blob: 8f7fc394f63672954cc4444da0f2158a42470b2c [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>
Andrea Arcangeliba761492011-01-13 15:46:58 -080015#include <linux/mm_inline.h>
16#include <linux/kthread.h>
17#include <linux/khugepaged.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080018#include <linux/freezer.h>
Andrea Arcangelia664b2d2011-01-13 15:47:17 -080019#include <linux/mman.h>
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080020#include <asm/tlb.h>
21#include <asm/pgalloc.h>
22#include "internal.h"
23
Andrea Arcangeliba761492011-01-13 15:46:58 -080024/*
25 * By default transparent hugepage support is enabled for all mappings
26 * and khugepaged scans all mappings. Defrag is only invoked by
27 * khugepaged hugepage allocations and by page faults inside
28 * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
29 * allocations.
30 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080031unsigned long transparent_hugepage_flags __read_mostly =
Andrea Arcangeli13ece882011-01-13 15:47:07 -080032#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
Andrea Arcangeliba761492011-01-13 15:46:58 -080033 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
Andrea Arcangeli13ece882011-01-13 15:47:07 -080034#endif
35#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
36 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
37#endif
Andrea Arcangelid39d33c2011-01-13 15:47:05 -080038 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
Andrea Arcangeliba761492011-01-13 15:46:58 -080039 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
40
41/* default scan 8*512 pte (or vmas) every 30 second */
42static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
43static unsigned int khugepaged_pages_collapsed;
44static unsigned int khugepaged_full_scans;
45static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
46/* during fragmentation poll the hugepage allocator once every minute */
47static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
48static struct task_struct *khugepaged_thread __read_mostly;
49static DEFINE_MUTEX(khugepaged_mutex);
50static DEFINE_SPINLOCK(khugepaged_mm_lock);
51static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
52/*
53 * default collapse hugepages if there is at least one pte mapped like
54 * it would have happened if the vma was large enough during page
55 * fault.
56 */
57static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
58
59static int khugepaged(void *none);
60static int mm_slots_hash_init(void);
61static int khugepaged_slab_init(void);
62static void khugepaged_slab_free(void);
63
64#define MM_SLOTS_HASH_HEADS 1024
65static struct hlist_head *mm_slots_hash __read_mostly;
66static struct kmem_cache *mm_slot_cache __read_mostly;
67
68/**
69 * struct mm_slot - hash lookup from mm to mm_slot
70 * @hash: hash collision list
71 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
72 * @mm: the mm that this information is valid for
73 */
74struct mm_slot {
75 struct hlist_node hash;
76 struct list_head mm_node;
77 struct mm_struct *mm;
78};
79
80/**
81 * struct khugepaged_scan - cursor for scanning
82 * @mm_head: the head of the mm list to scan
83 * @mm_slot: the current mm_slot we are scanning
84 * @address: the next address inside that to be scanned
85 *
86 * There is only the one khugepaged_scan instance of this cursor structure.
87 */
88struct khugepaged_scan {
89 struct list_head mm_head;
90 struct mm_slot *mm_slot;
91 unsigned long address;
H Hartley Sweeten2f1da642011-10-31 17:09:25 -070092};
93static struct khugepaged_scan khugepaged_scan = {
Andrea Arcangeliba761492011-01-13 15:46:58 -080094 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
95};
96
Andrea Arcangelif0005652011-01-13 15:47:04 -080097
98static int set_recommended_min_free_kbytes(void)
99{
100 struct zone *zone;
101 int nr_zones = 0;
102 unsigned long recommended_min;
103 extern int min_free_kbytes;
104
105 if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG,
106 &transparent_hugepage_flags) &&
107 !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
108 &transparent_hugepage_flags))
109 return 0;
110
111 for_each_populated_zone(zone)
112 nr_zones++;
113
114 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
115 recommended_min = pageblock_nr_pages * nr_zones * 2;
116
117 /*
118 * Make sure that on average at least two pageblocks are almost free
119 * of another type, one for a migratetype to fall back to and a
120 * second to avoid subsequent fallbacks of other types There are 3
121 * MIGRATE_TYPES we care about.
122 */
123 recommended_min += pageblock_nr_pages * nr_zones *
124 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
125
126 /* don't ever allow to reserve more than 5% of the lowmem */
127 recommended_min = min(recommended_min,
128 (unsigned long) nr_free_buffer_pages() / 20);
129 recommended_min <<= (PAGE_SHIFT-10);
130
131 if (recommended_min > min_free_kbytes)
132 min_free_kbytes = recommended_min;
133 setup_per_zone_wmarks();
134 return 0;
135}
136late_initcall(set_recommended_min_free_kbytes);
137
Andrea Arcangeliba761492011-01-13 15:46:58 -0800138static int start_khugepaged(void)
139{
140 int err = 0;
141 if (khugepaged_enabled()) {
142 int wakeup;
143 if (unlikely(!mm_slot_cache || !mm_slots_hash)) {
144 err = -ENOMEM;
145 goto out;
146 }
147 mutex_lock(&khugepaged_mutex);
148 if (!khugepaged_thread)
149 khugepaged_thread = kthread_run(khugepaged, NULL,
150 "khugepaged");
151 if (unlikely(IS_ERR(khugepaged_thread))) {
152 printk(KERN_ERR
153 "khugepaged: kthread_run(khugepaged) failed\n");
154 err = PTR_ERR(khugepaged_thread);
155 khugepaged_thread = NULL;
156 }
157 wakeup = !list_empty(&khugepaged_scan.mm_head);
158 mutex_unlock(&khugepaged_mutex);
159 if (wakeup)
160 wake_up_interruptible(&khugepaged_wait);
Andrea Arcangelif0005652011-01-13 15:47:04 -0800161
162 set_recommended_min_free_kbytes();
Andrea Arcangeliba761492011-01-13 15:46:58 -0800163 } else
164 /* wakeup to exit */
165 wake_up_interruptible(&khugepaged_wait);
166out:
167 return err;
168}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800169
170#ifdef CONFIG_SYSFS
Andrea Arcangeliba761492011-01-13 15:46:58 -0800171
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800172static ssize_t double_flag_show(struct kobject *kobj,
173 struct kobj_attribute *attr, char *buf,
174 enum transparent_hugepage_flag enabled,
175 enum transparent_hugepage_flag req_madv)
176{
177 if (test_bit(enabled, &transparent_hugepage_flags)) {
178 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
179 return sprintf(buf, "[always] madvise never\n");
180 } else if (test_bit(req_madv, &transparent_hugepage_flags))
181 return sprintf(buf, "always [madvise] never\n");
182 else
183 return sprintf(buf, "always madvise [never]\n");
184}
185static ssize_t double_flag_store(struct kobject *kobj,
186 struct kobj_attribute *attr,
187 const char *buf, size_t count,
188 enum transparent_hugepage_flag enabled,
189 enum transparent_hugepage_flag req_madv)
190{
191 if (!memcmp("always", buf,
192 min(sizeof("always")-1, count))) {
193 set_bit(enabled, &transparent_hugepage_flags);
194 clear_bit(req_madv, &transparent_hugepage_flags);
195 } else if (!memcmp("madvise", buf,
196 min(sizeof("madvise")-1, count))) {
197 clear_bit(enabled, &transparent_hugepage_flags);
198 set_bit(req_madv, &transparent_hugepage_flags);
199 } else if (!memcmp("never", buf,
200 min(sizeof("never")-1, count))) {
201 clear_bit(enabled, &transparent_hugepage_flags);
202 clear_bit(req_madv, &transparent_hugepage_flags);
203 } else
204 return -EINVAL;
205
206 return count;
207}
208
209static ssize_t enabled_show(struct kobject *kobj,
210 struct kobj_attribute *attr, char *buf)
211{
212 return double_flag_show(kobj, attr, buf,
213 TRANSPARENT_HUGEPAGE_FLAG,
214 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
215}
216static ssize_t enabled_store(struct kobject *kobj,
217 struct kobj_attribute *attr,
218 const char *buf, size_t count)
219{
Andrea Arcangeliba761492011-01-13 15:46:58 -0800220 ssize_t ret;
221
222 ret = double_flag_store(kobj, attr, buf, count,
223 TRANSPARENT_HUGEPAGE_FLAG,
224 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
225
226 if (ret > 0) {
227 int err = start_khugepaged();
228 if (err)
229 ret = err;
230 }
231
Andrea Arcangelif0005652011-01-13 15:47:04 -0800232 if (ret > 0 &&
233 (test_bit(TRANSPARENT_HUGEPAGE_FLAG,
234 &transparent_hugepage_flags) ||
235 test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
236 &transparent_hugepage_flags)))
237 set_recommended_min_free_kbytes();
238
Andrea Arcangeliba761492011-01-13 15:46:58 -0800239 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800240}
241static struct kobj_attribute enabled_attr =
242 __ATTR(enabled, 0644, enabled_show, enabled_store);
243
244static ssize_t single_flag_show(struct kobject *kobj,
245 struct kobj_attribute *attr, char *buf,
246 enum transparent_hugepage_flag flag)
247{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700248 return sprintf(buf, "%d\n",
249 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800250}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700251
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800252static ssize_t single_flag_store(struct kobject *kobj,
253 struct kobj_attribute *attr,
254 const char *buf, size_t count,
255 enum transparent_hugepage_flag flag)
256{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700257 unsigned long value;
258 int ret;
259
260 ret = kstrtoul(buf, 10, &value);
261 if (ret < 0)
262 return ret;
263 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800264 return -EINVAL;
265
Ben Hutchingse27e6152011-04-14 15:22:21 -0700266 if (value)
267 set_bit(flag, &transparent_hugepage_flags);
268 else
269 clear_bit(flag, &transparent_hugepage_flags);
270
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800271 return count;
272}
273
274/*
275 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
276 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
277 * memory just to allocate one more hugepage.
