blob: 60836870c6f7015e127f4eaf88805f099951a721 [file] [log] [blame]
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001/*
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>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080015#include <linux/shrinker.h>
Andrea Arcangeliba761492011-01-13 15:46:58 -080016#include <linux/mm_inline.h>
17#include <linux/kthread.h>
18#include <linux/khugepaged.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080019#include <linux/freezer.h>
Andrea Arcangelia664b2d2011-01-13 15:47:17 -080020#include <linux/mman.h>
Ralf Baechle325adeb2012-10-15 13:44:56 +020021#include <linux/pagemap.h>
Mel Gorman4daae3b2012-11-02 11:33:45 +000022#include <linux/migrate.h>
Sasha Levin43b5fbb2013-02-22 16:32:27 -080023#include <linux/hashtable.h>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080024
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080025#include <asm/tlb.h>
26#include <asm/pgalloc.h>
27#include "internal.h"
28
Andrea Arcangeliba761492011-01-13 15:46:58 -080029/*
30 * By default transparent hugepage support is enabled for all mappings
31 * and khugepaged scans all mappings. Defrag is only invoked by
32 * khugepaged hugepage allocations and by page faults inside
33 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
34 * allocations.
35 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080036unsigned long transparent_hugepage_flags __read_mostly =
Andrea Arcangeli13ece882011-01-13 15:47:07 -080037#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
Andrea Arcangeliba761492011-01-13 15:46:58 -080038 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
Andrea Arcangeli13ece882011-01-13 15:47:07 -080039#endif
40#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
41 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
42#endif
Andrea Arcangelid39d33c2011-01-13 15:47:05 -080043 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
Kirill A. Shutemov79da5402012-12-12 13:51:12 -080044 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
45 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
Andrea Arcangeliba761492011-01-13 15:46:58 -080046
47/* default scan 8*512 pte (or vmas) every 30 second */
48static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
49static unsigned int khugepaged_pages_collapsed;
50static unsigned int khugepaged_full_scans;
51static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
52/* during fragmentation poll the hugepage allocator once every minute */
53static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
54static struct task_struct *khugepaged_thread __read_mostly;
55static DEFINE_MUTEX(khugepaged_mutex);
56static DEFINE_SPINLOCK(khugepaged_mm_lock);
57static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
58/*
59 * default collapse hugepages if there is at least one pte mapped like
60 * it would have happened if the vma was large enough during page
61 * fault.
62 */
63static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
64
65static int khugepaged(void *none);
Andrea Arcangeliba761492011-01-13 15:46:58 -080066static int khugepaged_slab_init(void);
Andrea Arcangeliba761492011-01-13 15:46:58 -080067
Sasha Levin43b5fbb2013-02-22 16:32:27 -080068#define MM_SLOTS_HASH_BITS 10
69static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
70
Andrea Arcangeliba761492011-01-13 15:46:58 -080071static struct kmem_cache *mm_slot_cache __read_mostly;
72
73/**
74 * struct mm_slot - hash lookup from mm to mm_slot
75 * @hash: hash collision list
76 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
77 * @mm: the mm that this information is valid for
78 */
79struct mm_slot {
80 struct hlist_node hash;
81 struct list_head mm_node;
82 struct mm_struct *mm;
83};
84
85/**
86 * struct khugepaged_scan - cursor for scanning
87 * @mm_head: the head of the mm list to scan
88 * @mm_slot: the current mm_slot we are scanning
89 * @address: the next address inside that to be scanned
90 *
91 * There is only the one khugepaged_scan instance of this cursor structure.
92 */
93struct khugepaged_scan {
94 struct list_head mm_head;
95 struct mm_slot *mm_slot;
96 unsigned long address;
H Hartley Sweeten2f1da642011-10-31 17:09:25 -070097};
98static struct khugepaged_scan khugepaged_scan = {
Andrea Arcangeliba761492011-01-13 15:46:58 -080099 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
100};
101
Andrea Arcangelif0005652011-01-13 15:47:04 -0800102
103static int set_recommended_min_free_kbytes(void)
104{
105 struct zone *zone;
106 int nr_zones = 0;
107 unsigned long recommended_min;
Andrea Arcangelif0005652011-01-13 15:47:04 -0800108
Xiao Guangrong17c230a2012-10-08 16:29:56 -0700109 if (!khugepaged_enabled())
Andrea Arcangelif0005652011-01-13 15:47:04 -0800110 return 0;
111
112 for_each_populated_zone(zone)
113 nr_zones++;
114
115 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
116 recommended_min = pageblock_nr_pages * nr_zones * 2;
117
118 /*
119 * Make sure that on average at least two pageblocks are almost free
120 * of another type, one for a migratetype to fall back to and a
121 * second to avoid subsequent fallbacks of other types There are 3
122 * MIGRATE_TYPES we care about.
123 */
124 recommended_min += pageblock_nr_pages * nr_zones *
125 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
126
127 /* don't ever allow to reserve more than 5% of the lowmem */
128 recommended_min = min(recommended_min,
129 (unsigned long) nr_free_buffer_pages() / 20);
130 recommended_min <<= (PAGE_SHIFT-10);
131
132 if (recommended_min > min_free_kbytes)
133 min_free_kbytes = recommended_min;
134 setup_per_zone_wmarks();
135 return 0;
136}
137late_initcall(set_recommended_min_free_kbytes);
138
Andrea Arcangeliba761492011-01-13 15:46:58 -0800139static int start_khugepaged(void)
140{
141 int err = 0;
142 if (khugepaged_enabled()) {
Andrea Arcangeliba761492011-01-13 15:46:58 -0800143 if (!khugepaged_thread)
144 khugepaged_thread = kthread_run(khugepaged, NULL,
145 "khugepaged");
146 if (unlikely(IS_ERR(khugepaged_thread))) {
147 printk(KERN_ERR
148 "khugepaged: kthread_run(khugepaged) failed\n");
149 err = PTR_ERR(khugepaged_thread);
150 khugepaged_thread = NULL;
151 }
Xiao Guangrong911891a2012-10-08 16:29:41 -0700152
153 if (!list_empty(&khugepaged_scan.mm_head))
Andrea Arcangeliba761492011-01-13 15:46:58 -0800154 wake_up_interruptible(&khugepaged_wait);
Andrea Arcangelif0005652011-01-13 15:47:04 -0800155
156 set_recommended_min_free_kbytes();
Xiao Guangrong911891a2012-10-08 16:29:41 -0700157 } else if (khugepaged_thread) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700158 kthread_stop(khugepaged_thread);
159 khugepaged_thread = NULL;
160 }
Xiao Guangrong637e3a22012-10-08 16:29:38 -0700161
Andrea Arcangeliba761492011-01-13 15:46:58 -0800162 return err;
163}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800164
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800165static atomic_t huge_zero_refcount;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700166static struct page *huge_zero_page __read_mostly;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800167
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700168static inline bool is_huge_zero_page(struct page *page)
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800169{
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700170 return ACCESS_ONCE(huge_zero_page) == page;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800171}
172
173static inline bool is_huge_zero_pmd(pmd_t pmd)
174{
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700175 return is_huge_zero_page(pmd_page(pmd));
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800176}
177
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700178static struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800179{
180 struct page *zero_page;
181retry:
182 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700183 return ACCESS_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800184
185 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
186 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800187 if (!zero_page) {
188 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700189 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800190 }
191 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800192 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700193 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800194 preempt_enable();
195 __free_page(zero_page);
196 goto retry;
197 }
198
199 /* We take additional reference here. It will be put back by shrinker */
200 atomic_set(&huge_zero_refcount, 2);
201 preempt_enable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700202 return ACCESS_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800203}
204
205static void put_huge_zero_page(void)
206{
207 /*
208 * Counter should never go to zero here. Only shrinker can put
209 * last reference.
210 */
211 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
212}
213
214static int shrink_huge_zero_page(struct shrinker *shrink,
215 struct shrink_control *sc)
216{
217 if (!sc->nr_to_scan)
218 /* we can free zero page only if last reference remains */
219 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
220
221 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700222 struct page *zero_page = xchg(&huge_zero_page, NULL);
223 BUG_ON(zero_page == NULL);
224 __free_page(zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800225 }
226
227 return 0;
228}
229
230static struct shrinker huge_zero_page_shrinker = {
231 .shrink = shrink_huge_zero_page,
232 .seeks = DEFAULT_SEEKS,
233};
234
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800235#ifdef CONFIG_SYSFS
Andrea Arcangeliba761492011-01-13 15:46:58 -0800236
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800237static ssize_t double_flag_show(struct kobject *kobj,
238 struct kobj_attribute *attr, char *buf,
239 enum transparent_hugepage_flag enabled,
240 enum transparent_hugepage_flag req_madv)
241{
242 if (test_bit(enabled, &transparent_hugepage_flags)) {
243 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
244 return sprintf(buf, "[always] madvise never\n");
245 } else if (test_bit(req_madv, &transparent_hugepage_flags))
246 return sprintf(buf, "always [madvise] never\n");
247 else
248 return sprintf(buf, "always madvise [never]\n");
249}
250static ssize_t double_flag_store(struct kobject *kobj,
251 struct kobj_attribute *attr,
252 const char *buf, size_t count,
253 enum transparent_hugepage_flag enabled,
254 enum transparent_hugepage_flag req_madv)
255{
256 if (!memcmp("always", buf,
257 min(sizeof("always")-1, count))) {
258 set_bit(enabled, &transparent_hugepage_flags);
259 clear_bit(req_madv, &transparent_hugepage_flags);
260 } else if (!memcmp("madvise", buf,
261 min(sizeof("madvise")-1, count))) {
262 clear_bit(enabled, &transparent_hugepage_flags);
263 set_bit(req_madv, &transparent_hugepage_flags);
264 } else if (!memcmp("never", buf,
265 min(sizeof("never")-1, count))) {
266 clear_bit(enabled, &transparent_hugepage_flags);
267 clear_bit(req_madv, &transparent_hugepage_flags);
268 } else
269 return -EINVAL;
270
271 return count;
272}
273
274static ssize_t enabled_show(struct kobject *kobj,
275 struct kobj_attribute *attr, char *buf)
276{
277 return double_flag_show(kobj, attr, buf,
278 TRANSPARENT_HUGEPAGE_FLAG,
279 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
280}
281static ssize_t enabled_store(struct kobject *kobj,
282 struct kobj_attribute *attr,
283 const char *buf, size_t count)
284{
Andrea Arcangeliba761492011-01-13 15:46:58 -0800285 ssize_t ret;
286
287 ret = double_flag_store(kobj, attr, buf, count,
288 TRANSPARENT_HUGEPAGE_FLAG,
289 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
290
291 if (ret > 0) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700292 int err;
293
294 mutex_lock(&khugepaged_mutex);
295 err = start_khugepaged();
296 mutex_unlock(&khugepaged_mutex);
297
Andrea Arcangeliba761492011-01-13 15:46:58 -0800298 if (err)
299 ret = err;
300 }
301
302 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800303}
304static struct kobj_attribute enabled_attr =
305 __ATTR(enabled, 0644, enabled_show, enabled_store);
306
307static ssize_t single_flag_show(struct kobject *kobj,
308 struct kobj_attribute *attr, char *buf,
309 enum transparent_hugepage_flag flag)
310{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700311 return sprintf(buf, "%d\n",
312 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800313}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700314
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800315static ssize_t single_flag_store(struct kobject *kobj,
316 struct kobj_attribute *attr,
317 const char *buf, size_t count,
318 enum transparent_hugepage_flag flag)
319{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700320 unsigned long value;
321 int ret;
322
323 ret = kstrtoul(buf, 10, &value);
324 if (ret < 0)
325 return ret;
326 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800327 return -EINVAL;
328
Ben Hutchingse27e6152011-04-14 15:22:21 -0700329 if (value)
330 set_bit(flag, &transparent_hugepage_flags);
331 else
332 clear_bit(flag, &transparent_hugepage_flags);
333
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800334 return count;
335}
336
337/*
338 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
339 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
340 * memory just to allocate one more hugepage.
