blob: 1546655a2d78afe4dc1c954b5e0ccff77ba3437f [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/*
Jianguo Wu8bfa3f92013-11-12 15:07:16 -080030 * By default transparent hugepage support is disabled in order that avoid
31 * to risk increase the memory footprint of applications without a guaranteed
32 * benefit. When transparent hugepage support is enabled, is for all mappings,
33 * and khugepaged scans all mappings.
34 * Defrag is invoked by khugepaged hugepage allocations and by page faults
35 * for all hugepage allocations.
Andrea Arcangeliba761492011-01-13 15:46:58 -080036 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080037unsigned long transparent_hugepage_flags __read_mostly =
Andrea Arcangeli13ece882011-01-13 15:47:07 -080038#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
Andrea Arcangeliba761492011-01-13 15:46:58 -080039 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
Andrea Arcangeli13ece882011-01-13 15:47:07 -080040#endif
41#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
42 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
43#endif
Andrea Arcangelid39d33c2011-01-13 15:47:05 -080044 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
Kirill A. Shutemov79da5402012-12-12 13:51:12 -080045 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
46 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
Andrea Arcangeliba761492011-01-13 15:46:58 -080047
48/* default scan 8*512 pte (or vmas) every 30 second */
49static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
50static unsigned int khugepaged_pages_collapsed;
51static unsigned int khugepaged_full_scans;
52static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
53/* during fragmentation poll the hugepage allocator once every minute */
54static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
55static struct task_struct *khugepaged_thread __read_mostly;
56static DEFINE_MUTEX(khugepaged_mutex);
57static DEFINE_SPINLOCK(khugepaged_mm_lock);
58static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
59/*
60 * default collapse hugepages if there is at least one pte mapped like
61 * it would have happened if the vma was large enough during page
62 * fault.
63 */
64static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
65
66static int khugepaged(void *none);
Andrea Arcangeliba761492011-01-13 15:46:58 -080067static int khugepaged_slab_init(void);
Andrea Arcangeliba761492011-01-13 15:46:58 -080068
Sasha Levin43b5fbb2013-02-22 16:32:27 -080069#define MM_SLOTS_HASH_BITS 10
70static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
71
Andrea Arcangeliba761492011-01-13 15:46:58 -080072static struct kmem_cache *mm_slot_cache __read_mostly;
73
74/**
75 * struct mm_slot - hash lookup from mm to mm_slot
76 * @hash: hash collision list
77 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
78 * @mm: the mm that this information is valid for
79 */
80struct mm_slot {
81 struct hlist_node hash;
82 struct list_head mm_node;
83 struct mm_struct *mm;
84};
85
86/**
87 * struct khugepaged_scan - cursor for scanning
88 * @mm_head: the head of the mm list to scan
89 * @mm_slot: the current mm_slot we are scanning
90 * @address: the next address inside that to be scanned
91 *
92 * There is only the one khugepaged_scan instance of this cursor structure.
93 */
94struct khugepaged_scan {
95 struct list_head mm_head;
96 struct mm_slot *mm_slot;
97 unsigned long address;
H Hartley Sweeten2f1da642011-10-31 17:09:25 -070098};
99static struct khugepaged_scan khugepaged_scan = {
Andrea Arcangeliba761492011-01-13 15:46:58 -0800100 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
101};
102
Andrea Arcangelif0005652011-01-13 15:47:04 -0800103
104static int set_recommended_min_free_kbytes(void)
105{
106 struct zone *zone;
107 int nr_zones = 0;
108 unsigned long recommended_min;
Andrea Arcangelif0005652011-01-13 15:47:04 -0800109
Xiao Guangrong17c230a2012-10-08 16:29:56 -0700110 if (!khugepaged_enabled())
Andrea Arcangelif0005652011-01-13 15:47:04 -0800111 return 0;
112
113 for_each_populated_zone(zone)
114 nr_zones++;
115
116 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
117 recommended_min = pageblock_nr_pages * nr_zones * 2;
118
119 /*
120 * Make sure that on average at least two pageblocks are almost free
121 * of another type, one for a migratetype to fall back to and a
122 * second to avoid subsequent fallbacks of other types There are 3
123 * MIGRATE_TYPES we care about.
124 */
125 recommended_min += pageblock_nr_pages * nr_zones *
126 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
127
128 /* don't ever allow to reserve more than 5% of the lowmem */
129 recommended_min = min(recommended_min,
130 (unsigned long) nr_free_buffer_pages() / 20);
131 recommended_min <<= (PAGE_SHIFT-10);
132
Han Pingtian42aa83c2014-01-23 15:53:28 -0800133 if (recommended_min > min_free_kbytes) {
134 if (user_min_free_kbytes >= 0)
135 pr_info("raising min_free_kbytes from %d to %lu "
136 "to help transparent hugepage allocations\n",
137 min_free_kbytes, recommended_min);
138
Andrea Arcangelif0005652011-01-13 15:47:04 -0800139 min_free_kbytes = recommended_min;
Han Pingtian42aa83c2014-01-23 15:53:28 -0800140 }
Andrea Arcangelif0005652011-01-13 15:47:04 -0800141 setup_per_zone_wmarks();
142 return 0;
143}
144late_initcall(set_recommended_min_free_kbytes);
145
Andrea Arcangeliba761492011-01-13 15:46:58 -0800146static int start_khugepaged(void)
147{
148 int err = 0;
149 if (khugepaged_enabled()) {
Andrea Arcangeliba761492011-01-13 15:46:58 -0800150 if (!khugepaged_thread)
151 khugepaged_thread = kthread_run(khugepaged, NULL,
152 "khugepaged");
153 if (unlikely(IS_ERR(khugepaged_thread))) {
154 printk(KERN_ERR
155 "khugepaged: kthread_run(khugepaged) failed\n");
156 err = PTR_ERR(khugepaged_thread);
157 khugepaged_thread = NULL;
158 }
Xiao Guangrong911891a2012-10-08 16:29:41 -0700159
160 if (!list_empty(&khugepaged_scan.mm_head))
Andrea Arcangeliba761492011-01-13 15:46:58 -0800161 wake_up_interruptible(&khugepaged_wait);
Andrea Arcangelif0005652011-01-13 15:47:04 -0800162
163 set_recommended_min_free_kbytes();
Xiao Guangrong911891a2012-10-08 16:29:41 -0700164 } else if (khugepaged_thread) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700165 kthread_stop(khugepaged_thread);
166 khugepaged_thread = NULL;
167 }
Xiao Guangrong637e3a22012-10-08 16:29:38 -0700168
Andrea Arcangeliba761492011-01-13 15:46:58 -0800169 return err;
170}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800171
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800172static atomic_t huge_zero_refcount;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700173static struct page *huge_zero_page __read_mostly;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800174
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700175static inline bool is_huge_zero_page(struct page *page)
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800176{
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700177 return ACCESS_ONCE(huge_zero_page) == page;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800178}
179
180static inline bool is_huge_zero_pmd(pmd_t pmd)
181{
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700182 return is_huge_zero_page(pmd_page(pmd));
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800183}
184
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700185static struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800186{
187 struct page *zero_page;
188retry:
189 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700190 return ACCESS_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800191
192 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
193 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800194 if (!zero_page) {
195 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700196 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800197 }
198 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800199 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700200 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800201 preempt_enable();
202 __free_page(zero_page);
203 goto retry;
204 }
205
206 /* We take additional reference here. It will be put back by shrinker */
207 atomic_set(&huge_zero_refcount, 2);
208 preempt_enable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700209 return ACCESS_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800210}
211
212static void put_huge_zero_page(void)
213{
214 /*
215 * Counter should never go to zero here. Only shrinker can put
216 * last reference.
217 */
218 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
219}
220
Glauber Costa48896462013-08-28 10:18:15 +1000221static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
222 struct shrink_control *sc)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800223{
Glauber Costa48896462013-08-28 10:18:15 +1000224 /* we can free zero page only if last reference remains */
225 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
226}
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800227
Glauber Costa48896462013-08-28 10:18:15 +1000228static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
229 struct shrink_control *sc)
230{
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800231 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700232 struct page *zero_page = xchg(&huge_zero_page, NULL);
233 BUG_ON(zero_page == NULL);
234 __free_page(zero_page);
Glauber Costa48896462013-08-28 10:18:15 +1000235 return HPAGE_PMD_NR;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800236 }
237
238 return 0;
239}
240
241static struct shrinker huge_zero_page_shrinker = {
Glauber Costa48896462013-08-28 10:18:15 +1000242 .count_objects = shrink_huge_zero_page_count,
243 .scan_objects = shrink_huge_zero_page_scan,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800244 .seeks = DEFAULT_SEEKS,
245};
246
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800247#ifdef CONFIG_SYSFS
Andrea Arcangeliba761492011-01-13 15:46:58 -0800248
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800249static ssize_t double_flag_show(struct kobject *kobj,
250 struct kobj_attribute *attr, char *buf,
251 enum transparent_hugepage_flag enabled,
252 enum transparent_hugepage_flag req_madv)
253{
254 if (test_bit(enabled, &transparent_hugepage_flags)) {
255 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
256 return sprintf(buf, "[always] madvise never\n");
257 } else if (test_bit(req_madv, &transparent_hugepage_flags))
258 return sprintf(buf, "always [madvise] never\n");
259 else
260 return sprintf(buf, "always madvise [never]\n");
261}
262static ssize_t double_flag_store(struct kobject *kobj,
263 struct kobj_attribute *attr,
264 const char *buf, size_t count,
265 enum transparent_hugepage_flag enabled,
266 enum transparent_hugepage_flag req_madv)
267{
268 if (!memcmp("always", buf,
269 min(sizeof("always")-1, count))) {
270 set_bit(enabled, &transparent_hugepage_flags);
271 clear_bit(req_madv, &transparent_hugepage_flags);
272 } else if (!memcmp("madvise", buf,
273 min(sizeof("madvise")-1, count))) {
274 clear_bit(enabled, &transparent_hugepage_flags);
275 set_bit(req_madv, &transparent_hugepage_flags);
276 } else if (!memcmp("never", buf,
277 min(sizeof("never")-1, count))) {
278 clear_bit(enabled, &transparent_hugepage_flags);
279 clear_bit(req_madv, &transparent_hugepage_flags);
280 } else
281 return -EINVAL;
282
283 return count;
284}
285
286static ssize_t enabled_show(struct kobject *kobj,
287 struct kobj_attribute *attr, char *buf)
288{
289 return double_flag_show(kobj, attr, buf,
290 TRANSPARENT_HUGEPAGE_FLAG,
291 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
292}
293static ssize_t enabled_store(struct kobject *kobj,
294 struct kobj_attribute *attr,
295 const char *buf, size_t count)
296{
Andrea Arcangeliba761492011-01-13 15:46:58 -0800297 ssize_t ret;
298
299 ret = double_flag_store(kobj, attr, buf, count,
300 TRANSPARENT_HUGEPAGE_FLAG,
301 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
302
303 if (ret > 0) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700304 int err;
305
306 mutex_lock(&khugepaged_mutex);
307 err = start_khugepaged();
308 mutex_unlock(&khugepaged_mutex);
309
Andrea Arcangeliba761492011-01-13 15:46:58 -0800310 if (err)
311 ret = err;
312 }
313
314 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800315}
316static struct kobj_attribute enabled_attr =
317 __ATTR(enabled, 0644, enabled_show, enabled_store);
318
319static ssize_t single_flag_show(struct kobject *kobj,
320 struct kobj_attribute *attr, char *buf,
321 enum transparent_hugepage_flag flag)
322{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700323 return sprintf(buf, "%d\n",
324 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800325}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700326
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800327static ssize_t single_flag_store(struct kobject *kobj,
328 struct kobj_attribute *attr,
329 const char *buf, size_t count,
330 enum transparent_hugepage_flag flag)
331{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700332 unsigned long value;
333 int ret;
334
335 ret = kstrtoul(buf, 10, &value);
336 if (ret < 0)
337 return ret;
338 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800339 return -EINVAL;
340
Ben Hutchingse27e6152011-04-14 15:22:21 -0700341 if (value)
342 set_bit(flag, &transparent_hugepage_flags);
343 else
344 clear_bit(flag, &transparent_hugepage_flags);
345
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800346 return count;
347}
348
349/*
350 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
351 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
352 * memory just to allocate one more hugepage.
