mm: mmu_notifier: have mmu_notifiers use a global SRCU so they may safely schedule
[linux-2.6.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/mm_inline.h>
16 #include <linux/kthread.h>
17 #include <linux/khugepaged.h>
18 #include <linux/freezer.h>
19 #include <linux/mman.h>
20 #include <asm/tlb.h>
21 #include <asm/pgalloc.h>
22 #include "internal.h"
23
24 /*
25  * By default transparent hugepage support is enabled for all mappings
26  * and khugepaged scans all mappings. Defrag is only invoked by
27  * khugepaged hugepage allocations and by page faults inside
28  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
29  * allocations.
30  */
31 unsigned long transparent_hugepage_flags __read_mostly =
32 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
33         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
34 #endif
35 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
36         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
37 #endif
38         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
39         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
40
41 /* default scan 8*512 pte (or vmas) every 30 second */
42 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
43 static unsigned int khugepaged_pages_collapsed;
44 static unsigned int khugepaged_full_scans;
45 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
46 /* during fragmentation poll the hugepage allocator once every minute */
47 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
48 static struct task_struct *khugepaged_thread __read_mostly;
49 static DEFINE_MUTEX(khugepaged_mutex);
50 static DEFINE_SPINLOCK(khugepaged_mm_lock);
51 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
52 /*
53  * default collapse hugepages if there is at least one pte mapped like
54  * it would have happened if the vma was large enough during page
55  * fault.
56  */
57 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
58
59 static int khugepaged(void *none);
60 static int mm_slots_hash_init(void);
61 static int khugepaged_slab_init(void);
62 static void khugepaged_slab_free(void);
63
64 #define MM_SLOTS_HASH_HEADS 1024
65 static struct hlist_head *mm_slots_hash __read_mostly;
66 static struct kmem_cache *mm_slot_cache __read_mostly;
67
68 /**
69  * struct mm_slot - hash lookup from mm to mm_slot
70  * @hash: hash collision list
71  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
72  * @mm: the mm that this information is valid for
73  */
74 struct mm_slot {
75         struct hlist_node hash;
76         struct list_head mm_node;
77         struct mm_struct *mm;
78 };
79
80 /**
81  * struct khugepaged_scan - cursor for scanning
82  * @mm_head: the head of the mm list to scan
83  * @mm_slot: the current mm_slot we are scanning
84  * @address: the next address inside that to be scanned
85  *
86  * There is only the one khugepaged_scan instance of this cursor structure.
87  */
88 struct khugepaged_scan {
89         struct list_head mm_head;
90         struct mm_slot *mm_slot;
91         unsigned long address;
92 };
93 static struct khugepaged_scan khugepaged_scan = {
94         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
95 };
96
97
98 static int set_recommended_min_free_kbytes(void)
99 {
100         struct zone *zone;
101         int nr_zones = 0;
102         unsigned long recommended_min;
103         extern int min_free_kbytes;
104
105         if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG,
106                       &transparent_hugepage_flags) &&
107             !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
108                       &transparent_hugepage_flags))
109                 return 0;
110
111         for_each_populated_zone(zone)
112                 nr_zones++;
113
114         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
115         recommended_min = pageblock_nr_pages * nr_zones * 2;
116
117         /*
118          * Make sure that on average at least two pageblocks are almost free
119          * of another type, one for a migratetype to fall back to and a
120          * second to avoid subsequent fallbacks of other types There are 3
121          * MIGRATE_TYPES we care about.
122          */
123         recommended_min += pageblock_nr_pages * nr_zones *
124                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
125
126         /* don't ever allow to reserve more than 5% of the lowmem */
127         recommended_min = min(recommended_min,
128                               (unsigned long) nr_free_buffer_pages() / 20);
129         recommended_min <<= (PAGE_SHIFT-10);
130
131         if (recommended_min > min_free_kbytes)
132                 min_free_kbytes = recommended_min;
133         setup_per_zone_wmarks();
134         return 0;
135 }
136 late_initcall(set_recommended_min_free_kbytes);
137
138 static int start_khugepaged(void)
139 {
140         int err = 0;
141         if (khugepaged_enabled()) {
142                 int wakeup;
143                 if (unlikely(!mm_slot_cache || !mm_slots_hash)) {
144                         err = -ENOMEM;
145                         goto out;
146                 }
147                 mutex_lock(&khugepaged_mutex);
148                 if (!khugepaged_thread)
149                         khugepaged_thread = kthread_run(khugepaged, NULL,
150                                                         "khugepaged");
151                 if (unlikely(IS_ERR(khugepaged_thread))) {
152                         printk(KERN_ERR
153                                "khugepaged: kthread_run(khugepaged) failed\n");
154                         err = PTR_ERR(khugepaged_thread);
155                         khugepaged_thread = NULL;
156                 }
157                 wakeup = !list_empty(&khugepaged_scan.mm_head);
158                 mutex_unlock(&khugepaged_mutex);
159                 if (wakeup)
160                         wake_up_interruptible(&khugepaged_wait);
161
162                 set_recommended_min_free_kbytes();
163         } else
164                 /* wakeup to exit */
165                 wake_up_interruptible(&khugepaged_wait);
166 out:
167         return err;
168 }
169
170 #ifdef CONFIG_SYSFS
171
172 static ssize_t double_flag_show(struct kobject *kobj,
173                                 struct kobj_attribute *attr, char *buf,
174                                 enum transparent_hugepage_flag enabled,
175                                 enum transparent_hugepage_flag req_madv)
176 {
177         if (test_bit(enabled, &transparent_hugepage_flags)) {
178                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
179                 return sprintf(buf, "[always] madvise never\n");
180         } else if (test_bit(req_madv, &transparent_hugepage_flags))
181                 return sprintf(buf, "always [madvise] never\n");
182         else
183                 return sprintf(buf, "always madvise [never]\n");
184 }
185 static ssize_t double_flag_store(struct kobject *kobj,
186                                  struct kobj_attribute *attr,
187                                  const char *buf, size_t count,
188                                  enum transparent_hugepage_flag enabled,
189                                  enum transparent_hugepage_flag req_madv)
190 {
191         if (!memcmp("always", buf,
192                     min(sizeof("always")-1, count))) {
193                 set_bit(enabled, &transparent_hugepage_flags);
194                 clear_bit(req_madv, &transparent_hugepage_flags);
195         } else if (!memcmp("madvise", buf,
196                            min(sizeof("madvise")-1, count))) {
197                 clear_bit(enabled, &transparent_hugepage_flags);
198                 set_bit(req_madv, &transparent_hugepage_flags);
199         } else if (!memcmp("never", buf,
200                            min(sizeof("never")-1, count))) {
201                 clear_bit(enabled, &transparent_hugepage_flags);
202                 clear_bit(req_madv, &transparent_hugepage_flags);
203         } else
204                 return -EINVAL;
205
206         return count;
207 }
208
209 static ssize_t enabled_show(struct kobject *kobj,
210                             struct kobj_attribute *attr, char *buf)
211 {
212         return double_flag_show(kobj, attr, buf,
213                                 TRANSPARENT_HUGEPAGE_FLAG,
214                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
215 }
216 static ssize_t enabled_store(struct kobject *kobj,
217                              struct kobj_attribute *attr,
218                              const char *buf, size_t count)
219 {
220         ssize_t ret;
221
222         ret = double_flag_store(kobj, attr, buf, count,
223                                 TRANSPARENT_HUGEPAGE_FLAG,
224                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
225
226         if (ret > 0) {
227                 int err = start_khugepaged();
228                 if (err)
229                         ret = err;
230         }
231
232         if (ret > 0 &&
233             (test_bit(TRANSPARENT_HUGEPAGE_FLAG,
234                       &transparent_hugepage_flags) ||
235              test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
236                       &transparent_hugepage_flags)))
237                 set_recommended_min_free_kbytes();
238
239         return ret;
240 }
241 static struct kobj_attribute enabled_attr =
242         __ATTR(enabled, 0644, enabled_show, enabled_store);
243
244 static ssize_t single_flag_show(struct kobject *kobj,
245                                 struct kobj_attribute *attr, char *buf,
246                                 enum transparent_hugepage_flag flag)
247 {
248         return sprintf(buf, "%d\n",
249                        !!test_bit(flag, &transparent_hugepage_flags));
250 }
251
252 static ssize_t single_flag_store(struct kobject *kobj,
253                                  struct kobj_attribute *attr,
254                                  const char *buf, size_t count,
255                                  enum transparent_hugepage_flag flag)
256 {
257         unsigned long value;
258         int ret;
259
260         ret = kstrtoul(buf, 10, &value);
261         if (ret < 0)
262                 return ret;
263         if (value > 1)
264                 return -EINVAL;
265
266         if (value)
267                 set_bit(flag, &transparent_hugepage_flags);
268         else
269                 clear_bit(flag, &transparent_hugepage_flags);
270
271         return count;
272 }
273
274 /*
275  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
276  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
277  * memory just to allocate one more hugepage.
