blob: bccd5a628ea6765478d2fa45dc01390a83ee8a5b [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
133 if (recommended_min > min_free_kbytes)
134 min_free_kbytes = recommended_min;
135 setup_per_zone_wmarks();
136 return 0;
137}
138late_initcall(set_recommended_min_free_kbytes);
139
Andrea Arcangeliba761492011-01-13 15:46:58 -0800140static int start_khugepaged(void)
141{
142 int err = 0;
143 if (khugepaged_enabled()) {
Andrea Arcangeliba761492011-01-13 15:46:58 -0800144 if (!khugepaged_thread)
145 khugepaged_thread = kthread_run(khugepaged, NULL,
146 "khugepaged");
147 if (unlikely(IS_ERR(khugepaged_thread))) {
148 printk(KERN_ERR
149 "khugepaged: kthread_run(khugepaged) failed\n");
150 err = PTR_ERR(khugepaged_thread);
151 khugepaged_thread = NULL;
152 }
Xiao Guangrong911891a2012-10-08 16:29:41 -0700153
154 if (!list_empty(&khugepaged_scan.mm_head))
Andrea Arcangeliba761492011-01-13 15:46:58 -0800155 wake_up_interruptible(&khugepaged_wait);
Andrea Arcangelif0005652011-01-13 15:47:04 -0800156
157 set_recommended_min_free_kbytes();
Xiao Guangrong911891a2012-10-08 16:29:41 -0700158 } else if (khugepaged_thread) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700159 kthread_stop(khugepaged_thread);
160 khugepaged_thread = NULL;
161 }
Xiao Guangrong637e3a22012-10-08 16:29:38 -0700162
Andrea Arcangeliba761492011-01-13 15:46:58 -0800163 return err;
164}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800165
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800166static atomic_t huge_zero_refcount;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700167static struct page *huge_zero_page __read_mostly;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800168
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700169static inline bool is_huge_zero_page(struct page *page)
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800170{
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700171 return ACCESS_ONCE(huge_zero_page) == page;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800172}
173
174static inline bool is_huge_zero_pmd(pmd_t pmd)
175{
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700176 return is_huge_zero_page(pmd_page(pmd));
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800177}
178
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700179static struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800180{
181 struct page *zero_page;
182retry:
183 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700184 return ACCESS_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800185
186 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
187 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800188 if (!zero_page) {
189 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700190 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800191 }
192 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800193 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700194 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800195 preempt_enable();
196 __free_page(zero_page);
197 goto retry;
198 }
199
200 /* We take additional reference here. It will be put back by shrinker */
201 atomic_set(&huge_zero_refcount, 2);
202 preempt_enable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700203 return ACCESS_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800204}
205
206static void put_huge_zero_page(void)
207{
208 /*
209 * Counter should never go to zero here. Only shrinker can put
210 * last reference.
211 */
212 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
213}
214
Glauber Costa48896462013-08-28 10:18:15 +1000215static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
216 struct shrink_control *sc)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800217{
Glauber Costa48896462013-08-28 10:18:15 +1000218 /* we can free zero page only if last reference remains */
219 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
220}
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800221
Glauber Costa48896462013-08-28 10:18:15 +1000222static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
223 struct shrink_control *sc)
224{
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800225 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700226 struct page *zero_page = xchg(&huge_zero_page, NULL);
227 BUG_ON(zero_page == NULL);
228 __free_page(zero_page);
Glauber Costa48896462013-08-28 10:18:15 +1000229 return HPAGE_PMD_NR;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800230 }
231
232 return 0;
233}
234
235static struct shrinker huge_zero_page_shrinker = {
Glauber Costa48896462013-08-28 10:18:15 +1000236 .count_objects = shrink_huge_zero_page_count,
237 .scan_objects = shrink_huge_zero_page_scan,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800238 .seeks = DEFAULT_SEEKS,
239};
240
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800241#ifdef CONFIG_SYSFS
Andrea Arcangeliba761492011-01-13 15:46:58 -0800242
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800243static ssize_t double_flag_show(struct kobject *kobj,
244 struct kobj_attribute *attr, char *buf,
245 enum transparent_hugepage_flag enabled,
246 enum transparent_hugepage_flag req_madv)
247{
248 if (test_bit(enabled, &transparent_hugepage_flags)) {
249 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
250 return sprintf(buf, "[always] madvise never\n");
251 } else if (test_bit(req_madv, &transparent_hugepage_flags))
252 return sprintf(buf, "always [madvise] never\n");
253 else
254 return sprintf(buf, "always madvise [never]\n");
255}
256static ssize_t double_flag_store(struct kobject *kobj,
257 struct kobj_attribute *attr,
258 const char *buf, size_t count,
259 enum transparent_hugepage_flag enabled,
260 enum transparent_hugepage_flag req_madv)
261{
262 if (!memcmp("always", buf,
263 min(sizeof("always")-1, count))) {
264 set_bit(enabled, &transparent_hugepage_flags);
265 clear_bit(req_madv, &transparent_hugepage_flags);
266 } else if (!memcmp("madvise", buf,
267 min(sizeof("madvise")-1, count))) {
268 clear_bit(enabled, &transparent_hugepage_flags);
269 set_bit(req_madv, &transparent_hugepage_flags);
270 } else if (!memcmp("never", buf,
271 min(sizeof("never")-1, count))) {
272 clear_bit(enabled, &transparent_hugepage_flags);
273 clear_bit(req_madv, &transparent_hugepage_flags);
274 } else
275 return -EINVAL;
276
277 return count;
278}
279
280static ssize_t enabled_show(struct kobject *kobj,
281 struct kobj_attribute *attr, char *buf)
282{
283 return double_flag_show(kobj, attr, buf,
284 TRANSPARENT_HUGEPAGE_FLAG,
285 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
286}
287static ssize_t enabled_store(struct kobject *kobj,
288 struct kobj_attribute *attr,
289 const char *buf, size_t count)
290{
Andrea Arcangeliba761492011-01-13 15:46:58 -0800291 ssize_t ret;
292
293 ret = double_flag_store(kobj, attr, buf, count,
294 TRANSPARENT_HUGEPAGE_FLAG,
295 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
296
297 if (ret > 0) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700298 int err;
299
300 mutex_lock(&khugepaged_mutex);
301 err = start_khugepaged();
302 mutex_unlock(&khugepaged_mutex);
303
Andrea Arcangeliba761492011-01-13 15:46:58 -0800304 if (err)
305 ret = err;
306 }
307
308 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800309}
310static struct kobj_attribute enabled_attr =
311 __ATTR(enabled, 0644, enabled_show, enabled_store);
312
313static ssize_t single_flag_show(struct kobject *kobj,
314 struct kobj_attribute *attr, char *buf,
315 enum transparent_hugepage_flag flag)
316{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700317 return sprintf(buf, "%d\n",
318 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800319}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700320
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800321static ssize_t single_flag_store(struct kobject *kobj,
322 struct kobj_attribute *attr,
323 const char *buf, size_t count,
324 enum transparent_hugepage_flag flag)
325{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700326 unsigned long value;
327 int ret;
328
329 ret = kstrtoul(buf, 10, &value);
330 if (ret < 0)
331 return ret;
332 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800333 return -EINVAL;
334
Ben Hutchingse27e6152011-04-14 15:22:21 -0700335 if (value)
336 set_bit(flag, &transparent_hugepage_flags);
337 else
338 clear_bit(flag, &transparent_hugepage_flags);
339
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800340 return count;
341}
342
343/*
344 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
345 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
346 * memory just to allocate one more hugepage.
347 */
348static ssize_t defrag_show(struct kobject *kobj,
349 struct kobj_attribute *attr, char *buf)
350{
351 return double_flag_show(kobj, attr, buf,
352 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
353 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
354}
355static ssize_t defrag_store(struct kobject *kobj,
356 struct kobj_attribute *attr,
357 const char *buf, size_t count)
358{
359 return double_flag_store(kobj, attr, buf, count,
360 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
361 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
362}
363static struct kobj_attribute defrag_attr =
364 __ATTR(defrag, 0644, defrag_show, defrag_store);
365
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800366static ssize_t use_zero_page_show(struct kobject *kobj,
367 struct kobj_attribute *attr, char *buf)
368{
369 return single_flag_show(kobj, attr, buf,
370 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
371}
372static ssize_t use_zero_page_store(struct kobject *kobj,
373 struct kobj_attribute *attr, const char *buf, size_t count)
374{
375 return single_flag_store(kobj, attr, buf, count,
376 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
377}
378static struct kobj_attribute use_zero_page_attr =
379 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800380#ifdef CONFIG_DEBUG_VM
381static ssize_t debug_cow_show(struct kobject *kobj,
382 struct kobj_attribute *attr, char *buf)
383{
384 return single_flag_show(kobj, attr, buf,
385 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
386}
387static ssize_t debug_cow_store(struct kobject *kobj,
388 struct kobj_attribute *attr,
389 const char *buf, size_t count)
390{
391 return single_flag_store(kobj, attr, buf, count,
392 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
393}
394static struct kobj_attribute debug_cow_attr =
395 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
396#endif /* CONFIG_DEBUG_VM */
397
398static struct attribute *hugepage_attr[] = {
399 &enabled_attr.attr,
400 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800401 &use_zero_page_attr.attr,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800402#ifdef CONFIG_DEBUG_VM
403 &debug_cow_attr.attr,
404#endif
405 NULL,
406};
407
408static struct attribute_group hugepage_attr_group = {
409 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800410};
411
412static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
413 struct kobj_attribute *attr,
414 char *buf)
415{
416 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
417}
418
419static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
420 struct kobj_attribute *attr,
421 const char *buf, size_t count)
422{
423 unsigned long msecs;
424 int err;
425
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700426 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800427 if (err || msecs > UINT_MAX)
428 return -EINVAL;
429
430 khugepaged_scan_sleep_millisecs = msecs;
431 wake_up_interruptible(&khugepaged_wait);
432
433 return count;
434}
435static struct kobj_attribute scan_sleep_millisecs_attr =
436 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
437 scan_sleep_millisecs_store);
438
439static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
440 struct kobj_attribute *attr,
441 char *buf)
442{
443 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
444}
445
446static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
447 struct kobj_attribute *attr,
448 const char *buf, size_t count)
449{
450 unsigned long msecs;
451 int err;
452
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700453 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800454 if (err || msecs > UINT_MAX)
455 return -EINVAL;
456
457 khugepaged_alloc_sleep_millisecs = msecs;
458 wake_up_interruptible(&khugepaged_wait);
459
460 return count;
461}
462static struct kobj_attribute alloc_sleep_millisecs_attr =
463 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
464 alloc_sleep_millisecs_store);
465
466static ssize_t pages_to_scan_show(struct kobject *kobj,
467 struct kobj_attribute *attr,
468 char *buf)
469{
470 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
471}
472static ssize_t pages_to_scan_store(struct kobject *kobj,
473 struct kobj_attribute *attr,
474 const char *buf, size_t count)
475{
476 int err;
477 unsigned long pages;
478
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700479 err = kstrtoul(buf, 10, &pages);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800480 if (err || !pages || pages > UINT_MAX)
481 return -EINVAL;
482
483 khugepaged_pages_to_scan = pages;
484
485 return count;
486}
487static struct kobj_attribute pages_to_scan_attr =
488 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
489 pages_to_scan_store);
490
491static ssize_t pages_collapsed_show(struct kobject *kobj,
492 struct kobj_attribute *attr,
493 char *buf)
494{
495 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
496}
497static struct kobj_attribute pages_collapsed_attr =
498 __ATTR_RO(pages_collapsed);
499
500static ssize_t full_scans_show(struct kobject *kobj,
501 struct kobj_attribute *attr,
502 char *buf)
503{
504 return sprintf(buf, "%u\n", khugepaged_full_scans);
505}
506static struct kobj_attribute full_scans_attr =
507 __ATTR_RO(full_scans);
508
509static ssize_t khugepaged_defrag_show(struct kobject *kobj,
510 struct kobj_attribute *attr, char *buf)
511{
512 return single_flag_show(kobj, attr, buf,
513 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
514}
515static ssize_t khugepaged_defrag_store(struct kobject *kobj,
516 struct kobj_attribute *attr,
517 const char *buf, size_t count)
518{
519 return single_flag_store(kobj, attr, buf, count,
520 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
521}
522static struct kobj_attribute khugepaged_defrag_attr =
523 __ATTR(defrag, 0644, khugepaged_defrag_show,
524 khugepaged_defrag_store);
525
526/*
527 * max_ptes_none controls if khugepaged should collapse hugepages over
528 * any unmapped ptes in turn potentially increasing the memory
529 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
530 * reduce the available free memory in the system as it
531 * runs. Increasing max_ptes_none will instead potentially reduce the
532 * free memory in the system during the khugepaged scan.
