blob: e60837dc785c4ce35f579374527587c77665ecce [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
Andrew Mortonae3a8c12014-06-04 16:06:58 -07008#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080010#include <linux/mm.h>
11#include <linux/sched.h>
12#include <linux/highmem.h>
13#include <linux/hugetlb.h>
14#include <linux/mmu_notifier.h>
15#include <linux/rmap.h>
16#include <linux/swap.h>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080017#include <linux/shrinker.h>
Andrea Arcangeliba761492011-01-13 15:46:58 -080018#include <linux/mm_inline.h>
19#include <linux/kthread.h>
20#include <linux/khugepaged.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080021#include <linux/freezer.h>
Andrea Arcangelia664b2d2011-01-13 15:47:17 -080022#include <linux/mman.h>
Ralf Baechle325adeb2012-10-15 13:44:56 +020023#include <linux/pagemap.h>
Mel Gorman4daae3b2012-11-02 11:33:45 +000024#include <linux/migrate.h>
Sasha Levin43b5fbb2013-02-22 16:32:27 -080025#include <linux/hashtable.h>
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -080026
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080027#include <asm/tlb.h>
28#include <asm/pgalloc.h>
29#include "internal.h"
30
Andrea Arcangeliba761492011-01-13 15:46:58 -080031/*
Jianguo Wu8bfa3f92013-11-12 15:07:16 -080032 * By default transparent hugepage support is disabled in order that avoid
33 * to risk increase the memory footprint of applications without a guaranteed
34 * benefit. When transparent hugepage support is enabled, is for all mappings,
35 * and khugepaged scans all mappings.
36 * Defrag is invoked by khugepaged hugepage allocations and by page faults
37 * for all hugepage allocations.
Andrea Arcangeliba761492011-01-13 15:46:58 -080038 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -080039unsigned long transparent_hugepage_flags __read_mostly =
Andrea Arcangeli13ece882011-01-13 15:47:07 -080040#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
Andrea Arcangeliba761492011-01-13 15:46:58 -080041 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
Andrea Arcangeli13ece882011-01-13 15:47:07 -080042#endif
43#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
44 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
45#endif
Andrea Arcangelid39d33c2011-01-13 15:47:05 -080046 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
Kirill A. Shutemov79da5402012-12-12 13:51:12 -080047 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
48 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
Andrea Arcangeliba761492011-01-13 15:46:58 -080049
50/* default scan 8*512 pte (or vmas) every 30 second */
51static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
52static unsigned int khugepaged_pages_collapsed;
53static unsigned int khugepaged_full_scans;
54static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
55/* during fragmentation poll the hugepage allocator once every minute */
56static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
57static struct task_struct *khugepaged_thread __read_mostly;
58static DEFINE_MUTEX(khugepaged_mutex);
59static DEFINE_SPINLOCK(khugepaged_mm_lock);
60static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
61/*
62 * default collapse hugepages if there is at least one pte mapped like
63 * it would have happened if the vma was large enough during page
64 * fault.
65 */
66static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
67
68static int khugepaged(void *none);
Andrea Arcangeliba761492011-01-13 15:46:58 -080069static int khugepaged_slab_init(void);
Andrea Arcangeliba761492011-01-13 15:46:58 -080070
Sasha Levin43b5fbb2013-02-22 16:32:27 -080071#define MM_SLOTS_HASH_BITS 10
72static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
73
Andrea Arcangeliba761492011-01-13 15:46:58 -080074static struct kmem_cache *mm_slot_cache __read_mostly;
75
76/**
77 * struct mm_slot - hash lookup from mm to mm_slot
78 * @hash: hash collision list
79 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
80 * @mm: the mm that this information is valid for
81 */
82struct mm_slot {
83 struct hlist_node hash;
84 struct list_head mm_node;
85 struct mm_struct *mm;
86};
87
88/**
89 * struct khugepaged_scan - cursor for scanning
90 * @mm_head: the head of the mm list to scan
91 * @mm_slot: the current mm_slot we are scanning
92 * @address: the next address inside that to be scanned
93 *
94 * There is only the one khugepaged_scan instance of this cursor structure.
95 */
96struct khugepaged_scan {
97 struct list_head mm_head;
98 struct mm_slot *mm_slot;
99 unsigned long address;
H Hartley Sweeten2f1da642011-10-31 17:09:25 -0700100};
101static struct khugepaged_scan khugepaged_scan = {
Andrea Arcangeliba761492011-01-13 15:46:58 -0800102 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
103};
104
Andrea Arcangelif0005652011-01-13 15:47:04 -0800105
106static int set_recommended_min_free_kbytes(void)
107{
108 struct zone *zone;
109 int nr_zones = 0;
110 unsigned long recommended_min;
Andrea Arcangelif0005652011-01-13 15:47:04 -0800111
Xiao Guangrong17c230a2012-10-08 16:29:56 -0700112 if (!khugepaged_enabled())
Andrea Arcangelif0005652011-01-13 15:47:04 -0800113 return 0;
114
115 for_each_populated_zone(zone)
116 nr_zones++;
117
118 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
119 recommended_min = pageblock_nr_pages * nr_zones * 2;
120
121 /*
122 * Make sure that on average at least two pageblocks are almost free
123 * of another type, one for a migratetype to fall back to and a
124 * second to avoid subsequent fallbacks of other types There are 3
125 * MIGRATE_TYPES we care about.
126 */
127 recommended_min += pageblock_nr_pages * nr_zones *
128 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
129
130 /* don't ever allow to reserve more than 5% of the lowmem */
131 recommended_min = min(recommended_min,
132 (unsigned long) nr_free_buffer_pages() / 20);
133 recommended_min <<= (PAGE_SHIFT-10);
134
Han Pingtian42aa83c2014-01-23 15:53:28 -0800135 if (recommended_min > min_free_kbytes) {
136 if (user_min_free_kbytes >= 0)
137 pr_info("raising min_free_kbytes from %d to %lu "
138 "to help transparent hugepage allocations\n",
139 min_free_kbytes, recommended_min);
140
Andrea Arcangelif0005652011-01-13 15:47:04 -0800141 min_free_kbytes = recommended_min;
Han Pingtian42aa83c2014-01-23 15:53:28 -0800142 }
Andrea Arcangelif0005652011-01-13 15:47:04 -0800143 setup_per_zone_wmarks();
144 return 0;
145}
146late_initcall(set_recommended_min_free_kbytes);
147
Andrea Arcangeliba761492011-01-13 15:46:58 -0800148static int start_khugepaged(void)
149{
150 int err = 0;
151 if (khugepaged_enabled()) {
Andrea Arcangeliba761492011-01-13 15:46:58 -0800152 if (!khugepaged_thread)
153 khugepaged_thread = kthread_run(khugepaged, NULL,
154 "khugepaged");
155 if (unlikely(IS_ERR(khugepaged_thread))) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700156 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
Andrea Arcangeliba761492011-01-13 15:46:58 -0800157 err = PTR_ERR(khugepaged_thread);
158 khugepaged_thread = NULL;
159 }
Xiao Guangrong911891a2012-10-08 16:29:41 -0700160
161 if (!list_empty(&khugepaged_scan.mm_head))
Andrea Arcangeliba761492011-01-13 15:46:58 -0800162 wake_up_interruptible(&khugepaged_wait);
Andrea Arcangelif0005652011-01-13 15:47:04 -0800163
164 set_recommended_min_free_kbytes();
Xiao Guangrong911891a2012-10-08 16:29:41 -0700165 } else if (khugepaged_thread) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700166 kthread_stop(khugepaged_thread);
167 khugepaged_thread = NULL;
168 }
Xiao Guangrong637e3a22012-10-08 16:29:38 -0700169
Andrea Arcangeliba761492011-01-13 15:46:58 -0800170 return err;
171}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800172
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800173static atomic_t huge_zero_refcount;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700174static struct page *huge_zero_page __read_mostly;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800175
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700176static inline bool is_huge_zero_page(struct page *page)
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800177{
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700178 return ACCESS_ONCE(huge_zero_page) == page;
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800179}
180
181static inline bool is_huge_zero_pmd(pmd_t pmd)
182{
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700183 return is_huge_zero_page(pmd_page(pmd));
Kirill A. Shutemov4a6c1292012-12-12 13:50:47 -0800184}
185
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700186static struct page *get_huge_zero_page(void)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800187{
188 struct page *zero_page;
189retry:
190 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700191 return ACCESS_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800192
193 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
194 HPAGE_PMD_ORDER);
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800195 if (!zero_page) {
196 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700197 return NULL;
Kirill A. Shutemovd8a8e1f2012-12-12 13:51:09 -0800198 }
199 count_vm_event(THP_ZERO_PAGE_ALLOC);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800200 preempt_disable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700201 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800202 preempt_enable();
203 __free_page(zero_page);
204 goto retry;
205 }
206
207 /* We take additional reference here. It will be put back by shrinker */
208 atomic_set(&huge_zero_refcount, 2);
209 preempt_enable();
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700210 return ACCESS_ONCE(huge_zero_page);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800211}
212
213static void put_huge_zero_page(void)
214{
215 /*
216 * Counter should never go to zero here. Only shrinker can put
217 * last reference.
218 */
219 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
220}
221
Glauber Costa48896462013-08-28 10:18:15 +1000222static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
223 struct shrink_control *sc)
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800224{
Glauber Costa48896462013-08-28 10:18:15 +1000225 /* we can free zero page only if last reference remains */
226 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
227}
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800228
Glauber Costa48896462013-08-28 10:18:15 +1000229static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
230 struct shrink_control *sc)
231{
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800232 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700233 struct page *zero_page = xchg(&huge_zero_page, NULL);
234 BUG_ON(zero_page == NULL);
235 __free_page(zero_page);
Glauber Costa48896462013-08-28 10:18:15 +1000236 return HPAGE_PMD_NR;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800237 }
238
239 return 0;
240}
241
242static struct shrinker huge_zero_page_shrinker = {
Glauber Costa48896462013-08-28 10:18:15 +1000243 .count_objects = shrink_huge_zero_page_count,
244 .scan_objects = shrink_huge_zero_page_scan,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800245 .seeks = DEFAULT_SEEKS,
246};
247
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800248#ifdef CONFIG_SYSFS
Andrea Arcangeliba761492011-01-13 15:46:58 -0800249
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800250static ssize_t double_flag_show(struct kobject *kobj,
251 struct kobj_attribute *attr, char *buf,
252 enum transparent_hugepage_flag enabled,
253 enum transparent_hugepage_flag req_madv)
254{
255 if (test_bit(enabled, &transparent_hugepage_flags)) {
256 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
257 return sprintf(buf, "[always] madvise never\n");
258 } else if (test_bit(req_madv, &transparent_hugepage_flags))
259 return sprintf(buf, "always [madvise] never\n");
260 else
261 return sprintf(buf, "always madvise [never]\n");
262}
263static ssize_t double_flag_store(struct kobject *kobj,
264 struct kobj_attribute *attr,
265 const char *buf, size_t count,
266 enum transparent_hugepage_flag enabled,
267 enum transparent_hugepage_flag req_madv)
268{
269 if (!memcmp("always", buf,
270 min(sizeof("always")-1, count))) {
271 set_bit(enabled, &transparent_hugepage_flags);
272 clear_bit(req_madv, &transparent_hugepage_flags);
273 } else if (!memcmp("madvise", buf,
274 min(sizeof("madvise")-1, count))) {
275 clear_bit(enabled, &transparent_hugepage_flags);
276 set_bit(req_madv, &transparent_hugepage_flags);
277 } else if (!memcmp("never", buf,
278 min(sizeof("never")-1, count))) {
279 clear_bit(enabled, &transparent_hugepage_flags);
280 clear_bit(req_madv, &transparent_hugepage_flags);
281 } else
282 return -EINVAL;
283
284 return count;
285}
286
287static ssize_t enabled_show(struct kobject *kobj,
288 struct kobj_attribute *attr, char *buf)
289{
290 return double_flag_show(kobj, attr, buf,
291 TRANSPARENT_HUGEPAGE_FLAG,
292 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
293}
294static ssize_t enabled_store(struct kobject *kobj,
295 struct kobj_attribute *attr,
296 const char *buf, size_t count)
297{
Andrea Arcangeliba761492011-01-13 15:46:58 -0800298 ssize_t ret;
299
300 ret = double_flag_store(kobj, attr, buf, count,
301 TRANSPARENT_HUGEPAGE_FLAG,
302 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
303
304 if (ret > 0) {
Xiao Guangrong911891a2012-10-08 16:29:41 -0700305 int err;
306
307 mutex_lock(&khugepaged_mutex);
308 err = start_khugepaged();
309 mutex_unlock(&khugepaged_mutex);
310
Andrea Arcangeliba761492011-01-13 15:46:58 -0800311 if (err)
312 ret = err;
313 }
314
315 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800316}
317static struct kobj_attribute enabled_attr =
318 __ATTR(enabled, 0644, enabled_show, enabled_store);
319
320static ssize_t single_flag_show(struct kobject *kobj,
321 struct kobj_attribute *attr, char *buf,
322 enum transparent_hugepage_flag flag)
323{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700324 return sprintf(buf, "%d\n",
325 !!test_bit(flag, &transparent_hugepage_flags));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800326}
Ben Hutchingse27e6152011-04-14 15:22:21 -0700327
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800328static ssize_t single_flag_store(struct kobject *kobj,
329 struct kobj_attribute *attr,
330 const char *buf, size_t count,
331 enum transparent_hugepage_flag flag)
332{
Ben Hutchingse27e6152011-04-14 15:22:21 -0700333 unsigned long value;
334 int ret;
335
336 ret = kstrtoul(buf, 10, &value);
337 if (ret < 0)
338 return ret;
339 if (value > 1)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800340 return -EINVAL;
341
Ben Hutchingse27e6152011-04-14 15:22:21 -0700342 if (value)
343 set_bit(flag, &transparent_hugepage_flags);
344 else
345 clear_bit(flag, &transparent_hugepage_flags);
346
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800347 return count;
348}
349
350/*
351 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
352 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
353 * memory just to allocate one more hugepage.