278  */
279 static ssize_t defrag_show(struct kobject *kobj,
280                            struct kobj_attribute *attr, char *buf)
281 {
282         return double_flag_show(kobj, attr, buf,
283                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
284                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
285 }
286 static ssize_t defrag_store(struct kobject *kobj,
287                             struct kobj_attribute *attr,
288                             const char *buf, size_t count)
289 {
290         return double_flag_store(kobj, attr, buf, count,
291                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
292                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
293 }
294 static struct kobj_attribute defrag_attr =
295         __ATTR(defrag, 0644, defrag_show, defrag_store);
296
297 #ifdef CONFIG_DEBUG_VM
298 static ssize_t debug_cow_show(struct kobject *kobj,
299                                 struct kobj_attribute *attr, char *buf)
300 {
301         return single_flag_show(kobj, attr, buf,
302                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
303 }
304 static ssize_t debug_cow_store(struct kobject *kobj,
305                                struct kobj_attribute *attr,
306                                const char *buf, size_t count)
307 {
308         return single_flag_store(kobj, attr, buf, count,
309                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
310 }
311 static struct kobj_attribute debug_cow_attr =
312         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
313 #endif /* CONFIG_DEBUG_VM */
314
315 static struct attribute *hugepage_attr[] = {
316         &enabled_attr.attr,
317         &defrag_attr.attr,
318 #ifdef CONFIG_DEBUG_VM
319         &debug_cow_attr.attr,
320 #endif
321         NULL,
322 };
323
324 static struct attribute_group hugepage_attr_group = {
325         .attrs = hugepage_attr,
326 };
327
328 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
329                                          struct kobj_attribute *attr,
330                                          char *buf)
331 {
332         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
333 }
334
335 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
336                                           struct kobj_attribute *attr,
337                                           const char *buf, size_t count)
338 {
339         unsigned long msecs;
340         int err;
341
342         err = strict_strtoul(buf, 10, &msecs);
343         if (err || msecs > UINT_MAX)
344                 return -EINVAL;
345
346         khugepaged_scan_sleep_millisecs = msecs;
347         wake_up_interruptible(&khugepaged_wait);
348
349         return count;
350 }
351 static struct kobj_attribute scan_sleep_millisecs_attr =
352         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
353                scan_sleep_millisecs_store);
354
355 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
356                                           struct kobj_attribute *attr,
357                                           char *buf)
358 {
359         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
360 }
361
362 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
363                                            struct kobj_attribute *attr,
364                                            const char *buf, size_t count)
365 {
366         unsigned long msecs;
367         int err;
368
369         err = strict_strtoul(buf, 10, &msecs);
370         if (err || msecs > UINT_MAX)
371                 return -EINVAL;
372
373         khugepaged_alloc_sleep_millisecs = msecs;
374         wake_up_interruptible(&khugepaged_wait);
375
376         return count;
377 }
378 static struct kobj_attribute alloc_sleep_millisecs_attr =
379         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
380                alloc_sleep_millisecs_store);
381
382 static ssize_t pages_to_scan_show(struct kobject *kobj,
383                                   struct kobj_attribute *attr,
384                                   char *buf)
385 {
386         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
387 }
388 static ssize_t pages_to_scan_store(struct kobject *kobj,
389                                    struct kobj_attribute *attr,
390                                    const char *buf, size_t count)
391 {
392         int err;
393         unsigned long pages;
394
395         err = strict_strtoul(buf, 10, &pages);
396         if (err || !pages || pages > UINT_MAX)
397                 return -EINVAL;
398
399         khugepaged_pages_to_scan = pages;
400
401         return count;
402 }
403 static struct kobj_attribute pages_to_scan_attr =
404         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
405                pages_to_scan_store);
406
407 static ssize_t pages_collapsed_show(struct kobject *kobj,
408                                     struct kobj_attribute *attr,
409                                     char *buf)
410 {
411         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
412 }
413 static struct kobj_attribute pages_collapsed_attr =
414         __ATTR_RO(pages_collapsed);
415
416 static ssize_t full_scans_show(struct kobject *kobj,
417                                struct kobj_attribute *attr,
418                                char *buf)
419 {
420         return sprintf(buf, "%u\n", khugepaged_full_scans);
421 }
422 static struct kobj_attribute full_scans_attr =
423         __ATTR_RO(full_scans);
424
425 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
426                                       struct kobj_attribute *attr, char *buf)
427 {
428         return single_flag_show(kobj, attr, buf,
429                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
430 }
431 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
432                                        struct kobj_attribute *attr,
433                                        const char *buf, size_t count)
434 {
435         return single_flag_store(kobj, attr, buf, count,
436                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
437 }
438 static struct kobj_attribute khugepaged_defrag_attr =
439         __ATTR(defrag, 0644, khugepaged_defrag_show,
440                khugepaged_defrag_store);
441
442 /*
443  * max_ptes_none controls if khugepaged should collapse hugepages over
444  * any unmapped ptes in turn potentially increasing the memory
445  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
446  * reduce the available free memory in the system as it
447  * runs. Increasing max_ptes_none will instead potentially reduce the
448  * free memory in the system during the khugepaged scan.
449  */
450 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
451                                              struct kobj_attribute *attr,
452                                              char *buf)
453 {
454         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
455 }
456 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
457                                               struct kobj_attribute *attr,
458                                               const char *buf, size_t count)
459 {
460         int err;
461         unsigned long max_ptes_none;
462
463         err = strict_strtoul(buf, 10, &max_ptes_none);
464         if (err || max_ptes_none > HPAGE_PMD_NR-1)
465                 return -EINVAL;
466
467         khugepaged_max_ptes_none = max_ptes_none;
468
469         return count;
470 }
471 static struct kobj_attribute khugepaged_max_ptes_none_attr =
472         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
473                khugepaged_max_ptes_none_store);
474
475 static struct attribute *khugepaged_attr[] = {
476         &khugepaged_defrag_attr.attr,
477         &khugepaged_max_ptes_none_attr.attr,
478         &pages_to_scan_attr.attr,
479         &pages_collapsed_attr.attr,
480         &full_scans_attr.attr,
481         &scan_sleep_millisecs_attr.attr,
482         &alloc_sleep_millisecs_attr.attr,
483         NULL,
484 };
485
486 static struct attribute_group khugepaged_attr_group = {
487         .attrs = khugepaged_attr,
488         .name = "khugepaged",
489 };
490
491 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
492 {
493         int err;
494
495         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
496         if (unlikely(!*hugepage_kobj)) {
497                 printk(KERN_ERR "hugepage: failed kobject create\n");
498                 return -ENOMEM;
499         }
500
501         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
502         if (err) {
503                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
504                 goto delete_obj;
505         }
506
507         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
508         if (err) {
509                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
510                 goto remove_hp_group;
511         }
512
513         return 0;
514
515 remove_hp_group:
516         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
517 delete_obj:
518         kobject_put(*hugepage_kobj);
519         return err;
520 }
521
522 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
523 {
524         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
525         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
526         kobject_put(hugepage_kobj);
527 }
528 #else
529 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
530 {
531         return 0;
532 }
533
534 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
535 {
536 }
537 #endif /* CONFIG_SYSFS */
538
539 static int __init hugepage_init(void)
540 {
541         int err;
542         struct kobject *hugepage_kobj;
543
544         if (!has_transparent_hugepage()) {
545                 transparent_hugepage_flags = 0;
546                 return -EINVAL;
547         }
548
549         err = hugepage_init_sysfs(&hugepage_kobj);
550         if (err)
551                 return err;
552
553         err = khugepaged_slab_init();
554         if (err)
555                 goto out;
556
557         err = mm_slots_hash_init();
558         if (err) {
559                 khugepaged_slab_free();
560                 goto out;
561         }
562
563         /*
564          * By default disable transparent hugepages on smaller systems,
565          * where the extra memory used could hurt more than TLB overhead
566          * is likely to save.  The admin can still enable it through /sys.
567          */
568         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
569                 transparent_hugepage_flags = 0;
570
571         start_khugepaged();
572
573         set_recommended_min_free_kbytes();
574
575         return 0;
576 out:
577         hugepage_exit_sysfs(hugepage_kobj);
578         return err;
579 }
580 module_init(hugepage_init)
581
582 static int __init setup_transparent_hugepage(char *str)
583 {
584         int ret = 0;
585         if (!str)
586                 goto out;
587         if (!strcmp(str, "always")) {
588                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
589                         &transparent_hugepage_flags);
590                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
591                           &transparent_hugepage_flags);
592                 ret = 1;
593         } else if (!strcmp(str, "madvise")) {
594                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
595                           &transparent_hugepage_flags);
596                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
597                         &transparent_hugepage_flags);
598                 ret = 1;
599         } else if (!strcmp(str, "never")) {
600                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
601                           &transparent_hugepage_flags);
602                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
603                           &transparent_hugepage_flags);
604                 ret = 1;
605         }
606 out:
607         if (!ret)
608                 printk(KERN_WARNING
609                        "transparent_hugepage= cannot parse, ignored\n");
610         return ret;
611 }
612 __setup("transparent_hugepage=", setup_transparent_hugepage);
613
614 static void prepare_pmd_huge_pte(pgtable_t pgtable,
615                                  struct mm_struct *mm)
616 {
617         assert_spin_locked(&mm->page_table_lock);
618
619         /* FIFO */
620         if (!mm->pmd_huge_pte)
621                 INIT_LIST_HEAD(&pgtable->lru);
622         else
623                 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
624         mm->pmd_huge_pte = pgtable;
625 }
626
627 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
628 {
629         if (likely(vma->vm_flags & VM_WRITE))
630                 pmd = pmd_mkwrite(pmd);
631         return pmd;
632 }
633
634 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
635                                         struct vm_area_struct *vma,
636                                         unsigned long haddr, pmd_t *pmd,
637                                         struct page *page)
638 {
639         int ret = 0;
640         pgtable_t pgtable;
641
642         VM_BUG_ON(!PageCompound(page));
643         pgtable = pte_alloc_one(mm, haddr);
644         if (unlikely(!pgtable)) {
645                 mem_cgroup_uncharge_page(page);
646                 put_page(page);
647                 return VM_FAULT_OOM;
648         }
649
650         clear_huge_page(page, haddr, HPAGE_PMD_NR);
651         __SetPageUptodate(page);
652
653         spin_lock(&mm->page_table_lock);
654         if (unlikely(!pmd_none(*pmd))) {
655                 spin_unlock(&mm->page_table_lock);
656                 mem_cgroup_uncharge_page(page);
657                 put_page(page);
658                 pte_free(mm, pgtable);
659         } else {
660                 pmd_t entry;
661                 entry = mk_pmd(page, vma->vm_page_prot);
662                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
663                 entry = pmd_mkhuge(entry);
664                 /*
665                  * The spinlocking to take the lru_lock inside
666                  * page_add_new_anon_rmap() acts as a full memory
667                  * barrier to be sure clear_huge_page writes become
668                  * visible after the set_pmd_at() write.