278 */
279static ssize_t defrag_show(struct kobject *kobj,
280 struct kobj_attribute *attr, char *buf)
281{
282 return double_flag_show(kobj, attr, buf,
283 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
284 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
285}
286static ssize_t defrag_store(struct kobject *kobj,
287 struct kobj_attribute *attr,
288 const char *buf, size_t count)
289{
290 return double_flag_store(kobj, attr, buf, count,
291 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
292 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
293}
294static struct kobj_attribute defrag_attr =
295 __ATTR(defrag, 0644, defrag_show, defrag_store);
296
297#ifdef CONFIG_DEBUG_VM
298static ssize_t debug_cow_show(struct kobject *kobj,
299 struct kobj_attribute *attr, char *buf)
300{
301 return single_flag_show(kobj, attr, buf,
302 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
303}
304static ssize_t debug_cow_store(struct kobject *kobj,
305 struct kobj_attribute *attr,
306 const char *buf, size_t count)
307{
308 return single_flag_store(kobj, attr, buf, count,
309 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
310}
311static struct kobj_attribute debug_cow_attr =
312 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
313#endif /* CONFIG_DEBUG_VM */
314
315static struct attribute *hugepage_attr[] = {
316 &enabled_attr.attr,
317 &defrag_attr.attr,
318#ifdef CONFIG_DEBUG_VM
319 &debug_cow_attr.attr,
320#endif
321 NULL,
322};
323
324static struct attribute_group hugepage_attr_group = {
325 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800326};
327
328static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
329 struct kobj_attribute *attr,
330 char *buf)
331{
332 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
333}
334
335static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
336 struct kobj_attribute *attr,
337 const char *buf, size_t count)
338{
339 unsigned long msecs;
340 int err;
341
342 err = strict_strtoul(buf, 10, &msecs);
343 if (err || msecs > UINT_MAX)
344 return -EINVAL;
345
346 khugepaged_scan_sleep_millisecs = msecs;
347 wake_up_interruptible(&khugepaged_wait);
348
349 return count;
350}
351static struct kobj_attribute scan_sleep_millisecs_attr =
352 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
353 scan_sleep_millisecs_store);
354
355static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
356 struct kobj_attribute *attr,
357 char *buf)
358{
359 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
360}
361
362static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
363 struct kobj_attribute *attr,
364 const char *buf, size_t count)
365{
366 unsigned long msecs;
367 int err;
368
369 err = strict_strtoul(buf, 10, &msecs);
370 if (err || msecs > UINT_MAX)
371 return -EINVAL;
372
373 khugepaged_alloc_sleep_millisecs = msecs;
374 wake_up_interruptible(&khugepaged_wait);
375
376 return count;
377}
378static struct kobj_attribute alloc_sleep_millisecs_attr =
379 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
380 alloc_sleep_millisecs_store);
381
382static ssize_t pages_to_scan_show(struct kobject *kobj,
383 struct kobj_attribute *attr,
384 char *buf)
385{
386 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
387}
388static ssize_t pages_to_scan_store(struct kobject *kobj,
389 struct kobj_attribute *attr,
390 const char *buf, size_t count)
391{
392 int err;
393 unsigned long pages;
394
395 err = strict_strtoul(buf, 10, &pages);
396 if (err || !pages || pages > UINT_MAX)
397 return -EINVAL;
398
399 khugepaged_pages_to_scan = pages;
400
401 return count;
402}
403static struct kobj_attribute pages_to_scan_attr =
404 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
405 pages_to_scan_store);
406
407static ssize_t pages_collapsed_show(struct kobject *kobj,
408 struct kobj_attribute *attr,
409 char *buf)
410{
411 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
412}
413static struct kobj_attribute pages_collapsed_attr =
414 __ATTR_RO(pages_collapsed);
415
416static ssize_t full_scans_show(struct kobject *kobj,
417 struct kobj_attribute *attr,
418 char *buf)
419{
420 return sprintf(buf, "%u\n", khugepaged_full_scans);
421}
422static struct kobj_attribute full_scans_attr =
423 __ATTR_RO(full_scans);
424
425static ssize_t khugepaged_defrag_show(struct kobject *kobj,
426 struct kobj_attribute *attr, char *buf)
427{
428 return single_flag_show(kobj, attr, buf,
429 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
430}
431static ssize_t khugepaged_defrag_store(struct kobject *kobj,
432 struct kobj_attribute *attr,
433 const char *buf, size_t count)
434{
435 return single_flag_store(kobj, attr, buf, count,
436 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
437}
438static struct kobj_attribute khugepaged_defrag_attr =
439 __ATTR(defrag, 0644, khugepaged_defrag_show,
440 khugepaged_defrag_store);
441
442/*
443 * max_ptes_none controls if khugepaged should collapse hugepages over
444 * any unmapped ptes in turn potentially increasing the memory
445 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
446 * reduce the available free memory in the system as it
447 * runs. Increasing max_ptes_none will instead potentially reduce the
448 * free memory in the system during the khugepaged scan.
449 */
450static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
451 struct kobj_attribute *attr,
452 char *buf)
453{
454 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
455}
456static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
457 struct kobj_attribute *attr,
458 const char *buf, size_t count)
459{
460 int err;
461 unsigned long max_ptes_none;
462
463 err = strict_strtoul(buf, 10, &max_ptes_none);
464 if (err || max_ptes_none > HPAGE_PMD_NR-1)
465 return -EINVAL;
466
467 khugepaged_max_ptes_none = max_ptes_none;
468
469 return count;
470}
471static struct kobj_attribute khugepaged_max_ptes_none_attr =
472 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
473 khugepaged_max_ptes_none_store);
474
475static struct attribute *khugepaged_attr[] = {
476 &khugepaged_defrag_attr.attr,
477 &khugepaged_max_ptes_none_attr.attr,
478 &pages_to_scan_attr.attr,
479 &pages_collapsed_attr.attr,
480 &full_scans_attr.attr,
481 &scan_sleep_millisecs_attr.attr,
482 &alloc_sleep_millisecs_attr.attr,
483 NULL,
484};
485
486static struct attribute_group khugepaged_attr_group = {
487 .attrs = khugepaged_attr,
488 .name = "khugepaged",
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800489};
Shaohua Li569e5592012-01-12 17:19:11 -0800490
491static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
492{
493 int err;
494
495 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
496 if (unlikely(!*hugepage_kobj)) {
497 printk(KERN_ERR "hugepage: failed kobject create\n");
498 return -ENOMEM;
499 }
500
501 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
502 if (err) {
503 printk(KERN_ERR "hugepage: failed register hugeage group\n");
504 goto delete_obj;
505 }
506
507 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
508 if (err) {
509 printk(KERN_ERR "hugepage: failed register hugeage group\n");
510 goto remove_hp_group;
511 }
512
513 return 0;
514
515remove_hp_group:
516 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
517delete_obj:
518 kobject_put(*hugepage_kobj);
519 return err;
520}
521
522static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
523{
524 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
525 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
526 kobject_put(hugepage_kobj);
527}
528#else
529static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
530{
531 return 0;
532}
533
534static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
535{
536}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800537#endif /* CONFIG_SYSFS */
538
539static int __init hugepage_init(void)
540{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800541 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800542 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800543
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800544 if (!has_transparent_hugepage()) {
545 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800546 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800547 }
548
Shaohua Li569e5592012-01-12 17:19:11 -0800549 err = hugepage_init_sysfs(&hugepage_kobj);
550 if (err)
551 return err;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800552
553 err = khugepaged_slab_init();
554 if (err)
555 goto out;
556
557 err = mm_slots_hash_init();
558 if (err) {
559 khugepaged_slab_free();
560 goto out;
561 }
562
Rik van Riel97562cd2011-01-13 15:47:12 -0800563 /*
564 * By default disable transparent hugepages on smaller systems,
565 * where the extra memory used could hurt more than TLB overhead
566 * is likely to save. The admin can still enable it through /sys.
567 */
568 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
569 transparent_hugepage_flags = 0;
570
Andrea Arcangeliba761492011-01-13 15:46:58 -0800571 start_khugepaged();
572
Andrea Arcangelif0005652011-01-13 15:47:04 -0800573 set_recommended_min_free_kbytes();
574
Shaohua Li569e5592012-01-12 17:19:11 -0800575 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800576out:
Shaohua Li569e5592012-01-12 17:19:11 -0800577 hugepage_exit_sysfs(hugepage_kobj);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800578 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800579}
580module_init(hugepage_init)
581
582static int __init setup_transparent_hugepage(char *str)
583{
584 int ret = 0;
585 if (!str)
586 goto out;
587 if (!strcmp(str, "always")) {
588 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
589 &transparent_hugepage_flags);
590 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
591 &transparent_hugepage_flags);
592 ret = 1;
593 } else if (!strcmp(str, "madvise")) {
594 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
595 &transparent_hugepage_flags);
596 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
597 &transparent_hugepage_flags);
598 ret = 1;
599 } else if (!strcmp(str, "never")) {
600 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
601 &transparent_hugepage_flags);
602 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
603 &transparent_hugepage_flags);
604 ret = 1;
605 }
606out:
607 if (!ret)
608 printk(KERN_WARNING
609 "transparent_hugepage= cannot parse, ignored\n");
610 return ret;
611}
612__setup("transparent_hugepage=", setup_transparent_hugepage);
613
614static void prepare_pmd_huge_pte(pgtable_t pgtable,
615 struct mm_struct *mm)
616{
617 assert_spin_locked(&mm->page_table_lock);
618
619 /* FIFO */
620 if (!mm->pmd_huge_pte)
621 INIT_LIST_HEAD(&pgtable->lru);
622 else
623 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
624 mm->pmd_huge_pte = pgtable;
625}
626
627static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
628{
629 if (likely(vma->vm_flags & VM_WRITE))
630 pmd = pmd_mkwrite(pmd);
631 return pmd;
632}
633
634static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
635 struct vm_area_struct *vma,
636 unsigned long haddr, pmd_t *pmd,
637 struct page *page)
638{
639 int ret = 0;
640 pgtable_t pgtable;
641
642 VM_BUG_ON(!PageCompound(page));
643 pgtable = pte_alloc_one(mm, haddr);
644 if (unlikely(!pgtable)) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800645 mem_cgroup_uncharge_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800646 put_page(page);
647 return VM_FAULT_OOM;
648 }
649
650 clear_huge_page(page, haddr, HPAGE_PMD_NR);
651 __SetPageUptodate(page);
652
653 spin_lock(&mm->page_table_lock);
654 if (unlikely(!pmd_none(*pmd))) {
655 spin_unlock(&mm->page_table_lock);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800656 mem_cgroup_uncharge_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800657 put_page(page);
658 pte_free(mm, pgtable);
659 } else {
660 pmd_t entry;
661 entry = mk_pmd(page, vma->vm_page_prot);
662 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
663 entry = pmd_mkhuge(entry);
664 /*
665 * The spinlocking to take the lru_lock inside
666 * page_add_new_anon_rmap() acts as a full memory
667 * barrier to be sure clear_huge_page writes become
668 * visible after the set_pmd_at() write.