341 */
342static ssize_t defrag_show(struct kobject *kobj,
343 struct kobj_attribute *attr, char *buf)
344{
345 return double_flag_show(kobj, attr, buf,
346 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
347 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
348}
349static ssize_t defrag_store(struct kobject *kobj,
350 struct kobj_attribute *attr,
351 const char *buf, size_t count)
352{
353 return double_flag_store(kobj, attr, buf, count,
354 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
355 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
356}
357static struct kobj_attribute defrag_attr =
358 __ATTR(defrag, 0644, defrag_show, defrag_store);
359
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800360static ssize_t use_zero_page_show(struct kobject *kobj,
361 struct kobj_attribute *attr, char *buf)
362{
363 return single_flag_show(kobj, attr, buf,
364 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
365}
366static ssize_t use_zero_page_store(struct kobject *kobj,
367 struct kobj_attribute *attr, const char *buf, size_t count)
368{
369 return single_flag_store(kobj, attr, buf, count,
370 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
371}
372static struct kobj_attribute use_zero_page_attr =
373 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800374#ifdef CONFIG_DEBUG_VM
375static ssize_t debug_cow_show(struct kobject *kobj,
376 struct kobj_attribute *attr, char *buf)
377{
378 return single_flag_show(kobj, attr, buf,
379 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
380}
381static ssize_t debug_cow_store(struct kobject *kobj,
382 struct kobj_attribute *attr,
383 const char *buf, size_t count)
384{
385 return single_flag_store(kobj, attr, buf, count,
386 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
387}
388static struct kobj_attribute debug_cow_attr =
389 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
390#endif /* CONFIG_DEBUG_VM */
391
392static struct attribute *hugepage_attr[] = {
393 &enabled_attr.attr,
394 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800395 &use_zero_page_attr.attr,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800396#ifdef CONFIG_DEBUG_VM
397 &debug_cow_attr.attr,
398#endif
399 NULL,
400};
401
402static struct attribute_group hugepage_attr_group = {
403 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800404};
405
406static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
407 struct kobj_attribute *attr,
408 char *buf)
409{
410 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
411}
412
413static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
414 struct kobj_attribute *attr,
415 const char *buf, size_t count)
416{
417 unsigned long msecs;
418 int err;
419
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700420 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800421 if (err || msecs > UINT_MAX)
422 return -EINVAL;
423
424 khugepaged_scan_sleep_millisecs = msecs;
425 wake_up_interruptible(&khugepaged_wait);
426
427 return count;
428}
429static struct kobj_attribute scan_sleep_millisecs_attr =
430 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
431 scan_sleep_millisecs_store);
432
433static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
434 struct kobj_attribute *attr,
435 char *buf)
436{
437 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
438}
439
440static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
441 struct kobj_attribute *attr,
442 const char *buf, size_t count)
443{
444 unsigned long msecs;
445 int err;
446
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700447 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800448 if (err || msecs > UINT_MAX)
449 return -EINVAL;
450
451 khugepaged_alloc_sleep_millisecs = msecs;
452 wake_up_interruptible(&khugepaged_wait);
453
454 return count;
455}
456static struct kobj_attribute alloc_sleep_millisecs_attr =
457 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
458 alloc_sleep_millisecs_store);
459
460static ssize_t pages_to_scan_show(struct kobject *kobj,
461 struct kobj_attribute *attr,
462 char *buf)
463{
464 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
465}
466static ssize_t pages_to_scan_store(struct kobject *kobj,
467 struct kobj_attribute *attr,
468 const char *buf, size_t count)
469{
470 int err;
471 unsigned long pages;
472
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700473 err = kstrtoul(buf, 10, &pages);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800474 if (err || !pages || pages > UINT_MAX)
475 return -EINVAL;
476
477 khugepaged_pages_to_scan = pages;
478
479 return count;
480}
481static struct kobj_attribute pages_to_scan_attr =
482 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
483 pages_to_scan_store);
484
485static ssize_t pages_collapsed_show(struct kobject *kobj,
486 struct kobj_attribute *attr,
487 char *buf)
488{
489 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
490}
491static struct kobj_attribute pages_collapsed_attr =
492 __ATTR_RO(pages_collapsed);
493
494static ssize_t full_scans_show(struct kobject *kobj,
495 struct kobj_attribute *attr,
496 char *buf)
497{
498 return sprintf(buf, "%u\n", khugepaged_full_scans);
499}
500static struct kobj_attribute full_scans_attr =
501 __ATTR_RO(full_scans);
502
503static ssize_t khugepaged_defrag_show(struct kobject *kobj,
504 struct kobj_attribute *attr, char *buf)
505{
506 return single_flag_show(kobj, attr, buf,
507 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
508}
509static ssize_t khugepaged_defrag_store(struct kobject *kobj,
510 struct kobj_attribute *attr,
511 const char *buf, size_t count)
512{
513 return single_flag_store(kobj, attr, buf, count,
514 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
515}
516static struct kobj_attribute khugepaged_defrag_attr =
517 __ATTR(defrag, 0644, khugepaged_defrag_show,
518 khugepaged_defrag_store);
519
520/*
521 * max_ptes_none controls if khugepaged should collapse hugepages over
522 * any unmapped ptes in turn potentially increasing the memory
523 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
524 * reduce the available free memory in the system as it
525 * runs. Increasing max_ptes_none will instead potentially reduce the
526 * free memory in the system during the khugepaged scan.
527 */
528static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
529 struct kobj_attribute *attr,
530 char *buf)
531{
532 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
533}
534static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
535 struct kobj_attribute *attr,
536 const char *buf, size_t count)
537{
538 int err;
539 unsigned long max_ptes_none;
540
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700541 err = kstrtoul(buf, 10, &max_ptes_none);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800542 if (err || max_ptes_none > HPAGE_PMD_NR-1)
543 return -EINVAL;
544
545 khugepaged_max_ptes_none = max_ptes_none;
546
547 return count;
548}
549static struct kobj_attribute khugepaged_max_ptes_none_attr =
550 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
551 khugepaged_max_ptes_none_store);
552
553static struct attribute *khugepaged_attr[] = {
554 &khugepaged_defrag_attr.attr,
555 &khugepaged_max_ptes_none_attr.attr,
556 &pages_to_scan_attr.attr,
557 &pages_collapsed_attr.attr,
558 &full_scans_attr.attr,
559 &scan_sleep_millisecs_attr.attr,
560 &alloc_sleep_millisecs_attr.attr,
561 NULL,
562};
563
564static struct attribute_group khugepaged_attr_group = {
565 .attrs = khugepaged_attr,
566 .name = "khugepaged",
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800567};
Shaohua Li569e5592012-01-12 17:19:11 -0800568
569static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
570{
571 int err;
572
573 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
574 if (unlikely(!*hugepage_kobj)) {
Jeremy Eder2c797372012-12-20 15:05:32 -0800575 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800576 return -ENOMEM;
577 }
578
579 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
580 if (err) {
Jeremy Eder2c797372012-12-20 15:05:32 -0800581 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800582 goto delete_obj;
583 }
584
585 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
586 if (err) {
Jeremy Eder2c797372012-12-20 15:05:32 -0800587 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800588 goto remove_hp_group;
589 }
590
591 return 0;
592
593remove_hp_group:
594 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
595delete_obj:
596 kobject_put(*hugepage_kobj);
597 return err;
598}
599
600static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
601{
602 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
603 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
604 kobject_put(hugepage_kobj);
605}
606#else
607static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
608{
609 return 0;
610}
611
612static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
613{
614}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800615#endif /* CONFIG_SYSFS */
616
617static int __init hugepage_init(void)
618{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800619 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800620 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800621
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800622 if (!has_transparent_hugepage()) {
623 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800624 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800625 }
626
Shaohua Li569e5592012-01-12 17:19:11 -0800627 err = hugepage_init_sysfs(&hugepage_kobj);
628 if (err)
629 return err;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800630
631 err = khugepaged_slab_init();
632 if (err)
633 goto out;
634
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800635 register_shrinker(&huge_zero_page_shrinker);
636
Rik van Riel97562cd2011-01-13 15:47:12 -0800637 /*
638 * By default disable transparent hugepages on smaller systems,
639 * where the extra memory used could hurt more than TLB overhead
640 * is likely to save. The admin can still enable it through /sys.
641 */
642 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
643 transparent_hugepage_flags = 0;
644
Andrea Arcangeliba761492011-01-13 15:46:58 -0800645 start_khugepaged();
646
Shaohua Li569e5592012-01-12 17:19:11 -0800647 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800648out:
Shaohua Li569e5592012-01-12 17:19:11 -0800649 hugepage_exit_sysfs(hugepage_kobj);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800650 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800651}
652module_init(hugepage_init)
653
654static int __init setup_transparent_hugepage(char *str)
655{
656 int ret = 0;
657 if (!str)
658 goto out;
659 if (!strcmp(str, "always")) {
660 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
661 &transparent_hugepage_flags);
662 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
663 &transparent_hugepage_flags);
664 ret = 1;
665 } else if (!strcmp(str, "madvise")) {
666 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
667 &transparent_hugepage_flags);
668 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
669 &transparent_hugepage_flags);
670 ret = 1;
671 } else if (!strcmp(str, "never")) {
672 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
673 &transparent_hugepage_flags);
674 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
675 &transparent_hugepage_flags);
676 ret = 1;
677 }
678out:
679 if (!ret)
680 printk(KERN_WARNING
681 "transparent_hugepage= cannot parse, ignored\n");
682 return ret;
683}
684__setup("transparent_hugepage=", setup_transparent_hugepage);
685
Mel Gormanb32967f2012-11-19 12:35:47 +0000686pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800687{
688 if (likely(vma->vm_flags & VM_WRITE))
689 pmd = pmd_mkwrite(pmd);
690 return pmd;
691}
692
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700693static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
Bob Liub3092b32012-12-11 16:00:41 -0800694{
695 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700696 entry = mk_pmd(page, prot);
Bob Liub3092b32012-12-11 16:00:41 -0800697 entry = pmd_mkhuge(entry);
698 return entry;
699}
700
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800701static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
702 struct vm_area_struct *vma,
703 unsigned long haddr, pmd_t *pmd,
704 struct page *page)
705{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800706 pgtable_t pgtable;
707
708 VM_BUG_ON(!PageCompound(page));
709 pgtable = pte_alloc_one(mm, haddr);
David Rientjesedad9d22012-05-29 15:06:17 -0700710 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800711 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800712
713 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700714 /*
715 * The memory barrier inside __SetPageUptodate makes sure that
716 * clear_huge_page writes become visible before the set_pmd_at()
717 * write.
