drivers: dma-contiguous: refactor dma_alloc_from_contiguous()
[linux-3.10.git] / mm / huge_memory.c
1 /*
2  *  Copyright (C) 2009  Red Hat, Inc.
3  *
4  *  This work is licensed under the terms of the GNU GPL, version 2. See
5  *  the COPYING file in the top-level directory.
6  */
7
8 #include <linux/mm.h>
9 #include <linux/sched.h>
10 #include <linux/highmem.h>
11 #include <linux/hugetlb.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/rmap.h>
14 #include <linux/swap.h>
15 #include <linux/mm_inline.h>
16 #include <linux/kthread.h>
17 #include <linux/khugepaged.h>
18 #include <linux/freezer.h>
19 #include <linux/mman.h>
20 #include <asm/tlb.h>
21 #include <asm/pgalloc.h>
22 #include "internal.h"
23
24 /*
25  * By default transparent hugepage support is enabled for all mappings
26  * and khugepaged scans all mappings. Defrag is only invoked by
27  * khugepaged hugepage allocations and by page faults inside
28  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
29  * allocations.
30  */
31 unsigned long transparent_hugepage_flags __read_mostly =
32 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
33         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
34 #endif
35 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
36         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
37 #endif
38         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
39         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
40
41 /* default scan 8*512 pte (or vmas) every 30 second */
42 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
43 static unsigned int khugepaged_pages_collapsed;
44 static unsigned int khugepaged_full_scans;
45 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
46 /* during fragmentation poll the hugepage allocator once every minute */
47 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
48 static struct task_struct *khugepaged_thread __read_mostly;
49 static DEFINE_MUTEX(khugepaged_mutex);
50 static DEFINE_SPINLOCK(khugepaged_mm_lock);
51 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
52 /*
53  * default collapse hugepages if there is at least one pte mapped like
54  * it would have happened if the vma was large enough during page
55  * fault.
56  */
57 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
58
59 static int khugepaged(void *none);
60 static int mm_slots_hash_init(void);
61 static int khugepaged_slab_init(void);
62 static void khugepaged_slab_free(void);
63
64 #define MM_SLOTS_HASH_HEADS 1024
65 static struct hlist_head *mm_slots_hash __read_mostly;
66 static struct kmem_cache *mm_slot_cache __read_mostly;
67
68 /**
69  * struct mm_slot - hash lookup from mm to mm_slot
70  * @hash: hash collision list
71  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
72  * @mm: the mm that this information is valid for
73  */
74 struct mm_slot {
75         struct hlist_node hash;
76         struct list_head mm_node;
77         struct mm_struct *mm;
78 };
79
80 /**
81  * struct khugepaged_scan - cursor for scanning
82  * @mm_head: the head of the mm list to scan
83  * @mm_slot: the current mm_slot we are scanning
84  * @address: the next address inside that to be scanned
85  *
86  * There is only the one khugepaged_scan instance of this cursor structure.
87  */
88 struct khugepaged_scan {
89         struct list_head mm_head;
90         struct mm_slot *mm_slot;
91         unsigned long address;
92 };
93 static struct khugepaged_scan khugepaged_scan = {
94         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
95 };
96
97
98 static int set_recommended_min_free_kbytes(void)
99 {
100         struct zone *zone;
101         int nr_zones = 0;
102         unsigned long recommended_min;
103         extern int min_free_kbytes;
104
105         if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG,
106                       &transparent_hugepage_flags) &&
107             !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
108                       &transparent_hugepage_flags))
109                 return 0;
110
111         for_each_populated_zone(zone)
112                 nr_zones++;
113
114         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
115         recommended_min = pageblock_nr_pages * nr_zones * 2;
116
117         /*
118          * Make sure that on average at least two pageblocks are almost free
119          * of another type, one for a migratetype to fall back to and a
120          * second to avoid subsequent fallbacks of other types There are 3
121          * MIGRATE_TYPES we care about.
122          */
123         recommended_min += pageblock_nr_pages * nr_zones *
124                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
125
126         /* don't ever allow to reserve more than 5% of the lowmem */
127         recommended_min = min(recommended_min,
128                               (unsigned long) nr_free_buffer_pages() / 20);
129         recommended_min <<= (PAGE_SHIFT-10);
130
131         if (recommended_min > min_free_kbytes)
132                 min_free_kbytes = recommended_min;
133         setup_per_zone_wmarks();
134         return 0;
135 }
136 late_initcall(set_recommended_min_free_kbytes);
137
138 static int start_khugepaged(void)
139 {
140         int err = 0;
141         if (khugepaged_enabled()) {
142                 int wakeup;
143                 if (unlikely(!mm_slot_cache || !mm_slots_hash)) {
144                         err = -ENOMEM;
145                         goto out;
146                 }
147                 mutex_lock(&khugepaged_mutex);
148                 if (!khugepaged_thread)
149                         khugepaged_thread = kthread_run(khugepaged, NULL,
150                                                         "khugepaged");
151                 if (unlikely(IS_ERR(khugepaged_thread))) {
152                         printk(KERN_ERR
153                                "khugepaged: kthread_run(khugepaged) failed\n");
154                         err = PTR_ERR(khugepaged_thread);
155                         khugepaged_thread = NULL;
156                 }
157                 wakeup = !list_empty(&khugepaged_scan.mm_head);
158                 mutex_unlock(&khugepaged_mutex);
159                 if (wakeup)
160                         wake_up_interruptible(&khugepaged_wait);
161
162                 set_recommended_min_free_kbytes();
163         } else
164                 /* wakeup to exit */
165                 wake_up_interruptible(&khugepaged_wait);
166 out:
167         return err;
168 }
169
170 #ifdef CONFIG_SYSFS
171
172 static ssize_t double_flag_show(struct kobject *kobj,
173                                 struct kobj_attribute *attr, char *buf,
174                                 enum transparent_hugepage_flag enabled,
175                                 enum transparent_hugepage_flag req_madv)
176 {
177         if (test_bit(enabled, &transparent_hugepage_flags)) {
178                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
179                 return sprintf(buf, "[always] madvise never\n");
180         } else if (test_bit(req_madv, &transparent_hugepage_flags))
181                 return sprintf(buf, "always [madvise] never\n");
182         else
183                 return sprintf(buf, "always madvise [never]\n");
184 }
185 static ssize_t double_flag_store(struct kobject *kobj,
186                                  struct kobj_attribute *attr,
187                                  const char *buf, size_t count,
188                                  enum transparent_hugepage_flag enabled,
189                                  enum transparent_hugepage_flag req_madv)
190 {
191         if (!memcmp("always", buf,
192                     min(sizeof("always")-1, count))) {
193                 set_bit(enabled, &transparent_hugepage_flags);
194                 clear_bit(req_madv, &transparent_hugepage_flags);
195         } else if (!memcmp("madvise", buf,
196                            min(sizeof("madvise")-1, count))) {
197                 clear_bit(enabled, &transparent_hugepage_flags);
198                 set_bit(req_madv, &transparent_hugepage_flags);
199         } else if (!memcmp("never", buf,
200                            min(sizeof("never")-1, count))) {
201                 clear_bit(enabled, &transparent_hugepage_flags);
202                 clear_bit(req_madv, &transparent_hugepage_flags);
203         } else
204                 return -EINVAL;
205
206         return count;
207 }
208
209 static ssize_t enabled_show(struct kobject *kobj,
210                             struct kobj_attribute *attr, char *buf)
211 {
212         return double_flag_show(kobj, attr, buf,
213                                 TRANSPARENT_HUGEPAGE_FLAG,
214                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
215 }
216 static ssize_t enabled_store(struct kobject *kobj,
217                              struct kobj_attribute *attr,
218                              const char *buf, size_t count)
219 {
220         ssize_t ret;
221
222         ret = double_flag_store(kobj, attr, buf, count,
223                                 TRANSPARENT_HUGEPAGE_FLAG,
224                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
225
226         if (ret > 0) {
227                 int err = start_khugepaged();
228                 if (err)
229                         ret = err;
230         }
231
232         if (ret > 0 &&
233             (test_bit(TRANSPARENT_HUGEPAGE_FLAG,
234                       &transparent_hugepage_flags) ||
235              test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
236                       &transparent_hugepage_flags)))
237                 set_recommended_min_free_kbytes();
238
239         return ret;
240 }
241 static struct kobj_attribute enabled_attr =
242         __ATTR(enabled, 0644, enabled_show, enabled_store);
243
244 static ssize_t single_flag_show(struct kobject *kobj,
245                                 struct kobj_attribute *attr, char *buf,
246                                 enum transparent_hugepage_flag flag)
247 {
248         return sprintf(buf, "%d\n",
249                        !!test_bit(flag, &transparent_hugepage_flags));
250 }
251
252 static ssize_t single_flag_store(struct kobject *kobj,
253                                  struct kobj_attribute *attr,
254                                  const char *buf, size_t count,
255                                  enum transparent_hugepage_flag flag)
256 {
257         unsigned long value;
258         int ret;
259
260         ret = kstrtoul(buf, 10, &value);
261         if (ret < 0)
262                 return ret;
263         if (value > 1)
264                 return -EINVAL;
265
266         if (value)
267                 set_bit(flag, &transparent_hugepage_flags);
268         else
269                 clear_bit(flag, &transparent_hugepage_flags);
270
271         return count;
272 }
273
274 /*
275  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
276  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
277  * memory just to allocate one more hugepage.