354 */
355static ssize_t defrag_show(struct kobject *kobj,
356 struct kobj_attribute *attr, char *buf)
357{
358 return double_flag_show(kobj, attr, buf,
359 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
360 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
361}
362static ssize_t defrag_store(struct kobject *kobj,
363 struct kobj_attribute *attr,
364 const char *buf, size_t count)
365{
366 return double_flag_store(kobj, attr, buf, count,
367 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
368 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
369}
370static struct kobj_attribute defrag_attr =
371 __ATTR(defrag, 0644, defrag_show, defrag_store);
372
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800373static ssize_t use_zero_page_show(struct kobject *kobj,
374 struct kobj_attribute *attr, char *buf)
375{
376 return single_flag_show(kobj, attr, buf,
377 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
378}
379static ssize_t use_zero_page_store(struct kobject *kobj,
380 struct kobj_attribute *attr, const char *buf, size_t count)
381{
382 return single_flag_store(kobj, attr, buf, count,
383 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
384}
385static struct kobj_attribute use_zero_page_attr =
386 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800387#ifdef CONFIG_DEBUG_VM
388static ssize_t debug_cow_show(struct kobject *kobj,
389 struct kobj_attribute *attr, char *buf)
390{
391 return single_flag_show(kobj, attr, buf,
392 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
393}
394static ssize_t debug_cow_store(struct kobject *kobj,
395 struct kobj_attribute *attr,
396 const char *buf, size_t count)
397{
398 return single_flag_store(kobj, attr, buf, count,
399 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
400}
401static struct kobj_attribute debug_cow_attr =
402 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
403#endif /* CONFIG_DEBUG_VM */
404
405static struct attribute *hugepage_attr[] = {
406 &enabled_attr.attr,
407 &defrag_attr.attr,
Kirill A. Shutemov79da5402012-12-12 13:51:12 -0800408 &use_zero_page_attr.attr,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800409#ifdef CONFIG_DEBUG_VM
410 &debug_cow_attr.attr,
411#endif
412 NULL,
413};
414
415static struct attribute_group hugepage_attr_group = {
416 .attrs = hugepage_attr,
Andrea Arcangeliba761492011-01-13 15:46:58 -0800417};
418
419static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
420 struct kobj_attribute *attr,
421 char *buf)
422{
423 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
424}
425
426static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
427 struct kobj_attribute *attr,
428 const char *buf, size_t count)
429{
430 unsigned long msecs;
431 int err;
432
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700433 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800434 if (err || msecs > UINT_MAX)
435 return -EINVAL;
436
437 khugepaged_scan_sleep_millisecs = msecs;
438 wake_up_interruptible(&khugepaged_wait);
439
440 return count;
441}
442static struct kobj_attribute scan_sleep_millisecs_attr =
443 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
444 scan_sleep_millisecs_store);
445
446static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
447 struct kobj_attribute *attr,
448 char *buf)
449{
450 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
451}
452
453static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
454 struct kobj_attribute *attr,
455 const char *buf, size_t count)
456{
457 unsigned long msecs;
458 int err;
459
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700460 err = kstrtoul(buf, 10, &msecs);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800461 if (err || msecs > UINT_MAX)
462 return -EINVAL;
463
464 khugepaged_alloc_sleep_millisecs = msecs;
465 wake_up_interruptible(&khugepaged_wait);
466
467 return count;
468}
469static struct kobj_attribute alloc_sleep_millisecs_attr =
470 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
471 alloc_sleep_millisecs_store);
472
473static ssize_t pages_to_scan_show(struct kobject *kobj,
474 struct kobj_attribute *attr,
475 char *buf)
476{
477 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
478}
479static ssize_t pages_to_scan_store(struct kobject *kobj,
480 struct kobj_attribute *attr,
481 const char *buf, size_t count)
482{
483 int err;
484 unsigned long pages;
485
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700486 err = kstrtoul(buf, 10, &pages);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800487 if (err || !pages || pages > UINT_MAX)
488 return -EINVAL;
489
490 khugepaged_pages_to_scan = pages;
491
492 return count;
493}
494static struct kobj_attribute pages_to_scan_attr =
495 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
496 pages_to_scan_store);
497
498static ssize_t pages_collapsed_show(struct kobject *kobj,
499 struct kobj_attribute *attr,
500 char *buf)
501{
502 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
503}
504static struct kobj_attribute pages_collapsed_attr =
505 __ATTR_RO(pages_collapsed);
506
507static ssize_t full_scans_show(struct kobject *kobj,
508 struct kobj_attribute *attr,
509 char *buf)
510{
511 return sprintf(buf, "%u\n", khugepaged_full_scans);
512}
513static struct kobj_attribute full_scans_attr =
514 __ATTR_RO(full_scans);
515
516static ssize_t khugepaged_defrag_show(struct kobject *kobj,
517 struct kobj_attribute *attr, char *buf)
518{
519 return single_flag_show(kobj, attr, buf,
520 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
521}
522static ssize_t khugepaged_defrag_store(struct kobject *kobj,
523 struct kobj_attribute *attr,
524 const char *buf, size_t count)
525{
526 return single_flag_store(kobj, attr, buf, count,
527 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
528}
529static struct kobj_attribute khugepaged_defrag_attr =
530 __ATTR(defrag, 0644, khugepaged_defrag_show,
531 khugepaged_defrag_store);
532
533/*
534 * max_ptes_none controls if khugepaged should collapse hugepages over
535 * any unmapped ptes in turn potentially increasing the memory
536 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
537 * reduce the available free memory in the system as it
538 * runs. Increasing max_ptes_none will instead potentially reduce the
539 * free memory in the system during the khugepaged scan.
540 */
541static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
542 struct kobj_attribute *attr,
543 char *buf)
544{
545 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
546}
547static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
548 struct kobj_attribute *attr,
549 const char *buf, size_t count)
550{
551 int err;
552 unsigned long max_ptes_none;
553
Jingoo Han3dbb95f2013-09-11 14:20:25 -0700554 err = kstrtoul(buf, 10, &max_ptes_none);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800555 if (err || max_ptes_none > HPAGE_PMD_NR-1)
556 return -EINVAL;
557
558 khugepaged_max_ptes_none = max_ptes_none;
559
560 return count;
561}
562static struct kobj_attribute khugepaged_max_ptes_none_attr =
563 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
564 khugepaged_max_ptes_none_store);
565
566static struct attribute *khugepaged_attr[] = {
567 &khugepaged_defrag_attr.attr,
568 &khugepaged_max_ptes_none_attr.attr,
569 &pages_to_scan_attr.attr,
570 &pages_collapsed_attr.attr,
571 &full_scans_attr.attr,
572 &scan_sleep_millisecs_attr.attr,
573 &alloc_sleep_millisecs_attr.attr,
574 NULL,
575};
576
577static struct attribute_group khugepaged_attr_group = {
578 .attrs = khugepaged_attr,
579 .name = "khugepaged",
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800580};
Shaohua Li569e5592012-01-12 17:19:11 -0800581
582static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
583{
584 int err;
585
586 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
587 if (unlikely(!*hugepage_kobj)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700588 pr_err("failed to create transparent hugepage kobject\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800589 return -ENOMEM;
590 }
591
592 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
593 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700594 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800595 goto delete_obj;
596 }
597
598 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
599 if (err) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700600 pr_err("failed to register transparent hugepage group\n");
Shaohua Li569e5592012-01-12 17:19:11 -0800601 goto remove_hp_group;
602 }
603
604 return 0;
605
606remove_hp_group:
607 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
608delete_obj:
609 kobject_put(*hugepage_kobj);
610 return err;
611}
612
613static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
614{
615 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
616 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
617 kobject_put(hugepage_kobj);
618}
619#else
620static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
621{
622 return 0;
623}
624
625static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
626{
627}
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800628#endif /* CONFIG_SYSFS */
629
630static int __init hugepage_init(void)
631{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800632 int err;
Shaohua Li569e5592012-01-12 17:19:11 -0800633 struct kobject *hugepage_kobj;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800634
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800635 if (!has_transparent_hugepage()) {
636 transparent_hugepage_flags = 0;
Shaohua Li569e5592012-01-12 17:19:11 -0800637 return -EINVAL;
Andrea Arcangeli4b7167b2011-01-13 15:47:09 -0800638 }
639
Shaohua Li569e5592012-01-12 17:19:11 -0800640 err = hugepage_init_sysfs(&hugepage_kobj);
641 if (err)
642 return err;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800643
644 err = khugepaged_slab_init();
645 if (err)
646 goto out;
647
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800648 register_shrinker(&huge_zero_page_shrinker);
649
Rik van Riel97562cd2011-01-13 15:47:12 -0800650 /*
651 * By default disable transparent hugepages on smaller systems,
652 * where the extra memory used could hurt more than TLB overhead
653 * is likely to save. The admin can still enable it through /sys.
654 */
655 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
656 transparent_hugepage_flags = 0;
657
Andrea Arcangeliba761492011-01-13 15:46:58 -0800658 start_khugepaged();
659
Shaohua Li569e5592012-01-12 17:19:11 -0800660 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -0800661out:
Shaohua Li569e5592012-01-12 17:19:11 -0800662 hugepage_exit_sysfs(hugepage_kobj);
Andrea Arcangeliba761492011-01-13 15:46:58 -0800663 return err;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800664}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800665subsys_initcall(hugepage_init);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800666
667static int __init setup_transparent_hugepage(char *str)
668{
669 int ret = 0;
670 if (!str)
671 goto out;
672 if (!strcmp(str, "always")) {
673 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
674 &transparent_hugepage_flags);
675 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
676 &transparent_hugepage_flags);
677 ret = 1;
678 } else if (!strcmp(str, "madvise")) {
679 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
680 &transparent_hugepage_flags);
681 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
682 &transparent_hugepage_flags);
683 ret = 1;
684 } else if (!strcmp(str, "never")) {
685 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
686 &transparent_hugepage_flags);
687 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
688 &transparent_hugepage_flags);
689 ret = 1;
690 }
691out:
692 if (!ret)
Andrew Mortonae3a8c12014-06-04 16:06:58 -0700693 pr_warn("transparent_hugepage= cannot parse, ignored\n");
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800694 return ret;
695}
696__setup("transparent_hugepage=", setup_transparent_hugepage);
697
Mel Gormanb32967f2012-11-19 12:35:47 +0000698pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800699{
700 if (likely(vma->vm_flags & VM_WRITE))
701 pmd = pmd_mkwrite(pmd);
702 return pmd;
703}
704
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700705static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
Bob Liub3092b32012-12-11 16:00:41 -0800706{
707 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700708 entry = mk_pmd(page, prot);
Bob Liub3092b32012-12-11 16:00:41 -0800709 entry = pmd_mkhuge(entry);
710 return entry;
711}
712
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800713static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
714 struct vm_area_struct *vma,
715 unsigned long haddr, pmd_t *pmd,
716 struct page *page)
717{
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800718 pgtable_t pgtable;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800719 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800720
Sasha Levin309381fea2014-01-23 15:52:54 -0800721 VM_BUG_ON_PAGE(!PageCompound(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800722 pgtable = pte_alloc_one(mm, haddr);
David Rientjesedad9d22012-05-29 15:06:17 -0700723 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800724 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800725
726 clear_huge_page(page, haddr, HPAGE_PMD_NR);
Minchan Kim52f37622013-04-29 15:08:15 -0700727 /*
728 * The memory barrier inside __SetPageUptodate makes sure that
729 * clear_huge_page writes become visible before the set_pmd_at()
730 * write.