718 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800719 __SetPageUptodate(page);
720
721 spin_lock(&mm->page_table_lock);
722 if (unlikely(!pmd_none(*pmd))) {
723 spin_unlock(&mm->page_table_lock);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800724 mem_cgroup_uncharge_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800725 put_page(page);
726 pte_free(mm, pgtable);
727 } else {
728 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700729 entry = mk_huge_pmd(page, vma->vm_page_prot);
730 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800731 page_add_new_anon_rmap(page, vma, haddr);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700732 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800733 set_pmd_at(mm, haddr, pmd, entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800734 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Andrea Arcangeli1c641e82012-03-05 14:59:20 -0800735 mm->nr_ptes++;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800736 spin_unlock(&mm->page_table_lock);
737 }
738
David Rientjesaa2e8782012-05-29 15:06:17 -0700739 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800740}
741
Andi Kleencc5d4622011-03-22 16:33:13 -0700742static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800743{
Andi Kleencc5d4622011-03-22 16:33:13 -0700744 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800745}
746
747static inline struct page *alloc_hugepage_vma(int defrag,
748 struct vm_area_struct *vma,
Andi Kleencc5d4622011-03-22 16:33:13 -0700749 unsigned long haddr, int nd,
750 gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800751{
Andi Kleencc5d4622011-03-22 16:33:13 -0700752 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
Andi Kleen5c4b4be2011-03-04 17:36:32 -0800753 HPAGE_PMD_ORDER, vma, haddr, nd);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800754}
755
756#ifndef CONFIG_NUMA
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800757static inline struct page *alloc_hugepage(int defrag)
758{
Andi Kleencc5d4622011-03-22 16:33:13 -0700759 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800760 HPAGE_PMD_ORDER);
761}
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800762#endif
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800763
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800764static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800765 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700766 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800767{
768 pmd_t entry;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800769 if (!pmd_none(*pmd))
770 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700771 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800772 entry = pmd_wrprotect(entry);
773 entry = pmd_mkhuge(entry);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700774 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800775 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800776 mm->nr_ptes++;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800777 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800778}
779
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800780int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
781 unsigned long address, pmd_t *pmd,
782 unsigned int flags)
783{
784 struct page *page;
785 unsigned long haddr = address & HPAGE_PMD_MASK;
786 pte_t *pte;
787
788 if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
789 if (unlikely(anon_vma_prepare(vma)))
790 return VM_FAULT_OOM;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800791 if (unlikely(khugepaged_enter(vma)))
792 return VM_FAULT_OOM;
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800793 if (!(flags & FAULT_FLAG_WRITE) &&
794 transparent_hugepage_use_zero_page()) {
Kirill A. Shutemov80371952012-12-12 13:51:02 -0800795 pgtable_t pgtable;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700796 struct page *zero_page;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800797 bool set;
Kirill A. Shutemov80371952012-12-12 13:51:02 -0800798 pgtable = pte_alloc_one(mm, haddr);
799 if (unlikely(!pgtable))
800 return VM_FAULT_OOM;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700801 zero_page = get_huge_zero_page();
802 if (unlikely(!zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800803 pte_free(mm, pgtable);
804 count_vm_event(THP_FAULT_FALLBACK);
805 goto out;
806 }
Kirill A. Shutemov80371952012-12-12 13:51:02 -0800807 spin_lock(&mm->page_table_lock);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800808 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700809 zero_page);
Kirill A. Shutemov80371952012-12-12 13:51:02 -0800810 spin_unlock(&mm->page_table_lock);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800811 if (!set) {
812 pte_free(mm, pgtable);
813 put_huge_zero_page();
814 }
Kirill A. Shutemov80371952012-12-12 13:51:02 -0800815 return 0;
816 }
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800817 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
Andi Kleencc5d4622011-03-22 16:33:13 -0700818 vma, haddr, numa_node_id(), 0);
Andi Kleen81ab4202011-04-14 15:22:06 -0700819 if (unlikely(!page)) {
820 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800821 goto out;
Andi Kleen81ab4202011-04-14 15:22:06 -0700822 }
823 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800824 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
825 put_page(page);
826 goto out;
827 }
David Rientjesedad9d22012-05-29 15:06:17 -0700828 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
829 page))) {
830 mem_cgroup_uncharge_page(page);
831 put_page(page);
832 goto out;
833 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800834
David Rientjesedad9d22012-05-29 15:06:17 -0700835 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800836 }
837out:
838 /*
839 * Use __pte_alloc instead of pte_alloc_map, because we can't
840 * run pte_offset_map on the pmd, if an huge pmd could
841 * materialize from under us from a different thread.
842 */
Mel Gorman4fd01772011-10-12 21:06:51 +0200843 if (unlikely(pmd_none(*pmd)) &&
844 unlikely(__pte_alloc(mm, vma, pmd, address)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800845 return VM_FAULT_OOM;
846 /* if an huge pmd materialized from under us just retry later */
847 if (unlikely(pmd_trans_huge(*pmd)))
848 return 0;
849 /*
850 * A regular pmd is established and it can't morph into a huge pmd
851 * from under us anymore at this point because we hold the mmap_sem
852 * read mode and khugepaged takes it in write mode. So now it's
853 * safe to run pte_offset_map().
854 */
855 pte = pte_offset_map(pmd, address);
856 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
857}
858
859int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
860 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
861 struct vm_area_struct *vma)
862{
863 struct page *src_page;
864 pmd_t pmd;
865 pgtable_t pgtable;
866 int ret;
867
868 ret = -ENOMEM;
869 pgtable = pte_alloc_one(dst_mm, addr);
870 if (unlikely(!pgtable))
871 goto out;
872
873 spin_lock(&dst_mm->page_table_lock);
874 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
875
876 ret = -EAGAIN;
877 pmd = *src_pmd;
878 if (unlikely(!pmd_trans_huge(pmd))) {
879 pte_free(dst_mm, pgtable);
880 goto out_unlock;
881 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800882 /*
883 * mm->page_table_lock is enough to be sure that huge zero pmd is not
884 * under splitting since we don't split the page itself, only pmd to
885 * a page table.
886 */
887 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700888 struct page *zero_page;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800889 bool set;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800890 /*
891 * get_huge_zero_page() will never allocate a new page here,
892 * since we already have a zero page to copy. It just takes a
893 * reference.
894 */
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700895 zero_page = get_huge_zero_page();
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800896 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700897 zero_page);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800898 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800899 ret = 0;
900 goto out_unlock;
901 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800902 if (unlikely(pmd_trans_splitting(pmd))) {
903 /* split huge page running from under us */
904 spin_unlock(&src_mm->page_table_lock);
905 spin_unlock(&dst_mm->page_table_lock);
906 pte_free(dst_mm, pgtable);
907
908 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
909 goto out;
910 }
911 src_page = pmd_page(pmd);
912 VM_BUG_ON(!PageHead(src_page));
913 get_page(src_page);
914 page_dup_rmap(src_page);
915 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
916
917 pmdp_set_wrprotect(src_mm, addr, src_pmd);
918 pmd = pmd_mkold(pmd_wrprotect(pmd));
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700919 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800920 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Andrea Arcangeli1c641e82012-03-05 14:59:20 -0800921 dst_mm->nr_ptes++;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800922
923 ret = 0;
924out_unlock:
925 spin_unlock(&src_mm->page_table_lock);
926 spin_unlock(&dst_mm->page_table_lock);
927out:
928 return ret;
929}
930
Will Deacona1dd4502012-12-11 16:01:27 -0800931void huge_pmd_set_accessed(struct mm_struct *mm,
932 struct vm_area_struct *vma,
933 unsigned long address,
934 pmd_t *pmd, pmd_t orig_pmd,
935 int dirty)
936{
937 pmd_t entry;
938 unsigned long haddr;
939
940 spin_lock(&mm->page_table_lock);
941 if (unlikely(!pmd_same(*pmd, orig_pmd)))
942 goto unlock;
943
944 entry = pmd_mkyoung(orig_pmd);
945 haddr = address & HPAGE_PMD_MASK;
946 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
947 update_mmu_cache_pmd(vma, address, pmd);
948
949unlock:
950 spin_unlock(&mm->page_table_lock);
951}
952
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800953static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
954 struct vm_area_struct *vma, unsigned long address,
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800955 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800956{
957 pgtable_t pgtable;
958 pmd_t _pmd;
959 struct page *page;
960 int i, ret = 0;
961 unsigned long mmun_start; /* For mmu_notifiers */
962 unsigned long mmun_end; /* For mmu_notifiers */
963
964 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
965 if (!page) {
966 ret |= VM_FAULT_OOM;
967 goto out;
968 }
969
970 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
971 put_page(page);
972 ret |= VM_FAULT_OOM;
973 goto out;
974 }
975
976 clear_user_highpage(page, address);
977 __SetPageUptodate(page);
978
979 mmun_start = haddr;
980 mmun_end = haddr + HPAGE_PMD_SIZE;
981 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
982
983 spin_lock(&mm->page_table_lock);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800984 if (unlikely(!pmd_same(*pmd, orig_pmd)))
985 goto out_free_page;
986
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800987 pmdp_clear_flush(vma, haddr, pmd);
988 /* leave pmd empty until pte is filled */
989
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700990 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800991 pmd_populate(mm, &_pmd, pgtable);
992
993 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
994 pte_t *pte, entry;
995 if (haddr == (address & PAGE_MASK)) {
996 entry = mk_pte(page, vma->vm_page_prot);
997 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
998 page_add_new_anon_rmap(page, vma, haddr);
999 } else {
1000 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1001 entry = pte_mkspecial(entry);
1002 }
1003 pte = pte_offset_map(&_pmd, haddr);
1004 VM_BUG_ON(!pte_none(*pte));
1005 set_pte_at(mm, haddr, pte, entry);
1006 pte_unmap(pte);
1007 }
1008 smp_wmb(); /* make pte visible before pmd */
1009 pmd_populate(mm, pmd, pgtable);
1010 spin_unlock(&mm->page_table_lock);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001011 put_huge_zero_page();
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001012 inc_mm_counter(mm, MM_ANONPAGES);
1013
1014 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1015
1016 ret |= VM_FAULT_WRITE;
1017out:
1018 return ret;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -08001019out_free_page:
1020 spin_unlock(&mm->page_table_lock);
1021 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1022 mem_cgroup_uncharge_page(page);
1023 put_page(page);
1024 goto out;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001025}
1026
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001027static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1028 struct vm_area_struct *vma,
1029 unsigned long address,
1030 pmd_t *pmd, pmd_t orig_pmd,
1031 struct page *page,
1032 unsigned long haddr)
1033{
1034 pgtable_t pgtable;
1035 pmd_t _pmd;
1036 int ret = 0, i;
1037 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001038 unsigned long mmun_start; /* For mmu_notifiers */
1039 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001040
1041 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1042 GFP_KERNEL);
1043 if (unlikely(!pages)) {
1044 ret |= VM_FAULT_OOM;
1045 goto out;
1046 }
1047
1048 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -07001049 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1050 __GFP_OTHER_NODE,
Andi Kleen19ee1512011-03-04 17:36:31 -08001051 vma, address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001052 if (unlikely(!pages[i] ||
1053 mem_cgroup_newpage_charge(pages[i], mm,
1054 GFP_KERNEL))) {
1055 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001056 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001057 mem_cgroup_uncharge_start();
1058 while (--i >= 0) {
1059 mem_cgroup_uncharge_page(pages[i]);
1060 put_page(pages[i]);
1061 }
1062 mem_cgroup_uncharge_end();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001063 kfree(pages);
1064 ret |= VM_FAULT_OOM;
1065 goto out;
1066 }
1067 }
1068
1069 for (i = 0; i < HPAGE_PMD_NR; i++) {
1070 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001071 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001072 __SetPageUptodate(pages[i]);
1073 cond_resched();
1074 }
1075
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001076 mmun_start = haddr;
1077 mmun_end = haddr + HPAGE_PMD_SIZE;
1078 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1079
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001080 spin_lock(&mm->page_table_lock);
1081 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1082 goto out_free_pages;
1083 VM_BUG_ON(!PageHead(page));
1084
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001085 pmdp_clear_flush(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001086 /* leave pmd empty until pte is filled */
1087
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001088 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001089 pmd_populate(mm, &_pmd, pgtable);
1090
