mm: __count_immobile_pages(): make sure the node is online
[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                 spin_unlock(&mm->page_table_lock);
675         }
676
677         return ret;
678 }
679
680 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
681 {
682         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
683 }
684
685 static inline struct page *alloc_hugepage_vma(int defrag,
686                                               struct vm_area_struct *vma,
687                                               unsigned long haddr, int nd,
688                                               gfp_t extra_gfp)
689 {
690         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
691                                HPAGE_PMD_ORDER, vma, haddr, nd);
692 }
693
694 #ifndef CONFIG_NUMA
695 static inline struct page *alloc_hugepage(int defrag)
696 {
697         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
698                            HPAGE_PMD_ORDER);
699 }
700 #endif
701
702 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
703                                unsigned long address, pmd_t *pmd,
704                                unsigned int flags)
705 {
706         struct page *page;
707         unsigned long haddr = address & HPAGE_PMD_MASK;
708         pte_t *pte;
709
710         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
711                 if (unlikely(anon_vma_prepare(vma)))
712                         return VM_FAULT_OOM;
713                 if (unlikely(khugepaged_enter(vma)))
714                         return VM_FAULT_OOM;
715                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
716                                           vma, haddr, numa_node_id(), 0);
717                 if (unlikely(!page)) {
718                         count_vm_event(THP_FAULT_FALLBACK);
719                         goto out;
720                 }
721                 count_vm_event(THP_FAULT_ALLOC);
722                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
723                         put_page(page);
724                         goto out;
725                 }
726
727                 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
728         }
729 out:
730         /*
731          * Use __pte_alloc instead of pte_alloc_map, because we can't
732          * run pte_offset_map on the pmd, if an huge pmd could
733          * materialize from under us from a different thread.
734          */
735         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
736                 return VM_FAULT_OOM;
737         /* if an huge pmd materialized from under us just retry later */
738         if (unlikely(pmd_trans_huge(*pmd)))
739                 return 0;
740         /*
741          * A regular pmd is established and it can't morph into a huge pmd
742          * from under us anymore at this point because we hold the mmap_sem
743          * read mode and khugepaged takes it in write mode. So now it's
744          * safe to run pte_offset_map().
745          */
746         pte = pte_offset_map(pmd, address);
747         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
748 }
749
750 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
751                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
752                   struct vm_area_struct *vma)
753 {
754         struct page *src_page;
755         pmd_t pmd;
756         pgtable_t pgtable;
757         int ret;
758
759         ret = -ENOMEM;
760         pgtable = pte_alloc_one(dst_mm, addr);
761         if (unlikely(!pgtable))
762                 goto out;
763
764         spin_lock(&dst_mm->page_table_lock);
765         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
766
767         ret = -EAGAIN;
768         pmd = *src_pmd;
769         if (unlikely(!pmd_trans_huge(pmd))) {
770                 pte_free(dst_mm, pgtable);
771                 goto out_unlock;
772         }
773         if (unlikely(pmd_trans_splitting(pmd))) {
774                 /* split huge page running from under us */
775                 spin_unlock(&src_mm->page_table_lock);
776                 spin_unlock(&dst_mm->page_table_lock);
777                 pte_free(dst_mm, pgtable);
778
779                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
780                 goto out;
781         }
782         src_page = pmd_page(pmd);
783         VM_BUG_ON(!PageHead(src_page));
784         get_page(src_page);
785         page_dup_rmap(src_page);
786         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
787
788         pmdp_set_wrprotect(src_mm, addr, src_pmd);
789         pmd = pmd_mkold(pmd_wrprotect(pmd));
790         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
791         prepare_pmd_huge_pte(pgtable, dst_mm);
792
793         ret = 0;
794 out_unlock:
795         spin_unlock(&src_mm->page_table_lock);
796         spin_unlock(&dst_mm->page_table_lock);
797 out:
798         return ret;
799 }
800
801 /* no "address" argument so destroys page coloring of some arch */
802 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
803 {
804         pgtable_t pgtable;
805
806         assert_spin_locked(&mm->page_table_lock);
807
808         /* FIFO */
809         pgtable = mm->pmd_huge_pte;
810         if (list_empty(&pgtable->lru))
811                 mm->pmd_huge_pte = NULL;
812         else {
813                 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
814                                               struct page, lru);
815                 list_del(&pgtable->lru);
816         }
817         return pgtable;
818 }
819
820 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
821                                         struct vm_area_struct *vma,
822                                         unsigned long address,
823                                         pmd_t *pmd, pmd_t orig_pmd,
824                                         struct page *page,
825                                         unsigned long haddr)
826 {
827         pgtable_t pgtable;
828         pmd_t _pmd;
829         int ret = 0, i;
830         struct page **pages;
831
832         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
833                         GFP_KERNEL);
834         if (unlikely(!pages)) {
835                 ret |= VM_FAULT_OOM;
836                 goto out;
837         }
838
839         for (i = 0; i < HPAGE_PMD_NR; i++) {
840                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
841                                                __GFP_OTHER_NODE,
842                                                vma, address, page_to_nid(page));
843                 if (unlikely(!pages[i] ||
844                              mem_cgroup_newpage_charge(pages[i], mm,
845                                                        GFP_KERNEL))) {
846                         if (pages[i])
847                                 put_page(pages[i]);
848                         mem_cgroup_uncharge_start();
849                         while (--i >= 0) {
850                                 mem_cgroup_uncharge_page(pages[i]);
851                                 put_page(pages[i]);
852                         }
853                         mem_cgroup_uncharge_end();
854                         kfree(pages);
855                         ret |= VM_FAULT_OOM;
856                         goto out;
857                 }
858         }
859
860         for (i = 0; i < HPAGE_PMD_NR; i++) {
861                 copy_user_highpage(pages[i], page + i,
862                                    haddr + PAGE_SIZE * i, vma);
863                 __SetPageUptodate(pages[i]);
864                 cond_resched();
865         }
866
867         spin_lock(&mm->page_table_lock);
868         if (unlikely(!pmd_same(*pmd, orig_pmd)))
869                 goto out_free_pages;
870         VM_BUG_ON(!PageHead(page));
871
872         pmdp_clear_flush_notify(vma, haddr, pmd);
873         /* leave pmd empty until pte is filled */
874
875         pgtable = get_pmd_huge_pte(mm);
876         pmd_populate(mm, &_pmd, pgtable);
877
878         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
879                 pte_t *pte, entry;
880                 entry = mk_pte(pages[i], vma->vm_page_prot);
881                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
882                 page_add_new_anon_rmap(pages[i], vma, haddr);
883                 pte = pte_offset_map(&_pmd, haddr);
884                 VM_BUG_ON(!pte_none(*pte));
885                 set_pte_at(mm, haddr, pte, entry);