278  */
279 static ssize_t defrag_show(struct kobject *kobj,
280                            struct kobj_attribute *attr, char *buf)
281 {
282         return double_flag_show(kobj, attr, buf,
283                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
284                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
285 }
286 static ssize_t defrag_store(struct kobject *kobj,
287                             struct kobj_attribute *attr,
288                             const char *buf, size_t count)
289 {
290         return double_flag_store(kobj, attr, buf, count,
291                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
292                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
293 }
294 static struct kobj_attribute defrag_attr =
295         __ATTR(defrag, 0644, defrag_show, defrag_store);
296
297 #ifdef CONFIG_DEBUG_VM
298 static ssize_t debug_cow_show(struct kobject *kobj,
299                                 struct kobj_attribute *attr, char *buf)
300 {
301         return single_flag_show(kobj, attr, buf,
302                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
303 }
304 static ssize_t debug_cow_store(struct kobject *kobj,
305                                struct kobj_attribute *attr,
306                                const char *buf, size_t count)
307 {
308         return single_flag_store(kobj, attr, buf, count,
309                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
310 }
311 static struct kobj_attribute debug_cow_attr =
312         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
313 #endif /* CONFIG_DEBUG_VM */
314
315 static struct attribute *hugepage_attr[] = {
316         &enabled_attr.attr,
317         &defrag_attr.attr,
318 #ifdef CONFIG_DEBUG_VM
319         &debug_cow_attr.attr,
320 #endif
321         NULL,
322 };
323
324 static struct attribute_group hugepage_attr_group = {
325         .attrs = hugepage_attr,
326 };
327
328 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
329                                          struct kobj_attribute *attr,
330                                          char *buf)
331 {
332         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
333 }
334
335 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
336                                           struct kobj_attribute *attr,
337                                           const char *buf, size_t count)
338 {
339         unsigned long msecs;
340         int err;
341
342         err = strict_strtoul(buf, 10, &msecs);
343         if (err || msecs > UINT_MAX)
344                 return -EINVAL;
345
346         khugepaged_scan_sleep_millisecs = msecs;
347         wake_up_interruptible(&khugepaged_wait);
348
349         return count;
350 }
351 static struct kobj_attribute scan_sleep_millisecs_attr =
352         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
353                scan_sleep_millisecs_store);
354
355 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
356                                           struct kobj_attribute *attr,
357                                           char *buf)
358 {
359         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
360 }
361
362 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
363                                            struct kobj_attribute *attr,
364                                            const char *buf, size_t count)
365 {
366         unsigned long msecs;
367         int err;
368
369         err = strict_strtoul(buf, 10, &msecs);
370         if (err || msecs > UINT_MAX)
371                 return -EINVAL;
372
373         khugepaged_alloc_sleep_millisecs = msecs;
374         wake_up_interruptible(&khugepaged_wait);
375
376         return count;
377 }
378 static struct kobj_attribute alloc_sleep_millisecs_attr =
379         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
380                alloc_sleep_millisecs_store);
381
382 static ssize_t pages_to_scan_show(struct kobject *kobj,
383                                   struct kobj_attribute *attr,
384                                   char *buf)
385 {
386         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
387 }
388 static ssize_t pages_to_scan_store(struct kobject *kobj,
389                                    struct kobj_attribute *attr,
390                                    const char *buf, size_t count)
391 {
392         int err;
393         unsigned long pages;
394
395         err = strict_strtoul(buf, 10, &pages);
396         if (err || !pages || pages > UINT_MAX)
397                 return -EINVAL;
398
399         khugepaged_pages_to_scan = pages;
400
401         return count;
402 }
403 static struct kobj_attribute pages_to_scan_attr =
404         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
405                pages_to_scan_store);
406
407 static ssize_t pages_collapsed_show(struct kobject *kobj,
408                                     struct kobj_attribute *attr,
409                                     char *buf)
410 {
411         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
412 }
413 static struct kobj_attribute pages_collapsed_attr =
414         __ATTR_RO(pages_collapsed);
415
416 static ssize_t full_scans_show(struct kobject *kobj,
417                                struct kobj_attribute *attr,
418                                char *buf)
419 {
420         return sprintf(buf, "%u\n", khugepaged_full_scans);
421 }
422 static struct kobj_attribute full_scans_attr =
423         __ATTR_RO(full_scans);
424
425 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
426                                       struct kobj_attribute *attr, char *buf)
427 {
428         return single_flag_show(kobj, attr, buf,
429                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
430 }
431 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
432                                        struct kobj_attribute *attr,
433                                        const char *buf, size_t count)
434 {
435         return single_flag_store(kobj, attr, buf, count,
436                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
437 }
438 static struct kobj_attribute khugepaged_defrag_attr =
439         __ATTR(defrag, 0644, khugepaged_defrag_show,
440                khugepaged_defrag_store);
441
442 /*
443  * max_ptes_none controls if khugepaged should collapse hugepages over
444  * any unmapped ptes in turn potentially increasing the memory
445  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
446  * reduce the available free memory in the system as it
447  * runs. Increasing max_ptes_none will instead potentially reduce the
448  * free memory in the system during the khugepaged scan.
449  */
450 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
451                                              struct kobj_attribute *attr,
452                                              char *buf)
453 {
454         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
455 }
456 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
457                                               struct kobj_attribute *attr,
458                                               const char *buf, size_t count)
459 {
460         int err;
461         unsigned long max_ptes_none;
462
463         err = strict_strtoul(buf, 10, &max_ptes_none);
464         if (err || max_ptes_none > HPAGE_PMD_NR-1)
465                 return -EINVAL;
466
467         khugepaged_max_ptes_none = max_ptes_none;
468
469         return count;
470 }
471 static struct kobj_attribute khugepaged_max_ptes_none_attr =
472         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
473                khugepaged_max_ptes_none_store);
474
475 static struct attribute *khugepaged_attr[] = {
476         &khugepaged_defrag_attr.attr,
477         &khugepaged_max_ptes_none_attr.attr,
478         &pages_to_scan_attr.attr,
479         &pages_collapsed_attr.attr,
480         &full_scans_attr.attr,
481         &scan_sleep_millisecs_attr.attr,
482         &alloc_sleep_millisecs_attr.attr,
483         NULL,
484 };
485
486 static struct attribute_group khugepaged_attr_group = {
487         .attrs = khugepaged_attr,
488         .name = "khugepaged",
489 };
490
491 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
492 {
493         int err;
494
495         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
496         if (unlikely(!*hugepage_kobj)) {
497                 printk(KERN_ERR "hugepage: failed kobject create\n");
498                 return -ENOMEM;
499         }
500
501         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
502         if (err) {
503                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
504                 goto delete_obj;
505         }
506
507         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
508         if (err) {
509                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
510                 goto remove_hp_group;
511         }
512
513         return 0;
514
515 remove_hp_group:
516         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
517 delete_obj:
518         kobject_put(*hugepage_kobj);
519         return err;
520 }
521
522 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
523 {
524         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
525         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
526         kobject_put(hugepage_kobj);
527 }
528 #else
529 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
530 {
531         return 0;
532 }
533
534 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
535 {
536 }
537 #endif /* CONFIG_SYSFS */
538
539 static int __init hugepage_init(void)
540 {
541         int err;
542         struct kobject *hugepage_kobj;
543
544         if (!has_transparent_hugepage()) {
545                 transparent_hugepage_flags = 0;
546                 return -EINVAL;
547         }
548
549         err = hugepage_init_sysfs(&hugepage_kobj);
550         if (err)
551                 return err;
552
553         err = khugepaged_slab_init();
554         if (err)
555                 goto out;
556
557         err = mm_slots_hash_init();
558         if (err) {
559                 khugepaged_slab_free();
560                 goto out;
561         }
562
563         /*
564          * By default disable transparent hugepages on smaller systems,
565          * where the extra memory used could hurt more than TLB overhead
566          * is likely to save.  The admin can still enable it through /sys.
567          */
568         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
569                 transparent_hugepage_flags = 0;
570
571         start_khugepaged();
572
573         set_recommended_min_free_kbytes();
574
575         return 0;
576 out:
577         hugepage_exit_sysfs(hugepage_kobj);
578         return err;
579 }
580 module_init(hugepage_init)
581
582 static int __init setup_transparent_hugepage(char *str)
583 {
584         int ret = 0;
585         if (!str)
586                 goto out;
587         if (!strcmp(str, "always")) {
588                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
589                         &transparent_hugepage_flags);
590                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
591                           &transparent_hugepage_flags);
592                 ret = 1;
593         } else if (!strcmp(str, "madvise")) {
594                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
595                           &transparent_hugepage_flags);
596                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
597                         &transparent_hugepage_flags);
598                 ret = 1;
599         } else if (!strcmp(str, "never")) {
600                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
601                           &transparent_hugepage_flags);
602                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
603                           &transparent_hugepage_flags);
604                 ret = 1;
605         }
606 out:
607         if (!ret)
608                 printk(KERN_WARNING
609                        "transparent_hugepage= cannot parse, ignored\n");
610         return ret;
611 }
612 __setup("transparent_hugepage=", setup_transparent_hugepage);
613
614 static void prepare_pmd_huge_pte(pgtable_t pgtable,
615                                  struct mm_struct *mm)
616 {
617         assert_spin_locked(&mm->page_table_lock);
618
619         /* FIFO */
620         if (!mm->pmd_huge_pte)
621                 INIT_LIST_HEAD(&pgtable->lru);
622         else
623                 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
624         mm->pmd_huge_pte = pgtable;
625 }
626
627 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
628 {
629         if (likely(vma->vm_flags & VM_WRITE))
630                 pmd = pmd_mkwrite(pmd);
631         return pmd;
632 }
633
634 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
635                                         struct vm_area_struct *vma,
636                                         unsigned long haddr, pmd_t *pmd,
637                                         struct page *page)
638 {
639         pgtable_t pgtable;
640
641         VM_BUG_ON(!PageCompound(page));
642         pgtable = pte_alloc_one(mm, haddr);
643         if (unlikely(!pgtable))
644                 return VM_FAULT_OOM;
645
646         clear_huge_page(page, haddr, HPAGE_PMD_NR);
647         __SetPageUptodate(page);
648
649         spin_lock(&mm->page_table_lock);
650         if (unlikely(!pmd_none(*pmd))) {
651                 spin_unlock(&mm->page_table_lock);
652                 mem_cgroup_uncharge_page(page);
653                 put_page(page);
654                 pte_free(mm, pgtable);
655         } else {
656                 pmd_t entry;
657                 entry = mk_pmd(page, vma->vm_page_prot);
658                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
659                 entry = pmd_mkhuge(entry);
660                 /*
661                  * The spinlocking to take the lru_lock inside
662                  * page_add_new_anon_rmap() acts as a full memory
663                  * barrier to be sure clear_huge_page writes become
664                  * visible after the set_pmd_at() write.