1091 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1092 pte_t *pte, entry;
1093 entry = mk_pte(pages[i], vma->vm_page_prot);
1094 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1095 page_add_new_anon_rmap(pages[i], vma, haddr);
1096 pte = pte_offset_map(&_pmd, haddr);
1097 VM_BUG_ON(!pte_none(*pte));
1098 set_pte_at(mm, haddr, pte, entry);
1099 pte_unmap(pte);
1100 }
1101 kfree(pages);
1102
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001103 smp_wmb(); /* make pte visible before pmd */
1104 pmd_populate(mm, pmd, pgtable);
1105 page_remove_rmap(page);
1106 spin_unlock(&mm->page_table_lock);
1107
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001108 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1109
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001110 ret |= VM_FAULT_WRITE;
1111 put_page(page);
1112
1113out:
1114 return ret;
1115
1116out_free_pages:
1117 spin_unlock(&mm->page_table_lock);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001118 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001119 mem_cgroup_uncharge_start();
1120 for (i = 0; i < HPAGE_PMD_NR; i++) {
1121 mem_cgroup_uncharge_page(pages[i]);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001122 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001123 }
1124 mem_cgroup_uncharge_end();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001125 kfree(pages);
1126 goto out;
1127}
1128
1129int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1130 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1131{
1132 int ret = 0;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001133 struct page *page = NULL, *new_page;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001134 unsigned long haddr;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001135 unsigned long mmun_start; /* For mmu_notifiers */
1136 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001137
1138 VM_BUG_ON(!vma->anon_vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001139 haddr = address & HPAGE_PMD_MASK;
1140 if (is_huge_zero_pmd(orig_pmd))
1141 goto alloc;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001142 spin_lock(&mm->page_table_lock);
1143 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1144 goto out_unlock;
1145
1146 page = pmd_page(orig_pmd);
1147 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001148 if (page_mapcount(page) == 1) {
1149 pmd_t entry;
1150 entry = pmd_mkyoung(orig_pmd);
1151 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1152 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
David Millerb113da62012-10-08 16:34:25 -07001153 update_mmu_cache_pmd(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001154 ret |= VM_FAULT_WRITE;
1155 goto out_unlock;
1156 }
1157 get_page(page);
1158 spin_unlock(&mm->page_table_lock);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001159alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001160 if (transparent_hugepage_enabled(vma) &&
1161 !transparent_hugepage_debug_cow())
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001162 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
Andi Kleencc5d4622011-03-22 16:33:13 -07001163 vma, haddr, numa_node_id(), 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001164 else
1165 new_page = NULL;
1166
1167 if (unlikely(!new_page)) {
Andi Kleen81ab4202011-04-14 15:22:06 -07001168 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001169 if (is_huge_zero_pmd(orig_pmd)) {
1170 ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -08001171 address, pmd, orig_pmd, haddr);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001172 } else {
1173 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1174 pmd, orig_pmd, page, haddr);
1175 if (ret & VM_FAULT_OOM)
1176 split_huge_page(page);
1177 put_page(page);
1178 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001179 goto out;
1180 }
Andi Kleen81ab4202011-04-14 15:22:06 -07001181 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001182
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001183 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1184 put_page(new_page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001185 if (page) {
1186 split_huge_page(page);
1187 put_page(page);
1188 }
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001189 ret |= VM_FAULT_OOM;
1190 goto out;
1191 }
1192
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001193 if (is_huge_zero_pmd(orig_pmd))
1194 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1195 else
1196 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001197 __SetPageUptodate(new_page);
1198
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001199 mmun_start = haddr;
1200 mmun_end = haddr + HPAGE_PMD_SIZE;
1201 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1202
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001203 spin_lock(&mm->page_table_lock);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001204 if (page)
1205 put_page(page);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001206 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
David Rientjes6f60b692012-05-29 15:06:26 -07001207 spin_unlock(&mm->page_table_lock);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001208 mem_cgroup_uncharge_page(new_page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001209 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001210 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001211 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001212 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001213 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1214 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001215 pmdp_clear_flush(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001216 page_add_new_anon_rmap(new_page, vma, haddr);
1217 set_pmd_at(mm, haddr, pmd, entry);
David Millerb113da62012-10-08 16:34:25 -07001218 update_mmu_cache_pmd(vma, address, pmd);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001219 if (is_huge_zero_pmd(orig_pmd)) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001220 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001221 put_huge_zero_page();
1222 } else {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001223 VM_BUG_ON(!PageHead(page));
1224 page_remove_rmap(page);
1225 put_page(page);
1226 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001227 ret |= VM_FAULT_WRITE;
1228 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001229 spin_unlock(&mm->page_table_lock);
1230out_mn:
1231 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1232out:
1233 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001234out_unlock:
1235 spin_unlock(&mm->page_table_lock);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001236 return ret;
1237}
1238
David Rientjesb676b292012-10-08 16:34:03 -07001239struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001240 unsigned long addr,
1241 pmd_t *pmd,
1242 unsigned int flags)
1243{
David Rientjesb676b292012-10-08 16:34:03 -07001244 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001245 struct page *page = NULL;
1246
1247 assert_spin_locked(&mm->page_table_lock);
1248
1249 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1250 goto out;
1251
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001252 /* Avoid dumping huge zero page */
1253 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1254 return ERR_PTR(-EFAULT);
1255
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001256 page = pmd_page(*pmd);
1257 VM_BUG_ON(!PageHead(page));
1258 if (flags & FOLL_TOUCH) {
1259 pmd_t _pmd;
1260 /*
1261 * We should set the dirty bit only for FOLL_WRITE but
1262 * for now the dirty bit in the pmd is meaningless.
1263 * And if the dirty bit will become meaningful and
1264 * we'll only set it with FOLL_WRITE, an atomic
1265 * set_bit will be required on the pmd to set the
1266 * young bit, instead of the current set_pmd_at.
1267 */
1268 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
Aneesh Kumar K.V8663890a2013-06-06 00:20:34 -07001269 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1270 pmd, _pmd, 1))
1271 update_mmu_cache_pmd(vma, addr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001272 }
David Rientjesb676b292012-10-08 16:34:03 -07001273 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1274 if (page->mapping && trylock_page(page)) {
1275 lru_add_drain();
1276 if (page->mapping)
1277 mlock_vma_page(page);
1278 unlock_page(page);
1279 }
1280 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001281 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1282 VM_BUG_ON(!PageCompound(page));
1283 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001284 get_page_foll(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001285
1286out:
1287 return page;
1288}
1289
Mel Gormand10e63f2012-10-25 14:16:31 +02001290/* NUMA hinting page fault entry point for trans huge pmds */
Mel Gorman4daae3b2012-11-02 11:33:45 +00001291int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1292 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
Mel Gormand10e63f2012-10-25 14:16:31 +02001293{
Mel Gormanb32967f2012-11-19 12:35:47 +00001294 struct page *page;
Mel Gormand10e63f2012-10-25 14:16:31 +02001295 unsigned long haddr = addr & HPAGE_PMD_MASK;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001296 int target_nid;
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001297 int current_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001298 bool migrated;
Mel Gormand10e63f2012-10-25 14:16:31 +02001299
1300 spin_lock(&mm->page_table_lock);
1301 if (unlikely(!pmd_same(pmd, *pmdp)))
1302 goto out_unlock;
1303
1304 page = pmd_page(pmd);
Mel Gorman4daae3b2012-11-02 11:33:45 +00001305 get_page(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001306 current_nid = page_to_nid(page);
1307 count_vm_numa_event(NUMA_HINT_FAULTS);
1308 if (current_nid == numa_node_id())
1309 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Mel Gorman4daae3b2012-11-02 11:33:45 +00001310
1311 target_nid = mpol_misplaced(page, vma, haddr);
Mel Gormanb32967f2012-11-19 12:35:47 +00001312 if (target_nid == -1) {
1313 put_page(page);
Mel Gorman4daae3b2012-11-02 11:33:45 +00001314 goto clear_pmdnuma;
Mel Gormanb32967f2012-11-19 12:35:47 +00001315 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001316
Mel Gormanb32967f2012-11-19 12:35:47 +00001317 /* Acquire the page lock to serialise THP migrations */
1318 spin_unlock(&mm->page_table_lock);
1319 lock_page(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001320
Mel Gormanb32967f2012-11-19 12:35:47 +00001321 /* Confirm the PTE did not while locked */
1322 spin_lock(&mm->page_table_lock);
1323 if (unlikely(!pmd_same(pmd, *pmdp))) {
1324 unlock_page(page);
1325 put_page(page);
1326 goto out_unlock;
1327 }
1328 spin_unlock(&mm->page_table_lock);
1329
1330 /* Migrate the THP to the requested node */
1331 migrated = migrate_misplaced_transhuge_page(mm, vma,
Hugh Dickins340ef392013-02-22 16:34:33 -08001332 pmdp, pmd, addr, page, target_nid);
1333 if (!migrated)
1334 goto check_same;
Mel Gormanb32967f2012-11-19 12:35:47 +00001335
Hugh Dickins340ef392013-02-22 16:34:33 -08001336 task_numa_fault(target_nid, HPAGE_PMD_NR, true);
Mel Gorman4daae3b2012-11-02 11:33:45 +00001337 return 0;
1338
Hugh Dickins340ef392013-02-22 16:34:33 -08001339check_same:
1340 spin_lock(&mm->page_table_lock);
1341 if (unlikely(!pmd_same(pmd, *pmdp)))
1342 goto out_unlock;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001343clear_pmdnuma:
Mel Gormand10e63f2012-10-25 14:16:31 +02001344 pmd = pmd_mknonnuma(pmd);
1345 set_pmd_at(mm, haddr, pmdp, pmd);
1346 VM_BUG_ON(pmd_numa(*pmdp));
1347 update_mmu_cache_pmd(vma, addr, pmdp);
Mel Gormand10e63f2012-10-25 14:16:31 +02001348out_unlock:
1349 spin_unlock(&mm->page_table_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001350 if (current_nid != -1)
Hugh Dickins340ef392013-02-22 16:34:33 -08001351 task_numa_fault(current_nid, HPAGE_PMD_NR, false);
Mel Gormand10e63f2012-10-25 14:16:31 +02001352 return 0;
1353}
1354
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001355int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001356 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001357{
1358 int ret = 0;
1359
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001360 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1361 struct page *page;
1362 pgtable_t pgtable;
David Millerf5c8ad42012-10-08 16:34:26 -07001363 pmd_t orig_pmd;
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001364 /*
1365 * For architectures like ppc64 we look at deposited pgtable
1366 * when calling pmdp_get_and_clear. So do the
1367 * pgtable_trans_huge_withdraw after finishing pmdp related
1368 * operations.
1369 */
David Millerf5c8ad42012-10-08 16:34:26 -07001370 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001371 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001372 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001373 if (is_huge_zero_pmd(orig_pmd)) {
1374 tlb->mm->nr_ptes--;
1375 spin_unlock(&tlb->mm->page_table_lock);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001376 put_huge_zero_page();
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001377 } else {
1378 page = pmd_page(orig_pmd);
1379 page_remove_rmap(page);
1380 VM_BUG_ON(page_mapcount(page) < 0);
1381 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1382 VM_BUG_ON(!PageHead(page));
1383 tlb->mm->nr_ptes--;
1384 spin_unlock(&tlb->mm->page_table_lock);
1385 tlb_remove_page(tlb, page);
1386 }
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001387 pte_free(tlb->mm, pgtable);
1388 ret = 1;
1389 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001390 return ret;
1391}
1392
Johannes Weiner0ca16342011-01-13 15:47:02 -08001393int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1394 unsigned long addr, unsigned long end,
1395 unsigned char *vec)
1396{
1397 int ret = 0;
1398
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001399 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1400 /*
1401 * All logical pages in the range are present
1402 * if backed by a huge page.
1403 */
Johannes Weiner0ca16342011-01-13 15:47:02 -08001404 spin_unlock(&vma->vm_mm->page_table_lock);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001405 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1406 ret = 1;
1407 }
Johannes Weiner0ca16342011-01-13 15:47:02 -08001408
1409 return ret;
1410}
1411
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001412int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1413 unsigned long old_addr,
1414 unsigned long new_addr, unsigned long old_end,
1415 pmd_t *old_pmd, pmd_t *new_pmd)
1416{
1417 int ret = 0;
1418 pmd_t pmd;
1419
1420 struct mm_struct *mm = vma->vm_mm;
1421
1422 if ((old_addr & ~HPAGE_PMD_MASK) ||
1423 (new_addr & ~HPAGE_PMD_MASK) ||
1424 old_end - old_addr < HPAGE_PMD_SIZE ||
1425 (new_vma->vm_flags & VM_NOHUGEPAGE))
1426 goto out;
1427
1428 /*
1429 * The destination pmd shouldn't be established, free_pgtables()
1430 * should have release it.