353 */
354static ssize_t defrag_show(struct kobject *kobj,
355 struct kobj_attribute *attr, char *buf)
356{
357 return double_flag_show(kobj, attr, buf,
358 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
359 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
360}
361static ssize_t defrag_store(struct kobject *kobj,
362 struct kobj_attribute *attr,
363 const char *buf, size_t count)
364{
365 return double_flag_store(kobj, attr, buf, count,
366 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
367 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
368}
369static struct kobj_attribute defrag_attr =
370 __ATTR(defrag, 0644, defrag_show, defrag_store);
371
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800372static ssize_t use_zero_page_show(struct kobject *kobj,
373 struct kobj_attribute *attr, char *buf)
374{
375 return single_flag_show(kobj, attr, buf,
376 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
377}
378static ssize_t use_zero_page_store(struct kobject *kobj,
379 struct kobj_attribute *attr, const char *buf, size_t count)
380{
381 return single_flag_store(kobj, attr, buf, count,
382 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
383}
384static struct kobj_attribute use_zero_page_attr =
385 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800386#ifdef CONFIG_DEBUG_VM
387static ssize_t debug_cow_show(struct kobject *kobj,
388 struct kobj_attribute *attr, char *buf)
389{
390 return single_flag_show(kobj, attr, buf,
391 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
392}
393static ssize_t debug_cow_store(struct kobject *kobj,
394 struct kobj_attribute *attr,
395 const char *buf, size_t count)
396{
397 return single_flag_store(kobj, attr, buf, count,
398 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
399}
400static struct kobj_attribute debug_cow_attr =
401 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
402#endif /* CONFIG_DEBUG_VM */
403
404static struct attribute *hugepage_attr[] = {
405 &enabled_attr.attr,
406 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800407 &use_zero_page_attr.attr,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800408#ifdef CONFIG_DEBUG_VM
409 &debug_cow_attr.attr,
410#endif
411 NULL,
412};
413
414static struct attribute_group hugepage_attr_group = {
415 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800416};
417
418static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
419 struct kobj_attribute *attr,
420 char *buf)
421{
422 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
423}
424
425static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
426 struct kobj_attribute *attr,
427 const char *buf, size_t count)
428{
429 unsigned long msecs;
430 int err;
431
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700432 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800433 if (err || msecs > UINT_MAX)
434 return -EINVAL;
435
436 khugepaged_scan_sleep_millisecs = msecs;
437 wake_up_interruptible(&khugepaged_wait);
438
439 return count;
440}
441static struct kobj_attribute scan_sleep_millisecs_attr =
442 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
443 scan_sleep_millisecs_store);
444
445static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
446 struct kobj_attribute *attr,
447 char *buf)
448{
449 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
450}
451
452static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
453 struct kobj_attribute *attr,
454 const char *buf, size_t count)
455{
456 unsigned long msecs;
457 int err;
458
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700459 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800460 if (err || msecs > UINT_MAX)
461 return -EINVAL;
462
463 khugepaged_alloc_sleep_millisecs = msecs;
464 wake_up_interruptible(&khugepaged_wait);
465
466 return count;
467}
468static struct kobj_attribute alloc_sleep_millisecs_attr =
469 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
470 alloc_sleep_millisecs_store);
471
472static ssize_t pages_to_scan_show(struct kobject *kobj,
473 struct kobj_attribute *attr,
474 char *buf)
475{
476 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
477}
478static ssize_t pages_to_scan_store(struct kobject *kobj,
479 struct kobj_attribute *attr,
480 const char *buf, size_t count)
481{
482 int err;
483 unsigned long pages;
484
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700485 err = kstrtoul(buf, 10, &pages);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800486 if (err || !pages || pages > UINT_MAX)
487 return -EINVAL;
488
489 khugepaged_pages_to_scan = pages;
490
491 return count;
492}
493static struct kobj_attribute pages_to_scan_attr =
494 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
495 pages_to_scan_store);
496
497static ssize_t pages_collapsed_show(struct kobject *kobj,
498 struct kobj_attribute *attr,
499 char *buf)
500{
501 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
502}
503static struct kobj_attribute pages_collapsed_attr =
504 __ATTR_RO(pages_collapsed);
505
506static ssize_t full_scans_show(struct kobject *kobj,
507 struct kobj_attribute *attr,
508 char *buf)
509{
510 return sprintf(buf, "%u\n", khugepaged_full_scans);
511}
512static struct kobj_attribute full_scans_attr =
513 __ATTR_RO(full_scans);
514
515static ssize_t khugepaged_defrag_show(struct kobject *kobj,
516 struct kobj_attribute *attr, char *buf)
517{
518 return single_flag_show(kobj, attr, buf,
519 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
520}
521static ssize_t khugepaged_defrag_store(struct kobject *kobj,
522 struct kobj_attribute *attr,
523 const char *buf, size_t count)
524{
525 return single_flag_store(kobj, attr, buf, count,
526 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
527}
528static struct kobj_attribute khugepaged_defrag_attr =
529 __ATTR(defrag, 0644, khugepaged_defrag_show,
530 khugepaged_defrag_store);
531
532/*
533 * max_ptes_none controls if khugepaged should collapse hugepages over
534 * any unmapped ptes in turn potentially increasing the memory
535 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
536 * reduce the available free memory in the system as it
537 * runs. Increasing max_ptes_none will instead potentially reduce the
538 * free memory in the system during the khugepaged scan.
539 */
540static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
541 struct kobj_attribute *attr,
542 char *buf)
543{
544 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
545}
546static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
547 struct kobj_attribute *attr,
548 const char *buf, size_t count)
549{
550 int err;
551 unsigned long max_ptes_none;
552
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700553 err = kstrtoul(buf, 10, &max_ptes_none);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800554 if (err || max_ptes_none > HPAGE_PMD_NR-1)
555 return -EINVAL;
556
557 khugepaged_max_ptes_none = max_ptes_none;
558
559 return count;
560}
561static struct kobj_attribute khugepaged_max_ptes_none_attr =
562 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
563 khugepaged_max_ptes_none_store);
564
565static struct attribute *khugepaged_attr[] = {
566 &khugepaged_defrag_attr.attr,
567 &khugepaged_max_ptes_none_attr.attr,
568 &pages_to_scan_attr.attr,
569 &pages_collapsed_attr.attr,
570 &full_scans_attr.attr,
571 &scan_sleep_millisecs_attr.attr,
572 &alloc_sleep_millisecs_attr.attr,
573 NULL,
574};
575
576static struct attribute_group khugepaged_attr_group = {
577 .attrs = khugepaged_attr,
578 .name = "khugepaged",
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800579};
Shaohua Li569e5592012-01-12 17:19:11 -0800580
581static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
582{
583 int err;
584
585 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
586 if (unlikely(!*hugepage_kobj)) {
Jeremy Eder2c797372012-12-20 15:05:32 -0800587 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800588 return -ENOMEM;
589 }
590
591 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
592 if (err) {
Jeremy Eder2c797372012-12-20 15:05:32 -0800593 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800594 goto delete_obj;
595 }
596
597 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
598 if (err) {
Jeremy Eder2c797372012-12-20 15:05:32 -0800599 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800600 goto remove_hp_group;
601 }
602
603 return 0;
604
605remove_hp_group:
606 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
607delete_obj:
608 kobject_put(*hugepage_kobj);
609 return err;
610}
611
612static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
613{
614 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
615 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
616 kobject_put(hugepage_kobj);
617}
618#else
619static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
620{
621 return 0;
622}
623
624static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
625{
626}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800627#endif /* CONFIG_SYSFS */
628
629static int __init hugepage_init(void)
630{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800631 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800632 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800633
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800634 if (!has_transparent_hugepage()) {
635 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800636 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800637 }
638
Shaohua Li569e5592012-01-12 17:19:11 -0800639 err = hugepage_init_sysfs(&hugepage_kobj);
640 if (err)
641 return err;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800642
643 err = khugepaged_slab_init();
644 if (err)
645 goto out;
646
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800647 register_shrinker(&huge_zero_page_shrinker);
648
Rik van Riel97562cd2011-01-13 15:47:12 -0800649 /*
650 * By default disable transparent hugepages on smaller systems,
651 * where the extra memory used could hurt more than TLB overhead
652 * is likely to save. The admin can still enable it through /sys.
653 */
654 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
655 transparent_hugepage_flags = 0;
656
Andrea Arcangeliba761492011-01-13 15:46:58 -0800657 start_khugepaged();
658
Shaohua Li569e5592012-01-12 17:19:11 -0800659 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800660out:
Shaohua Li569e5592012-01-12 17:19:11 -0800661 hugepage_exit_sysfs(hugepage_kobj);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800662 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800663}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800664subsys_initcall(hugepage_init);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800665
666static int __init setup_transparent_hugepage(char *str)
667{
668 int ret = 0;
669 if (!str)
670 goto out;
671 if (!strcmp(str, "always")) {
672 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
673 &transparent_hugepage_flags);
674 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
675 &transparent_hugepage_flags);
676 ret = 1;
677 } else if (!strcmp(str, "madvise")) {
678 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
679 &transparent_hugepage_flags);
680 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
681 &transparent_hugepage_flags);
682 ret = 1;
683 } else if (!strcmp(str, "never")) {
684 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
685 &transparent_hugepage_flags);
686 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
687 &transparent_hugepage_flags);
688 ret = 1;
689 }
690out:
691 if (!ret)
692 printk(KERN_WARNING
693 "transparent_hugepage= cannot parse, ignored\n");
694 return ret;
695}
696__setup("transparent_hugepage=", setup_transparent_hugepage);
697
Mel Gormanb32967f2012-11-19 12:35:47 +0000698pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800699{
700 if (likely(vma->vm_flags & VM_WRITE))
701 pmd = pmd_mkwrite(pmd);
702 return pmd;
703}
704
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700705static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
Bob Liub3092b32012-12-11 16:00:41 -0800706{
707 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700708 entry = mk_pmd(page, prot);
Bob Liub3092b32012-12-11 16:00:41 -0800709 entry = pmd_mkhuge(entry);
710 return entry;
711}
712
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800713static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
714 struct vm_area_struct *vma,
715 unsigned long haddr, pmd_t *pmd,
716 struct page *page)
717{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800718 pgtable_t pgtable;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800719 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800720
Sasha Levin309381fea2014-01-23 15:52:54 -0800721 VM_BUG_ON_PAGE(!PageCompound(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800722 pgtable = pte_alloc_one(mm, haddr);
David Rientjesedad9d22012-05-29 15:06:17 -0700723 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800724 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800725
726 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700727 /*
728 * The memory barrier inside __SetPageUptodate makes sure that
729 * clear_huge_page writes become visible before the set_pmd_at()
730 * write.