665                  */
666                 page_add_new_anon_rmap(page, vma, haddr);
667                 set_pmd_at(mm, haddr, pmd, entry);
668                 prepare_pmd_huge_pte(pgtable, mm);
669                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
670                 mm->nr_ptes++;
671                 spin_unlock(&mm->page_table_lock);
672         }
673
674         return 0;
675 }
676
677 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
678 {
679         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
680 }
681
682 static inline struct page *alloc_hugepage_vma(int defrag,
683                                               struct vm_area_struct *vma,
684                                               unsigned long haddr, int nd,
685                                               gfp_t extra_gfp)
686 {
687         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
688                                HPAGE_PMD_ORDER, vma, haddr, nd);
689 }
690
691 #ifndef CONFIG_NUMA
692 static inline struct page *alloc_hugepage(int defrag)
693 {
694         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
695                            HPAGE_PMD_ORDER);
696 }
697 #endif
698
699 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
700                                unsigned long address, pmd_t *pmd,
701                                unsigned int flags)
702 {
703         struct page *page;
704         unsigned long haddr = address & HPAGE_PMD_MASK;
705         pte_t *pte;
706
707         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
708                 if (unlikely(anon_vma_prepare(vma)))
709                         return VM_FAULT_OOM;
710                 if (unlikely(khugepaged_enter(vma)))
711                         return VM_FAULT_OOM;
712                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
713                                           vma, haddr, numa_node_id(), 0);
714                 if (unlikely(!page)) {
715                         count_vm_event(THP_FAULT_FALLBACK);
716                         goto out;
717                 }
718                 count_vm_event(THP_FAULT_ALLOC);
719                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
720                         put_page(page);
721                         goto out;
722                 }
723                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
724                                                           page))) {
725                         mem_cgroup_uncharge_page(page);
726                         put_page(page);
727                         goto out;
728                 }
729
730                 return 0;
731         }
732 out:
733         /*
734          * Use __pte_alloc instead of pte_alloc_map, because we can't
735          * run pte_offset_map on the pmd, if an huge pmd could
736          * materialize from under us from a different thread.
737          */
738         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
739                 return VM_FAULT_OOM;
740         /* if an huge pmd materialized from under us just retry later */
741         if (unlikely(pmd_trans_huge(*pmd)))
742                 return 0;
743         /*
744          * A regular pmd is established and it can't morph into a huge pmd
745          * from under us anymore at this point because we hold the mmap_sem
746          * read mode and khugepaged takes it in write mode. So now it's
747          * safe to run pte_offset_map().
748          */
749         pte = pte_offset_map(pmd, address);
750         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
751 }
752
753 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
754                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
755                   struct vm_area_struct *vma)
756 {
757         struct page *src_page;
758         pmd_t pmd;
759         pgtable_t pgtable;
760         int ret;
761
762         ret = -ENOMEM;
763         pgtable = pte_alloc_one(dst_mm, addr);
764         if (unlikely(!pgtable))
765                 goto out;
766
767         spin_lock(&dst_mm->page_table_lock);
768         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
769
770         ret = -EAGAIN;
771         pmd = *src_pmd;
772         if (unlikely(!pmd_trans_huge(pmd))) {
773                 pte_free(dst_mm, pgtable);
774                 goto out_unlock;
775         }
776         if (unlikely(pmd_trans_splitting(pmd))) {
777                 /* split huge page running from under us */
778                 spin_unlock(&src_mm->page_table_lock);
779                 spin_unlock(&dst_mm->page_table_lock);
780                 pte_free(dst_mm, pgtable);
781
782                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
783                 goto out;
784         }
785         src_page = pmd_page(pmd);
786         VM_BUG_ON(!PageHead(src_page));
787         get_page(src_page);
788         page_dup_rmap(src_page);
789         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
790
791         pmdp_set_wrprotect(src_mm, addr, src_pmd);
792         pmd = pmd_mkold(pmd_wrprotect(pmd));
793         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
794         prepare_pmd_huge_pte(pgtable, dst_mm);
795         dst_mm->nr_ptes++;
796
797         ret = 0;
798 out_unlock:
799         spin_unlock(&src_mm->page_table_lock);
800         spin_unlock(&dst_mm->page_table_lock);
801 out:
802         return ret;
803 }
804
805 /* no "address" argument so destroys page coloring of some arch */
806 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
807 {
808         pgtable_t pgtable;
809
810         assert_spin_locked(&mm->page_table_lock);
811
812         /* FIFO */
813         pgtable = mm->pmd_huge_pte;
814         if (list_empty(&pgtable->lru))
815                 mm->pmd_huge_pte = NULL;
816         else {
817                 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
818                                               struct page, lru);
819                 list_del(&pgtable->lru);
820         }
821         return pgtable;
822 }
823
824 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
825                                         struct vm_area_struct *vma,
826                                         unsigned long address,
827                                         pmd_t *pmd, pmd_t orig_pmd,
828                                         struct page *page,
829                                         unsigned long haddr)
830 {
831         pgtable_t pgtable;
832         pmd_t _pmd;
833         int ret = 0, i;
834         struct page **pages;
835
836         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
837                         GFP_KERNEL);
838         if (unlikely(!pages)) {
839                 ret |= VM_FAULT_OOM;
840                 goto out;
841         }
842
843         for (i = 0; i < HPAGE_PMD_NR; i++) {
844                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
845                                                __GFP_OTHER_NODE,
846                                                vma, address, page_to_nid(page));
847                 if (unlikely(!pages[i] ||
848                              mem_cgroup_newpage_charge(pages[i], mm,
849                                                        GFP_KERNEL))) {
850                         if (pages[i])
851                                 put_page(pages[i]);
852                         mem_cgroup_uncharge_start();
853                         while (--i >= 0) {
854                                 mem_cgroup_uncharge_page(pages[i]);
855                                 put_page(pages[i]);
856                         }
857                         mem_cgroup_uncharge_end();
858                         kfree(pages);
859                         ret |= VM_FAULT_OOM;
860                         goto out;
861                 }
862         }
863
864         for (i = 0; i < HPAGE_PMD_NR; i++) {
865                 copy_user_highpage(pages[i], page + i,
866                                    haddr + PAGE_SIZE * i, vma);
867                 __SetPageUptodate(pages[i]);
868                 cond_resched();
869         }
870
871         spin_lock(&mm->page_table_lock);
872         if (unlikely(!pmd_same(*pmd, orig_pmd)))
873                 goto out_free_pages;
874         VM_BUG_ON(!PageHead(page));
875
876         pmdp_clear_flush_notify(vma, haddr, pmd);
877         /* leave pmd empty until pte is filled */
878
879         pgtable = get_pmd_huge_pte(mm);
880         pmd_populate(mm, &_pmd, pgtable);
881
882         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
883                 pte_t *pte, entry;
884                 entry = mk_pte(pages[i], vma->vm_page_prot);
885                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
886                 page_add_new_anon_rmap(pages[i], vma, haddr);
887                 pte = pte_offset_map(&_pmd, haddr);
888                 VM_BUG_ON(!pte_none(*pte));
889                 set_pte_at(mm, haddr, pte, entry);
890                 pte_unmap(pte);
891         }
892         kfree(pages);
893
894         smp_wmb(); /* make pte visible before pmd */
895         pmd_populate(mm, pmd, pgtable);
896         page_remove_rmap(page);
897         spin_unlock(&mm->page_table_lock);
898
899         ret |= VM_FAULT_WRITE;
900         put_page(page);
901
902 out:
903         return ret;
904
905 out_free_pages:
906         spin_unlock(&mm->page_table_lock);
907         mem_cgroup_uncharge_start();
908         for (i = 0; i < HPAGE_PMD_NR; i++) {
909                 mem_cgroup_uncharge_page(pages[i]);
910                 put_page(pages[i]);
911         }
912         mem_cgroup_uncharge_end();
913         kfree(pages);
914         goto out;
915 }
916
917 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
918                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
919 {
920         int ret = 0;
921         struct page *page, *new_page;
922         unsigned long haddr;
923
924         VM_BUG_ON(!vma->anon_vma);
925         spin_lock(&mm->page_table_lock);
926         if (unlikely(!pmd_same(*pmd, orig_pmd)))
927                 goto out_unlock;
928
929         page = pmd_page(orig_pmd);
930         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
931         haddr = address & HPAGE_PMD_MASK;
932         if (page_mapcount(page) == 1) {
933                 pmd_t entry;
934                 entry = pmd_mkyoung(orig_pmd);
935                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
936                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
937                         update_mmu_cache(vma, address, entry);
938                 ret |= VM_FAULT_WRITE;
939                 goto out_unlock;
940         }
941         get_page(page);
942         spin_unlock(&mm->page_table_lock);
943
944         if (transparent_hugepage_enabled(vma) &&
945             !transparent_hugepage_debug_cow())
946                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
947                                               vma, haddr, numa_node_id(), 0);
948         else
949                 new_page = NULL;
950
951         if (unlikely(!new_page)) {
952                 count_vm_event(THP_FAULT_FALLBACK);
953                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
954                                                    pmd, orig_pmd, page, haddr);
955                 if (ret & VM_FAULT_OOM)
956                         split_huge_page(page);
957                 put_page(page);
958                 goto out;
959         }
960         count_vm_event(THP_FAULT_ALLOC);
961
962         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
963                 put_page(new_page);
964                 split_huge_page(page);
965                 put_page(page);
966                 ret |= VM_FAULT_OOM;
967                 goto out;
968         }
969
970         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
971         __SetPageUptodate(new_page);
972
973         spin_lock(&mm->page_table_lock);
974         put_page(page);
975         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
976                 spin_unlock(&mm->page_table_lock);
977                 mem_cgroup_uncharge_page(new_page);
978                 put_page(new_page);
979                 goto out;
980         } else {
981                 pmd_t entry;
982                 VM_BUG_ON(!PageHead(page));
983                 entry = mk_pmd(new_page, vma->vm_page_prot);
984                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
985                 entry = pmd_mkhuge(entry);
986                 pmdp_clear_flush_notify(vma, haddr, pmd);
987                 page_add_new_anon_rmap(new_page, vma, haddr);
988                 set_pmd_at(mm, haddr, pmd, entry);
989                 update_mmu_cache(vma, address, entry);
990                 page_remove_rmap(page);
991                 put_page(page);
992                 ret |= VM_FAULT_WRITE;
993         }
994 out_unlock:
995         spin_unlock(&mm->page_table_lock);
996 out:
997         return ret;
998 }
999
1000 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
1001                                    unsigned long addr,
1002                                    pmd_t *pmd,
1003                                    unsigned int flags)
1004 {
1005         struct page *page = NULL;
1006
1007         assert_spin_locked(&mm->page_table_lock);
1008
1009         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1010                 goto out;
1011
1012         page = pmd_page(*pmd);
1013         VM_BUG_ON(!PageHead(page));
1014         if (flags & FOLL_TOUCH) {
1015                 pmd_t _pmd;
1016                 /*
1017                  * We should set the dirty bit only for FOLL_WRITE but
1018                  * for now the dirty bit in the pmd is meaningless.