1431 */
1432 if (WARN_ON(!pmd_none(*new_pmd))) {
1433 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1434 goto out;
1435 }
1436
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001437 ret = __pmd_trans_huge_lock(old_pmd, vma);
1438 if (ret == 1) {
1439 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1440 VM_BUG_ON(!pmd_none(*new_pmd));
Pavel Emelyanov0f8975e2013-07-03 15:01:20 -07001441 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001442 spin_unlock(&mm->page_table_lock);
1443 }
1444out:
1445 return ret;
1446}
1447
Johannes Weinercd7548a2011-01-13 15:47:04 -08001448int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001449 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001450{
1451 struct mm_struct *mm = vma->vm_mm;
1452 int ret = 0;
1453
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001454 if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1455 pmd_t entry;
1456 entry = pmdp_get_and_clear(mm, addr, pmd);
Hugh Dickinsa4f1de12012-12-16 18:56:58 -08001457 if (!prot_numa) {
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001458 entry = pmd_modify(entry, newprot);
Hugh Dickinsa4f1de12012-12-16 18:56:58 -08001459 BUG_ON(pmd_write(entry));
1460 } else {
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001461 struct page *page = pmd_page(*pmd);
1462
1463 /* only check non-shared pages */
1464 if (page_mapcount(page) == 1 &&
1465 !pmd_numa(*pmd)) {
1466 entry = pmd_mknuma(entry);
1467 }
1468 }
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001469 set_pmd_at(mm, addr, pmd, entry);
Johannes Weinercd7548a2011-01-13 15:47:04 -08001470 spin_unlock(&vma->vm_mm->page_table_lock);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001471 ret = 1;
1472 }
Johannes Weinercd7548a2011-01-13 15:47:04 -08001473
1474 return ret;
1475}
1476
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001477/*
1478 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1479 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1480 *
1481 * Note that if it returns 1, this routine returns without unlocking page
1482 * table locks. So callers must unlock them.
1483 */
1484int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1485{
1486 spin_lock(&vma->vm_mm->page_table_lock);
1487 if (likely(pmd_trans_huge(*pmd))) {
1488 if (unlikely(pmd_trans_splitting(*pmd))) {
1489 spin_unlock(&vma->vm_mm->page_table_lock);
1490 wait_split_huge_page(vma->anon_vma, pmd);
1491 return -1;
1492 } else {
1493 /* Thp mapped by 'pmd' is stable, so we can
1494 * handle it as it is. */
1495 return 1;
1496 }
1497 }
1498 spin_unlock(&vma->vm_mm->page_table_lock);
1499 return 0;
1500}
1501
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001502pmd_t *page_check_address_pmd(struct page *page,
1503 struct mm_struct *mm,
1504 unsigned long address,
1505 enum page_check_address_pmd_flag flag)
1506{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001507 pmd_t *pmd, *ret = NULL;
1508
1509 if (address & ~HPAGE_PMD_MASK)
1510 goto out;
1511
Bob Liu62190492012-12-11 16:00:37 -08001512 pmd = mm_find_pmd(mm, address);
1513 if (!pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001514 goto out;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001515 if (pmd_none(*pmd))
1516 goto out;
1517 if (pmd_page(*pmd) != page)
1518 goto out;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001519 /*
1520 * split_vma() may create temporary aliased mappings. There is
1521 * no risk as long as all huge pmd are found and have their
1522 * splitting bit set before __split_huge_page_refcount
1523 * runs. Finding the same huge pmd more than once during the
1524 * same rmap walk is not a problem.
1525 */
1526 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1527 pmd_trans_splitting(*pmd))
1528 goto out;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001529 if (pmd_trans_huge(*pmd)) {
1530 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1531 !pmd_trans_splitting(*pmd));
1532 ret = pmd;
1533 }
1534out:
1535 return ret;
1536}
1537
1538static int __split_huge_page_splitting(struct page *page,
1539 struct vm_area_struct *vma,
1540 unsigned long address)
1541{
1542 struct mm_struct *mm = vma->vm_mm;
1543 pmd_t *pmd;
1544 int ret = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001545 /* For mmu_notifiers */
1546 const unsigned long mmun_start = address;
1547 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001548
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001549 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001550 spin_lock(&mm->page_table_lock);
1551 pmd = page_check_address_pmd(page, mm, address,
1552 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1553 if (pmd) {
1554 /*
1555 * We can't temporarily set the pmd to null in order
1556 * to split it, the pmd must remain marked huge at all
1557 * times or the VM won't take the pmd_trans_huge paths
Ingo Molnar5a505082012-12-02 19:56:46 +00001558 * and it won't wait on the anon_vma->root->rwsem to
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001559 * serialize against split_huge_page*.
1560 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001561 pmdp_splitting_flush(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001562 ret = 1;
1563 }
1564 spin_unlock(&mm->page_table_lock);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001565 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001566
1567 return ret;
1568}
1569
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001570static void __split_huge_page_refcount(struct page *page,
1571 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001572{
1573 int i;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001574 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001575 struct lruvec *lruvec;
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001576 int tail_count = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001577
1578 /* prevent PageLRU to go away from under us, and freeze lru stats */
1579 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001580 lruvec = mem_cgroup_page_lruvec(page, zone);
1581
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001582 compound_lock(page);
KAMEZAWA Hiroyukie94c8a92012-01-12 17:18:20 -08001583 /* complete memcg works before add pages to LRU */
1584 mem_cgroup_split_huge_fixup(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001585
Shaohua Li45676882012-01-12 17:19:18 -08001586 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001587 struct page *page_tail = page + i;
1588
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001589 /* tail_page->_mapcount cannot change */
1590 BUG_ON(page_mapcount(page_tail) < 0);
1591 tail_count += page_mapcount(page_tail);
1592 /* check for overflow */
1593 BUG_ON(tail_count < 0);
1594 BUG_ON(atomic_read(&page_tail->_count) != 0);
1595 /*
1596 * tail_page->_count is zero and not changing from
1597 * under us. But get_page_unless_zero() may be running
1598 * from under us on the tail_page. If we used
1599 * atomic_set() below instead of atomic_add(), we
1600 * would then run atomic_set() concurrently with
1601 * get_page_unless_zero(), and atomic_set() is
1602 * implemented in C not using locked ops. spin_unlock
1603 * on x86 sometime uses locked ops because of PPro
1604 * errata 66, 92, so unless somebody can guarantee
1605 * atomic_set() here would be safe on all archs (and
1606 * not only on x86), it's safer to use atomic_add().
1607 */
1608 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1609 &page_tail->_count);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001610
1611 /* after clearing PageTail the gup refcount can be released */
1612 smp_mb();
1613
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001614 /*
1615 * retain hwpoison flag of the poisoned tail page:
1616 * fix for the unsuitable process killed on Guest Machine(KVM)
1617 * by the memory-failure.
1618 */
1619 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001620 page_tail->flags |= (page->flags &
1621 ((1L << PG_referenced) |
1622 (1L << PG_swapbacked) |
1623 (1L << PG_mlocked) |
Kirill A. Shutemove180cf82013-07-31 13:53:39 -07001624 (1L << PG_uptodate) |
1625 (1L << PG_active) |
1626 (1L << PG_unevictable)));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001627 page_tail->flags |= (1L << PG_dirty);
1628
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001629 /* clear PageTail before overwriting first_page */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001630 smp_wmb();
1631
1632 /*
1633 * __split_huge_page_splitting() already set the
1634 * splitting bit in all pmd that could map this
1635 * hugepage, that will ensure no CPU can alter the
1636 * mapcount on the head page. The mapcount is only
1637 * accounted in the head page and it has to be
1638 * transferred to all tail pages in the below code. So
1639 * for this code to be safe, the split the mapcount
1640 * can't change. But that doesn't mean userland can't
1641 * keep changing and reading the page contents while
1642 * we transfer the mapcount, so the pmd splitting
1643 * status is achieved setting a reserved bit in the
1644 * pmd, not by clearing the present bit.
1645 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001646 page_tail->_mapcount = page->_mapcount;
1647
1648 BUG_ON(page_tail->mapping);
1649 page_tail->mapping = page->mapping;
1650
Shaohua Li45676882012-01-12 17:19:18 -08001651 page_tail->index = page->index + i;
Mel Gorman22b751c2013-02-22 16:34:59 -08001652 page_nid_xchg_last(page_tail, page_nid_last(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001653
1654 BUG_ON(!PageAnon(page_tail));
1655 BUG_ON(!PageUptodate(page_tail));
1656 BUG_ON(!PageDirty(page_tail));
1657 BUG_ON(!PageSwapBacked(page_tail));
1658
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001659 lru_add_page_tail(page, page_tail, lruvec, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001660 }
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001661 atomic_sub(tail_count, &page->_count);
1662 BUG_ON(atomic_read(&page->_count) <= 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001663
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001664 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001665
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001666 ClearPageCompound(page);
1667 compound_unlock(page);
1668 spin_unlock_irq(&zone->lru_lock);
1669
1670 for (i = 1; i < HPAGE_PMD_NR; i++) {
1671 struct page *page_tail = page + i;
1672 BUG_ON(page_count(page_tail) <= 0);
1673 /*
1674 * Tail pages may be freed if there wasn't any mapping
1675 * like if add_to_swap() is running on a lru page that
1676 * had its mapping zapped. And freeing these pages
1677 * requires taking the lru_lock so we do the put_page
1678 * of the tail pages after the split is complete.
1679 */
1680 put_page(page_tail);
1681 }
1682
1683 /*
1684 * Only the head page (now become a regular page) is required
1685 * to be pinned by the caller.
1686 */
1687 BUG_ON(page_count(page) <= 0);
1688}
1689
1690static int __split_huge_page_map(struct page *page,
1691 struct vm_area_struct *vma,
1692 unsigned long address)
1693{
1694 struct mm_struct *mm = vma->vm_mm;
1695 pmd_t *pmd, _pmd;
1696 int ret = 0, i;
1697 pgtable_t pgtable;
1698 unsigned long haddr;
1699
1700 spin_lock(&mm->page_table_lock);
1701 pmd = page_check_address_pmd(page, mm, address,
1702 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1703 if (pmd) {
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001704 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001705 pmd_populate(mm, &_pmd, pgtable);
1706
Gerald Schaefere3ebcf642012-10-08 16:30:07 -07001707 haddr = address;
1708 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001709 pte_t *pte, entry;
1710 BUG_ON(PageCompound(page+i));
1711 entry = mk_pte(page + i, vma->vm_page_prot);
1712 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1713 if (!pmd_write(*pmd))
1714 entry = pte_wrprotect(entry);
1715 else
1716 BUG_ON(page_mapcount(page) != 1);
1717 if (!pmd_young(*pmd))
1718 entry = pte_mkold(entry);
Andrea Arcangeli1ba6e0b2012-10-04 01:51:06 +02001719 if (pmd_numa(*pmd))
1720 entry = pte_mknuma(entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001721 pte = pte_offset_map(&_pmd, haddr);
1722 BUG_ON(!pte_none(*pte));
1723 set_pte_at(mm, haddr, pte, entry);
1724 pte_unmap(pte);
1725 }
1726
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001727 smp_wmb(); /* make pte visible before pmd */
1728 /*
1729 * Up to this point the pmd is present and huge and
1730 * userland has the whole access to the hugepage
1731 * during the split (which happens in place). If we
1732 * overwrite the pmd with the not-huge version
1733 * pointing to the pte here (which of course we could
1734 * if all CPUs were bug free), userland could trigger
1735 * a small page size TLB miss on the small sized TLB
1736 * while the hugepage TLB entry is still established
1737 * in the huge TLB. Some CPU doesn't like that. See
1738 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1739 * Erratum 383 on page 93. Intel should be safe but is
1740 * also warns that it's only safe if the permission
1741 * and cache attributes of the two entries loaded in
1742 * the two TLB is identical (which should be the case
1743 * here). But it is generally safer to never allow
1744 * small and huge TLB entries for the same virtual
1745 * address to be loaded simultaneously. So instead of
1746 * doing "pmd_populate(); flush_tlb_range();" we first
1747 * mark the current pmd notpresent (atomically because
1748 * here the pmd_trans_huge and pmd_trans_splitting
1749 * must remain set at all times on the pmd until the
1750 * split is complete for this pmd), then we flush the
1751 * SMP TLB and finally we write the non-huge version
1752 * of the pmd entry with pmd_populate.