533 */
534static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
535 struct kobj_attribute *attr,
536 char *buf)
537{
538 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
539}
540static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
541 struct kobj_attribute *attr,
542 const char *buf, size_t count)
543{
544 int err;
545 unsigned long max_ptes_none;
546
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700547 err = kstrtoul(buf, 10, &max_ptes_none);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800548 if (err || max_ptes_none > HPAGE_PMD_NR-1)
549 return -EINVAL;
550
551 khugepaged_max_ptes_none = max_ptes_none;
552
553 return count;
554}
555static struct kobj_attribute khugepaged_max_ptes_none_attr =
556 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
557 khugepaged_max_ptes_none_store);
558
559static struct attribute *khugepaged_attr[] = {
560 &khugepaged_defrag_attr.attr,
561 &khugepaged_max_ptes_none_attr.attr,
562 &pages_to_scan_attr.attr,
563 &pages_collapsed_attr.attr,
564 &full_scans_attr.attr,
565 &scan_sleep_millisecs_attr.attr,
566 &alloc_sleep_millisecs_attr.attr,
567 NULL,
568};
569
570static struct attribute_group khugepaged_attr_group = {
571 .attrs = khugepaged_attr,
572 .name = "khugepaged",
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800573};
Shaohua Li569e5592012-01-12 17:19:11 -0800574
575static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
576{
577 int err;
578
579 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
580 if (unlikely(!*hugepage_kobj)) {
Jeremy Eder2c797372012-12-20 15:05:32 -0800581 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800582 return -ENOMEM;
583 }
584
585 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
586 if (err) {
Jeremy Eder2c797372012-12-20 15:05:32 -0800587 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800588 goto delete_obj;
589 }
590
591 err = sysfs_create_group(*hugepage_kobj, &khugepaged_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 remove_hp_group;
595 }
596
597 return 0;
598
599remove_hp_group:
600 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
601delete_obj:
602 kobject_put(*hugepage_kobj);
603 return err;
604}
605
606static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
607{
608 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
609 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
610 kobject_put(hugepage_kobj);
611}
612#else
613static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
614{
615 return 0;
616}
617
618static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
619{
620}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800621#endif /* CONFIG_SYSFS */
622
623static int __init hugepage_init(void)
624{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800625 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800626 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800627
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800628 if (!has_transparent_hugepage()) {
629 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800630 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800631 }
632
Shaohua Li569e5592012-01-12 17:19:11 -0800633 err = hugepage_init_sysfs(&hugepage_kobj);
634 if (err)
635 return err;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800636
637 err = khugepaged_slab_init();
638 if (err)
639 goto out;
640
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800641 register_shrinker(&huge_zero_page_shrinker);
642
Rik van Riel97562cd2011-01-13 15:47:12 -0800643 /*
644 * By default disable transparent hugepages on smaller systems,
645 * where the extra memory used could hurt more than TLB overhead
646 * is likely to save. The admin can still enable it through /sys.
647 */
648 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
649 transparent_hugepage_flags = 0;
650
Andrea Arcangeliba761492011-01-13 15:46:58 -0800651 start_khugepaged();
652
Shaohua Li569e5592012-01-12 17:19:11 -0800653 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800654out:
Shaohua Li569e5592012-01-12 17:19:11 -0800655 hugepage_exit_sysfs(hugepage_kobj);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800656 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800657}
658module_init(hugepage_init)
659
660static int __init setup_transparent_hugepage(char *str)
661{
662 int ret = 0;
663 if (!str)
664 goto out;
665 if (!strcmp(str, "always")) {
666 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
667 &transparent_hugepage_flags);
668 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
669 &transparent_hugepage_flags);
670 ret = 1;
671 } else if (!strcmp(str, "madvise")) {
672 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
673 &transparent_hugepage_flags);
674 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
675 &transparent_hugepage_flags);
676 ret = 1;
677 } else if (!strcmp(str, "never")) {
678 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
679 &transparent_hugepage_flags);
680 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
681 &transparent_hugepage_flags);
682 ret = 1;
683 }
684out:
685 if (!ret)
686 printk(KERN_WARNING
687 "transparent_hugepage= cannot parse, ignored\n");
688 return ret;
689}
690__setup("transparent_hugepage=", setup_transparent_hugepage);
691
Mel Gormanb32967f2012-11-19 12:35:47 +0000692pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800693{
694 if (likely(vma->vm_flags & VM_WRITE))
695 pmd = pmd_mkwrite(pmd);
696 return pmd;
697}
698
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700699static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
Bob Liub3092b32012-12-11 16:00:41 -0800700{
701 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700702 entry = mk_pmd(page, prot);
Bob Liub3092b32012-12-11 16:00:41 -0800703 entry = pmd_mkhuge(entry);
704 return entry;
705}
706
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800707static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
708 struct vm_area_struct *vma,
709 unsigned long haddr, pmd_t *pmd,
710 struct page *page)
711{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800712 pgtable_t pgtable;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800713 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800714
715 VM_BUG_ON(!PageCompound(page));
716 pgtable = pte_alloc_one(mm, haddr);
David Rientjesedad9d22012-05-29 15:06:17 -0700717 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800718 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800719
720 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700721 /*
722 * The memory barrier inside __SetPageUptodate makes sure that
723 * clear_huge_page writes become visible before the set_pmd_at()
724 * write.
725 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800726 __SetPageUptodate(page);
727
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800728 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800729 if (unlikely(!pmd_none(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800730 spin_unlock(ptl);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800731 mem_cgroup_uncharge_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800732 put_page(page);
733 pte_free(mm, pgtable);
734 } else {
735 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700736 entry = mk_huge_pmd(page, vma->vm_page_prot);
737 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800738 page_add_new_anon_rmap(page, vma, haddr);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700739 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800740 set_pmd_at(mm, haddr, pmd, entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800741 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800742 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800743 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800744 }
745
David Rientjesaa2e8782012-05-29 15:06:17 -0700746 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800747}
748
Andi Kleencc5d4622011-03-22 16:33:13 -0700749static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800750{
Andi Kleencc5d4622011-03-22 16:33:13 -0700751 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800752}
753
754static inline struct page *alloc_hugepage_vma(int defrag,
755 struct vm_area_struct *vma,
Andi Kleencc5d4622011-03-22 16:33:13 -0700756 unsigned long haddr, int nd,
757 gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800758{
Andi Kleencc5d4622011-03-22 16:33:13 -0700759 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
Andi Kleen5c4b4be2011-03-04 17:36:32 -0800760 HPAGE_PMD_ORDER, vma, haddr, nd);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800761}
762
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800763/* Caller must hold page table lock. */
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800764static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800765 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700766 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800767{
768 pmd_t entry;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800769 if (!pmd_none(*pmd))
770 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700771 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800772 entry = pmd_wrprotect(entry);
773 entry = pmd_mkhuge(entry);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700774 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800775 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800776 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800777 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800778}
779
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800780int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
781 unsigned long address, pmd_t *pmd,
782 unsigned int flags)
783{
784 struct page *page;
785 unsigned long haddr = address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800786
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700787 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700788 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700789 if (unlikely(anon_vma_prepare(vma)))
790 return VM_FAULT_OOM;
791 if (unlikely(khugepaged_enter(vma)))
792 return VM_FAULT_OOM;
793 if (!(flags & FAULT_FLAG_WRITE) &&
794 transparent_hugepage_use_zero_page()) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800795 spinlock_t *ptl;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700796 pgtable_t pgtable;
797 struct page *zero_page;
798 bool set;
799 pgtable = pte_alloc_one(mm, haddr);
800 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800801 return VM_FAULT_OOM;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700802 zero_page = get_huge_zero_page();
803 if (unlikely(!zero_page)) {
804 pte_free(mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700805 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700806 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700807 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800808 ptl = pmd_lock(mm, pmd);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700809 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
810 zero_page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800811 spin_unlock(ptl);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700812 if (!set) {
813 pte_free(mm, pgtable);
814 put_huge_zero_page();
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800815 }
David Rientjesedad9d22012-05-29 15:06:17 -0700816 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800817 }
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700818 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
819 vma, haddr, numa_node_id(), 0);
820 if (unlikely(!page)) {
821 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700822 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700823 }
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700824 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
825 put_page(page);
David Rientjes17766dd2013-09-12 15:14:06 -0700826 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700827 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700828 }
829 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
830 mem_cgroup_uncharge_page(page);
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
David Rientjes17766dd2013-09-12 15:14:06 -0700836 count_vm_event(THP_FAULT_ALLOC);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700837 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800838}
839
840int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
841 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
842 struct vm_area_struct *vma)
843{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800844 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800845 struct page *src_page;
846 pmd_t pmd;
847 pgtable_t pgtable;
848 int ret;
849
850 ret = -ENOMEM;
851 pgtable = pte_alloc_one(dst_mm, addr);
852 if (unlikely(!pgtable))
853 goto out;
854
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800855 dst_ptl = pmd_lock(dst_mm, dst_pmd);
856 src_ptl = pmd_lockptr(src_mm, src_pmd);
857 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800858
859 ret = -EAGAIN;
860 pmd = *src_pmd;
861 if (unlikely(!pmd_trans_huge(pmd))) {
862 pte_free(dst_mm, pgtable);
863 goto out_unlock;
864 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800865 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800866 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800867 * under splitting since we don't split the page itself, only pmd to
868 * a page table.
869 */
870 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700871 struct page *zero_page;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800872 bool set;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800873 /*
874 * get_huge_zero_page() will never allocate a new page here,
875 * since we already have a zero page to copy. It just takes a
876 * reference.
