */
2313         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2314         if (!new_page)
2315                 return;
2316
2317         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2318                 return;
2319
2320         /*
2321          * Prevent all access to pagetables with the exception of
2322          * gup_fast later hanlded by the ptep_clear_flush and the VM
2323          * handled by the anon_vma lock + PG_lock.
2324          */
2325         down_write(&mm->mmap_sem);
2326         if (unlikely(khugepaged_test_exit(mm)))
2327                 goto out;
2328
2329         vma = find_vma(mm, address);
2330         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2331         hend = vma->vm_end & HPAGE_PMD_MASK;
2332         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2333                 goto out;
2334         if (!hugepage_vma_check(vma))
2335                 goto out;
2336         pmd = mm_find_pmd(mm, address);
2337         if (!pmd)
2338                 goto out;
2339         if (pmd_trans_huge(*pmd))
2340                 goto out;
2341
2342         anon_vma_lock_write(vma->anon_vma);
2343
2344         pte = pte_offset_map(pmd, address);
2345         ptl = pte_lockptr(mm, pmd);
2346
2347         mmun_start = address;
2348         mmun_end   = address + HPAGE_PMD_SIZE;
2349         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2350         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2351         /*
2352          * After this gup_fast can't run anymore. This also removes
2353          * any huge TLB entry from the CPU so we won't allow
2354          * huge and small TLB entries for the same virtual address
2355          * to avoid the risk of CPU bugs in that area.
2356          */
2357         _pmd = pmdp_clear_flush(vma, address, pmd);
2358         spin_unlock(&mm->page_table_lock);
2359         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2360
2361         spin_lock(ptl);
2362         isolated = __collapse_huge_page_isolate(vma, address, pte);
2363         spin_unlock(ptl);
2364
2365         if (unlikely(!isolated)) {
2366                 pte_unmap(pte);
2367                 spin_lock(&mm->page_table_lock);
2368                 BUG_ON(!pmd_none(*pmd));
2369                 set_pmd_at(mm, address, pmd, _pmd);
2370                 spin_unlock(&mm->page_table_lock);
2371                 anon_vma_unlock(vma->anon_vma);
2372                 goto out;
2373         }
2374
2375         /*
2376          * All pages are isolated and locked so anon_vma rmap
2377          * can't run anymore.
2378          */
2379         anon_vma_unlock(vma->anon_vma);
2380
2381         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2382         pte_unmap(pte);
2383         __SetPageUptodate(new_page);
2384         pgtable = pmd_pgtable(_pmd);
2385
2386         _pmd = mk_huge_pmd(new_page, vma);
2387
2388         /*
2389          * spin_lock() below is not the equivalent of smp_wmb(), so
2390          * this is needed to avoid the copy_huge_page writes to become
2391          * visible after the set_pmd_at() write.
2392          */
2393         smp_wmb();
2394
2395         spin_lock(&mm->page_table_lock);
2396         BUG_ON(!pmd_none(*pmd));
2397         page_add_new_anon_rmap(new_page, vma, address);
2398         set_pmd_at(mm, address, pmd, _pmd);
2399         update_mmu_cache_pmd(vma, address, pmd);
2400         pgtable_trans_huge_deposit(mm, pgtable);
2401         spin_unlock(&mm->page_table_lock);
2402
2403         *hpage = NULL;
2404
2405         khugepaged_pages_collapsed++;
2406 out_up_write:
2407         up_write(&mm->mmap_sem);
2408         return;
2409
2410 out:
2411         mem_cgroup_uncharge_page(new_page);
2412         goto out_up_write;
2413 }
2414
2415 static int khugepaged_scan_pmd(struct mm_struct *mm,
2416                                struct vm_area_struct *vma,
2417                                unsigned long address,
2418                                struct page **hpage)
2419 {
2420         pmd_t *pmd;
2421         pte_t *pte, *_pte;
2422         int ret = 0, referenced = 0, none = 0;
2423         struct page *page;
2424         unsigned long _address;
2425         spinlock_t *ptl;
2426         int node = -1;
2427
2428         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2429
2430         pmd = mm_find_pmd(mm, address);
2431         if (!pmd)
2432                 goto out;
2433         if (pmd_trans_huge(*pmd))
2434                 goto out;
2435
2436         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2437         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2438              _pte++, _address += PAGE_SIZE) {
2439                 pte_t pteval = *_pte;
2440                 if (pte_none(pteval)) {
2441                         if (++none <= khugepaged_max_ptes_none)
2442                                 continue;
2443                         else
2444                                 goto out_unmap;
2445                 }
2446                 if (!pte_present(pteval) || !pte_write(pteval))
2447                         goto out_unmap;
2448                 page = vm_normal_page(vma, _address, pteval);
2449                 if (unlikely(!page))
2450                         goto out_unmap;
2451                 /*
2452                  * Chose the node of the first page. This could
2453                  * be more sophisticated and look at more pages,
2454                  * but isn't for now.