731 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800732 __SetPageUptodate(page);
733
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800734 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800735 if (unlikely(!pmd_none(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800736 spin_unlock(ptl);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800737 mem_cgroup_uncharge_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800738 put_page(page);
739 pte_free(mm, pgtable);
740 } else {
741 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700742 entry = mk_huge_pmd(page, vma->vm_page_prot);
743 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800744 page_add_new_anon_rmap(page, vma, haddr);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700745 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800746 set_pmd_at(mm, haddr, pmd, entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800747 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800748 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800749 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800750 }
751
David Rientjesaa2e8782012-05-29 15:06:17 -0700752 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800753}
754
Andi Kleencc5d4622011-03-22 16:33:13 -0700755static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800756{
Andi Kleencc5d4622011-03-22 16:33:13 -0700757 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800758}
759
760static inline struct page *alloc_hugepage_vma(int defrag,
761 struct vm_area_struct *vma,
Andi Kleencc5d4622011-03-22 16:33:13 -0700762 unsigned long haddr, int nd,
763 gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800764{
Andi Kleencc5d4622011-03-22 16:33:13 -0700765 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
Andi Kleen5c4b4be2011-03-04 17:36:32 -0800766 HPAGE_PMD_ORDER, vma, haddr, nd);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800767}
768
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800769/* Caller must hold page table lock. */
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800770static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800771 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700772 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800773{
774 pmd_t entry;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800775 if (!pmd_none(*pmd))
776 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700777 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800778 entry = pmd_wrprotect(entry);
779 entry = pmd_mkhuge(entry);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700780 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800781 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800782 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800783 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800784}
785
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800786int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
787 unsigned long address, pmd_t *pmd,
788 unsigned int flags)
789{
790 struct page *page;
791 unsigned long haddr = address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800792
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700793 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700794 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700795 if (unlikely(anon_vma_prepare(vma)))
796 return VM_FAULT_OOM;
797 if (unlikely(khugepaged_enter(vma)))
798 return VM_FAULT_OOM;
799 if (!(flags & FAULT_FLAG_WRITE) &&
800 transparent_hugepage_use_zero_page()) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800801 spinlock_t *ptl;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700802 pgtable_t pgtable;
803 struct page *zero_page;
804 bool set;
805 pgtable = pte_alloc_one(mm, haddr);
806 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800807 return VM_FAULT_OOM;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700808 zero_page = get_huge_zero_page();
809 if (unlikely(!zero_page)) {
810 pte_free(mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700811 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700812 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700813 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800814 ptl = pmd_lock(mm, pmd);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700815 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
816 zero_page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800817 spin_unlock(ptl);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700818 if (!set) {
819 pte_free(mm, pgtable);
820 put_huge_zero_page();
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800821 }
David Rientjesedad9d22012-05-29 15:06:17 -0700822 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800823 }
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700824 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
825 vma, haddr, numa_node_id(), 0);
826 if (unlikely(!page)) {
827 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700828 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700829 }
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700830 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
831 put_page(page);
David Rientjes17766dd2013-09-12 15:14:06 -0700832 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700833 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700834 }
835 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
836 mem_cgroup_uncharge_page(page);
837 put_page(page);
David Rientjes17766dd2013-09-12 15:14:06 -0700838 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700839 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700840 }
841
David Rientjes17766dd2013-09-12 15:14:06 -0700842 count_vm_event(THP_FAULT_ALLOC);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700843 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800844}
845
846int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
847 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
848 struct vm_area_struct *vma)
849{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800850 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800851 struct page *src_page;
852 pmd_t pmd;
853 pgtable_t pgtable;
854 int ret;
855
856 ret = -ENOMEM;
857 pgtable = pte_alloc_one(dst_mm, addr);
858 if (unlikely(!pgtable))
859 goto out;
860
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800861 dst_ptl = pmd_lock(dst_mm, dst_pmd);
862 src_ptl = pmd_lockptr(src_mm, src_pmd);
863 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800864
865 ret = -EAGAIN;
866 pmd = *src_pmd;
867 if (unlikely(!pmd_trans_huge(pmd))) {
868 pte_free(dst_mm, pgtable);
869 goto out_unlock;
870 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800871 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800872 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800873 * under splitting since we don't split the page itself, only pmd to
874 * a page table.
875 */
876 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700877 struct page *zero_page;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800878 bool set;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800879 /*
880 * get_huge_zero_page() will never allocate a new page here,
881 * since we already have a zero page to copy. It just takes a
882 * reference.
883 */
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700884 zero_page = get_huge_zero_page();
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800885 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700886 zero_page);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800887 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800888 ret = 0;
889 goto out_unlock;
890 }
Mel Gormande466bd2013-12-18 17:08:42 -0800891
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800892 if (unlikely(pmd_trans_splitting(pmd))) {
893 /* split huge page running from under us */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800894 spin_unlock(src_ptl);
895 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800896 pte_free(dst_mm, pgtable);
897
898 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
899 goto out;
900 }
901 src_page = pmd_page(pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -0800902 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800903 get_page(src_page);
904 page_dup_rmap(src_page);
905 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
906
907 pmdp_set_wrprotect(src_mm, addr, src_pmd);
908 pmd = pmd_mkold(pmd_wrprotect(pmd));
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700909 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800910 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800911 atomic_long_inc(&dst_mm->nr_ptes);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800912
913 ret = 0;
914out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800915 spin_unlock(src_ptl);
916 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800917out:
918 return ret;
919}
920
Will Deacona1dd4502012-12-11 16:01:27 -0800921void huge_pmd_set_accessed(struct mm_struct *mm,
922 struct vm_area_struct *vma,
923 unsigned long address,
924 pmd_t *pmd, pmd_t orig_pmd,
925 int dirty)
926{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800927 spinlock_t *ptl;
Will Deacona1dd4502012-12-11 16:01:27 -0800928 pmd_t entry;
929 unsigned long haddr;
930
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800931 ptl = pmd_lock(mm, pmd);
Will Deacona1dd4502012-12-11 16:01:27 -0800932 if (unlikely(!pmd_same(*pmd, orig_pmd)))
933 goto unlock;
934
935 entry = pmd_mkyoung(orig_pmd);
936 haddr = address & HPAGE_PMD_MASK;
937 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
938 update_mmu_cache_pmd(vma, address, pmd);
939
940unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800941 spin_unlock(ptl);
Will Deacona1dd4502012-12-11 16:01:27 -0800942}
943
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800944static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
945 struct vm_area_struct *vma, unsigned long address,
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800946 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800947{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800948 spinlock_t *ptl;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800949 pgtable_t pgtable;
950 pmd_t _pmd;
951 struct page *page;
952 int i, ret = 0;
953 unsigned long mmun_start; /* For mmu_notifiers */
954 unsigned long mmun_end; /* For mmu_notifiers */
955
956 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
957 if (!page) {
958 ret |= VM_FAULT_OOM;
959 goto out;
960 }
961
962 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
963 put_page(page);
964 ret |= VM_FAULT_OOM;
965 goto out;
966 }
967
968 clear_user_highpage(page, address);
969 __SetPageUptodate(page);
970
971 mmun_start = haddr;
972 mmun_end = haddr + HPAGE_PMD_SIZE;
973 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
974
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800975 ptl = pmd_lock(mm, pmd);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800976 if (unlikely(!pmd_same(*pmd, orig_pmd)))
977 goto out_free_page;
978
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800979 pmdp_clear_flush(vma, haddr, pmd);
980 /* leave pmd empty until pte is filled */
981
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700982 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800983 pmd_populate(mm, &_pmd, pgtable);
984
985 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
986 pte_t *pte, entry;
987 if (haddr == (address & PAGE_MASK)) {
988 entry = mk_pte(page, vma->vm_page_prot);
989 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
990 page_add_new_anon_rmap(page, vma, haddr);
991 } else {
992 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
993 entry = pte_mkspecial(entry);
994 }
995 pte = pte_offset_map(&_pmd, haddr);
996 VM_BUG_ON(!pte_none(*pte));
997 set_pte_at(mm, haddr, pte, entry);
998 pte_unmap(pte);
999 }
1000 smp_wmb(); /* make pte visible before pmd */
1001 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001002 spin_unlock(ptl);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001003 put_huge_zero_page();
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001004 inc_mm_counter(mm, MM_ANONPAGES);
1005
1006 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1007
1008 ret |= VM_FAULT_WRITE;
1009out:
1010 return ret;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -08001011out_free_page:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001012 spin_unlock(ptl);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -08001013 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1014 mem_cgroup_uncharge_page(page);
1015 put_page(page);
1016 goto out;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001017}
1018
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001019static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1020 struct vm_area_struct *vma,
1021 unsigned long address,
1022 pmd_t *pmd, pmd_t orig_pmd,
1023 struct page *page,
1024 unsigned long haddr)
1025{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001026 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001027 pgtable_t pgtable;
1028 pmd_t _pmd;
1029 int ret = 0, i;
1030 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001031 unsigned long mmun_start; /* For mmu_notifiers */
1032 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001033
1034 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1035 GFP_KERNEL);
1036 if (unlikely(!pages)) {
1037 ret |= VM_FAULT_OOM;
1038 goto out;
1039 }
1040
1041 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -07001042 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1043 __GFP_OTHER_NODE,
Andi Kleen19ee1512011-03-04 17:36:31 -08001044 vma, address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001045 if (unlikely(!pages[i] ||
1046 mem_cgroup_newpage_charge(pages[i], mm,
1047 GFP_KERNEL))) {
1048 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001049 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001050 mem_cgroup_uncharge_start();
1051 while (--i >= 0) {
1052 mem_cgroup_uncharge_page(pages[i]);
1053 put_page(pages[i]);
1054 }
1055 mem_cgroup_uncharge_end();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001056 kfree(pages);
1057 ret |= VM_FAULT_OOM;
1058 goto out;
1059 }
1060 }
1061
1062 for (i = 0; i < HPAGE_PMD_NR; i++) {
1063 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001064 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001065 __SetPageUptodate(pages[i]);
1066 cond_resched();
1067 }
1068
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001069 mmun_start = haddr;
1070 mmun_end = haddr + HPAGE_PMD_SIZE;
1071 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1072
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001073 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001074 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1075 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001076 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001077
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001078 pmdp_clear_flush(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001079 /* leave pmd empty until pte is filled */
1080
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001081 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001082 pmd_populate(mm, &_pmd, pgtable);
1083
1084 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1085 pte_t *pte, entry;
1086 entry = mk_pte(pages[i], vma->vm_page_prot);
1087 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1088 page_add_new_anon_rmap(pages[i], vma, haddr);
1089 pte = pte_offset_map(&_pmd, haddr);
1090 VM_BUG_ON(!pte_none(*pte));
1091 set_pte_at(mm, haddr, pte, entry);
1092 pte_unmap(pte);
1093 }
1094 kfree(pages);
1095
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001096 smp_wmb(); /* make pte visible before pmd */
1097 pmd_populate(mm, pmd, pgtable);
1098 page_remove_rmap(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001099 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001100
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001101 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1102
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001103 ret |= VM_FAULT_WRITE;
1104 put_page(page);
1105
1106out:
1107 return ret;
1108
1109out_free_pages:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001110 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001111 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001112 mem_cgroup_uncharge_start();
1113 for (i = 0; i < HPAGE_PMD_NR; i++) {
1114 mem_cgroup_uncharge_page(pages[i]);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001115 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001116 }
1117 mem_cgroup_uncharge_end();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001118 kfree(pages);
1119 goto out;
1120}
1121
1122int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1123 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1124{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001125 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001126 int ret = 0;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001127 struct page *page = NULL, *new_page;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001128 unsigned long haddr;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001129 unsigned long mmun_start; /* For mmu_notifiers */
1130 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001131
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001132 ptl = pmd_lockptr(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001133 VM_BUG_ON(!vma->anon_vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001134 haddr = address & HPAGE_PMD_MASK;
1135 if (is_huge_zero_pmd(orig_pmd))
1136 goto alloc;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001137 spin_lock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001138 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1139 goto out_unlock;
1140
1141 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001142 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001143 if (page_mapcount(page) == 1) {
1144 pmd_t entry;
1145 entry = pmd_mkyoung(orig_pmd);
1146 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1147 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
David Millerb113da62012-10-08 16:34:25 -07001148 update_mmu_cache_pmd(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001149 ret |= VM_FAULT_WRITE;
1150 goto out_unlock;
1151 }
1152 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001153 spin_unlock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001154alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001155 if (transparent_hugepage_enabled(vma) &&
1156 !transparent_hugepage_debug_cow())
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001157 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
Andi Kleencc5d4622011-03-22 16:33:13 -07001158 vma, haddr, numa_node_id(), 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001159 else
1160 new_page = NULL;
1161
1162 if (unlikely(!new_page)) {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001163 if (!page) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001164 ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -08001165 address, pmd, orig_pmd, haddr);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001166 } else {
1167 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1168 pmd, orig_pmd, page, haddr);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001169 if (ret & VM_FAULT_OOM) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001170 split_huge_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001171 ret |= VM_FAULT_FALLBACK;
1172 }
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001173 put_page(page);
1174 }
David Rientjes17766dd2013-09-12 15:14:06 -07001175 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001176 goto out;
1177 }
1178
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001179 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1180 put_page(new_page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001181 if (page) {
1182 split_huge_page(page);
1183 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001184 } else
1185 split_huge_page_pmd(vma, address, pmd);
1186 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001187 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001188 goto out;
1189 }
1190
David Rientjes17766dd2013-09-12 15:14:06 -07001191 count_vm_event(THP_FAULT_ALLOC);
1192
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001193 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001194 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
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001203 spin_lock(ptl);
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))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001207 spin_unlock(ptl);
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);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001219 if (!page) {
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 {
Sasha Levin309381fea2014-01-23 15:52:54 -08001223 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001224 page_remove_rmap(page);
1225 put_page(page);
1226 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001227 ret |= VM_FAULT_WRITE;
1228 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001229 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001230out_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:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001235 spin_unlock(ptl);
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
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001247 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001248
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
Mel Gorman2b4847e2013-12-18 17:08:32 -08001256 /* Full NUMA hinting faults to serialise migration in fault paths */
1257 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1258 goto out;
1259
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001260 page = pmd_page(*pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001261 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001262 if (flags & FOLL_TOUCH) {
1263 pmd_t _pmd;
1264 /*
1265 * We should set the dirty bit only for FOLL_WRITE but
1266 * for now the dirty bit in the pmd is meaningless.
