memcg: fix leak of accounting at failure path of hugepage collapsing
[linux-2.6.git] / mm / huge_memory.c
1 /*
2  *  Copyright (C) 2009  Red Hat, Inc.
3  *
4  *  This work is licensed under the terms of the GNU GPL, version 2. See
5  *  the COPYING file in the top-level directory.
6  */
7
8 #include <linux/mm.h>
9 #include <linux/sched.h>
10 #include <linux/highmem.h>
11 #include <linux/hugetlb.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/rmap.h>
14 #include <linux/swap.h>
15 #include <linux/mm_inline.h>
16 #include <linux/kthread.h>
17 #include <linux/khugepaged.h>
18 #include <linux/freezer.h>
19 #include <linux/mman.h>
20 #include <asm/tlb.h>
21 #include <asm/pgalloc.h>
22 #include "internal.h"
23
24 /*
25  * By default transparent hugepage support is enabled for all mappings
26  * and khugepaged scans all mappings. Defrag is only invoked by
27  * khugepaged hugepage allocations and by page faults inside
28  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
29  * allocations.
30  */
31 unsigned long transparent_hugepage_flags __read_mostly =
32 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
33         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
34 #endif
35 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
36         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
37 #endif
38         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
39         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
40
41 /* default scan 8*512 pte (or vmas) every 30 second */
42 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
43 static unsigned int khugepaged_pages_collapsed;
44 static unsigned int khugepaged_full_scans;
45 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
46 /* during fragmentation poll the hugepage allocator once every minute */
47 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
48 static struct task_struct *khugepaged_thread __read_mostly;
49 static DEFINE_MUTEX(khugepaged_mutex);
50 static DEFINE_SPINLOCK(khugepaged_mm_lock);
51 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
52 /*
53  * default collapse hugepages if there is at least one pte mapped like
54  * it would have happened if the vma was large enough during page
55  * fault.
56  */
57 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
58
59 static int khugepaged(void *none);
60 static int mm_slots_hash_init(void);
61 static int khugepaged_slab_init(void);
62 static void khugepaged_slab_free(void);
63
64 #define MM_SLOTS_HASH_HEADS 1024
65 static struct hlist_head *mm_slots_hash __read_mostly;
66 static struct kmem_cache *mm_slot_cache __read_mostly;
67
68 /**
69  * struct mm_slot - hash lookup from mm to mm_slot
70  * @hash: hash collision list
71  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
72  * @mm: the mm that this information is valid for
73  */
74 struct mm_slot {
75         struct hlist_node hash;
76         struct list_head mm_node;
77         struct mm_struct *mm;
78 };
79
80 /**
81  * struct khugepaged_scan - cursor for scanning
82  * @mm_head: the head of the mm list to scan
83  * @mm_slot: the current mm_slot we are scanning
84  * @address: the next address inside that to be scanned
85  *
86  * There is only the one khugepaged_scan instance of this cursor structure.
87  */
88 struct khugepaged_scan {
89         struct list_head mm_head;
90         struct mm_slot *mm_slot;
91         unsigned long address;
92 } khugepaged_scan = {
93         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
94 };
95
96
97 static int set_recommended_min_free_kbytes(void)
98 {
99         struct zone *zone;
100         int nr_zones = 0;
101         unsigned long recommended_min;
102         extern int min_free_kbytes;
103
104         if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG,
105                       &transparent_hugepage_flags) &&
106             !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
107                       &transparent_hugepage_flags))
108                 return 0;
109
110         for_each_populated_zone(zone)
111                 nr_zones++;
112
113         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
114         recommended_min = pageblock_nr_pages * nr_zones * 2;
115
116         /*
117          * Make sure that on average at least two pageblocks are almost free
118          * of another type, one for a migratetype to fall back to and a
119          * second to avoid subsequent fallbacks of other types There are 3
120          * MIGRATE_TYPES we care about.
121          */
122         recommended_min += pageblock_nr_pages * nr_zones *
123                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
124
125         /* don't ever allow to reserve more than 5% of the lowmem */
126         recommended_min = min(recommended_min,
127                               (unsigned long) nr_free_buffer_pages() / 20);
128         recommended_min <<= (PAGE_SHIFT-10);
129
130         if (recommended_min > min_free_kbytes)
131                 min_free_kbytes = recommended_min;
132         setup_per_zone_wmarks();
133         return 0;
134 }
135 late_initcall(set_recommended_min_free_kbytes);
136
137 static int start_khugepaged(void)
138 {
139         int err = 0;
140         if (khugepaged_enabled()) {
141                 int wakeup;
142                 if (unlikely(!mm_slot_cache || !mm_slots_hash)) {
143                         err = -ENOMEM;
144                         goto out;
145                 }
146                 mutex_lock(&khugepaged_mutex);
147                 if (!khugepaged_thread)
148                         khugepaged_thread = kthread_run(khugepaged, NULL,
149                                                         "khugepaged");
150                 if (unlikely(IS_ERR(khugepaged_thread))) {
151                         printk(KERN_ERR
152                                "khugepaged: kthread_run(khugepaged) failed\n");
153                         err = PTR_ERR(khugepaged_thread);
154                         khugepaged_thread = NULL;
155                 }
156                 wakeup = !list_empty(&khugepaged_scan.mm_head);
157                 mutex_unlock(&khugepaged_mutex);
158                 if (wakeup)
159                         wake_up_interruptible(&khugepaged_wait);
160
161                 set_recommended_min_free_kbytes();
162         } else
163                 /* wakeup to exit */
164                 wake_up_interruptible(&khugepaged_wait);
165 out:
166         return err;
167 }
168
169 #ifdef CONFIG_SYSFS
170
171 static ssize_t double_flag_show(struct kobject *kobj,
172                                 struct kobj_attribute *attr, char *buf,
173                                 enum transparent_hugepage_flag enabled,
174                                 enum transparent_hugepage_flag req_madv)
175 {
176         if (test_bit(enabled, &transparent_hugepage_flags)) {
177                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
178                 return sprintf(buf, "[always] madvise never\n");
179         } else if (test_bit(req_madv, &transparent_hugepage_flags))
180                 return sprintf(buf, "always [madvise] never\n");
181         else
182                 return sprintf(buf, "always madvise [never]\n");
183 }
184 static ssize_t double_flag_store(struct kobject *kobj,
185                                  struct kobj_attribute *attr,
186                                  const char *buf, size_t count,
187                                  enum transparent_hugepage_flag enabled,
188                                  enum transparent_hugepage_flag req_madv)
189 {
190         if (!memcmp("always", buf,
191                     min(sizeof("always")-1, count))) {
192                 set_bit(enabled, &transparent_hugepage_flags);
193                 clear_bit(req_madv, &transparent_hugepage_flags);
194         } else if (!memcmp("madvise", buf,
195                            min(sizeof("madvise")-1, count))) {
196                 clear_bit(enabled, &transparent_hugepage_flags);
197                 set_bit(req_madv, &transparent_hugepage_flags);
198         } else if (!memcmp("never", buf,
199                            min(sizeof("never")-1, count))) {
200                 clear_bit(enabled, &transparent_hugepage_flags);
201                 clear_bit(req_madv, &transparent_hugepage_flags);
202         } else
203                 return -EINVAL;
204
205         return count;
206 }
207
208 static ssize_t enabled_show(struct kobject *kobj,
209                             struct kobj_attribute *attr, char *buf)
210 {
211         return double_flag_show(kobj, attr, buf,
212                                 TRANSPARENT_HUGEPAGE_FLAG,
213                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
214 }
215 static ssize_t enabled_store(struct kobject *kobj,
216                              struct kobj_attribute *attr,
217                              const char *buf, size_t count)
218 {
219         ssize_t ret;
220
221         ret = double_flag_store(kobj, attr, buf, count,
222                                 TRANSPARENT_HUGEPAGE_FLAG,
223                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
224
225         if (ret > 0) {
226                 int err = start_khugepaged();
227                 if (err)
228                         ret = err;
229         }
230
231         if (ret > 0 &&
232             (test_bit(TRANSPARENT_HUGEPAGE_FLAG,
233                       &transparent_hugepage_flags) ||
234              test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
235                       &transparent_hugepage_flags)))
236                 set_recommended_min_free_kbytes();
237
238         return ret;
239 }
240 static struct kobj_attribute enabled_attr =
241         __ATTR(enabled, 0644, enabled_show, enabled_store);
242
243 static ssize_t single_flag_show(struct kobject *kobj,
244                                 struct kobj_attribute *attr, char *buf,
245                                 enum transparent_hugepage_flag flag)
246 {
247         if (test_bit(flag, &transparent_hugepage_flags))
248                 return sprintf(buf, "[yes] no\n");
249         else
250                 return sprintf(buf, "yes [no]\n");
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         if (!memcmp("yes", buf,
258                     min(sizeof("yes")-1, count))) {
259                 set_bit(flag, &transparent_hugepage_flags);
260         } else if (!memcmp("no", buf,
261                            min(sizeof("no")-1, count))) {
262                 clear_bit(flag, &transparent_hugepage_flags);
263         } else
264                 return -EINVAL;
265
266         return count;
267 }
268
269 /*
270  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
271  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
272  * memory just to allocate one more hugepage.