886                 pte_unmap(pte);
887         }
888         kfree(pages);
889
890         mm->nr_ptes++;
891         smp_wmb(); /* make pte visible before pmd */
892         pmd_populate(mm, pmd, pgtable);
893         page_remove_rmap(page);
894         spin_unlock(&mm->page_table_lock);
895
896         ret |= VM_FAULT_WRITE;
897         put_page(page);
898
899 out:
900         return ret;
901
902 out_free_pages:
903         spin_unlock(&mm->page_table_lock);
904         mem_cgroup_uncharge_start();
905         for (i = 0; i < HPAGE_PMD_NR; i++) {
906                 mem_cgroup_uncharge_page(pages[i]);
907                 put_page(pages[i]);
908         }
909         mem_cgroup_uncharge_end();
910         kfree(pages);
911         goto out;
912 }
913
914 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
915                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
916 {
917         int ret = 0;
918         struct page *page, *new_page;
919         unsigned long haddr;
920
921         VM_BUG_ON(!vma->anon_vma);
922         spin_lock(&mm->page_table_lock);
923         if (unlikely(!pmd_same(*pmd, orig_pmd)))
924                 goto out_unlock;
925
926         page = pmd_page(orig_pmd);
927         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
928         haddr = address & HPAGE_PMD_MASK;
929         if (page_mapcount(page) == 1) {
930                 pmd_t entry;
931                 entry = pmd_mkyoung(orig_pmd);
932                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
933                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
934                         update_mmu_cache(vma, address, entry);
935                 ret |= VM_FAULT_WRITE;
936                 goto out_unlock;
937         }
938         get_page(page);
939         spin_unlock(&mm->page_table_lock);
940
941         if (transparent_hugepage_enabled(vma) &&
942             !transparent_hugepage_debug_cow())
943                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
944                                               vma, haddr, numa_node_id(), 0);
945         else
946                 new_page = NULL;
947
948         if (unlikely(!new_page)) {
949                 count_vm_event(THP_FAULT_FALLBACK);
950                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
951                                                    pmd, orig_pmd, page, haddr);
952                 put_page(page);
953                 goto out;
954         }
955         count_vm_event(THP_FAULT_ALLOC);
956
957         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
958                 put_page(new_page);
959                 put_page(page);
960                 ret |= VM_FAULT_OOM;
961                 goto out;
962         }
963
964         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
965         __SetPageUptodate(new_page);
966
967         spin_lock(&mm->page_table_lock);
968         put_page(page);
969         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
970                 mem_cgroup_uncharge_page(new_page);
971                 put_page(new_page);
972         } else {
973                 pmd_t entry;
974                 VM_BUG_ON(!PageHead(page));
975                 entry = mk_pmd(new_page, vma->vm_page_prot);
976                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
977                 entry = pmd_mkhuge(entry);
978                 pmdp_clear_flush_notify(vma, haddr, pmd);
979                 page_add_new_anon_rmap(new_page, vma, haddr);
980                 set_pmd_at(mm, haddr, pmd, entry);
981                 update_mmu_cache(vma, address, entry);
982                 page_remove_rmap(page);
983                 put_page(page);
984                 ret |= VM_FAULT_WRITE;
985         }
986 out_unlock:
987         spin_unlock(&mm->page_table_lock);
988 out:
989         return ret;
990 }
991
992 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
993                                    unsigned long addr,
994                                    pmd_t *pmd,
995                                    unsigned int flags)
996 {
997         struct page *page = NULL;
998
999         assert_spin_locked(&mm->page_table_lock);
1000
1001         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1002                 goto out;
1003
1004         page = pmd_page(*pmd);
1005         VM_BUG_ON(!PageHead(page));
1006         if (flags & FOLL_TOUCH) {
1007                 pmd_t _pmd;
1008                 /*
1009                  * We should set the dirty bit only for FOLL_WRITE but
1010                  * for now the dirty bit in the pmd is meaningless.
1011                  * And if the dirty bit will become meaningful and
1012                  * we'll only set it with FOLL_WRITE, an atomic
1013                  * set_bit will be required on the pmd to set the
1014                  * young bit, instead of the current set_pmd_at.
1015                  */
1016                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1017                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1018         }
1019         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1020         VM_BUG_ON(!PageCompound(page));
1021         if (flags & FOLL_GET)
1022                 get_page_foll(page);
1023
1024 out:
1025         return page;
1026 }
1027
1028 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1029                  pmd_t *pmd, unsigned long addr)
1030 {
1031         int ret = 0;
1032
1033         spin_lock(&tlb->mm->page_table_lock);
1034         if (likely(pmd_trans_huge(*pmd))) {
1035                 if (unlikely(pmd_trans_splitting(*pmd))) {
1036                         spin_unlock(&tlb->mm->page_table_lock);
1037                         wait_split_huge_page(vma->anon_vma,
1038                                              pmd);
1039                 } else {
1040                         struct page *page;
1041                         pgtable_t pgtable;
1042                         pgtable = get_pmd_huge_pte(tlb->mm);
1043                         page = pmd_page(*pmd);
1044                         pmd_clear(pmd);
1045                         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1046                         page_remove_rmap(page);
1047                         VM_BUG_ON(page_mapcount(page) < 0);
1048                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1049                         VM_BUG_ON(!PageHead(page));
1050                         spin_unlock(&tlb->mm->page_table_lock);
1051                         tlb_remove_page(tlb, page);
1052                         pte_free(tlb->mm, pgtable);
1053                         ret = 1;
1054                 }
1055         } else
1056                 spin_unlock(&tlb->mm->page_table_lock);
1057
1058         return ret;
1059 }
1060
1061 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1062                 unsigned long addr, unsigned long end,
1063                 unsigned char *vec)
1064 {
1065         int ret = 0;
1066
1067         spin_lock(&vma->vm_mm->page_table_lock);
1068         if (likely(pmd_trans_huge(*pmd))) {
1069                 ret = !pmd_trans_splitting(*pmd);
1070                 spin_unlock(&vma->vm_mm->page_table_lock);
1071                 if (unlikely(!ret))
1072                         wait_split_huge_page(vma->anon_vma, pmd);
1073                 else {
1074                         /*
1075                          * All logical pages in the range are present
1076                          * if backed by a huge page.
1077                          */
1078                         memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1079                 }
1080         } else
1081                 spin_unlock(&vma->vm_mm->page_table_lock);
1082
1083         return ret;
1084 }
1085
1086 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1087                   unsigned long old_addr,
1088                   unsigned long new_addr, unsigned long old_end,
1089                   pmd_t *old_pmd, pmd_t *new_pmd)
1090 {
1091         int ret = 0;
1092         pmd_t pmd;
1093
1094         struct mm_struct *mm = vma->vm_mm;
1095
1096         if ((old_addr & ~HPAGE_PMD_MASK) ||
1097             (new_addr & ~HPAGE_PMD_MASK) ||
1098             old_end - old_addr < HPAGE_PMD_SIZE ||
1099             (new_vma->vm_flags & VM_NOHUGEPAGE))
1100                 goto out;
1101
1102         /*
1103          * The destination pmd shouldn't be established, free_pgtables()
1104          * should have release it.