1267 * And if the dirty bit will become meaningful and
1268 * we'll only set it with FOLL_WRITE, an atomic
1269 * set_bit will be required on the pmd to set the
1270 * young bit, instead of the current set_pmd_at.
1271 */
1272 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
Aneesh Kumar K.V8663890a2013-06-06 00:20:34 -07001273 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1274 pmd, _pmd, 1))
1275 update_mmu_cache_pmd(vma, addr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001276 }
David Rientjesb676b292012-10-08 16:34:03 -07001277 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1278 if (page->mapping && trylock_page(page)) {
1279 lru_add_drain();
1280 if (page->mapping)
1281 mlock_vma_page(page);
1282 unlock_page(page);
1283 }
1284 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001285 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Sasha Levin309381fea2014-01-23 15:52:54 -08001286 VM_BUG_ON_PAGE(!PageCompound(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001287 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001288 get_page_foll(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001289
1290out:
1291 return page;
1292}
1293
Mel Gormand10e63f2012-10-25 14:16:31 +02001294/* NUMA hinting page fault entry point for trans huge pmds */
Mel Gorman4daae3b2012-11-02 11:33:45 +00001295int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1296 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
Mel Gormand10e63f2012-10-25 14:16:31 +02001297{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001298 spinlock_t *ptl;
Mel Gormanb8916632013-10-07 11:28:44 +01001299 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001300 struct page *page;
Mel Gormand10e63f2012-10-25 14:16:31 +02001301 unsigned long haddr = addr & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001302 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001303 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001304 bool page_locked;
1305 bool migrated = false;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001306 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001307
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001308 ptl = pmd_lock(mm, pmdp);
Mel Gormand10e63f2012-10-25 14:16:31 +02001309 if (unlikely(!pmd_same(pmd, *pmdp)))
1310 goto out_unlock;
1311
Mel Gormande466bd2013-12-18 17:08:42 -08001312 /*
1313 * If there are potential migrations, wait for completion and retry
1314 * without disrupting NUMA hinting information. Do not relock and
1315 * check_same as the page may no longer be mapped.
1316 */
1317 if (unlikely(pmd_trans_migrating(*pmdp))) {
1318 spin_unlock(ptl);
1319 wait_migrate_huge_page(vma->anon_vma, pmdp);
1320 goto out;
1321 }
1322
Mel Gormand10e63f2012-10-25 14:16:31 +02001323 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001324 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001325 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001326 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001327 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001328 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001329 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001330 flags |= TNF_FAULT_LOCAL;
1331 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001332
Mel Gormanff9042b2013-10-07 11:28:43 +01001333 /*
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001334 * Avoid grouping on DSO/COW pages in specific and RO pages
1335 * in general, RO pages shouldn't hurt as much anyway since
1336 * they can be in shared cache state.
1337 */
1338 if (!pmd_write(pmd))
1339 flags |= TNF_NO_GROUP;
1340
1341 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001342 * Acquire the page lock to serialise THP migrations but avoid dropping
1343 * page_table_lock if at all possible
1344 */
Mel Gormanb8916632013-10-07 11:28:44 +01001345 page_locked = trylock_page(page);
1346 target_nid = mpol_misplaced(page, vma, haddr);
1347 if (target_nid == -1) {
1348 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001349 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001350 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001351 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001352
Mel Gormande466bd2013-12-18 17:08:42 -08001353 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001354 if (!page_locked) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001355 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001356 wait_on_page_locked(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001357 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001358 goto out;
1359 }
1360
Mel Gorman2b4847e2013-12-18 17:08:32 -08001361 /*
1362 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1363 * to serialises splits
1364 */
Mel Gormanb8916632013-10-07 11:28:44 +01001365 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001366 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001367 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001368
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001369 /* Confirm the PMD did not change while page_table_lock was released */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001370 spin_lock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001371 if (unlikely(!pmd_same(pmd, *pmdp))) {
1372 unlock_page(page);
1373 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001374 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001375 goto out_unlock;
1376 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001377
Mel Gormanc3a489c2013-12-18 17:08:38 -08001378 /* Bail if we fail to protect against THP splits for any reason */
1379 if (unlikely(!anon_vma)) {
1380 put_page(page);
1381 page_nid = -1;
1382 goto clear_pmdnuma;
1383 }
1384
Mel Gormana54a4072013-10-07 11:28:46 +01001385 /*
1386 * Migrate the THP to the requested node, returns with page unlocked
1387 * and pmd_numa cleared.
1388 */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001389 spin_unlock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001390 migrated = migrate_misplaced_transhuge_page(mm, vma,
Hugh Dickins340ef392013-02-22 16:34:33 -08001391 pmdp, pmd, addr, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001392 if (migrated) {
1393 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001394 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001395 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001396
Mel Gorman8191acb2013-10-07 11:28:45 +01001397 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001398clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001399 BUG_ON(!PageLocked(page));
Mel Gormand10e63f2012-10-25 14:16:31 +02001400 pmd = pmd_mknonnuma(pmd);
1401 set_pmd_at(mm, haddr, pmdp, pmd);
1402 VM_BUG_ON(pmd_numa(*pmdp));
1403 update_mmu_cache_pmd(vma, addr, pmdp);
Mel Gormana54a4072013-10-07 11:28:46 +01001404 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001405out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001406 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001407
1408out:
1409 if (anon_vma)
1410 page_unlock_anon_vma_read(anon_vma);
1411
Mel Gorman8191acb2013-10-07 11:28:45 +01001412 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001413 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001414
Mel Gormand10e63f2012-10-25 14:16:31 +02001415 return 0;
1416}
1417
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001418int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001419 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001420{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001421 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001422 int ret = 0;
1423
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001424 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001425 struct page *page;
1426 pgtable_t pgtable;
David Millerf5c8ad42012-10-08 16:34:26 -07001427 pmd_t orig_pmd;
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001428 /*
1429 * For architectures like ppc64 we look at deposited pgtable
1430 * when calling pmdp_get_and_clear. So do the
1431 * pgtable_trans_huge_withdraw after finishing pmdp related
1432 * operations.
1433 */
David Millerf5c8ad42012-10-08 16:34:26 -07001434 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001435 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001436 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001437 if (is_huge_zero_pmd(orig_pmd)) {
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -08001438 atomic_long_dec(&tlb->mm->nr_ptes);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001439 spin_unlock(ptl);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001440 put_huge_zero_page();
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001441 } else {
1442 page = pmd_page(orig_pmd);
1443 page_remove_rmap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001444 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001445 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
Sasha Levin309381fea2014-01-23 15:52:54 -08001446 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -08001447 atomic_long_dec(&tlb->mm->nr_ptes);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001448 spin_unlock(ptl);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001449 tlb_remove_page(tlb, page);
1450 }
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001451 pte_free(tlb->mm, pgtable);
1452 ret = 1;
1453 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001454 return ret;
1455}
1456
Johannes Weiner0ca16342011-01-13 15:47:02 -08001457int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1458 unsigned long addr, unsigned long end,
1459 unsigned char *vec)
1460{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001461 spinlock_t *ptl;
Johannes Weiner0ca16342011-01-13 15:47:02 -08001462 int ret = 0;
1463
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001464 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001465 /*
1466 * All logical pages in the range are present
1467 * if backed by a huge page.
1468 */
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001469 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001470 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1471 ret = 1;
1472 }
Johannes Weiner0ca16342011-01-13 15:47:02 -08001473
1474 return ret;
1475}
1476
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001477int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1478 unsigned long old_addr,
1479 unsigned long new_addr, unsigned long old_end,
1480 pmd_t *old_pmd, pmd_t *new_pmd)
1481{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001482 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001483 int ret = 0;
1484 pmd_t pmd;
1485
1486 struct mm_struct *mm = vma->vm_mm;
1487
1488 if ((old_addr & ~HPAGE_PMD_MASK) ||
1489 (new_addr & ~HPAGE_PMD_MASK) ||
1490 old_end - old_addr < HPAGE_PMD_SIZE ||
1491 (new_vma->vm_flags & VM_NOHUGEPAGE))
1492 goto out;
1493
1494 /*
1495 * The destination pmd shouldn't be established, free_pgtables()
1496 * should have release it.
1497 */
1498 if (WARN_ON(!pmd_none(*new_pmd))) {
1499 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1500 goto out;
1501 }
1502
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001503 /*
1504 * We don't have to worry about the ordering of src and dst
1505 * ptlocks because exclusive mmap_sem prevents deadlock.
1506 */
1507 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001508 if (ret == 1) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001509 new_ptl = pmd_lockptr(mm, new_pmd);
1510 if (new_ptl != old_ptl)
1511 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001512 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1513 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001514
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301515 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1516 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001517 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1518 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001519 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301520 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1521 if (new_ptl != old_ptl)
1522 spin_unlock(new_ptl);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001523 spin_unlock(old_ptl);
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001524 }
1525out:
1526 return ret;
1527}
1528
Mel Gormanf123d742013-10-07 11:28:49 +01001529/*
1530 * Returns
1531 * - 0 if PMD could not be locked
1532 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1533 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1534 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001535int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001536 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001537{
1538 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001539 spinlock_t *ptl;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001540 int ret = 0;
1541
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001542 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001543 pmd_t entry;
Mel Gormanf123d742013-10-07 11:28:49 +01001544 ret = 1;
Hugh Dickinsa4f1de12012-12-16 18:56:58 -08001545 if (!prot_numa) {
Mel Gormanf123d742013-10-07 11:28:49 +01001546 entry = pmdp_get_and_clear(mm, addr, pmd);
Mel Gorman16679182013-12-18 17:08:41 -08001547 if (pmd_numa(entry))
1548 entry = pmd_mknonnuma(entry);
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001549 entry = pmd_modify(entry, newprot);
Mel Gormanf123d742013-10-07 11:28:49 +01001550 ret = HPAGE_PMD_NR;
Aneesh Kumar K.V56eecdb2014-02-12 09:13:38 +05301551 set_pmd_at(mm, addr, pmd, entry);
Hugh Dickinsa4f1de12012-12-16 18:56:58 -08001552 BUG_ON(pmd_write(entry));
1553 } else {
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001554 struct page *page = pmd_page(*pmd);
1555
Mel Gormana1a46182013-10-07 11:28:50 +01001556 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001557 * Do not trap faults against the zero page. The
1558 * read-only data is likely to be read-cached on the
1559 * local CPU cache and it is less useful to know about
1560 * local vs remote hits on the zero page.
Mel Gormana1a46182013-10-07 11:28:50 +01001561 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001562 if (!is_huge_zero_page(page) &&
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001563 !pmd_numa(*pmd)) {
Aneesh Kumar K.V56eecdb2014-02-12 09:13:38 +05301564 pmdp_set_numa(mm, addr, pmd);
Mel Gormanf123d742013-10-07 11:28:49 +01001565 ret = HPAGE_PMD_NR;
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001566 }
1567 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001568 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001569 }
Johannes Weinercd7548a2011-01-13 15:47:04 -08001570
1571 return ret;
1572}
1573
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001574/*
1575 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1576 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1577 *
1578 * Note that if it returns 1, this routine returns without unlocking page
1579 * table locks. So callers must unlock them.