1019                  * And if the dirty bit will become meaningful and
1020                  * we'll only set it with FOLL_WRITE, an atomic
1021                  * set_bit will be required on the pmd to set the
1022                  * young bit, instead of the current set_pmd_at.
1023                  */
1024                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1025                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1026         }
1027         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1028         VM_BUG_ON(!PageCompound(page));
1029         if (flags & FOLL_GET)
1030                 get_page_foll(page);
1031
1032 out:
1033         return page;
1034 }
1035
1036 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1037                  pmd_t *pmd, unsigned long addr)
1038 {
1039         int ret = 0;
1040
1041         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1042                 struct page *page;
1043                 pgtable_t pgtable;
1044                 pgtable = get_pmd_huge_pte(tlb->mm);
1045                 page = pmd_page(*pmd);
1046                 pmd_clear(pmd);
1047                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1048                 page_remove_rmap(page);
1049                 VM_BUG_ON(page_mapcount(page) < 0);
1050                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1051                 VM_BUG_ON(!PageHead(page));
1052                 tlb->mm->nr_ptes--;
1053                 spin_unlock(&tlb->mm->page_table_lock);
1054                 tlb_remove_page(tlb, page);
1055                 pte_free(tlb->mm, pgtable);
1056                 ret = 1;
1057         }
1058         return ret;
1059 }
1060
1061 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1062                 unsigned long addr, unsigned long end,
1063                 unsigned char *vec)
1064 {
1065         int ret = 0;
1066
1067         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1068                 /*
1069                  * All logical pages in the range are present
1070                  * if backed by a huge page.
1071                  */
1072                 spin_unlock(&vma->vm_mm->page_table_lock);
1073                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1074                 ret = 1;
1075         }
1076
1077         return ret;
1078 }
1079
1080 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1081                   unsigned long old_addr,
1082                   unsigned long new_addr, unsigned long old_end,
1083                   pmd_t *old_pmd, pmd_t *new_pmd)
1084 {
1085         int ret = 0;
1086         pmd_t pmd;
1087
1088         struct mm_struct *mm = vma->vm_mm;
1089
1090         if ((old_addr & ~HPAGE_PMD_MASK) ||
1091             (new_addr & ~HPAGE_PMD_MASK) ||
1092             old_end - old_addr < HPAGE_PMD_SIZE ||
1093             (new_vma->vm_flags & VM_NOHUGEPAGE))
1094                 goto out;
1095
1096         /*
1097          * The destination pmd shouldn't be established, free_pgtables()
1098          * should have release it.
1099          */
1100         if (WARN_ON(!pmd_none(*new_pmd))) {
1101                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1102                 goto out;
1103         }
1104
1105         ret = __pmd_trans_huge_lock(old_pmd, vma);
1106         if (ret == 1) {
1107                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1108                 VM_BUG_ON(!pmd_none(*new_pmd));
1109                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1110                 spin_unlock(&mm->page_table_lock);
1111         }
1112 out:
1113         return ret;
1114 }
1115
1116 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1117                 unsigned long addr, pgprot_t newprot)
1118 {
1119         struct mm_struct *mm = vma->vm_mm;
1120         int ret = 0;
1121
1122         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1123                 pmd_t entry;
1124                 entry = pmdp_get_and_clear(mm, addr, pmd);
1125                 entry = pmd_modify(entry, newprot);
1126                 set_pmd_at(mm, addr, pmd, entry);
1127                 spin_unlock(&vma->vm_mm->page_table_lock);
1128                 ret = 1;
1129         }
1130
1131         return ret;
1132 }
1133
1134 /*
1135  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1136  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1137  *
1138  * Note that if it returns 1, this routine returns without unlocking page
1139  * table locks. So callers must unlock them.
1140  */
1141 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1142 {
1143         spin_lock(&vma->vm_mm->page_table_lock);
1144         if (likely(pmd_trans_huge(*pmd))) {
1145                 if (unlikely(pmd_trans_splitting(*pmd))) {
1146                         spin_unlock(&vma->vm_mm->page_table_lock);
1147                         wait_split_huge_page(vma->anon_vma, pmd);
1148                         return -1;
1149                 } else {
1150                         /* Thp mapped by 'pmd' is stable, so we can
1151                          * handle it as it is. */
1152                         return 1;
1153                 }
1154         }
1155         spin_unlock(&vma->vm_mm->page_table_lock);
1156         return 0;
1157 }
1158
1159 pmd_t *page_check_address_pmd(struct page *page,
1160                               struct mm_struct *mm,
1161                               unsigned long address,
1162                               enum page_check_address_pmd_flag flag)
1163 {
1164         pgd_t *pgd;
1165         pud_t *pud;
1166         pmd_t *pmd, *ret = NULL;
1167
1168         if (address & ~HPAGE_PMD_MASK)
1169                 goto out;
1170
1171         pgd = pgd_offset(mm, address);
1172         if (!pgd_present(*pgd))
1173                 goto out;
1174
1175         pud = pud_offset(pgd, address);
1176         if (!pud_present(*pud))
1177                 goto out;
1178
1179         pmd = pmd_offset(pud, address);
1180         if (pmd_none(*pmd))
1181                 goto out;
1182         if (pmd_page(*pmd) != page)
1183                 goto out;
1184         /*
1185          * split_vma() may create temporary aliased mappings. There is
1186          * no risk as long as all huge pmd are found and have their
1187          * splitting bit set before __split_huge_page_refcount
1188          * runs. Finding the same huge pmd more than once during the
1189          * same rmap walk is not a problem.
1190          */
1191         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1192             pmd_trans_splitting(*pmd))
1193                 goto out;
1194         if (pmd_trans_huge(*pmd)) {
1195                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1196                           !pmd_trans_splitting(*pmd));
1197                 ret = pmd;
1198         }
1199 out:
1200         return ret;
1201 }
1202
1203 static int __split_huge_page_splitting(struct page *page,
1204                                        struct vm_area_struct *vma,
1205                                        unsigned long address)
1206 {
1207         struct mm_struct *mm = vma->vm_mm;
1208         pmd_t *pmd;
1209         int ret = 0;
1210
1211         spin_lock(&mm->page_table_lock);
1212         pmd = page_check_address_pmd(page, mm, address,
1213                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1214         if (pmd) {
1215                 /*
1216                  * We can't temporarily set the pmd to null in order
1217                  * to split it, the pmd must remain marked huge at all
1218                  * times or the VM won't take the pmd_trans_huge paths
1219                  * and it won't wait on the anon_vma->root->mutex to
1220                  * serialize against split_huge_page*.
1221                  */
1222                 pmdp_splitting_flush_notify(vma, address, pmd);
1223                 ret = 1;
1224         }
1225         spin_unlock(&mm->page_table_lock);
1226
1227         return ret;
1228 }
1229
1230 static void __split_huge_page_refcount(struct page *page)
1231 {
1232         int i;
1233         struct zone *zone = page_zone(page);
1234         struct lruvec *lruvec;
1235         int tail_count = 0;
1236
1237         /* prevent PageLRU to go away from under us, and freeze lru stats */
1238         spin_lock_irq(&zone->lru_lock);
1239         lruvec = mem_cgroup_page_lruvec(page, zone);
1240
1241         compound_lock(page);
1242         /* complete memcg works before add pages to LRU */
1243         mem_cgroup_split_huge_fixup(page);
1244
1245         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1246                 struct page *page_tail = page + i;
1247
1248                 /* tail_page->_mapcount cannot change */
1249                 BUG_ON(page_mapcount(page_tail) < 0);
1250                 tail_count += page_mapcount(page_tail);
1251                 /* check for overflow */
1252                 BUG_ON(tail_count < 0);
1253                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1254                 /*
1255                  * tail_page->_count is zero and not changing from
1256                  * under us. But get_page_unless_zero() may be running
1257                  * from under us on the tail_page. If we used
1258                  * atomic_set() below instead of atomic_add(), we
1259                  * would then run atomic_set() concurrently with
1260                  * get_page_unless_zero(), and atomic_set() is
1261                  * implemented in C not using locked ops. spin_unlock
1262                  * on x86 sometime uses locked ops because of PPro
1263                  * errata 66, 92, so unless somebody can guarantee
1264                  * atomic_set() here would be safe on all archs (and
1265                  * not only on x86), it's safer to use atomic_add().