1753 */
Gerald Schaefer46dcde72012-10-08 16:30:09 -07001754 pmdp_invalidate(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001755 pmd_populate(mm, pmd, pgtable);
1756 ret = 1;
1757 }
1758 spin_unlock(&mm->page_table_lock);
1759
1760 return ret;
1761}
1762
Ingo Molnar5a505082012-12-02 19:56:46 +00001763/* must be called with anon_vma->root->rwsem held */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001764static void __split_huge_page(struct page *page,
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001765 struct anon_vma *anon_vma,
1766 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001767{
1768 int mapcount, mapcount2;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001769 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001770 struct anon_vma_chain *avc;
1771
1772 BUG_ON(!PageHead(page));
1773 BUG_ON(PageTail(page));
1774
1775 mapcount = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001776 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001777 struct vm_area_struct *vma = avc->vma;
1778 unsigned long addr = vma_address(page, vma);
1779 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001780 mapcount += __split_huge_page_splitting(page, vma, addr);
1781 }
Andrea Arcangeli05759d32011-01-13 15:46:53 -08001782 /*
1783 * It is critical that new vmas are added to the tail of the
1784 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1785 * and establishes a child pmd before
1786 * __split_huge_page_splitting() freezes the parent pmd (so if
1787 * we fail to prevent copy_huge_pmd() from running until the
1788 * whole __split_huge_page() is complete), we will still see
1789 * the newly established pmd of the child later during the
1790 * walk, to be able to set it as pmd_trans_splitting too.
1791 */
1792 if (mapcount != page_mapcount(page))
1793 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1794 mapcount, page_mapcount(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001795 BUG_ON(mapcount != page_mapcount(page));
1796
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001797 __split_huge_page_refcount(page, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001798
1799 mapcount2 = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001800 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001801 struct vm_area_struct *vma = avc->vma;
1802 unsigned long addr = vma_address(page, vma);
1803 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001804 mapcount2 += __split_huge_page_map(page, vma, addr);
1805 }
Andrea Arcangeli05759d32011-01-13 15:46:53 -08001806 if (mapcount != mapcount2)
1807 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1808 mapcount, mapcount2, page_mapcount(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001809 BUG_ON(mapcount != mapcount2);
1810}
1811
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001812/*
1813 * Split a hugepage into normal pages. This doesn't change the position of head
1814 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1815 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1816 * from the hugepage.
1817 * Return 0 if the hugepage is split successfully otherwise return 1.
1818 */
1819int split_huge_page_to_list(struct page *page, struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001820{
1821 struct anon_vma *anon_vma;
1822 int ret = 1;
1823
Kirill A. Shutemov5918d102013-04-29 15:08:44 -07001824 BUG_ON(is_huge_zero_page(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001825 BUG_ON(!PageAnon(page));
Mel Gorman062f1af2013-01-11 14:32:02 -08001826
1827 /*
1828 * The caller does not necessarily hold an mmap_sem that would prevent
1829 * the anon_vma disappearing so we first we take a reference to it
1830 * and then lock the anon_vma for write. This is similar to
1831 * page_lock_anon_vma_read except the write lock is taken to serialise
1832 * against parallel split or collapse operations.
1833 */
1834 anon_vma = page_get_anon_vma(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001835 if (!anon_vma)
1836 goto out;
Mel Gorman062f1af2013-01-11 14:32:02 -08001837 anon_vma_lock_write(anon_vma);
1838
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001839 ret = 0;
1840 if (!PageCompound(page))
1841 goto out_unlock;
1842
1843 BUG_ON(!PageSwapBacked(page));
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001844 __split_huge_page(page, anon_vma, list);
Andi Kleen81ab4202011-04-14 15:22:06 -07001845 count_vm_event(THP_SPLIT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001846
1847 BUG_ON(PageCompound(page));
1848out_unlock:
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08001849 anon_vma_unlock_write(anon_vma);
Mel Gorman062f1af2013-01-11 14:32:02 -08001850 put_anon_vma(anon_vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001851out:
1852 return ret;
1853}
1854
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001855#define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001856
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001857int hugepage_madvise(struct vm_area_struct *vma,
1858 unsigned long *vm_flags, int advice)
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001859{
Gerald Schaefer8e720332012-10-08 16:30:12 -07001860 struct mm_struct *mm = vma->vm_mm;
1861
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001862 switch (advice) {
1863 case MADV_HUGEPAGE:
1864 /*
1865 * Be somewhat over-protective like KSM for now!
1866 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001867 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001868 return -EINVAL;
Gerald Schaefer8e720332012-10-08 16:30:12 -07001869 if (mm->def_flags & VM_NOHUGEPAGE)
1870 return -EINVAL;
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001871 *vm_flags &= ~VM_NOHUGEPAGE;
1872 *vm_flags |= VM_HUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001873 /*
1874 * If the vma become good for khugepaged to scan,
1875 * register it here without waiting a page fault that
1876 * may not happen any time soon.
1877 */
1878 if (unlikely(khugepaged_enter_vma_merge(vma)))
1879 return -ENOMEM;
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001880 break;
1881 case MADV_NOHUGEPAGE:
1882 /*
1883 * Be somewhat over-protective like KSM for now!
1884 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001885 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001886 return -EINVAL;
1887 *vm_flags &= ~VM_HUGEPAGE;
1888 *vm_flags |= VM_NOHUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001889 /*
1890 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1891 * this vma even if we leave the mm registered in khugepaged if
1892 * it got registered before VM_NOHUGEPAGE was set.
1893 */
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001894 break;
1895 }
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001896
1897 return 0;
1898}
1899
Andrea Arcangeliba761492011-01-13 15:46:58 -08001900static int __init khugepaged_slab_init(void)
1901{
1902 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1903 sizeof(struct mm_slot),
1904 __alignof__(struct mm_slot), 0, NULL);
1905 if (!mm_slot_cache)
1906 return -ENOMEM;
1907
1908 return 0;
1909}
1910
Andrea Arcangeliba761492011-01-13 15:46:58 -08001911static inline struct mm_slot *alloc_mm_slot(void)
1912{
1913 if (!mm_slot_cache) /* initialization failed */
1914 return NULL;
1915 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1916}
1917
1918static inline void free_mm_slot(struct mm_slot *mm_slot)
1919{
1920 kmem_cache_free(mm_slot_cache, mm_slot);
1921}
1922
Andrea Arcangeliba761492011-01-13 15:46:58 -08001923static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1924{
1925 struct mm_slot *mm_slot;
Andrea Arcangeliba761492011-01-13 15:46:58 -08001926
Sasha Levinb67bfe02013-02-27 17:06:00 -08001927 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
Andrea Arcangeliba761492011-01-13 15:46:58 -08001928 if (mm == mm_slot->mm)
1929 return mm_slot;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08001930
Andrea Arcangeliba761492011-01-13 15:46:58 -08001931 return NULL;
1932}
1933
1934static void insert_to_mm_slots_hash(struct mm_struct *mm,
1935 struct mm_slot *mm_slot)
1936{
Andrea Arcangeliba761492011-01-13 15:46:58 -08001937 mm_slot->mm = mm;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08001938 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001939}
1940
1941static inline int khugepaged_test_exit(struct mm_struct *mm)
1942{
1943 return atomic_read(&mm->mm_users) == 0;
1944}
1945
1946int __khugepaged_enter(struct mm_struct *mm)
1947{
1948 struct mm_slot *mm_slot;
1949 int wakeup;
1950
1951 mm_slot = alloc_mm_slot();
1952 if (!mm_slot)
1953 return -ENOMEM;
1954
1955 /* __khugepaged_exit() must not run from under us */
1956 VM_BUG_ON(khugepaged_test_exit(mm));
1957 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1958 free_mm_slot(mm_slot);
1959 return 0;
1960 }
1961
1962 spin_lock(&khugepaged_mm_lock);
1963 insert_to_mm_slots_hash(mm, mm_slot);
1964 /*
1965 * Insert just behind the scanning cursor, to let the area settle
1966 * down a little.
1967 */
1968 wakeup = list_empty(&khugepaged_scan.mm_head);
1969 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1970 spin_unlock(&khugepaged_mm_lock);
1971
1972 atomic_inc(&mm->mm_count);
1973 if (wakeup)
1974 wake_up_interruptible(&khugepaged_wait);
1975
1976 return 0;
1977}
1978
1979int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1980{
1981 unsigned long hstart, hend;
1982 if (!vma->anon_vma)
1983 /*
1984 * Not yet faulted in so we will register later in the
1985 * page fault if needed.
1986 */
1987 return 0;
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001988 if (vma->vm_ops)
Andrea Arcangeliba761492011-01-13 15:46:58 -08001989 /* khugepaged not yet working on file or special mappings */
1990 return 0;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001991 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001992 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1993 hend = vma->vm_end & HPAGE_PMD_MASK;
1994 if (hstart < hend)
1995 return khugepaged_enter(vma);
1996 return 0;
1997}
1998
1999void __khugepaged_exit(struct mm_struct *mm)
2000{
2001 struct mm_slot *mm_slot;
2002 int free = 0;
2003
2004 spin_lock(&khugepaged_mm_lock);
2005 mm_slot = get_mm_slot(mm);
2006 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002007 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002008 list_del(&mm_slot->mm_node);
2009 free = 1;
2010 }
Chris Wrightd788e802011-07-25 17:12:14 -07002011 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002012
2013 if (free) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002014 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2015 free_mm_slot(mm_slot);
2016 mmdrop(mm);
2017 } else if (mm_slot) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002018 /*
2019 * This is required to serialize against
2020 * khugepaged_test_exit() (which is guaranteed to run
2021 * under mmap sem read mode). Stop here (after we
2022 * return all pagetables will be destroyed) until
2023 * khugepaged has finished working on the pagetables
2024 * under the mmap_sem.
2025 */
2026 down_write(&mm->mmap_sem);
2027 up_write(&mm->mmap_sem);
Chris Wrightd788e802011-07-25 17:12:14 -07002028 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002029}
2030
2031static void release_pte_page(struct page *page)
2032{
2033 /* 0 stands for page_is_file_cache(page) == false */
2034 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2035 unlock_page(page);
2036 putback_lru_page(page);
2037}
2038
2039static void release_pte_pages(pte_t *pte, pte_t *_pte)
2040{
2041 while (--_pte >= pte) {
2042 pte_t pteval = *_pte;
2043 if (!pte_none(pteval))
2044 release_pte_page(pte_page(pteval));
2045 }
2046}
2047
Andrea Arcangeliba761492011-01-13 15:46:58 -08002048static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2049 unsigned long address,
2050 pte_t *pte)
2051{
2052 struct page *page;
2053 pte_t *_pte;
Bob Liu344aa352012-12-11 16:00:34 -08002054 int referenced = 0, none = 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002055 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2056 _pte++, address += PAGE_SIZE) {
2057 pte_t pteval = *_pte;
2058 if (pte_none(pteval)) {
2059 if (++none <= khugepaged_max_ptes_none)
2060 continue;
Bob Liu344aa352012-12-11 16:00:34 -08002061 else
Andrea Arcangeliba761492011-01-13 15:46:58 -08002062 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002063 }
Bob Liu344aa352012-12-11 16:00:34 -08002064 if (!pte_present(pteval) || !pte_write(pteval))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002065 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002066 page = vm_normal_page(vma, address, pteval);
Bob Liu344aa352012-12-11 16:00:34 -08002067 if (unlikely(!page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002068 goto out;
Bob Liu344aa352012-12-11 16:00:34 -08002069
Andrea Arcangeliba761492011-01-13 15:46:58 -08002070 VM_BUG_ON(PageCompound(page));
2071 BUG_ON(!PageAnon(page));
2072 VM_BUG_ON(!PageSwapBacked(page));
2073
2074 /* cannot use mapcount: can't collapse if there's a gup pin */
Bob Liu344aa352012-12-11 16:00:34 -08002075 if (page_count(page) != 1)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002076 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002077 /*
2078 * We can do it before isolate_lru_page because the
2079 * page can't be freed from under us. NOTE: PG_lock
2080 * is needed to serialize against split_huge_page
2081 * when invoked from the VM.