669 */
670 page_add_new_anon_rmap(page, vma, haddr);
671 set_pmd_at(mm, haddr, pmd, entry);
672 prepare_pmd_huge_pte(pgtable, mm);
673 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Andrea Arcangeli1c641e82012-03-05 14:59:20 -0800674 mm->nr_ptes++;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800675 spin_unlock(&mm->page_table_lock);
676 }
677
678 return ret;
679}
680
Andi Kleencc5d4622011-03-22 16:33:13 -0700681static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800682{
Andi Kleencc5d4622011-03-22 16:33:13 -0700683 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800684}
685
686static inline struct page *alloc_hugepage_vma(int defrag,
687 struct vm_area_struct *vma,
Andi Kleencc5d4622011-03-22 16:33:13 -0700688 unsigned long haddr, int nd,
689 gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800690{
Andi Kleencc5d4622011-03-22 16:33:13 -0700691 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
Andi Kleen5c4b4be2011-03-04 17:36:32 -0800692 HPAGE_PMD_ORDER, vma, haddr, nd);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800693}
694
695#ifndef CONFIG_NUMA
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800696static inline struct page *alloc_hugepage(int defrag)
697{
Andi Kleencc5d4622011-03-22 16:33:13 -0700698 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800699 HPAGE_PMD_ORDER);
700}
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800701#endif
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800702
703int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
704 unsigned long address, pmd_t *pmd,
705 unsigned int flags)
706{
707 struct page *page;
708 unsigned long haddr = address & HPAGE_PMD_MASK;
709 pte_t *pte;
710
711 if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
712 if (unlikely(anon_vma_prepare(vma)))
713 return VM_FAULT_OOM;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800714 if (unlikely(khugepaged_enter(vma)))
715 return VM_FAULT_OOM;
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800716 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
Andi Kleencc5d4622011-03-22 16:33:13 -0700717 vma, haddr, numa_node_id(), 0);
Andi Kleen81ab4202011-04-14 15:22:06 -0700718 if (unlikely(!page)) {
719 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800720 goto out;
Andi Kleen81ab4202011-04-14 15:22:06 -0700721 }
722 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800723 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
724 put_page(page);
725 goto out;
726 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800727
728 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
729 }
730out:
731 /*
732 * Use __pte_alloc instead of pte_alloc_map, because we can't
733 * run pte_offset_map on the pmd, if an huge pmd could
734 * materialize from under us from a different thread.
735 */
736 if (unlikely(__pte_alloc(mm, vma, pmd, address)))
737 return VM_FAULT_OOM;
738 /* if an huge pmd materialized from under us just retry later */
739 if (unlikely(pmd_trans_huge(*pmd)))
740 return 0;
741 /*
742 * A regular pmd is established and it can't morph into a huge pmd
743 * from under us anymore at this point because we hold the mmap_sem
744 * read mode and khugepaged takes it in write mode. So now it's
745 * safe to run pte_offset_map().
746 */
747 pte = pte_offset_map(pmd, address);
748 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
749}
750
751int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
752 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
753 struct vm_area_struct *vma)
754{
755 struct page *src_page;
756 pmd_t pmd;
757 pgtable_t pgtable;
758 int ret;
759
760 ret = -ENOMEM;
761 pgtable = pte_alloc_one(dst_mm, addr);
762 if (unlikely(!pgtable))
763 goto out;
764
765 spin_lock(&dst_mm->page_table_lock);
766 spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
767
768 ret = -EAGAIN;
769 pmd = *src_pmd;
770 if (unlikely(!pmd_trans_huge(pmd))) {
771 pte_free(dst_mm, pgtable);
772 goto out_unlock;
773 }
774 if (unlikely(pmd_trans_splitting(pmd))) {
775 /* split huge page running from under us */
776 spin_unlock(&src_mm->page_table_lock);
777 spin_unlock(&dst_mm->page_table_lock);
778 pte_free(dst_mm, pgtable);
779
780 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
781 goto out;
782 }
783 src_page = pmd_page(pmd);
784 VM_BUG_ON(!PageHead(src_page));
785 get_page(src_page);
786 page_dup_rmap(src_page);
787 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
788
789 pmdp_set_wrprotect(src_mm, addr, src_pmd);
790 pmd = pmd_mkold(pmd_wrprotect(pmd));
791 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
792 prepare_pmd_huge_pte(pgtable, dst_mm);
Andrea Arcangeli1c641e82012-03-05 14:59:20 -0800793 dst_mm->nr_ptes++;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800794
795 ret = 0;
796out_unlock:
797 spin_unlock(&src_mm->page_table_lock);
798 spin_unlock(&dst_mm->page_table_lock);
799out:
800 return ret;
801}
802
803/* no "address" argument so destroys page coloring of some arch */
804pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
805{
806 pgtable_t pgtable;
807
808 assert_spin_locked(&mm->page_table_lock);
809
810 /* FIFO */
811 pgtable = mm->pmd_huge_pte;
812 if (list_empty(&pgtable->lru))
813 mm->pmd_huge_pte = NULL;
814 else {
815 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
816 struct page, lru);
817 list_del(&pgtable->lru);
818 }
819 return pgtable;
820}
821
822static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
823 struct vm_area_struct *vma,
824 unsigned long address,
825 pmd_t *pmd, pmd_t orig_pmd,
826 struct page *page,
827 unsigned long haddr)
828{
829 pgtable_t pgtable;
830 pmd_t _pmd;
831 int ret = 0, i;
832 struct page **pages;
833
834 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
835 GFP_KERNEL);
836 if (unlikely(!pages)) {
837 ret |= VM_FAULT_OOM;
838 goto out;
839 }
840
841 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -0700842 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
843 __GFP_OTHER_NODE,
Andi Kleen19ee1512011-03-04 17:36:31 -0800844 vma, address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800845 if (unlikely(!pages[i] ||
846 mem_cgroup_newpage_charge(pages[i], mm,
847 GFP_KERNEL))) {
848 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800849 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800850 mem_cgroup_uncharge_start();
851 while (--i >= 0) {
852 mem_cgroup_uncharge_page(pages[i]);
853 put_page(pages[i]);
854 }
855 mem_cgroup_uncharge_end();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800856 kfree(pages);
857 ret |= VM_FAULT_OOM;
858 goto out;
859 }
860 }
861
862 for (i = 0; i < HPAGE_PMD_NR; i++) {
863 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -0700864 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800865 __SetPageUptodate(pages[i]);
866 cond_resched();
867 }
868
869 spin_lock(&mm->page_table_lock);
870 if (unlikely(!pmd_same(*pmd, orig_pmd)))
871 goto out_free_pages;
872 VM_BUG_ON(!PageHead(page));
873
874 pmdp_clear_flush_notify(vma, haddr, pmd);
875 /* leave pmd empty until pte is filled */
876
877 pgtable = get_pmd_huge_pte(mm);
878 pmd_populate(mm, &_pmd, pgtable);
879
880 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
881 pte_t *pte, entry;
882 entry = mk_pte(pages[i], vma->vm_page_prot);
883 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
884 page_add_new_anon_rmap(pages[i], vma, haddr);
885 pte = pte_offset_map(&_pmd, haddr);
886 VM_BUG_ON(!pte_none(*pte));
887 set_pte_at(mm, haddr, pte, entry);
888 pte_unmap(pte);
889 }
890 kfree(pages);
891
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800892 smp_wmb(); /* make pte visible before pmd */
893 pmd_populate(mm, pmd, pgtable);
894 page_remove_rmap(page);
895 spin_unlock(&mm->page_table_lock);
896
897 ret |= VM_FAULT_WRITE;
898 put_page(page);
899
900out:
901 return ret;
902
903out_free_pages:
904 spin_unlock(&mm->page_table_lock);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800905 mem_cgroup_uncharge_start();
906 for (i = 0; i < HPAGE_PMD_NR; i++) {
907 mem_cgroup_uncharge_page(pages[i]);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800908 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800909 }
910 mem_cgroup_uncharge_end();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800911 kfree(pages);
912 goto out;
913}
914
915int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
916 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
917{
918 int ret = 0;
919 struct page *page, *new_page;
920 unsigned long haddr;
921
922 VM_BUG_ON(!vma->anon_vma);
923 spin_lock(&mm->page_table_lock);
924 if (unlikely(!pmd_same(*pmd, orig_pmd)))
925 goto out_unlock;
926
927 page = pmd_page(orig_pmd);
928 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
929 haddr = address & HPAGE_PMD_MASK;
930 if (page_mapcount(page) == 1) {
931 pmd_t entry;
932 entry = pmd_mkyoung(orig_pmd);
933 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
934 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
935 update_mmu_cache(vma, address, entry);
936 ret |= VM_FAULT_WRITE;
937 goto out_unlock;
938 }
939 get_page(page);
940 spin_unlock(&mm->page_table_lock);
941
942 if (transparent_hugepage_enabled(vma) &&
943 !transparent_hugepage_debug_cow())
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800944 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
Andi Kleencc5d4622011-03-22 16:33:13 -0700945 vma, haddr, numa_node_id(), 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800946 else
947 new_page = NULL;
948
949 if (unlikely(!new_page)) {
Andi Kleen81ab4202011-04-14 15:22:06 -0700950 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800951 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
952 pmd, orig_pmd, page, haddr);
953 put_page(page);
954 goto out;
955 }
Andi Kleen81ab4202011-04-14 15:22:06 -0700956 count_vm_event(THP_FAULT_ALLOC);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800957
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800958 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
959 put_page(new_page);
960 put_page(page);
961 ret |= VM_FAULT_OOM;
962 goto out;
963 }
964
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800965 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
966 __SetPageUptodate(new_page);
967
968 spin_lock(&mm->page_table_lock);
969 put_page(page);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800970 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
971 mem_cgroup_uncharge_page(new_page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800972 put_page(new_page);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800973 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800974 pmd_t entry;
975 VM_BUG_ON(!PageHead(page));
976 entry = mk_pmd(new_page, vma->vm_page_prot);
977 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
978 entry = pmd_mkhuge(entry);
979 pmdp_clear_flush_notify(vma, haddr, pmd);
980 page_add_new_anon_rmap(new_page, vma, haddr);
981 set_pmd_at(mm, haddr, pmd, entry);
982 update_mmu_cache(vma, address, entry);
983 page_remove_rmap(page);
984 put_page(page);
985 ret |= VM_FAULT_WRITE;
986 }
987out_unlock:
988 spin_unlock(&mm->page_table_lock);
989out:
990 return ret;
991}
992
993struct page *follow_trans_huge_pmd(struct mm_struct *mm,
994 unsigned long addr,
995 pmd_t *pmd,
996 unsigned int flags)
997{
998 struct page *page = NULL;
999
1000 assert_spin_locked(&mm->page_table_lock);
1001
1002 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1003 goto out;
1004
1005 page = pmd_page(*pmd);
1006 VM_BUG_ON(!PageHead(page));
1007 if (flags & FOLL_TOUCH) {
1008 pmd_t _pmd;
1009 /*
1010 * We should set the dirty bit only for FOLL_WRITE but
1011 * for now the dirty bit in the pmd is meaningless.