1105          */
1106         if (WARN_ON(!pmd_none(*new_pmd))) {
1107                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1108                 goto out;
1109         }
1110
1111         spin_lock(&mm->page_table_lock);
1112         if (likely(pmd_trans_huge(*old_pmd))) {
1113                 if (pmd_trans_splitting(*old_pmd)) {
1114                         spin_unlock(&mm->page_table_lock);
1115                         wait_split_huge_page(vma->anon_vma, old_pmd);
1116                         ret = -1;
1117                 } else {
1118                         pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1119                         VM_BUG_ON(!pmd_none(*new_pmd));
1120                         set_pmd_at(mm, new_addr, new_pmd, pmd);
1121                         spin_unlock(&mm->page_table_lock);
1122                         ret = 1;
1123                 }
1124         } else {
1125                 spin_unlock(&mm->page_table_lock);
1126         }
1127 out:
1128         return ret;
1129 }
1130
1131 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1132                 unsigned long addr, pgprot_t newprot)
1133 {
1134         struct mm_struct *mm = vma->vm_mm;
1135         int ret = 0;
1136
1137         spin_lock(&mm->page_table_lock);
1138         if (likely(pmd_trans_huge(*pmd))) {
1139                 if (unlikely(pmd_trans_splitting(*pmd))) {
1140                         spin_unlock(&mm->page_table_lock);
1141                         wait_split_huge_page(vma->anon_vma, pmd);
1142                 } else {
1143                         pmd_t entry;
1144
1145                         entry = pmdp_get_and_clear(mm, addr, pmd);
1146                         entry = pmd_modify(entry, newprot);
1147                         set_pmd_at(mm, addr, pmd, entry);
1148                         spin_unlock(&vma->vm_mm->page_table_lock);
1149                         ret = 1;
1150                 }
1151         } else
1152                 spin_unlock(&vma->vm_mm->page_table_lock);
1153
1154         return ret;
1155 }
1156
1157 pmd_t *page_check_address_pmd(struct page *page,
1158                               struct mm_struct *mm,
1159                               unsigned long address,
1160                               enum page_check_address_pmd_flag flag)
1161 {
1162         pgd_t *pgd;
1163         pud_t *pud;
1164         pmd_t *pmd, *ret = NULL;
1165
1166         if (address & ~HPAGE_PMD_MASK)
1167                 goto out;
1168
1169         pgd = pgd_offset(mm, address);
1170         if (!pgd_present(*pgd))
1171                 goto out;
1172
1173         pud = pud_offset(pgd, address);
1174         if (!pud_present(*pud))
1175                 goto out;
1176
1177         pmd = pmd_offset(pud, address);
1178         if (pmd_none(*pmd))
1179                 goto out;
1180         if (pmd_page(*pmd) != page)
1181                 goto out;
1182         /*
1183          * split_vma() may create temporary aliased mappings. There is
1184          * no risk as long as all huge pmd are found and have their
1185          * splitting bit set before __split_huge_page_refcount
1186          * runs. Finding the same huge pmd more than once during the
1187          * same rmap walk is not a problem.
1188          */
1189         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1190             pmd_trans_splitting(*pmd))
1191                 goto out;
1192         if (pmd_trans_huge(*pmd)) {
1193                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1194                           !pmd_trans_splitting(*pmd));
1195                 ret = pmd;
1196         }
1197 out:
1198         return ret;
1199 }
1200
1201 static int __split_huge_page_splitting(struct page *page,
1202                                        struct vm_area_struct *vma,
1203                                        unsigned long address)
1204 {
1205         struct mm_struct *mm = vma->vm_mm;
1206         pmd_t *pmd;
1207         int ret = 0;
1208
1209         spin_lock(&mm->page_table_lock);
1210         pmd = page_check_address_pmd(page, mm, address,
1211                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1212         if (pmd) {
1213                 /*
1214                  * We can't temporarily set the pmd to null in order
1215                  * to split it, the pmd must remain marked huge at all
1216                  * times or the VM won't take the pmd_trans_huge paths
1217                  * and it won't wait on the anon_vma->root->mutex to
1218                  * serialize against split_huge_page*.
1219                  */
1220                 pmdp_splitting_flush_notify(vma, address, pmd);
1221                 ret = 1;
1222         }
1223         spin_unlock(&mm->page_table_lock);
1224
1225         return ret;
1226 }
1227
1228 static void __split_huge_page_refcount(struct page *page)
1229 {
1230         int i;
1231         struct zone *zone = page_zone(page);
1232         int tail_count = 0;
1233
1234         /* prevent PageLRU to go away from under us, and freeze lru stats */
1235         spin_lock_irq(&zone->lru_lock);
1236         compound_lock(page);
1237         /* complete memcg works before add pages to LRU */
1238         mem_cgroup_split_huge_fixup(page);
1239
1240         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1241                 struct page *page_tail = page + i;
1242
1243                 /* tail_page->_mapcount cannot change */
1244                 BUG_ON(page_mapcount(page_tail) < 0);
1245                 tail_count += page_mapcount(page_tail);
1246                 /* check for overflow */
1247                 BUG_ON(tail_count < 0);
1248                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1249                 /*
1250                  * tail_page->_count is zero and not changing from
1251                  * under us. But get_page_unless_zero() may be running
1252                  * from under us on the tail_page. If we used
1253                  * atomic_set() below instead of atomic_add(), we
1254                  * would then run atomic_set() concurrently with
1255                  * get_page_unless_zero(), and atomic_set() is
1256                  * implemented in C not using locked ops. spin_unlock
1257                  * on x86 sometime uses locked ops because of PPro
1258                  * errata 66, 92, so unless somebody can guarantee
1259                  * atomic_set() here would be safe on all archs (and
1260                  * not only on x86), it's safer to use atomic_add().
1261                  */
1262                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1263                            &page_tail->_count);
1264
1265                 /* after clearing PageTail the gup refcount can be released */
1266                 smp_mb();
1267
1268                 /*
1269                  * retain hwpoison flag of the poisoned tail page:
1270                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1271                  *   by the memory-failure.
1272                  */
1273                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1274                 page_tail->flags |= (page->flags &
1275                                      ((1L << PG_referenced) |
1276                                       (1L << PG_swapbacked) |
1277                                       (1L << PG_mlocked) |
1278                                       (1L << PG_uptodate)));
1279                 page_tail->flags |= (1L << PG_dirty);
1280
1281                 /* clear PageTail before overwriting first_page */
1282                 smp_wmb();
1283
1284                 /*
1285                  * __split_huge_page_splitting() already set the
1286                  * splitting bit in all pmd that could map this
1287                  * hugepage, that will ensure no CPU can alter the
1288                  * mapcount on the head page. The mapcount is only
1289                  * accounted in the head page and it has to be
1290                  * transferred to all tail pages in the below code. So
1291                  * for this code to be safe, the split the mapcount
1292                  * can't change. But that doesn't mean userland can't
1293                  * keep changing and reading the page contents while
1294                  * we transfer the mapcount, so the pmd splitting
1295                  * status is achieved setting a reserved bit in the
1296                  * pmd, not by clearing the present bit.
1297                 */
1298                 page_tail->_mapcount = page->_mapcount;
1299
1300                 BUG_ON(page_tail->mapping);
1301                 page_tail->mapping = page->mapping;
1302
1303                 page_tail->index = page->index + i;
1304
1305                 BUG_ON(!PageAnon(page_tail));
1306                 BUG_ON(!PageUptodate(page_tail));
1307                 BUG_ON(!PageDirty(page_tail));
1308                 BUG_ON(!PageSwapBacked(page_tail));
1309
1310
1311                 lru_add_page_tail(zone, page, page_tail);
1312         }
1313         atomic_sub(tail_count, &page->_count);
1314         BUG_ON(atomic_read(&page->_count) <= 0);
1315
1316         __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1317         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1318
1319         ClearPageCompound(page);
1320         compound_unlock(page);
1321         spin_unlock_irq(&zone->lru_lock);
1322
1323         for (i = 1; i < HPAGE_PMD_NR; i++) {
1324                 struct page *page_tail = page + i;
1325                 BUG_ON(page_count(page_tail) <= 0);
1326                 /*
1327                  * Tail pages may be freed if there wasn't any mapping
1328                  * like if add_to_swap() is running on a lru page that
1329                  * had its mapping zapped. And freeing these pages
1330                  * requires taking the lru_lock so we do the put_page
1331                  * of the tail pages after the split is complete.
1332                  */
1333                 put_page(page_tail);
1334         }
1335
1336         /*
1337          * Only the head page (now become a regular page) is required
1338          * to be pinned by the caller.