731 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800732 __SetPageUptodate(page);
733
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800734 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800735 if (unlikely(!pmd_none(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800736 spin_unlock(ptl);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800737 mem_cgroup_uncharge_page(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800738 put_page(page);
739 pte_free(mm, pgtable);
740 } else {
741 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -0700742 entry = mk_huge_pmd(page, vma->vm_page_prot);
743 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800744 page_add_new_anon_rmap(page, vma, haddr);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700745 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800746 set_pmd_at(mm, haddr, pmd, entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800747 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800748 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800749 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800750 }
751
David Rientjesaa2e8782012-05-29 15:06:17 -0700752 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800753}
754
Andi Kleencc5d4622011-03-22 16:33:13 -0700755static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800756{
Andi Kleencc5d4622011-03-22 16:33:13 -0700757 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800758}
759
760static inline struct page *alloc_hugepage_vma(int defrag,
761 struct vm_area_struct *vma,
Andi Kleencc5d4622011-03-22 16:33:13 -0700762 unsigned long haddr, int nd,
763 gfp_t extra_gfp)
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800764{
Andi Kleencc5d4622011-03-22 16:33:13 -0700765 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
Andi Kleen5c4b4be2011-03-04 17:36:32 -0800766 HPAGE_PMD_ORDER, vma, haddr, nd);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -0800767}
768
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800769/* Caller must hold page table lock. */
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800770static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800771 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700772 struct page *zero_page)
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800773{
774 pmd_t entry;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800775 if (!pmd_none(*pmd))
776 return false;
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700777 entry = mk_pmd(zero_page, vma->vm_page_prot);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800778 entry = pmd_wrprotect(entry);
779 entry = pmd_mkhuge(entry);
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700780 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800781 set_pmd_at(mm, haddr, pmd, entry);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800782 atomic_long_inc(&mm->nr_ptes);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800783 return true;
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800784}
785
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800786int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
787 unsigned long address, pmd_t *pmd,
788 unsigned int flags)
789{
790 struct page *page;
791 unsigned long haddr = address & HPAGE_PMD_MASK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800792
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700793 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700794 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700795 if (unlikely(anon_vma_prepare(vma)))
796 return VM_FAULT_OOM;
797 if (unlikely(khugepaged_enter(vma)))
798 return VM_FAULT_OOM;
799 if (!(flags & FAULT_FLAG_WRITE) &&
800 transparent_hugepage_use_zero_page()) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800801 spinlock_t *ptl;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700802 pgtable_t pgtable;
803 struct page *zero_page;
804 bool set;
805 pgtable = pte_alloc_one(mm, haddr);
806 if (unlikely(!pgtable))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800807 return VM_FAULT_OOM;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700808 zero_page = get_huge_zero_page();
809 if (unlikely(!zero_page)) {
810 pte_free(mm, pgtable);
Andi Kleen81ab4202011-04-14 15:22:06 -0700811 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700812 return VM_FAULT_FALLBACK;
Andi Kleen81ab4202011-04-14 15:22:06 -0700813 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800814 ptl = pmd_lock(mm, pmd);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700815 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
816 zero_page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800817 spin_unlock(ptl);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700818 if (!set) {
819 pte_free(mm, pgtable);
820 put_huge_zero_page();
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800821 }
David Rientjesedad9d22012-05-29 15:06:17 -0700822 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800823 }
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700824 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
825 vma, haddr, numa_node_id(), 0);
826 if (unlikely(!page)) {
827 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700828 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700829 }
Michal Hockod715ae02014-04-07 15:37:46 -0700830 if (unlikely(mem_cgroup_charge_anon(page, mm, GFP_KERNEL))) {
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700831 put_page(page);
David Rientjes17766dd2013-09-12 15:14:06 -0700832 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700833 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700834 }
835 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
836 mem_cgroup_uncharge_page(page);
837 put_page(page);
David Rientjes17766dd2013-09-12 15:14:06 -0700838 count_vm_event(THP_FAULT_FALLBACK);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -0700839 return VM_FAULT_FALLBACK;
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700840 }
841
David Rientjes17766dd2013-09-12 15:14:06 -0700842 count_vm_event(THP_FAULT_ALLOC);
Kirill A. Shutemov128ec032013-09-12 15:14:03 -0700843 return 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800844}
845
846int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
847 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
848 struct vm_area_struct *vma)
849{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800850 spinlock_t *dst_ptl, *src_ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800851 struct page *src_page;
852 pmd_t pmd;
853 pgtable_t pgtable;
854 int ret;
855
856 ret = -ENOMEM;
857 pgtable = pte_alloc_one(dst_mm, addr);
858 if (unlikely(!pgtable))
859 goto out;
860
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800861 dst_ptl = pmd_lock(dst_mm, dst_pmd);
862 src_ptl = pmd_lockptr(src_mm, src_pmd);
863 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800864
865 ret = -EAGAIN;
866 pmd = *src_pmd;
867 if (unlikely(!pmd_trans_huge(pmd))) {
868 pte_free(dst_mm, pgtable);
869 goto out_unlock;
870 }
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800871 /*
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800872 * When page table lock is held, the huge zero pmd should not be
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800873 * under splitting since we don't split the page itself, only pmd to
874 * a page table.
875 */
876 if (is_huge_zero_pmd(pmd)) {
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700877 struct page *zero_page;
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800878 bool set;
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -0800879 /*
880 * get_huge_zero_page() will never allocate a new page here,
881 * since we already have a zero page to copy. It just takes a
882 * reference.
883 */
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700884 zero_page = get_huge_zero_page();
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800885 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
Kirill A. Shutemov5918d102013-04-29 15:08:44 -0700886 zero_page);
Kirill A. Shutemov3ea41e62012-12-12 13:51:14 -0800887 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
Kirill A. Shutemovfc9fe822012-12-12 13:50:51 -0800888 ret = 0;
889 goto out_unlock;
890 }
Mel Gormande466bd2013-12-18 17:08:42 -0800891
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800892 if (unlikely(pmd_trans_splitting(pmd))) {
893 /* split huge page running from under us */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800894 spin_unlock(src_ptl);
895 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800896 pte_free(dst_mm, pgtable);
897
898 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
899 goto out;
900 }
901 src_page = pmd_page(pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -0800902 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800903 get_page(src_page);
904 page_dup_rmap(src_page);
905 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
906
907 pmdp_set_wrprotect(src_mm, addr, src_pmd);
908 pmd = pmd_mkold(pmd_wrprotect(pmd));
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -0700909 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800910 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800911 atomic_long_inc(&dst_mm->nr_ptes);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800912
913 ret = 0;
914out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800915 spin_unlock(src_ptl);
916 spin_unlock(dst_ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800917out:
918 return ret;
919}
920
Will Deacona1dd4502012-12-11 16:01:27 -0800921void huge_pmd_set_accessed(struct mm_struct *mm,
922 struct vm_area_struct *vma,
923 unsigned long address,
924 pmd_t *pmd, pmd_t orig_pmd,
925 int dirty)
926{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800927 spinlock_t *ptl;
Will Deacona1dd4502012-12-11 16:01:27 -0800928 pmd_t entry;
929 unsigned long haddr;
930
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800931 ptl = pmd_lock(mm, pmd);
Will Deacona1dd4502012-12-11 16:01:27 -0800932 if (unlikely(!pmd_same(*pmd, orig_pmd)))
933 goto unlock;
934
935 entry = pmd_mkyoung(orig_pmd);
936 haddr = address & HPAGE_PMD_MASK;
937 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
938 update_mmu_cache_pmd(vma, address, pmd);
939
940unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800941 spin_unlock(ptl);
Will Deacona1dd4502012-12-11 16:01:27 -0800942}
943
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800944static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
945 struct vm_area_struct *vma,
946 unsigned long address,
947 pmd_t *pmd, pmd_t orig_pmd,
948 struct page *page,
949 unsigned long haddr)
950{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800951 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800952 pgtable_t pgtable;
953 pmd_t _pmd;
954 int ret = 0, i;
955 struct page **pages;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700956 unsigned long mmun_start; /* For mmu_notifiers */
957 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800958
959 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
960 GFP_KERNEL);
961 if (unlikely(!pages)) {
962 ret |= VM_FAULT_OOM;
963 goto out;
964 }
965
966 for (i = 0; i < HPAGE_PMD_NR; i++) {
Andi Kleencc5d4622011-03-22 16:33:13 -0700967 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
968 __GFP_OTHER_NODE,
Andi Kleen19ee1512011-03-04 17:36:31 -0800969 vma, address, page_to_nid(page));
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800970 if (unlikely(!pages[i] ||
Michal Hockod715ae02014-04-07 15:37:46 -0700971 mem_cgroup_charge_anon(pages[i], mm,
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800972 GFP_KERNEL))) {
973 if (pages[i])
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800974 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -0800975 mem_cgroup_uncharge_start();
976 while (--i >= 0) {
977 mem_cgroup_uncharge_page(pages[i]);
978 put_page(pages[i]);
979 }
980 mem_cgroup_uncharge_end();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800981 kfree(pages);
982 ret |= VM_FAULT_OOM;
983 goto out;
984 }
985 }
986
987 for (i = 0; i < HPAGE_PMD_NR; i++) {
988 copy_user_highpage(pages[i], page + i,
Hillf Danton0089e482011-10-31 17:09:38 -0700989 haddr + PAGE_SIZE * i, vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800990 __SetPageUptodate(pages[i]);
991 cond_resched();
992 }
993
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700994 mmun_start = haddr;
995 mmun_end = haddr + HPAGE_PMD_SIZE;
996 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
997
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800998 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800999 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1000 goto out_free_pages;
Sasha Levin309381fea2014-01-23 15:52:54 -08001001 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001002
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001003 pmdp_clear_flush(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001004 /* leave pmd empty until pte is filled */
1005
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001006 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001007 pmd_populate(mm, &_pmd, pgtable);
1008
1009 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1010 pte_t *pte, entry;
1011 entry = mk_pte(pages[i], vma->vm_page_prot);
1012 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1013 page_add_new_anon_rmap(pages[i], vma, haddr);
1014 pte = pte_offset_map(&_pmd, haddr);
1015 VM_BUG_ON(!pte_none(*pte));
1016 set_pte_at(mm, haddr, pte, entry);
1017 pte_unmap(pte);
1018 }
1019 kfree(pages);
1020
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001021 smp_wmb(); /* make pte visible before pmd */
1022 pmd_populate(mm, pmd, pgtable);
1023 page_remove_rmap(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001024 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001025
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001026 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1027
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001028 ret |= VM_FAULT_WRITE;
1029 put_page(page);
1030
1031out:
1032 return ret;
1033
1034out_free_pages:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001035 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001036 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001037 mem_cgroup_uncharge_start();
1038 for (i = 0; i < HPAGE_PMD_NR; i++) {
1039 mem_cgroup_uncharge_page(pages[i]);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001040 put_page(pages[i]);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001041 }
1042 mem_cgroup_uncharge_end();
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001043 kfree(pages);
1044 goto out;
1045}
1046
1047int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1048 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1049{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001050 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001051 int ret = 0;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001052 struct page *page = NULL, *new_page;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001053 unsigned long haddr;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001054 unsigned long mmun_start; /* For mmu_notifiers */
1055 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001056
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001057 ptl = pmd_lockptr(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001058 VM_BUG_ON(!vma->anon_vma);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001059 haddr = address & HPAGE_PMD_MASK;
1060 if (is_huge_zero_pmd(orig_pmd))
1061 goto alloc;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001062 spin_lock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001063 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1064 goto out_unlock;
1065
1066 page = pmd_page(orig_pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001067 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001068 if (page_mapcount(page) == 1) {
1069 pmd_t entry;
1070 entry = pmd_mkyoung(orig_pmd);
1071 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1072 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
David Millerb113da62012-10-08 16:34:25 -07001073 update_mmu_cache_pmd(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001074 ret |= VM_FAULT_WRITE;
1075 goto out_unlock;
1076 }
1077 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001078 spin_unlock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001079alloc:
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001080 if (transparent_hugepage_enabled(vma) &&
1081 !transparent_hugepage_debug_cow())
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001082 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
Andi Kleencc5d4622011-03-22 16:33:13 -07001083 vma, haddr, numa_node_id(), 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001084 else
1085 new_page = NULL;
1086
1087 if (unlikely(!new_page)) {
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001088 if (!page) {
Kirill A. Shutemove9b71ca2014-04-03 14:48:17 -07001089 split_huge_page_pmd(vma, address, pmd);
1090 ret |= VM_FAULT_FALLBACK;
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001091 } else {
1092 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1093 pmd, orig_pmd, page, haddr);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001094 if (ret & VM_FAULT_OOM) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001095 split_huge_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001096 ret |= VM_FAULT_FALLBACK;
1097 }
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001098 put_page(page);
1099 }
David Rientjes17766dd2013-09-12 15:14:06 -07001100 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001101 goto out;
1102 }
1103
Michal Hockod715ae02014-04-07 15:37:46 -07001104 if (unlikely(mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL))) {
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001105 put_page(new_page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001106 if (page) {
1107 split_huge_page(page);
1108 put_page(page);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08001109 } else
1110 split_huge_page_pmd(vma, address, pmd);
1111 ret |= VM_FAULT_FALLBACK;
David Rientjes17766dd2013-09-12 15:14:06 -07001112 count_vm_event(THP_FAULT_FALLBACK);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001113 goto out;
1114 }
1115
David Rientjes17766dd2013-09-12 15:14:06 -07001116 count_vm_event(THP_FAULT_ALLOC);
1117
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001118 if (!page)
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001119 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1120 else
1121 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001122 __SetPageUptodate(new_page);
1123
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001124 mmun_start = haddr;
1125 mmun_end = haddr + HPAGE_PMD_SIZE;
1126 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1127
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001128 spin_lock(ptl);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001129 if (page)
1130 put_page(page);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001131 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001132 spin_unlock(ptl);
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001133 mem_cgroup_uncharge_page(new_page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001134 put_page(new_page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001135 goto out_mn;
Andrea Arcangelib9bbfbe2011-01-13 15:46:57 -08001136 } else {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001137 pmd_t entry;
Kirill A. Shutemov31223592013-09-12 15:14:01 -07001138 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1139 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001140 pmdp_clear_flush(vma, haddr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001141 page_add_new_anon_rmap(new_page, vma, haddr);
1142 set_pmd_at(mm, haddr, pmd, entry);
David Millerb113da62012-10-08 16:34:25 -07001143 update_mmu_cache_pmd(vma, address, pmd);
Hugh Dickinseecc1e42014-01-12 01:25:21 -08001144 if (!page) {
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001145 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001146 put_huge_zero_page();
1147 } else {
Sasha Levin309381fea2014-01-23 15:52:54 -08001148 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemov93b47962012-12-12 13:50:54 -08001149 page_remove_rmap(page);
1150 put_page(page);
1151 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001152 ret |= VM_FAULT_WRITE;
1153 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001154 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001155out_mn:
1156 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1157out:
1158 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001159out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001160 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001161 return ret;
1162}
1163
David Rientjesb676b292012-10-08 16:34:03 -07001164struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001165 unsigned long addr,
1166 pmd_t *pmd,
1167 unsigned int flags)
1168{
David Rientjesb676b292012-10-08 16:34:03 -07001169 struct mm_struct *mm = vma->vm_mm;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001170 struct page *page = NULL;
1171
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001172 assert_spin_locked(pmd_lockptr(mm, pmd));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001173
1174 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1175 goto out;
1176
Kirill A. Shutemov85facf22013-02-04 14:28:42 -08001177 /* Avoid dumping huge zero page */
1178 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1179 return ERR_PTR(-EFAULT);
1180
Mel Gorman2b4847e2013-12-18 17:08:32 -08001181 /* Full NUMA hinting faults to serialise migration in fault paths */
1182 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1183 goto out;
1184
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001185 page = pmd_page(*pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08001186 VM_BUG_ON_PAGE(!PageHead(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001187 if (flags & FOLL_TOUCH) {
1188 pmd_t _pmd;
1189 /*
1190 * We should set the dirty bit only for FOLL_WRITE but
1191 * for now the dirty bit in the pmd is meaningless.