1012 * And if the dirty bit will become meaningful and
1013 * we'll only set it with FOLL_WRITE, an atomic
1014 * set_bit will be required on the pmd to set the
1015 * young bit, instead of the current set_pmd_at.
1016 */
1017 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1018 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1019 }
1020 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1021 VM_BUG_ON(!PageCompound(page));
1022 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001023 get_page_foll(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001024
1025out:
1026 return page;
1027}
1028
1029int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001030 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001031{
1032 int ret = 0;
1033
1034 spin_lock(&tlb->mm->page_table_lock);
1035 if (likely(pmd_trans_huge(*pmd))) {
1036 if (unlikely(pmd_trans_splitting(*pmd))) {
1037 spin_unlock(&tlb->mm->page_table_lock);
1038 wait_split_huge_page(vma->anon_vma,
1039 pmd);
1040 } else {
1041 struct page *page;
1042 pgtable_t pgtable;
1043 pgtable = get_pmd_huge_pte(tlb->mm);
1044 page = pmd_page(*pmd);
1045 pmd_clear(pmd);
Shaohua Lif21760b2012-01-12 17:19:16 -08001046 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001047 page_remove_rmap(page);
1048 VM_BUG_ON(page_mapcount(page) < 0);
1049 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1050 VM_BUG_ON(!PageHead(page));
Andrea Arcangeli1c641e82012-03-05 14:59:20 -08001051 tlb->mm->nr_ptes--;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001052 spin_unlock(&tlb->mm->page_table_lock);
1053 tlb_remove_page(tlb, page);
1054 pte_free(tlb->mm, pgtable);
1055 ret = 1;
1056 }
1057 } else
1058 spin_unlock(&tlb->mm->page_table_lock);
1059
1060 return ret;
1061}
1062
Johannes Weiner0ca16342011-01-13 15:47:02 -08001063int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1064 unsigned long addr, unsigned long end,
1065 unsigned char *vec)
1066{
1067 int ret = 0;
1068
1069 spin_lock(&vma->vm_mm->page_table_lock);
1070 if (likely(pmd_trans_huge(*pmd))) {
1071 ret = !pmd_trans_splitting(*pmd);
1072 spin_unlock(&vma->vm_mm->page_table_lock);
1073 if (unlikely(!ret))
1074 wait_split_huge_page(vma->anon_vma, pmd);
1075 else {
1076 /*
1077 * All logical pages in the range are present
1078 * if backed by a huge page.
1079 */
1080 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1081 }
1082 } else
1083 spin_unlock(&vma->vm_mm->page_table_lock);
1084
1085 return ret;
1086}
1087
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001088int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1089 unsigned long old_addr,
1090 unsigned long new_addr, unsigned long old_end,
1091 pmd_t *old_pmd, pmd_t *new_pmd)
1092{
1093 int ret = 0;
1094 pmd_t pmd;
1095
1096 struct mm_struct *mm = vma->vm_mm;
1097
1098 if ((old_addr & ~HPAGE_PMD_MASK) ||
1099 (new_addr & ~HPAGE_PMD_MASK) ||
1100 old_end - old_addr < HPAGE_PMD_SIZE ||
1101 (new_vma->vm_flags & VM_NOHUGEPAGE))
1102 goto out;
1103
1104 /*
1105 * The destination pmd shouldn't be established, free_pgtables()
1106 * should have release it.
1107 */
1108 if (WARN_ON(!pmd_none(*new_pmd))) {
1109 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1110 goto out;
1111 }
1112
1113 spin_lock(&mm->page_table_lock);
1114 if (likely(pmd_trans_huge(*old_pmd))) {
1115 if (pmd_trans_splitting(*old_pmd)) {
1116 spin_unlock(&mm->page_table_lock);
1117 wait_split_huge_page(vma->anon_vma, old_pmd);
1118 ret = -1;
1119 } else {
1120 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1121 VM_BUG_ON(!pmd_none(*new_pmd));
1122 set_pmd_at(mm, new_addr, new_pmd, pmd);
1123 spin_unlock(&mm->page_table_lock);
1124 ret = 1;
1125 }
1126 } else {
1127 spin_unlock(&mm->page_table_lock);
1128 }
1129out:
1130 return ret;
1131}
1132
Johannes Weinercd7548a2011-01-13 15:47:04 -08001133int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1134 unsigned long addr, pgprot_t newprot)
1135{
1136 struct mm_struct *mm = vma->vm_mm;
1137 int ret = 0;
1138
1139 spin_lock(&mm->page_table_lock);
1140 if (likely(pmd_trans_huge(*pmd))) {
1141 if (unlikely(pmd_trans_splitting(*pmd))) {
1142 spin_unlock(&mm->page_table_lock);
1143 wait_split_huge_page(vma->anon_vma, pmd);
1144 } else {
1145 pmd_t entry;
1146
1147 entry = pmdp_get_and_clear(mm, addr, pmd);
1148 entry = pmd_modify(entry, newprot);
1149 set_pmd_at(mm, addr, pmd, entry);
1150 spin_unlock(&vma->vm_mm->page_table_lock);
Johannes Weinercd7548a2011-01-13 15:47:04 -08001151 ret = 1;
1152 }
1153 } else
1154 spin_unlock(&vma->vm_mm->page_table_lock);
1155
1156 return ret;
1157}
1158
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001159pmd_t *page_check_address_pmd(struct page *page,
1160 struct mm_struct *mm,
1161 unsigned long address,
1162 enum page_check_address_pmd_flag flag)
1163{
1164 pgd_t *pgd;
1165 pud_t *pud;
1166 pmd_t *pmd, *ret = NULL;
1167
1168 if (address & ~HPAGE_PMD_MASK)
1169 goto out;
1170
1171 pgd = pgd_offset(mm, address);
1172 if (!pgd_present(*pgd))
1173 goto out;
1174
1175 pud = pud_offset(pgd, address);
1176 if (!pud_present(*pud))
1177 goto out;
1178
1179 pmd = pmd_offset(pud, address);
1180 if (pmd_none(*pmd))
1181 goto out;
1182 if (pmd_page(*pmd) != page)
1183 goto out;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001184 /*
1185 * split_vma() may create temporary aliased mappings. There is
1186 * no risk as long as all huge pmd are found and have their
1187 * splitting bit set before __split_huge_page_refcount
1188 * runs. Finding the same huge pmd more than once during the
1189 * same rmap walk is not a problem.
1190 */
1191 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1192 pmd_trans_splitting(*pmd))
1193 goto out;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001194 if (pmd_trans_huge(*pmd)) {
1195 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1196 !pmd_trans_splitting(*pmd));
1197 ret = pmd;
1198 }
1199out:
1200 return ret;
1201}
1202
1203static int __split_huge_page_splitting(struct page *page,
1204 struct vm_area_struct *vma,
1205 unsigned long address)
1206{
1207 struct mm_struct *mm = vma->vm_mm;
1208 pmd_t *pmd;
1209 int ret = 0;
1210
1211 spin_lock(&mm->page_table_lock);
1212 pmd = page_check_address_pmd(page, mm, address,
1213 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1214 if (pmd) {
1215 /*
1216 * We can't temporarily set the pmd to null in order
1217 * to split it, the pmd must remain marked huge at all
1218 * times or the VM won't take the pmd_trans_huge paths
Peter Zijlstra2b575eb2011-05-24 17:12:11 -07001219 * and it won't wait on the anon_vma->root->mutex to
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001220 * serialize against split_huge_page*.
1221 */
1222 pmdp_splitting_flush_notify(vma, address, pmd);
1223 ret = 1;
1224 }
1225 spin_unlock(&mm->page_table_lock);
1226
1227 return ret;
1228}
1229
1230static void __split_huge_page_refcount(struct page *page)
1231{
1232 int i;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001233 struct zone *zone = page_zone(page);
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001234 int tail_count = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001235
1236 /* prevent PageLRU to go away from under us, and freeze lru stats */
1237 spin_lock_irq(&zone->lru_lock);
1238 compound_lock(page);
KAMEZAWA Hiroyukie94c8a92012-01-12 17:18:20 -08001239 /* complete memcg works before add pages to LRU */
1240 mem_cgroup_split_huge_fixup(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001241
Shaohua Li45676882012-01-12 17:19:18 -08001242 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001243 struct page *page_tail = page + i;
1244
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001245 /* tail_page->_mapcount cannot change */
1246 BUG_ON(page_mapcount(page_tail) < 0);
1247 tail_count += page_mapcount(page_tail);
1248 /* check for overflow */
1249 BUG_ON(tail_count < 0);
1250 BUG_ON(atomic_read(&page_tail->_count) != 0);
1251 /*
1252 * tail_page->_count is zero and not changing from
1253 * under us. But get_page_unless_zero() may be running
1254 * from under us on the tail_page. If we used
1255 * atomic_set() below instead of atomic_add(), we
1256 * would then run atomic_set() concurrently with
1257 * get_page_unless_zero(), and atomic_set() is
1258 * implemented in C not using locked ops. spin_unlock
1259 * on x86 sometime uses locked ops because of PPro
1260 * errata 66, 92, so unless somebody can guarantee
1261 * atomic_set() here would be safe on all archs (and
1262 * not only on x86), it's safer to use atomic_add().
1263 */
1264 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1265 &page_tail->_count);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001266
1267 /* after clearing PageTail the gup refcount can be released */
1268 smp_mb();
1269
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001270 /*
1271 * retain hwpoison flag of the poisoned tail page:
1272 * fix for the unsuitable process killed on Guest Machine(KVM)
1273 * by the memory-failure.