877 */
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700878 zero_page = get_huge_zero_page();
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800879 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700880 zero_page);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800881 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800882 ret = 0;
883 goto out_unlock;
884 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800885 if (unlikely(pmd_trans_splitting(pmd))) {
886 /* split huge page running from under us */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800887 spin_unlock(src_ptl);
888 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800889 pte_free(dst_mm, pgtable);
890
891 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
892 goto out;
893 }
894 src_page = pmd_page(pmd);
895 VM_BUG_ON(!PageHead(src_page));
896 get_page(src_page);
897 page_dup_rmap(src_page);
898 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
899
900 pmdp_set_wrprotect(src_mm, addr, src_pmd);
901 pmd = pmd_mkold(pmd_wrprotect(pmd));
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700902 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800903 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800904 atomic_long_inc(&dst_mm->nr_ptes);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800905
906 ret = 0;
907out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800908 spin_unlock(src_ptl);
909 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800910out:
911 return ret;
912}
913
Will Deacona1dd4502012-12-11 16:01:27 -0800914void huge_pmd_set_accessed(struct mm_struct *mm,
915 struct vm_area_struct *vma,
916 unsigned long address,
917 pmd_t *pmd, pmd_t orig_pmd,
918 int dirty)
919{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800920 spinlock_t *ptl;
Will Deacona1dd4502012-12-11 16:01:27 -0800921 pmd_t entry;
922 unsigned long haddr;
923
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800924 ptl = pmd_lock(mm, pmd);
Will Deacona1dd4502012-12-11 16:01:27 -0800925 if (unlikely(!pmd_same(*pmd, orig_pmd)))
926 goto unlock;
927
928 entry = pmd_mkyoung(orig_pmd);
929 haddr = address & HPAGE_PMD_MASK;
930 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
931 update_mmu_cache_pmd(vma, address, pmd);
932
933unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800934 spin_unlock(ptl);
Will Deacona1dd4502012-12-11 16:01:27 -0800935}
936
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800937static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
938 struct vm_area_struct *vma, unsigned long address,
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800939 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800940{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800941 spinlock_t *ptl;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800942 pgtable_t pgtable;
943 pmd_t _pmd;
944 struct page *page;
945 int i, ret = 0;
946 unsigned long mmun_start; /* For mmu_notifiers */
947 unsigned long mmun_end; /* For mmu_notifiers */
948
949 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
950 if (!page) {
951 ret |= VM_FAULT_OOM;
952 goto out;
953 }
954
955 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
956 put_page(page);
957 ret |= VM_FAULT_OOM;
958 goto out;
959 }
960
961 clear_user_highpage(page, address);
962 __SetPageUptodate(page);
963
964 mmun_start = haddr;
965 mmun_end = haddr + HPAGE_PMD_SIZE;
966 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
967
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800968 ptl = pmd_lock(mm, pmd);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800969 if (unlikely(!pmd_same(*pmd, orig_pmd)))
970 goto out_free_page;
971
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800972 pmdp_clear_flush(vma, haddr, pmd);
973 /* leave pmd empty until pte is filled */
974
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700975 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800976 pmd_populate(mm, &_pmd, pgtable);
977
978 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
979 pte_t *pte, entry;
980 if (haddr == (address & PAGE_MASK)) {
981 entry = mk_pte(page, vma->vm_page_prot);
982 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
983 page_add_new_anon_rmap(page, vma, haddr);
984 } else {
985 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
986 entry = pte_mkspecial(entry);
987 }
988 pte = pte_offset_map(&_pmd, haddr);
989 VM_BUG_ON(!pte_none(*pte));
990 set_pte_at(mm, haddr, pte, entry);
991 pte_unmap(pte);
992 }
993 smp_wmb(); /* make pte visible before pmd */
994 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800995 spin_unlock(ptl);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800996 put_huge_zero_page();
Kirill A. Shutemov93b47962012-12-12 13:50:54 -0800997 inc_mm_counter(mm, MM_ANONPAGES);
998
999 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1000
1001 ret |= VM_FAULT_WRITE;
1002out:
1003 return ret;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -08001004out_free_page:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001005 spin_unlock(ptl);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -08001006 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1007 mem_cgroup_uncharge_page(page);
1008 put_page(page);
1009 goto out;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001010}
1011
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001012static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1013 struct vm_area_struct *vma,
1014 unsigned long address,
1015 pmd_t *pmd, pmd_t orig_pmd,
1016 struct page *page,
1017 unsigned long haddr)
1018{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001019 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001020 pgtable_t pgtable;
1021 pmd_t _pmd;
1022 int ret = 0, i;
1023 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001024 unsigned long mmun_start; /* For mmu_notifiers */
1025 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001026
1027 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1028 GFP_KERNEL);
1029 if (unlikely(!pages)) {
1030 ret |= VM_FAULT_OOM;
1031 goto out;
1032 }
1033
1034 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -07001035 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1036 __GFP_OTHER_NODE,
Andi Kleen19ee1512011-03-04 17:36:31 -08001037 vma, address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001038 if (unlikely(!pages[i] ||
1039 mem_cgroup_newpage_charge(pages[i], mm,
1040 GFP_KERNEL))) {
1041 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001042 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001043 mem_cgroup_uncharge_start();
1044 while (--i >= 0) {
1045 mem_cgroup_uncharge_page(pages[i]);
1046 put_page(pages[i]);
1047 }
1048 mem_cgroup_uncharge_end();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001049 kfree(pages);
1050 ret |= VM_FAULT_OOM;
1051 goto out;
1052 }
1053 }
1054
1055 for (i = 0; i < HPAGE_PMD_NR; i++) {
1056 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -07001057 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001058 __SetPageUptodate(pages[i]);
1059 cond_resched();
1060 }
1061
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001062 mmun_start = haddr;
1063 mmun_end = haddr + HPAGE_PMD_SIZE;
1064 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1065
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001066 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001067 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1068 goto out_free_pages;
1069 VM_BUG_ON(!PageHead(page));
1070
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001071 pmdp_clear_flush(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001072 /* leave pmd empty until pte is filled */
1073
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001074 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001075 pmd_populate(mm, &_pmd, pgtable);
1076
1077 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1078 pte_t *pte, entry;
1079 entry = mk_pte(pages[i], vma->vm_page_prot);
1080 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1081 page_add_new_anon_rmap(pages[i], vma, haddr);
1082 pte = pte_offset_map(&_pmd, haddr);
1083 VM_BUG_ON(!pte_none(*pte));
1084 set_pte_at(mm, haddr, pte, entry);
1085 pte_unmap(pte);
1086 }
1087 kfree(pages);
1088
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001089 smp_wmb(); /* make pte visible before pmd */
1090 pmd_populate(mm, pmd, pgtable);
1091 page_remove_rmap(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001092 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001093
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001094 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1095
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001096 ret |= VM_FAULT_WRITE;
1097 put_page(page);
1098
1099out:
1100 return ret;
1101
1102out_free_pages:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001103 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001104 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001105 mem_cgroup_uncharge_start();
1106 for (i = 0; i < HPAGE_PMD_NR; i++) {
1107 mem_cgroup_uncharge_page(pages[i]);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001108 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001109 }
1110 mem_cgroup_uncharge_end();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001111 kfree(pages);
1112 goto out;
1113}
1114
1115int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1116 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1117{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001118 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001119 int ret = 0;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001120 struct page *page = NULL, *new_page;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001121 unsigned long haddr;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001122 unsigned long mmun_start; /* For mmu_notifiers */
1123 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001124
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001125 ptl = pmd_lockptr(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001126 VM_BUG_ON(!vma->anon_vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001127 haddr = address & HPAGE_PMD_MASK;
1128 if (is_huge_zero_pmd(orig_pmd))
1129 goto alloc;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001130 spin_lock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001131 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1132 goto out_unlock;
1133
1134 page = pmd_page(orig_pmd);
1135 VM_BUG_ON(!PageCompound(page) || !PageHead(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001136 if (page_mapcount(page) == 1) {
1137 pmd_t entry;
1138 entry = pmd_mkyoung(orig_pmd);
1139 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1140 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
David Millerb113da62012-10-08 16:34:25 -07001141 update_mmu_cache_pmd(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001142 ret |= VM_FAULT_WRITE;
1143 goto out_unlock;
1144 }
1145 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001146 spin_unlock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001147alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001148 if (transparent_hugepage_enabled(vma) &&
1149 !transparent_hugepage_debug_cow())
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001150 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
Andi Kleencc5d4622011-03-22 16:33:13 -07001151 vma, haddr, numa_node_id(), 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001152 else
1153 new_page = NULL;
1154
1155 if (unlikely(!new_page)) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001156 if (is_huge_zero_pmd(orig_pmd)) {
1157 ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -08001158 address, pmd, orig_pmd, haddr);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001159 } else {
1160 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1161 pmd, orig_pmd, page, haddr);
1162 if (ret & VM_FAULT_OOM)
1163 split_huge_page(page);
1164 put_page(page);
1165 }
David Rientjes17766dd2013-09-12 15:14:06 -07001166 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001167 goto out;
1168 }
1169
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001170 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1171 put_page(new_page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001172 if (page) {
1173 split_huge_page(page);
1174 put_page(page);
1175 }
David Rientjes17766dd2013-09-12 15:14:06 -07001176 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001177 ret |= VM_FAULT_OOM;
1178 goto out;
1179 }
1180
David Rientjes17766dd2013-09-12 15:14:06 -07001181 count_vm_event(THP_FAULT_ALLOC);
1182
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001183 if (is_huge_zero_pmd(orig_pmd))
1184 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1185 else
1186 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001187 __SetPageUptodate(new_page);
1188
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001189 mmun_start = haddr;
1190 mmun_end = haddr + HPAGE_PMD_SIZE;
1191 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1192
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001193 spin_lock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001194 if (page)
1195 put_page(page);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001196 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001197 spin_unlock(ptl);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001198 mem_cgroup_uncharge_page(new_page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001199 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001200 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001201 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001202 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001203 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1204 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001205 pmdp_clear_flush(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001206 page_add_new_anon_rmap(new_page, vma, haddr);
1207 set_pmd_at(mm, haddr, pmd, entry);
David Millerb113da62012-10-08 16:34:25 -07001208 update_mmu_cache_pmd(vma, address, pmd);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001209 if (is_huge_zero_pmd(orig_pmd)) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001210 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001211 put_huge_zero_page();
1212 } else {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001213 VM_BUG_ON(!PageHead(page));
1214 page_remove_rmap(page);
1215 put_page(page);
1216 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001217 ret |= VM_FAULT_WRITE;
1218 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001219 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001220out_mn:
1221 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1222out:
1223 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001224out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001225 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001226 return ret;
1227}
1228
David Rientjesb676b292012-10-08 16:34:03 -07001229struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001230 unsigned long addr,
1231 pmd_t *pmd,
1232 unsigned int flags)
1233{
David Rientjesb676b292012-10-08 16:34:03 -07001234 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001235 struct page *page = NULL;
1236
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001237 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001238
1239 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1240 goto out;
1241
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001242 /* Avoid dumping huge zero page */
1243 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1244 return ERR_PTR(-EFAULT);
1245
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001246 page = pmd_page(*pmd);
1247 VM_BUG_ON(!PageHead(page));
1248 if (flags & FOLL_TOUCH) {
1249 pmd_t _pmd;
1250 /*
1251 * We should set the dirty bit only for FOLL_WRITE but
1252 * for now the dirty bit in the pmd is meaningless.
1253 * And if the dirty bit will become meaningful and
1254 * we'll only set it with FOLL_WRITE, an atomic
1255 * set_bit will be required on the pmd to set the
1256 * young bit, instead of the current set_pmd_at.
