thp: reduce khugepaged freezing latency
[linux-3.10.git] / mm / huge_memory.c
1 /*
2  *  Copyright (C) 2009  Red Hat, Inc.
3  *
4  *  This work is licensed under the terms of the GNU GPL, version 2. See
5  *  the COPYING file in the top-level directory.
6  */
7
8 #include <linux/mm.h>
9 #include <linux/sched.h>
10 #include <linux/highmem.h>
11 #include <linux/hugetlb.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/rmap.h>
14 #include <linux/swap.h>
15 #include <linux/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 #endif /* CONFIG_SYSFS */
491
492 static int __init hugepage_init(void)
493 {
494         int err;
495 #ifdef CONFIG_SYSFS
496         static struct kobject *hugepage_kobj;
497 #endif
498
499         err = -EINVAL;
500         if (!has_transparent_hugepage()) {
501                 transparent_hugepage_flags = 0;
502                 goto out;
503         }
504
505 #ifdef CONFIG_SYSFS
506         err = -ENOMEM;
507         hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
508         if (unlikely(!hugepage_kobj)) {
509                 printk(KERN_ERR "hugepage: failed kobject create\n");
510                 goto out;
511         }
512
513         err = sysfs_create_group(hugepage_kobj, &hugepage_attr_group);
514         if (err) {
515                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
516                 goto out;
517         }
518
519         err = sysfs_create_group(hugepage_kobj, &khugepaged_attr_group);
520         if (err) {
521                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
522                 goto out;
523         }
524 #endif
525
526         err = khugepaged_slab_init();
527         if (err)
528                 goto out;
529
530         err = mm_slots_hash_init();
531         if (err) {
532                 khugepaged_slab_free();
533                 goto out;
534         }
535
536         /*
537          * By default disable transparent hugepages on smaller systems,
538          * where the extra memory used could hurt more than TLB overhead
539          * is likely to save.  The admin can still enable it through /sys.
540          */
541         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
542                 transparent_hugepage_flags = 0;
543
544         start_khugepaged();
545
546         set_recommended_min_free_kbytes();
547
548 out:
549         return err;
550 }
551 module_init(hugepage_init)
552
553 static int __init setup_transparent_hugepage(char *str)
554 {
555         int ret = 0;
556         if (!str)
557                 goto out;
558         if (!strcmp(str, "always")) {
559                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
560                         &transparent_hugepage_flags);
561                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
562                           &transparent_hugepage_flags);
563                 ret = 1;
564         } else if (!strcmp(str, "madvise")) {
565                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
566                           &transparent_hugepage_flags);
567                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
568                         &transparent_hugepage_flags);
569                 ret = 1;
570         } else if (!strcmp(str, "never")) {
571                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
572                           &transparent_hugepage_flags);
573                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
574                           &transparent_hugepage_flags);
575                 ret = 1;
576         }
577 out:
578         if (!ret)
579                 printk(KERN_WARNING
580                        "transparent_hugepage= cannot parse, ignored\n");
581         return ret;
582 }
583 __setup("transparent_hugepage=", setup_transparent_hugepage);
584
585 static void prepare_pmd_huge_pte(pgtable_t pgtable,
586                                  struct mm_struct *mm)
587 {
588         assert_spin_locked(&mm->page_table_lock);
589
590         /* FIFO */
591         if (!mm->pmd_huge_pte)
592                 INIT_LIST_HEAD(&pgtable->lru);
593         else
594                 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
595         mm->pmd_huge_pte = pgtable;
596 }
597
598 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
599 {
600         if (likely(vma->vm_flags & VM_WRITE))
601                 pmd = pmd_mkwrite(pmd);
602         return pmd;
603 }
604
605 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
606                                         struct vm_area_struct *vma,
607                                         unsigned long haddr, pmd_t *pmd,
608                                         struct page *page)
609 {
610         int ret = 0;
611         pgtable_t pgtable;
612
613         VM_BUG_ON(!PageCompound(page));
614         pgtable = pte_alloc_one(mm, haddr);
615         if (unlikely(!pgtable)) {
616                 mem_cgroup_uncharge_page(page);
617                 put_page(page);
618                 return VM_FAULT_OOM;
619         }
620
621         clear_huge_page(page, haddr, HPAGE_PMD_NR);
622         __SetPageUptodate(page);
623
624         spin_lock(&mm->page_table_lock);
625         if (unlikely(!pmd_none(*pmd))) {
626                 spin_unlock(&mm->page_table_lock);
627                 mem_cgroup_uncharge_page(page);
628                 put_page(page);
629                 pte_free(mm, pgtable);
630         } else {
631                 pmd_t entry;
632                 entry = mk_pmd(page, vma->vm_page_prot);
633                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
634                 entry = pmd_mkhuge(entry);
635                 /*
636                  * The spinlocking to take the lru_lock inside
637                  * page_add_new_anon_rmap() acts as a full memory
638                  * barrier to be sure clear_huge_page writes become
639                  * visible after the set_pmd_at() write.
640                  */
641                 page_add_new_anon_rmap(page, vma, haddr);
642                 set_pmd_at(mm, haddr, pmd, entry);
643                 prepare_pmd_huge_pte(pgtable, mm);
644                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
645                 spin_unlock(&mm->page_table_lock);
646         }
647
648         return ret;
649 }
650
651 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
652 {
653         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
654 }
655
656 static inline struct page *alloc_hugepage_vma(int defrag,
657                                               struct vm_area_struct *vma,
658                                               unsigned long haddr, int nd,
659                                               gfp_t extra_gfp)
660 {
661         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
662                                HPAGE_PMD_ORDER, vma, haddr, nd);
663 }
664
665 #ifndef CONFIG_NUMA
666 static inline struct page *alloc_hugepage(int defrag)
667 {
668         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
669                            HPAGE_PMD_ORDER);
670 }
671 #endif
672
673 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
674                                unsigned long address, pmd_t *pmd,
675                                unsigned int flags)
676 {
677         struct page *page;
678         unsigned long haddr = address & HPAGE_PMD_MASK;
679         pte_t *pte;
680
681         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
682                 if (unlikely(anon_vma_prepare(vma)))
683                         return VM_FAULT_OOM;
684                 if (unlikely(khugepaged_enter(vma)))
685                         return VM_FAULT_OOM;
686                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
687                                           vma, haddr, numa_node_id(), 0);
688                 if (unlikely(!page)) {
689                         count_vm_event(THP_FAULT_FALLBACK);
690                         goto out;
691                 }
692                 count_vm_event(THP_FAULT_ALLOC);
693                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
694                         put_page(page);
695                         goto out;
696                 }
697
698                 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
699         }
700 out:
701         /*
702          * Use __pte_alloc instead of pte_alloc_map, because we can't
703          * run pte_offset_map on the pmd, if an huge pmd could
704          * materialize from under us from a different thread.
705          */
706         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
707                 return VM_FAULT_OOM;
708         /* if an huge pmd materialized from under us just retry later */
709         if (unlikely(pmd_trans_huge(*pmd)))
710                 return 0;
711         /*
712          * A regular pmd is established and it can't morph into a huge pmd
713          * from under us anymore at this point because we hold the mmap_sem
714          * read mode and khugepaged takes it in write mode. So now it's
715          * safe to run pte_offset_map().