669                  */
670                 page_add_new_anon_rmap(page, vma, haddr);
671                 set_pmd_at(mm, haddr, pmd, entry);
672                 prepare_pmd_huge_pte(pgtable, mm);
673                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
674                 mm->nr_ptes++;
675                 spin_unlock(&mm->page_table_lock);
676         }
677
678         return ret;
679 }
680
681 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
682 {
683         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
684 }
685
686 static inline struct page *alloc_hugepage_vma(int defrag,
687                                               struct vm_area_struct *vma,
688                                               unsigned long haddr, int nd,
689                                               gfp_t extra_gfp)
690 {
691         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
692                                HPAGE_PMD_ORDER, vma, haddr, nd);
693 }
694
695 #ifndef CONFIG_NUMA
696 static inline struct page *alloc_hugepage(int defrag)
697 {
698         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
699                            HPAGE_PMD_ORDER);
700 }
701 #endif
702
703 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
704                                unsigned long address, pmd_t *pmd,
705                                unsigned int flags)
706 {
707         struct page *page;
708         unsigned long haddr = address & HPAGE_PMD_MASK;
709         pte_t *pte;
710
711         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
712                 if (unlikely(anon_vma_prepare(vma)))
713                         return VM_FAULT_OOM;
714                 if (unlikely(khugepaged_enter(vma)))
715                         return VM_FAULT_OOM;
716                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
717                                           vma, haddr, numa_node_id(), 0);
718                 if (unlikely(!page)) {
719                         count_vm_event(THP_FAULT_FALLBACK);
720                         goto out;
721                 }
722                 count_vm_event(THP_FAULT_ALLOC);
723                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
724                         put_page(page);
725                         goto out;
726                 }
727
728                 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
729         }
730 out:
731         /*
732          * Use __pte_alloc instead of pte_alloc_map, because we can't
733          * run pte_offset_map on the pmd, if an huge pmd could
734          * materialize from under us from a different thread.
735          */
736         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
737                 return VM_FAULT_OOM;
738         /* if an huge pmd materialized from under us just retry later */
739         if (unlikely(pmd_trans_huge(*pmd)))
740                 return 0;
741         /*
742          * A regular pmd is established and it can't morph into a huge pmd
743          * from under us anymore at this point because we hold the mmap_sem
744          * read mode and khugepaged takes it in write mode. So now it's
745          * safe to run pte_offset_map().
746          */
747         pte = pte_offset_map(pmd, address);
748         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
749 }
750
751 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
752                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
753                   struct vm_area_struct *vma)
754 {
755         struct page *src_page;
756         pmd_t pmd;
757         pgtable_t pgtable;
758         int ret;
759
760         ret = -ENOMEM;
761         pgtable = pte_alloc_one(dst_mm, addr);
762         if (unlikely(!pgtable))
763                 goto out;
764
765         spin_lock(&dst_mm->page_table_lock);
766         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
767
768         ret = -EAGAIN;
769         pmd = *src_pmd;
770         if (unlikely(!pmd_trans_huge(pmd))) {
771                 pte_free(dst_mm, pgtable);
772                 goto out_unlock;
773         }
774         if (unlikely(pmd_trans_splitting(pmd))) {
775                 /* split huge page running from under us */
776                 spin_unlock(&src_mm->page_table_lock);
777                 spin_unlock(&dst_mm->page_table_lock);
778                 pte_free(dst_mm, pgtable);
779
780                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
781                 goto out;
782         }
783         src_page = pmd_page(pmd);
784         VM_BUG_ON(!PageHead(src_page));
785         get_page(src_page);
786         page_dup_rmap(src_page);
787         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
788
789         pmdp_set_wrprotect(src_mm, addr, src_pmd);
790         pmd = pmd_mkold(pmd_wrprotect(pmd));
791         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
792         prepare_pmd_huge_pte(pgtable, dst_mm);
793         dst_mm->nr_ptes++;
794
795         ret = 0;
796 out_unlock:
797         spin_unlock(&src_mm->page_table_lock);
798         spin_unlock(&dst_mm->page_table_lock);
799 out:
800         return ret;
801 }
802
803 /* no "address" argument so destroys page coloring of some arch */
804 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
805 {
806         pgtable_t pgtable;
807
808         assert_spin_locked(&mm->page_table_lock);
809
810         /* FIFO */
811         pgtable = mm->pmd_huge_pte;
812         if (list_empty(&pgtable->lru))
813                 mm->pmd_huge_pte = NULL;
814         else {
815                 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
816                                               struct page, lru);
817                 list_del(&pgtable->lru);
818         }
819         return pgtable;
820 }
821
822 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
823                                         struct vm_area_struct *vma,
824                                         unsigned long address,
825                                         pmd_t *pmd, pmd_t orig_pmd,
826                                         struct page *page,
827                                         unsigned long haddr)
828 {
829         pgtable_t pgtable;
830         pmd_t _pmd;
831         int ret = 0, i;
832         struct page **pages;
833
834         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
835                         GFP_KERNEL);
836         if (unlikely(!pages)) {
837                 ret |= VM_FAULT_OOM;
838                 goto out;
839         }
840
841         for (i = 0; i < HPAGE_PMD_NR; i++) {
842                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
843                                                __GFP_OTHER_NODE,
844                                                vma, address, page_to_nid(page));
845                 if (unlikely(!pages[i] ||
846                              mem_cgroup_newpage_charge(pages[i], mm,
847                                                        GFP_KERNEL))) {
848                         if (pages[i])
849                                 put_page(pages[i]);
850                         mem_cgroup_uncharge_start();
851                         while (--i >= 0) {
852                                 mem_cgroup_uncharge_page(pages[i]);
853                                 put_page(pages[i]);
854                         }
855                         mem_cgroup_uncharge_end();
856                         kfree(pages);
857                         ret |= VM_FAULT_OOM;
858                         goto out;
859                 }
860         }
861
862         for (i = 0; i < HPAGE_PMD_NR; i++) {
863                 copy_user_highpage(pages[i], page + i,
864                                    haddr + PAGE_SIZE * i, vma);
865                 __SetPageUptodate(pages[i]);
866                 cond_resched();
867         }
868
869         spin_lock(&mm->page_table_lock);
870         if (unlikely(!pmd_same(*pmd, orig_pmd)))
871                 goto out_free_pages;
872         VM_BUG_ON(!PageHead(page));
873
874         pmdp_clear_flush_notify(vma, haddr, pmd);
875         /* leave pmd empty until pte is filled */
876
877         pgtable = get_pmd_huge_pte(mm);
878         pmd_populate(mm, &_pmd, pgtable);
879
880         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
881                 pte_t *pte, entry;
882                 entry = mk_pte(pages[i], vma->vm_page_prot);
883                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
884                 page_add_new_anon_rmap(pages[i], vma, haddr);
885                 pte = pte_offset_map(&_pmd, haddr);
886                 VM_BUG_ON(!pte_none(*pte));
887                 set_pte_at(mm, haddr, pte, entry);
888                 pte_unmap(pte);
889         }
890         kfree(pages);
891
892         smp_wmb(); /* make pte visible before pmd */
893         pmd_populate(mm, pmd, pgtable);
894         page_remove_rmap(page);
895         spin_unlock(&mm->page_table_lock);
896
897         ret |= VM_FAULT_WRITE;
898         put_page(page);
899
900 out:
901         return ret;
902
903 out_free_pages:
904         spin_unlock(&mm->page_table_lock);
905         mem_cgroup_uncharge_start();
906         for (i = 0; i < HPAGE_PMD_NR; i++) {
907                 mem_cgroup_uncharge_page(pages[i]);
908                 put_page(pages[i]);
909         }
910         mem_cgroup_uncharge_end();
911         kfree(pages);
912         goto out;
913 }
914
915 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
916                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
917 {
918         int ret = 0;
919         struct page *page, *new_page;
920         unsigned long haddr;
921
922         VM_BUG_ON(!vma->anon_vma);
923         spin_lock(&mm->page_table_lock);
924         if (unlikely(!pmd_same(*pmd, orig_pmd)))
925                 goto out_unlock;
926
927         page = pmd_page(orig_pmd);
928         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
929         haddr = address & HPAGE_PMD_MASK;
930         if (page_mapcount(page) == 1) {
931                 pmd_t entry;
932                 entry = pmd_mkyoung(orig_pmd);
933                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
934                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
935                         update_mmu_cache(vma, address, entry);
936                 ret |= VM_FAULT_WRITE;
937                 goto out_unlock;
938         }
939         get_page(page);
940         spin_unlock(&mm->page_table_lock);
941
942         if (transparent_hugepage_enabled(vma) &&
943             !transparent_hugepage_debug_cow())
944                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
945                                               vma, haddr, numa_node_id(), 0);
946         else
947                 new_page = NULL;
948
949         if (unlikely(!new_page)) {
950                 count_vm_event(THP_FAULT_FALLBACK);
951                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
952                                                    pmd, orig_pmd, page, haddr);
953                 if (ret & VM_FAULT_OOM)
954                         split_huge_page(page);
955                 put_page(page);
956                 goto out;
957         }
958         count_vm_event(THP_FAULT_ALLOC);
959
960         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
961                 put_page(new_page);
962                 split_huge_page(page);
963                 put_page(page);
964                 ret |= VM_FAULT_OOM;
965                 goto out;
966         }
967
968         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
969         __SetPageUptodate(new_page);
970
971         spin_lock(&mm->page_table_lock);
972         put_page(page);
973         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
974                 mem_cgroup_uncharge_page(new_page);
975                 put_page(new_page);
976         } else {
977                 pmd_t entry;
978                 VM_BUG_ON(!PageHead(page));
979                 entry = mk_pmd(new_page, vma->vm_page_prot);
980                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
981                 entry = pmd_mkhuge(entry);
982                 pmdp_clear_flush_notify(vma, haddr, pmd);
983                 page_add_new_anon_rmap(new_page, vma, haddr);
984                 set_pmd_at(mm, haddr, pmd, entry);
985                 update_mmu_cache(vma, address, entry);
986                 page_remove_rmap(page);
987                 put_page(page);
988                 ret |= VM_FAULT_WRITE;
989         }
990 out_unlock:
991         spin_unlock(&mm->page_table_lock);
992 out:
993         return ret;
994 }
995
996 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
997                                    unsigned long addr,
998                                    pmd_t *pmd,
999                                    unsigned int flags)
1000 {
1001         struct page *page = NULL;
1002
1003         assert_spin_locked(&mm->page_table_lock);
1004
1005         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1006                 goto out;
1007
1008         page = pmd_page(*pmd);
1009         VM_BUG_ON(!PageHead(page));
1010         if (flags & FOLL_TOUCH) {
1011                 pmd_t _pmd;
1012                 /*
1013                  * We should set the dirty bit only for FOLL_WRITE but
1014                  * for now the dirty bit in the pmd is meaningless.