1339          */
1340         BUG_ON(page_count(page) <= 0);
1341 }
1342
1343 static int __split_huge_page_map(struct page *page,
1344                                  struct vm_area_struct *vma,
1345                                  unsigned long address)
1346 {
1347         struct mm_struct *mm = vma->vm_mm;
1348         pmd_t *pmd, _pmd;
1349         int ret = 0, i;
1350         pgtable_t pgtable;
1351         unsigned long haddr;
1352
1353         spin_lock(&mm->page_table_lock);
1354         pmd = page_check_address_pmd(page, mm, address,
1355                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1356         if (pmd) {
1357                 pgtable = get_pmd_huge_pte(mm);
1358                 pmd_populate(mm, &_pmd, pgtable);
1359
1360                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1361                      i++, haddr += PAGE_SIZE) {
1362                         pte_t *pte, entry;
1363                         BUG_ON(PageCompound(page+i));
1364                         entry = mk_pte(page + i, vma->vm_page_prot);
1365                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1366                         if (!pmd_write(*pmd))
1367                                 entry = pte_wrprotect(entry);
1368                         else
1369                                 BUG_ON(page_mapcount(page) != 1);
1370                         if (!pmd_young(*pmd))
1371                                 entry = pte_mkold(entry);
1372                         pte = pte_offset_map(&_pmd, haddr);
1373                         BUG_ON(!pte_none(*pte));
1374                         set_pte_at(mm, haddr, pte, entry);
1375                         pte_unmap(pte);
1376                 }
1377
1378                 mm->nr_ptes++;
1379                 smp_wmb(); /* make pte visible before pmd */
1380                 /*
1381                  * Up to this point the pmd is present and huge and
1382                  * userland has the whole access to the hugepage
1383                  * during the split (which happens in place). If we
1384                  * overwrite the pmd with the not-huge version
1385                  * pointing to the pte here (which of course we could
1386                  * if all CPUs were bug free), userland could trigger
1387                  * a small page size TLB miss on the small sized TLB
1388                  * while the hugepage TLB entry is still established
1389                  * in the huge TLB. Some CPU doesn't like that. See
1390                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1391                  * Erratum 383 on page 93. Intel should be safe but is
1392                  * also warns that it's only safe if the permission
1393                  * and cache attributes of the two entries loaded in
1394                  * the two TLB is identical (which should be the case
1395                  * here). But it is generally safer to never allow
1396                  * small and huge TLB entries for the same virtual
1397                  * address to be loaded simultaneously. So instead of
1398                  * doing "pmd_populate(); flush_tlb_range();" we first
1399                  * mark the current pmd notpresent (atomically because
1400                  * here the pmd_trans_huge and pmd_trans_splitting
1401                  * must remain set at all times on the pmd until the
1402                  * split is complete for this pmd), then we flush the
1403                  * SMP TLB and finally we write the non-huge version
1404                  * of the pmd entry with pmd_populate.
1405                  */
1406                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1407                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1408                 pmd_populate(mm, pmd, pgtable);
1409                 ret = 1;
1410         }
1411         spin_unlock(&mm->page_table_lock);
1412
1413         return ret;
1414 }
1415
1416 /* must be called with anon_vma->root->mutex hold */
1417 static void __split_huge_page(struct page *page,
1418                               struct anon_vma *anon_vma)
1419 {
1420         int mapcount, mapcount2;
1421         struct anon_vma_chain *avc;
1422
1423         BUG_ON(!PageHead(page));
1424         BUG_ON(PageTail(page));
1425
1426         mapcount = 0;
1427         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1428                 struct vm_area_struct *vma = avc->vma;
1429                 unsigned long addr = vma_address(page, vma);
1430                 BUG_ON(is_vma_temporary_stack(vma));
1431                 if (addr == -EFAULT)
1432                         continue;
1433                 mapcount += __split_huge_page_splitting(page, vma, addr);
1434         }
1435         /*
1436          * It is critical that new vmas are added to the tail of the
1437          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1438          * and establishes a child pmd before
1439          * __split_huge_page_splitting() freezes the parent pmd (so if
1440          * we fail to prevent copy_huge_pmd() from running until the
1441          * whole __split_huge_page() is complete), we will still see
1442          * the newly established pmd of the child later during the
1443          * walk, to be able to set it as pmd_trans_splitting too.
1444          */
1445         if (mapcount != page_mapcount(page))
1446                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1447                        mapcount, page_mapcount(page));
1448         BUG_ON(mapcount != page_mapcount(page));
1449
1450         __split_huge_page_refcount(page);
1451
1452         mapcount2 = 0;
1453         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1454                 struct vm_area_struct *vma = avc->vma;
1455                 unsigned long addr = vma_address(page, vma);
1456                 BUG_ON(is_vma_temporary_stack(vma));
1457                 if (addr == -EFAULT)
1458                         continue;
1459                 mapcount2 += __split_huge_page_map(page, vma, addr);
1460         }
1461         if (mapcount != mapcount2)
1462                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1463                        mapcount, mapcount2, page_mapcount(page));
1464         BUG_ON(mapcount != mapcount2);
1465 }
1466
1467 int split_huge_page(struct page *page)
1468 {
1469         struct anon_vma *anon_vma;
1470         int ret = 1;
1471
1472         BUG_ON(!PageAnon(page));
1473         anon_vma = page_lock_anon_vma(page);
1474         if (!anon_vma)
1475                 goto out;
1476         ret = 0;
1477         if (!PageCompound(page))
1478                 goto out_unlock;
1479
1480         BUG_ON(!PageSwapBacked(page));
1481         __split_huge_page(page, anon_vma);
1482         count_vm_event(THP_SPLIT);
1483
1484         BUG_ON(PageCompound(page));
1485 out_unlock:
1486         page_unlock_anon_vma(anon_vma);
1487 out:
1488         return ret;
1489 }
1490
1491 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1492                    VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1493
1494 int hugepage_madvise(struct vm_area_struct *vma,
1495                      unsigned long *vm_flags, int advice)
1496 {
1497         switch (advice) {
1498         case MADV_HUGEPAGE:
1499                 /*
1500                  * Be somewhat over-protective like KSM for now!
1501                  */
1502                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1503                         return -EINVAL;
1504                 *vm_flags &= ~VM_NOHUGEPAGE;
1505                 *vm_flags |= VM_HUGEPAGE;
1506                 /*
1507                  * If the vma become good for khugepaged to scan,
1508                  * register it here without waiting a page fault that
1509                  * may not happen any time soon.
1510                  */
1511                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1512                         return -ENOMEM;
1513                 break;
1514         case MADV_NOHUGEPAGE:
1515                 /*
1516                  * Be somewhat over-protective like KSM for now!
1517                  */
1518                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1519                         return -EINVAL;
1520                 *vm_flags &= ~VM_HUGEPAGE;
1521                 *vm_flags |= VM_NOHUGEPAGE;
1522                 /*
1523                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1524                  * this vma even if we leave the mm registered in khugepaged if
1525                  * it got registered before VM_NOHUGEPAGE was set.