1192 * And if the dirty bit will become meaningful and
1193 * we'll only set it with FOLL_WRITE, an atomic
1194 * set_bit will be required on the pmd to set the
1195 * young bit, instead of the current set_pmd_at.
1196 */
1197 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
Aneesh Kumar K.V8663890a2013-06-06 00:20:34 -07001198 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1199 pmd, _pmd, 1))
1200 update_mmu_cache_pmd(vma, addr, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001201 }
David Rientjesb676b292012-10-08 16:34:03 -07001202 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1203 if (page->mapping && trylock_page(page)) {
1204 lru_add_drain();
1205 if (page->mapping)
1206 mlock_vma_page(page);
1207 unlock_page(page);
1208 }
1209 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001210 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
Sasha Levin309381fea2014-01-23 15:52:54 -08001211 VM_BUG_ON_PAGE(!PageCompound(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001212 if (flags & FOLL_GET)
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001213 get_page_foll(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001214
1215out:
1216 return page;
1217}
1218
Mel Gormand10e63f2012-10-25 14:16:31 +02001219/* NUMA hinting page fault entry point for trans huge pmds */
Mel Gorman4daae3b2012-11-02 11:33:45 +00001220int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1221 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
Mel Gormand10e63f2012-10-25 14:16:31 +02001222{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001223 spinlock_t *ptl;
Mel Gormanb8916632013-10-07 11:28:44 +01001224 struct anon_vma *anon_vma = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001225 struct page *page;
Mel Gormand10e63f2012-10-25 14:16:31 +02001226 unsigned long haddr = addr & HPAGE_PMD_MASK;
Mel Gorman8191acb2013-10-07 11:28:45 +01001227 int page_nid = -1, this_nid = numa_node_id();
Peter Zijlstra90572892013-10-07 11:29:20 +01001228 int target_nid, last_cpupid = -1;
Mel Gorman8191acb2013-10-07 11:28:45 +01001229 bool page_locked;
1230 bool migrated = false;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001231 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02001232
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001233 ptl = pmd_lock(mm, pmdp);
Mel Gormand10e63f2012-10-25 14:16:31 +02001234 if (unlikely(!pmd_same(pmd, *pmdp)))
1235 goto out_unlock;
1236
Mel Gormande466bd2013-12-18 17:08:42 -08001237 /*
1238 * If there are potential migrations, wait for completion and retry
1239 * without disrupting NUMA hinting information. Do not relock and
1240 * check_same as the page may no longer be mapped.
1241 */
1242 if (unlikely(pmd_trans_migrating(*pmdp))) {
1243 spin_unlock(ptl);
1244 wait_migrate_huge_page(vma->anon_vma, pmdp);
1245 goto out;
1246 }
1247
Mel Gormand10e63f2012-10-25 14:16:31 +02001248 page = pmd_page(pmd);
Mel Gormana1a46182013-10-07 11:28:50 +01001249 BUG_ON(is_huge_zero_page(page));
Mel Gorman8191acb2013-10-07 11:28:45 +01001250 page_nid = page_to_nid(page);
Peter Zijlstra90572892013-10-07 11:29:20 +01001251 last_cpupid = page_cpupid_last(page);
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001252 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001253 if (page_nid == this_nid) {
Mel Gorman03c5a6e2012-11-02 14:52:48 +00001254 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001255 flags |= TNF_FAULT_LOCAL;
1256 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001257
Mel Gormanff9042b2013-10-07 11:28:43 +01001258 /*
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001259 * Avoid grouping on DSO/COW pages in specific and RO pages
1260 * in general, RO pages shouldn't hurt as much anyway since
1261 * they can be in shared cache state.
1262 */
1263 if (!pmd_write(pmd))
1264 flags |= TNF_NO_GROUP;
1265
1266 /*
Mel Gormanff9042b2013-10-07 11:28:43 +01001267 * Acquire the page lock to serialise THP migrations but avoid dropping
1268 * page_table_lock if at all possible
1269 */
Mel Gormanb8916632013-10-07 11:28:44 +01001270 page_locked = trylock_page(page);
1271 target_nid = mpol_misplaced(page, vma, haddr);
1272 if (target_nid == -1) {
1273 /* If the page was locked, there are no parallel migrations */
Mel Gormana54a4072013-10-07 11:28:46 +01001274 if (page_locked)
Mel Gormanb8916632013-10-07 11:28:44 +01001275 goto clear_pmdnuma;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001276 }
Mel Gorman4daae3b2012-11-02 11:33:45 +00001277
Mel Gormande466bd2013-12-18 17:08:42 -08001278 /* Migration could have started since the pmd_trans_migrating check */
Mel Gorman2b4847e2013-12-18 17:08:32 -08001279 if (!page_locked) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001280 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001281 wait_on_page_locked(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001282 page_nid = -1;
Mel Gormanb8916632013-10-07 11:28:44 +01001283 goto out;
1284 }
1285
Mel Gorman2b4847e2013-12-18 17:08:32 -08001286 /*
1287 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1288 * to serialises splits
1289 */
Mel Gormanb8916632013-10-07 11:28:44 +01001290 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001291 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001292 anon_vma = page_lock_anon_vma_read(page);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001293
Peter Zijlstrac69307d2013-10-07 11:28:41 +01001294 /* Confirm the PMD did not change while page_table_lock was released */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001295 spin_lock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001296 if (unlikely(!pmd_same(pmd, *pmdp))) {
1297 unlock_page(page);
1298 put_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001299 page_nid = -1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001300 goto out_unlock;
1301 }
Mel Gormanff9042b2013-10-07 11:28:43 +01001302
Mel Gormanc3a489c2013-12-18 17:08:38 -08001303 /* Bail if we fail to protect against THP splits for any reason */
1304 if (unlikely(!anon_vma)) {
1305 put_page(page);
1306 page_nid = -1;
1307 goto clear_pmdnuma;
1308 }
1309
Mel Gormana54a4072013-10-07 11:28:46 +01001310 /*
1311 * Migrate the THP to the requested node, returns with page unlocked
1312 * and pmd_numa cleared.
1313 */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001314 spin_unlock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001315 migrated = migrate_misplaced_transhuge_page(mm, vma,
Hugh Dickins340ef392013-02-22 16:34:33 -08001316 pmdp, pmd, addr, page, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001317 if (migrated) {
1318 flags |= TNF_MIGRATED;
Mel Gorman8191acb2013-10-07 11:28:45 +01001319 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001320 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001321
Mel Gorman8191acb2013-10-07 11:28:45 +01001322 goto out;
Mel Gorman4daae3b2012-11-02 11:33:45 +00001323clear_pmdnuma:
Mel Gormana54a4072013-10-07 11:28:46 +01001324 BUG_ON(!PageLocked(page));
Mel Gormand10e63f2012-10-25 14:16:31 +02001325 pmd = pmd_mknonnuma(pmd);
1326 set_pmd_at(mm, haddr, pmdp, pmd);
1327 VM_BUG_ON(pmd_numa(*pmdp));
1328 update_mmu_cache_pmd(vma, addr, pmdp);
Mel Gormana54a4072013-10-07 11:28:46 +01001329 unlock_page(page);
Mel Gormand10e63f2012-10-25 14:16:31 +02001330out_unlock:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001331 spin_unlock(ptl);
Mel Gormanb8916632013-10-07 11:28:44 +01001332
1333out:
1334 if (anon_vma)
1335 page_unlock_anon_vma_read(anon_vma);
1336
Mel Gorman8191acb2013-10-07 11:28:45 +01001337 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01001338 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
Mel Gorman8191acb2013-10-07 11:28:45 +01001339
Mel Gormand10e63f2012-10-25 14:16:31 +02001340 return 0;
1341}
1342
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001343int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
Shaohua Lif21760b2012-01-12 17:19:16 -08001344 pmd_t *pmd, unsigned long addr)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001345{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001346 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001347 int ret = 0;
1348
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001349 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001350 struct page *page;
1351 pgtable_t pgtable;
David Millerf5c8ad42012-10-08 16:34:26 -07001352 pmd_t orig_pmd;
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001353 /*
1354 * For architectures like ppc64 we look at deposited pgtable
1355 * when calling pmdp_get_and_clear. So do the
1356 * pgtable_trans_huge_withdraw after finishing pmdp related
1357 * operations.
1358 */
David Millerf5c8ad42012-10-08 16:34:26 -07001359 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001360 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
Aneesh Kumar K.Va6bf2bb2013-06-05 17:14:04 -07001361 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001362 if (is_huge_zero_pmd(orig_pmd)) {
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -08001363 atomic_long_dec(&tlb->mm->nr_ptes);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001364 spin_unlock(ptl);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08001365 put_huge_zero_page();
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001366 } else {
1367 page = pmd_page(orig_pmd);
1368 page_remove_rmap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001369 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001370 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
Sasha Levin309381fea2014-01-23 15:52:54 -08001371 VM_BUG_ON_PAGE(!PageHead(page), page);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -08001372 atomic_long_dec(&tlb->mm->nr_ptes);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001373 spin_unlock(ptl);
Kirill A. Shutemov479f0ab2012-12-12 13:50:50 -08001374 tlb_remove_page(tlb, page);
1375 }
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001376 pte_free(tlb->mm, pgtable);
1377 ret = 1;
1378 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001379 return ret;
1380}
1381
Johannes Weiner0ca16342011-01-13 15:47:02 -08001382int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1383 unsigned long addr, unsigned long end,
1384 unsigned char *vec)
1385{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001386 spinlock_t *ptl;
Johannes Weiner0ca16342011-01-13 15:47:02 -08001387 int ret = 0;
1388
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001389 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001390 /*
1391 * All logical pages in the range are present
1392 * if backed by a huge page.
1393 */
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001394 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001395 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1396 ret = 1;
1397 }
Johannes Weiner0ca16342011-01-13 15:47:02 -08001398
1399 return ret;
1400}
1401
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001402int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1403 unsigned long old_addr,
1404 unsigned long new_addr, unsigned long old_end,
1405 pmd_t *old_pmd, pmd_t *new_pmd)
1406{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001407 spinlock_t *old_ptl, *new_ptl;
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001408 int ret = 0;
1409 pmd_t pmd;
1410
1411 struct mm_struct *mm = vma->vm_mm;
1412
1413 if ((old_addr & ~HPAGE_PMD_MASK) ||
1414 (new_addr & ~HPAGE_PMD_MASK) ||
1415 old_end - old_addr < HPAGE_PMD_SIZE ||
1416 (new_vma->vm_flags & VM_NOHUGEPAGE))
1417 goto out;
1418
1419 /*
1420 * The destination pmd shouldn't be established, free_pgtables()
1421 * should have release it.
1422 */
1423 if (WARN_ON(!pmd_none(*new_pmd))) {
1424 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1425 goto out;
1426 }
1427
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001428 /*
1429 * We don't have to worry about the ordering of src and dst
1430 * ptlocks because exclusive mmap_sem prevents deadlock.
1431 */
1432 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001433 if (ret == 1) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001434 new_ptl = pmd_lockptr(mm, new_pmd);
1435 if (new_ptl != old_ptl)
1436 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001437 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1438 VM_BUG_ON(!pmd_none(*new_pmd));
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001439
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301440 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1441 pgtable_t pgtable;
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001442 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1443 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
Kirill A. Shutemov35928062013-12-12 17:12:33 -08001444 }
Aneesh Kumar K.Vb3084f42014-01-13 11:34:24 +05301445 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1446 if (new_ptl != old_ptl)
1447 spin_unlock(new_ptl);
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001448 spin_unlock(old_ptl);
Andrea Arcangeli37a1c492011-10-31 17:08:30 -07001449 }
1450out:
1451 return ret;
1452}
1453
Mel Gormanf123d742013-10-07 11:28:49 +01001454/*
1455 * Returns
1456 * - 0 if PMD could not be locked
1457 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1458 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1459 */
Johannes Weinercd7548a2011-01-13 15:47:04 -08001460int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001461 unsigned long addr, pgprot_t newprot, int prot_numa)
Johannes Weinercd7548a2011-01-13 15:47:04 -08001462{
1463 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001464 spinlock_t *ptl;
Johannes Weinercd7548a2011-01-13 15:47:04 -08001465 int ret = 0;
1466
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001467 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001468 pmd_t entry;
Mel Gormanf123d742013-10-07 11:28:49 +01001469 ret = 1;
Hugh Dickinsa4f1de12012-12-16 18:56:58 -08001470 if (!prot_numa) {
Mel Gormanf123d742013-10-07 11:28:49 +01001471 entry = pmdp_get_and_clear(mm, addr, pmd);
Mel Gorman16679182013-12-18 17:08:41 -08001472 if (pmd_numa(entry))
1473 entry = pmd_mknonnuma(entry);
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001474 entry = pmd_modify(entry, newprot);
Mel Gormanf123d742013-10-07 11:28:49 +01001475 ret = HPAGE_PMD_NR;
Aneesh Kumar K.V56eecdb2014-02-12 09:13:38 +05301476 set_pmd_at(mm, addr, pmd, entry);
Hugh Dickinsa4f1de12012-12-16 18:56:58 -08001477 BUG_ON(pmd_write(entry));
1478 } else {
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001479 struct page *page = pmd_page(*pmd);
1480
Mel Gormana1a46182013-10-07 11:28:50 +01001481 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001482 * Do not trap faults against the zero page. The
1483 * read-only data is likely to be read-cached on the
1484 * local CPU cache and it is less useful to know about
1485 * local vs remote hits on the zero page.