1015                  * And if the dirty bit will become meaningful and
1016                  * we'll only set it with FOLL_WRITE, an atomic
1017                  * set_bit will be required on the pmd to set the
1018                  * young bit, instead of the current set_pmd_at.
1019                  */
1020                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1021                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1022         }
1023         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1024         VM_BUG_ON(!PageCompound(page));
1025         if (flags & FOLL_GET)
1026                 get_page_foll(page);
1027
1028 out:
1029         return page;
1030 }
1031
1032 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1033                  pmd_t *pmd, unsigned long addr)
1034 {
1035         int ret = 0;
1036
1037         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1038                 struct page *page;
1039                 pgtable_t pgtable;
1040                 pgtable = get_pmd_huge_pte(tlb->mm);
1041                 page = pmd_page(*pmd);
1042                 pmd_clear(pmd);
1043                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1044                 page_remove_rmap(page);
1045                 VM_BUG_ON(page_mapcount(page) < 0);
1046                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1047                 VM_BUG_ON(!PageHead(page));
1048                 tlb->mm->nr_ptes--;
1049                 spin_unlock(&tlb->mm->page_table_lock);
1050                 tlb_remove_page(tlb, page);
1051                 pte_free(tlb->mm, pgtable);
1052                 ret = 1;
1053         }
1054         return ret;
1055 }
1056
1057 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1058                 unsigned long addr, unsigned long end,
1059                 unsigned char *vec)
1060 {
1061         int ret = 0;
1062
1063         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1064                 /*
1065                  * All logical pages in the range are present
1066                  * if backed by a huge page.
1067                  */
1068                 spin_unlock(&vma->vm_mm->page_table_lock);
1069                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1070                 ret = 1;
1071         }
1072
1073         return ret;
1074 }
1075
1076 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1077                   unsigned long old_addr,
1078                   unsigned long new_addr, unsigned long old_end,
1079                   pmd_t *old_pmd, pmd_t *new_pmd)
1080 {
1081         int ret = 0;
1082         pmd_t pmd;
1083
1084         struct mm_struct *mm = vma->vm_mm;
1085
1086         if ((old_addr & ~HPAGE_PMD_MASK) ||
1087             (new_addr & ~HPAGE_PMD_MASK) ||
1088             old_end - old_addr < HPAGE_PMD_SIZE ||
1089             (new_vma->vm_flags & VM_NOHUGEPAGE))
1090                 goto out;
1091
1092         /*
1093          * The destination pmd shouldn't be established, free_pgtables()
1094          * should have release it.
1095          */
1096         if (WARN_ON(!pmd_none(*new_pmd))) {
1097                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1098                 goto out;
1099         }
1100
1101         ret = __pmd_trans_huge_lock(old_pmd, vma);
1102         if (ret == 1) {
1103                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1104                 VM_BUG_ON(!pmd_none(*new_pmd));
1105                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1106                 spin_unlock(&mm->page_table_lock);
1107         }
1108 out:
1109         return ret;
1110 }
1111
1112 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1113                 unsigned long addr, pgprot_t newprot)
1114 {
1115         struct mm_struct *mm = vma->vm_mm;
1116         int ret = 0;
1117
1118         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1119                 pmd_t entry;
1120                 entry = pmdp_get_and_clear(mm, addr, pmd);
1121                 entry = pmd_modify(entry, newprot);
1122                 set_pmd_at(mm, addr, pmd, entry);
1123                 spin_unlock(&vma->vm_mm->page_table_lock);
1124                 ret = 1;
1125         }
1126
1127         return ret;
1128 }
1129
1130 /*
1131  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1132  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1133  *
1134  * Note that if it returns 1, this routine returns without unlocking page
1135  * table locks. So callers must unlock them.
1136  */
1137 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1138 {
1139         spin_lock(&vma->vm_mm->page_table_lock);
1140         if (likely(pmd_trans_huge(*pmd))) {
1141                 if (unlikely(pmd_trans_splitting(*pmd))) {
1142                         spin_unlock(&vma->vm_mm->page_table_lock);
1143                         wait_split_huge_page(vma->anon_vma, pmd);
1144                         return -1;
1145                 } else {
1146                         /* Thp mapped by 'pmd' is stable, so we can
1147                          * handle it as it is. */
1148                         return 1;
1149                 }
1150         }
1151         spin_unlock(&vma->vm_mm->page_table_lock);
1152         return 0;
1153 }
1154
1155 pmd_t *page_check_address_pmd(struct page *page,
1156                               struct mm_struct *mm,
1157                               unsigned long address,
1158                               enum page_check_address_pmd_flag flag)
1159 {
1160         pgd_t *pgd;
1161         pud_t *pud;
1162         pmd_t *pmd, *ret = NULL;
1163
1164         if (address & ~HPAGE_PMD_MASK)
1165                 goto out;
1166
1167         pgd = pgd_offset(mm, address);
1168         if (!pgd_present(*pgd))
1169                 goto out;
1170
1171         pud = pud_offset(pgd, address);
1172         if (!pud_present(*pud))
1173                 goto out;
1174
1175         pmd = pmd_offset(pud, address);
1176         if (pmd_none(*pmd))
1177                 goto out;
1178         if (pmd_page(*pmd) != page)
1179                 goto out;
1180         /*
1181          * split_vma() may create temporary aliased mappings. There is
1182          * no risk as long as all huge pmd are found and have their
1183          * splitting bit set before __split_huge_page_refcount
1184          * runs. Finding the same huge pmd more than once during the
1185          * same rmap walk is not a problem.
1186          */
1187         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1188             pmd_trans_splitting(*pmd))
1189                 goto out;
1190         if (pmd_trans_huge(*pmd)) {
1191                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1192                           !pmd_trans_splitting(*pmd));
1193                 ret = pmd;
1194         }
1195 out:
1196         return ret;
1197 }
1198
1199 static int __split_huge_page_splitting(struct page *page,
1200                                        struct vm_area_struct *vma,
1201                                        unsigned long address)
1202 {
1203         struct mm_struct *mm = vma->vm_mm;
1204         pmd_t *pmd;
1205         int ret = 0;
1206
1207         spin_lock(&mm->page_table_lock);
1208         pmd = page_check_address_pmd(page, mm, address,
1209                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1210         if (pmd) {
1211                 /*
1212                  * We can't temporarily set the pmd to null in order
1213                  * to split it, the pmd must remain marked huge at all
1214                  * times or the VM won't take the pmd_trans_huge paths
1215                  * and it won't wait on the anon_vma->root->mutex to
1216                  * serialize against split_huge_page*.
1217                  */
1218                 pmdp_splitting_flush_notify(vma, address, pmd);
1219                 ret = 1;
1220         }
1221         spin_unlock(&mm->page_table_lock);
1222
1223         return ret;
1224 }
1225
1226 static void __split_huge_page_refcount(struct page *page)
1227 {
1228         int i;
1229         struct zone *zone = page_zone(page);
1230         int tail_count = 0;
1231
1232         /* prevent PageLRU to go away from under us, and freeze lru stats */
1233         spin_lock_irq(&zone->lru_lock);
1234         compound_lock(page);
1235         /* complete memcg works before add pages to LRU */
1236         mem_cgroup_split_huge_fixup(page);
1237
1238         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1239                 struct page *page_tail = page + i;
1240
1241                 /* tail_page->_mapcount cannot change */
1242                 BUG_ON(page_mapcount(page_tail) < 0);
1243                 tail_count += page_mapcount(page_tail);
1244                 /* check for overflow */
1245                 BUG_ON(tail_count < 0);
1246                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1247                 /*
1248                  * tail_page->_count is zero and not changing from
1249                  * under us. But get_page_unless_zero() may be running
1250                  * from under us on the tail_page. If we used
1251                  * atomic_set() below instead of atomic_add(), we
1252                  * would then run atomic_set() concurrently with
1253                  * get_page_unless_zero(), and atomic_set() is
1254                  * implemented in C not using locked ops. spin_unlock
1255                  * on x86 sometime uses locked ops because of PPro
1256                  * errata 66, 92, so unless somebody can guarantee
1257                  * atomic_set() here would be safe on all archs (and
1258                  * not only on x86), it's safer to use atomic_add().
1259                  */
1260                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1261                            &page_tail->_count);
1262
1263                 /* after clearing PageTail the gup refcount can be released */
1264                 smp_mb();
1265
1266                 /*
1267                  * retain hwpoison flag of the poisoned tail page:
1268                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1269                  *   by the memory-failure.