1526                  */
1527                 break;
1528         }
1529
1530         return 0;
1531 }
1532
1533 static int __init khugepaged_slab_init(void)
1534 {
1535         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1536                                           sizeof(struct mm_slot),
1537                                           __alignof__(struct mm_slot), 0, NULL);
1538         if (!mm_slot_cache)
1539                 return -ENOMEM;
1540
1541         return 0;
1542 }
1543
1544 static void __init khugepaged_slab_free(void)
1545 {
1546         kmem_cache_destroy(mm_slot_cache);
1547         mm_slot_cache = NULL;
1548 }
1549
1550 static inline struct mm_slot *alloc_mm_slot(void)
1551 {
1552         if (!mm_slot_cache)     /* initialization failed */
1553                 return NULL;
1554         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1555 }
1556
1557 static inline void free_mm_slot(struct mm_slot *mm_slot)
1558 {
1559         kmem_cache_free(mm_slot_cache, mm_slot);
1560 }
1561
1562 static int __init mm_slots_hash_init(void)
1563 {
1564         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1565                                 GFP_KERNEL);
1566         if (!mm_slots_hash)
1567                 return -ENOMEM;
1568         return 0;
1569 }
1570
1571 #if 0
1572 static void __init mm_slots_hash_free(void)
1573 {
1574         kfree(mm_slots_hash);
1575         mm_slots_hash = NULL;
1576 }
1577 #endif
1578
1579 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1580 {
1581         struct mm_slot *mm_slot;
1582         struct hlist_head *bucket;
1583         struct hlist_node *node;
1584
1585         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1586                                 % MM_SLOTS_HASH_HEADS];
1587         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1588                 if (mm == mm_slot->mm)
1589                         return mm_slot;
1590         }
1591         return NULL;
1592 }
1593
1594 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1595                                     struct mm_slot *mm_slot)
1596 {
1597         struct hlist_head *bucket;
1598
1599         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1600                                 % MM_SLOTS_HASH_HEADS];
1601         mm_slot->mm = mm;
1602         hlist_add_head(&mm_slot->hash, bucket);
1603 }
1604
1605 static inline int khugepaged_test_exit(struct mm_struct *mm)
1606 {
1607         return atomic_read(&mm->mm_users) == 0;
1608 }
1609
1610 int __khugepaged_enter(struct mm_struct *mm)
1611 {
1612         struct mm_slot *mm_slot;
1613         int wakeup;
1614
1615         mm_slot = alloc_mm_slot();
1616         if (!mm_slot)
1617                 return -ENOMEM;
1618
1619         /* __khugepaged_exit() must not run from under us */
1620         VM_BUG_ON(khugepaged_test_exit(mm));
1621         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1622                 free_mm_slot(mm_slot);
1623                 return 0;
1624         }
1625
1626         spin_lock(&khugepaged_mm_lock);
1627         insert_to_mm_slots_hash(mm, mm_slot);
1628         /*
1629          * Insert just behind the scanning cursor, to let the area settle
1630          * down a little.
1631          */
1632         wakeup = list_empty(&khugepaged_scan.mm_head);
1633         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1634         spin_unlock(&khugepaged_mm_lock);
1635
1636         atomic_inc(&mm->mm_count);
1637         if (wakeup)
1638                 wake_up_interruptible(&khugepaged_wait);
1639
1640         return 0;
1641 }
1642
1643 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1644 {
1645         unsigned long hstart, hend;
1646         if (!vma->anon_vma)
1647                 /*
1648                  * Not yet faulted in so we will register later in the
1649                  * page fault if needed.
1650                  */
1651                 return 0;
1652         if (vma->vm_ops)
1653                 /* khugepaged not yet working on file or special mappings */
1654                 return 0;
1655         /*
1656          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1657          * true too, verify it here.
1658          */
1659         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1660         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1661         hend = vma->vm_end & HPAGE_PMD_MASK;
1662         if (hstart < hend)
1663                 return khugepaged_enter(vma);
1664         return 0;
1665 }
1666
1667 void __khugepaged_exit(struct mm_struct *mm)
1668 {
1669         struct mm_slot *mm_slot;
1670         int free = 0;
1671
1672         spin_lock(&khugepaged_mm_lock);
1673         mm_slot = get_mm_slot(mm);
1674         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1675                 hlist_del(&mm_slot->hash);
1676                 list_del(&mm_slot->mm_node);
1677                 free = 1;
1678         }
1679         spin_unlock(&khugepaged_mm_lock);
1680
1681         if (free) {
1682                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1683                 free_mm_slot(mm_slot);
1684                 mmdrop(mm);
1685         } else if (mm_slot) {
1686                 /*
1687                  * This is required to serialize against
1688                  * khugepaged_test_exit() (which is guaranteed to run
1689                  * under mmap sem read mode). Stop here (after we
1690                  * return all pagetables will be destroyed) until
1691                  * khugepaged has finished working on the pagetables
1692                  * under the mmap_sem.
1693                  */
1694                 down_write(&mm->mmap_sem);
1695                 up_write(&mm->mmap_sem);
1696         }
1697 }
1698
1699 static void release_pte_page(struct page *page)
1700 {
1701         /* 0 stands for page_is_file_cache(page) == false */
1702         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1703         unlock_page(page);
1704         putback_lru_page(page);
1705 }
1706
1707 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1708 {
1709         while (--_pte >= pte) {
1710                 pte_t pteval = *_pte;
1711                 if (!pte_none(pteval))
1712                         release_pte_page(pte_page(pteval));
1713         }
1714 }
1715
1716 static void release_all_pte_pages(pte_t *pte)
1717 {
1718         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1719 }
1720
1721 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1722                                         unsigned long address,
1723                                         pte_t *pte)
1724 {
1725         struct page *page;
1726         pte_t *_pte;
1727         int referenced = 0, isolated = 0, none = 0;
1728         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1729              _pte++, address += PAGE_SIZE) {
1730                 pte_t pteval = *_pte;
1731                 if (pte_none(pteval)) {
1732                         if (++none <= khugepaged_max_ptes_none)
1733                                 continue;
1734                         else {
1735                                 release_pte_pages(pte, _pte);
1736                                 goto out;
1737                         }
1738                 }
1739                 if (!pte_present(pteval) || !pte_write(pteval)) {
1740                         release_pte_pages(pte, _pte);
1741                         goto out;
1742                 }
1743                 page = vm_normal_page(vma, address, pteval);
1744                 if (unlikely(!page)) {
1745                         release_pte_pages(pte, _pte);
1746                         goto out;
1747                 }
1748                 VM_BUG_ON(PageCompound(page));
1749                 BUG_ON(!PageAnon(page));
1750                 VM_BUG_ON(!PageSwapBacked(page));
1751
1752                 /* cannot use mapcount: can't collapse if there's a gup pin */
1753                 if (page_count(page) != 1) {
1754                         release_pte_pages(pte, _pte);
1755                         goto out;
1756                 }
1757                 /*
1758                  * We can do it before isolate_lru_page because the
1759                  * page can't be freed from under us. NOTE: PG_lock
1760                  * is needed to serialize against split_huge_page
1761                  * when invoked from the VM.
1762                  */
1763                 if (!trylock_page(page)) {
1764                         release_pte_pages(pte, _pte);
1765                         goto out;
1766                 }
1767                 /*
1768                  * Isolate the page to avoid collapsing an hugepage
1769                  * currently in use by the VM.