716          */
717         pte = pte_offset_map(pmd, address);
718         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
719 }
720
721 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
722                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
723                   struct vm_area_struct *vma)
724 {
725         struct page *src_page;
726         pmd_t pmd;
727         pgtable_t pgtable;
728         int ret;
729
730         ret = -ENOMEM;
731         pgtable = pte_alloc_one(dst_mm, addr);
732         if (unlikely(!pgtable))
733                 goto out;
734
735         spin_lock(&dst_mm->page_table_lock);
736         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
737
738         ret = -EAGAIN;
739         pmd = *src_pmd;
740         if (unlikely(!pmd_trans_huge(pmd))) {
741                 pte_free(dst_mm, pgtable);
742                 goto out_unlock;
743         }
744         if (unlikely(pmd_trans_splitting(pmd))) {
745                 /* split huge page running from under us */
746                 spin_unlock(&src_mm->page_table_lock);
747                 spin_unlock(&dst_mm->page_table_lock);
748                 pte_free(dst_mm, pgtable);
749
750                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
751                 goto out;
752         }
753         src_page = pmd_page(pmd);
754         VM_BUG_ON(!PageHead(src_page));
755         get_page(src_page);
756         page_dup_rmap(src_page);
757         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
758
759         pmdp_set_wrprotect(src_mm, addr, src_pmd);
760         pmd = pmd_mkold(pmd_wrprotect(pmd));
761         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
762         prepare_pmd_huge_pte(pgtable, dst_mm);
763
764         ret = 0;
765 out_unlock:
766         spin_unlock(&src_mm->page_table_lock);
767         spin_unlock(&dst_mm->page_table_lock);
768 out:
769         return ret;
770 }
771
772 /* no "address" argument so destroys page coloring of some arch */
773 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
774 {
775         pgtable_t pgtable;
776
777         assert_spin_locked(&mm->page_table_lock);
778
779         /* FIFO */
780         pgtable = mm->pmd_huge_pte;
781         if (list_empty(&pgtable->lru))
782                 mm->pmd_huge_pte = NULL;
783         else {
784                 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
785                                               struct page, lru);
786                 list_del(&pgtable->lru);
787         }
788         return pgtable;
789 }
790
791 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
792                                         struct vm_area_struct *vma,
793                                         unsigned long address,
794                                         pmd_t *pmd, pmd_t orig_pmd,
795                                         struct page *page,
796                                         unsigned long haddr)
797 {
798         pgtable_t pgtable;
799         pmd_t _pmd;
800         int ret = 0, i;
801         struct page **pages;
802
803         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
804                         GFP_KERNEL);
805         if (unlikely(!pages)) {
806                 ret |= VM_FAULT_OOM;
807                 goto out;
808         }
809
810         for (i = 0; i < HPAGE_PMD_NR; i++) {
811                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
812                                                __GFP_OTHER_NODE,
813                                                vma, address, page_to_nid(page));
814                 if (unlikely(!pages[i] ||
815                              mem_cgroup_newpage_charge(pages[i], mm,
816                                                        GFP_KERNEL))) {
817                         if (pages[i])
818                                 put_page(pages[i]);
819                         mem_cgroup_uncharge_start();
820                         while (--i >= 0) {
821                                 mem_cgroup_uncharge_page(pages[i]);
822                                 put_page(pages[i]);
823                         }
824                         mem_cgroup_uncharge_end();
825                         kfree(pages);
826                         ret |= VM_FAULT_OOM;
827                         goto out;
828                 }
829         }
830
831         for (i = 0; i < HPAGE_PMD_NR; i++) {
832                 copy_user_highpage(pages[i], page + i,
833                                    haddr + PAGE_SIZE * i, vma);
834                 __SetPageUptodate(pages[i]);
835                 cond_resched();
836         }
837
838         spin_lock(&mm->page_table_lock);
839         if (unlikely(!pmd_same(*pmd, orig_pmd)))
840                 goto out_free_pages;
841         VM_BUG_ON(!PageHead(page));
842
843         pmdp_clear_flush_notify(vma, haddr, pmd);
844         /* leave pmd empty until pte is filled */
845
846         pgtable = get_pmd_huge_pte(mm);
847         pmd_populate(mm, &_pmd, pgtable);
848
849         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
850                 pte_t *pte, entry;
851                 entry = mk_pte(pages[i], vma->vm_page_prot);
852                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
853                 page_add_new_anon_rmap(pages[i], vma, haddr);
854                 pte = pte_offset_map(&_pmd, haddr);
855                 VM_BUG_ON(!pte_none(*pte));
856                 set_pte_at(mm, haddr, pte, entry);
857                 pte_unmap(pte);
858         }
859         kfree(pages);
860
861         mm->nr_ptes++;
862         smp_wmb(); /* make pte visible before pmd */
863         pmd_populate(mm, pmd, pgtable);
864         page_remove_rmap(page);
865         spin_unlock(&mm->page_table_lock);
866
867         ret |= VM_FAULT_WRITE;
868         put_page(page);
869
870 out:
871         return ret;
872
873 out_free_pages:
874         spin_unlock(&mm->page_table_lock);
875         mem_cgroup_uncharge_start();
876         for (i = 0; i < HPAGE_PMD_NR; i++) {
877                 mem_cgroup_uncharge_page(pages[i]);
878                 put_page(pages[i]);
879         }
880         mem_cgroup_uncharge_end();
881         kfree(pages);
882         goto out;
883 }
884
885 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
886                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
887 {
888         int ret = 0;
889         struct page *page, *new_page;
890         unsigned long haddr;
891
892         VM_BUG_ON(!vma->anon_vma);
893         spin_lock(&mm->page_table_lock);
894         if (unlikely(!pmd_same(*pmd, orig_pmd)))
895                 goto out_unlock;
896
897         page = pmd_page(orig_pmd);
898         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
899         haddr = address & HPAGE_PMD_MASK;
900         if (page_mapcount(page) == 1) {
901                 pmd_t entry;
902                 entry = pmd_mkyoung(orig_pmd);
903                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
904                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
905                         update_mmu_cache(vma, address, entry);
906                 ret |= VM_FAULT_WRITE;
907                 goto out_unlock;
908         }
909         get_page(page);
910         spin_unlock(&mm->page_table_lock);
911
912         if (transparent_hugepage_enabled(vma) &&
913             !transparent_hugepage_debug_cow())
914                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
915                                               vma, haddr, numa_node_id(), 0);
916         else
917                 new_page = NULL;
918
919         if (unlikely(!new_page)) {
920                 count_vm_event(THP_FAULT_FALLBACK);
921                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
922                                                    pmd, orig_pmd, page, haddr);
923                 put_page(page);
924                 goto out;
925         }
926         count_vm_event(THP_FAULT_ALLOC);
927
928         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
929                 put_page(new_page);
930                 put_page(page);
931                 ret |= VM_FAULT_OOM;
932                 goto out;
933         }
934
935         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
936         __SetPageUptodate(new_page);
937
938         spin_lock(&mm->page_table_lock);
939         put_page(page);
940         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
941                 mem_cgroup_uncharge_page(new_page);
942                 put_page(new_page);
943         } else {
944                 pmd_t entry;
945                 VM_BUG_ON(!PageHead(page));
946                 entry = mk_pmd(new_page, vma->vm_page_prot);
947                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
948                 entry = pmd_mkhuge(entry);
949                 pmdp_clear_flush_notify(vma, haddr, pmd);
950                 page_add_new_anon_rmap(new_page, vma, haddr);
951                 set_pmd_at(mm, haddr, pmd, entry);
952                 update_mmu_cache(vma, address, entry);
953                 page_remove_rmap(page);
954                 put_page(page);
955                 ret |= VM_FAULT_WRITE;
956         }
957 out_unlock:
958         spin_unlock(&mm->page_table_lock);
959 out:
960         return ret;
961 }
962
963 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
964                                    unsigned long addr,
965                                    pmd_t *pmd,
966                                    unsigned int flags)
967 {
968         struct page *page = NULL;
969
970         assert_spin_locked(&mm->page_table_lock);
971
972         if (flags & FOLL_WRITE && !pmd_write(*pmd))
973                 goto out;
974
975         page = pmd_page(*pmd);
976         VM_BUG_ON(!PageHead(page));
977         if (flags & FOLL_TOUCH) {
978                 pmd_t _pmd;
979                 /*
980                  * We should set the dirty bit only for FOLL_WRITE but
981                  * for now the dirty bit in the pmd is meaningless.
982                  * And if the dirty bit will become meaningful and
983                  * we'll only set it with FOLL_WRITE, an atomic
984                  * set_bit will be required on the pmd to set the
985                  * young bit, instead of the current set_pmd_at.