1270                  */
1271                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1272                 page_tail->flags |= (page->flags &
1273                                      ((1L << PG_referenced) |
1274                                       (1L << PG_swapbacked) |
1275                                       (1L << PG_mlocked) |
1276                                       (1L << PG_uptodate)));
1277                 page_tail->flags |= (1L << PG_dirty);
1278
1279                 /* clear PageTail before overwriting first_page */
1280                 smp_wmb();
1281
1282                 /*
1283                  * __split_huge_page_splitting() already set the
1284                  * splitting bit in all pmd that could map this
1285                  * hugepage, that will ensure no CPU can alter the
1286                  * mapcount on the head page. The mapcount is only
1287                  * accounted in the head page and it has to be
1288                  * transferred to all tail pages in the below code. So
1289                  * for this code to be safe, the split the mapcount
1290                  * can't change. But that doesn't mean userland can't
1291                  * keep changing and reading the page contents while
1292                  * we transfer the mapcount, so the pmd splitting
1293                  * status is achieved setting a reserved bit in the
1294                  * pmd, not by clearing the present bit.
1295                 */
1296                 page_tail->_mapcount = page->_mapcount;
1297
1298                 BUG_ON(page_tail->mapping);
1299                 page_tail->mapping = page->mapping;
1300
1301                 page_tail->index = page->index + i;
1302
1303                 BUG_ON(!PageAnon(page_tail));
1304                 BUG_ON(!PageUptodate(page_tail));
1305                 BUG_ON(!PageDirty(page_tail));
1306                 BUG_ON(!PageSwapBacked(page_tail));
1307
1308
1309                 lru_add_page_tail(zone, page, page_tail);
1310         }
1311         atomic_sub(tail_count, &page->_count);
1312         BUG_ON(atomic_read(&page->_count) <= 0);
1313
1314         __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1315         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1316
1317         ClearPageCompound(page);
1318         compound_unlock(page);
1319         spin_unlock_irq(&zone->lru_lock);
1320
1321         for (i = 1; i < HPAGE_PMD_NR; i++) {
1322                 struct page *page_tail = page + i;
1323                 BUG_ON(page_count(page_tail) <= 0);
1324                 /*
1325                  * Tail pages may be freed if there wasn't any mapping
1326                  * like if add_to_swap() is running on a lru page that
1327                  * had its mapping zapped. And freeing these pages
1328                  * requires taking the lru_lock so we do the put_page
1329                  * of the tail pages after the split is complete.
1330                  */
1331                 put_page(page_tail);
1332         }
1333
1334         /*
1335          * Only the head page (now become a regular page) is required
1336          * to be pinned by the caller.
1337          */
1338         BUG_ON(page_count(page) <= 0);
1339 }
1340
1341 static int __split_huge_page_map(struct page *page,
1342                                  struct vm_area_struct *vma,
1343                                  unsigned long address)
1344 {
1345         struct mm_struct *mm = vma->vm_mm;
1346         pmd_t *pmd, _pmd;
1347         int ret = 0, i;
1348         pgtable_t pgtable;
1349         unsigned long haddr;
1350
1351         spin_lock(&mm->page_table_lock);
1352         pmd = page_check_address_pmd(page, mm, address,
1353                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1354         if (pmd) {
1355                 pgtable = get_pmd_huge_pte(mm);
1356                 pmd_populate(mm, &_pmd, pgtable);
1357
1358                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1359                      i++, haddr += PAGE_SIZE) {
1360                         pte_t *pte, entry;
1361                         BUG_ON(PageCompound(page+i));
1362                         entry = mk_pte(page + i, vma->vm_page_prot);
1363                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1364                         if (!pmd_write(*pmd))
1365                                 entry = pte_wrprotect(entry);
1366                         else
1367                                 BUG_ON(page_mapcount(page) != 1);
1368                         if (!pmd_young(*pmd))
1369                                 entry = pte_mkold(entry);
1370                         pte = pte_offset_map(&_pmd, haddr);
1371                         BUG_ON(!pte_none(*pte));
1372                         set_pte_at(mm, haddr, pte, entry);
1373                         pte_unmap(pte);
1374                 }
1375
1376                 smp_wmb(); /* make pte visible before pmd */
1377                 /*
1378                  * Up to this point the pmd is present and huge and
1379                  * userland has the whole access to the hugepage
1380                  * during the split (which happens in place). If we
1381                  * overwrite the pmd with the not-huge version
1382                  * pointing to the pte here (which of course we could
1383                  * if all CPUs were bug free), userland could trigger
1384                  * a small page size TLB miss on the small sized TLB
1385                  * while the hugepage TLB entry is still established
1386                  * in the huge TLB. Some CPU doesn't like that. See
1387                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1388                  * Erratum 383 on page 93. Intel should be safe but is
1389                  * also warns that it's only safe if the permission
1390                  * and cache attributes of the two entries loaded in
1391                  * the two TLB is identical (which should be the case
1392                  * here). But it is generally safer to never allow
1393                  * small and huge TLB entries for the same virtual
1394                  * address to be loaded simultaneously. So instead of
1395                  * doing "pmd_populate(); flush_tlb_range();" we first
1396                  * mark the current pmd notpresent (atomically because
1397                  * here the pmd_trans_huge and pmd_trans_splitting
1398                  * must remain set at all times on the pmd until the
1399                  * split is complete for this pmd), then we flush the
1400                  * SMP TLB and finally we write the non-huge version
1401                  * of the pmd entry with pmd_populate.
1402                  */
1403                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1404                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1405                 pmd_populate(mm, pmd, pgtable);
1406                 ret = 1;
1407         }
1408         spin_unlock(&mm->page_table_lock);
1409
1410         return ret;
1411 }
1412
1413 /* must be called with anon_vma->root->mutex hold */
1414 static void __split_huge_page(struct page *page,
1415                               struct anon_vma *anon_vma)
1416 {
1417         int mapcount, mapcount2;
1418         struct anon_vma_chain *avc;
1419
1420         BUG_ON(!PageHead(page));
1421         BUG_ON(PageTail(page));
1422
1423         mapcount = 0;
1424         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1425                 struct vm_area_struct *vma = avc->vma;
1426                 unsigned long addr = vma_address(page, vma);
1427                 BUG_ON(is_vma_temporary_stack(vma));
1428                 if (addr == -EFAULT)
1429                         continue;
1430                 mapcount += __split_huge_page_splitting(page, vma, addr);
1431         }
1432         /*
1433          * It is critical that new vmas are added to the tail of the
1434          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1435          * and establishes a child pmd before
1436          * __split_huge_page_splitting() freezes the parent pmd (so if
1437          * we fail to prevent copy_huge_pmd() from running until the
1438          * whole __split_huge_page() is complete), we will still see
1439          * the newly established pmd of the child later during the
1440          * walk, to be able to set it as pmd_trans_splitting too.
1441          */
1442         if (mapcount != page_mapcount(page))
1443                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1444                        mapcount, page_mapcount(page));
1445         BUG_ON(mapcount != page_mapcount(page));
1446
1447         __split_huge_page_refcount(page);
1448
1449         mapcount2 = 0;
1450         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1451                 struct vm_area_struct *vma = avc->vma;
1452                 unsigned long addr = vma_address(page, vma);
1453                 BUG_ON(is_vma_temporary_stack(vma));
1454                 if (addr == -EFAULT)
1455                         continue;
1456                 mapcount2 += __split_huge_page_map(page, vma, addr);
1457         }
1458         if (mapcount != mapcount2)
1459                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1460                        mapcount, mapcount2, page_mapcount(page));
1461         BUG_ON(mapcount != mapcount2);
1462 }
1463
1464 int split_huge_page(struct page *page)
1465 {
1466         struct anon_vma *anon_vma;
1467         int ret = 1;
1468
1469         BUG_ON(!PageAnon(page));
1470         anon_vma = page_lock_anon_vma(page);
1471         if (!anon_vma)
1472                 goto out;
1473         ret = 0;
1474         if (!PageCompound(page))
1475                 goto out_unlock;
1476
1477         BUG_ON(!PageSwapBacked(page));
1478         __split_huge_page(page, anon_vma);
1479         count_vm_event(THP_SPLIT);
1480
1481         BUG_ON(PageCompound(page));
1482 out_unlock:
1483         page_unlock_anon_vma(anon_vma);
1484 out:
1485         return ret;
1486 }
1487
1488 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1489                    VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1490
1491 int hugepage_madvise(struct vm_area_struct *vma,
1492                      unsigned long *vm_flags, int advice)
1493 {
1494         switch (advice) {
1495         case MADV_HUGEPAGE:
1496                 /*
1497                  * Be somewhat over-protective like KSM for now!
1498                  */
1499                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1500                         return -EINVAL;
1501                 *vm_flags &= ~VM_NOHUGEPAGE;
1502                 *vm_flags |= VM_HUGEPAGE;
1503                 /*
1504                  * If the vma become good for khugepaged to scan,
1505                  * register it here without waiting a page fault that
1506                  * may not happen any time soon.
1507                  */
1508                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1509                         return -ENOMEM;
1510                 break;
1511         case MADV_NOHUGEPAGE:
1512                 /*
1513                  * Be somewhat over-protective like KSM for now!
1514                  */
1515                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1516                         return -EINVAL;
1517                 *vm_flags &= ~VM_HUGEPAGE;
1518                 *vm_flags |= VM_NOHUGEPAGE;
1519                 /*
1520                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1521                  * this vma even if we leave the mm registered in khugepaged if
1522                  * it got registered before VM_NOHUGEPAGE was set.