2455                  */
2456                 if (node == -1)
2457                         node = page_to_nid(page);
2458                 VM_BUG_ON(PageCompound(page));
2459                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2460                         goto out_unmap;
2461                 /* cannot use mapcount: can't collapse if there's a gup pin */
2462                 if (page_count(page) != 1)
2463                         goto out_unmap;
2464                 if (pte_young(pteval) || PageReferenced(page) ||
2465                     mmu_notifier_test_young(vma->vm_mm, address))
2466                         referenced = 1;
2467         }
2468         if (referenced)
2469                 ret = 1;
2470 out_unmap:
2471         pte_unmap_unlock(pte, ptl);
2472         if (ret)
2473                 /* collapse_huge_page will return with the mmap_sem released */
2474                 collapse_huge_page(mm, address, hpage, vma, node);
2475 out:
2476         return ret;
2477 }
2478
2479 static void collect_mm_slot(struct mm_slot *mm_slot)
2480 {
2481         struct mm_struct *mm = mm_slot->mm;
2482
2483         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2484
2485         if (khugepaged_test_exit(mm)) {
2486                 /* free mm_slot */
2487                 hlist_del(&mm_slot->hash);
2488                 list_del(&mm_slot->mm_node);
2489
2490                 /*
2491                  * Not strictly needed because the mm exited already.
2492                  *
2493                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2494                  */
2495
2496                 /* khugepaged_mm_lock actually not necessary for the below */
2497                 free_mm_slot(mm_slot);
2498                 mmdrop(mm);
2499         }
2500 }
2501
2502 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2503                                             struct page **hpage)
2504         __releases(&khugepaged_mm_lock)
2505         __acquires(&khugepaged_mm_lock)
2506 {
2507         struct mm_slot *mm_slot;
2508         struct mm_struct *mm;
2509         struct vm_area_struct *vma;
2510         int progress = 0;
2511
2512         VM_BUG_ON(!pages);
2513         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2514
2515         if (khugepaged_scan.mm_slot)
2516                 mm_slot = khugepaged_scan.mm_slot;
2517         else {
2518                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2519                                      struct mm_slot, mm_node);
2520                 khugepaged_scan.address = 0;
2521                 khugepaged_scan.mm_slot = mm_slot;
2522         }
2523         spin_unlock(&khugepaged_mm_lock);
2524
2525         mm = mm_slot->mm;
2526         down_read(&mm->mmap_sem);
2527         if (unlikely(khugepaged_test_exit(mm)))
2528                 vma = NULL;
2529         else
2530                 vma = find_vma(mm, khugepaged_scan.address);
2531
2532         progress++;
2533         for (; vma; vma = vma->vm_next) {
2534                 unsigned long hstart, hend;
2535
2536                 cond_resched();
2537                 if (unlikely(khugepaged_test_exit(mm))) {
2538                         progress++;
2539                         break;
2540                 }
2541                 if (!hugepage_vma_check(vma)) {
2542 skip:
2543                         progress++;
2544                         continue;
2545                 }
2546                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2547                 hend = vma->vm_end & HPAGE_PMD_MASK;
2548                 if (hstart >= hend)
2549                         goto skip;
2550                 if (khugepaged_scan.address > hend)
2551                         goto skip;
2552                 if (khugepaged_scan.address < hstart)
2553                         khugepaged_scan.address = hstart;
2554                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2555
2556                 while (khugepaged_scan.address < hend) {
2557                         int ret;
2558                         cond_resched();
2559                         if (unlikely(khugepaged_test_exit(mm)))
2560                                 goto breakouterloop;
2561
2562                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2563                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2564                                   hend);
2565                         ret = khugepaged_scan_pmd(mm, vma,
2566                                                   khugepaged_scan.address,
2567                                                   hpage);
2568                         /* move to next address */
2569                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2570                         progress += HPAGE_PMD_NR;
2571                         if (ret)
2572                                 /* we released mmap_sem so break loop */
2573                                 goto breakouterloop_mmap_sem;
2574                         if (progress >= pages)
2575                                 goto breakouterloop;
2576                 }
2577         }
2578 breakouterloop:
2579         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2580 breakouterloop_mmap_sem:
2581
2582         spin_lock(&khugepaged_mm_lock);
2583         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2584         /*
2585          * Release the current mm_slot if this mm is about to die, or
2586          * if we scanned all vmas of this mm.