986                  */
987                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
988                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
989         }
990         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
991         VM_BUG_ON(!PageCompound(page));
992         if (flags & FOLL_GET)
993                 get_page_foll(page);
994
995 out:
996         return page;
997 }
998
999 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1000                  pmd_t *pmd)
1001 {
1002         int ret = 0;
1003
1004         spin_lock(&tlb->mm->page_table_lock);
1005         if (likely(pmd_trans_huge(*pmd))) {
1006                 if (unlikely(pmd_trans_splitting(*pmd))) {
1007                         spin_unlock(&tlb->mm->page_table_lock);
1008                         wait_split_huge_page(vma->anon_vma,
1009                                              pmd);
1010                 } else {
1011                         struct page *page;
1012                         pgtable_t pgtable;
1013                         pgtable = get_pmd_huge_pte(tlb->mm);
1014                         page = pmd_page(*pmd);
1015                         pmd_clear(pmd);
1016                         page_remove_rmap(page);
1017                         VM_BUG_ON(page_mapcount(page) < 0);
1018                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1019                         VM_BUG_ON(!PageHead(page));
1020                         spin_unlock(&tlb->mm->page_table_lock);
1021                         tlb_remove_page(tlb, page);
1022                         pte_free(tlb->mm, pgtable);
1023                         ret = 1;
1024                 }
1025         } else
1026                 spin_unlock(&tlb->mm->page_table_lock);
1027
1028         return ret;
1029 }
1030
1031 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1032                 unsigned long addr, unsigned long end,
1033                 unsigned char *vec)
1034 {
1035         int ret = 0;
1036
1037         spin_lock(&vma->vm_mm->page_table_lock);
1038         if (likely(pmd_trans_huge(*pmd))) {
1039                 ret = !pmd_trans_splitting(*pmd);
1040                 spin_unlock(&vma->vm_mm->page_table_lock);
1041                 if (unlikely(!ret))
1042                         wait_split_huge_page(vma->anon_vma, pmd);
1043                 else {
1044                         /*
1045                          * All logical pages in the range are present
1046                          * if backed by a huge page.
1047                          */
1048                         memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1049                 }
1050         } else
1051                 spin_unlock(&vma->vm_mm->page_table_lock);
1052
1053         return ret;
1054 }
1055
1056 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1057                   unsigned long old_addr,
1058                   unsigned long new_addr, unsigned long old_end,
1059                   pmd_t *old_pmd, pmd_t *new_pmd)
1060 {
1061         int ret = 0;
1062         pmd_t pmd;
1063
1064         struct mm_struct *mm = vma->vm_mm;
1065
1066         if ((old_addr & ~HPAGE_PMD_MASK) ||
1067             (new_addr & ~HPAGE_PMD_MASK) ||
1068             old_end - old_addr < HPAGE_PMD_SIZE ||
1069             (new_vma->vm_flags & VM_NOHUGEPAGE))
1070                 goto out;
1071
1072         /*
1073          * The destination pmd shouldn't be established, free_pgtables()
1074          * should have release it.
1075          */
1076         if (WARN_ON(!pmd_none(*new_pmd))) {
1077                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1078                 goto out;
1079         }
1080
1081         spin_lock(&mm->page_table_lock);
1082         if (likely(pmd_trans_huge(*old_pmd))) {
1083                 if (pmd_trans_splitting(*old_pmd)) {
1084                         spin_unlock(&mm->page_table_lock);
1085                         wait_split_huge_page(vma->anon_vma, old_pmd);
1086                         ret = -1;
1087                 } else {
1088                         pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1089                         VM_BUG_ON(!pmd_none(*new_pmd));
1090                         set_pmd_at(mm, new_addr, new_pmd, pmd);
1091                         spin_unlock(&mm->page_table_lock);
1092                         ret = 1;
1093                 }
1094         } else {
1095                 spin_unlock(&mm->page_table_lock);
1096         }
1097 out:
1098         return ret;
1099 }
1100
1101 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1102                 unsigned long addr, pgprot_t newprot)
1103 {
1104         struct mm_struct *mm = vma->vm_mm;
1105         int ret = 0;
1106
1107         spin_lock(&mm->page_table_lock);
1108         if (likely(pmd_trans_huge(*pmd))) {
1109                 if (unlikely(pmd_trans_splitting(*pmd))) {
1110                         spin_unlock(&mm->page_table_lock);
1111                         wait_split_huge_page(vma->anon_vma, pmd);
1112                 } else {
1113                         pmd_t entry;
1114
1115                         entry = pmdp_get_and_clear(mm, addr, pmd);
1116                         entry = pmd_modify(entry, newprot);
1117                         set_pmd_at(mm, addr, pmd, entry);
1118                         spin_unlock(&vma->vm_mm->page_table_lock);
1119                         flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1120                         ret = 1;
1121                 }
1122         } else
1123                 spin_unlock(&vma->vm_mm->page_table_lock);
1124
1125         return ret;
1126 }
1127
1128 pmd_t *page_check_address_pmd(struct page *page,
1129                               struct mm_struct *mm,
1130                               unsigned long address,
1131                               enum page_check_address_pmd_flag flag)
1132 {
1133         pgd_t *pgd;
1134         pud_t *pud;
1135         pmd_t *pmd, *ret = NULL;
1136
1137         if (address & ~HPAGE_PMD_MASK)
1138                 goto out;
1139
1140         pgd = pgd_offset(mm, address);
1141         if (!pgd_present(*pgd))
1142                 goto out;
1143
1144         pud = pud_offset(pgd, address);
1145         if (!pud_present(*pud))
1146                 goto out;
1147
1148         pmd = pmd_offset(pud, address);
1149         if (pmd_none(*pmd))
1150                 goto out;
1151         if (pmd_page(*pmd) != page)
1152                 goto out;
1153         /*
1154          * split_vma() may create temporary aliased mappings. There is
1155          * no risk as long as all huge pmd are found and have their
1156          * splitting bit set before __split_huge_page_refcount
1157          * runs. Finding the same huge pmd more than once during the
1158          * same rmap walk is not a problem.
1159          */
1160         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1161             pmd_trans_splitting(*pmd))
1162                 goto out;
1163         if (pmd_trans_huge(*pmd)) {
1164                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1165                           !pmd_trans_splitting(*pmd));
1166                 ret = pmd;
1167         }
1168 out:
1169         return ret;
1170 }
1171
1172 static int __split_huge_page_splitting(struct page *page,
1173                                        struct vm_area_struct *vma,
1174                                        unsigned long address)
1175 {
1176         struct mm_struct *mm = vma->vm_mm;
1177         pmd_t *pmd;
1178         int ret = 0;
1179
1180         spin_lock(&mm->page_table_lock);
1181         pmd = page_check_address_pmd(page, mm, address,
1182                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1183         if (pmd) {
1184                 /*
1185                  * We can't temporarily set the pmd to null in order
1186                  * to split it, the pmd must remain marked huge at all
1187                  * times or the VM won't take the pmd_trans_huge paths
1188                  * and it won't wait on the anon_vma->root->mutex to
1189                  * serialize against split_huge_page*.
1190                  */
1191                 pmdp_splitting_flush_notify(vma, address, pmd);
1192                 ret = 1;
1193         }
1194         spin_unlock(&mm->page_table_lock);
1195
1196         return ret;
1197 }
1198
1199 static void __split_huge_page_refcount(struct page *page)
1200 {
1201         int i;
1202         unsigned long head_index = page->index;
1203         struct zone *zone = page_zone(page);
1204         int zonestat;
1205         int tail_count = 0;
1206
1207         /* prevent PageLRU to go away from under us, and freeze lru stats */
1208         spin_lock_irq(&zone->lru_lock);
1209         compound_lock(page);
1210
1211         for (i = 1; i < HPAGE_PMD_NR; i++) {
1212                 struct page *page_tail = page + i;
1213
1214                 /* tail_page->_mapcount cannot change */
1215                 BUG_ON(page_mapcount(page_tail) < 0);
1216                 tail_count += page_mapcount(page_tail);
1217                 /* check for overflow */
1218                 BUG_ON(tail_count < 0);
1219                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1220                 /*
1221                  * tail_page->_count is zero and not changing from
1222                  * under us. But get_page_unless_zero() may be running
1223                  * from under us on the tail_page. If we used
1224                  * atomic_set() below instead of atomic_add(), we
1225                  * would then run atomic_set() concurrently with
1226                  * get_page_unless_zero(), and atomic_set() is
1227                  * implemented in C not using locked ops. spin_unlock
1228                  * on x86 sometime uses locked ops because of PPro
1229                  * errata 66, 92, so unless somebody can guarantee
1230                  * atomic_set() here would be safe on all archs (and
1231                  * not only on x86), it's safer to use atomic_add().