273  */
274 static ssize_t defrag_show(struct kobject *kobj,
275                            struct kobj_attribute *attr, char *buf)
276 {
277         return double_flag_show(kobj, attr, buf,
278                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
279                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
280 }
281 static ssize_t defrag_store(struct kobject *kobj,
282                             struct kobj_attribute *attr,
283                             const char *buf, size_t count)
284 {
285         return double_flag_store(kobj, attr, buf, count,
286                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
287                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
288 }
289 static struct kobj_attribute defrag_attr =
290         __ATTR(defrag, 0644, defrag_show, defrag_store);
291
292 #ifdef CONFIG_DEBUG_VM
293 static ssize_t debug_cow_show(struct kobject *kobj,
294                                 struct kobj_attribute *attr, char *buf)
295 {
296         return single_flag_show(kobj, attr, buf,
297                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
298 }
299 static ssize_t debug_cow_store(struct kobject *kobj,
300                                struct kobj_attribute *attr,
301                                const char *buf, size_t count)
302 {
303         return single_flag_store(kobj, attr, buf, count,
304                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
305 }
306 static struct kobj_attribute debug_cow_attr =
307         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
308 #endif /* CONFIG_DEBUG_VM */
309
310 static struct attribute *hugepage_attr[] = {
311         &enabled_attr.attr,
312         &defrag_attr.attr,
313 #ifdef CONFIG_DEBUG_VM
314         &debug_cow_attr.attr,
315 #endif
316         NULL,
317 };
318
319 static struct attribute_group hugepage_attr_group = {
320         .attrs = hugepage_attr,
321 };
322
323 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
324                                          struct kobj_attribute *attr,
325                                          char *buf)
326 {
327         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
328 }
329
330 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
331                                           struct kobj_attribute *attr,
332                                           const char *buf, size_t count)
333 {
334         unsigned long msecs;
335         int err;
336
337         err = strict_strtoul(buf, 10, &msecs);
338         if (err || msecs > UINT_MAX)
339                 return -EINVAL;
340
341         khugepaged_scan_sleep_millisecs = msecs;
342         wake_up_interruptible(&khugepaged_wait);
343
344         return count;
345 }
346 static struct kobj_attribute scan_sleep_millisecs_attr =
347         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
348                scan_sleep_millisecs_store);
349
350 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
351                                           struct kobj_attribute *attr,
352                                           char *buf)
353 {
354         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
355 }
356
357 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
358                                            struct kobj_attribute *attr,
359                                            const char *buf, size_t count)
360 {
361         unsigned long msecs;
362         int err;
363
364         err = strict_strtoul(buf, 10, &msecs);
365         if (err || msecs > UINT_MAX)
366                 return -EINVAL;
367
368         khugepaged_alloc_sleep_millisecs = msecs;
369         wake_up_interruptible(&khugepaged_wait);
370
371         return count;
372 }
373 static struct kobj_attribute alloc_sleep_millisecs_attr =
374         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
375                alloc_sleep_millisecs_store);
376
377 static ssize_t pages_to_scan_show(struct kobject *kobj,
378                                   struct kobj_attribute *attr,
379                                   char *buf)
380 {
381         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
382 }
383 static ssize_t pages_to_scan_store(struct kobject *kobj,
384                                    struct kobj_attribute *attr,
385                                    const char *buf, size_t count)
386 {
387         int err;
388         unsigned long pages;
389
390         err = strict_strtoul(buf, 10, &pages);
391         if (err || !pages || pages > UINT_MAX)
392                 return -EINVAL;
393
394         khugepaged_pages_to_scan = pages;
395
396         return count;
397 }
398 static struct kobj_attribute pages_to_scan_attr =
399         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
400                pages_to_scan_store);
401
402 static ssize_t pages_collapsed_show(struct kobject *kobj,
403                                     struct kobj_attribute *attr,
404                                     char *buf)
405 {
406         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
407 }
408 static struct kobj_attribute pages_collapsed_attr =
409         __ATTR_RO(pages_collapsed);
410
411 static ssize_t full_scans_show(struct kobject *kobj,
412                                struct kobj_attribute *attr,
413                                char *buf)
414 {
415         return sprintf(buf, "%u\n", khugepaged_full_scans);
416 }
417 static struct kobj_attribute full_scans_attr =
418         __ATTR_RO(full_scans);
419
420 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
421                                       struct kobj_attribute *attr, char *buf)
422 {
423         return single_flag_show(kobj, attr, buf,
424                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
425 }
426 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
427                                        struct kobj_attribute *attr,
428                                        const char *buf, size_t count)
429 {
430         return single_flag_store(kobj, attr, buf, count,
431                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
432 }
433 static struct kobj_attribute khugepaged_defrag_attr =
434         __ATTR(defrag, 0644, khugepaged_defrag_show,
435                khugepaged_defrag_store);
436
437 /*
438  * max_ptes_none controls if khugepaged should collapse hugepages over
439  * any unmapped ptes in turn potentially increasing the memory
440  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
441  * reduce the available free memory in the system as it
442  * runs. Increasing max_ptes_none will instead potentially reduce the
443  * free memory in the system during the khugepaged scan.
444  */
445 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
446                                              struct kobj_attribute *attr,
447                                              char *buf)
448 {
449         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
450 }
451 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
452                                               struct kobj_attribute *attr,
453                                               const char *buf, size_t count)
454 {
455         int err;
456         unsigned long max_ptes_none;
457
458         err = strict_strtoul(buf, 10, &max_ptes_none);
459         if (err || max_ptes_none > HPAGE_PMD_NR-1)
460                 return -EINVAL;
461
462         khugepaged_max_ptes_none = max_ptes_none;
463
464         return count;
465 }
466 static struct kobj_attribute khugepaged_max_ptes_none_attr =
467         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
468                khugepaged_max_ptes_none_store);
469
470 static struct attribute *khugepaged_attr[] = {
471         &khugepaged_defrag_attr.attr,
472         &khugepaged_max_ptes_none_attr.attr,
473         &pages_to_scan_attr.attr,
474         &pages_collapsed_attr.attr,
475         &full_scans_attr.attr,
476         &scan_sleep_millisecs_attr.attr,
477         &alloc_sleep_millisecs_attr.attr,
478         NULL,
479 };
480
481 static struct attribute_group khugepaged_attr_group = {
482         .attrs = khugepaged_attr,
483         .name = "khugepaged",
484 };
485 #endif /* CONFIG_SYSFS */
486
487 static int __init hugepage_init(void)
488 {
489         int err;
490 #ifdef CONFIG_SYSFS
491         static struct kobject *hugepage_kobj;
492 #endif
493
494         err = -EINVAL;
495         if (!has_transparent_hugepage()) {
496                 transparent_hugepage_flags = 0;
497                 goto out;
498         }
499
500 #ifdef CONFIG_SYSFS
501         err = -ENOMEM;
502         hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
503         if (unlikely(!hugepage_kobj)) {
504                 printk(KERN_ERR "hugepage: failed kobject create\n");
505                 goto out;
506         }
507
508         err = sysfs_create_group(hugepage_kobj, &hugepage_attr_group);
509         if (err) {
510                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
511                 goto out;
512         }
513
514         err = sysfs_create_group(hugepage_kobj, &khugepaged_attr_group);
515         if (err) {
516                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
517                 goto out;
518         }
519 #endif
520
521         err = khugepaged_slab_init();
522         if (err)
523                 goto out;
524
525         err = mm_slots_hash_init();
526         if (err) {
527                 khugepaged_slab_free();
528                 goto out;
529         }
530
531         /*
532          * By default disable transparent hugepages on smaller systems,
533          * where the extra memory used could hurt more than TLB overhead
534          * is likely to save.  The admin can still enable it through /sys.
535          */
536         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
537                 transparent_hugepage_flags = 0;
538
539         start_khugepaged();
540
541         set_recommended_min_free_kbytes();
542
543 out:
544         return err;
545 }
546 module_init(hugepage_init)
547
548 static int __init setup_transparent_hugepage(char *str)
549 {
550         int ret = 0;
551         if (!str)
552                 goto out;
553         if (!strcmp(str, "always")) {
554                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
555                         &transparent_hugepage_flags);
556                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
557                           &transparent_hugepage_flags);
558                 ret = 1;
559         } else if (!strcmp(str, "madvise")) {
560                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
561                           &transparent_hugepage_flags);
562                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
563                         &transparent_hugepage_flags);
564                 ret = 1;
565         } else if (!strcmp(str, "never")) {
566                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
567                           &transparent_hugepage_flags);
568                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
569                           &transparent_hugepage_flags);
570                 ret = 1;
571         }
572 out:
573         if (!ret)
574                 printk(KERN_WARNING
575                        "transparent_hugepage= cannot parse, ignored\n");
576         return ret;
577 }
578 __setup("transparent_hugepage=", setup_transparent_hugepage);
579
580 static void prepare_pmd_huge_pte(pgtable_t pgtable,
581                                  struct mm_struct *mm)
582 {
583         assert_spin_locked(&mm->page_table_lock);
584
585         /* FIFO */
586         if (!mm->pmd_huge_pte)
587                 INIT_LIST_HEAD(&pgtable->lru);
588         else
589                 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
590         mm->pmd_huge_pte = pgtable;
591 }
592
593 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
594 {
595         if (likely(vma->vm_flags & VM_WRITE))
596                 pmd = pmd_mkwrite(pmd);
597         return pmd;
598 }
599
600 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
601                                         struct vm_area_struct *vma,
602                                         unsigned long haddr, pmd_t *pmd,
603                                         struct page *page)
604 {
605         int ret = 0;
606         pgtable_t pgtable;
607
608         VM_BUG_ON(!PageCompound(page));
609         pgtable = pte_alloc_one(mm, haddr);
610         if (unlikely(!pgtable)) {
611                 mem_cgroup_uncharge_page(page);
612                 put_page(page);
613                 return VM_FAULT_OOM;
614         }
615
616         clear_huge_page(page, haddr, HPAGE_PMD_NR);
617         __SetPageUptodate(page);
618
619         spin_lock(&mm->page_table_lock);
620         if (unlikely(!pmd_none(*pmd))) {
621                 spin_unlock(&mm->page_table_lock);
622                 mem_cgroup_uncharge_page(page);
623                 put_page(page);
624                 pte_free(mm, pgtable);
625         } else {
626                 pmd_t entry;
627                 entry = mk_pmd(page, vma->vm_page_prot);
628                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
629                 entry = pmd_mkhuge(entry);
630                 /*
631                  * The spinlocking to take the lru_lock inside
632                  * page_add_new_anon_rmap() acts as a full memory
633                  * barrier to be sure clear_huge_page writes become
634                  * visible after the set_pmd_at() write.