1266                  */
1267                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1268                            &page_tail->_count);
1269
1270                 /* after clearing PageTail the gup refcount can be released */
1271                 smp_mb();
1272
1273                 /*
1274                  * retain hwpoison flag of the poisoned tail page:
1275                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1276                  *   by the memory-failure.
1277                  */
1278                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1279                 page_tail->flags |= (page->flags &
1280                                      ((1L << PG_referenced) |
1281                                       (1L << PG_swapbacked) |
1282                                       (1L << PG_mlocked) |
1283                                       (1L << PG_uptodate)));
1284                 page_tail->flags |= (1L << PG_dirty);
1285
1286                 /* clear PageTail before overwriting first_page */
1287                 smp_wmb();
1288
1289                 /*
1290                  * __split_huge_page_splitting() already set the
1291                  * splitting bit in all pmd that could map this
1292                  * hugepage, that will ensure no CPU can alter the
1293                  * mapcount on the head page. The mapcount is only
1294                  * accounted in the head page and it has to be
1295                  * transferred to all tail pages in the below code. So
1296                  * for this code to be safe, the split the mapcount
1297                  * can't change. But that doesn't mean userland can't
1298                  * keep changing and reading the page contents while
1299                  * we transfer the mapcount, so the pmd splitting
1300                  * status is achieved setting a reserved bit in the
1301                  * pmd, not by clearing the present bit.
1302                 */
1303                 page_tail->_mapcount = page->_mapcount;
1304
1305                 BUG_ON(page_tail->mapping);
1306                 page_tail->mapping = page->mapping;
1307
1308                 page_tail->index = page->index + i;
1309
1310                 BUG_ON(!PageAnon(page_tail));
1311                 BUG_ON(!PageUptodate(page_tail));
1312                 BUG_ON(!PageDirty(page_tail));
1313                 BUG_ON(!PageSwapBacked(page_tail));
1314
1315                 lru_add_page_tail(page, page_tail, lruvec);
1316         }
1317         atomic_sub(tail_count, &page->_count);
1318         BUG_ON(atomic_read(&page->_count) <= 0);
1319
1320         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1321         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1322
1323         ClearPageCompound(page);
1324         compound_unlock(page);
1325         spin_unlock_irq(&zone->lru_lock);
1326
1327         for (i = 1; i < HPAGE_PMD_NR; i++) {
1328                 struct page *page_tail = page + i;
1329                 BUG_ON(page_count(page_tail) <= 0);
1330                 /*
1331                  * Tail pages may be freed if there wasn't any mapping
1332                  * like if add_to_swap() is running on a lru page that
1333                  * had its mapping zapped. And freeing these pages
1334                  * requires taking the lru_lock so we do the put_page
1335                  * of the tail pages after the split is complete.
1336                  */
1337                 put_page(page_tail);
1338         }
1339
1340         /*
1341          * Only the head page (now become a regular page) is required
1342          * to be pinned by the caller.
1343          */
1344         BUG_ON(page_count(page) <= 0);
1345 }
1346
1347 static int __split_huge_page_map(struct page *page,
1348                                  struct vm_area_struct *vma,
1349                                  unsigned long address)
1350 {
1351         struct mm_struct *mm = vma->vm_mm;
1352         pmd_t *pmd, _pmd;
1353         int ret = 0, i;
1354         pgtable_t pgtable;
1355         unsigned long haddr;
1356
1357         spin_lock(&mm->page_table_lock);
1358         pmd = page_check_address_pmd(page, mm, address,
1359                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1360         if (pmd) {
1361                 pgtable = get_pmd_huge_pte(mm);
1362                 pmd_populate(mm, &_pmd, pgtable);
1363
1364                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1365                      i++, haddr += PAGE_SIZE) {
1366                         pte_t *pte, entry;
1367                         BUG_ON(PageCompound(page+i));
1368                         entry = mk_pte(page + i, vma->vm_page_prot);
1369                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1370                         if (!pmd_write(*pmd))
1371                                 entry = pte_wrprotect(entry);
1372                         else
1373                                 BUG_ON(page_mapcount(page) != 1);
1374                         if (!pmd_young(*pmd))
1375                                 entry = pte_mkold(entry);
1376                         pte = pte_offset_map(&_pmd, haddr);
1377                         BUG_ON(!pte_none(*pte));
1378                         set_pte_at(mm, haddr, pte, entry);
1379                         pte_unmap(pte);
1380                 }
1381
1382                 smp_wmb(); /* make pte visible before pmd */
1383                 /*
1384                  * Up to this point the pmd is present and huge and
1385                  * userland has the whole access to the hugepage
1386                  * during the split (which happens in place). If we
1387                  * overwrite the pmd with the not-huge version
1388                  * pointing to the pte here (which of course we could
1389                  * if all CPUs were bug free), userland could trigger
1390                  * a small page size TLB miss on the small sized TLB
1391                  * while the hugepage TLB entry is still established
1392                  * in the huge TLB. Some CPU doesn't like that. See
1393                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1394                  * Erratum 383 on page 93. Intel should be safe but is
1395                  * also warns that it's only safe if the permission
1396                  * and cache attributes of the two entries loaded in
1397                  * the two TLB is identical (which should be the case
1398                  * here). But it is generally safer to never allow
1399                  * small and huge TLB entries for the same virtual
1400                  * address to be loaded simultaneously. So instead of
1401                  * doing "pmd_populate(); flush_tlb_range();" we first
1402                  * mark the current pmd notpresent (atomically because
1403                  * here the pmd_trans_huge and pmd_trans_splitting
1404                  * must remain set at all times on the pmd until the
1405                  * split is complete for this pmd), then we flush the
1406                  * SMP TLB and finally we write the non-huge version
1407                  * of the pmd entry with pmd_populate.
1408                  */
1409                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1410                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1411                 pmd_populate(mm, pmd, pgtable);
1412                 ret = 1;
1413         }
1414         spin_unlock(&mm->page_table_lock);
1415
1416         return ret;
1417 }
1418
1419 /* must be called with anon_vma->root->mutex hold */
1420 static void __split_huge_page(struct page *page,
1421                               struct anon_vma *anon_vma)
1422 {
1423         int mapcount, mapcount2;
1424         struct anon_vma_chain *avc;
1425
1426         BUG_ON(!PageHead(page));
1427         BUG_ON(PageTail(page));
1428
1429         mapcount = 0;
1430         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1431                 struct vm_area_struct *vma = avc->vma;
1432                 unsigned long addr = vma_address(page, vma);
1433                 BUG_ON(is_vma_temporary_stack(vma));
1434                 if (addr == -EFAULT)
1435                         continue;
1436                 mapcount += __split_huge_page_splitting(page, vma, addr);
1437         }
1438         /*
1439          * It is critical that new vmas are added to the tail of the
1440          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1441          * and establishes a child pmd before
1442          * __split_huge_page_splitting() freezes the parent pmd (so if
1443          * we fail to prevent copy_huge_pmd() from running until the
1444          * whole __split_huge_page() is complete), we will still see
1445          * the newly established pmd of the child later during the
1446          * walk, to be able to set it as pmd_trans_splitting too.
1447          */
1448         if (mapcount != page_mapcount(page))
1449                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1450                        mapcount, page_mapcount(page));
1451         BUG_ON(mapcount != page_mapcount(page));
1452
1453         __split_huge_page_refcount(page);
1454
1455         mapcount2 = 0;
1456         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1457                 struct vm_area_struct *vma = avc->vma;
1458                 unsigned long addr = vma_address(page, vma);
1459                 BUG_ON(is_vma_temporary_stack(vma));
1460                 if (addr == -EFAULT)
1461                         continue;
1462                 mapcount2 += __split_huge_page_map(page, vma, addr);
1463         }
1464         if (mapcount != mapcount2)
1465                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1466                        mapcount, mapcount2, page_mapcount(page));
1467         BUG_ON(mapcount != mapcount2);
1468 }
1469
1470 int split_huge_page(struct page *page)
1471 {
1472         struct anon_vma *anon_vma;
1473         int ret = 1;
1474
1475         BUG_ON(!PageAnon(page));
1476         anon_vma = page_lock_anon_vma(page);
1477         if (!anon_vma)
1478                 goto out;
1479         ret = 0;
1480         if (!PageCompound(page))
1481                 goto out_unlock;
1482
1483         BUG_ON(!PageSwapBacked(page));
1484         __split_huge_page(page, anon_vma);
1485         count_vm_event(THP_SPLIT);
1486
1487         BUG_ON(PageCompound(page));
1488 out_unlock:
1489         page_unlock_anon_vma(anon_vma);
1490 out:
1491         return ret;
1492 }
1493
1494 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1495                    VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1496
1497 int hugepage_madvise(struct vm_area_struct *vma,
1498                      unsigned long *vm_flags, int advice)
1499 {
1500         switch (advice) {
1501         case MADV_HUGEPAGE:
1502                 /*
1503                  * Be somewhat over-protective like KSM for now!
1504                  */
1505                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1506                         return -EINVAL;
1507                 *vm_flags &= ~VM_NOHUGEPAGE;
1508                 *vm_flags |= VM_HUGEPAGE;
1509                 /*
1510                  * If the vma become good for khugepaged to scan,
1511                  * register it here without waiting a page fault that
1512                  * may not happen any time soon.
1513                  */
1514                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1515                         return -ENOMEM;
1516                 break;
1517         case MADV_NOHUGEPAGE:
1518                 /*
1519                  * Be somewhat over-protective like KSM for now!
1520                  */
1521                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1522                         return -EINVAL;
1523                 *vm_flags &= ~VM_HUGEPAGE;
1524                 *vm_flags |= VM_NOHUGEPAGE;
1525                 /*
1526                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1527                  * this vma even if we leave the mm registered in khugepaged if
1528                  * it got registered before VM_NOHUGEPAGE was set.