1257 */
1258 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
Aneesh Kumar K.V8663890a2013-06-06 00:20:34 -07001259 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1260 pmd, _pmd, 1))
1261 update_mmu_cache_pmd(vma, addr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001262 }
David Rientjesb676b292012-10-08 16:34:03 -07001263 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1264 if (page->mapping && trylock_page(page)) {
1265 lru_add_drain();
1266 if (page->mapping)
1267 mlock_vma_page(page);
1268 unlock_page(page);
1269 }
1270 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001271 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1272 VM_BUG_ON(!PageCompound(page));
1273 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001274 get_page_foll(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001275
1276out:
1277 return page;
1278}
1279
Mel Gormand10e63f2012-10-25 14:16:31 +02001280/* NUMA hinting page fault entry point for trans huge pmds */
Mel Gorman4daae3b2012-11-02 11:33:45 +00001281int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1282 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
Mel Gormand10e63f2012-10-25 14:16:31 +02001283{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001284 spinlock_t *ptl;
Mel Gormanb8916632013-10-07 11:28:44 +01001285 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001286 struct page *page;
Mel Gormand10e63f2012-10-25 14:16:31 +02001287 unsigned long haddr = addr & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001288 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001289 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001290 bool page_locked;
1291 bool migrated = false;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001292 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001293
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001294 ptl = pmd_lock(mm, pmdp);
Mel Gormand10e63f2012-10-25 14:16:31 +02001295 if (unlikely(!pmd_same(pmd, *pmdp)))
1296 goto out_unlock;
1297
1298 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001299 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001300 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001301 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001302 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001303 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001304 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001305 flags |= TNF_FAULT_LOCAL;
1306 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001307
Mel Gormanff9042b2013-10-07 11:28:43 +01001308 /*
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001309 * Avoid grouping on DSO/COW pages in specific and RO pages
1310 * in general, RO pages shouldn't hurt as much anyway since
1311 * they can be in shared cache state.
1312 */
1313 if (!pmd_write(pmd))
1314 flags |= TNF_NO_GROUP;
1315
1316 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001317 * Acquire the page lock to serialise THP migrations but avoid dropping
1318 * page_table_lock if at all possible
1319 */
Mel Gormanb8916632013-10-07 11:28:44 +01001320 page_locked = trylock_page(page);
1321 target_nid = mpol_misplaced(page, vma, haddr);
1322 if (target_nid == -1) {
1323 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001324 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001325 goto clear_pmdnuma;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001326
Mel Gormana54a4072013-10-07 11:28:46 +01001327 /*
1328 * Otherwise wait for potential migrations and retry. We do
1329 * relock and check_same as the page may no longer be mapped.
1330 * As the fault is being retried, do not account for it.
1331 */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001332 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001333 wait_on_page_locked(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001334 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001335 goto out;
1336 }
1337
1338 /* Page is misplaced, serialise migrations and parallel THP splits */
1339 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001340 spin_unlock(ptl);
Mel Gormana54a4072013-10-07 11:28:46 +01001341 if (!page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001342 lock_page(page);
Mel Gormanb8916632013-10-07 11:28:44 +01001343 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001344
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001345 /* Confirm the PMD did not change while page_table_lock was released */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001346 spin_lock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001347 if (unlikely(!pmd_same(pmd, *pmdp))) {
1348 unlock_page(page);
1349 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001350 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001351 goto out_unlock;
1352 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001353
Mel Gormana54a4072013-10-07 11:28:46 +01001354 /*
1355 * Migrate the THP to the requested node, returns with page unlocked
1356 * and pmd_numa cleared.
1357 */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001358 spin_unlock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001359 migrated = migrate_misplaced_transhuge_page(mm, vma,
Hugh Dickins340ef392013-02-22 16:34:33 -08001360 pmdp, pmd, addr, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001361 if (migrated) {
1362 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001363 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001364 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001365
Mel Gorman8191acb2013-10-07 11:28:45 +01001366 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001367clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001368 BUG_ON(!PageLocked(page));
Mel Gormand10e63f2012-10-25 14:16:31 +02001369 pmd = pmd_mknonnuma(pmd);
1370 set_pmd_at(mm, haddr, pmdp, pmd);
1371 VM_BUG_ON(pmd_numa(*pmdp));
1372 update_mmu_cache_pmd(vma, addr, pmdp);
Mel Gormana54a4072013-10-07 11:28:46 +01001373 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001374out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001375 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001376
1377out:
1378 if (anon_vma)
1379 page_unlock_anon_vma_read(anon_vma);
1380
Mel Gorman8191acb2013-10-07 11:28:45 +01001381 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001382 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001383
Mel Gormand10e63f2012-10-25 14:16:31 +02001384 return 0;
1385}
1386
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001387int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001388 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001389{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001390 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001391 int ret = 0;
1392
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001393 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001394 struct page *page;
1395 pgtable_t pgtable;
David Millerf5c8ad42012-10-08 16:34:26 -07001396 pmd_t orig_pmd;
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001397 /*
1398 * For architectures like ppc64 we look at deposited pgtable
1399 * when calling pmdp_get_and_clear. So do the
1400 * pgtable_trans_huge_withdraw after finishing pmdp related
1401 * operations.
1402 */
David Millerf5c8ad42012-10-08 16:34:26 -07001403 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001404 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001405 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001406 if (is_huge_zero_pmd(orig_pmd)) {
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -08001407 atomic_long_dec(&tlb->mm->nr_ptes);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001408 spin_unlock(ptl);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001409 put_huge_zero_page();
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001410 } else {
1411 page = pmd_page(orig_pmd);
1412 page_remove_rmap(page);
1413 VM_BUG_ON(page_mapcount(page) < 0);
1414 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1415 VM_BUG_ON(!PageHead(page));
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -08001416 atomic_long_dec(&tlb->mm->nr_ptes);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001417 spin_unlock(ptl);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001418 tlb_remove_page(tlb, page);
1419 }
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001420 pte_free(tlb->mm, pgtable);
1421 ret = 1;
1422 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001423 return ret;
1424}
1425
Johannes Weiner0ca16342011-01-13 15:47:02 -08001426int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1427 unsigned long addr, unsigned long end,
1428 unsigned char *vec)
1429{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001430 spinlock_t *ptl;
Johannes Weiner0ca16342011-01-13 15:47:02 -08001431 int ret = 0;
1432
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001433 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001434 /*
1435 * All logical pages in the range are present
1436 * if backed by a huge page.
1437 */
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001438 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001439 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1440 ret = 1;
1441 }
Johannes Weiner0ca16342011-01-13 15:47:02 -08001442
1443 return ret;
1444}
1445
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001446int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1447 unsigned long old_addr,
1448 unsigned long new_addr, unsigned long old_end,
1449 pmd_t *old_pmd, pmd_t *new_pmd)
1450{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001451 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001452 int ret = 0;
1453 pmd_t pmd;
1454
1455 struct mm_struct *mm = vma->vm_mm;
1456
1457 if ((old_addr & ~HPAGE_PMD_MASK) ||
1458 (new_addr & ~HPAGE_PMD_MASK) ||
1459 old_end - old_addr < HPAGE_PMD_SIZE ||
1460 (new_vma->vm_flags & VM_NOHUGEPAGE))
1461 goto out;
1462
1463 /*
1464 * The destination pmd shouldn't be established, free_pgtables()
1465 * should have release it.
1466 */
1467 if (WARN_ON(!pmd_none(*new_pmd))) {
1468 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1469 goto out;
1470 }
1471
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001472 /*
1473 * We don't have to worry about the ordering of src and dst
1474 * ptlocks because exclusive mmap_sem prevents deadlock.
1475 */
1476 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001477 if (ret == 1) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001478 new_ptl = pmd_lockptr(mm, new_pmd);
1479 if (new_ptl != old_ptl)
1480 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001481 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1482 VM_BUG_ON(!pmd_none(*new_pmd));
Pavel Emelyanov0f8975e2013-07-03 15:01:20 -07001483 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001484 if (new_ptl != old_ptl)
1485 spin_unlock(new_ptl);
1486 spin_unlock(old_ptl);
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001487 }
1488out:
1489 return ret;
1490}
1491
Mel Gormanf123d742013-10-07 11:28:49 +01001492/*
1493 * Returns
1494 * - 0 if PMD could not be locked
1495 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1496 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1497 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001498int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001499 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001500{
1501 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001502 spinlock_t *ptl;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001503 int ret = 0;
1504
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001505 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001506 pmd_t entry;
Mel Gormanf123d742013-10-07 11:28:49 +01001507 ret = 1;
Hugh Dickinsa4f1de12012-12-16 18:56:58 -08001508 if (!prot_numa) {
Mel Gormanf123d742013-10-07 11:28:49 +01001509 entry = pmdp_get_and_clear(mm, addr, pmd);
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001510 entry = pmd_modify(entry, newprot);
Mel Gormanf123d742013-10-07 11:28:49 +01001511 ret = HPAGE_PMD_NR;
Hugh Dickinsa4f1de12012-12-16 18:56:58 -08001512 BUG_ON(pmd_write(entry));
1513 } else {
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001514 struct page *page = pmd_page(*pmd);
1515
Mel Gormana1a46182013-10-07 11:28:50 +01001516 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001517 * Do not trap faults against the zero page. The
1518 * read-only data is likely to be read-cached on the
1519 * local CPU cache and it is less useful to know about
1520 * local vs remote hits on the zero page.
Mel Gormana1a46182013-10-07 11:28:50 +01001521 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001522 if (!is_huge_zero_page(page) &&
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001523 !pmd_numa(*pmd)) {
Mel Gormanf123d742013-10-07 11:28:49 +01001524 entry = pmdp_get_and_clear(mm, addr, pmd);
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001525 entry = pmd_mknuma(entry);
Mel Gormanf123d742013-10-07 11:28:49 +01001526 ret = HPAGE_PMD_NR;
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001527 }
1528 }
Mel Gormanf123d742013-10-07 11:28:49 +01001529
1530 /* Set PMD if cleared earlier */
1531 if (ret == HPAGE_PMD_NR)
1532 set_pmd_at(mm, addr, pmd, entry);
1533
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001534 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001535 }
Johannes Weinercd7548a2011-01-13 15:47:04 -08001536
1537 return ret;
1538}
1539
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001540/*
1541 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1542 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1543 *
1544 * Note that if it returns 1, this routine returns without unlocking page
1545 * table locks. So callers must unlock them.
1546 */
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001547int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1548 spinlock_t **ptl)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001549{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001550 *ptl = pmd_lock(vma->vm_mm, pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001551 if (likely(pmd_trans_huge(*pmd))) {
1552 if (unlikely(pmd_trans_splitting(*pmd))) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001553 spin_unlock(*ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001554 wait_split_huge_page(vma->anon_vma, pmd);
1555 return -1;
1556 } else {
1557 /* Thp mapped by 'pmd' is stable, so we can
1558 * handle it as it is. */
1559 return 1;
1560 }
1561 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001562 spin_unlock(*ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001563 return 0;
1564}
1565
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001566/*
1567 * This function returns whether a given @page is mapped onto the @address
1568 * in the virtual space of @mm.
1569 *
1570 * When it's true, this function returns *pmd with holding the page table lock
1571 * and passing it back to the caller via @ptl.
1572 * If it's false, returns NULL without holding the page table lock.
1573 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001574pmd_t *page_check_address_pmd(struct page *page,
1575 struct mm_struct *mm,
1576 unsigned long address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001577 enum page_check_address_pmd_flag flag,
1578 spinlock_t **ptl)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001579{
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001580 pmd_t *pmd;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001581
1582 if (address & ~HPAGE_PMD_MASK)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001583 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001584
Bob Liu62190492012-12-11 16:00:37 -08001585 pmd = mm_find_pmd(mm, address);
1586 if (!pmd)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001587 return NULL;
1588 *ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001589 if (pmd_none(*pmd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001590 goto unlock;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001591 if (pmd_page(*pmd) != page)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001592 goto unlock;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001593 /*
1594 * split_vma() may create temporary aliased mappings. There is
1595 * no risk as long as all huge pmd are found and have their
1596 * splitting bit set before __split_huge_page_refcount
1597 * runs. Finding the same huge pmd more than once during the
1598 * same rmap walk is not a problem.