1580 */
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001581int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1582 spinlock_t **ptl)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001583{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001584 *ptl = pmd_lock(vma->vm_mm, pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001585 if (likely(pmd_trans_huge(*pmd))) {
1586 if (unlikely(pmd_trans_splitting(*pmd))) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001587 spin_unlock(*ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001588 wait_split_huge_page(vma->anon_vma, pmd);
1589 return -1;
1590 } else {
1591 /* Thp mapped by 'pmd' is stable, so we can
1592 * handle it as it is. */
1593 return 1;
1594 }
1595 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001596 spin_unlock(*ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001597 return 0;
1598}
1599
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001600/*
1601 * This function returns whether a given @page is mapped onto the @address
1602 * in the virtual space of @mm.
1603 *
1604 * When it's true, this function returns *pmd with holding the page table lock
1605 * and passing it back to the caller via @ptl.
1606 * If it's false, returns NULL without holding the page table lock.
1607 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001608pmd_t *page_check_address_pmd(struct page *page,
1609 struct mm_struct *mm,
1610 unsigned long address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001611 enum page_check_address_pmd_flag flag,
1612 spinlock_t **ptl)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001613{
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001614 pmd_t *pmd;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001615
1616 if (address & ~HPAGE_PMD_MASK)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001617 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001618
Bob Liu62190492012-12-11 16:00:37 -08001619 pmd = mm_find_pmd(mm, address);
1620 if (!pmd)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001621 return NULL;
1622 *ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001623 if (pmd_none(*pmd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001624 goto unlock;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001625 if (pmd_page(*pmd) != page)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001626 goto unlock;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001627 /*
1628 * split_vma() may create temporary aliased mappings. There is
1629 * no risk as long as all huge pmd are found and have their
1630 * splitting bit set before __split_huge_page_refcount
1631 * runs. Finding the same huge pmd more than once during the
1632 * same rmap walk is not a problem.
1633 */
1634 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1635 pmd_trans_splitting(*pmd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001636 goto unlock;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001637 if (pmd_trans_huge(*pmd)) {
1638 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1639 !pmd_trans_splitting(*pmd));
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001640 return pmd;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001641 }
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001642unlock:
1643 spin_unlock(*ptl);
1644 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001645}
1646
1647static int __split_huge_page_splitting(struct page *page,
1648 struct vm_area_struct *vma,
1649 unsigned long address)
1650{
1651 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001652 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001653 pmd_t *pmd;
1654 int ret = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001655 /* For mmu_notifiers */
1656 const unsigned long mmun_start = address;
1657 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001658
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001659 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001660 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001661 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001662 if (pmd) {
1663 /*
1664 * We can't temporarily set the pmd to null in order
1665 * to split it, the pmd must remain marked huge at all
1666 * times or the VM won't take the pmd_trans_huge paths
Ingo Molnar5a505082012-12-02 19:56:46 +00001667 * and it won't wait on the anon_vma->root->rwsem to
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001668 * serialize against split_huge_page*.
1669 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001670 pmdp_splitting_flush(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001671 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001672 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001673 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001674 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001675
1676 return ret;
1677}
1678
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001679static void __split_huge_page_refcount(struct page *page,
1680 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001681{
1682 int i;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001683 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001684 struct lruvec *lruvec;
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001685 int tail_count = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001686
1687 /* prevent PageLRU to go away from under us, and freeze lru stats */
1688 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001689 lruvec = mem_cgroup_page_lruvec(page, zone);
1690
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001691 compound_lock(page);
KAMEZAWA Hiroyukie94c8a92012-01-12 17:18:20 -08001692 /* complete memcg works before add pages to LRU */
1693 mem_cgroup_split_huge_fixup(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001694
Shaohua Li45676882012-01-12 17:19:18 -08001695 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001696 struct page *page_tail = page + i;
1697
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001698 /* tail_page->_mapcount cannot change */
1699 BUG_ON(page_mapcount(page_tail) < 0);
1700 tail_count += page_mapcount(page_tail);
1701 /* check for overflow */
1702 BUG_ON(tail_count < 0);
1703 BUG_ON(atomic_read(&page_tail->_count) != 0);
1704 /*
1705 * tail_page->_count is zero and not changing from
1706 * under us. But get_page_unless_zero() may be running
1707 * from under us on the tail_page. If we used
1708 * atomic_set() below instead of atomic_add(), we
1709 * would then run atomic_set() concurrently with
1710 * get_page_unless_zero(), and atomic_set() is
1711 * implemented in C not using locked ops. spin_unlock
1712 * on x86 sometime uses locked ops because of PPro
1713 * errata 66, 92, so unless somebody can guarantee
1714 * atomic_set() here would be safe on all archs (and
1715 * not only on x86), it's safer to use atomic_add().
1716 */
1717 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1718 &page_tail->_count);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001719
1720 /* after clearing PageTail the gup refcount can be released */
1721 smp_mb();
1722
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001723 /*
1724 * retain hwpoison flag of the poisoned tail page:
1725 * fix for the unsuitable process killed on Guest Machine(KVM)
1726 * by the memory-failure.
1727 */
1728 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001729 page_tail->flags |= (page->flags &
1730 ((1L << PG_referenced) |
1731 (1L << PG_swapbacked) |
1732 (1L << PG_mlocked) |
Kirill A. Shutemove180cf82013-07-31 13:53:39 -07001733 (1L << PG_uptodate) |
1734 (1L << PG_active) |
1735 (1L << PG_unevictable)));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001736 page_tail->flags |= (1L << PG_dirty);
1737
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001738 /* clear PageTail before overwriting first_page */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001739 smp_wmb();
1740
1741 /*
1742 * __split_huge_page_splitting() already set the
1743 * splitting bit in all pmd that could map this
1744 * hugepage, that will ensure no CPU can alter the
1745 * mapcount on the head page. The mapcount is only
1746 * accounted in the head page and it has to be
1747 * transferred to all tail pages in the below code. So
1748 * for this code to be safe, the split the mapcount
1749 * can't change. But that doesn't mean userland can't
1750 * keep changing and reading the page contents while
1751 * we transfer the mapcount, so the pmd splitting
1752 * status is achieved setting a reserved bit in the
1753 * pmd, not by clearing the present bit.
1754 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001755 page_tail->_mapcount = page->_mapcount;
1756
1757 BUG_ON(page_tail->mapping);
1758 page_tail->mapping = page->mapping;
1759
Shaohua Li45676882012-01-12 17:19:18 -08001760 page_tail->index = page->index + i;
Peter Zijlstra90572892013-10-07 11:29:20 +01001761 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001762
1763 BUG_ON(!PageAnon(page_tail));
1764 BUG_ON(!PageUptodate(page_tail));
1765 BUG_ON(!PageDirty(page_tail));
1766 BUG_ON(!PageSwapBacked(page_tail));
1767
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001768 lru_add_page_tail(page, page_tail, lruvec, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001769 }
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001770 atomic_sub(tail_count, &page->_count);
1771 BUG_ON(atomic_read(&page->_count) <= 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001772
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001773 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001774
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001775 ClearPageCompound(page);
1776 compound_unlock(page);
1777 spin_unlock_irq(&zone->lru_lock);
1778
1779 for (i = 1; i < HPAGE_PMD_NR; i++) {
1780 struct page *page_tail = page + i;
1781 BUG_ON(page_count(page_tail) <= 0);
1782 /*
1783 * Tail pages may be freed if there wasn't any mapping
1784 * like if add_to_swap() is running on a lru page that
1785 * had its mapping zapped. And freeing these pages
1786 * requires taking the lru_lock so we do the put_page
1787 * of the tail pages after the split is complete.
1788 */
1789 put_page(page_tail);
1790 }
1791
1792 /*
1793 * Only the head page (now become a regular page) is required
1794 * to be pinned by the caller.
1795 */
1796 BUG_ON(page_count(page) <= 0);
1797}
1798
1799static int __split_huge_page_map(struct page *page,
1800 struct vm_area_struct *vma,
1801 unsigned long address)
1802{
1803 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001804 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001805 pmd_t *pmd, _pmd;
1806 int ret = 0, i;
1807 pgtable_t pgtable;
1808 unsigned long haddr;
1809
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001810 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001811 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001812 if (pmd) {
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001813 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001814 pmd_populate(mm, &_pmd, pgtable);
1815
Gerald Schaefere3ebcf62012-10-08 16:30:07 -07001816 haddr = address;
1817 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001818 pte_t *pte, entry;
1819 BUG_ON(PageCompound(page+i));
1820 entry = mk_pte(page + i, vma->vm_page_prot);
1821 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1822 if (!pmd_write(*pmd))
1823 entry = pte_wrprotect(entry);
1824 else
1825 BUG_ON(page_mapcount(page) != 1);
1826 if (!pmd_young(*pmd))
1827 entry = pte_mkold(entry);
Andrea Arcangeli1ba6e0b2012-10-04 01:51:06 +02001828 if (pmd_numa(*pmd))
1829 entry = pte_mknuma(entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001830 pte = pte_offset_map(&_pmd, haddr);
1831 BUG_ON(!pte_none(*pte));
1832 set_pte_at(mm, haddr, pte, entry);
1833 pte_unmap(pte);
1834 }
1835
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001836 smp_wmb(); /* make pte visible before pmd */
1837 /*
1838 * Up to this point the pmd is present and huge and
1839 * userland has the whole access to the hugepage
1840 * during the split (which happens in place). If we
1841 * overwrite the pmd with the not-huge version
1842 * pointing to the pte here (which of course we could
1843 * if all CPUs were bug free), userland could trigger
1844 * a small page size TLB miss on the small sized TLB
1845 * while the hugepage TLB entry is still established
1846 * in the huge TLB. Some CPU doesn't like that. See
1847 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1848 * Erratum 383 on page 93. Intel should be safe but is
1849 * also warns that it's only safe if the permission
1850 * and cache attributes of the two entries loaded in
1851 * the two TLB is identical (which should be the case
1852 * here). But it is generally safer to never allow
1853 * small and huge TLB entries for the same virtual
1854 * address to be loaded simultaneously. So instead of
1855 * doing "pmd_populate(); flush_tlb_range();" we first
1856 * mark the current pmd notpresent (atomically because
1857 * here the pmd_trans_huge and pmd_trans_splitting
1858 * must remain set at all times on the pmd until the
1859 * split is complete for this pmd), then we flush the
1860 * SMP TLB and finally we write the non-huge version
1861 * of the pmd entry with pmd_populate.
1862 */
Gerald Schaefer46dcde72012-10-08 16:30:09 -07001863 pmdp_invalidate(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001864 pmd_populate(mm, pmd, pgtable);
1865 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001866 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001867 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001868
1869 return ret;
1870}
1871
Ingo Molnar5a505082012-12-02 19:56:46 +00001872/* must be called with anon_vma->root->rwsem held */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001873static void __split_huge_page(struct page *page,
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001874 struct anon_vma *anon_vma,
1875 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001876{
1877 int mapcount, mapcount2;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001878 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001879 struct anon_vma_chain *avc;
1880
1881 BUG_ON(!PageHead(page));
1882 BUG_ON(PageTail(page));
1883
1884 mapcount = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001885 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001886 struct vm_area_struct *vma = avc->vma;
1887 unsigned long addr = vma_address(page, vma);
1888 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001889 mapcount += __split_huge_page_splitting(page, vma, addr);
1890 }
Andrea Arcangeli05759d32011-01-13 15:46:53 -08001891 /*
1892 * It is critical that new vmas are added to the tail of the
1893 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1894 * and establishes a child pmd before
1895 * __split_huge_page_splitting() freezes the parent pmd (so if
1896 * we fail to prevent copy_huge_pmd() from running until the
1897 * whole __split_huge_page() is complete), we will still see
1898 * the newly established pmd of the child later during the
1899 * walk, to be able to set it as pmd_trans_splitting too.