635                  */
636                 page_add_new_anon_rmap(page, vma, haddr);
637                 set_pmd_at(mm, haddr, pmd, entry);
638                 prepare_pmd_huge_pte(pgtable, mm);
639                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
640                 spin_unlock(&mm->page_table_lock);
641         }
642
643         return ret;
644 }
645
646 static inline gfp_t alloc_hugepage_gfpmask(int defrag)
647 {
648         return GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT);
649 }
650
651 static inline struct page *alloc_hugepage_vma(int defrag,
652                                               struct vm_area_struct *vma,
653                                               unsigned long haddr)
654 {
655         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag),
656                                HPAGE_PMD_ORDER, vma, haddr);
657 }
658
659 #ifndef CONFIG_NUMA
660 static inline struct page *alloc_hugepage(int defrag)
661 {
662         return alloc_pages(alloc_hugepage_gfpmask(defrag),
663                            HPAGE_PMD_ORDER);
664 }
665 #endif
666
667 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
668                                unsigned long address, pmd_t *pmd,
669                                unsigned int flags)
670 {
671         struct page *page;
672         unsigned long haddr = address & HPAGE_PMD_MASK;
673         pte_t *pte;
674
675         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
676                 if (unlikely(anon_vma_prepare(vma)))
677                         return VM_FAULT_OOM;
678                 if (unlikely(khugepaged_enter(vma)))
679                         return VM_FAULT_OOM;
680                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
681                                           vma, haddr);
682                 if (unlikely(!page))
683                         goto out;
684                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
685                         put_page(page);
686                         goto out;
687                 }
688
689                 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
690         }
691 out:
692         /*
693          * Use __pte_alloc instead of pte_alloc_map, because we can't
694          * run pte_offset_map on the pmd, if an huge pmd could
695          * materialize from under us from a different thread.
696          */
697         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
698                 return VM_FAULT_OOM;
699         /* if an huge pmd materialized from under us just retry later */
700         if (unlikely(pmd_trans_huge(*pmd)))
701                 return 0;
702         /*
703          * A regular pmd is established and it can't morph into a huge pmd
704          * from under us anymore at this point because we hold the mmap_sem
705          * read mode and khugepaged takes it in write mode. So now it's
706          * safe to run pte_offset_map().
707          */
708         pte = pte_offset_map(pmd, address);
709         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
710 }
711
712 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
713                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
714                   struct vm_area_struct *vma)
715 {
716         struct page *src_page;
717         pmd_t pmd;
718         pgtable_t pgtable;
719         int ret;
720
721         ret = -ENOMEM;
722         pgtable = pte_alloc_one(dst_mm, addr);
723         if (unlikely(!pgtable))
724                 goto out;
725
726         spin_lock(&dst_mm->page_table_lock);
727         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
728
729         ret = -EAGAIN;
730         pmd = *src_pmd;
731         if (unlikely(!pmd_trans_huge(pmd))) {
732                 pte_free(dst_mm, pgtable);
733                 goto out_unlock;
734         }
735         if (unlikely(pmd_trans_splitting(pmd))) {
736                 /* split huge page running from under us */
737                 spin_unlock(&src_mm->page_table_lock);
738                 spin_unlock(&dst_mm->page_table_lock);
739                 pte_free(dst_mm, pgtable);
740
741                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
742                 goto out;
743         }
744         src_page = pmd_page(pmd);
745         VM_BUG_ON(!PageHead(src_page));
746         get_page(src_page);
747         page_dup_rmap(src_page);
748         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
749
750         pmdp_set_wrprotect(src_mm, addr, src_pmd);
751         pmd = pmd_mkold(pmd_wrprotect(pmd));
752         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
753         prepare_pmd_huge_pte(pgtable, dst_mm);
754
755         ret = 0;
756 out_unlock:
757         spin_unlock(&src_mm->page_table_lock);
758         spin_unlock(&dst_mm->page_table_lock);
759 out:
760         return ret;
761 }
762
763 /* no "address" argument so destroys page coloring of some arch */
764 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
765 {
766         pgtable_t pgtable;
767
768         assert_spin_locked(&mm->page_table_lock);
769
770         /* FIFO */
771         pgtable = mm->pmd_huge_pte;
772         if (list_empty(&pgtable->lru))
773                 mm->pmd_huge_pte = NULL;
774         else {
775                 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
776                                               struct page, lru);
777                 list_del(&pgtable->lru);
778         }
779         return pgtable;
780 }
781
782 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
783                                         struct vm_area_struct *vma,
784                                         unsigned long address,
785                                         pmd_t *pmd, pmd_t orig_pmd,
786                                         struct page *page,
787                                         unsigned long haddr)
788 {
789         pgtable_t pgtable;
790         pmd_t _pmd;
791         int ret = 0, i;
792         struct page **pages;
793
794         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
795                         GFP_KERNEL);
796         if (unlikely(!pages)) {
797                 ret |= VM_FAULT_OOM;
798                 goto out;
799         }
800
801         for (i = 0; i < HPAGE_PMD_NR; i++) {
802                 pages[i] = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
803                                           vma, address);
804                 if (unlikely(!pages[i] ||
805                              mem_cgroup_newpage_charge(pages[i], mm,
806                                                        GFP_KERNEL))) {
807                         if (pages[i])
808                                 put_page(pages[i]);
809                         mem_cgroup_uncharge_start();
810                         while (--i >= 0) {
811                                 mem_cgroup_uncharge_page(pages[i]);
812                                 put_page(pages[i]);
813                         }
814                         mem_cgroup_uncharge_end();
815                         kfree(pages);
816                         ret |= VM_FAULT_OOM;
817                         goto out;
818                 }
819         }
820
821         for (i = 0; i < HPAGE_PMD_NR; i++) {
822                 copy_user_highpage(pages[i], page + i,
823                                    haddr + PAGE_SHIFT*i, vma);
824                 __SetPageUptodate(pages[i]);
825                 cond_resched();
826         }
827
828         spin_lock(&mm->page_table_lock);
829         if (unlikely(!pmd_same(*pmd, orig_pmd)))
830                 goto out_free_pages;
831         VM_BUG_ON(!PageHead(page));
832
833         pmdp_clear_flush_notify(vma, haddr, pmd);
834         /* leave pmd empty until pte is filled */
835
836         pgtable = get_pmd_huge_pte(mm);
837         pmd_populate(mm, &_pmd, pgtable);
838
839         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
840                 pte_t *pte, entry;
841                 entry = mk_pte(pages[i], vma->vm_page_prot);
842                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
843                 page_add_new_anon_rmap(pages[i], vma, haddr);
844                 pte = pte_offset_map(&_pmd, haddr);
845                 VM_BUG_ON(!pte_none(*pte));
846                 set_pte_at(mm, haddr, pte, entry);
847                 pte_unmap(pte);
848         }
849         kfree(pages);
850
851         mm->nr_ptes++;
852         smp_wmb(); /* make pte visible before pmd */
853         pmd_populate(mm, pmd, pgtable);
854         page_remove_rmap(page);
855         spin_unlock(&mm->page_table_lock);
856
857         ret |= VM_FAULT_WRITE;
858         put_page(page);
859
860 out:
861         return ret;
862
863 out_free_pages:
864         spin_unlock(&mm->page_table_lock);
865         mem_cgroup_uncharge_start();
866         for (i = 0; i < HPAGE_PMD_NR; i++) {
867                 mem_cgroup_uncharge_page(pages[i]);
868                 put_page(pages[i]);
869         }
870         mem_cgroup_uncharge_end();
871         kfree(pages);
872         goto out;
873 }
874
875 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
876                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
877 {
878         int ret = 0;
879         struct page *page, *new_page;
880         unsigned long haddr;
881
882         VM_BUG_ON(!vma->anon_vma);
883         spin_lock(&mm->page_table_lock);
884         if (unlikely(!pmd_same(*pmd, orig_pmd)))
885                 goto out_unlock;
886
887         page = pmd_page(orig_pmd);
888         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
889         haddr = address & HPAGE_PMD_MASK;
890         if (page_mapcount(page) == 1) {
891                 pmd_t entry;
892                 entry = pmd_mkyoung(orig_pmd);
893                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
894                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
895                         update_mmu_cache(vma, address, entry);
896                 ret |= VM_FAULT_WRITE;
897                 goto out_unlock;
898         }
899         get_page(page);
900         spin_unlock(&mm->page_table_lock);
901
902         if (transparent_hugepage_enabled(vma) &&
903             !transparent_hugepage_debug_cow())
904                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
905                                               vma, haddr);
906         else
907                 new_page = NULL;
908
909         if (unlikely(!new_page)) {
910                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
911                                                    pmd, orig_pmd, page, haddr);
912                 put_page(page);
913                 goto out;
914         }
915
916         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
917                 put_page(new_page);
918                 put_page(page);
919                 ret |= VM_FAULT_OOM;
920                 goto out;
921         }
922
923         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
924         __SetPageUptodate(new_page);
925
926         spin_lock(&mm->page_table_lock);
927         put_page(page);
928         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
929                 mem_cgroup_uncharge_page(new_page);
930                 put_page(new_page);
931         } else {
932                 pmd_t entry;
933                 VM_BUG_ON(!PageHead(page));
934                 entry = mk_pmd(new_page, vma->vm_page_prot);
935                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
936                 entry = pmd_mkhuge(entry);
937                 pmdp_clear_flush_notify(vma, haddr, pmd);
938                 page_add_new_anon_rmap(new_page, vma, haddr);
939                 set_pmd_at(mm, haddr, pmd, entry);
940                 update_mmu_cache(vma, address, entry);
941                 page_remove_rmap(page);
942                 put_page(page);
943                 ret |= VM_FAULT_WRITE;
944         }
945 out_unlock:
946         spin_unlock(&mm->page_table_lock);
947 out:
948         return ret;
949 }
950
951 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
952                                    unsigned long addr,
953                                    pmd_t *pmd,
954                                    unsigned int flags)
955 {
956         struct page *page = NULL;
957
958         assert_spin_locked(&mm->page_table_lock);
959
960         if (flags & FOLL_WRITE && !pmd_write(*pmd))
961                 goto out;
962
963         page = pmd_page(*pmd);
964         VM_BUG_ON(!PageHead(page));
965         if (flags & FOLL_TOUCH) {
966                 pmd_t _pmd;
967                 /*
968                  * We should set the dirty bit only for FOLL_WRITE but
969                  * for now the dirty bit in the pmd is meaningless.