1232                  */
1233                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1234                            &page_tail->_count);
1235
1236                 /* after clearing PageTail the gup refcount can be released */
1237                 smp_mb();
1238
1239                 /*
1240                  * retain hwpoison flag of the poisoned tail page:
1241                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1242                  *   by the memory-failure.
1243                  */
1244                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1245                 page_tail->flags |= (page->flags &
1246                                      ((1L << PG_referenced) |
1247                                       (1L << PG_swapbacked) |
1248                                       (1L << PG_mlocked) |
1249                                       (1L << PG_uptodate)));
1250                 page_tail->flags |= (1L << PG_dirty);
1251
1252                 /* clear PageTail before overwriting first_page */
1253                 smp_wmb();
1254
1255                 /*
1256                  * __split_huge_page_splitting() already set the
1257                  * splitting bit in all pmd that could map this
1258                  * hugepage, that will ensure no CPU can alter the
1259                  * mapcount on the head page. The mapcount is only
1260                  * accounted in the head page and it has to be
1261                  * transferred to all tail pages in the below code. So
1262                  * for this code to be safe, the split the mapcount
1263                  * can't change. But that doesn't mean userland can't
1264                  * keep changing and reading the page contents while
1265                  * we transfer the mapcount, so the pmd splitting
1266                  * status is achieved setting a reserved bit in the
1267                  * pmd, not by clearing the present bit.
1268                 */
1269                 page_tail->_mapcount = page->_mapcount;
1270
1271                 BUG_ON(page_tail->mapping);
1272                 page_tail->mapping = page->mapping;
1273
1274                 page_tail->index = ++head_index;
1275
1276                 BUG_ON(!PageAnon(page_tail));
1277                 BUG_ON(!PageUptodate(page_tail));
1278                 BUG_ON(!PageDirty(page_tail));
1279                 BUG_ON(!PageSwapBacked(page_tail));
1280
1281                 mem_cgroup_split_huge_fixup(page, page_tail);
1282
1283                 lru_add_page_tail(zone, page, page_tail);
1284         }
1285         atomic_sub(tail_count, &page->_count);
1286         BUG_ON(atomic_read(&page->_count) <= 0);
1287
1288         __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1289         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1290
1291         /*
1292          * A hugepage counts for HPAGE_PMD_NR pages on the LRU statistics,
1293          * so adjust those appropriately if this page is on the LRU.
1294          */
1295         if (PageLRU(page)) {
1296                 zonestat = NR_LRU_BASE + page_lru(page);
1297                 __mod_zone_page_state(zone, zonestat, -(HPAGE_PMD_NR-1));
1298         }
1299
1300         ClearPageCompound(page);
1301         compound_unlock(page);
1302         spin_unlock_irq(&zone->lru_lock);
1303
1304         for (i = 1; i < HPAGE_PMD_NR; i++) {
1305                 struct page *page_tail = page + i;
1306                 BUG_ON(page_count(page_tail) <= 0);
1307                 /*
1308                  * Tail pages may be freed if there wasn't any mapping
1309                  * like if add_to_swap() is running on a lru page that
1310                  * had its mapping zapped. And freeing these pages
1311                  * requires taking the lru_lock so we do the put_page
1312                  * of the tail pages after the split is complete.
1313                  */
1314                 put_page(page_tail);
1315         }
1316
1317         /*
1318          * Only the head page (now become a regular page) is required
1319          * to be pinned by the caller.
1320          */
1321         BUG_ON(page_count(page) <= 0);
1322 }
1323
1324 static int __split_huge_page_map(struct page *page,
1325                                  struct vm_area_struct *vma,
1326                                  unsigned long address)
1327 {
1328         struct mm_struct *mm = vma->vm_mm;
1329         pmd_t *pmd, _pmd;
1330         int ret = 0, i;
1331         pgtable_t pgtable;
1332         unsigned long haddr;
1333
1334         spin_lock(&mm->page_table_lock);
1335         pmd = page_check_address_pmd(page, mm, address,
1336                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1337         if (pmd) {
1338                 pgtable = get_pmd_huge_pte(mm);
1339                 pmd_populate(mm, &_pmd, pgtable);
1340
1341                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1342                      i++, haddr += PAGE_SIZE) {
1343                         pte_t *pte, entry;
1344                         BUG_ON(PageCompound(page+i));
1345                         entry = mk_pte(page + i, vma->vm_page_prot);
1346                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1347                         if (!pmd_write(*pmd))
1348                                 entry = pte_wrprotect(entry);
1349                         else
1350                                 BUG_ON(page_mapcount(page) != 1);
1351                         if (!pmd_young(*pmd))
1352                                 entry = pte_mkold(entry);
1353                         pte = pte_offset_map(&_pmd, haddr);
1354                         BUG_ON(!pte_none(*pte));
1355                         set_pte_at(mm, haddr, pte, entry);
1356                         pte_unmap(pte);
1357                 }
1358
1359                 mm->nr_ptes++;
1360                 smp_wmb(); /* make pte visible before pmd */
1361                 /*
1362                  * Up to this point the pmd is present and huge and
1363                  * userland has the whole access to the hugepage
1364                  * during the split (which happens in place). If we
1365                  * overwrite the pmd with the not-huge version
1366                  * pointing to the pte here (which of course we could
1367                  * if all CPUs were bug free), userland could trigger
1368                  * a small page size TLB miss on the small sized TLB
1369                  * while the hugepage TLB entry is still established
1370                  * in the huge TLB. Some CPU doesn't like that. See
1371                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1372                  * Erratum 383 on page 93. Intel should be safe but is
1373                  * also warns that it's only safe if the permission
1374                  * and cache attributes of the two entries loaded in
1375                  * the two TLB is identical (which should be the case
1376                  * here). But it is generally safer to never allow
1377                  * small and huge TLB entries for the same virtual
1378                  * address to be loaded simultaneously. So instead of
1379                  * doing "pmd_populate(); flush_tlb_range();" we first
1380                  * mark the current pmd notpresent (atomically because
1381                  * here the pmd_trans_huge and pmd_trans_splitting
1382                  * must remain set at all times on the pmd until the
1383                  * split is complete for this pmd), then we flush the
1384                  * SMP TLB and finally we write the non-huge version
1385                  * of the pmd entry with pmd_populate.
1386                  */
1387                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1388                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1389                 pmd_populate(mm, pmd, pgtable);
1390                 ret = 1;
1391         }
1392         spin_unlock(&mm->page_table_lock);
1393
1394         return ret;
1395 }
1396
1397 /* must be called with anon_vma->root->mutex hold */
1398 static void __split_huge_page(struct page *page,
1399                               struct anon_vma *anon_vma)
1400 {
1401         int mapcount, mapcount2;
1402         struct anon_vma_chain *avc;
1403
1404         BUG_ON(!PageHead(page));
1405         BUG_ON(PageTail(page));
1406
1407         mapcount = 0;
1408         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1409                 struct vm_area_struct *vma = avc->vma;
1410                 unsigned long addr = vma_address(page, vma);
1411                 BUG_ON(is_vma_temporary_stack(vma));
1412                 if (addr == -EFAULT)
1413                         continue;
1414                 mapcount += __split_huge_page_splitting(page, vma, addr);
1415         }
1416         /*
1417          * It is critical that new vmas are added to the tail of the
1418          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1419          * and establishes a child pmd before
1420          * __split_huge_page_splitting() freezes the parent pmd (so if
1421          * we fail to prevent copy_huge_pmd() from running until the
1422          * whole __split_huge_page() is complete), we will still see
1423          * the newly established pmd of the child later during the
1424          * walk, to be able to set it as pmd_trans_splitting too.