Mel Gormana1a46182013-10-07 11:28:50 +01001486 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001487 if (!is_huge_zero_page(page) &&
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001488 !pmd_numa(*pmd)) {
Aneesh Kumar K.V56eecdb2014-02-12 09:13:38 +05301489 pmdp_set_numa(mm, addr, pmd);
Mel Gormanf123d742013-10-07 11:28:49 +01001490 ret = HPAGE_PMD_NR;
Mel Gorman4b10e7d2012-10-25 14:16:32 +02001491 }
1492 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001493 spin_unlock(ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001494 }
Johannes Weinercd7548a2011-01-13 15:47:04 -08001495
1496 return ret;
1497}
1498
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001499/*
1500 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1501 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1502 *
1503 * Note that if it returns 1, this routine returns without unlocking page
1504 * table locks. So callers must unlock them.
1505 */
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001506int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1507 spinlock_t **ptl)
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001508{
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001509 *ptl = pmd_lock(vma->vm_mm, pmd);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001510 if (likely(pmd_trans_huge(*pmd))) {
1511 if (unlikely(pmd_trans_splitting(*pmd))) {
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001512 spin_unlock(*ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001513 wait_split_huge_page(vma->anon_vma, pmd);
1514 return -1;
1515 } else {
1516 /* Thp mapped by 'pmd' is stable, so we can
1517 * handle it as it is. */
1518 return 1;
1519 }
1520 }
Kirill A. Shutemovbf929152013-11-14 14:30:54 -08001521 spin_unlock(*ptl);
Naoya Horiguchi025c5b22012-03-21 16:33:57 -07001522 return 0;
1523}
1524
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001525/*
1526 * This function returns whether a given @page is mapped onto the @address
1527 * in the virtual space of @mm.
1528 *
1529 * When it's true, this function returns *pmd with holding the page table lock
1530 * and passing it back to the caller via @ptl.
1531 * If it's false, returns NULL without holding the page table lock.
1532 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001533pmd_t *page_check_address_pmd(struct page *page,
1534 struct mm_struct *mm,
1535 unsigned long address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001536 enum page_check_address_pmd_flag flag,
1537 spinlock_t **ptl)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001538{
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001539 pgd_t *pgd;
1540 pud_t *pud;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001541 pmd_t *pmd;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001542
1543 if (address & ~HPAGE_PMD_MASK)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001544 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001545
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001546 pgd = pgd_offset(mm, address);
1547 if (!pgd_present(*pgd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001548 return NULL;
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001549 pud = pud_offset(pgd, address);
1550 if (!pud_present(*pud))
1551 return NULL;
1552 pmd = pmd_offset(pud, address);
1553
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001554 *ptl = pmd_lock(mm, pmd);
Kirill A. Shutemovb5a8cad2014-04-18 15:07:25 -07001555 if (!pmd_present(*pmd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001556 goto unlock;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001557 if (pmd_page(*pmd) != page)
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001558 goto unlock;
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08001559 /*
1560 * split_vma() may create temporary aliased mappings. There is
1561 * no risk as long as all huge pmd are found and have their
1562 * splitting bit set before __split_huge_page_refcount
1563 * runs. Finding the same huge pmd more than once during the
1564 * same rmap walk is not a problem.
1565 */
1566 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1567 pmd_trans_splitting(*pmd))
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001568 goto unlock;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001569 if (pmd_trans_huge(*pmd)) {
1570 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1571 !pmd_trans_splitting(*pmd));
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001572 return pmd;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001573 }
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001574unlock:
1575 spin_unlock(*ptl);
1576 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001577}
1578
1579static int __split_huge_page_splitting(struct page *page,
1580 struct vm_area_struct *vma,
1581 unsigned long address)
1582{
1583 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001584 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001585 pmd_t *pmd;
1586 int ret = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001587 /* For mmu_notifiers */
1588 const unsigned long mmun_start = address;
1589 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001590
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001591 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001592 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001593 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001594 if (pmd) {
1595 /*
1596 * We can't temporarily set the pmd to null in order
1597 * to split it, the pmd must remain marked huge at all
1598 * times or the VM won't take the pmd_trans_huge paths
Ingo Molnar5a505082012-12-02 19:56:46 +00001599 * and it won't wait on the anon_vma->root->rwsem to
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001600 * serialize against split_huge_page*.
1601 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001602 pmdp_splitting_flush(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001603 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001604 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001605 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001606 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001607
1608 return ret;
1609}
1610
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001611static void __split_huge_page_refcount(struct page *page,
1612 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001613{
1614 int i;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001615 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001616 struct lruvec *lruvec;
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001617 int tail_count = 0;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001618
1619 /* prevent PageLRU to go away from under us, and freeze lru stats */
1620 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001621 lruvec = mem_cgroup_page_lruvec(page, zone);
1622
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001623 compound_lock(page);
KAMEZAWA Hiroyukie94c8a92012-01-12 17:18:20 -08001624 /* complete memcg works before add pages to LRU */
1625 mem_cgroup_split_huge_fixup(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001626
Shaohua Li45676882012-01-12 17:19:18 -08001627 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001628 struct page *page_tail = page + i;
1629
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001630 /* tail_page->_mapcount cannot change */
1631 BUG_ON(page_mapcount(page_tail) < 0);
1632 tail_count += page_mapcount(page_tail);
1633 /* check for overflow */
1634 BUG_ON(tail_count < 0);
1635 BUG_ON(atomic_read(&page_tail->_count) != 0);
1636 /*
1637 * tail_page->_count is zero and not changing from
1638 * under us. But get_page_unless_zero() may be running
1639 * from under us on the tail_page. If we used
1640 * atomic_set() below instead of atomic_add(), we
1641 * would then run atomic_set() concurrently with
1642 * get_page_unless_zero(), and atomic_set() is
1643 * implemented in C not using locked ops. spin_unlock
1644 * on x86 sometime uses locked ops because of PPro
1645 * errata 66, 92, so unless somebody can guarantee
1646 * atomic_set() here would be safe on all archs (and
1647 * not only on x86), it's safer to use atomic_add().
1648 */
1649 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1650 &page_tail->_count);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001651
1652 /* after clearing PageTail the gup refcount can be released */
1653 smp_mb();
1654
Jin Dongminga6d30dd2011-02-01 15:52:40 -08001655 /*
1656 * retain hwpoison flag of the poisoned tail page:
1657 * fix for the unsuitable process killed on Guest Machine(KVM)
1658 * by the memory-failure.
1659 */
1660 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001661 page_tail->flags |= (page->flags &
1662 ((1L << PG_referenced) |
1663 (1L << PG_swapbacked) |
1664 (1L << PG_mlocked) |
Kirill A. Shutemove180cf82013-07-31 13:53:39 -07001665 (1L << PG_uptodate) |
1666 (1L << PG_active) |
1667 (1L << PG_unevictable)));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001668 page_tail->flags |= (1L << PG_dirty);
1669
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001670 /* clear PageTail before overwriting first_page */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001671 smp_wmb();
1672
1673 /*
1674 * __split_huge_page_splitting() already set the
1675 * splitting bit in all pmd that could map this
1676 * hugepage, that will ensure no CPU can alter the
1677 * mapcount on the head page. The mapcount is only
1678 * accounted in the head page and it has to be
1679 * transferred to all tail pages in the below code. So
1680 * for this code to be safe, the split the mapcount
1681 * can't change. But that doesn't mean userland can't
1682 * keep changing and reading the page contents while
1683 * we transfer the mapcount, so the pmd splitting
1684 * status is achieved setting a reserved bit in the
1685 * pmd, not by clearing the present bit.
1686 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001687 page_tail->_mapcount = page->_mapcount;
1688
1689 BUG_ON(page_tail->mapping);
1690 page_tail->mapping = page->mapping;
1691
Shaohua Li45676882012-01-12 17:19:18 -08001692 page_tail->index = page->index + i;
Peter Zijlstra90572892013-10-07 11:29:20 +01001693 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001694
1695 BUG_ON(!PageAnon(page_tail));
1696 BUG_ON(!PageUptodate(page_tail));
1697 BUG_ON(!PageDirty(page_tail));
1698 BUG_ON(!PageSwapBacked(page_tail));
1699
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001700 lru_add_page_tail(page, page_tail, lruvec, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001701 }
Andrea Arcangeli70b50f92011-11-02 13:36:59 -07001702 atomic_sub(tail_count, &page->_count);
1703 BUG_ON(atomic_read(&page->_count) <= 0);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001704
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001705 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001706
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001707 ClearPageCompound(page);
1708 compound_unlock(page);
1709 spin_unlock_irq(&zone->lru_lock);
1710
1711 for (i = 1; i < HPAGE_PMD_NR; i++) {
1712 struct page *page_tail = page + i;
1713 BUG_ON(page_count(page_tail) <= 0);
1714 /*
1715 * Tail pages may be freed if there wasn't any mapping
1716 * like if add_to_swap() is running on a lru page that
1717 * had its mapping zapped. And freeing these pages
1718 * requires taking the lru_lock so we do the put_page
1719 * of the tail pages after the split is complete.
1720 */
1721 put_page(page_tail);
1722 }
1723
1724 /*
1725 * Only the head page (now become a regular page) is required
1726 * to be pinned by the caller.
1727 */
1728 BUG_ON(page_count(page) <= 0);
1729}
1730
1731static int __split_huge_page_map(struct page *page,
1732 struct vm_area_struct *vma,
1733 unsigned long address)
1734{
1735 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001736 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001737 pmd_t *pmd, _pmd;
1738 int ret = 0, i;
1739 pgtable_t pgtable;
1740 unsigned long haddr;
1741
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001742 pmd = page_check_address_pmd(page, mm, address,
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001743 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001744 if (pmd) {
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07001745 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001746 pmd_populate(mm, &_pmd, pgtable);
1747
Gerald Schaefere3ebcf642012-10-08 16:30:07 -07001748 haddr = address;
1749 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001750 pte_t *pte, entry;
1751 BUG_ON(PageCompound(page+i));
1752 entry = mk_pte(page + i, vma->vm_page_prot);
1753 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1754 if (!pmd_write(*pmd))
1755 entry = pte_wrprotect(entry);
1756 else
1757 BUG_ON(page_mapcount(page) != 1);
1758 if (!pmd_young(*pmd))
1759 entry = pte_mkold(entry);
Andrea Arcangeli1ba6e0b2012-10-04 01:51:06 +02001760 if (pmd_numa(*pmd))
1761 entry = pte_mknuma(entry);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001762 pte = pte_offset_map(&_pmd, haddr);
1763 BUG_ON(!pte_none(*pte));
1764 set_pte_at(mm, haddr, pte, entry);
1765 pte_unmap(pte);
1766 }
1767
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001768 smp_wmb(); /* make pte visible before pmd */
1769 /*
1770 * Up to this point the pmd is present and huge and
1771 * userland has the whole access to the hugepage
1772 * during the split (which happens in place). If we
1773 * overwrite the pmd with the not-huge version
1774 * pointing to the pte here (which of course we could
1775 * if all CPUs were bug free), userland could trigger
1776 * a small page size TLB miss on the small sized TLB
1777 * while the hugepage TLB entry is still established
1778 * in the huge TLB. Some CPU doesn't like that. See
1779 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1780 * Erratum 383 on page 93. Intel should be safe but is
1781 * also warns that it's only safe if the permission
1782 * and cache attributes of the two entries loaded in
1783 * the two TLB is identical (which should be the case
1784 * here). But it is generally safer to never allow
1785 * small and huge TLB entries for the same virtual
1786 * address to be loaded simultaneously. So instead of
1787 * doing "pmd_populate(); flush_tlb_range();" we first
1788 * mark the current pmd notpresent (atomically because
1789 * here the pmd_trans_huge and pmd_trans_splitting
1790 * must remain set at all times on the pmd until the
1791 * split is complete for this pmd), then we flush the
1792 * SMP TLB and finally we write the non-huge version
1793 * of the pmd entry with pmd_populate.
1794 */
Gerald Schaefer46dcde72012-10-08 16:30:09 -07001795 pmdp_invalidate(vma, address, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001796 pmd_populate(mm, pmd, pgtable);
1797 ret = 1;
Kirill A. Shutemov117b0792013-11-14 14:30:56 -08001798 spin_unlock(ptl);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001799 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001800
1801 return ret;
1802}
1803
Ingo Molnar5a505082012-12-02 19:56:46 +00001804/* must be called with anon_vma->root->rwsem held */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001805static void __split_huge_page(struct page *page,
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001806 struct anon_vma *anon_vma,
1807 struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001808{
1809 int mapcount, mapcount2;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001810 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001811 struct anon_vma_chain *avc;
1812
1813 BUG_ON(!PageHead(page));
1814 BUG_ON(PageTail(page));
1815
1816 mapcount = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001817 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001818 struct vm_area_struct *vma = avc->vma;
1819 unsigned long addr = vma_address(page, vma);
1820 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001821 mapcount += __split_huge_page_splitting(page, vma, addr);
1822 }
Andrea Arcangeli05759d32011-01-13 15:46:53 -08001823 /*
1824 * It is critical that new vmas are added to the tail of the
1825 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1826 * and establishes a child pmd before
1827 * __split_huge_page_splitting() freezes the parent pmd (so if
1828 * we fail to prevent copy_huge_pmd() from running until the
1829 * whole __split_huge_page() is complete), we will still see
1830 * the newly established pmd of the child later during the
1831 * walk, to be able to set it as pmd_trans_splitting too.