1900 */
1901 if (mapcount != page_mapcount(page))
1902 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1903 mapcount, page_mapcount(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001904 BUG_ON(mapcount != page_mapcount(page));
1905
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001906 __split_huge_page_refcount(page, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001907
1908 mapcount2 = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001909 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001910 struct vm_area_struct *vma = avc->vma;
1911 unsigned long addr = vma_address(page, vma);
1912 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001913 mapcount2 += __split_huge_page_map(page, vma, addr);
1914 }
Andrea Arcangeli05759d32011-01-13 15:46:53 -08001915 if (mapcount != mapcount2)
1916 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1917 mapcount, mapcount2, page_mapcount(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001918 BUG_ON(mapcount != mapcount2);
1919}
1920
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001921/*
1922 * Split a hugepage into normal pages. This doesn't change the position of head
1923 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1924 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1925 * from the hugepage.
1926 * Return 0 if the hugepage is split successfully otherwise return 1.
1927 */
1928int split_huge_page_to_list(struct page *page, struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001929{
1930 struct anon_vma *anon_vma;
1931 int ret = 1;
1932
Kirill A. Shutemov5918d102013-04-29 15:08:44 -07001933 BUG_ON(is_huge_zero_page(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001934 BUG_ON(!PageAnon(page));
Mel Gorman062f1af2013-01-11 14:32:02 -08001935
1936 /*
1937 * The caller does not necessarily hold an mmap_sem that would prevent
1938 * the anon_vma disappearing so we first we take a reference to it
1939 * and then lock the anon_vma for write. This is similar to
1940 * page_lock_anon_vma_read except the write lock is taken to serialise
1941 * against parallel split or collapse operations.
1942 */
1943 anon_vma = page_get_anon_vma(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001944 if (!anon_vma)
1945 goto out;
Mel Gorman062f1af2013-01-11 14:32:02 -08001946 anon_vma_lock_write(anon_vma);
1947
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001948 ret = 0;
1949 if (!PageCompound(page))
1950 goto out_unlock;
1951
1952 BUG_ON(!PageSwapBacked(page));
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001953 __split_huge_page(page, anon_vma, list);
Andi Kleen81ab4202011-04-14 15:22:06 -07001954 count_vm_event(THP_SPLIT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001955
1956 BUG_ON(PageCompound(page));
1957out_unlock:
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08001958 anon_vma_unlock_write(anon_vma);
Mel Gorman062f1af2013-01-11 14:32:02 -08001959 put_anon_vma(anon_vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001960out:
1961 return ret;
1962}
1963
Vlastimil Babka9050d7e2014-03-03 15:38:27 -08001964#define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001965
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001966int hugepage_madvise(struct vm_area_struct *vma,
1967 unsigned long *vm_flags, int advice)
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001968{
Gerald Schaefer8e720332012-10-08 16:30:12 -07001969 struct mm_struct *mm = vma->vm_mm;
1970
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001971 switch (advice) {
1972 case MADV_HUGEPAGE:
1973 /*
1974 * Be somewhat over-protective like KSM for now!
1975 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001976 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001977 return -EINVAL;
Gerald Schaefer8e720332012-10-08 16:30:12 -07001978 if (mm->def_flags & VM_NOHUGEPAGE)
1979 return -EINVAL;
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001980 *vm_flags &= ~VM_NOHUGEPAGE;
1981 *vm_flags |= VM_HUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001982 /*
1983 * If the vma become good for khugepaged to scan,
1984 * register it here without waiting a page fault that
1985 * may not happen any time soon.
1986 */
1987 if (unlikely(khugepaged_enter_vma_merge(vma)))
1988 return -ENOMEM;
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001989 break;
1990 case MADV_NOHUGEPAGE:
1991 /*
1992 * Be somewhat over-protective like KSM for now!
1993 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001994 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001995 return -EINVAL;
1996 *vm_flags &= ~VM_HUGEPAGE;
1997 *vm_flags |= VM_NOHUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001998 /*
1999 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2000 * this vma even if we leave the mm registered in khugepaged if
2001 * it got registered before VM_NOHUGEPAGE was set.
2002 */
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08002003 break;
2004 }
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08002005
2006 return 0;
2007}
2008
Andrea Arcangeliba761492011-01-13 15:46:58 -08002009static int __init khugepaged_slab_init(void)
2010{
2011 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2012 sizeof(struct mm_slot),
2013 __alignof__(struct mm_slot), 0, NULL);
2014 if (!mm_slot_cache)
2015 return -ENOMEM;
2016
2017 return 0;
2018}
2019
Andrea Arcangeliba761492011-01-13 15:46:58 -08002020static inline struct mm_slot *alloc_mm_slot(void)
2021{
2022 if (!mm_slot_cache) /* initialization failed */
2023 return NULL;
2024 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2025}
2026
2027static inline void free_mm_slot(struct mm_slot *mm_slot)
2028{
2029 kmem_cache_free(mm_slot_cache, mm_slot);
2030}
2031
Andrea Arcangeliba761492011-01-13 15:46:58 -08002032static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2033{
2034 struct mm_slot *mm_slot;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002035
Sasha Levinb67bfe02013-02-27 17:06:00 -08002036 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002037 if (mm == mm_slot->mm)
2038 return mm_slot;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002039
Andrea Arcangeliba761492011-01-13 15:46:58 -08002040 return NULL;
2041}
2042
2043static void insert_to_mm_slots_hash(struct mm_struct *mm,
2044 struct mm_slot *mm_slot)
2045{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002046 mm_slot->mm = mm;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002047 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002048}
2049
2050static inline int khugepaged_test_exit(struct mm_struct *mm)
2051{
2052 return atomic_read(&mm->mm_users) == 0;
2053}
2054
2055int __khugepaged_enter(struct mm_struct *mm)
2056{
2057 struct mm_slot *mm_slot;
2058 int wakeup;
2059
2060 mm_slot = alloc_mm_slot();
2061 if (!mm_slot)
2062 return -ENOMEM;
2063
2064 /* __khugepaged_exit() must not run from under us */
2065 VM_BUG_ON(khugepaged_test_exit(mm));
2066 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2067 free_mm_slot(mm_slot);
2068 return 0;
2069 }
2070
2071 spin_lock(&khugepaged_mm_lock);
2072 insert_to_mm_slots_hash(mm, mm_slot);
2073 /*
2074 * Insert just behind the scanning cursor, to let the area settle
2075 * down a little.
2076 */
2077 wakeup = list_empty(&khugepaged_scan.mm_head);
2078 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2079 spin_unlock(&khugepaged_mm_lock);
2080
2081 atomic_inc(&mm->mm_count);
2082 if (wakeup)
2083 wake_up_interruptible(&khugepaged_wait);
2084
2085 return 0;
2086}
2087
2088int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2089{
2090 unsigned long hstart, hend;
2091 if (!vma->anon_vma)
2092 /*
2093 * Not yet faulted in so we will register later in the
2094 * page fault if needed.
2095 */
2096 return 0;
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07002097 if (vma->vm_ops)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002098 /* khugepaged not yet working on file or special mappings */
2099 return 0;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002100 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002101 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2102 hend = vma->vm_end & HPAGE_PMD_MASK;
2103 if (hstart < hend)
2104 return khugepaged_enter(vma);
2105 return 0;
2106}
2107
2108void __khugepaged_exit(struct mm_struct *mm)
2109{
2110 struct mm_slot *mm_slot;
2111 int free = 0;
2112
2113 spin_lock(&khugepaged_mm_lock);
2114 mm_slot = get_mm_slot(mm);
2115 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002116 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002117 list_del(&mm_slot->mm_node);
2118 free = 1;
2119 }
Chris Wrightd788e802011-07-25 17:12:14 -07002120 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002121
2122 if (free) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002123 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2124 free_mm_slot(mm_slot);
2125 mmdrop(mm);
2126 } else if (mm_slot) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002127 /*
2128 * This is required to serialize against
2129 * khugepaged_test_exit() (which is guaranteed to run
2130 * under mmap sem read mode). Stop here (after we
2131 * return all pagetables will be destroyed) until
2132 * khugepaged has finished working on the pagetables
2133 * under the mmap_sem.
2134 */
2135 down_write(&mm->mmap_sem);
2136 up_write(&mm->mmap_sem);
Chris Wrightd788e802011-07-25 17:12:14 -07002137 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002138}
2139
2140static void release_pte_page(struct page *page)
2141{
2142 /* 0 stands for page_is_file_cache(page) == false */
2143 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2144 unlock_page(page);
2145 putback_lru_page(page);
2146}
2147
2148static void release_pte_pages(pte_t *pte, pte_t *_pte)
2149{
2150 while (--_pte >= pte) {
2151 pte_t pteval = *_pte;
2152 if (!pte_none(pteval))
2153 release_pte_page(pte_page(pteval));
2154 }
2155}
2156
Andrea Arcangeliba761492011-01-13 15:46:58 -08002157static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2158 unsigned long address,
2159 pte_t *pte)
2160{
2161 struct page *page;
2162 pte_t *_pte;
Bob Liu344aa352012-12-11 16:00:34 -08002163 int referenced = 0, none = 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002164 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2165 _pte++, address += PAGE_SIZE) {
2166 pte_t pteval = *_pte;
2167 if (pte_none(pteval)) {
2168 if (++none <= khugepaged_max_ptes_none)
2169 continue;
Bob Liu344aa352012-12-11 16:00:34 -08002170 else
Andrea Arcangeliba761492011-01-13 15:46:58 -08002171 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002172 }
Bob Liu344aa352012-12-11 16:00:34 -08002173 if (!pte_present(pteval) || !pte_write(pteval))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002174 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002175 page = vm_normal_page(vma, address, pteval);
Bob Liu344aa352012-12-11 16:00:34 -08002176 if (unlikely(!page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002177 goto out;
Bob Liu344aa352012-12-11 16:00:34 -08002178
Sasha Levin309381fea2014-01-23 15:52:54 -08002179 VM_BUG_ON_PAGE(PageCompound(page), page);
2180 VM_BUG_ON_PAGE(!PageAnon(page), page);
2181 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002182
2183 /* cannot use mapcount: can't collapse if there's a gup pin */
Bob Liu344aa352012-12-11 16:00:34 -08002184 if (page_count(page) != 1)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002185 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002186 /*
2187 * We can do it before isolate_lru_page because the
2188 * page can't be freed from under us. NOTE: PG_lock
2189 * is needed to serialize against split_huge_page
2190 * when invoked from the VM.
2191 */
Bob Liu344aa352012-12-11 16:00:34 -08002192 if (!trylock_page(page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002193 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002194 /*
2195 * Isolate the page to avoid collapsing an hugepage
2196 * currently in use by the VM.
2197 */
2198 if (isolate_lru_page(page)) {
2199 unlock_page(page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002200 goto out;
2201 }
2202 /* 0 stands for page_is_file_cache(page) == false */
2203 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
Sasha Levin309381fea2014-01-23 15:52:54 -08002204 VM_BUG_ON_PAGE(!PageLocked(page), page);
2205 VM_BUG_ON_PAGE(PageLRU(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002206
2207 /* If there is no mapped pte young don't collapse the page */
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002208 if (pte_young(pteval) || PageReferenced(page) ||
2209 mmu_notifier_test_young(vma->vm_mm, address))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002210 referenced = 1;
2211 }
Bob Liu344aa352012-12-11 16:00:34 -08002212 if (likely(referenced))
2213 return 1;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002214out:
Bob Liu344aa352012-12-11 16:00:34 -08002215 release_pte_pages(pte, _pte);
2216 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002217}
2218
2219static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2220 struct vm_area_struct *vma,
2221 unsigned long address,
2222 spinlock_t *ptl)
2223{
2224 pte_t *_pte;
2225 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2226 pte_t pteval = *_pte;
2227 struct page *src_page;
2228
2229 if (pte_none(pteval)) {
2230 clear_user_highpage(page, address);
2231 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2232 } else {
2233 src_page = pte_page(pteval);
2234 copy_user_highpage(page, src_page, address, vma);
Sasha Levin309381fea2014-01-23 15:52:54 -08002235 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002236 release_pte_page(src_page);
2237 /*
2238 * ptl mostly unnecessary, but preempt has to
2239 * be disabled to update the per-cpu stats
2240 * inside page_remove_rmap().
2241 */
2242 spin_lock(ptl);
2243 /*
2244 * paravirt calls inside pte_clear here are
2245 * superfluous.