1425          */
1426         if (mapcount != page_mapcount(page))
1427                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1428                        mapcount, page_mapcount(page));
1429         BUG_ON(mapcount != page_mapcount(page));
1430
1431         __split_huge_page_refcount(page);
1432
1433         mapcount2 = 0;
1434         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1435                 struct vm_area_struct *vma = avc->vma;
1436                 unsigned long addr = vma_address(page, vma);
1437                 BUG_ON(is_vma_temporary_stack(vma));
1438                 if (addr == -EFAULT)
1439                         continue;
1440                 mapcount2 += __split_huge_page_map(page, vma, addr);
1441         }
1442         if (mapcount != mapcount2)
1443                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1444                        mapcount, mapcount2, page_mapcount(page));
1445         BUG_ON(mapcount != mapcount2);
1446 }
1447
1448 int split_huge_page(struct page *page)
1449 {
1450         struct anon_vma *anon_vma;
1451         int ret = 1;
1452
1453         BUG_ON(!PageAnon(page));
1454         anon_vma = page_lock_anon_vma(page);
1455         if (!anon_vma)
1456                 goto out;
1457         ret = 0;
1458         if (!PageCompound(page))
1459                 goto out_unlock;
1460
1461         BUG_ON(!PageSwapBacked(page));
1462         __split_huge_page(page, anon_vma);
1463         count_vm_event(THP_SPLIT);
1464
1465         BUG_ON(PageCompound(page));
1466 out_unlock:
1467         page_unlock_anon_vma(anon_vma);
1468 out:
1469         return ret;
1470 }
1471
1472 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1473                    VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1474
1475 int hugepage_madvise(struct vm_area_struct *vma,
1476                      unsigned long *vm_flags, int advice)
1477 {
1478         switch (advice) {
1479         case MADV_HUGEPAGE:
1480                 /*
1481                  * Be somewhat over-protective like KSM for now!
1482                  */
1483                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1484                         return -EINVAL;
1485                 *vm_flags &= ~VM_NOHUGEPAGE;
1486                 *vm_flags |= VM_HUGEPAGE;
1487                 /*
1488                  * If the vma become good for khugepaged to scan,
1489                  * register it here without waiting a page fault that
1490                  * may not happen any time soon.
1491                  */
1492                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1493                         return -ENOMEM;
1494                 break;
1495         case MADV_NOHUGEPAGE:
1496                 /*
1497                  * Be somewhat over-protective like KSM for now!
1498                  */
1499                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1500                         return -EINVAL;
1501                 *vm_flags &= ~VM_HUGEPAGE;
1502                 *vm_flags |= VM_NOHUGEPAGE;
1503                 /*
1504                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1505                  * this vma even if we leave the mm registered in khugepaged if
1506                  * it got registered before VM_NOHUGEPAGE was set.
1507                  */
1508                 break;
1509         }
1510
1511         return 0;
1512 }
1513
1514 static int __init khugepaged_slab_init(void)
1515 {
1516         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1517                                           sizeof(struct mm_slot),
1518                                           __alignof__(struct mm_slot), 0, NULL);
1519         if (!mm_slot_cache)
1520                 return -ENOMEM;
1521
1522         return 0;
1523 }
1524
1525 static void __init khugepaged_slab_free(void)
1526 {
1527         kmem_cache_destroy(mm_slot_cache);
1528         mm_slot_cache = NULL;
1529 }
1530
1531 static inline struct mm_slot *alloc_mm_slot(void)
1532 {
1533         if (!mm_slot_cache)     /* initialization failed */
1534                 return NULL;
1535         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1536 }
1537
1538 static inline void free_mm_slot(struct mm_slot *mm_slot)
1539 {
1540         kmem_cache_free(mm_slot_cache, mm_slot);
1541 }
1542
1543 static int __init mm_slots_hash_init(void)
1544 {
1545         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1546                                 GFP_KERNEL);
1547         if (!mm_slots_hash)
1548                 return -ENOMEM;
1549         return 0;
1550 }
1551
1552 #if 0
1553 static void __init mm_slots_hash_free(void)
1554 {
1555         kfree(mm_slots_hash);
1556         mm_slots_hash = NULL;
1557 }
1558 #endif
1559
1560 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1561 {
1562         struct mm_slot *mm_slot;
1563         struct hlist_head *bucket;
1564         struct hlist_node *node;
1565
1566         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1567                                 % MM_SLOTS_HASH_HEADS];
1568         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1569                 if (mm == mm_slot->mm)
1570                         return mm_slot;
1571         }
1572         return NULL;
1573 }
1574
1575 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1576                                     struct mm_slot *mm_slot)
1577 {
1578         struct hlist_head *bucket;
1579
1580         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1581                                 % MM_SLOTS_HASH_HEADS];
1582         mm_slot->mm = mm;
1583         hlist_add_head(&mm_slot->hash, bucket);
1584 }
1585
1586 static inline int khugepaged_test_exit(struct mm_struct *mm)
1587 {
1588         return atomic_read(&mm->mm_users) == 0;
1589 }
1590
1591 int __khugepaged_enter(struct mm_struct *mm)
1592 {
1593         struct mm_slot *mm_slot;
1594         int wakeup;
1595
1596         mm_slot = alloc_mm_slot();
1597         if (!mm_slot)
1598                 return -ENOMEM;
1599
1600         /* __khugepaged_exit() must not run from under us */
1601         VM_BUG_ON(khugepaged_test_exit(mm));
1602         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1603                 free_mm_slot(mm_slot);
1604                 return 0;
1605         }
1606
1607         spin_lock(&khugepaged_mm_lock);
1608         insert_to_mm_slots_hash(mm, mm_slot);
1609         /*
1610          * Insert just behind the scanning cursor, to let the area settle
1611          * down a little.
1612          */
1613         wakeup = list_empty(&khugepaged_scan.mm_head);
1614         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1615         spin_unlock(&khugepaged_mm_lock);
1616
1617         atomic_inc(&mm->mm_count);
1618         if (wakeup)
1619                 wake_up_interruptible(&khugepaged_wait);
1620
1621         return 0;
1622 }
1623
1624 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1625 {
1626         unsigned long hstart, hend;
1627         if (!vma->anon_vma)
1628                 /*
1629                  * Not yet faulted in so we will register later in the
1630                  * page fault if needed.
1631                  */
1632                 return 0;
1633         if (vma->vm_ops)
1634                 /* khugepaged not yet working on file or special mappings */
1635                 return 0;
1636         /*
1637          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1638          * true too, verify it here.
1639          */
1640         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1641         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1642         hend = vma->vm_end & HPAGE_PMD_MASK;
1643         if (hstart < hend)
1644                 return khugepaged_enter(vma);
1645         return 0;
1646 }
1647
1648 void __khugepaged_exit(struct mm_struct *mm)
1649 {
1650         struct mm_slot *mm_slot;
1651         int free = 0;
1652
1653         spin_lock(&khugepaged_mm_lock);
1654         mm_slot = get_mm_slot(mm);
1655         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1656                 hlist_del(&mm_slot->hash);
1657                 list_del(&mm_slot->mm_node);
1658                 free = 1;
1659         }
1660         spin_unlock(&khugepaged_mm_lock);
1661
1662         if (free) {
1663                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1664                 free_mm_slot(mm_slot);
1665                 mmdrop(mm);
1666         } else if (mm_slot) {
1667                 /*
1668                  * This is required to serialize against
1669                  * khugepaged_test_exit() (which is guaranteed to run
1670                  * under mmap sem read mode). Stop here (after we
1671                  * return all pagetables will be destroyed) until
1672                  * khugepaged has finished working on the pagetables
1673                  * under the mmap_sem.