1529                  */
1530                 break;
1531         }
1532
1533         return 0;
1534 }
1535
1536 static int __init khugepaged_slab_init(void)
1537 {
1538         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1539                                           sizeof(struct mm_slot),
1540                                           __alignof__(struct mm_slot), 0, NULL);
1541         if (!mm_slot_cache)
1542                 return -ENOMEM;
1543
1544         return 0;
1545 }
1546
1547 static void __init khugepaged_slab_free(void)
1548 {
1549         kmem_cache_destroy(mm_slot_cache);
1550         mm_slot_cache = NULL;
1551 }
1552
1553 static inline struct mm_slot *alloc_mm_slot(void)
1554 {
1555         if (!mm_slot_cache)     /* initialization failed */
1556                 return NULL;
1557         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1558 }
1559
1560 static inline void free_mm_slot(struct mm_slot *mm_slot)
1561 {
1562         kmem_cache_free(mm_slot_cache, mm_slot);
1563 }
1564
1565 static int __init mm_slots_hash_init(void)
1566 {
1567         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1568                                 GFP_KERNEL);
1569         if (!mm_slots_hash)
1570                 return -ENOMEM;
1571         return 0;
1572 }
1573
1574 #if 0
1575 static void __init mm_slots_hash_free(void)
1576 {
1577         kfree(mm_slots_hash);
1578         mm_slots_hash = NULL;
1579 }
1580 #endif
1581
1582 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1583 {
1584         struct mm_slot *mm_slot;
1585         struct hlist_head *bucket;
1586         struct hlist_node *node;
1587
1588         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1589                                 % MM_SLOTS_HASH_HEADS];
1590         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1591                 if (mm == mm_slot->mm)
1592                         return mm_slot;
1593         }
1594         return NULL;
1595 }
1596
1597 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1598                                     struct mm_slot *mm_slot)
1599 {
1600         struct hlist_head *bucket;
1601
1602         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1603                                 % MM_SLOTS_HASH_HEADS];
1604         mm_slot->mm = mm;
1605         hlist_add_head(&mm_slot->hash, bucket);
1606 }
1607
1608 static inline int khugepaged_test_exit(struct mm_struct *mm)
1609 {
1610         return atomic_read(&mm->mm_users) == 0;
1611 }
1612
1613 int __khugepaged_enter(struct mm_struct *mm)
1614 {
1615         struct mm_slot *mm_slot;
1616         int wakeup;
1617
1618         mm_slot = alloc_mm_slot();
1619         if (!mm_slot)
1620                 return -ENOMEM;
1621
1622         /* __khugepaged_exit() must not run from under us */
1623         VM_BUG_ON(khugepaged_test_exit(mm));
1624         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1625                 free_mm_slot(mm_slot);
1626                 return 0;
1627         }
1628
1629         spin_lock(&khugepaged_mm_lock);
1630         insert_to_mm_slots_hash(mm, mm_slot);
1631         /*
1632          * Insert just behind the scanning cursor, to let the area settle
1633          * down a little.
1634          */
1635         wakeup = list_empty(&khugepaged_scan.mm_head);
1636         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1637         spin_unlock(&khugepaged_mm_lock);
1638
1639         atomic_inc(&mm->mm_count);
1640         if (wakeup)
1641                 wake_up_interruptible(&khugepaged_wait);
1642
1643         return 0;
1644 }
1645
1646 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1647 {
1648         unsigned long hstart, hend;
1649         if (!vma->anon_vma)
1650                 /*
1651                  * Not yet faulted in so we will register later in the
1652                  * page fault if needed.
1653                  */
1654                 return 0;
1655         if (vma->vm_ops)
1656                 /* khugepaged not yet working on file or special mappings */
1657                 return 0;
1658         /*
1659          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1660          * true too, verify it here.
1661          */
1662         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1663         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1664         hend = vma->vm_end & HPAGE_PMD_MASK;
1665         if (hstart < hend)
1666                 return khugepaged_enter(vma);
1667         return 0;
1668 }
1669
1670 void __khugepaged_exit(struct mm_struct *mm)
1671 {
1672         struct mm_slot *mm_slot;
1673         int free = 0;
1674
1675         spin_lock(&khugepaged_mm_lock);
1676         mm_slot = get_mm_slot(mm);
1677         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1678                 hlist_del(&mm_slot->hash);
1679                 list_del(&mm_slot->mm_node);
1680                 free = 1;
1681         }
1682         spin_unlock(&khugepaged_mm_lock);
1683
1684         if (free) {
1685                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1686                 free_mm_slot(mm_slot);
1687                 mmdrop(mm);
1688         } else if (mm_slot) {
1689                 /*
1690                  * This is required to serialize against
1691                  * khugepaged_test_exit() (which is guaranteed to run
1692                  * under mmap sem read mode). Stop here (after we
1693                  * return all pagetables will be destroyed) until
1694                  * khugepaged has finished working on the pagetables
1695                  * under the mmap_sem.
1696                  */
1697                 down_write(&mm->mmap_sem);
1698                 up_write(&mm->mmap_sem);
1699         }
1700 }
1701
1702 static void release_pte_page(struct page *page)
1703 {
1704         /* 0 stands for page_is_file_cache(page) == false */
1705         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1706         unlock_page(page);
1707         putback_lru_page(page);
1708 }
1709
1710 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1711 {
1712         while (--_pte >= pte) {
1713                 pte_t pteval = *_pte;
1714                 if (!pte_none(pteval))
1715                         release_pte_page(pte_page(pteval));
1716         }
1717 }
1718
1719 static void release_all_pte_pages(pte_t *pte)
1720 {
1721         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1722 }
1723
1724 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1725                                         unsigned long address,
1726                                         pte_t *pte)
1727 {
1728         struct page *page;
1729         pte_t *_pte;
1730         int referenced = 0, isolated = 0, none = 0;
1731         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1732              _pte++, address += PAGE_SIZE) {
1733                 pte_t pteval = *_pte;
1734                 if (pte_none(pteval)) {
1735                         if (++none <= khugepaged_max_ptes_none)
1736                                 continue;
1737                         else {
1738                                 release_pte_pages(pte, _pte);
1739                                 goto out;
1740                         }
1741                 }
1742                 if (!pte_present(pteval) || !pte_write(pteval)) {
1743                         release_pte_pages(pte, _pte);
1744                         goto out;
1745                 }
1746                 page = vm_normal_page(vma, address, pteval);
1747                 if (unlikely(!page)) {
1748                         release_pte_pages(pte, _pte);
1749                         goto out;
1750                 }
1751                 VM_BUG_ON(PageCompound(page));
1752                 BUG_ON(!PageAnon(page));
1753                 VM_BUG_ON(!PageSwapBacked(page));
1754
1755                 /* cannot use mapcount: can't collapse if there's a gup pin */
1756                 if (page_count(page) != 1) {
1757                         release_pte_pages(pte, _pte);
1758                         goto out;
1759                 }
1760                 /*
1761                  * We can do it before isolate_lru_page because the
1762                  * page can't be freed from under us. NOTE: PG_lock
1763                  * is needed to serialize against split_huge_page
1764                  * when invoked from the VM.
1765                  */
1766                 if (!trylock_page(page)) {
1767                         release_pte_pages(pte, _pte);
1768                         goto out;
1769                 }
1770                 /*
1771                  * Isolate the page to avoid collapsing an hugepage
1772                  * currently in use by the VM.
1773                  */
1774                 if (isolate_lru_page(page)) {
1775                         unlock_page(page);
1776                         release_pte_pages(pte, _pte);
1777                         goto out;
1778                 }
1779                 /* 0 stands for page_is_file_cache(page) == false */
1780                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1781                 VM_BUG_ON(!PageLocked(page));
1782                 VM_BUG_ON(PageLRU(page));
1783
1784                 /* If there is no mapped pte young don't collapse the page */
1785                 if (pte_young(pteval) || PageReferenced(page) ||
1786                     mmu_notifier_test_young(vma->vm_mm, address))
1787                         referenced = 1;
1788         }
1789         if (unlikely(!referenced))
1790                 release_all_pte_pages(pte);
1791         else
1792                 isolated = 1;
1793 out:
1794         return isolated;
1795 }
1796
1797 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1798                                       struct vm_area_struct *vma,
1799                                       unsigned long address,
1800                                       spinlock_t *ptl)
1801 {
1802         pte_t *_pte;
1803         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1804                 pte_t pteval = *_pte;
1805                 struct page *src_page;
1806
1807                 if (pte_none(pteval)) {
1808                         clear_user_highpage(page, address);
1809                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1810                 } else {
1811                         src_page = pte_page(pteval);
1812                         copy_user_highpage(page, src_page, address, vma);
1813                         VM_BUG_ON(page_mapcount(src_page) != 1);
1814                         release_pte_page(src_page);
1815                         /*
1816                          * ptl mostly unnecessary, but preempt has to
1817                          * be disabled to update the per-cpu stats
1818                          * inside page_remove_rmap().
1819                          */
1820                         spin_lock(ptl);
1821                         /*
1822                          * paravirt calls inside pte_clear here are
1823                          * superfluous.
1824                          */
1825                         pte_clear(vma->vm_mm, address, _pte);
1826                         page_remove_rmap(src_page);
1827                         spin_unlock(ptl);
1828                         free_page_and_swap_cache(src_page);
1829                 }
1830
1831                 address += PAGE_SIZE;
1832                 page++;
1833         }
1834 }
1835
1836 static void collapse_huge_page(struct mm_struct *mm,
1837                                unsigned long address,
1838                                struct page **hpage,
1839                                struct vm_area_struct *vma,
1840                                int node)
1841 {
1842         pgd_t *pgd;
1843         pud_t *pud;
1844         pmd_t *pmd, _pmd;
1845         pte_t *pte;
1846         pgtable_t pgtable;
1847         struct page *new_page;
1848         spinlock_t *ptl;
1849         int isolated;
1850         unsigned long hstart, hend;
1851
1852         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1853 #ifndef CONFIG_NUMA
1854         up_read(&mm->mmap_sem);
1855         VM_BUG_ON(!*hpage);
1856         new_page = *hpage;
1857 #else
1858         VM_BUG_ON(*hpage);
1859         /*
1860          * Allocate the page while the vma is still valid and under
1861          * the mmap_sem read mode so there is no memory allocation
1862          * later when we take the mmap_sem in write mode. This is more
1863          * friendly behavior (OTOH it may actually hide bugs) to
1864          * filesystems in userland with daemons allocating memory in
1865          * the userland I/O paths.  Allocating memory with the
1866          * mmap_sem in read mode is good idea also to allow greater
1867          * scalability.