2246 */
2247 pte_clear(vma->vm_mm, address, _pte);
2248 page_remove_rmap(src_page);
2249 spin_unlock(ptl);
2250 free_page_and_swap_cache(src_page);
2251 }
2252
2253 address += PAGE_SIZE;
2254 page++;
2255 }
2256}
2257
Xiao Guangrong26234f32012-10-08 16:29:51 -07002258static void khugepaged_alloc_sleep(void)
2259{
2260 wait_event_freezable_timeout(khugepaged_wait, false,
2261 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2262}
2263
Bob Liu9f1b8682013-11-12 15:07:37 -08002264static int khugepaged_node_load[MAX_NUMNODES];
2265
Xiao Guangrong26234f32012-10-08 16:29:51 -07002266#ifdef CONFIG_NUMA
Bob Liu9f1b8682013-11-12 15:07:37 -08002267static int khugepaged_find_target_node(void)
2268{
2269 static int last_khugepaged_target_node = NUMA_NO_NODE;
2270 int nid, target_node = 0, max_value = 0;
2271
2272 /* find first node with max normal pages hit */
2273 for (nid = 0; nid < MAX_NUMNODES; nid++)
2274 if (khugepaged_node_load[nid] > max_value) {
2275 max_value = khugepaged_node_load[nid];
2276 target_node = nid;
2277 }
2278
2279 /* do some balance if several nodes have the same hit record */
2280 if (target_node <= last_khugepaged_target_node)
2281 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2282 nid++)
2283 if (max_value == khugepaged_node_load[nid]) {
2284 target_node = nid;
2285 break;
2286 }
2287
2288 last_khugepaged_target_node = target_node;
2289 return target_node;
2290}
2291
Xiao Guangrong26234f32012-10-08 16:29:51 -07002292static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2293{
2294 if (IS_ERR(*hpage)) {
2295 if (!*wait)
2296 return false;
2297
2298 *wait = false;
Xiao Guangronge3b41262012-10-08 16:32:57 -07002299 *hpage = NULL;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002300 khugepaged_alloc_sleep();
2301 } else if (*hpage) {
2302 put_page(*hpage);
2303 *hpage = NULL;
2304 }
2305
2306 return true;
2307}
2308
2309static struct page
2310*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2311 struct vm_area_struct *vma, unsigned long address,
2312 int node)
2313{
Sasha Levin309381fea2014-01-23 15:52:54 -08002314 VM_BUG_ON_PAGE(*hpage, *hpage);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002315 /*
2316 * Allocate the page while the vma is still valid and under
2317 * the mmap_sem read mode so there is no memory allocation
2318 * later when we take the mmap_sem in write mode. This is more
2319 * friendly behavior (OTOH it may actually hide bugs) to
2320 * filesystems in userland with daemons allocating memory in
2321 * the userland I/O paths. Allocating memory with the
2322 * mmap_sem in read mode is good idea also to allow greater
2323 * scalability.
2324 */
Bob Liu9f1b8682013-11-12 15:07:37 -08002325 *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2326 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002327 /*
2328 * After allocating the hugepage, release the mmap_sem read lock in
2329 * preparation for taking it in write mode.
2330 */
2331 up_read(&mm->mmap_sem);
2332 if (unlikely(!*hpage)) {
2333 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2334 *hpage = ERR_PTR(-ENOMEM);
2335 return NULL;
2336 }
2337
2338 count_vm_event(THP_COLLAPSE_ALLOC);
2339 return *hpage;
2340}
2341#else
Bob Liu9f1b8682013-11-12 15:07:37 -08002342static int khugepaged_find_target_node(void)
2343{
2344 return 0;
2345}
2346
Bob Liu10dc4152013-11-12 15:07:35 -08002347static inline struct page *alloc_hugepage(int defrag)
2348{
2349 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2350 HPAGE_PMD_ORDER);
2351}
2352
Xiao Guangrong26234f32012-10-08 16:29:51 -07002353static struct page *khugepaged_alloc_hugepage(bool *wait)
2354{
2355 struct page *hpage;
2356
2357 do {
2358 hpage = alloc_hugepage(khugepaged_defrag());
2359 if (!hpage) {
2360 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2361 if (!*wait)
2362 return NULL;
2363
2364 *wait = false;
2365 khugepaged_alloc_sleep();
2366 } else
2367 count_vm_event(THP_COLLAPSE_ALLOC);
2368 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2369
2370 return hpage;
2371}
2372
2373static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2374{
2375 if (!*hpage)
2376 *hpage = khugepaged_alloc_hugepage(wait);
2377
2378 if (unlikely(!*hpage))
2379 return false;
2380
2381 return true;
2382}
2383
2384static struct page
2385*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2386 struct vm_area_struct *vma, unsigned long address,
2387 int node)
2388{
2389 up_read(&mm->mmap_sem);
2390 VM_BUG_ON(!*hpage);
2391 return *hpage;
2392}
2393#endif
2394
Bob Liufa475e52012-12-11 16:00:39 -08002395static bool hugepage_vma_check(struct vm_area_struct *vma)
2396{
2397 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2398 (vma->vm_flags & VM_NOHUGEPAGE))
2399 return false;
2400
2401 if (!vma->anon_vma || vma->vm_ops)
2402 return false;
2403 if (is_vma_temporary_stack(vma))
2404 return false;
2405 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2406 return true;
2407}
2408
Andrea Arcangeliba761492011-01-13 15:46:58 -08002409static void collapse_huge_page(struct mm_struct *mm,
Xiao Guangrong26234f32012-10-08 16:29:51 -07002410 unsigned long address,
2411 struct page **hpage,
2412 struct vm_area_struct *vma,
2413 int node)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002414{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002415 pmd_t *pmd, _pmd;
2416 pte_t *pte;
2417 pgtable_t pgtable;
2418 struct page *new_page;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002419 spinlock_t *pmd_ptl, *pte_ptl;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002420 int isolated;
2421 unsigned long hstart, hend;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002422 unsigned long mmun_start; /* For mmu_notifiers */
2423 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002424
2425 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002426
Xiao Guangrong26234f32012-10-08 16:29:51 -07002427 /* release the mmap_sem read lock. */
2428 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2429 if (!new_page)
Andrea Arcangelice83d212011-01-13 15:47:06 -08002430 return;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002431
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002432 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002433 return;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002434
2435 /*
2436 * Prevent all access to pagetables with the exception of
2437 * gup_fast later hanlded by the ptep_clear_flush and the VM
2438 * handled by the anon_vma lock + PG_lock.
2439 */
2440 down_write(&mm->mmap_sem);
2441 if (unlikely(khugepaged_test_exit(mm)))
2442 goto out;
2443
2444 vma = find_vma(mm, address);
Libina8f531eb2013-09-11 14:20:38 -07002445 if (!vma)
2446 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002447 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2448 hend = vma->vm_end & HPAGE_PMD_MASK;
2449 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2450 goto out;
Bob Liufa475e52012-12-11 16:00:39 -08002451 if (!hugepage_vma_check(vma))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002452 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002453 pmd = mm_find_pmd(mm, address);
2454 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002455 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002456 if (pmd_trans_huge(*pmd))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002457 goto out;
2458
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +00002459 anon_vma_lock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002460
2461 pte = pte_offset_map(pmd, address);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002462 pte_ptl = pte_lockptr(mm, pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002463
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002464 mmun_start = address;
2465 mmun_end = address + HPAGE_PMD_SIZE;
2466 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002467 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002468 /*
2469 * After this gup_fast can't run anymore. This also removes
2470 * any huge TLB entry from the CPU so we won't allow
2471 * huge and small TLB entries for the same virtual address
2472 * to avoid the risk of CPU bugs in that area.
2473 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002474 _pmd = pmdp_clear_flush(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002475 spin_unlock(pmd_ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002476 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002477
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002478 spin_lock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002479 isolated = __collapse_huge_page_isolate(vma, address, pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002480 spin_unlock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002481
2482 if (unlikely(!isolated)) {
Johannes Weiner453c7192011-01-20 14:44:18 -08002483 pte_unmap(pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002484 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002485 BUG_ON(!pmd_none(*pmd));
Aneesh Kumar K.V7c342512013-05-24 15:55:21 -07002486 /*
2487 * We can only use set_pmd_at when establishing
2488 * hugepmds and never for establishing regular pmds that
2489 * points to regular pagetables. Use pmd_populate for that
2490 */
2491 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002492 spin_unlock(pmd_ptl);
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002493 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002494 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002495 }
2496
2497 /*
2498 * All pages are isolated and locked so anon_vma rmap
2499 * can't run anymore.
2500 */
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002501 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002502
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002503 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
Johannes Weiner453c7192011-01-20 14:44:18 -08002504 pte_unmap(pte);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002505 __SetPageUptodate(new_page);
2506 pgtable = pmd_pgtable(_pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002507
Kirill A. Shutemov31223592013-09-12 15:14:01 -07002508 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2509 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002510
2511 /*
2512 * spin_lock() below is not the equivalent of smp_wmb(), so
2513 * this is needed to avoid the copy_huge_page writes to become
2514 * visible after the set_pmd_at() write.
2515 */
2516 smp_wmb();
2517
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002518 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002519 BUG_ON(!pmd_none(*pmd));
2520 page_add_new_anon_rmap(new_page, vma, address);
Aneesh Kumar K.Vfce144b2013-06-05 17:14:06 -07002521 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002522 set_pmd_at(mm, address, pmd, _pmd);
David Millerb113da62012-10-08 16:34:25 -07002523 update_mmu_cache_pmd(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002524 spin_unlock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002525
2526 *hpage = NULL;
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002527
Andrea Arcangeliba761492011-01-13 15:46:58 -08002528 khugepaged_pages_collapsed++;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002529out_up_write:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002530 up_write(&mm->mmap_sem);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002531 return;
2532
Andrea Arcangelice83d212011-01-13 15:47:06 -08002533out:
KAMEZAWA Hiroyuki678ff892011-02-10 15:01:36 -08002534 mem_cgroup_uncharge_page(new_page);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002535 goto out_up_write;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002536}
2537
2538static int khugepaged_scan_pmd(struct mm_struct *mm,
2539 struct vm_area_struct *vma,
2540 unsigned long address,
2541 struct page **hpage)
2542{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002543 pmd_t *pmd;
2544 pte_t *pte, *_pte;
2545 int ret = 0, referenced = 0, none = 0;
2546 struct page *page;
2547 unsigned long _address;
2548 spinlock_t *ptl;
David Rientjes00ef2d22013-02-22 16:35:36 -08002549 int node = NUMA_NO_NODE;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002550
2551 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2552
Bob Liu62190492012-12-11 16:00:37 -08002553 pmd = mm_find_pmd(mm, address);
2554 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002555 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002556 if (pmd_trans_huge(*pmd))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002557 goto out;
2558
Bob Liu9f1b8682013-11-12 15:07:37 -08002559 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002560 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2561 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2562 _pte++, _address += PAGE_SIZE) {
2563 pte_t pteval = *_pte;
2564 if (pte_none(pteval)) {
2565 if (++none <= khugepaged_max_ptes_none)
2566 continue;
2567 else
2568 goto out_unmap;
2569 }
2570 if (!pte_present(pteval) || !pte_write(pteval))
2571 goto out_unmap;
2572 page = vm_normal_page(vma, _address, pteval);
2573 if (unlikely(!page))
2574 goto out_unmap;
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002575 /*
Bob Liu9f1b8682013-11-12 15:07:37 -08002576 * Record which node the original page is from and save this
2577 * information to khugepaged_node_load[].
2578 * Khupaged will allocate hugepage from the node has the max
2579 * hit record.