2082 */
Bob Liu344aa352012-12-11 16:00:34 -08002083 if (!trylock_page(page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002084 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002085 /*
2086 * Isolate the page to avoid collapsing an hugepage
2087 * currently in use by the VM.
2088 */
2089 if (isolate_lru_page(page)) {
2090 unlock_page(page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002091 goto out;
2092 }
2093 /* 0 stands for page_is_file_cache(page) == false */
2094 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2095 VM_BUG_ON(!PageLocked(page));
2096 VM_BUG_ON(PageLRU(page));
2097
2098 /* If there is no mapped pte young don't collapse the page */
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002099 if (pte_young(pteval) || PageReferenced(page) ||
2100 mmu_notifier_test_young(vma->vm_mm, address))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002101 referenced = 1;
2102 }
Bob Liu344aa352012-12-11 16:00:34 -08002103 if (likely(referenced))
2104 return 1;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002105out:
Bob Liu344aa352012-12-11 16:00:34 -08002106 release_pte_pages(pte, _pte);
2107 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002108}
2109
2110static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2111 struct vm_area_struct *vma,
2112 unsigned long address,
2113 spinlock_t *ptl)
2114{
2115 pte_t *_pte;
2116 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2117 pte_t pteval = *_pte;
2118 struct page *src_page;
2119
2120 if (pte_none(pteval)) {
2121 clear_user_highpage(page, address);
2122 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2123 } else {
2124 src_page = pte_page(pteval);
2125 copy_user_highpage(page, src_page, address, vma);
2126 VM_BUG_ON(page_mapcount(src_page) != 1);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002127 release_pte_page(src_page);
2128 /*
2129 * ptl mostly unnecessary, but preempt has to
2130 * be disabled to update the per-cpu stats
2131 * inside page_remove_rmap().
2132 */
2133 spin_lock(ptl);
2134 /*
2135 * paravirt calls inside pte_clear here are
2136 * superfluous.
2137 */
2138 pte_clear(vma->vm_mm, address, _pte);
2139 page_remove_rmap(src_page);
2140 spin_unlock(ptl);
2141 free_page_and_swap_cache(src_page);
2142 }
2143
2144 address += PAGE_SIZE;
2145 page++;
2146 }
2147}
2148
Xiao Guangrong26234f32012-10-08 16:29:51 -07002149static void khugepaged_alloc_sleep(void)
2150{
2151 wait_event_freezable_timeout(khugepaged_wait, false,
2152 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2153}
2154
2155#ifdef CONFIG_NUMA
2156static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2157{
2158 if (IS_ERR(*hpage)) {
2159 if (!*wait)
2160 return false;
2161
2162 *wait = false;
Xiao Guangronge3b41262012-10-08 16:32:57 -07002163 *hpage = NULL;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002164 khugepaged_alloc_sleep();
2165 } else if (*hpage) {
2166 put_page(*hpage);
2167 *hpage = NULL;
2168 }
2169
2170 return true;
2171}
2172
2173static struct page
2174*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2175 struct vm_area_struct *vma, unsigned long address,
2176 int node)
2177{
2178 VM_BUG_ON(*hpage);
2179 /*
2180 * Allocate the page while the vma is still valid and under
2181 * the mmap_sem read mode so there is no memory allocation
2182 * later when we take the mmap_sem in write mode. This is more
2183 * friendly behavior (OTOH it may actually hide bugs) to
2184 * filesystems in userland with daemons allocating memory in
2185 * the userland I/O paths. Allocating memory with the
2186 * mmap_sem in read mode is good idea also to allow greater
2187 * scalability.
2188 */
2189 *hpage = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2190 node, __GFP_OTHER_NODE);
2191
2192 /*
2193 * After allocating the hugepage, release the mmap_sem read lock in
2194 * preparation for taking it in write mode.
2195 */
2196 up_read(&mm->mmap_sem);
2197 if (unlikely(!*hpage)) {
2198 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2199 *hpage = ERR_PTR(-ENOMEM);
2200 return NULL;
2201 }
2202
2203 count_vm_event(THP_COLLAPSE_ALLOC);
2204 return *hpage;
2205}
2206#else
2207static struct page *khugepaged_alloc_hugepage(bool *wait)
2208{
2209 struct page *hpage;
2210
2211 do {
2212 hpage = alloc_hugepage(khugepaged_defrag());
2213 if (!hpage) {
2214 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2215 if (!*wait)
2216 return NULL;
2217
2218 *wait = false;
2219 khugepaged_alloc_sleep();
2220 } else
2221 count_vm_event(THP_COLLAPSE_ALLOC);
2222 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2223
2224 return hpage;
2225}
2226
2227static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2228{
2229 if (!*hpage)
2230 *hpage = khugepaged_alloc_hugepage(wait);
2231
2232 if (unlikely(!*hpage))
2233 return false;
2234
2235 return true;
2236}
2237
2238static struct page
2239*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2240 struct vm_area_struct *vma, unsigned long address,
2241 int node)
2242{
2243 up_read(&mm->mmap_sem);
2244 VM_BUG_ON(!*hpage);
2245 return *hpage;
2246}
2247#endif
2248
Bob Liufa475e52012-12-11 16:00:39 -08002249static bool hugepage_vma_check(struct vm_area_struct *vma)
2250{
2251 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2252 (vma->vm_flags & VM_NOHUGEPAGE))
2253 return false;
2254
2255 if (!vma->anon_vma || vma->vm_ops)
2256 return false;
2257 if (is_vma_temporary_stack(vma))
2258 return false;
2259 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2260 return true;
2261}
2262
Andrea Arcangeliba761492011-01-13 15:46:58 -08002263static void collapse_huge_page(struct mm_struct *mm,
Xiao Guangrong26234f32012-10-08 16:29:51 -07002264 unsigned long address,
2265 struct page **hpage,
2266 struct vm_area_struct *vma,
2267 int node)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002268{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002269 pmd_t *pmd, _pmd;
2270 pte_t *pte;
2271 pgtable_t pgtable;
2272 struct page *new_page;
2273 spinlock_t *ptl;
2274 int isolated;
2275 unsigned long hstart, hend;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002276 unsigned long mmun_start; /* For mmu_notifiers */
2277 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002278
2279 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002280
Xiao Guangrong26234f32012-10-08 16:29:51 -07002281 /* release the mmap_sem read lock. */
2282 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2283 if (!new_page)
Andrea Arcangelice83d212011-01-13 15:47:06 -08002284 return;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002285
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002286 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002287 return;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002288
2289 /*
2290 * Prevent all access to pagetables with the exception of
2291 * gup_fast later hanlded by the ptep_clear_flush and the VM
2292 * handled by the anon_vma lock + PG_lock.
2293 */
2294 down_write(&mm->mmap_sem);
2295 if (unlikely(khugepaged_test_exit(mm)))
2296 goto out;
2297
2298 vma = find_vma(mm, address);
Libina8f531eb2013-09-11 14:20:38 -07002299 if (!vma)
2300 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002301 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2302 hend = vma->vm_end & HPAGE_PMD_MASK;
2303 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2304 goto out;
Bob Liufa475e52012-12-11 16:00:39 -08002305 if (!hugepage_vma_check(vma))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002306 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002307 pmd = mm_find_pmd(mm, address);
2308 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002309 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002310 if (pmd_trans_huge(*pmd))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002311 goto out;
2312
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +00002313 anon_vma_lock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002314
2315 pte = pte_offset_map(pmd, address);
2316 ptl = pte_lockptr(mm, pmd);
2317
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002318 mmun_start = address;
2319 mmun_end = address + HPAGE_PMD_SIZE;
2320 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002321 spin_lock(&mm->page_table_lock); /* probably unnecessary */
2322 /*
2323 * After this gup_fast can't run anymore. This also removes
2324 * any huge TLB entry from the CPU so we won't allow
2325 * huge and small TLB entries for the same virtual address
2326 * to avoid the risk of CPU bugs in that area.
2327 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002328 _pmd = pmdp_clear_flush(vma, address, pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002329 spin_unlock(&mm->page_table_lock);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002330 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002331
2332 spin_lock(ptl);
2333 isolated = __collapse_huge_page_isolate(vma, address, pte);
2334 spin_unlock(ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002335
2336 if (unlikely(!isolated)) {
Johannes Weiner453c7192011-01-20 14:44:18 -08002337 pte_unmap(pte);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002338 spin_lock(&mm->page_table_lock);
2339 BUG_ON(!pmd_none(*pmd));
Aneesh Kumar K.V7c342512013-05-24 15:55:21 -07002340 /*
2341 * We can only use set_pmd_at when establishing
2342 * hugepmds and never for establishing regular pmds that
2343 * points to regular pagetables. Use pmd_populate for that
2344 */
2345 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002346 spin_unlock(&mm->page_table_lock);
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002347 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002348 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002349 }
2350
2351 /*
2352 * All pages are isolated and locked so anon_vma rmap
2353 * can't run anymore.
2354 */
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002355 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002356
2357 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
Johannes Weiner453c7192011-01-20 14:44:18 -08002358 pte_unmap(pte);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002359 __SetPageUptodate(new_page);
2360 pgtable = pmd_pgtable(_pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002361
Kirill A. Shutemov31223592013-09-12 15:14:01 -07002362 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2363 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002364
2365 /*
2366 * spin_lock() below is not the equivalent of smp_wmb(), so
2367 * this is needed to avoid the copy_huge_page writes to become
2368 * visible after the set_pmd_at() write.
2369 */
2370 smp_wmb();
2371
2372 spin_lock(&mm->page_table_lock);
2373 BUG_ON(!pmd_none(*pmd));
2374 page_add_new_anon_rmap(new_page, vma, address);
Aneesh Kumar K.Vfce144b2013-06-05 17:14:06 -07002375 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002376 set_pmd_at(mm, address, pmd, _pmd);
David Millerb113da62012-10-08 16:34:25 -07002377 update_mmu_cache_pmd(vma, address, pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002378 spin_unlock(&mm->page_table_lock);
2379
2380 *hpage = NULL;
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002381
Andrea Arcangeliba761492011-01-13 15:46:58 -08002382 khugepaged_pages_collapsed++;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002383out_up_write:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002384 up_write(&mm->mmap_sem);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002385 return;
2386
Andrea Arcangelice83d212011-01-13 15:47:06 -08002387out:
KAMEZAWA Hiroyuki678ff892011-02-10 15:01:36 -08002388 mem_cgroup_uncharge_page(new_page);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002389 goto out_up_write;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002390}
2391
2392static int khugepaged_scan_pmd(struct mm_struct *mm,
2393 struct vm_area_struct *vma,
2394 unsigned long address,
2395 struct page **hpage)
2396{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002397 pmd_t *pmd;
2398 pte_t *pte, *_pte;
2399 int ret = 0, referenced = 0, none = 0;
2400 struct page *page;
2401 unsigned long _address;
2402 spinlock_t *ptl;
David Rientjes00ef2d22013-02-22 16:35:36 -08002403 int node = NUMA_NO_NODE;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002404
2405 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2406
Bob Liu62190492012-12-11 16:00:37 -08002407 pmd = mm_find_pmd(mm, address);
2408 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002409 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002410 if (pmd_trans_huge(*pmd))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002411 goto out;
2412
2413 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2414 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2415 _pte++, _address += PAGE_SIZE) {
2416 pte_t pteval = *_pte;
2417 if (pte_none(pteval)) {
2418 if (++none <= khugepaged_max_ptes_none)
2419 continue;
2420 else
2421 goto out_unmap;
2422 }
2423 if (!pte_present(pteval) || !pte_write(pteval))
2424 goto out_unmap;
2425 page = vm_normal_page(vma, _address, pteval);
2426 if (unlikely(!page))
2427 goto out_unmap;
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002428 /*
2429 * Chose the node of the first page. This could
2430 * be more sophisticated and look at more pages,
2431 * but isn't for now.
