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