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002580 */
Bob Liu9f1b8682013-11-12 15:07:37 -08002581 node = page_to_nid(page);
2582 khugepaged_node_load[node]++;
Sasha Levin309381fea2014-01-23 15:52:54 -08002583 VM_BUG_ON_PAGE(PageCompound(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002584 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2585 goto out_unmap;
2586 /* cannot use mapcount: can't collapse if there's a gup pin */
2587 if (page_count(page) != 1)
2588 goto out_unmap;
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002589 if (pte_young(pteval) || PageReferenced(page) ||
2590 mmu_notifier_test_young(vma->vm_mm, address))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002591 referenced = 1;
2592 }
2593 if (referenced)
2594 ret = 1;
2595out_unmap:
2596 pte_unmap_unlock(pte, ptl);
Bob Liu9f1b8682013-11-12 15:07:37 -08002597 if (ret) {
2598 node = khugepaged_find_target_node();
Andrea Arcangelice83d212011-01-13 15:47:06 -08002599 /* collapse_huge_page will return with the mmap_sem released */
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002600 collapse_huge_page(mm, address, hpage, vma, node);
Bob Liu9f1b8682013-11-12 15:07:37 -08002601 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002602out:
2603 return ret;
2604}
2605
2606static void collect_mm_slot(struct mm_slot *mm_slot)
2607{
2608 struct mm_struct *mm = mm_slot->mm;
2609
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002610 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002611
2612 if (khugepaged_test_exit(mm)) {
2613 /* free mm_slot */
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002614 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002615 list_del(&mm_slot->mm_node);
2616
2617 /*
2618 * Not strictly needed because the mm exited already.
2619 *
2620 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2621 */
2622
2623 /* khugepaged_mm_lock actually not necessary for the below */
2624 free_mm_slot(mm_slot);
2625 mmdrop(mm);
2626 }
2627}
2628
2629static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2630 struct page **hpage)
H Hartley Sweeten2f1da642011-10-31 17:09:25 -07002631 __releases(&khugepaged_mm_lock)
2632 __acquires(&khugepaged_mm_lock)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002633{
2634 struct mm_slot *mm_slot;
2635 struct mm_struct *mm;
2636 struct vm_area_struct *vma;
2637 int progress = 0;
2638
2639 VM_BUG_ON(!pages);
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002640 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002641
2642 if (khugepaged_scan.mm_slot)
2643 mm_slot = khugepaged_scan.mm_slot;
2644 else {
2645 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2646 struct mm_slot, mm_node);
2647 khugepaged_scan.address = 0;
2648 khugepaged_scan.mm_slot = mm_slot;
2649 }
2650 spin_unlock(&khugepaged_mm_lock);
2651
2652 mm = mm_slot->mm;
2653 down_read(&mm->mmap_sem);
2654 if (unlikely(khugepaged_test_exit(mm)))
2655 vma = NULL;
2656 else
2657 vma = find_vma(mm, khugepaged_scan.address);
2658
2659 progress++;
2660 for (; vma; vma = vma->vm_next) {
2661 unsigned long hstart, hend;
2662
2663 cond_resched();
2664 if (unlikely(khugepaged_test_exit(mm))) {
2665 progress++;
2666 break;
2667 }
Bob Liufa475e52012-12-11 16:00:39 -08002668 if (!hugepage_vma_check(vma)) {
2669skip:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002670 progress++;
2671 continue;
2672 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002673 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2674 hend = vma->vm_end & HPAGE_PMD_MASK;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002675 if (hstart >= hend)
2676 goto skip;
2677 if (khugepaged_scan.address > hend)
2678 goto skip;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002679 if (khugepaged_scan.address < hstart)
2680 khugepaged_scan.address = hstart;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002681 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002682
2683 while (khugepaged_scan.address < hend) {
2684 int ret;
2685 cond_resched();
2686 if (unlikely(khugepaged_test_exit(mm)))
2687 goto breakouterloop;
2688
2689 VM_BUG_ON(khugepaged_scan.address < hstart ||
2690 khugepaged_scan.address + HPAGE_PMD_SIZE >
2691 hend);
2692 ret = khugepaged_scan_pmd(mm, vma,
2693 khugepaged_scan.address,
2694 hpage);
2695 /* move to next address */
2696 khugepaged_scan.address += HPAGE_PMD_SIZE;
2697 progress += HPAGE_PMD_NR;
2698 if (ret)
2699 /* we released mmap_sem so break loop */
2700 goto breakouterloop_mmap_sem;
2701 if (progress >= pages)
2702 goto breakouterloop;
2703 }
2704 }
2705breakouterloop:
2706 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2707breakouterloop_mmap_sem:
2708
2709 spin_lock(&khugepaged_mm_lock);
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002710 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002711 /*
2712 * Release the current mm_slot if this mm is about to die, or
2713 * if we scanned all vmas of this mm.
2714 */
2715 if (khugepaged_test_exit(mm) || !vma) {
2716 /*
2717 * Make sure that if mm_users is reaching zero while
2718 * khugepaged runs here, khugepaged_exit will find
2719 * mm_slot not pointing to the exiting mm.
2720 */
2721 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2722 khugepaged_scan.mm_slot = list_entry(
2723 mm_slot->mm_node.next,
2724 struct mm_slot, mm_node);
2725 khugepaged_scan.address = 0;
2726 } else {
2727 khugepaged_scan.mm_slot = NULL;
2728 khugepaged_full_scans++;
2729 }
2730
2731 collect_mm_slot(mm_slot);
2732 }
2733
2734 return progress;
2735}
2736
2737static int khugepaged_has_work(void)
2738{
2739 return !list_empty(&khugepaged_scan.mm_head) &&
2740 khugepaged_enabled();
2741}
2742
2743static int khugepaged_wait_event(void)
2744{
2745 return !list_empty(&khugepaged_scan.mm_head) ||
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002746 kthread_should_stop();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002747}
2748
Xiao Guangrongd5169042012-10-08 16:29:48 -07002749static void khugepaged_do_scan(void)
2750{
2751 struct page *hpage = NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002752 unsigned int progress = 0, pass_through_head = 0;
2753 unsigned int pages = khugepaged_pages_to_scan;
Xiao Guangrongd5169042012-10-08 16:29:48 -07002754 bool wait = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002755
2756 barrier(); /* write khugepaged_pages_to_scan to local stack */
2757
2758 while (progress < pages) {
Xiao Guangrong26234f32012-10-08 16:29:51 -07002759 if (!khugepaged_prealloc_page(&hpage, &wait))
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002760 break;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002761
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002762 cond_resched();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002763
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002764 if (unlikely(kthread_should_stop() || freezing(current)))
2765 break;
2766
Andrea Arcangeliba761492011-01-13 15:46:58 -08002767 spin_lock(&khugepaged_mm_lock);
2768 if (!khugepaged_scan.mm_slot)
2769 pass_through_head++;
2770 if (khugepaged_has_work() &&
2771 pass_through_head < 2)
2772 progress += khugepaged_scan_mm_slot(pages - progress,
Xiao Guangrongd5169042012-10-08 16:29:48 -07002773 &hpage);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002774 else
2775 progress = pages;
2776 spin_unlock(&khugepaged_mm_lock);
2777 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002778
Xiao Guangrongd5169042012-10-08 16:29:48 -07002779 if (!IS_ERR_OR_NULL(hpage))
2780 put_page(hpage);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002781}
2782
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002783static void khugepaged_wait_work(void)
2784{
2785 try_to_freeze();
2786
2787 if (khugepaged_has_work()) {
2788 if (!khugepaged_scan_sleep_millisecs)
2789 return;
2790
2791 wait_event_freezable_timeout(khugepaged_wait,
2792 kthread_should_stop(),
2793 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2794 return;
2795 }
2796
2797 if (khugepaged_enabled())
2798 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2799}
2800
Andrea Arcangeliba761492011-01-13 15:46:58 -08002801static int khugepaged(void *none)
2802{
2803 struct mm_slot *mm_slot;
2804
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002805 set_freezable();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002806 set_user_nice(current, 19);
2807
Xiao Guangrongb7231782012-10-08 16:29:54 -07002808 while (!kthread_should_stop()) {
2809 khugepaged_do_scan();
2810 khugepaged_wait_work();
2811 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002812
2813 spin_lock(&khugepaged_mm_lock);
2814 mm_slot = khugepaged_scan.mm_slot;
2815 khugepaged_scan.mm_slot = NULL;
2816 if (mm_slot)
2817 collect_mm_slot(mm_slot);
2818 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002819 return 0;
2820}
2821
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002822static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2823 unsigned long haddr, pmd_t *pmd)
2824{
2825 struct mm_struct *mm = vma->vm_mm;
2826 pgtable_t pgtable;
2827 pmd_t _pmd;
2828 int i;
2829
2830 pmdp_clear_flush(vma, haddr, pmd);
2831 /* leave pmd empty until pte is filled */
2832
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07002833 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002834 pmd_populate(mm, &_pmd, pgtable);
2835
2836 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2837 pte_t *pte, entry;
2838 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2839 entry = pte_mkspecial(entry);
2840 pte = pte_offset_map(&_pmd, haddr);
2841 VM_BUG_ON(!pte_none(*pte));
2842 set_pte_at(mm, haddr, pte, entry);
2843 pte_unmap(pte);
2844 }
2845 smp_wmb(); /* make pte visible before pmd */
2846 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08002847 put_huge_zero_page();
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002848}
2849
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002850void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2851 pmd_t *pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002852{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002853 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002854 struct page *page;
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002855 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002856 unsigned long haddr = address & HPAGE_PMD_MASK;
2857 unsigned long mmun_start; /* For mmu_notifiers */
2858 unsigned long mmun_end; /* For mmu_notifiers */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002859
2860 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002861
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002862 mmun_start = haddr;
2863 mmun_end = haddr + HPAGE_PMD_SIZE;
Hugh Dickins750e8162013-10-16 13:47:08 -07002864again:
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002865 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002866 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002867 if (unlikely(!pmd_trans_huge(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002868 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002869 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2870 return;
2871 }
2872 if (is_huge_zero_pmd(*pmd)) {
2873 __split_huge_zero_page_pmd(vma, haddr, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002874 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002875 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002876 return;
2877 }
2878 page = pmd_page(*pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08002879 VM_BUG_ON_PAGE(!page_count(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002880 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002881 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002882 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002883
2884 split_huge_page(page);
2885
2886 put_page(page);
Hugh Dickins750e8162013-10-16 13:47:08 -07002887
2888 /*
2889 * We don't always have down_write of mmap_sem here: a racing
2890 * do_huge_pmd_wp_page() might have copied-on-write to another
2891 * huge page before our split_huge_page() got the anon_vma lock.
2892 */
2893 if (unlikely(pmd_trans_huge(*pmd)))
2894 goto again;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002895}
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002896
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002897void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2898 pmd_t *pmd)
2899{
2900 struct vm_area_struct *vma;
2901
2902 vma = find_vma(mm, address);
2903 BUG_ON(vma == NULL);
2904 split_huge_page_pmd(vma, address, pmd);
2905}
2906
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002907static void split_huge_page_address(struct mm_struct *mm,
2908 unsigned long address)
2909{
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002910 pmd_t *pmd;
2911
2912 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2913
Bob Liu62190492012-12-11 16:00:37 -08002914 pmd = mm_find_pmd(mm, address);
2915 if (!pmd)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002916 return;
2917 /*
2918 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2919 * materialize from under us.
2920 */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002921 split_huge_page_pmd_mm(mm, address, pmd);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002922}
2923
2924void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2925 unsigned long start,
2926 unsigned long end,
2927 long adjust_next)
2928{
2929 /*
2930 * If the new start address isn't hpage aligned and it could
2931 * previously contain an hugepage: check if we need to split
2932 * an huge pmd.
2933 */
2934 if (start & ~HPAGE_PMD_MASK &&
2935 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2936 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2937 split_huge_page_address(vma->vm_mm, start);
2938
2939 /*
2940 * If the new end address isn't hpage aligned and it could
2941 * previously contain an hugepage: check if we need to split
2942 * an huge pmd.
2943 */
2944 if (end & ~HPAGE_PMD_MASK &&
2945 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2946 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2947 split_huge_page_address(vma->vm_mm, end);
2948
2949 /*
2950 * If we're also updating the vma->vm_next->vm_start, if the new
2951 * vm_next->vm_start isn't page aligned and it could previously
2952 * contain an hugepage: check if we need to split an huge pmd.
2953 */
2954 if (adjust_next > 0) {
2955 struct vm_area_struct *next = vma->vm_next;
2956 unsigned long nstart = next->vm_start;
2957 nstart += adjust_next << PAGE_SHIFT;
2958 if (nstart & ~HPAGE_PMD_MASK &&
2959 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2960 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2961 split_huge_page_address(next->vm_mm, nstart);
2962 }
2963}