1770                  */
1771                 if (isolate_lru_page(page)) {
1772                         unlock_page(page);
1773                         release_pte_pages(pte, _pte);
1774                         goto out;
1775                 }
1776                 /* 0 stands for page_is_file_cache(page) == false */
1777                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1778                 VM_BUG_ON(!PageLocked(page));
1779                 VM_BUG_ON(PageLRU(page));
1780
1781                 /* If there is no mapped pte young don't collapse the page */
1782                 if (pte_young(pteval) || PageReferenced(page) ||
1783                     mmu_notifier_test_young(vma->vm_mm, address))
1784                         referenced = 1;
1785         }
1786         if (unlikely(!referenced))
1787                 release_all_pte_pages(pte);
1788         else
1789                 isolated = 1;
1790 out:
1791         return isolated;
1792 }
1793
1794 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1795                                       struct vm_area_struct *vma,
1796                                       unsigned long address,
1797                                       spinlock_t *ptl)
1798 {
1799         pte_t *_pte;
1800         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1801                 pte_t pteval = *_pte;
1802                 struct page *src_page;
1803
1804                 if (pte_none(pteval)) {
1805                         clear_user_highpage(page, address);
1806                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1807                 } else {
1808                         src_page = pte_page(pteval);
1809                         copy_user_highpage(page, src_page, address, vma);
1810                         VM_BUG_ON(page_mapcount(src_page) != 1);
1811                         VM_BUG_ON(page_count(src_page) != 2);
1812                         release_pte_page(src_page);
1813                         /*
1814                          * ptl mostly unnecessary, but preempt has to
1815                          * be disabled to update the per-cpu stats
1816                          * inside page_remove_rmap().
1817                          */
1818                         spin_lock(ptl);
1819                         /*
1820                          * paravirt calls inside pte_clear here are
1821                          * superfluous.
1822                          */
1823                         pte_clear(vma->vm_mm, address, _pte);
1824                         page_remove_rmap(src_page);
1825                         spin_unlock(ptl);
1826                         free_page_and_swap_cache(src_page);
1827                 }
1828
1829                 address += PAGE_SIZE;
1830                 page++;
1831         }
1832 }
1833
1834 static void collapse_huge_page(struct mm_struct *mm,
1835                                unsigned long address,
1836                                struct page **hpage,
1837                                struct vm_area_struct *vma,
1838                                int node)
1839 {
1840         pgd_t *pgd;
1841         pud_t *pud;
1842         pmd_t *pmd, _pmd;
1843         pte_t *pte;
1844         pgtable_t pgtable;
1845         struct page *new_page;
1846         spinlock_t *ptl;
1847         int isolated;
1848         unsigned long hstart, hend;
1849
1850         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1851 #ifndef CONFIG_NUMA
1852         up_read(&mm->mmap_sem);
1853         VM_BUG_ON(!*hpage);
1854         new_page = *hpage;
1855 #else
1856         VM_BUG_ON(*hpage);
1857         /*
1858          * Allocate the page while the vma is still valid and under
1859          * the mmap_sem read mode so there is no memory allocation
1860          * later when we take the mmap_sem in write mode. This is more
1861          * friendly behavior (OTOH it may actually hide bugs) to
1862          * filesystems in userland with daemons allocating memory in
1863          * the userland I/O paths.  Allocating memory with the
1864          * mmap_sem in read mode is good idea also to allow greater
1865          * scalability.
1866          */
1867         new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1868                                       node, __GFP_OTHER_NODE);
1869
1870         /*
1871          * After allocating the hugepage, release the mmap_sem read lock in
1872          * preparation for taking it in write mode.
1873          */
1874         up_read(&mm->mmap_sem);
1875         if (unlikely(!new_page)) {
1876                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1877                 *hpage = ERR_PTR(-ENOMEM);
1878                 return;
1879         }
1880 #endif
1881
1882         count_vm_event(THP_COLLAPSE_ALLOC);
1883         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1884 #ifdef CONFIG_NUMA
1885                 put_page(new_page);
1886 #endif
1887                 return;
1888         }
1889
1890         /*
1891          * Prevent all access to pagetables with the exception of
1892          * gup_fast later hanlded by the ptep_clear_flush and the VM
1893          * handled by the anon_vma lock + PG_lock.
1894          */
1895         down_write(&mm->mmap_sem);
1896         if (unlikely(khugepaged_test_exit(mm)))
1897                 goto out;
1898
1899         vma = find_vma(mm, address);
1900         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1901         hend = vma->vm_end & HPAGE_PMD_MASK;
1902         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1903                 goto out;
1904
1905         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1906             (vma->vm_flags & VM_NOHUGEPAGE))
1907                 goto out;
1908
1909         if (!vma->anon_vma || vma->vm_ops)
1910                 goto out;
1911         if (is_vma_temporary_stack(vma))
1912                 goto out;
1913         /*
1914          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1915          * true too, verify it here.
1916          */
1917         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1918
1919         pgd = pgd_offset(mm, address);
1920         if (!pgd_present(*pgd))
1921                 goto out;
1922
1923         pud = pud_offset(pgd, address);
1924         if (!pud_present(*pud))
1925                 goto out;
1926
1927         pmd = pmd_offset(pud, address);
1928         /* pmd can't go away or become huge under us */
1929         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1930                 goto out;
1931
1932         anon_vma_lock(vma->anon_vma);
1933
1934         pte = pte_offset_map(pmd, address);
1935         ptl = pte_lockptr(mm, pmd);
1936
1937         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1938         /*
1939          * After this gup_fast can't run anymore. This also removes
1940          * any huge TLB entry from the CPU so we won't allow
1941          * huge and small TLB entries for the same virtual address
1942          * to avoid the risk of CPU bugs in that area.
1943          */
1944         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1945         spin_unlock(&mm->page_table_lock);
1946
1947         spin_lock(ptl);
1948         isolated = __collapse_huge_page_isolate(vma, address, pte);
1949         spin_unlock(ptl);
1950
1951         if (unlikely(!isolated)) {
1952                 pte_unmap(pte);
1953                 spin_lock(&mm->page_table_lock);
1954                 BUG_ON(!pmd_none(*pmd));
1955                 set_pmd_at(mm, address, pmd, _pmd);
1956                 spin_unlock(&mm->page_table_lock);
1957                 anon_vma_unlock(vma->anon_vma);
1958                 goto out;
1959         }
1960
1961         /*
1962          * All pages are isolated and locked so anon_vma rmap
1963          * can't run anymore.
1964          */
1965         anon_vma_unlock(vma->anon_vma);
1966
1967         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1968         pte_unmap(pte);
1969         __SetPageUptodate(new_page);
1970         pgtable = pmd_pgtable(_pmd);
1971         VM_BUG_ON(page_count(pgtable) != 1);
1972         VM_BUG_ON(page_mapcount(pgtable) != 0);
1973
1974         _pmd = mk_pmd(new_page, vma->vm_page_prot);
1975         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1976         _pmd = pmd_mkhuge(_pmd);
1977
1978         /*
1979          * spin_lock() below is not the equivalent of smp_wmb(), so
1980          * this is needed to avoid the copy_huge_page writes to become
1981          * visible after the set_pmd_at() write.