1274 */
1275 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001276 page_tail->flags |= (page->flags &
1277 ((1L << PG_referenced) |
1278 (1L << PG_swapbacked) |
1279 (1L << PG_mlocked) |
1280 (1L << PG_uptodate)));
1281 page_tail->flags |= (1L << PG_dirty);
1282
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001283 /* clear PageTail before overwriting first_page */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001284 smp_wmb();
1285
1286 /*
1287 * __split_huge_page_splitting() already set the
1288 * splitting bit in all pmd that could map this
1289 * hugepage, that will ensure no CPU can alter the
1290 * mapcount on the head page. The mapcount is only
1291 * accounted in the head page and it has to be
1292 * transferred to all tail pages in the below code. So
1293 * for this code to be safe, the split the mapcount
1294 * can't change. But that doesn't mean userland can't
1295 * keep changing and reading the page contents while
1296 * we transfer the mapcount, so the pmd splitting
1297 * status is achieved setting a reserved bit in the
1298 * pmd, not by clearing the present bit.
1299 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001300 page_tail->_mapcount = page->_mapcount;
1301
1302 BUG_ON(page_tail->mapping);
1303 page_tail->mapping = page->mapping;
1304
Shaohua Li45676882012-01-12 17:19:18 -08001305 page_tail->index = page->index + i;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001306
1307 BUG_ON(!PageAnon(page_tail));
1308 BUG_ON(!PageUptodate(page_tail));
1309 BUG_ON(!PageDirty(page_tail));
1310 BUG_ON(!PageSwapBacked(page_tail));
1311
KAMEZAWA Hiroyukica3e0212011-01-20 14:44:24 -08001312
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001313 lru_add_page_tail(zone, page, page_tail);
1314 }
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001315 atomic_sub(tail_count, &page->_count);
1316 BUG_ON(atomic_read(&page->_count) <= 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001317
Andrea Arcangeli79134172011-01-13 15:46:58 -08001318 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1319 __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1320
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001321 ClearPageCompound(page);
1322 compound_unlock(page);
1323 spin_unlock_irq(&zone->lru_lock);
1324
1325 for (i = 1; i < HPAGE_PMD_NR; i++) {
1326 struct page *page_tail = page + i;
1327 BUG_ON(page_count(page_tail) <= 0);
1328 /*
1329 * Tail pages may be freed if there wasn't any mapping
1330 * like if add_to_swap() is running on a lru page that
1331 * had its mapping zapped. And freeing these pages
1332 * requires taking the lru_lock so we do the put_page
1333 * of the tail pages after the split is complete.
1334 */
1335 put_page(page_tail);
1336 }
1337
1338 /*
1339 * Only the head page (now become a regular page) is required
1340 * to be pinned by the caller.
1341 */
1342 BUG_ON(page_count(page) <= 0);
1343}
1344
1345static int __split_huge_page_map(struct page *page,
1346 struct vm_area_struct *vma,
1347 unsigned long address)
1348{
1349 struct mm_struct *mm = vma->vm_mm;
1350 pmd_t *pmd, _pmd;
1351 int ret = 0, i;
1352 pgtable_t pgtable;
1353 unsigned long haddr;
1354
1355 spin_lock(&mm->page_table_lock);
1356 pmd = page_check_address_pmd(page, mm, address,
1357 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1358 if (pmd) {
1359 pgtable = get_pmd_huge_pte(mm);
1360 pmd_populate(mm, &_pmd, pgtable);
1361
1362 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1363 i++, haddr += PAGE_SIZE) {
1364 pte_t *pte, entry;
1365 BUG_ON(PageCompound(page+i));
1366 entry = mk_pte(page + i, vma->vm_page_prot);
1367 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1368 if (!pmd_write(*pmd))
1369 entry = pte_wrprotect(entry);
1370 else
1371 BUG_ON(page_mapcount(page) != 1);
1372 if (!pmd_young(*pmd))
1373 entry = pte_mkold(entry);
1374 pte = pte_offset_map(&_pmd, haddr);
1375 BUG_ON(!pte_none(*pte));
1376 set_pte_at(mm, haddr, pte, entry);
1377 pte_unmap(pte);
1378 }
1379
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001380 smp_wmb(); /* make pte visible before pmd */
1381 /*
1382 * Up to this point the pmd is present and huge and
1383 * userland has the whole access to the hugepage
1384 * during the split (which happens in place). If we
1385 * overwrite the pmd with the not-huge version
1386 * pointing to the pte here (which of course we could
1387 * if all CPUs were bug free), userland could trigger
1388 * a small page size TLB miss on the small sized TLB
1389 * while the hugepage TLB entry is still established
1390 * in the huge TLB. Some CPU doesn't like that. See
1391 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1392 * Erratum 383 on page 93. Intel should be safe but is
1393 * also warns that it's only safe if the permission
1394 * and cache attributes of the two entries loaded in
1395 * the two TLB is identical (which should be the case
1396 * here). But it is generally safer to never allow
1397 * small and huge TLB entries for the same virtual
1398 * address to be loaded simultaneously. So instead of
1399 * doing "pmd_populate(); flush_tlb_range();" we first
1400 * mark the current pmd notpresent (atomically because
1401 * here the pmd_trans_huge and pmd_trans_splitting
1402 * must remain set at all times on the pmd until the
1403 * split is complete for this pmd), then we flush the
1404 * SMP TLB and finally we write the non-huge version
1405 * of the pmd entry with pmd_populate.
1406 */
1407 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1408 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1409 pmd_populate(mm, pmd, pgtable);
1410 ret = 1;
1411 }
1412 spin_unlock(&mm->page_table_lock);
1413
1414 return ret;
1415}
1416
Peter Zijlstra2b575eb2011-05-24 17:12:11 -07001417/* must be called with anon_vma->root->mutex hold */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001418static void __split_huge_page(struct page *page,
1419 struct anon_vma *anon_vma)
1420{
1421 int mapcount, mapcount2;
1422 struct anon_vma_chain *avc;
1423
1424 BUG_ON(!PageHead(page));
1425 BUG_ON(PageTail(page));
1426
1427 mapcount = 0;
1428 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1429 struct vm_area_struct *vma = avc->vma;
1430 unsigned long addr = vma_address(page, vma);
1431 BUG_ON(is_vma_temporary_stack(vma));
1432 if (addr == -EFAULT)
1433 continue;
1434 mapcount += __split_huge_page_splitting(page, vma, addr);
1435 }
Andrea Arcangeli05759d32011-01-13 15:46:53 -08001436 /*
1437 * It is critical that new vmas are added to the tail of the
1438 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1439 * and establishes a child pmd before
1440 * __split_huge_page_splitting() freezes the parent pmd (so if
1441 * we fail to prevent copy_huge_pmd() from running until the
1442 * whole __split_huge_page() is complete), we will still see
1443 * the newly established pmd of the child later during the
1444 * walk, to be able to set it as pmd_trans_splitting too.
1445 */
1446 if (mapcount != page_mapcount(page))
1447 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1448 mapcount, page_mapcount(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001449 BUG_ON(mapcount != page_mapcount(page));
1450
1451 __split_huge_page_refcount(page);
1452
1453 mapcount2 = 0;
1454 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1455 struct vm_area_struct *vma = avc->vma;
1456 unsigned long addr = vma_address(page, vma);
1457 BUG_ON(is_vma_temporary_stack(vma));
1458 if (addr == -EFAULT)
1459 continue;
1460 mapcount2 += __split_huge_page_map(page, vma, addr);
1461 }
Andrea Arcangeli05759d32011-01-13 15:46:53 -08001462 if (mapcount != mapcount2)
1463 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1464 mapcount, mapcount2, page_mapcount(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001465 BUG_ON(mapcount != mapcount2);
1466}
1467
1468int split_huge_page(struct page *page)
1469{
1470 struct anon_vma *anon_vma;
1471 int ret = 1;
1472
1473 BUG_ON(!PageAnon(page));
1474 anon_vma = page_lock_anon_vma(page);
1475 if (!anon_vma)
1476 goto out;
1477 ret = 0;
1478 if (!PageCompound(page))
1479 goto out_unlock;
1480
1481 BUG_ON(!PageSwapBacked(page));
1482 __split_huge_page(page, anon_vma);
Andi Kleen81ab4202011-04-14 15:22:06 -07001483 count_vm_event(THP_SPLIT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001484
1485 BUG_ON(PageCompound(page));
1486out_unlock:
1487 page_unlock_anon_vma(anon_vma);
1488out:
1489 return ret;
1490}
1491
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001492#define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1493 VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1494
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001495int hugepage_madvise(struct vm_area_struct *vma,
1496 unsigned long *vm_flags, int advice)
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001497{
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001498 switch (advice) {
1499 case MADV_HUGEPAGE:
1500 /*
1501 * Be somewhat over-protective like KSM for now!
1502 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001503 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001504 return -EINVAL;
1505 *vm_flags &= ~VM_NOHUGEPAGE;
1506 *vm_flags |= VM_HUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001507 /*
1508 * If the vma become good for khugepaged to scan,
1509 * register it here without waiting a page fault that
1510 * may not happen any time soon.
1511 */
1512 if (unlikely(khugepaged_enter_vma_merge(vma)))
1513 return -ENOMEM;
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001514 break;
1515 case MADV_NOHUGEPAGE:
1516 /*
1517 * Be somewhat over-protective like KSM for now!
1518 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001519 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001520 return -EINVAL;
1521 *vm_flags &= ~VM_HUGEPAGE;
1522 *vm_flags |= VM_NOHUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001523 /*
1524 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1525 * this vma even if we leave the mm registered in khugepaged if
1526 * it got registered before VM_NOHUGEPAGE was set.