970                  * And if the dirty bit will become meaningful and
971                  * we'll only set it with FOLL_WRITE, an atomic
972                  * set_bit will be required on the pmd to set the
973                  * young bit, instead of the current set_pmd_at.
974                  */
975                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
976                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
977         }
978         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
979         VM_BUG_ON(!PageCompound(page));
980         if (flags & FOLL_GET)
981                 get_page(page);
982
983 out:
984         return page;
985 }
986
987 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
988                  pmd_t *pmd)
989 {
990         int ret = 0;
991
992         spin_lock(&tlb->mm->page_table_lock);
993         if (likely(pmd_trans_huge(*pmd))) {
994                 if (unlikely(pmd_trans_splitting(*pmd))) {
995                         spin_unlock(&tlb->mm->page_table_lock);
996                         wait_split_huge_page(vma->anon_vma,
997                                              pmd);
998                 } else {
999                         struct page *page;
1000                         pgtable_t pgtable;
1001                         pgtable = get_pmd_huge_pte(tlb->mm);
1002                         page = pmd_page(*pmd);
1003                         pmd_clear(pmd);
1004                         page_remove_rmap(page);
1005                         VM_BUG_ON(page_mapcount(page) < 0);
1006                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1007                         VM_BUG_ON(!PageHead(page));
1008                         spin_unlock(&tlb->mm->page_table_lock);
1009                         tlb_remove_page(tlb, page);
1010                         pte_free(tlb->mm, pgtable);
1011                         ret = 1;
1012                 }
1013         } else
1014                 spin_unlock(&tlb->mm->page_table_lock);
1015
1016         return ret;
1017 }
1018
1019 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1020                 unsigned long addr, unsigned long end,
1021                 unsigned char *vec)
1022 {
1023         int ret = 0;
1024
1025         spin_lock(&vma->vm_mm->page_table_lock);
1026         if (likely(pmd_trans_huge(*pmd))) {
1027                 ret = !pmd_trans_splitting(*pmd);
1028                 spin_unlock(&vma->vm_mm->page_table_lock);
1029                 if (unlikely(!ret))
1030                         wait_split_huge_page(vma->anon_vma, pmd);
1031                 else {
1032                         /*
1033                          * All logical pages in the range are present
1034                          * if backed by a huge page.
1035                          */
1036                         memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1037                 }
1038         } else
1039                 spin_unlock(&vma->vm_mm->page_table_lock);
1040
1041         return ret;
1042 }
1043
1044 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1045                 unsigned long addr, pgprot_t newprot)
1046 {
1047         struct mm_struct *mm = vma->vm_mm;
1048         int ret = 0;
1049
1050         spin_lock(&mm->page_table_lock);
1051         if (likely(pmd_trans_huge(*pmd))) {
1052                 if (unlikely(pmd_trans_splitting(*pmd))) {
1053                         spin_unlock(&mm->page_table_lock);
1054                         wait_split_huge_page(vma->anon_vma, pmd);
1055                 } else {
1056                         pmd_t entry;
1057
1058                         entry = pmdp_get_and_clear(mm, addr, pmd);
1059                         entry = pmd_modify(entry, newprot);
1060                         set_pmd_at(mm, addr, pmd, entry);
1061                         spin_unlock(&vma->vm_mm->page_table_lock);
1062                         flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1063                         ret = 1;
1064                 }
1065         } else
1066                 spin_unlock(&vma->vm_mm->page_table_lock);
1067
1068         return ret;
1069 }
1070
1071 pmd_t *page_check_address_pmd(struct page *page,
1072                               struct mm_struct *mm,
1073                               unsigned long address,
1074                               enum page_check_address_pmd_flag flag)
1075 {
1076         pgd_t *pgd;
1077         pud_t *pud;
1078         pmd_t *pmd, *ret = NULL;
1079
1080         if (address & ~HPAGE_PMD_MASK)
1081                 goto out;
1082
1083         pgd = pgd_offset(mm, address);
1084         if (!pgd_present(*pgd))
1085                 goto out;
1086
1087         pud = pud_offset(pgd, address);
1088         if (!pud_present(*pud))
1089                 goto out;
1090
1091         pmd = pmd_offset(pud, address);
1092         if (pmd_none(*pmd))
1093                 goto out;
1094         if (pmd_page(*pmd) != page)
1095                 goto out;
1096         /*
1097          * split_vma() may create temporary aliased mappings. There is
1098          * no risk as long as all huge pmd are found and have their
1099          * splitting bit set before __split_huge_page_refcount
1100          * runs. Finding the same huge pmd more than once during the
1101          * same rmap walk is not a problem.
1102          */
1103         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1104             pmd_trans_splitting(*pmd))
1105                 goto out;
1106         if (pmd_trans_huge(*pmd)) {
1107                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1108                           !pmd_trans_splitting(*pmd));
1109                 ret = pmd;
1110         }
1111 out:
1112         return ret;
1113 }
1114
1115 static int __split_huge_page_splitting(struct page *page,
1116                                        struct vm_area_struct *vma,
1117                                        unsigned long address)
1118 {
1119         struct mm_struct *mm = vma->vm_mm;
1120         pmd_t *pmd;
1121         int ret = 0;
1122
1123         spin_lock(&mm->page_table_lock);
1124         pmd = page_check_address_pmd(page, mm, address,
1125                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1126         if (pmd) {
1127                 /*
1128                  * We can't temporarily set the pmd to null in order
1129                  * to split it, the pmd must remain marked huge at all
1130                  * times or the VM won't take the pmd_trans_huge paths
1131                  * and it won't wait on the anon_vma->root->lock to
1132                  * serialize against split_huge_page*.
1133                  */
1134                 pmdp_splitting_flush_notify(vma, address, pmd);
1135                 ret = 1;
1136         }
1137         spin_unlock(&mm->page_table_lock);
1138
1139         return ret;
1140 }
1141
1142 static void __split_huge_page_refcount(struct page *page)
1143 {
1144         int i;
1145         unsigned long head_index = page->index;
1146         struct zone *zone = page_zone(page);
1147         int zonestat;
1148
1149         /* prevent PageLRU to go away from under us, and freeze lru stats */
1150         spin_lock_irq(&zone->lru_lock);
1151         compound_lock(page);
1152
1153         for (i = 1; i < HPAGE_PMD_NR; i++) {
1154                 struct page *page_tail = page + i;
1155
1156                 /* tail_page->_count cannot change */
1157                 atomic_sub(atomic_read(&page_tail->_count), &page->_count);
1158                 BUG_ON(page_count(page) <= 0);
1159                 atomic_add(page_mapcount(page) + 1, &page_tail->_count);
1160                 BUG_ON(atomic_read(&page_tail->_count) <= 0);
1161
1162                 /* after clearing PageTail the gup refcount can be released */
1163                 smp_mb();
1164
1165                 /*
1166                  * retain hwpoison flag of the poisoned tail page:
1167                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1168                  *   by the memory-failure.
1169                  */
1170                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1171                 page_tail->flags |= (page->flags &
1172                                      ((1L << PG_referenced) |
1173                                       (1L << PG_swapbacked) |
1174                                       (1L << PG_mlocked) |
1175                                       (1L << PG_uptodate)));
1176                 page_tail->flags |= (1L << PG_dirty);
1177
1178                 /*
1179                  * 1) clear PageTail before overwriting first_page
1180                  * 2) clear PageTail before clearing PageHead for VM_BUG_ON
1181                  */
1182                 smp_wmb();
1183
1184                 /*
1185                  * __split_huge_page_splitting() already set the
1186                  * splitting bit in all pmd that could map this
1187                  * hugepage, that will ensure no CPU can alter the
1188                  * mapcount on the head page. The mapcount is only
1189                  * accounted in the head page and it has to be
1190                  * transferred to all tail pages in the below code. So
1191                  * for this code to be safe, the split the mapcount
1192                  * can't change. But that doesn't mean userland can't
1193                  * keep changing and reading the page contents while
1194                  * we transfer the mapcount, so the pmd splitting
1195                  * status is achieved setting a reserved bit in the
1196                  * pmd, not by clearing the present bit.