1523                  */
1524                 break;
1525         }
1526
1527         return 0;
1528 }
1529
1530 static int __init khugepaged_slab_init(void)
1531 {
1532         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1533                                           sizeof(struct mm_slot),
1534                                           __alignof__(struct mm_slot), 0, NULL);
1535         if (!mm_slot_cache)
1536                 return -ENOMEM;
1537
1538         return 0;
1539 }
1540
1541 static void __init khugepaged_slab_free(void)
1542 {
1543         kmem_cache_destroy(mm_slot_cache);
1544         mm_slot_cache = NULL;
1545 }
1546
1547 static inline struct mm_slot *alloc_mm_slot(void)
1548 {
1549         if (!mm_slot_cache)     /* initialization failed */
1550                 return NULL;
1551         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1552 }
1553
1554 static inline void free_mm_slot(struct mm_slot *mm_slot)
1555 {
1556         kmem_cache_free(mm_slot_cache, mm_slot);
1557 }
1558
1559 static int __init mm_slots_hash_init(void)
1560 {
1561         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1562                                 GFP_KERNEL);
1563         if (!mm_slots_hash)
1564                 return -ENOMEM;
1565         return 0;
1566 }
1567
1568 #if 0
1569 static void __init mm_slots_hash_free(void)
1570 {
1571         kfree(mm_slots_hash);
1572         mm_slots_hash = NULL;
1573 }
1574 #endif
1575
1576 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1577 {
1578         struct mm_slot *mm_slot;
1579         struct hlist_head *bucket;
1580         struct hlist_node *node;
1581
1582         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1583                                 % MM_SLOTS_HASH_HEADS];
1584         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1585                 if (mm == mm_slot->mm)
1586                         return mm_slot;
1587         }
1588         return NULL;
1589 }
1590
1591 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1592                                     struct mm_slot *mm_slot)
1593 {
1594         struct hlist_head *bucket;
1595
1596         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1597                                 % MM_SLOTS_HASH_HEADS];
1598         mm_slot->mm = mm;
1599         hlist_add_head(&mm_slot->hash, bucket);
1600 }
1601
1602 static inline int khugepaged_test_exit(struct mm_struct *mm)
1603 {
1604         return atomic_read(&mm->mm_users) == 0;
1605 }
1606
1607 int __khugepaged_enter(struct mm_struct *mm)
1608 {
1609         struct mm_slot *mm_slot;
1610         int wakeup;
1611
1612         mm_slot = alloc_mm_slot();
1613         if (!mm_slot)
1614                 return -ENOMEM;
1615
1616         /* __khugepaged_exit() must not run from under us */
1617         VM_BUG_ON(khugepaged_test_exit(mm));
1618         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1619                 free_mm_slot(mm_slot);
1620                 return 0;
1621         }
1622
1623         spin_lock(&khugepaged_mm_lock);
1624         insert_to_mm_slots_hash(mm, mm_slot);
1625         /*
1626          * Insert just behind the scanning cursor, to let the area settle
1627          * down a little.
1628          */
1629         wakeup = list_empty(&khugepaged_scan.mm_head);
1630         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1631         spin_unlock(&khugepaged_mm_lock);
1632
1633         atomic_inc(&mm->mm_count);
1634         if (wakeup)
1635                 wake_up_interruptible(&khugepaged_wait);
1636
1637         return 0;
1638 }
1639
1640 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1641 {
1642         unsigned long hstart, hend;
1643         if (!vma->anon_vma)
1644                 /*
1645                  * Not yet faulted in so we will register later in the
1646                  * page fault if needed.
1647                  */
1648                 return 0;
1649         if (vma->vm_ops)
1650                 /* khugepaged not yet working on file or special mappings */
1651                 return 0;
1652         /*
1653          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1654          * true too, verify it here.
1655          */
1656         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1657         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1658         hend = vma->vm_end & HPAGE_PMD_MASK;
1659         if (hstart < hend)
1660                 return khugepaged_enter(vma);
1661         return 0;
1662 }
1663
1664 void __khugepaged_exit(struct mm_struct *mm)
1665 {
1666         struct mm_slot *mm_slot;
1667         int free = 0;
1668
1669         spin_lock(&khugepaged_mm_lock);
1670         mm_slot = get_mm_slot(mm);
1671         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1672                 hlist_del(&mm_slot->hash);
1673                 list_del(&mm_slot->mm_node);
1674                 free = 1;
1675         }
1676         spin_unlock(&khugepaged_mm_lock);
1677
1678         if (free) {
1679                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1680                 free_mm_slot(mm_slot);
1681                 mmdrop(mm);
1682         } else if (mm_slot) {
1683                 /*
1684                  * This is required to serialize against
1685                  * khugepaged_test_exit() (which is guaranteed to run
1686                  * under mmap sem read mode). Stop here (after we
1687                  * return all pagetables will be destroyed) until
1688                  * khugepaged has finished working on the pagetables
1689                  * under the mmap_sem.
1690                  */
1691                 down_write(&mm->mmap_sem);
1692                 up_write(&mm->mmap_sem);
1693         }
1694 }
1695
1696 static void release_pte_page(struct page *page)
1697 {
1698         /* 0 stands for page_is_file_cache(page) == false */
1699         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1700         unlock_page(page);
1701         putback_lru_page(page);
1702 }
1703
1704 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1705 {
1706         while (--_pte >= pte) {
1707                 pte_t pteval = *_pte;
1708                 if (!pte_none(pteval))
1709                         release_pte_page(pte_page(pteval));
1710         }
1711 }
1712
1713 static void release_all_pte_pages(pte_t *pte)
1714 {
1715         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1716 }
1717
1718 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1719                                         unsigned long address,
1720                                         pte_t *pte)
1721 {
1722         struct page *page;
1723         pte_t *_pte;
1724         int referenced = 0, isolated = 0, none = 0;
1725         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1726              _pte++, address += PAGE_SIZE) {
1727                 pte_t pteval = *_pte;
1728                 if (pte_none(pteval)) {
1729                         if (++none <= khugepaged_max_ptes_none)
1730                                 continue;
1731                         else {
1732                                 release_pte_pages(pte, _pte);
1733                                 goto out;
1734                         }
1735                 }
1736                 if (!pte_present(pteval) || !pte_write(pteval)) {
1737                         release_pte_pages(pte, _pte);
1738                         goto out;
1739                 }
1740                 page = vm_normal_page(vma, address, pteval);
1741                 if (unlikely(!page)) {
1742                         release_pte_pages(pte, _pte);
1743                         goto out;
1744                 }
1745                 VM_BUG_ON(PageCompound(page));
1746                 BUG_ON(!PageAnon(page));
1747                 VM_BUG_ON(!PageSwapBacked(page));
1748
1749                 /* cannot use mapcount: can't collapse if there's a gup pin */
1750                 if (page_count(page) != 1) {
1751                         release_pte_pages(pte, _pte);
1752                         goto out;
1753                 }
1754                 /*
1755                  * We can do it before isolate_lru_page because the
1756                  * page can't be freed from under us. NOTE: PG_lock
1757                  * is needed to serialize against split_huge_page
1758                  * when invoked from the VM.
1759                  */
1760                 if (!trylock_page(page)) {
1761                         release_pte_pages(pte, _pte);
1762                         goto out;
1763                 }
1764                 /*
1765                  * Isolate the page to avoid collapsing an hugepage
1766                  * currently in use by the VM.
1767                  */
1768                 if (isolate_lru_page(page)) {
1769                         unlock_page(page);
1770                         release_pte_pages(pte, _pte);
1771                         goto out;
1772                 }
1773                 /* 0 stands for page_is_file_cache(page) == false */
1774                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1775                 VM_BUG_ON(!PageLocked(page));
1776                 VM_BUG_ON(PageLRU(page));
1777
1778                 /* If there is no mapped pte young don't collapse the page */
1779                 if (pte_young(pteval) || PageReferenced(page) ||
1780                     mmu_notifier_test_young(vma->vm_mm, address))
1781                         referenced = 1;
1782         }
1783         if (unlikely(!referenced))
1784                 release_all_pte_pages(pte);
1785         else
1786                 isolated = 1;
1787 out:
1788         return isolated;
1789 }
1790
1791 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1792                                       struct vm_area_struct *vma,
1793                                       unsigned long address,
1794                                       spinlock_t *ptl)
1795 {
1796         pte_t *_pte;
1797         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1798                 pte_t pteval = *_pte;
1799                 struct page *src_page;
1800
1801                 if (pte_none(pteval)) {
1802                         clear_user_highpage(page, address);
1803                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1804                 } else {
1805                         src_page = pte_page(pteval);
1806                         copy_user_highpage(page, src_page, address, vma);
1807                         VM_BUG_ON(page_mapcount(src_page) != 1);
1808                         VM_BUG_ON(page_count(src_page) != 2);
1809                         release_pte_page(src_page);
1810                         /*
1811                          * ptl mostly unnecessary, but preempt has to
1812                          * be disabled to update the per-cpu stats
1813                          * inside page_remove_rmap().
1814                          */
1815                         spin_lock(ptl);
1816                         /*
1817                          * paravirt calls inside pte_clear here are
1818                          * superfluous.
1819                          */
1820                         pte_clear(vma->vm_mm, address, _pte);
1821                         page_remove_rmap(src_page);
1822                         spin_unlock(ptl);
1823                         free_page_and_swap_cache(src_page);
1824                 }
1825
1826                 address += PAGE_SIZE;
1827                 page++;
1828         }
1829 }
1830
1831 static void collapse_huge_page(struct mm_struct *mm,
1832                                unsigned long address,
1833                                struct page **hpage,
1834                                struct vm_area_struct *vma,
1835                                int node)
1836 {
1837         pgd_t *pgd;
1838         pud_t *pud;
1839         pmd_t *pmd, _pmd;
1840         pte_t *pte;
1841         pgtable_t pgtable;
1842         struct page *new_page;
1843         spinlock_t *ptl;
1844         int isolated;
1845         unsigned long hstart, hend;
1846
1847         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1848 #ifndef CONFIG_NUMA
1849         up_read(&mm->mmap_sem);
1850         VM_BUG_ON(!*hpage);
1851         new_page = *hpage;
1852 #else
1853         VM_BUG_ON(*hpage);
1854         /*
1855          * Allocate the page while the vma is still valid and under
1856          * the mmap_sem read mode so there is no memory allocation
1857          * later when we take the mmap_sem in write mode. This is more
1858          * friendly behavior (OTOH it may actually hide bugs) to
1859          * filesystems in userland with daemons allocating memory in
1860          * the userland I/O paths.  Allocating memory with the
1861          * mmap_sem in read mode is good idea also to allow greater
1862          * scalability.