1527 */
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001528 break;
1529 }
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001530
1531 return 0;
1532}
1533
Andrea Arcangeliba761492011-01-13 15:46:58 -08001534static int __init khugepaged_slab_init(void)
1535{
1536 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1537 sizeof(struct mm_slot),
1538 __alignof__(struct mm_slot), 0, NULL);
1539 if (!mm_slot_cache)
1540 return -ENOMEM;
1541
1542 return 0;
1543}
1544
1545static void __init khugepaged_slab_free(void)
1546{
1547 kmem_cache_destroy(mm_slot_cache);
1548 mm_slot_cache = NULL;
1549}
1550
1551static inline struct mm_slot *alloc_mm_slot(void)
1552{
1553 if (!mm_slot_cache) /* initialization failed */
1554 return NULL;
1555 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1556}
1557
1558static inline void free_mm_slot(struct mm_slot *mm_slot)
1559{
1560 kmem_cache_free(mm_slot_cache, mm_slot);
1561}
1562
1563static int __init mm_slots_hash_init(void)
1564{
1565 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1566 GFP_KERNEL);
1567 if (!mm_slots_hash)
1568 return -ENOMEM;
1569 return 0;
1570}
1571
1572#if 0
1573static void __init mm_slots_hash_free(void)
1574{
1575 kfree(mm_slots_hash);
1576 mm_slots_hash = NULL;
1577}
1578#endif
1579
1580static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1581{
1582 struct mm_slot *mm_slot;
1583 struct hlist_head *bucket;
1584 struct hlist_node *node;
1585
1586 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1587 % MM_SLOTS_HASH_HEADS];
1588 hlist_for_each_entry(mm_slot, node, bucket, hash) {
1589 if (mm == mm_slot->mm)
1590 return mm_slot;
1591 }
1592 return NULL;
1593}
1594
1595static void insert_to_mm_slots_hash(struct mm_struct *mm,
1596 struct mm_slot *mm_slot)
1597{
1598 struct hlist_head *bucket;
1599
1600 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1601 % MM_SLOTS_HASH_HEADS];
1602 mm_slot->mm = mm;
1603 hlist_add_head(&mm_slot->hash, bucket);
1604}
1605
1606static inline int khugepaged_test_exit(struct mm_struct *mm)
1607{
1608 return atomic_read(&mm->mm_users) == 0;
1609}
1610
1611int __khugepaged_enter(struct mm_struct *mm)
1612{
1613 struct mm_slot *mm_slot;
1614 int wakeup;
1615
1616 mm_slot = alloc_mm_slot();
1617 if (!mm_slot)
1618 return -ENOMEM;
1619
1620 /* __khugepaged_exit() must not run from under us */
1621 VM_BUG_ON(khugepaged_test_exit(mm));
1622 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1623 free_mm_slot(mm_slot);
1624 return 0;
1625 }
1626
1627 spin_lock(&khugepaged_mm_lock);
1628 insert_to_mm_slots_hash(mm, mm_slot);
1629 /*
1630 * Insert just behind the scanning cursor, to let the area settle
1631 * down a little.
1632 */
1633 wakeup = list_empty(&khugepaged_scan.mm_head);
1634 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1635 spin_unlock(&khugepaged_mm_lock);
1636
1637 atomic_inc(&mm->mm_count);
1638 if (wakeup)
1639 wake_up_interruptible(&khugepaged_wait);
1640
1641 return 0;
1642}
1643
1644int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1645{
1646 unsigned long hstart, hend;
1647 if (!vma->anon_vma)
1648 /*
1649 * Not yet faulted in so we will register later in the
1650 * page fault if needed.
1651 */
1652 return 0;
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001653 if (vma->vm_ops)
Andrea Arcangeliba761492011-01-13 15:46:58 -08001654 /* khugepaged not yet working on file or special mappings */
1655 return 0;
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001656 /*
1657 * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1658 * true too, verify it here.
1659 */
1660 VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001661 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1662 hend = vma->vm_end & HPAGE_PMD_MASK;
1663 if (hstart < hend)
1664 return khugepaged_enter(vma);
1665 return 0;
1666}
1667
1668void __khugepaged_exit(struct mm_struct *mm)
1669{
1670 struct mm_slot *mm_slot;
1671 int free = 0;
1672
1673 spin_lock(&khugepaged_mm_lock);
1674 mm_slot = get_mm_slot(mm);
1675 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1676 hlist_del(&mm_slot->hash);
1677 list_del(&mm_slot->mm_node);
1678 free = 1;
1679 }
Chris Wrightd788e802011-07-25 17:12:14 -07001680 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001681
1682 if (free) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08001683 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1684 free_mm_slot(mm_slot);
1685 mmdrop(mm);
1686 } else if (mm_slot) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08001687 /*
1688 * This is required to serialize against
1689 * khugepaged_test_exit() (which is guaranteed to run
1690 * under mmap sem read mode). Stop here (after we
1691 * return all pagetables will be destroyed) until
1692 * khugepaged has finished working on the pagetables
1693 * under the mmap_sem.
1694 */
1695 down_write(&mm->mmap_sem);
1696 up_write(&mm->mmap_sem);
Chris Wrightd788e802011-07-25 17:12:14 -07001697 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08001698}
1699
1700static void release_pte_page(struct page *page)
1701{
1702 /* 0 stands for page_is_file_cache(page) == false */
1703 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1704 unlock_page(page);
1705 putback_lru_page(page);
1706}
1707
1708static void release_pte_pages(pte_t *pte, pte_t *_pte)
1709{
1710 while (--_pte >= pte) {
1711 pte_t pteval = *_pte;
1712 if (!pte_none(pteval))
1713 release_pte_page(pte_page(pteval));
1714 }
1715}
1716
1717static void release_all_pte_pages(pte_t *pte)
1718{
1719 release_pte_pages(pte, pte + HPAGE_PMD_NR);
1720}
1721
1722static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1723 unsigned long address,
1724 pte_t *pte)
1725{
1726 struct page *page;
1727 pte_t *_pte;
1728 int referenced = 0, isolated = 0, none = 0;
1729 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1730 _pte++, address += PAGE_SIZE) {
1731 pte_t pteval = *_pte;
1732 if (pte_none(pteval)) {
1733 if (++none <= khugepaged_max_ptes_none)
1734 continue;
1735 else {
1736 release_pte_pages(pte, _pte);
1737 goto out;
1738 }
1739 }
1740 if (!pte_present(pteval) || !pte_write(pteval)) {
1741 release_pte_pages(pte, _pte);
1742 goto out;
1743 }
1744 page = vm_normal_page(vma, address, pteval);
1745 if (unlikely(!page)) {
1746 release_pte_pages(pte, _pte);
1747 goto out;
1748 }
1749 VM_BUG_ON(PageCompound(page));
1750 BUG_ON(!PageAnon(page));
1751 VM_BUG_ON(!PageSwapBacked(page));
1752
1753 /* cannot use mapcount: can't collapse if there's a gup pin */
1754 if (page_count(page) != 1) {
1755 release_pte_pages(pte, _pte);
1756 goto out;
1757 }
1758 /*
1759 * We can do it before isolate_lru_page because the
1760 * page can't be freed from under us. NOTE: PG_lock
1761 * is needed to serialize against split_huge_page
1762 * when invoked from the VM.
1763 */
1764 if (!trylock_page(page)) {
1765 release_pte_pages(pte, _pte);
1766 goto out;
1767 }
1768 /*
1769 * Isolate the page to avoid collapsing an hugepage
1770 * currently in use by the VM.
1771 */
1772 if (isolate_lru_page(page)) {
1773 unlock_page(page);
1774 release_pte_pages(pte, _pte);
1775 goto out;
1776 }
1777 /* 0 stands for page_is_file_cache(page) == false */
1778 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1779 VM_BUG_ON(!PageLocked(page));
1780 VM_BUG_ON(PageLRU(page));
1781
1782 /* If there is no mapped pte young don't collapse the page */
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08001783 if (pte_young(pteval) || PageReferenced(page) ||
1784 mmu_notifier_test_young(vma->vm_mm, address))
Andrea Arcangeliba761492011-01-13 15:46:58 -08001785 referenced = 1;
1786 }
1787 if (unlikely(!referenced))
1788 release_all_pte_pages(pte);
1789 else
1790 isolated = 1;
1791out:
1792 return isolated;
1793}
1794
1795static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1796 struct vm_area_struct *vma,
1797 unsigned long address,
1798 spinlock_t *ptl)
1799{
1800 pte_t *_pte;
1801 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1802 pte_t pteval = *_pte;
1803 struct page *src_page;
1804
1805 if (pte_none(pteval)) {
1806 clear_user_highpage(page, address);
1807 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1808 } else {
1809 src_page = pte_page(pteval);
1810 copy_user_highpage(page, src_page, address, vma);
1811 VM_BUG_ON(page_mapcount(src_page) != 1);
1812 VM_BUG_ON(page_count(src_page) != 2);
1813 release_pte_page(src_page);
1814 /*
1815 * ptl mostly unnecessary, but preempt has to
1816 * be disabled to update the per-cpu stats
1817 * inside page_remove_rmap().
1818 */
1819 spin_lock(ptl);
1820 /*
1821 * paravirt calls inside pte_clear here are
1822 * superfluous.
1823 */
1824 pte_clear(vma->vm_mm, address, _pte);
1825 page_remove_rmap(src_page);
1826 spin_unlock(ptl);
1827 free_page_and_swap_cache(src_page);
1828 }
1829
1830 address += PAGE_SIZE;
1831 page++;
1832 }
1833}
1834
1835static void collapse_huge_page(struct mm_struct *mm,
1836 unsigned long address,
Andrea Arcangelice83d212011-01-13 15:47:06 -08001837 struct page **hpage,
Andi Kleen5c4b4be2011-03-04 17:36:32 -08001838 struct vm_area_struct *vma,
1839 int node)
Andrea Arcangeliba761492011-01-13 15:46:58 -08001840{
Andrea Arcangeliba761492011-01-13 15:46:58 -08001841 pgd_t *pgd;
1842 pud_t *pud;
1843 pmd_t *pmd, _pmd;
1844 pte_t *pte;
1845 pgtable_t pgtable;
1846 struct page *new_page;
1847 spinlock_t *ptl;
1848 int isolated;
1849 unsigned long hstart, hend;
1850
1851 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001852#ifndef CONFIG_NUMA
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07001853 up_read(&mm->mmap_sem);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001854 VM_BUG_ON(!*hpage);
Andrea Arcangelice83d212011-01-13 15:47:06 -08001855 new_page = *hpage;
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001856#else
1857 VM_BUG_ON(*hpage);
Andrea Arcangelice83d212011-01-13 15:47:06 -08001858 /*
1859 * Allocate the page while the vma is still valid and under
1860 * the mmap_sem read mode so there is no memory allocation
1861 * later when we take the mmap_sem in write mode. This is more
1862 * friendly behavior (OTOH it may actually hide bugs) to
1863 * filesystems in userland with daemons allocating memory in
1864 * the userland I/O paths. Allocating memory with the
1865 * mmap_sem in read mode is good idea also to allow greater
1866 * scalability.
1867 */
Andi Kleen5c4b4be2011-03-04 17:36:32 -08001868 new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
Andi Kleencc5d4622011-03-22 16:33:13 -07001869 node, __GFP_OTHER_NODE);
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07001870
1871 /*
1872 * After allocating the hugepage, release the mmap_sem read lock in
1873 * preparation for taking it in write mode.