1197                 */
1198                 BUG_ON(page_mapcount(page_tail));
1199                 page_tail->_mapcount = page->_mapcount;
1200
1201                 BUG_ON(page_tail->mapping);
1202                 page_tail->mapping = page->mapping;
1203
1204                 page_tail->index = ++head_index;
1205
1206                 BUG_ON(!PageAnon(page_tail));
1207                 BUG_ON(!PageUptodate(page_tail));
1208                 BUG_ON(!PageDirty(page_tail));
1209                 BUG_ON(!PageSwapBacked(page_tail));
1210
1211                 mem_cgroup_split_huge_fixup(page, page_tail);
1212
1213                 lru_add_page_tail(zone, page, page_tail);
1214         }
1215
1216         __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1217         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1218
1219         /*
1220          * A hugepage counts for HPAGE_PMD_NR pages on the LRU statistics,
1221          * so adjust those appropriately if this page is on the LRU.
1222          */
1223         if (PageLRU(page)) {
1224                 zonestat = NR_LRU_BASE + page_lru(page);
1225                 __mod_zone_page_state(zone, zonestat, -(HPAGE_PMD_NR-1));
1226         }
1227
1228         ClearPageCompound(page);
1229         compound_unlock(page);
1230         spin_unlock_irq(&zone->lru_lock);
1231
1232         for (i = 1; i < HPAGE_PMD_NR; i++) {
1233                 struct page *page_tail = page + i;
1234                 BUG_ON(page_count(page_tail) <= 0);
1235                 /*
1236                  * Tail pages may be freed if there wasn't any mapping
1237                  * like if add_to_swap() is running on a lru page that
1238                  * had its mapping zapped. And freeing these pages
1239                  * requires taking the lru_lock so we do the put_page
1240                  * of the tail pages after the split is complete.
1241                  */
1242                 put_page(page_tail);
1243         }
1244
1245         /*
1246          * Only the head page (now become a regular page) is required
1247          * to be pinned by the caller.
1248          */
1249         BUG_ON(page_count(page) <= 0);
1250 }
1251
1252 static int __split_huge_page_map(struct page *page,
1253                                  struct vm_area_struct *vma,
1254                                  unsigned long address)
1255 {
1256         struct mm_struct *mm = vma->vm_mm;
1257         pmd_t *pmd, _pmd;
1258         int ret = 0, i;
1259         pgtable_t pgtable;
1260         unsigned long haddr;
1261
1262         spin_lock(&mm->page_table_lock);
1263         pmd = page_check_address_pmd(page, mm, address,
1264                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1265         if (pmd) {
1266                 pgtable = get_pmd_huge_pte(mm);
1267                 pmd_populate(mm, &_pmd, pgtable);
1268
1269                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1270                      i++, haddr += PAGE_SIZE) {
1271                         pte_t *pte, entry;
1272                         BUG_ON(PageCompound(page+i));
1273                         entry = mk_pte(page + i, vma->vm_page_prot);
1274                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1275                         if (!pmd_write(*pmd))
1276                                 entry = pte_wrprotect(entry);
1277                         else
1278                                 BUG_ON(page_mapcount(page) != 1);
1279                         if (!pmd_young(*pmd))
1280                                 entry = pte_mkold(entry);
1281                         pte = pte_offset_map(&_pmd, haddr);
1282                         BUG_ON(!pte_none(*pte));
1283                         set_pte_at(mm, haddr, pte, entry);
1284                         pte_unmap(pte);
1285                 }
1286
1287                 mm->nr_ptes++;
1288                 smp_wmb(); /* make pte visible before pmd */
1289                 /*
1290                  * Up to this point the pmd is present and huge and
1291                  * userland has the whole access to the hugepage
1292                  * during the split (which happens in place). If we
1293                  * overwrite the pmd with the not-huge version
1294                  * pointing to the pte here (which of course we could
1295                  * if all CPUs were bug free), userland could trigger
1296                  * a small page size TLB miss on the small sized TLB
1297                  * while the hugepage TLB entry is still established
1298                  * in the huge TLB. Some CPU doesn't like that. See
1299                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1300                  * Erratum 383 on page 93. Intel should be safe but is
1301                  * also warns that it's only safe if the permission
1302                  * and cache attributes of the two entries loaded in
1303                  * the two TLB is identical (which should be the case
1304                  * here). But it is generally safer to never allow
1305                  * small and huge TLB entries for the same virtual
1306                  * address to be loaded simultaneously. So instead of
1307                  * doing "pmd_populate(); flush_tlb_range();" we first
1308                  * mark the current pmd notpresent (atomically because
1309                  * here the pmd_trans_huge and pmd_trans_splitting
1310                  * must remain set at all times on the pmd until the
1311                  * split is complete for this pmd), then we flush the
1312                  * SMP TLB and finally we write the non-huge version
1313                  * of the pmd entry with pmd_populate.
1314                  */
1315                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1316                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1317                 pmd_populate(mm, pmd, pgtable);
1318                 ret = 1;
1319         }
1320         spin_unlock(&mm->page_table_lock);
1321
1322         return ret;
1323 }
1324
1325 /* must be called with anon_vma->root->lock hold */
1326 static void __split_huge_page(struct page *page,
1327                               struct anon_vma *anon_vma)
1328 {
1329         int mapcount, mapcount2;
1330         struct anon_vma_chain *avc;
1331
1332         BUG_ON(!PageHead(page));
1333         BUG_ON(PageTail(page));
1334
1335         mapcount = 0;
1336         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1337                 struct vm_area_struct *vma = avc->vma;
1338                 unsigned long addr = vma_address(page, vma);
1339                 BUG_ON(is_vma_temporary_stack(vma));
1340                 if (addr == -EFAULT)
1341                         continue;
1342                 mapcount += __split_huge_page_splitting(page, vma, addr);
1343         }
1344         /*
1345          * It is critical that new vmas are added to the tail of the
1346          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1347          * and establishes a child pmd before
1348          * __split_huge_page_splitting() freezes the parent pmd (so if
1349          * we fail to prevent copy_huge_pmd() from running until the
1350          * whole __split_huge_page() is complete), we will still see
1351          * the newly established pmd of the child later during the
1352          * walk, to be able to set it as pmd_trans_splitting too.
1353          */
1354         if (mapcount != page_mapcount(page))
1355                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1356                        mapcount, page_mapcount(page));
1357         BUG_ON(mapcount != page_mapcount(page));
1358
1359         __split_huge_page_refcount(page);
1360
1361         mapcount2 = 0;
1362         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1363                 struct vm_area_struct *vma = avc->vma;
1364                 unsigned long addr = vma_address(page, vma);
1365                 BUG_ON(is_vma_temporary_stack(vma));
1366                 if (addr == -EFAULT)
1367                         continue;
1368                 mapcount2 += __split_huge_page_map(page, vma, addr);
1369         }
1370         if (mapcount != mapcount2)
1371                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1372                        mapcount, mapcount2, page_mapcount(page));
1373         BUG_ON(mapcount != mapcount2);
1374 }
1375
1376 int split_huge_page(struct page *page)
1377 {
1378         struct anon_vma *anon_vma;
1379         int ret = 1;
1380
1381         BUG_ON(!PageAnon(page));
1382         anon_vma = page_lock_anon_vma(page);
1383         if (!anon_vma)
1384                 goto out;
1385         ret = 0;
1386         if (!PageCompound(page))
1387                 goto out_unlock;
1388
1389         BUG_ON(!PageSwapBacked(page));
1390         __split_huge_page(page, anon_vma);
1391
1392         BUG_ON(PageCompound(page));
1393 out_unlock:
1394         page_unlock_anon_vma(anon_vma);
1395 out:
1396         return ret;
1397 }
1398
1399 int hugepage_madvise(struct vm_area_struct *vma,
1400                      unsigned long *vm_flags, int advice)
1401 {
1402         switch (advice) {
1403         case MADV_HUGEPAGE:
1404                 /*
1405                  * Be somewhat over-protective like KSM for now!
1406                  */
1407                 if (*vm_flags & (VM_HUGEPAGE |
1408                                  VM_SHARED   | VM_MAYSHARE   |
1409                                  VM_PFNMAP   | VM_IO      | VM_DONTEXPAND |
1410                                  VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1411                                  VM_MIXEDMAP | VM_SAO))
1412                         return -EINVAL;
1413                 *vm_flags &= ~VM_NOHUGEPAGE;
1414                 *vm_flags |= VM_HUGEPAGE;
1415                 /*
1416                  * If the vma become good for khugepaged to scan,
1417                  * register it here without waiting a page fault that
1418                  * may not happen any time soon.
1419                  */
1420                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1421                         return -ENOMEM;
1422                 break;
1423         case MADV_NOHUGEPAGE:
1424                 /*
1425                  * Be somewhat over-protective like KSM for now!
1426                  */
1427                 if (*vm_flags & (VM_NOHUGEPAGE |
1428                                  VM_SHARED   | VM_MAYSHARE   |
1429                                  VM_PFNMAP   | VM_IO      | VM_DONTEXPAND |
1430                                  VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1431                                  VM_MIXEDMAP | VM_SAO))
1432                         return -EINVAL;
1433                 *vm_flags &= ~VM_HUGEPAGE;
1434                 *vm_flags |= VM_NOHUGEPAGE;
1435                 /*
1436                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1437                  * this vma even if we leave the mm registered in khugepaged if
1438                  * it got registered before VM_NOHUGEPAGE was set.