1599 */
1600 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1601 pmd_trans_splitting(*pmd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001602 goto unlock;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001603 if (pmd_trans_huge(*pmd)) {
1604 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1605 !pmd_trans_splitting(*pmd));
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001606 return pmd;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001607 }
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001608unlock:
1609 spin_unlock(*ptl);
1610 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001611}
1612
1613static int __split_huge_page_splitting(struct page *page,
1614 struct vm_area_struct *vma,
1615 unsigned long address)
1616{
1617 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001618 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001619 pmd_t *pmd;
1620 int ret = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001621 /* For mmu_notifiers */
1622 const unsigned long mmun_start = address;
1623 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001624
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001625 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001626 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001627 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001628 if (pmd) {
1629 /*
1630 * We can't temporarily set the pmd to null in order
1631 * to split it, the pmd must remain marked huge at all
1632 * times or the VM won't take the pmd_trans_huge paths
Ingo Molnar5a505082012-12-02 19:56:46 +00001633 * and it won't wait on the anon_vma->root->rwsem to
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001634 * serialize against split_huge_page*.
1635 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001636 pmdp_splitting_flush(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001637 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001638 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001639 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001640 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001641
1642 return ret;
1643}
1644
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001645static void __split_huge_page_refcount(struct page *page,
1646 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001647{
1648 int i;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001649 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001650 struct lruvec *lruvec;
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001651 int tail_count = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001652
1653 /* prevent PageLRU to go away from under us, and freeze lru stats */
1654 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001655 lruvec = mem_cgroup_page_lruvec(page, zone);
1656
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001657 compound_lock(page);
KAMEZAWA Hiroyukie94c8a92012-01-12 17:18:20 -08001658 /* complete memcg works before add pages to LRU */
1659 mem_cgroup_split_huge_fixup(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001660
Shaohua Li45676882012-01-12 17:19:18 -08001661 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001662 struct page *page_tail = page + i;
1663
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001664 /* tail_page->_mapcount cannot change */
1665 BUG_ON(page_mapcount(page_tail) < 0);
1666 tail_count += page_mapcount(page_tail);
1667 /* check for overflow */
1668 BUG_ON(tail_count < 0);
1669 BUG_ON(atomic_read(&page_tail->_count) != 0);
1670 /*
1671 * tail_page->_count is zero and not changing from
1672 * under us. But get_page_unless_zero() may be running
1673 * from under us on the tail_page. If we used
1674 * atomic_set() below instead of atomic_add(), we
1675 * would then run atomic_set() concurrently with
1676 * get_page_unless_zero(), and atomic_set() is
1677 * implemented in C not using locked ops. spin_unlock
1678 * on x86 sometime uses locked ops because of PPro
1679 * errata 66, 92, so unless somebody can guarantee
1680 * atomic_set() here would be safe on all archs (and
1681 * not only on x86), it's safer to use atomic_add().
1682 */
1683 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1684 &page_tail->_count);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001685
1686 /* after clearing PageTail the gup refcount can be released */
1687 smp_mb();
1688
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001689 /*
1690 * retain hwpoison flag of the poisoned tail page:
1691 * fix for the unsuitable process killed on Guest Machine(KVM)
1692 * by the memory-failure.
1693 */
1694 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001695 page_tail->flags |= (page->flags &
1696 ((1L << PG_referenced) |
1697 (1L << PG_swapbacked) |
1698 (1L << PG_mlocked) |
Kirill A. Shutemove180cf82013-07-31 13:53:39 -07001699 (1L << PG_uptodate) |
1700 (1L << PG_active) |
1701 (1L << PG_unevictable)));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001702 page_tail->flags |= (1L << PG_dirty);
1703
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001704 /* clear PageTail before overwriting first_page */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001705 smp_wmb();
1706
1707 /*
1708 * __split_huge_page_splitting() already set the
1709 * splitting bit in all pmd that could map this
1710 * hugepage, that will ensure no CPU can alter the
1711 * mapcount on the head page. The mapcount is only
1712 * accounted in the head page and it has to be
1713 * transferred to all tail pages in the below code. So
1714 * for this code to be safe, the split the mapcount
1715 * can't change. But that doesn't mean userland can't
1716 * keep changing and reading the page contents while
1717 * we transfer the mapcount, so the pmd splitting
1718 * status is achieved setting a reserved bit in the
1719 * pmd, not by clearing the present bit.
1720 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001721 page_tail->_mapcount = page->_mapcount;
1722
1723 BUG_ON(page_tail->mapping);
1724 page_tail->mapping = page->mapping;
1725
Shaohua Li45676882012-01-12 17:19:18 -08001726 page_tail->index = page->index + i;
Peter Zijlstra90572892013-10-07 11:29:20 +01001727 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001728
1729 BUG_ON(!PageAnon(page_tail));
1730 BUG_ON(!PageUptodate(page_tail));
1731 BUG_ON(!PageDirty(page_tail));
1732 BUG_ON(!PageSwapBacked(page_tail));
1733
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001734 lru_add_page_tail(page, page_tail, lruvec, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001735 }
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001736 atomic_sub(tail_count, &page->_count);
1737 BUG_ON(atomic_read(&page->_count) <= 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001738
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001739 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001740
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001741 ClearPageCompound(page);
1742 compound_unlock(page);
1743 spin_unlock_irq(&zone->lru_lock);
1744
1745 for (i = 1; i < HPAGE_PMD_NR; i++) {
1746 struct page *page_tail = page + i;
1747 BUG_ON(page_count(page_tail) <= 0);
1748 /*
1749 * Tail pages may be freed if there wasn't any mapping
1750 * like if add_to_swap() is running on a lru page that
1751 * had its mapping zapped. And freeing these pages
1752 * requires taking the lru_lock so we do the put_page
1753 * of the tail pages after the split is complete.
1754 */
1755 put_page(page_tail);
1756 }
1757
1758 /*
1759 * Only the head page (now become a regular page) is required
1760 * to be pinned by the caller.
1761 */
1762 BUG_ON(page_count(page) <= 0);
1763}
1764
1765static int __split_huge_page_map(struct page *page,
1766 struct vm_area_struct *vma,
1767 unsigned long address)
1768{
1769 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001770 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001771 pmd_t *pmd, _pmd;
1772 int ret = 0, i;
1773 pgtable_t pgtable;
1774 unsigned long haddr;
1775
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001776 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001777 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001778 if (pmd) {
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001779 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001780 pmd_populate(mm, &_pmd, pgtable);
1781
Gerald Schaefere3ebcf642012-10-08 16:30:07 -07001782 haddr = address;
1783 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001784 pte_t *pte, entry;
1785 BUG_ON(PageCompound(page+i));
1786 entry = mk_pte(page + i, vma->vm_page_prot);
1787 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1788 if (!pmd_write(*pmd))
1789 entry = pte_wrprotect(entry);
1790 else
1791 BUG_ON(page_mapcount(page) != 1);
1792 if (!pmd_young(*pmd))
1793 entry = pte_mkold(entry);
Andrea Arcangeli1ba6e0b2012-10-04 01:51:06 +02001794 if (pmd_numa(*pmd))
1795 entry = pte_mknuma(entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001796 pte = pte_offset_map(&_pmd, haddr);
1797 BUG_ON(!pte_none(*pte));
1798 set_pte_at(mm, haddr, pte, entry);
1799 pte_unmap(pte);
1800 }
1801
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001802 smp_wmb(); /* make pte visible before pmd */
1803 /*
1804 * Up to this point the pmd is present and huge and
1805 * userland has the whole access to the hugepage
1806 * during the split (which happens in place). If we
1807 * overwrite the pmd with the not-huge version
1808 * pointing to the pte here (which of course we could
1809 * if all CPUs were bug free), userland could trigger
1810 * a small page size TLB miss on the small sized TLB
1811 * while the hugepage TLB entry is still established
1812 * in the huge TLB. Some CPU doesn't like that. See
1813 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1814 * Erratum 383 on page 93. Intel should be safe but is
1815 * also warns that it's only safe if the permission
1816 * and cache attributes of the two entries loaded in
1817 * the two TLB is identical (which should be the case
1818 * here). But it is generally safer to never allow
1819 * small and huge TLB entries for the same virtual
1820 * address to be loaded simultaneously. So instead of
1821 * doing "pmd_populate(); flush_tlb_range();" we first
1822 * mark the current pmd notpresent (atomically because
1823 * here the pmd_trans_huge and pmd_trans_splitting
1824 * must remain set at all times on the pmd until the
1825 * split is complete for this pmd), then we flush the
1826 * SMP TLB and finally we write the non-huge version
1827 * of the pmd entry with pmd_populate.
1828 */
Gerald Schaefer46dcde72012-10-08 16:30:09 -07001829 pmdp_invalidate(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001830 pmd_populate(mm, pmd, pgtable);
1831 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001832 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001833 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001834
1835 return ret;
1836}
1837
Ingo Molnar5a505082012-12-02 19:56:46 +00001838/* must be called with anon_vma->root->rwsem held */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001839static void __split_huge_page(struct page *page,
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001840 struct anon_vma *anon_vma,
1841 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001842{
1843 int mapcount, mapcount2;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001844 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001845 struct anon_vma_chain *avc;
1846
1847 BUG_ON(!PageHead(page));
1848 BUG_ON(PageTail(page));
1849
1850 mapcount = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001851 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001852 struct vm_area_struct *vma = avc->vma;
1853 unsigned long addr = vma_address(page, vma);
1854 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001855 mapcount += __split_huge_page_splitting(page, vma, addr);
1856 }
Andrea Arcangeli05759d32011-01-13 15:46:53 -08001857 /*
1858 * It is critical that new vmas are added to the tail of the
1859 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1860 * and establishes a child pmd before
1861 * __split_huge_page_splitting() freezes the parent pmd (so if
1862 * we fail to prevent copy_huge_pmd() from running until the
1863 * whole __split_huge_page() is complete), we will still see
1864 * the newly established pmd of the child later during the
1865 * walk, to be able to set it as pmd_trans_splitting too.
1866 */
1867 if (mapcount != page_mapcount(page))
1868 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1869 mapcount, page_mapcount(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001870 BUG_ON(mapcount != page_mapcount(page));
1871
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001872 __split_huge_page_refcount(page, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001873
1874 mapcount2 = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001875 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001876 struct vm_area_struct *vma = avc->vma;
1877 unsigned long addr = vma_address(page, vma);
1878 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001879 mapcount2 += __split_huge_page_map(page, vma, addr);
1880 }
Andrea Arcangeli05759d32011-01-13 15:46:53 -08001881 if (mapcount != mapcount2)
1882 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1883 mapcount, mapcount2, page_mapcount(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001884 BUG_ON(mapcount != mapcount2);
1885}
1886
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001887/*
1888 * Split a hugepage into normal pages. This doesn't change the position of head
1889 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1890 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1891 * from the hugepage.
1892 * Return 0 if the hugepage is split successfully otherwise return 1.
1893 */
1894int split_huge_page_to_list(struct page *page, struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001895{
1896 struct anon_vma *anon_vma;
1897 int ret = 1;
1898
Kirill A. Shutemov5918d102013-04-29 15:08:44 -07001899 BUG_ON(is_huge_zero_page(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001900 BUG_ON(!PageAnon(page));
Mel Gorman062f1af2013-01-11 14:32:02 -08001901
1902 /*
1903 * The caller does not necessarily hold an mmap_sem that would prevent
1904 * the anon_vma disappearing so we first we take a reference to it
1905 * and then lock the anon_vma for write. This is similar to
1906 * page_lock_anon_vma_read except the write lock is taken to serialise
1907 * against parallel split or collapse operations.