1832 */
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001833 if (mapcount != page_mapcount(page)) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -07001834 pr_err("mapcount %d page_mapcount %d\n",
1835 mapcount, page_mapcount(page));
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001836 BUG();
1837 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001838
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001839 __split_huge_page_refcount(page, list);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001840
1841 mapcount2 = 0;
Michel Lespinassebf181b92012-10-08 16:31:39 -07001842 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001843 struct vm_area_struct *vma = avc->vma;
1844 unsigned long addr = vma_address(page, vma);
1845 BUG_ON(is_vma_temporary_stack(vma));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001846 mapcount2 += __split_huge_page_map(page, vma, addr);
1847 }
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001848 if (mapcount != mapcount2) {
Andrew Mortonae3a8c12014-06-04 16:06:58 -07001849 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1850 mapcount, mapcount2, page_mapcount(page));
Kirill A. Shutemovff9e43e2014-06-04 16:06:57 -07001851 BUG();
1852 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001853}
1854
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001855/*
1856 * Split a hugepage into normal pages. This doesn't change the position of head
1857 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1858 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1859 * from the hugepage.
1860 * Return 0 if the hugepage is split successfully otherwise return 1.
1861 */
1862int split_huge_page_to_list(struct page *page, struct list_head *list)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001863{
1864 struct anon_vma *anon_vma;
1865 int ret = 1;
1866
Kirill A. Shutemov5918d102013-04-29 15:08:44 -07001867 BUG_ON(is_huge_zero_page(page));
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001868 BUG_ON(!PageAnon(page));
Mel Gorman062f1af2013-01-11 14:32:02 -08001869
1870 /*
1871 * The caller does not necessarily hold an mmap_sem that would prevent
1872 * the anon_vma disappearing so we first we take a reference to it
1873 * and then lock the anon_vma for write. This is similar to
1874 * page_lock_anon_vma_read except the write lock is taken to serialise
1875 * against parallel split or collapse operations.
1876 */
1877 anon_vma = page_get_anon_vma(page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001878 if (!anon_vma)
1879 goto out;
Mel Gorman062f1af2013-01-11 14:32:02 -08001880 anon_vma_lock_write(anon_vma);
1881
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001882 ret = 0;
1883 if (!PageCompound(page))
1884 goto out_unlock;
1885
1886 BUG_ON(!PageSwapBacked(page));
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001887 __split_huge_page(page, anon_vma, list);
Andi Kleen81ab4202011-04-14 15:22:06 -07001888 count_vm_event(THP_SPLIT);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001889
1890 BUG_ON(PageCompound(page));
1891out_unlock:
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08001892 anon_vma_unlock_write(anon_vma);
Mel Gorman062f1af2013-01-11 14:32:02 -08001893 put_anon_vma(anon_vma);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001894out:
1895 return ret;
1896}
1897
Vlastimil Babka9050d7e2014-03-03 15:38:27 -08001898#define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001899
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001900int hugepage_madvise(struct vm_area_struct *vma,
1901 unsigned long *vm_flags, int advice)
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001902{
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001903 switch (advice) {
1904 case MADV_HUGEPAGE:
Alex Thorlton1e1836e2014-04-07 15:37:09 -07001905#ifdef CONFIG_S390
1906 /*
1907 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1908 * can't handle this properly after s390_enable_sie, so we simply
1909 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1910 */
1911 if (mm_has_pgste(vma->vm_mm))
1912 return 0;
1913#endif
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001914 /*
1915 * Be somewhat over-protective like KSM for now!
1916 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001917 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001918 return -EINVAL;
1919 *vm_flags &= ~VM_NOHUGEPAGE;
1920 *vm_flags |= VM_HUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001921 /*
1922 * If the vma become good for khugepaged to scan,
1923 * register it here without waiting a page fault that
1924 * may not happen any time soon.
1925 */
1926 if (unlikely(khugepaged_enter_vma_merge(vma)))
1927 return -ENOMEM;
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001928 break;
1929 case MADV_NOHUGEPAGE:
1930 /*
1931 * Be somewhat over-protective like KSM for now!
1932 */
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07001933 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001934 return -EINVAL;
1935 *vm_flags &= ~VM_HUGEPAGE;
1936 *vm_flags |= VM_NOHUGEPAGE;
Andrea Arcangeli60ab3242011-01-13 15:47:18 -08001937 /*
1938 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1939 * this vma even if we leave the mm registered in khugepaged if
1940 * it got registered before VM_NOHUGEPAGE was set.
1941 */
Andrea Arcangelia664b2d2011-01-13 15:47:17 -08001942 break;
1943 }
Andrea Arcangeli0af4e982011-01-13 15:46:55 -08001944
1945 return 0;
1946}
1947
Andrea Arcangeliba761492011-01-13 15:46:58 -08001948static int __init khugepaged_slab_init(void)
1949{
1950 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1951 sizeof(struct mm_slot),
1952 __alignof__(struct mm_slot), 0, NULL);
1953 if (!mm_slot_cache)
1954 return -ENOMEM;
1955
1956 return 0;
1957}
1958
Andrea Arcangeliba761492011-01-13 15:46:58 -08001959static inline struct mm_slot *alloc_mm_slot(void)
1960{
1961 if (!mm_slot_cache) /* initialization failed */
1962 return NULL;
1963 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1964}
1965
1966static inline void free_mm_slot(struct mm_slot *mm_slot)
1967{
1968 kmem_cache_free(mm_slot_cache, mm_slot);
1969}
1970
Andrea Arcangeliba761492011-01-13 15:46:58 -08001971static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1972{
1973 struct mm_slot *mm_slot;
Andrea Arcangeliba761492011-01-13 15:46:58 -08001974
Sasha Levinb67bfe02013-02-27 17:06:00 -08001975 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
Andrea Arcangeliba761492011-01-13 15:46:58 -08001976 if (mm == mm_slot->mm)
1977 return mm_slot;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08001978
Andrea Arcangeliba761492011-01-13 15:46:58 -08001979 return NULL;
1980}
1981
1982static void insert_to_mm_slots_hash(struct mm_struct *mm,
1983 struct mm_slot *mm_slot)
1984{
Andrea Arcangeliba761492011-01-13 15:46:58 -08001985 mm_slot->mm = mm;
Sasha Levin43b5fbb2013-02-22 16:32:27 -08001986 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
Andrea Arcangeliba761492011-01-13 15:46:58 -08001987}
1988
1989static inline int khugepaged_test_exit(struct mm_struct *mm)
1990{
1991 return atomic_read(&mm->mm_users) == 0;
1992}
1993
1994int __khugepaged_enter(struct mm_struct *mm)
1995{
1996 struct mm_slot *mm_slot;
1997 int wakeup;
1998
1999 mm_slot = alloc_mm_slot();
2000 if (!mm_slot)
2001 return -ENOMEM;
2002
2003 /* __khugepaged_exit() must not run from under us */
2004 VM_BUG_ON(khugepaged_test_exit(mm));
2005 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2006 free_mm_slot(mm_slot);
2007 return 0;
2008 }
2009
2010 spin_lock(&khugepaged_mm_lock);
2011 insert_to_mm_slots_hash(mm, mm_slot);
2012 /*
2013 * Insert just behind the scanning cursor, to let the area settle
2014 * down a little.
2015 */
2016 wakeup = list_empty(&khugepaged_scan.mm_head);
2017 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2018 spin_unlock(&khugepaged_mm_lock);
2019
2020 atomic_inc(&mm->mm_count);
2021 if (wakeup)
2022 wake_up_interruptible(&khugepaged_wait);
2023
2024 return 0;
2025}
2026
2027int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2028{
2029 unsigned long hstart, hend;
2030 if (!vma->anon_vma)
2031 /*
2032 * Not yet faulted in so we will register later in the
2033 * page fault if needed.
2034 */
2035 return 0;
Andrea Arcangeli78f11a22011-04-27 15:26:45 -07002036 if (vma->vm_ops)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002037 /* khugepaged not yet working on file or special mappings */
2038 return 0;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07002039 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002040 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2041 hend = vma->vm_end & HPAGE_PMD_MASK;
2042 if (hstart < hend)
2043 return khugepaged_enter(vma);
2044 return 0;
2045}
2046
2047void __khugepaged_exit(struct mm_struct *mm)
2048{
2049 struct mm_slot *mm_slot;
2050 int free = 0;
2051
2052 spin_lock(&khugepaged_mm_lock);
2053 mm_slot = get_mm_slot(mm);
2054 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002055 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002056 list_del(&mm_slot->mm_node);
2057 free = 1;
2058 }
Chris Wrightd788e802011-07-25 17:12:14 -07002059 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002060
2061 if (free) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002062 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2063 free_mm_slot(mm_slot);
2064 mmdrop(mm);
2065 } else if (mm_slot) {
Andrea Arcangeliba761492011-01-13 15:46:58 -08002066 /*
2067 * This is required to serialize against
2068 * khugepaged_test_exit() (which is guaranteed to run
2069 * under mmap sem read mode). Stop here (after we
2070 * return all pagetables will be destroyed) until
2071 * khugepaged has finished working on the pagetables
2072 * under the mmap_sem.
2073 */
2074 down_write(&mm->mmap_sem);
2075 up_write(&mm->mmap_sem);
Chris Wrightd788e802011-07-25 17:12:14 -07002076 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002077}
2078
2079static void release_pte_page(struct page *page)
2080{
2081 /* 0 stands for page_is_file_cache(page) == false */
2082 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2083 unlock_page(page);
2084 putback_lru_page(page);
2085}
2086
2087static void release_pte_pages(pte_t *pte, pte_t *_pte)
2088{
2089 while (--_pte >= pte) {
2090 pte_t pteval = *_pte;
2091 if (!pte_none(pteval))
2092 release_pte_page(pte_page(pteval));
2093 }
2094}
2095
Andrea Arcangeliba761492011-01-13 15:46:58 -08002096static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2097 unsigned long address,
2098 pte_t *pte)
2099{
2100 struct page *page;
2101 pte_t *_pte;
Bob Liu344aa352012-12-11 16:00:34 -08002102 int referenced = 0, none = 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002103 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2104 _pte++, address += PAGE_SIZE) {
2105 pte_t pteval = *_pte;
2106 if (pte_none(pteval)) {
2107 if (++none <= khugepaged_max_ptes_none)
2108 continue;
Bob Liu344aa352012-12-11 16:00:34 -08002109 else
Andrea Arcangeliba761492011-01-13 15:46:58 -08002110 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002111 }
Bob Liu344aa352012-12-11 16:00:34 -08002112 if (!pte_present(pteval) || !pte_write(pteval))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002113 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002114 page = vm_normal_page(vma, address, pteval);
Bob Liu344aa352012-12-11 16:00:34 -08002115 if (unlikely(!page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002116 goto out;
Bob Liu344aa352012-12-11 16:00:34 -08002117
Sasha Levin309381fea2014-01-23 15:52:54 -08002118 VM_BUG_ON_PAGE(PageCompound(page), page);
2119 VM_BUG_ON_PAGE(!PageAnon(page), page);
2120 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002121
2122 /* cannot use mapcount: can't collapse if there's a gup pin */
Bob Liu344aa352012-12-11 16:00:34 -08002123 if (page_count(page) != 1)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002124 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002125 /*
2126 * We can do it before isolate_lru_page because the
2127 * page can't be freed from under us. NOTE: PG_lock
2128 * is needed to serialize against split_huge_page
2129 * when invoked from the VM.
2130 */
Bob Liu344aa352012-12-11 16:00:34 -08002131 if (!trylock_page(page))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002132 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002133 /*
2134 * Isolate the page to avoid collapsing an hugepage
2135 * currently in use by the VM.
2136 */
2137 if (isolate_lru_page(page)) {
2138 unlock_page(page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002139 goto out;
2140 }
2141 /* 0 stands for page_is_file_cache(page) == false */
2142 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
Sasha Levin309381fea2014-01-23 15:52:54 -08002143 VM_BUG_ON_PAGE(!PageLocked(page), page);
2144 VM_BUG_ON_PAGE(PageLRU(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002145
2146 /* If there is no mapped pte young don't collapse the page */
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002147 if (pte_young(pteval) || PageReferenced(page) ||
2148 mmu_notifier_test_young(vma->vm_mm, address))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002149 referenced = 1;
2150 }
Bob Liu344aa352012-12-11 16:00:34 -08002151 if (likely(referenced))
2152 return 1;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002153out:
Bob Liu344aa352012-12-11 16:00:34 -08002154 release_pte_pages(pte, _pte);
2155 return 0;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002156}
2157
2158static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2159 struct vm_area_struct *vma,
2160 unsigned long address,
2161 spinlock_t *ptl)
2162{
2163 pte_t *_pte;
2164 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2165 pte_t pteval = *_pte;
2166 struct page *src_page;
2167
2168 if (pte_none(pteval)) {
2169 clear_user_highpage(page, address);
2170 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2171 } else {
2172 src_page = pte_page(pteval);
2173 copy_user_highpage(page, src_page, address, vma);
Sasha Levin309381fea2014-01-23 15:52:54 -08002174 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002175 release_pte_page(src_page);
2176 /*
2177 * ptl mostly unnecessary, but preempt has to
2178 * be disabled to update the per-cpu stats
2179 * inside page_remove_rmap().