1439                  */
1440                 break;
1441         }
1442
1443         return 0;
1444 }
1445
1446 static int __init khugepaged_slab_init(void)
1447 {
1448         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1449                                           sizeof(struct mm_slot),
1450                                           __alignof__(struct mm_slot), 0, NULL);
1451         if (!mm_slot_cache)
1452                 return -ENOMEM;
1453
1454         return 0;
1455 }
1456
1457 static void __init khugepaged_slab_free(void)
1458 {
1459         kmem_cache_destroy(mm_slot_cache);
1460         mm_slot_cache = NULL;
1461 }
1462
1463 static inline struct mm_slot *alloc_mm_slot(void)
1464 {
1465         if (!mm_slot_cache)     /* initialization failed */
1466                 return NULL;
1467         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1468 }
1469
1470 static inline void free_mm_slot(struct mm_slot *mm_slot)
1471 {
1472         kmem_cache_free(mm_slot_cache, mm_slot);
1473 }
1474
1475 static int __init mm_slots_hash_init(void)
1476 {
1477         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1478                                 GFP_KERNEL);
1479         if (!mm_slots_hash)
1480                 return -ENOMEM;
1481         return 0;
1482 }
1483
1484 #if 0
1485 static void __init mm_slots_hash_free(void)
1486 {
1487         kfree(mm_slots_hash);
1488         mm_slots_hash = NULL;
1489 }
1490 #endif
1491
1492 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1493 {
1494         struct mm_slot *mm_slot;
1495         struct hlist_head *bucket;
1496         struct hlist_node *node;
1497
1498         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1499                                 % MM_SLOTS_HASH_HEADS];
1500         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1501                 if (mm == mm_slot->mm)
1502                         return mm_slot;
1503         }
1504         return NULL;
1505 }
1506
1507 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1508                                     struct mm_slot *mm_slot)
1509 {
1510         struct hlist_head *bucket;
1511
1512         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1513                                 % MM_SLOTS_HASH_HEADS];
1514         mm_slot->mm = mm;
1515         hlist_add_head(&mm_slot->hash, bucket);
1516 }
1517
1518 static inline int khugepaged_test_exit(struct mm_struct *mm)
1519 {
1520         return atomic_read(&mm->mm_users) == 0;
1521 }
1522
1523 int __khugepaged_enter(struct mm_struct *mm)
1524 {
1525         struct mm_slot *mm_slot;
1526         int wakeup;
1527
1528         mm_slot = alloc_mm_slot();
1529         if (!mm_slot)
1530                 return -ENOMEM;
1531
1532         /* __khugepaged_exit() must not run from under us */
1533         VM_BUG_ON(khugepaged_test_exit(mm));
1534         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1535                 free_mm_slot(mm_slot);
1536                 return 0;
1537         }
1538
1539         spin_lock(&khugepaged_mm_lock);
1540         insert_to_mm_slots_hash(mm, mm_slot);
1541         /*
1542          * Insert just behind the scanning cursor, to let the area settle
1543          * down a little.
1544          */
1545         wakeup = list_empty(&khugepaged_scan.mm_head);
1546         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1547         spin_unlock(&khugepaged_mm_lock);
1548
1549         atomic_inc(&mm->mm_count);
1550         if (wakeup)
1551                 wake_up_interruptible(&khugepaged_wait);
1552
1553         return 0;
1554 }
1555
1556 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1557 {
1558         unsigned long hstart, hend;
1559         if (!vma->anon_vma)
1560                 /*
1561                  * Not yet faulted in so we will register later in the
1562                  * page fault if needed.
1563                  */
1564                 return 0;
1565         if (vma->vm_file || vma->vm_ops)
1566                 /* khugepaged not yet working on file or special mappings */
1567                 return 0;
1568         VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
1569         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1570         hend = vma->vm_end & HPAGE_PMD_MASK;
1571         if (hstart < hend)
1572                 return khugepaged_enter(vma);
1573         return 0;
1574 }
1575
1576 void __khugepaged_exit(struct mm_struct *mm)
1577 {
1578         struct mm_slot *mm_slot;
1579         int free = 0;
1580
1581         spin_lock(&khugepaged_mm_lock);
1582         mm_slot = get_mm_slot(mm);
1583         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1584                 hlist_del(&mm_slot->hash);
1585                 list_del(&mm_slot->mm_node);
1586                 free = 1;
1587         }
1588
1589         if (free) {
1590                 spin_unlock(&khugepaged_mm_lock);
1591                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1592                 free_mm_slot(mm_slot);
1593                 mmdrop(mm);
1594         } else if (mm_slot) {
1595                 spin_unlock(&khugepaged_mm_lock);
1596                 /*
1597                  * This is required to serialize against
1598                  * khugepaged_test_exit() (which is guaranteed to run
1599                  * under mmap sem read mode). Stop here (after we
1600                  * return all pagetables will be destroyed) until
1601                  * khugepaged has finished working on the pagetables
1602                  * under the mmap_sem.
1603                  */
1604                 down_write(&mm->mmap_sem);
1605                 up_write(&mm->mmap_sem);
1606         } else
1607                 spin_unlock(&khugepaged_mm_lock);
1608 }
1609
1610 static void release_pte_page(struct page *page)
1611 {
1612         /* 0 stands for page_is_file_cache(page) == false */
1613         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1614         unlock_page(page);
1615         putback_lru_page(page);
1616 }
1617
1618 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1619 {
1620         while (--_pte >= pte) {
1621                 pte_t pteval = *_pte;
1622                 if (!pte_none(pteval))
1623                         release_pte_page(pte_page(pteval));
1624         }
1625 }
1626
1627 static void release_all_pte_pages(pte_t *pte)
1628 {
1629         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1630 }
1631
1632 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1633                                         unsigned long address,
1634                                         pte_t *pte)
1635 {
1636         struct page *page;
1637         pte_t *_pte;
1638         int referenced = 0, isolated = 0, none = 0;
1639         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1640              _pte++, address += PAGE_SIZE) {
1641                 pte_t pteval = *_pte;
1642                 if (pte_none(pteval)) {
1643                         if (++none <= khugepaged_max_ptes_none)
1644                                 continue;
1645                         else {
1646                                 release_pte_pages(pte, _pte);
1647                                 goto out;
1648                         }
1649                 }
1650                 if (!pte_present(pteval) || !pte_write(pteval)) {
1651                         release_pte_pages(pte, _pte);
1652                         goto out;
1653                 }
1654                 page = vm_normal_page(vma, address, pteval);
1655                 if (unlikely(!page)) {
1656                         release_pte_pages(pte, _pte);
1657                         goto out;
1658                 }
1659                 VM_BUG_ON(PageCompound(page));
1660                 BUG_ON(!PageAnon(page));
1661                 VM_BUG_ON(!PageSwapBacked(page));
1662
1663                 /* cannot use mapcount: can't collapse if there's a gup pin */
1664                 if (page_count(page) != 1) {
1665                         release_pte_pages(pte, _pte);
1666                         goto out;
1667                 }
1668                 /*
1669                  * We can do it before isolate_lru_page because the
1670                  * page can't be freed from under us. NOTE: PG_lock
1671                  * is needed to serialize against split_huge_page
1672                  * when invoked from the VM.
1673                  */
1674                 if (!trylock_page(page)) {
1675                         release_pte_pages(pte, _pte);
1676                         goto out;
1677                 }
1678                 /*
1679                  * Isolate the page to avoid collapsing an hugepage
1680                  * currently in use by the VM.
1681                  */
1682                 if (isolate_lru_page(page)) {
1683                         unlock_page(page);
1684                         release_pte_pages(pte, _pte);
1685                         goto out;
1686                 }
1687                 /* 0 stands for page_is_file_cache(page) == false */
1688                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1689                 VM_BUG_ON(!PageLocked(page));
1690                 VM_BUG_ON(PageLRU(page));
1691
1692                 /* If there is no mapped pte young don't collapse the page */
1693                 if (pte_young(pteval) || PageReferenced(page) ||
1694                     mmu_notifier_test_young(vma->vm_mm, address))
1695                         referenced = 1;
1696         }
1697         if (unlikely(!referenced))
1698                 release_all_pte_pages(pte);
1699         else
1700                 isolated = 1;
1701 out:
1702         return isolated;
1703 }
1704
1705 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1706                                       struct vm_area_struct *vma,
1707                                       unsigned long address,
1708                                       spinlock_t *ptl)
1709 {
1710         pte_t *_pte;
1711         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1712                 pte_t pteval = *_pte;
1713                 struct page *src_page;
1714
1715                 if (pte_none(pteval)) {
1716                         clear_user_highpage(page, address);
1717                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1718                 } else {
1719                         src_page = pte_page(pteval);
1720                         copy_user_highpage(page, src_page, address, vma);
1721                         VM_BUG_ON(page_mapcount(src_page) != 1);
1722                         VM_BUG_ON(page_count(src_page) != 2);
1723                         release_pte_page(src_page);
1724                         /*
1725                          * ptl mostly unnecessary, but preempt has to
1726                          * be disabled to update the per-cpu stats
1727                          * inside page_remove_rmap().
1728                          */
1729                         spin_lock(ptl);
1730                         /*
1731                          * paravirt calls inside pte_clear here are
1732                          * superfluous.
1733                          */
1734                         pte_clear(vma->vm_mm, address, _pte);
1735                         page_remove_rmap(src_page);
1736                         spin_unlock(ptl);
1737                         free_page_and_swap_cache(src_page);
1738                 }
1739
1740                 address += PAGE_SIZE;
1741                 page++;
1742         }
1743 }
1744
1745 static void collapse_huge_page(struct mm_struct *mm,
1746                                unsigned long address,
1747                                struct page **hpage,
1748                                struct vm_area_struct *vma)
1749 {
1750         pgd_t *pgd;
1751         pud_t *pud;
1752         pmd_t *pmd, _pmd;
1753         pte_t *pte;
1754         pgtable_t pgtable;
1755         struct page *new_page;
1756         spinlock_t *ptl;
1757         int isolated;
1758         unsigned long hstart, hend;
1759
1760         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1761 #ifndef CONFIG_NUMA
1762         VM_BUG_ON(!*hpage);
1763         new_page = *hpage;
1764 #else
1765         VM_BUG_ON(*hpage);
1766         /*
1767          * Allocate the page while the vma is still valid and under
1768          * the mmap_sem read mode so there is no memory allocation
1769          * later when we take the mmap_sem in write mode. This is more
1770          * friendly behavior (OTOH it may actually hide bugs) to
1771          * filesystems in userland with daemons allocating memory in
1772          * the userland I/O paths.  Allocating memory with the
1773          * mmap_sem in read mode is good idea also to allow greater
1774          * scalability.