2587          */
2588         if (khugepaged_test_exit(mm) || !vma) {
2589                 /*
2590                  * Make sure that if mm_users is reaching zero while
2591                  * khugepaged runs here, khugepaged_exit will find
2592                  * mm_slot not pointing to the exiting mm.
2593                  */
2594                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2595                         khugepaged_scan.mm_slot = list_entry(
2596                                 mm_slot->mm_node.next,
2597                                 struct mm_slot, mm_node);
2598                         khugepaged_scan.address = 0;
2599                 } else {
2600                         khugepaged_scan.mm_slot = NULL;
2601                         khugepaged_full_scans++;
2602                 }
2603
2604                 collect_mm_slot(mm_slot);
2605         }
2606
2607         return progress;
2608 }
2609
2610 static int khugepaged_has_work(void)
2611 {
2612         return !list_empty(&khugepaged_scan.mm_head) &&
2613                 khugepaged_enabled();
2614 }
2615
2616 static int khugepaged_wait_event(void)
2617 {
2618         return !list_empty(&khugepaged_scan.mm_head) ||
2619                 kthread_should_stop();
2620 }
2621
2622 static void khugepaged_do_scan(void)
2623 {
2624         struct page *hpage = NULL;
2625         unsigned int progress = 0, pass_through_head = 0;
2626         unsigned int pages = khugepaged_pages_to_scan;
2627         bool wait = true;
2628
2629         barrier(); /* write khugepaged_pages_to_scan to local stack */
2630
2631         while (progress < pages) {
2632                 if (!khugepaged_prealloc_page(&hpage, &wait))
2633                         break;
2634
2635                 cond_resched();
2636
2637                 if (unlikely(kthread_should_stop() || freezing(current)))
2638                         break;
2639
2640                 spin_lock(&khugepaged_mm_lock);
2641                 if (!khugepaged_scan.mm_slot)
2642                         pass_through_head++;
2643                 if (khugepaged_has_work() &&
2644                     pass_through_head < 2)
2645                         progress += khugepaged_scan_mm_slot(pages - progress,
2646                                                             &hpage);
2647                 else
2648                         progress = pages;
2649                 spin_unlock(&khugepaged_mm_lock);
2650         }
2651
2652         if (!IS_ERR_OR_NULL(hpage))
2653                 put_page(hpage);
2654 }
2655
2656 static void khugepaged_wait_work(void)
2657 {
2658         try_to_freeze();
2659
2660         if (khugepaged_has_work()) {
2661                 if (!khugepaged_scan_sleep_millisecs)
2662                         return;
2663
2664                 wait_event_freezable_timeout(khugepaged_wait,
2665                                              kthread_should_stop(),
2666                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2667                 return;
2668         }
2669
2670         if (khugepaged_enabled())
2671                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2672 }
2673
2674 static int khugepaged(void *none)
2675 {
2676         struct mm_slot *mm_slot;
2677
2678         set_freezable();
2679         set_user_nice(current, 19);
2680
2681         while (!kthread_should_stop()) {
2682                 khugepaged_do_scan();
2683                 khugepaged_wait_work();
2684         }
2685
2686         spin_lock(&khugepaged_mm_lock);
2687         mm_slot = khugepaged_scan.mm_slot;
2688         khugepaged_scan.mm_slot = NULL;
2689         if (mm_slot)
2690                 collect_mm_slot(mm_slot);
2691         spin_unlock(&khugepaged_mm_lock);
2692         return 0;
2693 }
2694
2695 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2696                 unsigned long haddr, pmd_t *pmd)
2697 {
2698         struct mm_struct *mm = vma->vm_mm;
2699         pgtable_t pgtable;
2700         pmd_t _pmd;
2701         int i;
2702
2703         pmdp_clear_flush(vma, haddr, pmd);
2704         /* leave pmd empty until pte is filled */
2705
2706         pgtable = pgtable_trans_huge_withdraw(mm);
2707         pmd_populate(mm, &_pmd, pgtable);
2708
2709         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2710                 pte_t *pte, entry;
2711                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2712                 entry = pte_mkspecial(entry);