1863          */
1864         new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1865                                       node, __GFP_OTHER_NODE);
1866
1867         /*
1868          * After allocating the hugepage, release the mmap_sem read lock in
1869          * preparation for taking it in write mode.
1870          */
1871         up_read(&mm->mmap_sem);
1872         if (unlikely(!new_page)) {
1873                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1874                 *hpage = ERR_PTR(-ENOMEM);
1875                 return;
1876         }
1877 #endif
1878
1879         count_vm_event(THP_COLLAPSE_ALLOC);
1880         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1881 #ifdef CONFIG_NUMA
1882                 put_page(new_page);
1883 #endif
1884                 return;
1885         }
1886
1887         /*
1888          * Prevent all access to pagetables with the exception of
1889          * gup_fast later hanlded by the ptep_clear_flush and the VM
1890          * handled by the anon_vma lock + PG_lock.
1891          */
1892         down_write(&mm->mmap_sem);
1893         if (unlikely(khugepaged_test_exit(mm)))
1894                 goto out;
1895
1896         vma = find_vma(mm, address);
1897         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1898         hend = vma->vm_end & HPAGE_PMD_MASK;
1899         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1900                 goto out;
1901
1902         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1903             (vma->vm_flags & VM_NOHUGEPAGE))
1904                 goto out;
1905
1906         if (!vma->anon_vma || vma->vm_ops)
1907                 goto out;
1908         if (is_vma_temporary_stack(vma))
1909                 goto out;
1910         /*
1911          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1912          * true too, verify it here.
1913          */
1914         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1915
1916         pgd = pgd_offset(mm, address);
1917         if (!pgd_present(*pgd))
1918                 goto out;
1919
1920         pud = pud_offset(pgd, address);
1921         if (!pud_present(*pud))
1922                 goto out;
1923
1924         pmd = pmd_offset(pud, address);
1925         /* pmd can't go away or become huge under us */
1926         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1927                 goto out;
1928
1929         anon_vma_lock(vma->anon_vma);
1930
1931         pte = pte_offset_map(pmd, address);
1932         ptl = pte_lockptr(mm, pmd);
1933
1934         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1935         /*
1936          * After this gup_fast can't run anymore. This also removes
1937          * any huge TLB entry from the CPU so we won't allow
1938          * huge and small TLB entries for the same virtual address
1939          * to avoid the risk of CPU bugs in that area.
1940          */
1941         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1942         spin_unlock(&mm->page_table_lock);
1943
1944         spin_lock(ptl);
1945         isolated = __collapse_huge_page_isolate(vma, address, pte);
1946         spin_unlock(ptl);
1947
1948         if (unlikely(!isolated)) {
1949                 pte_unmap(pte);
1950                 spin_lock(&mm->page_table_lock);
1951                 BUG_ON(!pmd_none(*pmd));
1952                 set_pmd_at(mm, address, pmd, _pmd);
1953                 spin_unlock(&mm->page_table_lock);
1954                 anon_vma_unlock(vma->anon_vma);
1955                 goto out;
1956         }
1957
1958         /*
1959          * All pages are isolated and locked so anon_vma rmap
1960          * can't run anymore.
1961          */
1962         anon_vma_unlock(vma->anon_vma);
1963
1964         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1965         pte_unmap(pte);
1966         __SetPageUptodate(new_page);
1967         pgtable = pmd_pgtable(_pmd);
1968         VM_BUG_ON(page_count(pgtable) != 1);
1969         VM_BUG_ON(page_mapcount(pgtable) != 0);
1970
1971         _pmd = mk_pmd(new_page, vma->vm_page_prot);
1972         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1973         _pmd = pmd_mkhuge(_pmd);
1974
1975         /*
1976          * spin_lock() below is not the equivalent of smp_wmb(), so
1977          * this is needed to avoid the copy_huge_page writes to become
1978          * visible after the set_pmd_at() write.
1979          */
1980         smp_wmb();
1981
1982         spin_lock(&mm->page_table_lock);
1983         BUG_ON(!pmd_none(*pmd));
1984         page_add_new_anon_rmap(new_page, vma, address);
1985         set_pmd_at(mm, address, pmd, _pmd);
1986         update_mmu_cache(vma, address, _pmd);
1987         prepare_pmd_huge_pte(pgtable, mm);
1988         spin_unlock(&mm->page_table_lock);
1989
1990 #ifndef CONFIG_NUMA
1991         *hpage = NULL;
1992 #endif
1993         khugepaged_pages_collapsed++;
1994 out_up_write:
1995         up_write(&mm->mmap_sem);
1996         return;
1997
1998 out:
1999         mem_cgroup_uncharge_page(new_page);
2000 #ifdef CONFIG_NUMA
2001         put_page(new_page);
2002 #endif
2003         goto out_up_write;
2004 }
2005
2006 static int khugepaged_scan_pmd(struct mm_struct *mm,
2007                                struct vm_area_struct *vma,
2008                                unsigned long address,
2009                                struct page **hpage)
2010 {
2011         pgd_t *pgd;
2012         pud_t *pud;
2013         pmd_t *pmd;
2014         pte_t *pte, *_pte;
2015         int ret = 0, referenced = 0, none = 0;
2016         struct page *page;
2017         unsigned long _address;
2018         spinlock_t *ptl;
2019         int node = -1;
2020
2021         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2022
2023         pgd = pgd_offset(mm, address);
2024         if (!pgd_present(*pgd))
2025                 goto out;
2026
2027         pud = pud_offset(pgd, address);
2028         if (!pud_present(*pud))
2029                 goto out;
2030
2031         pmd = pmd_offset(pud, address);
2032         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2033                 goto out;
2034
2035         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2036         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2037              _pte++, _address += PAGE_SIZE) {
2038                 pte_t pteval = *_pte;
2039                 if (pte_none(pteval)) {
2040                         if (++none <= khugepaged_max_ptes_none)
2041                                 continue;
2042                         else
2043                                 goto out_unmap;
2044                 }
2045                 if (!pte_present(pteval) || !pte_write(pteval))
2046                         goto out_unmap;
2047                 page = vm_normal_page(vma, _address, pteval);
2048                 if (unlikely(!page))
2049                         goto out_unmap;
2050                 /*
2051                  * Chose the node of the first page. This could
2052                  * be more sophisticated and look at more pages,
2053                  * but isn't for now.
2054                  */
2055                 if (node == -1)
2056                         node = page_to_nid(page);
2057                 VM_BUG_ON(PageCompound(page));
2058                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2059                         goto out_unmap;
2060                 /* cannot use mapcount: can't collapse if there's a gup pin */
2061                 if (page_count(page) != 1)
2062                         goto out_unmap;
2063                 if (pte_young(pteval) || PageReferenced(page) ||
2064                     mmu_notifier_test_young(vma->vm_mm, address))
2065                         referenced = 1;
2066         }
2067         if (referenced)
2068                 ret = 1;
2069 out_unmap:
2070         pte_unmap_unlock(pte, ptl);
2071         if (ret)
2072                 /* collapse_huge_page will return with the mmap_sem released */
2073                 collapse_huge_page(mm, address, hpage, vma, node);
2074 out:
2075         return ret;
2076 }
2077
2078 static void collect_mm_slot(struct mm_slot *mm_slot)
2079 {
2080         struct mm_struct *mm = mm_slot->mm;
2081
2082         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2083
2084         if (khugepaged_test_exit(mm)) {
2085                 /* free mm_slot */
2086                 hlist_del(&mm_slot->hash);
2087                 list_del(&mm_slot->mm_node);
2088
2089                 /*
2090                  * Not strictly needed because the mm exited already.
2091                  *
2092                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2093                  */
2094
2095                 /* khugepaged_mm_lock actually not necessary for the below */
2096                 free_mm_slot(mm_slot);
2097                 mmdrop(mm);
2098         }
2099 }
2100
2101 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2102                                             struct page **hpage)
2103         __releases(&khugepaged_mm_lock)
2104         __acquires(&khugepaged_mm_lock)
2105 {
2106         struct mm_slot *mm_slot;
2107         struct mm_struct *mm;
2108         struct vm_area_struct *vma;
2109         int progress = 0;
2110
2111         VM_BUG_ON(!pages);
2112         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2113
2114         if (khugepaged_scan.mm_slot)
2115                 mm_slot = khugepaged_scan.mm_slot;
2116         else {
2117                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2118                                      struct mm_slot, mm_node);
2119                 khugepaged_scan.address = 0;
2120                 khugepaged_scan.mm_slot = mm_slot;
2121         }
2122         spin_unlock(&khugepaged_mm_lock);
2123
2124         mm = mm_slot->mm;
2125         down_read(&mm->mmap_sem);
2126         if (unlikely(khugepaged_test_exit(mm)))
2127                 vma = NULL;
2128         else
2129                 vma = find_vma(mm, khugepaged_scan.address);
2130
2131         progress++;
2132         for (; vma; vma = vma->vm_next) {
2133                 unsigned long hstart, hend;
2134
2135                 cond_resched();
2136                 if (unlikely(khugepaged_test_exit(mm))) {
2137                         progress++;
2138                         break;
2139                 }
2140
2141                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2142                      !khugepaged_always()) ||
2143                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2144                 skip:
2145                         progress++;
2146                         continue;
2147                 }
2148                 if (!vma->anon_vma || vma->vm_ops)
2149                         goto skip;
2150                 if (is_vma_temporary_stack(vma))
2151                         goto skip;
2152                 /*
2153                  * If is_pfn_mapping() is true is_learn_pfn_mapping()
2154                  * must be true too, verify it here.