1775          */
1776         new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address);
1777         if (unlikely(!new_page)) {
1778                 up_read(&mm->mmap_sem);
1779                 *hpage = ERR_PTR(-ENOMEM);
1780                 return;
1781         }
1782 #endif
1783         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1784                 up_read(&mm->mmap_sem);
1785                 put_page(new_page);
1786                 return;
1787         }
1788
1789         /* after allocating the hugepage upgrade to mmap_sem write mode */
1790         up_read(&mm->mmap_sem);
1791
1792         /*
1793          * Prevent all access to pagetables with the exception of
1794          * gup_fast later hanlded by the ptep_clear_flush and the VM
1795          * handled by the anon_vma lock + PG_lock.
1796          */
1797         down_write(&mm->mmap_sem);
1798         if (unlikely(khugepaged_test_exit(mm)))
1799                 goto out;
1800
1801         vma = find_vma(mm, address);
1802         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1803         hend = vma->vm_end & HPAGE_PMD_MASK;
1804         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1805                 goto out;
1806
1807         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1808             (vma->vm_flags & VM_NOHUGEPAGE))
1809                 goto out;
1810
1811         /* VM_PFNMAP vmas may have vm_ops null but vm_file set */
1812         if (!vma->anon_vma || vma->vm_ops || vma->vm_file)
1813                 goto out;
1814         VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
1815
1816         pgd = pgd_offset(mm, address);
1817         if (!pgd_present(*pgd))
1818                 goto out;
1819
1820         pud = pud_offset(pgd, address);
1821         if (!pud_present(*pud))
1822                 goto out;
1823
1824         pmd = pmd_offset(pud, address);
1825         /* pmd can't go away or become huge under us */
1826         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1827                 goto out;
1828
1829         anon_vma_lock(vma->anon_vma);
1830
1831         pte = pte_offset_map(pmd, address);
1832         ptl = pte_lockptr(mm, pmd);
1833
1834         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1835         /*
1836          * After this gup_fast can't run anymore. This also removes
1837          * any huge TLB entry from the CPU so we won't allow
1838          * huge and small TLB entries for the same virtual address
1839          * to avoid the risk of CPU bugs in that area.
1840          */
1841         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1842         spin_unlock(&mm->page_table_lock);
1843
1844         spin_lock(ptl);
1845         isolated = __collapse_huge_page_isolate(vma, address, pte);
1846         spin_unlock(ptl);
1847
1848         if (unlikely(!isolated)) {
1849                 pte_unmap(pte);
1850                 spin_lock(&mm->page_table_lock);
1851                 BUG_ON(!pmd_none(*pmd));
1852                 set_pmd_at(mm, address, pmd, _pmd);
1853                 spin_unlock(&mm->page_table_lock);
1854                 anon_vma_unlock(vma->anon_vma);
1855                 goto out;
1856         }
1857
1858         /*
1859          * All pages are isolated and locked so anon_vma rmap
1860          * can't run anymore.
1861          */
1862         anon_vma_unlock(vma->anon_vma);
1863
1864         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1865         pte_unmap(pte);
1866         __SetPageUptodate(new_page);
1867         pgtable = pmd_pgtable(_pmd);
1868         VM_BUG_ON(page_count(pgtable) != 1);
1869         VM_BUG_ON(page_mapcount(pgtable) != 0);
1870
1871         _pmd = mk_pmd(new_page, vma->vm_page_prot);
1872         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1873         _pmd = pmd_mkhuge(_pmd);
1874
1875         /*
1876          * spin_lock() below is not the equivalent of smp_wmb(), so
1877          * this is needed to avoid the copy_huge_page writes to become
1878          * visible after the set_pmd_at() write.
1879          */
1880         smp_wmb();
1881
1882         spin_lock(&mm->page_table_lock);
1883         BUG_ON(!pmd_none(*pmd));
1884         page_add_new_anon_rmap(new_page, vma, address);
1885         set_pmd_at(mm, address, pmd, _pmd);
1886         update_mmu_cache(vma, address, entry);
1887         prepare_pmd_huge_pte(pgtable, mm);
1888         mm->nr_ptes--;
1889         spin_unlock(&mm->page_table_lock);
1890
1891 #ifndef CONFIG_NUMA
1892         *hpage = NULL;
1893 #endif
1894         khugepaged_pages_collapsed++;
1895 out_up_write:
1896         up_write(&mm->mmap_sem);
1897         return;
1898
1899 out:
1900         mem_cgroup_uncharge_page(new_page);
1901 #ifdef CONFIG_NUMA
1902         put_page(new_page);
1903 #endif
1904         goto out_up_write;
1905 }
1906
1907 static int khugepaged_scan_pmd(struct mm_struct *mm,
1908                                struct vm_area_struct *vma,
1909                                unsigned long address,
1910                                struct page **hpage)
1911 {
1912         pgd_t *pgd;
1913         pud_t *pud;
1914         pmd_t *pmd;
1915         pte_t *pte, *_pte;
1916         int ret = 0, referenced = 0, none = 0;
1917         struct page *page;
1918         unsigned long _address;
1919         spinlock_t *ptl;
1920
1921         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1922
1923         pgd = pgd_offset(mm, address);
1924         if (!pgd_present(*pgd))
1925                 goto out;
1926
1927         pud = pud_offset(pgd, address);
1928         if (!pud_present(*pud))
1929                 goto out;
1930
1931         pmd = pmd_offset(pud, address);
1932         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1933                 goto out;
1934
1935         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1936         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1937              _pte++, _address += PAGE_SIZE) {
1938                 pte_t pteval = *_pte;
1939                 if (pte_none(pteval)) {
1940                         if (++none <= khugepaged_max_ptes_none)
1941                                 continue;
1942                         else
1943                                 goto out_unmap;
1944                 }
1945                 if (!pte_present(pteval) || !pte_write(pteval))
1946                         goto out_unmap;
1947                 page = vm_normal_page(vma, _address, pteval);
1948                 if (unlikely(!page))
1949                         goto out_unmap;
1950                 VM_BUG_ON(PageCompound(page));
1951                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
1952                         goto out_unmap;
1953                 /* cannot use mapcount: can't collapse if there's a gup pin */
1954                 if (page_count(page) != 1)
1955                         goto out_unmap;
1956                 if (pte_young(pteval) || PageReferenced(page) ||
1957                     mmu_notifier_test_young(vma->vm_mm, address))
1958                         referenced = 1;
1959         }
1960         if (referenced)
1961                 ret = 1;
1962 out_unmap:
1963         pte_unmap_unlock(pte, ptl);
1964         if (ret)
1965                 /* collapse_huge_page will return with the mmap_sem released */
1966                 collapse_huge_page(mm, address, hpage, vma);
1967 out:
1968         return ret;
1969 }
1970
1971 static void collect_mm_slot(struct mm_slot *mm_slot)
1972 {
1973         struct mm_struct *mm = mm_slot->mm;
1974
1975         VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
1976
1977         if (khugepaged_test_exit(mm)) {
1978                 /* free mm_slot */
1979                 hlist_del(&mm_slot->hash);
1980                 list_del(&mm_slot->mm_node);
1981
1982                 /*
1983                  * Not strictly needed because the mm exited already.
1984                  *
1985                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1986                  */
1987
1988                 /* khugepaged_mm_lock actually not necessary for the below */
1989                 free_mm_slot(mm_slot);
1990                 mmdrop(mm);
1991         }
1992 }
1993
1994 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1995                                             struct page **hpage)
1996 {
1997         struct mm_slot *mm_slot;
1998         struct mm_struct *mm;
1999         struct vm_area_struct *vma;
2000         int progress = 0;
2001
2002         VM_BUG_ON(!pages);
2003         VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock));
2004
2005         if (khugepaged_scan.mm_slot)
2006                 mm_slot = khugepaged_scan.mm_slot;
2007         else {
2008                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2009                                      struct mm_slot, mm_node);
2010                 khugepaged_scan.address = 0;
2011                 khugepaged_scan.mm_slot = mm_slot;
2012         }
2013         spin_unlock(&khugepaged_mm_lock);
2014
2015         mm = mm_slot->mm;
2016         down_read(&mm->mmap_sem);
2017         if (unlikely(khugepaged_test_exit(mm)))
2018                 vma = NULL;
2019         else
2020                 vma = find_vma(mm, khugepaged_scan.address);
2021
2022         progress++;
2023         for (; vma; vma = vma->vm_next) {
2024                 unsigned long hstart, hend;
2025
2026                 cond_resched();
2027                 if (unlikely(khugepaged_test_exit(mm))) {
2028                         progress++;
2029                         break;
2030                 }
2031
2032                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2033                      !khugepaged_always()) ||
2034                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2035                         progress++;
2036                         continue;
2037                 }
2038
2039                 /* VM_PFNMAP vmas may have vm_ops null but vm_file set */
2040                 if (!vma->anon_vma || vma->vm_ops || vma->vm_file) {
2041                         khugepaged_scan.address = vma->vm_end;
2042                         progress++;
2043                         continue;
2044                 }
2045                 VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma));
2046
2047                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2048                 hend = vma->vm_end & HPAGE_PMD_MASK;
2049                 if (hstart >= hend) {
2050                         progress++;
2051                         continue;
2052                 }
2053                 if (khugepaged_scan.address < hstart)
2054                         khugepaged_scan.address = hstart;
2055                 if (khugepaged_scan.address > hend) {
2056                         khugepaged_scan.address = hend + HPAGE_PMD_SIZE;
2057                         progress++;
2058                         continue;
2059                 }
2060                 BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2061
2062                 while (khugepaged_scan.address < hend) {
2063                         int ret;
2064                         cond_resched();
2065                         if (unlikely(khugepaged_test_exit(mm)))
2066                                 goto breakouterloop;
2067
2068                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2069                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2070                                   hend);
2071                         ret = khugepaged_scan_pmd(mm, vma,
2072                                                   khugepaged_scan.address,
2073                                                   hpage);
2074                         /* move to next address */
2075                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2076                         progress += HPAGE_PMD_NR;
2077                         if (ret)
2078                                 /* we released mmap_sem so break loop */
2079                                 goto breakouterloop_mmap_sem;
2080                         if (progress >= pages)
2081                                 goto breakouterloop;
2082                 }
2083         }
2084 breakouterloop:
2085         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2086 breakouterloop_mmap_sem:
2087
2088         spin_lock(&khugepaged_mm_lock);
2089         BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2090         /*
2091          * Release the current mm_slot if this mm is about to die, or
2092          * if we scanned all vmas of this mm.