1674                  */
1675                 down_write(&mm->mmap_sem);
1676                 up_write(&mm->mmap_sem);
1677         }
1678 }
1679
1680 static void release_pte_page(struct page *page)
1681 {
1682         /* 0 stands for page_is_file_cache(page) == false */
1683         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1684         unlock_page(page);
1685         putback_lru_page(page);
1686 }
1687
1688 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1689 {
1690         while (--_pte >= pte) {
1691                 pte_t pteval = *_pte;
1692                 if (!pte_none(pteval))
1693                         release_pte_page(pte_page(pteval));
1694         }
1695 }
1696
1697 static void release_all_pte_pages(pte_t *pte)
1698 {
1699         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1700 }
1701
1702 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1703                                         unsigned long address,
1704                                         pte_t *pte)
1705 {
1706         struct page *page;
1707         pte_t *_pte;
1708         int referenced = 0, isolated = 0, none = 0;
1709         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1710              _pte++, address += PAGE_SIZE) {
1711                 pte_t pteval = *_pte;
1712                 if (pte_none(pteval)) {
1713                         if (++none <= khugepaged_max_ptes_none)
1714                                 continue;
1715                         else {
1716                                 release_pte_pages(pte, _pte);
1717                                 goto out;
1718                         }
1719                 }
1720                 if (!pte_present(pteval) || !pte_write(pteval)) {
1721                         release_pte_pages(pte, _pte);
1722                         goto out;
1723                 }
1724                 page = vm_normal_page(vma, address, pteval);
1725                 if (unlikely(!page)) {
1726                         release_pte_pages(pte, _pte);
1727                         goto out;
1728                 }
1729                 VM_BUG_ON(PageCompound(page));
1730                 BUG_ON(!PageAnon(page));
1731                 VM_BUG_ON(!PageSwapBacked(page));
1732
1733                 /* cannot use mapcount: can't collapse if there's a gup pin */
1734                 if (page_count(page) != 1) {
1735                         release_pte_pages(pte, _pte);
1736                         goto out;
1737                 }
1738                 /*
1739                  * We can do it before isolate_lru_page because the
1740                  * page can't be freed from under us. NOTE: PG_lock
1741                  * is needed to serialize against split_huge_page
1742                  * when invoked from the VM.
1743                  */
1744                 if (!trylock_page(page)) {
1745                         release_pte_pages(pte, _pte);
1746                         goto out;
1747                 }
1748                 /*
1749                  * Isolate the page to avoid collapsing an hugepage
1750                  * currently in use by the VM.
1751                  */
1752                 if (isolate_lru_page(page)) {
1753                         unlock_page(page);
1754                         release_pte_pages(pte, _pte);
1755                         goto out;
1756                 }
1757                 /* 0 stands for page_is_file_cache(page) == false */
1758                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1759                 VM_BUG_ON(!PageLocked(page));
1760                 VM_BUG_ON(PageLRU(page));
1761
1762                 /* If there is no mapped pte young don't collapse the page */
1763                 if (pte_young(pteval) || PageReferenced(page) ||
1764                     mmu_notifier_test_young(vma->vm_mm, address))
1765                         referenced = 1;
1766         }
1767         if (unlikely(!referenced))
1768                 release_all_pte_pages(pte);
1769         else
1770                 isolated = 1;
1771 out:
1772         return isolated;
1773 }
1774
1775 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1776                                       struct vm_area_struct *vma,
1777                                       unsigned long address,
1778                                       spinlock_t *ptl)
1779 {
1780         pte_t *_pte;
1781         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1782                 pte_t pteval = *_pte;
1783                 struct page *src_page;
1784
1785                 if (pte_none(pteval)) {
1786                         clear_user_highpage(page, address);
1787                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1788                 } else {
1789                         src_page = pte_page(pteval);
1790                         copy_user_highpage(page, src_page, address, vma);
1791                         VM_BUG_ON(page_mapcount(src_page) != 1);
1792                         VM_BUG_ON(page_count(src_page) != 2);
1793                         release_pte_page(src_page);
1794                         /*
1795                          * ptl mostly unnecessary, but preempt has to
1796                          * be disabled to update the per-cpu stats
1797                          * inside page_remove_rmap().
1798                          */
1799                         spin_lock(ptl);
1800                         /*
1801                          * paravirt calls inside pte_clear here are
1802                          * superfluous.
1803                          */
1804                         pte_clear(vma->vm_mm, address, _pte);
1805                         page_remove_rmap(src_page);
1806                         spin_unlock(ptl);
1807                         free_page_and_swap_cache(src_page);
1808                 }
1809
1810                 address += PAGE_SIZE;
1811                 page++;
1812         }
1813 }
1814
1815 static void collapse_huge_page(struct mm_struct *mm,
1816                                unsigned long address,
1817                                struct page **hpage,
1818                                struct vm_area_struct *vma,
1819                                int node)
1820 {
1821         pgd_t *pgd;
1822         pud_t *pud;
1823         pmd_t *pmd, _pmd;
1824         pte_t *pte;
1825         pgtable_t pgtable;
1826         struct page *new_page;
1827         spinlock_t *ptl;
1828         int isolated;
1829         unsigned long hstart, hend;
1830
1831         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1832 #ifndef CONFIG_NUMA
1833         up_read(&mm->mmap_sem);
1834         VM_BUG_ON(!*hpage);
1835         new_page = *hpage;
1836 #else
1837         VM_BUG_ON(*hpage);
1838         /*
1839          * Allocate the page while the vma is still valid and under
1840          * the mmap_sem read mode so there is no memory allocation
1841          * later when we take the mmap_sem in write mode. This is more
1842          * friendly behavior (OTOH it may actually hide bugs) to
1843          * filesystems in userland with daemons allocating memory in
1844          * the userland I/O paths.  Allocating memory with the
1845          * mmap_sem in read mode is good idea also to allow greater
1846          * scalability.
1847          */
1848         new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1849                                       node, __GFP_OTHER_NODE);
1850
1851         /*
1852          * After allocating the hugepage, release the mmap_sem read lock in
1853          * preparation for taking it in write mode.
1854          */
1855         up_read(&mm->mmap_sem);
1856         if (unlikely(!new_page)) {
1857                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1858                 *hpage = ERR_PTR(-ENOMEM);
1859                 return;
1860         }
1861 #endif
1862
1863         count_vm_event(THP_COLLAPSE_ALLOC);
1864         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1865 #ifdef CONFIG_NUMA
1866                 put_page(new_page);
1867 #endif
1868                 return;
1869         }
1870
1871         /*
1872          * Prevent all access to pagetables with the exception of
1873          * gup_fast later hanlded by the ptep_clear_flush and the VM
1874          * handled by the anon_vma lock + PG_lock.
1875          */
1876         down_write(&mm->mmap_sem);
1877         if (unlikely(khugepaged_test_exit(mm)))
1878                 goto out;
1879
1880         vma = find_vma(mm, address);
1881         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1882         hend = vma->vm_end & HPAGE_PMD_MASK;
1883         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1884                 goto out;
1885
1886         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1887             (vma->vm_flags & VM_NOHUGEPAGE))
1888                 goto out;
1889
1890         if (!vma->anon_vma || vma->vm_ops)
1891                 goto out;
1892         if (is_vma_temporary_stack(vma))
1893                 goto out;
1894         /*
1895          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1896          * true too, verify it here.
1897          */
1898         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1899
1900         pgd = pgd_offset(mm, address);
1901         if (!pgd_present(*pgd))
1902                 goto out;
1903
1904         pud = pud_offset(pgd, address);
1905         if (!pud_present(*pud))
1906                 goto out;
1907
1908         pmd = pmd_offset(pud, address);
1909         /* pmd can't go away or become huge under us */
1910         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1911                 goto out;
1912
1913         anon_vma_lock(vma->anon_vma);
1914
1915         pte = pte_offset_map(pmd, address);
1916         ptl = pte_lockptr(mm, pmd);
1917
1918         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1919         /*
1920          * After this gup_fast can't run anymore. This also removes
1921          * any huge TLB entry from the CPU so we won't allow
1922          * huge and small TLB entries for the same virtual address
1923          * to avoid the risk of CPU bugs in that area.
1924          */
1925         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1926         spin_unlock(&mm->page_table_lock);
1927
1928         spin_lock(ptl);
1929         isolated = __collapse_huge_page_isolate(vma, address, pte);
1930         spin_unlock(ptl);
1931
1932         if (unlikely(!isolated)) {
1933                 pte_unmap(pte);
1934                 spin_lock(&mm->page_table_lock);
1935                 BUG_ON(!pmd_none(*pmd));
1936                 set_pmd_at(mm, address, pmd, _pmd);
1937                 spin_unlock(&mm->page_table_lock);
1938                 anon_vma_unlock(vma->anon_vma);
1939                 goto out;
1940         }
1941
1942         /*
1943          * All pages are isolated and locked so anon_vma rmap
1944          * can't run anymore.