1874 */
1875 up_read(&mm->mmap_sem);
Andrea Arcangelice83d212011-01-13 15:47:06 -08001876 if (unlikely(!new_page)) {
Andi Kleen81ab4202011-04-14 15:22:06 -07001877 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
Andrea Arcangelice83d212011-01-13 15:47:06 -08001878 *hpage = ERR_PTR(-ENOMEM);
1879 return;
1880 }
Hugh Dickins2fbfac42011-03-14 01:08:47 -07001881#endif
Andrea Arcangelice83d212011-01-13 15:47:06 -08001882
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07001883 count_vm_event(THP_COLLAPSE_ALLOC);
1884 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1885#ifdef CONFIG_NUMA
1886 put_page(new_page);
1887#endif
1888 return;
1889 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08001890
1891 /*
1892 * Prevent all access to pagetables with the exception of
1893 * gup_fast later hanlded by the ptep_clear_flush and the VM
1894 * handled by the anon_vma lock + PG_lock.
1895 */
1896 down_write(&mm->mmap_sem);
1897 if (unlikely(khugepaged_test_exit(mm)))
1898 goto out;
1899
1900 vma = find_vma(mm, address);
1901 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1902 hend = vma->vm_end & HPAGE_PMD_MASK;
1903 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1904 goto out;
1905
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001906 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1907 (vma->vm_flags & VM_NOHUGEPAGE))
Andrea Arcangeliba761492011-01-13 15:46:58 -08001908 goto out;
1909
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001910 if (!vma->anon_vma || vma->vm_ops)
Andrea Arcangeliba761492011-01-13 15:46:58 -08001911 goto out;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01001912 if (is_vma_temporary_stack(vma))
1913 goto out;
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001914 /*
1915 * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1916 * true too, verify it here.
1917 */
1918 VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001919
1920 pgd = pgd_offset(mm, address);
1921 if (!pgd_present(*pgd))
1922 goto out;
1923
1924 pud = pud_offset(pgd, address);
1925 if (!pud_present(*pud))
1926 goto out;
1927
1928 pmd = pmd_offset(pud, address);
1929 /* pmd can't go away or become huge under us */
1930 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1931 goto out;
1932
Andrea Arcangeliba761492011-01-13 15:46:58 -08001933 anon_vma_lock(vma->anon_vma);
1934
1935 pte = pte_offset_map(pmd, address);
1936 ptl = pte_lockptr(mm, pmd);
1937
1938 spin_lock(&mm->page_table_lock); /* probably unnecessary */
1939 /*
1940 * After this gup_fast can't run anymore. This also removes
1941 * any huge TLB entry from the CPU so we won't allow
1942 * huge and small TLB entries for the same virtual address
1943 * to avoid the risk of CPU bugs in that area.
1944 */
1945 _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1946 spin_unlock(&mm->page_table_lock);
1947
1948 spin_lock(ptl);
1949 isolated = __collapse_huge_page_isolate(vma, address, pte);
1950 spin_unlock(ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001951
1952 if (unlikely(!isolated)) {
Johannes Weiner453c7192011-01-20 14:44:18 -08001953 pte_unmap(pte);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001954 spin_lock(&mm->page_table_lock);
1955 BUG_ON(!pmd_none(*pmd));
1956 set_pmd_at(mm, address, pmd, _pmd);
1957 spin_unlock(&mm->page_table_lock);
1958 anon_vma_unlock(vma->anon_vma);
Andrea Arcangelice83d212011-01-13 15:47:06 -08001959 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08001960 }
1961
1962 /*
1963 * All pages are isolated and locked so anon_vma rmap
1964 * can't run anymore.
1965 */
1966 anon_vma_unlock(vma->anon_vma);
1967
1968 __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
Johannes Weiner453c7192011-01-20 14:44:18 -08001969 pte_unmap(pte);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001970 __SetPageUptodate(new_page);
1971 pgtable = pmd_pgtable(_pmd);
1972 VM_BUG_ON(page_count(pgtable) != 1);
1973 VM_BUG_ON(page_mapcount(pgtable) != 0);
1974
1975 _pmd = mk_pmd(new_page, vma->vm_page_prot);
1976 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1977 _pmd = pmd_mkhuge(_pmd);
1978
1979 /*
1980 * spin_lock() below is not the equivalent of smp_wmb(), so
1981 * this is needed to avoid the copy_huge_page writes to become
1982 * visible after the set_pmd_at() write.
1983 */
1984 smp_wmb();
1985
1986 spin_lock(&mm->page_table_lock);
1987 BUG_ON(!pmd_none(*pmd));
1988 page_add_new_anon_rmap(new_page, vma, address);
1989 set_pmd_at(mm, address, pmd, _pmd);
Hillf Danton35d8c7ad722011-10-31 17:09:40 -07001990 update_mmu_cache(vma, address, _pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001991 prepare_pmd_huge_pte(pgtable, mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001992 spin_unlock(&mm->page_table_lock);
1993
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001994#ifndef CONFIG_NUMA
Andrea Arcangeliba761492011-01-13 15:46:58 -08001995 *hpage = NULL;
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001996#endif
Andrea Arcangeliba761492011-01-13 15:46:58 -08001997 khugepaged_pages_collapsed++;
Andrea Arcangelice83d212011-01-13 15:47:06 -08001998out_up_write:
Andrea Arcangeliba761492011-01-13 15:46:58 -08001999 up_write(&mm->mmap_sem);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002000 return;
2001
Andrea Arcangelice83d212011-01-13 15:47:06 -08002002out:
KAMEZAWA Hiroyuki678ff892011-02-10 15:01:36 -08002003 mem_cgroup_uncharge_page(new_page);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002004#ifdef CONFIG_NUMA
2005 put_page(new_page);
2006#endif
Andrea Arcangelice83d212011-01-13 15:47:06 -08002007 goto out_up_write;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002008}
2009
2010static int khugepaged_scan_pmd(struct mm_struct *mm,
2011 struct vm_area_struct *vma,
2012 unsigned long address,
2013 struct page **hpage)
2014{
2015 pgd_t *pgd;
2016 pud_t *pud;
2017 pmd_t *pmd;
2018 pte_t *pte, *_pte;
2019 int ret = 0, referenced = 0, none = 0;
2020 struct page *page;
2021 unsigned long _address;
2022 spinlock_t *ptl;
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002023 int node = -1;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002024
2025 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2026
2027 pgd = pgd_offset(mm, address);
2028 if (!pgd_present(*pgd))
2029 goto out;
2030
2031 pud = pud_offset(pgd, address);
2032 if (!pud_present(*pud))
2033 goto out;
2034
2035 pmd = pmd_offset(pud, address);
2036 if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2037 goto out;
2038
2039 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2040 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2041 _pte++, _address += PAGE_SIZE) {
2042 pte_t pteval = *_pte;
2043 if (pte_none(pteval)) {
2044 if (++none <= khugepaged_max_ptes_none)
2045 continue;
2046 else
2047 goto out_unmap;
2048 }
2049 if (!pte_present(pteval) || !pte_write(pteval))
2050 goto out_unmap;
2051 page = vm_normal_page(vma, _address, pteval);
2052 if (unlikely(!page))
2053 goto out_unmap;
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002054 /*
2055 * Chose the node of the first page. This could
2056 * be more sophisticated and look at more pages,
2057 * but isn't for now.
2058 */
2059 if (node == -1)
2060 node = page_to_nid(page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002061 VM_BUG_ON(PageCompound(page));
2062 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2063 goto out_unmap;
2064 /* cannot use mapcount: can't collapse if there's a gup pin */
2065 if (page_count(page) != 1)
2066 goto out_unmap;
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002067 if (pte_young(pteval) || PageReferenced(page) ||
2068 mmu_notifier_test_young(vma->vm_mm, address))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002069 referenced = 1;
2070 }
2071 if (referenced)
2072 ret = 1;
2073out_unmap:
2074 pte_unmap_unlock(pte, ptl);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002075 if (ret)
2076 /* collapse_huge_page will return with the mmap_sem released */
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002077 collapse_huge_page(mm, address, hpage, vma, node);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002078out:
2079 return ret;
2080}
2081
2082static void collect_mm_slot(struct mm_slot *mm_slot)
2083{
2084 struct mm_struct *mm = mm_slot->mm;
2085
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002086 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002087
2088 if (khugepaged_test_exit(mm)) {
2089 /* free mm_slot */
2090 hlist_del(&mm_slot->hash);
2091 list_del(&mm_slot->mm_node);
2092
2093 /*
2094 * Not strictly needed because the mm exited already.
2095 *
2096 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2097 */
2098
2099 /* khugepaged_mm_lock actually not necessary for the below */
2100 free_mm_slot(mm_slot);
2101 mmdrop(mm);
2102 }
2103}
2104
2105static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2106 struct page **hpage)
H Hartley Sweeten2f1da642011-10-31 17:09:25 -07002107 __releases(&khugepaged_mm_lock)
2108 __acquires(&khugepaged_mm_lock)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002109{
2110 struct mm_slot *mm_slot;
2111 struct mm_struct *mm;
2112 struct vm_area_struct *vma;
2113 int progress = 0;
2114
2115 VM_BUG_ON(!pages);
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002116 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002117
2118 if (khugepaged_scan.mm_slot)
2119 mm_slot = khugepaged_scan.mm_slot;
2120 else {
2121 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2122 struct mm_slot, mm_node);
2123 khugepaged_scan.address = 0;
2124 khugepaged_scan.mm_slot = mm_slot;
2125 }
2126 spin_unlock(&khugepaged_mm_lock);
2127
2128 mm = mm_slot->mm;
2129 down_read(&mm->mmap_sem);
2130 if (unlikely(khugepaged_test_exit(mm)))
2131 vma = NULL;
2132 else
2133 vma = find_vma(mm, khugepaged_scan.address);
2134
2135 progress++;
2136 for (; vma; vma = vma->vm_next) {
2137 unsigned long hstart, hend;
2138
2139 cond_resched();
2140 if (unlikely(khugepaged_test_exit(mm))) {
2141 progress++;
2142 break;
2143 }
2144
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08002145 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2146 !khugepaged_always()) ||
2147 (vma->vm_flags & VM_NOHUGEPAGE)) {
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002148 skip:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002149 progress++;
2150 continue;
2151 }
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07002152 if (!vma->anon_vma || vma->vm_ops)
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002153 goto skip;
2154 if (is_vma_temporary_stack(vma))
2155 goto skip;
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07002156 /*
2157 * If is_pfn_mapping() is true is_learn_pfn_mapping()
2158 * must be true too, verify it here.