2093          */
2094         if (khugepaged_test_exit(mm) || !vma) {
2095                 /*
2096                  * Make sure that if mm_users is reaching zero while
2097                  * khugepaged runs here, khugepaged_exit will find
2098                  * mm_slot not pointing to the exiting mm.
2099                  */
2100                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2101                         khugepaged_scan.mm_slot = list_entry(
2102                                 mm_slot->mm_node.next,
2103                                 struct mm_slot, mm_node);
2104                         khugepaged_scan.address = 0;
2105                 } else {
2106                         khugepaged_scan.mm_slot = NULL;
2107                         khugepaged_full_scans++;
2108                 }
2109
2110                 collect_mm_slot(mm_slot);
2111         }
2112
2113         return progress;
2114 }
2115
2116 static int khugepaged_has_work(void)
2117 {
2118         return !list_empty(&khugepaged_scan.mm_head) &&
2119                 khugepaged_enabled();
2120 }
2121
2122 static int khugepaged_wait_event(void)
2123 {
2124         return !list_empty(&khugepaged_scan.mm_head) ||
2125                 !khugepaged_enabled();
2126 }
2127
2128 static void khugepaged_do_scan(struct page **hpage)
2129 {
2130         unsigned int progress = 0, pass_through_head = 0;
2131         unsigned int pages = khugepaged_pages_to_scan;
2132
2133         barrier(); /* write khugepaged_pages_to_scan to local stack */
2134
2135         while (progress < pages) {
2136                 cond_resched();
2137
2138 #ifndef CONFIG_NUMA
2139                 if (!*hpage) {
2140                         *hpage = alloc_hugepage(khugepaged_defrag());
2141                         if (unlikely(!*hpage))
2142                                 break;
2143                 }
2144 #else
2145                 if (IS_ERR(*hpage))
2146                         break;
2147 #endif
2148
2149                 if (unlikely(kthread_should_stop() || freezing(current)))
2150                         break;
2151
2152                 spin_lock(&khugepaged_mm_lock);
2153                 if (!khugepaged_scan.mm_slot)
2154                         pass_through_head++;
2155                 if (khugepaged_has_work() &&
2156                     pass_through_head < 2)
2157                         progress += khugepaged_scan_mm_slot(pages - progress,
2158                                                             hpage);
2159                 else
2160                         progress = pages;
2161                 spin_unlock(&khugepaged_mm_lock);
2162         }
2163 }
2164
2165 static void khugepaged_alloc_sleep(void)
2166 {
2167         DEFINE_WAIT(wait);
2168         add_wait_queue(&khugepaged_wait, &wait);
2169         schedule_timeout_interruptible(
2170                 msecs_to_jiffies(
2171                         khugepaged_alloc_sleep_millisecs));
2172         remove_wait_queue(&khugepaged_wait, &wait);
2173 }
2174
2175 #ifndef CONFIG_NUMA
2176 static struct page *khugepaged_alloc_hugepage(void)
2177 {
2178         struct page *hpage;
2179
2180         do {
2181                 hpage = alloc_hugepage(khugepaged_defrag());
2182                 if (!hpage)
2183                         khugepaged_alloc_sleep();
2184         } while (unlikely(!hpage) &&
2185                  likely(khugepaged_enabled()));
2186         return hpage;
2187 }
2188 #endif
2189
2190 static void khugepaged_loop(void)
2191 {
2192         struct page *hpage;
2193
2194 #ifdef CONFIG_NUMA
2195         hpage = NULL;
2196 #endif
2197         while (likely(khugepaged_enabled())) {
2198 #ifndef CONFIG_NUMA
2199                 hpage = khugepaged_alloc_hugepage();
2200                 if (unlikely(!hpage))
2201                         break;
2202 #else
2203                 if (IS_ERR(hpage)) {
2204                         khugepaged_alloc_sleep();
2205                         hpage = NULL;
2206                 }
2207 #endif
2208
2209                 khugepaged_do_scan(&hpage);
2210 #ifndef CONFIG_NUMA
2211                 if (hpage)
2212                         put_page(hpage);
2213 #endif
2214                 try_to_freeze();
2215                 if (unlikely(kthread_should_stop()))
2216                         break;
2217                 if (khugepaged_has_work()) {
2218                         DEFINE_WAIT(wait);
2219                         if (!khugepaged_scan_sleep_millisecs)
2220                                 continue;
2221                         add_wait_queue(&khugepaged_wait, &wait);
2222                         schedule_timeout_interruptible(
2223                                 msecs_to_jiffies(
2224                                         khugepaged_scan_sleep_millisecs));
2225                         remove_wait_queue(&khugepaged_wait, &wait);
2226                 } else if (khugepaged_enabled())
2227                         wait_event_freezable(khugepaged_wait,
2228                                              khugepaged_wait_event());
2229         }
2230 }
2231
2232 static int khugepaged(void *none)
2233 {
2234         struct mm_slot *mm_slot;
2235
2236         set_freezable();
2237         set_user_nice(current, 19);
2238
2239         /* serialize with start_khugepaged() */
2240         mutex_lock(&khugepaged_mutex);
2241
2242         for (;;) {
2243                 mutex_unlock(&khugepaged_mutex);
2244                 BUG_ON(khugepaged_thread != current);
2245                 khugepaged_loop();
2246                 BUG_ON(khugepaged_thread != current);
2247
2248                 mutex_lock(&khugepaged_mutex);
2249                 if (!khugepaged_enabled())
2250                         break;
2251                 if (unlikely(kthread_should_stop()))
2252                         break;
2253         }
2254
2255         spin_lock(&khugepaged_mm_lock);
2256         mm_slot = khugepaged_scan.mm_slot;
2257         khugepaged_scan.mm_slot = NULL;
2258         if (mm_slot)
2259                 collect_mm_slot(mm_slot);
2260         spin_unlock(&khugepaged_mm_lock);
2261
2262         khugepaged_thread = NULL;
2263         mutex_unlock(&khugepaged_mutex);
2264
2265         return 0;
2266 }
2267
2268 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2269 {
2270         struct page *page;
2271
2272         spin_lock(&mm->page_table_lock);
2273         if (unlikely(!pmd_trans_huge(*pmd))) {
2274                 spin_unlock(&mm->page_table_lock);
2275                 return;
2276         }
2277         page = pmd_page(*pmd);
2278         VM_BUG_ON(!page_count(page));
2279         get_page(page);
2280         spin_unlock(&mm->page_table_lock);
2281
2282         split_huge_page(page);
2283
2284         put_page(page);
2285         BUG_ON(pmd_trans_huge(*pmd));
2286 }
2287
2288 static void split_huge_page_address(struct mm_struct *mm,
2289                                     unsigned long address)
2290 {
2291         pgd_t *pgd;
2292         pud_t *pud;
2293         pmd_t *pmd;
2294
2295         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2296
2297         pgd = pgd_offset(mm, address);
2298         if (!pgd_present(*pgd))
2299                 return;
2300
2301         pud = pud_offset(pgd, address);
2302         if (!pud_present(*pud))
2303                 return;
2304
2305         pmd = pmd_offset(pud, address);
2306         if (!pmd_present(*pmd))
2307                 return;
2308         /*
2309          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2310          * materialize from under us.
2311          */
2312         split_huge_page_pmd(mm, pmd);
2313 }
2314
2315 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2316                              unsigned long start,
2317                              unsigned long end,
2318                              long adjust_next)
2319 {
2320         /*
2321          * If the new start address isn't hpage aligned and it could
2322          * previously contain an hugepage: check if we need to split
2323          * an huge pmd.
2324          */
2325         if (start & ~HPAGE_PMD_MASK &&
2326             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2327             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2328                 split_huge_page_address(vma->vm_mm, start);
2329
2330         /*
2331          * If the new end address isn't hpage aligned and it could
2332          * previously contain an hugepage: check if we need to split
2333          * an huge pmd.
2334          */
2335         if (end & ~HPAGE_PMD_MASK &&
2336             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2337             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2338                 split_huge_page_address(vma->vm_mm, end);
2339
2340         /*
2341          * If we're also updating the vma->vm_next->vm_start, if the new
2342          * vm_next->vm_start isn't page aligned and it could previously
2343          * contain an hugepage: check if we need to split an huge pmd.
2344          */
2345         if (adjust_next > 0) {
2346                 struct vm_area_struct *next = vma->vm_next;
2347                 unsigned long nstart = next->vm_start;
2348                 nstart += adjust_next << PAGE_SHIFT;
2349                 if (nstart & ~HPAGE_PMD_MASK &&
2350                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2351                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2352                         split_huge_page_address(next->vm_mm, nstart);
2353         }
2354 }