1868          */
1869         new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1870                                       node, __GFP_OTHER_NODE);
1871
1872         /*
1873          * After allocating the hugepage, release the mmap_sem read lock in
1874          * preparation for taking it in write mode.
1875          */
1876         up_read(&mm->mmap_sem);
1877         if (unlikely(!new_page)) {
1878                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1879                 *hpage = ERR_PTR(-ENOMEM);
1880                 return;
1881         }
1882 #endif
1883
1884         count_vm_event(THP_COLLAPSE_ALLOC);
1885         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1886 #ifdef CONFIG_NUMA
1887                 put_page(new_page);
1888 #endif
1889                 return;
1890         }
1891
1892         /*
1893          * Prevent all access to pagetables with the exception of
1894          * gup_fast later hanlded by the ptep_clear_flush and the VM
1895          * handled by the anon_vma lock + PG_lock.
1896          */
1897         down_write(&mm->mmap_sem);
1898         if (unlikely(khugepaged_test_exit(mm)))
1899                 goto out;
1900
1901         vma = find_vma(mm, address);
1902         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1903         hend = vma->vm_end & HPAGE_PMD_MASK;
1904         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1905                 goto out;
1906
1907         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1908             (vma->vm_flags & VM_NOHUGEPAGE))
1909                 goto out;
1910
1911         if (!vma->anon_vma || vma->vm_ops)
1912                 goto out;
1913         if (is_vma_temporary_stack(vma))
1914                 goto out;
1915         /*
1916          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1917          * true too, verify it here.
1918          */
1919         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1920
1921         pgd = pgd_offset(mm, address);
1922         if (!pgd_present(*pgd))
1923                 goto out;
1924
1925         pud = pud_offset(pgd, address);
1926         if (!pud_present(*pud))
1927                 goto out;
1928
1929         pmd = pmd_offset(pud, address);
1930         /* pmd can't go away or become huge under us */
1931         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1932                 goto out;
1933
1934         anon_vma_lock(vma->anon_vma);
1935
1936         pte = pte_offset_map(pmd, address);
1937         ptl = pte_lockptr(mm, pmd);
1938
1939         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1940         /*
1941          * After this gup_fast can't run anymore. This also removes
1942          * any huge TLB entry from the CPU so we won't allow
1943          * huge and small TLB entries for the same virtual address
1944          * to avoid the risk of CPU bugs in that area.
1945          */
1946         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1947         spin_unlock(&mm->page_table_lock);
1948
1949         spin_lock(ptl);
1950         isolated = __collapse_huge_page_isolate(vma, address, pte);
1951         spin_unlock(ptl);
1952
1953         if (unlikely(!isolated)) {
1954                 pte_unmap(pte);
1955                 spin_lock(&mm->page_table_lock);
1956                 BUG_ON(!pmd_none(*pmd));
1957                 set_pmd_at(mm, address, pmd, _pmd);
1958                 spin_unlock(&mm->page_table_lock);
1959                 anon_vma_unlock(vma->anon_vma);
1960                 goto out;
1961         }
1962
1963         /*
1964          * All pages are isolated and locked so anon_vma rmap
1965          * can't run anymore.
1966          */
1967         anon_vma_unlock(vma->anon_vma);
1968
1969         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1970         pte_unmap(pte);
1971         __SetPageUptodate(new_page);
1972         pgtable = pmd_pgtable(_pmd);
1973         VM_BUG_ON(page_count(pgtable) != 1);
1974         VM_BUG_ON(page_mapcount(pgtable) != 0);
1975
1976         _pmd = mk_pmd(new_page, vma->vm_page_prot);
1977         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1978         _pmd = pmd_mkhuge(_pmd);
1979
1980         /*
1981          * spin_lock() below is not the equivalent of smp_wmb(), so
1982          * this is needed to avoid the copy_huge_page writes to become
1983          * visible after the set_pmd_at() write.
1984          */
1985         smp_wmb();
1986
1987         spin_lock(&mm->page_table_lock);
1988         BUG_ON(!pmd_none(*pmd));
1989         page_add_new_anon_rmap(new_page, vma, address);
1990         set_pmd_at(mm, address, pmd, _pmd);
1991         update_mmu_cache(vma, address, _pmd);
1992         prepare_pmd_huge_pte(pgtable, mm);
1993         spin_unlock(&mm->page_table_lock);
1994
1995 #ifndef CONFIG_NUMA
1996         *hpage = NULL;
1997 #endif
1998         khugepaged_pages_collapsed++;
1999 out_up_write:
2000         up_write(&mm->mmap_sem);
2001         return;
2002
2003 out:
2004         mem_cgroup_uncharge_page(new_page);
2005 #ifdef CONFIG_NUMA
2006         put_page(new_page);
2007 #endif
2008         goto out_up_write;
2009 }
2010
2011 static int khugepaged_scan_pmd(struct mm_struct *mm,
2012                                struct vm_area_struct *vma,
2013                                unsigned long address,
2014                                struct page **hpage)
2015 {
2016         pgd_t *pgd;
2017         pud_t *pud;
2018         pmd_t *pmd;
2019         pte_t *pte, *_pte;
2020         int ret = 0, referenced = 0, none = 0;
2021         struct page *page;
2022         unsigned long _address;
2023         spinlock_t *ptl;
2024         int node = -1;
2025
2026         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2027
2028         pgd = pgd_offset(mm, address);
2029         if (!pgd_present(*pgd))
2030                 goto out;
2031
2032         pud = pud_offset(pgd, address);
2033         if (!pud_present(*pud))
2034                 goto out;
2035
2036         pmd = pmd_offset(pud, address);
2037         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2038                 goto out;
2039
2040         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2041         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2042              _pte++, _address += PAGE_SIZE) {
2043                 pte_t pteval = *_pte;
2044                 if (pte_none(pteval)) {
2045                         if (++none <= khugepaged_max_ptes_none)
2046                                 continue;
2047                         else
2048                                 goto out_unmap;
2049                 }
2050                 if (!pte_present(pteval) || !pte_write(pteval))
2051                         goto out_unmap;
2052                 page = vm_normal_page(vma, _address, pteval);
2053                 if (unlikely(!page))
2054                         goto out_unmap;
2055                 /*
2056                  * Chose the node of the first page. This could
2057                  * be more sophisticated and look at more pages,
2058                  * but isn't for now.
2059                  */
2060                 if (node == -1)
2061                         node = page_to_nid(page);
2062                 VM_BUG_ON(PageCompound(page));
2063                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2064                         goto out_unmap;
2065                 /* cannot use mapcount: can't collapse if there's a gup pin */
2066                 if (page_count(page) != 1)
2067                         goto out_unmap;
2068                 if (pte_young(pteval) || PageReferenced(page) ||
2069                     mmu_notifier_test_young(vma->vm_mm, address))
2070                         referenced = 1;
2071         }
2072         if (referenced)
2073                 ret = 1;
2074 out_unmap:
2075         pte_unmap_unlock(pte, ptl);
2076         if (ret)
2077                 /* collapse_huge_page will return with the mmap_sem released */
2078                 collapse_huge_page(mm, address, hpage, vma, node);
2079 out:
2080         return ret;
2081 }
2082
2083 static void collect_mm_slot(struct mm_slot *mm_slot)
2084 {
2085         struct mm_struct *mm = mm_slot->mm;
2086
2087         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2088
2089         if (khugepaged_test_exit(mm)) {
2090                 /* free mm_slot */
2091                 hlist_del(&mm_slot->hash);
2092                 list_del(&mm_slot->mm_node);
2093
2094                 /*
2095                  * Not strictly needed because the mm exited already.
2096                  *
2097                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2098                  */
2099
2100                 /* khugepaged_mm_lock actually not necessary for the below */
2101                 free_mm_slot(mm_slot);
2102                 mmdrop(mm);
2103         }
2104 }
2105
2106 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2107                                             struct page **hpage)
2108         __releases(&khugepaged_mm_lock)
2109         __acquires(&khugepaged_mm_lock)
2110 {
2111         struct mm_slot *mm_slot;
2112         struct mm_struct *mm;
2113         struct vm_area_struct *vma;
2114         int progress = 0;
2115
2116         VM_BUG_ON(!pages);
2117         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2118
2119         if (khugepaged_scan.mm_slot)
2120                 mm_slot = khugepaged_scan.mm_slot;
2121         else {
2122                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2123                                      struct mm_slot, mm_node);
2124                 khugepaged_scan.address = 0;
2125                 khugepaged_scan.mm_slot = mm_slot;
2126         }
2127         spin_unlock(&khugepaged_mm_lock);
2128
2129         mm = mm_slot->mm;
2130         down_read(&mm->mmap_sem);
2131         if (unlikely(khugepaged_test_exit(mm)))
2132                 vma = NULL;
2133         else
2134                 vma = find_vma(mm, khugepaged_scan.address);
2135
2136         progress++;
2137         for (; vma; vma = vma->vm_next) {
2138                 unsigned long hstart, hend;
2139
2140                 cond_resched();
2141                 if (unlikely(khugepaged_test_exit(mm))) {
2142                         progress++;
2143                         break;
2144                 }
2145
2146                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2147                      !khugepaged_always()) ||
2148                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2149                 skip:
2150                         progress++;
2151                         continue;
2152                 }
2153                 if (!vma->anon_vma || vma->vm_ops)
2154                         goto skip;
2155                 if (is_vma_temporary_stack(vma))
2156                         goto skip;
2157                 /*
2158                  * If is_pfn_mapping() is true is_learn_pfn_mapping()
2159                  * must be true too, verify it here.