2713                 pte = pte_offset_map(&_pmd, haddr);
2714                 VM_BUG_ON(!pte_none(*pte));
2715                 set_pte_at(mm, haddr, pte, entry);
2716                 pte_unmap(pte);
2717         }
2718         smp_wmb(); /* make pte visible before pmd */
2719         pmd_populate(mm, pmd, pgtable);
2720         put_huge_zero_page();
2721 }
2722
2723 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2724                 pmd_t *pmd)
2725 {
2726         struct page *page;
2727         struct mm_struct *mm = vma->vm_mm;
2728         unsigned long haddr = address & HPAGE_PMD_MASK;
2729         unsigned long mmun_start;       /* For mmu_notifiers */
2730         unsigned long mmun_end;         /* For mmu_notifiers */
2731
2732         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2733
2734         mmun_start = haddr;
2735         mmun_end   = haddr + HPAGE_PMD_SIZE;
2736         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2737         spin_lock(&mm->page_table_lock);
2738         if (unlikely(!pmd_trans_huge(*pmd))) {
2739                 spin_unlock(&mm->page_table_lock);
2740                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2741                 return;
2742         }
2743         if (is_huge_zero_pmd(*pmd)) {
2744                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2745                 spin_unlock(&mm->page_table_lock);
2746                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2747                 return;
2748         }
2749         page = pmd_page(*pmd);
2750         VM_BUG_ON(!page_count(page));
2751         get_page(page);
2752         spin_unlock(&mm->page_table_lock);
2753         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2754
2755         split_huge_page(page);
2756
2757         put_page(page);
2758         BUG_ON(pmd_trans_huge(*pmd));
2759 }
2760
2761 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2762                 pmd_t *pmd)
2763 {
2764         struct vm_area_struct *vma;
2765
2766         vma = find_vma(mm, address);
2767         BUG_ON(vma == NULL);
2768         split_huge_page_pmd(vma, address, pmd);
2769 }
2770
2771 static void split_huge_page_address(struct mm_struct *mm,
2772                                     unsigned long address)
2773 {
2774         pmd_t *pmd;
2775
2776         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2777
2778         pmd = mm_find_pmd(mm, address);
2779         if (!pmd)
2780                 return;
2781         /*
2782          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2783          * materialize from under us.
2784          */
2785         split_huge_page_pmd_mm(mm, address, pmd);
2786 }
2787
2788 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2789                              unsigned long start,
2790                              unsigned long end,
2791                              long adjust_next)
2792 {
2793         /*
2794          * If the new start address isn't hpage aligned and it could
2795          * previously contain an hugepage: check if we need to split
2796          * an huge pmd.
2797          */
2798         if (start & ~HPAGE_PMD_MASK &&
2799             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2800             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2801                 split_huge_page_address(vma->vm_mm, start);
2802
2803         /*
2804          * If the new end 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 (end & ~HPAGE_PMD_MASK &&
2809             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2810             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2811                 split_huge_page_address(vma->vm_mm, end);
2812
2813         /*
2814          * If we're also updating the vma->vm_next->vm_start, if the new
2815          * vm_next->vm_start isn't page aligned and it could previously
2816          * contain an hugepage: check if we need to split an huge pmd.
2817          */
2818         if (adjust_next > 0) {
2819                 struct vm_area_struct *next = vma->vm_next;
2820                 unsigned long nstart = next->vm_start;
2821                 nstart += adjust_next << PAGE_SHIFT;
2822                 if (nstart & ~HPAGE_PMD_MASK &&
2823                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2824                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2825                         split_huge_page_address(next->vm_mm, nstart);
2826         }
2827 }