1908 */
1909 anon_vma = page_get_anon_vma(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001910 if (!anon_vma)
1911 goto out;
Mel Gorman062f1af2013-01-11 14:32:02 -08001912 anon_vma_lock_write(anon_vma);
1913
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001914 ret = 0;
1915 if (!PageCompound(page))
1916 goto out_unlock;
1917
1918 BUG_ON(!PageSwapBacked(page));
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001919 __split_huge_page(page, anon_vma, list);
Andi Kleen81ab4202011-04-14 15:22:06 -07001920 count_vm_event(THP_SPLIT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001921
1922 BUG_ON(PageCompound(page));
1923out_unlock:
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08001924 anon_vma_unlock_write(anon_vma);
Mel Gorman062f1af2013-01-11 14:32:02 -08001925 put_anon_vma(anon_vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001926out:
1927 return ret;
1928}
1929
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001930#define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001931
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001932int hugepage_madvise(struct vm_area_struct *vma,
1933 unsigned long *vm_flags, int advice)
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001934{
Gerald Schaefer8e720332012-10-08 16:30:12 -07001935 struct mm_struct *mm = vma->vm_mm;
1936
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001937 switch (advice) {
1938 case MADV_HUGEPAGE:
1939 /*
1940 * Be somewhat over-protective like KSM for now!
1941 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001942 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001943 return -EINVAL;
Gerald Schaefer8e720332012-10-08 16:30:12 -07001944 if (mm->def_flags & VM_NOHUGEPAGE)
1945 return -EINVAL;
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001946 *vm_flags &= ~VM_NOHUGEPAGE;
1947 *vm_flags |= VM_HUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001948 /*
1949 * If the vma become good for khugepaged to scan,
1950 * register it here without waiting a page fault that
1951 * may not happen any time soon.
1952 */
1953 if (unlikely(khugepaged_enter_vma_merge(vma)))
1954 return -ENOMEM;
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001955 break;
1956 case MADV_NOHUGEPAGE:
1957 /*
1958 * Be somewhat over-protective like KSM for now!
1959 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001960 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001961 return -EINVAL;
1962 *vm_flags &= ~VM_HUGEPAGE;
1963 *vm_flags |= VM_NOHUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001964 /*
1965 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1966 * this vma even if we leave the mm registered in khugepaged if
1967 * it got registered before VM_NOHUGEPAGE was set.
1968 */
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001969 break;
1970 }
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001971
1972 return 0;
1973}
1974
Andrea Arcangeliba761492011-01-13 15:46:58 -08001975static int __init khugepaged_slab_init(void)
1976{
1977 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1978 sizeof(struct mm_slot),
1979 __alignof__(struct mm_slot), 0, NULL);
1980 if (!mm_slot_cache)
1981 return -ENOMEM;
1982
1983 return 0;
1984}
1985
Andrea Arcangeliba761492011-01-13 15:46:58 -08001986static inline struct mm_slot *alloc_mm_slot(void)
1987{
1988 if (!mm_slot_cache) /* initialization failed */
1989 return NULL;
1990 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1991}
1992
1993static inline void free_mm_slot(struct mm_slot *mm_slot)
1994{
1995 kmem_cache_free(mm_slot_cache, mm_slot);
1996}
1997
Andrea Arcangeliba761492011-01-13 15:46:58 -08001998static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1999{
2000 struct mm_slot *mm_slot;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002001
Sasha Levinb67bfe02013-02-27 17:06:00 -08002002 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002003 if (mm == mm_slot->mm)
2004 return mm_slot;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002005
Andrea Arcangeliba761492011-01-13 15:46:58 -08002006 return NULL;
2007}
2008
2009static void insert_to_mm_slots_hash(struct mm_struct *mm,
2010 struct mm_slot *mm_slot)
2011{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002012 mm_slot->mm = mm;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002013 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002014}
2015
2016static inline int khugepaged_test_exit(struct mm_struct *mm)
2017{
2018 return atomic_read(&mm->mm_users) == 0;
2019}
2020
2021int __khugepaged_enter(struct mm_struct *mm)
2022{
2023 struct mm_slot *mm_slot;
2024 int wakeup;
2025
2026 mm_slot = alloc_mm_slot();
2027 if (!mm_slot)
2028 return -ENOMEM;
2029
2030 /* __khugepaged_exit() must not run from under us */
2031 VM_BUG_ON(khugepaged_test_exit(mm));
2032 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2033 free_mm_slot(mm_slot);
2034 return 0;
2035 }
2036
2037 spin_lock(&khugepaged_mm_lock);
2038 insert_to_mm_slots_hash(mm, mm_slot);
2039 /*
2040 * Insert just behind the scanning cursor, to let the area settle
2041 * down a little.
2042 */
2043 wakeup = list_empty(&khugepaged_scan.mm_head);
2044 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2045 spin_unlock(&khugepaged_mm_lock);
2046
2047 atomic_inc(&mm->mm_count);
2048 if (wakeup)
2049 wake_up_interruptible(&khugepaged_wait);
2050
2051 return 0;
2052}
2053
2054int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2055{
2056 unsigned long hstart, hend;
2057 if (!vma->anon_vma)
2058 /*
2059 * Not yet faulted in so we will register later in the
2060 * page fault if needed.
2061 */
2062 return 0;
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07002063 if (vma->vm_ops)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002064 /* khugepaged not yet working on file or special mappings */
2065 return 0;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002066 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002067 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2068 hend = vma->vm_end & HPAGE_PMD_MASK;
2069 if (hstart < hend)
2070 return khugepaged_enter(vma);
2071 return 0;
2072}
2073
2074void __khugepaged_exit(struct mm_struct *mm)
2075{
2076 struct mm_slot *mm_slot;
2077 int free = 0;
2078
2079 spin_lock(&khugepaged_mm_lock);
2080 mm_slot = get_mm_slot(mm);
2081 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002082 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002083 list_del(&mm_slot->mm_node);
2084 free = 1;
2085 }
Chris Wrightd788e802011-07-25 17:12:14 -07002086 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002087
2088 if (free) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002089 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2090 free_mm_slot(mm_slot);
2091 mmdrop(mm);
2092 } else if (mm_slot) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002093 /*
2094 * This is required to serialize against
2095 * khugepaged_test_exit() (which is guaranteed to run
2096 * under mmap sem read mode). Stop here (after we
2097 * return all pagetables will be destroyed) until
2098 * khugepaged has finished working on the pagetables
2099 * under the mmap_sem.
2100 */
2101 down_write(&mm->mmap_sem);
2102 up_write(&mm->mmap_sem);
Chris Wrightd788e802011-07-25 17:12:14 -07002103 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002104}
2105
2106static void release_pte_page(struct page *page)
2107{
2108 /* 0 stands for page_is_file_cache(page) == false */
2109 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2110 unlock_page(page);
2111 putback_lru_page(page);
2112}
2113
2114static void release_pte_pages(pte_t *pte, pte_t *_pte)
2115{
2116 while (--_pte >= pte) {
2117 pte_t pteval = *_pte;
2118 if (!pte_none(pteval))
2119 release_pte_page(pte_page(pteval));
2120 }
2121}
2122
Andrea Arcangeliba761492011-01-13 15:46:58 -08002123static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2124 unsigned long address,
2125 pte_t *pte)
2126{
2127 struct page *page;
2128 pte_t *_pte;
Bob Liu344aa352012-12-11 16:00:34 -08002129 int referenced = 0, none = 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002130 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2131 _pte++, address += PAGE_SIZE) {
2132 pte_t pteval = *_pte;
2133 if (pte_none(pteval)) {
2134 if (++none <= khugepaged_max_ptes_none)
2135 continue;
Bob Liu344aa352012-12-11 16:00:34 -08002136 else
Andrea Arcangeliba761492011-01-13 15:46:58 -08002137 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002138 }
Bob Liu344aa352012-12-11 16:00:34 -08002139 if (!pte_present(pteval) || !pte_write(pteval))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002140 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002141 page = vm_normal_page(vma, address, pteval);
Bob Liu344aa352012-12-11 16:00:34 -08002142 if (unlikely(!page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002143 goto out;
Bob Liu344aa352012-12-11 16:00:34 -08002144
Andrea Arcangeliba761492011-01-13 15:46:58 -08002145 VM_BUG_ON(PageCompound(page));
2146 BUG_ON(!PageAnon(page));
2147 VM_BUG_ON(!PageSwapBacked(page));
2148
2149 /* cannot use mapcount: can't collapse if there's a gup pin */
Bob Liu344aa352012-12-11 16:00:34 -08002150 if (page_count(page) != 1)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002151 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002152 /*
2153 * We can do it before isolate_lru_page because the
2154 * page can't be freed from under us. NOTE: PG_lock
2155 * is needed to serialize against split_huge_page
2156 * when invoked from the VM.
2157 */
Bob Liu344aa352012-12-11 16:00:34 -08002158 if (!trylock_page(page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002159 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002160 /*
2161 * Isolate the page to avoid collapsing an hugepage
2162 * currently in use by the VM.
2163 */
2164 if (isolate_lru_page(page)) {
2165 unlock_page(page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002166 goto out;
2167 }
2168 /* 0 stands for page_is_file_cache(page) == false */
2169 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2170 VM_BUG_ON(!PageLocked(page));
2171 VM_BUG_ON(PageLRU(page));
2172
2173 /* If there is no mapped pte young don't collapse the page */
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002174 if (pte_young(pteval) || PageReferenced(page) ||
2175 mmu_notifier_test_young(vma->vm_mm, address))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002176 referenced = 1;
2177 }
Bob Liu344aa352012-12-11 16:00:34 -08002178 if (likely(referenced))
2179 return 1;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002180out:
Bob Liu344aa352012-12-11 16:00:34 -08002181 release_pte_pages(pte, _pte);
2182 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002183}
2184
2185static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2186 struct vm_area_struct *vma,
2187 unsigned long address,
2188 spinlock_t *ptl)
2189{
2190 pte_t *_pte;
2191 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2192 pte_t pteval = *_pte;
2193 struct page *src_page;
2194
2195 if (pte_none(pteval)) {
2196 clear_user_highpage(page, address);
2197 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2198 } else {
2199 src_page = pte_page(pteval);
2200 copy_user_highpage(page, src_page, address, vma);
2201 VM_BUG_ON(page_mapcount(src_page) != 1);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002202 release_pte_page(src_page);
2203 /*
2204 * ptl mostly unnecessary, but preempt has to
2205 * be disabled to update the per-cpu stats
2206 * inside page_remove_rmap().
2207 */
2208 spin_lock(ptl);
2209 /*
2210 * paravirt calls inside pte_clear here are
2211 * superfluous.