1945          */
1946         anon_vma_unlock(vma->anon_vma);
1947
1948         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1949         pte_unmap(pte);
1950         __SetPageUptodate(new_page);
1951         pgtable = pmd_pgtable(_pmd);
1952         VM_BUG_ON(page_count(pgtable) != 1);
1953         VM_BUG_ON(page_mapcount(pgtable) != 0);
1954
1955         _pmd = mk_pmd(new_page, vma->vm_page_prot);
1956         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1957         _pmd = pmd_mkhuge(_pmd);
1958
1959         /*
1960          * spin_lock() below is not the equivalent of smp_wmb(), so
1961          * this is needed to avoid the copy_huge_page writes to become
1962          * visible after the set_pmd_at() write.
1963          */
1964         smp_wmb();
1965
1966         spin_lock(&mm->page_table_lock);
1967         BUG_ON(!pmd_none(*pmd));
1968         page_add_new_anon_rmap(new_page, vma, address);
1969         set_pmd_at(mm, address, pmd, _pmd);
1970         update_mmu_cache(vma, address, _pmd);
1971         prepare_pmd_huge_pte(pgtable, mm);
1972         mm->nr_ptes--;
1973         spin_unlock(&mm->page_table_lock);
1974
1975 #ifndef CONFIG_NUMA
1976         *hpage = NULL;
1977 #endif
1978         khugepaged_pages_collapsed++;
1979 out_up_write:
1980         up_write(&mm->mmap_sem);
1981         return;
1982
1983 out:
1984         mem_cgroup_uncharge_page(new_page);
1985 #ifdef CONFIG_NUMA
1986         put_page(new_page);
1987 #endif
1988         goto out_up_write;
1989 }
1990
1991 static int khugepaged_scan_pmd(struct mm_struct *mm,
1992                                struct vm_area_struct *vma,
1993                                unsigned long address,
1994                                struct page **hpage)
1995 {
1996         pgd_t *pgd;
1997         pud_t *pud;
1998         pmd_t *pmd;
1999         pte_t *pte, *_pte;
2000         int ret = 0, referenced = 0, none = 0;
2001         struct page *page;
2002         unsigned long _address;
2003         spinlock_t *ptl;
2004         int node = -1;
2005
2006         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2007
2008         pgd = pgd_offset(mm, address);
2009         if (!pgd_present(*pgd))
2010                 goto out;
2011
2012         pud = pud_offset(pgd, address);
2013         if (!pud_present(*pud))
2014                 goto out;
2015
2016         pmd = pmd_offset(pud, address);
2017         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2018                 goto out;
2019
2020         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2021         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2022              _pte++, _address += PAGE_SIZE) {
2023                 pte_t pteval = *_pte;
2024                 if (pte_none(pteval)) {
2025                         if (++none <= khugepaged_max_ptes_none)
2026                                 continue;
2027                         else
2028                                 goto out_unmap;
2029                 }
2030                 if (!pte_present(pteval) || !pte_write(pteval))
2031                         goto out_unmap;
2032                 page = vm_normal_page(vma, _address, pteval);
2033                 if (unlikely(!page))
2034                         goto out_unmap;
2035                 /*
2036                  * Chose the node of the first page. This could
2037                  * be more sophisticated and look at more pages,
2038                  * but isn't for now.
2039                  */
2040                 if (node == -1)
2041                         node = page_to_nid(page);
2042                 VM_BUG_ON(PageCompound(page));
2043                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2044                         goto out_unmap;
2045                 /* cannot use mapcount: can't collapse if there's a gup pin */
2046                 if (page_count(page) != 1)
2047                         goto out_unmap;
2048                 if (pte_young(pteval) || PageReferenced(page) ||
2049                     mmu_notifier_test_young(vma->vm_mm, address))
2050                         referenced = 1;
2051         }
2052         if (referenced)
2053                 ret = 1;
2054 out_unmap:
2055         pte_unmap_unlock(pte, ptl);
2056         if (ret)
2057                 /* collapse_huge_page will return with the mmap_sem released */
2058                 collapse_huge_page(mm, address, hpage, vma, node);
2059 out:
2060         return ret;
2061 }
2062
2063 static void collect_mm_slot(struct mm_slot *mm_slot)
2064 {
2065         struct mm_struct *mm = mm_slot->mm;
2066
2067         VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
2068
2069         if (khugepaged_test_exit(mm)) {
2070                 /* free mm_slot */
2071                 hlist_del(&mm_slot->hash);
2072                 list_del(&mm_slot->mm_node);
2073
2074                 /*
2075                  * Not strictly needed because the mm exited already.
2076                  *
2077                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2078                  */
2079
2080                 /* khugepaged_mm_lock actually not necessary for the below */
2081                 free_mm_slot(mm_slot);
2082                 mmdrop(mm);
2083         }
2084 }
2085
2086 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2087                                             struct page **hpage)
2088         __releases(&khugepaged_mm_lock)
2089         __acquires(&khugepaged_mm_lock)
2090 {
2091         struct mm_slot *mm_slot;
2092         struct mm_struct *mm;
2093         struct vm_area_struct *vma;
2094         int progress = 0;
2095
2096         VM_BUG_ON(!pages);
2097         VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
2098
2099         if (khugepaged_scan.mm_slot)
2100                 mm_slot = khugepaged_scan.mm_slot;
2101         else {
2102                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2103                                      struct mm_slot, mm_node);
2104                 khugepaged_scan.address = 0;
2105                 khugepaged_scan.mm_slot = mm_slot;
2106         }
2107         spin_unlock(&khugepaged_mm_lock);
2108
2109         mm = mm_slot->mm;
2110         down_read(&mm->mmap_sem);
2111         if (unlikely(khugepaged_test_exit(mm)))
2112                 vma = NULL;
2113         else
2114                 vma = find_vma(mm, khugepaged_scan.address);
2115
2116         progress++;
2117         for (; vma; vma = vma->vm_next) {
2118                 unsigned long hstart, hend;
2119
2120                 cond_resched();
2121                 if (unlikely(khugepaged_test_exit(mm))) {
2122                         progress++;
2123                         break;
2124                 }
2125
2126                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2127                      !khugepaged_always()) ||
2128                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2129                 skip:
2130                         progress++;
2131                         continue;
2132                 }
2133                 if (!vma->anon_vma || vma->vm_ops)
2134                         goto skip;
2135                 if (is_vma_temporary_stack(vma))
2136                         goto skip;
2137                 /*
2138                  * If is_pfn_mapping() is true is_learn_pfn_mapping()
2139                  * must be true too, verify it here.
2140                  */
2141                 VM_BUG_ON(is_linear_pfn_mapping(vma) ||
2142                           vma->vm_flags & VM_NO_THP);
2143
2144                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2145                 hend = vma->vm_end & HPAGE_PMD_MASK;
2146                 if (hstart >= hend)
2147                         goto skip;
2148                 if (khugepaged_scan.address > hend)
2149                         goto skip;
2150                 if (khugepaged_scan.address < hstart)
2151                         khugepaged_scan.address = hstart;
2152                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2153
2154                 while (khugepaged_scan.address < hend) {
2155                         int ret;
2156                         cond_resched();
2157                         if (unlikely(khugepaged_test_exit(mm)))
2158                                 goto breakouterloop;
2159
2160                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2161                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2162                                   hend);
2163                         ret = khugepaged_scan_pmd(mm, vma,
2164                                                   khugepaged_scan.address,
2165                                                   hpage);
2166                         /* move to next address */
2167                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2168                         progress += HPAGE_PMD_NR;
2169                         if (ret)
2170                                 /* we released mmap_sem so break loop */
2171                                 goto breakouterloop_mmap_sem;
2172                         if (progress >= pages)
2173                                 goto breakouterloop;
2174                 }
2175         }
2176 breakouterloop:
2177         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2178 breakouterloop_mmap_sem:
2179
2180         spin_lock(&khugepaged_mm_lock);
2181         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2182         /*
2183          * Release the current mm_slot if this mm is about to die, or
2184          * if we scanned all vmas of this mm.
2185          */
2186         if (khugepaged_test_exit(mm) || !vma) {
2187                 /*
2188                  * Make sure that if mm_users is reaching zero while
2189                  * khugepaged runs here, khugepaged_exit will find
2190                  * mm_slot not pointing to the exiting mm.