2432 */
David Rientjes00ef2d22013-02-22 16:35:36 -08002433 if (node == NUMA_NO_NODE)
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002434 node = page_to_nid(page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002435 VM_BUG_ON(PageCompound(page));
2436 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2437 goto out_unmap;
2438 /* cannot use mapcount: can't collapse if there's a gup pin */
2439 if (page_count(page) != 1)
2440 goto out_unmap;
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002441 if (pte_young(pteval) || PageReferenced(page) ||
2442 mmu_notifier_test_young(vma->vm_mm, address))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002443 referenced = 1;
2444 }
2445 if (referenced)
2446 ret = 1;
2447out_unmap:
2448 pte_unmap_unlock(pte, ptl);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002449 if (ret)
2450 /* collapse_huge_page will return with the mmap_sem released */
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002451 collapse_huge_page(mm, address, hpage, vma, node);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002452out:
2453 return ret;
2454}
2455
2456static void collect_mm_slot(struct mm_slot *mm_slot)
2457{
2458 struct mm_struct *mm = mm_slot->mm;
2459
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002460 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002461
2462 if (khugepaged_test_exit(mm)) {
2463 /* free mm_slot */
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002464 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002465 list_del(&mm_slot->mm_node);
2466
2467 /*
2468 * Not strictly needed because the mm exited already.
2469 *
2470 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2471 */
2472
2473 /* khugepaged_mm_lock actually not necessary for the below */
2474 free_mm_slot(mm_slot);
2475 mmdrop(mm);
2476 }
2477}
2478
2479static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2480 struct page **hpage)
H Hartley Sweeten2f1da642011-10-31 17:09:25 -07002481 __releases(&khugepaged_mm_lock)
2482 __acquires(&khugepaged_mm_lock)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002483{
2484 struct mm_slot *mm_slot;
2485 struct mm_struct *mm;
2486 struct vm_area_struct *vma;
2487 int progress = 0;
2488
2489 VM_BUG_ON(!pages);
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002490 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002491
2492 if (khugepaged_scan.mm_slot)
2493 mm_slot = khugepaged_scan.mm_slot;
2494 else {
2495 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2496 struct mm_slot, mm_node);
2497 khugepaged_scan.address = 0;
2498 khugepaged_scan.mm_slot = mm_slot;
2499 }
2500 spin_unlock(&khugepaged_mm_lock);
2501
2502 mm = mm_slot->mm;
2503 down_read(&mm->mmap_sem);
2504 if (unlikely(khugepaged_test_exit(mm)))
2505 vma = NULL;
2506 else
2507 vma = find_vma(mm, khugepaged_scan.address);
2508
2509 progress++;
2510 for (; vma; vma = vma->vm_next) {
2511 unsigned long hstart, hend;
2512
2513 cond_resched();
2514 if (unlikely(khugepaged_test_exit(mm))) {
2515 progress++;
2516 break;
2517 }
Bob Liufa475e52012-12-11 16:00:39 -08002518 if (!hugepage_vma_check(vma)) {
2519skip:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002520 progress++;
2521 continue;
2522 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002523 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2524 hend = vma->vm_end & HPAGE_PMD_MASK;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002525 if (hstart >= hend)
2526 goto skip;
2527 if (khugepaged_scan.address > hend)
2528 goto skip;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002529 if (khugepaged_scan.address < hstart)
2530 khugepaged_scan.address = hstart;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002531 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002532
2533 while (khugepaged_scan.address < hend) {
2534 int ret;
2535 cond_resched();
2536 if (unlikely(khugepaged_test_exit(mm)))
2537 goto breakouterloop;
2538
2539 VM_BUG_ON(khugepaged_scan.address < hstart ||
2540 khugepaged_scan.address + HPAGE_PMD_SIZE >
2541 hend);
2542 ret = khugepaged_scan_pmd(mm, vma,
2543 khugepaged_scan.address,
2544 hpage);
2545 /* move to next address */
2546 khugepaged_scan.address += HPAGE_PMD_SIZE;
2547 progress += HPAGE_PMD_NR;
2548 if (ret)
2549 /* we released mmap_sem so break loop */
2550 goto breakouterloop_mmap_sem;
2551 if (progress >= pages)
2552 goto breakouterloop;
2553 }
2554 }
2555breakouterloop:
2556 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2557breakouterloop_mmap_sem:
2558
2559 spin_lock(&khugepaged_mm_lock);
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002560 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002561 /*
2562 * Release the current mm_slot if this mm is about to die, or
2563 * if we scanned all vmas of this mm.
2564 */
2565 if (khugepaged_test_exit(mm) || !vma) {
2566 /*
2567 * Make sure that if mm_users is reaching zero while
2568 * khugepaged runs here, khugepaged_exit will find
2569 * mm_slot not pointing to the exiting mm.
2570 */
2571 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2572 khugepaged_scan.mm_slot = list_entry(
2573 mm_slot->mm_node.next,
2574 struct mm_slot, mm_node);
2575 khugepaged_scan.address = 0;
2576 } else {
2577 khugepaged_scan.mm_slot = NULL;
2578 khugepaged_full_scans++;
2579 }
2580
2581 collect_mm_slot(mm_slot);
2582 }
2583
2584 return progress;
2585}
2586
2587static int khugepaged_has_work(void)
2588{
2589 return !list_empty(&khugepaged_scan.mm_head) &&
2590 khugepaged_enabled();
2591}
2592
2593static int khugepaged_wait_event(void)
2594{
2595 return !list_empty(&khugepaged_scan.mm_head) ||
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002596 kthread_should_stop();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002597}
2598
Xiao Guangrongd5169042012-10-08 16:29:48 -07002599static void khugepaged_do_scan(void)
2600{
2601 struct page *hpage = NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002602 unsigned int progress = 0, pass_through_head = 0;
2603 unsigned int pages = khugepaged_pages_to_scan;
Xiao Guangrongd5169042012-10-08 16:29:48 -07002604 bool wait = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002605
2606 barrier(); /* write khugepaged_pages_to_scan to local stack */
2607
2608 while (progress < pages) {
Xiao Guangrong26234f32012-10-08 16:29:51 -07002609 if (!khugepaged_prealloc_page(&hpage, &wait))
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002610 break;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002611
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002612 cond_resched();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002613
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002614 if (unlikely(kthread_should_stop() || freezing(current)))
2615 break;
2616
Andrea Arcangeliba761492011-01-13 15:46:58 -08002617 spin_lock(&khugepaged_mm_lock);
2618 if (!khugepaged_scan.mm_slot)
2619 pass_through_head++;
2620 if (khugepaged_has_work() &&
2621 pass_through_head < 2)
2622 progress += khugepaged_scan_mm_slot(pages - progress,
Xiao Guangrongd5169042012-10-08 16:29:48 -07002623 &hpage);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002624 else
2625 progress = pages;
2626 spin_unlock(&khugepaged_mm_lock);
2627 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002628
Xiao Guangrongd5169042012-10-08 16:29:48 -07002629 if (!IS_ERR_OR_NULL(hpage))
2630 put_page(hpage);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002631}
2632
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002633static void khugepaged_wait_work(void)
2634{
2635 try_to_freeze();
2636
2637 if (khugepaged_has_work()) {
2638 if (!khugepaged_scan_sleep_millisecs)
2639 return;
2640
2641 wait_event_freezable_timeout(khugepaged_wait,
2642 kthread_should_stop(),
2643 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2644 return;
2645 }
2646
2647 if (khugepaged_enabled())
2648 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2649}
2650
Andrea Arcangeliba761492011-01-13 15:46:58 -08002651static int khugepaged(void *none)
2652{
2653 struct mm_slot *mm_slot;
2654
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002655 set_freezable();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002656 set_user_nice(current, 19);
2657
Xiao Guangrongb7231782012-10-08 16:29:54 -07002658 while (!kthread_should_stop()) {
2659 khugepaged_do_scan();
2660 khugepaged_wait_work();
2661 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002662
2663 spin_lock(&khugepaged_mm_lock);
2664 mm_slot = khugepaged_scan.mm_slot;
2665 khugepaged_scan.mm_slot = NULL;
2666 if (mm_slot)
2667 collect_mm_slot(mm_slot);
2668 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002669 return 0;
2670}
2671
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002672static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2673 unsigned long haddr, pmd_t *pmd)
2674{
2675 struct mm_struct *mm = vma->vm_mm;
2676 pgtable_t pgtable;
2677 pmd_t _pmd;
2678 int i;
2679
2680 pmdp_clear_flush(vma, haddr, pmd);
2681 /* leave pmd empty until pte is filled */
2682
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07002683 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002684 pmd_populate(mm, &_pmd, pgtable);
2685
2686 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2687 pte_t *pte, entry;
2688 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2689 entry = pte_mkspecial(entry);
2690 pte = pte_offset_map(&_pmd, haddr);
2691 VM_BUG_ON(!pte_none(*pte));
2692 set_pte_at(mm, haddr, pte, entry);
2693 pte_unmap(pte);
2694 }
2695 smp_wmb(); /* make pte visible before pmd */
2696 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08002697 put_huge_zero_page();
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002698}
2699
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002700void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2701 pmd_t *pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002702{
2703 struct page *page;
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002704 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002705 unsigned long haddr = address & HPAGE_PMD_MASK;
2706 unsigned long mmun_start; /* For mmu_notifiers */
2707 unsigned long mmun_end; /* For mmu_notifiers */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002708
2709 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002710
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002711 mmun_start = haddr;
2712 mmun_end = haddr + HPAGE_PMD_SIZE;
2713 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002714 spin_lock(&mm->page_table_lock);
2715 if (unlikely(!pmd_trans_huge(*pmd))) {
2716 spin_unlock(&mm->page_table_lock);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002717 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2718 return;
2719 }
2720 if (is_huge_zero_pmd(*pmd)) {
2721 __split_huge_zero_page_pmd(vma, haddr, pmd);
2722 spin_unlock(&mm->page_table_lock);
2723 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002724 return;
2725 }
2726 page = pmd_page(*pmd);
2727 VM_BUG_ON(!page_count(page));
2728 get_page(page);
2729 spin_unlock(&mm->page_table_lock);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002730 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002731
2732 split_huge_page(page);
2733
2734 put_page(page);
2735 BUG_ON(pmd_trans_huge(*pmd));
2736}
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002737
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002738void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2739 pmd_t *pmd)
2740{
2741 struct vm_area_struct *vma;
2742
2743 vma = find_vma(mm, address);
2744 BUG_ON(vma == NULL);
2745 split_huge_page_pmd(vma, address, pmd);
2746}
2747
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002748static void split_huge_page_address(struct mm_struct *mm,
2749 unsigned long address)
2750{
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002751 pmd_t *pmd;
2752
2753 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2754
Bob Liu62190492012-12-11 16:00:37 -08002755 pmd = mm_find_pmd(mm, address);
2756 if (!pmd)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002757 return;
2758 /*
2759 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2760 * materialize from under us.
2761 */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002762 split_huge_page_pmd_mm(mm, address, pmd);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002763}
2764
2765void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2766 unsigned long start,
2767 unsigned long end,
2768 long adjust_next)
2769{
2770 /*
2771 * If the new start address isn't hpage aligned and it could
2772 * previously contain an hugepage: check if we need to split
2773 * an huge pmd.
2774 */
2775 if (start & ~HPAGE_PMD_MASK &&
2776 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2777 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2778 split_huge_page_address(vma->vm_mm, start);
2779
2780 /*
2781 * If the new end address isn't hpage aligned and it could
2782 * previously contain an hugepage: check if we need to split
2783 * an huge pmd.
2784 */
2785 if (end & ~HPAGE_PMD_MASK &&
2786 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2787 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2788 split_huge_page_address(vma->vm_mm, end);
2789
2790 /*
2791 * If we're also updating the vma->vm_next->vm_start, if the new
2792 * vm_next->vm_start isn't page aligned and it could previously
2793 * contain an hugepage: check if we need to split an huge pmd.
2794 */
2795 if (adjust_next > 0) {
2796 struct vm_area_struct *next = vma->vm_next;
2797 unsigned long nstart = next->vm_start;
2798 nstart += adjust_next << PAGE_SHIFT;
2799 if (nstart & ~HPAGE_PMD_MASK &&
2800 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2801 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2802 split_huge_page_address(next->vm_mm, nstart);
2803 }
2804}