2155                  */
2156                 VM_BUG_ON(is_linear_pfn_mapping(vma) ||
2157                           vma->vm_flags & VM_NO_THP);
2158
2159                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2160                 hend = vma->vm_end & HPAGE_PMD_MASK;
2161                 if (hstart >= hend)
2162                         goto skip;
2163                 if (khugepaged_scan.address > hend)
2164                         goto skip;
2165                 if (khugepaged_scan.address < hstart)
2166                         khugepaged_scan.address = hstart;
2167                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2168
2169                 while (khugepaged_scan.address < hend) {
2170                         int ret;
2171                         cond_resched();
2172                         if (unlikely(khugepaged_test_exit(mm)))
2173                                 goto breakouterloop;
2174
2175                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2176                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2177                                   hend);
2178                         ret = khugepaged_scan_pmd(mm, vma,
2179                                                   khugepaged_scan.address,
2180                                                   hpage);
2181                         /* move to next address */
2182                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2183                         progress += HPAGE_PMD_NR;
2184                         if (ret)
2185                                 /* we released mmap_sem so break loop */
2186                                 goto breakouterloop_mmap_sem;
2187                         if (progress >= pages)
2188                                 goto breakouterloop;
2189                 }
2190         }
2191 breakouterloop:
2192         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2193 breakouterloop_mmap_sem:
2194
2195         spin_lock(&khugepaged_mm_lock);
2196         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2197         /*
2198          * Release the current mm_slot if this mm is about to die, or
2199          * if we scanned all vmas of this mm.
2200          */
2201         if (khugepaged_test_exit(mm) || !vma) {
2202                 /*
2203                  * Make sure that if mm_users is reaching zero while
2204                  * khugepaged runs here, khugepaged_exit will find
2205                  * mm_slot not pointing to the exiting mm.
2206                  */
2207                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2208                         khugepaged_scan.mm_slot = list_entry(
2209                                 mm_slot->mm_node.next,
2210                                 struct mm_slot, mm_node);
2211                         khugepaged_scan.address = 0;
2212                 } else {
2213                         khugepaged_scan.mm_slot = NULL;
2214                         khugepaged_full_scans++;
2215                 }
2216
2217                 collect_mm_slot(mm_slot);
2218         }
2219
2220         return progress;
2221 }
2222
2223 static int khugepaged_has_work(void)
2224 {
2225         return !list_empty(&khugepaged_scan.mm_head) &&
2226                 khugepaged_enabled();
2227 }
2228
2229 static int khugepaged_wait_event(void)
2230 {
2231         return !list_empty(&khugepaged_scan.mm_head) ||
2232                 !khugepaged_enabled();
2233 }
2234
2235 static void khugepaged_do_scan(struct page **hpage)
2236 {
2237         unsigned int progress = 0, pass_through_head = 0;
2238         unsigned int pages = khugepaged_pages_to_scan;
2239
2240         barrier(); /* write khugepaged_pages_to_scan to local stack */
2241
2242         while (progress < pages) {
2243                 cond_resched();
2244
2245 #ifndef CONFIG_NUMA
2246                 if (!*hpage) {
2247                         *hpage = alloc_hugepage(khugepaged_defrag());
2248                         if (unlikely(!*hpage)) {
2249                                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2250                                 break;
2251                         }
2252                         count_vm_event(THP_COLLAPSE_ALLOC);
2253                 }
2254 #else
2255                 if (IS_ERR(*hpage))
2256                         break;
2257 #endif
2258
2259                 if (unlikely(kthread_should_stop() || freezing(current)))
2260                         break;
2261
2262                 spin_lock(&khugepaged_mm_lock);
2263                 if (!khugepaged_scan.mm_slot)
2264                         pass_through_head++;
2265                 if (khugepaged_has_work() &&
2266                     pass_through_head < 2)
2267                         progress += khugepaged_scan_mm_slot(pages - progress,
2268                                                             hpage);
2269                 else
2270                         progress = pages;
2271                 spin_unlock(&khugepaged_mm_lock);
2272         }
2273 }
2274
2275 static void khugepaged_alloc_sleep(void)
2276 {
2277         wait_event_freezable_timeout(khugepaged_wait, false,
2278                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2279 }
2280
2281 #ifndef CONFIG_NUMA
2282 static struct page *khugepaged_alloc_hugepage(void)
2283 {
2284         struct page *hpage;
2285
2286         do {
2287                 hpage = alloc_hugepage(khugepaged_defrag());
2288                 if (!hpage) {
2289                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2290                         khugepaged_alloc_sleep();
2291                 } else
2292                         count_vm_event(THP_COLLAPSE_ALLOC);
2293         } while (unlikely(!hpage) &&
2294                  likely(khugepaged_enabled()));
2295         return hpage;
2296 }
2297 #endif
2298
2299 static void khugepaged_loop(void)
2300 {
2301         struct page *hpage;
2302
2303 #ifdef CONFIG_NUMA
2304         hpage = NULL;
2305 #endif
2306         while (likely(khugepaged_enabled())) {
2307 #ifndef CONFIG_NUMA
2308                 hpage = khugepaged_alloc_hugepage();
2309                 if (unlikely(!hpage))
2310                         break;
2311 #else
2312                 if (IS_ERR(hpage)) {
2313                         khugepaged_alloc_sleep();
2314                         hpage = NULL;
2315                 }
2316 #endif
2317
2318                 khugepaged_do_scan(&hpage);
2319 #ifndef CONFIG_NUMA
2320                 if (hpage)
2321                         put_page(hpage);
2322 #endif
2323                 try_to_freeze();
2324                 if (unlikely(kthread_should_stop()))
2325                         break;
2326                 if (khugepaged_has_work()) {
2327                         if (!khugepaged_scan_sleep_millisecs)
2328                                 continue;
2329                         wait_event_freezable_timeout(khugepaged_wait, false,
2330                             msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2331                 } else if (khugepaged_enabled())
2332                         wait_event_freezable(khugepaged_wait,
2333                                              khugepaged_wait_event());
2334         }
2335 }
2336
2337 static int khugepaged(void *none)
2338 {
2339         struct mm_slot *mm_slot;
2340
2341         set_freezable();
2342         set_user_nice(current, 19);
2343
2344         /* serialize with start_khugepaged() */
2345         mutex_lock(&khugepaged_mutex);
2346
2347         for (;;) {
2348                 mutex_unlock(&khugepaged_mutex);
2349                 VM_BUG_ON(khugepaged_thread != current);
2350                 khugepaged_loop();
2351                 VM_BUG_ON(khugepaged_thread != current);
2352
2353                 mutex_lock(&khugepaged_mutex);
2354                 if (!khugepaged_enabled())
2355                         break;
2356                 if (unlikely(kthread_should_stop()))
2357                         break;
2358         }
2359
2360         spin_lock(&khugepaged_mm_lock);
2361         mm_slot = khugepaged_scan.mm_slot;
2362         khugepaged_scan.mm_slot = NULL;
2363         if (mm_slot)
2364                 collect_mm_slot(mm_slot);
2365         spin_unlock(&khugepaged_mm_lock);
2366
2367         khugepaged_thread = NULL;
2368         mutex_unlock(&khugepaged_mutex);
2369
2370         return 0;
2371 }
2372
2373 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2374 {
2375         struct page *page;
2376
2377         spin_lock(&mm->page_table_lock);
2378         if (unlikely(!pmd_trans_huge(*pmd))) {
2379                 spin_unlock(&mm->page_table_lock);
2380                 return;
2381         }
2382         page = pmd_page(*pmd);
2383         VM_BUG_ON(!page_count(page));
2384         get_page(page);
2385         spin_unlock(&mm->page_table_lock);
2386
2387         split_huge_page(page);
2388
2389         put_page(page);
2390         BUG_ON(pmd_trans_huge(*pmd));
2391 }
2392
2393 static void split_huge_page_address(struct mm_struct *mm,
2394                                     unsigned long address)
2395 {
2396         pgd_t *pgd;
2397         pud_t *pud;
2398         pmd_t *pmd;
2399
2400         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2401
2402         pgd = pgd_offset(mm, address);
2403         if (!pgd_present(*pgd))
2404                 return;
2405
2406         pud = pud_offset(pgd, address);
2407         if (!pud_present(*pud))
2408                 return;
2409
2410         pmd = pmd_offset(pud, address);
2411         if (!pmd_present(*pmd))
2412                 return;
2413         /*
2414          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2415          * materialize from under us.
2416          */
2417         split_huge_page_pmd(mm, pmd);
2418 }
2419
2420 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2421                              unsigned long start,
2422                              unsigned long end,
2423                              long adjust_next)
2424 {
2425         /*
2426          * If the new start address isn't hpage aligned and it could
2427          * previously contain an hugepage: check if we need to split
2428          * an huge pmd.
2429          */
2430         if (start & ~HPAGE_PMD_MASK &&
2431             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2432             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2433                 split_huge_page_address(vma->vm_mm, start);
2434
2435         /*
2436          * If the new end address isn't hpage aligned and it could
2437          * previously contain an hugepage: check if we need to split
2438          * an huge pmd.
2439          */
2440         if (end & ~HPAGE_PMD_MASK &&
2441             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2442             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2443                 split_huge_page_address(vma->vm_mm, end);
2444
2445         /*
2446          * If we're also updating the vma->vm_next->vm_start, if the new
2447          * vm_next->vm_start isn't page aligned and it could previously
2448          * contain an hugepage: check if we need to split an huge pmd.
2449          */
2450         if (adjust_next > 0) {
2451                 struct vm_area_struct *next = vma->vm_next;
2452                 unsigned long nstart = next->vm_start;
2453                 nstart += adjust_next << PAGE_SHIFT;
2454                 if (nstart & ~HPAGE_PMD_MASK &&
2455                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2456                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2457                         split_huge_page_address(next->vm_mm, nstart);
2458         }
2459 }