2160                  */
2161                 VM_BUG_ON(is_linear_pfn_mapping(vma) ||
2162                           vma->vm_flags & VM_NO_THP);
2163
2164                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2165                 hend = vma->vm_end & HPAGE_PMD_MASK;
2166                 if (hstart >= hend)
2167                         goto skip;
2168                 if (khugepaged_scan.address > hend)
2169                         goto skip;
2170                 if (khugepaged_scan.address < hstart)
2171                         khugepaged_scan.address = hstart;
2172                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2173
2174                 while (khugepaged_scan.address < hend) {
2175                         int ret;
2176                         cond_resched();
2177                         if (unlikely(khugepaged_test_exit(mm)))
2178                                 goto breakouterloop;
2179
2180                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2181                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2182                                   hend);
2183                         ret = khugepaged_scan_pmd(mm, vma,
2184                                                   khugepaged_scan.address,
2185                                                   hpage);
2186                         /* move to next address */
2187                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2188                         progress += HPAGE_PMD_NR;
2189                         if (ret)
2190                                 /* we released mmap_sem so break loop */
2191                                 goto breakouterloop_mmap_sem;
2192                         if (progress >= pages)
2193                                 goto breakouterloop;
2194                 }
2195         }
2196 breakouterloop:
2197         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2198 breakouterloop_mmap_sem:
2199
2200         spin_lock(&khugepaged_mm_lock);
2201         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2202         /*
2203          * Release the current mm_slot if this mm is about to die, or
2204          * if we scanned all vmas of this mm.
2205          */
2206         if (khugepaged_test_exit(mm) || !vma) {
2207                 /*
2208                  * Make sure that if mm_users is reaching zero while
2209                  * khugepaged runs here, khugepaged_exit will find
2210                  * mm_slot not pointing to the exiting mm.
2211                  */
2212                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2213                         khugepaged_scan.mm_slot = list_entry(
2214                                 mm_slot->mm_node.next,
2215                                 struct mm_slot, mm_node);
2216                         khugepaged_scan.address = 0;
2217                 } else {
2218                         khugepaged_scan.mm_slot = NULL;
2219                         khugepaged_full_scans++;
2220                 }
2221
2222                 collect_mm_slot(mm_slot);
2223         }
2224
2225         return progress;
2226 }
2227
2228 static int khugepaged_has_work(void)
2229 {
2230         return !list_empty(&khugepaged_scan.mm_head) &&
2231                 khugepaged_enabled();
2232 }
2233
2234 static int khugepaged_wait_event(void)
2235 {
2236         return !list_empty(&khugepaged_scan.mm_head) ||
2237                 !khugepaged_enabled();
2238 }
2239
2240 static void khugepaged_do_scan(struct page **hpage)
2241 {
2242         unsigned int progress = 0, pass_through_head = 0;
2243         unsigned int pages = khugepaged_pages_to_scan;
2244
2245         barrier(); /* write khugepaged_pages_to_scan to local stack */
2246
2247         while (progress < pages) {
2248                 cond_resched();
2249
2250 #ifndef CONFIG_NUMA
2251                 if (!*hpage) {
2252                         *hpage = alloc_hugepage(khugepaged_defrag());
2253                         if (unlikely(!*hpage)) {
2254                                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2255                                 break;
2256                         }
2257                         count_vm_event(THP_COLLAPSE_ALLOC);
2258                 }
2259 #else
2260                 if (IS_ERR(*hpage))
2261                         break;
2262 #endif
2263
2264                 if (unlikely(kthread_should_stop() || freezing(current)))
2265                         break;
2266
2267                 spin_lock(&khugepaged_mm_lock);
2268                 if (!khugepaged_scan.mm_slot)
2269                         pass_through_head++;
2270                 if (khugepaged_has_work() &&
2271                     pass_through_head < 2)
2272                         progress += khugepaged_scan_mm_slot(pages - progress,
2273                                                             hpage);
2274                 else
2275                         progress = pages;
2276                 spin_unlock(&khugepaged_mm_lock);
2277         }
2278 }
2279
2280 static void khugepaged_alloc_sleep(void)
2281 {
2282         wait_event_freezable_timeout(khugepaged_wait, false,
2283                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2284 }
2285
2286 #ifndef CONFIG_NUMA
2287 static struct page *khugepaged_alloc_hugepage(void)
2288 {
2289         struct page *hpage;
2290
2291         do {
2292                 hpage = alloc_hugepage(khugepaged_defrag());
2293                 if (!hpage) {
2294                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2295                         khugepaged_alloc_sleep();
2296                 } else
2297                         count_vm_event(THP_COLLAPSE_ALLOC);
2298         } while (unlikely(!hpage) &&
2299                  likely(khugepaged_enabled()));
2300         return hpage;
2301 }
2302 #endif
2303
2304 static void khugepaged_loop(void)
2305 {
2306         struct page *hpage;
2307
2308 #ifdef CONFIG_NUMA
2309         hpage = NULL;
2310 #endif
2311         while (likely(khugepaged_enabled())) {
2312 #ifndef CONFIG_NUMA
2313                 hpage = khugepaged_alloc_hugepage();
2314                 if (unlikely(!hpage))
2315                         break;
2316 #else
2317                 if (IS_ERR(hpage)) {
2318                         khugepaged_alloc_sleep();
2319                         hpage = NULL;
2320                 }
2321 #endif
2322
2323                 khugepaged_do_scan(&hpage);
2324 #ifndef CONFIG_NUMA
2325                 if (hpage)
2326                         put_page(hpage);
2327 #endif
2328                 try_to_freeze();
2329                 if (unlikely(kthread_should_stop()))
2330                         break;
2331                 if (khugepaged_has_work()) {
2332                         if (!khugepaged_scan_sleep_millisecs)
2333                                 continue;
2334                         wait_event_freezable_timeout(khugepaged_wait, false,
2335                             msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2336                 } else if (khugepaged_enabled())
2337                         wait_event_freezable(khugepaged_wait,
2338                                              khugepaged_wait_event());
2339         }
2340 }
2341
2342 static int khugepaged(void *none)
2343 {
2344         struct mm_slot *mm_slot;
2345
2346         set_freezable();
2347         set_user_nice(current, 19);
2348
2349         /* serialize with start_khugepaged() */
2350         mutex_lock(&khugepaged_mutex);
2351
2352         for (;;) {
2353                 mutex_unlock(&khugepaged_mutex);
2354                 VM_BUG_ON(khugepaged_thread != current);
2355                 khugepaged_loop();
2356                 VM_BUG_ON(khugepaged_thread != current);
2357
2358                 mutex_lock(&khugepaged_mutex);
2359                 if (!khugepaged_enabled())
2360                         break;
2361                 if (unlikely(kthread_should_stop()))
2362                         break;
2363         }
2364
2365         spin_lock(&khugepaged_mm_lock);
2366         mm_slot = khugepaged_scan.mm_slot;
2367         khugepaged_scan.mm_slot = NULL;
2368         if (mm_slot)
2369                 collect_mm_slot(mm_slot);
2370         spin_unlock(&khugepaged_mm_lock);
2371
2372         khugepaged_thread = NULL;
2373         mutex_unlock(&khugepaged_mutex);
2374
2375         return 0;
2376 }
2377
2378 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2379 {
2380         struct page *page;
2381
2382         spin_lock(&mm->page_table_lock);
2383         if (unlikely(!pmd_trans_huge(*pmd))) {
2384                 spin_unlock(&mm->page_table_lock);
2385                 return;
2386         }
2387         page = pmd_page(*pmd);
2388         VM_BUG_ON(!page_count(page));
2389         get_page(page);
2390         spin_unlock(&mm->page_table_lock);
2391
2392         split_huge_page(page);
2393
2394         put_page(page);
2395         BUG_ON(pmd_trans_huge(*pmd));
2396 }
2397
2398 static void split_huge_page_address(struct mm_struct *mm,
2399                                     unsigned long address)
2400 {
2401         pgd_t *pgd;
2402         pud_t *pud;
2403         pmd_t *pmd;
2404
2405         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2406
2407         pgd = pgd_offset(mm, address);
2408         if (!pgd_present(*pgd))
2409                 return;
2410
2411         pud = pud_offset(pgd, address);
2412         if (!pud_present(*pud))
2413                 return;
2414
2415         pmd = pmd_offset(pud, address);
2416         if (!pmd_present(*pmd))
2417                 return;
2418         /*
2419          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2420          * materialize from under us.
2421          */
2422         split_huge_page_pmd(mm, pmd);
2423 }
2424
2425 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2426                              unsigned long start,
2427                              unsigned long end,
2428                              long adjust_next)
2429 {
2430         /*
2431          * If the new start address isn't hpage aligned and it could
2432          * previously contain an hugepage: check if we need to split
2433          * an huge pmd.
2434          */
2435         if (start & ~HPAGE_PMD_MASK &&
2436             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2437             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2438                 split_huge_page_address(vma->vm_mm, start);
2439
2440         /*
2441          * If the new end address isn't hpage aligned and it could
2442          * previously contain an hugepage: check if we need to split
2443          * an huge pmd.
2444          */
2445         if (end & ~HPAGE_PMD_MASK &&
2446             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2447             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2448                 split_huge_page_address(vma->vm_mm, end);
2449
2450         /*
2451          * If we're also updating the vma->vm_next->vm_start, if the new
2452          * vm_next->vm_start isn't page aligned and it could previously
2453          * contain an hugepage: check if we need to split an huge pmd.
2454          */
2455         if (adjust_next > 0) {
2456                 struct vm_area_struct *next = vma->vm_next;
2457                 unsigned long nstart = next->vm_start;
2458                 nstart += adjust_next << PAGE_SHIFT;
2459                 if (nstart & ~HPAGE_PMD_MASK &&
2460                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2461                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2462                         split_huge_page_address(next->vm_mm, nstart);
2463         }
2464 }