1982          */
1983         smp_wmb();
1984
1985         spin_lock(&mm->page_table_lock);
1986         BUG_ON(!pmd_none(*pmd));
1987         page_add_new_anon_rmap(new_page, vma, address);
1988         set_pmd_at(mm, address, pmd, _pmd);
1989         update_mmu_cache(vma, address, _pmd);
1990         prepare_pmd_huge_pte(pgtable, mm);
1991         mm->nr_ptes--;
1992         spin_unlock(&mm->page_table_lock);
1993
1994 #ifndef CONFIG_NUMA
1995         *hpage = NULL;
1996 #endif
1997         khugepaged_pages_collapsed++;
1998 out_up_write:
1999         up_write(&mm->mmap_sem);
2000         return;
2001
2002 out:
2003         mem_cgroup_uncharge_page(new_page);
2004 #ifdef CONFIG_NUMA
2005         put_page(new_page);
2006 #endif
2007         goto out_up_write;
2008 }
2009
2010 static int khugepaged_scan_pmd(struct mm_struct *mm,
2011                                struct vm_area_struct *vma,
2012                                unsigned long address,
2013                                struct page **hpage)
2014 {
2015         pgd_t *pgd;
2016         pud_t *pud;
2017         pmd_t *pmd;
2018         pte_t *pte, *_pte;
2019         int ret = 0, referenced = 0, none = 0;
2020         struct page *page;
2021         unsigned long _address;
2022         spinlock_t *ptl;
2023         int node = -1;
2024
2025         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2026
2027         pgd = pgd_offset(mm, address);
2028         if (!pgd_present(*pgd))
2029                 goto out;
2030
2031         pud = pud_offset(pgd, address);
2032         if (!pud_present(*pud))
2033                 goto out;
2034
2035         pmd = pmd_offset(pud, address);
2036         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2037                 goto out;
2038
2039         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2040         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2041              _pte++, _address += PAGE_SIZE) {
2042                 pte_t pteval = *_pte;
2043                 if (pte_none(pteval)) {
2044                         if (++none <= khugepaged_max_ptes_none)
2045                                 continue;
2046                         else
2047                                 goto out_unmap;
2048                 }
2049                 if (!pte_present(pteval) || !pte_write(pteval))
2050                         goto out_unmap;
2051                 page = vm_normal_page(vma, _address, pteval);
2052                 if (unlikely(!page))
2053                         goto out_unmap;
2054                 /*
2055                  * Chose the node of the first page. This could
2056                  * be more sophisticated and look at more pages,
2057                  * but isn't for now.
2058                  */
2059                 if (node == -1)
2060                         node = page_to_nid(page);
2061                 VM_BUG_ON(PageCompound(page));
2062                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2063                         goto out_unmap;
2064                 /* cannot use mapcount: can't collapse if there's a gup pin */
2065                 if (page_count(page) != 1)
2066                         goto out_unmap;
2067                 if (pte_young(pteval) || PageReferenced(page) ||
2068                     mmu_notifier_test_young(vma->vm_mm, address))
2069                         referenced = 1;
2070         }
2071         if (referenced)
2072                 ret = 1;
2073 out_unmap:
2074         pte_unmap_unlock(pte, ptl);
2075         if (ret)
2076                 /* collapse_huge_page will return with the mmap_sem released */
2077                 collapse_huge_page(mm, address, hpage, vma, node);
2078 out:
2079         return ret;
2080 }
2081
2082 static void collect_mm_slot(struct mm_slot *mm_slot)
2083 {
2084         struct mm_struct *mm = mm_slot->mm;
2085
2086         VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
2087
2088         if (khugepaged_test_exit(mm)) {
2089                 /* free mm_slot */
2090                 hlist_del(&mm_slot->hash);
2091                 list_del(&mm_slot->mm_node);
2092
2093                 /*
2094                  * Not strictly needed because the mm exited already.
2095                  *
2096                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2097                  */
2098
2099                 /* khugepaged_mm_lock actually not necessary for the below */
2100                 free_mm_slot(mm_slot);
2101                 mmdrop(mm);
2102         }
2103 }
2104
2105 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2106                                             struct page **hpage)
2107         __releases(&khugepaged_mm_lock)
2108         __acquires(&khugepaged_mm_lock)
2109 {
2110         struct mm_slot *mm_slot;
2111         struct mm_struct *mm;
2112         struct vm_area_struct *vma;
2113         int progress = 0;
2114
2115         VM_BUG_ON(!pages);
2116         VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
2117
2118         if (khugepaged_scan.mm_slot)
2119                 mm_slot = khugepaged_scan.mm_slot;
2120         else {
2121                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2122                                      struct mm_slot, mm_node);
2123                 khugepaged_scan.address = 0;
2124                 khugepaged_scan.mm_slot = mm_slot;
2125         }
2126         spin_unlock(&khugepaged_mm_lock);
2127
2128         mm = mm_slot->mm;
2129         down_read(&mm->mmap_sem);
2130         if (unlikely(khugepaged_test_exit(mm)))
2131                 vma = NULL;
2132         else
2133                 vma = find_vma(mm, khugepaged_scan.address);
2134
2135         progress++;
2136         for (; vma; vma = vma->vm_next) {
2137                 unsigned long hstart, hend;
2138
2139                 cond_resched();
2140                 if (unlikely(khugepaged_test_exit(mm))) {
2141                         progress++;
2142                         break;
2143                 }
2144
2145                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2146                      !khugepaged_always()) ||
2147                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2148                 skip:
2149                         progress++;
2150                         continue;
2151                 }
2152                 if (!vma->anon_vma || vma->vm_ops)
2153                         goto skip;
2154                 if (is_vma_temporary_stack(vma))
2155                         goto skip;
2156                 /*
2157                  * If is_pfn_mapping() is true is_learn_pfn_mapping()
2158                  * must be true too, verify it here.
2159                  */
2160                 VM_BUG_ON(is_linear_pfn_mapping(vma) ||
2161                           vma->vm_flags & VM_NO_THP);
2162
2163                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2164                 hend = vma->vm_end & HPAGE_PMD_MASK;
2165                 if (hstart >= hend)
2166                         goto skip;
2167                 if (khugepaged_scan.address > hend)
2168                         goto skip;
2169                 if (khugepaged_scan.address < hstart)
2170                         khugepaged_scan.address = hstart;
2171                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2172
2173                 while (khugepaged_scan.address < hend) {
2174                         int ret;
2175                         cond_resched();
2176                         if (unlikely(khugepaged_test_exit(mm)))
2177                                 goto breakouterloop;
2178
2179                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2180                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2181                                   hend);
2182                         ret = khugepaged_scan_pmd(mm, vma,
2183                                                   khugepaged_scan.address,
2184                                                   hpage);
2185                         /* move to next address */
2186                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2187                         progress += HPAGE_PMD_NR;
2188                         if (ret)
2189                                 /* we released mmap_sem so break loop */
2190                                 goto breakouterloop_mmap_sem;
2191                         if (progress >= pages)
2192                                 goto breakouterloop;
2193                 }
2194         }
2195 breakouterloop:
2196         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2197 breakouterloop_mmap_sem:
2198
2199         spin_lock(&khugepaged_mm_lock);
2200         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2201         /*
2202          * Release the current mm_slot if this mm is about to die, or
2203          * if we scanned all vmas of this mm.
2204          */
2205         if (khugepaged_test_exit(mm) || !vma) {
2206                 /*
2207                  * Make sure that if mm_users is reaching zero while
2208                  * khugepaged runs here, khugepaged_exit will find
2209                  * mm_slot not pointing to the exiting mm.