2212 */
2213 pte_clear(vma->vm_mm, address, _pte);
2214 page_remove_rmap(src_page);
2215 spin_unlock(ptl);
2216 free_page_and_swap_cache(src_page);
2217 }
2218
2219 address += PAGE_SIZE;
2220 page++;
2221 }
2222}
2223
Xiao Guangrong26234f32012-10-08 16:29:51 -07002224static void khugepaged_alloc_sleep(void)
2225{
2226 wait_event_freezable_timeout(khugepaged_wait, false,
2227 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2228}
2229
Bob Liu9f1b8682013-11-12 15:07:37 -08002230static int khugepaged_node_load[MAX_NUMNODES];
2231
Xiao Guangrong26234f32012-10-08 16:29:51 -07002232#ifdef CONFIG_NUMA
Bob Liu9f1b8682013-11-12 15:07:37 -08002233static int khugepaged_find_target_node(void)
2234{
2235 static int last_khugepaged_target_node = NUMA_NO_NODE;
2236 int nid, target_node = 0, max_value = 0;
2237
2238 /* find first node with max normal pages hit */
2239 for (nid = 0; nid < MAX_NUMNODES; nid++)
2240 if (khugepaged_node_load[nid] > max_value) {
2241 max_value = khugepaged_node_load[nid];
2242 target_node = nid;
2243 }
2244
2245 /* do some balance if several nodes have the same hit record */
2246 if (target_node <= last_khugepaged_target_node)
2247 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2248 nid++)
2249 if (max_value == khugepaged_node_load[nid]) {
2250 target_node = nid;
2251 break;
2252 }
2253
2254 last_khugepaged_target_node = target_node;
2255 return target_node;
2256}
2257
Xiao Guangrong26234f32012-10-08 16:29:51 -07002258static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2259{
2260 if (IS_ERR(*hpage)) {
2261 if (!*wait)
2262 return false;
2263
2264 *wait = false;
Xiao Guangronge3b41262012-10-08 16:32:57 -07002265 *hpage = NULL;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002266 khugepaged_alloc_sleep();
2267 } else if (*hpage) {
2268 put_page(*hpage);
2269 *hpage = NULL;
2270 }
2271
2272 return true;
2273}
2274
2275static struct page
2276*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2277 struct vm_area_struct *vma, unsigned long address,
2278 int node)
2279{
2280 VM_BUG_ON(*hpage);
2281 /*
2282 * Allocate the page while the vma is still valid and under
2283 * the mmap_sem read mode so there is no memory allocation
2284 * later when we take the mmap_sem in write mode. This is more
2285 * friendly behavior (OTOH it may actually hide bugs) to
2286 * filesystems in userland with daemons allocating memory in
2287 * the userland I/O paths. Allocating memory with the
2288 * mmap_sem in read mode is good idea also to allow greater
2289 * scalability.
2290 */
Bob Liu9f1b8682013-11-12 15:07:37 -08002291 *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2292 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002293 /*
2294 * After allocating the hugepage, release the mmap_sem read lock in
2295 * preparation for taking it in write mode.
2296 */
2297 up_read(&mm->mmap_sem);
2298 if (unlikely(!*hpage)) {
2299 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2300 *hpage = ERR_PTR(-ENOMEM);
2301 return NULL;
2302 }
2303
2304 count_vm_event(THP_COLLAPSE_ALLOC);
2305 return *hpage;
2306}
2307#else
Bob Liu9f1b8682013-11-12 15:07:37 -08002308static int khugepaged_find_target_node(void)
2309{
2310 return 0;
2311}
2312
Bob Liu10dc4152013-11-12 15:07:35 -08002313static inline struct page *alloc_hugepage(int defrag)
2314{
2315 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2316 HPAGE_PMD_ORDER);
2317}
2318
Xiao Guangrong26234f32012-10-08 16:29:51 -07002319static struct page *khugepaged_alloc_hugepage(bool *wait)
2320{
2321 struct page *hpage;
2322
2323 do {
2324 hpage = alloc_hugepage(khugepaged_defrag());
2325 if (!hpage) {
2326 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2327 if (!*wait)
2328 return NULL;
2329
2330 *wait = false;
2331 khugepaged_alloc_sleep();
2332 } else
2333 count_vm_event(THP_COLLAPSE_ALLOC);
2334 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2335
2336 return hpage;
2337}
2338
2339static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2340{
2341 if (!*hpage)
2342 *hpage = khugepaged_alloc_hugepage(wait);
2343
2344 if (unlikely(!*hpage))
2345 return false;
2346
2347 return true;
2348}
2349
2350static struct page
2351*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2352 struct vm_area_struct *vma, unsigned long address,
2353 int node)
2354{
2355 up_read(&mm->mmap_sem);
2356 VM_BUG_ON(!*hpage);
2357 return *hpage;
2358}
2359#endif
2360
Bob Liufa475e52012-12-11 16:00:39 -08002361static bool hugepage_vma_check(struct vm_area_struct *vma)
2362{
2363 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2364 (vma->vm_flags & VM_NOHUGEPAGE))
2365 return false;
2366
2367 if (!vma->anon_vma || vma->vm_ops)
2368 return false;
2369 if (is_vma_temporary_stack(vma))
2370 return false;
2371 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2372 return true;
2373}
2374
Andrea Arcangeliba761492011-01-13 15:46:58 -08002375static void collapse_huge_page(struct mm_struct *mm,
Xiao Guangrong26234f32012-10-08 16:29:51 -07002376 unsigned long address,
2377 struct page **hpage,
2378 struct vm_area_struct *vma,
2379 int node)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002380{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002381 pmd_t *pmd, _pmd;
2382 pte_t *pte;
2383 pgtable_t pgtable;
2384 struct page *new_page;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002385 spinlock_t *pmd_ptl, *pte_ptl;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002386 int isolated;
2387 unsigned long hstart, hend;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002388 unsigned long mmun_start; /* For mmu_notifiers */
2389 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002390
2391 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002392
Xiao Guangrong26234f32012-10-08 16:29:51 -07002393 /* release the mmap_sem read lock. */
2394 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2395 if (!new_page)
Andrea Arcangelice83d212011-01-13 15:47:06 -08002396 return;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002397
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002398 if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002399 return;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002400
2401 /*
2402 * Prevent all access to pagetables with the exception of
2403 * gup_fast later hanlded by the ptep_clear_flush and the VM
2404 * handled by the anon_vma lock + PG_lock.
2405 */
2406 down_write(&mm->mmap_sem);
2407 if (unlikely(khugepaged_test_exit(mm)))
2408 goto out;
2409
2410 vma = find_vma(mm, address);
Libina8f531eb2013-09-11 14:20:38 -07002411 if (!vma)
2412 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002413 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2414 hend = vma->vm_end & HPAGE_PMD_MASK;
2415 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2416 goto out;
Bob Liufa475e52012-12-11 16:00:39 -08002417 if (!hugepage_vma_check(vma))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002418 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002419 pmd = mm_find_pmd(mm, address);
2420 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002421 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002422 if (pmd_trans_huge(*pmd))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002423 goto out;
2424
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +00002425 anon_vma_lock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002426
2427 pte = pte_offset_map(pmd, address);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002428 pte_ptl = pte_lockptr(mm, pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002429
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002430 mmun_start = address;
2431 mmun_end = address + HPAGE_PMD_SIZE;
2432 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002433 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002434 /*
2435 * After this gup_fast can't run anymore. This also removes
2436 * any huge TLB entry from the CPU so we won't allow
2437 * huge and small TLB entries for the same virtual address
2438 * to avoid the risk of CPU bugs in that area.
2439 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002440 _pmd = pmdp_clear_flush(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002441 spin_unlock(pmd_ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002442 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002443
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002444 spin_lock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002445 isolated = __collapse_huge_page_isolate(vma, address, pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002446 spin_unlock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002447
2448 if (unlikely(!isolated)) {
Johannes Weiner453c7192011-01-20 14:44:18 -08002449 pte_unmap(pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002450 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002451 BUG_ON(!pmd_none(*pmd));
Aneesh Kumar K.V7c342512013-05-24 15:55:21 -07002452 /*
2453 * We can only use set_pmd_at when establishing
2454 * hugepmds and never for establishing regular pmds that
2455 * points to regular pagetables. Use pmd_populate for that
2456 */
2457 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002458 spin_unlock(pmd_ptl);
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002459 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002460 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002461 }
2462
2463 /*
2464 * All pages are isolated and locked so anon_vma rmap
2465 * can't run anymore.
2466 */
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002467 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002468
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002469 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
Johannes Weiner453c7192011-01-20 14:44:18 -08002470 pte_unmap(pte);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002471 __SetPageUptodate(new_page);
2472 pgtable = pmd_pgtable(_pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002473
Kirill A. Shutemov31223592013-09-12 15:14:01 -07002474 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2475 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002476
2477 /*
2478 * spin_lock() below is not the equivalent of smp_wmb(), so
2479 * this is needed to avoid the copy_huge_page writes to become
2480 * visible after the set_pmd_at() write.
2481 */
2482 smp_wmb();
2483
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));
2486 page_add_new_anon_rmap(new_page, vma, address);
Aneesh Kumar K.Vfce144b2013-06-05 17:14:06 -07002487 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002488 set_pmd_at(mm, address, pmd, _pmd);
David Millerb113da62012-10-08 16:34:25 -07002489 update_mmu_cache_pmd(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002490 spin_unlock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002491
2492 *hpage = NULL;
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002493
Andrea Arcangeliba761492011-01-13 15:46:58 -08002494 khugepaged_pages_collapsed++;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002495out_up_write:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002496 up_write(&mm->mmap_sem);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002497 return;
2498
Andrea Arcangelice83d212011-01-13 15:47:06 -08002499out:
KAMEZAWA Hiroyuki678ff892011-02-10 15:01:36 -08002500 mem_cgroup_uncharge_page(new_page);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002501 goto out_up_write;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002502}
2503
2504static int khugepaged_scan_pmd(struct mm_struct *mm,
2505 struct vm_area_struct *vma,
2506 unsigned long address,
2507 struct page **hpage)
2508{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002509 pmd_t *pmd;
2510 pte_t *pte, *_pte;
2511 int ret = 0, referenced = 0, none = 0;
2512 struct page *page;
2513 unsigned long _address;
2514 spinlock_t *ptl;
David Rientjes00ef2d22013-02-22 16:35:36 -08002515 int node = NUMA_NO_NODE;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002516
2517 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2518
Bob Liu62190492012-12-11 16:00:37 -08002519 pmd = mm_find_pmd(mm, address);
2520 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002521 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002522 if (pmd_trans_huge(*pmd))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002523 goto out;
2524
Bob Liu9f1b8682013-11-12 15:07:37 -08002525 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002526 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2527 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2528 _pte++, _address += PAGE_SIZE) {
2529 pte_t pteval = *_pte;
2530 if (pte_none(pteval)) {
2531 if (++none <= khugepaged_max_ptes_none)
2532 continue;
2533 else
2534 goto out_unmap;
2535 }
2536 if (!pte_present(pteval) || !pte_write(pteval))
2537 goto out_unmap;
2538 page = vm_normal_page(vma, _address, pteval);
2539 if (unlikely(!page))
2540 goto out_unmap;
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002541 /*
Bob Liu9f1b8682013-11-12 15:07:37 -08002542 * Record which node the original page is from and save this
2543 * information to khugepaged_node_load[].
2544 * Khupaged will allocate hugepage from the node has the max
2545 * hit record.
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002546 */
Bob Liu9f1b8682013-11-12 15:07:37 -08002547 node = page_to_nid(page);
2548 khugepaged_node_load[node]++;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002549 VM_BUG_ON(PageCompound(page));
2550 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2551 goto out_unmap;
2552 /* cannot use mapcount: can't collapse if there's a gup pin */
2553 if (page_count(page) != 1)
2554 goto out_unmap;
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002555 if (pte_young(pteval) || PageReferenced(page) ||
2556 mmu_notifier_test_young(vma->vm_mm, address))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002557 referenced = 1;
2558 }
2559 if (referenced)
2560 ret = 1;
2561out_unmap:
2562 pte_unmap_unlock(pte, ptl);
Bob Liu9f1b8682013-11-12 15:07:37 -08002563 if (ret) {
2564 node = khugepaged_find_target_node();
Andrea Arcangelice83d212011-01-13 15:47:06 -08002565 /* collapse_huge_page will return with the mmap_sem released */
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002566 collapse_huge_page(mm, address, hpage, vma, node);
Bob Liu9f1b8682013-11-12 15:07:37 -08002567 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002568out:
2569 return ret;
2570}
2571
2572static void collect_mm_slot(struct mm_slot *mm_slot)
2573{
2574 struct mm_struct *mm = mm_slot->mm;
2575
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002576 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002577
2578 if (khugepaged_test_exit(mm)) {
2579 /* free mm_slot */
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002580 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002581 list_del(&mm_slot->mm_node);
2582
2583 /*
2584 * Not strictly needed because the mm exited already.