2191                  */
2192                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2193                         khugepaged_scan.mm_slot = list_entry(
2194                                 mm_slot->mm_node.next,
2195                                 struct mm_slot, mm_node);
2196                         khugepaged_scan.address = 0;
2197                 } else {
2198                         khugepaged_scan.mm_slot = NULL;
2199                         khugepaged_full_scans++;
2200                 }
2201
2202                 collect_mm_slot(mm_slot);
2203         }
2204
2205         return progress;
2206 }
2207
2208 static int khugepaged_has_work(void)
2209 {
2210         return !list_empty(&khugepaged_scan.mm_head) &&
2211                 khugepaged_enabled();
2212 }
2213
2214 static int khugepaged_wait_event(void)
2215 {
2216         return !list_empty(&khugepaged_scan.mm_head) ||
2217                 !khugepaged_enabled();
2218 }
2219
2220 static void khugepaged_do_scan(struct page **hpage)
2221 {
2222         unsigned int progress = 0, pass_through_head = 0;
2223         unsigned int pages = khugepaged_pages_to_scan;
2224
2225         barrier(); /* write khugepaged_pages_to_scan to local stack */
2226
2227         while (progress < pages) {
2228                 cond_resched();
2229
2230 #ifndef CONFIG_NUMA
2231                 if (!*hpage) {
2232                         *hpage = alloc_hugepage(khugepaged_defrag());
2233                         if (unlikely(!*hpage)) {
2234                                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2235                                 break;
2236                         }
2237                         count_vm_event(THP_COLLAPSE_ALLOC);
2238                 }
2239 #else
2240                 if (IS_ERR(*hpage))
2241                         break;
2242 #endif
2243
2244                 if (unlikely(kthread_should_stop() || freezing(current)))
2245                         break;
2246
2247                 spin_lock(&khugepaged_mm_lock);
2248                 if (!khugepaged_scan.mm_slot)
2249                         pass_through_head++;
2250                 if (khugepaged_has_work() &&
2251                     pass_through_head < 2)
2252                         progress += khugepaged_scan_mm_slot(pages - progress,
2253                                                             hpage);
2254                 else
2255                         progress = pages;
2256                 spin_unlock(&khugepaged_mm_lock);
2257         }
2258 }
2259
2260 static void khugepaged_alloc_sleep(void)
2261 {
2262         wait_event_freezable_timeout(khugepaged_wait, false,
2263                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2264 }
2265
2266 #ifndef CONFIG_NUMA
2267 static struct page *khugepaged_alloc_hugepage(void)
2268 {
2269         struct page *hpage;
2270
2271         do {
2272                 hpage = alloc_hugepage(khugepaged_defrag());
2273                 if (!hpage) {
2274                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2275                         khugepaged_alloc_sleep();
2276                 } else
2277                         count_vm_event(THP_COLLAPSE_ALLOC);
2278         } while (unlikely(!hpage) &&
2279                  likely(khugepaged_enabled()));
2280         return hpage;
2281 }
2282 #endif
2283
2284 static void khugepaged_loop(void)
2285 {
2286         struct page *hpage;
2287
2288 #ifdef CONFIG_NUMA
2289         hpage = NULL;
2290 #endif
2291         while (likely(khugepaged_enabled())) {
2292 #ifndef CONFIG_NUMA
2293                 hpage = khugepaged_alloc_hugepage();
2294                 if (unlikely(!hpage))
2295                         break;
2296 #else
2297                 if (IS_ERR(hpage)) {
2298                         khugepaged_alloc_sleep();
2299                         hpage = NULL;
2300                 }
2301 #endif
2302
2303                 khugepaged_do_scan(&hpage);
2304 #ifndef CONFIG_NUMA
2305                 if (hpage)
2306                         put_page(hpage);
2307 #endif
2308                 try_to_freeze();
2309                 if (unlikely(kthread_should_stop()))
2310                         break;
2311                 if (khugepaged_has_work()) {
2312                         if (!khugepaged_scan_sleep_millisecs)
2313                                 continue;
2314                         wait_event_freezable_timeout(khugepaged_wait, false,
2315                             msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2316                 } else if (khugepaged_enabled())
2317                         wait_event_freezable(khugepaged_wait,
2318                                              khugepaged_wait_event());
2319         }
2320 }
2321
2322 static int khugepaged(void *none)
2323 {
2324         struct mm_slot *mm_slot;
2325
2326         set_freezable();
2327         set_user_nice(current, 19);
2328
2329         /* serialize with start_khugepaged() */
2330         mutex_lock(&khugepaged_mutex);
2331
2332         for (;;) {
2333                 mutex_unlock(&khugepaged_mutex);
2334                 VM_BUG_ON(khugepaged_thread != current);
2335                 khugepaged_loop();
2336                 VM_BUG_ON(khugepaged_thread != current);
2337
2338                 mutex_lock(&khugepaged_mutex);
2339                 if (!khugepaged_enabled())
2340                         break;
2341                 if (unlikely(kthread_should_stop()))
2342                         break;
2343         }
2344
2345         spin_lock(&khugepaged_mm_lock);
2346         mm_slot = khugepaged_scan.mm_slot;
2347         khugepaged_scan.mm_slot = NULL;
2348         if (mm_slot)
2349                 collect_mm_slot(mm_slot);
2350         spin_unlock(&khugepaged_mm_lock);
2351
2352         khugepaged_thread = NULL;
2353         mutex_unlock(&khugepaged_mutex);
2354
2355         return 0;
2356 }
2357
2358 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2359 {
2360         struct page *page;
2361
2362         spin_lock(&mm->page_table_lock);
2363         if (unlikely(!pmd_trans_huge(*pmd))) {
2364                 spin_unlock(&mm->page_table_lock);
2365                 return;
2366         }
2367         page = pmd_page(*pmd);
2368         VM_BUG_ON(!page_count(page));
2369         get_page(page);
2370         spin_unlock(&mm->page_table_lock);
2371
2372         split_huge_page(page);
2373
2374         put_page(page);
2375         BUG_ON(pmd_trans_huge(*pmd));
2376 }
2377
2378 static void split_huge_page_address(struct mm_struct *mm,
2379                                     unsigned long address)
2380 {
2381         pgd_t *pgd;
2382         pud_t *pud;
2383         pmd_t *pmd;
2384
2385         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2386
2387         pgd = pgd_offset(mm, address);
2388         if (!pgd_present(*pgd))
2389                 return;
2390
2391         pud = pud_offset(pgd, address);
2392         if (!pud_present(*pud))
2393                 return;
2394
2395         pmd = pmd_offset(pud, address);
2396         if (!pmd_present(*pmd))
2397                 return;
2398         /*
2399          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2400          * materialize from under us.
2401          */
2402         split_huge_page_pmd(mm, pmd);
2403 }
2404
2405 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2406                              unsigned long start,
2407                              unsigned long end,
2408                              long adjust_next)
2409 {
2410         /*
2411          * If the new start address isn't hpage aligned and it could
2412          * previously contain an hugepage: check if we need to split
2413          * an huge pmd.
2414          */
2415         if (start & ~HPAGE_PMD_MASK &&
2416             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2417             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2418                 split_huge_page_address(vma->vm_mm, start);
2419
2420         /*
2421          * If the new end address isn't hpage aligned and it could
2422          * previously contain an hugepage: check if we need to split
2423          * an huge pmd.
2424          */
2425         if (end & ~HPAGE_PMD_MASK &&
2426             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2427             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2428                 split_huge_page_address(vma->vm_mm, end);
2429
2430         /*
2431          * If we're also updating the vma->vm_next->vm_start, if the new
2432          * vm_next->vm_start isn't page aligned and it could previously
2433          * contain an hugepage: check if we need to split an huge pmd.
2434          */
2435         if (adjust_next > 0) {
2436                 struct vm_area_struct *next = vma->vm_next;
2437                 unsigned long nstart = next->vm_start;
2438                 nstart += adjust_next << PAGE_SHIFT;
2439                 if (nstart & ~HPAGE_PMD_MASK &&
2440                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2441                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2442                         split_huge_page_address(next->vm_mm, nstart);
2443         }
2444 }