2210                  */
2211                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2212                         khugepaged_scan.mm_slot = list_entry(
2213                                 mm_slot->mm_node.next,
2214                                 struct mm_slot, mm_node);
2215                         khugepaged_scan.address = 0;
2216                 } else {
2217                         khugepaged_scan.mm_slot = NULL;
2218                         khugepaged_full_scans++;
2219                 }
2220
2221                 collect_mm_slot(mm_slot);
2222         }
2223
2224         return progress;
2225 }
2226
2227 static int khugepaged_has_work(void)
2228 {
2229         return !list_empty(&khugepaged_scan.mm_head) &&
2230                 khugepaged_enabled();
2231 }
2232
2233 static int khugepaged_wait_event(void)
2234 {
2235         return !list_empty(&khugepaged_scan.mm_head) ||
2236                 !khugepaged_enabled();
2237 }
2238
2239 static void khugepaged_do_scan(struct page **hpage)
2240 {
2241         unsigned int progress = 0, pass_through_head = 0;
2242         unsigned int pages = khugepaged_pages_to_scan;
2243
2244         barrier(); /* write khugepaged_pages_to_scan to local stack */
2245
2246         while (progress < pages) {
2247                 cond_resched();
2248
2249 #ifndef CONFIG_NUMA
2250                 if (!*hpage) {
2251                         *hpage = alloc_hugepage(khugepaged_defrag());
2252                         if (unlikely(!*hpage)) {
2253                                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2254                                 break;
2255                         }
2256                         count_vm_event(THP_COLLAPSE_ALLOC);
2257                 }
2258 #else
2259                 if (IS_ERR(*hpage))
2260                         break;
2261 #endif
2262
2263                 if (unlikely(kthread_should_stop() || freezing(current)))
2264                         break;
2265
2266                 spin_lock(&khugepaged_mm_lock);
2267                 if (!khugepaged_scan.mm_slot)
2268                         pass_through_head++;
2269                 if (khugepaged_has_work() &&
2270                     pass_through_head < 2)
2271                         progress += khugepaged_scan_mm_slot(pages - progress,
2272                                                             hpage);
2273                 else
2274                         progress = pages;
2275                 spin_unlock(&khugepaged_mm_lock);
2276         }
2277 }
2278
2279 static void khugepaged_alloc_sleep(void)
2280 {
2281         wait_event_freezable_timeout(khugepaged_wait, false,
2282                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2283 }
2284
2285 #ifndef CONFIG_NUMA
2286 static struct page *khugepaged_alloc_hugepage(void)
2287 {
2288         struct page *hpage;
2289
2290         do {
2291                 hpage = alloc_hugepage(khugepaged_defrag());
2292                 if (!hpage) {
2293                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2294                         khugepaged_alloc_sleep();
2295                 } else
2296                         count_vm_event(THP_COLLAPSE_ALLOC);
2297         } while (unlikely(!hpage) &&
2298                  likely(khugepaged_enabled()));
2299         return hpage;
2300 }
2301 #endif
2302
2303 static void khugepaged_loop(void)
2304 {
2305         struct page *hpage;
2306
2307 #ifdef CONFIG_NUMA
2308         hpage = NULL;
2309 #endif
2310         while (likely(khugepaged_enabled())) {
2311 #ifndef CONFIG_NUMA
2312                 hpage = khugepaged_alloc_hugepage();
2313                 if (unlikely(!hpage))
2314                         break;
2315 #else
2316                 if (IS_ERR(hpage)) {
2317                         khugepaged_alloc_sleep();
2318                         hpage = NULL;
2319                 }
2320 #endif
2321
2322                 khugepaged_do_scan(&hpage);
2323 #ifndef CONFIG_NUMA
2324                 if (hpage)
2325                         put_page(hpage);
2326 #endif
2327                 try_to_freeze();
2328                 if (unlikely(kthread_should_stop()))
2329                         break;
2330                 if (khugepaged_has_work()) {
2331                         if (!khugepaged_scan_sleep_millisecs)
2332                                 continue;
2333                         wait_event_freezable_timeout(khugepaged_wait, false,
2334                             msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2335                 } else if (khugepaged_enabled())
2336                         wait_event_freezable(khugepaged_wait,
2337                                              khugepaged_wait_event());
2338         }
2339 }
2340
2341 static int khugepaged(void *none)
2342 {
2343         struct mm_slot *mm_slot;
2344
2345         set_freezable();
2346         set_user_nice(current, 19);
2347
2348         /* serialize with start_khugepaged() */
2349         mutex_lock(&khugepaged_mutex);
2350
2351         for (;;) {
2352                 mutex_unlock(&khugepaged_mutex);
2353                 VM_BUG_ON(khugepaged_thread != current);
2354                 khugepaged_loop();
2355                 VM_BUG_ON(khugepaged_thread != current);
2356
2357                 mutex_lock(&khugepaged_mutex);
2358                 if (!khugepaged_enabled())
2359                         break;
2360                 if (unlikely(kthread_should_stop()))
2361                         break;
2362         }
2363
2364         spin_lock(&khugepaged_mm_lock);
2365         mm_slot = khugepaged_scan.mm_slot;
2366         khugepaged_scan.mm_slot = NULL;
2367         if (mm_slot)
2368                 collect_mm_slot(mm_slot);
2369         spin_unlock(&khugepaged_mm_lock);
2370
2371         khugepaged_thread = NULL;
2372         mutex_unlock(&khugepaged_mutex);
2373
2374         return 0;
2375 }
2376
2377 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2378 {
2379         struct page *page;
2380
2381         spin_lock(&mm->page_table_lock);
2382         if (unlikely(!pmd_trans_huge(*pmd))) {
2383                 spin_unlock(&mm->page_table_lock);
2384                 return;
2385         }
2386         page = pmd_page(*pmd);
2387         VM_BUG_ON(!page_count(page));
2388         get_page(page);
2389         spin_unlock(&mm->page_table_lock);
2390
2391         split_huge_page(page);
2392
2393         put_page(page);
2394         BUG_ON(pmd_trans_huge(*pmd));
2395 }
2396
2397 static void split_huge_page_address(struct mm_struct *mm,
2398                                     unsigned long address)
2399 {
2400         pgd_t *pgd;
2401         pud_t *pud;
2402         pmd_t *pmd;
2403
2404         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2405
2406         pgd = pgd_offset(mm, address);
2407         if (!pgd_present(*pgd))
2408                 return;
2409
2410         pud = pud_offset(pgd, address);
2411         if (!pud_present(*pud))
2412                 return;
2413
2414         pmd = pmd_offset(pud, address);
2415         if (!pmd_present(*pmd))
2416                 return;
2417         /*
2418          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2419          * materialize from under us.
2420          */
2421         split_huge_page_pmd(mm, pmd);
2422 }
2423
2424 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2425                              unsigned long start,
2426                              unsigned long end,
2427                              long adjust_next)
2428 {
2429         /*
2430          * If the new start address isn't hpage aligned and it could
2431          * previously contain an hugepage: check if we need to split
2432          * an huge pmd.
2433          */
2434         if (start & ~HPAGE_PMD_MASK &&
2435             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2436             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2437                 split_huge_page_address(vma->vm_mm, start);
2438
2439         /*
2440          * If the new end address isn't hpage aligned and it could
2441          * previously contain an hugepage: check if we need to split
2442          * an huge pmd.
2443          */
2444         if (end & ~HPAGE_PMD_MASK &&
2445             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2446             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2447                 split_huge_page_address(vma->vm_mm, end);
2448
2449         /*
2450          * If we're also updating the vma->vm_next->vm_start, if the new
2451          * vm_next->vm_start isn't page aligned and it could previously
2452          * contain an hugepage: check if we need to split an huge pmd.
2453          */
2454         if (adjust_next > 0) {
2455                 struct vm_area_struct *next = vma->vm_next;
2456                 unsigned long nstart = next->vm_start;
2457                 nstart += adjust_next << PAGE_SHIFT;
2458                 if (nstart & ~HPAGE_PMD_MASK &&
2459                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2460                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2461                         split_huge_page_address(next->vm_mm, nstart);
2462         }
2463 }