2180 */
2181 spin_lock(ptl);
2182 /*
2183 * paravirt calls inside pte_clear here are
2184 * superfluous.
2185 */
2186 pte_clear(vma->vm_mm, address, _pte);
2187 page_remove_rmap(src_page);
2188 spin_unlock(ptl);
2189 free_page_and_swap_cache(src_page);
2190 }
2191
2192 address += PAGE_SIZE;
2193 page++;
2194 }
2195}
2196
Xiao Guangrong26234f32012-10-08 16:29:51 -07002197static void khugepaged_alloc_sleep(void)
2198{
2199 wait_event_freezable_timeout(khugepaged_wait, false,
2200 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2201}
2202
Bob Liu9f1b8682013-11-12 15:07:37 -08002203static int khugepaged_node_load[MAX_NUMNODES];
2204
Xiao Guangrong26234f32012-10-08 16:29:51 -07002205#ifdef CONFIG_NUMA
Bob Liu9f1b8682013-11-12 15:07:37 -08002206static int khugepaged_find_target_node(void)
2207{
2208 static int last_khugepaged_target_node = NUMA_NO_NODE;
2209 int nid, target_node = 0, max_value = 0;
2210
2211 /* find first node with max normal pages hit */
2212 for (nid = 0; nid < MAX_NUMNODES; nid++)
2213 if (khugepaged_node_load[nid] > max_value) {
2214 max_value = khugepaged_node_load[nid];
2215 target_node = nid;
2216 }
2217
2218 /* do some balance if several nodes have the same hit record */
2219 if (target_node <= last_khugepaged_target_node)
2220 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2221 nid++)
2222 if (max_value == khugepaged_node_load[nid]) {
2223 target_node = nid;
2224 break;
2225 }
2226
2227 last_khugepaged_target_node = target_node;
2228 return target_node;
2229}
2230
Xiao Guangrong26234f32012-10-08 16:29:51 -07002231static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2232{
2233 if (IS_ERR(*hpage)) {
2234 if (!*wait)
2235 return false;
2236
2237 *wait = false;
Xiao Guangronge3b41262012-10-08 16:32:57 -07002238 *hpage = NULL;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002239 khugepaged_alloc_sleep();
2240 } else if (*hpage) {
2241 put_page(*hpage);
2242 *hpage = NULL;
2243 }
2244
2245 return true;
2246}
2247
2248static struct page
2249*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2250 struct vm_area_struct *vma, unsigned long address,
2251 int node)
2252{
Sasha Levin309381fea2014-01-23 15:52:54 -08002253 VM_BUG_ON_PAGE(*hpage, *hpage);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002254 /*
2255 * Allocate the page while the vma is still valid and under
2256 * the mmap_sem read mode so there is no memory allocation
2257 * later when we take the mmap_sem in write mode. This is more
2258 * friendly behavior (OTOH it may actually hide bugs) to
2259 * filesystems in userland with daemons allocating memory in
2260 * the userland I/O paths. Allocating memory with the
2261 * mmap_sem in read mode is good idea also to allow greater
2262 * scalability.
2263 */
Bob Liu9f1b8682013-11-12 15:07:37 -08002264 *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2265 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
Xiao Guangrong26234f32012-10-08 16:29:51 -07002266 /*
2267 * After allocating the hugepage, release the mmap_sem read lock in
2268 * preparation for taking it in write mode.
2269 */
2270 up_read(&mm->mmap_sem);
2271 if (unlikely(!*hpage)) {
2272 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2273 *hpage = ERR_PTR(-ENOMEM);
2274 return NULL;
2275 }
2276
2277 count_vm_event(THP_COLLAPSE_ALLOC);
2278 return *hpage;
2279}
2280#else
Bob Liu9f1b8682013-11-12 15:07:37 -08002281static int khugepaged_find_target_node(void)
2282{
2283 return 0;
2284}
2285
Bob Liu10dc4152013-11-12 15:07:35 -08002286static inline struct page *alloc_hugepage(int defrag)
2287{
2288 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2289 HPAGE_PMD_ORDER);
2290}
2291
Xiao Guangrong26234f32012-10-08 16:29:51 -07002292static struct page *khugepaged_alloc_hugepage(bool *wait)
2293{
2294 struct page *hpage;
2295
2296 do {
2297 hpage = alloc_hugepage(khugepaged_defrag());
2298 if (!hpage) {
2299 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2300 if (!*wait)
2301 return NULL;
2302
2303 *wait = false;
2304 khugepaged_alloc_sleep();
2305 } else
2306 count_vm_event(THP_COLLAPSE_ALLOC);
2307 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2308
2309 return hpage;
2310}
2311
2312static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2313{
2314 if (!*hpage)
2315 *hpage = khugepaged_alloc_hugepage(wait);
2316
2317 if (unlikely(!*hpage))
2318 return false;
2319
2320 return true;
2321}
2322
2323static struct page
2324*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2325 struct vm_area_struct *vma, unsigned long address,
2326 int node)
2327{
2328 up_read(&mm->mmap_sem);
2329 VM_BUG_ON(!*hpage);
2330 return *hpage;
2331}
2332#endif
2333
Bob Liufa475e52012-12-11 16:00:39 -08002334static bool hugepage_vma_check(struct vm_area_struct *vma)
2335{
2336 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2337 (vma->vm_flags & VM_NOHUGEPAGE))
2338 return false;
2339
2340 if (!vma->anon_vma || vma->vm_ops)
2341 return false;
2342 if (is_vma_temporary_stack(vma))
2343 return false;
2344 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2345 return true;
2346}
2347
Andrea Arcangeliba761492011-01-13 15:46:58 -08002348static void collapse_huge_page(struct mm_struct *mm,
Xiao Guangrong26234f32012-10-08 16:29:51 -07002349 unsigned long address,
2350 struct page **hpage,
2351 struct vm_area_struct *vma,
2352 int node)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002353{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002354 pmd_t *pmd, _pmd;
2355 pte_t *pte;
2356 pgtable_t pgtable;
2357 struct page *new_page;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002358 spinlock_t *pmd_ptl, *pte_ptl;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002359 int isolated;
2360 unsigned long hstart, hend;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002361 unsigned long mmun_start; /* For mmu_notifiers */
2362 unsigned long mmun_end; /* For mmu_notifiers */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002363
2364 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002365
Xiao Guangrong26234f32012-10-08 16:29:51 -07002366 /* release the mmap_sem read lock. */
2367 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2368 if (!new_page)
Andrea Arcangelice83d212011-01-13 15:47:06 -08002369 return;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002370
Michal Hockod715ae02014-04-07 15:37:46 -07002371 if (unlikely(mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL)))
Andrea Arcangeli692e0b32011-05-24 17:12:14 -07002372 return;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002373
2374 /*
2375 * Prevent all access to pagetables with the exception of
2376 * gup_fast later hanlded by the ptep_clear_flush and the VM
2377 * handled by the anon_vma lock + PG_lock.
2378 */
2379 down_write(&mm->mmap_sem);
2380 if (unlikely(khugepaged_test_exit(mm)))
2381 goto out;
2382
2383 vma = find_vma(mm, address);
Libina8f531eb2013-09-11 14:20:38 -07002384 if (!vma)
2385 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002386 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2387 hend = vma->vm_end & HPAGE_PMD_MASK;
2388 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2389 goto out;
Bob Liufa475e52012-12-11 16:00:39 -08002390 if (!hugepage_vma_check(vma))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002391 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002392 pmd = mm_find_pmd(mm, address);
2393 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002394 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002395 if (pmd_trans_huge(*pmd))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002396 goto out;
2397
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +00002398 anon_vma_lock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002399
2400 pte = pte_offset_map(pmd, address);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002401 pte_ptl = pte_lockptr(mm, pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002402
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002403 mmun_start = address;
2404 mmun_end = address + HPAGE_PMD_SIZE;
2405 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002406 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
Andrea Arcangeliba761492011-01-13 15:46:58 -08002407 /*
2408 * After this gup_fast can't run anymore. This also removes
2409 * any huge TLB entry from the CPU so we won't allow
2410 * huge and small TLB entries for the same virtual address
2411 * to avoid the risk of CPU bugs in that area.
2412 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002413 _pmd = pmdp_clear_flush(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002414 spin_unlock(pmd_ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002415 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002416
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002417 spin_lock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002418 isolated = __collapse_huge_page_isolate(vma, address, pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002419 spin_unlock(pte_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002420
2421 if (unlikely(!isolated)) {
Johannes Weiner453c7192011-01-20 14:44:18 -08002422 pte_unmap(pte);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002423 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002424 BUG_ON(!pmd_none(*pmd));
Aneesh Kumar K.V7c342512013-05-24 15:55:21 -07002425 /*
2426 * We can only use set_pmd_at when establishing
2427 * hugepmds and never for establishing regular pmds that
2428 * points to regular pagetables. Use pmd_populate for that
2429 */
2430 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002431 spin_unlock(pmd_ptl);
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002432 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002433 goto out;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002434 }
2435
2436 /*
2437 * All pages are isolated and locked so anon_vma rmap
2438 * can't run anymore.
2439 */
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -08002440 anon_vma_unlock_write(vma->anon_vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002441
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002442 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
Johannes Weiner453c7192011-01-20 14:44:18 -08002443 pte_unmap(pte);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002444 __SetPageUptodate(new_page);
2445 pgtable = pmd_pgtable(_pmd);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002446
Kirill A. Shutemov31223592013-09-12 15:14:01 -07002447 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2448 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002449
2450 /*
2451 * spin_lock() below is not the equivalent of smp_wmb(), so
2452 * this is needed to avoid the copy_huge_page writes to become
2453 * visible after the set_pmd_at() write.
2454 */
2455 smp_wmb();
2456
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002457 spin_lock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002458 BUG_ON(!pmd_none(*pmd));
2459 page_add_new_anon_rmap(new_page, vma, address);
Aneesh Kumar K.Vfce144b2013-06-05 17:14:06 -07002460 pgtable_trans_huge_deposit(mm, pmd, pgtable);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002461 set_pmd_at(mm, address, pmd, _pmd);
David Millerb113da62012-10-08 16:34:25 -07002462 update_mmu_cache_pmd(vma, address, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002463 spin_unlock(pmd_ptl);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002464
2465 *hpage = NULL;
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002466
Andrea Arcangeliba761492011-01-13 15:46:58 -08002467 khugepaged_pages_collapsed++;
Andrea Arcangelice83d212011-01-13 15:47:06 -08002468out_up_write:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002469 up_write(&mm->mmap_sem);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002470 return;
2471
Andrea Arcangelice83d212011-01-13 15:47:06 -08002472out:
KAMEZAWA Hiroyuki678ff892011-02-10 15:01:36 -08002473 mem_cgroup_uncharge_page(new_page);
Andrea Arcangelice83d212011-01-13 15:47:06 -08002474 goto out_up_write;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002475}
2476
2477static int khugepaged_scan_pmd(struct mm_struct *mm,
2478 struct vm_area_struct *vma,
2479 unsigned long address,
2480 struct page **hpage)
2481{
Andrea Arcangeliba761492011-01-13 15:46:58 -08002482 pmd_t *pmd;
2483 pte_t *pte, *_pte;
2484 int ret = 0, referenced = 0, none = 0;
2485 struct page *page;
2486 unsigned long _address;
2487 spinlock_t *ptl;
David Rientjes00ef2d22013-02-22 16:35:36 -08002488 int node = NUMA_NO_NODE;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002489
2490 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2491
Bob Liu62190492012-12-11 16:00:37 -08002492 pmd = mm_find_pmd(mm, address);
2493 if (!pmd)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002494 goto out;
Bob Liu62190492012-12-11 16:00:37 -08002495 if (pmd_trans_huge(*pmd))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002496 goto out;
2497
Bob Liu9f1b8682013-11-12 15:07:37 -08002498 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002499 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2500 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2501 _pte++, _address += PAGE_SIZE) {
2502 pte_t pteval = *_pte;
2503 if (pte_none(pteval)) {
2504 if (++none <= khugepaged_max_ptes_none)
2505 continue;
2506 else
2507 goto out_unmap;
2508 }
2509 if (!pte_present(pteval) || !pte_write(pteval))
2510 goto out_unmap;
2511 page = vm_normal_page(vma, _address, pteval);
2512 if (unlikely(!page))
2513 goto out_unmap;
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002514 /*
Bob Liu9f1b8682013-11-12 15:07:37 -08002515 * Record which node the original page is from and save this
2516 * information to khugepaged_node_load[].
2517 * Khupaged will allocate hugepage from the node has the max
2518 * hit record.
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002519 */
Bob Liu9f1b8682013-11-12 15:07:37 -08002520 node = page_to_nid(page);
2521 khugepaged_node_load[node]++;
Sasha Levin309381fea2014-01-23 15:52:54 -08002522 VM_BUG_ON_PAGE(PageCompound(page), page);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002523 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2524 goto out_unmap;
2525 /* cannot use mapcount: can't collapse if there's a gup pin */
2526 if (page_count(page) != 1)
2527 goto out_unmap;
Andrea Arcangeli8ee53822011-01-13 15:47:10 -08002528 if (pte_young(pteval) || PageReferenced(page) ||
2529 mmu_notifier_test_young(vma->vm_mm, address))
Andrea Arcangeliba761492011-01-13 15:46:58 -08002530 referenced = 1;
2531 }
2532 if (referenced)
2533 ret = 1;
2534out_unmap:
2535 pte_unmap_unlock(pte, ptl);
Bob Liu9f1b8682013-11-12 15:07:37 -08002536 if (ret) {
2537 node = khugepaged_find_target_node();
Andrea Arcangelice83d212011-01-13 15:47:06 -08002538 /* collapse_huge_page will return with the mmap_sem released */
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002539 collapse_huge_page(mm, address, hpage, vma, node);
Bob Liu9f1b8682013-11-12 15:07:37 -08002540 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002541out:
2542 return ret;
2543}
2544
2545static void collect_mm_slot(struct mm_slot *mm_slot)
2546{
2547 struct mm_struct *mm = mm_slot->mm;
2548
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002549 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002550
2551 if (khugepaged_test_exit(mm)) {
2552 /* free mm_slot */
Sasha Levin43b5fbb2013-02-22 16:32:27 -08002553 hash_del(&mm_slot->hash);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002554 list_del(&mm_slot->mm_node);
2555
2556 /*
2557 * Not strictly needed because the mm exited already.