2585 *
2586 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2587 */
2588
2589 /* khugepaged_mm_lock actually not necessary for the below */
2590 free_mm_slot(mm_slot);
2591 mmdrop(mm);
2592 }
2593}
2594
2595static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2596 struct page **hpage)
H Hartley Sweeten2f1da642011-10-31 17:09:25 -07002597 __releases(&khugepaged_mm_lock)
2598 __acquires(&khugepaged_mm_lock)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002599{
2600 struct mm_slot *mm_slot;
2601 struct mm_struct *mm;
2602 struct vm_area_struct *vma;
2603 int progress = 0;
2604
2605 VM_BUG_ON(!pages);
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002606 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002607
2608 if (khugepaged_scan.mm_slot)
2609 mm_slot = khugepaged_scan.mm_slot;
2610 else {
2611 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2612 struct mm_slot, mm_node);
2613 khugepaged_scan.address = 0;
2614 khugepaged_scan.mm_slot = mm_slot;
2615 }
2616 spin_unlock(&khugepaged_mm_lock);
2617
2618 mm = mm_slot->mm;
2619 down_read(&mm->mmap_sem);
2620 if (unlikely(khugepaged_test_exit(mm)))
2621 vma = NULL;
2622 else
2623 vma = find_vma(mm, khugepaged_scan.address);
2624
2625 progress++;
2626 for (; vma; vma = vma->vm_next) {
2627 unsigned long hstart, hend;
2628
2629 cond_resched();
2630 if (unlikely(khugepaged_test_exit(mm))) {
2631 progress++;
2632 break;
2633 }
Bob Liufa475e52012-12-11 16:00:39 -08002634 if (!hugepage_vma_check(vma)) {
2635skip:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002636 progress++;
2637 continue;
2638 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002639 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2640 hend = vma->vm_end & HPAGE_PMD_MASK;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002641 if (hstart >= hend)
2642 goto skip;
2643 if (khugepaged_scan.address > hend)
2644 goto skip;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002645 if (khugepaged_scan.address < hstart)
2646 khugepaged_scan.address = hstart;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002647 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002648
2649 while (khugepaged_scan.address < hend) {
2650 int ret;
2651 cond_resched();
2652 if (unlikely(khugepaged_test_exit(mm)))
2653 goto breakouterloop;
2654
2655 VM_BUG_ON(khugepaged_scan.address < hstart ||
2656 khugepaged_scan.address + HPAGE_PMD_SIZE >
2657 hend);
2658 ret = khugepaged_scan_pmd(mm, vma,
2659 khugepaged_scan.address,
2660 hpage);
2661 /* move to next address */
2662 khugepaged_scan.address += HPAGE_PMD_SIZE;
2663 progress += HPAGE_PMD_NR;
2664 if (ret)
2665 /* we released mmap_sem so break loop */
2666 goto breakouterloop_mmap_sem;
2667 if (progress >= pages)
2668 goto breakouterloop;
2669 }
2670 }
2671breakouterloop:
2672 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2673breakouterloop_mmap_sem:
2674
2675 spin_lock(&khugepaged_mm_lock);
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002676 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002677 /*
2678 * Release the current mm_slot if this mm is about to die, or
2679 * if we scanned all vmas of this mm.
2680 */
2681 if (khugepaged_test_exit(mm) || !vma) {
2682 /*
2683 * Make sure that if mm_users is reaching zero while
2684 * khugepaged runs here, khugepaged_exit will find
2685 * mm_slot not pointing to the exiting mm.
2686 */
2687 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2688 khugepaged_scan.mm_slot = list_entry(
2689 mm_slot->mm_node.next,
2690 struct mm_slot, mm_node);
2691 khugepaged_scan.address = 0;
2692 } else {
2693 khugepaged_scan.mm_slot = NULL;
2694 khugepaged_full_scans++;
2695 }
2696
2697 collect_mm_slot(mm_slot);
2698 }
2699
2700 return progress;
2701}
2702
2703static int khugepaged_has_work(void)
2704{
2705 return !list_empty(&khugepaged_scan.mm_head) &&
2706 khugepaged_enabled();
2707}
2708
2709static int khugepaged_wait_event(void)
2710{
2711 return !list_empty(&khugepaged_scan.mm_head) ||
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002712 kthread_should_stop();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002713}
2714
Xiao Guangrongd5169042012-10-08 16:29:48 -07002715static void khugepaged_do_scan(void)
2716{
2717 struct page *hpage = NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002718 unsigned int progress = 0, pass_through_head = 0;
2719 unsigned int pages = khugepaged_pages_to_scan;
Xiao Guangrongd5169042012-10-08 16:29:48 -07002720 bool wait = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002721
2722 barrier(); /* write khugepaged_pages_to_scan to local stack */
2723
2724 while (progress < pages) {
Xiao Guangrong26234f32012-10-08 16:29:51 -07002725 if (!khugepaged_prealloc_page(&hpage, &wait))
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002726 break;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002727
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002728 cond_resched();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002729
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002730 if (unlikely(kthread_should_stop() || freezing(current)))
2731 break;
2732
Andrea Arcangeliba761492011-01-13 15:46:58 -08002733 spin_lock(&khugepaged_mm_lock);
2734 if (!khugepaged_scan.mm_slot)
2735 pass_through_head++;
2736 if (khugepaged_has_work() &&
2737 pass_through_head < 2)
2738 progress += khugepaged_scan_mm_slot(pages - progress,
Xiao Guangrongd5169042012-10-08 16:29:48 -07002739 &hpage);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002740 else
2741 progress = pages;
2742 spin_unlock(&khugepaged_mm_lock);
2743 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002744
Xiao Guangrongd5169042012-10-08 16:29:48 -07002745 if (!IS_ERR_OR_NULL(hpage))
2746 put_page(hpage);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002747}
2748
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002749static void khugepaged_wait_work(void)
2750{
2751 try_to_freeze();
2752
2753 if (khugepaged_has_work()) {
2754 if (!khugepaged_scan_sleep_millisecs)
2755 return;
2756
2757 wait_event_freezable_timeout(khugepaged_wait,
2758 kthread_should_stop(),
2759 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2760 return;
2761 }
2762
2763 if (khugepaged_enabled())
2764 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2765}
2766
Andrea Arcangeliba761492011-01-13 15:46:58 -08002767static int khugepaged(void *none)
2768{
2769 struct mm_slot *mm_slot;
2770
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002771 set_freezable();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002772 set_user_nice(current, 19);
2773
Xiao Guangrongb7231782012-10-08 16:29:54 -07002774 while (!kthread_should_stop()) {
2775 khugepaged_do_scan();
2776 khugepaged_wait_work();
2777 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002778
2779 spin_lock(&khugepaged_mm_lock);
2780 mm_slot = khugepaged_scan.mm_slot;
2781 khugepaged_scan.mm_slot = NULL;
2782 if (mm_slot)
2783 collect_mm_slot(mm_slot);
2784 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002785 return 0;
2786}
2787
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002788static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2789 unsigned long haddr, pmd_t *pmd)
2790{
2791 struct mm_struct *mm = vma->vm_mm;
2792 pgtable_t pgtable;
2793 pmd_t _pmd;
2794 int i;
2795
2796 pmdp_clear_flush(vma, haddr, pmd);
2797 /* leave pmd empty until pte is filled */
2798
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07002799 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002800 pmd_populate(mm, &_pmd, pgtable);
2801
2802 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2803 pte_t *pte, entry;
2804 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2805 entry = pte_mkspecial(entry);
2806 pte = pte_offset_map(&_pmd, haddr);
2807 VM_BUG_ON(!pte_none(*pte));
2808 set_pte_at(mm, haddr, pte, entry);
2809 pte_unmap(pte);
2810 }
2811 smp_wmb(); /* make pte visible before pmd */
2812 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08002813 put_huge_zero_page();
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002814}
2815
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002816void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2817 pmd_t *pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002818{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002819 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002820 struct page *page;
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002821 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002822 unsigned long haddr = address & HPAGE_PMD_MASK;
2823 unsigned long mmun_start; /* For mmu_notifiers */
2824 unsigned long mmun_end; /* For mmu_notifiers */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002825
2826 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002827
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002828 mmun_start = haddr;
2829 mmun_end = haddr + HPAGE_PMD_SIZE;
Hugh Dickins750e8162013-10-16 13:47:08 -07002830again:
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002831 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002832 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002833 if (unlikely(!pmd_trans_huge(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002834 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002835 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2836 return;
2837 }
2838 if (is_huge_zero_pmd(*pmd)) {
2839 __split_huge_zero_page_pmd(vma, haddr, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002840 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002841 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002842 return;
2843 }
2844 page = pmd_page(*pmd);
2845 VM_BUG_ON(!page_count(page));
2846 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002847 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002848 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002849
2850 split_huge_page(page);
2851
2852 put_page(page);
Hugh Dickins750e8162013-10-16 13:47:08 -07002853
2854 /*
2855 * We don't always have down_write of mmap_sem here: a racing
2856 * do_huge_pmd_wp_page() might have copied-on-write to another
2857 * huge page before our split_huge_page() got the anon_vma lock.
2858 */
2859 if (unlikely(pmd_trans_huge(*pmd)))
2860 goto again;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002861}
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002862
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002863void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2864 pmd_t *pmd)
2865{
2866 struct vm_area_struct *vma;
2867
2868 vma = find_vma(mm, address);
2869 BUG_ON(vma == NULL);
2870 split_huge_page_pmd(vma, address, pmd);
2871}
2872
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002873static void split_huge_page_address(struct mm_struct *mm,
2874 unsigned long address)
2875{
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002876 pmd_t *pmd;
2877
2878 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2879
Bob Liu62190492012-12-11 16:00:37 -08002880 pmd = mm_find_pmd(mm, address);
2881 if (!pmd)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002882 return;
2883 /*
2884 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2885 * materialize from under us.
2886 */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002887 split_huge_page_pmd_mm(mm, address, pmd);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002888}
2889
2890void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2891 unsigned long start,
2892 unsigned long end,
2893 long adjust_next)
2894{
2895 /*
2896 * If the new start address isn't hpage aligned and it could
2897 * previously contain an hugepage: check if we need to split
2898 * an huge pmd.
2899 */
2900 if (start & ~HPAGE_PMD_MASK &&
2901 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2902 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2903 split_huge_page_address(vma->vm_mm, start);
2904
2905 /*
2906 * If the new end address isn't hpage aligned and it could
2907 * previously contain an hugepage: check if we need to split
2908 * an huge pmd.
2909 */
2910 if (end & ~HPAGE_PMD_MASK &&
2911 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2912 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2913 split_huge_page_address(vma->vm_mm, end);
2914
2915 /*
2916 * If we're also updating the vma->vm_next->vm_start, if the new
2917 * vm_next->vm_start isn't page aligned and it could previously
2918 * contain an hugepage: check if we need to split an huge pmd.
2919 */
2920 if (adjust_next > 0) {
2921 struct vm_area_struct *next = vma->vm_next;
2922 unsigned long nstart = next->vm_start;
2923 nstart += adjust_next << PAGE_SHIFT;
2924 if (nstart & ~HPAGE_PMD_MASK &&
2925 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2926 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2927 split_huge_page_address(next->vm_mm, nstart);
2928 }
2929}