2159 */
2160 VM_BUG_ON(is_linear_pfn_mapping(vma) ||
2161 vma->vm_flags & VM_NO_THP);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002162
2163 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2164 hend = vma->vm_end & HPAGE_PMD_MASK;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002165 if (hstart >= hend)
2166 goto skip;
2167 if (khugepaged_scan.address > hend)
2168 goto skip;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002169 if (khugepaged_scan.address < hstart)
2170 khugepaged_scan.address = hstart;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002171 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002172
2173 while (khugepaged_scan.address < hend) {
2174 int ret;
2175 cond_resched();
2176 if (unlikely(khugepaged_test_exit(mm)))
2177 goto breakouterloop;
2178
2179 VM_BUG_ON(khugepaged_scan.address < hstart ||
2180 khugepaged_scan.address + HPAGE_PMD_SIZE >
2181 hend);
2182 ret = khugepaged_scan_pmd(mm, vma,
2183 khugepaged_scan.address,
2184 hpage);
2185 /* move to next address */
2186 khugepaged_scan.address += HPAGE_PMD_SIZE;
2187 progress += HPAGE_PMD_NR;
2188 if (ret)
2189 /* we released mmap_sem so break loop */
2190 goto breakouterloop_mmap_sem;
2191 if (progress >= pages)
2192 goto breakouterloop;
2193 }
2194 }
2195breakouterloop:
2196 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2197breakouterloop_mmap_sem:
2198
2199 spin_lock(&khugepaged_mm_lock);
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002200 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002201 /*
2202 * Release the current mm_slot if this mm is about to die, or
2203 * if we scanned all vmas of this mm.
2204 */
2205 if (khugepaged_test_exit(mm) || !vma) {
2206 /*
2207 * Make sure that if mm_users is reaching zero while
2208 * khugepaged runs here, khugepaged_exit will find
2209 * mm_slot not pointing to the exiting mm.
2210 */
2211 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2212 khugepaged_scan.mm_slot = list_entry(
2213 mm_slot->mm_node.next,
2214 struct mm_slot, mm_node);
2215 khugepaged_scan.address = 0;
2216 } else {
2217 khugepaged_scan.mm_slot = NULL;
2218 khugepaged_full_scans++;
2219 }
2220
2221 collect_mm_slot(mm_slot);
2222 }
2223
2224 return progress;
2225}
2226
2227static int khugepaged_has_work(void)
2228{
2229 return !list_empty(&khugepaged_scan.mm_head) &&
2230 khugepaged_enabled();
2231}
2232
2233static int khugepaged_wait_event(void)
2234{
2235 return !list_empty(&khugepaged_scan.mm_head) ||
2236 !khugepaged_enabled();
2237}
2238
2239static void khugepaged_do_scan(struct page **hpage)
2240{
2241 unsigned int progress = 0, pass_through_head = 0;
2242 unsigned int pages = khugepaged_pages_to_scan;
2243
2244 barrier(); /* write khugepaged_pages_to_scan to local stack */
2245
2246 while (progress < pages) {
2247 cond_resched();
2248
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002249#ifndef CONFIG_NUMA
Andrea Arcangeliba761492011-01-13 15:46:58 -08002250 if (!*hpage) {
2251 *hpage = alloc_hugepage(khugepaged_defrag());
Andi Kleen81ab4202011-04-14 15:22:06 -07002252 if (unlikely(!*hpage)) {
2253 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002254 break;
Andi Kleen81ab4202011-04-14 15:22:06 -07002255 }
2256 count_vm_event(THP_COLLAPSE_ALLOC);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002257 }
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002258#else
2259 if (IS_ERR(*hpage))
2260 break;
2261#endif
Andrea Arcangeliba761492011-01-13 15:46:58 -08002262
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002263 if (unlikely(kthread_should_stop() || freezing(current)))
2264 break;
2265
Andrea Arcangeliba761492011-01-13 15:46:58 -08002266 spin_lock(&khugepaged_mm_lock);
2267 if (!khugepaged_scan.mm_slot)
2268 pass_through_head++;
2269 if (khugepaged_has_work() &&
2270 pass_through_head < 2)
2271 progress += khugepaged_scan_mm_slot(pages - progress,
2272 hpage);
2273 else
2274 progress = pages;
2275 spin_unlock(&khugepaged_mm_lock);
2276 }
2277}
2278
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002279static void khugepaged_alloc_sleep(void)
2280{
Andrea Arcangeli1dfb0592011-12-08 14:33:57 -08002281 wait_event_freezable_timeout(khugepaged_wait, false,
2282 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002283}
2284
2285#ifndef CONFIG_NUMA
Andrea Arcangeliba761492011-01-13 15:46:58 -08002286static struct page *khugepaged_alloc_hugepage(void)
2287{
2288 struct page *hpage;
2289
2290 do {
2291 hpage = alloc_hugepage(khugepaged_defrag());
Andi Kleen81ab4202011-04-14 15:22:06 -07002292 if (!hpage) {
2293 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002294 khugepaged_alloc_sleep();
Andi Kleen81ab4202011-04-14 15:22:06 -07002295 } else
2296 count_vm_event(THP_COLLAPSE_ALLOC);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002297 } while (unlikely(!hpage) &&
2298 likely(khugepaged_enabled()));
2299 return hpage;
2300}
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002301#endif
Andrea Arcangeliba761492011-01-13 15:46:58 -08002302
2303static void khugepaged_loop(void)
2304{
2305 struct page *hpage;
2306
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002307#ifdef CONFIG_NUMA
2308 hpage = NULL;
2309#endif
Andrea Arcangeliba761492011-01-13 15:46:58 -08002310 while (likely(khugepaged_enabled())) {
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002311#ifndef CONFIG_NUMA
Andrea Arcangeliba761492011-01-13 15:46:58 -08002312 hpage = khugepaged_alloc_hugepage();
Andrea Arcangelif300ea42011-06-15 15:08:08 -07002313 if (unlikely(!hpage))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002314 break;
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002315#else
2316 if (IS_ERR(hpage)) {
2317 khugepaged_alloc_sleep();
2318 hpage = NULL;
2319 }
2320#endif
Andrea Arcangeliba761492011-01-13 15:46:58 -08002321
2322 khugepaged_do_scan(&hpage);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002323#ifndef CONFIG_NUMA
Andrea Arcangeliba761492011-01-13 15:46:58 -08002324 if (hpage)
2325 put_page(hpage);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002326#endif
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002327 try_to_freeze();
2328 if (unlikely(kthread_should_stop()))
2329 break;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002330 if (khugepaged_has_work()) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002331 if (!khugepaged_scan_sleep_millisecs)
2332 continue;
Andrea Arcangeli1dfb0592011-12-08 14:33:57 -08002333 wait_event_freezable_timeout(khugepaged_wait, false,
2334 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002335 } else if (khugepaged_enabled())
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002336 wait_event_freezable(khugepaged_wait,
2337 khugepaged_wait_event());
Andrea Arcangeliba761492011-01-13 15:46:58 -08002338 }
2339}
2340
2341static int khugepaged(void *none)
2342{
2343 struct mm_slot *mm_slot;
2344
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002345 set_freezable();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002346 set_user_nice(current, 19);
2347
2348 /* serialize with start_khugepaged() */
2349 mutex_lock(&khugepaged_mutex);
2350
2351 for (;;) {
2352 mutex_unlock(&khugepaged_mutex);
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002353 VM_BUG_ON(khugepaged_thread != current);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002354 khugepaged_loop();
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002355 VM_BUG_ON(khugepaged_thread != current);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002356
2357 mutex_lock(&khugepaged_mutex);
2358 if (!khugepaged_enabled())
2359 break;
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002360 if (unlikely(kthread_should_stop()))
2361 break;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002362 }
2363
2364 spin_lock(&khugepaged_mm_lock);
2365 mm_slot = khugepaged_scan.mm_slot;
2366 khugepaged_scan.mm_slot = NULL;
2367 if (mm_slot)
2368 collect_mm_slot(mm_slot);
2369 spin_unlock(&khugepaged_mm_lock);
2370
2371 khugepaged_thread = NULL;
2372 mutex_unlock(&khugepaged_mutex);
2373
2374 return 0;
2375}
2376
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002377void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2378{
2379 struct page *page;
2380
2381 spin_lock(&mm->page_table_lock);
2382 if (unlikely(!pmd_trans_huge(*pmd))) {
2383 spin_unlock(&mm->page_table_lock);
2384 return;
2385 }
2386 page = pmd_page(*pmd);
2387 VM_BUG_ON(!page_count(page));
2388 get_page(page);
2389 spin_unlock(&mm->page_table_lock);
2390
2391 split_huge_page(page);
2392
2393 put_page(page);
2394 BUG_ON(pmd_trans_huge(*pmd));
2395}
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002396
2397static void split_huge_page_address(struct mm_struct *mm,
2398 unsigned long address)
2399{
2400 pgd_t *pgd;
2401 pud_t *pud;
2402 pmd_t *pmd;
2403
2404 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2405
2406 pgd = pgd_offset(mm, address);
2407 if (!pgd_present(*pgd))
2408 return;
2409
2410 pud = pud_offset(pgd, address);
2411 if (!pud_present(*pud))
2412 return;
2413
2414 pmd = pmd_offset(pud, address);
2415 if (!pmd_present(*pmd))
2416 return;
2417 /*
2418 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2419 * materialize from under us.
2420 */
2421 split_huge_page_pmd(mm, pmd);
2422}
2423
2424void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2425 unsigned long start,
2426 unsigned long end,
2427 long adjust_next)
2428{
2429 /*
2430 * If the new start address isn't hpage aligned and it could
2431 * previously contain an hugepage: check if we need to split
2432 * an huge pmd.
2433 */
2434 if (start & ~HPAGE_PMD_MASK &&
2435 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2436 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2437 split_huge_page_address(vma->vm_mm, start);
2438
2439 /*
2440 * If the new end address isn't hpage aligned and it could
2441 * previously contain an hugepage: check if we need to split
2442 * an huge pmd.
2443 */
2444 if (end & ~HPAGE_PMD_MASK &&
2445 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2446 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2447 split_huge_page_address(vma->vm_mm, end);
2448
2449 /*
2450 * If we're also updating the vma->vm_next->vm_start, if the new
2451 * vm_next->vm_start isn't page aligned and it could previously
2452 * contain an hugepage: check if we need to split an huge pmd.
2453 */
2454 if (adjust_next > 0) {
2455 struct vm_area_struct *next = vma->vm_next;
2456 unsigned long nstart = next->vm_start;
2457 nstart += adjust_next << PAGE_SHIFT;
2458 if (nstart & ~HPAGE_PMD_MASK &&
2459 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2460 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2461 split_huge_page_address(next->vm_mm, nstart);
2462 }
2463}