2558 *
2559 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2560 */
2561
2562 /* khugepaged_mm_lock actually not necessary for the below */
2563 free_mm_slot(mm_slot);
2564 mmdrop(mm);
2565 }
2566}
2567
2568static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2569 struct page **hpage)
H Hartley Sweeten2f1da642011-10-31 17:09:25 -07002570 __releases(&khugepaged_mm_lock)
2571 __acquires(&khugepaged_mm_lock)
Andrea Arcangeliba761492011-01-13 15:46:58 -08002572{
2573 struct mm_slot *mm_slot;
2574 struct mm_struct *mm;
2575 struct vm_area_struct *vma;
2576 int progress = 0;
2577
2578 VM_BUG_ON(!pages);
Hugh Dickinsb9980cd2012-02-08 17:13:40 -08002579 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
Andrea Arcangeliba761492011-01-13 15:46:58 -08002580
2581 if (khugepaged_scan.mm_slot)
2582 mm_slot = khugepaged_scan.mm_slot;
2583 else {
2584 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2585 struct mm_slot, mm_node);
2586 khugepaged_scan.address = 0;
2587 khugepaged_scan.mm_slot = mm_slot;
2588 }
2589 spin_unlock(&khugepaged_mm_lock);
2590
2591 mm = mm_slot->mm;
2592 down_read(&mm->mmap_sem);
2593 if (unlikely(khugepaged_test_exit(mm)))
2594 vma = NULL;
2595 else
2596 vma = find_vma(mm, khugepaged_scan.address);
2597
2598 progress++;
2599 for (; vma; vma = vma->vm_next) {
2600 unsigned long hstart, hend;
2601
2602 cond_resched();
2603 if (unlikely(khugepaged_test_exit(mm))) {
2604 progress++;
2605 break;
2606 }
Bob Liufa475e52012-12-11 16:00:39 -08002607 if (!hugepage_vma_check(vma)) {
2608skip:
Andrea Arcangeliba761492011-01-13 15:46:58 -08002609 progress++;
2610 continue;
2611 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002612 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2613 hend = vma->vm_end & HPAGE_PMD_MASK;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002614 if (hstart >= hend)
2615 goto skip;
2616 if (khugepaged_scan.address > hend)
2617 goto skip;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002618 if (khugepaged_scan.address < hstart)
2619 khugepaged_scan.address = hstart;
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002620 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002621
2622 while (khugepaged_scan.address < hend) {
2623 int ret;
2624 cond_resched();
2625 if (unlikely(khugepaged_test_exit(mm)))
2626 goto breakouterloop;
2627
2628 VM_BUG_ON(khugepaged_scan.address < hstart ||
2629 khugepaged_scan.address + HPAGE_PMD_SIZE >
2630 hend);
2631 ret = khugepaged_scan_pmd(mm, vma,
2632 khugepaged_scan.address,
2633 hpage);
2634 /* move to next address */
2635 khugepaged_scan.address += HPAGE_PMD_SIZE;
2636 progress += HPAGE_PMD_NR;
2637 if (ret)
2638 /* we released mmap_sem so break loop */
2639 goto breakouterloop_mmap_sem;
2640 if (progress >= pages)
2641 goto breakouterloop;
2642 }
2643 }
2644breakouterloop:
2645 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2646breakouterloop_mmap_sem:
2647
2648 spin_lock(&khugepaged_mm_lock);
Andrea Arcangelia7d6e4e2011-02-15 19:02:45 +01002649 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002650 /*
2651 * Release the current mm_slot if this mm is about to die, or
2652 * if we scanned all vmas of this mm.
2653 */
2654 if (khugepaged_test_exit(mm) || !vma) {
2655 /*
2656 * Make sure that if mm_users is reaching zero while
2657 * khugepaged runs here, khugepaged_exit will find
2658 * mm_slot not pointing to the exiting mm.
2659 */
2660 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2661 khugepaged_scan.mm_slot = list_entry(
2662 mm_slot->mm_node.next,
2663 struct mm_slot, mm_node);
2664 khugepaged_scan.address = 0;
2665 } else {
2666 khugepaged_scan.mm_slot = NULL;
2667 khugepaged_full_scans++;
2668 }
2669
2670 collect_mm_slot(mm_slot);
2671 }
2672
2673 return progress;
2674}
2675
2676static int khugepaged_has_work(void)
2677{
2678 return !list_empty(&khugepaged_scan.mm_head) &&
2679 khugepaged_enabled();
2680}
2681
2682static int khugepaged_wait_event(void)
2683{
2684 return !list_empty(&khugepaged_scan.mm_head) ||
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002685 kthread_should_stop();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002686}
2687
Xiao Guangrongd5169042012-10-08 16:29:48 -07002688static void khugepaged_do_scan(void)
2689{
2690 struct page *hpage = NULL;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002691 unsigned int progress = 0, pass_through_head = 0;
2692 unsigned int pages = khugepaged_pages_to_scan;
Xiao Guangrongd5169042012-10-08 16:29:48 -07002693 bool wait = true;
Andrea Arcangeliba761492011-01-13 15:46:58 -08002694
2695 barrier(); /* write khugepaged_pages_to_scan to local stack */
2696
2697 while (progress < pages) {
Xiao Guangrong26234f32012-10-08 16:29:51 -07002698 if (!khugepaged_prealloc_page(&hpage, &wait))
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002699 break;
Xiao Guangrong26234f32012-10-08 16:29:51 -07002700
Xiao Guangrong420256ef2012-10-08 16:29:49 -07002701 cond_resched();
Andrea Arcangeliba761492011-01-13 15:46:58 -08002702
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002703 if (unlikely(kthread_should_stop() || freezing(current)))
2704 break;
2705
Andrea Arcangeliba761492011-01-13 15:46:58 -08002706 spin_lock(&khugepaged_mm_lock);
2707 if (!khugepaged_scan.mm_slot)
2708 pass_through_head++;
2709 if (khugepaged_has_work() &&
2710 pass_through_head < 2)
2711 progress += khugepaged_scan_mm_slot(pages - progress,
Xiao Guangrongd5169042012-10-08 16:29:48 -07002712 &hpage);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002713 else
2714 progress = pages;
2715 spin_unlock(&khugepaged_mm_lock);
2716 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002717
Xiao Guangrongd5169042012-10-08 16:29:48 -07002718 if (!IS_ERR_OR_NULL(hpage))
2719 put_page(hpage);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002720}
2721
Xiao Guangrong2017c0b2012-10-08 16:29:44 -07002722static void khugepaged_wait_work(void)
2723{
2724 try_to_freeze();
2725
2726 if (khugepaged_has_work()) {
2727 if (!khugepaged_scan_sleep_millisecs)
2728 return;
2729
2730 wait_event_freezable_timeout(khugepaged_wait,
2731 kthread_should_stop(),
2732 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2733 return;
2734 }
2735
2736 if (khugepaged_enabled())
2737 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2738}
2739
Andrea Arcangeliba761492011-01-13 15:46:58 -08002740static int khugepaged(void *none)
2741{
2742 struct mm_slot *mm_slot;
2743
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002744 set_freezable();
Dongsheng Yang8698a742014-03-11 18:09:12 +08002745 set_user_nice(current, MAX_NICE);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002746
Xiao Guangrongb7231782012-10-08 16:29:54 -07002747 while (!kthread_should_stop()) {
2748 khugepaged_do_scan();
2749 khugepaged_wait_work();
2750 }
Andrea Arcangeliba761492011-01-13 15:46:58 -08002751
2752 spin_lock(&khugepaged_mm_lock);
2753 mm_slot = khugepaged_scan.mm_slot;
2754 khugepaged_scan.mm_slot = NULL;
2755 if (mm_slot)
2756 collect_mm_slot(mm_slot);
2757 spin_unlock(&khugepaged_mm_lock);
Andrea Arcangeliba761492011-01-13 15:46:58 -08002758 return 0;
2759}
2760
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002761static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2762 unsigned long haddr, pmd_t *pmd)
2763{
2764 struct mm_struct *mm = vma->vm_mm;
2765 pgtable_t pgtable;
2766 pmd_t _pmd;
2767 int i;
2768
2769 pmdp_clear_flush(vma, haddr, pmd);
2770 /* leave pmd empty until pte is filled */
2771
Aneesh Kumar K.V6b0b50b2013-06-05 17:14:02 -07002772 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002773 pmd_populate(mm, &_pmd, pgtable);
2774
2775 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2776 pte_t *pte, entry;
2777 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2778 entry = pte_mkspecial(entry);
2779 pte = pte_offset_map(&_pmd, haddr);
2780 VM_BUG_ON(!pte_none(*pte));
2781 set_pte_at(mm, haddr, pte, entry);
2782 pte_unmap(pte);
2783 }
2784 smp_wmb(); /* make pte visible before pmd */
2785 pmd_populate(mm, pmd, pgtable);
Kirill A. Shutemov97ae1742012-12-12 13:51:06 -08002786 put_huge_zero_page();
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002787}
2788
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002789void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2790 pmd_t *pmd)
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002791{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002792 spinlock_t *ptl;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002793 struct page *page;
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002794 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002795 unsigned long haddr = address & HPAGE_PMD_MASK;
2796 unsigned long mmun_start; /* For mmu_notifiers */
2797 unsigned long mmun_end; /* For mmu_notifiers */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002798
2799 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002800
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002801 mmun_start = haddr;
2802 mmun_end = haddr + HPAGE_PMD_SIZE;
Hugh Dickins750e8162013-10-16 13:47:08 -07002803again:
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002804 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002805 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002806 if (unlikely(!pmd_trans_huge(*pmd))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002807 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002808 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2809 return;
2810 }
2811 if (is_huge_zero_pmd(*pmd)) {
2812 __split_huge_zero_page_pmd(vma, haddr, pmd);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002813 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002814 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002815 return;
2816 }
2817 page = pmd_page(*pmd);
Sasha Levin309381fea2014-01-23 15:52:54 -08002818 VM_BUG_ON_PAGE(!page_count(page), page);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002819 get_page(page);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08002820 spin_unlock(ptl);
Kirill A. Shutemovc5a647d2012-12-12 13:51:00 -08002821 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002822
2823 split_huge_page(page);
2824
2825 put_page(page);
Hugh Dickins750e8162013-10-16 13:47:08 -07002826
2827 /*
2828 * We don't always have down_write of mmap_sem here: a racing
2829 * do_huge_pmd_wp_page() might have copied-on-write to another
2830 * huge page before our split_huge_page() got the anon_vma lock.
2831 */
2832 if (unlikely(pmd_trans_huge(*pmd)))
2833 goto again;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08002834}
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002835
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002836void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2837 pmd_t *pmd)
2838{
2839 struct vm_area_struct *vma;
2840
2841 vma = find_vma(mm, address);
2842 BUG_ON(vma == NULL);
2843 split_huge_page_pmd(vma, address, pmd);
2844}
2845
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002846static void split_huge_page_address(struct mm_struct *mm,
2847 unsigned long address)
2848{
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002849 pmd_t *pmd;
2850
2851 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2852
Bob Liu62190492012-12-11 16:00:37 -08002853 pmd = mm_find_pmd(mm, address);
2854 if (!pmd)
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002855 return;
2856 /*
2857 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2858 * materialize from under us.
2859 */
Kirill A. Shutemove1803772012-12-12 13:50:59 -08002860 split_huge_page_pmd_mm(mm, address, pmd);
Andrea Arcangeli94fcc582011-01-13 15:47:08 -08002861}
2862
2863void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2864 unsigned long start,
2865 unsigned long end,
2866 long adjust_next)
2867{
2868 /*
2869 * If the new start address isn't hpage aligned and it could
2870 * previously contain an hugepage: check if we need to split
2871 * an huge pmd.
2872 */
2873 if (start & ~HPAGE_PMD_MASK &&
2874 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2875 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2876 split_huge_page_address(vma->vm_mm, start);
2877
2878 /*
2879 * If the new end address isn't hpage aligned and it could
2880 * previously contain an hugepage: check if we need to split
2881 * an huge pmd.
2882 */
2883 if (end & ~HPAGE_PMD_MASK &&
2884 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2885 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2886 split_huge_page_address(vma->vm_mm, end);
2887
2888 /*
2889 * If we're also updating the vma->vm_next->vm_start, if the new
2890 * vm_next->vm_start isn't page aligned and it could previously
2891 * contain an hugepage: check if we need to split an huge pmd.
2892 */
2893 if (adjust_next > 0) {
2894 struct vm_area_struct *next = vma->vm_next;
2895 unsigned long nstart = next->vm_start;
2896 nstart += adjust_next << PAGE_SHIFT;
2897 if (nstart & ~HPAGE_PMD_MASK &&
2898 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2899 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2900 split_huge_page_address(next->vm_mm, nstart);
2901 }
2902}